Waves in Neural Media

Lecture Notes on Mathematical Modelling
in the Life Sciences
Paul C. Bressloff
Waves in
Neural Media
From Single Neurons to Neural Fields
Lecture Notes on Mathematical Modelling

in the Life Sciences
Series Editors
Angela Stevens
Michael C. Mackey
For further volumes:

http://www.springer.com/series/10049
Paul C. Bressloff
Waves in Neural Media

From Single Neurons to Neural Fields
123
Paul C. Bressloff
University of Utah
Department of Mathematics
Salt Lake City, UT, USA
ISSN 2193-4789
ISSN 2193-4797 (electronic)
ISBN 978-1-4614-8865-1
ISBN 978-1-4614-8866-8 (eBook)
DOI 10.1007/978-1-4614-8866-8
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2013948234
Mathematics Subject Classification (2010): 35C07, 82C31, 92C20, 92C30, 92C37, 82C31
Springer Science+Business Media New York 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publishers location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any
errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect
to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
To the Shmu
Preface
This is a book on mathematical neuroscience, in which the unifying theme is

wavelike phenomena at multiple spatial and temporal scales. There are already several excellent books on applications of dynamical systems theory and bifurcation
theory to neuroscience, but these are mainly concerned with mathematical models of neurons and networks that take the form of systems of ordinary differential
equations (ODEs). Spatial aspects are usually restricted to standard partial differential equation (PDE) models of action potential propagation along axons and the
spatial spread of voltage along dendrites (the cable equation). However, there are
many other interesting spatiotemporal phenomena that are not usually addressed in
any detail, at least within a single book. These include electrical and chemical waves
along spiny dendrites, calcium waves in astrocytes, excitable and oscillatory waves
in cortical and subcortical structures, binocular rivalry waves and ambiguous perception, oscillatory waves and swimming locomotion, cytoskeletal waves and neuritogenesis, spreading depression and migraine auras, the spread of prions and other
neurodegenerative diseases within the brain, and epileptic waves in complex neural
networks. All of these phenomena are linked by the notion of the time-dependent
propagation or spread of some quantity in space, whether it is chemical concentration, a protein aggregate, voltage, or spiking activity of a population of neurons.
They should be distinguished from another important class of spatial phenomena,
namely, pattern-forming instabilities, which could be the subject of a future book!
In order to develop the themes of this book, I have endeavored to give a pedagogical introduction to the mathematical modeling of biological neurons and networks,
as well as a toolbox of analytical techniques used to study wavelike phenomena,
including the construction of traveling wave solutions of PDES and nonlocal equations, stability analysis, phase-plane analysis, singular perturbation theory, phasereduction and averaging methods, and homogenization theory. A common thread
throughout the book is an emphasis on stochasticity. In recent years, there has been
an explosion of interest in the role of noise in biological systems, partly driven
by the presence of low copy numbers in gene and biochemical networks. Noise is
also ubiquitous in the brain, ranging from fluctuations in gene expression and the
opening of ion-channel proteins to the Poisson-like spiking activity of individual
vii
viii
Preface
cortical neurons, the presence of fluctuations during perceptual decision making and
bistable perception, and the generation of spontaneous activity during resting brain
states. Hence, the book contains an introduction to the theory of stochastic processes
and how it can be used to study the effects of noise on wave propagation and wave
initiation. In the latter case, one typically has to solve an escape problem, in which
noise drives some quantity such as voltage, calcium concentration, or population
spiking activity above a threshold. This results, for example, in the generation of a
spontaneous action potential, a calcium spark, or a transition between cortical up
and down states, respectively.
Before giving a more detailed description of the contents of this book, I thought
that I would elaborate a little on my own approach to neuroscience. This is not
to favor one approach over anothera diversity of methodological and conceptual frameworks is crucial in trying to understand such a complex system as the
brainbut rather to highlight my own tastes and biases. Let me begin with a story
(perhaps apocryphal). Several years ago a well-known mathematical biology journal was relaunched with a new editorial board. At the final organizational meeting of the editors in chief, after some good food and wine, a satisfied glow spread
around the room as the list of associate editors was drawn up, covering the major
areas of biology. However, someone then sheepishly pointed out that one area was
missingneuroscience! I suspect that neuroscience has a slightly complicated relationship with the rest of theoretical biology due to the ambiguous notion of computation. One sense of the word, which applies to all areas of biology, refers to
the high level of biological detail and computational complexity in a model, which
means that it is usually less amenable to mathematical analysis than simpler reduced models. The focus tends to be on issues such as parameter searches and data
fitting, sensitivity analysis, numerical convergence, and computational efficiency. It
follows that one way to characterize a modeling approach is where it lies on the
mathematicalcomputational biology axis. Within this context, the present book is
on mathematical rather than computational neuroscience, since I consider reduced,
mathematically tractable models of wavelike phenomena.
The other sense of the word computation refers to a functional interpretation
of the brain as a biological computer that processes sensory inputs via a sequence of
internal brain states to produce a motor output or a change in cognitive state. A common assumption is that if we had a complete understanding of the physiology of the
heart at the molecular, cellular, and systems levels, then we would understand the
function of the heart, whereas the same is not true of the brain. That is, what makes
neuroscience so challenging, beyond the incredible complexity of the brain as a biological organ, is how one links physiological brain states to behavior and cognition.
Consequently, a major drive in computational neuroscience (in the second sense) is
to link physiological descriptions of neurons and networks with computation and
information processing. Popular themes are spike trains and neural coding, statistical correlations, the role of feedback in predictive coding, learning algorithms,
Bayesian statistics, and maximum entropy. The current book is not concerned with
this form of computational neuroscience either, since I do not really discuss the
possible functional role of wavelike neural phenomena, although the role of waves
Preface
ix
in development and disease is discussed at length. There have been some attempts
to interpret the computational role of waves, based on the observation that cortical
waves often occur between sensory processing events. However, such an observation is correlative at best and provides very little insight into the causal mechanisms
of perception. Unfortunately, the same could be said of many of the overhyped results coming from functional magnetic resonance imaging (fMRI), where parts of
the brain light up when someone is telling a joke or planning to commit a crime.
In the media lately, a number of distinguished physicists have had some harsh
words to say about philosophy and its relevance to scientific questions. I personally
disagree with such attitudes, partly due to my own interests in philosophy and also
because I have found it useful to stand back occasionally in order to place my own
work in a broader context. I think that this is particularly important in a field such
as neuroscience, with all of its potential clinical, psychological, ethical, sociological, and cultural ramifications. Of course, reading philosophy does not make me a
philosopher, so I proceed with caution. Nevertheless, I am a naturalist, in the sense
that I do not believe the brain is made of different stuff or obeys different physical
laws from planets, sushi, and hot tubs. On the other hand, I tend not to view the
brain in functionalist terms, that is, as the hardware to the minds software. I favor a
more direct biological interpretation in terms of an evolved and adaptive system that
enables an organism to carry out goal-directed behavior in a changing physical and
social environment (although many people believe that this can be reformulated in
terms of algorithms and computation). A related idea is that meaning or semantics
cannot be generated from a purely syntactic process, since one needs to take into account the evolutionary and biographical history of an embodied mind acting within
a societal and cultural context. Another reason I tend to steer clear of computational
approaches is that I fear that using such a paradigm is in danger of imposing order, rationality, and engineering design principles on a rather messy and haphazard
system. I tend to agree with the idea that the brain is a kludge, in which modules
that evolved to solve simpler problems have been thrown together suboptimally as
the complexity of an organism has increased. I also feel that randomness plays a
significant role in how we think, which counters the idea of following a computer
program or algorithm. I do not mean randomness in the sense of exploiting stochasticity to enhance performance (e.g., simulated annealing or random search), which
itself could be algorithmic, but randomness as irrationality. People commit suicide,
skateboard down Everest on a piano, and drink root beer (an alien concept to an
Englishman).
The elephant in the room is the so-called hard problem of consciousness, that
is, of explaining how and why we have qualia or phenomenal experiences. (This
is a controversial issue, since some view qualia as an illusion based on misguided
folk psychology or poor intuition pumps.) A related question is how one goes from
objective brain states to subjective mental states. I would describe myself as a nonreductionist, in the sense that I believe mental properties supervene on (but are distinct from) physical properties and are an emergent feature of the physical brain.
However, a systematic framework for understanding this emergence is currently
beyond our grasp. An analogy is how one goes from microscopic descriptions of
Preface
molecules to macroscopic descriptions of states of matter. The latter also supervene

on the former, but in this case we do have a framework for crossing the divide,
namely, statistical mechanics. Hence, I do not believe we will discover a statistical
mechanics of the brain, at least not during the funding periods of the European
Human Brain Project and the USAs BRAIN Initiative! An interesting alternative to emergentism is neutral monism, which posits that protean versions of mental
properties inhere in all physical objectsa form of panpsychism.
Having elucidated what this book is not about and why, let me now turn to its
contents. The book is divided into three parts: Neurons (Chaps. 14), Networks
(Chaps. 58), and Development and disease (Chap. 9). Chapter 1 provides a detailed
introduction to the working parts of a neuron, including conductance-based models
of action potential generation, synaptic and dendritic processing, and ion channels.
Two important mathematical topics are also introduced. First, the dynamics of a
periodically forced neural oscillator is used to introduce phase-reduction and averaging methods, phase-resetting curves, and synchronization. These are later applied
to the study of waves in oscillatory neural media (Chap. 5). Second, a detailed account of stochastic ion channels and membrane voltage fluctuations is given, which
also provides background material on stochastic processes. A major theme is how to
model and analyze stochastic hybrid systems, in which a continuous variable (e.g.,
voltage) couples to a discrete jump Markov process (e.g., number of open ion channels). Spontaneous action potential generation is formulated as a first passage time
problem, which is solved using perturbation methods such as WKB and matched
asymptotics. These methods are later used to analyze related problems such as the
generation of calcium sparks (Chap. 4) and bistability in populations of spiking neurons (Chap. 6). Chapter 2 covers the classical problem of waves in one-dimensional
excitable media, as exemplified by the FitzHughNagumo model of action potential
propagation along an axon. Standard methods for analyzing front and pulse solutions of PDEs are described, including phase-plane analysis, singular perturbation
methods and slowfast analysis, and Evans functions for wave stability. In addition,
the problem of wave propagation failure in myelinated axons is considered, where
an averaging method is used to determine the effects of spatial discreteness on wave
speed. This method is later used to analyze wave propagation failure in inhomogeneous neural fields (Chap. 7). Finally, stochastic traveling waves are considered,
where formal perturbation methods are used to show how to separate out fast fluctuations of the wave profile from the slow diffusive-like wandering of the wave.
Chapter 3 presents two different models of traveling waves along spiny dendrites: a spikediffusespike model of propagating voltage spikes mediated by active dendritic spines and a reactiondiffusion model of Ca2+ /calmodulin-dependent
protein kinase II (CaMKII) translocation waves. The former model introduces methods that are also used to analyze solitary waves propagating in spiking neural networks (Chap. 5). The latter model turns out to be identical in form to the diffusive
susceptible-infected (SI) model of the spread of epidemics, which is a generalization
of the scalar FisherKPP equation of population genetics. One characteristic feature
of such equations is that they support traveling fronts propagating into an unstable
steady state, in which the wave speed and longtime asymptotics are determined by
Preface
xi
the dynamics in the leading edge of the wavethe so-called pulled fronts. In particular, a sufficiently localized initial perturbation will asymptotically approach the
traveling front solution that has the minimum possible wave speed. Hence, pulled
fronts have very different properties from those considered in Chap. 2. Homogenization methods are also presented, which allow one to approximate the discrete
distribution of spines by a smooth distribution. Chapter 4 gives a comprehensive
review of calcium wave modeling, with an emphasis on their role in neuronal calcium signaling. Two models of intracellular waves are considered in some detail:
a reactiondiffusion model of calcium dynamics and the firediffusefire model of
calcium release. The latter is formally very similar to the spikediffusespike model
of spiny dendrites and is analyzed accordingly. Stochastic models of spontaneous
calcium release (calcium puffs and sparks) are then analyzed using the stochastic methods introduced in Chap. 1. Finally, several models of intercellular calcium
waves in astrocytes are presented. Traditionally, astrocytes were thought to be physiologically passive cells that only play a supporting role in the central nervous system by regulating and optimizing the environment within which neurons operate.
However, there is an increasing amount of empirical data indicating that astrocytes
play a more active role in modulating synaptic transmission and neuronal signal
processing.
In the second part of the book, we consider a variety of neural network models
that have been used to investigate the mechanisms and properties of wave propagation in cortex and other parts of the nervous system. There has been a rapid
increase in the number of computational studies of network dynamics, which are
based on biophysically detailed conductance-based models of synaptically (and possibly electrically) coupled neurons. These models provide considerable insights into
the role of ionic currents, synaptic processing, and network structure on spatiotemporal dynamics, but they tend to be analytically intractable. This has motivated an
alternative approach to network dynamics, involving simplified neuron models that
hopefully capture important aspects of wave phenomena, while allowing a more
concise mathematical treatment. In the case of oscillatory networks, such a simplification can be achieved by reducing a conductance-based neuron model to a phase
model along the lines of Chap. 1. Alternatively, one can use a simplified spiking
neuron model such as integrate-and-fire in order to investigate waves in excitable
and oscillatory neural media. Both of these approaches are considered in Chap. 5,
which also provides a summary of various wave phenomena in cortical and subcortical structures. Chapter 6 describes the construction of population-based rate models
under the assumption that the spiking of individual neurons is unimportant. The issue of how stochasticity at the single-cell level manifests itself at the population
level is discussed, introducing topics such as balanced networks, Poisson statistics,
and asynchronous states. Stochastic methods are then used to analyze bistability in
a stochastic population model. Finally, the transition from spatially structured neural networks to continuum neural fields is highlighted. The latter take the form of
nonlocal integrodifferential equations, in which the integral kernel represents the
distribution of synaptic connections.
xii
Preface
Chapter 7 develops the theory of waves in excitable neural fields, where the fundamental network element is a local population of cells rather than a single neuron.
It is shown how many of the PDE methods and results from the analysis of waves
in reactiondiffusion equations considered in Part I can be extended to the nonlocal
equations of neural field theory. First, the existence and stability of solitary traveling
fronts and pulses in one-dimensional excitatory neural fields are considered. In the
case of traveling pulses, it is necessary to include some form of local negative feedback mechanism such as synaptic depression or spike frequency adaptation. Two
approaches to analyzing wave propagation failure in inhomogeneous neural media
are then presented: one based on averaging methods and the other on interfacial dynamics. Finally, wave propagation in stochastic neural fields is analyzed and oscillatory waves in two-dimensional neural media are briefly discussed. In Chap. 8, neural
field theory is used to model binocular rivalry waves. During binocular rivalry, visual perception switches back and forth between different images presented to the
two eyes. The resulting fluctuations in perceptual dominance and suppression provide a basis for noninvasive studies of the human visual system and the identification
of possible neural mechanisms underlying conscious visual awareness. Various psychophysical experiments have demonstrated that the switch between a dominant and
suppressed visual percept propagates as a traveling front for each eye. In addition
to considering the particular problem of binocular rivalry waves, the more general
issue of how to develop neural field models of the functional architecture of primary
visual cortex (V1) is discussed.
In the final part of the book, Chap. 9, a variety of topics regarding wavelike phenomena in the developing and diseased brain are presented. First, the possible role of
calcium and retinal waves in early development is summarized. There is then a detailed description and analysis of cytoskeletal waves involved in neurite growth and
cell polarization. This introduces another interesting phenomenon, namely, wave
pinning. Three distinct examples of waves in the diseased brain are considered:
spreading depression and migraine auras, epileptic waves, and the spread of neurodegenerative waves due to protein aggregation. In order to tackle the latter two
phenomena, a review of complex network theory is included, covering topics such
as small-world networks, scale-free networks, neuronal avalanches, branching processes, and epidemics on networks.
I end this introduction by noting that I sometimes have to deal with comments
such as neuroscience is pre-Newtonian or systems neuroscience lacks a Navier
Stokes equation. My reaction to such comments is that it is for these very reasons
that neuroscience is so exciting. One does not have to be as smart as Isaac Newton to
work at the frontiers of a twenty-first-century science and have the chance to make a
significant contribution. Of course, it also means that the contents of this book could
one day be irrelevant and pass into dust the way of my flesh. If this does happen dear
reader, please do not ask for your money back!
Salt Lake City, UT, USA
Paul C. Bressloff
Acknowledgements
I would first like to thank my past graduate students: Pete Roper, Stefan Folias, Andy
Oster, Berton Earnshaw, Will Nesse, Zack Kilpatrick, Jay Newby, Yi-Ming Lai, and
Matt Webber. I have really enjoyed mentoring all of my students and collaborating
on so many fun projects. There are a number of people who have inspired my own
work in spatiotemporal aspects of mathematical neuroscience, either independently
or in collaboration: Shun-Ichi Amari, Carson Chow, Jack Cowan, Steve Coombes,
Bard Ermentrout, Olivier Faugeras, Marty Golubitsky, David Holcman, Jim Keener,
Carlo Laing, John Rinzel, Jon Rubin, Dave Terman, Bill Troy, and Hugh Wilson. I
would like to give a particular thanks to Jack Cowan, for his encouragement and
guidance, and Steve Coombes for keeping the fire burning in the UK. I thank my
parents Andy and Jackie for their unstinting love and support, Neil and his family
for putting up with the crazy brother, and, last but not least, my wife Alessandra,
who gave me the greatest gift of all, our son Luca. Ti amo e grazie mille per tutto.
xiii
Contents
Part I Neurons
1
Single Neuron Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1 Conductance-Based Model of a Neuron . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Periodically Forced Neural Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Isochrones and Phase-Resetting Curves . . . . . . . . . . . . . . . . . .
1.2.2 Phase-Locking and Synchronization . . . . . . . . . . . . . . . . . . . .
1.3 Synaptic Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1 Excitatory and Inhibitory Synapses . . . . . . . . . . . . . . . . . . . . .
1.3.2 Synaptic Depression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.3 Kinetic Model of a Synapse . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Dendritic Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 The Cable Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2 Dendritic Filtering of Synaptic Inputs . . . . . . . . . . . . . . . . . . .
1.4.3 Active Dendrites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Stochastic Ion Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.1 Ensemble of Two-State Ion Channels . . . . . . . . . . . . . . . . . . .
1.5.2 Diffusion Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.3 Membrane Voltage Fluctuations . . . . . . . . . . . . . . . . . . . . . . . .
1.5.4 First Passage Time Problem . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Stochastic Hybrid System with Fast Kinetics . . . . . . . . . . . . . . . . . . . .
1.6.1 First Passage Time Problem and the Projection Method . . . .
1.6.2 The WKB Method and the Quasistationary Density . . . . . . . .
1.6.3 Calculation of the Principal Eigenvalue . . . . . . . . . . . . . . . . . .
1.7 Appendix: Stochastic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7.1 Ito Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7.2 Itos Formula and the FokkerPlanck Equation . . . . . . . . . . .
1.7.3 Multiplicative Noise and Stratonovich Integrals . . . . . . . . . . .
3
3
10
10
14
17
18
20
21
24
24
29
32
36
36
38
41
43
46
48
51
53
57
57
60
61
xv
xvi
Contents
Traveling Waves in One-Dimensional Excitable Media . . . . . . . . . . . . .

2.1 Excitable Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Traveling Fronts in a Scalar Bistable Equation . . . . . . . . . . . . . . . . . .
2.3 Traveling Pulses in the FitzHughNagumo Equations . . . . . . . . . . . .
2.3.1 Periodic Wave Trains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Wave Stability and Evans Functions . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.1 Stability of Fronts in the Bistable Equation . . . . . . . . . . . . . . .
2.4.2 The Evans Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Myelinated Axons and Discrete Diffusion . . . . . . . . . . . . . . . . . . . . . .
2.5.1 The Discrete Bistable Equation . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.2 Propagation Failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 Stochastic Traveling Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7 Appendix: Linear Differential Operators . . . . . . . . . . . . . . . . . . . . . . .
2.7.1 Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.2 Fredholm Alternative Theorem . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.3 Spectrum of a Linear Differential Operator . . . . . . . . . . . . . . .
63
64
69
72
74
76
77
78
80
83
84
86
91
95
95
97
97
Wave Propagation Along Spiny Dendrites . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.1 Solitary Waves in a SpikeDiffuseSpike Model of Active
Spines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.1.1 Existence of a Solitary Wave for a Uniform Density . . . . . . . 104
3.1.2 Linear Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.1.3 Discrete Distribution of Spines . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.2 ReactionDiffusion Model of CaMKII Translocation Waves . . . . . . . 113
3.2.1 Translocation Waves for a Uniform Distribution
of Spines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.2.2 Pulsating Waves in the Presence of Discretely
Distributed Spines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.2.3 Homogenization of Pulsating Waves for a Fast Periodic
Modulation of Spine Density . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.2.4 Wave Speed for a Slowly Modulated Spine Density . . . . . . . . 126
3.3 Appendix: Pulled and Pushed Fronts . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.3.1 The Linear Spreading Velocity . . . . . . . . . . . . . . . . . . . . . . . . . 131
3.3.2 Variational Method for Front Velocity Selection . . . . . . . . . . . 135
Calcium Waves and Sparks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

4.1 Calcium Signaling in Neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.2 ReactionDiffusion Models of Ca2+ Dynamics . . . . . . . . . . . . . . . . . . 141
4.3 Discrete Calcium Sources and the FireDiffuseFire Model . . . . . . . 146
4.3.1 Discrete Calcium Sources and Averaging . . . . . . . . . . . . . . . . 146
4.3.2 The FireDiffuseFire Model of Ca2+ Release . . . . . . . . . . . . 150
4.4 Stochastic Models of Ca2+ Release . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4.4.1 Stochastic Model of Ca2+ Puffs in a Cluster of IP3 Rs . . . . . . 153
4.4.2 Stochastic Model of Ca2+ Sparks in a Cluster of RyRs . . . . . 155
Contents
xvii
4.4.3 WKB Method and Quasistationary Approximation . . . . . . . . 160

4.4.4 Stochastic Phase-Plane Analysis . . . . . . . . . . . . . . . . . . . . . . . . 165
4.4.5 Whole-Cell Model of Ca2+ Sparks . . . . . . . . . . . . . . . . . . . . . . 167
4.5 Intercellular Calcium Waves in Astrocytes . . . . . . . . . . . . . . . . . . . . . . 168
4.5.1 Astrocytes and the Tripartite Synapse . . . . . . . . . . . . . . . . . . . 169
4.5.2 Mechanisms and Models of Intercellular Calcium Waves . . . 173
4.6 Appendix: Homogenization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Part II Networks
5
Waves in Synaptically Coupled Spiking Networks . . . . . . . . . . . . . . . . . . 185

5.1 Waves in Cortical and Subcortical Systems . . . . . . . . . . . . . . . . . . . . . 186
5.2 Phase Reduction of Synaptically Coupled Neural Oscillators . . . . . . 195
5.2.1 Phase-Locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
5.2.2 Traveling Phase Waves on a Chain . . . . . . . . . . . . . . . . . . . . . . 201
5.2.3 Ring of Phase Oscillators and the Sync Basin . . . . . . . . . . . . . 204
5.2.4 Two-Dimensional Network of Phase Oscillators . . . . . . . . . . 208
5.3 Synaptically Coupled Integrate-and-Fire (IF) Oscillators . . . . . . . . . . 212
5.4 Spiking Network Models of Cortical and Subcortical Waves . . . . . . . 219
5.4.1 Traveling Pulses in an Excitatory Spiking Network . . . . . . . . 219
5.4.2 Traveling Pulses in Networks of Excitatory and Inhibitory
Spiking Neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
5.4.3 Spiking Model of Thalamic Spindle Waves . . . . . . . . . . . . . . . 228
Population Models and Neural Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

6.1 Population Averaging and Rate Models . . . . . . . . . . . . . . . . . . . . . . . . 233
6.2 EI Oscillator Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
6.3 Firing Rates, Asynchronous States, and Spiking Statistics . . . . . . . . . 242
6.3.1 The Asynchronous State in a Homogeneous Spiking
Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
6.3.2 Balanced Networks and Neural Variability . . . . . . . . . . . . . . . 246
6.3.3 Spike Statistics and the Poisson Process . . . . . . . . . . . . . . . . . 247
6.4 Stochastic Population Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
6.4.1 Neural Master Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
6.4.2 Metastability in a One-Population Model . . . . . . . . . . . . . . . . 256
6.4.3 Metastability in a Two-Population Model . . . . . . . . . . . . . . . . 260
6.5 Spatially Structured Networks and Neural Fields . . . . . . . . . . . . . . . . 264
Waves in Excitable Neural Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

7.1 Traveling Fronts in a Scalar Neural Field . . . . . . . . . . . . . . . . . . . . . . . 272
7.1.1 Propagating Fronts in a Bistable Neural Field . . . . . . . . . . . . . 272
7.1.2 Wave Stability and Evans Functions . . . . . . . . . . . . . . . . . . . . . 274
7.1.3 Pulled Fronts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
7.2 Traveling Pulses in Adaptive Neural Fields . . . . . . . . . . . . . . . . . . . . . 282
7.2.1 Exact Traveling Pulse Solution . . . . . . . . . . . . . . . . . . . . . . . . . 283
7.2.2 Singularly Perturbed Pulse Solution . . . . . . . . . . . . . . . . . . . . . 287
xviii
Contents
7.3 Wave Propagation in Heterogeneous Neural Fields . . . . . . . . . . . . . . . 289

7.3.1 Averaging Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
7.3.2 Interfacial Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
7.3.3 HamiltonJacobi Dynamics and Slow Spatial
Heterogeneities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
7.4 Wave Propagation in Stochastic Neural Fields . . . . . . . . . . . . . . . . . . . 303
7.4.1 Spontaneous Front Propagation . . . . . . . . . . . . . . . . . . . . . . . . 304
7.4.2 Stimulus-Locked Fronts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
7.4.3 Stochastic Pulled Fronts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
7.5 Traveling Waves in 2D Oscillatory Neural Fields . . . . . . . . . . . . . . . . 315
8
Neural Field Model of Binocular Rivalry Waves . . . . . . . . . . . . . . . . . . . 319

8.1 Binocular Rivalry Waves and the Structure of Primary
Visual Cortex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
8.1.1 Functional Architecture of V1 . . . . . . . . . . . . . . . . . . . . . . . . . 319
8.1.2 Binocular Rivalry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
8.2 Competitive Network Models of Binocular Rivalry . . . . . . . . . . . . . . 325
8.3 Neural Field Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
8.3.1 Rivalry Waves in a Deterministic Neural Field Model . . . . . . 331
8.3.2 Effects of Extrinsic Noise in the Fast Activity Variables . . . . 334
8.3.3 Explicit Results for a Heaviside Rate Function . . . . . . . . . . . . 336
8.4 Modeling Primary Visual Cortex Using 2D Neural Fields . . . . . . . . . 341
8.4.1 2D Planar Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
8.4.2 Coupled Hypercolumn Model . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Part III Development and Disease

9
Waves in the Developing and the Diseased Brain . . . . . . . . . . . . . . . . . . . 349

9.1 Intercellular Waves in Neural Development . . . . . . . . . . . . . . . . . . . . . 350
9.1.1 Calcium Waves in Early Development . . . . . . . . . . . . . . . . . . . 350
9.1.2 Retinal Waves in Later Development . . . . . . . . . . . . . . . . . . . . 353
9.2 Traveling Waves, the Cytoskeleton, and Cell Polarization . . . . . . . . . 356
9.2.1 Neuronal Polarization and the Cytoskeleton . . . . . . . . . . . . . . 356
9.2.2 Wave-Pinning and Cell Polarity in a Bistable
ReactionDiffusion Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
9.2.3 Actin Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.2.4 Diffusion-Based Model of Neurite Symmetry Breaking . . . . 366
9.3 Spreading Depression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
9.3.1 TuckwellMiura Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
9.3.2 Brain Tissue as a Porous Medium . . . . . . . . . . . . . . . . . . . . . . 375
9.3.3 Migraine Auras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
9.4 Epilepsy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
9.4.1 Basics of Network Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
9.4.2 Epilepsy and Network Structure . . . . . . . . . . . . . . . . . . . . . . . . 388
Contents
xix
9.5 Neuronal Avalanches, Criticality, and Branching Processes . . . . . . . . 392

9.6 Protein Aggregate Spreading in Neurodegenerative Diseases . . . . . . 399
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
Part I
Neurons
Chapter 1
Single Neuron Modeling
In this introductory chapter we describe conductance-based models of single

neurons, which take into account action potential generation, the effects of external stimuli, synaptic and dendritic processing, and ion channel fluctuations.
In the appendix we provide an informal introduction to stochastic calculus. In
Chap. 2 we turn to the classical theory of waves in one-dimensional neural media,
as exemplified by action potential propagation along axons. This will introduce
some of the techniques used throughout the book, including phase-plane analysis, slowfast systems, stability theory and Evans functions, averaging theory, and
stochastic wave propagation. We then consider two examples of intracellular waves
propagating along spiny dendrites (Chap. 3): the spikediffusespike model of
calcium-mediated spikes and a reactiondiffusion model of CaMKII translocation
waves. A number of additional methods will be presented, including the analysis of
waves in spiking models, homogenization theory, and the theory of pulled fronts.
Finally, in Chap. 4 we consider both intracellular and intercellular calcium waves,
emphasizing aspects that are specific to neurons. In particular, we highlight the
important role of calcium signaling in astrocytes.
1.1 Conductance-Based Model of a Neuron

Cortical neurons typically consist of a cell body (or soma) where the nucleus containing DNA is located, a branching output structure known as the axon and a
branching input structure known as the dendritic tree; see Fig. 1.1. Neurons mainly
communicate with each other by sending electrical impulses or spikes (action potentials) along their axons. (Some neurons are also coupled diffusively via gap junctions
[126].) These axons make contacts on the dendrites of other neurons via microscopic
junctions known as synapses. The arrival of an electrical spike at a synaptic junction leads to the flow of electrical current along the dendritic tree of the stimulated
neuron. If the total synaptic current from all of the activated synapses forces the
electrical potential within the cell body to cross some threshold, then the neuron
P.C. Bressloff, Waves in Neural Media: From Single Neurons to Neural Fields,
Lecture Notes on Mathematical Modelling in the Life Sciences,
DOI 10.1007/978-1-4614-8866-8 1, Springer Science+Business Media New York 2014
1 Single Neuron Modeling
fires a spike. The standard biophysical model for describing the dynamics of a
single neuron with somatic membrane potential v is based upon conservation of
electric charge:
dv
(1.1)
C = Icon + u + Iext,
dt
where C is the cell capacitance, Icon is the membrane current, u denotes the sum
of synaptic currents entering the cell body, and Iext describes any externally injected
currents. Ions can diffuse in and out of the cell through ion specific channels embedded in the cell membrane. Ion pumps within the cell membrane maintain concentration gradients, such that there is a higher concentration of Na+ and Ca2+ outside the
cell and a higher concentration of K+ inside the cell. The membrane current through
a specific channel varies approximately linearly with changes in the potential v relative to some equilibrium or reversal potential, which is the potential at which there is
a balance between the opposing effects of diffusion and electrical forces. Summing
over all channel types, the total membrane current (flow of positive ions) leaving the
cell through the cell membrane is
Icon = gs (v Vs ),
(1.2)
where gs is the conductance due to channels of type s and Vs is the corresponding

reversal potential. In the case of a channel selective to a single ion, Vs satisfies the
Nernst equation

[outside]s
kB T
ln
Vs =
,
(1.3)
q
[inside]s
where q is the charge of the ion, kB is the Boltzmann constant, T is temperature (in
degrees Kelvin), and [outside]s , [inside]s denote the extracellular and intracellular
concentrations of the given ion. Typical values for the common ion species are VK
75 mV, VNa 50 mV, VCa 150 mV, and VCl 60 mV (which is close to the
resting potential of the cell).
The generation and propagation of an action potential arises from nonlinearities
associated with active membrane conductances. Recordings of the current flowing
through single channels indicate that channels fluctuate rapidly between open and
closed states in a stochastic fashion. Nevertheless, most models of a neuron use
deterministic descriptions of conductance changes, under the assumption that there
are a large number of approximately independent channels of each type. It then
follows from the law of large numbers that the fraction of channels open at any
given time is approximately equal to the probability that any one channel is in an
open state. The conductance gs for ion channels of type s is thus taken to be the
product gs = gs Ps where gs is equal to the density of channels in the membrane
multiplied by the conductance of a single channel and Ps is the fraction of open
channels. The voltage dependence of the probabilities Ps in the case of a delayedrectifier K+ current and a fast Na+ current was originally obtained by Hodgkin and
Huxley [279] as part of their Nobel Prize winning work on the generation of action
potentials in the squid giant axon. The delayed-rectifier K+ current is responsible for
dendrite
Iext
soma
u(t)
v(t)
synapse
spike train
axon
a(t)
Fig. 1.1 Basic structure of a neuron. (Inset shows a synaptic connection from an upstream or
presynaptic neuron and a downstream or postsynaptic neuron.) See text for details
terminating an action potential by repolarizing a neuron. One finds that opening of

the K+ channel requires structural changes in 4 identical and independent subunits
so that PK = n4 where n is the probability that any one gate subunit has opened. In
the case of the fast Na+ current, which is responsible for the rapid depolarization
of a cell leading to action potential generation, the probability of an open channel
takes the form PNa = m3 h where m3 is the probability that an activating gate is open
and h is the probability that an inactivating gate is open. Depolarization causes m to
increase and h to decrease, whereas hyperpolarization has the opposite effect.
The dynamics of the gating variables m, n, h are usually formulated in terms of
a simple kinetic scheme that describes voltage-dependent transitions of each gating
subunit between open and closed states. More specifically, for each X {m, n, h}
dX
= X (v)(1 X) X (v)X,
dt
(1.4)
where X (v) is the rate of the transition closed open and X (v) is the rate of the
reverse transition open closed. Equation (1.4) can be rewritten in the alternative
form
dX
= X (v) X,
X (v)
with X {m, n, h},
(1.5)
dt
where
1
,
X (v) = X (v)X (v).
X (v) =
X (v) + X (v)
It follows that the conductance variables m, n, and h approach the asymptotic values
m (v), n (v), and h (v) exponentially with time constants m (v), n (v), and h (v),
respectively. From basic thermodynamic arguments, the opening and closing rates
are expected to be exponential functions of the voltage. Hodgkin and Huxley [279]
fitted exponential-like functions to the experimental data obtained from the squid
axon:
0.1(v + 40)
1 exp[0.1(v + 40)]
0.01(v + 55)
n =
1 exp[0.1(v + 55)]
1
h =
1 + exp[0.1(v + 35)]
m =
h = 0.07 exp[0.05(v + 65)],

m = 4.0 exp[0.556(v + 65)],
n = 0.125 exp[0.125(v + 65)].
All potentials are measured in mV, all times in ms, and all currents in A/cm2 . The
corresponding asymptotic functions X (v) and time constants X (v) are plotted in
Fig. 1.2.
1.0
10
h
0.6
(ms)
0.8
0.4
0.2
6
4
2
m
-80
-40
v (mV)
-80
-40
v (mV)
Fig. 1.2 Voltage-dependent steady-state levels of activation and inactivation (left panel) and
voltage-dependent time constants (right panel) for the HodgkinHuxley model
We can now write down the HodgkinHuxley model for the generation of an action potential, which takes the membrane current to be the sum of a leakage current,
a delayed-rectifier K+ current, and a fast Na+ current,
C
dv
= f (v, m, n, h) + Iext ,
dt
(1.6)
with
f (v, m, n, h) = gNa m3 h(v VNa ) gK n4 (v VK ) g L(v VL).
(1.7)
The maximal conductances and reversal potentials used in the original model are
gL = 0.003 ms/mm2 , gK = 0.36 mS/mm2 , gNa = 1.2 mS/mm2 , VL = 54.387 mV,
VK = 77 mV, and VNa = 50 mV. Note that the leakage current groups together various voltage-independent processes such as the currents carried by ion pumps that
maintain the concentration gradients. The variables m, n, h evolve according to (1.4).
The temporal evolution of the variables v, f , m, n, h during a single action potential is
shown in Fig. 1.3. Injection of a depolarizing current induces a rapid increase in the
m variable describing activation of the Na+ current. Since the slower h variable is
initially around 0.6, there is a large influx of Na+ ions, producing a sharp downward
spike in the membrane current and a rapid depolarization through positive feedback.
However, the rise in the membrane potential causes the Na+ conductance to inactivate by driving h towards zero. In addition, the depolarization activates the K+
conductance, resulting in a subsequent hyperpolarization.
1
m 0.5
0
1
v (mV)
50
0
h 0.5
-50
f (A/mm 2)
0
5
n 0.5
-5
(ms)
10
0
15
(ms)
10
15
Fig. 1.3 The dynamics of v, f , n, m, h in the HodgkinHuxley model during the firing of an action
potential induced by a current injection at t = 5 ms
Once a neuron has generated an action potential, it propagates as a traveling pulse

along the axon of the neuron. In order to model a propagating action potential, it is
necessary to combine the HodgkinHuxley conductance-based model given by (1.4)
and (1.6) with a one-dimensional cable equation describing passive voltage changes
along an axon or dendrite. The result is a partial differential equation of the form
v
2v
= K 2 + f (v, m, n, h),
(1.8)
t
x
where K is related to the membrane space constant of the cable; see Sect. 1.4, and
X = m, n, h evolve according to (1.4). Equation (1.8) is an example of a nonlinear reactiondiffusion equation used to model wave propagation in an excitable
medium; one characteristic of an excitable medium is that it cannot support the
passing of another wave until a certain amount of time has passed (known as the refractory period). A rigorous proof of the existence of traveling wave solutions of the
spatially extended HodgkinHuxley equations has been developed [104, 267]. However, following standard treatments of waves in excitable media [242, 322, 444],
we will develop the theory by considering the simpler FitzHughNagumo model
[192, 446]; see Chap. 2.
There is an ongoing debate about how best to characterize the output of a neuron,
from either a dynamical systems or an information processing perspective. It is generally agreed that the detailed shape of an action potential is usually unimportant,
so, at the fine temporal scale of a few milliseconds, the output of a neuron can be
C
represented in terms of the times at which the neuron fires an action potential (or
spike). Given the conductance-based model (1.1), one typically identifies a firing
threshold such that if v(t) crosses the threshold from below at time t = T , then
the neuron fires an action potential. If T m denotes the mth firing time of the neuron
since t = 0, say, then we have the threshold condition
T m = inf{t,t > T m1 |v(t) = , v(t)
> 0}.
(1.9)
As an alternative to a spike timing representation of neuronal output, one can consider a rate-based representation that is obtained by filtering the spike train with
some causal integral kernel (t), (t) = 0 for t < 0:
z(t) = (t T ) =
m
where
(t )a( )d ,
a(t) = (t T m ).
(1.10)
(1.11)
For example, if (t) = T 1 for 0 t < T and is zero otherwise, then z(t) simply counts the number of spikes within the time interval [t T,t]. In the special
case of a regular spike train with T m+1 T m = 0 for all m, z(t) = 1/0 in the
limit T .
Suppose, for the moment, that we ignore synaptic currents and consider what
happens as the external input Iext to a neuron is increased. Experimentally it is
found that most cortical neurons switch from a resting state characterized by a low
rate of (noise-driven) spontaneous firing to an active state characterized by either
tonic (regular, repetitive) firing or bursting [131]. There has been considerable theoretical work on the transitions from resting to active states in conductance-based
models based on bifurcation theory; see [173, 301] for excellent reviews. We will
focus on tonic firing neurons, since these comprise the majority of cells in cortical
networks. In the case of constant input Iext = I, the firing rate z (mean number of
spikes per second) of the neuron is typically found to be a nonlinear function of
the input:
z = F(I)
(1.12)
with the form of F depending on the nature of the bifurcation from the stable resting state to repetitive firing. A common bifurcation scenario in conductance-based
models of cortical neurons is a saddlenode on an invariant circle [173, 301], which
is classified as type I excitability. Close to the bifurcation point (see Fig. 1.4), we
have
(1.13)
F(I) = F0 I Ic ,
where Ic is the critical current for onset of regular spiking. (Another common form
of excitability is type II, in which the loss of stability of the resting state and the
transition to repetitive firing occur via a Hopf bifurcation. There are also more exotic
forms of excitability as detailed elsewhere [201, 423].) If one includes stochastic

effects arising from synaptic and membrane noise, for example, then the effective
mean firing rate becomes a smooth sigmoid-like function of injected current,
F(I) =
F0
,
1 + e (I )
(1.14)
where is the gain and is the firing threshold. In the high-gain limit , this
reduces to a Heaviside firing rate function

F0 if I >
F(I) = F0 H(I ) =
(1.15)
0 if I < .
Yet another commonly used firing rate function is the piecewise linear function
0, I < ,
F(I) = (u ), < I < + 1 ,

(1.16)
1, I > + 1 .
This preserves the hard threshold of the saddlenode on a limit cycle bifurcation but
ensures that the firing rate saturates at high input currents.
1
0.8
0.6
F(I)
0.4
0.2
0
-0.1
0.1
0.2
I
0.3
0.4
0.5
Fig. 1.4 Various forms of the nonlinear firing rate function F(I). Sigmoid function (black curve)
and Heaviside function (dashed curve) have a threshold = 0.3, whereas the square root function
(gray curve) has a critical current Ic = 0.05
The issue of how to represent a single neuron becomes even more salient
when considering synaptically coupled spiking networks (Chap. 5) and neural fields
(Chap. 6). In order to make analytical progress, it is necessary either to consider a
very simple model of a spiking neuron such as integrate-and-fire [323] (Sect. 5.3)
or to carry out some form of reduction of a conductance-based model. In the case
of weakly coupled neural oscillators, one can reduce the dynamics to a network of
coupled phase oscillators; see Chap. 5. The basic ideas underlying so-called phase
reduction methods can be understood by considering the simpler case of a single,
periodically forced neural oscillator.
10
1.2 Periodically Forced Neural Oscillator

A conductance-based model of a neuron with constant input current can be formulated as an M-dimensional (M 2) system of ODEs
dx
= f(x),
dt
x = (x1 , . . . xM ).
(1.17)
Here x1 , say, represents the membrane potential of the neuron (treated as a point
processor) and xm , m > 1, represent various ionic channel gating variables. Suppose
that the neuron has a stable periodic solution x(t) = x(t + 0 ) where 0 = 2 /0
is the natural frequency of the oscillator. In phase space the solution is an isolated
attractive trajectory called a limit cycle. The dynamics on the limit cycle can be
described by a uniformly rotating phase such that
d
= 0 ,
dt
(1.18)
and x(t) = g( (t)) with g a 2 -periodic function. Note that the phase is neutrally
stable with respect to perturbations along the limit cyclethis reflects invariance of
an autonomous dynamical system with respect to time shifts. Now suppose that a
small external periodic input is applied to the oscillator such that
dx
= f(x) + p(x,t),
dt
(1.19)
where p(x,t) = p(x,t + ) and = 2 / is the forcing frequency. If the amplitude

is sufficiently small and the cycle is stable, then deviations transverse to the limit
cycle are small so that the main effect of the perturbation is to induce shifts in the
phase. Therefore, we need to extend the definition of phase to a neighborhood of the
limit cycle. This leads to the notion of an isochrone [221, 350, 679].
1.2.1 Isochrones and Phase-Resetting Curves

Suppose that we observe the unperturbed system stroboscopically at time intervals
of length 0 . This leads to a Poincare mapping
x(t) x(t + 0 ) P(x(t)).
This mapping has all points on the limit cycle as fixed points. Choose a point x on
the cycle and consider all points in the vicinity of x that are attracted to it under
the action of P. They form an (M 1)-dimensional hypersurface I , called an
isochrone, crossing the limit cycle at x (see Fig. 1.5). A unique isochrone can be
drawn through each point on the limit cycle so we can parameterize the isochrones
by the phase, I = I ( ). Finally, we extend the definition of phase by taking all
11
points x I ( ) to have the same phase, (x) = , which then rotates at the natural
frequency 0 (in the unperturbed case). Hence, for an unperturbed oscillator in the
vicinity of the limit cycle, we have
0 =
dxk
d (x)
=
=
fk (x).
dt
k xk dt
k xk
Now consider the perturbed system but with the unperturbed definition of the phase:
d (x)
=
( fk (x) + pk (x,t)) = 0 +
pk (x,t).
dt
k xk
k xk
Fig. 1.5 Isochrones in the neighborhood of a stable limit cycle
To a first approximation we can neglect deviations of x from the limit cycle which
we denote by x :
d (x)
(x )
= 0 +
pk (x ,t).
dt
x
k
k
Finally, since points on the limit cycle are in 1:1 correspondence with the phase ,
we obtain the closed phase equation
d
= 0 + Q( ,t),
dt
where
Q( ,t) =
k
(x ( ))
pk (x ( ),t)
xk
is a 2 -periodic function of and a -periodic function of t.
(1.20)
(1.21)
12
Consider, as an example, the complex amplitude equation that arises for a limit
cycle oscillator close to a Hopf bifurcation [248]:
dA
= (1 + i )A (1 + i )|A|2A,
dt
A C.
(1.22)
In polar coordinates A = Rei ,

dR
= R(1 R2),
dt
d
= R2 .
dt
The solution for arbitrary initial data R(0) = R0 , (0) = 0 is

1/2
1 R20 2t
R(t) = 1 +
e
,
R20
(1.23)
(t) = 0 + 0t log(R20 + (1 R20)e2t ),

2
where 0 = is the natural frequency of the stable limit cycle at R = 1. Strobing
the solution at times t = n0 , we see that
lim (n0 ) = 0 ln R0 .
Hence, we can define a phase on the whole plane
(R, ) = ln R.
(1.24)
It follows that the isochrones are logarithmic spirals with ln R = constant.

Now rewrite (1.22) in Cartesian coordinates
dx
= x y (x2 + y2 )(x y) + cos t,
dt
dy
= y + y (x2 + y2 )(y + x),
dt
Tn
T n+1
phase-retarded
Z()
T n T n+1
phase-advanced
Fig. 1.6 Phase shift Z( ) induced by a small perturbation of the membrane potential applied at
time t = 0 when the phase of the limit cycle is . Black (gray) dots represent points on the unperturbed (perturbed) trajectory. The phase shift induces a corresponding shift in successive firing
times
13
where we have added a periodic modulation in the x-direction. Rewrite the phase
(1.24) as
y
= tan1 log(x2 + y2 ),
x 2
so that
y
x
= 2
2
.
x
x + y2
x + y2
On the limit cycle x0 ( ) = (cos , sin ), we have
(x0 )
= sin cos .
x
It follows that the corresponding phase equation is
d
= 0 ( cos + sin ) cos t.
dt
The phase reduction method is particularly useful because the function Q( ,t)
can be related to an easily measurable property of a neural oscillator, namely, its
phase-resetting curve (PRC), which we denote by the 2 -periodic function Z( ).
The PRC is found experimentally (or numerically) by perturbing the oscillator with
a brief depolarizing voltage stimulus of size V at different times in its cycle and
measuring the resulting phase shift from the unperturbed system [221, 679]; see
Fig. 1.6. Taking the coordinate x1 as the membrane potential, it follows from (1.20)
that
d
(x ( ))
= 0 + x1
(t t0 ).
(1.25)
dt
x1
Integrating this equation over a small interval around t0 , we see that the impulse
induces a phase shift = ( x1 )Z(0 ) where Z( ) = (x ( ))/ x1 and 0 =
(t0 ). Thus comparing the phase at large times for the unperturbed and perturbed
cases generates the PRC. Given the PRC Z( ), the response of the neuron to a
more general time-dependent voltage perturbation P(t) is determined by the phase
equation
d
= 0 + Z( )P(t).
(1.26)
dt
We can also express the PRC in terms of the firing times of a neuron (assuming fast reconvergence to the limit cycle). Suppose that there exists a well-defined
threshold signaling the onset of fast somatic membrane depolarization and the
subsequent firing of an action potential spike. Let T n denote the nth firing time of
the neuron as defined by (1.9). Since the membrane voltage v(t) = x1 ( (t)), the
threshold corresponds to a particular phase of the limit cycle, which we choose to
be = 0. In the absence of perturbations, we have (t) = 2 t/0, so that the firing
times are T n = n0 where 0 is the natural period of oscillation. On the other hand,
a small perturbation applied at the point on the limit cycle at time t, T n < t < T n+1 ,
induces a phase shift that changes the next time of firing according to (see Fig. 1.6)
14
b
400
0.2
Type I
200
0.1
Z()
0
-100
Z()
Type II
Type II
0
-0.1
0
Fig. 1.7 (a) MorrisLecar model showing two different response types. In both cases vk = 0.7,
uL = 0.5, uCa = 1, gK = 2, gL = 0.5, v1 = 0.01, v2 = 0.15. For a type I response, gCa = 1.33,
v3 = 0.1, v4 = 0.145, = 1/3, and I = 0.0695. For a type II response, gCa = 1.1, v3 = 0, v4 = 0.3,
= 0.2, and I = 0.25. Responses have been scaled to the same ranges. (b) HodgkinHuxley model
with external drive I = 10 showing type II phase-resetting curve
( V )Z( )
T n+1 T n
= 1
.
0
2
(1.27)
For certain types of neuron a depolarizing stimulus always advances the onset of
the next spike, that is, the PRC is always positive, whereas for others the stimulus
may also delay the next spike. Oscillators with a strictly positive PRC are called
type I whereas those for which the PRC has a negative regime are called type II.
A numerical example illustrating both types of PRC is shown in Fig. 1.7a for the
MorrisLecar model of a neuron, which was originally introduced to describe how
under constant current injection barnacle muscle fibers respond with a host of oscillatory voltage waveforms [440]. It takes the form
dv
= I gL (v vL ) gK w(v vK ) gCa m (v)(v vCa ),
dt
dw
= (v)(w (v) w),
dt
(1.28)
with
m (v) = 0.5(1 + tanh[(v v1)/v2 ]),
w (v) = 0.5(1 + tanh[(v v3)/v4 ]),
(v) = cosh[(v v3 )/(2v4 )].

Here, gL is the leakage conductance, gK , gCa are potassium and calcium conductances, vL , vK , vCa are corresponding reversal potentials, m (v), w (v) are voltagedependent gating functions, and (v) is a voltage-dependent rate. The type II PRC
for the HodgkinHuxley model is shown in Fig. 1.7b.
1.2.2 Phase-Locking and Synchronization

Now suppose that Q( ,t) in (1.20) is expanded as a double Fourier series
15
Q( ,t) = al,k eik +il t .

l,k
Substitute for using the zero-order approximation = 0t + 0 :

Q( ,t) = al,k eik0 +i(k0 +l )t .
l,k
It follows that Q contains fast oscillating terms (compared to the time scale 0 / )
together with slowly varying terms that satisfy the resonance condition
k0 + l 0.
(1.29)
Only the latter will lead to large variations in the phase, so we can average the
forcing term Q keeping only the resonant terms. The simplest case is 0 for
which the resonant terms satisfy l = k and
Q( ,t) ak,k eik( t) = q( t).
(1.30)
The phase equation then becomes

d
= 0 + q( t).
dt
The phase difference between the oscillator and external drive, = t, then
satisfies the equation
d
= + q( ),
(1.31)
dt
where = 0 is the degree of frequency detuning. Similarly, if m0 /n,
then
Q( ,t) an j,m j ei j(m n t) = q(m
n t),
(1.32)
k
and
d
= m0 n + mq(
),
dt
(1.33)
where = m n t.
The above is an example of an application of the averaging theorem [248].
Assuming that = 0 = O( ) and defining = t, we have
d
= + Q( + t,t) = O( ).
dt
Define
1
T T
and consider the averaged equation
q( ) = lim
T
0
Q( + t,t)dt,
(1.34)
(1.35)
16
d
= + q( ).
dt
(1.36)
It is easy to establish that q only contains the resonant terms of Q as above. The
averaging theorem ensures that there exists a change of variables that maps solutions
of the full equation to those of the averaged equation to leading order in . The
question then remains as to what extent solutions of the averaged equations are a
good approximation to the solutions of the full equation. In general, one can only
establish that a solution of the full equation is -close to a corresponding solution
of the averaged system for times of O( 1 ). No such problem occurs however for
hyperbolic fixed points corresponding to phase-locked states.
Suppose that the 2 -periodic function q( ) has a unique maximum qmax and a
unique minimum qmin in the interval [0, 2 ). We can then distinguish between two
regimes [502]:
Synchronization regime: If the degree of detuning for a given drive amplitude is
sufficiently small,
qmin < < qmax ,
Fig. 1.8 Saddlenode bifurcation signaling a transition from a synchronized to a drifting state
as the size of frequency detuning | | increases (a) Synchronization regime. (b) Saddlenode
bifurcation. (c) Drift regime
then there exists at least one pair of stable/unstable fixed points (s , u ). (This
follows from the fact that q( ) is 2 -periodic and continuous so it has to cross
any horizontal line an even number of times.) The system evolves to the synchronized state
(t) = t + s ,
in which the oscillator is phase-locked to the external drive and is frequency
entrained. Note that the stability of a phase-locked state is determined by the
sign of q
( ) with q
(s ) < 0 and q
(u ) > 0 (see Fig. 1.8a).
Drift regime: As | | increases, it approaches one of the critical values qmin,max
where the two fixed points coalesce in a saddlenode bifurcation and phaselocking disappears; see Fig. 1.8b, c. Hence, if the degree of tuning is large, then
d /dt never changes sign and the oscillation frequency differs from the drive
frequency . The phase (t) rotates through 2 with period
1.3 Synaptic Processing
17

2

d
.
T =
q( )
0
(1.37)
The mean frequency of rotation is thus = + where = 2 /T is

known as the beat frequency. One is often interested in how the behavior varies
in the ( , )-plane (see Fig. 1.9). First the boundary between the two regimes
consists of the two straight lines = qmax,min . Second, close to the boundary
has a characteristic form. Suppose, for example, that max is small
for fixed with max = qmax . The integral in (1.37) is then dominated by a
small region around max . Expanding q( ) as a Taylor series,
1

d
2

=
2

(
2 (
T
q
)
max
max

= |q
(max )|( max ).
(1.38)
0
Fig. 1.9 (a) Synchronization regime (shaded) in ( , )-plane. (b) Variation of beat frequency
with for fixed

In the conductance-based model given by (1.6), we decomposed the total input current to the soma into an external part Iext (t) and a synaptic part u(t). In this section,
we consider the current generated at a single synapse and the sequence of events underlying conductance changes in the postsynaptic membrane due to the arrival of an
action potential at the presynaptic terminal. We then show how these conductance
changes can be modeled in terms of a kinetic scheme describing the opening and
closing of ion channels in the postsynaptic membrane.
18
1.3.1 Excitatory and Inhibitory Synapses

The basic stages of synaptic processing induced by the arrival of an action potential
at an axon terminal are shown in Fig. 1.10. (See [99] for a more detailed description.)
An action potential arriving at the terminal of a presynaptic axon causes voltagegated Ca2+ channels within an active zone to open. The influx of Ca2+ produces
a high concentration of Ca2+ near the active zone [45, 195], which in turn causes
vesicles containing neurotransmitter to fuse with the presynaptic cell membrane and
release their contents into the synaptic cleft (a process known as exocytosis). The
released neurotransmitter molecules then diffuse across the synaptic cleft and bind
to specific receptors on the postsynaptic membrane. These receptors cause ion channels to open, thereby changing the membrane conductance and membrane potential
of the postsynaptic cell. A single synaptic event due to the arrival of an action potential at time T induces a synaptic current of the form
Isyn (t) = gsyn (t T )(Vsyn v(t)),
(1.39)
where v is the voltage of the postsynaptic neuron, Vsyn is the synaptic reversal potential, and gsyn (t) is the change in synaptic conductance with gsyn (t) = 0 for t < 0.
The sign of Vsyn relative to the resting potential Vrest (typically Vrest 65 mV) determines whether the synapse is excitatory (Vsyn > Vrest ) or inhibitory (Vsyn < Vrest ).
For simplicity, it is often assumed that a neuron spends most of its time close to rest
such that Vsyn v Vsyn Vrest , with the factor Vsyn Vrest absorbed into gsyn . One is
then effectively taking the arrival of a spike as generating a synaptic current rather
than a change in conductance.
voltage-gated
Ca++ channel
synaptic
vesicle
neurotransmitter
axon
terminal
reuptake
pump
postsynaptic
density
receptors
synaptic
cleft
dendritic
spine
Fig. 1.10 Basic stages of synaptic processing shown for an excitatory synapse. See text for details
The predominant fast, excitatory neurotransmitter of the vertebrate central nervous system is the amino acid glutamate, whereas in the peripheral nervous system, it is acetylcholine. Glutamate-sensitive receptors in the postsynaptic membrane can be subdivided into two major types, namely, NMDA and AMPA [99].
19
At an AMPA receptor the postsynaptic channels open very rapidly. The resulting
increase in conductance peaks within a few hundred microseconds, with an exponential decay of around 1 ms. The time course of the synaptic conductance change
can be modeled in terms of an nth state Markov process [154] (see Sect. 1.3.3).
Usually a simplified representation of gsyn (t) is used that is given by the difference
of exponentials
1
1
gsyn (t) = g
2 1
(et/1 et/2 )H(t),
(1.40)
with H(t) the Heaviside function. In many cases, the rise time is much shorter than
t/d .
the fall time (1 2 ) so that we have an exponential synapse with gsyn (t) = ge
In the limit 2 1 = 1 , (1.40) reduces to the well-known function
gsyn (t) = g 2te t H(t).
(1.41)
These expressions for the conductance are also used for GABA inhibitory synapses
(see below). In contrast to an AMPA receptor, the NMDA receptor operates about
ten times slower and the amplitude of the conductance change depends on the postsynaptic membrane potential. If the postsynaptic potential is at rest and glutamate is
bound to the NMDA receptor, then the channel opens, but it is physically obstructed
by Mg2+ ions. As the membrane is depolarized, the Mg2+ ions move out and the
channel becomes permeable to Na+ and Ca2+ ions. The NMDA conductance can
be derived from a model in which the binding rate constant of Mg2+ varies as an
exponential function of the local voltage v [304]:
gsyn (t) = g
et/1 et/2
,
1 + [Mg2+]e v(t)
t > 0,
(1.42)
where [X] denotes concentration of X and is a rate constant. The rapid influx of
calcium ions due to the opening of NMDA channels is thought to be the critical trigger for the onset of long-term potentiation or LTP, a major component of synaptic
plasticity (see also Sect. 4.1).
The most common inhibitory neurotransmitter in the central nervous system of
both vertebrates and invertebrates appears to be GABA. There are two major forms
of postsynaptic receptors termed A and B. The GABAA receptors open channels
selective to chloride ions, whose reversal potential Vsyn = 70 mV is close to that
of the resting potential. The postsynaptic conductance change is quite fast, rising
within 1 ms and decaying within 1020 ms. GABAB receptors are at least 10 times
slower and open ion channels selective for K + ions. Thus they tend to be considerably more hyperpolarizing with Vsyn 100 mV. The two receptor classes tend to
be segregated with GABAA occurring at or close to the soma and GABAB further
out on the dendrites. Another way to distinguish between GABAA and GABAB receptors is that the former are ionotropic (as are NMDA and AMPA) while the latter
are metabotropic [99]. Neurotransmitter binding to an ionotropic receptor directly
opens an ion channel through a series of conformational changes of the receptor. On
20
the other hand, neurotransmitter binding to a metabotropic receptor indirectly opens

an ion channel elsewhere in the membrane through a sequence of biochemical steps
mediated by G proteins.
1.3.2 Synaptic Depression

A single synaptic event due to the arrival of an action potential at time T induces a
synaptic current of the form (1.39). As a crude approximation we might try summing
individual responses to model the synaptic current arising from a train of action
potentials arriving at times T m , integer m:
Isyn (t) = gsyn (t T m )(Vsyn v(t)).
(1.43)
Note that this sum only includes spikes for which T m < t since gsyn (t) = 0 for t < 0
(causality condition). For many synapses such a simple ansatz does not hold, since
some form of short-term synaptic depression causes the amplitude of the response to
depend on the previous history of presynaptic firing [4, 405]. One way to incorporate
this history-dependent effect is to take [1]

m
m
(1.44)
Isyn (t) = q(T )gsyn (t T ) (Vsyn v(t)),
m
where the factor q(T m ) reduces the response evoked by an action potential by an
amount that depends upon the details of the previous spike train data. One interpretation of the factor q is that it represents a short-term (reversible) reduction in
the release probability for synaptic transmission due to a depletion in the number
of vesicles that can readily fuse with the cell membrane [700]. In certain cases, it is
also possible for a synapse to undergo a temporary facilitation in response to activation, which may be due to the presence of residual calcium in the axonal terminal
[700].
A common phenomenological model of synaptic depression is to assume that
between spikes q(t) relaxes at a rate q to its steady-state value of one, but that
directly after the arrival of a spike it changes discontinuously, that is, q q with
< 1. The depression time constant q can vary between around 100 ms and a few
seconds [4]. The model for synaptic depression may be written succinctly as
dq (1 q)
=
(1 ) q(T n ) (t T n ),
dt
q
n
q(0) = 1,
which has the solution of the form

q(T m ) = 1 (1 ) [mn1] e(T
n<m
m T n )/
(1.45)
21
Assuming a regular sequence of incoming spikes T n T n1 = for all n we find

that the asymptotic amplitude q ( ) limm q(T m ) is given by
q ( ) =
1 e /q
.
1 e /q
(1.46)
One possible computational role for synaptic depression is as a mechanism for cortical gain control [4]. The basic idea can be understood from the dependence of
the asymptotic amplitude q ( ) on the stimulus frequency f = 1 . Assuming
that q , we can Taylor expand q in (1.46) to find that q ( f ) / f , where
= q /(1 ). The main point to note is that the postsynaptic response per unit
time is approximately independent of f (assuming that each spike elicits the same
response in the steady state). This means that the synapse is very sensitive to changes
in the stimulus frequency. The instantaneous response to a rapid increase f in the
stimulus rate is given by f / f . In other words, the synapse responds to relative
rather than absolute changes in the rate of input stimulation.
1.3.3 Kinetic Model of a Synapse

Let gsyn (t) s(t) where s(t) is the fraction of synaptic receptor channels that are
in an open conducting state. The probability of being in an open state depends on
the presence and concentration T of neurotransmitter released by the presynaptic
neuron. Assuming a first-order kinetic scheme, in which a closed receptor in the
presence of a concentration of neurotransmitter T equilibrates with the open receptor state, we have
r1 (v,T )
C O,
r2 (v)
where C and O represent the closed and open states of the channel and r1 (v, T ) and
r2 (v) are the associated rate constants. However, in many cases synaptic channels
are found to have time-dependent properties that are more accurately modeled with
a second-order kinetic scheme. In fact the presence of one or more receptor sites
on a channel allows the possibility of transitions to desensitized states. Such states
are equivalent to the inactivated states of voltage-dependent ion channels. The addition of such a desensitized state to the first-order process generates a second-order
scheme:
ds
= r1 (v, T )(1 s z) [r2(v) + r3 (v)]s + r4 (v)z,
dt
dz
= r6 (v, T )(1 s z) [r4(v) + r5 (v)]z + r3 (v)s,
dt
(1.47)
where z is the fraction of channels in the desensitized state. All neurotransmitterdependent rate constants have the form ri (v, T ) = ri (v)T . It is common for detailed
Markov models of voltage-gated channels to assume that the voltage dependence
22
of all rates takes a simple exponential form. However, it has been shown that the
number of states needed by a model to more accurately reproduce the behavior of a
channel may be reduced by adopting sigmoidal functions for the voltage-dependent
transition rates (see Destexhe et al. [154] for a discussion), so that we write
ri (v) =
ai
.
1 + exp[(v ci)/bi ]
(1.48)
The ai set the maximum transition rate, bi the steepness of the voltage dependence, and ci the voltage at which the half-maximal rate is reached. Furthermore,
the concentration of neurotransmitter can often be successfully approximated by a
sigmoidal function of the presynaptic potential vpre :
T (vpre ) =
Tmax
.
1 + exp[(vpre v )/ ]
(1.49)
Here, Tmax is the maximal concentration of transmitter in the synaptic cleft, vpre
is the presynaptic voltage, gives the steepness, and v sets the value at which
the function is half activated. It is common to take = 5 mV and v = 2 mV. One
of the main advantages of using an expression such as (1.49) is that it provides
a smooth transformation between presynaptic voltage and transmitter concentration from which postsynaptic currents can easily be calculated from (1.39), (1.47),
(1.48), and (1.49).
Now consider the following second-order gating scheme
r1 (T )
C C1

r2
r4

r3
O
where C and C1 are the closed forms of the receptor, O is the open (conducting)
form, and the ri are voltage-independent transition rates. Under certain assumptions
it may be shown that this particular second-order scheme describes the so-called
alpha function response commonly used in synaptic modeling. The following approximations are required: (i) The transmitter concentration T occurs as a pulse
(t t0 ) for a release event occurring at time t = t0 , that is, r1 (T ) = r1 (t t0 );
(ii) The fraction of channels in C is considered constant and 1. The kinetic equation (1.47) then reduce to
dc(t)
= Qc(t) + I(t),
dt
(assuming c(0) = 0), where

11 0
r1 (t t0)
z
Q=
,
I(t)
=
,
c
=
,
0
s
r3 1
2
23
and 1 = 1/(r2 + r3 ), 2 = 1/r4 . Here z and s represent the fraction of receptors in

the forms C1 and O, respectively. This Markov chain system has a solution of the
form
t
c(t) =
G(t s)I(s)ds, G(t) = etQ .
0
The eigenvectors of Q are (1, r3 /(21 11 )) and (0, 1) with associated eigenvalues
1/1 and 1/2 , respectively. Hence, one finds that

s(t) = r1 r3
1
1
2 1
1
(e(tt0 )/1 e(tt0 )/2 ),
t > t0 .
In the limit 2 1 s this reduces to an alpha function

s(t) = r1 r3 (t t0 )e(tt0 )/s ,
t > t0 .
This kinetic derivation of the alpha function only holds for s 1 in order to remain
consistent with condition (ii).
The time course of some ion-channel open and closed states seems to follow a
power law rather than multiexponential law at large times [429]. In order to understand such power-law behavior, consider an ion channel with N closed states such
that the transition to an open state can only take place from state 1 at one end of a
chain
N1
N1
01 2 . . . N.
The corresponding kinetic equations are
dc1
= 1 c2 (1 + )c1 ,
dt
dcn
= n1 cn1 + ncn+1 (n + n1)cn ,
dt
dcN
= N1 cN1 N1 cN .
dt
1 < n < N,
In the following we take n = n = 1 for all n and = 1, so that the system of equations describes a discrete diffusion process along a chain with a reflecting boundary
at n = N and an absorbing boundary at n = 0. In the large N limit, it can be shown
that given the initial condition pn (0) = n,1 , the exact solution is
cn (t) = e2t [In1 (t) In+1(t)],
where In (t) is the modified Bessel function of integer order:
In (t) =
eink e2t cos(k)
dk
.
2
(1.50)
24
By carrying out an asymptotic expansion for large t, it can be shown that

cn (t)
n
.
2 1/2t 3/2
Define F(t) to be the total probability of finding the system in a closed state:
F(t) =
cn (t).
n=1
It follows that dF/dt = c1 and, hence, F(t) ( t)1/2 for large N,t and = 1.
More recently, it has been suggested that synapses with multiple states, which exhibit dynamics over a wide range of time scales and show power-law-like behavior,
could have some interesting computational properties [208, 219]. For example, it
has been suggested that such synapses could provide a way of combining high levels of memory storage with long retention times [208].
1.4 Dendritic Processing

Typically, a single neuron in cerebral cortex has up to 10,000 synapses, which are
spatially distributed along the dendritic tree (and perhaps on the cell body and proximal part of the axon). In order to find the total synaptic current u(t) entering the cell
body, it is necessary to determine how the various local currents flow along the dendritic tree and combine at the soma. We will show that if the dendrites are modeled
as passive electrical cables, then the dendritic tree acts as a linear spatiotemporal
filter of synaptic currents.
1.4.1 The Cable Equation

Neurons display a wide range of dendritic morphologies, ranging from compact arborizations to elaborate branching patterns. At the simplest level, the dendritic tree
can be treated as a passive electrical medium that filters incoming synaptic stimuli
in a diffusive manner. The current flow and potential changes along a branch of the
tree may be described with a second-order, linear partial differential equation commonly known as the cable equation. (The application of cable theory to the study of
passive, spatially extended dendrites was pioneered by Wilfrid Rall in the 1960s and
1970s. For more recent accounts of this work see [515] and the annotated collection
of papers edited by Segev, Rinzel, and Shepherd [299].) The cable equation is based
on a number of approximations: (1) magnetic fields due to the movement of electric
charge can be neglected, (2) changes in ionic concentrations are sufficiently small
so that Ohms law holds, (3) radial and angular components of voltage can be ignored so that the cable can be treated as one-dimensional medium, and (4) dendritic
membrane properties are voltage-independent, that is, there are no active elements.
25
A nerve cable consists of a long thin, electrically conducting core surrounded by a

thin membrane whose resistance to transmembrane current flow is much greater than
that of either the internal core or the surrounding medium. Injected current can travel
long distances along the dendritic core before a significant fraction leaks out across
the highly resistive cell membrane. Linear cable theory expresses conservation of
electric current in infinitesimal cylindrical elements of nerve fiber modeled using
the equivalent circuit shown in Fig. 1.11. Define v(x,t) as the membrane potential
at position x along a cable at time t (measured relative to the resting potential of
the membrane). Let Cm be the capacitance per unit area of the cell membrane, R
the resistivity of the intracellular fluid (in units of resistance length), Rm the cell
membrane resistance (in units of resistance area), and a the cable radius. Note that
Cm , Rm , R are independent of cable radiusthe corresponding quantities expressed
per unit length of cable are
r=
1
2 a
=
,
rm
Rm
R
,
a2
cm = 2Cm a.
(1.51)
I ext(x,t)
Il(x,t)
Im(x,t)
cm
v(x,t)
rm
xx
x+x
Fig. 1.11 Equivalent circuit for successive cylindrical segments of passive dendritic membrane
Current conservation implies that (see Fig. 1.11)

I ext (x,t) Im (x,t) =
Il (x,t) Il (x x,t) Il (x,t)
,
x
x
(1.52)
where I ext (x,t) is an external injected current density. From Ohms law we also have
Il (x,t)r =
and
v(x,t)
v(x,t) v(x + x,t)
,
x
x
(1.53)
26
Im (x,t) = cm
v(x,t) v(x,t)
+
.
t
rm
(1.54)
Combining these three equations yields the uniform cable equation
v(x,t)
2 v(x,t)
= v(x,t) + m2
+ rm Iext (x,t),
t
x2
t 0,
(1.55)
where m = RmCm is the membrane time constant and m = (Rm a/2R)1/2 is the
membrane space constant. (It follows that the coupling constant appearing in the
HodgkinHuxley equation (1.8) is K = m2 /Rm .)
Infinite Cable. In the case of an infinite uniform cable, x R, we can solve (1.55)
by Fourier transforming with respect to x. That is, define the Fourier transform of v
(and other quantities) as
v(k,t)
=
with inverse transform
v(x,t) =
eikx v(x,t)dx,
eikx v(k,t)
dk
.
2
Then
v(k,t)
= v(k,t)
m2 k2 v(k,t)
+ rm Iext (k,t),
t
x R,
t 0.
(1.56)
This first-order ODE can now be solved straightforwardly as

rm
v(k,t)
=
m
t
G 0 (k,t t
)Iext (k,t
)dt
,
(1.57)
where G 0 (k,t) = e(1+m k )t/m . Taking the inverse Fourier transform and using the
convolution theorem shows that
t
v(x,t) = rm
G0 (x x
,t t
)I ext (x
,t
)dx
dt
,
(1.58)
2 2
where we have absorbed m into rm and

G0 (x,t) =

dk ikx (1+m2 k2 )t/m
e e
2
2
2
1

=
et/m em x /4mt .
2m t/m
(1.59)
The function G0 (x,t) is the fundamental solution or Greens function for the cable equation with unbounded domain. It is positive and symmetric and satisfies the
homogeneous cable equation
27

2
2
m + 1 m 2 G0 (x,t) = 0,
t
x
(1.60)
G0 (x, 0) = (x).
(1.61)
with initial condition
Moreover, for any 0 < s < t, it satisfies the Markovian property

G0 (x y,t) =
G0 (x z, s)G0 (z y,t s)dz.
(1.62)
The Greens function is plotted as a function of time in Fig. 1.12a for a range of
separations x.
Semi-infinite Cable. Using Fourier cosine or sine transforms with respect to x it is
straightforward to determine the Greens function for a semi-infinite cable x [0, )
with either an open-end boundary condition
v(0,t) = 0
or a closed-end boundary condition (zero current flow)

v(x,t)
=0
x x=0
One finds that
v(x,t) = rm
t
G (x, x
,t t
)I ext (x
,t
)dx
dt
(1.63)
(1.64)
(1.65)
where
G (x, y,t) = G0 (x y,t) G0(x + y,t)
(1.66)
for the open () and closed (+) cases. For a discussion of finite-length cables see
[342].
Single Branching Node. Let us now consider a single branching node and label
each semi-infinite segment by the index i = 1, . . . , N. (Typically N = 3.) We shall
assume that the cables only differ in their radius ai . In order to simplify the analysis
we will measure the
distance along the ith cable from the branch point at x = 0
in units of m,i = Rm ai /2R such that the cable equation on each branch can be
written as
vi (X,t)
2 vi (x,t)
= vi (x,t) +
m
+ Ii (x,t).
(1.67)
t
x2
The boundary conditions are continuity of the potential at the node
vi (0,t) = v j (0,t),
and conservation of current
(1.68)
28
x=1
0.2
vi
Gik(x,y,t)
x
0.15
G0(x,t)
x=2
0.1
vk
vj
x=3
0.05
x=4
x=6
5
10
time t
20
15
G0(x+y,t)
Fig. 1.12 (a) Greens function of an infinite cable as a function of time t (in units of m ) for various
values of distance x (in units of ). (b) Branching dendritic tree of an idealized single neuron and
an equivalent cylinder representation
aj
3/2
j=1

v j (x,t)
= 0.
x x=0
(1.69)
3/2
The factor ai arises from the fact that we have rescaled length and that the longitudinal resistance varies as the cross-sectional area of the cable. The solution of
these equations is
Gi j (x, x
,t t
)I j (x
,t
)dx
dt
,
(1.70)
Gi j (x, y,t) = i, j G0 (x y,t) + (2p j i, j )Go (x + y,t),
(1.71)
vi (x,t) =
where
t
j=1
and
3/2
pk =
ak
3/2
m am
(1.72)
It is interesting to note that if pk = 1/2 for a certain segment, then the response to
current injected into that segment can be represented by a single equivalent cylinder;
see Fig. 1.12b. That is, if i = j, then Gi j (x, y,t) = G0 (x + y), where x + y is the
distance of the direct path between x and y on the tree. Hence, the node in some
sense acts no differently than an ordinary point on a single cable. This is the basis
of Ralls equivalent cylinder concept [515].
Dendritic Tree. An arbitrary dendritic tree may be considered as a set of branching nodes linked by finite pieces of uniform cable. Given an external current density I ext (x,t) the voltage response has the formal solution (absorbing rm into the
definition of G)
29
v(x,t) =
t
G(x, y,t s)Iext (y, s)dy ds.
(1.73)
The associated Greens function G(x, y,t s) satisfies the homogeneous cable equation on each segment together with boundary conditions at the branching and terminal nodes of the tree. Rules for constructing such a Greens function have been
developed by Butz and Cowan [95] using a graphical calculus and by Abbott et al.
[3] using path-summing methods. The latter approach can be understood in terms
of a compartmental model obtained by spatially discretizing the cable equation. The
construction of the discretized Greens function involves summing over paths of
a random walk on the tree with the corresponding Greens function for the cable
equation recovered in the continuum limit [77].
1.4.2 Dendritic Filtering of Synaptic Inputs

So far we have considered the linear response of a dendritic cable to external current injection as determined by the Greens function or transfer function. Suppose
that we replace the external current by a synaptic current of the form discussed in
Sect. 1.3. That is, Iext (x,t) I(x,t), where I(x,t) is the synaptic current density at
location x at time t:
I(x,t) = (x) gsyn (t T m (x))[Vsyn v(x,t)] g(x,t)[Vsyn v(x,t)],
(1.74)
where g(x,t) = (x) m gsyn (t T m (x)). Here (x) is the density of synapses (assuming that they have identical properties) and {T m (x)} is the sequence of spikes
arriving into the synapses located at x. In the case of a discrete set of identical
synapses at dendritic locations {x j , j = 1, . . . , M}, we have (x) = j (x x j ) and
T m (x j ) = T jm . The formal solution for the membrane potential is now [see (1.73)]
v(x,t) =
t
G(x, x
,t t
)g(x
,t
)[Vsyn v(x
,t
)]]dx
dt
(1.75)
which is a Volterra integral equation of the second kind. In order to solve this integral
equation, we introduce the convolution operator ,
t
[G f ](x,t) :=
G(x, x ,t t ) f (x ,t )dx dt
(1.76)
for any function f (x,t). We can then iterate (1.75) to obtain a series solution for v:
v = Vsyn G g G (gv)
= Vsyn G g VsynG (g[VsynG g G (gv)])
2
= Vsyn G g Vsyn
G [gG g] + G [gG (gv)]
30

2
3
= Vsyn G g Vsyn
G (gG g) + Vsyn
G (gG [gG g]) . . .
2
= Vsyn (G VsynG gG + Vsyn
G gG gG . . .) g
= Vsyn G g,
(1.77)
where
2
G gG gG . . .
G := G VsynG gG + Vsyn
(1.78)
which is conis a Neumann series expansion for the effective Greens function G,
vergent for a passive cable [344]. More explicitly, we can write the solution as
t
G(x,t;
x
,t
)g(x
,t
)dx
dt
,
(1.79)
v(x,t) = Vsyn
u(t)
synapse
dendritic cable
V(t)
g(x,t)
soma
Fig. 1.13 Schematic diagram of a neuron consisting of a soma resistively coupled to one end of a
dendritic cable. A synaptic conductance change g(x,t) at position x on the cable induces a synaptic
current into the soma at x = 0
with G satisfying the Volterra integral equation
G(x,t;
x
,t
) = G(x, x
,t t
)
(1.80)
t

,t
; x
,t
)dx
dt
.
Vsyn
G(x, x
,t t
)g(x
,t
)G(x
t
The
One can check that iteration of this equation recovers the series expansion for G.
major point to note is that the resulting series involves interactions between synaptic
conductances at different points on the dendritic cable. For simplicity, however, we
shall assume that Vsyn v(x,t) so that G G and synaptic interactions become
negligible.
Given a distribution of synaptic inputs innervating the dendritic tree, what is
the net synaptic current u(t) entering the soma? In order to address this problem,
consider a semi-infinite uniform dendritic cable, 0 x < , with the soma located
at the end x = 0. The soma is modeled as a conductance-based point process that is
passively coupled to the dendritic cable as illustrated in Fig. 1.13:
C
dv
= Icon + [v(0,t) v(t)],
dt
(1.81)
31
and
v(x,t)
2 v(x,t)
= v(x,t) + m2
+ rmVsyn g(x,t).
(1.82)
t
x2
Here u(t) = [v(0,t) v(t)] is the net current density flowing into the soma from
the dendritic cable at x = 0. (We are neglecting direct synaptic inputs into the soma.)
Current conservation implies the boundary condition
m
1 v
(0,t) = [v(0,t) v(t)].
r x
(1.83)
Since we can eliminate the term v(t) in (1.81) by shifting the linear term in v, it
follows that the total synaptic current into the soma is u(t) = v(0,t).
ai(t)
Iext,i
vi(t)
a(t)
i * ai
v(t)
Fig. 1.14 Schematic diagram of a feedforward network showing a set of afferent neurons labeled
by i synaptically driving another neuron
The inhomogeneous boundary problem for v(0,t) can be solved formally using
the Greens function G+ for the semi-infinite cable with a closed boundary:
t
v(0,t) = rmVsyn
G+ (0, x ,t t )g(x ,t )dx dt
t
G+ (0, 0,t t
)[v(0,t
) v(t
)]dt
(1.84)
This shows that the effective synaptic current u(t) flowing into the soma will itself
be affected by the cell firing an action potential, due to the dependence of v(0,t)
on the somatic potential v(t). However, suppose that the second term on the r.h.s. is
negligible compared to the first term arising from synaptic inputs. This approximation corresponds to imposing the homogeneous boundary condition v/ x(0,t) = 0.
It then follows that the total synaptic input into the soma is
t
u(t) = rmVsyn
G+ (0, x
,t t
)g(x
,t
)dx
dt
(1.85)
A similar analysis can also be carried out for more general dendritic topologies with
the soma coupled to one of the terminals of the tree. We conclude that under the
32
given approximations, the passive dendritic tree acts like a spatiotemporal linear
filter on incoming spike trains, whose properties are determined by the underlying
Greens function on the tree.
Recall from (1.74) that g(x,t) = (x) m gsyn (t T m (x)). Suppose that there exists a discrete set of synapses along the dendrite so that (x) = j (x x j ) and
T m (x j ) = T jm . Substituting into (1.85) then gives
u(t) = j (t T jm ) =
j m
where
j (t) = rmVsyn
and
t
0
t
j (t t
)a j (t
)dt
,
G+ (0, x j ,t )gsyn ( )d ,
a j (t) = (t T jm ).
(1.86)
(1.87)
(1.88)
Note that a j (t) represents the spike train arriving into the jth synapse in terms of
a sum of Dirac delta functions. Hence, under our various approximations, we can
view the total synaptic input u(t) as the sum of linearly filtered spike trains, with the
kernel of each filter determined by synaptic and dendritic processing. Now suppose
that each incoming spike train is associated with a distinct afferent neuron, so that
there is a one-to-one correspondence between synaptic and afferent neuron labels;
see Fig. 1.14. Then T jm = Tjm + j , where Tjm is the mth firing time of the jth afferent neuron and j is an axonal propagation time delay. Unless stated otherwise,
we will ignore axonal propagation delays and set T jm = T jm . Let us also assume that
each afferent neuron is described by a conductance-based model of the form
Ci
dvi
= Icon,i + Iext,i
dt
(1.89)
where vi is the somatic membrane potential of the ith afferent neuron, each of which
is driven by an external input Iext,i . If we associate with each neuron a firing threshold , then the spike times Tim are determined according to
Tim = inf{t,t > Tim1 |vi (t) = , vi (t) > 0}.
(1.90)
Finally, given u(t), the spike train a(t) of the output neuron is determined by (1.1)
and (1.9). In summary, the feedforward network involves a mapping {Iext,i (t)}
{Tim } {T m }.
1.4.3 Active Dendrites

It has been known for more than twenty years that the dendrites of cortical neurons do not simply act as passive electrical cables but also exhibit a variety of
33
active physiological processes [608]. For example, thick apical dendrites of pyramidal neurons express voltage-gated Na+ , K+ , and Ca2+ channels, which support
the back propagation of action potentials (APs) from the soma into the dendritic
tree [397, 606]; back-propagating APs are thought to play an important role in
spike-timing-dependent synaptic plasticity (STDP) [585]. In addition, sufficient local stimulation of active apical dendrites can initiate regenerative membrane depolarizations known as dendritic spikes [333, 555]. Some dendritic spikes are restricted
to the local initiation zone rather than invading the cell body and are thus well placed
to mediate the long-term potentiation of synaptic inputs in the absence of output
spiking of the neuron [226]. On the other hand, Ca2+ action potentials initiated in
apical dendrites can propagate towards the soma, which provides a mechanism for
actively amplifying the effects of distal synapses on AP generation in the cell body
[361]. Following advances in imaging techniques and methods of dendritic stimulation, Schiller et al. [556] established in vitro that active processes can also occur in
thin basal and apical dendritic branches of pyramidal neurons, where the majority
of synapses occur; see Fig. 1.15a. In particular, they found stimulus-evoked dendritic spikes whose major ionic component involved ligand-gated and voltage-gated
N-methyl-D-aspartate receptor (NMDAR) channels; see also [362, 400, 520] and
the review [12]. When glutamate binds to an NMDAR, it modifies the voltage sensitivity of the corresponding ion-channel current, which develops a negative slope
conductance due to removal of a magnesium block [414, 466]. This means that in
the presence of high levels of glutamate, the currentvoltage (IV) characteristics of
an NMDAR channel are very similar to the voltage-gated Na channel. Hence, during
thin apical
tufts
thin oblique
tufts
b
Na+ spikelet
plateau potential
thick apical
branch
NMDA spike
20 mV
subthreshold EPSP
50 ms
axon
thin basal
branches
Fig. 1.15 (a) Schematic illustration of a pyramidal neuron showing the thick apical dendrite and
various thin dendrites. The latter support the initiation of dendritic NMDA spikes. (b) Typical
waveform of a dendritic NMDA spike. Weak glutamatergic inputs generate EPSP-like (subthreshold) depolarizations. A stronger input can trigger a dendritic plateau potential, consisting of a rapid
onset that is often associated with a Na spikelet, a long-lasting plateau phase that can have a duration of several hundred ms, and a sudden collapse at the end of the plateau phase. The plateau
potential consists of several dendritic conductances, the most predominant being due to NDMAR
channels. Pharmacologically blocking Na and Ca2+ channels reveals the pure dendritic NMDA
spike [556]
34
strong stimulation of a thin dendrite due to the local uncaging of glutamate or high
frequency stimulation of a cluster of synapses, the NMDARs can fire a regenerative
dendritic spike, just as Na channels support the initiation of an action potential following membrane depolarization. However, the duration of the dendritic spike is of
the order 100 ms rather than 1 ms; see Fig. 1.15b. Finally, active processes can also
be found in dendritic spines, which can support the propagation of saltatory waves;
see Sect. 3.1
For relatively small deviations of the membrane potential from some constant
value, a linearization of the channel kinetics can be performed. The resulting system has a membrane impedance that displays resonant-like behavior due to the
additional presence of effective inductances [61, 141, 341, 342]. We sketch how
inductance-like behavior can arise from active neural membrane by considering a
generic ion current of the form I(v, x1 , . . . , xM ), where v is membrane voltage and xk
are gating variables that satisfy
k (v)
dxk
= xk, (v) xk ,
dt
k = 1, . . . , M.
(1.91)
It is convenient to set k (v) = (k (v) + k (v))1 and xk, (v) = k (v)k (v). Linearizing around a fixed point z = z with vectors defined by z = (v, x1 , . . . , xM )T and
z = (v , x1, (v ), . . . , xM, (v )), we have

M
v
I
+
I =
x ,
(1.92)

k
R
k=1 xk
z=z
where R is an effective resistance such that R1 = I/ V |z=z . From (1.91) it follows that

d k
d
d[k + k ]
+ k + k xk =
xk,
Vk .
(1.93)
dt
dV
dV
Combining (1.92) and (1.93) we arrive at the following equation for the first-order
variation of the current:
M
V
+ Ik ,
I =
(1.94)
R
k=1

d
Ik = V,
rk + Lk
dt
where
and
rk1
I
= k
xk

d k
d[k + k ]
xk,

dV
dV
z=z
Lk = k rk
(1.95)
(1.96)
(1.97)
35
Hence, for a small perturbation around the steady state, the current I responds as
though the resistance R is in parallel with M impedance lines, each of which is
a resistance rk that is itself in series with an inductance Lk (see Fig. 1.16). Such
inductive terms account for the oscillatory overshoot commonly seen in response to
depolarising current steps or even after the firing of an action potential. This form
of equivalent linear membrane circuit is typically called quasi-active in order to
distinguish it from a truly active (i.e. nonlinear) membrane [341].
rh
rn
rm
R
Ln
Cm
Lh
Fig. 1.16 Electrical LRC circuit representing the linearized response of the HodgkinHuxley
equations
Quasilinear membrane can be incorporated into the cable equation (1.55) by introducing the space-dependent voltage v(x,t) and currents Ik (x,t), x R, such that
V
V
2V
1
= + Dm 2
t
m
x
Cm
Lk
Ik Iext
(1.98a)
dIk
= rk Ik + V.
dt
(1.98b)
Here Dm = m2 /m . Laplace transforming (1.98) with v(x, 0) = 0, Ik (x, 0) = 0 yields

the ODE
dV 2
(1.99)
2 + 2 ( )V = I,
dx
) = Iext (x, )/Cm , and
with V = V (x, ), I = I(x,

1
1
1
1
( ) =
+ +
.
Dm m
Cm
k rk + Lk
2
(1.100)
It follows that in Laplace space,

V (x, ) =

0
y, )I(y,
)dy,
G(x
( )|x|
) = e
,
G(x,
2Dm ( )
(1.101)
36
where G is the Laplace transform of the Greens function on an infinite quasi-active

cable. Having obtained G, it is then possible to extend Greens function methods
outlined for passive branching cables to the quasi-active case [141].
1.5 Stochastic Ion Channels

In the standard conductance-based model of a neuron (1.6), it is assumed that the
number of voltage-gated ion channels is sufficiently large so that one can represent
the opening and closing of the channels in terms of deterministic kinetic equations.
These keep track of the fraction of open and closed channels as a function of time.
However, the opening and closing of a single channel is a stochastic process. This
then raises the important issue of how such stochasticity manifests itself when there
are relatively few ion channels. (A similar issue applies to synapses with a small
number of receptor-mediated ion channels.) In this section, we develop the theory
of stochastic ion channels in some detail, since it provides an excellent platform
for introducing various methods and ideas in stochastic processes that will be used
throughout the book. For reviews on noise in ion channels see [227, 588, 670].
1.5.1 Ensemble of Two-State Ion Channels

First, consider a single ion channel that can exist either in a closed state (C) or an
open state (O). Transitions between the two states are governed by a continuoustime jump Markov process
(v)
C(closed) O(open).
(v)
(1.102)
with voltage-dependent transition rates (v), (v). For the moment, we assume that
v is fixed. In order to understand what such a process means, let Z(t) be a discrete
random variable taking values Z {C, O} and set Pz (t) = Prob [Z(t) = z]. From
conservation of probability,
PC (t) + PO (t) = 1.
The transition rates then determine the probability of jumping from one state to the
other in a small interval t:
t = Prob [Z(t + t) = O|Z(t) = C],
t = Prob [Z(t + t) = C|Z(t) = O].
It follows that there are two possible ways for the ion channel to enter or leave the
closed state:
PC (t + t) = PC (t) PC (t) t + PO(t) t.
37
Writing down a similar equation for the open state, dividing by t, and taking the
limit t 0 leads to the pair of equations
dPC
= PC + PO
dt
dPO
= PC PO .
dt
(1.103a)
(1.103b)
Now suppose that there are N identical, independent two-state ion channels. In the
limit N we can reinterpret PC and PO as the mean fraction of closed and
open ion channels within the population, and fluctuations can be neglected. After setting PO = X and PC = 1 X, we recover the kinetic equation (1.4). (An
identical argument can be applied to the kinetic model of a synapse considered
in Sect. 1.3.)
In order to take into account fluctuations in the case of finite N, it is necessary
to keep track of the probability P(n,t) that there are n open channels at time t,
0 n N. (If there are n open channels, then it immediately follows that there
are N n closed channels, so we do not need to keep track of the latter as well.)
Consider a time interval [t,t + t] with t sufficiently small so that only one channel
has a significant probability of making a C O or O C transition. There are four
possible events that can influence P(n,t) during this interval, two of which involve
transitions into the state of n open ion channels, and two of which involve transitions
out of the state. Collecting these terms and taking the limit t 0 leads to the
master equation
d
P(n,t) = (N n + 1)P(n 1,t) + (n + 1)P(n + 1,t) [ (N n) + n]P(n,t).
dt
(1.104)
The first term on the right-hand side represents the probability flux that one of
N (n 1) closed channels undergoes the transition C O, whereas the second
term represents the probability flux that one of n + 1 open channels undergoes the
transition O C. The last two terms represent transitions n n 1. Define the
mean number of open channels at time t by
n(t) =
nP(n,t).
n=0
By differentiating both sides of this equation with respect to t and using the master
equation (1.104), it can be shown that in the limit of large N (where the upper limit
in the sum can be taken to be ) we recover the kinetic equation (1.4) with X = n/N.
The steady-state solution Ps (n) of the master equation (1.104) satisfies J(n) =
J(n + 1) with
J(n) = (n)Ps (n) + (n 1)Ps(n 1),
38
and
+ (n) = (N n) ,
(n) = n .
Using the fact that n is a nonnegative integer, that is, Ps (n) = 0 for n < 0, it follows
that J(n) = 0 for all n. Hence, by iteration,

n
n
+ (m 1)
N!
= Ps (0)
.
(1.105)
Ps (n) = Ps (0)
(m)
n!(N
n)!
m=1
Taking logs of both sides of this equation and using Stirlings formula log(n!)
n log n n it can be shown that for large n, N, Ps (n) ps (x) where x = n/N,
ps (x) = N eN (x)
(1.106)
with normalization factor N and (x) is the effective potential
(x) = x log( / ) + x log(x) + (1 x) log(1 x).
(1.107)
Let x be the unique critical point of the effective potential (x), that is,
(x ) = 0.
Note that x coincides with the fixed point of the corresponding deterministic kinetic
equations,
.
(1.108)
x =
+
Since N is large, we can make the Gaussian approximation

ps (x) p(0) exp N (x ) N
(x )(x x )2 /2 .
(1.109)
Under this approximation, the mean and variance of the fraction of open channels
are given by
n
(n n)
2 x (1 x)
= x =
.
(1.110)
,
=
2
N
+
N
N
It is clear that fluctuations become negligible in the large-N limit.
1.5.2 Diffusion Approximation

A useful diffusion approximation of the master equation (1.104) for large but finite
N can be obtained by carrying out a KramersMoyal or system-size expansion to
second order in N 1 [210, 309], which was originally applied to ion-channel models
by Fox and Lu [200]. This yields a FokkerPlanck (FP) equation describing the evolution of the probability density of a corresponding continuous stochastic process
that is the solution to a stochastic differential equation (SDE), which in the physics
literature is often called a Langevin equation; we will take these terms to be interchangeable. Further details concerning SDEs can be found in appendix section 1.7
and in various references [210, 476]. Moreover, a rigorous analysis of the diffusion
approximation and its relationship to the system-size expansion has been carried out
by Kurtz [352].
39
First, introduce the rescaled variable x = n/N and transition rates N (x) =
(Nx). Equation (1.104) can then be rewritten in the form
d p(x,t)
= N[+ (x 1/N)p(x 1/N,t) + (x + 1/N)p(x + 1/N,t)
dt
(+ (x) + (x))p(x,t)].
Treating x, 0 x 1, as a continuous variable and Taylor expanding terms on the
right-hand side to second order in N 1 leads to the FP equation
1 2
p(x,t)
= [A(x)p(x,t)] +
[B(x)p(x,t)]
t
x
2N x2
(1.111)
A(x) = + (x) (x) ( + )x,
(1.112a)
B(x) = + (x) + (x) + ( )x.
(1.112b)
with
The FP equation takes the form of a conservation equation
p
J
= ,
t
x
(1.113)
where J(x,t) is the probability flux,

J(x,t) =
1
[B(x)p(x,t)] + A(x)p(x,t).
2N x
(1.114)
The FP equation is supplemented by the no-flux or reflecting boundary conditions

at the ends x = 0, 1 and a normalization condition,
J(0,t) = J(1,t) = 0,
1
0
p(x,t)dx = 1.
(1.115)
The FP equation has a unique steady-state solution obtained by setting J(x,t) = 0

for all 0 x 1. The resulting first-order ODE can be solved to give a steady-state
probability density of the form (1.107), with corresponding potential
FP (x) = 2
x
A(x
)
B(x
)
dx
= 2
x
+ (x
) (x
)
+ (x
) + (x
)
dx
.
(1.116)
The mean and variance of the fraction of open channels close to the fixed point x
can again be determined by carrying out a Gaussian approximation, and the results
agree with those obtained from the steady-state solution of the master equation.
An alternative way of calculating the mean and variance is to note that the solution
to the FP equation (1.111) determines the probability density function for a corresponding stochastic process X(t), which evolves according to the SDE or Langevin
equation [210]
40
1
dX = A(X)dt + b(X)dW (t).
N
(1.117)
with b(x)2 = B(x). Here W (t) denotes a Wiener process with dW (t) distributed
according to a Gaussian process with mean and covariance
dW (t) = 0,
dW (t)dW (s) = (t s)dtds.
(1.118)
Note that the noise term in (1.117) is multiplicative, since it depends on the current state X(t). It is well known that there is an ambiguity in how one integrates
multiplicative noise terms, which relates to the issue of Ito versus Stratonovich
versions of stochastic calculus [210]; see appendix section1.7. However, for this
particular example, based on the reduction of a master equation, the explicit form of
the corresponding FP equation (1.111) ensures that the noise should be interpreted
in the sense of Ito.
Thus, one can view the SDE as describing a stochastic path in phase space that
involves a deterministic trajectory converging to the unique

stable fixed point x
thatis perturbed by Gaussian-like fluctuations of order 1/ N. Substituting X x =
Y/ N
into the SDE equation (1.117) and formally Taylor expanding to lowest order
in 1/ N yields the so-called linear noise approximation
dY = kY dt + b(x)dW (t),
with
k A (x ) = + ,

b(x ) = B(x ) =
(1.119)

2
.
+
This takes the form of an OrnsteinUhlenbeck equation [210], which can be

solved as
t
Y (t) = ekt Y0 + b(x ) ek(tt ) dW (t

),
(1.120)
0
given the initial condition Y (0) = Y0 . It follows that the mean and covariance
of Y (t) are
Y (t) = ekt Y0 ,
(1.121)
and
cov(Y (t),Y (s)) [Y (t) Y (t)][Y (s) Y (s)]
(1.122)

t

t
= b(x )2
ek(tt ) dW (t
)
ek(st ) dW (t
)
0
= b(x )2 ek(t+s)
Here
s
0
e2kt dt
=
b(x )2
2k
b(x )2
=
= x (1 x),
2k
( + )2

ek(ts) 1 e2ks .
41
and, without loss of generality, we have assumed that t s. Thus, in the stationary
limit t ,
b(x )2 k|ts|
e
Y (t) 0, cov(Y (t),Y (s))
.
2k
Since Y (t)/ N = X(t) x, we recover the results of (1.110).

Note that (1.116) differs from the effective potential (1.107) obtained directly
from the master equation for large N, given that the latter can be rewritten in the
form
x
(x
)
dx .
(x) =
ln
(1.123)
+ (x
)
Although, this discrepancy is not much of an issue when the underlying kinetic
equations have a unique fixed point, it can lead to exponentially large errors when
there are multiple stable fixed points. The Langevin or diffusion approximation still
accounts for the effects of fluctuations well within the basin of attraction of a locally stable fixed point. However, there is now a small probability that there is a
noise-induced transition to the basin of attraction of another fixed point. Since the
probability of such a transition is usually of order e N with = O(1), except close
to the boundary of the basin of attraction, such a contribution cannot be analyzed
accurately using standard FokkerPlanck methods [309]. These exponentially small
transitions play a crucial role in allowing the network to approach the unique stationary state (if it exists) in the asymptotic limit t . In other words, for multistable
systems, the limits t and N do not commute [24, 256, 641]. One example
where multistability is important is when considering the effects of stochastic ion
channels on membrane voltage fluctuations (see below).
1.5.3 Membrane Voltage Fluctuations

Let us now return to a conductance-based model of a neuron, in which the stochastic
opening of ion channels generates a stochastic ionic current that drives the membrane voltage. We are particularly interested in how fluctuations affect the initiation
of an action potential due to the opening of a finite number of Na+ channels. Therefore, we imagine freezing the slow K+ channels, so that they effectively act as a
leak current and simplify the sodium channels by treating each as a single activating
subunit. The stochastic membrane voltage then evolves according to the piecewise
deterministic equation
1
dV
= F(V, n) f (V )n(t) g(V),
dt
N
(1.124)
where f (v) = gNa (VNa v) represents the gated sodium currents, g(v) = geff [Veff
v] I represents the sum of effective leakage currents and external inputs I, and
n(t) is the number of open sodium channels. Note that (1.124) only holds between jumps in the number of open ion channels, with the latter described by the
42
master equation (1.104). The stochastic process defined by (1.104) and (1.124) is
an example of a so-called stochastic hybrid system with piecewise deterministic dynamics. There has been a lot of recent interest in such systems, particularly within
the context of conductance-based models [88, 321, 484, 654]. The associated probability density p(v, n,t), which is defined according to
p(v, n,t)dv = Prob[n(t) = n, v V (t) v + dv],
given an initial condition V (0) = V0 , n(0) = n0 , satisfies the differential Chapman
Kolmogorov (CK) equation

p
n
=
f (v) g(v) p + + (v, n 1)p(v, n 1,t)
(1.125)
t
v N
+ (v, n + 1)p(v, n + 1,t) [+(v, n) + (v, n)]p(v, n,t),
with
+ (v, n) = (v)(N n),
(v, n) = (v)n.
(1.126)
Note that the right-hand side of (1.124) is negative for large V and positive for small
V , which implies that the voltage V is confined to some bounded domain [V1 ,V2 ].
In order to investigate action potential initiation, we will assume that N is sufficiently large so that we can approximate the jump Markov process for the ion channels by a continuous Markov process using a diffusion approximation, and (ii) we
assume that the transitions between different discrete states is much faster than the
voltage dynamics so we can assume that, for fixed v, the number of open ion channels is close to the quasi-equilibrium x (v) = (v)/( (v)+ (v)). This limiting case
was originally considered by Chow and White [116]. Under these approximations,
the voltage dynamics is described by an SDE of the form [see (1.119)]

dV = f (V )(x + Y (t)/ N) g(V) dt, dY = kY dt + b(x)dW (t). (1.127)

Thus the stochastic voltage is coupled to a fast OrnsteinUhlenbeck or colored noise
process Y (t). If we make the further assumption that the latter is in quasi-equilibrium
for a given V (fast ion channels), Y (t)dt k1 b(x )dW (t), then we obtain a scalar
SDE for the voltage:
1
dV = [ f (V )x (V ) g(V )] dt + (V ) f (V )dW (t),
N
where
1
b(x (V ))
=
(V ) =
k(V )
(V ) + (V )
2 (V ) (V )
.
(V ) + (V )
(1.128)
(1.129)
Taking , = O(1/ ) for some dimensionless parameter 0 < 1, we see that

(V ) = O( 1/2 ). In deriving (1.128), we have effectively taken a zero correlation
limit of a colored noise process. It can be shown that the multiplicative noise term
should be interpreted in the sense of Stratonovich, and the associated FP equation is
given by [210, 321]
43
1
p(v,t)
= [A (v)p(v,t)] +
B(v) p(v,t) ,
t
v
N v
v

(1.130)
with
A (v) = f (v)x (v) g(v),
B(v) = [ (v) f (v)]2 /2.
We have ignored an O( ) contribution to the drift term of the form B

(v)/N. The
FP equation is supplemented by reflecting boundary conditions at v = V1 ,V2 :
J(V1 ,t) = J(V2 ,t) = 0,
with
J(v,t) = A (v)p(v,t)
B(v)
p(v,t).
N v
(1.131)
(1.132)
1.5.4 First Passage Time Problem

A key property that one would like to calculate is the mean time to fire an action
potential (MFPT) as a function of the stimulus current I. In the absence of noise, the
system evolves according to the deterministic equation
d 0 (v)
(v)
dv
= A (v) =
f (v) g(v)
,
dt
(v) + (v)
dv
(1.133)
where 0 (v) is a deterministic potential. In Fig. 1.17, we plot 0 (v) as a function of

v for various values of the external input current and the particular transition rates

2(v v1)
(v) = exp
, = constant.
v2
The minima and maxima of the potential correspond to stable and unstable fixed
points of the deterministic dynamics, respectively. It can be seen that below a threshold current I , I < I , there exist two stable fixed points v (minima) separated by
an unstable fixed point at v0 (maximum), that is, the system exhibits bistability. The
left-hand fixed point represents the resting state, whereas the right-hand fixed point
corresponds to an excited state. Thus, in the bistable regime the deterministic system requires an external perturbation in order to generate an action potential starting
from the resting state. On the other hand, for the stochastic system it is possible that
fluctuations in the opening and closing of Na+ ion channels induce a transition from
the resting state to the excited state by crossing over the potential hill at v0 . This is
directly analogous to a diffusing particle escaping from the left to the right well in
a double well potential, which is a classical example of a first passage time (FPT)
problem in physics [210, 257]. (Of course, once such an event occurs, one has to take
into account the K + dynamics in order to incorporate the effects of repolarization
44
that return the system to the resting state. If one includes the slow opening and closing of these channels, then the underlying deterministic system becomes excitable
rather than bistable; see Sect. 2.1. For the moment, we will assume that this does not
significantly affect the noise-induced initiation of an action potential. It turns out
that such an assumption breaks down if fluctuations in the opening and closing of
K+ channels become significant [456].)
We now outline the basic calculation of the mean time to escape from the resting
state using the diffusion approximation. Since the voltage will rapidly approach the
excited state v+ once it has passed the maximum at v0 , the major contribution to the
escape time will be due to the fluctuation-driven transition from v to v0 . We can
model this process by replacing the reflecting boundary condition at v = V2 with an
absorbing boundary condition at v = v0 < V2 :
p(v0 ,t) = 0.
v0
v-
I < I*
I = I*
0(v)
v+
-100 -80 -60 -40 -20 0 20 40 60 80 100

v [mV]
Fig. 1.17 Plot of deterministic potential 0 (v) as a function of voltage v for different values of
the external stimulus current I. Parameter values are N = 10, vNa = 120 mV, veff = 62.3 mV,
gNa = 4.4 mS/cm2 , geff = 2.2 mS/cm2 , = 0.8 s1 , and (v) = exp[(v + 1.2)/9]
We also shift the voltage v so that the left-hand boundary is at V1 = 0. Let T (v)
denote the stochastic time for the particle to exit the right-hand boundary at v0 ,
given that it starts at location v [0, v0 ] at time t = 0. As a first step, we introduce
the survival probability P(v,t) that the particle has not yet exited at time t:
P(v,t) =
v0
p(x,t|v, 0)dx.
(1.134)
It follows that Prob[T (v) t] = 1 P(v,t) and we can define the FPT density according to
P(v,t)
.
(1.135)
f (v,t) =
t
45
It can be shown from (1.130) and the Markovian nature of the stochastic process
that the FPT density satisfies a backward FP equation of the form [210]

P(v,t)
P(v,t)
= A (v)
+
B(v) P(v,t) ,
t
v
v
v
(1.136)
where we have absorbed the factor of 1/N into B.

A quantity of particular interest is the mean first passage time (MFPT) (v) defined according to
(v) = T (v)

0
f (v,t)tdt
(1.137)

P(v,t)
dt =

0
P(v,t)dt,
after integration by parts. Hence, integrating both sides of (1.136) shows that the
MFPT satisfies the ODE

d2
d (v)
[A (v) + B (v)]
+ B(v) 2 (v) = 1.
(1.138)
dv
dv
Equation (1.138) is supplemented by reflecting and absorbing boundary conditions
for the backward FP equation:

(0) = 0,
(v0 ) = 0.
(1.139)
It is straightforward to solve (1.138) by direct integration [210]. First, introducing

an integration factor and integrating once gives
(v)
(v) =
where
(v) =
v (v
)
e
0
B(v
)
dv
,
v
A (v
) + B
(v
)
B(v
)
dv
.
(1.140)
and we have used the boundary condition

(0) = 0. Integrating once more with
respect to v and using (v0 ) = 0 then gives
(v) =
v0
v
e (v ) dv
v
(v
)
e
0
B(v
dv
(1.141)
There is now a standard procedure for approximating this double integral based on
Kramers reaction rate theory [210, 257]. We simply quote the result here: (v ) =
1/ where

v
0 A (v)
B(v0 ) A
(v ) A
(v0 )
exp
dv.
.
(1.142)
B(v ) B(v0 )
v B(v)
46
MFPT (ms)
1010
diffusion approximation
I < I*
I > I*
105
exact
1
50
applied current I (mA)
100
Fig. 1.18 Schematic diagram comparing MFPT calculated using the diffusion approximation with
the MFPT of the full system. The scales of the axes are based on numerical results carried out in
[321] for N = 10. Other parameter values as in Fig. 1.17
Keener and Newby [321] calculated the MFPT using the diffusion approximation
and compared it with Monte Carlo simulations of the full stochastic model whose
probability density evolves according to the CK equation (1.126). A summary of
their findings is shown schematically in Fig. 1.18. The main observation is that although the Gaussian-like diffusion approximation does well in the superthreshold
regime (I > I ), it deviates significantly from the full model results in the subthreshold regime (I < I ), where it overestimates the mean time to spike. This is related
to the fact that the effective potential of the steady-state density under the diffusion
approximation generates exponentially large errors in the MFPT, as discussed at the
end of Sect. 1.5.2. Finally, note that the model of stochastic ion channels and membrane voltage fluctuations presented above is based on a number of simplifications.
First, only the initiation of an action potential was considered, which involved the
opening of sodium channels, while the termination of the action potential due to
Na+ inactivation and K + activation was ignored. Indeed, in stochastic versions of
the HodgkinHuxley equations spontaneous activity can be observed due to fluctuations in K + ion channels [116, 456]. That is, if some K + channels spontaneously
close, then a regenerative Na+ current can induce an action potential. The second
simplification was to treat each ion channel as a single unit rather than as a cluster
of subunits. In other words, the Markov chain of events associated with opening and
closing of an ion channel involves transitions between more than two internal states.
1.6 Stochastic Hybrid System with Fast Kinetics

As shown by Keener and Newby [321], it is possible to obtain a much more accurate
analytical expression for the MFPT in the subthreshold regime by analyzing the CK
equation (1.126) in the limit of fast sodium channels. This analysis applies equally
47
well to small and large N. First, it is convenient to rewrite the CK equation in a more
compact and general form:
p
[F(v, n)p(v, n,t)] 1 N
=
+ A(n, m; v)p(v, m,t)
t
v
m=0
(1.143)
with F given by (1.124) and A a tridiagonal matrix (for fixed v): A(n, n 1; v) =
+ (v, n 1), A(n, n; v) = + (v, n) (n), A(n, n + 1; v) = (n + 1). We are
now making explicit the condition that the open and closing of Na+ channels occurs
on a much faster time scale than the voltage dynamics by scaling the transition rates
according to the small parameter 1. In the limit 0, (1.124) reduces to the
deterministic or mean field (1.133) with A (v) = n F(v, n) (v, n), where (v, n) is
the unique steady-state density satisfying m A(n, m; v) (v, m) = 0 (see (1.105)):
(v, n) =
N!
a(v)n b(v)Nn ,
(N n)!n!
a(v) =
(v)
, b(v) =
. (1.144)
(v) +
(v) +
The mean number of open channels is n = Nn=1 n (v, n) = Na(v). In the regime
0 < 1, for which there are typically a large number of transitions between
different channel states n while the voltage v hardly changes at all. This suggests
that the system rapidly converges to the (quasi-)steady-state (v, n), which will then
be perturbed as v slowly evolves. The resulting perturbations can be analyzed using
a quasi-steady-state (QSS) diffusion or adiabatic approximation [210, 454, 487], in
which the CK equation (1.143) is approximated by a FokkerPlanck (FP) equation;
see also Sect. 6.4
The QSS approximation was first developed from a probabilistic perspective by
Papanicolaou [487]; see also [210]. It has subsequently been applied to a wide range
of problems in biology, including cell movement [274, 481], wavelike behavior in
models of slow axonal transport [206, 207, 518], and molecular motor-based models
of random intermittent search [453, 454]. The basic idea of the QSS reduction is to
decompose the probability density as
p(v, n,t) = C(v,t) (v, n) + w(v, n,t),
(1.145)
where n p(v, n,t) = C(v,t) and n w(v, n,t) = 0. Carrying out an asymptotic expansion in , it can be shown that C evolves according to the FP equation [81, 453, 454]

C
= (A C) +
D
,
(1.146)
t
v
v
v
with the drift term given by (1.133), and diffusion coefficient
D(v) =
Z(v, n)F(v, n),
n=0
where Z(v, n) is the unique solution to
(1.147)
48
A(n, m; v)Z(v, m) = [A (v) F(v, n)] (v, n)
(1.148)
with m Z(v, m) = 0. As in the case of the diffusion approximation for large N, the
FP equation (1.146) captures the Gaussian-like fluctuations within the basin of attraction of a fixed point of the mean field (1.133), but breaks down when considering
rare event transitions between metastable states since it yields exponentially large errors in the escape rates. Therefore, one has to deal with the full CK equation (1.143).
1.6.1 First Passage Time Problem and the Projection Method

In order to revisit the FPT problem considered in Sect. 1.5.4, we supplement (1.143)
with the following absorbing boundary conditions at v0 :
p(v0 , n,t) = 0, for all n = {0, . . . , k 1}.
(1.149)
Here, denotes the set of integers for which F(v0 , n) < 0. The initial condition is
taken to be
(1.150)
p(v, n, 0) = (v v )n,n0 .
Let T denote the (stochastic) FPT for which the system first reaches v0 , given that it
started at v . The distribution of FPTs is related to the survival probability that the
system has not yet reached v0 :
S(t)
v0 N
p(v, n,t)dv.
(1.151)
n=0
That is, Prob{t > T } = S(t) and the FPT density is

f (t) =
dS
=
dt
v0 N
p
0
n=0
(v, n,t)dv.
Substituting for p/ t using the CK equation (1.143) shows that

v0 N
N
[F(v, n)p(v, n,t)]
dv = p(v0 , n,t)F(v0 , n).
f (t) =
v
0
n=0
n=0
(1.152)
(1.153)
We have used n A(n, m; v) = 0 and limv F(v, n)p(v, n,t) = 0. The FPT density
can thus be interpreted as the probability flux J(v,t) at the absorbing boundary, since
we have the conservation law
p(v, n,t)
J(v,t)
=
,
t
v
n=0
N
J(v,t) =
F(v, n)p(v, n,t).
n=0
(1.154)
49
We wish to solve the FPT problem in the weak noise limit 1. One of the characteristic features of the weak noise limit is that the flux through the absorbing boundary and the inverse of the MFPT T are exponentially small, that is, T eC/
for some constant C. This means that standard singular perturbation theory cannot
be used to solve the resulting boundary value problem, in which one matches inner and outer solutions of a boundary layer around the point v = v0 . Instead, one
proceeds by finding a quasistationary solution using a WentzelKramersBrillouin
(WKB) approximation. Recently, this approach has been extended by Keener and
Newby [321] to a CK equation of the form (1.143), using a so-called projection
method [660].
In order to apply the projection method, it is necessary to assume certain properties of the non-self-adjoint linear operator L on the right-hand side of (1.143)
with respect to the Hilbert space of functions h : [0, v0 ] {0, . . ., N} R with inner
product defined according to
h, g =
v0 N
0
h(v, n)g(v, n)dv.
(1.155)
n=0
1. L has a complete set of eigenfunctions r with

d
1 N
L r (v, n) (F(v, n)r (v, n)) A(n, m; v)r (v, m) = r r (v, n), (1.156)
dv
m=0
together with the boundary conditions
r (v0 , n) = 0, for n = 0, . . . , k 1.
(1.157)
2. The real part of each eigenvalue r is positive definite and the smallest eigenvalue
0 is real and simple. Thus we can introduce the ordering 0 < 0 < Re[1 ]
Re[2 ] . . ..
3. 0 is exponentially small, 0 eC/ , whereas Re[r ] = O(1) for r 1. In particular, lim 0 0 = 0 and lim 0 0 (v, n) = (v, n).
Under the above assumptions, we can introduce the eigenfunction expansion
p(v, n,t) =
Cr ert r (v, n), .
(1.158)
r=0
with 0 Re[r ] for all r 1. Thus, at large times we have the quasistationary
approximation
(1.159)
p(v, n,t) C0 e0t 0 (v, n).
Substituting such an approximation into (1.153) implies that
f (t) e0t
0 (v0 , n)F(v0 , n),
n=0
1t 1.
(1.160)
50
Equation (1.156) implies that

N
v0
n=0 0

L0 (v, n)dv
F(v0 , n)0 (v0 , n)
n=0
= 0
In other words,
0 =
v0
n=0 0
0 (v, n)dv.
Nn=0 F(v0 , n)0 (v0 , n)

.
1, 0
(1.161)
Combining (1.161) and the quasistationary approximation (1.160) shows that the
(normalized) FPT density reduces to
f (t) 0 e0t
(1.162)
and, hence, T = 0 t f (t)dt 1/0.

It remains to obtain an approximation of the principal eigenfunction 0 , which
can be achieved using the WKB method as described in Sect. 1.6.2. This yields a
quasistationary density that approximates 0 up to exponentially small terms at the
boundary, that is,
(1.163)
L = 0, (u , n) = O(eC/ ).
In order to express 0 in terms of the quasistationary density , we consider the
eigenfunctions of the adjoint operator, which satisfy the equation
d r (v, n) 1
A(m, n; v)r (v, m) = r r (v, n),
L r (v, n) F(v, n)
dv
m
(1.164)
and the boundary conditions
r (v0 , n) = 0,
n k.
(1.165)
The two sets of eigenfunctions {r } and {r } form a biorthogonal set with respect
to the underlying inner product,
r , s = r,s .
(1.166)
, L 0 = 0 , 0 .
(1.167)
Now consider the identity
Integrating by parts the left-hand side of (1.164) picks up a boundary term so that
0 =
Nn=0 (v0 , n)0 (v0 , n)F(v0 , n)

.
, 0
(1.168)
51
The calculation of the principal eigenvalue 0 thus reduces to the problem of determining the quasistationary density and the adjoint eigenfunction 0 .
1.6.2 The WKB Method and the Quasistationary Density

We now show how the WKB method [160, 256, 398, 445, 559] can be used to
compute the quasistationary density . We seek a solution of the form

(v)
(v, n) R(v, n) exp
,
(1.169)
where (v) is a scalar potential. Substituting into L = 0 gives

dF(v, n)R(v, n)
, (1.170)
A(n, m; v) +
(v)n,m F(v, m) R(v, m) =
dx
m=0
N
where
= d /dx. Introducing the asymptotic expansions R R(0) + R(1) and
0 + 1 , the leading-order equation is
N
A(n, m; v)R(0) (v, m) = 0

(v)F(v, n)R(0)(v, n).
(1.171)
m=0
(Note that since F(v, n) is nonzero almost everywhere for v < v0 , we can identify
m; v) = A(n, m; v)/F(v, n)
0
and R(0) as an eigenpair of the matrix operator A(n,
for fixed v.) Positivity of the probability density requires positivity of the corresponding solution R(0) . One positive solution is R(0) = , for which 0
= 0. However, such a solution is not admissible since 0 = constant. It can be proven using
linear algebra that if F(v, n) for fixed v < v0 changes sign as n increases from zero,
then there exists one other positive solution, which also has the appropriate functional form to describe the potential well. That is, 0
(x) has the correct sign and
vanishes at the fixed points. Hence, it can be identified as the appropriate WKB
solution.
Proceeding to the next order in the asymptotic expansion of (1.170), we have
N

A(n, m; v) + 0
(v)n,m F(v, m) R(1) (v, m)
m=0
dF(v, n)R(0) (v, n)

1
(v)F(v, n)R(0) (v, n).
dx
(1.172)
m; v) = A(n, m; v) +
For fixed v and WKB potential 0 , the matrix operator A(n,
0 (v)n,m F(v, m) on the left-hand side of this equation has a one-dimensional null
52
space spanned by the positive WKB solution R(0) . The Fredholm alternative theorem1
then implies that the right-hand side of (1.172) is orthogonal to the left null vector
That is, we have the solvability condition
S of A.

N
dF(v, n)R(0) (v, n)
(0)
1 (v)F(v, n)R (v, n) = 0,
(1.173)
S(v, n)
dv
n=0
with S satisfying
N
S(v, n)

A(n, m; v) + 0
(v)n,m F(v, m) = 0.
(1.174)
n=0
Given R(0) , S and 0 , the solvability condition yields the following equation for 1 :
1
(x) =
Nn=0 S(v, n)[F(v, n)R(0) (v, n)]
.
Nn=0 S(v, n)F(v, n)R(0) (v, n)
Combining the various results, and defining

v
k(v) = exp
1 (y)dy ,
(1.175)
(1.176)
gives to leading order in ,

0 (v)
(v, n) N k(v) exp
R(0) (v, n),
(1.177)
where we choose n R(0) (v, n) = 1 for all v and N is the normalization factor,
v

0
0 (v) 1
N =
k(v) exp
.
(1.178)
0
The latter can be approximated using Laplaces method to give

|0
(v )|
0 (v )
1
exp
N
.
k(v )
2
(1.179)
Consider an M-dimensional linear inhomogeneous system Ax = b with x, b RM . Suppose that

the M M matrix A has a nontrivial null space and let v be a null vector of the adjoint matrix A ,
that is, A v = 0. The Fredholm alternative theorem states that the inhomogeneous equation has a
(nonunique) solution if and only if v b = 0 for all null vectors v.
53
1.6.3 Calculation of the Principal Eigenvalue

In order to evaluate the principal eigenvalue 0 satisfying (1.168), it is necessary to
determine the adjoint eigenfunction 0 . Following [321, 455], this can be achieved
using singular perturbation methods. Since 0 is exponentially small in , (1.164)
yields the leading-order equation
F(v, n)
N
d 0 (v, n)
+ A(m, n; v)0 (v, m) = 0,
dx
m=0
(1.180)
supplemented by the absorbing boundary condition
0 (v0 , n) = 0,
n k.
(1.181)
A first attempt at obtaining an approximate solution that also satisfies the boundary
conditions is to construct a boundary layer in a neighborhood of the unstable fixed
point v0 by performing the change of variables v = v0 z and setting Q(z, n) =
0 (v0 z). Equation (1.180) then becomes
F(v0 , n)
N
dQ(z, n)
+ A(m, n; v0 )Q(z, m) = 0.
dz
m=0
(1.182)
This inner solution has to be matched with the outer solution 0 = 1, which means
that
lim Q(z, n) = 1
(1.183)
z
for all n. Consider the eigenvalue equation

N
(A(n, m; v) r (v)n,m F(v, m)) Sr (v, n) = 0,
(1.184)
n=0
with r = 0, . . . , N. We take S0 (v, n) = 1 so that 0 = 0 and set S1 (v, n) =

S(v, n), 1 (v) = 0
(v), where S satisfies (1.174). We then introduce the eigenfunction expansion
N
Q(z, n) = c0 + cr Sr (v0 , n)er (v0 )z .
(1.185)
r=1
In order that the solution remains bounded as z we require that cr = 0

if Re[r (v0 )] < 0. The boundary conditions (1.181) generate a system of linear
equations for the coefficients cr with codimension k. One of the unknowns is determined by matching the outer solution, which suggests that there are k 1 eigenvalues with negative real part. The eigenvalues are ordered so that Re[r (v0 )] < 0
for r > N + 1 k.
There is, however, one problem with the above eigenfunction expansion, namely,
that 1 (v0 ) 0
(v0 ) = 0 so that the zero eigenvalue is degenerated at v = v0 .
Hence, the solution needs to include a secular term involving the generalized eigen-
54
vector S,
N
0 , n) = F(v0 , m).
A(n, m; v0)S(v
(1.186)
n=0
The Fredholm alternative theorem ensures that S exists and is unique, since
the stationary density (v0 , m) is the right null vector of A(n, m; v0 ) and
n (v0 , n)F(v0 , n) A (v0 ) = 0; see (1.133). The solution for Q(z) is now
0 , n) z) +
Q(z, n) = c0 + c1 (S(v
N+1k
cr Sr (v0 , n)er (v0 )z .
(1.187)
r=2
The presence of the secular term means that the solution is unbounded in the limit
z , which means that the inner solution cannot be matched with the outer solution. One way to remedy this situation is to introduce an alternative scaling in
the boundary layer of the form v = v0 + 1/2 z, as detailed in [455]. One can then
eliminate the secular term c1 z and show that

2|0
(v0 )|
+ O( 1/2 ), cr = O( 1/2 ) for r 2
(1.188)
c1
It turns out that we only require the first coefficient c1 in order to evaluate the
principal eigenvalue 0 using (1.168). This follows from (1.171) and (1.184) and
m; v) =
the observation that the left and right eigenvectors of the matrix A(n,
A(n, m; v)/F(v, n) are biorthogonal. In particular, since the quasistationary approximation is proportional to R(0) (see (1.177)), it follows that is orthogonal to all
eigenvectors Sr , r = 1. Simplifying the denominator of (1.168) by using the outer
solution 0 1, we obtain
n 0 (v0 , n)F(v0 , n) (v0 , n)
, 1

k(v0 )B(v0 ) |
(v )|
0 (v0 ) 0 (v )
exp
c1
,
k(v )
2
(1.189)
with
0 , n)F(v0 , n) (v0 , n).

B(v0 ) = S(v
(1.190)
n=0
Substituting for c1

1 k(v0 )B(v0 )
0 (v0 ) 0 (v )
0
0 (v )|0 (v0 )| exp
. (1.191)
k(v )
Finally, comparison of (1.186) and (1.190) with (1.148) and (1.147) establishes that
B(v0 ) D(v0 ).
55
The above analysis holds for any CK equation of the form (1.143). There
are essentially three basic steps needed in order to evaluate the escape rate formula (1.191), which we now apply to the specific model of membrane voltage
fluctuations.
1. Find the unique nontrivial positive eigenfunction n (v) = R(0) (v, n) and associated eigenvalue (v) = 0
(v). In the case of the stochastic ion-channel model,
(1.171) takes the explicit form
(N n + 1)n1 [n + (N n) ]n + (n + 1) n+1
n

f (v) g(v) n
=
N
(1.192)
Motivated by the form of the stationary density (v, n), we try the solution
n (v) =
(v)n
,
(N n)!n!
(1.193)
which yields the following equation relating and :

n

n
f (v) g(v) .
+ (N n) n (N n) =
N
We now fix in terms of so that the terms linear in n vanish:

1
N
=
+ 1 ( + 1) .
f
Eliminating then shows that
(v) =
We deduce that
(v) = N
g(v)
.
f (v) g(v)
(v) f (v) ( (v) + )g(v)

,
g(v)( f (v) g(v))
(1.194)
and the normalized eigenvector is
n (v) =
( f (v) g(v))Nn g(v)n

N!
.
(N n)!n!
f (v)N
(1.195)
Note that (v) vanishes at the fixed points v , v0 of the mean field (1.133) with
(v) > 0 for 0 < v < v and (v) > 0 for v < v < v0 . Moreover, n (v) =
(v, n) at the fixed points v0 , v . In conclusion R(0) (v, n) = n (v) and the effective
potential 0 is given by
0 (v) =
v
(y)dy.
(1.196)
56
The effective potential is defined up to an arbitrary constant, which has been fixed
by setting 0 (v ) = 0.
2. Determine the null eigenfunction n (v) = S(v, n) of (1.174), which becomes
(N m)m+1 [(N m) + m ]m + m m1
m

f (v) g(v) m .
=
N
(1.197)
Trying a solution of the form m (v) = (v)m yields

m
f (v) g(v) .
(N m) ((N m) + m ) + m 1 =
N
(1.198)
is then determined by canceling terms independent of m, which gives

n
b(v)g(v)
n (v) =
.
(1.199)
a(v)( f (v) g(v)))
The prefactor k(v) may now be determined using (1.175) and (1.176).
0 , n) of (1.186), which reduces to
3. Calculate the generalized eigenvector n = S(v
(N n) (v0 )n+1 + n n1 ((N n) (v0 ) + n )n = g(v0 )
n
f (v0 ).
N
(1.200)
It is straightforward to show that this has the solution
n =
f (v0 )
n.
N( (v0 ) + )
(1.201)
It follows that the factor B(v0 ) defined by (1.190) is

N
f (v0 )
f (v0 ) 2
n
B(v0 ) =
(v
,
n)
g(v
)n
+
0
0
N( (v0 ) + ) n=0
N

f (v0 ) 2
f (v0 )
n
=
g(v0 )n +
N( (v0 ) + )
N
=
f (v0 )2 (v0 )
,
N( (v0 ) + )3
(1.202)
where have used the fixed point condition g(v0 ) = f (v0 )a(v0 ).
Keener and Newby [321] showed that the WKB and asymptotic methods outlined
above yield a MFPT that was in excellent agreement with numerical simulations in
both the superthreshold and subthreshold regimes.
1.7 Appendix: Stochastic Calculus
57

In this appendix we give an informal introduction to stochastic calculus, following
along the lines of Jacobs [302]. A more detailed treatment can be found in Gardiner [210], and a rigorous mathematical account can be found in [476]. The basic
approach is to treat a continuous-time stochastic process as the limit of a discrete
time process. That is, an SDE prescribes how a stochastic variable X(t) changes
in each infinitesimal time step dt. Determining changes over finite times then requires evaluating an associated stochastic integral. In order to make sense of this,
we discretize time into small, but finite, intervals of duration t and consider a corresponding stochastic difference equation for Xn = X(n t).
1.7.1 Ito Calculus

Suppose that we divide the time interval [0, T ] into N increments of size t = T /N
and set tn = n t. Consider the stochastic difference equation
X(tn ) X(tn+1 ) X(tn) = Wn ,

where each Wn , n = 0, . . . , N 1, is an independent, identically distributed (i.i.d)
Gaussian variable with zero mean and variance 2 = t:
2
1
P( W ) =
e( W ) /2 t .
2 t
(1.203)
Iterating the difference equation starting from X(0) = 0 yields

Xn X(n t) =
n1
Wj.
j=0
Using the fact that the sum of Gaussian random variables is also a Gaussian, it
follows that the probability density for Xn is a Gaussian. Thus, we only need to
determine its mean and variance. Since the W j are all independent, we have
Xn =
n1
W j = 0,
Var(Xn ) =
j=0
and
n1
Var( W j ) = N t,
j=0
1
2
P(Xn ) =
eXn /(2n t) .
2 n t
We can now construct a corresponding continuous-time process by taking the limit

N such that t 0 with N T = T fixed. In particular,
X(T ) = lim
N1
j=0
Wj
T
0
dW (t) W (T ),
58
where W (T ) is identified as a Wiener process. (A rigorous treatment would be more

precise with regard to what is meant by the convergence of random variables.) It is
still a Gaussian, whose mean and variance are obtained by taking the limit N
of the results for Xn . We deduce that W (t) has the Gaussian probability density
1 w(t)2 /2t
e
.
P(w(t)) =
2 t
Now consider the modified stochastic difference equation
Xn+1 Xn = f (tn ) Wn ,
where f (t) is a deterministic function of time. Once again Xn is a Gaussian random
variable, with
Xn =
n1
f (t j ) W j = 0,
Var(Xn ) =
j=0
n1
n1
j=0
j=0
Var( f (t j ) W j ) = f (t j )2 t.
Taking the continuum limit along identical lines to the previous case yields the
continuous-time Gaussian variable
X(T ) = lim
N1
f (t j ) W j
T
j=0
f (t)dW (t),
(1.204)
with zero mean and variance

Var(X(T )) =
T
0
f (s)2 ds.
(1.205)
Substituting for X(T ) into this equation gives

T
T
f (t)dW (t)
f (s)dW (s) =
0
f (s)2 ds,
which can be captured by the rule

dW (t)dW (s) = (t s)dt ds.
(1.206)
However, care must be taken with this rule when (t s) appears inside an integral
having t or s as one of its limits. For example, consider the double stochastic integral

T
f (s)dW (s) g(t)dW (t) lim
N1
n1
n=0
m=0
f (tm )dWm g(tn )dWn .
We see that there are no terms in the double sum on the right-hand side that have a
product of Wiener increments in the same time interval. Thus, taking the expectation
of both sides,

59

T
0

f (s)dW (s) g(t)dW (t) = 0.
Hence, we require
t
f (s) (t s)ds = 0,
t
f (s) (s)ds = f (0).
(1.207)
Following the previous examples, let us turn to a discretized version of the general
SDE for X(t),
dX = a(X,t)dt + b(X,t)dW(t),
(1.208)
which takes the form
Xn+1 Xn = a(Xn ,tn ) t + b(Xn,tn ) Wn .
(1.209)
Iterating this equation starting from a fixed X(0) = x0 yields

XN = x0 +
N1
N1
b(Xn,tn ) Wn .
a(Xn ,tn ) t +
n=0
n=0
The continuum limit then gives the stochastic integral equation

X(T ) = x0 +
T
0
a(X(t),t)dt +
T
0
b(X(t),t)dW (t),
(1.210)
with the final term defined as the Ito stochastic integral

T
0
b(X(t),t)dW (t) = lim
N1
b(Xn,tn ) Wn .
(1.211)
n=0
The integral equation is not very useful for generating an explicit solution for X(t).
However, from the definition of the Ito stochastic integral, it immediately follows
that

T
b(X(t),t)dW (t) = 0,
(1.212)
0
since Xn is a function of previous Wiener increments Wn1 , . . . , W0 so it is uncorrelated with Wn . The stochastic difference equation (1.209) is the starting point
for developing numerical schemes for solving an SDE. However, if one is interested
in carrying out explicit calculations, it is usually more useful to go to the associated FokkerPlanck equation for the probability density. In order to derive the FP
equation from the corresponding SDE, we first need to consider the object (dW )2 .
In terms of Wiener increments,
T
0
(dW (t))2 = lim
N1
( Wn )2 .
n=0
60
Taking the expectation of both sides and using the fact that each Wn is an i.i.d.,
gives
T
T
T
2
2
(dW (t)) =
(dW (t)) =
dt = T.
(1.213)
0
What about the variance? Using the Gaussian probability density (1.203), it is simple to show that
Var[( W )2 ] = 2( t)2 = 2T 2 /N 2 .
Hence,

Var
0
(dW (t))
= lim Var
N
= lim
N1
( Wn )
N1
Var
N
= lim
n=0
( Wn )2
n=0
2T
= 0.
N
We thus obtain the surprising result that the integral of (dW )2 is deterministic and
thus equal to its mean:
T
0
(dW (t))2 = T =
T
dt.
(1.214)
In other words, we can set (dW )2 = dt, a result known as Itos rule. Using similar
arguments, it can also be shown that dW m = 0 for m > 2.
1.7.2 Itos Formula and the FokkerPlanck Equation

The result dW (t)2 = dt has important implications for how one carries out a change
of variables in stochastic calculus. This is most directly established by considering
the SDE for an arbitrary function f (X(t)) with X(t) evolving according to (1.208):
d f (X(t)) = f (X(t) + dX(t)) f (X(t))
1
= f
(X(t))dX(t) + f
(X(t))dX(t)2 + . . .
2
1
= f (X(t))[a(X,t)dt + b(X,t)dW(t)] + f
(X(t))b(X,t)2 dW (t)2 ,
2
where all terms of higher order than dt have been dropped. Now using dW (t)2 = dt,
we obtain the following SDE for f , which is known as Itos formula:

1
d f (X(t)) = a(X(t),t) f
(X(t)) + b(X,t)2 f
(X(t)) dt + b(X,t) f
(X(t))dW (t).
2
(1.215)
Hence, changing variables in Ito calculus is not given by ordinary calculus unless f
is a constant or a linear function.
61
We can now use Itos formula to derive the FP equation for an Ito SDE. First,

d f (X(t))
1
= a(X(t),t) f
(X(t)) + b(X(t),t)2 f
(X(t))
dt
2

1
=
a(x,t) f
(x) + b(x,t)2 f
(x) p(x,t)dx,
2

1 2
2
= f (x) (a(x,t)p(x,t)) +
(b(x,t)
p(x,t))
dx. (1.216)
x
2 x2
after integration by parts, where p(x,t) is the probability density of the stochastic
process X(t) under the initial condition X(t0 ) = x0 . However, we also have

d f (X(t))
d f (X(t))
=
dt
dt
d
= f (X(t),t)
dt
= f (x) p(x,t)dx.
(1.217)
t
Comparing (1.216) and (1.217) and using the fact that f (x) is arbitrary, we obtain
the Ito version of the FP equation
1 2
p(x,t) = (a(x,t)p(x,t)) +
(b(x,t)2 p(x,t)).
t
x
2 x2
(1.218)
1.7.3 Multiplicative Noise and Stratonovich Integrals

It turns out that there is more than one way to define a stochastic difference equation
driven by an incremental Wiener process and thus more than one way to obtain an
SDE in the continuum limit. This issue only arises in the case of multiplicative noise,
that is, when the term multiplying dW (t) depends on the state variable X(t). Recall
that in the Ito integral (1.211), it is the value of b(x,t) at the start of the nth time
step that multiplies Wn , so that there are no contributions of the form ( Wn )2 . An
alternative definition of a stochastic integral is the Stratonovich integral
T
0

Xn+1 + Xn
,tn Wn ,
b(X(t),t)dW (t) = lim b
N
2
n=0
N1
(1.219)
where we have used to distinguish it from the Ito integral. Now b depends on the
value Xn+1 at the end of the nth time step, which means there will be an extra term
involving ( Wn )2 . In order to compare the Ito and Stratonovich integrals, suppose
that Xn evolves according to the stochastic difference equation (1.209). Thus, in the
62
continuum limit X(t) is the solution to an Ito SDE. Suppose that we Taylor expand
the nth term in the sum defining the Stratonovich integral about the point Xn and set
bn = b(Xn ,tn ):

Xn+1 + Xn
Xn bn 1 Xn 2 2 bn
,tn = bn +
+
+ ...
b
2
2 x
2
2
x2
Substituting for Xn using (1.209) and dropping terms that are higher order than t
shows that

Xn+1 + Xn
an bn b2n 2 bn
bn bn
,tn = bn +
+
b
t +
Wn .
2
2 x
8 x2
2 x
Applying this result to the sum appearing in the definition of the Stratonovich integral, (1.219), and again dropping higher-order terms in t, yields the result
N1
n=0

N1
N1
Xn+1 + Xn
bn bn
,tn Wn = bn Wn +
( Wn )2 .
2
2
x
n=0
n=0
Finally, taking the continuum limit with dW (t)2 = dt, we have

T
0
b(X(t),t)dW (t) =
T
0
b(X(t),t)dW (t) +
1
2
T
b(X(t),t)
0
b(X(t),t)dt.
(1.220)
Now suppose that Y (t) is a stochastic process evolving according to the Stratonovich
SDE
dY = a(Y,t) + b(Y,t)dW(t).
(1.221)
This means that the integral equation satisfied by Y (t) is based on the Stratonovich
integral, that is,
Y (t) = y0 +
t
0
a(Y (s), s)ds +
t
0
b(Y (t),t)dW (t).
(1.222)
Using (1.220), we can rewrite the solution in terms of an Ito integral according to
t
t
1 b(Y (s), s)
Y (t) = y0 +
g(Y (s), s) ds + b(Y (s), s)dW (s).
a(Y (s), s) +
2
y
0
0
(1.223)
The latter is the solution to an equivalent Ito SDE of the form

b(Y (t),t) b(Y (t),t)
dt + b(Y (t),t)dW (t).
(1.224)
dY = a(Y (t),t) +
2
y
Finally, given that we know the FP equation corresponding to an Ito SDE, we can
immediately write down the FP equation corresponding to the Stratonovich SDE
equation (1.221):

1
p(y,t) = (a(y,t)p(y,t)) +
(1.225)
b(y,t) [b(y,t)p(y,t)] .
t
y
2 y
y
Chapter 2
Traveling Waves in One-Dimensional

Excitable Media
We now consider our first example of wave propagation in neural media, namely, the
propagation of an action potential along the axon of a neuron. Such a phenomenon
can be formulated mathematically in terms of finding a traveling pulse solution of
the spatially extended HodgkinHuxley equations (1.8) and (1.4). Formally speaking, a traveling wave is a solution of a PDE on an infinite domain that travels at
constant velocity and fixed shape. For one-dimensional systems, one can distinguish
two types of solitary traveling wave: a traveling front linking a stable resting state to
a stable excited state and a traveling pulse that begins and ends at the resting state;
see Fig. 2.1. For the HodgkinHuxley model, a traveling front would occur if the
recovery variable n associated with K+ channels were frozen; otherwise repolarization ensures that the trailing edge of the wave returns to the resting state. In order to
develop the basic theory of wave propagation in one-dimensional excitable media,
we will consider a simplified version of the HodgkinHuxley equations given by the
FitzHughNagumo (FN) equations [192, 446]:
v(x)
v(x)
x
x
Fig. 2.1 Schematic illustration of (a) a traveling front, and (b) a traveling pulse
63
64
2 Traveling Waves in One-Dimensional Excitable Media
v 2v
= 2 + f (v) w f (v, w),
t
x
w
= (v w) g(v, w).
t
with 0 < 1 and
f (v) = v(v a)(1 v).
(2.1a)
(2.1b)
(2.2)
Here v represents a fast voltage variable and w is a slow recovery variable. The time
and space scales have been non-dimensionalized so that the effective diffusivity of
the cable is unity. A number of excellent reviews of waves in excitable systems can
be found elsewhere [242, 316, 322, 444].
2.1 Excitable Systems

Before consider traveling wave solutions, it is instructive to consider the excitable
behavior of the space-clamped (x-independent) FN model. The space-clamped FN
model takes the form of a planar dynamical system
dv
= f (v, w),
dt
dw
= g(v, w).
dt
(2.3a)
(2.3b)
The fast variable has a cubic nullcline (along which v = 0) and the slow variable
has a monotonically increasing nullcline (along which w = 0). It is assumed that the
nullclines have a single intersection point at (v , w ). This corresponds to a fixed
point of the system, which we identify with the resting state. A schematic diagram
of the phase plane is shown in Fig. 2.2. For a finite range of values of w, there
exist three solutions v = v(w) of the equation f (v, w) = 0, which we denote by
V (w),V0 (w), and V+ (w). Whenever these solutions coexist, we have the ordering
V (w) V0 (w) V+ (w). Let W denote the minimal value of w for which V (w)
exists, and let W denote the maximal value of w for which V+ (w) exists.
First, suppose that the fixed point is located on the left-hand branch close to the
minimum of the cubic. It is straightforward to show that the fixed point is linearly
stable by evaluating the eigenvalues of the corresponding Jacobian. Moreover, the
system is excitable in the sense that sufficiently large perturbations of the resting
state result in a time-dependent trajectory taking a prolonged excursion through
state space before returning to the resting state; see Fig. 2.3. Such a trajectory rapidly
transitions to the right branch V+ , after which it slowly moves upward in a neighborhood of the branch before reaching the maximum. It then rapidly transitions back to
the left branch V followed by a slow returns to the resting state along this branch.
The time-dependent plot of the variable v can be interpreted as an action potential.
Since the resting state is linearly stable, small perturbations simply result in small
65
g(v,w)=0
f(v,w)=0
W*
V+(w)
V-(w)
V0(w)
v
W
*
Fig. 2.2 Schematic diagram of the phase plane for the FitzHughNagumo equations
excursions that decay exponentially in time. Hence, there is effectively a threshold

phenomenon in which subthreshold perturbations result in a simple return to the
resting state, whereas superthreshold perturbations generate an action potential.
f(v,w)=0
g(v,w)=0
time
v
action potential
Fig. 2.3 Schematic diagram illustrating the trajectory of a single action potential in the phase
plane for the FitzHughNagumo equations. The unique rest point is stable. Inset shows the action
potential as a function of time
A more mathematical description of the above events can be developed in terms

of singular perturbation theory [242, 322]. Due to the separation of time scales with
1, the fast variable v rapidly adjusts whenever it can to maintain the quasiequilibrium f (v, w) = 0. This can be captured by introducing the slow time scale
= t such that (2.3) become
dv
= f (v, w),
d
dw
= g(v, w).
d
(2.4)
66
Now setting = 0 and assuming that v is moving along the stable branches V (w)
of f (v, w) = 0, the dynamics of the recovery variable reduces to
dw
= g(V (w), w) G (w).
d
(2.5)
In the case of rapid transitions between the left and right branches, the dynamics
with respect to the fast time scale can be approximated by setting = 0 in (2.3),
dv
= f (v, w),
dt
dw
= 0.
dt
(2.6)
Thus, on this time scale, w is constant and v converges to a stable solution of

f (v, w) = 0. Suppose that the system starts from a superthreshold initial condition
(v0 , w0 ) such that v0 > V0 (w0 ). After rapidly reaching the right branch, it takes a
finite time to reach the upper knee of the nullcline f (v, w) and is obtained by
integrating (2.5):
W
dw
.
(2.7)
Te =
w0 G+ (w)
On the other hand, the time taken to return to the resting state along the left branch
is infinite, since G (w) vanishes at the fixed point.
It is possible to convert the FN equations from an excitable to an oscillatory
regime by adding a constant external current Iext to the right-hand side of the voltage equation in (2.3). For an intermediate range of values of Iext one finds that the
fixed point shifts to the middle branch V0 (w) where it is unstable. The fixed point
now coexists with a limit cycle, along which the trajectory alternates periodically between the left and right branches, while w varies between W and W ; see Fig. 2.4.
The resulting limit cycle behavior with fast jumps alternating with slow dynamics
is known as a relaxation oscillator. For small , the period T of the oscillator is
dominated by the times to move along the left and right branches. Hence

W
1
1
T=
dw,
(2.8)
G+ (w) G (w)
W
with G+ > 0 and G < 0.
Another well-known planar model of an excitable neuron is the MorrisLecar
(ML) model [440] (see Eq. (1.28)) which we write in the form
dv
= a(v) fNa (v) + w fK (v) g(v)
dt
(2.9a)
dw w (v) w
=
,
dt
w (v)
(2.9b)
where fi (v) = gi (vi v) and w represents the fraction of open K+ channels. The fraction of Na+ channels (or Ca2+ channels in the original formulation of the model) is
assumed to be in quasi steady state. Again we can analyze the generation of action
67
f(v,w)=0
g(v,w)=0
time
v
limit cycle
Fig. 2.4 Schematic diagram illustrating the trajectory of a globally stable periodic orbit in the
phase plane for the FitzHughNagumo equations. The unique rest point is unstable. Inset shows
the periodic orbit as a function of time
potentials using a slow/fast analysis of the deterministic system. However, it turns

out that this adiabatic approximation breaks down when stochastic fluctuations in
the opening and closing of K+ channels are taken into account. This can be established by considering a stochastic version of the ML model [456] consisting of N
sodium and M potassium channels (see also Sect. 1.5):
n
m
dv
= F(v, m, n) fNa (v) + fK (v) g(v).
dt
N
M
(2.10)
We assume that each channel can either be open or closed and can switch between
each state according to the kinetic scheme
C
i (v)
i (v)
O,
i = Na, K,
(2.11)
The Na+ channels open and close rapidly relative to the voltage and K+ dynamics. The probability density function p(v, m, n,t) of the resulting stochastic hybrid
system (see Sect. 1.6) evolves according to the differential ChapmanKolmogorov
(CK) equation,
p
(F p)
=
+ LK p + LNa p.
(2.12)
t
v
The jump operators L j , j = Na, K, are defined according to
+
+
L j = (E
n 1) j (n, v) + (En 1) j (n, v),
(2.13)
+
with E
n f (n) = f (n 1), j (n, v) = n j (v) and j (n, v) = (N n) j (v).
Introducing the small parameter with Na , Na , M = O(1/ ), one can extend
the WKB approximation method of Sect. 1.6 to analyze noise-induced transitions
in the phase plane [456]. The WKB potential can be interpreted as the action
of an effective Hamiltonian dynamical system whose solutions determine charac-
68
Fig. 2.5 Illustration of spontaneous action potentials (SAPs) for the stochastic ML model (2.10)
with a finite number of sodium and potassium ion channels [456]. Orange curves are SAP trajectories, shown until they reach the effective metastable separatrix (S). The dashed red curve is a SAP
that reaches S near the metastable saddle (SN). All of the SAP trajectories in the shaded region
containing the most probable, observable SAP trajectoriesare visually indistinguishable from the
dashed red line before crossing S. Deterministic trajectories are shown as black streamlines. Also
shown are a caustic (C), caustic formation point (CP), v nullcline (VN), and w nullcline (WN).
Parameter values are N = M = 10 and = 0.1
teristic paths in the phase plane (see also Sect. 4.4). The latter correspond to the
paths a stochastic trajectory is most likely to follow during a metastable transition
(i.e., a path of maximum likelihood [160]). Based on the fast/slow analysis of the
deterministic system (2.9), one might expect w to be approximately constant along
a stochastic trajectory that jumps between the left and right branches of the voltage
nullcline, since the K+ channels open and close slowly. In fact this does not hold
for spontaneous action potentials arising from K+ channel fluctuations [456] (see
Fig. 2.5), which is in contrast to the effects of noise in the voltage or fast sodium
channels. In general, it is difficult to solve FPT problems in more than one dimension. In the case of a metastable state with a well-defined basin of attraction, one has
to calculate the MFPT to cross the separatrices forming the boundary of the basin
of attraction. There is an additional level of complexity for an excitable system, due
to the fact that there is no well-defined deterministic separatrix. Interestingly, as illustrated in Fig. 2.5, the stochastic ML model has an effective separatrix that any
stochastic trajectory has to cross in order to generate a stochastic action potential
[456]; see also [327]. Another commonly observed feature of the WKB approximation in two or more dimensions is the formation of caustics, where characteristic
projections of the Hamiltonian intersect. There is now quite an extensive literature
on the effects of noise in excitable systems, as reviewed in [383]. Most of these studies consider extrinsic Gaussian noise in the voltage dynamics and phenomena such
as stochastic and coherence resonance. In Sect. 2.6 we will consider the effects of
2.2 Traveling Fronts in a Scalar Bistable Equation
69
Gaussian noise on wave propagation in a spatially extended excitable system. One

can motivate this form of noise by considering diffusion approximations of models
of stochastic ion channels. However, as the above example shows, certain caution
must be exercised when considering such approximations.
Finally, we note that fast/slow decomposition has been applied extensively in recent years to the study of rhythmic activity patterns in single neurons and in synaptically coupled relaxation oscillators (see also Sect. 5). In the latter case, if synapses
turn on and off on the fast time scale, then geometric singular perturbation theory
can be used to investigate how synaptic inputs modify geometric structures such as
null surfaces in the phase spaces of individual neurons, as reviewed by Rubin and
Terman [537]; see also Chap. 9 of [173]. Moreover, such methods have been used to
study a three-dimensional version of the HodgkinHuxley model, in which there is
one fast variable and two slow variables [539, 540]. In addition to the transition from
excitable to regular oscillatory behavior, as observed in classical relaxation oscillators, the model neuron also exhibits more complex dynamics such as mixed-mode
oscillations that are associated with slow action potentials.

In the absence of a recovery variable, the FN equations reduce to the so-called scalar
bistable equation
v 2v
= 2 + f (v), < x <
(2.14)
t
x
with f (v) given by the cubic (2.2). For such a choice of nonlinearity, the corresponding ODE, dv/dt = f (v), has stable equilibria at v = 0, 1 separated by an unstable
equilibrium at x = a. We define a traveling front solution according to
v(x,t) = v(x ct) = V ( ),
= x ct
(2.15)
for some yet to be determined wave speed c, supplemented by asymptotic boundary conditions ensuring that the front links the two stable fixed points of the xindependent system. For concreteness, we take
V ( ) 1 as ,
V ( ) 0 as .
(2.16)
Substituting the traveling front solution into the bistable Eq. (2.14) yields the ODE
V + cV + f (V ) = 0,
(2.17)
where V = dV /d .
Classical phase-plane analysis can be used to find a traveling front solution by
rewriting the second-order equation in the form
V = Z,
Z = cZ f (V ).
(2.18)
70

c << 1
c >> 1
stable
manifold
unstable
manifold
V
V
unstable
manifold
stable
manifold
c = c*
V
unstable
manifold
stable
manifold
Fig. 2.6 Shooting method for constructing a front solution in the (V, Z) phase plane with Z = V .
See text for details
One now has to look for a heteroclinic connection that links the excited state
(V, Z) = (1, 0) at to the resting state at . This can be achieved using
a geometric argument based on a shooting method, as illustrated in Fig. 2.6. For the
sake of illustration, suppose that 0 < a < 1/2 so that c > 0 (see below). First note
that irrespective of the speed c, the fixed points (1, 0) and (0, 0) are saddles, each
with one-dimensional stable and unstable manifolds. By looking at trajectories in
the phase plane, it is straightforward to see that when c 1, the unstable manifold
of (1, 0) lies below the stable manifold of (0, 0) when 0 < V < 1, whereas the opposite holds when c is very large. Since these manifolds depend continuously on c,
it follows that there must exist at least one value of c for which the manifolds cross,
and this corresponds to the heteroclinic connection that represents the traveling front
solution. It can also be established that this front is unique. A useful formula for determining the sign of the wave speed can be obtained by multiplying both sides of
(2.17) by V and integrating with respect to :

(V )2 d =
=
1
0
V f (V ( ))d
f (V )dV,
V V d ,
(2.19)
d[V 2 /2]
since V ( ) is monotone, and

V V d =
d d = 0. As the integral
on the left-hand side is positive, it follows that the sign of c is determined by the
71
sign of the area of f between the two stable equilibria. If 0 < a < 1/2, then the
latter is positive and the wave moves to the right, converting the medium from the
resting state to the excited state. On the other hand, a left-moving front occurs when
1/2 < a < 1, converting the medium from the excited state to the resting state. If the
negative and positive areas exactly cancel, then the front is stationary.
The numerical construction of the traveling front using a shooting method does
not depend on the precise form of f . However, if f is given by the cubic (2.2),
then it is possible to construct the front explicitly. That is, we substitute the ansatz
a) = 0.
Z = AV (1 V ) into (2.17) to obtain the condition A2 (2V 1) + cA (V
Collecting terms
linear
in
V
and
terms
independent
of
V
requires
A
=
1/
2 and
c = (1 2a)/ 2. This immediately establishes that c switches sign at a = 1/2.

Since W = V , it follows that the corresponding wave profile is
V ( ) =

1
1 tanh( /2 2) .
2
Finally, recall that we have non-dimensionalized the units of space and time in the
bistable equation by setting the membrane time and space constants of the cable to
unity (m = 1, m = 1); see (1.55). Hence, in physical units, the speed of the wave is

c m
d
c
c =
,
(2.20)
=
m
2Cm Rm R
where
d is the cable diameter. Based on empirical estimates, one finds that c
d mm/sec. In the case of a squid axon of diameter d = 500 m, the estimated

propagation speed is of around 20 mm/ms.
Another choice of nonlinearity for which an explicit front can be calculated is
the piecewise linear function
f (v) = v + H(v a).
(2.21)
Substituting into Eq. (2.17) gives

V + cV V + H(V a) = 0.
(2.22)
Translation symmetry of the system means that we are free to choose V to cross the
threshold a at = 0 so that V ( ) > a for < 0 and V ( ) < a for > 0. Solving the
resulting linear equation on either side of the threshold point = 0 and imposing
the threshold condition V (0) = a yields the explicit solution

ae ,
>0
V ( ) =
(2.23)
1 + (a 1)e+ , < 0,
where are the roots of the characteristic equation 2 + c 1 = 0. The wave
speed is then obtained by imposing continuity of V at = 0, (a 1)+ = a ,
which after rearranging gives
72
1 2a
c=
.
a a2
(2.24)
2.3 Traveling Pulses in the FitzHughNagumo Equations

The bistable equation cannot support a traveling pulse solution because there is
no recovery variable, that is, it does describe an excitable system. In order to
obtain traveling pulse solutions, it is necessary to consider the full FitzHugh
Nagumo equations (2.1). Suppose that the unique fixed point (v , w ) lies on the
left-hand branch as in Fig. 2.2. Assume a traveling wave solution of the form
(v(x,t), w(x,t)) = (V ( ),W ( )) with = x ct so that (2.1) reduce to the system
of ODEs
V + cV + f (V,W ) = 0,
(2.25a)
cW + g(V,W ) = 0.
(2.25b)
These are supplemented by the asymptotic boundary conditions

lim (V ( ),W ( )) = (v , w ).
(2.26)
Mathematically speaking, one needs to find a trajectory in the phase space (V, Z,W )
with Z = V that is homoclinic to the resting state (v , 0, w ). (Although the resting state is stable in the space-clamped system, it is a saddle in the phase plane
(V, Z,W ) of the spatially extended system.) The existence of such an orbit can be
demonstrated using geometric singular perturbation theory [104, 267]. The basic
idea is to formally set = 0 and construct a singular homoclinic orbit. This will
consist of four parts: the jump-up from the resting state to the right branch V+ (w),
an active phase along V+ (w), the jump-down to the left branch V (w), and a quiescent phase as it returns to the resting state along V (w); see Fig. 2.7. Given the
existence of a singular homoclinic orbit, one can then prove that such an orbit persists for > 0 provided that is sufficiently small. In this section we will focus
on the construction of the singular solution, exploiting the fact that the jumps take
place on a fast spatial scale , whereas the active and quiescent phases occur on a
slow spatial scale z = .x
In order to analyze the jump-up from the resting state to the active phase, we set
= 0 in (2.25) to obtain the reduced system
V + cV + f (V,W ) = 0,
(2.27a)
W = 0.
(2.27b)
Thus the recovery variable is a constant w and V evolves according to the bistable
equation
V + cV + f (V, w) = 0,
(2.28)

w
73
g(v,w)=0
f(v,w)=0
wave back
wt = G+(w)
wt = G-(w)
wave front
v
Fig. 2.7 Phase portrait of the fast traveling solitary pulse for the FN equations in the singular limit
0. The arrow directions are for increasing time (decreasing )
for f (V, w) = v(v a)(1 v) w. Following Sect. 2.2, we know that for fixed w
there exists a unique traveling front solution of (2.28) with associated wave speed
c = c(w) that is a heteroclinic connection with V VL (w) as and v VR (w)
as . Moreover,
V+ (w)
c(w) =
f (V, w)dV
V (w)
2
V d
(2.29)
Setting w = w with VL (w ) = v we obtain the wave speed c = c(w ). Next consider the slow active phase with c = c . Introducing the slow time scale z = we
have
2Vzz + c Vz + f (V,W ) = 0,
cWz + g(V,W ) = 0.
(2.30a)
(2.30b)
Setting = 0 then leads to the reduced system

f (V,W ) = 0,
1
Wz = g(V,W ).
c
(2.31a)
(2.31b)
Taking the solution V = VR (W ) of f (V,W ) = 0, it follows that the trajectory moves

along the right branch at a rate determined by G+ (W ) = g(VR (W ),W ).
Now suppose that the trajectory leaves the branch at some value W = Wd < W
(where W is at the maximum of the cubic f (V,W ) = 0) and jumps back down to the
left branch. Similar to the jump-up phase, the recovery variable is constant, and V
evolves according to the bistable Eq. (2.28) with w = Wd . Again using Sect. 2.2,
we can construct a unique traveling wave solution with associated wave speed
c(Wd ) < 0 that is a heteroclinic connection from VR (Wd ) to VL (Wd ). The wave speed
is negative, since the jump-down starts from an active state rather than a quiescent
74
state; it is thus referred to as a wave back. The requirement that the solution be a
steadily propagating pulse means that the speeds of the jump-up and jump-down
solutions must be the same, which implies that
c(w ) = c(Wd ).
(2.32)
This condition uniquely determines the transition point Wd . Finally, the trajectory
returns to the resting state by moving along the left branch at a rate determined by
G (W )/c .
A number of comments are in order. First, from the perspective of matched
asymptotics in the slow variable z, the active and quiescent phases correspond to
outer solutions that are slowly varying, whereas the jump-up and jump-down phases
correspond to transition layers or inner solutions. Denote the wave-front and waveback solutions by V ( ). Matching the inner and outer solutions then leads to the
following asymptotic conditions:
lim V+ ( ) = VR (W (0)),
W (0) = w
lim V+ ( ) = lim VL (W (z)) = v
lim V ( ) = VL (W (zT )),
W (zT ) = Wd
lim V ( ) = VR (W (zT ))
The location zT where the jump-down occurs is determined from the slow dynamics
according to
zT = c
Wd
w
dW
dW.
G+ (W )
(2.33)
One can interpret zT as the width of the pulse. Second, it may be the case that there is
no solution of c(w ) = c(Wd ) such that Wd < W . The jump-down transition then
occurs at the upper knee, and the solution is referred to as a phase wave, that is, the
jump occurs at a time or phase determined solely by the outer dynamics. The wave
behavior is then qualitatively different, since the wave can travel at any speed above
some minimum, analogous to the well-known FisherKPP equation of population
genetics [191, 345]; see Sect. 3.3.
2.3.1 Periodic Wave Trains

One of the characteristic features of excitable systems is that they exhibit refractoriness. That is, once the system has responded to a superthreshold stimulus by
generating an action potential, there is a refractory period during which no subsequent stimuli can be evoked. From the singular construction of a traveling pulse, the
refractory period can be estimated as follows: after the jump-down there is a range
of values of the recovery variable, W0 w Wd for which the front solution to the
75
bistable Eq. (2.2) has negative wave speed, c(w) 0, with c(W0 ) = 0. The time
taken to reach W0 from Wd along the left branch is (for fixed x and rescaled time)
Tref =
Wd
W0
dW
.
|G (W )|
(2.34)
We can identify Tref as the refractory period. Once excitability has been restored, it is
possible to evoke another wave of excitation. However, the speed of the subsequent
pulse is expected to be slower due to the residual refractoriness of the system. Now
suppose that we periodically initiate action potentials at one end of a semi-infinite
cable. This will asymptotically produce a periodic wave train with the time between
successive pulses equal to the forcing period T . Assuming that the wave train travels
at a speed c, the spatial separation or wavelength of the pulses will be = cT .
The speed of the wave train is expected to be smaller than the speed of an isolated
pulse due to refractoriness, although this effect should decrease as T increases. In
other words, there exists a dispersion curve c = c(T ) with c(T ), a monotonically
increasing function of T .
w
g(v,w)=0
f(v,w)=0
wave back
wt = G+(w)
wt = G-(w)
wave front
Fig. 2.8 Phase portrait for the fast periodic wave train for the FN equations in the singular limit
0
It is possible to estimate the dispersion curve for the FitzHughNagumo equations using the previous singular construction. A periodic wave train consists of
an alternating sequence of jump-ups and jump-downs, separated by regions of slow
dynamics. A phase portrait for such a solution is shown in Fig. 2.8. The major difference from an isolated pulse (see Fig. 2.7) is that the jump-up occurs before reaching
the resting state, with W = WP > w . Denoting the corresponding value at the jumpdown by WQ , we require that the speeds of the corresponding wave front and wave
back are the same, that is, c(WP ) = c(WQ ). Since the time taken for the jumps is
negligible, the major contributions to the period T come from the time spent traversing the right and left branches:
T=
WQ
WP
dW
+
G+ (W )
WP
WQ
dW
.
G (W )
(2.35)
76
speed c
2
= 0.1
1
= 0.01
0
0
4
6
period T
10
Fig. 2.9 Illustration of typical dispersion curves c = c(T ) for T -periodic wave-train solutions of
the FN equations. A stable fast branch coexists with an unstable slow branch. They annihilate in a
saddlenode bifurcation
Solving for WQ in terms of WP using the matching of speeds and inverting the relation c = c(WP ) then generate the dispersion curve c(T ). It turns out that the dispersion curve breaks down when the wavelength cT = O( ), because the transitional fronts and backs become arbitrarily close together so that it is no longer
possible to carry out matched asymptotics. In order to construct the true dispersion
curve, it is necessary to carry out a regular perturbation analysis in with c = O( )
[156]. One then finds that the dispersion curve consists of two branches that annihilate in a saddlenode bifurcation at T = Tc ; see Fig 2.9. Hence there are no traveling
pulses for T < Tc . It can be shown that the upper branch of fast traveling pulses
is stable, whereas the lower branch of slow pulses is unstable. (The issue of wave
stability will be addressed in Sect. 2.4.) Note that as T , c c on the upper
branch, where c is the speed of an isolated pulse. Interestingly, c also approaches
a finite limit as T on the lower branch, suggesting that there also exists a slow
unstable isolated pulse; this is indeed found to be the case.
2.3.2 Kinematics
It is also possible to have wave trains consisting of action potentials that are irregularly spaced and move at different speeds. Rinzel and Maginu [523] developed a
kinematic theory of wave propagation that uses the dispersion relation to determine
the instantaneous speed of a pulse. That is, suppose an initial pulse is generated
at x = 0 at time t1 . The time at which the pulse reaches a given point x will be
T1 (x) = t1 + x/c , where c is the speed of a solitary pulse. Suppose that a second
spike is initiated at x = 0 at time t2 . The instantaneous speed of the new pulse at x
will depend on the time difference T2 (x) T1 (x) due to refractoriness. The Rinzel
and Maginu approximation is to take the instantaneous speed to be c(T2 (x) T1 (x)),
where c(T ) is the dispersion curve for a periodic wave train. It then follows that
2.4 Wave Stability and Evans Functions
77
1
dT2 (x)
=
,
dx
c(T2 (x) T1 (x))
(2.36)
and the time interval = T2 (x) T1 (x) between two action potentials initiated at
x = 0 will evolve in space according to
1
1
d
=
( ).
dx
c( ) c
(2.37)
If the function ( ) has one or more zeroes , then the phase difference between
the two action potentials will lock at ; the phase-locked state will be stable if

( ) < 0. Note the kinematic approximation can be extended to multiple action
potentials under the assumption that the instantaneous speed only depends on the
relative phase of the preceding action potential. If Tn+1 (x) is the arrival time of the
n + 1th action potential, then
1
dTn+1 (x)
=
.
dx
c(Tn+1 (x) Tn(x))
(2.38)
Note that an explicit version of kinematic theory can be derived using the singular
solution of the FN equations [322], assuming that recovery always occurs via a
phase wave at the value W . Suppose that wn (x) is the value of the recovery variable
along the wave front of the nth pulse when it is located at x. The instantaneous speed
of the front is thus c(wn ). The time between fronts of successive action potentials is
then given by
Tn+1 (x) Tn (x) =
w
wn
dw
+
G+ (w)
wn+1
w
dw
.
G (w)
(2.39)
Differentiating both sides with respect to x and using dTn /dx = 1/c(wn ) yields
1
1
1
1
dwn+1
dwn
=
+
.
G (wn+1 ) dx
G (wn ) dx
c(wn+1 ) c(wn )
(2.40)
This generates an iterative equation for wn (x), which can be solved to determine the
speed and arrival time of each successive action potential.

This section requires some basic definitions and results in functional analysis, in
particular, with regard to Banach spaces and the spectrum of linear differential operators; see appendix section 2.7.
78
2.4.1 Stability of Fronts in the Bistable Equation

In order to introduce some of the basic principles of wave stability, let us return to
the simple case of the scalar bistable Eq. (2.14). Let V ( ), = x ct, denote the
unique traveling front solution with speed c such that V ( ) 1 as and
V ( ) 0 as . In order to investigate the linear stability of such a solution,
we set
(2.41)
v(x,t) = V ( ) + ( ,t),
where is some small perturbation belonging to an appropriately defined Banach
space B (complete, normed vector space). It is convenient to use the moving coordinate so that we may see how the perturbation evolves in the moving frame of
the front. Substituting for v in (2.14) and keeping only terms linear in gives
= L
+c
+ f
(V ) ,
t
2
R,
t > 0.
(2.42)
Equation (2.42) takes the form of a linear equation with associated linear differential
operator L : D(L) B with domain D(L) B. Determining the linear stability of
the front thus reduces to the problem of calculating the spectrum (L) of L. That is,
the front will be asymptotically stable if 0 as t for all B, with
the norm on B. This is guaranteed if (L) lies strictly in the left-hand side of the
complex plane, that is, there exists > 0 such that Re ( ) for all (L).
The longtime asymptotics is then e t . However, differentiating both sides
of (2.17) with respect to gives
V + cV + f
(V )V LV = 0,
which implies that zero is an eigenvalue of L with associated eigenfunction V .
This is not a major issue, once one notices that V is the generator of infinitesimal
translations of the front solution:
V ( + h) = V ( ) + hV ( ) + O(h2).
Hence, such perturbations only cause a phase shift of the original front and can
thus be discounted. This motivates defining stability of the solution V in terms of
the stability of the family of waves obtained by rigid translations of V . In other
words, V is said to be stable if and only if v(x,t) = V ( ) + ( ,t) converges to
V ( + h) for some constant, finite h as t . This will hold provided that zero
is a simple eigenvalue of L and the remainder of the spectrum lies in a half-space
{ , Re ( ) } for some real > 0. It is important to note that the spectrum of
L will depend on the choice of Banach space B. Restricting the class of admissible
functions can push the spectrum to the left-half complex plane. However, this may
exclude classes of perturbations that are physically relevant. A common choice is
thus L2 (R), which includes all normalizable, continuous functions on R with respect
to the L2 norm:
=
79
| ( )|2 d < .
We now turn to the spectrum of the differential operator appearing in (2.42). As

explained in appendix section 2.7, (L) consists of isolated eigenvalues of finite
multiplicity combined with the essential spectrum, which includes any continuous
and residual spectrum. L is of the general form of the second-order operator considered in appendix section 2.7 (see (2.106)) with constant coefficient p = c and
x-dependent coefficient q(x) = f
(V (x)). Moreover, q(x) q as x with
q+ = f
(0) = a and q = f
(1) = 1 + a. Since the essential spectrum is located
to the left of the parabolas defined by (2.108), and q < 0, we deduce that the essential spectrum is bounded to the left of Re( ) = min{a, a 1} and thus does
not contribute to any instabilities. The stability of the front thus depends on the
eigenvalues of L, where
L + c + f
(V ) = ,
(2.43)
with L2 (R). Suppose that Re( ) 0 so ( ) e as with c.

(This follows from noting f
(V ) a as and analyzing the resulting
constant-coefficient characteristic equation.) Performing the change of variables
( ) = ( )ec /2 yields the modified eigenvalue problem

c2
= ,
(2.44)
L1 + f (V )
4
with L2 (R), since it also decays exponentially as | | . The useful feature of
the modified operator is that it is self-adjoint, implying that any eigenvalues in the
right-half complex plane are real. Multiplying both sides of the self-adjoint eigenvalue Eq. (2.44) by and integrating over R, we have

c2
2d =
2 f
(V )
2 d .
(2.45)
4
Recall that V is an eigenfunction of L with = 0, so that if ( ) = V ( )ec /2 ,

then + ( f
( ) c2/4) = 0. Hence, (2.45) can be rewritten as
d =
2
2 +

2 2
=
2
+
d
2

d
2
=
( / ) d .
d

2
d
80
This last result implies that 0, and if = 0, then = V . We conclude

that there are no eigenvalues in the right-half complex plane and = 0 is a simple
eigenvalue. Thus the traveling front of the scalar bistable equation is stable.
2.4.2 The Evans Function

Determining the stability of traveling pulse solutions of the FN equations (2.1) or
the HodgkinHuxley equations (1.8) is much more complicated. One general result, however, is that the discrete spectrum of the differential operator obtained by
linearizing about a traveling wave solution may be associated with the zeros of a
complex analytic function known as the Evans function. Indeed, Evans [177] originally developed the formalism within the context of the stability of solitary pulses
in HodgkinHuxley-type equations for action potential propagation. Since then, the
Evans function construction has been extended to a wide range of PDEs; see the review [551]. The basic construction of the Evans function can be illustrated relatively
easily by considering a higher-dimensional version of the bistable equation [173].
Consider a general class of reactiondiffusion equations of the form
u
2u
= D 2 + F(u),
t
x
(2.46)
where u(x,t) RN and F : RN RN . Moreover, D is assumed to be a diagonal

matrix with positive definite entries corresponding to the diffusion coefficients of
the various component fields. Suppose that the system exhibits bistability, that is,
there are two stable fixed points u = u j , j = 1, 2, with F(u j ) = 0. We will assume
that there exists a traveling front solution U( ) with speed c that connects u1 and
u2 . Linearizing about the wave solution along identical lines to the scalar case by
setting u(x,t) = U( ) + p( )e t leads to the eigenvalue problem
Lp Dp + cp + F(U)p = p,
(2.47)
where F(U) denotes the matrix with components Fi / U j . It is convenient to

rewrite this as a system of 2N first-order equations

p
0
In
p
=
.
(2.48)
q
q
D1 ( F(U)) cD1
Any eigensolution of this equation must satisfy the asymptotic conditions
lim (p( ), q( )) = (0, 0).
Setting z = (p, q)T Rn , n = 2N, the associated ODE takes the general form
T ( )z( )
dz
A( ; )z = 0,
d
(2.49)
81
with A( ; ) = A( ) + B( ). Thus we have a family of linear operators T ( ) :

D B parameterized by . We take B = L2 (R, Rn ) and D to be the space of
admissible functions such that z B, T ( )z B. The basic form of the linear
Eq. (2.49) holds for a wide range of PDEs supporting solitary traveling waves [551].
The discrete spectrum of the operator L thus corresponds to the values of for
which T ( ) is not invertible.
An important concept for analyzing (2.49) is that of exponential dichotomies.
First, consider the linear constant-coefficient equation
dz
= A( )z,
d
(2.50)
for which A( ) is independent of . Suppose that the matrix A( ) is hyperbolic,

that is, all its eigenvalues have nonzero real part. We can then decompose Rn (or its
complexification) as
Rn = Es ( ) Eu( ),
where Es,u ( ) are the generalized stable and unstable eigenspaces of the matrix
A( ). Thus Es is spanned by n+ ( ) eigenfunctions that decay exponentially as
and Eu is spanned by n ( ) eigenfunctions that decay exponentially as
with n+ ( ) + n( ) = n. The notion of exponential dichotomies can now be
extended to (2.49) by noting that
lim A( ; ) A ( )
(2.51)
where A correspond to the matrix appearing in (2.48) in the limits U( ) u1

and U( ) u2 , respectively. Moreover, the spectral properties of T ( ) can be
expressed in terms of these exponential dichotomies. We summarize the main results
below:
Equation (2.49) is said to have an exponential dichotomy on R+ if and only if the
matrix A+ ( ) is hyperbolic. Let V+ ( ) denote the linear subspace spanned by
solutions of (2.49) that decay as . The codimension of V+ ( ) is defined
to be the Morse index i+ ( ) of the exponential dichotomy on R+ , and i+ ( ) =
dim Eu+ ( ).
Equation (2.49) is said to have an exponential dichotomy on R if and only if
the matrix A ( ) is hyperbolic. Let V ( ) denote the linear subspace spanned
by solutions of (2.49) that decay as . The dimension of V ( ) is defined
to be the Morse index i ( ) of the exponential dichotomy on R , and i ( ) =
dim Eu ( ).
is in the discrete spectrum if and only if A ( ) are both hyperbolic with the
same Morse index i+ ( ) = i ( ) such that V+ ( ) V ( ) = {0}.
is in the essential spectrum if either at least one of the two asymptotic matrices
A ( ) is not hyperbolic or else if both are hyperbolic but their Morse indices
differ.
82
In most applications the essential spectrum ess lies in the left-half complex plane
and thus does not contribute to instabilities of a wave solution. Therefore, suppose
that
/ ess . It then follows that i+ ( ) = i ( ) = k, say. (For the multicomponent
bistable equation k = N = n/2.) In order to construct the Evans function, introduce
a basis for the subspaces V ( ) according to
V ( ) = span {Q
1 , Q2 , . . . , Qk },
+
+
V+ ( ) = span {Q+
1 , Q2 , . . . , Qnk }
where each Qj ( ) is an n-dimensional basis vector. Now form the n n matrix

M ( ), in which the first k columns are given by the vectors Qj ( ), j = 1, . . . , k,
and the next n k columns are given by Q+j , j = 1, . . . , n k. The Evans function is
then defined according to
(2.52)
E ( ) = det M (0 ),
for an arbitrary point 0 which can be taken to be zero. The Evans function has a
number of important properties. First, is an eigenvalue if and only if E ( ) = 0.
Second, if is a zero of E ( ), then the order of this zero is equal to the algebraic
multiplicity of viewed as an eigenvalue. Third, the Evans function is analytic. The
first property is simple to establish. For E ( ) = 0 if and only if det M (0 ) = 0, and
the latter holds if and only if there exist constant coefficients c+

i , c j such that
k
nk
j=1
i=1
cj Qj (0 ) + c+i Q+i (0 ) = 0,
that is,
k
nk
j=1
i=1
cj Qj (0 ) = c+i Q+i (0 ).
Hence, is a zero of E ( ) if and only if V+ ( ) and V ( ) have a nonzero intersection, which means that is an eigenvalue, since the corresponding eigensolution
decays at both = and = . Finally, note that one of the powerful features
of the Evans function construction is that it can be applied to a wide variety of wave
phenomena beyond fronts, including pulses, periodic wave trains, and multi-bump
pulses [551]. In a sense, a single pulse is a special case of a front, since the exponential dichotomies on R+ and R are the same, that is,
lim A( ; ) A0 ( ),
with A0 ( ) evaluated at the same resting state.

Finally, a word of caution: linear stability does not necessarily imply nonlinear
stability. When considering perturbations about a traveling wave solution of a nonlinear PDE, p(x,t) = u(x,t) U( ), one can decompose the PDE as
p
= Lp + N (p),
t
(2.53)
2.5 Myelinated Axons and Discrete Diffusion
83
where N (p) = O(|p|2 ). Determining linear stability of the traveling wave in terms
of the spectrum (L) assumes that the perturbation p remains small with respect
to the given norm. A challenging mathematical problem is determining whether or
not this is true. In the simpler case of ODEs one can use the stable manifold the-
Vn
In
Vn+1
In+1
nodes of Ranvier
Fig. 2.10 Schematic diagram of a myelinated axon. Currents in myelinated region are confined to
the axial direction. Potentials at the nodes are governed by active currents
orem to show that linear (in)stability implies nonlinear (in)stability in the case of
hyperbolic fixed points. In the case of non-hyperbolic fixed points one has to use
the center manifold theorem and bifurcation theory. A number of techniques have
been used to study nonlinear stability of traveling waves, including center manifold
reductions, Lyapunov functionals, and energy methods [551]. For example, the nonlinear stability of traveling wave solutions of the reactiondiffusion system (2.46)
can be proven using a center manifold reduction. Nonlinear stability theory is also
important if one wishes to determine what happens when a traveling wave becomes
unstable, since perturbations grow and the linear approximation breaks down.

Many vertebrate axons are coated with a lipid material called myelin, which is made
up of the membranes of glial cells (see Sect. 4.5) that wrap around the axon many
times to form a thick insulating layer. This wrapping increases the effective membrane resistance and decreases the membrane capacitance by a factor of around
100. At regularly spaced intervals the axon is exposed to the extracellular medium
at the so-called nodes of Ranvier, where there is a high density of sodium channels;
see Fig. 2.10. The length L of myelin sheath between successive nodes is typically
12 mm, and the width l of a single node of Ranvier is around 1 m. Propagation
of an action potential along a myelinated axon is considerably faster than along a
nonmyelinated axon. In terms of the cable equation, this can be understood as consequence of the fact that the transmembrane currents in the myelinated sections are
84
negligible so that the myelinated sheath acts like a simple resistor. In effect, the
action potential does not propagate continuously along the axon but rather jumps
from node to node as a saltatory (leaping) wave.
Wave propagation along a myelinated axon can be modeled by spatially discretizing the diffusion term in the spatially extended HodgkinHuxley model (1.8).
Suppose that the membrane voltage does not vary within each node of Ranvier, that
is, the nodes are isopotential and denote the voltage of the nth node by Vn . Treating each myelin sheath as a pure Ohmic resistor with resistance rL, where r is the
intracellular resistance per unit length, the axial current between nodes n and n + 1 is
In+1 =
1
(Vn+1 Vn).
rL
(2.54)
Conservation of current at the nth node of Ranvier then implies that the total transmembrane current into the node is

Vn
1
+ Iion = In In+1 = (Vn+1 2Vn + Vn1),
(2.55)
2 al Cm
t
rL
where a is the radius of the cable. It follows that
Vn
= Iion + D(Vn+1 2Vn + Vn1),
t
(2.56)
with coupling coefficient

D=
2
Rm
= m .
(2 ar)lLm
lLm
We have used R = a2 r, m = RmCm and m = (Rm a/2R)1/2; see Sect. 1.4.1.
2.5.1 The Discrete Bistable Equation

In order to investigate the effects of myelination on propagation speed, let us consider the simpler case of the spatially discrete bistable equation
Vn
= f (Vn ) + D(Vn+1 2Vn + Vn1),
t
(2.57)
where f (V ) = V (V a)(1 V). Proving the existence of traveling wave solutions

of (2.57) is nontrivial [699]. However, assuming that such a solution does exist, one
can use perturbation methods to estimate the wave speed [317, 321]. A traveling
wave solution is defined according to the iterative relationship Vn+1 (t) = Vn (t + d )
where d represents the time delay for the wave to jump between successive nodes.
The corresponding invariant wave profile (t) satisfies a delay differential equation
that is obtained by substituting Vn (t) = (t) into (2.57):
d
= D( (t d ) 2 (t) + (t + d )) + f ( (t)).
dt
(2.58)
85
Suppose that we fix the time units by setting m = 1. If the function (t) is
sufficiently smooth and d is sufficiently small, then we can expand (t d ) as
a Taylor series in d . Keeping only terms up to d2 yields the leading-order equation
Dd2 tt t + f ( ) = 0.
(2.59)
This is identical in form to the bistable Eq. (2.17) for a traveling front of (dimensionless) wave speed c, provided that we set = ct and Dd2 = c2 . It follows
is the traveling front solution of the continuous bistable equation and
that ( /c)
c = (1 2a)/ 2. The wave speed of the saltatory wave (in physical time units) is
then

D
L+l
L + l c m
c =
= (L + l)c
=
.
(2.60)
d
m
lL m
Comparison
with (2.20) shows that myelination increases wave speed by a factor
(L + l)/ lL 10 assuming that L 100l.
Higher-order corrections to the wave speed can be determined using regular perturbation theory [317, 321]. Thus, in terms of the small parameter = D1 , we
introduce the series expansions (with m = 1)
(t) = 0 (t) + 1 (t) + . . .,
d 2 =
+ 2 1 + . . .
c2
(2.61)
Substituting into the discrete bistable Eq. (2.58) and collecting terms of equal
powers in yields a hierarchy of equations for n , with 0 satisfying (2.59) for
Dd2 = c2 , and
L1
(4)
1 1
+ f
(0 )1 = 0 4 1 0
.
2
c
12c
(2.62)
Here L denotes a linear differential operator acting on the space L2 (R) of squareintegrable function on R. The operator L is not invertible, since L0
= 0, which
follows from differentiating (2.58) with respect to t. (A similar operator arose in our
discussion of wave stability in Sect. 2.4.) It follows from the Fredholm alternative
(see appendix section 2.7) that a solution for 1 exists if and only if the right-hand
side of (2.62) is orthogonal to the null-space of the adjoint operator L . The latter is
L V =
V + V
+ f
(0 )V ,
c2
(2.63)
which has a one-dimensional null-space spanned by V (t) = ec t 0

(t). We thus
have a solvability condition for the leading-order correction 1 to the delay:
(4)

0
c2 t
e
0 (t)
+ 1 0 dt = 0.
12c4
86
Once 1 has been determined, the propagation speed is (in physical time units)

D
1 c2
c = c(L + l)
+ O([c2 /Dm ]2 ) .
(2.64)
1
m
2Dm
2.5.2 Propagation Failure

The above perturbation analysis establishes that for sufficiently large coupling D,
there are traveling front solutions that approach the speed of the appropriately scaled
continuous model. Another important property of the discrete bistable equation
is that if D is sufficiently small, then wave propagation failure occurs,
reflecting

the fact that there are nontrivial standing front solutions even when 01 f (v)dv > 0
[165, 183, 317, 318]. Here we describe one method for estimating the critical value
of coupling Dc below which propagation fail, which has been developed by Keener
[318] using averaging theory. The first step is to rewrite the discrete bistable equation in the form
(2.65)
(1 + g
(x))[vt f (v)] = Dvxx ,
where g(x) is the periodic sawtooth function
g(x) =
and
L
x,
2
0 < x < L,
g(x + nL) = g(x),
1 + g
(x) = L (x nL).
(2.66)
(2.67)
Equation (2.65) implies that vxx = 0 between nodes, that is, v(x,t) = An x + Bn for
nL < x < (n + 1)L. Matching v(x,t) with Vn (t) and Vn+1 (t) at both ends shows that
An = L1 [Vn+1 Vn ] and hence vx ((n + 1)L,t) = L1 (Vn+1 Vn). If we now integrate
(2.65) over the interval [nL, (n + 1)L), then we recover the discrete bistable equation
t Vn f (Vn ) =
D
[vx ((n + 1)L,t) vx(nL,t)] = D[Vn+1 2Vn + Vn1].
L
In order to proceed, we will smooth out the function g(x) by taking

L
(x nL)2
exp
1 + g
(x) =
,
2 2
2 2 n
(2.68)
(2.69)
so that 1 + g
(x) > 0 for all x. At the end of the calculation we will take the limit
0 to recover the sawtooth function. We then have the scalar bistable equation
with an inhomogeneous diffusion coefficient,
vt = f (v) +
D
vxx .
1 + g
(x)
87
Following Keener [318], it is convenient to carry out the coordinate transformation

y = y(x) x + g(x) L/2 so that

vt = f (v) + y [1 + g
(x)]y v .
If g(x) is a sawtooth-like function, then y is a steplike function with y = nL for
nL < x < (n+1)L in the limit 0. Hence 1+g
(y(x)) = n (y(x)nL) blows up
for all nL < x < (n + 1)L, whereas 1 + g
(y(x)) = 0 when x = nL. That is, 1 + g
(x) =
1/(1 + g
(y(x))), so that (after rewriting y as x),

D
v
.
vt = f (v) + x
x
1 + g
(x)
Fixing the spatial units by setting L = 1 (having already non-dimensionalized time),
and using the fact that waves exist
for sufficiently large D, we introduce the
small
dimensionless parameter = 1/ D. Rescaling space according to x x/ D, we
finally obtain the modified bistable equation

1
vx .
vt = f (v) + x
(2.70)
1 + g
(x/ )
Thus the problem of wave propagation failure for the discrete bistable equation has
been reduced to the problem of calculating the mean wave speed c(
) of a wavelike
solution of an inhomogeneous continuous bistable equation and determining how
c(
) vanishes as increases (the coupling D decreases).
Equation (2.70) is now in a suitable form to apply the averaging method of
Keener [318]; see below. The basic result is that the wavelike solution takes the
form
(2.71)
v(x,t) = V (x (t)) + O( ),
where V (x) is the wave solution in the homogeneous case (g = 0). The phase (t)
evolves according to
d
= c ( / )
(2.72)
dt
with c the speed of the unmodulated wave,
( / ) =
and
g
([ + ]/ )V
( )V
( )ec d ,

ec V
( )2 d .
(2.73)
(2.74)
Equation (2.72) implies that the solution is not translationally invariant; rather, it
moves with a time-dependent velocity
. If c ( / ) is strictly positive, then

(t) is a positive, periodic function of t with period
88
T=

0
d
.
c ( / )
(2.75)
The mean speed of the wave is c = /T . On the other hand, if c ( / ) vanishes

for some , then propagation failure is expected to occur:
Averaging method for discrete bistable equation. Standard applications of the averaging theorem in dynamical systems theory [248] only apply to structurally stable solutions, whereas
traveling wave solutions are not structurally stable. Therefore, it is necessary to consider a
modified averaging procedure as outlined by Keener [318, 319]. The first step is to rewrite
(2.70) as the second-order system
vx = (1 + g
(x/ ))u,
ux = vt f (v).
(2.76)
Introducing the exact change of variables v = w + ug(x/ ) yields the new system
wx = u ux g(x/ ),
ux = wt + ut g(x/ ) f (w + ug(x/ )).
(2.77)
It can be seen that if we ignore terms of order , then the system of equations is independent
of x/ . This lowest-order averaged system reduces to the standard bistable equation, which
we know supports a traveling front solution V (x ct). Including the inhomogeneous factors g(x/ ) means that the system is no longer translationally invariant. However, we can
look for solutions that are in some sense close to a traveling front by going to a traveling
coordinate system by setting = x (t), with
w = u u g([ + ]/ )
(2.78a)
u = w u g([ + ]/ ) f (w + ug([ + ]/ ).
(2.78b)
We now seek a perturbative solution of the form

w = w0 ( ) + w1 ( ) + . . .,
v( ) = v0 ( ) + v1 ( ) + . . .,

(t) = c + 1
(t) + . . .
Substituting into (2.78) and collecting terms of equal powers in yields a hierarchy of
equations, the first two of which are
w0 = u0 ,
u0 = cu0 f (w0 ),
(2.79)
which recovers the homogeneous bistable equation for w0 , and
w1 u1 = ( u0 )g([ + ]/ ),
u1 + f
(w0 )w1 + cu1 = 1
u0 f
(w0 )u0 g([ + ]/ ).
(2.80a)
(2.80b)
Let us take the solution of the lowest-order equation to be the traveling front solution V ( )
constructed in Sect. 2.2: w0 = V, u0 = V
. The next-order system of Eq. (2.80) can be rewritten in the vector form

1
w1
hw
,
(2.81)
L
=
, L=
u1
1V + hu
f (V ) + c
with hw and hu determined by inhomogeneous terms on the right-hand side of (2.80a) and
(2.80b), respectively. Following our analysis of the linear operator (2.62), we know that
the matrix operator appearing in (2.81) has a null-space spanned by (V
,V
). Similarly, the
adjoint operator

f
(w0 )
L =
1
+ c
89
has the null vector ec (V
,V
). Hence, applying the Fredholm alternative, we see that

(t) must satisfy
1
(t)
ec V
( )2 d =
=
=
ec [V
( )hw ( ) +V
( )hu ( )d
ec [V
( )2 f
(V )V
( )2 ]g([ + ]/ )d
g([ + ]/ )

d
V ( )V
( )ec d ,
d
g
([ + ]/ )V
( )V
( )ec d ,
after using V
+ cV
+ f
(V )V
= 0 and performing integration by parts. We thus obtain
the phase Eq. (2.72) with (t) = c + 1 (t).
It turns out that solving the phase equation in the case of a cubic nonlinearity
is rather involved [318]. Therefore, for the sake of illustration, we will consider
the simpler case of the piecewise linear function (2.21). There then exists a unique
traveling front solution of the homogeneous bistable equation given by (2.23) with
corresponding wave speed (2.24). Substituting (2.23) into (2.74) gives

ec e2 d + ([a 1]+)2
0

1
1
2
= (a )
c + 2+ c + 2

= 2 a a2(a a2),
= (a )2
0
ec e2+ d
where we have used the results a = (a 1)+,

1
,
c + 2 = c2 + 4 =
a a2

1
a1
c + c2 + 4 =
.
2
a a2
Similarly, substituting (2.23) into (2.73) gives, to leading order in ,

1 2 3
a
g ([ + ]/ )ec e2 d
0
0
1
+ [a 1]2+3
g
([ + ]/ )ec e2+ d

0
1 2 3
g ([ + ]/ )d + [a 1]2+3
g
([ + ]/ )d
a
2
2
= (a ) [+ ] g( / ) + O( )
g( / ) + O( 2 ).
=
2[a a2]
( / ) =
Applying this to the sawtooth function (2.66), the phase Eq. (2.72) reduces to
90
0.3
_
avergae wave speed c
c = 0.3
0.2
c = 0.2
0.1
c = 0.1
0.1
0.2
0.3
0.4
Fig. 2.11 Illustrative sketch of mean wave speed as a function of for the discrete bistable
Eq. (2.65) with L = 2

1
d
= c
,
dt
2[a a2] 2
and
T = 2 ln
c + /
c /

,
= 4[a a2].
We deduce that the mean wave speed is

c =
1
.

2 ln c+ /
(2.82)
c /
This predicts propagation failure when c. In particular, propagation failure is

more likely to occur as increases, which is equivalent to reducing the coupling
strength D. Finally, note that the cubic nonlinearity yields different behavior, both
qualitatively and quantitatively. The main reason for this is that the above averaging
procedure results in an expression for the mean wave speed that involves exponentially small terms of the form e / [318]. This has two major implications. First,
it is necessary to include higher-order terms in the perturbation analysis in order to
obtain sufficient accuracy. Second, the rapid increase in e / as increases can
result in a sharp transition to propagation failure at relatively small values of , as
illustrated in Fig. 2.11.
2.6 Stochastic Traveling Waves
91

In Sect. 1.5 we showed how the stochastic opening and closing of a finite number of ion channels can generate a multiplicative noise term in a space-clamped,
conductance-based model of a neuron. The membrane voltage then evolves according to a stochastic differential equation (SDE) such as (1.128). This suggests that
ion channel fluctuations could also affect the propagation of action potentials in a
spatially extended model. Such an observation motivates the more general issue of
how to analyze traveling wave solutions of stochastic partial differential equations
(SPDEs). In this section we review a formal perturbation method for analyzing traveling front solutions in the case of weak noise [14, 494, 546, 557], which we apply
to a stochastic version of the bistable equation. (The rigorous treatment of SPDEs is
much more complicated than SDEs because one has to keep track of the regularity
of solutions with respect to both time and space. Nevertheless, there have been some
recent mathematical studies of a stochastic version of spatially extended excitable
membranes based on the HodgkinHuxley equations [17, 88].)
Consider a scalar SPDE of the form
2
V (x,t) = 2 V (x,t) + f (V (x,t)) + g(V (x,t)) (x,t),
t
x
(2.83)
Here (x,t) is a Gaussian random function with zero mean and correlation
(x,t) (x
,t
) = 2C([x x
]/ ) (t t
)
(2.84)
The parameter is the spatial correlation length of the noise such that C(x/ )
(x) in the limit 0, and determines the strength of the noise, which is assumed
to be weak. Note that we can formally set (x,t)dt = dW (x,t), where W (x,t) is a
space-dependent Wiener process with zero mean and dW (x,t)dW (x
,t) = 2C([x
x
]/ ). We have also included a multiplicative noise factor g(V ), and based on our
analysis of channel fluctuations, we treat the multiplicative noise in the sense of
Stratonovich (see Sect. 1.7). The starting point of the perturbation method is the
observation that multiplicative noise in the Stratonovich sense leads to a systematic
shift in the speed of the front (assuming a front of speed c exists when = 0) [14].
This is a consequence of the fact that g(V ) = 0 even though = 0. The former
average can be calculated using Novikovs theorem [465, 486]:
1/2 g(U) = C(0)g

(U)g(U)dt,
(2.85)
Note that in the limit 0, C(0) 1/ x where x is a lattice cutoff, which can
be identified with the step size of the spatial discretization scheme used in numerical
simulations.
Novikovs theorem. Suppose that X(s) is a Gaussian random function with zero mean and
correlation
X(s)X(s
) = C(s, s
),
then for any functional L [h],
92

L [X]
C(s, s
)
X(s)L [X] =
ds
.
X(s
)
Applying this theorem to g(U) , we have

g(V ) = 2

V (x,t)
C([x x
]/ ) g
(V (x,t))
dx
.
(x
,t)
The stochastic voltage is a functional of the noise term (x,t), as can be seen by formally
integrating Eq. (2.83) with respect to time:
2
V (x,t) =
V
(x,t
)
+
f
(V
(x,t
))
+
g(V
(x,t
))
(x,t
)
H(t t
)dt
.
2
x
It follows that
V (x,t)
= g(V (x,t
))H(t t
) (x x
),
(x
,t
)
which yields (2.85) after using H(0) = 1/2.

An alternative derivation of (2.85) is based on Fourier transforming (2.83) [546]. It is convenient to restrict x to a bounded domain, L/2 x L/2, and to impose periodic boundary
conditions. We can then introduce the discrete Fourier series
V (x,t) =
1
eikn xVn (t),
L
n
W (x,t) =
1
eikn xWn (t)
L
n
(2.86)
with kn = 2 n/L and Wn (t), an independent Wiener process, such that

dWn (t) = 0,
dWn (t)dWm (t) = 2Lm+n,0 dt.
(2.87)
Fourier transforming (2.83) gives in differential form

dUn (t) = [kn2Vn (t) + Fn (t)]dt +
1/2
L
gnm (t)dWm (t),
(2.88)
where Fn , gn are the Fourier coefficients of the time-dependent functions F U(t) and
g U(t), respectively. The associated Stratonovich FokkerPlanck equation takes the form
[209] (see also Sect. 1.7)
=
[(kl2Vn (t) + Fl (t))P] +
glq
gm+q P.
t
L l,m,q Vl
Vm
l ul
(2.89)
Multiplying both sides of this equation by Vn and integrating with respect to Vm , integer m,
leads to the following evolution equation for the mean:

gnq
dVn
= kn2 Vn + Fn +

gm+q .
(2.90)
dt
L m,q Vm
Finally, taking the inverse transform of (2.90) gives
#
2
"
dU(x,t)
=
g(U(x,t))g
(U(x,t)) ,
V (x,t) + F(V (x,t)) +
dt
x2
x
(2.91)
where we have used the result gn / Um = [g

(U)]nm . Note that it is necessary to introduce
a cutoff in the frequencies, which is equivalent to introducing a fundamental lattice spacing
of x. Alternatively, the multiplicative noise can be taken to have a small but finite corre-
93
lation length in space so that C(0) = 1/ x. Comparison of (2.90) with the mean of (2.83)
yields the desired result.
Following [14], it is convenient to rewrite (2.83) so that the fluctuating term has
zero mean:
where
and
2
V (x,t) = 2 V (x,t) + h(V (x,t)) + R(V, x,t),
t
x
(2.92)
h(V ) = f (V ) + C(0)g
(V )g(V )
(2.93)
R(V, x,t) = g(V ) (x,t) 1/2C(0)g

(U)g(U).
(2.94)
The stochastic process R has zero mean (so does not contribute to the effective wave
speed) and correlation:
R(V, x,t)R(V, x
,t
) = g(V (x,t)) (x,t)g(V (x
,t
) (x
,t
) + O( 1/2).
(2.95)
The next step in the analysis is to assume that the fluctuating term in (2.92) generates two distinct phenomena that occur on different time scales: a diffusive-like
displacement of the front from its uniformly translating position at long time scales
and fluctuations in the front profile around its instantaneous position at short time
scales [14, 494, 546, 557]. In particular, following [14], we express the solution V
of (2.92) as a combination of a fixed wave profile V0 that is displaced by an amount
(t) from its uniformly translating position = x ct
and a time-dependent fluctuation in the front shape about the instantaneous position of the front:
V (x,t) = V0 ( (t)) + 1/2 ( (t),t).
(2.96)
The wave profile V0 and associated wave speed c are obtained by solving the modified deterministic equation
c
dV0 d 2V02
+
+ h(V0( )) = 0.
d
d 2
(2.97)
Both c and V0 depend nontrivially on the noise strength due to the dependence of
the function h; see (2.93). As an example, suppose that f (V ) = V (V a)(1 V ) and
g(V ) = V (1 V ) [14]. The form of multiplicative noise is chosen so that it preserves
the stationary states V = 0, 1. Hence, the noise is most important in regions close to
the front but vanishes asymptotically at . The effective nonlinearity h is also a
cubic with
h(V ) = V (1 V )(a
k
V ),
a
= a C(0),
k
= 1 2 C(0).
Thus, from the

analysis of the bistable equation in Sect. 2.2, we find that
c = (k
2a
)/ 2k
.
94
It turns out that if V0 is chosen to satisfy (2.97) then to leading order, the stochastic variable (t) undergoes unbiased Brownian motion (a Wiener process):
(t) = 0,
(t)2 = 2D( )t
(2.98)
with a diffusion coefficient D( ) = O( ) (see below). Thus (t) represents the effects of slow fluctuations, whereas represents the effects of fast fluctuations. Note
that since (t) = O( 1/2 ), (2.96) implies that V (x,t) = V0 (x ct)+
O( 1/2 ). Hence,
+ O( 1/2 ).
averaging with respect to the noise shows that V (x,t) = V0 (x ct)
Thus, in the case of weak noise, averaging over many realizations of the stochastic
wave front generates a mean front whose speed is approximately equal to c.
Calculation of diffusion coefficient. Substitute the decomposition (2.96) into (2.92) and expand to first order in O( 1/2 ) (exploiting the fact that the usual rules of calculus apply in
the case of Stratonovich noise, see Sect. 1.7):

[c + (t)]V0
(t ) + 1/2 t (t ,t) [c + (t)] (t ,t)
= h(V0 (t ) + 1/2 h
(V0 (t )) (t ,t)

+2 V0 (t ) + 1/2 (t ,t) + 1/2 R(V0 (t ), x,t) + O( ),
where t (t). Imposing (2.97) and dividing through by 1/2 then gives
( ,t)
+ L ( ,t) = 1/2V0
( ) (t) + R(V0 ( ), ,t) + O( 1/2),
t
(2.99)
where L is the non-self-adjoint linear operator

( ) + h
(V0 ( ))A( )
L A( ) = A
( ) + cA
(2.100)
for any function A( ) L2 (R). We have also made the approximation t , since
(t) = O( 1/2 ). The linear differential operator L has a zero eigenvalue with associated
eigenfunction V0
( ), which can be seen by differentiating (2.97) with respect to , and reflects the fact that the underlying system is equivariant with respect to uniform translations.
We then have the solvability condition for the existence of a bounded solution of (2.99),
namely, that the inhomogeneous part on the right-hand side is orthogonal to all elements
of the null-space of the adjoint operator L . The latter is defined with respect to the inner
product

B( )L A( )d =
(2.101)
L B( ) A( )d
where A( ) and B( ) are arbitrary integrable functions. Hence,

( ) + h
(V0 ( ))B( ).
L B( ) = B
(
) cB
(2.102)
The linear operator L also has a zero eigenvalue, with corresponding eigenfunction
V ( ) = ec V0
( ). Thus taking the inner product of both sides of (2.99) with respect to
V ( ) leads to the solvability condition

V ( ) V0
( ) (t) + 1/2 R(V0 , ,t) d = 0,
which implies that (t) satisfies the stochastic differential equation (SDE)
(2.103)
2.7 Appendix: Linear Differential Operators
95
d (t) = 1/2
V ( )dR(V0 , ,t)d
V ( )V0
( )d
(2.104)
Using the lowest-order approximation dR(V0 , ,t) = g(V0 ( ))dW ( ,t), we deduce that
(for (0) = 0) (t) is a Wiener process with diffusion coefficient

D( ) =
V ( )V (
)g(V0 ( ))g(V0 (
))C([
]/ )d d
. (2.105)

2
V ( )V0
( )d
Although the above analysis is based on a formal perturbation calculation, rather

than rigorous analysis, it does appear to capture well the effects of weak external
noise on front propagation in a variety of reactiondiffusion models [486, 546]. In
Sect. 7.4, we will show how the method can be extended to study stochastic traveling
waves in nonlocal neural field equations, which represent large-scale continuum
models of spatially structured neural networks. Note, however, that one class of front
solution where the method breaks down is a so-called pulled front, which propagates
into an unstable rather than a metastable state and whose dynamics is dominated by
the linear spreading of small perturbations within the leading edge of the front [544].
One well-known reactiondiffusion model that supports pulled fronts is the Fisher
KPP equation [191, 345]. As we will describe later, pulled fronts also arise in a
PDE model of CaMKII translocation waves along spiny dendrites [72, 161] (Sect.
3.2), in certain neural field models (Sect. 7.4), and in a model of protein aggregation
(Sect. 9.6).

Throughout this book, we will encounter linear operators acting on some function
space. As already demonstrated in this chapter, linear differential operators arise
when analyzing the stability of a traveling wave solution of some PDE, or when
carrying out a formal perturbation expansion. In this appendix, we summarize some
of the basic results regarding linear differential operators acting on a function space,
viewed from the perspective of a linear map acting on an infinite-dimensional vector
space. For simplicity, we will restrict ourselves to real-valued functions f : R R,
although it is straightforward to generalize the results to complex-valued functions.
2.7.1 Function Spaces

Consider the set of all real functions f (x) on the interval [a, b]. This is a vector
space over the set of real numbers: given two functions f1 (x), f2 (x) and two real
96
numbers a1 , a2 , we can form the sum f (x) = a1 f1 (x) + a2 f2 (x) such that f (x) is also
a function on [a, b]. Either on physical grounds or for mathematical convenience,
we usually restrict ourselves to a subspace of functions that are differentiable to
some given order. For example, the space of functions on [a, b] with n continuous
derivatives is denoted by Cn [a, b], and the space of analytic functions (those whose
Taylor expansion converges to the given function) is denoted by C [a, b].
In order to describe the convergence of a sequence of functions fn , n = 1, 2, . . . to
a limit function f , we need to introduce the concept of a norm, which is a generalization of the usual measure of the length of a finite-dimensional vector. The norm
f of a function f is a real number with the following properties:
(i) Positivity: f 0, and f = 0 if and only if f = 0
(ii) The triangle inequality: f + g f + g
(iii) Linearity: f = | | f for R
Common examples of norms are the sup norm
f = sup | f (x)|,
x[a,b]
and the L p norm

f p =

a
1/p
| f (x)| dx
p
Given the L p norm, we can introduce another important function space L p [a, b],
which is the space of real-valued functions on [a, b] for which f p < . However,
there is one subtlety here, namely, that it is possible for f = 0 without f being
identically zero. For example, f may vanish at all but a finite set of points (set of
measure zero). This violates the positivity property of a norm. Therefore, one should
really treat elements of L p [a, b] as equivalence classes of functions, where functions
differing on a set of measure zero are identified.
Given a normed function space, convergence of a sequence fn f can be expressed as
lim fn f = 0.
n
In the case of the sup norm, fn is said to converge uniformly to f , whereas for
the L1 norm, it is said to converge in the mean. An important property of a function
space is that of being complete. First, consider the following definition of a Cauchy
sequence: A sequence fn in a normed vector space is Cauchy if for any > 0, we
can find an integer N such that n, m > N implies that fm fn < . In other words,
elements of the sequence become arbitrarily close together as n . A normed
vector space is then complete with respect to its norm if every Cauchy sequence
converges to some element in the space. A complete normed vector space is called a
Banach space B. In many applications, the norm of the function space is taken to be
the so-called natural norm obtained from an underlying inner product. For example,
if we define an inner product for L2 [a, b] according to
f , g =
97
b
a
f (x)g(x)dx,
then the L2 [a, b] norm can be written as

f 2 =

f , f .
A Banach space with an inner product is called a Hilbert space H .
2.7.2 Fredholm Alternative Theorem

In the case of one-dimensional traveling wave solutions, it is more natural to consider functions on the real line R rather than a finite interval. Suppose that L is a
linear differential operator acting on a subspace of L2 (R), which we denote by the
domain D(L). Linearity of the operator means that for f1 , f2 D(L) and a1 , a2 R,
L(a1 f1 + a2 f2 ) = a1 L f1 + a2 L f2 .
Given the standard inner product on L2 (R), we define the adjoint linear operator L
according to
f , Lg = L f , g,
f , g D(L).
The operator is said to be self-adjoint if L = L. Note that, in practice, one determines L using integration by parts. For functions defined on finite intervals, this
generates boundary terms that only vanish if appropriate boundary conditions are
imposed. In general, this can result in different domains for L and L . Therefore,
the condition for self-adjointness becomes L = L and D(L) = D(L ). Given a differential operator L on L2 (R), we can now state the Fredholm alternative theorem:
The inhomogeneous equation
Lf = h
has a solution if and only if
h, v = 0
for all v satisfying L v = 0.
2.7.3 Spectrum of a Linear Differential Operator

Let B be a Banach space and L : D(L) B be a linear operator with domain
D(L) B. For any complex number , introduce the new operator
L = L I,
98
where I is the identity operator on B. If L has an inverse, then R (L) = L1

is
called the resolvent of L. Given these definitions, is said to be a regular point for
L if the following hold:
(i) R exists.
(ii) R is bounded.
(iii) R is defined on a dense subset of B.
The spectrum (L) is then the set of points that are not regular, which generally
consists of three disjoint parts:
(a) The point spectrum of eigenvalues is the set of values of for which R does
not exist.
(b) The continuous spectrum is the set of values of for which R exists but is
unbounded.
(c) The residual spectrum is the set of values of for which R exists, is bounded,
but is not defined on a dense subset of B.
The continuous spectrum and residual spectrum are contained in the essential spectrum, which is any point in (L) that is not an isolated eigenvalue of finite multiplicity.
We will illustrate how to calculate the essential spectrum of a simple linear operator acting on B = L2 (R) [242]:
Lu = uxx + pux qu
(2.106)
for constant positive coefficients p, q and D(L) = {u : u L2 (R), Lu L2 (R)}.

Firstly, suppose that L is not invertible for some . This means that there exists
B such that L = 0. The latter equation is a linear second-order ODE with
constant coefficients and thus has solutions of the form e x with , the roots of the
characteristic polynomial 2 + p (q + ) = 0. Such a solution cannot decay at
both x = and so does not belong to B. It follows that L has no eigenvalues and
the resolvent R exists. We can then represent R in terms of the Greens function
G defined according to L G(x x
) = (x x
), where L is the adjoint of L with
respect to the standard inner product on L2 (R):
L u = uxx pux (q + )u.
For any h D(R ) B we can express the solution u = R h to the inhomogeneous
equation L u = h as
u(x) =
h(y)G(y x)dy.
For constant coefficients, the Greens function can be solved explicitly according to
y
e +
y0
G(y) =
e y y 0,
99
where are the roots of the characteristic polynomial

P( ) = 2 p ( + q),
and is chosen such that 1 = (+ ).
If P( ) has one root + with positive real part and one root with negative
real part, then clearly G L1 (R) so that R is bounded with dense domain equal
to B. This situation holds, for example, when is real and > p. The roots of
P( ) vary continuously with in the complex plane. Hence, the boundedness of
R will break down when one of the roots crosses the imaginary axis at ik, say,
with = q k2 ika. This is a parabola in the complex plane (r , i ) given by
r = q i2/p2 . If r is to the right of this parabola,
r > q
i2
,
p2
then P( ) has a root on either side of the imaginary axis and R is bounded. We
conclude that the essential spectrum lies to the left of the parabola,
(L) { : Re( ) q Im( )2 /p2 }.
(2.107)
It can be shown that the essential spectrum includes the parabola itself. It immediately follows that the essential spectrum lies in the left-half complex plane if q > 0.
In Sect. 2.4, we considered the linear stability of a traveling front, which required
finding the spectrum of a second-order linear operator with nonconstant coefficients;
see (2.42). It turns out that one can generalize the above analysis to an operator with
x-dependent coefficients p(x), q(x). Suppose that p(x), q(x) p , q as x .
Introduce the parabolas
S = { : = q k2 ikp }.
(2.108)
Let A denote the union of the regions to the left of the curves S that includes the
curves themselves. Then the essential spectrum of L lies in A and includes S .
Chapter 3
Wave Propagation Along Spiny Dendrites
The majority of excitatory synapses that occur in the cerebral cortex are located
on tiny specialized protoplasmic protuberances called dendritic spines [607]; see
Fig. 3.1. They typically occupy 2070 % of the total dendritic membrane. Since the
input impedance of a spine head is typically large, a small excitatory synaptic current can produce a large local depolarization. Moreover, the thin stem neck connecting the spine to the main body of the dendrite provides an axial resistance that
partially decouples the spine-head dynamics from the dendritic tree. Hence, it has
long been theorized that the dendritic spine is a favorable site for the initiation of
an action potential [427, 572] and is thus a possible substrate for local computations [571]. Modeling studies also suggest that if the heads of dendritic spines have
excitable membrane properties, then the spread of current from one spine along
the dendrites could bring adjacent spines to their thresholds for impulse generation. The result would be a sequence of spine-head action potentials, representing
a saltatory propagating wave in the distal dendritic branches [19, 129]. Calciumimaging experiments provide strong evidence that the spine heads are endowed with
voltage-dependent Na+ and Ca2+ channels that can indeed support an all-or-nothing
response to an excitatory synaptic input [694].
Early theoretical studies of spines also considered their potential role in synaptic
plasticity and Hebbian learning. This was motivated by the idea that small changes
in spine morphology, such as changes in the width of the spine neck, could lead
to large changes in the amplitude of response to excitatory synaptic inputs on to
the spine. It is now known that spines are rich in actin filaments, which have the
capacity to drive such changes in spine shape [413]. Moreover, there is increasing
experimental evidence that the growth and removal of spines provides an important substrate for structural changes during brain development [410, 480, 693]. It
is less clear whether changes in spine morphology play a significant role in adult
plasticity. Nevertheless, the basic geometry of a spine does provide an isolated biochemical microenvironment for Ca2+ to accumulate, and Ca2+ is thought to be a
major chemical signal for the induction of synaptic plasticity [343, 462]. The dynamics of calcium diffusion in dendritic spines has been explored in a number of
computational models [202, 283].
101
102
3 Wave Propagation Along Spiny Dendrites
Fig. 3.1 An example of a piece of spine-studded dendritic tissue (from rat hippocampal region
CA1 stratum radiatum). Magnified view on right-hand side shows a dendrite 5 m in length.
Taken with permission from SynapseWeb, Kristen M. Harris, PI, http://synapses.clm.utexas.edu/
Another important signaling molecule involved in the induction of synaptic

plasticity is Ca2+ calmodulin-dependent protein kinase II (CaMKII) [295, 386].
CaMKII is also found to be abundant within dendritic spines, where it can detect changes in the local levels of Ca2+ entering the synapse following plasticityinducing stimuli, via binding of CaMKII to Ca2+ /CaM. Confinement of CaMKII
within spines arises from the geometry of the spine and through interactions with
protein receptors and cytoskeletal elements within the postsynaptic density (PSD),
which is the protein-rich region at the tip of the spine head. Activated CaMKII phosphorylates substrates responsible for the expression of synaptic plasticity, namely,
the number and the conductivity of synaptic AMPA receptors [151, 370]. Moreover, once activated, CaMKII can transition into a Ca2+ /CaM-independent, hyperactivated state via the autophosphorylation of neighboring enzymatic subunits and
thus continue to phosphorylate its substrates even after the plasticity-inducing Ca2+
signal has ended [261, 427, 547, 689]. Experimentally, translocation of CaMKII
into spines can be induced in a local region of dendrite by exposing it to a puff
of glutamate, and this can initiate a wave of CaMKII translocation that spreads
towards the distal end of the dendrite with an average speed of 1 m/s [532].
It is found that the wave is preceded by a much faster Ca2+ -mediated spike
that propagates along the dendrite (see above), which could provide a mechanism for priming CaMKII outside the stimulus region for diffusion-based activation. Moreover, the CaMKII translocation wave is associated with an increase
in AMPA receptor numbers at both stimulated and nonstimulated synapses [532].
This suggests that it could provide a possible molecular substrate for heterosynaptic
plasticity.
3.1 Solitary Waves in a SpikeDiffuseSpike Model of Active Spines
103
In this chapter we consider two mathematical models of wave propagation along

spiny dendrites: (I) a spikediffusespike model of spine-mediated spikes [129, 139]
and (II) a reactiondiffusion model of CaMKII translocation waves [72, 161]. The
former model introduces methods for analyzing solitary waves propagating in spiking networks; see also Sect. 5.4. The latter model turns out to be identical in form to
the diffusive SI model introduced by Noble [464] to explain the spread of bubonic
plague through Europe in the fourteenth century. This, in turn, is a generalization
of the scalar FisherKPP equation [191, 345] which was originally introduced to
model the invasion of a gene into a population. One characteristic feature of such
equations is that they support traveling fronts propagating into an unstable steady
state, in which the wave speed and longtime asymptotics are determined by the
dynamics in the leading edge of the waveso-called pulled fronts [544]. In particular, a sufficiently localized initial perturbation will asymptotically approach the
traveling front solution that has the minimum possible wave speed. Hence, pulled
fronts have very different properties from those found for the bistable equation in
Sect. 2.2. Another important topic raised by these models is how to use homogenization methods to approximate the discrete distribution of spines by a smooth
distribution.
3.1 Solitary Waves in a SpikeDiffuseSpike Model

of Active Spines
The first theoretical study of active wave propagation along a dendritic cable that
is mediated by dendritic spines was carried out by Baer and Rinzel [19]. They considered a continuum model of a dendritic tree coupled to a distribution of excitable
dendritic spines. The active spine-head dynamics is modeled with HodgkinHuxley
kinetics, while the (distal) dendritic tissue is modeled with the cable equation. The
spine head is coupled to the cable via a spine-stem resistance that delivers a current proportional to the number of spines at the contact point. There is no direct
coupling between neighboring spines; voltage spread along the cable is the only
way for spines to interact. Numerical studies of the BaerRinzel model [19] show
both smooth and saltatory traveling wave solutions, the former arising in the case
of uniform spine distributions and the latter when spines are clustered in groups.
The saltatory nature of a propagating wave may be directly attributed to the fact that
active spine clusters are physically separated. In this section we describe an alternative, analytically tractable treatment of saltatory waves based on the so-called spike
diffusespike (SDS) model of active dendritic spines [129, 138, 624, 625], which
reduces the spine-head dynamics to an all-or-nothing action potential response.
In order to formulate the model, we first consider a continuum of spines with
(x) representing the spine density per unit length along a uniform, passive dendritic
cable. Denoting the voltage at position x on the cable at time t by V = V (x,t), the
associated cable equation is given by (see Sect. 1.4)
V
V
2V
a V
= + Dm 2 + (x)
,
t
m
x
Cm rs
(3.1)
104
U
^
Cs
rs
active spines
rs
xm
dendritic cable V
Fig. 3.2 Schematic diagram of a dendritic cable with active spines
where Dm = m2 /m with m and m the membrane time and space constants of the
cable, Cm is the membrane capacitance per unit length, and rs is the spine-stem resistance of an individual spine. In the SDS model, the function a(x,t) represents the sequence of action potentials generated in the spine head at x whenever the associated
subthreshold spine-head potential U(x,t), driven by current from the shaft, crosses
some threshold h. Given the high resistance of the spine stem, we neglect subthreshold currents into the cable. The voltage U evolves according to the integrate-and-fire
(IF) equation (see Sect. 5.3)
U V U
U
= +
,
Cs
t
rs
rs
(3.2)
such that whenever U crosses the threshold h it is immediately reset to zero. Here
Cs and rs are the membrane capacitance and resistance of the spine head; see
Fig. 3.2. Let T j (x) denote the jth firing time of the spine head at position x such
that U(x, T j (x)) = h. Then
a(x,t) = (t T j (x)),
(3.3)
with (t) = 0 for all t < 0. The shape of the action potential is specified by the
function (t), which can be fitted to the universal shape of an action potential.
3.1.1 Existence of a Solitary Wave for a Uniform Density

Let us begin by considering the simplest case of a uniform spine density (x) = 0
and a rectangular pulse shape of height 0 and duration d . We define a solitary
wave as one that causes the spine head at x to reach threshold only once at the
time t = T (x) x/c. We recognize c as the speed of the wave so that a(x,t) = (t
x/c), which suggests adopting a moving coordinate frame = ct x. Equation (3.1)
becomes
(3.4)
DmV ( ) cV ( ) (m1 + )V ( ) = a( /c),
105
where = 0 /(Cm rs ), and
0
a( ) = 0
< < 0
0 < < cd
> cd .
(3.5)
If one is looking for traveling pulses which satisfy lim V ( ) = 0, then the
solution to (3.4) takes the form
exp(m+ ),
< < 0
1
0
V ( ) = 2 exp(m+ ) + 3 exp(m ) + 1 , 0 < < cd
(3.6)
m +
4 exp(m ),
> cd
with
m =

c2 + 4Dm (m1 + )
2Dm
(3.7)
By ensuring the continuity of the solution and its first derivative at = 0 and = cd
one may solve for the unknowns 1 . . . 4 as
1 = 3
m
[1 exp(m+ cd )],
m+
(3.8)
m
exp(m+ cd ),
m+
(3.9)
2 = 3
3 =
0
m+
,
m1 + (m m+ )
4 = 3 [1 exp(m cd )].
(3.10)
(3.11)
As yet the speed of the pulse is undetermined. However, by demanding that

the IF process in the spine head reaches threshold at = 0 and that lim
U( ) = 0, one can determine a self-consistent value for the speed of the traveling
wave along similar lines to the analysis of solitary waves in one-dimensional networks of IF neurons [62, 63, 167]; see Sect. 5.4. In the traveling coordinate frame,
(3.2) becomes
V
(3.12)
cCsU = gsU +
rs
with U(0) = h and gs = rs1 + rs1 . This first-order system may be solved as

0
1
V (
) exp(
/[c])d
,
U( ) = exp( /[c]) h
(3.13)
cCs rs
106
wave speed c
1.2
0.8
0.4
0
20
40
60
80
spine density 0
100
Fig. 3.3 Speed of a traveling pulse as a function of uniform spine density 0 for m = Dm = 1,
h = 2.5, rs = 2, rs = 0.8, Cs = 1, Cm = 1, d = 2, 0 = 100. The crosses show the results of a direct
numerical simulation of the SDS model with N = 200 discrete spines uniformly distributed along
a cable of length L = 10
where = Cs /gs . In order for this to be bounded as , the term inside the
large parentheses must vanish as . This yields the dispersion relationship
for the speed of the pulse as a function of system parameters:
h=
1
1
1
Cs rs + cm+
(3.14)
In Fig. 3.3 we plot dispersion curves for the speed c of a traveling pulse as a function
of the spine density 0 , which are obtained by numerically solving (3.14). Note that
there are two solution branches for a given density 0 . Direct simulations suggest
that it is the upper (faster) branch that is stable, which can be confirmed analytically
using linear stability analysis (see below). Figure 3.3 also shows that for a wide
range of 0 the speed of the stable wave is approximately m /m in physical units,
in agreement with the original observations of Baer and Rinzel [19].
One useful feature of the SDS model is that it yields an exact expression for the
speed of the wave that can easily be solved to obtain the dependence in terms of
other system parameters such as the spine-stem resistance. Hence, one can find the
minimum spine density capable of supporting a traveling pulse as well as extract information about how the speed decays as a function of spine density. In Fig. 3.4(a)
we plot the speed of a traveling pulse as a function of the spine-stem resistance rs .
It is clear that for realistic choices of the biophysical parameters in the model, that
propagation failure can occur for too large a choice of the spine-stem resistance.
Moreover, for small rs , the speed of a stable pulse is very sensitive to rs , demonstrating that a modifiable value of the spine-stem resistance could have important
ramifications for neural processing. Finally in Fig. 3.4(b) we show the dependence
of the wave speed on the width, d , of a rectangular pulse generated in the spine
head. Interestingly, for a fixed pulse height, there is a minimum duration time below
which propagation cannot occur. This highlights the fact that it is crucial to model
the shape of an action potential in the reduced IF model with biologically realistic
107
choices for the amplitude and duration of the spine-head pulse. For large values of
d the speed of the wave approaches a constant value (i.e., the speed of the wave
becomes insensitive to the precise choice of d ).
3.1.2 Linear Stability Analysis

For the purposes of linear stability analysis it is more convenient to work in terms
of the original variables (x,t) rather than in the moving frame. Since the shape of
the traveling pulse is fixed by the function (t), it is natural to consider local perturbations of the firing times given by T (x) = x/c + (x). A similar approach is used
to analyze the stability of traveling waves in IF networks with synaptic and passive
dendritic interactions [62, 63]; see Sect. 5.4. The membrane potential V (x,t) satisfying (3.1) with a(t) = (t T (x)) can be evaluated in terms of the Greens function
G0 for the infinite cable equation. That is, setting Cm = 1 and taking m 0 /rs ,

0 t
G
(x
y,t
s)
(s
T
(y))dy
ds
(3.15)
V (x,t) =
0
rs
with G0 given by (1.59). If we now demand that the IF process (3.2) driven by the
potential V (x,t) reaches threshold at time T (x), then we obtain the following selfconsistency condition for a traveling pulse (with Cs = 1):
h = U(x, T (x)) =
wave speed c
1
rs
0
et/ V (x,t + T (x))dt.
(3.16)
1.2
1.5
0.8
0.4
0.5
0
2
3
4
5
spine resistance rs
0.5
1
1.5
pulse width d
Fig. 3.4 (a) Speed of a traveling pulse as a function of spine-stem resistance rs with 0 = 25 and
d = 2. (b) Speed of a traveling pulse as a function of pulse width d with 0 = 25 and rs = 2. All
other parameters are as Fig. 3.3. Note that for sufficiently large rs or small d solitary waves cannot
propagate
We now expand (3.15) and (3.16) in powers of the perturbation (x). The zerothorder term generates the self-consistency condition for the speed c of the unperturbed traveling pulse:
108

1 0
h=
where
0
V (x,t) =
rs
rs
(3.17)
t
et/ V (x,t + x/c)dt,
G0 (x y,t s) (s y/c)dy ds.
(3.18)
We can evaluate (3.18) using Fourier transforms without needing to restrict to be

a rectangular pulse. That is, expand (s) as
(t) =
ei t ( )
d
,
2
(3.19)
and then perform the integrations over y and s with

eikx G0 (x,t)dx = e (k)t ,
(k) = m1 + Dm k2 ,
(3.20)
to obtain
V (x,t) =
0
rs
ei (tx/c)
( )
d
.
( /c) + i 2
(3.21)
Substitution of (3.21) into (3.17) finally gives

h=
0
rs2
d
( )
.
1
+ i ] 2
[ ( /c) + i ][
(3.22)
When one considers a rectangular pulse shape for the action potential waveform of
height 0 and duration d such that
( ) = 0
1 eid
,
i
(3.23)
then it is a simple matter to check that the dispersion relationship (3.22) reduces to
(3.14).
The first-order term in the expansion of (3.16) yields a linear equation for the perturbations (x) from which the linear stability of the traveling pulse can be deduced.
This linear equation takes the form
0=
1
rs
0
et/ V (x,t + x/c)dt,
(3.24)
where (after integration by parts)

0 t
V (x,t) =
G0 (x y,t s)
(s y/c)[ (x) (y)]dy ds.
rs
(3.25)
109
Equations (3.24) and (3.25) have solutions of the form (x) = e x with satisfying
the characteristic equation
I( )
d
ik ( )
I(0) = 0.
1
+ i ] 2
[ ( /c + i ) + i ][
(3.26)
Asymptotic stability holds if all nonzero solutions of the characteristic equation

have negative real part. (The existence of a solution = 0 reflects the translation
invariance of the underlying system; see Sect. 2.4.) Equation (3.26) can be evaluated
by closing the contour in the lower-half complex -plane. Since (s) = 0 for s < 0
it follows that any poles of ( ) lie in the upper-half complex plane so that we only
have to consider poles arising from the zeros of the function ( /c + i ) + i . The
latter are given explicitly by = i ( ) where

( )
c
= +
(3.27)
R( ),
c
2Dm
where
R( ) =
c2
+ c /Dm + 1/(Dm m ).
4D2m
(3.28)
Let us decompose into real and imaginary parts according to = + i . Then
( ) = u ( , ) iv ( , ),
(3.29)
with
u ( , )
c
=+
A( , ),
c
2Dm
v ( , )
= B( , ),
c
(3.30)
and (for > 0)

1
A( , ) =
R( ) + R( )2 + c2 2 /D2m ,
2

1
B( , ) =
R( ) + R( )2 + c2 2 /D2m .
2
(3.31)
One may now determine the linear stability of a solitary pulse by simultaneously
solving Re[I( )] = Re[I(0)] = 0 and Im[I( )] = Im[I(0)] = 0 for and , with c determined by (3.22) (or equivalently (3.14) for the specific case of a rectangular pulse
shape). This is illustrated in Fig. 3.5 for a rectangular pulse, which shows solutions
with = 0 and < 0 along the upper branch of Fig. 3.3. Moreover, changes sign
as it passes through the point where d 0 /dc = 0 in the (c, 0 ) plane while moving
from the upper solution branch to the lower. Hence, of the two possible traveling
110
wave solutions, the faster one is stable. Other solutions with < 0 and > 0 are
also found for both the fast and slow branches but do not affect the above stability
argument.
0
-2
Re( )
-4
-6
-8
20
40
60
80
100
spine density 0
Fig. 3.5 A plot of the eigenvalues arising in the linearization of the SDS model shows that solutions
with < 0 and = 0 can be found for the branch of solutions with greatest speed. The graph
above shows the behavior of = Re( ) with = 0 for the upper solution branch of Fig. 3.3. Note
that = 0 at the point in the (c, ) plane at which the speed of the two possible traveling waves
becomes equal. For the slower branch one finds that there exists solutions in the ( , ) plane with
> 0 indicating that the slower branch is unstable
3.1.3 Discrete Distribution of Spines

So far we have assumed that the spine density is uniform. Although wave propagation failure is known to occur if the spine density is below some critical level, the
numerical studies of Baer and Rinzel [19] suggest that propagation may be recovered by redistributing the spines into equally spaced dense clusters. Since interspine
distances are of the order of m and electronic length m is typically measured in
mm, we shall consider spine-head voltage at a cluster site to be the local spatial
average of membrane potential in adjacent spines. Hence, we consider a discrete
distribution of spines for which
(x) = n (x xm ),
(3.32)
where xm is the location of the mth spine cluster and n is the number of spines in
a cluster. Such a distribution breaks continuous translation symmetry so that saltatory or lurching waves are expected rather than traveling waves of constant profile.
(Saltatory waves are also found in myelinated axons; see Sect. 2.5.) We define a
saltatory wave as an ordered sequence of firing times . . . Tm1 < Tm < Tm+1 in which
each spine cluster only fires once. Substituting the discrete density (3.32) into (3.1)
with
(x)a(x,t) = n (t Tm ) (x xm ),
m
111
and making the approximation n rs so that the term (x)V /rs can be neglected,
we obtain the equation (using physical units)
V
V
2V
n
= + Dm 2 + (x xn ) (t Tn ),
t
m
x
rs n
(3.33)
where we have absorbed a factor of rm /m into and Dm = m2 /m . This can be

solved using the Greens functions G0 of the cable equation:
V (x,t) =
n
H(x xm,t Tm ),
rs m=0
(3.34)
where we assumed the initial condition V (x, 0) = 0 for all x.

H(x,t) =
t
0
2
1
G0 (x,t) =
et/m x /4Dmt .
4 Dm t
G0 (x,t s) (s)ds,
(3.35)
Suppose that the spine clusters are uniformly distributed along the cable such
that xm = md, where d is the spacing between clusters. We will derive a condition
for the existence of a saltatory wave solution given by Tm = m for large m. The
parameter measures the time between successive threshold crossings at adjacent
spine-head clusters such that the speed c = d/ . First, using the causality condition
H(x,t) = 0 for t < 0, it follows from (3.34) that
V (Nd, N ) =
n N
H(nd, n ),
rs n=1
(3.36)
The wave speed of a saltatory wave (if it exists) is then determined self-consistently
from the threshold condition
h = lim V (Nd, N ) =
N
n
H(nd, n ).
rs n=1
(3.37)
In order to calculate the wave speed, it is useful to rewrite (3.35) in the form
H(x,t) =

dk ikx (k)t
e e
(k,t),
(3.38)
where (k,t) = 0t (s)e (k)s ds. One may then exploit the convolution structure of
(3.38) to evaluate it in closed form for a given (t). For the sake of illustration,
consider again the rectangular pulse shape , for which
(k,t) = 0 (e (k) min(t,d ) 1)/ (k),

so that H(x,t) = 0 [A(x,t min(t, d )) A(x,t)], with
A(x,t) =

dk eikx (k)t
(k)
(3.39)
112
This is a standard integral given explicitly by

t
1 m
|x|
A(x,t) =
+
e|x|/ m Dm erfc
4 Dm
m
4Dmt
%
|x|
t
+ e|x|/ m Dm erfc
+
.
m
4Dmt
(3.40)
Finally, we may write the threshold condition in the form

h=
n
H(md, m ),
rs2 m=1
(3.41)
where H(x,t)
= 0 [A(x,t
d ) A(x,t)],
and
A(x,t)
=
0
es/ A(x,t s)ds.
(3.42)
The sum in (3.41) can then be performed numerically to obtain the speed of a saltatory wave c = c(d, h).
In Fig. 3.6(a) we plot the speed c as a function of cluster spacing d for fixed
threshold h, which shows that if the spine clusters are separated beyond some critical spacing, on the order of the electronic length m = 1, a saltatory pulse will
fail to propagate. Again linear stability analysis establishes that, as in the continuum model, it is the faster of the two branches that is stable. It is also instructive
to consider the region in the (d, h) parameter plane where saltatory pulses exist.
This may be obtained by continuing the limit point defining propagation failure of
a saltatory pulse in the (d, c) plane as a function of h. The resulting phase diagram
is shown in Fig. 3.6(b) and establishes that with increasing d the critical threshold
for propagation failure decreases. Interestingly, the minimum wave speed of a stable
saltatory pulse is found to be relatively insensitive to variation in cluster spacing d
and threshold h. Now that the speed of a saltatory pulse has been determined as a
function of system parameters it is possible to close the expression for the shape of a
solitary pulse given by V (x,t) = (n/rs ) m H(x md,t md/v). A plot of this analytical expression is shown in Fig. 3.7, which clearly demonstrates that the saltatory
pulse has a nonconstant profile.
Finally, note that there have been a number of recent extensions of the spike
diffusespike model. These include taking into account the branching structure of
the dendritic tree [625], in which the Greens function G0 is replaced by the Greens
function of the tree, and incorporating active processes within the dendritic cable.
In the latter case, each infinitesimal cable compartment is modeled as a quasi-linear
LRC circuit, where the membrane resistance rm is in series with an inductance L
[624]; see Fig. 1.11. It should also be pointed out that from a mathematical perspective, the spikediffusespike model is very similar in structure to the firediffuse
fire model of calcium puffs; see Sect. 4.3.
3.2 ReactionDiffusion Model of CaMKII Translocation Waves
b 25
threshold h
wave speed c
a 4
2
1
0
113
20
15
no saltatory wave
10
saltatory wave
5
0
0.4
0.8
1.2
cluster spacing d
1.6
0.2 0.4 0.6 0.8

cluster spacing d
Fig. 3.6 (a) Plot of wave speed c for a saltatory pulse as a function of cluster spacing d. Here,
= d = 1, h = 2.5, rs = 2, 0 = 100, and n = 2. The upper (lower) branch is stable (unstable).
(b) Continuation of the limit point in (a) showing the region in the (d, h)-plane where stable saltatory traveling waves exist
Fig. 3.7 Plot of the analytically obtained saltatory solution V (x,t) in the dendritic cable with
parameters as in Fig. 3.6 and d = 1. The x axis covers 10 lattice sites and the t axis 10d/c

As we indicated at the beginning of this chapter, CaMKII (Ca2+ calmodulindependent protein kinase II) is a key regulator of glutamatergic synapses and plays
an essential role in many forms of synaptic plasticity. It has recently been observed
experimentally that chemically stimulating a local region of dendrite not only induces the local translocation of CaMKII from the dendritic shaft to synaptic targets
within spines, but also initiates a wave of CaMKII translocation that spreads distally
through the dendrite with an average speed of order 1 m/s [532]. In Fig. 3.8, we
provide a cartoon of the mechanism for translocation waves hypothesized by Rose
et al. [532]. Before local stimulation using a glutamate/glycine puff, the majority of
CaMKII is in an inactive state and distributed uniformly throughout the dendrite.
Upon stimulation, all CaMKII in the region of the puff (30 m of dendrite) is
converted to an active state, probably the autonomous state of CaMKII (see
Fig. 3.8a), and begins translocating into spines. Simultaneously, a Ca2+ spike is
114
initiated and rapidly travels the length of the dendrite (as modeled in Sect. 3.1),
causing CaMKII to bind Ca2+ /CaM along the way. In this primed or partially phosphorylated state, CaMKII does not yet translocate into spines. In the meantime, a
portion of the activated CaMKII from the stimulated region diffuses into the region
of primed CaMKII and the two types interact, with the result that primed CaMKII
is activated. Some of these newly activated holoenzymes translocate into spines
while others diffuse into more distal regions of the dendrite containing primed
CaMKII, and the wave proceeds in this fashion. In certain cases one also finds a
second wave propagating proximally from the stimulated site to the soma [532].
A schematic diagram illustrating the progression of a translocation wave along a
dendrite following the rapid priming phase is shown in Fig. 3.8b.
2 x
a
soma
dendrite
x=0
x=L
Fig. 3.8 Proposed mechanism of CaMKII translocation waves. (a) A glutamate/glycine puff activates CaMKII locally and initiates a fast Ca2+ spike that propagates distally (indicated by larger
horizontal arrow) and primes CaMKII in the remainder of the dendrite. In certain cases one also
finds a second wave propagating proximally from the stimulated site to the soma (indicated by
smaller horizontal arrow). (b) Activated CaMKII (gray dots) both translocates into spines and diffuses into distal regions of the dendrite where it activates primed CaMKII (black dots). The net
effect is a wave of translocated CaMKII propagating along the dendrite
A simple mathematical model of the above mechanism can be constructed using a system of reactiondiffusion equations for the concentrations of activated and
primed CaMKII in the dendrite and spines [72, 161]. These equations incorporate
three major components of the dynamics: diffusion of CaMKII along the dendrite,
activation of primed CaMKII, and translocation of activated CaMKII from the dendrite to spines. For simplicity, consider a uniform one-dimensional, nonbranching
dendritic cable as shown in Fig. 3.8a. Suppose that a region of width 30 m is stimulated with a glutamate/glycine puff at time t = 0. The center of the stimulated
region is taken to be at x = 0 and the distal end of the dendrite is at x = L = 150 m.
The diffusion, activation, and translocation of CaMKII along the dendrite following
stimulation are modeled according to the following system of equations:
P
2P
= D 2 k0 AP
t
x
(3.43a)
115
A
2A
= D 2 + k0 AP hA
t
x
(3.43b)
S
= hA,
t
(3.43c)
where D is the diffusivity of CaMKII within the cytosol. Here P(x,t) and A(x,t)
denote the concentration of primed and activated CaMKII at time t > 0 and location
x along the dendrite. S(x,t) denotes the corresponding concentration of CaMKII in
the population of spines at the same time and distance. For simplicity, all parameters
are constant in space and time. The reaction term kAP represents the conversion of
CaMKII from its primed to active state based on the irreversible first-order reaction
scheme
A + P 2A
with mass action kinetics, where k0 is the rate at which primed CaMKII is activated
per unit concentration of activated CaMKII. The decay term hA represents the loss
of activated CaMKII from the dendrite due to translocation into a uniform distribution of spines at a rate h. The model assumes that translocation is irreversible
over the time scale of simulations, which is reasonable given that activated CaMKII
accumulation at synapses can persist for several minutes [569].
As a further simplification we will only consider the distal transport of CaMKII
from the stimulated region by taking 0 x L and imposing closed or reflecting boundary conditions at the ends x = 0, L. Hence, no CaMKII can escape from
the ends. In reality activated CaMKII could also diffuse in the proximal direction
and trigger a second proximal translocation wave. However, the choice of boundary
condition has little effect on the properties of the wave. Taking the distal half of
the stimulated region to be 0 x 15 m, consider the following initial conditions:
P(x, 0) = 0 and A(x, 0) = P0 for 0 x 15 m, whereas P(x, 0) = P0 and A(x, 0) = 0
for x 15 m, where P0 is the uniform resting concentration of CaMKII in the dendrite. Typical values of C range from 0.1 to 30M [605], covering two orders of
magnitude. We also set S(x, 0) = 0 everywhere. In other words, we assume that
all the CaMKII is activated within the stimulated region at t = 0, but none has yet
translocated into spines nor diffused into the nonstimulated region. We also neglect
any delays associated with priming CaMKII along the dendrite. This is a reasonable
approximation, since the Ca2+ spike travels much faster than the CaMKII translocation wave [532]; see Sect. 3.1. Thus by the time a significant amount of activated
CaMKII has diffused into nonstimulated regions of the dendrite, any CaMKII encountered there will already be primed. The benefit of this assumption is that it
eliminates the need to model the Ca2+ spike. However, a more detailed model that
takes into account the initial transient associated with the priming phase could be
constructed by coupling the reactiondiffusion equations with the spikediffuse
spike model of Sect. 3.1.
Note that the system of equations (3.43) is identical in form to the diffusive SI
model introduced by Noble [464] to explain the spread of bubonic plague through
Europe in the fourteenth century. In the latter model, P(x,t) and A(x,t) would rep-
116
resent the densities of susceptible and infective people at spatial location x at time t,
respectively; k0 would be the transmission rate and h the death rate. In the absence
of translocation into spines (h = 0), the total amount of CaMKII is conserved so that
A(x,t) + P(x,t) = P0 for all x and t 0. Equations (3.43) then reduce to the scalar
FisherKPP equation
A
2A
= D 2 + k0 A(P0 A),
(3.44)
t
x
which was originally introduced to model the invasion of a gene into a population.
The FisherKPP equation and its generalizations have been widely used to describe
the spatial spread of invading species including plants, insects, genes, and diseases;
see, for example, [285, 444, 575, 648] and references therein. One characteristic
feature of such equations is that they support traveling fronts propagating into an
unstable steady state, in which the wave speed and longtime asymptotics are determined by the dynamics in the leading edge of the waveso-called pulled fronts
[544]. In particular, a sufficiently localized initial perturbation (such as the stimulus
used to generate CaMKII waves) will asymptotically approach the traveling front
solution that has the minimum possible wave speed. (If we perform the change of
variables Q = P0 P in the CaMKII model, then the traveling wave solution constructed below propagates into the unstable state A = 0, Q = 0.) An overview of the
theory of pulled fronts is presented in appendix section 3.3.
3.2.1 Translocation Waves for a Uniform Distribution of Spines

A traveling wave solution of (3.43a) and (3.43b) is P(x,t) = P( ) and A(x,t) =
A( ), = x ct, where c, c > 0, is the wave speed, such that
P( ) P0 ,
and
P( ) P1 < P0 ,
A( ) 0 as
A( ) 0
as .
Here P1 is the residual concentration of primed CaMKII following translocation of

activated CaMKII into spines. The minimum wave speed can be calculated by substituting the traveling wave solution into (3.43a) and (3.43b) and linearizing near the
leading edge of the wave where P P0 and A 0. In the traveling wave coordinate
frame (3.43a) and (3.43b) are transformed to
dP
d2P
= D 2 k0 AP
d
d
dA
d2A
c
= D 2 + k0 AP hA
d
d
(3.45a)
(3.45b)
This is a system of two second-order ordinary differential equations in the

variable . A global view of the nature of traveling wave solutions can be ob-
117
tained by identifying (3.45a) and (3.45b) with the equation of motion of a classical
particle in two spatial dimensions undergoing damping due to friction and subject
to an external force. Thus we identify A and P with the spatial coordinates
of the particle, with the corresponding time coordinate, and the speed c as a
friction coefficient. If we ignore boundary effects by taking < < , then
we can view a traveling wave solution as a particle trajectory that connects the point
(P, A) = (0, 0) at = to the point (P, A) = (P0 , 0) at = . A restriction on
the allowed values of c can now be obtained by investigating how the point (1, 0) is
approached in the large- limit.
Linearizing Eqs. (3.45a) and (3.45b) about the point (P, A) = (P0 , 0) we obtain a
pair of second-order linear equations, which have solutions of the form (P P0, A) =
Ve where and V satisfy the matrix equation

D 2
k
c V =
V,
(3.46)
0 D 2 + k h
where k = k0 P0 . Solving for the eigenvalue leads to the four solutions

c c2 4D(k h)
c
,
= 0,
D
2D
(3.47)
and these, along with their corresponding eigenvectors V, determine the shape of
the wave as it approaches the point (1, 0). Note that the last two eigenvalues have a
nonzero imaginary part when c2 < 4D(k h), implying that as becomes large the
wave oscillates about the point (1, 0). This cannot be allowed since it would imply
that the activated CaMKII concentration A takes on negative values (inspection of
the corresponding eigenvectors shows that their components in the A-direction are
nonzero and so A would indeed oscillate). Therefore, we must have

(3.48)
c cmin = 2 D(k h),
which implies that k > h. Note that the minimum wave speed can be identified with
the linear spreading velocity of a pulled front; see appendix section 3.3. This then
yields a more direct method for obtaining the minimum wave speed. That is, the
characteristic equation obtained from (3.46) yields the dispersion relation
c( ) = D +
kh
.
(3.49)
The theory of pulled fronts shows that the minimum wave speed is obtained by
minimizing c(
). The equation c
( ) = 0 gives D = (k h)/ , which has the
solution 0 = (k h)/D, so that cmin = c(0 ) = 2D0 = 2 D(k h).
An example of a numerically determined traveling wave solution with minimal
speed cmin is shown in Fig. 3.9 for parameter values consistent with experimental
studies of CaMKII , which is one of the two main isoforms of CaMKII. In its inactive state CaMKII tends to be located in the cytosol, whereas the other isoform,
CaMKII , is weakly actin bound [570]. One finds empirically that D 1m2 /s,
h 0.05 s1 and c 1 m/s [532, 569, 570] for CaMKII . (CaMKII has a
diffusivity and translocation rate an order of magnitude smaller but exhibits compa-
118
CaMKII [norm. conc.]
1.2
10 sec
65 sec
0.8
activated
CaMKII
0.6
primed CaMKII
0.4
0.2
0
30
60 90
x [m]
120 150
1.2
30
60 90
x [m]
120 150
140 sec
1
0.8
0.6
0.4
0.2
0
30
60 90
x [m]
120 150
Fig. 3.9 Three successive snapshots of a numerically simulated translocation wave propagating
along a homogeneous dendrite. Solutions of (3.43a) and (3.43b) are plotted for parameter values
consistent with experimental data on CaMKII [532, 569, 570]. The translocation rate h = 0.05 s,
diffusivity D = 1 m2 /s and the activation rate k0 P0 = 0.21 s. At time t = 0 all of the CaMKII
within the stimulated region (indicated by thick bar) is in the activated state, whereas all of the
CaMKII within the nonstimulated region is in the primed state. Concentrations are normalized
with respect to the initial concentration of primed CaMKII. Composite wave consists of a pulse
of activated CaMKII (gray curve) moving at the same speed as a front of primed CaMKII (black
curve). Also shown is the total CaMKII concentration along the dendrite (dashed black curve),
which decreases with time due to translocation into spines. As indicated in the center plot, the
front forms an interface between a quiescent region containing a uniform concentration of primed
CaMKII and a region dominated by translocation of activated CaMKII into spines. The dynamics
in the interfacial (shaded) region is dominated by diffusionactivation of primed CaMKII
rable wave speeds.) The formula for the wave speed then gives an estimate for the
unknown activation rate, k 0.2 s1 . It can be seen in Fig. 3.9 that the wave profile
of primed CaMKII is in the form of a front, whereas the co-moving wave profile of
activated CaMKII is a localized pulse.
The above analysis predicts wave propagation failure when the translocation rate
h is greater than the effective activation rate k. Experimentally, h is determined by
globally activating CaMKII along a dendrite and determining the rate at which the
level of CaMKII decays [569, 570]. The detailed microscopic mechanism whereby
CaMKII is translocated into spines is currently not known, so it is difficult to relate h to individual spine properties. A simple hypothesis is that the translocation
rate depends on the spine density according to h = 0 0 , where 0 is an effective
velocity associated with translocation into an individual spine. Since the activa-
119
tion rate k = k0 P0 , where P0 is the initial concentration of primed CaMKII in the

nonstimulated region of the dendrite, the model predicts that CaMKII translocation
waves will fail to propagate when
0 0 > k0 P0 .
(3.50)
For example, this inequality predicts that dendrites with a high density of spines are
less likely to exhibit translocation waves than those with a low spine density. It also
predicts that dendrites with a larger initial concentration of primed CaMKII in the
shaft are more likely to exhibit translocation waves than those with a smaller initial
concentration. Since the initial concentration P0 of primed CaMKII depends on the
effectiveness of the Ca2+ spike in both propagating along the dendrite and priming
the inactive state, the model agrees with the experimental finding that translocation
waves fail to propagate when L-type Ca2+ channels are blocked [532]. One also
finds that Ca2+ spikes are less likely to propagate towards the soma, which could
explain why translocation waves are more often observed propagating towards the
distal end of a dendrite.
3.2.2 Pulsating Waves in the Presence of Discretely

Distributed Spines
One of the major simplifications of the above model is that the discrete nature of
dendritic spines is ignored by treating the spine density and, hence, the translocation rate h as uniform. Given the fact that the radius of the spine neck is typically
at the submicron level, which is much smaller than any other length scale of the
system, one can take into account the discreteness of spines by representing the
spine density as the sum of Dirac delta functions (3.32), which represents clusters
of n spines at discrete locations xn ; see Sect. 3.1. It immediately follows that the
translocation rate h is itself space-dependent and (3.43a) and (3.43b) become heterogeneous. For the sake of illustration, consider the case of a set of spine clusters
that are uniformly spaced with xn = n where is the spine cluster spacing. In order to separate the effects of discreteness from the effects of spine density, we will
assume that the size of a cluster scales with so that n = 0 with 0 fixed. Thus,
setting h = 0 0 , we have the space-dependent translocation rate
h(x) = h
(x n ),
(3.51)
nZ
such that L1 0L h(x)dx = h for L .

In recent years, there has been an increasing interest in studying biological invasion in heterogeneous environments using reactiondiffusion equations
[40, 102, 212, 335, 574, 575, 587, 664, 687]. Heterogeneity is often incorporated
by assuming that the diffusion coefficient and the growth rate of a population are
periodically varying functions of space. One of the simplest examples of a single population model in a periodic environment was proposed by Shigesada et al.
120
[574, 575], in which two different homogeneous patches are arranged alternately
in one-dimensional space so that the diffusion coefficient and the growth rate are
given by periodic step functions. The authors showed numerically that an invading
population starting from a localized perturbation evolves to a traveling periodic
wave in the form of a pulsating front. The population density u(x,t) of such a wave
is defined by the condition u(x,t) = u(x + ,t + T ) such that limx u(x,t) = 0 and
limx u(x,t) = p(x), where p(x) is a spatially periodic stationary solution of the
corresponding reactiondiffusion equation. This form of solution repeats itself in
a time interval T if it is observed at two successive points separated by a distance
. The speed of the wave is then taken to be c = /T . Shigesada et al. [574] also
used linearized information within the leading edge of the pulsating front to derive
wave speed estimates, generalizing the analysis of pulled fronts in homogeneous
media [544]. An interesting recent extension of this approach has been used to
study pulsating fronts in periodically modulated nonlocal neural field equations
[132]. The theory of pulsating fronts has also been developed in a more general and
rigorous setting [40, 212, 587, 664, 687].
The analysis of CaMKII translocation waves developed in [72] follows the basic
formulation of Shigesada et. al. [574] by linearizing (3.43b) at the leading edge of
the wave where A(x,t) 0 and P(x,t) P0 :
A
2A
= D 2 + kA h(x)A,
t
x
(3.52)
with h(x) given by the -periodic function (3.51). Assume a solution of the form
A(x,t) = a( )P(x), = x ct, and set
c ,
t
+
.
x
x
Substitution into (3.52) then gives

cP(x)a
( )=D a
( )P(x)+2a
( )P
(x)+a( )P
(x) +[kh(x)]a( )P(x).

Dividing through by a( )P(x) and rearranging yields

a ( )
a
( )
P
(x)
P
(x)
D
+ 2D
+c
= D
k + h(x).
a( )
P(x)
a( )
P(x)
(3.53)
Applying the operator x to both sides of (3.53) implies that either P

(x)/P(x)
is a constant or a
( )/a( ) is a constant. Only the latter condition is consistent
with P(x) being a periodic function. Thus, a( ) = A0 e with determined by
solutions to the damped Hill equation

k h(x) c
P
(x) 2 P
(x) + 2 +
P(x) = 0.
(3.54)
D
Note that if P(x) = e xU(x) then U(x) satisfies the undamped Hill equation
121
DU
(x) + [k h(x) c ]U(x) = 0.
(3.55)
In order to determine the minimal wave speed cmin , it is necessary to find a bounded
periodic solution P(x) of (3.54), which yields a corresponding dispersion relation
c = c( ), whose minimum with respect to can then be determined (assuming it
exists). Unfortunately, for general periodic functions h(x), it is not possible to solve
(3.54) explicitly, and some form of approximation scheme is required as described
in the next section.
3.2.3 Homogenization of Pulsating Waves for a Fast Periodic

Modulation of Spine Density
Since the spine cluster spacing is at least an order of magnitude smaller than
the width of the traveling wave of the homogeneous system, one can use homogenization theory to approximate the discrete effects of spines by a corresponding
continuum model [72]. Such a method has also been applied to studying variations in electrical voltage/conductance [424] and the distribution of protein receptors [68] along spiny dendrites. Interestingly, Smaily et al. [587] independently
applied the same homogenization procedure to analyze wave speed in the population model of Shigesada et. al. [574]. A more general discussion of homogenization techniques applied to traveling fronts can be found in the review by
Xin [687].
As a first step, we introduce a macroscopic length scale and set = with
1. We identify with the effective width of the primed CaMKII front, which
turns out to be around 2030 m in the given parameter regimes. Equation (3.55)
can then be rewritten in the form
where
x
d 2U
U(x) = 0,
+
2
dx
(3.56)
= [k c h]/D
(3.57)
and
h
(y) =
D
(y n ) 1
(3.58)
nZ
such that (y) is a -periodic function of y. Applying the homogenization procedure outlined below leads to the result

U(x) = U(x) 1 + 2 U(x/ ) + O( 3 ),
(3.59)
122
where U satisfies the averaged equation

d 2 U
+ U = 0,
dx2
and
(3.60)
h
D

y2
1
h 2 y
U(y) =
2
.
D 2 2
12
= 22 + O( 3 ),
2 =
1
12
2
,
(3.61)
(3.62)
Homogenization method for discrete spines. The basic idea of multi-scale homogenization
is to expand the solution of Eq. (3.56) as a power series in , with each term in the expansion
depending explicitly on the slow (macroscopic) variable x and the fast (microscopic)
variable y = x/ [68, 424, 587]:
U(x, y) = U0 (x) + U1 (x, y) + 2U2 (x, y) + . . .,
(3.63)
where U j (x, y), j = 1, . . . are -periodic in y. The perturbation series expansion is then substituted into Eq. (3.56) with x, y treated as independent variables so that derivatives with
respect to x are modified according to x x + 1 y . This generates a hierarchy of equations corresponding to successive powers of :
2U1
=0
y2
2
2
2

U1 U2
d U0
+
+2
+ (y) U0 = 0
dx2
x y
y2
(3.64)
(3.65)
at powers 1 , 1 and

2Un
2Un+1 2Un+2
+
+2
+ (y) Un = 0
x2
x y
y2
(3.66)
at O( n ), n 1.
Equation (3.64) and boundedness of U1 imply that U1 is independent of y and can thus be
absorbed into U0 (x). Thus the leading-order corrections arising from small-scale fluctuations in the spine density occur at O( 2 ). Define the spatial average of a periodic function
F(y), denoted by F, according to
F =
F(y)dy.
(3.67)
Taking the spatial average of (3.65) with U0 = U0 then gives

d 2U0
+ U0 = 0.
dx2
(3.68)
We have exploited the fact that U2 is periodic in y so 2U2 / y2 = 0. In order to calculate

U2 , we first subtract the averaged Eq. (3.68) from (3.65) to obtain
2U2
= (y)U0 (x).
y2
(3.69)
123
It follows that U2 (x, y) = U0 (x) (y) with
(y) = (y) and

a -periodic function of
y. Integrating once with respect to y gives
(y) =
(0) + 0y (z)dz. We can eliminate
the unknown
! y (0) by spatially averaging with respect to y and using = 0. This gives
(y) = 0 (z)dz with

y

y
y
f (z)dz
f (z)dz
f (z)dz
(3.70)
0
for any integrable function f . Another integration with respect to y shows that
(y) = (0) +
y
(z)dzdy
.
Spatially averaging this equation with respect to y in order to express (0) in terms of
and multiplying through by U0 (x) finally gives
U2 (x, y) U2 (x, y) U2 (x)

= U0 (x)
y
0
(3.71)
(z)dzdy
.
It remains to determine the equation satisfied by U2 . Spatially averaging Eq. 3.66) for
n = 2 gives
d 2 U2
+ U2 = (y)U2 (x, y).
dx2
(3.72)
Substituting (3.71) into (3.72) and reordering the resulting multiple integral yields the result
&
2 '
y
d 2 U2
U0 (x).
+
U

=
(z)dz
(3.73)
2
dx2
0
Finally, writing U = U0 + 2 U2 + . . . we obtain the homogenized version of the Hill
equation (3.55):
d 2 U
+ U = 0,
(3.74)
dx2
where
&
2 '
y
2
3
.
(3.75)
= + 2 + O( ), 2 =
(z)dz
0
It is straightforward to calculate the integrals in (3.71) and (3.75) for a periodic spine density
with (y) given by (3.58) [424]:
&
2 =
y
y
0
y
0
2
2 '
1 h
=
(z)dz
,
12 D
(z)dzdy
=
y2
1
h 2 y
2
.
D 2 2
12
(3.76)
(3.77)
We thus obtain (3.59) and (3.60).

Note that it is possible to extend the above homogenization scheme to the case of randomly rather than periodically distributed spines, provided that the resulting heterogeneous
medium is ergodic. That is, the result of averaging over all realizations of the ensemble of
spine distributions is equivalent to averaging over the length L of the dendrite in the infinite-
124
L limit. If such an ergodic hypothesis holds and L is sufficiently large so that boundary
terms can be neglected, then the above analysis carries over with now denoting ensemble averaging [424].
Recall from our discussion of (3.54) that the solution P(x) = e xU(x) has to be a
bounded -periodic function of x. It follows from (3.59) and (3.60) that e x U(x)
should be a finite constant. Writing the solution of (3.60) as U(x) e x ,

yields the characteristic equation

c k + h
22 ,
=
(3.78)
D
where we have substituted for using (3.57). Squaring both sides and rearranging
thus leads to the following dispersion relation for the wave speed c:
c = c( ) D +
k h + 2 D2
.
Minimizing with respect then shows that

cmin = 2 D(k h) + 2 D22 ,
(3.79)
(3.80)
For sufficiently small , we can Taylor expand (3.80) to obtain

cmin c +
2D22 2
,
c
(3.81)

with c = 2 D(k h) the wave speed of the corresponding homogeneous distribution of spines. Hence, a periodic variation in the spine density due to clustering leads
to an O( 2 ) increase in the wave speed. An analogous result was obtained by Smaily
et. al. [587] for the Shigesada et. al. model [574]. Note that (3.59) also implies that
there are small-scale fluctuations of the wave profile in the leading edge of the wave
P(x/ )
= 2 U(x/ ) + O( 3 ).
P
(3.82)
Since = / , it follows from (3.62) that fluctuations in the wave profile vary
between h 2 /(12D) at spine clusters and h 2 /(24D) between spine clusters. In
terms of physical parameters, the minimum wave speed is

2
(3.83)
cmin = 2 D(k h) + 2 h /12.
It immediately follows that for fixed h, k, D (and hence c), spine clustering increases
the speed of a translocation wave. This is illustrated in Fig. 3.10, where we plot the
minimal wave speed cmin given by (3.83) as a function of the activation rate k for
various values of the cluster spacing . An additional important consequence of
clustering is that it reduces the threshold for the existence of a translocation wave.
125
That is, there exists a critical value of the activation rate, k = ( ), below which
translocation waves do not exist and ( ) is a decreasing function of . In the
homogenization limit 0, we recover the result = h.
The existence of a pulsating wave due to spine clustering and the associated increase (decrease) in the minimal speed (threshold) of the wave can also be confirmed
numerically. For the sake of illustration, consider a dendrite of length L = 300 m
with reflecting boundary conditions at both ends x = 0, L. The initial conditions are
taken to be
/ [0, L]
P(x, 0) = P0 , A(x, 0) = 0 for all x
P(x, 0) = 0, A(x, 0) = P0 for all x [0, L],
with L = 15 m. We discretize space by setting x = m x, where x is the step
length and m = 0, 1, . . . , M with M = L/ x. In discrete spatial units the spine cluster
spacing is taken to be = P x. The spine cluster distribution is then represented
numerically by the discrete sum
(m x) =
1 M/P
m, jP ,
x j=0
where m, j is the Kronecker delta and x is chosen so that M, P and M/P are integers. An example of a pulsating wave is shown in Fig. 3.11. Comparison between
waves for a spatially uniform distribution of spines and a spatially discrete distribution of spine clusters shows that the wave is periodically modulated and faster in
the latter case. This is a consequence of the fact that translocation is less effective
1.4
= 8m
wavespeed cmin [m/s]
1.2
= 4m
hom
0.8
0.6
0.4
0.2
0
0
_
h
( )
0.1
0.2
0.3
activation rate k
0.4
0.5
[sec-1]
Fig. 3.10 Plot of minimal wave speed cmin as a function of activation rate k = k0 P0 for various values of the spine cluster spacing . Also shown is the corresponding wave speed for a homogeneous
spine distribution (black curve). Other parameters are h = 0.1 s1 and D = 1 m2 /s. Note that wave
propagation failure occurs as k ( ) from above where ( ) is the propagation threshold
126
1.2
141 sec
0.8
0.6
0.4
0.2
0
10 sec
60
120 180 240 300

x [m]
60
120 180 240 300

x [m]
1.2
350 sec
242 sec
1
0.8
0.6
0.4
0.2
0
60
120 180 240 300

x [m]
60
120 180 240 300

x [m]
Fig. 3.11 Numerical traveling wave solution of (3.43a) and (3.43b) for an inhomogeneous distribution of spine clusters with = 8 m. The translocation rate h = 0.1 s, diffusivity D = 1 m2 /s
and the activation rate k = 0.19 s. At time t = 0 all of the CaMKII within the stimulated region (indicated by thick bar) is in the activated state, whereas all of the CaMKII within the nonstimulated
region is in the primed state. Concentrations are normalized with respect to the initial concentration of primed CaMKII. The resulting wave profiles for activated (gray curve) and primed (black
curve) CaMKII along the dendrite are shown at four successive snapshots in time. The numerically
estimated wave speed cmin 0.66 m/s, which is faster than the wave speed c = 0.6 m/s of the
corresponding uniform spine distribution
in the presence of spine clusters. Although doubling the degree of clustering only
leads to a change in wave speed of order 0.05m/s (consistent with the analytical
calculations), it leads to a significant difference in propagation times along a 300 m
dendrite.
3.2.4 Wave Speed for a Slowly Modulated Spine Density

So far we have considered the effects of heterogeneity at a microscopic length scale
comparable to the spacing of individual spines. In particular, we took the homogenized translocation rate h to be constant over the length of a dendrite. However,
it is found experimentally that there is a slow proximal to distal variation in the
density of spines [22, 346]. An illustration of a typical spine density found in pyramidal neurons of mouse cortex [22] is shown in Fig. 3.12. Such a variation in spine
127
density can be incorporated into (3.43a) and (3.43b) by setting h = h + h( x),

where h denotes the translocation rate at the initiation point x0 of the wave and
h( x) represents the slow modulation of the (homogenized) translocation rate over
the length of a dendrite with 1. The general problem of biological invasion in
slowly modulated heterogeneous environments can be analyzed using a Hamilton
Jacobi method for front velocity selection [421, 687]; see Sect. 3.3. This method
was originally applied to homogeneous media by Freidlin using large deviation theory [203, 204, 212] and was subsequently formulated in terms of PDEs by Evans
and Sougandis [178]. We will illustrate the method by applying it to the reaction
diffusion model of CaMKII translocation waves with slow periodic modulation; see
also [72]
number of spines
25
20
15
10
5
10
30
50
90
110
70
130
distance from soma [m]
150
170
Fig. 3.12 Illustrative example of the spine density variation along the basal dendrite of a pyramidal
cell in mouse cortex (black curve). Density is calculated as the number of spines per 10 m segment
of the dendrite from the soma to the tip of the dendrite. Abstracted from experimental data in [22].
Also shown is a simplified piecewise linear approximation of the spine density variation (gray
curve)
The first step in the analysis is to rescale space and time in (3.43a) and (3.43b)
according to t t/ and x x/ ; see [178, 204, 421]:
2P
P
= D 2 2 k0 AP
t
x
2A
A
= D 2 2 + k0 AP [h + h(x)]A.
t
x
(3.84a)
(3.84b)
Under the spatial rescaling the front region where A (P) rapidly increases (decreases)
as x decreases from infinity becomes a step as 0. This motivates the introduction
of solutions of the form

P(x,t) P0 1 eG (x,t)/ , A(x,t) A0 (x)eG (x,t)/
(3.85)
with G (x,t) > 0 for all x > x(t) and G (x(t),t) = 0. The point x(t) determines the
location of the front and c = x.
Substituting (3.85) into (3.84a) and (3.84b) gives
128

G
G
2 G
G (x,t)/
=D
D
k
A
(x)
1
e
0
0
t
x
x2

2

G
G
2 G
= A0 (x) D
A0 (x)
D
+k0 P0 1eG (x,t)/ [h+ h(x)]
t
x
x2
+ 2 A
0 (x)G 2 A
0 (x)
G
.
x
Since eG (x,t)/ 0 as 0 for G > 0, it follows that the limiting function

G(x,t) = lim 0 G (x,t) satisfies

G
G 2
=D
k0 A0 (x)
t
x

G
G 2
=D
+ k [h + h(x)],
t
x
where k = ko P0 as before. It immediately follows that

k h h(x)
A0 (x) =
P0 .
k
(3.86a)
(3.86b)
(3.87)
The remaining equation (3.86b) can be analyzed along identical lines to a previous study of the heterogeneous FisherKPP equation [421]. Formally comparing (3.86b) with the HamiltonJacobi equation t G + H(x G, x) = 0, we define the
Hamiltonian
(3.88)
H = Dp2 + k [h + h(x)],
where p = x G is interpreted as the conjugate momentum of x. It now follows that
(3.86b) can be solved in terms of the Hamilton equations
dx
= 2Dp,
ds
d p d h
=
.
ds
dx
(3.89)
Combining these equations yields the second-order ODE

x 2D h(x)
= 0.
(3.90)
This takes the form of a Newtonian particle moving in a potential V (x) =

2D h(x). Given the solution x(s) = (s; x,t) with (0; x,t) = x0 and (t; x,t) = x,
we can then determine G(x,t) according to
G(x,t) = E(x,t)t +
Here
1
2D
t
0
(s; x,t)2 ds.
E(x,t) = H( (s; x,t)/2D, (s; x,t)),
which is independent of s due to conservation of energy.
(3.91)
(3.92)
129
For certain choices of the modulation function h(x), (3.90) can be solved
explicitly [421]. In particular, suppose that the spine density curve in Fig. 3.3 is
approximated by a piecewise linear function, in which the density increases linearly
with distance from the soma to some intermediate location along the dendrite and
then decreases linearly towards the distal end. Assuming that the right-moving wave
is initiated beyond the point , x0 > , then we can simply take h(x) = (x x0)
for > 0. Substituting into (3.90) and integrating twice with respect to s using the
Cauchy conditions gives
(s; x,t) = x0 + (x x0)s/t + D ts D s2.
(3.93)
The corresponding energy function is then

E(x,t) =
2
(x x0)2
+ k h + (x x0) + Dt 2
2
4Dt
2
4
(3.94)
1.3
1.2
wavespeed c
1.1
1
0.9
0.8
k=0.4
k=0.35
k=0.28
0.7
0.6
0.5
0.4
0
10 20 30 40 50 60 70 80 90 100
time t [sec]
Fig. 3.13 Plot of time-dependent variation in wave speed c given by (3.96) for various values
of the
activation rate k. Other parameters are h = 0.24 s1 and D = 1 m2 /s. At t = 0, c(0) = 2 D(k h)
and (3.91) shows that

G(x,t) =
2
(x x0 )2
[k h]t (x x0)t Dt 3 .
4Dt
2
12
(3.95)
We can now determine the wave speed c by imposing the condition G(x(t),t) = 0.
This leads to a quadratic equation with positive solution
130
kh 2 2
+ t
x(t) = x0 + D t 2 + 2Dt
D
3

2
2
2
4 D t
= x0 + ct 1 +
+ D t 2
3c2
with c = 2

D(k h). Finally, differentiating both sides with respect to t yields
c x(t)
=c

c0 2t 2
1 + 0 2t 2 +
+ 2D t,
1 + 0 2t 2
(3.96)
where 0 = 4D2 /(3c2 ). For sufficiently small times such that D t 1, we have the
approximation
c c + 2D t +
2(D t)2
.
c
(3.97)
Figure 3.13 shows example plots of the time-dependent wave speed for various
choices of the activation rate k. It can be seen that there are significant changes
in speed over a time course of 100 s, which is comparable to the time a wave would
travel along a dendrite of a few hundred microns. In principle, it should be possible
to test experimentally the predictions of the above analysis by initiating a translocation wave at different points along a dendrite and determining the corresponding
wave speed.
3.3 Appendix: Pulled and Pushed Fronts

In this appendix, we review some of the basic properties of fronts propagating into
unstable states. For a much more detailed account, see the review by van Saarloos
[544] and Chap. 4 of [422]. For concreteness, we focus on a slight generalization of
the FisherKPP equation
u 2u
= 2 + f (u),
t
x
f C1 [0, 1],
f (0) = f (1) = 0,
(3.98)
for which the homogeneous fixed point u = 0 is unstable ( f

(0) > 0) and u = 1 is
stable ( f
(1) < 0). We also assume that f (u) > 0 for all u (0, 1). We are interested
in determining the longtime asymptotics of a front propagating to the right into the
unstable state u = 0, given initial conditions for which u(x, 0) = 0 for sufficiently
large x. It is not possible to carry out an asymptotic analysis by simply moving to
a traveling coordinate frame, since there is a continuous family of front solutions.
However, considerable insight into the evolution of a localized initial condition can
be obtained by linearizing about the unstable state.
131
u(x,t)
t3
t2
speed
t1
c*
u=a
t0
xa(t1) xa(t2)
xa(t3)
Fig. 3.14 Illustrative sketch of the growth and spreading of a solution u(x,t) of the Fisher equation
linearized about the unstable state u = 0, given a localized initial condition u(x,t0 )
3.3.1 The Linear Spreading Velocity

Linearizing (3.98) about u = 0,
u 2u
= 2 + f
(0)u.
t
x
(3.99)
Substitution of the Fourier mode ei t+ikx gives the dispersion relation
(k) = i( f
(0) k2).
(3.100)
The state u = 0 is then said to be linearly unstable if Im[ (k)] > 0 for some range
of k-values. In the case of the FisherKPP equation, after writing k = kr + iki , this
will occur when f
(0) + ki2 > kr2 . Consider some generic initial condition u(x, 0)
that is sufficiently localized in space (to be made precise later). Since there exists
a range of unstable linear eigenmodes, we expect the localized initial condition to
grow and spread out within the linear regime, as illustrated in Fig. 3.14. Tracking
the evolution of a level curve xa (t) with u(xa (t),t) = a, the linear spreading velocity
c is defined to be the asymptotic speed of the point xa (t) in the rightward moving
edge (assuming it exists):
dxa (t)
c = lim
.
(3.101)
t dt
The linearity of the underlying evolution equation (3.99) means that c is independent of the value a. (Note that for an isotropic medium, the leftward moving edge
moves with the same asymptotic speed but in the opposite direction.) Suppose that
c is finite. If we were to move in the traveling coordinate frame = x ct, then the
leading rightward edge would neither grow nor decay exponentially. Imposing this
condition on the Fourier expansion of the solution u(x,t) then determines c in terms
of the dispersion curve (k). More specifically, denoting the Fourier transform of
the initial condition u0 (x) = u(x, 0) by u0 (k), we have
132
C*
Im( (k))
k*
k0
Re( (k))
Fig. 3.15 Method of steepest descents. Sketch of Re( ) contours in the complex k-plane for an
analytic function in the region of a saddle point at k . The integration contour C is deformed so
that it passes through the saddle point
u(x,t) =
=
=
u0 (k)ei[kx (k)t])
dk
2
k]t)
u0 (k)eik ei[ (k)c

u0 (k)eik e (k)t
dk
2
dk
,
2
(3.102)
where
(k) = i[ (k) c k] = i (k) c ki i[r (k) c kr ]
(3.103)
In the limit t with finite, we can approximate this integral using steepest descents. For the moment, we assume that u(k)
is an entire function (analytic in every
finite region of the complex plane) so that we can deform the contour in the complex
k-plane, that is, (, ) C, with C linking points at infinity in the complex plane
where Re( ) < 0.
Method of steepest descents. We briefly describe the method of steepest descents for a general analytic function (k); see also Fig. 3.15 and [275]. First, one would like to choose C so
that the maximum of r Re( ) along the contour at k0 , say, is as large as possible, since
this point will dominate the integral. Recall, however, that one of the CauchyRiemann
conditions on an analytic function is that 2 (Re( )) = 0, which means that Re( ) cannot
have any maxima or minima (except at singularities or branch points where would be
nonanalytic). Therefore (Re( )) = 0 only at saddle points. Second, for a general integration contour, evaluating the integral in a neighborhood of the point k0 will overestimate the
value of the integral, since it does not take into account cancellations due to the rapidly oscillating function eiIm( (k)) . The latter issue can be eliminated by choosing the contour that
is the path of steepest ascent to a saddle point and the steepest descent away from the saddle.
(If there exists more than one saddle, then one chooses the highest one). By construction,
the path is parallel to Re( ). Hence, from the CauchyRiemann conditions, (Im( )) = 0
133
so that i Im( ) is constant along the contour. In other words, there are no fast oscillations along the path of steepest ascent and descent and one can obtain a good estimate of
the integral by Taylor expanding about the saddle point. Thus, taking i (k) = i (k ) along
C , we have

dk
2

dk
=
u0 (k)eik e (k)t
u0 (k)eik e (k)t
u0 (k )eii (k
)t
eik er (k)t
dk
.
2
Finally, we Taylor expand r (k) to second order in k = k k , noting that

(k ) = 0 and
r
(k ) < 0 at the saddle,
r (k) = r (k ) +
( k)2

r (k ),
2
and approximate the remaining contour integral by a Gaussian. This then gives
I
u0 (k )ei[k +i (k )t] e /4Dt er (k )t ,

4 Dt
(3.104)
where D = r
(k )/2.
Let us now apply steepest descent to the integral (3.102) for (k) given by
(3.103) such that r (k) = i (k) c ki and i (k) = [r (k) c kr ]. At the (unique)
saddle point k at which
(k ) = 0, we have

d (k)
c =
.
(3.105)
dk k=k
Moreover, equation (3.104) becomes
1

2
I
u0 (k )ei[k [r (k )c kr ]t] e /4Dt ,
4 Dt
(3.106)
where
1
D = i
(k ).
(3.107)
2
Finally, we can determine the linear spreading velocity c by requiring that the
asymptotic solution neither grows nor decays with time, r (k ) = 0, which implies
c =
i (k )
.
ki
(3.108)
134
Note that equating real and imaginary parts in (3.105) and combining with (3.108)
means that we have three equations in the three unknowns c , kr , ki . In the particular
case of the FisherKPP equation (3.98),

kr = 0, ki = f
(0), c = 2 f
(0), D = 1.
(3.109)
Since kr = 0, we can combine (3.105) and (3.108) into the single condition c =
c( ), = ki , where

i (i )
dc( )
,
= 0.
(3.110)
c( ) =
d =
It can also be seen that the modulus of u falls off as
2
1
|u(x,t)| e e /4t .
t
(3.111)
A more direct way to derive the function c( ) is to consider the linearized PDE in
the moving frame = x ct,
c
dU
d 2U
=
+ f
(0)U,
d
d 2
and to assume the leading-edge solution U( ) e .

In the above analysis, it was assumed that the Fourier transform of the initial condition u0 (x) was an entire function. This would apply to cases for which u0 (x) is a
Dirac delta function, has compact support, or decays faster than any exponential for
large enough x (e.g., a Gaussian). Now suppose that u0 (x) falls off exponentially for
large x, u0 (x) e x for some . Then u0 (k) has a pole in the upper-half complex
plane at k = k
with Im(k
) = . It follows that when deforming the contour C in
the complex k-plane in order to perform steepest descents, we pick up a contribution from the pole. Taking this into account, it can be shown that, within the linear
regime, initial conditions whose exponential decay rate > lead to profiles that
asymptotically spread with the linear spreading velocity v . On the other hand, if
< then the profile advances at a speed faster than c [544].
So far we have investigated the evolution of a localized initial condition in the
linear regime. It still remains to determine whether or not there are classes of initial
conditions under which the full nonlinear system converges to a unique asymptotic
front solution and how the speed of the front c is related to the linear spreading
velocity c . It turns out that for front propagation into a linearly unstable state, there
are only two possibilities when starting from sufficiently steep initial conditions,
that is, initial conditions that fall off faster than e x [544]:
Pulled front: c = c so that the front dynamics is determined by the behavior in the
leading edge of the front where u(x,t) 0, that is, the front is pulled along by the
linear spreading of small perturbations into the linearly unstable state.
135
Pushed front: c > c so that nonlinearities play an important role in determining the
velocity of the front, pushing it into the unstable state.
In the special case of initial conditions with compact support, it can be proven that
the solution evolves into a front that propagates at the minimal possible wave speed
cmin , which is bounded above and below [15]:

f (u)
c = 2 f (0) cmin < 2 sup

.
(3.112)
u
u
For any concave function, f (u) u f
(0), the lower and upper bounds coincide and
we have a pulled front; this applies to the standard FisherKPP equation where
f (u) = u(1 u). On the other, the upper and lower bounds do not coincide for
concave f (u). The minimal front velocity can then be larger than the linear velocity
indicative of a pushed front. An example of the latter is the GinzburgLandau term
f (u) = u(1 u)(1 + u) with > 0. One finds that for compact initial conditions, a
pulled front is selected when 2, whereas a pushed front is selected when > 2
[33].
3.3.2 Variational Method for Front Velocity Selection

Over recent years a number of methods have been developed to tackle the general
problem of front velocity selection, that is, which of the infinity of possible wave
speeds is selected by a given initial condition. One method analyzes the dynamics
of the front position using HamiltonJacobi theory. Although it is only applicable
to pulled fronts, it has the advantage of being able to tackle the effects of heterogeneities, as illustrated in Sect. 3.2.4 for CaMKII translocation waves. Here we
briefly describe another method due to Benguria and Depassier [36], which is based
on a variational principle. The latter can be applied to both pulled and pushed fronts,
although in many cases it only provides lower and upper bounds for the front velocity. Consider a front solution of the reactiondiffusion Eq. (3.98) in traveling wave
coordinates,
(3.113)
u + cu + f (u) = 0,
with lim u( ) = 1 and lim u( ) = 0, = x ct. Set q(u) = u > 0 for
u (0, 1), so that (3.113) becomes
q(u)
dq
cq(u) + f (u) = 0,
du
q(0) = 0,
q(1) = 0.
(3.114)

Let g C1 [0, 1] be a positive, monotonically decreasing function with 01 g(u)du <

, and set h = g
> 0. We denote the space of admissible functions g by the domain D. Multiplying (3.114) by g/q and integrating with respect to u gives (after
integration by parts)

1
1
f (u)
g(u) du = c
g(u)du.
(3.115)
h(u)q(u) +
u
0
0
136
For fixed u, the functional
f
[q] = qh + g
q

has a minimum at q0 = f g/h with [q0 ] = 2 f gh. It follows that

1
c
0
g(u)du =
1
0
[q(u)]du 2
1
0
f (u)g(u)h(u)du,
that is,
1
c2
f (u)g(u)h(u)du
1
0
g(u)du
I[g].
(3.116)
As we show below, there exists a function g D that maximizes the lower bound
for c in such a way that the equality holds. Hence, we have the variational principle

1
f (u)g(u)h(u)du
0
c = max 2
,
(3.117)
1
gD
0 g(u)du
which can be used to estimate c using a parameterized set of trial functions. This
result can also be used to derive the upper bound of (3.112) [36].
In order to establish the existence of the function g,
we first require [q] = [q0 ]
for all u, that is, hq = f g/q.
Combined with (3.114) this gives
c q
= ,
g
q q
which can be integrated to yield

g(u)
= q(u) exp
u0
u
c
du
q(u
)

(3.118)
for some u0 , 0 < u0 < 1. Since g > 0 on (0, 1) and h = g f /q2 > 0 we deduce that g

C1 [0, 1] and is a positive, decreasing function. It remains to check that 01 g(u)du
<
, which requires determining the behavior near u =
0. Linearizing (3.113) around
u = 0 shows that if u e for with = (c+ c2 4 f
(0))/2, then q u
for u 0. The integral solution for g then implies that
g
1
,
uc/ 1
u 0.

1
< and g D. Finally, in

Therefore, if c < 2 , that
is, c > 2 f (0), then 0 g(u)du
the special case c = 2 f (0), one can take the set of

trial functions g (u) = u 1
with g
D for 0 < < 1 and
show that I[g ] 2 f (0) as 0. This means
c = 2 f
(0) maxg I[g] 2 f
(0), that is, c = maxg I[g].
Chapter 4
Calcium Waves and Sparks
Calcium (Ca2+ ) is one of the most important and well-studied cellular signaling
molecules. From a modeling perspective, it attracts a great deal of interest due to
the fact that calcium signaling often involves complex spatiotemporal dynamics,
including oscillations and waves. There are a number of excellent general reviews
on the modeling of calcium dynamics within cells, including Chap. 7 of Keener
and Sneyd [322] and Falcke [181]. In this chapter, we focus on the mathematical
modeling and analysis of calcium waves and their potential significance in neuronal
calcium signaling; see also the reviews by Berridge [43] and Ross [534]. Although
there are a number of very detailed whole-cell models of calcium signaling, we
will focus on simplified models that are analytically tractable. In particular, we will
highlight some of the mathematical methods used to characterize calcium waves.
4.1 Calcium Signaling in Neurons

In vertebrates, most of the Ca2+ is stored in bones, from where it can be released by
hormonal stimulation to maintain a high extracellular Ca2+ concentration (around
1 mM). On the other hand, active ion pumps and exchangers maintain the cytoplasmic Ca2+ concentration at relatively low levels (around 10100 nM). The resulting
steep concentration gradient across the plasma membrane means that cells are able
to increase their cytoplasmic Ca2+ concentration rapidly by opening either voltagegated or ligand-gated Ca2+ ion channels. Within the context of neuronal processing,
the activation of voltage-gated calcium channels (VGCCs) at axon terminals triggers the release of synaptic neurotransmitter (see Sect. 1.3); VGCCs also contribute
to the generation of dendritic action potentials (see Sects. 1.4.3 and 3.1). A classical
example of ligand gating in neurons is the entry of Ca2+ through NMDA receptors
on postsynaptic dendritic spines. The resulting transient Ca2+ signal is thought to
play a major role in the induction of changes in synaptic strength; see below.
Another major mechanism whereby cells, including neurons, regulate their cytoplasmic Ca2+ concentration is via the intracellular supply of Ca2+ from internal
137
138
4 Calcium Waves and Sparks
neurotransmitters
LGCC
VGCC
mGluR
PLC
ATPase
pump
endoplasmic
reticulum
buffered
Ca2+
RYR
G
plasma
membrane
Ca2+
IP3
IP3 R
Fig. 4.1 Neural calcium signaling. The entry of Ca2+ from outside the cell is mediated by voltagegated (VGCC) and ligand-gated (LGCC) calcium channels. Stimulation of metabotropic glutamate
receptors (mGluRs) produces IP3 second messengers that bind to IP3 receptors (IP3 Rs), which
subsequently release Ca2+ from the endoplasmic reticulum. Both IP3 Rs and Ryanodine receptors (RyRs) are sensitive to Ca2+ , resulting in Ca2+ -induced Ca2+ -release (CICR). The latter can
sometimes result in the propagation of a Ca2+ wave along the dendrites
stores such as the endoplasmic reticulum (ER) and mitochondria. Inositol (1,4,5)trisphosphate (IP3 ) receptors and Ryanodine (Ry) receptors distributed throughout
the ER, for example, mediate the release of Ca2+ into the cytoplasm, whereas Ca2+
ion pumps maintain the relatively high Ca2+ concentration within the ER. The Ry
receptor plays a critical role in excitationcontraction coupling in skeletal and cardiac muscle cells, but is also found in nonmuscle cells such as neurons. One important feature of Ry receptors is that they can undergo Ca2+ -induced Ca2+ -release
(CICR), in which elevated cytoplasmic Ca2+ activates Ry receptors that release further Ca2+ , which then activates other Ry receptors, resulting in a nonlinear regenerative feedback mechanism. The IP3 receptor is similar in structure to the Ry receptor, but is found predominantly in nonmuscle cells and is sensitive to the second
messenger IP3 . The binding of an extracellular ligand such as a hormone or a neurotransmitter to a metabotropic receptor results in the activation of a G-protein and the
subsequent activation of phospholipase C (PLC). This then cleaves phosphatidylinositol bisphosphate (PIP2 ) into diacylglycerol (DAG) and IP3 . The water soluble
IP3 is free to diffuse throughout the cell cytoplasm and bind to IP3 receptors located
on the ER membrane, which then open and release Ca2+ from the ER. The opening
and closing of an IP3 receptor is also modulated by the concentration of cytoplasmic
Ca2+ , so it too can undergo CICR. Yet another mechanism for controlling cytoplasmic Ca2+ is through buffering (binding) to large proteins. Indeed, it is estimated that
4.1 Calcium Signaling in Neurons
139
spine
apparatus
dendrites
ER
nucleus
subsurface
cistern
Fig. 4.2 The endoplasmic reticulum (ER) is a single and continuous membrane-bound organelle
that is distributed throughout the dendrites and axon (not shown) of a neuron [516]. It is responsible for the synthesis and posttranslational modification of most secretory and membrane proteins,
as well as the regulation of Ca2+ levels. The shape of the ER is heterogeneous, but can be divided into three domains: the nuclear envelope, the ribosome-bound rough ER (RER), and the
ribosome-free smooth ER (SER). The RER is constituted primarily by sheets or cisternae, whereas
the SER is predominantly composed of 60100 nm diameter tubules that form irregular polygons
with a common luminal space connected via three-way junctions. The ER present in the soma and
proximal dendritic compartment is rich in ribosomes, corresponding to the RER, whereas the ER
distributed in distal dendrites and axons corresponds mostly to SER and contains only a few sparse
ribosomes. The ER comes into close proximity with the plasma membrane at various locations,
forming subsurface cisternae within the soma and axon initial segment, and the spine apparatus
within dendritic spines
at least 99 % of the total cytoplasmic Ca2+ is bound to buffers. A summary of the

basic extracellular and intracellular mechanisms for controlling the concentration of
cytoplasmic Ca2+ is shown in Fig. 4.1.
The signaling mechanisms and dynamics of Ca2+ release from the ER in neurons
have been much less well studied than extracellular sources of Ca2+ . However, the
ER is a continuous network that is distributed throughout a neuron, extending all
the way into the dendrites and axon and coming into close contact with the plasma
membrane and synapses at many locations; see Fig. 4.2. Hence, as suggested by
Berridge [43], it is likely that the integrative and regenerative properties of the ER
and plasma membrane provide a binary membrane system that regulates a variety
of neuronal process via Ca2+ signaling, including excitability, neurotransmitter release, synaptic plasticity, and gene transcription.
(a) Within the soma and initial axonal segment, the ER forms flattened sheets
known as subsurface cisternae that come into close contact with the plasma
membrane. These internal structures are thought to play an important role in
regulating neuronal excitability. For example, many neurons exhibit significant
after-hyperpolarizations (AHPs) following either a single action potential or a
140
burst of action potentials, which suppress subsequent firing of the cell (spike
frequency adaptation). Often AHPs have both a fast component and a slow
component, which arise from the opening of two separate Ca2+ -activated K+
channels (gKCa1 and gKCa2 ). The fast activation of gKCa1 is due to the entry of
extracellular Ca2+ via VCCNs during the course of an action potential, whereas
the slow activation of gKCa2 is probably due to CICR from Ry receptors located
in the subsurface cisternae.
(b) Within the axon, the ER consists of connecting tubules that run parallel along
the axon. The axonal ER network extends into a synapse, coming into close
contact with the plasma membrane and associated neurotransmitter secretory
vesicles. It is thought that calcium release from Ry and IP3 receptors contributes
to the formation of localized high-concentration bursts of Ca2+ necessary for
the exocytosis (transport and fusion) of vesicles to the plasma membrane.
(c) The ER network is also distributed throughout the dendritic tree and into spines,
where it terminates at the so-called spine apparatus. The latter consists of
closely packed plates of ER that are located within the middle of the spine
and is thus well placed to contribute to Ca2+ signaling during synaptic plasticity. The latter refers to the rapid and long-lasting changes in synaptic strength
that are thought to provide the subcellular substrate for learning and memory. Much of the information regarding synaptic plasticity in mammalian neurons has been obtained from studies of hippocampal and cerebellar neurons
[120, 123, 385, 402]. In particular, it has been found that the same synapses
can be switched rapidly and reversibly from an enhanced state of long-term potentiation (LTP) to a reduced state of long-term depression (LTD). Interestingly,
both LTP and LTD are induced by a transient increase in the local Ca2+ concentration within a spine. It was originally thought that the bidirectional switch only
depended on the amplitude of the Ca2+ signal, with a stronger stimulus inducing LTP and a weaker one LTD. Now, however, it appears likely that the detailed
spatial and temporal structure of the calcium signal may be important. Although
the influx of extracellular Ca2+ through NMDA receptors or voltage-gated
channels is still the major contributor to the induction of LTP and LTD, there
is increasing evidence that there is also a component arising from intracellular
stores [428, 447, 534]. Activation of the ER could be mediated either by Ca2+
itself via CICR or by stimulation of metabotropic glutamate (mGlu) receptors
in the plasma membrane, resulting in the production of IP3 . The fact that CICR
requires both cytoplasmic IP3 and Ca2+ suggests that IP3 receptors could act
as coincidence detectors for mGluR activation and postsynaptic Ca2+ signaling
arising from back-propagating action potentials or dendritic Ca2+ spikes [658].
(d) One of the most dramatic consequences of CICR is the propagation of intracellular Ca2+ waves mediated primarily by the opening of IP3 receptors. These
waves were first observed in nonneuronal cells such as Xenopus laevis oocytes
[367, 490], where the resulting changes in Ca2+ concentration across the whole
cell provided a developmental signal. The discovery of Ca2+ waves in neurons
is more recent [303, 428, 447, 534, 658]. Since these cells have extensive den-
4.2 ReactionDiffusion Models of Ca2+ Dynamics
141
dritic and axonal arborizations, the distribution of the components of the Ca2+
signaling apparatus within the cell plays a crucial role in determining whether
or not local Ca2+ release transforms into a propagating wave, and how far it
travels. For example, in pyramidal neurons of the hippocampus and cortex,
Ca2+ waves are usually observed in the primary apical dendrite and perhaps the
soma, rarely reaching beyond the point where thick dendrites begin to branch.
Since the ER network is distributed throughout the cell, this suggests that there
is a heterogeneous distribution of mGlu receptors and IP3 receptors along the
dendrite. Ca2+ waves rarely propagate from the dendrite to the soma, unless
there is a sufficiently strong and sustained stimulation of a neuron that results
in high levels of cytoplasmic IP3 or higher concentrations of internally stored
Ca2+ . If a Ca2+ wave did reach the soma, then it would generate a strong Ca2+
signal in the soma and nucleus due to the high concentration of IP3 receptors
in these regions. It has been suggested that a Ca2+ wave could allow a strongly
stimulated synapse in the dendrites to signal to the nucleus, where the large
increase in Ca2+ concentration could activate gene transcription, which is a
necessary step in more persistent forms of synaptic plasticity such as late longterm potentiation (L-LTP). However, the role of Ca2+ waves in initiating gene
transcription is still controversial. On the other hand, the likely contribution of
IP3 receptor-mediated Ca2+ release during normal LTP and LTD indicates that
Ca2+ waves could be important in determining the spatiotemporal pattern of
synaptic changes.
(e) Many nonneuronal cell types exhibit spontaneous localized Ca2+ release events
known as sparks [110]. Calcium sparks, which are thought to be the building
blocks of the large regenerative Ca2+ signal that controls contraction in cardiac and skeletal muscle cells, arise from the opening of clusters of RyRs by
local CICR. The frequency of calcium spark events is sensitive to changes in
membrane potential, although they rarely induce calcium waves [111]. Similar
events known as calcium puffs have been found in Xenopus laevis oocytes
[489, 490, 690]. They are also spatially and temporally localized, are fast, and
occur stochastically. However, in contrast to sparks, they are mediated by IP3 Rs
and coalesce more easily to form calcium waves. Localized events in neurons
with spark-like and puff-like properties have recently been observed in slice
preparations, and their frequency can be modulated by synaptic activity and
changes in membrane potential [389, 403]. Although it is likely that they also
occur in vivo, it is not yet clear whether or not they have a specific function.

One of the major challenges in modeling calcium waves is that the ER and cytoplasm comprise two highly interconnected three-dimensional domains (ignoring, for
simplicity, other intracellular compartments such as the mitochondria). Assuming
142
that Ca2+ undergoes normal diffusion within each compartment, together with
buffering, we can write down the following pair of diffusion equations for the Ca2+
concentrations c, ce in the cytoplasm and ER [322]:
c
= (Dc c) Jon + Joff ,
t
r c
(4.1)
and
ce
= (De ce ) Jon + Joff , r e ,
(4.2)
t
where c and e denote the cytoplasmic and ER domains, respectively, Dc and De
are the corresponding diffusivities in the two domains, and the on/off fluxes take into
account the binding/unbinding of Ca2+ to buffering proteins. These equations are
supplemented by various boundary conditions that take into account the exchange
of Ca2+ between the cytoplasm and the plasma membrane or ER. Let Jin denote the
total inward flux of Ca2+ from the plasma membrane to the cytoplasm via voltagegated and ligand-gated ion channels, and let Jp1 be the outward flux due to ATPase
ion pumps in the plasma membrane. Then
Dc c n = Jin Jp1,
r c,m ,
(4.3)
where c,m is the surface of the plasma membrane with unit normal n. Similarly,
let JIP and JRy denote the inward fluxes from the ER to the cytoplasm via IP3 Rs and
RyRs, respectively, and let Jp2 denote the outward flux due to ion pumps in the ER.
This yields
Dc c n = De ce n = JIP + JRy Jp2 ,
r e ,
(4.4)
A schematic diagram of
where e is the boundary of the ER with unit normal n.
all the fluxes is shown in Fig. 4.3.
Let us first consider a space-clamped version of this model, in which one treats
the cytoplasm and ER as two well-mixed homogeneous compartments. In that case,
the various surface and volume fluxes are combined as follows:
and
| c,m |
| e |
dc
= Jon + Joff +
[Jin Jp1 ] +
[JIP + JRy Jp2]
dt
| c |
| c |
(4.5)
| e |
dce
= Jon + Joff
[JIP + JRy Jp2 ].
dt
| e |
(4.6)
Each of the fluxes in these equations corresponds to a component of the so-called

Ca2+ -signaling toolkit [44]. Thus, the equations for Ca2+ dynamics have to be coupled to additional biochemical equations describing the dynamics of ion pumps, the
Ca2+ -dependent gating of the Ry and IP3 receptors, and how the latter are regulated by the second messenger IP3 . Often the external flux Jin is simply taken to be
a linear increasing function of the IP3 concentration. There have been many studies of space-clamped Ca2+ models with particular focus on calcium oscillations
143
plasma membrane
cytoplasm (c)
ER (ce)
^
n
^
Jon
^
Joff
JRy
Jp1
Jon
JIP
Joff
Jin
Jp2
Fig. 4.3 Schematic diagram showing various fluxes in reactiondiffusion model of Ca2+ dynamics
[181, 322]. Although the models differ in their degree of complexity with regards
the component fluxes, they can often be reduced to simpler models that have similar
structure to the FitzHughNagumo (FN) model of excitability. For example, suppose that we ignore fluxes through the plasma membrane so that the total intracellular Ca2+ is conserved, that is, |c |c + |e |ce is a constant. Furthermore, suppose
that there are only two intracellular fluxes, JIP and Jp2 , and the IP3 concentration is
fixed. Ignoring the effects of buffering, one can write down an analogous equation
to (2.3) of the form
dc
= f (c, h),
dt
(c)
dh
= h (c) h,
dt
(4.7)
where c is the concentration of free cytosolic Ca2+ , f (c, h) represents the net flux
into the cytoplasm from the ER, and h is an inactivation variable that plays a similar
role to the recovery variable w of the original FN equations. One interpretation of h
is the fraction of IP3 receptors that are not inactivated by bound Ca2+ . Li and Rinzel
[381] derived such a two-variable model by reducing the classical kinetic model of
Ca2+ -gated IP3 receptors due to De Young and Keizer [692]; see Sect. 4.4.1.
Unfortunately, the complex geometry and heterogeneity of the ER means that
the spatially distributed model given by (4.1)(4.4) is far too complicated to analyze, even when simplified models of the various fluxes are considered. A common
strategy is to ignore the detailed structure of the ER by assuming that the concentrations c and ce coexist at every point in space. This is motivated by the idea that
since diffusion is fast over short distances, local variations due to heterogeneities
are smoothed out. Phenomenologically speaking, this leads to the calcium bidomain
equations
c
= (Dc c) Jon + Joff + c [JIP + JRy Jp2]
(4.8)
t
and
144
ce
= (De ce ) Jon + Joff e [JIP + JRy Jp2].
t
(4.9)
Here c = | e |/|c |, e = | e |/|e | are surface to volume ratios, and Dc , De

are effective diffusion coefficients. Note, however, that in order to determine the
effective diffusion coefficients, which will depend on the geometry of the ER, and
justify the precise form of the factors c , e , it is necessary to derive these equations
from first principles using homogenization theory [224, 322]. Elements of the theory
are described in appendix section 4.6. Now suppose that only the fluxes JIP and Jp2
are included along the lines of the space-clamped model. However, in contrast to
the latter, it is now necessary to keep track of both the cytoplasmic and ER Ca2+
concentrations, since they are spatially varying. A one-dimensional bidomain model
with constant diffusion coefficients then takes the form
c
2c
= Dc 2 + f (c, ce , h)
t
x
ce
2c
= De 2 e f (c, ce , h)
t
x
h
= h (c) h,
(c)
t
(4.10a)
(4.10b)
(4.10c)
where f (c, ce , h) represents the net Ca2+ flux from the ER to the cytoplasm the
spatial dimensions have been fixed such that c = 1. Analysis of traveling wave
solutions of (4.10) proceeds along very similar lines to that of the HodgkinHuxley
equations for membrane voltage excitability; see Chap. 2. Depending on parameter
values, the model system can support solitary waves when the physiological state of
the cytoplasm is excitable, periodic waves when it is oscillatory, and traveling fronts
when it is bistable [322].
However, one important feature that needs to be taken into account in the case
of calcium excitability is buffering. Insights into the effects of buffering can be
obtained by looking at reduced models of excitability such as the bistable equation
or the FN equations [592, 635, 636, 653]. Following [592, 653], consider a simple
extension of the bistable equation (Sect. 2.2). First, note that the basic chemical
reaction for Ca2+ buffering takes the form
k+
P + Ca2+ B,
k
where P is the buffering protein and B is the protein/Ca2+ complex. Let c and b
denote, respectively, the concentration of free cytosolic Ca2+ and buffered Ca2+ .
The extended bistable equation is then (in one spatial dimension)
c
2c
= Dc 2 + f (c) + kb k+c(bt b),
t
x
b
2b
= Db 2 k b + k+c(bt b).
t
x
(4.11a)
(4.11b)
145
Here bt is the total buffer concentration, f (c) takes into account all Ca2+ fluxes, and
the diffusion coefficients of free and buffered Ca2+ are assumed to be constants.
Suppose that the buffer reaction rates k are faster than the time constants of Ca2+
dynamics. The buffer concentration b can be taken to be in the quasi steady state
k b k+c(bt b) = 0, that is,
b=
bt c
,
K+c
K=
k
.
k+
(4.12)
Adding (4.11a) and (4.11b) and eliminating b using (4.12) gives [653]

2
1
c
w(c)
=
+ f (c)
t
1 + (c)
x2
2
Dc + Db (c) 2 c
c
2Db (c)
f (c)
=
,
+
1 + (c) x2 (K + c)(1 + (c)) x
1 + (c)
where
w(c) = Dc c + Db bt
c
,
K+c
(c) =
bt K
.
(K + c)2
(4.13)
It is clear that a mobile buffer (Db > 0) generates a nonlinear advectiondiffusion

equation for c; the advection term vanishes if the buffer is immobile (Db = 0).
The form of the effective reactiondiffusion equation for c suggests making then
change of variables w = w(c), with w, a monotone increasing function of c since
w
(c) = Dc + Db (c) > 0. Denoting the unique inverse of the function w(c) by
c = (w), we can write

w Dc + Db (w) 2 w
=
+
f
(
(w))
,
(4.14)
t
1 + (w)
x2
where (w) = bt K/(K + (w))2 . Now suppose that f (c) has two stable zeros c
separated by an unstable zero c0 with c < c0 < c+ . It immediately follows that
f ( (w)) has three corresponding zeros w , w0 with w < w0 < w+ and w stable.
A traveling front solution of wave speed v, w(x,t) = W (x vt), satisfies the ODE
V +
v
W + f ( (W )).
De f f (W )
(4.15)
Proceeding as in Sect. 2.2, multiply both sides of this equation by W and integrate
to show that

w+
W2
d =
v
f ( (w))dw.
w
Deff (W ( ))
Since the integral on the left-hand side is positive definite, the sign of v is determined
by the sign of the integral on right-hand side. In particular, a right-moving wave
(v > 0) exists provided that
c+
c
(Dc + Db (c)) f (c)dc > 0,
(4.16)
146
where we have converted back to the original variable c. It can be seen that immobile
buffers (Db = 0) have no effect on the existence of a traveling wave; this result also
holds true for wave stability and uniqueness. On the other hand, mobile buffers can
eliminate a traveling wave when they are of high affinity (small K ) or have a large
diffusion coefficient (large Db ).
4.3 Discrete Calcium Sources and the FireDiffuseFire Model

4.3.1 Discrete Calcium Sources and Averaging
The above reactiondiffusion models assumed that Ca2+ release from the ER is homogeneous in space. This neglects the fact that IP3 receptors (as well as Ry receptors) tend to be arranged in clusters. For example, in Xenopus oocytes, the density
of clusters is around 1 per 30 m2 , with each cluster containing about 25 IP3 receptors [322]. Consequently, the propagation of Ca2+ waves tends to be saltatory,
jumping from one cluster to the next. (Clustering of IP3 receptors is also thought to
occur along dendrites [534]). We have previously encountered saltatory waves due
to discreteness effects, namely, action potential propagation along myelinated axons
(Sect. 2.5) and solitary Ca2+ spikes propagating along spiny dendrites (Sect. 3.1).
The latter waves depend on the action of VGCCs and other voltage-gated ion channels and are thus distinct from the Ca2+ waves considered in this chapter, which
depend on Ca2+ release from internal pools. Nevertheless, the analytical techniques
used to study these other examples can be adapted to the present case. First, we
will discuss how the averaging method of Sect. 2.5 can be used to study wave propagation failure in the presence of discrete release sites, following [318]. We will
then describe an alternative, firediffusefire (FDF) model of saltatory Ca2+ waves,
which is identical in structure to the spikediffusespike (SDS) model of Sect. 3.1.
(In contrast to other chapters, we will use c to denote cytosolic Ca2+ concentration
and v to denote wave speed.)
Consider a one-dimensional model of Ca2+ diffusion and release of the
form [318]
c
2c
= kc c + Dc 2 + L (x nL) f (c),
(4.17)
t
x
n
where f (c) represents the Ca2+ -dependent rate of release from each cluster (taken
to be identical), and the decay term represents the effects of ion pumps. The clusters are assumed to be evenly spaced with spatial separation L. For simplicity, the
ER concentration ce and IP3 concentration p are taken to be fixed. It is tempting to
apply homogenization theory to this problem (see appendix section 4.6), assuming
diffusion is fast on the length scale L, that is, L2 kc /Dc 1. Intuitively, one would
expectthe L-periodic function g(x) = L n (x nL) to be replaced by its spatial average 0L g(x)dx/L = 1. However, as highlighted by Keener [318, 319] and discussed
in Sect. 2.5, the homogenized system cannot account for the fact that discreteness
147
effects can lead to wave propagation failure in regimes that the continuous model
would predict traveling waves exist. This reflects the fact that wave solutions are
not structurally stable. The occurrence of propagation failure can be established by
constructing standing wave-front solutions [318, 322]. The latter are stationary solutions of (4.17). On the intervals nL < x < (n + 1)L, this gives
0 = kc c + Dc
2c
,
x2
(4.18)
which is supplemented by jump conditions at x = nL,

c nL+
Dc + L f (cn ) = 0.
x nL
(4.19)
The latter is obtained by integrating the stationary equation on either side of x = nL.
The general solution of (4.18) is
c(x) = cn cosh( (x nL)/L) + cn sinh( (x nL)/L),
nL < x < (n + 1)L,
where 2 = kc L2 /Dc , cn = c(nL) and cn /L = c

(nL+ ). Continuity of the solution
at x = (n + 1)L shows that
cn cosh( ) + cn sinh( ) = cn+1 ,
that is,
cn =
cn+1 cn cosh( )
.
sinh( )
It follows that
c
(nL+ ) = (cn+1 cn cosh( ))
L sinh( )
and
c
(nL ) = (cn1 cn cosh( ))
.
L sinh( )
Substituting these results into the jump condition yields the difference equation
ks
(cn+1 2cn cosh( ) + cn1) + f (cn ) = 0.
sinh( )
(4.20)
It is usually not possible to obtain an explicit solution of a nonlinear difference

equation. One exception is if f (c) is piecewise linear,
f (c) = f0 H(c c ),
(4.21)
where c is a release threshold. Without loss of generality, suppose that the standing
wave solution crosses threshold from below in the interval 0 < x < L. Try a solution
of the form cn = A n for n 0 and cn = C A+ n for n 0 with 0 < < 1.
The constant C is determined by taking the limit n in (4.20) with cn C > c :
148
ks
(C 2C cosh( ) + C) + f0 = 0,
sinh( )
so that
C=
sinh( ) f0
cosh( /2) f0
=
.
2 cosh( ) 2 kc
2 sinh( /2) kc
The constant then satisfies the quadratic equation 2 + 1= 0 with =

cosh( ). Choosing the root that is less than one, we have = 2 1 = e .
Finally, the factors A can be determined by considering the difference equation for
n = 0, 1, which gives [322]
A ( ) =
( )C( )
,
1 + ( )
A+ ( ) = e A ( ).
Given that C, A and are all functions of (assuming that f0 /kc is fixed), the
conditions for the existence of a standing wave-front solution are
A ( ) < c ,
C( ) A+( ) ( ) = e A ( ) c .
(4.22)
For a given , a traveling wave exists if the threshold c lies beneath the curve
A ( ), whereas a standing wave exists if c lies between A ( ) and e A ( ); in
the latter case propagation failure occurs. Note, in particular, that increasing the
threshold c or increasing makes propagation failure more likely. Moreover, increasing corresponds to increasing the cluster spacing L or decreasing the Ca2+
diffusion coefficient Dc .
Having established that wave propagation failure can occur, we now describe
how to calculate the wave speed as a function of discreteness using the averaging
method of Keener [318, 319]. The analysis proceeds along identical lines to the
study of inhomogeneous diffusivity in the discrete bistable equation; see Sect. 2.5.
First, rewrite (4.17) in the dimensionless form
c
2c
= c + 2 + (1 + g
(x/ )) f (c),
t
x
(4.23)

where = L2 kc /D 1, a factor of kc1 has been absorbed into f , and g(x) is the
periodic sawtooth function, g(x) = 1/2 x for 0 < x < 1 and g(x + n) = g(x). (More
precisely, g(x) is assumed to be the sum of Gaussians (2.69) of width , with the
limit 0 taken at the end of the calculation.) Suppose that the nonlinear function
f (c) is one for which there exists a traveling front solution c(x,t) = C(x vt) of
the homogeneous equation (g 0). The main result of the averaging method is that
for sufficiently small, there exists a wavelike solution or pulsating wave of the
form [318]
(4.24)
c(x,t) = C(x (t)) + O( ),
where the phase (t) evolves according to [see also (2.72)]
d
= v ( / )
dt
(4.25)
with
( / ) =
149
g
([ + ]/ ) f (C( ))C
( )ev d ,
and
(4.26)
ev C
( )2 d .
(4.27)
Equation (4.25) implies that the solution is not translationally invariant, rather it
moves with a time-dependent velocity
. If v ( / ) is strictly positive then

(t) is a positive, periodic function of t with period
T=

0
d
.
v ( / )
(4.28)
The mean speed of the wave is v = /T . On the other hand. if v ( / ) vanishes

for some , then propagation failure is expected to occur.
Averaging method for discrete Ca2+ release sites. The first step is to rewrite (4.23) as the
first-order system
(4.29)
cx = u, ux = ct (1 + g
(x/ )) f (c) + c.
Introducing the exact change of variables u = w g(x/ ) f (u) yields the new system
cx = w g(x/ ) f (c),
wx = ct f (c) + c + g(x/ ) f
(c)cx .
(4.30)
It can be seen that if the O(1) system of equations is independent of x/ , then the lowestorder averaged system reduces to the homogeneous equation, which we are assuming supports a traveling front solution C(x vt). Including the inhomogeneous factors g(x/ )
means that the system is no longer translationally invariant. However, we can look for solutions that are close to a traveling front by going to a traveling coordinate system with
= x (t) to give [318]
c w = g([ + ]/ ) f (c)
(4.31a)
ct + w + c + f (c) c = g([ + ]/ ) f (c)c .
(4.31b)
We now seek a perturbative solution of the form

c = C( ) + c1 ( ,t) + . . .,
w( ) = C
( ) + w1 ( ,t) + . . .,

(t) = v + 1
(t) + . . .
Substituting into equations (4.31) and collecting terms of equal powers in yields a hierarchy of equations, the first of which is
c1 w1 = g([ + ]/ ) f (C)
(4.32a)
t c1 + w1 + vw1 + f
(C)c1 c1 = vg([ + ]/ ) f (C) 1
(t)C
+ g([ + ]/ ) f
(C)C
.
(4.32b)
Equations (4.32) can be rewritten as

0
c1
hc
+L
=
,
t c1
w1
1
C
+ hw

L=
f
(C) c
1
+ c

,
(4.33)
with hc and hw determined by inhomogeneous terms on the right-hand side of equations

(4.32a) and (4.32b), respectively. Following our analysis of the linear operator (2.62), we
know that the matrix operator appearing in equation (4.33) has a null space spanned by
(C
,C
). Similarly, the adjoint operator
150

L =
f
(C) C
+ c
has the null vector ec (C
,C
). Hence, applying the Fredholm alternative, it follows that

(t) must satisfy the phase equation (4.25) with (t) = v + 1 (t).
For the particular choice f (c) = H(c c) used in the analysis of standing waves
with f0 /kc = 1, the homogeneous equation is identical to the one analyzed at the end
of Sect. 2.2. Thus, there exists a unique traveling front solution of the form (2.23)
with corresponding wave speed (2.24). The mean wave speed can then be calculated
explicitly along similar lines to Sect. 2.5, so that
v =
.

2 ln v+ /
(4.34)
v /

with v = (1 2c)/ c c 2 and = 4[c c 2 ]. This would predict propagation
failure when v, that is,

4(1 2c)( c c2 ).
(4.35)
Since the analysis is based on the assumption that is small, we expect this
condition to be reasonable when c 0.5. Taylor expanding about this point the condition for propagation failure is approximately > 2(1 2c ). Similarly, the condition (4.22) based on the standing wave analysis predicts propagation failure when
/(2(e 1)) < c , which reduces to > 2(1 2c) when c 1/2. Thus the two
approaches agree for small and yield qualitatively similar results for large . Finally, as in the analysis of myelinated axons, a cubic nonlinearity for the release
function f (c) yields an expression for the mean wave speed that involves exponentially small terms so it is necessary to include higher-order terms in the perturbation
analysis in order to obtain sufficient accuracy [318].
4.3.2 The FireDiffuseFire Model of Ca2+ Release

An alternative approach to studying the effects of discrete Ca2+ release sites is to
consider the so-called FDF model [127, 137, 140, 325, 497, 511]. The basic idea of
the model is that once the Ca2+ concentration reaches a threshold value c at a given
release site, that site fires and instantaneously releases a fixed amount of Ca2+
into the cytoplasm [325, 497, 511]. It is also straightforward to take into account the
effects of ion pumps by including a decay term and replacing instantaneous release
by a fixed pulse of finite width [127]. The FDF model is then almost identical to
the SDS model of Ca2+ spikes propagating along spiny dendrites [129, 130]. Since
the latter model was analyzed in detail in Sect. 3.1, we can carry over the previous
results after minor modifications. Therefore, consider the following version of the
FDF model [127]:
c
c
2c
= + Dc 2 + (x xn ) (t Tn ),
t
c
x
n
151
(4.36)
where Tn is the firing time of the site xn , which only fires once:
Tn = inf{t | c(xn ,t) c }.
For concreteness, let
(4.37)
(t) = ( /d )H(t)H(d t),
so that each Ca puff or spark is of size and duration d . Suppose that the Ca2+
release sites are regularly spaced, xm = mL, and consider a saltatory wave solution
Tm = m with speed v = L/ . The existence of such a wave can be determined
straightforwardly by noting that (4.36) is identical to (3.33) under the mapping
2+
n
0 .
rs
d
V c,
The one major difference between the two models is that thresholding in the SDS
model is mediated by an IF process (3.2), which introduces an additional time constant . Therefore, in order to apply the analysis of saltatory waves along a spiny
dendrite to the FDF model (see in Sect. 3.1.3), we need to take the limit 0. This
then gives the following threshold condition for the existence of a saltatory wave:
c =
H(nL, n ),
(4.38)
n=1
with
H(x,t) =
d
0
G0 (x,t s)ds,
G0 (x,t) =
2
1
et/c x /4Dct .
4 Dc t
(4.39)
In the limit d 0, H(x,t) G0 (x,t) and the threshold condition becomes

c =
G0 (nL, n )
n=1
2

Lt
1
exp n
+
=
g( ).
4Dc c
n=1 4 Dc n
(4.40)
In the additional limit c (i.e., neglecting the flux due to ion pumps), the results
of [497] are recovered. In particular, g( ) becomes a monotone function of , and
the speed of the traveling waves scales linearly as Dc /L. The latter follows after introducing the dimensionless threshold c = c L/ and noting that Lg/ is then only
a function of the dimensionless quantity Dc /L2 . However, this result is inconsistent with the analysis of standing waves, which showed that propagation failure can
occur by decreasing the diffusion coefficient. The discrepancy is resolved by taking
into account the effects of ion pumps, that is, taking c to be finite. Moreover, g( )
152
is now non-monotone and consequently one finds two solution branches with the
faster one corresponding to stable waves [127]. This result also holds in the case of
finite pulse width d , for which
H(x,t) =
[A(x,t min(t, d )) A(x,t)] ,

d
(4.41)
with A(x,t) given by (3.40). In summary, the main qualitative results of the SDS
model carry over to the FDF model, for example, one finds that propagation failure
occurs if the spacing between release sites becomes too large.
One of the limitations of the above FDF model is that it neglects variations in the
Ca2+ concentration within the ER. However, in many cells, Ca2+ release can result
in a significant local depletion of ER Ca2+ . In order to take this into account, the
FDF model has been extended in the form of a bidomain threshold-release model
[621, 622]:
c
2c
= Dc 2 + Jrel (c, ce ) J2(c, ce )
t
x
ce
2c
= De 2 e [JIP (c, ce ) J2(c, ce )] ,
t
x
(4.42a)
(4.42b)
where the release and uptake fluxes, Jrel (c, ce ) J2(c, ce ), are functions of the cytosolic and ER Ca2+ concentrations. For simplicity, assume that there is a continuous and uniform distribution of Ca2+ release sites and that the release flux depends
on the difference in the cytosolic and ER Ca2+ concentrations:
Jrel (x,t) Jrel (c(x,t), ce (x,t)) = (ce (x,t) c(x,t)) (t T (x)),
with
T (x) = inf{t | c(x,t) c }.
(4.43)
(4.44)
The release sites could be either RyRs or IP3 Rs; in the latter case the IP3 concentration p is held fixed. Finally, the reuptake flux due to the action of ion pumps is taken
to have the linear form
c ce
(4.45)
J2 (c, ce ) = .
c e
The existence and stability of a solitary pulse solution of (4.42) can be investigated
along similar lines to Sect. 3.1.1.
4.4 Stochastic Models of Ca2+ Release

The fluorescent imaging of localized Ca2+ puffs and sparks has established that
Ca2+ release is a stochastic process that occurs at spatially discrete sites consisting
of clusters of IP3 Rs and RyRs, respectively. In Xenopus oocytes, Ca2+ puffs have
an amplitude ranging from around 50 to 600 nM, a spatial spread of approximately
153
6 m and a typical duration of 1 s [489, 490, 690]. For sufficiently high levels of
IP3 concentration, the amplification of Ca2+ puffs by CICR can lead to the formation of Ca2+ waves. Ca2+ sparks in heart and skeletal muscle tend to be of shorter
duration and less spatial spread and are less likely to result in wave propagation
[110, 111]. As we have already mentioned Ca2+ puffs and sparks have also been
observed in neurons [534], suggesting that one needs to take into account stochastic release events when modeling Ca2+ waves. Typically, one combines a stochastic
model of localized release through RyRs or IP3 Rs with a deterministic reaction
diffusion model of Ca2+ waves such as the FDF model [137, 320]. The stochastic
modeling of the Ca2+ -dependent opening and closing of these receptors proceeds
along analogous lines to the modeling of voltage-gated ion channels considered in
Sect. 1.5.
4.4.1 Stochastic Model of Ca2+ Puffs in a Cluster of IP3 Rs

Stochastic models of Ca2+ puffs typically treat a cluster of IP3 Rs as a set of N channels that open and close independently, but are indirectly coupled by the common
cytoplasmic Ca2+ concentration [180, 182, 579, 612]. Models differ in the level of
detail regarding individual receptors. The first deterministic kinetic model of Ca2+ gated IP3 Rs was proposed by De Young and Keizer, in their study of agonist-induced
Ca2+ oscillations. This model assumes that the IP3 receptor consists of three equivalent receptor subunits, all of which have to be in a conducting state in order to
generate a Ca2+ flux. Each subunit is taken to have an IP3 -binding site, an activating Ca2+ -binding site, and an inactivating Ca2+ -binding site; the conducting state
corresponds to the state in which all subunits have the first two binding sites occupied but the third unoccupied. Although the De YoungKeizer model is simple
to describe, it involves a relatively large number of variables that have to be coupled to the Ca2+ and IP3 concentrations. A simplified version of the model was
subsequently developed by Li and Rinzel [381]. They exploited the fact that the
binding of IP3 and activating Ca2+ are fast relative to inactivating Ca2+ and used
a quasi-steady-state argument to reduce the eight-state subunit model to a model
that simply keeps track of whether or not the inactivating Ca2+ -binding site of a
subunit is occupied. More specifically, the LiRinzel model is a two-variable model
given by
dc
= JIP + Jleak Jp2
dt
dh
= h (1 h) hh,
dt
(4.46a)
(4.46b)
where c is the cytoplasmic Ca2+ concentration, h is the fraction of receptors in a

cluster not inactivated by Ca2+ , and p is the IP3 concentration, which is assumed
fixed. The three Ca2+ fluxes included in the model are the channel flux JIP and
154
leakage flux Jleak from the ER to the cytoplasm, and the flux Jp2 pumped back into
the ER; see also Fig. 4.2. The expressions for the various fluxes are
JIP = f (c, p)h3 [ce c],
Jleak = 0 [ce c],
JP2 =
1 c2
,
2
k3 + c2
(4.47)
with

f (c, p) =
p
p + k4
3
c
c + k5
3
h = 2
p + k4
,
p + k6
h = 3 c.
(4.48)
The various cubic terms reflect the existence of three subunits. Parameter values
of the model can be found in [381]. We see that the simplified model resembles
the HodgkinHuxley model (see Sect. 1.1), after replacing Ca2+ concentration c by
membrane voltage v and ce by a reversal potential.
We now describe a stochastic version of the LiRinzel model for a cluster of
IP3 Rs due to Shuai and Jung [579]. For stochastic versions of the full De Young
Keizer model, see, for example, [180, 182, 244, 612]. The deterministic equations
(4.46) describe the mean behavior of a large cluster of Ca2+ channels, just as
the HodgkinHuxley equations for membrane voltage apply to a large number of
voltage-gated ion channels. As we discussed in Sect. 1.5, if the number of channels
is relatively small, then it is necessary to take into account thermally driven fluctuations in the opening and closing of individual channels. In the case of the LiRinzel
model, one only needs to consider the opening and closing process for the gating
variable h of each subunit. The latter is modeled as the two-state Markov process
h
C(closed) O(open).
h (c)
(4.49)
Suppose that there are N independent IP3 Rs, each with three independent subunits
labeled i = 1, 2, 3 that are described by the above two-state Markov process. Let
Ni (t) (i = 1, 2, 3) denote the number of receptors at time t that have the ith subunit
open. Under the adiabatic assumption that the Ca2+ concentration c evolves much
more slowly than the state transitions of the channels, we can write down a master
equation for the probability P(ni ,t) = Prob[Ni (t) = ni |Ni (0) = n0 ] according to
dP(ni ,t)
= (N ni + 1)h P(ni 1,t) + (ni + 1)hP(ni + 1,t)
dt
(ni h + (N ni )h )P(ni ,t), i = 1, 2, 3.
(4.50)
As with voltage-gated ion channels, we have a stochastic hybrid system, since the
transition rate h depends on the Ca2+ concentration c(t), which evolves according
to a piecewise deterministic equation of the form (4.46a). The latter, in turn, couples
to the discrete stochastic variables Ni (t) through the flux
Ni (t)
.
i=1 N
3
JIP = f (c(t), p)[ce c(t)]
(4.51)
155
[Note that one should really write down a differential ChapmanKolmogorov

equation for the joint probability density p(n1 , n2 , n3 , c,t) along the lines of
(1.126)]. Finally, for large N, one can obtain a further simplification by carrying out a KramersMoyal expansion of the master equation (4.50) along the lines
of Sect. 1.5.2. This yields the following SDE for Hi (t) = Ni (t)/N with Hi treated as
a continuous stochastic variable:
1
dHi = h (1 Hi ) hHi + b(Hi )dWi ,
N
where
b(Hi ) =
(4.52)

h (1 Hi) + hHi ,
and Wi (t) is an independent Wiener process with

dWi (t) = 0,
dWi (t)dW j (t
) = (t t
)dt dt
i, j .
Shuai and Jung [579] simulated the stochastic LiRinzel model in order to investigate the effects of noise on Ca2+ oscillations in a space-clamped model. They
assumed that the deterministic system (4.46) was monostable at low and high IP3
concentrations and exhibited limit cycle oscillations (occurring via a Hopf bifurcation) at intermediate concentrations. They showed that noise can enlarge the range
of IP3 concentrations over which oscillations occuran effect known as coherence
resonance. They also found a broad distribution of puff amplitudes, lifetimes, and
interpuff intervals. In particular, at low IP3 concentrations, the amplitude distribution is a monotonically decaying function, whereas at higher concentrations, it is
unimodal. This suggests that Ca2+ puffs become more significant as IP3 concentration is increased and hence could impact the spontaneous generation of Ca2+ waves.
This issue was investigated numerically by Falcke [180] using a stochastic version of
the De YoungKeizer model that was incorporated into a reactiondiffusion model
of spatially distributed channel clusters. He showed that there is indeed a transition from Ca2+ puffs to waves as the IP3 concentration is increased. At low concentrations, only puffs occur, since there is not enough Ca2+ released to stimulate
neighboring clusters, which means that the response is purely local. However, as IP3
concentration increases, global Ca2+ waves can emerge from local nucleation sites
of high Ca2+ concentration. At intermediate levels of IP3 , global events are rare
and waves only progress a short distance before dying out. On the other hand, for
higher IP3 concentrations, global waves occur regularly with a well-defined period.
Again this oscillatory-like behavior can occur in parameter regimes for which the
deterministic model is non-oscillatory.
4.4.2 Stochastic Model of Ca2+ Sparks in a Cluster of RyRs

We now turn to a stochastic model of Ca2+ sparks due to Hinch [276]; see also [241].
This model was originally developed for cardiac myocytes (heart muscle cells) and
156
includes details of the geometry of Ca2+ release units, in particular, the narrow
junctional gap known as the diadic space that separates the sarcoplasmic reticulum
(SR) from the plasma membrane; see Fig. 4.4. (In smooth muscle cells the smooth
ER is referred to as the sarcoplasmic reticulum.) However, it is possible that a similar mechanism occurs in neurons at the specialized subsurface cisternae of the ER,
which also form narrow junctions with the plasma membrane; see Fig. 4.2. In a typical myocyte, there could be up to 10,000 Ca2+ release units, each one containing
a cluster of around N = 50 RyRs on the surface of the SR. The cluster of RyRs
is apposed to L-type Ca2+ channels located on so-called t-tubules, which are invaginations of the plasma membrane into the myocyte. (The Ca2+ channels are not
involved in the spontaneous generation of Ca2+ sparks so are ignored in the model.)
The diadic space separating the SR from the t-tubules is a region of the mytoplasm
(intracellular fluid of myocytes), which is approximately cylindrical in shape with
width 10 nm and radius 100 nm. Since the diadic space is a small enclosed volume,
it supports an elevation in Ca2+ concentration relative to the bulk mytoplasm following the release of Ca2+ from an RyR. Such a local elevation plays a crucial role
in the Ca2+ -induced Ca2+ -release (CICR) that results in a Ca2+ spark. The SR in a
neighborhood of the RyRs is known as the junctional SR (JSR), which may have a
different Ca2+ concentration from the bulk or network SR (NSR).
We now briefly introduce the model of Hinch in nondimensional form; details of
model approximations and estimates of experimentally based model parameters can
be found in [276]. First, the diadic space is modeled as a single compartment with
Ca2+ concentration c satisfying the current conservation equation
dc
= JRyR JD .
dt
(4.53)
Here D is a time constant, JRyR is the total Ca2+ current through the RyRs, and JD
is the diffusive current from the diadic space to the bulk mytoplasm. The latter is
modeled as the Fickian current
JD = c cm,
(4.54)
where cm is the bulk mytoplasm Ca2+ concentration. The total current through the
RyRs is taken to be proportional to the number n of open RyRs times the Ca2+
concentration csr in the JSR:
n
(4.55)
JRyR = csr x, x = ,
N
with N the total number of RyRs in the cluster. Each RyR has Ca2+ -binding sites,
which can be activating or deactivating. When an RyR is in an activated state it is
promoted to a mode where it continuously opens and closes according to a Markov
process (see Sect. 1.5), with a mean open time of 1 ms [695]. The opening of an
RyR channel results in an extra Ca2+ current flowing into the diadic space, which
increases the rate at which Ca2+ binds to the other RyRs via CICR, thus creating
a positive feedback loop. This feedback loop provides a mechanism for bistability.
157
Note that the RyRs also contain inactivating Ca2+ -binding sites, but these do not
play a role in initiating a Ca2+ spark so are not included in the model. For simplicity,
the RyRs are modeled using a two-state Markov process involving a single closed
state and a single open state (see also [324]):
k+ (c)
C(closed) O(open),
(4.56)
with transition rates

k+ (c) =
1 c
,
ko c + 1
k =
1
.
o
(4.57)
t-tubules
JD
JD
JRyR
diadic
space
RyRs
junctional SR
bulk
mytoplasm
JSR
network SR
Fig. 4.4 Schematic diagram of a Ca2+ release unit in a cardiac myocyte. It is divided into four
compartments: the network SR, the junctional SR, the diadic space, and the bulk mytoplasm. See
text for details
Here o is the mean open time of an RyR, is the number of Ca2+ ions that are
needed to open an RyR, and k determines the proportion of time the RyRs are open.
It is assumed that the RyRs are gated independently of each other. They are, however, indirectly coupled via the Ca2+ concentration in the diadic space. The time
constant D of diffusive flux from the diadic space is several orders of magnitude
smaller than the mean open time o of an RyR, that is, D 3 s whereas o 1 ms.
Therefore, the Ca2+ concentration in the diadic space can be taken to be in quasiequilibrium, D 0, so that
c = cm + csr x.
(4.58)
It follows that the transition rate can be reexpressed as a function of the fraction of
open channels and the Ca2+ concentration in the SR, k+ = k+ (cm + csr x).
Now consider N independent RyRs within a Ca2+ release unit, each described
by the above two-state Markov process. Let N(t) be the number of open channels at
158
time t and set pn (t) = Prob[N(t) = n|N(0) = n0 ]. Following along similar lines to
Sect. 1.5.1, it can be shown that pn (t) evolves according to a master equation with
transitions rates (n/N) : n n 1:
+ (x) = N(1 x)
(cm + csr x)
,
k((cm + csr x) + 1)
(x) = Nx.
(4.59)
For the moment, it is assumed that cm and csr are fixed so that + can be treated as
a function of x alone. (In Sect. 4.4.4, the dynamics of csr following initiation of a
Ca2+ spark will also be taken into account.) The units of time are fixed by setting
o = 1. The master equation for the RyR cluster is then
d pn
= + ((n 1)/N)pn1 + ((n + 1)/N)pn+1 (+ (n/N) + (n/N))pn ,
dt
(4.60)
with boundary conditions pN+1 = p1 0. Multiplying both sides of the master
equation (4.60) by n/N and summing over n gives
dn/N
= + (n/N) (n/N),
dt
(4.61)
where = N , and the brackets . . . denote a time-dependent ensemble averaging over realizations of the stochastic dynamics, that is, A(n/N) = n pn A(n/N)
for any function of state A(n/N). In the thermodynamic limit N (infinite cluster size), statistical correlations can be ignored so that one can take the mean-field
limit (n/N) (n/N). This then yields a deterministic equation for the
fraction x of open RyRs:
dx
(cm + csr x)
= + (x) (x) = (1 x)
x.
dt
k((cm + csr x) + 1)
(4.62)
It can be shown that, for physiologically reasonable parameter values, this equation
exhibits bistability [276], that is, there exists a pair of stable fixed points x separated by an unstable fixed point x0 . The fixed point x 0 represents a quiescent
state, whereas the other fixed point x+ represents a Ca2+ spark in which a significant fraction of RyRs are in the active mode and can be interpreted as a burst phase.
Hence, this model operates in a completely different regime to the model considered by Shuai and Jung [579], which was in a monostable state (a fixed point or
limit cycle) in the deterministic limit. Bistability of the Hinch model means that a
Ca2+ spark is a distinct event rather than a broad distribution of events.
Noise-induced transitions from x to x+ determine the distribution of inter-spark
intervals, just as noise-induced transitions from x+ to x determine the distribution of spark lifetimes. Hence, estimating the mean time for the occurrence of a
spark event reduces to the problem of calculating the MFPT to reach x+ , starting
from a neighborhood of x , by crossing x0 . (The dominant contribution to this
MFPT is the time to reach x0 , since the system then quickly relaxes to x+ .) We
encountered an analogous problem in Sect. 1.5, where we considered the mean time
159
for spontaneous action potential initiation due to ion-channel fluctuations. We first

estimated the MFPT by carrying out a diffusion approximation of the underlying
master equation (1.104), and calculating the MFPT of the corresponding Fokker
Planck equation. However, the diffusion approximation resulted in exponentially
large errors. Therefore, we considered an alternative approach in Sect. 1.6, which
was based on WKB methods and asymptotic analysis. Here we consider a corresponding WKB approximation of the master equation (4.60). Such an approach has
been used increasingly to analyze escape problems in chemical and biological systems [160, 164, 175, 176, 256, 339, 340], including the Ca2+ release model of Hinch
[277]. Note, however, that in contrast to the analysis of membrane voltage fluctuations in Sect. 1.6, the analysis of calcium sparks is performed with respect to the
small parameter 1/N rather than . That is, we carry out a system-size expansion
rather than assuming fast kinetics. For the moment, we simply state the results and
leave the details to Sect. 4.4.3.
If x is not on the boundary x = 0 then a standard analysis shows that the mean
time i to initiate a Ca2+ spark starting from the quiescent state x is
i =
with
2

eN[ (x0 ) (x )] .
+ (x ) | (x0 )| (x )

(x) +
(x)
,
(x) =
(x) + (x)
(x) =
x
(4.63)

(y)
ln
dy.
+ (y)
Similarly, the mean duration f of a spark, which corresponds to the mean time to
transition back from x+ to x , is given by
f =
2

eN[ (x0 ) (x+ )] .
+ (x+ ) | (x0 )| (x+ )
(4.64)
Note that at a fixed point x j , + (x j ) = (x j ) so that sign( (x j )) = sign(

(x j )
+
(x j )). Thus, (x0 ) < 0 (unstable) and (x ) > 0 (stable). One also finds that
(x0 ) > (x ) so that i , f are exponentially large. It turns out that in the case of
Ca2+ release, the quiescent state x is in an O(1/N) neighborhood of the boundary x = 0, so that the prefactor of the MFPT has to be modified accordingly; see
[276, 277] for details. Nevertheless, the leading order exponential is unchanged.
From the perspective of modeling stochastic Ca2+ waves using the FDF model of
Sect. 4.3.2, the main result that emerges from the analysis of Ca2+ sparks is that the
deterministic threshold mechanism needs to be replaced by a probabilistic mechanism. In the deterministic case, a cluster of receptors fires whenever the intracellular
Ca2+ crosses a fixed threshold. On the other hand, in a stochastic FDF model, Ca2+
release is a probabilistic process. Given the MFPT i , the probability of a spark event
in a time interval t is given by
Pi (t) = 1 et/i .
(4.65)
160
Recall that the transition rates , and thus i , are functions of the Ca2+
concentration cm of the mytoplasm, which has so far been assumed fixed. Indeed,
one finds that the probability of release in a given time interval t is a sigmoidal
function of cm [276]. This stochastic release process can now be incorporated into
a stochastic version of the FDF model by taking the sequence of release times Tn
in (4.36) to be generated according to a sigmoidal probability distribution that is
parameterized by the Ca2+ concentration in the mytoplasm, which itself evolves
according to (4.36); see [137, 320] for details. The resulting model can be used to
simulate the spontaneous occurrence of Ca2+ waves.
4.4.3 WKB Method and Quasistationary Approximation

We now present the details of how to calculate the rate of escape from a metastable
state. Although, we focus here on the particular problem of Ca2+ release, the basic approach applies to the master equation of any bistable system in the large N
(weak noise) limit. Throughout the analysis we will switch between n/N and x,
with x treated as a continuous variable; this is a reasonable approximation when
N is large. First, note that the master equation (4.60) with the given reflecting
boundary conditions at x = 0, 1 has a unique stationary state given by (see also
Sect. 1.5.1)

n
n
+ ((m 1)/N)
+ ((m 1)/N)
= pk exp ln
pn = pk
. (4.66)
(m/N)
(m/N)
m=1
m=k+1
The functions (x) are smooth and the second derivatives of ln (x) are bounded
(except in the limits x 0 and x 1). Therefore, away from the boundaries, the
sums can be approximated by the trapezium rule [276]
n
m=k
so that
ln( f (m/N)) =
ln( f (k/N) + ln f (n/N)

+N
2
N
eN (x) ,
pn =
+ (x) (x)
n/N
k/N
ln( f (y))dy + O(1/N),
x = n/N,
(4.67)
where (x) is the effective potential
(x) =
x

(y)
ln
dy,
+ (y)
(4.68)
and N is a normalization factor such that Nn=0 pn = 1.

Now suppose that we place an absorbing boundary at the unstable fixed point
x0 = n0 /N. Although there no longer exists a stationary solution, the flux through
the absorbing boundary is exponentially small for large N, so that we can use a
161
spectral projection method analogous to Sect. 1.6. First, rewrite the master equation
(4.60) for n = 0, . . . , n0 as the linear system
dp
= Qp,
dt
(4.69)
where p = (p0 (t), p1 (t), . . . , pn0 (t))T , Q is the matrix of transition rates, and
pn0 (t) = 0 (absorbing boundary condition). Suppose that the eigenvalues of Q
are ordered according to 0 > 0 Re[1 ] Re[2 ] . . . with corresponding eigen( j)
functions n , and consider the eigenfunction expansion
pn (t) =
n0
Cr ert n
(r)
(4.70)
r=0
We assume that |0 | e N for = O(1), whereas r for r > 0 are only weakly
dependent on N. It follows that all other eigenmodes decay to zero much faster than
the perturbed stationary density. Thus at large times, we have the quasistationary
approximation
(0)
(4.71)
pn (t) C0 e0t n .
One can now use a WKB ansatz (see below) to generate a quasistationary solution n for which Q = 0 and n0 O(e N ). Since the WKB solution does not
satisfy the absorbing boundary condition, it is necessary to perform an asymptotic
expansion in order to match the quasistationary solution with the solution in a neighborhood of x0 . In the process this determines 0 , whose inverse can be identified as
the MFPT to escape from the basin of attraction of the metastable state x .
Dropping exponentially small terms and writing n = (x) with x treated as a
continuous variable, we have
0 = + (x 1/N) (x 1/N)
(4.72)
+ (x + 1/N) (x + 1/N) (+(x) + (x)) (x).

We seek a WKB solution of the form
(x) K(x; )e (x)/ ,
(4.73)
m
with K(x; )
m=0 Km (x). Substituting (4.73) into (4.72), Taylor expanding
with respect to , and collecting the O(1) terms gives
+ (x)(e
(x)
1) + (x)(e
(x)
1) = 0,
(4.74)
where
= d /dx. Solving this quadratic equation in e shows that
x
ln
(y)
dy or = constant.
+ (y)
(4.75)
162
Proceeding to the next level, equating terms at O( ) gives

K

K0
+ e 0 +
+
+ e
+
e + e = 0.
K0
2
K0
2
Substituting for using (4.74) and solving for K0 yields the following leading order
forms for :
A
(x) =
(4.76)
eN (x) .
+ (x) (x)
with given by (4.75), which is sometimes called the activation solution, and
(x) =
B
,
+ (x) (x)
(4.77)
which is sometimes called the relaxation solution. The constants A, B are determined
by matching solutions around x0 . Clearly, (4.77) is singular at any fixed point x j ,
where + (x j ) = (x j ), so is not a valid solution for the required quasistationary
density. On the other hand, it does have an interpretation in terms of a corresponding
HamiltonianJacobi formulation of the WKB ansatz.
The WKB ansatz was also used in the analysis of metastability in stochastic
ion-channel models in Sect. 1.6 and in the analysis of sharp interfaces for slowly
modulated pulled fronts; see Sect. 3.2.4. Following along similar lines to the latter
example, we can formally interpret (4.74) as a stationary HamiltonJacobi equation
H(x, (x)) = 0 for , with Hamiltonian

H(x, p) =
r (x) [erp 1].
(4.78)
r=
This suggests a corresponding classical mechanical interpretation, in which H determines the motion of a particle with position x and conjugate momentum p.
A trajectory of the particle is given by the solution of Hamiltons equations
x =
H
= rr (x)erp
p r=1
p =
H
r
=
(x) [erp 1].
x r=1
x
(4.79)
(4.80)
Here the time t should be viewed as a parameterization of paths rather than as a real
time variable. Introducing the Lagrangian
L(x, x)
= p x H(x, p),
(4.81)
= 0 corresponds to the classical action evaluated

it follows that (x) with (x)
along the least-action trajectory from x to x:
(x) =
T
inf
x(t0 )=x,x(T
)=x 0
L(x, x)dt.
(4.82)
163
6
5
4
3
activation trajectory
p
2
relaxation trajectory
1
x-
x+
x0
1
0
0.2
0.4
0.6
0.8
x
Fig. 4.5 Phase portrait of Hamiltonian equations of motion for = /N given by equation
(4.59) with cm = 0.01, ce = 4, = 4 and k = 0.8. The zero-energy solutions are shown as thicker
curves
In terms of the underlying stochastic process X(t), the least-action path can be
interpreted as the most probable fluctuational path from x to x (in the large-N limit)
[205, 398]. Since p = S
everywhere along this path, we have
(x) =
x
x
p(x
)dx
,
(4.83)
with the integral taken along the trajectory. It follows that the leading order term in
the WKB approximation is determined by finding zero-energy solutions p = p(x)
such that H(x, p(x)) = 0. One solution is p = 0 or = constant, which represents
the classical action along a deterministic (or relaxation) trajectory. For example,
once the system escapes from the metastable state x , it tends to rapidly converge to
the other metastable state x+ along such a deterministic path. (The contribution of
relaxation trajectory to the mean escape time is usually neglected.) Another solution
for is (4.75), which can be interpreted as the action along a non-deterministic
path that represents the most probable path of escape from x to x0 [160, 164, 175,
176]. In Fig. 4.5 we illustrate the Hamiltonian phase space for the model showing
the constant energy solutions of the Hamiltonian given by (4.78); the zero-energy
activation and relaxation trajectories through the fixed points of the deterministic
system are highlighted as thicker curves
Given the quasistationary approximation, the rate of escape from the metastable
state centered about x = x can be calculated by matching it with an appropriate
inner solution in a neighborhood of the point x = x0 [160, 164, 175, 256, 277]. This
is necessary since the quasistationary solution (4.76) does not satisfy the absorbing
164
boundary condition at the point x0 separating the two metastable states. There are a
number of different ways of carrying out the matched asymptotics; see, for example, [277]. Here we will follow an approach based on fixing the probability flux J0
through x0 and then matching the activation solution for x < x0 with the relaxation
solution for x > x0 using a diffusion approximation of the full master equation (4.60)
in the vicinity of x0 [164, 175, 256]. The latter yields the FokkerPlanck equation
(1.111), which can be rewritten in the form of a conservation equation
P(x,t) = J(x,t)
t
x
(4.84)
with
J(x,t) = (+ (x) (x))P(x,t)
1
[(+ (x) + (x))P(x,t)] .
2N x
Substituting the quasistationary solution p(x,t) = C0 e0t (x) into (4.84) and using
the fact that 0 is exponentially small give
J0 = (+ (x) (x)) (x)
1
[(+ (x) + (x)) (x)] ,
2N x
where J0 is the constant flux through x0 . In a neighborhood of x0 , this equation can

be Taylor expanded to leading order in x x0 and integrated to obtain the solution
(x) =
J0 N (xx0 )2 / 2
e
+ (x0 )

where
e(yx0 )
2 / 2
dy,
(4.85)
2+ (x0 )
N[+ (x0 ) (x0 )]
(4.86)
determines the size of the boundary layer around x0 .

In order to match the activation and relaxation solutions, the following asymptotic behavior of the inner solution (4.85) is used:
NJ0 2
, x x0
(x x0 )+ (x0 )
(4.87)
(x) =
2
2
NJ
e(xx0 ) / , x0 x .
+ (x0 )
The solution to the right of the saddle matches the relaxation solution (4.77) since
+ (x) (x) (x x0 )[+
(x0 )
(x0 )] for x x0 such that B = J0 . In order
to match the solution on the left-hand side of x0 with the activation solution (4.76),
Taylor expand (x) about x0 using (x0 ) = 0 and (x0 ) = 2/N 2 . It follows that

| (x0 )| N (x0 )
A+ (x0 )
J0 =
e
.
(4.88)
2 N
+ (x0 ) (x0 )
165
The final step in the analysis is to link the flux J0 with the escape rate 0 . This is
achieved by substituting the quasistationary solution into the continuity equation
(4.84) and integrating over the interval x [0, x0 ] with a reflecting boundary condition at x = 0:

1 x0
1
=
(y)dy.
(4.89)
0 J0 0
Since the activation solution is strongly peaked around the fixed point x , a Gaussian
approximation of (x) around x yields the final result
0 =
+ (x )
2
| (x0 )| (x )eN[ (x0 ) (x )] .
(4.90)
Hence, we obtain (4.63) with i = 01 and

(x) +
(x)
(x)
d
ln
= (x).

(x) =
=
dx
+ (x)
(x) + (x)
Similarly, we can obtain (4.64) for the mean time f to terminate a spark.
4.4.4 Stochastic Phase-Plane Analysis

In the above analysis of Ca2+ sparks, the concentration csr in the JSR was held fixed.
This is a reasonable approximation when considering the initiation of a Ca2+ spark.
However, following Ca2+ release from the RyRs, the Ca2+ concentration csr slowly
changes according to
sr
dcsr
= csr x + ksr [c0 csr ].
dt
(4.91)
where sr o D . The first term on the right-hand side is the loss of Ca2+
through the RyRs, whereas the second terms are the influx JSR of Ca2+ from the
NSR with fixed Ca2+ concentration c0 ; see Fig. 4.4. The variation of csr means that
one has to modify the analysis of the time to terminate the Ca2+ spark. Following
Hinch [276], this can be achieved by combining the theory of stochastic transitions
outlined in Sect. 4.4.2 with the classical phase-plane analysis of slowfast excitable
systems such as the FitzHughNagumo equations (see Sect. 2.1). That is, (4.62) and
(4.91) form an excitable system with the fraction x of open RyRs acting as the fast
variable and csr acting as the slow variable. In Fig. 4.6 we sketch the nullclines of
the deterministic system in a parameter regime where there is a single, stable fixed
point (x , csr ). In the full stochastic model, the initiation of a Ca2+ spark induces a
transition to the right-hand x-nullcline according to x (csr ) x+ (csr ) as outlined
in Sect. 4.4.2. The slow variable then moves down the right-hand nullcline x+ (csr )
according to the equation
sr
dcsr
= csr x+ (csr ) + ksr [c0 csr ].
dt
(4.92)
166

4
x-
x0
x+
nullcline of x
fluctuations
about nullcline
csr
nullcline of csr
0
0
x
2+
Fig. 4.6 Sketch of nullclines in the deterministic planar Ca spark model with x denoting the
fraction of open RyRs and cSR the Ca2+ concentration in the junctional SR. The csr nullcline is a
monotonically decreasing function x(csr ), whereas the x nullcline is cubic-like with three branches
x (csr ) and x0 (csr ). (Note that the branch x (csr ) 0; we have moved it away from the vertical
axis for the sake of illustration.) In the given diagram there is a single, stable fixed point on the lefthand branch. In the stochastic version of the model a Ca2+ spark initiates a jump to the right-hand
branch x+ (csr ) along the lines outlined in Sect. 4.4.2. This is followed by a stochastic trajectory
in which the slow variable csr (t) moves down the nullcline until it undergoes a noise-induced
transition back to the left-hand branch before the knee at x = xc . In the deterministic case, the
return transition occurs at the knee (dashed curve)
That is, although x is a stochastic variable, it fluctuates much faster than the dynamics of csr so one can substitute a time-averaged value of x in (4.91).
Suppose that csr (t) is the solution of (4.92) with csr (0) = csr , that is, the Ca2+
spark occurs at t = 0. In principle, the spark can terminate at any time t > 0 due to
fluctuations in the number of open RyRs. Again using a separation of time scales,
we can estimate the rate of transition back to the left-hand branch at time t using

+ (x+ (t)) | (x0 (t))| (x+ (t)) N[ (x+ (t)) (x0 (t))]
(t) =
e
,
(4.93)
2
where x+ (t) = x+ (csr (t)), etc. One can now calculate the distribution of spark durations T . Let P( ) = P(T > ) and introduce the spark duration probability density
p( ) =
dP
.
d
The probability that a spark terminates in an infinitesimal time interval is

( ) , so that
P( + ) = P( )(1 ( ) ).
167

Taking the limit 0 and integrating gives P( ) = exp 0 (t)dt , and hence

p( ) = ( ) exp
(t)dt .
(4.94)
0
4.4.5 Whole-Cell Model of Ca2+ Sparks

One of the major simplifications of the Hinch model [276] is that the background
Ca2+ concentrations in the mytoplasm (cm ) and the NSR (cnsr ) are held fixed. It thus
fails to capture the collective behavior of a large population of Ca2+ release units
(CaRUs) that are coupled via global changes in these background concentrations
(assuming diffusion can be neglected on the relevant time scales). This has motivated the development of a whole-cell model of calcium-induced calcium release
in cardiac myocytes, based on a system of N globally coupled CaRUs [672, 673].
We will describe a simplified version of the model in which Ca2+ pumps, leakage
currents, and voltage-gated Ca2+ channels are not modeled explicitly. Let c j and
c j denote the Ca2+ concentration in the dyadic space and JSR of the jth CaRU,
j = 1, . . . , N . Then
dc j
j
j
= JRyR JD ,
dt
sr
d c j
j
= Jsrj JRyR ,
dt
(4.95)
where JRyR is the total Ca2+ current through the RyRs, JD is the diffusive current
from the diadic space of the jth CaRU to the bulk mytoplasm, and Jsrj is the current
from the NSR to the JSR of the jth CaRU. The various fluxes are given by
JDj = c j cm ,
j
JRyR
= x j c j ,
Jsrj = ksr [cnsr c j ],
(4.96)
where x j is the fraction of open RyRs in the jth CaRU. Finally, from conservation
of Ca2+ ions, the bulk concentrations cm and cnsr evolve according to
N
dcm
= JDj + Jm ,
dt
j=1
sr
N
dcnsr
= Jnsr Jsrj .
dt
j=1
(4.97)
Here Jm and Jnsr are the total external currents into the mytoplasm and NSR, respectively, arising from Ca2+ pumps and other factors.
When the number N of CaRUs is large, one can develop a probability density version of the above model, which takes the form of a differential Chapman
Kolmogorov (CK) equation describing the evolution of a single equivalent stochastic CaRU [672]. The basic idea is to introduce a probability density p(c, c,
n,t),
with p(c, c,
n,t)dcd c interpreted as the fraction of CaRUs at time t that have Ca2+
concentrations in the range c c(t) c + dc, c c(t)
c + d c and are in the internal state n, with the latter determined by the number of open RyRs, for example.
Introduce the single CaRU equations
168
n
dc
F(c, c,
n) = c + cm c
dt
N
n
d c
c,
n) = ksr [cnsr c]
sr F(
c.
dt
N
(4.98)
The probability density p then evolves as
p
(F p) (F p)
=
+ A(n, n
; c, c)p(c,
c,
n
,t),
t
c
c
n
(4.99)
with A(n, n
, c, c)
constructed from the transition rates of the master equation describing the opening and closing of RyRs within each CaRU. Note that the CK
equation (4.99) is formally similar to the CK equation (1.143) for membrane voltage fluctuations considered; see Sects. 1.5 and 1.6. However, the solution of (1.126)
represents the distribution of state trajectories with respect to repeated trails of a
single neuron, whereas the solution of (4.99) represents the distribution of state trajectories with respect to a large ensemble of identical CaRUs. Moreover, the latter
are globally coupled due to the dependence of F, F and A on the Ca2+ concentrations
cm , cnsr . Within the probabilistic framework these evolve according to the equations

dcm
=
(c, c,t)[c
m c]dcd c + Jm
dt
0
0

dcnsr
= Jnsr ksr
sr
(c, c,t)[c
c,
nsr c]dcd
dt
0
0
(4.100)
(4.101)
with (c, c,t)

= n p(c, c,
n,t). The probability density formulation of CaRUs can
increase computational efficiency compared to Monte Carlo simulations of the full
ODE model [672, 673].
4.5 Intercellular Calcium Waves in Astrocytes

So far we have considered intracellular Ca2+ waves that propagate within a single cell. However, one also finds intercellular Ca2+ waves (ICWs) consisting of
increases in cytoplasmic Ca2+ concentration that are communicated between cells
and appear as waves that spread out from an initiating or trigger cell. The speed and
size of ICWs depend on the initiating stimulus and the particular propagation mechanism. An ICW often propagates at a speed of 1020 m/s and lasts for periods of up
to tens of seconds, indicating that it can involve the recruitment of hundreds of contiguous cells. ICWs were first observed in cultures of astrocytes (discussed below)
in response to extracellular glutamate [143], and in airway epithelial cells following
mechanical stimulation [550]. They have subsequently been found in a wide range
of cell types and under a variety of stimulus conditions, as reviewed in [380, 554].
169
4.5.1 Astrocytes and the Tripartite Synapse

Within the context of the central nervous system (CNS), ICWs are emerging as a
major long-range signaling mechanism of a particular type of glial (nonneuronal)
cell known as an astrocyte; see Fig. 4.7. Other types of glial cells include oligodendrocytes and radial glial cells. The former are responsible for myelination in the
CNS just as Schwann cells are responsible for myelination in the peripheral nervous system; see also Sect. 2.5. On the other hand, radial glial cells play a pivotal
role during development, with regard to both neuronal migration and neurogenesis [101]; see Sect. 9.1. Radial glial cells usually disappear post-development, but
some cells, such as Bergmann glia in the cerebellum and Muller glia in the retina,
maintain their radial glia-like properties into adulthood.
Traditionally, astrocytes were thought to be physiologically passive cells that
only play a supporting role in the CNS by regulating and optimizing the environment within which neurons operate. Their housekeeping duties include maintaining
local ion and pH homeostasis, delivering glucose and providing metabolic support
via contacts with the vasculature, and clearing neuronal waste such as metabolic
products and excess neurotransmitters in the synaptic cleft. However, over the past
20 years there has been an increasing amount of empirical data indicating that astrocytes play an active role in modulating synaptic transmission [253, 268, 449].
That is, astrocytes respond to synaptic neurotransmitters such as glutamate by elevating their intracellular Ca2+ levels. This then results in the release of glutamate
and other signaling molecules from the astrocytes (gliotransmission), which in turn
influences synaptic activity. Thus, astrocytes could potentially contribute to higher
brain function and, consequently, behavior. This is consistent with the fact that the
number of astrocytes relative to the number of neurons increases dramatically with
Fig. 4.7 Left: Astrocyte in vitro stained with GFAP to show filaments. Right: Astrocytes stained for
GFAP, with end-feet ensheathing blood vessels [Public domain figure from Wikipedia Commons]
170
brain complexity, ranging from 1:25 in the leech, 1:6 in C elegans, 1:3 in lower
mammals, to 1.4:1 in the human cortex. Conversely, it is known that following injury to the nervous system or under pathological conditions such as Alzheimers
disease, epilepsy, and schizophrenia, the structure and protein expression of astrocytes are altered [564]. One interpretation of these structural changes is that they are
indicative of the supportive action of astrocytes following injury. On the other hand,
given that many CNS disorders are correlated with alterations in synaptic transmission and gliotransmission regulates synaptic processing, it is possible that changes
in gliotransmission contribute to these disorders; see Chap. 9.
In many higher organisms, astrocytes establish nonoverlapping anatomical and
functional domains of size around 100 m, and their distribution is highly organized.
neuron
astrocyte
blood
vessel
astrocyte
AMPAR
glutamate
A
B
C
Ca2+
glutamate
tripartite
synapse
Fig. 4.8 Astrocytes modulate synaptic transmission and blood flow via Ca2+ signaling. Inset: the
tripartite synapse. Synaptic transmission of glutamate (A) activates not only postsynaptic receptors such as AMPA but also astrocytic receptors that elevate astrocytic Ca2+ levels (B). This then
induces the local release of glutamate from the astrocyte (C), which can modulate the synapse. Astrocytes also form a link between a neuron and the local vasculature. Increases in astrocytic Ca2+
concentration propagate via astrocyte end feet to small blood vessels, where they control vessel
diameter and blood flow
171
The resulting astrocytic network parcellates the whole neutrophil such that a single
astrocyte can make contact with thousands of synapses. The processes of a single
astrocyte make contact with a synapse by wrapping around both the presynaptic and
postsynaptic domains to form what is known as a tripartite synapse [253, 268, 449];
see Fig. 4.8. Hence, from a geometrical perspective, it is well placed to modulate synaptic transmission. The first component of glialneuronal interactions at a
synapse, namely, the metabotropic glutamate receptor 5 (mGluR5)-dependent elevation in astrocytic Ca2+ concentration in response to the release of glutamate neurotransmitters, was discovered in the early 1990s using Ca2+ imaging to study cultured
glial cells [106, 143]. Both intracellular Ca2+ oscillations and ICWs propagating between neighboring astrocytes were observed. Subsequent studies have established
that neuroglial signaling also occurs in astrocytes of isolated retina and brain slices
[265, 460]. (Note, however, that a recent study of adult rodents has called into question the link between neuroglial Ca2+ signaling at the tripartite synapse [610]. It
would appear that the astrocytic expression of mGluR5 receptors is developmentally regulated and, in particular, is undetectable in mature rodent brains; almost all
previous studies of the tripartite synapse have been in immature animals.) Following the observation that neuronal activity increases Ca2+ levels in astrocytes, it was
subsequently discovered that the resulting Ca2+ elevation induces gliotransmission
of glutamate and a variety of other signaling molecules including ATP [249, 491].
Gliotransmission has been demonstrated to modulate synaptic transmission presynaptically and postsynaptically. For example, glutamate can enhance neurotransmitter release presynaptically by binding to metabotropic (mGlu) receptors or kainate
receptors [189, 388]. Another example of enhanced synaptic transmission is the
binding of ATP to postsynaptic receptors, which induces elevations in Ca2+ . This,
in turn, can drive the insertion of additional AMPA receptors, thus temporarily increasing synaptic efficacy [236]. (The binding of ATP to postsynaptic receptors also
provides a mechanism for the propagation of ICWs via paracrine signaling; see below.) On the other hand, hydrolysis of ATP results in the accumulation of adenosine,
causing a reduction of synaptic transmission [493].
In addition to directly modulating synaptic transmission, astrocytes appear to regulate NMDA-receptor function. Recall from Sect. 1.3 that NMDA receptors are crucial for the induction of long-term synaptic plasticity, requiring both glutamate and
postsynaptic depolarization in order to activate. The opening of an NMDA channel
leads to a temporary increase of intracellular Ca2+ within the postsynaptic domain,
whose amplitude and temporal profile determine whether the synapse is potentiated
or depressed. It turns out that there is an additional glycine-binding site that regulates NMDA-receptor activity, whose co-agonist D-serine is released by astrocytes
during gliotransmission. Increases in the level of D-serine appear to shift the balance point of long-term plasticity from depression to potentiation [485], a process
known as metaplasticity. Finally, not only does astrocytic Ca2+ signaling modulate
synaptic transmission and plasticity, it also controls blood flow, since intracellular
waves propagate to small blood vessels via astrocyte end feet, where they control
vessel diameter; see Fig. 4.8.
172
Following the original studies of ICWs in astrocyte cultures [106, 143], intracellular Ca2+ wave propagation between astrocytes has been observed in vitro in
brain slices and whole retina preparations [265, 460, 492]. These empirical findings,
combined with the structure of the tripartite synapse, have led to the suggestion that
networks of astrocytes could form an extraneuronal pathway for rapid long-distance
signaling within the CNS, which interacts reciprocally with neuronal networks. In
particular, ICWs could provide a potential mechanism for coordinating and synchronizing the activity of a large group of neuronal and nonneuronal cells. It should be
pointed out, however, that the function of ICWs in astrocytes remains controversial,
since there have been very few in vivo observations of ICWs in the CNS, excluding
the case of spontaneous ICWs during early development where astrocytes play a
role in the generation, differentiation, and migration of neurons [666]; see Sect. 9.1.
One recent exception is a study demonstrating in vivo spontaneous ICWs (termed
glissandi) in mouse hippocampal astrocytes. (ICWs have also been observed in vivo
in other glial networks, e.g., Bergmann glia in the cerebellum of rodents [287].) Another concern regarding the functional role of ICWs, either in vitro or in vivo, is
whether they require strong inducing stimuli that do not have a normal physiological counterpart. On the other hand, certain pathological conditions such as brain
trauma, brain ischemia (oxygen deprivation), and seizures are often associated with
strong stimuli. They are all coupled with the release of ATP and glutamate and reductions in the level of extracellular Ca2+ that collectively could promote ICW generation. Such ICWs could exacerbate an injury by propagating signals that initiate
local
stimulation
ATP
hemichannels
G
endoplasmic
reticulum
ATP
PLC
IP3
PLC
IP3
Ca2+
Ca2+
IP3 R
Fig. 4.9 Two mechanisms for the propagation of ICWs. Local stimulation of a single cell induces
an elevation in intracellular IP3 that generates an intracellular Ca2+ wave. Diffusion of IP3 through
gap-junction coupling with an adjacent cell initiates a second intracellular Ca2+ wave. In addition, or alternatively, stimulation of the first cell releases ATP via hemichannels in the plasma
membrane or vesicular release; this could also be dependent on intracellular Ca2+ . The extracellular diffusion of ATP to neighboring cells can activate metabotropic receptors that then stimulate
the production of IP3 and the production of downstream intracellular Ca2+ waves
173
cell apoptosis (cell death) in surrounding cells [681]. There is also growing evidence
that ICWs provide a source of excitation during epileptic episodes [564, 647, 668];
see Sect. 9.4. For example, studies of a brain slice that exhibits epileptiform activity, showed that there was also a corresponding increase in the frequency of Ca2+
oscillations in astrocytes [623]. Moreover, the release of glutamate from astrocytes
has been observed to cause synchronous neuronal depolarizations [186].
4.5.2 Mechanisms and Models of Intercellular Calcium Waves

There are thought to be two basic mechanisms for the propagation of ICWs between
adjacent cells that occur in astrocytes and other cell types: (a) diffusion of the second messenger IP3 and, to a lesser extent, Ca2+ via gap junctions; (b) paracrine
signaling via a diffusing extracellular messenger such as ATP. These mechanisms
are illustrated in Fig. 4.9.
Gap-Junction Signaling
Gap junctions are arrays of transmembrane channels that connect the cytoplasm of
two neighboring cells and thus provide a direct diffusion pathway between the cells.
Cells sharing a gap-junction channel each provides a hemichannel (also known as
a connexon) that connect head-to-head [545]. Each hemichannel is composed of
proteins called connexins that exist as various isoforms named Cx23 through Cx62,
with Cx43 being the most common. The physiological properties of a gap junction,
including its permeability and gating characteristics, are determined by the particular connexins forming the channel. Although gap junctions are readily permeable
to Ca2+ and IP3 , the diffusion of Ca2+ through a channel is greatly restricted due
to buffering, so that the propagation of ICWs is primarily mediated by the diffusion of IP3 when gap junctions are the main mechanism. The possible role of gap
junctions in the propagation of ICWs was initially suggested by the observation in
epithelial cells that waves only spread in adjacent cells from localized membrane
sites in direct contact with the initiating cell. Moreover, the ICWs were not influenced by fluid flow over the apical surface of cells, that would have interfered with
extracellular signals [259]. Further evidence for the role of gap junctions came from
the study of glioma cells that lacked gap junctions [107]. These cells only supported
ICWs when they were transfected with connexins; however, the presence of connexins could also enhance ATP and thus contribute to paracrine signaling (see below).
Subsequent experimental and modeling studies have established that the diffusion
of IP3 through gap junctions can support the propagation of ICWs [379, 591]. In
particular, ICWs could be initiated by an elevation of IP3 in a single cell followed
by diffusion to other cells without any regeneration of IP3. On the other hand, the
range of the waves tends to be limited to only a few cells, suggesting that some other
mechanism may be needed in order to generate longer-range waves.
174
The first model of ICWs mediated by gap junctions was based on studies of
mechanical stimulation of cultured epithelial cells [590, 591]. It has subsequently
been extended to models of ICWs in networks of astrocytes of varying complexity
[163, 281, 310]. The models all have the same basic structure, but differ in terms of
whether or not regenerative mechanisms or extracellular signaling mechanisms are
included. A simple 1D version of an ICW model is shown in Fig. 4.10. Each cell is
treated as a square of size L with local coordinates (x, y), 0 x, y L. There are N
cells labeled n = 1, . . . , N arranged on a regular lattice with nearest neighbor coupling. The intracellular dynamics within each cell is described by reactiondiffusion
equations of the form given in Sect. 4.2. For example, in the original model of Sneyd
et al. [591], this included equations for cytoplasmic Ca2+ concentration (cn ), IP3
concentration pn , and a slow variable hn representing the fraction of IP3 receptors
that have not been inactivated by Ca2+ . An example of an intracellular model for
the nth cell is (see, e.g., [280])
pn
= D p 2 pn k p pn ,
t
cn
= Dc 2 pn kc cn + f (cn , hn , pn ),
t
dhn
= h (cn ) hn.
h
dt
(4.102a)
(4.102b)
(4.102c)
L
cn, pn, hn
IP3
input
Jn
x
n-1
n
2+
Fig. 4.10 Schematic diagram of a 1D model of intercellular Ca
n+1
waves mediated by gap junctions
Here the nonlinear function f represents the net flux due to Ca2+ release from IP3
receptors. It is usually assumed that D p Dc due to Ca2+ buffering. Coupling
between adjacent cells in incorporated into the model using the boundary conditions
for IP3 . Thus along the common vertical edge between cells n 1 and n, 1 < n < N,
the IP3 flux Jn is taken to be of the form

pn1 (x, y)
Dp

x
x=L

pn (x, y)
= Dp
x
K [pn1 (L, y) pn (0, y)]. (4.103)

x=0
These are supplemented by no-flux boundary conditions for IP3 across all other
edges; all cell borders are assumed to be impermeable to Ca2+ . In a 2D version of
the model arranged on a square grid, IP3 can also flow across common horizontal
175
edges of neighboring cells, while no-flux boundary conditions are imposed along the
exterior boundary of the grid. A typical simulation of such a model is to inject IP3
in the cell at one end of the lattice, which is then allowed to diffuse from cell to cell,
resulting in the propagation of an intracellular calcium wave. It is important to note
that the ICW is not a traveling wave in the mathematical sense used throughout this
book. Rather, the spread of Ca2+ is mediated by passive diffusion of intracellular
Ca2+ and intra-/intercellular IP3 . Thus the rate of spread decreases with distance
from the stimulated cell, and the arrival time of the wave increases exponentially
with distance. In order to create more persistent ICWs, it is necessary to incorporate
some regenerative mechanism. One example of such a mechanism is to assume that
the rate of production of IP3 can be amplified by intracellular Ca2+ so that (4.102a)
becomes [281]
pn
c2n
= D p 2 pn +
k p pn .
t
K + c2n
Hofer et al. [281] showed that for an intermediate range of values of , partial regeneration of IP3 supports ICWs that propagate further than those relying solely
on passive diffusion of IP3 . Finally, an alternative mechanism for increasing the
range of ICWs, which does not require some form of regeneration or active process,
is to combine gap-junction coupling with an extracellular pathway involving ATP
[163, 310]; see below.
Paracrine Signaling
Early evidence for the involvement of an extracellular component to ICWs arose
from the observation that ICWs can propagate between disconnected groups of astrocytes, provided that the degree of separation does not exceed around 100 m
[266]. More recently, advanced imaging techniques have revealed that ICWs are
associated with a cloud of extracellular ATP [249, 457]. There are a variety of
pathways that could mediate ATP release, including hemichannels in the plasma
membrane and vesicular discharge. Modeling studies suggest that a combination of
gap-junction and paracrine signaling can significantly increase the range of ICWs
[298]. It is also possible that there is a regenerative mechanism that boosts the propagation of ICWs. One candidate for active rather than passive (diffusive) wave propagation is Ca2+ -dependent ATP release via connexin hemichannels [652]. In the
case of astrocytes, the dominant propagation mechanism depends on where they
are located within the nervous system. For example, ICWs in the neocortex tend to
be mediated by gap junctions, whereas in the hippocampus, both gap-junction and
paracrine signaling play a role.
Here we will describe a mathematical model of paracrine signaling in astrocyte
networks due to Bennett et al. [37]. Suppose that a(t) denotes the extracellular concentration of ATP in the vicinity of a single cell. ATP binds to metabotropic P2Y
receptors in the cell membrane according to the first-order kinetic scheme
176

k+
AT P + R R.
k
In the case of fast binding kinetics, the fraction of P2Y receptors that are bound by
ATP is given by the quasi-equilibrium solution
r=
a
,
KR + a
KR =
k
.
k+
(4.104)
This then drives activation of G-proteins according to the kinetic scheme

G0
ka (r+r0 )

kd
G1 ,
where G0,1 represent the inactive and active states of the G-protein. Denoting the
amount of activated protein by g and the total G-protein by gT , we have
dg
= ka (r + r0 )(gT g) kd g.
dt
(4.105)
Again assuming fast kinetics, the quasi-equilibrium solution is

g = gT
r + r0
,
KG + r + r0
KG =
kd
.
ka
(4.106)
A major simplification of the model is to assume that the release of IP3 within the
cell is mainly driven by activation of G-protein so that Ca2+ dependence is ignored.
If p is the intracellular IP3 concentration then
dp
= g kdeg p,
dt
(4.107)
for constants , kdeg . The next stage of paracrine signaling, namely, the release of
ATP by astrocytes is not yet well understood. However, IP3 as a triggering mechanism, possibly via Ca2+ release from intracellular stores, is probably involved. In
the model of Bennett et al. [37], IP3 is assumed to directly cause ATP release from
the cell surface into the extracellular space at the rate

p pmin
JATP (p,t) = ATP (t)
.
(4.108)
Krel + p 0
Here ATP and Krel are constants, [z]0 = z for z > 0 and [z]0 = 0 for z < 0, and (t)
takes into account depletion of ATP stores within the cell. That is (0) = 1 and
d
= JATP (t).
dt
(4.109)
The rate of ATP release is thresholded in order to insure that very small levels of
ATP are not amplified, and the depletion factor (t) is included in order to terminate
177
ATP release; the latter is motivated by the observation that ATP release is suppressed
by repeated stimulation of a cell.
Finally, coupling between cells is modeled in terms of a diffusion equation for the
extracellular ATP concentration that takes into account the IP3 -mediated release of
ATP from the astrocytes. One way to proceed would be to use a continuum bidomain
model, in which astrocytes and extracellular ATP coexist at every point in space.
Taking a = a(x,t), p = p(x,t) and denoting the density of astrocytes by (x), we
would then have
a
a
= DATP 2 a + (x)JATP(p,t)
.
t
Kdeg + a
(4.110)
Alternatively, one could explicitly distinguish between extracellular space and the
intracellular space of astrocytes, in which case the fluxes JATP only exist at the
boundaries of the cells. The latter approach is used in the computer simulations of
Bennett et al. [37], who take a simplified cubic geometry for the astrocytes and also
allow for intracellular diffusion of IP3 . Note that under a quasi-steady-state approximation, (4.107) and (4.110) form a closed system of differential equations, since
(4.104) and (4.107) can be used to express g in terms of a. Hence, one can solve for
the ATP and IP3 concentrations independently of Ca2+ and then incorporate ICWs
by using the IP3 to drive a standard model of Ca2+ release from intracellular stores
such as LiRinzel. However, if there is regenerative Ca2+ -dependent production of
IP3 or a Ca2+ -dependent mechanism for gliotransmission of ATP, then it is necessary to couple the Ca2+ dynamics to that of IP3 and ATP.
Fig. 4.11 Illustration of a 2D periodic heterogeneous medium consisting of a cytosolic domain c

and an ER domain e separated by a boundary
178
4.6 Appendix: Homogenization

As we noted in Sect. 4.2, it is not practical to try modeling the detailed microstructure of the cytoplasmic and ER boundaries, which are highly interleaved. This motivates consideration of a mean-field description of the Ca2+ concentrations, in which
the concentrations c, ce are taken to coexist at every point in space with effective diffusion coefficients. As shown by Goel et al. [224], the resulting bidomain model can
be derived from first principles using homogenization theory. In this appendix we
sketch the basic steps in their derivation. Note that we also encountered a 1D example of homogenization theory in Sect. 3.2.3, where we spatially averaged over a
discrete distribution of dendritic spines.
Consider a domain R3 in which the ER is treated as a periodic network that
occupies a fraction of , as illustrated in Fig. 4.11. For simplicity, assume that the
unit cell is a cube of length . The ER Ca2+ occupies a connected domain e , and
the complementary connected domain c = /e is occupied by cytosolic Ca2+ .
The concentrations c and ce satisfy the 3D diffusion equation in their respective
domains:
c
= (A (x/ )c(x)), x c
t
ce
= (B (x/ )ce (x)), x e ,
t
(4.111a)
(4.111b)
where A and B are the diffusion tensors for Ca2+ in the cytosol and ER, respectively. The boundary conditions on the membrane separating the cytosol from the
ER take the form
A (x/ )c(x) nc = f (c , ce ),
B (x/ )ce (x) nc
= f (c
, ce ),
on
(4.112a)
on ,
(4.112b)
where nc , ne denote the unit exterior normals to the boundaries c and e , respectively, satisfying nc = ne on , and f determines the net flux from the ER to
the cytosol.
The basic idea of the homogenization method is to supplement the macroscopic spatial variables x = (x1 , x2 , x3 ) with a set of microscopic variables y =
(y1 , y2 , y3 ), 0 yi 1, that are the coordinates of a unit cube [496, 630]. Denote
by c (e ) the set of points y = x/ in the unit cube for which x c (x e ).
The concentrations c , ce are then taken to be functions of x and y with x , y c
for c , and x , y e for ce :
c = c(x, y,t),
ce = ce (x, y,t)
(4.113)
with y = x/ . The next step is to introduce formal asymptotic expansion for c and
ce of the form
c = c0 (x, y,t) + c1 (x, y,t) + 2 c2 (x, y,t) + . . .
(4.114a)
ce
(4.114b)
c0e (x, y,t) + c1e (x, y,t) + 2c2e (x, y,t) + . . .,
179
where ck and cke are 1-periodic functions of y. Setting

u = x u(x, y) + 1y u(x, y),
it follows that (4.111) becomes

c 2
= y (A (y)y ) + 1[x (A (y)y ) + y (A (y)x )] + A (y)2x c
t
for x , y c
(4.115a)

ce 2
= y (B (y)y ) + 1[x (B (y)y ) + y (B (y)x )] + A (y)2x ce
t
for x , y e ,
(4.115b)
and the boundary conditions (4.112) become

A (y) x + 1 y c nc = f (c , ce ),

B (y) x + 1 y ce nc = f (c , ce ),
for x , y
(4.116a)
for x , y .
(4.116b)
Substituting the asymptotic expansions for c and ce and collecting terms of the
same order in then yields a hierarchy of equations for ck , cke . At O( 2 ), we have
y (A (y)y )c0 = 0,
y c
(4.117a)
y (B (y)y )c0e
y e .
(4.117b)
= 0,

y ,
A (y)y c n = 0 = B (y)y c0e n,
(4.118)
where n is a unit normal on . For fixed x, the only periodic solution to these equations is c0 = constant, c0e = constant, that is,
c0 = c0 (x,t),
c0e = c0e (x,t).
The O( 1 ) equations are

y A (y) x c0 + y c1 = 0,

y B (y) x c0e + y c1e = 0,
(4.119)
y c
(4.120a)
y e ,
(4.120b)

A (y)[x c0 + y c1 ] n = 0 = B (y)[x c0e + y c1e ] n,
y .
(4.121)
The solutions for c1 and c1e can be written in the form

3
c1 (x, y,t) = Vic (y)

i=1
c0 (x,t)
+ c1 (x,t),
xi
(4.122a)
180

3
c1e (x, y,t) = Vie (y)

i=1
c0e (x,t)
+ c1e (x,t),
xi
(4.122b)
with the 1-periodic vectors Vc and Ve satisfying the cell equations

y (A (y) [yVkc (y) + ek ]) = 0,
y (B (y) [yVke (y) + ek ]) = 0,
y c
y e ,
(4.123a)
(4.123b)
and
y ,
A (y)[yVkc (y) + ek ] n = 0 = B (y)[yVke (y) + ek ] n,
(4.124)
for k = 1, 2, 3, where ek is the vector with components e jk = jk .

In order to derive mean-field equations for the homogenized concentrations
c0 , c0e , it is necessary to proceed to O(1), for which

c0
,
x A (y)(x c0 + y c1 ) + y A (y)(x c1 + y c2 ) =
t
y c

x B (y)(x c0e + y c1e ) + y B (y)(x c1e + y c2e ) =
y e
c0e
t
(4.125a)
(4.125b)
with boundary conditions

A (y)[x c1 + y c2 ] n = f (c0 , c0e ),
B (y)[x c + y c ] n = f (c
1
, c0e ),
(4.126a)
y .
(4.126b)
Integrating (4.125a) over c gives

c0
,
x A (y)(x c0 + y c1 ) + y A (y)(x c1 + y c2 ) dy = |c |
t
c
(4.127)
where |c | is the volume fraction of the unit cell occupied by the cytosol. Using
(4.122a), the first term on the left-hand side becomes

x A (y)(x c0 + y c1 ) dy
c

=
x A (y)(x c + y
0
[A ]ik =
[A (y)]i j
c j
i=1

= x A x c0 ,
where
c
Vic (y)

xi
Vkc (y)
+ j,k dy.
yj
dy
(4.128)
181
Finally, applying the divergence theorem to the second term on the left-hand side of
(4.127),

c

y A (y)(x c1 + y c2 ) dy = A (y)(x c1 + y c2 ) nd (y)
f (c0 , c0e )d (y)
= | | f (c0 , c0e ).
Combining all of the results, we obtain a macroscopic equation for c0 given by
|c |

c0
= x A x c0 + | | f (c0 , c0e ).
t
(4.129)
Proceeding in an identical fashion, we also obtain a macroscopic equation for c0e :

|e |

c0e
= x B x c0e | | f (c0 , c0e ),
t
(4.130)
where B is defined according to (4.128) with A replaced by B on the right-hand

side. Equations (4.129) and (4.130) constitute a bidomain model with homogenized
diffusion tensors A and B . The latter can be determined by solving the cell equations (4.123) and (4.124) for Vc and Ve . Goel et al. [224] calculated effective diffusion tensors for several different ER microstructures. They found that diffusion in
the cytoplasm could be halved as the ER volume fraction increased from 0 to 0.9.
Part II
Networks
Chapter 5
Waves in Synaptically Coupled Spiking

Networks
In the second part of this book, we consider a variety of neural network models
that have been used to investigate the mechanisms and properties of wave propagation in the cortex and other parts of the nervous system. There has been a rapid
increase in the number of computational studies of network spatiotemporal dynamics, including oscillations and waves, which are based on biophysically detailed
conductance-based models of synaptically (and possibly electrically) coupled neurons; see, for example, [153]. These models provide considerable insights into the
role of ionic currents, synaptic processing, and network structure on spatiotemporal dynamics, but they tend to be analytically intractable. This has motivated an
alternative approach to network dynamics, involving simplified neuron models that
hopefully capture important aspects of wave phenomena, while allowing a more
concise mathematical treatment. In the case of oscillatory networks, such a simplification can be achieved by reducing a conductance-based neuron model to a phase
model along the lines of Sect. 1.2. Alternatively, one can use a simplified spiking
neuron model such as integrate-and-fire (Sect. 3.1) in order to investigate waves in
excitable and oscillatory neural media. Both of these approaches are considered
in this chapter. (Note that we do not discuss another important method for analyzing the dynamics of synaptically coupled spiking networks, namely, fast/slow
decomposition and geometric singular perturbation theory, which is reviewed extensively elsewhere [173, 537, 538].) In Chaps. 6 and 7 we turn to population-based
approaches, for which the fundamental active element is taken to be a local population of cells rather than a single neuron, and the cortex is treated as a continuum
neural field. The mathematical advantage of neural fields is that many of the PDE
methods from Part I can be adapted to study traveling wave solutions. An illustrative application of neural field theory to binocular rivalry waves and models of
visual cortex is presented in Chap. 8.
185
186
5 Waves in Synaptically Coupled Spiking Networks
5.1 Waves in Cortical and Subcortical Systems

We begin by summarizing the basic structure of the brain and then describe various
experimental studies of waves in different regions of the brain, focusing on cortical and subcortical structures. The brain can be divided into three main parts: the
forebrain, midbrain, and hindbrain; see Fig. 5.1. The forebrain consists of the cerebral cortex, thalamus, and limbic system, the midbrain consists of the tectum and
tegmentum, and the hindbrain is made of the cerebellum, pons, and medulla. Often
the midbrain, pons, and medulla are referred to collectively as the brainstem, which
is responsible for vital life functions such as breathing, blood pressure, and heart
pumping. The cerebral cortex is the largest part of the human brain and is responsible for higher brain function. The cortex is subdivided into four main sections
or lobes: (i) the frontal lobe, associated with planning, rational and emotional
thinking, and problem solving; (ii) the parietal lobe, associated with motor control
and somatosensory processing; (iii) the occipital lobe, associated with visual processing; (iv) the temporal lobe, associated with auditory processing, language, and
memory. Cortical folding increases the effective size of the cortex, which when unfolded forms a thin sheet of neural tissue only a few millimeters thick. Neurons
are arranged into six distinct layers through the depth of the cortex, each with its
own characteristic set of interneurons, input sources (afferents), and output targets
(efferents). For example, layer IV is the primary input layer of the sensory cortex,
receiving sensory information via the thalamus, superficial layers IIII mediate intracortical connections, while deep layers V and VI send feedback to subcortical
regions including the thalamus; see Fig. 5.2. For a more detailed description of the
Fig. 5.1 Schematic illustration of the brain, showing cortical, thalamic, and hippocampal regions.
See text for details [Public domain figure from Wikipedia Commons]
187
functional structure of primary visual cortex, see Sect. 8.1. The limbic system is a
complex set of brain structures, including the hippocampus and amygdala, which is
located on either side of the thalamus and adjacent to the inner border of the cortex. The limbic system appears to be primarily involved in memory and emotions.
For example, the hippocampus is required for converting short-term memories to
long-term memories and is also implicated in maintaining cognitive maps for spatial
navigation. The amygdala, on the other hand, is involved in processing emotionally
significant stimuli such as those related to fear and reward. Finally, moving on from
the limbic system, the cerebellum of the hindbrain is associated with the regulation
and coordination of fine movements, as well as posture and balance.
from other cortical

areas
to other cortical
areas
from thalamus
to thalamus
Fig. 5.2 Schematic illustration of the layered structure of primate sensory cortex. See text for
details [Public domain figure from Wikipedia Commons]
Cortical Waves
A common in vitro method for studying the propagation of electrical activity in the
cortex (or hippocampus) is to remove a vertical brain slice and bathe it in a pharmacological medium that blocks the effects of inhibition. This type of slice preparation
188
Fig. 5.3 Propagating wave of activity in a brain slice preparation in which inhibition has been
blocked. (a) Schematic of a slice removed from the somatosensory cortex of the rat. The cortex is a
thin (submillimeter) layered sheet of neurons. A slice approximately 500 m thickness is obtained
by cutting vertically through the cortical layers. (b) A multielectrode array placed into layers II/III
of the slice records extracellularly the propagation of a wave. (c) Pseudocolor plot of propagating
activity (Adapted from Pinto et al. [507])
was originally developed as an experimental model of epilepsy. Synchronized discharges can be evoked by a weak electrical stimulus to a local site on the slice and
each discharge propagates away from the stimulus as a continuous traveling pulse
at a characteristic speed of about 10 cm/s [112, 228, 507, 521, 685], as illustrated in
Fig. 5.3. The underlying mechanism for the propagation of such waves is synaptic
in origin rather than diffusive, in contrast to the much faster propagation of action
potentials along the axons of individual neurons; see Chap. 2. The pharmacological
dependence of propagating activity in vitro has been well characterized and implicates polysynaptic fast glutamatergic transmission (AMPA) as the major factor in
sustaining propagating activity. On the other hand, if a cortical slice is bathed in
normal cerebral fluid so that inhibition is present, then the speed of propagating activity is slowed by a factor of 10, and the propagating pulse of ensemble activity
has significantly more fluctuations consistent with the complex oscillation profile
seen in the corresponding local field potential [124, 549, 685]. This type of oscillatory wave consists of several oscillation cycles embedded in a single propagating wave envelope. The resulting activity is non-epileptic, with individual neurons
firing at a relatively low rate of around 10 Hz, and is mediated by slower NMDAbased excitation. It has been suggested that these waves are the in vitro analog of
189
low-frequency oscillations observed during slow-wave sleep [602]; see below. An

alternative type of oscillatory wave, which is often referred to as a one-cycle-onewave, has been observed in disinhibited cortical slices or in slices for which NMDA
receptor-mediated excitation has been enhanced. In this case, the cortical slice acts
like an oscillatory medium with a frequency around 10 Hz, and the traveling pulse
takes the form of a phase wave [684]; see Sect. 5.2.2. In two-dimensional (tangential) slices, these oscillations can develop into two-dimensional waves, including
spiral waves [292].
Fig. 5.4 Propagating wave of activity in the central olfactory lobe of the mollusk Limax. Successive images of the membrane potential in the olfactory lobe shown for one cycle. A band of
depolarization can be seen propagating from the distal to the proximal end of the lobe. The time
scale is 112 ms/frame and the length-scale bar is 100 m (Adapted from Kleinfeld et al. [338])
A variety of sensory stimuli have been linked to propagating waves in vivo [442,
553, 683]. For example, a number of studies of vertebrate and invertebrate olfactory bulbs have shown that they support propagating phase waves [149, 338, 357];
see Fig. 5.4. These waves occur spontaneously, but transiently switch to standing
oscillations during the presentation of an odorant stimulus, suggesting that the spatiotemporal patterns may play a role in olfactory processing. Propagating waves consisting of low-frequency oscillations superimposed on a traveling pulse have been
observed in the visual cortex of anesthetized rats [255, 688], cats [39], and primates
[243]. One feature of these visually evoked waves is that they exhibit anisotropy.
For example, Xu et al. [688] found that the stimulus-evoked wave propagated fast
in V1 but then slowed down as it approached the V1/V2 border, causing wave compression. This was followed by the initiation of a reflected wave that propagated
backward into V1 and a transmitted wave propagating across V2. In addition to
these single-cycle events, it is also possible to observe traveling phase waves in the
visual cortex [512]. The precise functional role of spontaneous and stimulus-evoked
waves in the cortex is not currently known. However, one possible contribution has
been proposed in the case of sensory brain regions [170, 683]. More specifically,
synchrony could represent a state of perceptual recognition, while waves could correspond to a state of perceptual scanning. In the latter case, a phase wave would
provide a background depolarization in successive cortical regions, in which neurons have an increased probability of firing and a higher synaptic transmission ef-
190
ficacy (enhanced sensitivity). Thus, rotating waves such as spiral waves could act
as an organizing center that periodically sensitizes an area without the need for any
pacemaker. Interestingly, it has recently been hypothesized that in the motor cortex,
synchrony and phase waves switch roles, with synchrony encoding the absence of
motor action while waves with different frequencies and directions of propagation
encode different motor actions [269].
SleepWake Cycle
Fig. 5.5 EEG recordings from cortex at different stages of the sleepwake cycle [Adapted from
wiki.brown.edu]
Cortical oscillations and waves as measured using electroencephalogram (EEG)

and microelectrode recordings have different signatures depending on the general
state of the brain [153, 272, 442, 602]; see Fig. 5.5. Such rhythms either are generated directly in cortex or emerge from interactions with the thalamus and with
a sheet of inhibitory neurons between the thalamus and cortex known as the thalamic reticular nucleus (TRN); see Fig. 5.6. The waking state and the rapid eye
movement (REM) stage of sleep are both characterized by low-amplitude fast oscillations that tend to have a low level of spatiotemporal coherence. On the other
hand, during non-REM sleep and anesthesia, the brain exhibits significant oscillatory activity at a variety of frequencies and with striking long-range synchrony. Of
191
IV
V
cortex
VI
TRN
+
+
+
thalamus
Fig. 5.6 Three principal regions thought to be responsible for the various rhythms and waves
observed during the sleepwake cycle. Surrounding the thalamus and interposed between the thalamus and cortex is a thin layer of GABAergic neurons known as the thalamic reticular nucleus
(TRN). These neurons are reciprocally coupled to excitatory neurons in the thalamus, forming
the reticulothalamic loop. Thalamic neurons innervate pyramidal cells in input layer IV of cortex,
while pyramidal cells in cortical layer VI send feedback connections to thalamus, resulting in the
thalamocortical loop. There are also axon collaterals from layer VI neurons to neurons in the TRN
particular significance is a slow oscillation (< 1 Hz) of the membrane potential of

cortical neurons, comprising a hyperpolarization phase or down state, during which
the majority of neurons are silent for several seconds, followed by a depolarization phase or up state of smaller duration in which neurons fire at a rate of around
10 Hz. Cortical slice studies [549] combined with computational modeling [27, 122]
suggest that the transition to the up state is due to the spontaneous activation of a
small subset of kindling neurons in deep cortical layers, followed by propagation of
activity vertically through the layers and horizontally across the cortical network.
The transition to the down state involves some combination of synaptic depression
and the opening of a slow Na+ -dependent K+ conductance. Superimposed on the
slow oscillations is another rhythm known as spindle waves, which appear in the
EEG during non-REM sleep as 714 Hz oscillations that wax and wane over a period of a few seconds. Unlike the slow oscillations, which are cortical in origin, the
spindle waves are generated by de-inactivation of a low-threshold Ca2+ current in
thalamic cells following inhibition from TRN cells, which is sufficient to generate
a rebound burst of action potentials that then excites the TRN cells; see Sect. 5.4.3.
Propagating spindle waves with speeds of 1 cm/s can be observed in thalamic slices
following local current injection [334]. (Note that various EEG rhythms are referred
to as waves even though this does not necessarily imply that they are actually
192
propagating as wavesthey can be standing oscillations, e.g., spindle waves, sharp

waves.) Transitions between non-REM sleep, REM sleep, and arousal are controlled
by a variety of neurotransmitters released from brainstem nuclei including acetylcholine, norepinephrine, and serotonin.
Hippocampal Waves
The primary regions of the hippocampus, which is shaped like a rams horn (cornu
ammonis), are CA1, CA3, the dentate gyrus, and the subiculum; see Fig. 5.7. The
main input pathway (perforant path) to the hippocampus comes from the entorhinal
cortex (EC), which is strongly and reciprocally connected to many other parts of
the cerebral cortex. Within the hippocampus, the flow of information from the EC
is largely unidirectional (see inset of Fig. 5.7), with signals propagating through
a series of densely packed cell regions: the dentate gyrus CA3 via the mossy
fibers CA1 via the Shaffer collaterals the subiculum EC. Each of these
regions also contains complex intrinsic circuitry and extensive lateral connections.
In particular, both the dentate gyrus and CA3 have extensive recurrent connections
that are thought to be important in the storage of memories within the hippocampus
[10]. The hippocampus exhibits two major modes of physiological activity, known
as the theta rhythm and large irregular activity (LIA), each of which corresponds
to distinct patterns of population activity and EEG rhythms [98]. The properties of
these activity states have been determined primarily from studies of the rat, although
analogous states have been found in primates.
Fig. 5.7 Modified drawing of the neural circuitry of the rodent hippocampus by Santiago Ramn y
Cajal. Histologie du Systeme Nerveux de lHomme et des Vertebretes, Vols. 1 and 2. A. Maloine.
Paris. 1911. DG: dentate gyrus. Sub: subiculum. EC: entorhinal cortex [Public domain figure from
Wikipedia Commons]
193
During the theta phase, the EEG is dominated by large regular oscillations with
a frequency range of 69 Hz [97]. It is associated with sparse population activity,
in which the vast majority of neurons are silent within any short time window, with
the small remaining fraction of cells firing at relatively high rates (up to 50 Hz)
for around a second. Theta rhythms occur during REM sleep and during periods
of active, alert behavior (particularly locomotion), where cell activity is primarily
determined by the spatial location of the animal, so that the set of active neurons
is constantly changing. This is exemplified by the selective activation of CA1 pyramidal cells at particular locations in space called place fields [474]. A given configuration of place cells provides an internal representation of the environment that
affords an animal knowledge of its position relative to important locations. A key
observation is that the spike activity of a hippocampal CA1 cell advances to earlier
phases of the theta cycle as the animal passes through the cells place field [475].
One possible interpretation of this phase advance is that it reflects a prediction of the
sequence of upcoming locations based on the rats current position, that is, it is part
of a path integration system [384]. It was originally thought that theta oscillations
were synchronized across the hippocampus, but it has subsequently been shown that
theta oscillations in freely behaving rats propagate as traveling waves [393].
The LIA mode occurs during non-REM sleep and during periods of waking
immobility such as resting or eating. The pattern of irregular activity is occasionally interrupted by large amplitude surges called sharp waves [96], which coincide
with 50100 ms bursts of spiking activity in pyramidal cells of CA3 and CA1. Superimposed on the sharp waves are short-lasting high-frequency oscillations called
ripples, with frequencies in the range 150200 Hz in rats. An in vitro study indicates that the sharp waveripple complexes (SPW-R) propagate from CA3 to CA1
and then to the subiculum with a speed around 5 cm s1 [399]. One of the interesting
features of sharp waves is that they have been implicated in memory consolidation
during sleep [611]. For example, place cells exhibit experience-dependent reactivation during sleep that is representative of previous behavior [677]. More specifically,
neurons with overlapping place fields during spatial exploration within an environment show increased coactivity during subsequent sleep, firing in the same sequence
as they fired when the rodent was exploring its surroundings in the awake state.
Pathological States
Propagating waves of activity are also associated with a variety of pathological
brain states, including cortical spreading depression and epileptic seizures. Cortical spreading depression is a complex pulse-like wave of transient depolarization of
neurons and glial cells that propagates slowly across cortical and subcortical gray
matter at speeds of 15 mm/min [596]. Spreading depression, which is thought to
play a major role in migraine with aura [358, 639], appears to involve the diffusion
and active transport of extracellular ions such as K+ and is thus modeled in terms of
reactiondiffusion equations rather than spiking networks or neural field equations;
194
see Sect. 9.3. On the other hand, the latter models have been used to study the initiation, propagation, and termination of epileptiform activity [158, 507]. Epilepsy
is characterized by spontaneously occurring seizures that may involve a significant
portion of the brain at onset (primary generalized seizures) or originate from a circumscribed brain region and may or may not spread to other regions (focal seizures)
[125, 348, 415, 430]; see Sect. 9.4. In the case of focal seizures, a localized region
of pacemaker neurons within deep layers of cortex can drive hypersynchronous activity that rapidly spreads vertically through the layers and then spreads horizontally
as a traveling pulse. There is also a relationship between epilepsy and sleep, which
is not surprising given that sleep also involves changes in levels of synchrony. For
example, the increased synchronization during non-REM sleep might facilitate the
abnormal synchronization of the epileptic brain, while the relatively desynchronized
state during REM sleep or waking might suppress it. Unfortunately, the relationship
between sleep and epilepsy is far more complex and heterogeneous [584].
=2 c D
2D
= 2 c D
3D
= function ()
Fig. 5.8 Three mechanisms for generating and sustaining propagating waves. (a) Fictive waves
due to a local oscillator driving adjacent cortical regions with increasing time delays D . (b) A
local oscillator drives a chain of excitable cells. (c) Traveling phase wave in a chain of neuronal
oscillators with local coupling parameter
Basic Wave Propagation Mechanisms

Ermentrout and Kleinfeld [170] have highlighted three basic mechanisms for wave
propagation in the cortex and other neural circuits; see Fig. 5.8:
1. A single oscillator directly excites neighboring areas with increasing time delays
that generates a fictive propagating wave. This has been suggested as one mechanism for traveling waves in anesthetized visual cortex, with the spread of the
wave determined by axonal propagation delays in long-range horizontal connections [39, 243]. The predominant wave mechanism is monosynaptic.
5.2 Phase Reduction of Synaptically Coupled Neural Oscillators
195
2. A single pacemaker oscillator (or local group of oscillators) drives an excitable

neural medium. Each time the pacemaker fires, it generates a solitary propagating wave of activity. If the pacemaker ceases to be active, then waves no longer
occur. This polysynaptic mechanism underlies the propagation of solitary pulses
in disinhibited cortical slices and is one framework for understanding the spread
of synchronous activity in focal epilepsy or during slow-wave sleep. We will
consider traveling waves in excitable spiking networks in Sect. 5.4.
3. The whole medium is in an oscillatory state so that wave propagation represents
a spatial gradient in the phase of the oscillators, as has been observed in various
sensory cortices as well as central pattern generators (CPGs) for locomotion. In
these cases one can interpret the resulting wave as a phase-locked solution of a
coupled phase oscillator model; see Sect. 5.2.
One important issue is determining what the fundamental oscillatory unit is in each
of these mechanismsa single neuron as considered here or a local population of
neurons (see Chap. 6). A related issue is the time scale over which phase-locking
occursdoes it occur at the level of individual spikes, or bursts of action potentials,
or population firing rates?

In Sect. 1.4 we considered a feedforward neuronal network in which a single output
neuron was synaptically driven by a set of afferent neurons that were themselves
driven by external inputs; see Fig. 1.14. However, in order to model the dynamics
of neuronal networks in the cerebral cortex, for example, it is necessary to take into
account the fact that around 80 % of inputs into a cortical neuron are from other
cortical neurons, indicative of a highly recurrent network architecture. This means
that the output spike train of one neuron feeds back as input to a subset of other
neurons that themselves innervate the given neuron; see Fig. 5.9. Therefore, consider
a network of N synaptically coupled cortical neurons labeled i = 1, . . . , N. Let vi (t)
denote the somatic membrane potential of the ith neuron, evolving according to a
conductance-based model of the form
C
dvi
= Icon,i (vi , . . .) + ui(t),
dt
(5.1)
supplemented by additional equations for the ionic gating variables. (For the
moment we are ignoring any external currents.) The total synaptic current ui (t)
entering the soma is generated by the incoming spike trains from other neurons
in the network that are filtered by synaptic and dendritic processing. Denote the
sequence of firing times of the jth neuron by {T jm , m Z}, with
T jm = inf{t,t > T jm1 |v j (t) = , v j (t) > 0},
(5.2)
196
where is the firing threshold. Following Sect. 1.4, the net synaptic current into
postsynaptic neuron i due to innervation by the spike train from presynaptic neuron j is taken to have the general form m i j (t T jm ), where i j (t) represents
the temporal filtering effects of synaptic and dendritic processing on the spike train
from neuron j innervating neuron i. (For the moment we ignore short-term synaptic
depression and the effects of axonal propagation delays.) Assuming that all synaptic inputs sum linearly, we find that the total synaptic input to the soma of the ith
neuron is
ui (t) = Nj=1 m i j (t T jm ) = Nj=1
t
i j (t t
)a
j (t
)dt
,
(5.3)
where we have set

a j (t) =
(t Tjm ).
(5.4)
mZ
That is, a j (t) represents the output spike train of the jth neuron in terms of a sum
of Dirac delta functions. In general, (5.1)(5.3) are difficult to analyze, particularly
when N is large, so that some form of approximation scheme is needed. One approach is to use the phase reduction procedure of Sect. 1.2.
aj(t)
vi(t)
vj(t)
ij * aj(t)
Fig. 5.9 Schematic diagram of a recurrent neuronal network. Not all connections are shown
Consider a network of N identical conductance-based neurons, each of which is

acting as a limit cycle oscillator. (It is straightforward to extend the analysis to nonidentical oscillators with different intrinsic frequencies, e.g.) If the synaptic interactions between the neurons are sufficiently weak, then the phase reduction method
outlined in Sect. 1.2 can be applied at the network level. This leads to a system of
coupled equations for the phases of the neural oscillators in the network, in which
the effective phase interaction function H depends on both the nature of the synaptic
interactions and the intrinsic properties of each oscillator as expressed through its
phase-resetting curve (see Fig. 5.10). Denoting the state of the ith neuron by xi Rk ,
with x1,i = vi , and taking i j (t) = wi j (t), the network dynamics given by (5.1) and
(5.3) can be written as the coupled system of ODEs
197
N
dx ,i
= f (xi ) + ,1 wi j (t T jm )
dt
m
j=1
(5.5)
for = 1, . . . , k. Suppose that in the absence of interactions each neuron has a stable
limit cycle X with natural frequency 0 , and introduce the phase variable i such
that
d i
= 0 .
(5.6)
dt
There is a one-to-one mapping i X(i ) on the limit cycle. Treating pi (t) =
Nj=1 wi j m (t T jm ) as a weak time-dependent perturbation of the voltage of
the ith oscillator, we can use the method of isochrones in Sect. 1.2, to obtain the
phase equation
N
d i
= 0 + wi j Z(i (t)) (t T jm ),
(5.7)
dt
m
j=1
where Z(i ) is the phase-resetting curve (PRC)more precisely, the component of
the PRC with respect to voltage perturbations.
i
H( j
i)
Fig. 5.10 A system of two synaptically coupled limit cycle oscillators can be reduced to a pair of
phase equations involving an effective phase interaction function H( j i ). See text for details
In order to obtain a closed system of equations, it is necessary to express the firing

times T jm in terms of the phase j (t). Assume that the jth neuron fires whenever the
phase j crosses zero. The firing times then satisfy the condition
j (T jm ) = 2 m,
j (T jm ) > 0.
(5.8)
If we now set j (t) = j (t) 0t then i is a slowly varying function of time (for
small ) so that to a first approximation
T jm = (2 m j (t))/0 = t + (2 m j (t))/0 ,
(5.9)
that is, the current phase determines the most recent sequence of firing times. Hence,
(5.7) reduces to the closed system of equations
198

N
d i
= 0 + wi j Z(i (t))P( j (t)),
dt
j=1
where
(5.10)
P( j (t)) = [( j (t) 2 m)/0].

m
The final step of the analysis is to use the method of averaging to obtain effective
interactions that depend only on phase differences. Substituting i = i 0t into
the phase equation (5.10) gives
d i
=
dt
wi j Z(i (t) + 0t)P( j (t) + 0t).
(5.11)
j=1
The corresponding averaged equation is then

d i
=
dt
wi j H( j i),
(5.12)
j=1
where
1 0
Z(i + 0t)P( j + 0t])dt
H( j i ) =
0 0
2
1
Z( + i j )P( )d ,
=
2 0
(5.13)
with 0 = 2 /0. We have exploited the periodicity of the functions R, P. A useful

alternative expression for H is obtained by substituting for P:
1 2
Z( ) ([ 2 m]/0)d
2 0
m

1
=
Z( ) ( /0 )d .
2 0
H( ) =
(5.14)
5.2.1 Phase-Locking
We define a 1:1 phase-locked solution of the phase equation (5.12) to be of the form
i (t) = ( 0 )t + i ,
(5.15)
where is the collective frequency of the coupled oscillators and = ( 1 , . . . , n )

is a set of constant phases. Substitution into (5.12) leads to the fixed-point equations
= 0 + wi j H( j i ),
j
i = 1, . . . , N.
(5.16)
199
After choosing some reference oscillator, the N phase equations determine the
collective period and N 1 relative phases with the latter independent of .
The local stability of a phase-locked solution can be analyzed by setting i (t) =
i + [ 0 ]t + i (t) and expanding to first order in i (t):
d i
=
dt
Hi j ( ) j ,
(5.17)
j=1
where
Hi j ( ) = wi j H
( j i ) i, j wik H
( k i ),
(5.18)
and H
( ) = dH( )/d . One of the eigenvalues of the Jacobian H is always
zero, and the corresponding eigenvector points in the direction of the flow, that is
(1, 1, . . . , 1). The phase-locked solution will be stable provided that all other eigenvalues have a negative real part [166].
Consider, for example, two identical oscillators (N = 2) that are symmetrically
coupled, with w12 = w21 = 1 and w11 = w22 = 0. The phase difference = 2 1
satisfies the simple equation
d
= H ( ),
(5.19)
dt
where H ( ) = H( ) H( ). Phase-locked states are then given by the zeros
of the odd function H :
H ( ) = 0.
(5.20)
A given phase-locked state is stable provided that

dH ( )
< 0.
d
(5.21)
Note that by symmetry both the inphase ( = 0) and antiphase ( = ) states are
guaranteed to exist. However, there may be additional fixed points depending on the
parameters. The collective frequency of oscillation is determined by the even part
of H:
= 0 + H + ( ).
(5.22)
2
It turns out that the stability of the inphase state for a pair of neurons depends on
a number of factors [258, 651]: whether the synapses are excitatory or inhibitory,
whether the kinetics are fast or slow, whether the phase-resetting curve is type I or
type II, and the size of any axonal delays. We shall illustrate this by considering a
simple example.
Example: Pair of HodgkinHuxley neurons. In Fig. 5.11 we show a typical phaseresetting curve R( ) for the HodgkinHuxley model (1.6) together with the resulting
phase interaction function H( ) for (t) given by an function:
200
b
0.2
R() 0.1
H()
1
0
0
-0.1
-0.5
Fig. 5.11 Plot of type II phase-resetting curve R( ) for the HodgkinHuxley model together with
effective phase interaction function H( )
(t) = 2te t .
(5.23)
If we approximate the PRC by Z( ) = sin then

1 2
Z( )P( )d
2 0

2 d
=
sin( )e /0 d
2 d 0

2
d
=
Im
ei( ) e /0 d .
2 d 0
H( ) =
Now
and so
d
d

0
ei e /0 d =
0
( i0 )2
H( ) = A sin B cos ,
where
A = [ 2 02 ]K,
B = 20 K,
K=
(5.24)
2 0
.
2 [ 2 + 02]2
The odd part is then H ( ) = A sin with A = 0 provided that = 0 , and then
there are phase-locked states at = 0, as expected. The synchronous solution is
stable if H (0) < 0. This implies that synchronization will occur for excitatory
coupling if > 0 and for inhibitory coupling if < 0 . Thus, inhibition rather
than excitation synchronizes the neurons when the synapses have sufficiently slow
kinetics. The reverse is true for the antiphase solution = . If higher-order terms
in the Fourier series expansion of H are included, H( ) = n hn ein , then the transition from synchronous to antisynchronous behavior is smoother, that is, the system
no longer makes a sudden jump from one to the other at a critical value of .
201
motor
output
mutual
inhibition
motor
output
0
time [ms]
150
Fig. 5.12 (a) Swimming locomotion is associated with a phase gradient along the spinal cord
of a lamprey. (Gray bars indicate active phase of a given neuron at different spatial locations or
segments along the spinal cord.) (b) Half-center oscillator consisting of a pair of neurons that have
reciprocal inhibition and fire in antiphase
5.2.2 Traveling Phase Waves on a Chain

The idea that a network of synaptically coupled spiking neurons can be reduced
to a corresponding network of coupled phase oscillators is the starting point for
considering a different type of wave than previous chapters, which can arise when
the network is endowed with spatial structure. That is, suppose that the oscillators are arranged on a regular 1D lattice with sites labeled n = 1, . . . N, and assume
that the synaptic weight wnm of the connection from lattice site m to lattice site
n depends on the spatial separation of the two sites, wnm = w(n m). It is then
possible for the lattice of oscillators to support a phase-locked solution in which
there is an approximately constant phase gradient along the bulk of the lattice.
In other words n (t) t + n . Such a solution is known as a phase wave. Considerable insight into the conditions under which spatial phase gradients occur have
been obtained in studies of central pattern generators (CPGs) for swimming locomotion in invertebrates such as the leech and primitive vertebrates such as the lamprey
[121, 273, 404, 443]; see Fig. 5.12. At the simplest level, this may be modeled in
terms of a chain of N phase oscillators n with nearest-neighbor synaptic coupling
as shown in Fig. 5.13. Each oscillator represents the CPG of an individual segment
of the spinal cord. In the case of the lamprey, each oscillatory unit is a so-called halfcenter oscillator; see Fig. 5.12b. A half-center oscillator consists of two neurons that
are reciprocally coupled by (intrasegmental) inhibition and consequently oscillate
in antiphase. The swimming movement is thus produced by alternating neural activity between the left and right side of the body that propagates along the body as
a phase wave to produce a side-to-side undulating wave; see Fig. 5.12a. Note that
the individual CPG neurons of a half-center oscillator need not be intrinsic oscillators, in which case reciprocal inhibition is required in order to generate the rhythm,
for example, via post-inhibitory rebound. In the case of the lamprey, however, it
202
now appears that excitation from an external drive combined with membrane properties is sufficient for rhythm generation in an individual CPG neuron. In particular,
the primary mechanism for terminating a burst is the activation of fast and slow
Ca2+ -dependent K+ channels [273, 443]. In the case of the leech, intrasegmental
neurons on either side of the body axis fire synchronously, which is consistent with
the fact that the undulatory motion during swimming is dorsoventral rather than
side-to-side.
H+
H+
H+
H+
n
n-1
H-
H-
H+
H-
n+1
H-
H-
Fig. 5.13 A chain of N phase oscillators
During swimming, approximately one full body-length wave is maintained by

period-independent intersegmental phase lags; see Fig. 5.12a. That is, there is a
phase wave that propagates along the spinal cord such that the phase lag =
n+1 n between adjacent segments is a fixed percentage of the cycle period. For
example, a lamprey typically has 100 segments so that the phase lag is around 1 %.
At least two symmetry-breaking mechanisms have been identified for maintaining
a phase gradient along a chain of oscillators [171, 347, 443]:
(a) Asymmetry in the form of a gradient in intrinsic frequencies of the individual
neurons, that is, n = 0 + n1 where 0,1 are constants, will induce a gradient in the relative phases of the coupled system. The resulting phase wave
propagates opposite to the direction of the gradient, that is, from high to low
frequencies. Such a frequency gradient has been found in the lamprey CPG for
swimming, with frequency decreasing from the head (rostral end) to the tail
(caudal end). On the other hand, the distribution of intrinsic frequencies appears
to be U-shaped in the leech, with the highest frequencies occurring mid-body
[443]. Interestingly, a spatial gradient of frequencies has also been observed experimentally in slices of molluscan olfactory lobe, with the wave propagating
from high to low frequencies as predicted by theory [167]; see Sect. 7.5.
(b) Asymmetry between the ascending and descending synaptic interactions so that
H+ = H . In this case, a constant phase lag can be obtained in the bulk of the
chain provided that H+ ( ) or H ( ) vanishes at a nonzero value of , which is
known as the edge condition [171, 347]; see below. Asymmetric coupling has
been found in both the leech and the lamprey. In the latter case, the descending interactions have stronger functional coupling than ascending connections,
whereas in the leech, the coupling strength is symmetric while the specific pattern of connections is asymmetric. More specifically, leech CPG neurons within
a hemi-segment (segment on one side of the body) are active at three different
phases separated by 120, and there are asymmetries between ascending and
descending fibers associated with a given phase group.
203
We will illustrate the above two mechanisms using a simple model of a chain with
H ( ) = W H( ). Such a network is described by the following system of coupled
phase equations:
1 = 1 + W+ H(2 1 )
n = n + W+ H(n+1 n ) + W H(n1 n ),
N = N + W H(N1 N ).
n = 2, . . . , N 1
(5.25)
Introducing the intersegmental phase differences n = n+1 n leads to the N 1dimensional system of equations
n = n +W+ [H(n+1 )H(n )]+W [H(n )H(n1)],
n = 1, . . ., N 1
(5.26)
with boundary conditions

H(0 ) = 0 = H(N )
(5.27)
and n = n+1 n . Suppose that there is a gradient of frequencies along the

chain, that is, n has the same sign for all n. Also take an isotropic, odd interaction
function, W = 1 and H( ) = H( ). Then
n = n + H(m+1 ) + H(n1) 2H(n),
n = 1, . . . , N 1.
(5.28)
The phase-locked solutions = (1 , . . . , N1 ) then satisfy the matrix equation

H( ) = A1 D,
(5.29)
where H( ) = (H(1 ), . . . , H(N1 ))T , D = ( 1 , . . . N1 ) and A is a tridiagonal matrix with elements Ann = 2, An,n+1 = 1 = An+1,n . Suppose, for concreteness,
that H( ) = sin . Then a solution will exist only if every component of A1 D
lies between 1. Let a0 = max{|A1 Di |}. If a0 < 1, then for each n = 1, . . . , N 1
there are two distinct solutions n in the interval [0, 2 ) with H
(n ) > 0 and
H
(n+ ) < 0. In other words, there are 2N phase-locked solutions. Linearizing about
each phase-locked solution and exploiting the structure of the matrix A, it can be
) is stable [171]. Assuming that

proven that only the solution = (1 , . . . N1
the frequency gradient is monotonic, this solution corresponds to a stable traveling
wave. When the frequency gradient becomes too steep to allow phase-locking, that
is, a0 > 1, two or more pools of oscillators (frequency plateaus) tend to form and
oscillate at different frequencies. Waves produced by a frequency gradient do not
have a constant speed or, equivalently, constant phase lags along the chain.
Constant speed waves can be generated from (5.26) by considering phase-locked
solutions defined by n = for all n with a collective period of oscillation determined using n = such that
= 1 + W+ H( ).
The steady-state solutions are then
(5.30)
204
1 + W H( ) = 0
n = 0,
N1 W+ H( ) = 0.
n = 2, . . . , N 2
(5.31)
This implies that all internal oscillators (i = 2, . . . , N 2) have the same intrinsic
frequency. If we further assume that w1 = 0 = N1 , then a constant speed wave
will exist provided that, (i) the interactions are unidirectional with W = 0 and (ii)
H satisfies the edge Property H( ) = 0 for some = 0. For example, if H( ) =
sin( + ) for a fixed phase shift , then the steady-state phase solution is =
. It also follows that although the speed of the wave can change by changing
the intrinsic frequency 1 , the phase lag remains the same. This is consistent
with the swimming motion of invertebrates. Finally, note that if coupling is in both
directions, as in the lamprey CPG, one finds that for sufficiently long chains and
H + = H , the system typically behaves like the unidirectional case except that there
is a small boundary layer close to one end [347].
Although the mathematical analysis of phase waves along a chain of oscillators with nearest-neighbor coupling has provided important insights into CPGs and
swimming locomotion, there are still a number of outstanding issues. One of the
most important is understanding how the phase lags maintain a one body-length
wave irrespective of the cycle period. Is this somehow built into the phase interaction function or the distribution of intrinsic frequencies, or does sensory feedback
play an important role? For example, it is known that leeches and lampreys can
sense their own body movements via proprioceptors and can adjust their cycle period and phase lags accordingly. Moreover, their swimming behavior is strongly
affected by changes in their fluid environment, consistent with the idea that fluid
resistance forces provided by water are also sensed by proprioceptors [443]. Yet
another factor in determining phase lags is the range of synaptic interactions. The
chain model assumed nearest-neighbor coupling, which is consistent with experimental observations that short-range coupling plays an important role in determining the phase. However, longer-range projections do also exist, which in lamprey
can extend up to 20 segments.
5.2.3 Ring of Phase Oscillators and the Sync Basin

In the above example, a stable phase wave solution was generated through a combination of network heterogeneities (in natural frequencies or synaptic interactions)
and boundary effects. An alternative network architecture that supports traveling
phase waves is a ring of identical oscillators [166, 671]. However, in this case, one
or more stable phase waves can coexist with a stable synchronous state. This then
raises the interesting question as to the size of the basin of attraction of the synchronous state (the sync basin) or a particular phase wave as network parameters
change. In other words, what is the set of initial conditions that induces a particular
phase-locked state? From a dynamical systems perspective, finding such basins of
205
attraction is nontrivial. However, Wiley et al. [671] have made some progress by
analyzing a ring of N identical phase oscillators, each of which is coupled to its k
nearest neighbors on either side using a sinusoidal interaction function:
i+k
i = +
sin( j i ),
i = 1, . . . , N,
(5.32)
j=ik
where N 1 and the index i is taken to be periodic mod N. In the case k = N/2
(or its integer part), (5.32) reduces to the Kuramoto model of globally coupled oscillators [350]. It is straightforward to show that the system (5.32) always supports
a stable synchronous state i (t) = t for all i = 1, . . . , N. Moreover, numerically it
is found that the synchronous state is a global attractor for sufficiently long-range
coupling, that is, for k > kc 0.34N [671]. However, as k decreases below kc , other
competing attractors emerge, which take the form of traveling phase waves (also
known as splay states or uniformly twisted waves):
j = t +
2 q j
+ C,
N
j = 1, . . . , N.
(5.33)
Here C is a constant and the integer q is a winding number that specifies the number
of complete rotations or twists in phase that occur as the ring is traversed once. We
summarize the analysis of (5.32) carried out by Wiley et al. [671].
Moving to a rotating frame by setting i = i t and allowing for a nonuniform
coupling W (i j) between oscillators i and j that depends on the distance |i j|
gives
i =
i+k
W (i j) sin( j i ),
i = 1, . . . , N.
(5.34)
j=ik
It is assumed that W (s) is a positive, even, monotonically decreasing function of |s|.

Equation (5.34) is an example of a gradient system, since it can be rewritten in the
form
1 N
V
i =
, V = W (i j) cos( j i ),
(5.35)
i
2 i, j=1
where V is a potential function. It follows that all trajectories of (5.34) flow monotonically downhill on the resulting potential surface V (1 , . . . , N ), and asymptotically approaches fixed points. In other words the only attractors of the system are
phase-locked states. Since N is assumed to be large, it is convenient to consider a
spatially continuous version of (5.34), which takes the form (see also Sect. 5.4.1)
=
t
w(x y) sin[ (y,t) (x,t)]dy.
(5.36)
This can be derived by setting = /N, x = (i N/2) , i = (x), W (i j) =

w(x y) and taking the limit N . For convenience, w is normalized to unity.
It is simple to show that (x,t) = qx, integer q, is a phase-locked solution of (5.36)
206
by direct substitution:
0=
w(x y) sin[q(x y)]dy,
which holds since w is an even function while sine is an odd function. The ring
topology requires that q is an integer, with q = 0 corresponding to a synchronous
state and q = 0 to a phase wave with winding number q.
The linear stability of a phase-locked state is determined by setting (x,t) =
qx + (x,t) and keeping only linear terms in :
=
t
w(x y) cos[q(x y)][ (y,t) (x,t)]dy.
(5.37)
The corresponding eigenvalue equation is obtained by setting (x,t) = e t eimx for

integer m:
w(q
+ m) + w(q
m) 2w(q)
,
(5.38)
= (m, q) =
2
where w is the Fourier transform of the weight distribution:
w(q)
w(y)eiqy dy.
(5.39)
From the properties of w, all eigenvalues are real and (m, q) = (m, q) so that
the condition for stability of the q-phase wave is
w(q
+ m) + w(q
m) 2w(q)
<0
(5.40)
for all m = 1, 2, . . .. Note that the existence of a zero eigenvalue ( (0, q) = 0) reflects the fact the system is time-translation invariant, that is, it corresponds to a
perturbation consisting of a uniform shift of all phases. In general, it is difficult to
establish that (5.40) holds for all positive integers m even when q is fixed. However, for a particular choice of w, which yields the continuum version of (5.32), the
countably infinite set of stability conditions reduces to a single condition [671].That
is, consider the weight distribution

(2 r)1 , |x| r
(5.41)
w(x) =
0,
|x| > r
for 0 < r < 1, which means that each oscillator is coupled to all oscillators within a
fraction r of the ring with coupling strength 1/(2 r). The Fourier transform of w is
w(q)
= f ( qr)
so that (5.40) becomes
sin( qr)
,
qr
(5.42)
207
1
SQ (M) [ f (Q + M) + f (Q M)] f (Q) < 0,
2
M = r, 2r, . . . ,
(5.43)
where Q = qr, M = mr. Treating M as a continuous, real variable and plotting the
function SQ (M) for various Q, one finds that SQ (M) is negative for all M > 0 when
Q is sufficiently small, which ensures that the corresponding phase waves are stable.
At a critical value Q = , SQ
(0) changes sign from negative to positive; SQ
(0) =
f
(Q) so is the smallest positive root of f
(Q) and can be obtained by solving

the transcendental equation
tan( ) =
2
,
2 ( )2
which yields 0.66. Since SQ (0) = 0, it follows that the continuous function
SQ (M) becomes positive in a neighborhood of M = 0 when Q > , and one can no
longer guarantee stability.
b
M
increasing r
SQ(M)
r > rc
r < rc
1
basin size
0.2
kc
probability
0.25
-4
Fig. 5.14 Phase-locked states on a ring of N identical phase oscillators. (a) Graphical construction
to determine the necessary condition for the stability of a q-phase wave with q fixed. A sketch of
the function SQ (M), Q = qr, is shown for r < rc or Q < (gray curve) and r > rc or Q > Qc (black
curve). The points M = nr, integer n are also shown for r < rc (gray dots) and r > rc (black dots).
(b) Variation of the size of the sync basin as a function of k for large N. One way to numerically
estimate the basin size is to consider the fraction of uniformly random initial conditions that
converge to the synchronous state. For k > kc , the synchronous state becomes a global attractor
with = 1. (c) For k = 1, the probability that a uniformly random condition converges to a qphase wave can be fitted by a Gaussian function of q
The above analysis provides a sufficient condition for stability of a q-phase wave
assuming a given range of coupling r [671]: the q-phase wave is stable if |q|r <
. The stability condition is not necessary, however, because one has to take into
account the fact that M is actually discrete. That is, it is possible that the discrete
points M = r, 2r, . . . , lie outside the interval where SQ (M) is positive when Q > . A
simple graphical argument then establishes that for fixed q, increasing r from near
zero means that Q crosses from below, resulting in an increasing interval over
208
which SQ (M) is positive. Even though the points M = r, 2r, . . . , are also moving
away from the origin as r increases, the expanding interval eventually encloses the
point M = r at a critical value rc with Qc = qrc (q) > ; see Fig. 5.14a. This argument
also establishes that the first eigenmode to become unstable is M = r. In summary,
given r > 0, the q-phase wave is stable if and only if [671]
1
[ f (qr + r) + f (qr r)] < f (qr).
2
(5.44)
For a given choice of q, the critical range of coupling rc (q) where the q-phase wave
becomes unstable is the solution to the equation
1
[ f (qrc + rc ) + f (qrc rc )] = f (qrc ).
2
(5.45)
Numerically evaluating rc (q), one finds that rc (q) is a decreasing function of q. In

particular, rc (1) 0.68, which means that all phase waves are unstable for r > rc (1).
In terms of the original discrete network (5.32) with r = 2k/N, this means that the
synchronous solution is globally stable when k > kc (N) 0.34N; see Fig. 5.14b.
(To be more careful, it has not been established that there are no other phase-locked
states other than synchrony and phase waves. However, numerical simulations suggest that any additional states are unstable [671].) Finally, one can estimate the
relative size of the basin of attraction of a phase-locked state by determining the
fraction of uniformly random initial conditions that converge to that particular state;
see Fig. 5.14b, c.
Clearly the idealized ring model considered by Wiley et al. [671] is not biologically realistic. However, it provides a useful framework for exploring an issue that
has been neglected in almost all studies of oscillations and waves in neural networks. That is, in addition to determining existence and stability, it is also important
to consider the types of initial conditions that generate such solutions, particularly
when trying to relate models to experiments in which oscillations and waves are
initiated by specific environmental conditions or external stimuli such as locally
injected currents.
5.2.4 Two-Dimensional Network of Phase Oscillators

Traveling phase waves have recently garnered attention beyond the context of central pattern generators. In particular, Ermentrout and Kleinfeld [170] have argued
that the theory of phase waves in spatially structured networks provides a basis for
understanding traveling electrical waves that have been observed across a variety
of sensory cortices [149, 292, 357, 683, 684]; see Fig. 5.4. The waves are often
expressed during periods of rest, whereas synchronous activity tends to dominate
during the presentation of a strong stimulus. On the other hand, it has been suggested that in motor cortex it is phase waves that encode motor actions, whereas
209
synchrony represents an idling state [269]. In contrast to CPGs, the cortex is effectively a two-dimensional oscillatory medium (assuming activity through the cortical
layers can be lumped together). There have been relatively few studies of coupled
phase oscillators in discrete or continuum two-dimensional arrays [173, 495]. Here
we follow Heitmann et al. [269] and consider a two-dimensional sheet of identical,
locally coupled Kuramoto oscillators:
=+
t
R2
w(|r r
|) sin[ (r
,t) (r,t)]dr
.
(5.46)
For simplicity, the weight kernel is assumed to be isotropic and the domain is
taken to be unbounded. Following the analysis of the ring network in Sect. 5.2.3,
one expects planar phase wave solutions of the form (r,t) = t + k r. This can
be verified by substituting such a solution into (5.46) and using polar coordinates
r = r(cos , sin ), k = k(cos k , sin k ) together with the JacobiAnger expansion
of Bessel functions
eiz cos =
in Jn (z)ein .
(5.47)
n=
Thus
w(|r r
|) sin[k (r
r)]dr
2

= Im
rw(r)
eikr cos( k ) d dr
0=
R2
= Im

0
rw(r)J0 (kr),
which holds since the right-hand side integral is real.

Since the domain is unbounded we expect the spectrum of the linear operator
obtained by linearizing about the phase wave solution to be continuous. That is,
writing (r,t) = t + k r + (r,t) and Taylor expanding (5.46) to first in order in
yields the linear equation
=
t

R2
w(|r r
|) cos[k (r
r)]( (r
,t) (r,t))dr
.
(5.48)
This has solutions of the form
(r,t) = e t eiq.r ,
(5.49)
with q = q(cos q , sin q ) and satisfying

R2

w(|r r
|) cos[k (r
r)] eiq(r r) 1 dr
.
(5.50)
210
Again using polar coordinates and the JacobiAnger expansion, we a obtain a dispersion relation for the continuous spectrum

2

1
n
in( k )
(q) =
rw(r)
+ c.c.
i Jn (kr)e
2 0
0
n=

im Jm (qr)eim( q ) 1 d dr
m=
1
=
2
rw(r)
0

0

Jn (kr)Jn (qr)ein(q k ) + Jn (kr)Jn (qr)ein(q k ) dr
n=
rw(r)J0 (kr).
Using the identity Jn (z) = (1)n Jn (z) then gives

(q) =
rw(r)
0
(1 + (1) )Jn (kr)Jn (qr) cos(n ) + J0(kr)(J0 (qr) 1)

n
dr,
n=1
(5.51)
where = q k . It immediately follows that (q) is real for all q, the spectrum
is independent of the direction of k (reflecting isotropy of the system), and (0) = 0
(reflecting time-translation symmetry). The phase wave of wave number k is then
stable (modulo uniform phase shifts) if (q) < 0 for all q > 0 and [0, 2 ].
Note that it is not sufficient to analyze stability by considering a reduced 1D ring
network as was assumed by Heitmann et al. [269], even if one were to consider a
bounded 2D domain.
In the case of the synchronous solution k = 0, (5.51) reduces to
(q) =

0
rw(r)(J0 (qr) 1)dr,
(5.52)
since Jn (0) = n,0 . Given that J0 (z) 1 for all z, it follows that (q) < 0 for all
q > 0 when the coupling is purely excitatory, that is, w(r) 0 for all r. Hence
the synchronous state of a 2D network of Kuramoto phase oscillators with excitatory coupling is stable. Heitmann et al. used numerical simulations to explore
pattern forming instabilities in the case of center-surround coupling for which (see
Fig. 5.15a)

br2
br2 1 2 4
2
b r br .
+ 4he
(5.53)
w(r) = e
3
They found that when h = 0 (no surround inhibition), the synchronous state is stable,
but increasing h destabilizes the synchronous state resulting in regular or irregular
phase waves. Performing a numerical continuation of stable solutions as h varied,
Heitmann et al. [269] also observed hysteresis in the level of phase coherence R as
a function of h, with the latter order parameter defined according to [350]
R(t)ei (t) =

R2
211
ei (r,t) dr.
(5.54)
A sketch of a numerically determined hysteresis curve is shown in Fig. 5.15b. Synchronous states are identified by high phase coherence (R 1), while phase waves
are identified by low phase coherence (R 0). We summarize the main findings
of Heitmann et al. [269]. Starting at low levels of inhibition (small h), the system
is in a synchronous state (upper branch of Fig. 5.15b), which then develops small
amplitude spatial modulations or ripple patterns as h increases, with intermediate
phase coherence, until the ripples destabilize and collapse to a full wave solution
(lower branch of Fig. 5.15b). The transition from the lower to upper branch occurs
at a smaller value of h then the opposite transition, that is, there is hysteresis. It immediately follows that there exists a range of h values over which there is bistability
between the synchronous state and phase waves. Moreover, in the bistable regime, it
is found numerically that phase waves tend to develop spatial heterogeneities, which
can destabilize the phase waves at lower values of h than regular planar waves.
w(r)
synchrony
ripple
waves
coherence R
waves
weight parameter h
Fig. 5.15 (a) Radially symmetric center-surround weight kernel w(r). (b) Sketch of hysteresis
curve for phase coherence R as a function of weight parameter h, indicating transitions between
the synchronous state (R 1) and phase waves (R 0)
Under the hypothesis that synchrony represents an idling state, whereas a phase
wave somehow encodes a voluntary movement, there has to be some mechanism for
switching between the two modes. Heitmann et al. [269] exploit the bistability of
the model system by proposing that such a switch could be achieved in terms of a
state-dependent perturbation that repels each oscillator from the global mean phase
(see (5.54)) according to
(x) = k sin( (x) ).
(5.55)
One possible physiological implementation of such a perturbation could be a temporary activation of a dense network of inhibitory neurons within the motor cortex
by excitatory thalamocortical projections. Interestingly, it is known that voluntary
movement is initiated in the cortex when these projections are disinhibited by the
basal ganglia [6]. A partial validation of the model was achieved by comparing the
output of the network under perturbation-driven transitions with movement-related
212
modulations of beta oscillations (2025 Hz), which were obtained from magnetoencephalogram (MEG) recordings of human motor cortex during a repetitive fingertapping task [54]. The network output was generated by band-pass filtering the real
part of the order parameter (5.54). It would be interesting to understand some of
these numerical results analytical, using the dispersion relation (5.51), for example,
as well as investigating the effects of boundaries (where the spectrum becomes discrete), network heterogeneities such as spatial variations in the natural frequencies
of the oscillators, and anistropic weight distributions. As might be expected, the
latter tends to select planar waves in particular directions [269]. It would also be
interesting to further explore the similarities and differences between phase waves
in sensory and motor cortices.
5.3 Synaptically Coupled Integrate-and-Fire (IF) Oscillators

The integrate-and-fire (IF) model of a spiking neuron provides a caricature of the capacitative nature of the cell membrane at the expense of a detailed model of the spike
generation process [323]. The basic circuit of the IF model consists of a capacitor
C in parallel with a resistor R driven by a current Ii (t), where i is a neuron label.
The latter is decomposed as Ii (t) = Iext,i (t) + ui (t) where Iext,i (t) is an external current and ui (t) is the synaptic current (5.3) generated by the arrival of spikes from
other neurons. The voltage vi (t) of an IF neuron evolves according to the first-order
equation
C
vi (t)
dvi
=
+ Ii (t).
dt
R
(5.56)
The form of the action potential is not described explicitly. Spikes are formal events
characterized by the ordered sequence of firing times {Tim , m Z} determined by
the threshold crossing conditions (5.2). However, immediately after firing, the potential is reset to a value vr < ,
lim vi (t) = vr .
t[Tim ]+
(5.57)
For simplicity, we set = 1, vr = 0 and C = R = 1. We also take Iext,i (t) = I0 , where

I0 is a constant background input. A network of synaptically coupled IF neurons is
then given by (5.56), (5.2), and (5.3) together with the reset condition (5.57). Taking
the synaptic kernel i j (t) = wi j (t), ui (t) satisfies the equation
M
dui
= ui + I0 + wi j (t T jm ),
dt
j=1
mZ
(5.58)
for Tin < t < Tin+1 . Two network operating regimes can be distinguished according
to the value of the background input relative to the threshold = 1. If I0 > 1, then
213
the network is in an oscillatory regime, since each IF neuron acts as an intrinsic

oscillator when = 0. On the other hand, if I0 < 1, then the network is in an excitable regime, since synaptic inputs are necessary in order to push a neuron over
the threshold. We will consider excitable IF networks in Sect. 5.4
Suppose that I0 > 1. For sufficiently weak coupling , the oscillatory IF network
can be reduced to a coupled phase model following Sect. 5.2. All that is required is
to determine the PRC of a single IF neuron. Therefore, set Ii (t) = I0 + I(t) in (5.56)
and perform the change of variables
i (t) mod 1 =
1
0
v(t)
0
1
dV
I0
=
,
log
I0 V
0
I0 v(t)
(5.59)
where 0 = log(I0 /[I0 1]) is the natural period of oscillation. [It is convenient to
take [0, 1) rather than [0, 2 ).] Differentiating both sides with respect to t shows
that
I(t)
e (t)0
d
= 0 +
= 0 + I(t)
,
dt
I0 v(t)
I0 0
(5.60)
where 0 = 1/0. Thus the phase-resetting curve is the periodic function

Z( ) =
e 0
,
I0 0
0 < 1.
(5.61)
The phase interaction function may now be calculated using (5.13), rewritten in the
form

H( ) =
Z( )P( + )d .
(5.62)
Here P( ) = m ([ m]0 ) is a periodic function of that in the case of the

alpha function (5.23) can be summed as a geometric series to give

2 0 e 0
e0
P( ) =
(5.63)
, 0 < < 1.
0
0
1 e0
(1 e0 )
Using the above phase interaction function, the -dependence of phase-locked
solutions in the case of a symmetric pair of excitatory or inhibitory IF neurons can
be determined [651]. The results are summarized in Fig. 5.16a. For excitatory coupling ( > 0) the synchronous state is unstable for all 0 < < . On the other hand,
the antisynchronous solution is stable for < c but loses stability when > c
with the creation of two stable, partially synchronized or asynchronous states. The
emergence of these two additional states can be understood in terms of the behavior
of the odd part of the interaction function H ( ) as shown in Fig. 5.16b. In the limit
the two asynchronous states asymptotically approach the synchronous state
so that the neurons are almost perfectly synchronized. This is consistent with the
analysis of Mirollo and Strogatz [431] who proved rigorously that globally coupled
IF oscillators almost always synchronize in the presence of instantaneous excita-
214
1.0
0.8
0.6
0.4
0.2
0.0
10
12
14
16
18
20
b
_
H ( )
0
synchronous
anti-synchronous
Fig. 5.16 (a) Relative phase = 2 1 for a pair of IF oscillators with symmetric excitatory
coupling as a function of with I0 = 2. In each case the antiphase state undergoes a bifurcation at
a critical value of = c where it becomes unstable and two additional stable solutions , 1
are created. The synchronous state remains unstable for all . (b) Odd part of the phase interaction
function for just above the critical value c showing two new phase-locked solutions
tory interactions. The stability properties of all solutions are reversed in the case
of inhibitory coupling ( < 0) so that, in particular, the synchronous state is now
stable for all . Introduction of a discrete axonal delay a produces a checkerboard
pattern of alternating stable/unstable solution branches that can overlap to produce
multistable solutions [133].
One of the useful features of the IF model (5.58) is that it is possible to derive
phase-locking equations without the assumption of weak coupling [62, 73, 651].
This can be achieved by solving equation (5.58) directly under the ansatz that the
firing times are of the form T jn = (n j ) for some self-consistent period and
constant phases j . Integrating over the interval t ( i , i ) and incorporating the reset condition leads to the result
1 = (1 e )I0 +
wi j H( j i ),
(5.64)
(t + (k + ) ).
(5.65)
j=1
with
H( ) = e

0
dt et
kZ
215
Note that up to a positive constant factor, H( ) reduces to the weak coupling phase
interaction function of equation (5.62) for 0 . It can be seen that (5.64) has
an identical structure to that of (5.16). However, the former is exact whereas the
latter is only strictly valid to O( ), since it is derived under the assumption of weak
coupling. Moreover, the collective period of oscillations must be determined selfconsistently in (5.64), since H depends on rather than the natural period 0 . It
is also possible to study the stability of phase-locked solutions directly in terms of
the firing times [73, 215, 649]. In the case of slow synapses, one can carry out a
perturbation analysis in terms of the small rate parameter . One then finds that
there are at least two distinct types of local instabilities: (i) instabilities that are
independent of the strength of coupling | |, which also occur in the corresponding
phase reduced mode; (ii) strong coupling instabilities that occur at some critical
value of the coupling strength | | [73]. (Note that there could be additional global
instabilities arising from discontinuous bifurcations [138].)
Linear stability analysis of phase-locked states. For the sake of illustration, consider the
linear stability of the synchronous solution i = 0 for all i. In order to ensure that such a
solution exists, we impose the condition j wi j = . Defining perturbed firing times according to Tin = n + in , we integrate (5.58) from Tin to Tin+1 using the reset condition. This
leads to a mapping of the firing times that can be expanded to first order in the perturbations
in [73, 215, 649]:

{I0 1 + P(0)} in+1 in =
j=1
m=
wi j

Gm (0) jnm in ,
(5.66)
where P( ) = m ([ m] ) and
Gm ( ) = e

0
et
(t + (m + ) )dt.
(5.67)
Note that in contrast to the weak coupling regime, P now depends on the collective period
rather than the natural period 0 The resulting linear delay-difference equation (5.66)
has solutions of the form jn = en j with 0 Im( ) < 2 . Taking (1 , . . . N ) to be an
eigenvector of the weight matrix w with associated eigenvalue p yields the characteristic
equation
) p G(0)],
[I0 1 + P(0)] (e 1) = [G(
(5.68)
with
) =
G(
em Gm (0, ).
(5.69)
m=
One solution to (5.68) is = 0 with i = for all i = 1, . . ., N. This reflects the invariance
of the dynamics with respect to uniform phase shifts in the firing times, Tin Tin + . Thus
the condition for linear stability of a phase-locked state is that all remaining solutions
of (5.68) have negative real part. This ensures that jn 0 as n and, hence, that the
synchronous solution is asymptotically stable. [Note that the above analysis does not take
into account the possibility of instabilities arising from discontinuous bifurcations. These
additional instabilities shift the stability boundaries calculated using linear theory [138].]
216
It is not possible to obtain exact solutions of (5.68). However, considerable insight into
the nature of phase-locking instabilities can be obtained by taking (t) to be the alpha
function (5.23) and performing a perturbation expansion of equation (5.68) with respect to
the inverse rise-time , that is, by working in the regime of slow synapses. First, using the
Fourier transform of (t) gives
P( ) =
eim
1
= + O( 2 ),
2
(
+
i
m
mZ
(5.70)
and
2
im + /T
1
) = e
G(
+ O( 3 ),
2
mZ ( + im + / ) (1 + im + / )
(5.71)
with m = 2 m/ . Substitute these into the eigenvalue equation (5.68) and expand as
a power series in . Interestingly, one finds that there exist two distinct types of solution
for , one of which is O( ) and the other is O( 2 ). This is illustrated schematically in
Fig. 5.17.
Im ()
O()
O(2)
Re ()
Fig. 5.17 Schematic diagram illustrating the spectrum of the linearized firing time map in the
limit of slow synapses. There are N eigenvalues of O( 2 ), one of which is a zero eigenvalue
associated with phase-shift invariance, and 2N eigenvalues s of O( ). (The case N = 3 is shown.)
Destabilization of the synchronous state can occur in the strong coupling regime due to a complex
conjugate pair of O( ) eigenvalues crossing the imaginary axis signaling the onset of a discrete
Hopf bifurcation in the firing times
= O( 2 ) and weak coupling instabilities. Performing a perturbation expansion of equation (5.68) under the assumption = O( 2 ) yields the N eigenvalues
w ( p ) [ p ]G(0),
(5.72)
p = 1, . . ., N. Hence, to the lowest order in , a necessary condition for stability of the

synchronous state at a given coupling strength is Re w ( p ) < 0 for all p = 1, . . ., N 1
(with N = ). In the weak coupling regime this stability condition is also sufficient, as
can be shown by performing a perturbation expansion of equation (5.68) with respect to the
coupling strength rather than . Comparison of (5.67) and (5.69) with (5.65) shows that
G(0)
H
(0)). It follows that the O( 2 ) stability condition is equivalent to the one based
217
on the Jacobian of the phase-averaged model, (5.17). That is, H

(0)[Re p ] < 0, p =
1, . . ., N 1 (with N = ). Such a stability condition only depends on the sign of . We
wish to determine whether or not this stability condition breaks down as | | is increased
(with the sign of fixed).
= O( ) and strong coupling instabilities. Performing a perturbation expansion of equation (5.68) with = O( ) gives the 2N eigenvalues

p [1 e ]
s
( p ) = 1
,
(5.73)
(I0 1 + / ) 2
p = 1, . . ., N. In contrast to the previous case, these solutions are guaranteed to lie in the
left-half complex plane when is sufficiently small so they do not play a role in weak coupling instabilities. However, as is increased from zero one or more eigenvalues may cross
the imaginary axis from left to right in the complex -plane signaling the onset of a discrete
Hopf bifurcation in the firing times and the breakdown of 1:1 frequency-locking. The corresponding Hopf frequency is of the form H = 1 + 2 2 + O( 3 ) where w1 = Imcs at
the bifurcation point and 2 is determined by higher-order terms in the perturbation expansion. If 1 = 0, then the instability induces a time periodic variation in the instantaneous
firing rates of the neurons, whereas if 1 = 0, then the firing rates remain constant but may
differ between neurons.
It turns out that the strong coupling instabilities for slow synapses can be understood in terms of a corresponding rate model [73]. First, rewrite (5.58) as
dvi
= vi + I0 + ui (t)
dt
with
ui (t) = wi j
j

0
( )a j (t )d ,
a j (t) =
(t Tjm ).
mZ
If the synapses are sufficiently slow (small ) then (t) is effectively performing a
short-term temporal average over the incoming spike trains, and the synaptic current
varies slowly in time. Thus, over time intervals of O( 1 ), we can treat the total
input to each IF neuron as approximately constant, so that the neuron acts like a
regular oscillator with instantaneous firing rate
F(ui ) = [log(I0 + ui)/(I0 + ui 1))]1.
(5.74)
Hence, under the approximation

0
( )a j (t )d

0
( )F(u j (t ))d ,
we obtain the closed set of rate equations

ui (t) = wi j
j

0
( )F(u j (t )d .
(5.75)
218
Finally, introducing the differential operator L for which Li (t) = (t), we obtain
the second-order equation
1 d 2 ui (t) 1 dui (t)
+ ui (t) = wi j F(u j (t)).
+
2 dt 2
dt
j
(5.76)
(A related derivation of a rate model based on population averaging will be considered in Sect. 6.1.) Under the condition j wi j = for all i, there exists a homogeneous fixed point ui (t) = u such that u = F(I0 + u)
. Linearizing (5.76) about
the fixed point and substituting a solution of the form ui (t) = e t i , where is an
eigenvector of the weight matrix w, we obtain the set of eigenvalues

( p )
= 1 F
(u)
p,
p = 1, . . . , N.
(5.77)
The fixed point will be asymptotically stable if and only if Re ( p ) < 0 for
all p. It is now straightforward to establish that the O( ) eigenvalues associated with the linear stability of the synchronous state (see (5.73)) are equivalent
to those obtained from the linear stability analysis of the homogeneous state in
the firing rate model (5.77). For a direct comparison, take the firing rate to be
= 1/ so that / = u and
the same in the two models by setting F(I0 + u)
1 e = 1/(u + I0 ). Differentiating F(I0 + u)
with respect to u then shows that
2 /F
(I0 + u)
= (u + I0 )(u + I0 1) and the result follows.
F(I0 + u)
The above analysis establishes that, in the limit of slow synapses, a strong coupling instability in the synchronous state of the spiking model is reflected by a corresponding instability in the homogeneous fixed point of the rate model. As | | is
increased from zero, a codimension one bifurcation of the rate model occurs in two
distinct ways. If a single real eigenvalue crosses the origin in the complex -plane
(1 = 0) then a static bifurcation can occur, leading to the emergence of additional
fixed-point solutions with inhomogeneous firing rates. On the other hand, if a pair
of complex conjugate eigenvalues crosses the imaginary axis from left to right in
the complex plane (1 = 0), then a continuous Hopf bifurcation can occur, leading
to the formation of periodic solutions, that is, time-dependent firing rates. This suggests that the spiking model undergoes a corresponding transition to a non-phaselocked state, consisting of spatiotemporal modulations of the interspike intervals
that are well described by the firing-rate model, at least in the small regime.
Moreover, the fine temporal structure of these modulations, which are averaged out
in the firing-rate model, should become significant with increasing in the spiking
model. This is indeed found to be the case [73].
In Fig. 5.18 we illustrate a strong coupling instability in an excitatoryinhibitory
pair of IF neurons. It is straightforward to show that the rate-based version of this
network exhibits periodic variations in the mean firing rates in the strong coupling
regime. It can be established from (5.72) and (5.73) that the synchronous state of
the IF network is stable for sufficiently weak coupling but destabilizes due to a Hopf
bifurcation in the firing time map at a critical coupling c ( ). A direct numerical
simulation of the system shows that beyond the bifurcation point, the neurons jumps
5.4 Spiking Network Models of Cortical and Subcortical Waves
219
to a state in which they exhibit periodic bursting patterns. This can be understood in
terms of mode-locking associated with periodic variations of the interspike intervals
(ISIs) kn = Tkn+1 Tkn , k = E, I, on closed attracting orbits. Suppose that the kth
n+pMk
oscillator has a periodic solution of length Mk so that k
= kn for all integers p.
1
n
If k k for all n = 2, . . . , Mk , say, then the resulting spike train exhibits bursting
with the interburst interval equal to k1 and the number of spikes per burst equal
to Mk . Although both oscillators have different interburst intervals (11 = 21 ) and
numbers of spikes per burst (M1 = M2 ), their spike trains have the same total period,
M2
n
n
1
that is, M
n=1 1 = n=1 2 . Due to the periodicity of the activity, the ISIs only
fall on a number of discrete points on the orbit. The variation of the ISIs in with
n is compared directly with the corresponding variation of the firing rates of the
reduced model in Fig. 5.18b, which shows that the instantaneous firing rate captures
the modulation in the ISIs.
wEI
wEE
c
0.3
wII
wIE
n
k
k=2
n
k
0.2
0.2
2
0
120
130
140
150
t (in units of d )
160
n-1 0.3
k
k=1
0
n-1
k
Fig. 5.18 Strong coupling instability in an excitatoryinhibitory pair of IF neurons. (a) Basic network. Synaptic weights are wEE = wII = 0, wIE = 1 and wEI = 2. The synaptic rate constant is
= 1 and the coupling strength is = 2. (b) Spike train dynamics of the two neurons. For the sake
of illustration the firing times of the two oscillators are represented with lines of different heights
(marked with a +). Smooth curves represent variations in the firing rates of the corresponding rate
model. (c) A plot of the interspike intervals (ISIs) (kn1 , kn ), k = E, I, of the spike trains shown
in (b) showing how they lie on closed periodic orbits. Points on an orbit are connected by lines.
Each triangular region is associated with only one of the neurons, highlighting the difference in
interburst intervals. The inset is a blowup of orbit points for one of the neurons within a burst

5.4.1 Traveling Pulses in an Excitatory Spiking Network
In the previous section we showed how oscillations and phase-locking in a synaptically coupled IF network can be analyzed in terms of an underlying firing time
map. It turns out that the firing times can also be used to investigate solitary wave
propagation in an excitable IF network, for which I0 < 1 [62, 63, 167]. Consider a
220
network of IF neurons distributed along an infinite 1D lattice:

dv j
= v j + I0 + g W ( j k) (t Tk ),
dt
kZ
(5.78)
with the threshold condition v j (T j ) = 1 and the reset condition lim 0+ v j (T j + ) =

0. For simplicity, we are assuming that each neuron fires only once during passage
of a traveling pulse. (Numerical simulations indicate that wave propagation is not
dramatically altered if each neuron fires multiple spikes [167].) We now define a
traveling pulse solution as one for which there is a constant time lag between the
firing of adjacent neurons, that is, Tk = k (up to an arbitrary constant). (The same
definition was used for intracellular Ca2+ waves in the firediffusefire model of
Sect. 4.3.2.) Substituting into (5.78), gives
dv j
= v j + I0 + g W ( j k) (t k ).
dt
kZ
(5.79)
Analytically speaking, it is convenient to consider a continuum version of the model.

Therefore, introduce a lattice spacing d and set v(x,t) = vx/d (t), w(x y) = W ([x
y]/d), T (y) = Ty/d . Moreover, rescale the coupling constant according to g gd.
Taking the continuum limit d 0 and 0 such that the speed d/ = c is fixed,
then T (y) = y/c and we obtain the continuum IF equation
v(x,t)
= I0 v(x,t) + g
t
w(x y) (t y/c)dy.
(5.80)
Equation (5.80) is supplemented by the reset condition v(x,t + ) = 0 whenever

v(x,t) = 1. In the following, we set I0 = 0, assume that g is a positive coupling
parameter,
and W (x) is a positive, symmetric, monotonically decreasing function of
|x| with 0 W (x)dx < . Typical choices for such a weight function are
2
2
1
ex /2
w(x) =
2
2
(Gaussian),
w(x) =
1 |x|/
e
2
(exponential).
(5.81)
For concreteness, we choose the exponential function, although the basic results will
not depend on the precise form of w(x).
In order to determine the velocity c of a solitary pulse, multiply both sides of
(5.80) by et and integrate over the interval < t T (x) = x/c using the threshold
condition v(x, T (x)) = 1. This generates a self-consistency condition for c of the
form [62, 167]
x/c
et
w(x y) (t y/c)dy dt
1=g
=g

0
w(x)ex/c
x/c
0
et (t)dtdx.
(5.82)
221
The second line is obtained by noting that (t y/c) = 0 for y > ct, reversing the
order of integration,
x/c
ct
dt
dy
x
x/c
dy
dy,
y/c
and performing a change of variables. In the case of an exponential weight function, the integration over x and t can be performed explicitly
Fourier expanding
iby
(t) and using the analyticity properties of ( ) =
e t (t)dt, namely, that
( ) is analytic in the lower-half complex -plane since (t) = 0 for t 0. After
performing a contour integral, we obtain the result
gc
(ic/ ).
2( + c)
1=
(5.83)
Suppose that (t) is given by the function (5.23) together with a finite axonal
propagation delay a :
(t) = 2 (t a )e (ta ) H(t a ),
(5.84)
where H is the Heaviside function. Substitution into (5.83) leads to the following
implicit equation for the propagation velocity of a solitary pulse:
1=
g 2 c
eca /
.
2( + c) ( + c/ )2
(5.85)
In the absence of axonal

delays (a = 0) and large c, the velocity scales according
to a power law c g/2 [167]. This form of power law is consistent with the
behavior observed in more detailed computational models [228]. The velocity c is
plotted as a function of the coupling g in Fig. 5.19. It can be seen that for fixed
and a there exists a critical coupling gs ( , a ) such that there are no traveling pulse
solutions for g < gs and two solutions for g > gs . Using linear stability analysis (see
below), one finds that for the given parameter values, the lower (upper) solution
branch is unstable (stable).
The (local) asymptotic stability of a solitary wave can be determined by considering perturbations of the firing times. For simplicity, we will only consider linear
stability analysis based upon perturbations traveling at the same speed as the wave
itself, that is, perturbations of the single neuron firing times T (x). It is possible that
there are further instabilities associated with global bifurcations where, for example,
certain neurons fire more than once, fail to fire, or fire out of sequence. Suppose that
T (x) = x/c+ (x) with (x) = 0 (x) for x 0 and 0 (x) a prescribed, bounded function on (, 0]. Asymptotic stability then corresponds to the condition (x) 0 as
x for arbitrary nonuniform initial data 0 (x). Integrating (5.80) over the interval
(, T (x)] gives
ex/c+ (x) = g
x/c+ (x)
et
w(x x
) (t x
/c (x
))dx
dt.
222

2
a=0
a=1
1.5
c
=2
= 0.25
a=0
0.5
a=1
10
20
30
40
50
60
70
80
Fig. 5.19 Variation of the speed c of a traveling pulse as a function of the synaptic coupling g for
different values of the inverse rise-time and axonal delay a . Solid (dashed) lines correspond to
stable (unstable) branches. The external input is taken to be I0 = 0 and the range of the exponential
weight function is = 1
Expanding to first order in (x) and imposing (5.82) leads to the linear integral
equation
0

et
w(x x
)
(t + [x x
]/c) (x) (x
) dx
dt = 0. (5.86)
Substitution of a solution of the form (x) = e x then yields the characteristic equation (after reversing the order of integration and performing a change of variables)

H ( )
e x 1 w(x) f (x/c)dx = 0,
(5.87)
0
with
f ( ) =

0
e( t)
(t)dt.
(5.88)
Asymptotic stability then holds if all nonzero solutions have negative real part.
Such solutions depend indirectly on the coupling g via the speed c, which satisfies (5.82). Equation (5.87) can be evaluated using Fourier analysis and, up to an
irrelevant constant factor,
H ( ) = P(c[ 1 + ]) P(c/ ),
P(z) =
z
(iz).
2(1 + z)
(5.89)
(5.90)
Note that P(z) is real when z is real, and P(z) 0 as |z| . The stability of the
fast and slow solution branches displayed in Fig 5.19. can be determined from the
following theorem [62]:
223
Theorem 5.1. Let C+ and C denote, respectively, the fast and slow solution branches
c = c (g), g gs , of the self-consistency condition (5.83) for the velocity of a solitary
pulse, assuming a unimodal response function (t). Here gs is the critical coupling for the
existence of a solitary wave. Denote the time to peak of (t) by max . The following stability
results then hold: (i) The branch C is unstable for all max and a . (ii) The branch C+ is
stable for all max in the case of zero axonal delays, a = 0. (iii) In the case of large axonal
delays and fast synapses (small max ), there exists a Hopf bifurcation point gh such that
c+ (g) is stable (unstable) for g > gh (gs g < gh ).
We sketch a proof. First, note that (5.83) can be written in the more compact form
1 = g(c)P(c/ ),
(5.91)
where the dependence on the coupling as a function of c has been made explicit. Now
decompose = a + ib and differentiate (5.89) with respect to a for b = 0:
H (a, 0)/ a|a=0 = 1 P

(c/ ).
Similarly, differentiating (5.91) with respect to c shows that
0 = g
(c)P(c/ ) + g(c)P
(c/ ).
It follows that
sign [ H (a, 0)/ a|a=0 ] = sign[g

(c)].
Hence, H(a, 0)/ a|a=0 > 0 when dg/dc < 0, that is, when c C . Since lima H (a, 0) =
P(c) < 0, it follows that H(a, 0) must cross the positive a-axis at least once when c C .
This implies that H(a, 0) = 0 has at least one solution for a > 0 and, hence, that the branch
C is unstable. Second, assuming that the delay kernel (t) is unimodal with a maximum
at max , it follows from (5.88) that f ( ) > 0 for 0 < < max . The real part of (5.89) can
then be decomposed as
Re[H ( )] =
cmax
0

eax 1 w(x) f (x/c)dx +
cmax

e x 1 w(x) f (x/c)dx.
with the first term on the right-hand side positive definite. Since w(x) is a monotonically
decreasing function of x, it follows that the second term becomes negligible when cmax is
sufficiently large. In other words, H ( ) = 0 and hence the upper branch C+ is stable in the
large-c limit. It can also be established that H (a, 0) = 0 does not have any positive definite
solutions when c C+ so that the upper branch is stable with respect to static bifurcations.
The proof is completed by establishing conditions for the occurrence of a Hopf instability
on the upper branch. This requires searching for solutions a = 0, b = 0 of (5.89).
It turns out that for the parameter values chosen in Fig. 5.19, the upper branch
is stable. However, one finds that for larger values of the axonal delay a , a continuous solitary pulse can become unstable as the axonal delay is increased. This
is illustrated in Fig. 5.20, where the fast and slow solution branches are plotted as
a function of a . Interestingly, Golomb and Ermentrout [230] have demonstrated
that in the regime where the continuous wave is unstable, lurching (pulsating)
waves propagate with discontinuous, periodic spatiotemporal characteristics; see
Fig 5.21. Space is now spontaneously divided into discrete spatial intervals of length
x, and
T (x + x) = T (x) + t,
(5.92)
224
2
stable
unstable
1.5
c 1
c+
0.5
0
c0
10
a
Fig. 5.20 Fast and slow velocity branches as a function of the axonal delay a for fixed and g.
The stable slow branch undergoes a Hopf instability at a critical value of axonal delay
small delay
large delay
x
x
Fig. 5.21 Schematic illustration of the difference between continuous and lurching pulses in an
excitatory IF network with delays. Each IF neuron only fires a single spike at a time as indicated
by the solid curves
where t is the period of a lurching cycle. Hence, the average speed of the pulse is
c = x/ t. Suppose that the lurching wave starts at x = 0 and T (0) = 0. The firing
time of the neuron at x is then
T (x) = n t + f (x),
(5.93)
where n is the integer part of T (x)/ t and x = x n x. The function f specifies

how the firing times vary within a single lurch, and is defined on the interval [0.L)
with f (0) = 0. One also finds that lurching waves can persist in regions where the
continuous wave does not exist and, for certain choices of weight kernel w(x) such as
a step function, the two types of wave may coexist (bistability) [230]. It follows that
T (x) x/c = f (x)
x/c
is a x-periodic function of x. Interestingly, the possible
correlation between the nature of wave propagation and the effective size of discrete
delays is suggestive of the observed differences between the spread of activity in
cortical and thalamic slices; see Sect. 5.4.3.
225
There have also been a number of studies of 2D spiking networks. For example, Kistler et al. analyzed solitary waves and periodic wave trains in an excitable
network based on the so-called spike response model [337], which is essentially an
integrated version of the IF model. However, there were a few important differences
between their treatment and the analysis of IF networks reviewed above. First, they
effectively took the activity of a solitary pulse propagating in the x-direction, say,
to be c1 (t x/c) rather than (ct x). The additional factor of c1 results in
only one solution branch for c = c(g) rather than two as found in [62, 167]. It is
not clear that there is any biophysical justification for this additional factor, given
the interpretation of (t) as a synaptic impulse response function; see Sect. 1.4.
Second, their stability analysis was mainly based on studying changes in local field
potentials, which is not sufficient to guarantee asymptotic stability with respect to
more general perturbations of the firing times as considered above. In particular, it
does not take into account the full spectrum of the linearized firing time map, which
is crucial in establishing the occurrence of Hopf instabilities. The numerical results
presented in [337] did, however, reveal a rich repertoire of geometrical waves such
as spirals and target patterns. Spontaneous spiral waves have also been observed
in a stochastic IF network [117]. Finally, the existence of stable traveling pulses in
a synaptically coupled IF network has been linked to the empirical observation of
traveling pulses in disinhibited brain slices and with epileptiform activity following brain trauma. However, as in the case of oscillatory IF networks, it is not clear
what the appropriate level of modeling should be. That is, whether the fundamental
physiological element of a synaptically coupled network, discrete or continuous, is
a single neuron or a local population of neurons. We will consider the latter case in
Chaps. 6 and 7.
5.4.2 Traveling Pulses in Networks of Excitatory and Inhibitory

Spiking Neurons
It is also possible to extend the above analysis to a 1D network of excitatory (E) and
inhibitory IF neurons, in which there are fast (AMPA) and slow (NMDA) excitatory
synapses [229]; see Fig. 5.22. The generalization of the continuum model (5.80)
takes the form
v (x,t)
= I v (x,t) + g E
t
= f ,s
g I
w E (x y) E (t TE (y))dy
w I (x y) I (t TI (y))dy
(5.94)
for = E, I. Here v (x,t) describes the time-dependent membrane potential of the

-type neuron at x, w (x) represents the synaptic weight density from neurons of
type to neurons of type , and the synaptic filters are taken to be exponentials
226

excitation
excitatory (E) neurons
excitation
inhibition
inhibitory (I) neurons

inhibition
Fig. 5.22 Cortical network model consisting of a 1D layer of excitatory neurons reciprocally coupled to a 1D layer of inhibitory layers
1
exp(t/ )H(t).
(5.95)
The additional superscript for excitatory synapses specifies whether the excitatory
synapse is fast ( = f ) or slow ( = s). For concreteness, the synaptic weights are
also taken to be exponentials,
w =
1
exp(|x|/ ).
(5.96)
Again it is assumed that each neuron at x only fires once and the corresponding
firing times are denoted by T (x) for = E, I. The spatial and temporal scales are
fixed by taking EE = 1 and the membrane time constants to be unity. Numerical
simulations establish that there are two types of stable traveling pulses. One is a fast
pulse that is similar to the continuous pulse found in the excitatory networkits
speed increases as inhibition decreases. The other is a continuous slow pulse that
is characterized by a much smaller velocity and by the fact that the inhibitory cells
fire before their neighboring excitatory cells [229]. Moreover, in a narrow range of
parameters, the two solutions coexist (bistability).
The analysis of existence and stability of traveling pulses proceeds along similar
lines to the excitatory case [229], so we will only summarize the results. Integrating
(5.94) with respect to t and imposing the threshold conditions v (T ) = gives
the pair of equations
= f ,s
g E
g I
w E (x
)G E [T (x) TE (x x
)]dx
w I (x
)G I [T (x) TI (x x
)]dx
,
where
G (t) =
t
0
(s)ds.
(5.97)
(5.98)
227
It is also required that the voltage be below threshold prior to firing, V (x,t) <
for t < T (x). A traveling pulse solution with velocity c is defined according to
x
TE (x) = ,
c
TI (x) =
x
+,
c
(5.99)
where is the shift in the firing times of the inhibitory neurons relative to the excitatory neurons, so that < 0 means that the I cells fire before the corresponding
E cells. Substituting such a solution into (5.97) and using the explicit expressions
for w and G yield two algebraic equations for the two unknowns c, . Existence of
a traveling pulse then reduces to the problem of solving for c, . Similarly, stability
of a solution is determined by substituting
TE (x) =
x
+ E (x),
c
TI (x) =
x
+ + I (x),
c
(5.100)
into (5.97), expanding to first order in E,I and analyzing the spectrum of the resulting linear operator. The quantitative results that emerge from such an analysis are
summarized in Fig. 5.23. A number of additional points are as follows [229]:
1. The fast pulses are the continuation of the waves found in excitatory networks.
The E neuron typically fires before or just after neighboring I cells. They are
robust solutions that can be generated by a strong localized current injection.
2. Slow waves are characterized by I cells firing significantly before neighboring
E cells. Although slow waves can exist in the absence of slow excitation, the
presence of the latter greatly expands the basin of attraction and the possibility
of inducing a slow wave with a localized perturbation.
3. There exists a bistable regime and a regime in which there is a continuous transition from fast to slow waves.
4. It is also possible to find parameter regimes in which there are lurching waves
and other more irregular solutions.
The possibility for a cortical network to sustain two propagation velocities has
also been observed in a much more detailed conductance-based model [122], with
slow waves observed in the presence of inhibition and fast waves in the absence
of inhibition. The more detailed model also showed that inhibitory neurons always
fired before excitatory neurons during the propagation of slow waves. One limitation of the IF network study was that, in order to make the mathematical analysis
tractable, it was necessary to impose the constraint that each neuron fired only one
spike. On the other hand, the biophysical model can simulate the propagation of
a pulse that represents the onset of a long-lasting active spiking (up) state and the
transition to an inactive (down) state. Neurons within the pulse fire coherently at
a rate of around 10 Hz. Repetitive stimulus-invoked or spontaneous activation of a
local population resulted in the rhythmic slow oscillations (< 1 Hz) between up and
down states observed in EEG recordings during slow-wave sleep [602] and in vitro
slice studies [549]. There is still a debate regarding the precise mechanisms for the
spontaneous transitions between the up and down states, although they are likely to
involve some combination of intrinsic membrane properties and slow/fast synaptic
interactions [27, 122, 419, 488, 549].
228

fast
c
SN
SN
20
slow
SN
g EI
SN
g EI
slow wave
fast wave
0
B
f
g IE
continuous transitions
g EI
40
Fig. 5.23 Sketch of phase diagram showing the existence of stable fast and slow waves in the
f
f
(gIE
, gEI ) plane for gsEE = 10, gEE
= 12, and gsIE = 0. The ranges of the synaptic weights are
taken to be EE = 1 = IE , II = 0.5 = IE , and the thresholds are E = I = 1. Assuming the
s = s = 5/3, f = f = 1/12, and = = 4/15. The
membrane time constants are unity, EE
II
EI
IE
EE
IE
solid lines show saddlenode (SN) bifurcation curves: along AB a stable/unstable pair of fast waves
coalesce, along CD a stable/unstable pair of slow waves coalesce, and along OB the unstable fast
wave coalesces with the stable slow wave. A stable fast wave exists in the dark-shaded domain,
a stable slow wave exists in the light-shaded domain, and the waves coexist in the region OCB
(bistability). The points B,C are codimension two-cusp bifurcation points. The insets show the
f
bifurcation diagram with respect to gEI for two different values of gIE (not shown to scale). In
particular, the right-hand inset illustrates the case of a continuous transition from a fast to a slow
wave with the latter disappearing in an SN. The parameter regime over which slow waves exists
shrinks significantly in the absence of slow excitation (gsIE 0), partly due to the fact that the slow
wave can destabilize via a Hopf bifurcation
5.4.3 Spiking Model of Thalamic Spindle Waves

As we have already highlighted, computational and experimental studies of disinhibited neocortical slices indicate that neuronal discharges propagate continuously
at a velocity c 10 cm/s [228]. Axonal delays are relatively small. On the other
hand, in models of thalamic slices, composed of excitatory thalamocortical (TC)
neurons and inhibitory reticular (RE) thalamic neurons, waves propagate in a lurching manner at a velocity c 1 cm/s [232, 334, 349]. This is thought to form the
basic mechanism for the generation of 714 Hz spindle oscillations during the onset
of sleep. Each recruitment cycle of the lurching waves has two stages (see Fig. 5.24):
(i) a new group of inhibitory RE cells is excited by synapses from TC cells, and this
229
RE group then inhibits a new group of TC cells; (ii) the new recruited TC cells rebound from hyperpolarization and fire a burst of spikes, which further recruit more
RE cells during the next cycle. One can reduce the two-layer thalamic model to a
single-layer excitatory network with a large effective delay (a 100 ms) caused
by the time needed for a TC cell to rebound from inhibition [230, 525]. If such
a network is modeled in terms of an IF network with delays, then one can obtain
lurching waves, as discussed at the end of Sect. 5.3. A more detailed analysis of the
post-inhibitory rebound mechanism underlying the generation of continuous and
lurching waves in thalamic networks has been carried out using conductance-based
[153, 232, 620] and integrate-and-fire-or-burst models [128]. We briefly describe
the latter model.
TE cells
excitation
inhibition
RE cells
inhibition
Fig. 5.24 Two-layer thalamic model. Inhibitory reticular (RE) cells inhibit excitatory thalamocortical (TC) cells, and TC cells excite RE cells. RE cells also mutually inhibit each other but this has
a relatively small effect on discharge propagation
Several models of bursting neurons include a low-threshold calcium T -current

[657]. Such a current is slowly activated when the neuron is sufficiently hyperpolarized so that it can cause the neuron to fire a burst of action potentials when released
from sustained inhibition, so-called post-inhibitory rebound. Recently the integrateand-fire model has been extended to incorporate this slow acting T -current leading
to the aptly named integrate-and-fire-or-burst (IFB) model [128, 138, 589]. The
current balance equation for the model is
C
dv
= I IL IT ,
dt
(5.101)
where C is a membrane capacitance, I represents external or synaptic current, and

gL (v vL ) is a leakage current with constant conductance gL and reversal potential
vL . The low-threshold Ca2+ current is given by
IT = gT h(t)(v vT )H(v vh ),
(5.102)
where H is a Heaviside function and the slow variable h has the dynamics

dh
h/ , v vh
=
(5.103)
(1
h)/+, v < vh .
dt
230
The variable h, which represents slow de-inactivation of the low-threshold Ca2+

conductance, relaxes to zero with time constant when v vh and relaxes to unity
with time constant + when v < vh . Since vh < vr < , where is the firing threshold
and vr is the rest level, it follows that sustained hyperpolarization is required in order
that the potential crosses the vh boundary from above, resulting in an increase in h.
If the cell is then released from inhibition so that v recrosses the threshold vh from
below, a nonzero IT current is generated that can depolarize the cell and produce a
burst of spikes. In the absence of such hyperpolarization, the system acts as a regular integrate-and-fire neuron. For an appropriate choice of parameters, it has been
shown that the IFB model under periodic stimulation exhibits excellent agreement
with experimental results obtained from intracellular recordings of thalamic relay
cells [589]. In the case of a constant external input, it is possible to obtain an exact
solution of the IFB model in terms of characteristic curves [105]. Taking the ratio of
(5.101) and (5.103) in the Ca2+ de-inactivated region (v < vh ), we have (for C = 1)
I0 + gL (vL v)
dv
= +
,
dh
1h
(5.104)
whereas, in the Ca2+ inactivated region (v < vh ),

dv
I0 + gL (vL v) + gT h(vT v)
=
.
dh
h
(5.105)
A sketch of a typical solution for v = v(h) in both regimes is shown in Fig. 5.25; see
[105] for a more detailed analysis.
1.0
(v*,h* )
0.0
vh
vr
Fig. 5.25 Illustrative sketch of typical characteristic curves for an IFB neuron driven by a subthreshold constant input. A burst of four spikes arises when the neuron crosses the threshold vh
from below. Bursting ceases when h has decreased sufficiently so that the voltage v can no longer
cross the firing threshold . Once v decreases below vh , h increases to its equilibrium value h = 1
where v = v < vh
231
A network version of the IFB model can be constructed along similar lines to a
standard IF model [128]. Let (vi , hi ) denote the voltage and slow inactivation variable of neuron i. Equation (5.101) becomes (for C = 1)
dvi
= ui (t) gL (vi vL ) gT hi (t)(vi vT )H(vi vh ),
(5.106)
dt
where hi evolves according to (5.103) and ui (t) is the total synaptic current into
neuron i:
(5.107)
ui (t) = wi j (t T jm ),
j
T jm
with the firing times

determined by (5.2). Equation (5.106) is supplemented by
the reset condition (5.57). In principle, one could now investigate the existence and
stability of continuous and lurching waves in a one-dimensional network of IFB
neurons by taking a continuum limit along the lines of Sect. 5.4.1. However, the
analysis is still nontrivial unless some further approximation is made such as taking
the synapses to be slow so that the spiking model can be reduced to a rate-based
model [128]. Nevertheless, it can be shown that a single-layer inhibitory network of
IFB neurons with vh below the resting potential can support lurching waves, whereas
a corresponding excitatory network with vh above the resting potential supports continuous waves. In terms of the two-layer network shown in Fig. 5.24, the former can
be interpreted as a thalamic network with TRN-mediated inhibition while the former
corresponds to a TRN network with thalamus-mediated excitation.
Chapter 6
Population Models and Neural Fields
In this chapter we show how to construct population-based models of synaptically

coupled neural networks, distinguishing between voltage-based and activity-based
versions [71, 168]. We also consider the important issue of how noise at the singlecell level (Poisson-like spiking statistics) relates to noise at the population level. In
particular, we describe a neural master equation formulation of stochastic population
dynamics. One biological motivation for population-based models is that neurons in
cortex tend to be organized into tightly coupled groups known as cortical columns,
in which the cells share similar functional properties (Sects. 5.1 and 8.1). Another
important feature of cortex is that it is spatially organized at both the anatomical and
functional levels, which can be modeled in terms of spatially structured networks of
interacting populations. In the continuum limit, such models can be described in
terms of neural field equations. The advantage of a continuum rather than a discrete
representation of spatially structured networks is that various techniques from the
analysis of PDEs presented in Part I can be adapted to the study of neural field
models of cortex (Chaps. 7 and 8).
6.1 Population Averaging and Rate Models

Suppose that a network of synaptically coupled spiking neurons is partitioned into
a set of P homogeneous populations labeled = 1, . . . , P, with N = N neurons in each population. (It is straightforward to relax this assumption by taking
N = O(N).) Let denote the population function that maps the single neuron index
i = 1, . . . , N to the population index to which neuron i belongs: (i) = . Furthermore, suppose the synaptic interactions between populations are the same for all
neuron pairs. (Relaxing this assumption can lead to additional sources of stochasticity as explored in [184, 632].) Denote the sequence of firing times of the jth neuron
by {T jm , m Z}. The net synaptic current into postsynaptic neuron i due to innervation by the spike train from presynaptic neuron j, with (i) = , ( j) = , is
taken to have the general form N 1 m (t T jm ), where N 1 (t) represents
233
234
6 Population Models and Neural Fields
the temporal filtering effects of synaptic and dendritic processing of inputs from any
neuron of population to any neuron of population ; see Sect. 1.4.2. Assuming
that all synaptic inputs sum linearly, the total synaptic input to the soma of the ith
neuron is
ui (t) =
1
N
j; ( j)=
(t T jm ) =
t
(t t
) N
j; ( j)=
a j (t
)dt
(6.1)
for all (i) = , where a j (t) = mZ (t T jm ). That is, a j (t) represents the output
spike train of the jth neuron in terms of a sum of Dirac delta functions. In order
to obtain a closed set of equations, we have to determine the firing times Tim given
by Eq. (5.2), where vi (t) evolves according to conductance-based model (5.1). It
follows that, since the right-hand side of (6.1) is independent of i, we can set ui (t) =
u (t) for all p(i) = with
u (t) =
t
=1
(t t
)a (t
)dt
,
(6.2)
and a (t) is the output activity of the th population:

a (t) =
1
N
j; ( j)=
a j (t).
(6.3)
We now assume that each homogeneous population is close to a so-called asynchronous state, in which the spike trains of different neurons within a homogeneous
population are uncorrelated (see Sect. 6.3.1). The population activity a is then
approximately constant, which means that the synaptic currents are also slowly
varying functions of time. It follows that a can track the input current according to a (t) = F(u (t)), where F is a population gain function [2, 86, 87, 213].
Substituting this approximation into (6.2) leads to the closed system of equations
u (t) =
t
=1
(t t
)F[u (t
)]dt
,
(6.4)
(Note, however, the asynchronous state only exists in the thermodynamic limit
so that for finite N we expect statistical fluctuations around the asynchronous
state; see Sect. 6.4.) A rate equation identical in form to (6.4) was derived in
Sect. 5.3 for an IF network with slow synapses, except that it involved single neuron
labels rather than the population labels , . As highlighted elsewhere [71, 168],
Eq. (6.4) can be reduced to a system of ordinary differential equations provided that
we place rather mild restrictions on the time dependence of (t). First, suppose
that (t) = w (t) where w denotes the mean synaptic strength of connections from population to neuron and (t) determines the time course of the
input, which is assumed to depend only on properties of the postsynaptic population
235
. Furthermore, suppose that there exists a differential operator L such that (see
also (5.76))
L (t) = (t).
(6.5)
Applying the operator L to both sides of equation (6.4) then leads to a system of
differential equations for the population currents u (t):
L u (t) =
w F (u (t)).
(6.6)
=1
Note that we could easily convert the synaptic current u (t) into an input voltage v (t) = u (t)/ using an effective input conductance , for example. Thus
Eq. (6.6) is often referred to as a voltage equation and forms the basis of most classical neural networks such as the Hopfield model [288]. On the other hand, if the time
course of the inputs depends only on presynaptic parameters, (t) = w (t),
with having inverse differential operator L , then we obtain a system of differential equations for the so-called synaptic drives:
z (t) =
t
(t t
)F (u (t
))dt
.
(6.7)
That is, applying the differential operator L to Eq. (6.7) and using u (t) =
P =1 w z (t) leads to the activity-based model

L z (t) = F
w z (t)
=1
(6.8)
The differential operator L appearing in Eqs. (6.6) and (6.8) is often taken to be
first order in time:
1
+ ,
L =
(6.9)
t
with inverse kernel (t) = H(t)et/ . In order to relate the effective time constant
to membrane and synaptic time constants, let us assume for simplicity that all
synapses are sufficiently close to the soma so that the dendrite simply acts as a
first-order low-pass filter and set (with Vrest = 0)
(t) = rm, Vsyn, g H(t)

with
h (s) =
t
0
e(ts)/m, h (s)ds,
d,
(es/d, es/r, ).
d, r,
We have made explicit that the reversal potential Vsyn and synaptic rise/fall times
r,d only depend on the particular class of synapses innervated by the presynaptic
population , whereas the membrane time constant m , resistance rm , and conductance are solely properties of the postsynaptic population . Only the maximum
236
conductance g is specific to the particular synapses . The various constant factors can be combined to define the synaptic weight w . In particular, w Vsyn,
so that the sign of Vsyn, (relative to the resting potential) determines whether the
synapse is excitatory or inhibitory. If m r , d , then the time course is effectively
independent of the presynaptic label , and we have the voltage-based model (6.6)
with first-order operator L and = m . On the other hand, if d m , r , then
we obtain the activity-based model with = d . Although the reduction to a ratebased model is a major simplification of the full conductance-based model, it is still
possible to incorporate various additional physiological features.
1. Synaptic depression. In order to incorporate synaptic depression we need to
return to the single neuron level. Equation (6.1) is modified according to
ui (t) =
1
N
j; ( j)=
t
(t T jm )qi j (t T jm )

1
=
(t t ) N

j; ( j)=
a j (t )qi j (t ) dt
with qi j evolving according to an equation of the form (1.45), which we write as

dqi j
1 qi j
=
(1 )qi j (t)a j (t).
dt
q
Averaging the latter equation with respect to j, ( j) = , and introducing the
compact notation
1
f (t) =
f j (t),
N j;
( j)=
we have the pair of equations
ui (t) =
and
t
(t t
)a(t
)qi (t
) dt
(6.10)
dqi
1 qi
=
(1 )a(t)qi(t) .
dt
q
We now make the mean-field approximation

a(t)qi (t) = a (t)qi (t)
(6.11)
Since all averaged depression variables qi (t) , i = 1, . . . , N} for fixed now
have a common input drive a (t), it follows that
d(qi (t) qi

(t) )
= [qi (t) qi
(t) ],
dt
237
and thus qi (t) qi

(t) = q (t) for all i, i
= 1, . . . , N. In other words,
after an initial transient of duration q , we can identify all depression variables associated with a given presynaptic population . If we now assume that
(t) = w (t), we can introduce the synaptic drives (6.7) and derive the
modified activity-based model [25, 617, 637, 644]:

P
w q (t)z (t)
L z (t) = F
=1
with
dq
1 q (t)
=
(1 )q (t)F
dt
q
(6.12)
w q (t)z (t)
=1
(6.13)
The corresponding voltage-based model is

P
w q (t)F (u (t)).
L u (t) =
with
(6.14)
=1
dq
1 q (t)
=
(1 )q (t)F (u (t)) .
dt
q
(6.15)
2. Axonal propagation delays. In the above derivation of rate-based models, we

have assumed that the spiking of a presynaptic neuron has an instantaneous effect
on downstream postsynaptic neurons. This neglects the fact that action potentials
take time to propagate along an axon to innervate a synaptic terminal. Let us
denote the corresponding axonal delay of a synapse by . The integral
equation (6.2) is modified according to
u (t) =
t
=1
(t t
)a (t
)dt
.
(6.16)
The corresponding voltage-based model then takes the form of a system of delay
differential equations,
L u (t) =
w F (u (t )),
(6.17)
=1
and similarly for the activity-based model.

3. Adaptive threshold dynamics. Another biophysical process that can be incorporated into rate-based models is spike frequency adaptation. Spike frequency
adaptation causes a neurons firing rate to decay to a submaximal level and
occurs when a potassium current, presumably activated by elevated intracellular
238
calcium, hyperpolarizes the membrane voltage [35, 396, 604]. This afterhyperpolarization current has a time constant of around 40120 ms. Spike frequency
adaptation can be introduced as a negative current ci on the right-hand side of
the conductance-based model equation (5.1). Assuming that ci (t) varies slowly
relative to the voltage vi (t), it can be shown that ci effectively acts as an adaptive
threshold that varies linearly with the firing rate [35]. In the case of a network
of homogeneous populations, each neuron within a given population fires at the
same mean rate so that we can identify ci = c for all i such that (i) = . Hence,
the voltage-based model becomes
with
w F (u (t) c (t)).
(6.18)
dc
c (t)
=
+ c F (u (t) c (t)).
dt
c
(6.19)
L u (t) =
=1
6.2 EI Oscillator Networks

One of the simplest population-based networks, which is often taken to be a fundamental module in large-scale models of cortex, is a pair of mutually coupled
local populations of excitatory and inhibitory neurons known as an EI network;
see Fig. 6.1. An EI network has the important property of exhibiting limit cycle
oscillations and can thus act as a basic oscillatory element in network models of
cortical phase waves, as an alternative to single spiking neurons (Sect. 5.2). An
activity-based version of an EI network takes the form (for first-order synapses)
daE
= aE + F (wEE aE wEI aI + hE )
dt
daI
= aI + F (wIE aE wII aI + hI ) ,
dt
hI
wEI
wEE
I
wII
wIE
hE
Fig. 6.1 Two-population EI network
(6.20)
239
where hE , hI represent constant background inputs. For simplicity, we set E = I =

1. The bifurcation structure of the two-population model given by Eq. (6.20) has
been analyzed in detail elsewhere [55]. An equilibrium (aE , aI ) is obtained as a
solution of the pair of equations
aE = F (wEE aE wEI aI + hE )
aI = F (wIE aE wII aI + hI ) .
(6.21)
These can be inverted to yield

hE = F 1 (aE ) wEE aE + wEI aI
hI = F 1 (aI ) wIE aE + wII aI .
(6.22)
As a further simplification, take the gain function F to be the simple sigmoid F(u) =
(1 + eu )1 . Using the fact that the sigmoid function then satisfies F = F(1
F) and applying the fixed-point equations allows one to represent the associated
Jacobian in the form

1 + wEE aE (1 aE ) wEI aE (1 aE )
=
.
wIE aI (1 aI )
1 wII aI (1 aI )
An equilibrium will be stable provided that the eigenvalues of have negative
real parts, where

1
2
=
(6.23)
Tr [Tr ] 4Det .
2
This leads to the stability conditions Tr < 0 and Det > 0. In order to construct
a phase diagram in the (hE , hI )-plane for a fixed weight matrix w, we express aI as
a function of aE by imposing a constraint on the eigenvalues and then substitute
the resulting function into Eq. (6.22). This yields bifurcation curves in the (hE , hI )plane that are parameterized by aE , 0 < aE < 1; see Fig. 6.2. For example, the
constraint
Tr 2 + wEE aE (1 aE ) wII aI (1 aI ) = 0
(6.24)
with Det > 0 determines Hopf bifurcation curves where a pair of complex conjugate eigenvalues cross the imaginary axis. Since the trace is a quadratic function
of aE , aI , we obtain two Hopf branches. Similarly, the constraint Det = 0 with
Tr < 0 determines saddlenode or fold bifurcation curves where a single real
eigenvalue crosses zero. The saddlenode curves have to be determined numerically, since the determinant is a quartic function of aE , aI .
Now consider a network of synaptically coupled EI modules or subnetworks,
with each module labeled by the discrete index n:

anE
danE
m
m
= + F [wEE (n, m)aE wEI (n, m)aI + hE ]
(6.25a)
dt
E
m
240
Fold
hI
Hopf
-5
Fold
-10
-6
-4
-2
hE
Fig. 6.2 Phase diagram of two-population WilsonCowan model (6.20) for fixed set of weights
wEE = 11.5, wIE = wEI = 10, , wII = 2. The dots correspond to TakensBogdanov bifurcation
points
an
danI
= I +F
dt
I

m
m
[w
(n,
m)a
w
(n,
m)a
+
h
]
,
II
I
IE
E
I
(6.25b)
Suppose that synaptic interactions within a local EI network are stronger than those
between EI networks. That is, we write
wab (n, m) = wab nm + Jab(n, m)(1 nm )
with 1. Substituting into (6.25) and Taylor expanding to O( ) then gives
danE
= anE + F ([wEE anE wEI anI + hE ])
(6.26a)
dt
m
+ F
([wEE anE wEI anI + hE ]) [JEE (n, m)am
E JEI (n, m)aI ]
m=n
danI
= anI + F ([wIE anE wII anI + hI ])
(6.26b)
dt
m
+ F
([wIE anE wII anI + hE ]) [JIE (n, m)am
E JII (n, m)aI ]
m=n
One can view the synaptic interaction between EI modules as weak perturbations of the underlying limit cycle oscillators, which suggests carrying out a phase
reduction of (6.26) along the lines of Sects. 1.2 and 5.2. In order to proceed, it is
first necessary to determine the phase-resetting curve (PRC) of an individual EI
network. Suppose that each uncoupled EI network operates in a parameter regime
where the mean-field equations (6.20) support a stable limit cycle. For concreteness,
take a point in parameter space between the two Hopf curves in Fig. 6.2, namely,
241
(hE , hI ) = (0, 4). A plot of the oscillation in phase space is shown in Fig. 6.3(a)
and the components ZE , ZI of the corresponding PRC are shown in Fig. 6.3(b). Note
that both components are approximately sinusoidal so that the EI network acts as
a type II oscillator.
activity
0.8
phase resetting curve
0.9
xI
xE
0.7
0.6
0.5
0.4
15
ZE
10
ZI
5
0
-5
-10
-15
phase
time t
Fig. 6.3 (a) Trajectories along the limit cycle of an EI network: xE (t) (solid curve) and xI (t)
(dashed curve). Parameters are wEE = 11.5, wIE = wEI = 10, wII = 2, hE = 0, and hI = 4.
Also F(u) = 1/(1 + eu ). (b) Corresponding components ZE and ZI of the phase-resetting curve
The next step is to rewrite (6.26) in the more compact form

dan
= an + f(an ) + g(an)T J(n, m)am ,
dt
m=n
(6.27)
where an = (anE , anI )T , f = ( fE , fI ), g = (gE , gI ),

f p (a) = F(w pE aE w pI aI + h p),
g p (a) = F
(w pE aE w pI aI + h p),
p = E, I.
Phase reduction now proceeds using the method of isochrones as described in

Sect. 1.2. Let (a) represent the isochronal mapping in a neighborhood of an uncoupled limit cycle oscillator and denote the natural frequency by 0 . Then
ZE ( ) =
(a ( ))
,
aE
ZI ( ) =
(a ( ))
,
aI
(6.28)
where a is a point on the limit cycle. Applying the phase reduction procedure to
(6.27) with n = (an ), we have
d n
= 0 + Z p (n )g p (n ) J pq (n, m)aq (m ).
dt
m
p,q=E,I
(6.29)
Here, all quantities are evaluated on the limit cycles so that g p (n ) = g p (a (n )) etc.
Finally, averaging over one period 0 = 2 /0 gives the phase equations

d n
= 0 + J pq (n, m)H pq (m n ) ,
(6.30)
dt
p,q=E,I m
242
with phase interaction functions

H pq ( ) =
1
2
2
0
Z p ( )g p ( )aq ( )d .
(6.31)
Given the phase equations, one can investigate the existence and stability of phaselocked states, including phase waves, along identical lines to the analysis of waves
on a chain in Sect. 5.2.2. Note that phase-coupled EI networks have been used to
study the phenomenon of stimulus-induced oscillations and synchrony in primary
visual cortex model, where each EI network is interpreted as a cortical column
consisting of reciprocally coupled populations of excitatory and inhibitory neurons
[239, 560].
6.3 Firing Rates, Asynchronous States, and Spiking Statistics

6.3.1 The Asynchronous State in a Homogeneous Spiking Network
One of the major assumptions in the derivation of rate-based population models in
Sect. 6.1 was that each homogeneous population is in an asynchronous state. Here
we consider the existence and stability of an asynchronous state in a large, globally
coupled network of IF neurons [2, 213]. Consider the following synaptically coupled
network of nonlinear IF neurons (see also Sect. 5.3):
dvi
= G(vi ) +
dt
N
j=1
(t
d ) (t t
T jm )dt
,
(6.32)
with given by the alpha function (5.23), d a discrete axonal delay and determines the strength of coupling. We take a threshold v = 1 and a reset vr = 0. In the
case of global coupling, we can express the sum over delta functions in terms of the
population activity variable
a(t) =
(t Tjm )
(6.33)
j=1 m
so that
dvi
= G(vi ) + u(t),
(6.34)
dt

where u(t) =
(t
d )a(t t
)dt
. Now suppose that there exists an asynchronous state a(t) = a0 . (Strictly speaking, such a state only exists in the thermodynamic limit N .) Since is normalized to unity it follows that u(t) = a0 as
well (ignoring transients). An implicit equation for a0 is then obtained by integrating
equation (6.34) between successive firing times:
1
=
a0
1
0
243
du
.
G(u) + a0
(6.35)
We will assume that there exists a unique solution to this equation for given G and .
If G depends on a uniform external input I, then this yields the population gain
function F with a0 = F(I).
In order to study the stability of the asynchronous state, it is convenient to carry
out the change of variables
yi = a 0
vi
0
du
,
G(u) + a0
(6.36)
with 0 < yi < 1 such that Eq. (6.34) becomes

dyi
= a0 + (yi )[u(t) a0]
dt
and
(y) =
a0
.
G(y) + a0
(6.37)
(6.38)
We also incorporate the effects of synaptic noise by including an additive white

noise term i (t),
dyi
= a0 + (yi )[u(t) a0] + i(t),
(6.39)
dt
with
i (t) = 0,
i (t) j (t
) = 2 i j (t t
).
(6.40)
(Note that diffusive fluctuations of the membrane potential due to stochastic background activity would lead to an additive white noise term in Eq.(6.34) rather than
in Eq. (6.40). The corresponding stochastic equation for yi would then involve multiplicative noise, which is much harder to analyze.) In the presence of noise the
variable yi can become negative so < yi < 1. The Langevin equation (6.40) has
an associated FokkerPlanck equation
p(y,t) = J(y,t),
t
y
(6.41)
where J(y,t) is the probability flux

J(y,t) = [a0 + (y)[u(t) a0]] p(y,t)
2
p(y,t).
2 y
(6.42)
This is supplemented by the boundary conditions arising from reset:

p(1,t) = 0,
p(0+ ,t) = p(0 ,t),
J(1,t) = a(t),
J(0+ ,t) J(0,t) = a(t).
(6.43)
(6.44)
244
We also require p(,t) = 0 and J(,t) = 0. The steady-state solution of the

FokkerPlanck equation is J(y,t) = a0 and p(y,t) = p0 (y) with
(
2
2
e2a0 y/ e2a0 (y1)/ , y < 0
p0 (y) =
.
(6.45)
2
0<y<1
1 e2a0 (y1)/ ,
The stability of the steady state can be determined by setting
p(y,t) = p0 (y) + (y)e t ,
a(t) = a0 + a1e t
(6.46)
and expanding to first order in , a1 . This gives the eigenvalue equation
(y) =
2 2
(y) a0 (y) a1 ( ) [ (y)p0 (y)],

2 y2
y
y
(6.47)
where ( ) is the Laplace transform
( ) =

0
(t d )e t =
2
e d .
( + )2
(6.48)
Defining the function

h(y) = a1 ( )
[ (y)p0 (y)],
y
we can write Eq. (6.47) as the inhomogeneous equation

2 2
a
(y) = 1] (y) + h(y; ).
[L (y)
0
2 y2
y
(6.49)
(6.50)
This inhomogeneous equation can be solved in terms of the associated onedimensional Greens function satisfying [L 1]G (y, y
; ) = (y y
) and
G (1, y
, ) = 0:
(y) =
1
G (y, y
; )h(y
; )dy
a1G (y, 0; )
(6.51)
with

A e+ ( )[y1] e ( )[y1] e ( )[y
1] , y
< y < 1

G (y, y
; ) =
A e ( )[y
1] e+ ( )[y
1] e+ ( )[y1] , y < y
,
where
2
1
,
2 + +

1
2
2
( ) = 2
a0 + 2 a0 .
A=
(6.52)
(6.53)
(6.54)
245
Note that the term a1 G (y, 0; ) ensures that the flux discontinuity at y = 0 is satisfied.
Finally, an implicit equation for the eigenvalues can be obtained by substituting
Eq. (6.49) into (6.51) and imposing the boundary condition J(1,t) = a(t), which
corresponds to the following first-order condition:
2
(y,t)|y=1 = a1 .
2 y
(6.55)
The resulting characteristic equation is [2]

e ( ) 1 = ( ) ( )
p0 (y) (y)e ( )y dy.
(6.56)
In the zero noise limit 0, we have ( ) /a0 and p0 (y) 1 for 0 <
y < 1 and is zero otherwise. Thus, Eq. (6.56) becomes
1

(y)e y/a0 dy.
e /a0 1 = ( )
a0
0
(6.57)
In the weak coupling regime, solutions of Eq. (6.57) are of the form = 2 ina0 +
n for integer n with n = O( ). The term n can be calculated by performing a
perturbation expansion in the coupling . The lowest-order contribution is simply
determined by setting = 2 ina0 on the right-hand side of Eq. (6.57). In the case
of a linear IF model with G(u) = I u, we have (y) = ey/a0 so that

2 ina0
n =
(2 ina0) + O( 2 ).
(6.58)
1 + 2 ina0
We then have the following stability results in the absence of noise [2, 213]:
(i) For zero axonal delays (d = 0) and excitatory coupling ( > 0), the asynchronous state is stable with respect to excitation of the nth mode if and only
if < n where

n = 1 + 1 + 4n2 2 a20
(6.59)
Hence, it is stable for sufficiently slow synapses, that is, < 1 . The asynchronous state is always unstable in the case of inhibitory coupling since the
condition for stability with respect to the nth harmonic is now > n , which
cannot be satisfied for all n.
(ii) The asynchronous state is almost always unstable for nonzero delays (in the
noise-free case).
(iii) For large n, |n | 1/n2 so that higher harmonics grow or decay slowly.
Note that although the zero delay case is a singular limit in the absence of noise,
it becomes non-singular for arbitrarily small amounts of noise, where instabilities
246
with respect to higher harmonics are suppressed [2, 213]. One finds that for sufficiently high noise levels the asynchronous state is always stable. Reducing the noise
for fixed delay induces an instability due to excitation of one of the harmonic modes
with frequency n = 2 na0. A bifurcation at 1 implies that the period of
the resulting collective oscillation is identical to the period of the individual oscillators. Higher harmonics correspond to instabilities of the asynchronous state that
lead to the formation of cluster states [213, 231]: each neuron fires with mean rate
a0 , but the population of neurons splits up into several groups that fire in sequence so
that the overall activity exhibits faster oscillations. Finally, note that fast oscillations
are also found in sparsely connected random networks [87].
6.3.2 Balanced Networks and Neural Variability

It is well known that the spike trains of individual cortical neurons in vivo tend
to be very noisy, having interspike interval (ISI) distributions that are close to
Poisson [179, 593]. Indeed, one observes trial-to-trial variability in spike trains,
even across trials in which external stimuli are identical. This raises a number of
important issues. First, neurons are continuously bombarded by thousands of synaptic inputs, many of which are uncorrelated, so that an application of the law of large
numbers would suggest that total input fluctuations are small. This would make it
difficult to account for the Poisson-like behavior of individual neurons, even when
stochastic ion channel fluctuations (Sect. 1.5) or random synaptic background activity is taken into account. For example, in the homogeneous spiking network analyzed above, the spike trains of individual neurons can be quite regular even when
the population activity is asynchronous. Conversely, irregularity in spiking at the
single-cell level can coexist with regular firing rates at the population level. However, there is growing evidence that noise can play a significant role at the network
level. For example, noise appears to be present during perceptual decision making
[656] and bistable perception, the latter being exemplified by perceptual switching
during binocular rivalry [436, 578, 663]; see Sect. 8. Noise also contributes to the
generation of spontaneous activity during resting states [148, 152]. At the level of
large-scale neural systems, as measured with functional MRI (fMRI) imaging, this
ongoing spontaneous activity reflects the organization of a series of highly coherent
functional networks that may play an important role in cognition. Another issue is
the possible computational role of pairwise and higher-order statistical correlations
between the spike trains of distinct neurons, that is, whether or not it is necessary to
go beyond firing rates.
One paradigm for exploring these various issues is the so-called balanced network [567, 645, 650]. In such networks, each neuron is driven by a combination
of strong excitation and strong inhibition, which mainly cancel each other out, so
that the remaining fluctuations occasionally and irregularly push the neuron over the
firing threshold. Even in the absence of any external sources of noise, the resulting
deterministic dynamics is chaotic and neural outputs are Poisson-like. Interestingly,
247
firing rates
spikes
there is some experimental evidence that cortical networks can operate in a balanced regime [391]. Another emergent feature of balanced networks is that they
can support an asynchronous state characterized by large variability in single neuron spiking and yet arbitrarily small pairwise correlations even in the presence of
substantial amounts of shared inputs [519]. Thus there is a growing consensus that
the trial-to-trial irregularity in the spiking of individual neurons is often unimportant, and that information is typically encoded in firing rates. (Such rates could be
a measure of population activity as in Sect. 6.1 or correspond to the rates of individual neurons modeled as inhomogeneous Poisson processes; see below.) There
is then another level of neural variability, namely, trail-to-trial variations in the firing rates themselves. Recent physiological data shows that the onset of a stimulus
reduces firing rate fluctuations in cortical neurons while having little or no effect
on spiking variability [119]; see Fig. 6.4. Litwin-Kumar and Doiron have recently
shown how these two levels of stochastic variability can emerge in a chaotic balanced network of randomly connected spiking neurons, in which a small amount of
clustered connections induces firing rate fluctuations superimposed on spontaneous
spike fluctuations [387].
stimulus onset
Fig. 6.4 Neurons can exhibit double stochasticity. Trial-to-trial variability in spike trains (shown
as raster plots) can coexist with trial-to-trial variability in the firing rate (shown as continuous gray
curves). Trial-averaged firing rate is the black curve. Following stimulus onset, the variability in
the firing rate is reduced
6.3.3 Spike Statistics and the Poisson Process

Given the ubiquity of Poisson processes in spiking neuron models and its relevance to neural master equations (see Sect. 6.4), it is useful to consider these processes in a little more detail. Suppose that a neuron emits a sequence of spikes at
times t1 ,t2 , . . . ,tn . One way to characterize this sequence statistically is in terms
of the probability density (t1 , . . .tn ) of finding such a sequence over many experimental trials. In other words, the probability of having a sequence of n spikes
in the interval [0, T ] with the ith spike falling between the times ti and ti + t
248
is P[t1 , . . .tn ] = (t1 , . . .tn ) t n . In principle, the probability of an event occurring,

namely, the firing of the next spike, could depend on the entire history of past spikes.
If this dependence only extends to the previous spike so that the interspike intervals
are statistically independent, then the stochastic process is said to be a renewal process. If there is no dependence at all on preceding spikes so that the firing times are
themselves independent, then we have a Poisson process.
Consider a sequence of spikes generated by a homogeneous Poisson process,
that is, one with a time-independent mean firing rate r. Divide a given time interval
T into M bins of size t = T /M and assume that T is small enough so that the
probability of finding two spikes within any one bin can be neglected. Then the
probability PT [n] of finding n spikes over the interval T is given by
PT [n] = lim
t0
M!
(r t)n (1 r t)Mn
(M n)!n!
This consists of the probability (r t)n of finding n spikes in n specific bins multiplied by the probability (1 r t)Mn of not finding spikes in the remaining bins.
The binomial factor is the number of ways of choosing n out of M bins with spikes.
Using the approximation M n M = T / t and defining = r t, we have that

rT
lim (1 r t)Mn = lim (1 + )1/
= erT .
t0
For large M, M!/(M n)! M n = (T / t)n , so that we obtain the Poisson distribution
(rT )n rT
e
PT [n] =
n!
Given that there are n independent spikes over the interval [0, T ], the probability that
these spikes lie within specified bins of size t is n!( t/T )n . Hence
n
1
(t1 , . . .tn ) = n!
PT [n] = rn erT
(6.60)
T
Also note that the fastest way to generate a sequence of Poisson spikes for constant r
is to iterate the firing times tn+1 = tn log(xrand )/r with xrand uniformly distributed
over [0, 1].
A simple method for calculating the moments of the Poisson distribution is to
introduce the moment-generating function
g(s) =
PT [n]esn
n=0
Differentiating with respect to s shows that

d k g(s)
= nk
dsk s=0
249
The generating function for the Poisson process can be evaluated explicitly as
g(s) = exp(rT ) exp(rT es )
from which we deduce that
n = rT,
n2 = rT
Another useful quantity is the interspike interval (ISI) distribution. Suppose that
a spike was last fired at time T n . The probability of a homogeneous Poisson process
generating the next spike in the interval T n + T n+1 T n + + is equal to
the probability that no spike is fired for a time , which is P [0] = er multiplied
by the probability r of generating a spike within the following interval :
Pr[ T n+1 T n + ] = r er
The ISI probability density is thus an exponential, ( ) = rer . It follows that the
mean interspike interval is
=
and the variance is

0
rer d =
1
r
1
r2
The ratio of the standard deviation to the mean is called the coefficient of variation
2 =
rer 2 d 2 =
CV =

It follows that for a homogeneous Poisson process CV = 1.

How well do Poisson statistics describe experimentally measured spike trains?
One often finds that for ISIs longer than about 10 msec, the ISI distribution is indeed
approximately exponential. However, for shorter intervals, there is a rapid decrease
in the distribution reflecting the fact that neurons are refractory immediately after
Fig. 6.5 (a) Interspike interval distribution from a neuron from the MT cortical visual area of a
macaque monkey responding to a moving image. (b) Interspike interval generated with a Poisson
model with a stochastic refractory period. Redrawn from [20]
250
firing. This is illustrated in Fig. 6.5. The data can be fitted more accurately by a
gamma distribution
r(r )k et
(6.61)
p[ ] =
k!
Alternatively, one can introduce a refractory period into a standard Poisson model.
Note that CV values extracted from the spike trains of cortical neurons also take
values around unity, provided that the mean interspike interval is not too small [593].
It is possible to generalize the above Poisson model to the case of a timedependent rate r(t). The simplest way to analyze this inhomogeneous Poisson process is to consider the probability distribution (t1 , . . .tn ). This is given by the
product of the probabilities r(ti ) t that the neuron fires within the time intervals
ti t ti + t and the probabilities of not firing during the interspike intervals.
The latter is given by
Pr[no spikes in (ti ,ti+1 )] =
(1 r(ti + m t) t)
m=1
where we have partitioned the interval (ti ,ti+1 ) into M bins of size t. Taking the
logarithm,
log Pr[no spikes in (ti ,ti+1 )] =
log(1 r(ti + m t) t)
m=1
r(ti + m t) t
m=1
Taking the limit t 0 and exponentiating again shows that

t

i+1
Pr[no spikes in (ti ,ti+1 )] = exp
r(t)dt
ti
Hence

(t1 , . . .tn ) = r(ti ) exp
n
i=1
r(t)dt
(6.62)
In the case of a time-dependent rate, one generates xrand at each time step and a
spike is fired if r(t) t > xrand .
6.4 Stochastic Population Models

The discussion of neural variability in Sect. 6.3 motivates the incorporation of noise
directly into rate-based models, without explicitly modeling spike variability. One
approach is to introduce noise into a rate-based network model using extrinsic noise
sources [21, 84, 184, 297]. An alternative approach is to assume that noise arises intrinsically as a collective population effect. A number of methods involve carrying
out some form of dimension reduction of a network of synaptically coupled spiking
neurons. These include population density methods [395, 468, 477], mean-field the-
251
ories [21, 86, 87], and Boltzmann-like kinetic theories [90, 100, 517]. However, such
methods tend to consider either fully connected or sparsely connected networks and
simplified models of spiking neurons such as the integrate-and-fire (IF) model. Nevertheless, as discussed in Sect. 6.3, one important feature of spiking networks is that
they can support an asynchronous state in which individual neurons exhibit Poissonlike statistics, whereas the total population activity can be expressed in terms of a
characteristic activation or gain function [2, 86, 87, 213, 519]. Formally speaking,
the asynchronous state only exists in the thermodynamic limit N , where N determines the size of the population. This then suggests a possible source of intrinsic
noise at the network level arises from fluctuations about the asynchronous state due
to finite size effects [57, 220, 412, 425, 597]; this is distinct from intrinsic noise
at the single-cell level due to channel fluctuations and it is assumed that the latter
is negligible at the population level. The presence of finite size effects has motivated the development of a closer analogy between intrinsic noise in biochemical
networks (including ion channel kinetics) and neural networks [68, 69, 72], based
on extensions of the neural master equation introduced by Buice et al. [91, 92]; see
also [473]. In this section, we describe the master equation framework for stochastic
neural population dynamics.
6.4.1 Neural Master Equations

One way to incorporate intrinsic noise at the population level is to treat the output
activity of a local homogeneous population as a discrete stochastic variable A (t)
rather than the instantaneous firing rate a = F(u ) [68, 91, 92]:
A (t) =
N (t)
,
N t
(6.63)
where N (t) is the number of neurons in the th population that fired in the time
interval [t t,t] and t is the width of a sliding window that counts spikes. The
discrete stochastic variables N (t) are taken to evolve according to a one-step jump
Markov process:
N (t) N (t) 1 :
transition rate (U (t), N (t)),
(6.64)
with the synaptic current U (t) given by (for exponential synapses)

dU (t) = U (t) +
w A (t)
dt,
(6.65)
=1
where, for convenience, we have rescaled the weights according to w w / .

The transition rates are taken to be (cf. [68])
252
+ (u , n ) + (u ) =
N t
F(u ),
a
(u , n ) (n ) =
n
.
a
(6.66)
We see that the transition rate + depends on U , with the latter itself coupled to
the associated jump Markov according to Eq. (6.65), which is only defined between
jumps, during which U (t) evolves deterministically. Hence, the resulting stochastic
process defined by Eqs. (6.63), (6.64), (6.65) and (6.66) provides an example of a
stochastic hybrid system at the population neuron level, just as stochastic ion channel gating provides an example at the molecular level (Sect. 1.5). It is important to
note that the time constant a cannot be identified directly with membrane or synaptic time constants. Instead, it determines the relaxation rate of a local population to
the instantaneous firing rate.
A. Case 0 (fast synapses). In the limit 0, Eq. (6.65) implies that the continuous variables U (t) can be eliminated by setting U (t) = w A (t). This
then leads to a pure birthdeath process for the discrete variables N (t). That is,
let P(n,t) = Prob[N(t) = n] denote the probability that the network of interacting
populations has configuration n = (n1 , n2 , . . . , nM ) at time t,t > 0, given some initial distribution P(n, 0) with 0 n N. The probability distribution then evolves
according to the birthdeath master equation [68, 91, 92]

+

dP(n,t)
= (T 1) (n)P(n,t) + (T1
1) (n)P(n,t) , (6.67)
dt
where
N t
+ (n) =
F
a
w n /N t
(n) =
n
,
a
(6.68)
and T is a translation operator: T1

F(n) = F(n ) for any function F with n
denoting the configuration with n replaced by n 1. Equation (6.67) is supplemented by the boundary conditions P(n,t) 0 if n = N + 1 or n = 1 for some
. The neural master equation (6.67) has two versions depending on the choice of
t. First, fixing t = 1 leads to the Bressloff version of the master equation. For
large but finite N, the master equation (6.67) can be approximated by a Fokker
Planck equation using a KramersMoyal or system-size expansion, so that the population activity A evolves according to a Langevin equation [68]. Introduce the
rescaled variables a = n /N and

corresponding transition rates (a) = a /a
+
1
and (a) = F w a . Carrying out a KramersMoyal expansion to second order then leads to the multivariate FP equation
M
P(a,t)
1 M 2
=
[A (a)P(a,t)] +
[B (a)P(a,t)] (6.69)
2
t
2N
=1 a
=1 a
with
A (a) = + (a) (a),
B (a) = + (a) + (a).
(6.70)
253
The solution to the FokkerPlanck equation (6.69) determines the probability density function for a corresponding stochastic process A(t)=(A1 (t), . . . , AM (t)), which
evolves according to a neural Langevin equation or SDE of the form
1
dA = A (A)dt + b (A)dW (t).
N
(6.71)
with b (x)2 = B (x). Here W (t) denotes an independent Wiener process such that
dW (t) = 0,
dW (t)dW (s) = , (t s)dtds.
(6.72)
In the thermodynamic limit N , we recover the activity-based mean-field

equation
da
= A (a) = a (t) + F( w a (t)).
(6.73)
dt
Note that the multiplicative noise in Eq. (6.71) is interpreted in the sense of Ito,
which follows from the form of the FP equation (6.69); see Sect. 1.7.
A rigorous probabilistic treatment of the thermodynamic limit of the neural master equation has also been developed [89], extending previous work on chemical
master equations [351]. However, as we noted within the context of stochastic ion
channels in Sect. 1.5, the diffusion approximation breaks down when considering
noise-induced transitions between multistage states. It is then necessary to use a
WKB approximation of the master equation together with matched asymptotics
[69]. The second version of the neural master equation, which was introduced by
Buice et al. [91, 92], is obtained by taking the limit N , t 0 such that
N t = 1. In this case there is no small parameter that allows one to construct a
Langevin approximation to the master equation. Nevertheless, it is possible to determine the moment hierarchy of the master equation using path integral methods or
factorial moments, based on the observation that the network operates in a Poissonlike regime. The role of the sliding window size t is crucial in understanding the
difference between the two versions of the master equation. First, it should be emphasized that the stochastic models are keeping track of changes in population spiking activity. If the network is operating close to an asynchronous state for large N,
then one-step changes in population activity could occur relatively slowly, so there
is no need to take the limit t 0. On the other hand, if population activity is characterized by a Poisson process, then it is necessary to take the limit t 0 in order
to maintain a one-step process. However, given the existence of an arbitrarily small
time scale t, it is no longer clear that one is justified in ignoring synaptic dynamics
by taking the limit 0 in Eq. (6.65).
B. Case a > 0 (stochastic hybrid system). Now suppose that > 0 in
the full stochastic model given by (6.63)(6.66), with N , t 0 such that
N t = 1. Denote the random state of the full model at time t by {(U (t), N (t)); =
1, . . . , M}. Introduce the corresponding probability density
254
Prob{U (t) (u , u + du), N (t) = n ; = 1, . . . , M} = p(u, n,t)du,
(6.74)
with n = (n1 , . . . , nM ) and u = (u1 , . . . , uM ). It follows from Eqs. (6.63)(6.66) that

the probability density evolves according to the ChapmanKolmogorov (CK) equation [82]
p 1 [v (u, n)p(u, n, t)]

+
(6.75)
t
u

1
= (T 1) ( (n )p(u, n,t)) + (T1
1) (+ (u )p(u, n,t)) ,
a
with
+ (u ) = F(u ),
(n ) = n ,
v (u, n) = u + w n .
(6.76)
Equation (6.75) can be rewritten in the more compact form (see also (1.143))
1 M
p
1
=
(v (u, n)p(u, n,t)) + A(n, m; u)p(u, m,t).
t
=1 u
a m
(6.77)
The drift velocities v (u, n) for fixed n represent the piecewise deterministic
synaptic dynamics according to
du
= v (u, n),
dt
= 1, . . . , M,
(6.78)
and A represents the u-dependent transition matrix for the jump Markov process.
For fixed u, the matrix A(n, m; u) is irreducible (which means that there is a
nonzero probability of transitioning, possibly in more than one step, from any state
to any other state in the jump Markov process) and has a simple zero eigenvalue.
In particular, n A(n, m; u) = 0 for all m, that is, n = (1, 1, . . . , 1)T is the left null
vector of A. The PerronFrobenius theorem (for large but finite N) ensures that all
other eigenvalues of A are negative and the continuous-time Markov process for
fixed u,
1
d p(u, n,t)
=
A(n, m; u)p(u, m,t),
dt
a mI
has a globally attracting steady-state (u, n) such that p(u, n,t) (u, n) as t .
The steady-state equation is
0 = A(n, m; u) (u, m)
=
m
M
[(n + 1) (u, n + e ) n (u, n) + F(u )( (u, n e ) (u, n))] ,
=1
255
where [e ] = , . The solution can be factorized as (u, n) = M

=1 1 (u , n )
with

0=
1(u , n )
=1 =
where
[J(u , n + 1) J(u , n )] ,
J(u, n) = n1 (u, n) F(u)1 (u, n 1).
Since 1 (u, 1) 0, it follows that J(u, n) = 0 for all n. Hence,

n
F(u)
F(u)n
= (u, 0)
,
n!
m=1 m
1 (u, n) = (u, 0)
(6.79)
and the corresponding normalized density is a Poisson process with rate F(u)
1 (u, n) = eF(u)
F(u)n
.
n!
(6.80)
There are two time scales in the CK equation (6.75), the synaptic time constant
and the time constant a , which characterizes the relaxation rate of population activity. In the limit 0 for fixed a , Eq. (6.75) reduces to the neural master equation
(6.67) with u = u(n) such that v (u(n), n) = 0. On the other hand, if a 0 for
fixed , then we obtain deterministic voltage- or current-based mean-field equations
du
= v (u(t)) v (u(t), n) (u(t), n)
dt
n
= u (t) +
w n (u(t), n).
=1
(6.81)
Since (u, n) is given by product of independent Poisson processes with rates

F(u ), consistent with the operating regime of the Buice et. al. master equation
[91, 92], it follows that
n = F(u ),
(6.82)
and (6.81) reduces to the standard voltage- or current-based activity equation. Now
suppose that the network operates in the regime 0 < a / 1, for which there
are typically a large number of transitions between different firing states n while
the synaptic currents u hardly change at all. This suggests that the system rapidly
converges to the (quasi) steady state (u, n), which will then be perturbed as u
slowly evolves. The resulting perturbations can be analyzed using a quasi-steadystate (QSS) diffusion or adiabatic approximation, in which the CK equation (6.75)
is approximated by a FokkerPlanck equation [82]; see Sect. 1.6. However, when
considering escape from a metastable state, it is necessary to use the WKB method
outlined in Sect. 1.6, as will be illustrated below.
256
6.4.2 Metastability in a One-Population Model

One of the useful features of the master equation formulation of stochastic population dynamics is that one can apply methods previously used to analyze jump
Markov processes at the cellular level. We will illustrate this by considering a
first passage time (FPT) problem for a one-population version of the CK equation
(6.75), which exhibits bistability in the deterministic limit [82]. This is analogous
to the problem of spontaneous action potential generation presented in Sect. 1.6.
In the case of a single homogeneous population of excitatory neurons, (6.75)
becomes
p [v(u, n)p(u, n,t)] 1

+
= A(n, m; u)p(u, m,t)
t
u
m
(6.83)
with drift term

v(u, n) = u + n,
(6.84)
A(n, n; u) = F(u) n,
A(n, n + 1; u) = n + 1. (6.85)
and tridiagonal transition matrix

A(n, n 1; u) = F(u),
As shown in Sect. 6.4.1, the steady-state density for A is given by a Poisson process,
(u, n) =
[F(u)]n eF(u)
,
n!
(6.86)
and the mean-field equation obtained in the 0 limit is
du
= v(u, n) (u, n) = u + F(u).
dt
n=0
(6.87)
If F is given by the sigmoid (1.14), it is straightforward to show that (6.87) exhibits

bistability for a range of thresholds and gains; see Fig. 6.6.
The CK equation (6.83) is identical in form to (1.143) under the change of variables u v, where v is voltage, and v(u, n) F(v, n). It follows that the general
analysis developed in Sect. 1.6 can be applied directly to the stochastic neural population model. In particular, the escape rate 0 from the low activity state u is given
by (1.191). As in the case of stochastic ion channels, there are three basic steps
needed in order to evaluate 0 using the particular form of the drift v and transition
matrix A [72]:
(1) Find the unique nontrivial positive eigenfunction n (u) = R(0) (u, n) and associated eigenvalue (u) = 0
(u). In the case of the neural population model,
Eq. (1.171) takes the explicit form
F(u)n1 (u) (F(u) + n)n(u) + (n + 1)n+1(u) = (u + n)n(u).
(6.88)
257
2.5
u+
2
1.5
F(u)
=1
u0
= 0.8
u-
= 0.6
0.5
0
0.5
1.5
2.5
Fig. 6.6 Bistability in the deterministic network satisfying u = u + F(u) with F given by the
sigmoid (1.14) for gain = 4 and threshold = 1.0, F0 = 2. There exist two stable fixed points
u separated by an unstable fixed point u0 . As the threshold is reduced the network switches to
a monostable regime
Trying a solution for of the form
n (u) =
(u)n
n!
(6.89)
yields the following equation relating and the corresponding eigenvalue :

F(u)
1 n + F(u) = (u)(u + n).
We now collect terms independent of n and linear in n, respectively, to obtain the

pair of equations

F(u)
=
1 , = F(u) u.
We deduce that
= u,
F(u)
1 ,
u
(6.90)
and the normalized eigenfunction is
n (u) =
un u
e .
n!
(6.91)
Note that (u) vanishes at the fixed points u , u of the mean-field equation (6.87)
with (u) > 0 for 0 < u < u and (u) < 0 for u < u < u . Moreover, comparing
Eq. (6.86) with (6.91) establishes that n (u) = (u, n) at the fixed points u , u . In
conclusion R(0) (u, n) = n (u) and the effective potential 0 is given by
258
0 (u) =
u
u
(y)dy.
(6.92)
The effective potential is defined up to an arbitrary constant, which has been fixed
by setting 0 (u ) = 0.
(2) Determine the null eigenfunction n (u) = S(u, n) of Eq. (1.174), which becomes
F(u)m+1 (F(u) + m)m + mm1 = (u)[u + m]m.
(6.93)
Trying a solution of the form m = m yields

(F(u)) (F(u) + m) + m 1 = (u)[u + m].
is then determined by canceling terms linear in m, which finally gives

n
u
n (u) =
.
F(u)
(6.94)
(6.95)
, n) of Eq. (1.186), which

(3) Calculate the generalized eigenvector n = S(u
reduces to
(6.96)
F(u )n+1 + nn1 (F(u ) + n)n = u n.
It is straightforward to show that this has the solution n = n (up to an arbitrary constant that does not contribute to the principal eigenvalue). It follows from Eq. (6.86)
that the factor B(u ) defined by (1.190) is
B(u ) =
(u, n)
u n + n2
n=0

= un + n2 .
(6.97)
Recall that (u, n) is given by a Poisson density with rate F(u), which implies that
n2 = n + n2 with n = F(u). Therefore,
B(u ) = F(u ) [2F(u ) u] ,
(6.98)
which reduces to B(u ) = F(u )2 since u = F(u ) at a fixed point.

It is instructive to compare the effective potential 0 obtained using the WKB
approximation with the potential obtained using the FP equation (1.146) based on
the QSS approximation. First, substitute (6.86), (6.84), and (6.85) into Eqs. (1.147)
and (1.148) under the change of variables u v and v(u, n) F(v, n). We find that
Z(u, n) = n (u, n) so that
D(u) = [un + wn2] = B(u).
(6.99)
The steady-state solution of the FP equation (1.146) takes the form C(u)
exp0 (u)/ with stochastic potential
0 (u) =
259
u
F (y)
D(y)
dy =
u
y + F(y)
dy.
wF(y)[2F(y) y]
(6.100)
Note that 0 differs from the potential 0 , Eq. (6.92), obtained using the more
accurate WKB method. Equations (6.90) and (6.92) show that the latter has the
integral form
u
1 wF(y)
1 dy.
(6.101)
0 (u) =
w
y
Thus, there will be exponentially large differences between the steady states for
small .
Fig. 6.7 Comparison of the double-well potentials 0 (u) and 0 (u) obtained using the quasistationary approximation and the QSS diffusion approximation, respectively [82]. Parameter values
are chosen so that deterministic network is bistable: F0 = 2.3, = 4, and = 1
In Fig. 6.7, we plot the potential function 0 of Eq. (6.101), which is obtained
using the quasistationary approximation in a parameter regime for which the underlying deterministic network is bistable. We also plot the corresponding potential
function 0 of Eq. (6.100), under the QSS diffusion approximation. The differences
between the two lead to exponentially large differences in estimates for the mean
exit times when is small. The mean exit time from the left and right well is shown
in Fig. 6.8. Solid curves show the analytical approximation T 1/0, where 0 is
given by (1.191), as a function of 1/ . For comparison, the mean exit time computed from averaged Monte Carlo simulations of the full stochastic system is shown
as symbols. As expected, the log of the mean exit time is an asymptotically linear
function of 1/ , and this is confirmed by Monte Carlo simulations. The slope is
determined by the depth of the potential well, and the vertical shift is determined
by the prefactor. Also shown is the corresponding MFPT calculated using the QSS
diffusion approximation (dashed curves), which is typically several orders of magnitude out and validates the relative accuracy of the quasistationary approximation.
One of the applications of noise-induced transitions between metastable states
in a stochastic population model is to the study of switching between up and down
states during slow-wave sleep; see Sect. 5.1. For example, Holcman and Tsodyks
260
Fig. 6.8 Mean exit time from the left and right well calculated using the quasistationary approximation (solid line) and the QSS diffusion approximation (dashed line) [82]. The open circles
represent data points obtained by numerically solving the corresponding jump velocity Markov
process using the Gillespie algorithm. Parameter values are the same as in Fig. 6.7
[282] consider an extension of the deterministic one-population model (6.87) by

including the effects of synaptic depression and extrinsic noise. The voltage-based
model equations take the form
du
= u + qw F(u) + (t) + I(t)
dt
1q
dq
=
qF(V ),
dt
q
(6.102a)
(6.102b)
where q(t) is the depression variable, (t) is a white noise term, I(t) is an external
input, and the constant is known as a utilization parameter. The firing rate function
is taken to be a linear threshold function F(V ) = [u ]+. In the absence of noise
( = 0) and external inputs (I = 0), the deterministic system undergoes a series
of bifurcations as the weight w is increased. For sufficiently small w, there exists
a globally attracting stable fixed point which is a low activity or down state. At a
critical value of w, a saddle and unstable node appear, with the latter undergoing a
subcritical Hopf bifurcation at a second critical value of wthis generates a stable
up state inside an unstable limit cycle. The system then exhibits bistability such
that noise-induced transitions from the down to the up state have to cross both the
separatrix of the saddle and the unstable limit cycle. The resulting fluctuations in
activity are suggestive of slow-wave oscillations observed in cortical slices [549].
6.4.3 Metastability in a Two-Population Model

It turns out that the analysis of metastability in a one-population model can be extended to higher-dimensional networks [72]. Consider, for example, the EI network
of Sect. 6.2. The corresponding CK equation (6.75) becomes
261
1
p
= (vp) (vp)
+ A(n, m; x)p(x, n)
t
x
y
m
(6.103)
where x = (x, y), n = (nx , ny ), the drift terms are

v(x, nx , ny ) = x + [wEE nx wEI ny ] ,
(6.104)
v(y,
nx , ny ) = y + [wIE nx wII ny ] ,
(6.105)
and A has the nonzero components

A(nx , ny , nx 1, ny ; x) = F(x),
A(nx , ny , nx , ny 1; x) = F(y),
(6.106)
A(nx , ny , nx + 1, ny ; x) = nx + 1,
A(nx , ny , nx , ny + 1; x) = ny + 1,
(6.107)
and
A(nx , ny , nx , ny ; x) = [F(x) + F(y) + nx + ny ].
(6.108)
Here x and y denote the excitatory and inhibitory population variables aE , aI , respectively. In the limit 0, we recover the deterministic equations (6.20). However,
in contrast to Sect. 6.2, it is now assumed that the EI network operates in a bistable
regime as illustrated in Fig. 6.9.
2.5
-0.8
0.8
1.6
2.4
3.2
4.8
5.6
Fig. 6.9 Bistability in an EI network. The x-nullcline through the saddle is its stable manifold
and acts as the separatrix between the two stable fixed points. Two deterministic trajectories
are shown (directed black curves), starting from either side of the unstable saddle and ending at
a stable fixed point. Parameter values are F0 = 1, = 3, = 2, wEE = 5, wEI = 1, wIE = 9, and
wII = 6
Again the MFPT can be identified as the inverse of the principal eigenvalue 0
of the associated linear operator on the right-hand side of (6.103). However, now
the analysis is more complicated, since one has to consider stochastic trajectories
crossing different points along the separatrix between the two metastable states.
262
Nevertheless, the various steps in the calculation of 0 proceed along similar lines
to the one-population model. That is, a spectral projection method can be used to
express 0 in terms of the inner product of a quasistationary density with a corresponding adjoint eigenfunction. The former can be calculated using a WKB approximation, except that now one has to determine the stochastic potential 0 by solving
an equation of the form [82]
H xPx yPy F(x) F(y) + x (x, Px , Py ) + y (y, Px , Py ) = 0, (6.109)
where
Px =
0
x
Py =
0
,
y
(6.110)
y =
F(y)
1 + wEI Px + wII Py
(6.111)
and
x =
F(x)
,
1 wEE Px wIE Py
Mathematically speaking, Eq. (6.109) is identical to a stationary HamiltonJacobi

equation for a classical particle with H identified as the Hamiltonian (see also
Sect. 4.4). A trajectory of the particle is given by the solution of Hamiltons
equations
H
H
dy
dx
=
=
,
,
dt
Px
dt
Py
dPy
H
H
dPx
=
,
=
dt
x
dt
y
(6.112)
Here t is treated as a parameterization of trajectories rather than as a real-time variable. Given a solution curve (x(t), y(t)), known as a ray, the potential 0 can be
determined along the ray by solving the equation
0 dx 0 dy
dx
dy
d 0
+
= Px + Py .
dt
x dt
y dt
dt
dt
(6.113)
Thus, 0 can be identified as the action along a zero-energy trajectory. One numerically solves for 0 by considering Cauchy data in a neighborhood of the stable fixed
point (x , y ) [82].
The rays (x(t), y(t)) (i.e., solutions to Hamiltons equations (6.112) in the (x, y)
plane) have an important physical meaning. The trajectory of the ray is the most
likely trajectory or path leading away from a point in the neighborhood of a stable
fixed point [160]. The rays shown in Fig. 6.10 are obtained by integrating Hamiltons equations (6.112) . These trajectories are only valid in one direction: away
from the stable fixed points. For parameter values considered in Fig. 6.10, rays originating from the neighborhood of each stable fixed point cover separate regions, so
that the most likely paths between points in each region are connected by deterministic trajectories starting at the boundary between the two regions. Note that this
boundary is not the separatrix (gray curve). For example, a trajectory initially at the
left fixed point which crosses the separatrix at the saddle would most likely follow
263
Fig. 6.10 (a) Characteristic paths of maximum likelihood for the 2D model. Rays originating from
the left (right) stable fixed point are shown in orange (cyan), with the ray connecting to the saddle
shown in red (blue). The gray curve is the separatrix . Level curves of constant 0 (x, y) are
shown as black dots. Each ray has four dots for different values of 0 (x, y). Rays originating from
the left fixed point have dots at = 0.1, 0.2, + 0.01, + 0.02, and rays originating from the
right fixed point have dots at 0 = 0.19, 0.23, 0.28, 0.30, where = (x , y ) = 0.28. All rays
terminate at 0 = + 0.02. (b) Sample trajectories of the two-population velocity jump Markov
process, whose associated probability density evolves according to (6.83), are computed using the
Gillespie algorithm with = 0.05 and N t = 1. (The maximum likelihood paths are independent
of .) Other parameter values are the same as in Fig. 6.9 [72]
a ray towards the saddle and then follow a deterministic trajectory to the right fixed
point. If a trajectory crosses the separatrix away from the saddle, it is most likely to
cross the separatrix above the saddle when starting from the left fixed point and below the saddle when starting from the right fixed point. In Fig. 6.11, the probability
density function for the y coordinate of the point on the separatrix reached by an exit
trajectory is shown for each well (square symbols show the histogram for exit from
the left well and likewise, o symbols for the right well). Each density function is
264
Fig. 6.11 The probability density for the exit point (y coordinate) where the separatrix is crossed by
an exiting trajectory. Results are obtained by 102 Monte Carlo simulation with the same parameters
as used in Fig. 6.9, with = 0.08. The square symbols show trajectories from the left well, and o
symbols show trajectories from the right well
peaked away from the saddle point, showing a phenomena known as saddle point
avoidance [398, 559]. As 0, the two peaks merge at the saddle point. Although
one might expect the saddle point to be the most likely exit point, since it is the point
on the separatrix where the potential 0 takes its minimum value, Fig. 6.11 shows
that this is not necessarily true. Even though the most likely exit point is shifted
from the saddle, the value of the potential around the saddle point still dominates
the mean first exit time.
6.5 Spatially Structured Networks and Neural Fields

So far we have not made any assumptions about the topology of the underlying neural network, that is, the structure of the weight matrix W with components wi j . If one
looks at a region of cortex such as primary visual cortex (V1), one finds that it has
a characteristic spatial structure, in which a high density of neurons (105 per mm3
in primates) are distributed according to an approximately two-dimensional (2D)
architecture. That is, the physical location of a vertical column of neurons within
the two-dimensional cortical sheet often reflects the specific information processing
role of that population of neurons. For example, in V1 there is an orderly retinotopic mapping of the visual field onto the cortical surface, with left and right halves
of the visual field mapped onto right and left visual cortices, respectively. Superimposed upon this are additional two-dimensional maps reflecting the fact that neurons
respond preferentially to stimuli with particular features such as local orientation
[613]. (A more detailed description of the functional architecture of V1 is given in
Sect. 8.1.) This suggests labeling neurons according to their spatial location in cortex. We now give a heuristic argument for how such labeling leads to a continuum
neural field model of cortex, following along similar lines to Gerstner and Kistler
[214].
265
For simplicity, consider a population of neurons distributed along a

one-dimensional axis. (Extensions to higher dimensions proceed in a similar fashion.) Suppose that we partition space into segments of length d such that the number
of neurons in segment [nd, (n + 1)d] is N = d where is the cell density. We treat
neurons in that interval as a homogeneous population of cells (cortical column)
labeled by the integer n and assume that synaptic interactions between the nth and
mth populations only depend on the discrete locations of the populations on the
line. Writing nm (t) = d (nd, md,t) and un (t) = u(nd,t), Eq. (6.4) becomes
u(nd,t) = d
m
t
(nd, md,t t
)F(u(md,t
))dt
.
Taking the limit d 0, the summation on the right-hand side can be replaced by an
integral to give
u(x,t) =
t

(x, y,t t
)F(u(y,t
))dt
dy,
(6.114)
where we have absorbed the factor into . Following our derivation of the discrete
voltage-based model (6.6), suppose that we can decompose the integral kernel as
(x, y,t) = w(x, y) (t),
(t) = et/ H(t).
That is, we assume that there is only one type of neuron so that the temporal kernel
(t) is independent of the presynaptic label y and the postsynaptic label x. Applying
the differential operator Lt = t + 1 to the integral equation for u then leads to the
scalar neural field equation
u(x,t)
+
u(x,t) =
t
w(x, y)F(u(y,t))dy.
(6.115)
Alternatively, we could have applied the differential operator Lt to the correspondt

ing synaptic drive z(x,t) =
(t t
)F(u(x,t
))dt
to obtain the activity-based
neural field model

z(x,t)
z(x,t) =
+F
w(x, y)z(y,t)dy .
(6.116)
t
Following the same basic procedure, it is straightforward to incorporate into the

neural field equation (6.115) or (6.116) additional features such as synaptic depression [329331], adaptive thresholds [135, 329], and axonal propagation delays
[16, 139, 296, 307, 382, 528, 603]. For example, a voltage-based neural field equation with synaptic depression takes the form
u(x,t)
u(x,t) =
+
w(x, y)q(y,t)F(u(y,t))dy,
t
1 q(x,t)
q(x,t) =
q(x,t)F (u(x,t)) ,
dt
q
(6.117)
266
with = 1 . In the case of axonal delays, i j (x, y) in the continuum limit.

Assuming that an action potential propagates with constant speed v along the axon,
then (x, y) = |x y|/v so that the voltage-based equation (6.115) becomes
u(x,t)
u(x,t) =
+
t
w(x, y)F(u(y,t |x y|/v))dy.
(6.118)
Two-dimensional versions of these various models are obtained by taking x x =

(x1 , x2 ) and y y = (y1 , y2 ) with dy = dy1 dy2 .
Now suppose that there are M classes of neuron distributed along the line labeled
by the population index a = 1, . . . M. Equation (6.114) then generalizes to the multipopulation integral equation
ua (x,t) =
t
ab (x, y,t t
)Fb(ub (y,t
|x y|/vab))dt
dy. (6.119)
b=1
We have included axonal delays with vab the conduction velocity along axonal projections from neurons of type b to neurons of type a. Assuming that ab (x, y,t) =
wab (x, y) (t) with (t) = et/ H(t), we obtain multi-population neural field
equations:
M
ua (x,t)
ua
+
=
t
b=1
and
za (x,t)
za
=
+ Fb
t
wab (x, y)Fb (ub (y,t |x y|/vab))dy,
b=1
(6.120)

wab (x, y)zb (y,t |x y|/vab )dy
(6.121)
for a = 1, . . . , M. The latter is a version of the WilsonCowan equations for cortical

dynamics [675, 676]. Note that all synapses innervated by a particular type of neuron have the same sign. That is, if type b neurons are excitatory (inhibitory), then
wab (x, y) 0 (wab (x, y) 0) for all a = 1, . . . , M and (x, y). Thus, one of the major
reasons for considering more than one class of neuron is to incorporate both excitatory and inhibitory synapses. It can be argued that since excitatory and inhibitory
synapses tend to have different time courses in response to action potentials, one
should take ab (x, y,t) = wab (x, y)b (t), suggesting that the activity-based model
(6.121) with a is more biologically realistic than the voltage-based model, at
least in the case of excitatoryinhibitory networks [173]. However, in practice, both
versions of the neural field equations are used to model cortical dynamics. Since
both versions exhibit very similar types of solution, and since most analytical results have been obtained for voltage-based neural fields, we will mainly focus on
the latter.
Under certain additional simplifying assumptions, it is possible to incorporate inhibition into the scalar neural field equations (6.115) or (6.116) [505]. For
example, consider a two-population model (M = 2) of excitatory (a = E) and
267
inhibitory (a = I) neurons evolving according to the pair of continuum voltage-based

equations
uE (x,t)
uE
=
+
t
E
uI (x,t)
uI
=
+
t
I
wEE (x, y)FE (uE (y,t))dy +
wIE (x, y)FE (uE (y,t))dy +
wEI (x, y)FI (uI (y,t))dy

(6.122a)
wII (x, y)FI (uI (y,t))dy,

(6.122b)
with wEE , wIE 0 and wEI , wII 0. Now suppose that wII 0, FI (uI ) = uI /I and
I E . It follows that we can eliminate uI by setting
uI (x) I
wIE (x, y)FE (uE (y,t))dy,
which leads to a scalar equation for uE of the form (6.115) with effective weight
distribution

w(x, y) = wEE (x, y) +
wEI (x, y
)wIE (y
, y)dy
.
(6.123)
It is then possible for w(x, y) to change sign as a function of x, y. (Often w is modeled as a difference of Gaussians or exponentialsthe so-called Mexican hat weight
distribution.) The reduced model can be used to investigate the effects of inhibition
on stationary solutions and propagating waves. However, in contrast to the full twopopulation model, it does not support oscillatory solutions (in the absence of axonal
delays, higher-order synapses, or some form of adaptation such as synaptic depression).
It is important to emphasize that there does not currently exist a multi-scale
analysis of conductance-based neural networks that provides a rigorous derivation of neural field equations, although some progress has been made in this direction [100, 147, 184, 307, 675, 676]. One crucial step in the derivation of neural field equations presented here was the assumption of slowly varying synaptic
currents, which is related to the assumption that there is not significant coherent
activity at the level of individual spikes. This allowed us to treat the output of a
neuron (or population of neurons) as an instantaneous firing rate. A more rigorous derivation would need to incorporate the mean-field analysis of local populations of stochastic spiking neurons into a larger-scale cortical model and to carry
out a systematic form of coarse graining or homogenization in order to generate
a continuum neural field model. Nevertheless, the heuristic approach does provide
a framework for relating parameters of neural field equations to biophysical parameters such as membrane/synaptic time constants and axonal delays and also
prescribes how to incorporate additional physiological processes such as synaptic
depression and spike frequency adaptation. Moreover, neural field models make it
possible to explore the dependence of cortical dynamics on the detailed anatomy of
local and long-range synaptic connections. It is often assumed that w depends on
the Euclidean distance between interacting cells within the 2D cortical sheet so that
268
w(x, y) = w(|x y|). However, this is an oversimplification of the detailed architecture of cortex [6466, 308, 526]; see Sect. 8.1. A related simplification is to take
axonal delays to depend on Euclidean distance according to |x y|/v, where v is the
speed of propagation.
It is also possible to construct continuum neural field equations for stochastic
population models. For the sake of illustration, consider the Langevin equation
(6.71), except we simplify the multiplicative noise by additive noise dW for constant noise strength . The continuum limit of Eq. (6.71) proceeds as follows. First,
set A (t) = A( d,t) and w = dw( d, d) where is a synaptic density and d is an infinitesimal length
scale. We also assume that the noise strength
= / d and define W (t)/ d = W ( d,t). Taking the limit d 0 with
= for all gives

dA(x,t) = A(x,t) + F( w(x y)A(y,t)dy) dt + dW (x,t)
(6.124)
with
dW (x,t) = 0,
dW (x,t)dW (y,t) = (x y)dt.
(6.125)
Similarly, we can write down a stochastic version of a voltage-based neural field

equation, namely,
dU(x,t) = [U(x,t) +
w(x y)F(U(y,t))dy]dt + dW (x,t)
(6.126)
From a numerical perspective, any computer simulation would involve rediscretizing space and then solving a time-discretized version of the resulting stochastic
differential equation. On the other hand, in order to investigate analytically the effects of noise on spatiotemporal dynamics, it is more useful to work directly with
stochastic neural fields. One can then adapt various PDE methods for studying noise
in spatially extended systems [546], as illustrated in Sect. 7.4. Recently, Buice and
Cowan [91] have used path integral methods and renormalization group theory to
establish that a stochastic neural field model based on a continuum version of a
birthdeath master equation belongs to the universality class of directed percolation
and consequently exhibits power law behavior, suggestive of many measurements of
spontaneous cortical activity in vitro and in vivo [30, 508]; see Sect. 9.5. Although
the existence of power law behavior is still controversial [28], the application of
path integral methods provides yet another example of how analytical techniques
familiar in the study of PDEs are being adapted to studies of continuum neural
fields. (For reviews on path integral methods for stochastic differential equations
see [115, 618, 698].)
Finally, note that several groups have constructed equivalent PDE models for
neural fields with axonal propagation delays, which take the form of damped inhomogeneous wave equations [307, 382, 467, 527, 528, 603]. The basic idea is to
assume a particular form for the synaptic weight distribution and to use Fourier
transforms. Consider, for example, a 2D version of the multi-population integral
269
equation (6.119). Suppose that ab (x, y,t) = wab (|x y|) (t) and introduce the
auxiliary field
ab (x,t) =
wab (|x y|)Fb(y,t |x y|/vab)dy,
R2
where we have set Fb (y,t) = Fb (ub (y,t)). Fourier transforming this equation with
ab (k, ) =
R2
ei(kr+ t)ab (x,t)dt dx,
and
wab (r) =
we find that
ab (k, ) = w0ab
w0ab r/ab
e
,
2
Aab ( )
Fb (k, )
(Aab ( )2 + k2 )3/2
with Aab ( ) = 1/ab + i /vab. If one now makes a long-wavelength approximation by Taylor expanding the denominator of the above equation about k = 0 and
rearranging to give (Aab ( )2 + 3k2 /2)ab(k, ) = Fb (k, ), one can then apply the
inverse Fourier transform to derive the damped wave equation

2
1
1
3 2
+
t ab (x,t) = w0ab Fb (ub (x,t)).
ab vab
2
The current ua (x,t) is then related to the field ab (x,t) according to
ua (x,t) =
t
(t t
) ab (x,t
)dt
,
b
which can itself be converted to a PDE by applying the inverse operator Lt . There
have also been various extensions of the PDE theory including improvements upon
the long-wavelength approximation [142] and incorporation of network inhomogeneities [526]. The damped wave equation (6.127) and its generalizations have
been used extensively to study large-scale phenomena such as EEG rhythms, where
axonal propagation delays are important [60, 467]; see also Sect. 9.4. PDE models
have also been used to study single and multi-bump stationary solutions of scalar
neural field equations, in which the time-independent equations reduce to fourthorder differential equations with a Hamiltonian structure [355, 356].
Chapter 7
Waves in Excitable Neural Fields
In Sect. 5, we showed how both solitary pulses and oscillatory phase waves could
occur in a synaptically coupled network of spiking neurons, where the fundamental
element of the network was a single neuron. Hence, whether the network acted as
an excitable or an oscillatory medium depended primarily on the intrinsic properties of the individual neurons. In this chapter, we focus on waves in excitable neural
fields, where the fundamental network element is a local population of cells (see
Sect. 6), and show how many of the PDE methods and results for the analysis of
waves in reactiondiffusion equations (see part I) can be extended to the nonlocal
integrodifferential equations of neural field theory. We begin by analyzing the existence and stability of solitary traveling fronts in a 1D scalar neural field. (Since
there is strong vertical coupling between layers of a cortical column, it is possible
to treat a thin vertical cortical slice as an effective 1D medium.) In order to relate
the models to experiments on disinhibited cortical slices (Sect. 5.1), we assume
that the weight distribution is purely excitatory. This is also motivated by the observation that epileptic seizures are often associated with greatly enhanced levels
of recurrent excitation [430] (Sect. 9.4). We then extend the analysis to the case of
traveling pulses, which requires the inclusion of some form of local negative feedback mechanism such as synaptic depression or spike frequency adaptation. Next
we describe two approaches to analyzing wave propagation failure in inhomogeneous neural media: one based on homogenization theory [79, 332] and the other on
interfacial dynamics [132]. This is followed by a discussion of wave propagation in
stochastic neural fields.
It is useful to emphasize here that there are two main approaches to analyzing
the spatiotemporal dynamics of neural field equations. The first method is based on
the original work of Amari [8], in which one establishes the existence of nonlinear
traveling wave solutions by explicit construction. This is possible if one takes the
firing rate function F to be the Heaviside (1.15). It is also possible to study the linear stability of such solutions by constructing an associated Evans function, whose
zeros determine the spectrum of the resulting linear operator [134, 552, 696]. The
constructive approach of Amari [8] has been particularly useful in providing explicit insights into how spatiotemporal network dynamics depends on the structure
271
272
7 Waves in Excitable Neural Fields
of the synaptic weight kernel as well as various physiological parameters. Moreover, in certain cases, it is possible to use singular perturbation methods [504, 505]
or fixed-point theorems [172, 336] to extend results for neural fields with Heaviside
nonlinearities to those with more realistic sigmoidal nonlinearities; see also [136].
The second method is based on bifurcation theory, following the original work of
Ermentrout and Cowan [169], in which one investigates the emergence of spatially
periodic stationary and oscillatory patterns through a combination of linear stability
analysis, weakly nonlinear analysis, symmetric bifurcation theory, and numerical
simulations, as reviewed in [67, 71, 167]. Rigorous functional analytical techniques
combined with numerical bifurcation schemes have also been used to study the existence and (absolute) stability of stationary solutions for a general class of neural
field models with smooth firing rate functions [185, 642]. As far as we are aware,
these methods have not yet been applied to traveling wave solutions of neural field
equations.
7.1 Traveling Fronts in a Scalar Neural Field

7.1.1 Propagating Fronts in a Bistable Neural Field
We begin by using Amaris constructive method [8] to analyze the existence of
traveling front solutions in a scalar neural field equation. Similar analyses are found
in Refs. [76, 134, 504]. We assume a Heaviside rate function (1.15) and an excitatory
weight distribution of the form w(x, y) = w(x y) with w(x) 0 and w(x) = w(x).
We also assume that w(x) is a monotonically decreasing function of x for x 0. A
common choice is the exponential weight distribution
1 |x|/
e
,
(7.1)
w(x) =
2
where determines the range of synaptic connections. The latter tends to range
from 100 m to 1 mm. The resulting neural field equation is
u(x,t)
= u(x,t) +
w(x x
)F(u(x
,t))dx
,
(7.2)
t
with F(u) = H(u ). We have fixed the units of time by setting = 1. If is

interpreted as a membrane time constant, then 10 msec. In order to construct
a traveling front solution of (7.2), we introduce the traveling wave coordinate =
xct, where c denotes the wave speed, and set u(x,t) = U( ) with lim U( ) =
U+ > 0 and lim U( ) = 0 such that U( ) only crosses the threshold once.
Here U+ =
w(y)dy is a spatially uniform fixed-point solution of (7.2). Since
Eq. (7.2) is equivariant with respect to uniform translations, we are free to take the
threshold crossing point to be at the origin, U(0) = , so that U( ) < for > 0
and U( ) > for < 0. Substituting this traveling front solution into Eq. (7.2)
then gives
cU
( ) + U( ) =
0
273
w(
)d
=
w(x)dx W ( ),
(7.3)
where U
( ) = dU/d . Multiplying both sides of the above equation by e /c and
integrating with respect to leads to the solution

1 y/c
/c
e
W (y)dy .
(7.4)
U( ) = e
c 0
Finally, requiring the solution to remain bounded as ( ) for c > 0 (for
c < 0) implies that must satisfy the condition
1
|c|

0
ey/|c|W (sign(c)y)dy.
(7.5)
Thus, one of the useful aspects of the constructive method is that it allows us to
derive an explicit expression for the wave speed as a function of physiological parameters such as firing threshold and range of synaptic connections. In the case of
the exponential weight distribution (7.1), the relationship between wave speed c and
threshold is
c=
[1 2 ] (for < 0.5),

2
c=
1 2
(for 0.5 < < 1).
2 1
(7.6)
This establishes the existence of a unique front solution for fixed , which travels to
the right (c > 0) when < 0.5 and travels to the left (c < 0) when > 0.5. As we
will show below, the traveling front is stable.
Given the existence of a traveling front solution for a Heaviside rate function, it
is possible to extend the analysis to a smooth sigmoid nonlinearity using a continuation method [172]. We briefly summarize the main result. Consider the scalar neural
field equation (7.2) with F given by the sigmoid function (1.14) and w(x) nonneg
w(x)dx = 1. Suppose that the function
ative and symmetric with normalization
F(u)
= u + F(u) has precisely three zeros at u = U ,U0 with U < U0 < U+
and F
(U ) < 0. It can then be shown that (modulo uniform translations) there exists a unique traveling front solution u(x,t) = U( ), = x ct, with
cU
( ) + U( ) =
w(
)F(U(
))d
,
(7.7)
and U( ) U as [172]. Moreover, the speed of the wave satisfies

c =
U
( )2 F
(U( ))d
(7.8)
where F
(U) = dF/dU and
U+
U
F(U)dU.
(7.9)
274
Since the denominator of Eq. (7.8) is positive definite, the sign of c is determined by
the sign of the coefficient . In particular, if the threshold = 0.5 and the gain of the
sigmoid > 4 (see Eq. (1.14), then there exists a pair of stable homogeneous fixed
points with U = U+ , which in turn implies that = 0 and the front solution
is stationary. Note that this analysis has been extended to a more general form of
nonlocal equations by Chen [108].
7.1.2 Wave Stability and Evans Functions

Suppose that the scalar neural field equation (7.2) has a traveling wave solution
u(x,t) = U( ), = x ct with c > 0. Following Coombes and Owen [134], it is
convenient to rewrite the neural field equation in the integral form
u(x,t) =

0
w(y) (s)F (u(x y,t s))dsdy,
(7.10)
with (t) = et H(t). For this representation, the front solution satisfies
U( ) =

0
w(y) (s)F (U( y + cs)dsdy.
(7.11)
In order to determine the stability of the front solutions, we transform to traveling

wave coordinates by setting u(x,t) = U( ,t) = U( ) + ( ,t), and Taylor expand
to first order in . This leads to the linear integral equation
( ,t) =

0
w(y) (s)F
(U( y + cs)) ( y + cs,t s)dsdy. (7.12)
We now seek solutions of Eq. (7.12) of the form ( ,t) = ( )e t , C, which

leads to the eigenvalue equation = L( ) . That is,
( ) =

y
w(y) ((s + y )/c)e (s+y )/cF

(U(s)) (s)
ds
dy, (7.13)
c
where we have performed the change of variables cs+ y s. The linear stability
of the traveling front can then be determined in terms of the spectrum (L( )).
Following appendix section 2.7, we assume that the eigenfunctions L2 (R)
and introduce the resolvent operator R( ) = [L( ) I]1 , where I denotes the
identity operator. We can then decompose the spectrum (L ) into the disjoint sum
of the discrete spectrum and the essential spectrum. Given the spectrum of the linear
operator defined by Eq. (7.13), the traveling wave is said to be linearly stable if [551]
max{Re( ) : (L ), = 0} K
(7.14)
275
for some K > 0, and = 0 is a simple eigenvalue of . The existence of at

least one zero eigenvalue is a consequence of translation invariance. Indeed, differentiating equation (7.11) with respect to shows that ( ) = U
( ) is an
eigenfunction solution of Eq. (7.13) with = 0. As in the case of PDEs (see
Sect. 2.4), the discrete spectrum may be associated with the zeros of an Evans function. A number of authors have applied the Evans function construction to neural field equations [134, 198, 506, 536, 552, 696], as well as more general nonlocal problems [314]. Moreover, for neural fields with Heaviside firing rate functions, the Evans function can be calculated explicitly. This was first carried out by
Zhang [696], who applied the method of variation of parameters to the linearized
version of the integrodifferential Eq. (7.2), and was subsequently extended using a
more direct integral formulation by Coombes and Owen [134].
Construction of Evans function. Setting F(U) = H(U ) in Eq. (7.12) and using the
identity
( )
(7.15)
H
(U( ) ) = (U( ) ) =
|U (0)|
gives
( ) =
(0)
c|U
(0)|
w(y) ((y )/c)e (y )/c dy.
(7.16)
In order to obtain a self-consistent solution at = 0, we require that
(0) =
(0)
c|U
(0)|

0
w(y) (y/c)e y/c dy,
(7.17)
We have used the fact that (y) = 0 for y < 0, which is a consequence of causality. Hence,
a nontrivial solution exists provided that E ( ) = 0, where
E ( ) = 1
1
c|U
(0)|

0
w(y) (y/c)e y/c dy.
(7.18)
Equation (7.18) can be identified with the Evans function for the traveling front solution
of the scalar neural field equation (7.10). It is real valued if is real. Furthermore, (i) the
complex number is an eigenvalue of the operator L if and only if E ( ) = 0, and (ii) the
algebraic multiplicity of an eigenvalue is equal to the order of the zero of the Evans function [134, 552, 696]. We briefly indicate the proof of (i) for (t) = et H(t). Equation (7.16)
becomes
(0) ( +1) /c
e
w(y)e( +1)y/c dy,
c|U
(0)|

1
( +1)y/c
w(y)e
dy
e( +1) /c ,
= (0) 1
c|U
(0)| 0
( ) =
which in the limit gives

lim ( ) = (0)E ( ) lim e( +1) /c .
Assuming that Re > 1 (which turns out to be to the right of the essential spectrum), then
( ) will be unbounded as unless E ( ) = 0. That is, if E ( ) = 0, then ( ) is
normalizable, the resolvent operator is not invertible and is an eigenvalue.
276
It is also straightforward to show that E (0) = 0, which we expect from translation invariance. First, setting F(U) = H(U ) in Eq. (7.11) and differentiating with respect to
show that

1
U
( ) =
w(y) ((y )/c)dy.
(7.19)
c
Thus, defining
H ( ) =
w(y) (y/c)e y/c dy,
(7.20)
H ( )
.
H (0)
(7.21)
we see that c|U

(0)| = H (0) and, hence,
E ( ) = 1
It immediately follows that E (0) = 0.
In order to determine the essential spectrum, consider the inhomogeneous equation
( )
(0)
c|U
(0)|
w(y) ((y )/c)e (y )/c dy = h( )
(7.22)
for some normalizable smooth function h on R. Assuming that does not belong to the
discrete spectrum, E ( ) = 0, we can expresses the constant (0) in terms of h(0) by setting
= 0 in Eq. (7.22): (0) = h(0)/E ( ). Thus,
( ) = h( ) +
h(0)
1
E ( ) c|U
(0)|
w(y) ((y )/c)e (y )/c dy.
(7.23)
Fourier transforming this equation using the convolution theorem gives

+
(k) = h(k)
h(0)
1
(k) (kc + i ),
E ( ) c|U
(0)|
where
(k) =
(y)eiky dy
(7.24)
(7.25)
etc. Now suppose that for a given value of k, there exists = (k) for which [ (kc +
i (k))]1 = 0. It follows that the right-hand side of Eq. (7.24) blows up if = (k), that
is, the dispersion curve belongs to the essential spectrum.
For the sake of illustration, let us calculate the zeros of the Evans function in the
case of the exponential weight function (7.1). Substituting (t) = et and w(y) =
e|y|/ /2 in Eq. (7.20) gives
H ( ) =
1
1
2 1 + /c + 1/c
so that [134]
E ( ) =
.
c/ + 1 +
(7.26)
It follows that = 0 is the only zero of the Evans function and it is a simple
root (since E
(0) > 0). Furthermore, in the particular case (t) = et , we have
[ (kc + i )]1 = 1 ikc + so that the essential spectrum is (k) = 1 + ikc,
that is, a vertical line in the complex plane at Re = 1. It follows that the
277
corresponding traveling front (it it exists) is stable. This example illustrates one
of the powerful features of the constructive method based on Heavisides. Not only
is it possible to construct exact traveling wave solutions and derive formulae for
the speed of the wave, but one can also explicitly construct the Evans function that
determines wave stability. The method extends to multi-population neural field models, neural fields with axonal propagation delays, and adaptive neural fields [134].
(Although taking the high-gain limit of a smooth firing rate function is not very
realistic from a biological perspective, one finds that many of the basic features of
traveling waves persist for finite gain.) In the particular case of axonal delays, it
can be shown that delays reduce the speed of a wave but do not affect its stability
properties. For example, given a right-moving traveling front solution of the scalar
neural field equation (6.118) with = 1 and exponential weights, one finds that the
speed of the wave is [134, 139]
c=
1 2
,
2 + (1 2 )/v
where v is the propagation speed along an axon, and the Evans function is
E ( ) =
.
c/ + (1 c/v) +
7.1.3 Pulled Fronts

So far we have assumed that the scalar neural field operates in a bistable regime analogous to the FitzHughNagumo equations; see Sect. 2. However, as we explored
within the context of CaMKII translocation waves (Sect. 3.2), Fisher-like reaction
diffusion equations support traveling waves propagating into unstable states, resulting in pulled fronts (Sect. 3.3). It turns out that it is also possible to observe pulled
fronts in an activity-based version of a scalar neural field equation [70, 132]:

a(x,t)
= a(x,t) + F
w(x x )a(x ,t)dx .

(7.27)
t
with a(x,t) 0 for all (x,t). Note that the restriction to positive values of a is a
feature shared with population models in ecology or evolutionary biology, for example, where the corresponding dependent variables represent number densities. Indeed, Eq. (7.27) has certain similarities with a nonlocal version of the FisherKPP
equation, which takes the form [238]

p(x,t)
2 p(x,t)
=D
+ p(x,t) 1
K(x x )p(x ,t)dx . (7.28)
t
x2
278
One major difference from a mathematical perspective is that Eq. (7.28) supports
traveling fronts even when the range of the interaction kernel K goes to zero, that is,
K(x) (x), since we recover the standard local FisherKPP equation (3.44) [191,
345]. In particular, as the nonlocal interactions appear nonlinearly in Eq. (7.28), they
do not contribute to the linear spreading velocity in the leading edge of the front.
On the other hand, nonlocal interactions play a necessary role in the generation of
fronts in the neural field equation (7.27).
Suppose that F(a) in Eq. (7.27) is a positive, bounded, monotonically increasing
function of a with F(0) = 0, lima0+ F
(a) = 1 and lima F(a) = for some
positive constant . For concreteness, we take
0, a 0
F(a) = a, 0 < a
(7.29)
, a > .
A homogeneous fixed-point solution a of Eq. (7.27) satisfies
a = F(W0 a ),
W0 =
w(y)dy.
(7.30)
In the case of the given piecewise linear firing rate function, we find that if W0 > 1,
then there exists an unstable fixed point at a = 0 (absorbing state) and a stable fixed
point at a = ; see Fig. 7.1(a). The construction of a front solution linking the stable
and unstable fixed points differs considerably from that considered in neural fields
with sigmoidal or Heaviside nonlinearities [8, 167], where the front propagates into
a metastable state; see Fig 7.1(b). Following the PDE theory of fronts propagating
into unstable states [544] (see Sect. 3.3), we expect there to be a continuum of front
velocities and associated traveling wave solutions.
Recall that a conceptual framework for studying such solutions is the linear
spreading velocity c , which is the asymptotic rate with which an initial localized
perturbation spreads into an unstable state based on the linear equations obtained by
linearizing the full nonlinear equations about the unstable state. Therefore, linearizing equation (7.27) about a = 0 gives
b
F(W0a)
F(W0a)
/W0 activity a
/W0
activity a
Fig. 7.1 Plots of firing rate function. Intercepts of y = F(W0 a) with the straight line y = a determine
homogeneous fixed points. (a) Piecewise linear rate function (7.29) showing the existence of an
unstable fixed point at a = 0 and a stable fixed point at a = . (b) Sigmoidal rate function F(a) =
2/(1 + e2[a ] ) showing the existence of two stable fixed points separated by an unstable fixed
point
279
a(x,t)
= a(x,t) +
t
w(x x
)a(x
,t)dx
.
(7.31)
Note that if a(x, 0) 0 for all x, then Eq. (7.31) ensures that a(x,t) 0 for all
x and t > 0. One way
to see this is to note from Eq. (7.31) that a(x,t + t) =

(1 t)a(x,t) + t
w(x x
)a(x
,t)dx
. Assuming positivity of the solution at
time t and using the fact that the neural field is purely excitatory (w(x) 0 for all
x), it follows that a(x,t + t) is also positive. An arbitrary initial condition a(x, 0)
will evolve under Eq. (7.31) as
a(x,t) =
G(x y,t)a(y, 0)dy,
(7.32)
where G(x,t) is Greens function

G(x,t) =
eikxi (k)t
dk
,
2
(k) = i[w(k)
1],
(7.33)
and w(k)
is the Fourier transform of the weight distribution w(x). Hence, the solution
can be written in the form of Eq. (3.102):
a(x,t) =
a0 (k)ei[kx (k)t])
dk
.
2
(7.34)
with a0 , the Fourier transform of the initial condition a(x, 0).

Given a sufficiently steep initial condition, for which the Fourier transform a(k)
is analytic, the asymptotic behavior of a(x,t) can be obtained from the large-time
asymptotics of G(x,t) based on steepest descents. It immediately follows from the
analysis of the Fisher equation in Sect. 3.3 that the linear spreading velocity c is
given by c = c( ) where

Im( (i ))
dc( )
c( ) =
,
= 0.
(7.35)
d =
Using the fact that the Fourier transform of the weight distribution is real valued,
we find that
1
c( ) = [W ( ) 1],
(7.36)
( ) + W
( ) and W ( ) is the Laplace transform of w(x):
where W ( ) = W
W ( ) =
w(y)e y dy.
(7.37)
We are assuming that w(y) decays sufficiently fast as |y| so that the Laplace
transform W ( ) exists for bounded, negative values of . This holds in the case of
a Gaussian weight distribution
280

2
2
W0
w(x) =
ex /2 ,
2
2
(7.38)
since
W ( ) =
W0
w(y)e y dy =
2 2
Hence,
W0 e
2 /2 2
ey
e y dy = W0 e
2 2 /2
2 2 /2
1
.
(7.39)
If W0 > 1 (necessary for the zero activity state to be unstable), then c( ) is a positive
unimodal function with c( ) as 0 or and a unique minimum at
= 0 with 0 the solution to the implicit equation
c( ) =
0 2 =
W0 e0
2W0
2 2
/2
(7.40)
Example dispersion curves are shown in Fig. 7.2(a) for various values of the
Gaussian weight amplitude W0 . Combining Eqs. (7.39) and (7.40) shows that
2 2
c0
= 2W0 e0 /2 = 2 (0 c0 + 1),
0
so that

1
1
4
1
0 =
+
.
+
2
c0
c0 2 2
(7.41)
(7.42)
Assuming that the full nonlinear system supports a pulled front (see Sect. 3.3), then
a sufficiently localized initial perturbation (one that decays faster than e0 x ) will
asymptotically approach the traveling front solution with the minimum wave speed
c0 = c(0 ). Note that c0 and 0 1 . In Fig. 7.2(b), we show an asymptotic front profile obtained by numerically solving the neural field equation (7.27)
when W0 = 1.2. The corresponding displacement of the front is a linear function
of time with a slope consistent with the minimal wave speed c0 0.7 of the corresponding dispersion curve shown in Fig. 7.2(a). This wave speed is independent
of .
The asymptotic analysis of the linear equation (7.31) also shows that, given a
sufficiently localized initial condition, |a(x,t)| e ( ,t) as t , where =

x ct and the leading-edge variable ( ,t) is given by
e /(4Dt)
( ,t)
4 Dt
(7.43)
(i ) d 2 c( )
=
.
2
2 d 2
(7.44)
with
D =
281
b
6
0.5
W0 =3
t=0
c0
0.4
t = 12
t = 24
c() 3
0.2
2
W0 =1.2
1
0
t = 36
0.4
0.8
1.2
1.6
0.0
50
x
100
Fig. 7.2 (a) Velocity dispersion curves c = c( ) for a pulled front solution of the neural field
equation (7.27) with piecewise linear firing rate function (7.29) and a Gaussian weight distribution
with amplitude W0 and width . Here = 1.0, = 0.4 and W0 = 1, 2, 1.5, 2.0, 2.5, 3.0. Black dots
indicate minimum wave speed c0 for each value of W0 . (b) Snapshots of the front profile evolving
from an initial condition consisting of a steep sigmoid function of unit amplitude (gray curve).
Here W0 = 1.2
Positivity of D follows from the fact that is a minimum of c( ). However, as

shown by Ebert and van Saarloos [162], although the spreading of the leading edge
under linearization gives the right qualitative behavior, it fails to match correctly the
traveling front solution of the full nonlinear system. In particular, the asymptotic
front profile takes the form A ( ) e for 1. The factor of reflects

the fact that at the saddle point the two branches of the velocity dispersion curve
c( ) meet, indicating a degeneracy. In order to match the e asymptotics of the

front solution with the leading-edge solution, it is necessary to take the leading-edge
function (x,t) to be the so-called dipole solution of the diffusion equation [162]:
e /(4Dt)
e /(4Dt)
=
(x,t) =
.
4 Dt
2 (2Dt)3/2
2
(7.45)
Putting all of this together, if the neural field equation supports a pulled front, then
the leading edge should relax asymptotically as
|a| e e
with =
x ct.
2 /(4Dt)
t 3/2
(7.46)
Finally, writing
3
lnt
(7.47)
2
suggests that to leading order, the velocity relaxes to the pulled velocity c according
to (see also [162])
e t 3/2 = e
[xv tX(t)]
X(t) =
= c 3 + h.o.t.
v(t) = c + X(t)
2 t
(7.48)
282
7.2 Traveling Pulses in Adaptive Neural Fields

Traveling fronts are not particularly realistic, since populations of cells do not
stay in the excited state forever. Hence, rather than a traveling front, propagating activity in cortex is usually better described as a traveling pulse. (One example where fronts rather than pulses occur is wave propagation during binocular rivalry [83, 312, 369, 678]; see Sect. 8.) One way to generate a traveling pulse is
to include some form of synaptic inhibition, provided that it is not too strong [8].
However, even in the absence of synaptic inhibition, most neurons possess intrinsic
negative feedback mechanisms that slowly bring the cell back to resting voltages
after periods of high activity. Possible nonlinear mechanisms include synaptic depression or spike frequency adaptation as discussed in Sect. 6.1. However, most
analytical studies of traveling pulses in neural field models have been based on a
simpler linear form of adaptation introduced by Pinto and Ermentrout [504]. (For
an analysis of waves in neural fields with nonlinear adaptation, see, e.g., [135, 329].)
The linear adaptation model is given by
u(x,t)
= u(x,t) +
w(x x
)F(u(x
,t))dx
q(x,t)
t
1 q(x,t)
= q(x,t) + u(x,t),
t
(7.49a)
(7.49b)
where and determine the rate and amplitude of linear adaptation. We first
show how to construct a traveling pulse solution of Eq. (7.49) in the case of a
Heaviside rate function F(u) = H(u ), following the particular formulation
of [198, 696]. We then indicate how singular perturbation methods can be used
to construct a traveling pulse for smooth F, as carried out by Pinto and Ermentrout [504]. The introduction of adaptation means that the neural field can support fronts or pulses, depending on whether there exist one or two stable homogeneous fixed points; see Fig. 7.3. We will focus on the latter here. Note, however, that linear (or nonlinear) adaptation can have a nontrivial effect on the propagation of traveling fronts [76, 80]. This is due to the occurrence of a symmetry breaking front bifurcation analogous to that found in reactiondiffusion systems [251, 252, 524, 561]. That is, a stationary front can undergo a supercritical
pitchfork bifurcation at a critical rate of adaptation, leading to bidirectional front
propagation. As in the case of reactiondiffusion systems, the front bifurcation acts
as an organizing center for a variety of nontrivial dynamics including the formation of oscillatory fronts or breathers. The latter can occur, for example, through a
Hopf bifurcation from a stationary front in the presence of a weak stationary input
inhomogeneity [76].
283
0.6
= 1.0
0.4
0.2
= 2.5
q 0
-0.2
-0.4
-0.6
-0.2
0.2
0.4
0.6
Fig. 7.3 Plot of nullclines for space-clamped planar system u = u + F(u) u, 1 q = q +

u with F(u) = 1/(1 + e (u ) ). Nullcline q = u + F(u)]/ for = 1.0 ( = 2.5) intercepts
straight nullcline q = u at three fixed points (one fixed point) and the corresponding spatially
extended network acts as a bistable (excitable) medium. Other parameters are = 20, = 0.25
7.2.1 Exact Traveling Pulse Solution

Without loss of generality, let us consider a right-moving traveling pulse solution
of the form (u(x,t), q(x,t)) = (U(x ct), Q(x ct)) with U(), Q() = 0 and
U( ) = U(0) = ; see Fig. 2.1(b). Here c, denote the speed and width of the
wave, respectively. We also assume that U( ) > for ( , 0) and U( ) <
for < and > 0. Substituting this solution into Eq. (7.49) with = x ct
then gives
cU
( ) + U( ) + Q( ) =
0
w(
)d
cQ
( ) + [Q( ) U( )] = 0.
It is useful to rewrite Eq. (7.50) in the matrix form

1
1
U
U
c
= [W ( ) W ( + )]
0

Q
Q
(7.50)
(7.51)
with W ( ) = w(x)dx. We proceed by diagonalizing the left-hand side of

Eq. (7.51) using the right eigenvectors v of the matrix

1
M=
.
(7.52)

These are given by v = ( , )T with corresponding eigenvalues

1
=
1 + (1 + )2 4 (1 + ) .
2
(7.53)
284
We will assume that is sufficiently small so that < (1 )2 /4 and consequently

are real. (For a discussion of the effects of complex eigenvalues see [580].)
Note that v e /c are the corresponding null vectors of the linear operator on the
left-hand side of Eq. (7.51). Performing the transformation

U
1 U
=
T
(7.54)
, T = v+ v ,
Q
Q
then gives the pair of equations
c U + +U = + [W ( ) W ( + )]
c Q + Q = [W ( ) W ( + )]
with = 1/(+ ). Integrating the equation for U from to , we have

)e + /c + e+
/c [W (
) W (
+ )]d
.
) = e+ /c U(
U(
c
Finiteness of U in the limit requires the term in square brackets to cancel.
) to obtain the result
Hence, we can eliminate U(
) = +
U(
c
e+ /c [W (
+ ) W(
+ + )]d
.
(7.55)
Similarly,
) =
Q(
/c
[W (
+ ) W(
+ + )]d
.
(7.56)
Performing the inverse transformation U = ( + )U + ( )Q we have

U( ) =
1
c

0
+ e +
/c
+ e
/c
[W (
+ ) W(
+ + )]d
,
(7.57)
with = ( ) . The threshold conditions U( ) = and U(0) = then

yield a pair of equations whose solutions determine existence curves relating the
speed c and width of a pulse to the threshold [134, 198, 504].
For the sake of illustration, let w be given by the exponential function (7.1). In
the domain > 0, there is a common factor of e / in the integrand of Eq. (7.57)
so that U( ) = e / for > 0 provided that
(c + )(1 e / )
1
.
2
2 c + c (1 + ) + 2 (1 + )
(7.58)
(Note that for zero negative feedback ( = 0), Eq. (7.58) reduces to the formula
for wave speed of a front in the limit .) On the other hand, when < 0,
one has to partition the integral of Eq. (7.57) into the separate domains
> | |,
30
1
wave speed c
pulse width
285
20
10
0.2
0.4
threshold
0.6
0.8
0.6
0.4
0.2
0
0.2
0.4
threshold
0.6
Fig. 7.4 Existence of right-moving traveling pulses in the case of the excitatory network (7.49)
with linear adaptation for an exponential weight distribution (7.1). Here = 1, = 0.01 and =
2.5. (a) Plot of pulse width against threshold . (b) Plot of wave speed c against threshold .
Stable (unstable) branches indicated by black (gray) curves
| | <
< | | and
< | | . This then determines the second threshold
condition as well as the asymptotic behavior of U( ) in the limit :
U( ) = A+ e+ /c + A e /c + A0 e .
(7.59)
where the amplitudes A and A0 can be determined from matching conditions at

the threshold crossing points [198, 504]. Note that the leading edge of the pulse
is positive, whereas the trailing edge is negative due to the effects of adaptation.
One finds that for sufficiently slow negative feedback (small ) and large there
exist two pulse solutions: one narrow and slow and the other wide and fast. This is
illustrated in Fig. 7.4. Note that a numerical value of c 1 in dimensionless units
( = = 1) translates into a physical speed of 6090 mm/s if the membrane time
constant = 10 msec and the range of synaptic connections is = 600900 m.
Numerically, the fast solution is found to be stable [504], and this can be confirmed
analytically using an Evans function construction [134, 198, 507]; see below. Finally, note that one of the predictions of the neural field model is that the speed of
wave propagation should increase as the threshold decreases [504]. Interestingly,
this has been confirmed experimentally by applying electric fields to a disinhibited
rat cortical slice [521]. The experimental setup is shown in Fig. 7.5. A positive (negative) electric field increases (decreases) the speed of wave propagation by altering
the effective excitability of layer V pyramidal neurons. Such neurons have long apical dendrites and are easily polarizable by the electric field.
Construction of Evans function. Rewrite the neural field equation (7.49) in the integral form
u(x,t) =

0
w(y) (s)F (u(x y,t s))dsdy

0
(s)u(x,t s)ds,
(7.60)
286
global field
field direction
local field
FE1
R1
R2
R2
+
FE1
R3
R1
SE
SE
FE2
Positive
Field
R1
R2
Zero
Field
40 ms
200 V
Negative
Field
Fig. 7.5 (a) Rat cortical slices are bathed in picrotoxin (a GABAA blocker) and a stimulation
electrode (SE) is placed in layers 56 to initiate epileptiform bursts. An electric field is applied
globally or locally across the slice using Ag/AgCl electrodes (FE1,FE2). Layer five neurons have
long apical dendrites and are easily polarizable by an electric field, which controls the effective
firing threshold of the neuron. (b) The time for an activity pulse to travel between two recording
electrodes R1 and R2 depends on the applied electric field, reflecting the dependence of wave speed
on the effective firing threshold. [Adapted from Richardson, Schiff and Gluckman [521]]

with (t) = et H(t) and (t) = 0t (s)e (ts) ds. Linearizing about the pulse solution by
setting u(x,t) = U( ) + ( )e t gives
( ) =

y

w(y) ((s + y )/c)e (s+y )/c F

(U(s)) (s)
((s )/c)e (s )/c (s)
ds
.
c
ds
dy
c
(7.61)
Proceeding along similar lines to the analysis of front stability in Sect. 7.1, we set F(U) =
H(U ) and use the identity
H
(U( ) ) = (U( ) ) =
( )
( + )
+
.
|U
(0)| |U
( )|
(7.62)
287
This gives
( ) +
=
((s )/c)e (s )/c (s)
ds
c
(7.63)
(0)
( )
H ( , ) +
H ( , + )
c|U
(0)|
c|U
( )|
where
H ( , ) =
w(y) ((y )/c)e (y )/c dy.
(7.64)
Let H ( , k) denote the Fourier transform of H ( , ) and G( , k) denote the Fourier transform of ( /c)e /c . Using Fourier transforms and the convolution theorem, Eq. (7.63)
can then be rewritten as
( ) =
(0)
( )
B( , ) +
B( , + ),
c|U
(0)|
c|U
( )|
(7.65)
with B( , ) the inverse transform of

, k) =
B(
H ( , k)
.
[1 + G( , k)/c]
(7.66)
Finally, the eigenvalues are determined by setting = 0, and solving the resulting
matrix equation f = M ( )f with f = ( (0), ( )) and
1
M ( ) =
c
B( ,0)
B( , )
|U
(1 )| |U
( )|
B( , ) B( ,0)
|U
(0)|
|U
( )|

.
(7.67)
It follows that the eigenvalues are zeros of the Evans function

E ( ) = Det[1 M ( )],
(7.68)
where 1 denotes the identity matrix.
7.2.2 Singularly Perturbed Pulse Solution

In the case of slow adaptation ( 1), Pinto and Ermentrout [504] showed how to
construct a traveling pulse solution of Eq. (7.49) for a smooth firing rate function F
by exploiting the existence of traveling front solutions of the corresponding scalar
equation (7.2). The method is analogous to the construction of traveling pulses in
the FitzHughNagumo equation [316]; see Sect.2.3. The basic idea is to analyze
separately the fast and slow time behavior of solutions to Eq. (7.49) expressed in
traveling wave coordinates:
c
dU( )
= U( ) Q( ) +
d
dQ( )
= [Q( ) + U( )].
d
w(
)F(U(
))d
,
(7.69)
(7.70)
288
We will assume the normalization w(y)dy = 1. In the case of fast time, the slow
adaptation is taken to be constant by setting = 0 so that we have the inner layer
equations
dU( )
= U Q0 +
d
dQ( )
= 0.
d
w(
)F(U(
))d
,
(7.71)
(7.72)
Since Q( ) = Q0 is a constant, the term Q0 can be absorbed into the threshold of

the firing rate function F by making the shift U( ) U( )+ Q0 . Hence Eq. (3.55)
is equivalent to the scalar equation (7.7), which supports the propagation of traveling
fronts. In the case of slow time, we introduce the compressed variable = so
that

dU( )
1
= U( ) Q( ) +
w([
]/ )F(U(
))d
, (7.73)
c
d

c
dQ( )
= Q( ) + U( ).
d
(7.74)
In the limit 0, we have

1
w([
]/ ) (
)
so that first equation becomes
Q( ) = U( ) + F(U( ))
(7.75)
Inverting this equation yields two branches U = g (Q). Hence we obtain a slow
time or outer layer equation on each branch (see Fig. 7.6):
dQ 1
= [Q g(Q)]
d
c
(7.76)
The construction of the traveling pulse now proceeds by matching inner and outer
solutions [504]. This can be visualized by considering the nullclines of the spaceclamped version of Eq. (7.49); see Fig. 7.6. We assume that the gain of F and the
strength of adaptation are such that there is only a single fixed point of the spaceclamped system:
I Starting at the unique fixed point, use the fast inner equations and the existence
results of [172] to construct a leading front solution at Q = Q0 with speed c0
and matching conditions lim U( ) = g (Q0 ).
II Use the slow outer equations to determine dynamics of Q along upper branch
U = g+ (Q)
III The solution leaves upper branch at some point Q1 . Once again use the fast
inner equations and [172] to construct a trailing front solution with speed c1
and matching conditions
lim U( ) = g (Q1 )
7.3 Wave Propagation in Heterogeneous Neural Fields
289
IV Finally, use the slow outer equations to determine the return to the fixed point
along the lower branch.
In order to establish the existence of a traveling pulse solution, it remains to find a
value Q1 for which c1 = c0 so that the leading and trailing edges of the pulse move
at the same speed and thus the pulse maintains its shape as it propagates. (Since Q0
is known, so is c0 .) Adapting the formula for the wave speed obtained in [172], we
have
2 ( )F
(U( ))d
U
c1 =
g+ (Q1 )
g (Q1 )
[U Q1 + F(U)]dU. (7.77)
Unfortunately, it is not possible to derive a closed form expression for the wave
speed. However, the existence of a matching speed can be established provided that
certain additional assumptions are made regarding the shape of the firing rate function; see [504] for more details.
g+(Q0)
U = g-(Q)
II
g+(Q1)
Q = Q1
III
IV
I
I
g-(Q0)
II
Q = Q0
III
0
IV
g-(Q1)
U = g+(Q)
Fig. 7.6 Singular perturbation construction of a traveling pulse in (a) the phase plane and (b)
traveling wave coordinates. See text for details

Most studies of neural field theory assume that the synaptic weight distribution
only depends upon the distance between interacting populations, that is, w(x, y) =
w(|x y|). This implies translation symmetry of the underlying integrodifferential
equations (in an unbounded or periodic domain) and an excitatory network can support the propagation of solitary traveling waves. However, if one looks more closely
at the anatomy of cortex, it is clear that its detailed microstructure is far from homogeneous. For example, to a first approximation, primary visual cortex (V1) has a
periodic-like microstructure on the millimeter length scale, reflecting the existence
of various stimulus feature maps; see Sect. 8.1. This has motivated a number of
290
studies concerned with the effects of a periodically modulated weight distribution

on wave propagation in neural fields [64, 132, 332].
We first consider the voltage-based neural field equation (6.115) with periodically modulated weight distribution
w(x, y) = w(x y)[1 + K(y/ )],
(7.78)
where is the amplitude of the periodic modulation and is the period with
K(x) = K(x + 1) for all x. It will also be assumed that if = 0 (no periodic modulation), then the resulting homogeneous network supports a traveling front solution of
speed c0 as analyzed in Sect. 7.1.1. We will describe two alternative methods for analyzing the effects of periodic wave modulation: one based on averaging theory for
small [79], which adapts the method use to study the propagation failure in myelinated axons (Sect. 2.5), and the other based on analyzing interfacial dynamics [132].
Both approaches make use of the observation that for sufficiently small , numerical
simulations of the inhomogeneous network show a front-like wave separating high
and low activity states. However, the wave does not propagate with constant speed,
but oscillates periodically in an appropriately chosen moving frame. This pulsating
front solution satisfies the periodicity condition u(x,t) = u(x + ,t + T ) so that we
can define the mean speed of the wave to be c = /T . We will then consider the
effects of periodically modulated weights on the propagation of pulled fronts in an
activity-based neural field equation, extending the HamiltonJacobi method used in
the analysis of CaMKII waves in Sect. 3.2.
7.3.1 Averaging Theory

Suppose that the period of weight modulations is much smaller than the range
of synaptic interactions . (We fix the length scales by setting = 1.) Following the analysis of saltatory waves along myelinated axons (Sect. 2.5) and [318,
319], we want any inhomogeneous terms to be O( ). Therefore, after substituting
Eq. (7.78) into (6.115), we integrate by parts to obtain the equation
u(x,t)
= u(x,t) +
t
w(x x
)F(u(x
,t))dx
(7.79)

F(u(x
,t))
K (x
/ ) w
(x x
)F(u(x
,t)) w(x x
)
+
dx
.
x
Here K
(x) = K(x) with K only having to be defined up to an arbitrary constant.
Motivated by the existence of pulsating front solutions, we perform the change of
variables = x (t) and = t. Equation (7.79) becomes
u
= u( , )+
291
u( , )
w(
)F(u(
, ))d
+
(7.80)
F(u(
, ))
w
(
)F(u(
, ))w(
)
d
.
Next perform the perturbation expansions

u( , ) = U( ) + u1( , ) + 2 u2 ( , ) + . . .,
(7.81)

( ) = c0 + 1
( )
(7.82)
where U( ) is the unique traveling wave solution of the corresponding homogeneous equation (7.7) with unperturbed wave speed c = c0 . The first-order term u1
satisfies the inhomogeneous linear equation
u1 ( , )
+ Lu1 ( , ) = 1
( )U
( ) + h1( , / )
(7.83)
where
Lu( ) = c0
and
h1 =

K
du( )
u( ) +
d
w(
)F
(U(
))u(
)d
w
(
)F(U(
)) + w(
)
(7.84)
dF(U(
))
d
.
d
(7.85)
The linear operator L has a one-dimensional null-space spanned by U

. The existence of U
as a null vector follows immediately from differentiating both sides of
Eq. (7.7) with respect to , whereas its uniqueness can be shown using properties
of positive linear operators [172]. Therefore, a bounded solution of Eq. (7.83) with
respect to and will only exist if the right-hand side of Eq. (7.83) is orthogonal to
all elements of the null-space of the adjoint operator L . The latter is defined with
respect to the inner product

u( )Lv( )d =

L u( ) v( )d
(7.86)
where u( ) and v( ) are arbitrary integrable functions. Hence,

L u( ) = c
du( )
u( ) + F
(U( ))
d
w(
)u(
)d
.
(7.87)
292
It can be proven that L also has a one-dimensional null-space [172], that is, it is
spanned by some function V ( ). Equation (7.83) thus has a bounded solution if and
only if

B0 1
( ) =
where
B0 =
V ( )h1 ( , / )d
(7.88)
V ( )U
( )d .
(7.89)
Note that B0 is strictly positive since V and U

can be chosen to have the same
sign [172]. Substituting for h1 using Eqs. (7.85) and (7.82) and performing an integration by parts leads to a differential equation for the phase :

d
= c + 1
,
(7.90)
d
where

+
1
1
w( )K
=
B0

dF(U(

))
d
d .
V ( )F(U( )) + V ( )
d
(7.91)
The phase equation (7.90) is identical in form to the one derived in Sect. 2.5 for
wave propagation along myelinated axons; see Eq. (2.72). It implies that there are
two distinct types of behavior. If the right-hand side of Eq. (7.90) is strictly positive,
then there exists a pulsating front of the approximate form U(x (t)) and the
average speed of propagation is c = /T with
T=

0
d
c + 1
.
(7.92)
On the other hand, if the right-hand side of Eq. (7.90) vanishes for some , then
there is wave propagation failure.
In the case of a Heaviside firing rate function F(u) = H(u ), it is possible
to derive an explicit expression for the wave speed c [64]. The solution for the
unperturbed wave front U( ) was derived in Sect. 7.1, so it is only necessary to
determine the solution V ( ) of the adjoint equation (7.87), which becomes
cV
( ) + V ( ) =
( )
U
(0)
w(
)V (
)d
.
(7.93)
This can be integrated to give

V ( ) = H( )e /c .
(7.94)
293
Given the solutions for U( ) and V ( ), it can then be shown that (7.91) reduces to
the form
B0 1

W ( )
w( ) d , (7.95)
K
= W (0)K
+
c
0
where
W ( ) =

0
ey/c0 w(y + )dy cU

( ),
and
B0 =
1
c0

0
e /c0 W ( )d .
(7.96)
(7.97)
Keeping only the lowest-order contribution to 1 , Eq. (7.92) reduces to

T=

0
d
c0 + (c0 )A

(7.98)
with (c0 ) = W (0)/B0 . For the sake of illustration, suppose that the periodic modulation functions K and A are pure sinusoids. Setting A(x) = sin(2 x)/(2 ) in
Eq. (7.98) we find that
(7.99)
T=
c20 2 2 (c0 )2
and, hence,
c=

c20 2 2 (c0 )2 /(2 )2 .
(7.100)
This establishes that a sinusoidally varying heterogeneous neural medium only supports a propagating wave if the velocity c0 of the (unique) solution of the corresponding homogeneous medium satisfies the inequality
c0 (c0 ).
(7.101)
For the particular example of an exponential distribution (7.1) with = 1, we have

c0 = (1 2 )/(2 ) and (c0 ) = 1 + c0 so that
c = c0

1 0 2 2 ,
0 =
1
.
2 (2 1)
(7.102)
The above averaging method can also be extended to the case of periodically
modulated traveling pulses (pulsating pulses) (see [332]), in which there are two
294
a 1.5
1.5
1
u(x)
u(x)
1
0.5
0
-0.5
-40
0.5
0
-20
0
x
20
40
-0.5
-30 -20 -10
0
x
10
20
30
Fig. 7.7 Pulsating pulse solutions in a 1D excitatory neural field with linear adaptation and
Heaviside firing rate function; see Eq. (7.49). The threshold is = 0.2, strength of adaptation is = 2.0, and adaptation rate constant is = 0.04. The weight distribution is given
by w(x, y) = w(x y) sin(2 x/ ) with 2 = 0.3 and w(x) an exponential weight function.
(a) Single-bump solution for = 0.3. The interior of the pulse consists of non-propagating, transient ripples. (b) Multi-bump solution for = 0.8. The solitary pulse corresponds to the envelope
of a multiple bump solution, in which individual bumps are non-propagating and transient. The
disappearance of bumps at one end and the emergence of new bumps at the other end generate the
propagation of activity [332]
threshold crossing points. One simplifying assumption of the analysis is that, in the
presence of periodically modulated weights, additional threshold crossing points
do not occur. However, numerical solutions of a neural field equation with linear
adaptation have shown that in the case of large amplitude modulations, a pulsating
pulse can develop multiple threshold crossing points [332]. That is, the traveling
wave represents the envelope of a multi-bump solution, in which individual bumps
are non-propagating and transient; see Fig. 7.7. The appearance (disappearance) of
bumps at the leading (trailing) edge of the pulse generates the coherent propagation of the pulse. Wave propagation failure occurs when activity is insufficient to
maintain bumps at the leading edge.
7.3.2 Interfacial Dynamics

The averaging method provides a reasonable estimate for the mean wave speed and
the critical amplitude for wave propagation failure, provided that the spatial period
1. As shown by Coombes and Laing [132] in the case of a Heaviside firing rate
function, a more accurate estimate for the wave speed for larger values of can be
obtained by analyzing the dynamics of the interface between high and low activity
states, provided that the amplitude of periodic modulations is not too large [132].
The basic idea is to change to a co-moving frame of the unperturbed system, u =
u( ,t) with = x c0t such that Eq. (6.115) becomes
c0 u + ut = u +
w( + c0t, y)F(u(y c0t,t)dy,
(7.103)
295
with w given by Eq. (7.78) and F(u) = H(u ). The moving interface (level set)
is then defined according to the threshold condition
u(0 (t),t) = .
(7.104)
Differentiating with respect to t then determines the velocity of the interface in the
co-moving frame according to
ut (0 (t),t)
d 0
=
.
dt
u (0 (t),t)
(7.105)
As in the previous averaging method, suppose that for = 0, there exists a traveling
front solution U( ) of the homogeneous equation (7.7) with speed c0 . Now make
the approximation u (0 (t),t) = U
(0), which is based on the assumption that for
small amplitudes , the slope of the traveling front varies sufficiently slowly. Setting
= 0 (t) in Eq. (7.103) and using Eq. (7.3), it is then straightforward to show
that [132]

d 0
= c0
dt
w(y)K(0 + c0t y)
(7.106)
w(y)dy
0
In order to match up with the previous method, let K(x) = sin(2 x/ ) and w(x) =
e|x| /2. Then c0 = (1 2 )/(2 ) and [132]

2
d 0
(0 (t) + c0t) + 0 ( ) ,
= c0 ( ) sin
dt
(7.107)
with
( ) =
1
1

,
2 1 1 + (2 / )2
tan 0 ( ) =
2
.
(7.108)
The final step is to look for a T -periodic solution of Eq. (7.107) such that 0 (t) =
0 (t + T ). Setting x0 = 0 + c0t with x0 [0, ] and integrating gives
x0
0
dx
= c0t.
1 + sin(2 x/ + )
(7.109)
This may be evaluated using a half-angle substitution,

z0 (t)+

1
z

1
c0 t =
tan
,

2
2
1 2 2
1 z (0)+
0
(7.110)
296
where z0 (t) = tan[(2 x0 (t)/ + )/2] and x0 (0) = 0. A self-consistent pulsating

front solution is then obtained by imposing the condition = x0 (T ), which then
determines the effective speed c = /T to be
c = c0

1 2 ( )2 .
(7.111)
Note that on Taylor expanding ( ) to first order in , Eq. (7.111) recovers the corresponding result (7.102) obtained using averaging theory. However, the expression
derived using interfacial dynamics is more accurate when the period increases,
provided that the amplitude does not become too large.
7.3.3 HamiltonJacobi Dynamics and Slow Spatial Heterogeneities

We now turn to the effects of periodically modulated weights on the propagation of
pulled front solutions of the activity-based neural field equation (7.27). In the case
of high-frequency modulation, Coombes and Laing [132] adapted previous work by
Shigesada et al. on pulsating fronts in reactiondiffusion models of the spatial spread
of invading species into heterogeneous environments [574, 575]. (In Sect. 3.2.2 we
applied the theory of pulsating fronts to CaMKII translocation waves along spiny
dendrites.) We briefly sketch the basic steps in the analysis. First, substitute the
periodically modulated weight distribution (7.78) into Eq. (7.27) and linearize about
the leading edge of the wave where a(x,t) 0:
a(x,t)
= a(x,t) +
t
w(x y)[1 + K(y/ )]a(y,t)dy.
(7.112)
Now assume a solution of the form a(x,t) = A( )P(x), = x ct with A( ) 0

as and P(x + 2 ) = P(x). Substitution into Eq. (7.112) then gives
cP(x)A
( ) = P(x)A( ) +
w(x y)[1 + K(y/ )]P(y)A( [x y])dy.

(7.113)
Taking A( ) e and substituting into the above equation yields a nonlocal version of the Hill equation:
(1 + c )P(x) =
e [xy] w(x y)[1 + K(y/ )]P(y)dy.
(7.114)
In order to determine the minimal wave speed, it is necessary to find a bounded

periodic solution P(x) of Eq. (7.114), which yields a corresponding dispersion relation c = c( ), whose minimum with respect to can then be determined (assuming it exists). One way to obtain an approximate solution to Eq. (7.114) is to use
Fourier methods to derive an infinite matrix equation for the Fourier coefficients of
80
60
time t [units of ]
time t [units of ]
40
20
297
80
60
40
20
20
40
60
space x [units of ]
80
100
0
0
20
40
60
space x [units of ]
80
100
Fig. 7.8 Spacetime contour plots of a pulsating front solution of the neural field equation (7.112)
with piecewise linear firing rate function (7.29), Gaussian weight distribution (7.38), and a 2 periodic modulation of the synaptic weights, K(x) = cos(x/ ). (a) = 0.5 and (b) = 0.8. Other
parameters are W0 = 1.2, = 1.0, and = 0.4
the periodic function P(x) and then to numerically solve a finite truncated version
of the matrix equation. This is the approach followed in [132]. The matrix equation
takes the form
(1 + c )Pm = W ( im/ )Pm + W ( im/ ) Kl Pml ,
(7.115)
(p) with
where K(x/ ) = n Kn eimx/ , P(x) = n Pn eimx/ , and W (p) = W (p) + W
W (p), the Laplace transform of w(x). One finds that the mean velocity of a pulsating front increases with the period 2 of the synaptic modulations [132]. This is
illustrated in Fig. 7.8, which shows spacetime plots of a pulsating front for = 0.5
and = 0.8.
Now suppose that there is a slowly varying spatial modulation of the synaptic
weight distribution (relative to the range of synaptic interactions). (Although we do
not have a specific example of long-wavelength modulations in mind, we conjecture
that these might be associated with inter-area cortical connections. For example, it
has been shown that heterogeneities arise as one approaches the V1/V2 border in
visual cortex, which has a number of effects including the generation of reflected
waves [688].) In the case of slow modulations, one can extend the HamiltonJacobi
theory of sharp interfaces developed originally for PDEs (see [178, 203, 204, 212,
421] and Sect. 3.3.2) to the case of neural fields [70]. In order to illustrate this,
consider a heterogeneous version of the activity-based neural field equation (7.27)
of the form

a(x,t)
= a(x,t) + F
w(x x )J( x )a(x ,t)dx ,
(7.116)
t
in which there is a slow (nonperiodic) spatial modulation J( x) of the synaptic

weight distribution with 1.
298
Recall from Sect. 3.3.2 that the first step in the HamiltonJacobi method is to
rescale space and time in Eq. (7.116) according to t t/ and x x/ [178, 204,
421]:

1
a(x,t)
= a(x,t) + F
w([x x ]/ )J(x )a(x ,t)dx .

t

(7.117)
Under this hyperbolic rescaling, the front region where the activity a(x,t) rapidly
increases as x decreases from infinity becomes a step as 0; see Fig. 7.2(b). This
motivates introducing the WKB approximation
a(x,t) eG(x,t)/
(7.118)
with G(x,t) > 0 for all x > x(t) and G(x(t),t) = 0. The point x(t) determines the
location of the front and c = x.
Substituting (3.85) into Eq. (7.117) gives
t G(x,t) = 1 +
w([x x
]/ )J(x
)e[G(x ,t)G(x,t)]/ dx
.
(7.119)
We have used the fact that for x > x(t) and 1, the solution is in the leading edge
of the front so that F can be linearized. Equation (7.119) can be simplified using the
method of steepest descents [70]; see below. This yields the equation
t G(x,t) = 1 + w(i
x G(x,t))J(x),
(7.120)
where w(k)
is the Fourier transform of w(x):

w(x) =
ikx
w(k)e
dk
.
2
(7.121)
Equation (7.120) is formally equivalent to the HamiltonJacobi equation
t G + H(x G, x) = 0
(7.122)
with corresponding Hamiltonian

H(p, x) = 1 + w(ip)J(x)
(7.123)
where p = x G is interpreted as the conjugate momentum of x, and w(ip)
= W (p).
It follows that the HamiltonJacobi equation (7.122) can be solved in terms of the
Hamilton equations
dx H

(p)]
=
= J(x)W
(p) = J(x)[W
(p) W
ds
p
dp
H
=
= J
(x)W (p).
ds
x
(7.124)
(7.125)
299
Let X(s; x,t), P(s; x,t) denote the solution with x(0) = 0 and x(t) = x. We can then
determine G(x,t) according to
G(x,t) = E(x,t)t +
t
x,t)ds.
P(s; x,t)X(s;
(7.126)
Here
E(x,t) = H(P(s; x,t), X(s; x,t)),
(7.127)
which is independent of s due to conservation of energy, that is, the Hamiltonian

is not an explicit function of time.
Steepest descent calculation of G. The derivation of Eq. (7.120) using steepest descents proceeds as follows. First, substituting the Fourier transfer (7.121) into (7.119) and reversing
the order of integration gives
t G(x,t) = 1 +
where

S(k,x ;x,t)/
w(k)J(x
)e
dx
dk
,
2
S(k, x
; x,t) = ik(x
x) + G(x
,t) G(x,t).
(7.128)
(7.129)
Exploiting the fact that is small, we perform steepest descents with respect to the x
variable with (k, x,t) fixed. Let x
= z(k,t) denote the stationary point for which S/ x
= 0,
which is given by the solution to the implicit equation
ik + x G(x
,t) = 0.
(7.130)
Taylor expanding S about this point (assuming it is unique) gives to second order

1 2 S
S(k, x
; x,t) S(k, z(k,t); x,t) +
(x
z(k,t))2
2 x
2 x
=z(k,t)
= ik[z(k,t) x] + G(z(k,t),t) G(x,t)
1
xx G(z(k,t),t)(x
z(k,t))2 .
2
(7.131)
Substituting into Eq. (7.128) and performing the resulting Gaussian integral with respect to
x
yields the result

2
1
t G(x,t) = 1 +
w(k)J(z(k,t))
xx G(z(k,t),t)
dk
e (ik[z(k,t) x] + G(z(k,t),t) G(x,t))/
.
2
(7.132)
This can be rewritten in the form

t G(x,t) = 1 +
1
2
S(k;x,t)/
w(k)J(z(k,t))e
dk,
(7.133)
where
x,t) = ik[z(k,t) x] + G(z(k,t),t) G(x,t) + ln xx G(z(k,t),t).
S(k;
2
(7.134)
300
The integral over k can also be evaluated using steepest descents. Thus, Taylor expand S to
second order about the stationary point k = k(x,t), which is the solution to the equation
0=
S
z(k,t)
xxx G(z(k,t),t)
= i[z(k,t) x] +
ik + x G(z(k,t),t) +
.
k
k
2 xx G(z(k,t),t)
(7.135)
It follows from Eqs. (7.130) and (7.135) that z(k(x,t),t) = x + O( ) and so
Moreover,
k(x,t) = ix G(x,t) + O( ).
(7.136)

2
x,t) 1 S
(k k(x,t))2 .
S(k;
2 k2 k=k(x,t)
(7.137)
Substituting into Eq. (7.133) and performing the Gaussian integral with respect to k gives
to leading order in
1
w(k(x,t))J(x).
t G(x,t) = 1 +
ixx G(x,t)k z(k(x,t),t)
(7.138)
Finally, setting x
= z(k,t) in Eq. (7.130) and differentiating with respect to k show that
xx G(z(k,t),t)k z(k,t) = i and we obtain Eq. (7.120).
Given G(x,t), the location x(t) of the front at time t is determined by the equation
G(x(t),t) = 0. Differentiating with respect to t shows that xx G + t G = 0. Let us
begin by rederiving the wave speed for a homogeneous neural field by setting J(x)
1. In this case, d p/ds = 0 so that p = 0 independently of s. Hence, x(s) = xs/t,
which implies that
x =
dx
= W
(0 ).
ds
(7.139)
On the other hand,

x =
t G 1 + W (0 )
=
.
x G
0
(7.140)
Combining these two results means that 0 is given by the minimum of the function
c( ) =
1 + W ( )
(7.141)
and c0 = c(0 ). This recovers the result of Sect. 7.1.3. Thus, in the case of a Gaussian weight distribution, 0 is related to c0 according to Eq. (7.42). Now suppose that there exists a small-amplitude, slow modulation of the synaptic weights
J(x) = 1 + f (x) with 1. We can then obtain an approximate solution of
Hamiltons Eqs. (7.124) and (7.125) and the corresponding wave speed using regular perturbation theory along analogous lines to a previous study of the FKPP
equation [421]. We find (see below) that
x(t) = c0t +

W (0 ) c0 t
c0 0
301
f (y)dy + O( 2 ).
(7.142)
Here c0 is the wave speed of the homogeneous neural field ( = 0), which is given by
c0 = c(0 ) with 0 obtained by minimizing the function c( ) defined by Eq. (7.141);
see Eq. (7.42). Finally, differentiating both sides with respect to t and inverting the
hyperbolic scaling yields
c x(t)
= c0 +
W (0 )
f ( c0 t) + O( 2).
0
(7.143)
The analytical results agree reasonably well with numerical simulations, provided that is sufficiently small [70]. In Fig. 7.9(a) we show snapshots of a pulled
front in the case of a homogeneous network with Gaussian weights (7.38) and piecewise linear firing rate function (7.29). Space and time units are fixed by setting the
range of synaptic weights = 1 and the time constant = 1. A corresponding
spacetime plot is given in Fig. 7.9(b), which illustrates that the speed of the front
asymptotically approaches the calculated minimal wave speed c0 . (Note that pulled
fronts take an extremely long time to approach the minimal wave speed at high levels
of numerical accuracy, since the asymptotics are algebraic rather than exponential
in time [162].) In Figs. 7.9(c,d) we plot the corresponding results in the case of an
inhomogeneous network. For the sake of illustration, the synaptic heterogeneity is
taken to be a linear function of displacement, that is, J(x) = 1 + (x l), and = .
Equation (7.142) implies that
2 W (0 )
[(c0t)2 2c0lt]
x(t) = l + c0t +
2c0 0

2 l(c0 0 + 1)
2 c0 (c0 0 + 1) 2
= l + c0
t ,
t+
0
20
(7.144)
where we have used Eq. (7.141) and assumed that the initial position of the front
is x(0) = l. Hence, the perturbation theory predicts that a linearly increasing modulation in synaptic weights results in the leading edge of the front tracing out a
downward parabola in a spacetime plot for times t O(1/ 2 ). This is consistent
with numerical simulations for 2 = 0.005, as can be seen in the spacetime plot of
Fig. 7.9(d).
Perturbation calculation of wave speed. Introduce the perturbation expansions
x(s) = x0 (s) + x1 (s) + O( 2 ),
p(s) = p0 (s) + p1 (s) + O( 2 )
(7.145)
and substitute into Eqs. (7.124) and (7.125). Taylor expanding the nonlinear function f (x)
about x0 and W (p) = W (p) + W (p) about p0 then leads to a hierarchy of equations, the
first two of which are
(7.146)
p0 (s) = 0, x0 (s) = W
(p0 ),
and
p1 (s) = f
(x0 )W (p0 ),
x1 (s) = W
(p0 )p1 (s) + f (x0 )W

(p0 ),
(7.147)
302
40
time t [units of ]
0.4
0.3
t=40
a
t=0
0.2
0.1
0
30
20
10
0
10
20
30
40
space x [units of ]
50
10
20
30
40
space x [units of ]
50
40
0.3
time t [units of ]
0.4
t=40
a
0.2
t=0
0.1
0
0
10
20
30
40
space x [units of ]
50
30
20
10
0
10
20
30
40
space x [units of ]
50
Fig. 7.9 (a) Propagating front in a homogeneous network with Gaussian weights (7.38) and piecewise linear rate function (7.29). Parameter values are W0 = 1.2, = 1, = 0.4. The initial condition is taken to be a steep sigmoid a(x, 0) = 0.5/(1 + exp( (x l))) with = 5 and l = 10.
(a) Snapshots of wave profile at time intervals of width t = 5 from t = 10 to t = 40. (b) Space
time contour plot. Wave speed asymptotically approaches the minimum c0 of the velocity dispersion curve given by Eq. (7.39). (c,d) Propagating front in a network with a linear heterogeneity in
the synaptic weights, J(x) = 1 + (x l), l = 10, and 2 = 0.005. Other parameters as in (a,b).
(c) Snapshots of wave profile at time intervals of width t = 5 from t = 10 to t = 40. (d) Space
time contour plot. Wave speed increases approximately linearly with time, so the position x(t) of
front evolves according to a downward parabola. Theoretical curve based on the perturbation calculation is shown by the solid curve. The trajectory of the front in the corresponding homogeneous
case is indicated by the dashed curve
These are supplemented by the Cauchy conditions x0 (0) = 0, x0 (t) = x and xn (0) = xn (t) = 0
for all integers n 1. Equations (7.146) have solutions of the form
p0 (s) = ,
x0 (s) = W
( )s + B0
(7.148)
with , B0 independent of s. Imposing the Cauchy data then implies that B0 = 0 and
satisfies the equation
(7.149)
W
( ) = x/t.
7.4 Wave Propagation in Stochastic Neural Fields
303
At the next order,

t
p1 (s) = W ( ) f (xs/t) + A1 ,
x

t 2 xs/t
x1 (s) = W
( )W ( ) 2
f (y)dy
x 0
+
xs/t
0
(7.150)
f (y)dy + W
( )A1 s + B1 ,
(7.151)
with A1 , B1 independent of s. Imposing the Cauchy data then implies that B1 = 0 and

x
t x
1
f (y)dy
f (y)dy.
(7.152)
A1 = A1 (x,t) = W ( ) 2
x 0
tW ( ) 0
Given these solutions, the energy function E(x,t) is
E(x,t) = 1 + [1 + f (x0 + x1 + . . .)]W ( + p1 + . . .)
= 1 + W ( ) + [W
( )p1 (s) + f (x0 (s))W ( )] + O( 2 ).
(7.153)
Substituting for x0 (s) and p1 (s) and using the condition W

( ) = x/t, we find that
x
(7.154)
E(x,t) = 1 + W ( ) + A1 (x,t) + O( 2 ),
t
which is independent of s as expected. Similarly,
t
0
p(s)x(s)ds
= x + W
( )
t
0
p1 (s)ds + O( 2 )

W
( ) t
x1 (s) W
( ) f (W
( )s) ds + O( 2 )
W ( ) 0

W
( ) x
= x
f (y)dy + O( 2 ).
(7.155)
W ( ) 0
= x+
Hence, to first order in ,

G(x,t) = t W ( )t + x W ( )
t
x
x
0
f (y)dy.
(7.156)
We can now determine the wave speed c by imposing the condition G(x(t),t) = 0 and performing the perturbation expansions x(t) = x0 (t) + x1 (t) + O( 2 ) and = 0 + 1 +
O( 2 ). Substituting into Eq. (7.156) and collecting terms at O(1) and O( ) leads to
Eq. (7.142).

In Sect. 6.4 we constructed stochastic neuronal population models based on a master equation formulation. However, continuum versions of these models are difficult
to analyze even under a diffusion approximation, due to the nonlocal nature of the
multiplicative noise terms; see Eq. (6.71). Therefore, in this section, we analyze the
effects of noise on wave propagation in stochastic neural fields with local multiplicative noise, extending the PDE methods outlined in Sect. 2.6. This form of noise can
also be interpreted in terms of parametric fluctuations in the firing threshold [58].
304
7.4.1 Spontaneous Front Propagation

Consider the following stochastic neural field equation: U(x,t)

dU(x,t) = U(x,t) +
w(x y)F(U(y,t))dy dt + 1/2 g(U(x,t))dW (x,t).
(7.157)
We assume that dW (x,t) represents an independent Wiener process such that
dW (x,t)dW (x
,t
) = 2C([x x
]/ ) (t t
)dtdt
,
dW (x,t) = 0,
(7.158)
where denotes averaging with respect to the Wiener process. Here is the spatial correlation length of the noise such that C(x/ ) (x) in the limit 0,
and determines the strength of the noise, which is assumed to be weak. Moreover,
the multiplicative noise term is taken to be of Stratonovich form. The analysis of
Eq. (7.157) proceeds along similar lines to the analysis of the stochastic bistable
equation in Sect. 2.6; see also [70]. First, using Novikovs theorem, we rewrite
Eq. (7.157) so that the fluctuating term has zero mean:

dU(x,t) = h(U(x,t)) +
w(x y)F(U(y,t))dy dt + 1/2 dR(U, x,t), (7.159)
where
and
h(U) = U + C(0)g
(U)g(U)
(7.160)
dR(U, x,t) = g(U)dW (x,t) 1/2C(0)g

(U)g(U)dt.
(7.161)
The stochastic process R has the variance

dR(U, x,t)dR(U, x
,t) = g(U(x,t))dW (x,t)g(U(x
,t)dW (x
,t) + O( 1/2).
(7.162)
The next step in the analysis is to express the solution U of Eq. (7.159) as a combination of a fixed wave profile U0 that is displaced by an amount (t) from its
uniformly translating position = x c t and a time-dependent fluctuation in
the front shape about the instantaneous position of the front:
U(x,t) = U0 ( (t)) + 1/2 ( (t),t).
(7.163)
The wave profile U0 and associated wave speed c are obtained by solving the modified deterministic equation
c
dU0
h(U0( )) =
d
w(
)F(U0 (
))d
,
(7.164)
305
As in Sect. 2.6, Eq. (7.164) is chosen so that that to leading order, the stochastic
variable (t) undergoes unbiased Brownian motion with
(t) = 0,
(t)2 = 2D( )t
(7.165)
where the diffusion coefficient D( ) = O( ) can be calculated using perturbation

analysis (see below).
Perturbation calculation of diffusivity D( ). Substitute the decomposition (7.163) into
Eq. (7.159) and expand to first order in O( 1/2 ):

c U0
( (t))dt U0
( (t))d (t) + 1/2 d ( (t),t) c
( (t),t)dt
1/2
( (t),t)d (t)
= h(U0 ( (t)))dt + h
(U0 ( (t))) 1/2 ( (t),t)dt

w(
) F(U0 (
(t))) + F
(U0 (
(t))) 1/2 (
(t),t) d
dt
+

+ 1/2 dR(U0 ( (t)), ,t) + O( ).

Imposing Eq. (7.164), after shifting (t), and dividing through by 1/2 then gives
d ( (t),t) = L ( (t),t)dt + 1/2U0
( (t))d (t)
+dR(U0 ( (t), ,t) + O( 1/2 ),
(7.166)
where L is the non-self-adjoint linear operator

L A( ) = c
dA( )
+ h
(U0 ( ))A( ) +
d
w(
)F
(U0 (
))A(
)d
(7.167)
for any function A( ) L2 (R). Finally, for all terms in Eq. (7.166) to be of the same order,
we require that (t) = O( 1/2 ). It then follows that U0 ( (t)) = U0 ( ) + O( 1/2 ) and
Eq. (7.166) reduces to
d ( ,t) = L ( ,t)dt + 1/2U0
( )d (t) + dR(U0 ( ), ,t) + O( 1/2 ). (7.168)
It can be shown that for a sigmoid firing rate function and exponential weight distribution,
the operator L has a 1D null-space spanned by U0
( ) [172]. (The fact that U0
( ) belongs
to the null-space follows immediately from differentiating equation (7.164) with respect
to .) We then have the solvability condition for the existence of a nontrivial solution of
Eq. (7.168), namely, that the inhomogeneous part is orthogonal to all elements of the nullspace of the adjoint operator L . The latter is almost identical in form to Eq. (7.87):
L B( ) = c
dB( )
+ h
(U0 ( ))B( ) + F
(U0 ( ))
d
w(
)B(
)d
.
(7.169)
Hence, L has a one-dimensional null-space that is spanned by some function V ( ). Taking the inner product of both sides of Eq. (7.168) with respect to V ( ) then leads to the
solvability condition

V ( ) U0
( )d (t) + 1/2 dR(U0 , ,t) d = 0.
(7.170)
306
Thus (t) satisfies the stochastic differential equation (SDE)

d (t) = 1/2
V ( )dR(U0 , ,t)d
V ( )U0
( )d
(7.171)
Using the lowest-order approximation dR(U0 , ,t) = g(U0 ( ))dW ( ,t), we deduce that
(t) is a Wiener process with effective diffusion coefficient

D( ) =
V ( )V (
)g(U0 ( ))g(U0 (
))dW ( ,t)dW (
,t)d d
2
V ( )U0
( )d
(7.172)
In the case of a Heaviside rate function F(U) = H(U ) and multiplicative

noise g(U) = g0U, the effective speed c and diffusion
coefficient D( ) can be calculated explicitly [70]. (The constant g0 has units of length/time.) The deterministic equation (7.164) for the fixed profile U0 then reduces to
c
dU0
+ U0 ( ) ( ) =
d
with
w(
)H(U0 (
) )d
,
( ) = (1 g20C(0)),
(7.173)
(7.174)
This is identical in structure to Eq. (7.3) for the deterministic neural field modulo
the rescaling of the decay term. The analysis of the wave speeds proceeds along
similar lines to Sect. 7.1. Thus, multiplying both sides of Eq. (7.173) by e ( )/c
and integrating with respect to gives

1 y ( )/c
( )/c
e
W (y)dy .
(7.175)
U0 ( ) = e
c 0
Finally, requiring the solution to remain bounded as ( ) for c > 0
(for c < 0) implies that must satisfy the condition
1
|c |

0
ey ( )/|c |W (sign(c )y)dy.
(7.176)
Hence, in the case of the exponential weight distribution (7.1), we have
[1 2 ( )]
2
(7.177)
( ) 1 2 ( )
,
2 1 ( )
(7.178)
c =
for c > 0, and
c =
1.2
1.0
U 0.6
1.5
1.0
0.4
0.5
0.2
2.5
2.0
0.8
0.0
0
307
10
20
30
40
0.0
50
2.0
20
10
20
30
40
50
30
40
50
2.0
1.5
1.5
U 1.0
U 1.0
0.5
0.5
0.0
0
10
0.0
10
20
30
40
50
Fig. 7.10 Numerical simulation showing the propagation of a front solution of the stochastic neural
field equation (7.157) for Heaviside weight function F(U) = H(U ) with = 0.35, exponential
weight function (7.1) with = 2, and multiplicative noise g(U) = U. Noise strength = 0.005
and C(0) = 10. The wave profile is shown at successive times (a) t = 0 (b) t = 12 (c) t = 18
and (d) t = 24, with the initial profile at t = 0 given by the solution to Eq. (7.164). In numerical
simulations we take the discrete space and time steps x = 0.1, t = 0.01. The deterministic part
U0 of the stochastic wave is shown by the solid gray curves and the corresponding solution in the
absence of noise ( = 0) is shown by the dashed gray curves
for c < 0. Assuming that 0 ( ) 1, we see that multiplicative noise shifts the
effective velocity of front propagation in the positive direction.
In order to calculate the diffusion coefficient, it is first necessary to determine the
null vector V ( ) of the adjoint linear operator L defined by Eq. (7.169). Setting
F(U) = H(U ) and g(U) = g0U, we obtain an adjoint equation almost identical
to (7.93):

( )
c V
( ) + ( )V ( ) =
w(
)V (
)d
.
(7.179)
U0 (0)
Hence, this has the solution
V ( ) = H( ) exp( ( ) ) ,
( ) =
( )
,
c
(7.180)
308
40
_
X(t)
30
35
0.8
0.6
2 (t)
X
0.4
25
20
0.2
15
10
0
10
15 20
time t
25
30
0.0
10
15 20
time t
25
30
Fig. 7.11 Plot of (a) mean X(t) and (b) variance X2 (t) of front position as a function of time,
averaged over N = 4096 trials. Same parameter values as Fig. 7.10
and Eq. (7.172) reduces to the form

e2 ( ) U0 ( )2 d
D( ) =
2 .
( )
e
U0 ( )d
0
(7.181)
In the case of an exponential weight distribution, U0 ( ) has the explicit form
U0 ( ) =
1 e /
2c 1 + ( )
( )
/
e
2e
2

2c ( ) 1 + 2 ( )2 + ( ) + 1 ( ) < 0,
(7.182)
and the integrals in (7.181) can be evaluated explicitly to give

1
D( ) = g20 (1 + ( )).
2
(7.183)
In Fig. 7.10 we show the temporal evolution of a single stochastic wave front,
which is obtained by numerically solving the stochastic neural field equation (7.157)
for F(U) = H(U ), g(U) = U and an exponential weight distribution w. In order
to numerically calculate the mean location of the front as a function of time, we carry
out a large number of level set position measurements. That is, we determine the positions Xa (t) such that U(Xa (t),t) = a, for various level set values a (0.5 , 1.3 )
and then define the mean location to be X(t) = E[Xa (t)], where the expectation is
first taken with respect to the sampled values a and then averaged over N trials.
2 ]. In Fig. 7.11 we
The corresponding variance is given by X2 (t) = E[(Xa (t) X(t))

2
plot X(t) and X (t) as a function of t. It can be seen that both vary linearly with
t, consistent with the assumption that there is a diffusive-like displacement of the
front from its uniformly translating position at long time scales. The slopes of these
curves then determine the effective wave speed and diffusion coefficient according
b
8
309
0.12
D()
6
0.08
4
0.04
2
0.0
0.2
0.3
0.4
threshold
0.5
0.0
0.2
0.3
0.4
threshold
0.5
Fig. 7.12 Plot of (a) wave speed c and (b) diffusion coefficient D( ) as a function of threshold
. Numerical results (solid dots) are obtained by averaging over N = 4, 096 trials starting from the
initial condition given by Eq. (7.182 ). Corresponding theoretical predictions (solid curves) for c
and D( ) are based on Eqs. (7.177) and (7.181), respectively. Other parameters as in Fig. 7.10
to X(t) c t and X2 (t) 2D( )t. In Fig. 7.12 we plot the numerically estimated
speed and diffusion coefficient for various values of the threshold and compare
these to the corresponding theoretical curves obtained using the above analysis. It
can be seen that there is excellent agreement with our theoretical predictions provided that is not too large. As 0.5, the wave speed decreases towards zero so
that the assumption of relatively slow diffusion breaks down.
7.4.2 Stimulus-Locked Fronts

So far we have assumed that the underlying deterministic neural field equation is
homogeneous in space so that there exists a family of traveling front solutions related by a uniform shift. Now suppose that there exists an external front-like input that propagates at a uniform speed v, so that the deterministic equation (7.2)
becomes
u(x,t)
= u(x,t) +
t
w(x x
)F(u(x
,t))dx
+ I(x vt),
(7.184)
where the input is taken to be a positive, bounded, monotonically decreasing function of amplitude I0 = I() I(). The resulting inhomogeneous neural field
equation can support a traveling front that locks to the stimulus, provided that the
amplitude of the stimulus is sufficiently large [198]. Consider, in particular, the case
of a Heaviside firing rate function F(u) = H(u ). (See [174] for an extension
to the case of a smooth sigmoid function F.) We seek a traveling wave solution
u(x,t) = U ( ) where = x vt and U (0 ) = at a single threshold crossing point
0 R. The front is assumed to travel at the same speed as the input (stimulus-locked
front). If I0 = 0, then we recover the homogeneous equation (7.2) and 0 becomes
a free parameter, whereas the wave propagates at the natural speed c( ) given by
Eq. (7.6). Substituting the front solution into Eq. (7.184) yields
310
dU ( )
= U ( ) +
d
0
w(
)d
+ I( ).
(7.185)
This can be solved for v > 0 by multiplying both sides by the integrating factor
v1 ev and integrating over the interval [ , ) with U( ) 0 as to give
U ( ) =
1
v
e( )/v [W (
0) + I(
)]d
,
with W ( ) defined according to Eq. (7.3). Similarly, for v < 0, we multiply by the
same integrating factor and then integrate over (, ] with U( ) W0 as
to find

1 (
)/v
U ( ) =
e
[W (
0 ) + I(
)]d
.
v
The threshold crossing condition U (0 ) = then determines the position 0 of the
front relative to the input as a function of speed v, input amplitude I0 , and threshold .
One of the interesting features of stimulus-locked fronts is that they are much
more robust to noise [70]. In order to show this, consider the following stochastic
version of Eq. (7.184):

w(x y)F(U(y,t))dy + I(x vt) dt
dU(x,t) = U(x,t) +
1/2
g(U(x,t))dW (x,t).
(7.186)
Proceeding along identical lines to the case of freely propagating fronts, Eq. (7.186)
is first rewritten so that the fluctuating term has zero mean:

dU(x,t) = h(U(x,t))+
w(xy)F(U(y,t))dy+I(xvt) dt+ 1/2 dR(U, x,t),
(7.187)
and h and R are given by Eqs. (7.160) and (7.161), respectively. The stochastic field
U(x,t) is then decomposed according to Eq. (7.163) with U0 a front solution of
v
dU0
h(U0( )) I( ) =
d
w(
)F(U0 (
))d
.
(7.188)
It is assumed that the fixed profile U0 is locked to the stimulus (has speed v). However, multiplicative noise still has an effect on U0 by generating an -dependent
threshold crossing point such that U0 ( ) = .
Proceeding to the next order and imposing equation (7.188), we find that (t) =
O( 1/2 ) and
311
d ( ,t) = L ( ,t)dt + 1/2U0

( )d (t) + dR(U0, ,t) + 1/2I
( ) (t)dt
(7.189)
where L is the non-self-adjoint linear operator (7.167) with c v. The last term
on the right-hand side of Eq. (7.189) arises from the fact that in Eq. (7.163), U0
and are expressed as functions of (t) and I( ) = I( (t) + (t))
I( (t)) + I
( (t)) (t). A nontrivial solution of Eq. (7.189) exists if and
only if the inhomogeneous part is orthogonal to the null vector V ( ) of the adjoint operator L defined by Eq. (7.169) with c v. Taking the inner product
of both sides of Eq. (7.189) with respect to V ( ) thus leads to the solvability
condition

V ( ) U0
( )d (t) + I
( ) (t)dt + 1/2dR(U0 , ,t) d = 0.
(7.190)
It follows that, to leading order, (t) satisfies the OrnsteinUhlenbeck equation

d (t) + A (t)dt = dW (t),

where
A =
and
W (t) = 1/2
V ( )I
( )d
V ( )U0
( )d
V ( )g(U0 ( ))W ( ,t)d

(7.191)
V ( )U0
( )d
(7.192)
(7.193)
Note that A > 0 for I0 > 0, since both U0 ( ) and I( ) are monotonically decreasing
functions of . Moreover
dW (t) = 0,
dW (t)dW (t) = 2D( )dt
(7.194)
with D( ) given by Eq. (7.172). Using standard properties of an Ornstein

Uhlenbeck process [210], we conclude that

D( )
1 e2At .
(7.195)
A
In particular, the variance approaches a constant D( )/A in the large t limit, rather
than increasing linearly with time as found for freely propagating fronts.
In order to illustrate the above analysis, take g(U) = g0U for the multiplicative
noise term and set F(U) = H(U ). The deterministic Eq. (7.188) for the profile
U0 then reduces to
(t) = (0)eAt ,
(t)2 (t)2 =
dU0
+U0 ( )[1 g20C(0)] + I( ) =
d
w(
)H(U0 (
) )d
. (7.196)
312
b
1.2
amplitude I0
amplitude I0
1.2
0.8
0.4
0.0
-40
-20
0
20
stimulus velocity v
0.8
0.4
0.0
-40
40
-20
0
20
stimulus velocity v
40
Fig. 7.13 Plot of existence regions of a stimulus-locked front without noise ( = 1, dark gray) and
in the presence of noise ( = 0.9, light gray) with overlapping regions indicated by medium gray.
Stimulus taken to be of the form I(x,t) = I0 H( ), = x vt with amplitude I0 and speed v. Other
parameter values as in Fig. 7.10. (a) = 0.95: spontaneous fronts exist in the absence of a stimulus
(I0 = 0). (b) = 1.25: there are no spontaneous fronts
2.0
3.0
1.5
U 1.0
1.0
0.5
0.0
0
2.0
10
20
30
40
0.0
50
4.0
3.0
U 2.0
U 2.0
1.0
1.0
0
10
20
30
40
50
20
10
20
30
40
50
30
40
50
4.0
3.0
0.0
10
0.0
Fig. 7.14 Numerical simulation showing the propagation of a stimulus-locked wave-front solution
(black curves) of the stochastic neural field equation (7.186) for Heaviside weight function F(U) =
H(U ) with = 0.35, exponential weight function (7.1) with = 2, and multiplicative noise
g(U) = U. The external input (gray curves) is taken to be of the form I(x,t) = I0 Erfc[x vt] with
amplitude I0 = 0.4 and speed v = 1.5. Noise strength = 0.005 and C(0) = 10. The wave profile
is shown at successive times (a) t = 0 (b) t = 6 (c) t = 12 and (d) t = 24, with the initial profile at
t = 0 given by the solution U0 of Eq. (7.188). In numerical simulations we take the discrete space
and time steps x = 0.1, t = 0.01
313
40
_ 30
X(t)
20
x103
4
2
X(t)
2
10
0
0
10
15
time t
20
25
0 2 4 6
8 10 12 14 16 18
time t
Fig. 7.15 Plot of (a) mean X(t) and (b) variance X2 (t) of the position of a stimulus-locked front
as a function of time, averaged over N = 4096 trials. Smooth gray curve in (b) indicates theoretical
prediction of variance. Stimulus taken to be of the form I(x,t) = I(x ct) = I0 Erfc[x vt] with
amplitude I 0 = 0.4 and speed v = 1.5. Other parameter values as in Fig. 7.10
Existence of front solution proceeds along identical lines to Sect. 7.1, except now
the speed v is fixed, whereas the threshold crossing point 0 , say, is no longer arbitrary due to the breaking of translation symmetry. The point 0 is determined by
the threshold condition U0 (0 ) = and will depend on the noise strength . In
Fig. 7.13 we show existence regions in the (v, I0 )-plane for stimulus-locked fronts
when I( ) = I0 H( ), that is, for a step function input of speed v and amplitude
I0 . This illustrates the fact that multiplicative noise leads to an -dependent shift
in the existence regions. In Fig. 7.14 we show the temporal evolution of a single
stimulus-locked front, which is obtained by numerically solving the Langevin equation (7.186) for F(U) = H(U ), g(U) = U and an exponential weight distribution w. Numerically speaking, it is convenient to avoid discontinuities in the input
by taking I(x,t) = I0 Erfc[x vt] rather than a Heaviside. The corresponding mean
X(t) and variance X2 (t) of the position of the front, which are obtained by averaging over level sets as outlined in Sect. 7.4.1, are shown in Fig. 7.15. It can be
seen that, as predicted by the analysis, X(t) varies linearly with t with a slope equal
to the stimulus speed v = 1.5. Moreover, the variance X2 (t) approaches a constant
value as t , which is comparable to the theoretical value D( )/A evaluated for
the given input. Thus, we find that stimulus-locked fronts are much more robust to
noise than freely propagating fronts, since the variance of the mean position saturates as t . Consequently, stimulus locking persists in the presence of noise over
most of the parameter range for which stimulus locking is predicted to occur.
7.4.3 Stochastic Pulled Fronts

In the case of the FKPP equation with multiplicative noise, one finds that the
stochastic wandering of a pulled front about its mean position is subdiffusive with
var (t) t 1/2 , in contrast to the diffusive wandering of a front propagating into a
314
0.25
slope = 1.0
-1
-2
log[2X(t)]
2X(t)
0.20
0.15
0.10
slope = 0.5
-3
-4
-5
-6
-7
-8
0.05
10
20
30
time t
40
50
60
-1
2
log[t]
Fig. 7.16 (a) Plot of variance X2 (t) of the position of a stochastic pulled front as a function of
time. (b) Log-log plot of variance X2 (t) as a function of time t. Noise amplitude = 0.005 and
= 0.8. Other parameter values as in Fig. 7.2
metastable state for which var (t) t [530]. Such scaling is a consequence of the
asymptotic relaxation of the leading edge of the deterministic pulled front. Since
pulled front solutions of the neural field equation (7.27) exhibit similar asymptotic
dynamics (see Eq. (7.46)), it suggests that there will also be subdiffusive wandering of these fronts in the presence of multiplicative noise. In order to illustrate this,
consider the stochastic neural field equation

dA(x,t) = A(x,t)+F
w(xy)A(y,t)dy dt+ 1/2 g0 A(x,t)dW (x,t) (7.197)
with W (x,t) a Wiener process satisfying Eq. (2.84). Note that the noise term has to
vanish when A(x,t) = 0, since the firing rate A is restricted to be positive. Hence, the
noise has to be multiplicative. Formally speaking, one can carry over the analysis of
the Langevin equation (7.157). First, decompose the solution along similar lines to
Eq. (2.96):
(7.198)
A(x,t) = A0 ( (t)) + 1/2 ( (t),t)
with = x c t, with the fixed front profile A0 satisfying the deterministic equation

dA0
+ A0 ( )[1 g20C(0)] = F
w(
)A0 (
)d
.
(7.199)
c
d
The effective velocity c of the front is given by the minimum of the dispersion
curve

1
( ) [1 g20C(0)] .
c ( ) =
W ( ) + W
(7.200)
Fluctuations thus shift the dispersion curve to higher velocities. However, it is no

longer possible to derive an expression for the diffusion coefficient D( ) along the
lines of Eq. (7.172), since both numerator and denominator would diverge for a
pulled front. This reflects the asymptotic behavior of the leading edge of the front.
7.5 Traveling Waves in 2D Oscillatory Neural Fields
315
It is also a consequence of the fact that there is no characteristic time scale for the
convergence of the front velocity to its asymptotic value, which means that it is not
possible to separate the fluctuations into a slow wandering of front position and fast
fluctuations of the front shape [162, 486]. Nevertheless, numerical simulations of
Eq. (7.197) with F given by the piecewise linear firing rate (7.29) are consistent
with subdiffusive wandering of the front. In Fig. 7.16(a), we plot the variance X2 (t)
of the position of a pulled front solution of Eq. (7.197), which are obtained by
averaging over level sets along identical lines to Sect. 7.4.1. It can be seen that the
variance appears to exhibit subdiffusive behavior over longer time scales. This is
further illustrated by plotting a log-log plot of X2 (t) against time t; see Fig. 7.16(b).
At intermediate time scales, the slope of the curve is approximately equal to one,
consistent with normal diffusion, but at later times the slope decreases, indicating
subdiffusive behavior.

Troy and Shusterman [580, 634] have shown how a neural field model with strong
linear adaptation (see Eq. (7.49)) can act as an oscillatory network that supports
2D target patterns and spiral waves consistent with experimental studies of tangential cortical slices [292]. (For the analysis of spiral waves in the corresponding
excitable regime, see [354].) However, since the linear form of adaptation used in
these studies is not directly related to physiological models of adaptation, it is difficult to ascertain whether or not the strength of adaptation required is biologically
reasonable. This motivated a more recent study of spiral waves in a 2D neural
0.4
stable focus
0.4
0.3
limit cycle
u
0.2
limit cycle
0.3
unstable focus
0.1
q
0.2
q nullcline
saddle
0
1.6
2.4
stable node
3.2 4.0 4.8 5.6
b
u nullcline
6.4
0.1
0.1
0. 2
0.3
Fig. 7.17 Limit cycle oscillations in the space-clamped system (7.202) for a piecewise linear firing
rate function (1.16) with threshold = 0.01 and gain = 4. (a) Bifurcation diagram showing fixed
points u of the system as a function of for q = 80. (b) Corresponding phase-plane plot of q versus
u (gray curve) for = 4, showing that the system supports a stable limit cycle [329]
316

t=0
t = 30
t = 60
t = 75
t = 90
t = 150
200
100
y
-100
-200
200
100
y
-100
-200
-200 -100 0
x
100 200
-200 -100 0
x
100 200
-200 -100 0
x
100 200
Fig. 7.18 Target patterns in a 2D neural field with synaptic depression induced by an initial conditional stimulus specified by Eq. (7.203) at t = 0, where = 1 and = 25. Initially, an activated
state spreads radially outward, across the entire medium as a traveling front. Then, the localized
oscillating core of activity emits a target wave with each oscillation cycle. Eventually, these target
waves fill the domain. Each target wave can be considered as a phase shift in space of the oscillation throughout the medium; they travel with the same speed as the initial front. Parameters are
q = 80, = 4, = 4, and = 0.01 [330]
medium involving a nonlinear, physiologically based form of adaptation, namely,

synaptic depression [330]. The latter model is given by
u(r,t)
= u(r,t) + w(|r r
|)q(r
,t)F(u(r
,t))dr
t
1 q(r,t)
q(r,t)
=
q(r,t)F(u(r,t)).
t
q
(7.201)
The radially symmetric excitatory weight distribution is taken to be an exponential,

w(r) = er /2 . It can be shown that the space-clamped model
u(t)
= u(t) + q(t)F(u(t)),
q(t)
=
1 q(t)
q(t)F(u(t)),
q
(7.202)
supports limit cycle oscillations provided that the firing rate function has finite
gain. For example, in the case of the piecewise linear firing rate function (1.16),
oscillations arise via a subcritical Hopf bifurcation of a high activity fixed point;
see Fig. 7.17. One then finds that the full network model (7.201) supports a spatially localized oscillating core that periodically emits traveling pulses [330]. Such
dynamics can be induced by taking an initial condition of the form
317
t=0
t=9
t = 18
t = 27
t = 36
t = 45
60
30
y 0
-30
-60
60
30
y 0
-30
-60
-120 -90 -60 -30
x
-120 -90 -60 -30

x
-120 -90 -60 -30

x
Fig. 7.19 Spiral wave generated by shifting the phase of the top and bottom halves of the target pattern shown in Fig. 7.18. The period of the spiral wave oscillation is roughly the same as the period
of the oscillation in the space-clamped system. All patches of neurons are oscillating at the same
frequency, but phase-shifted as coordinates are rotated about the central phase singularity [330]
2 +y2 )/ 2
(u(r, 0), q(r, 0)) = ( e(x
, 1),
(7.203)
where and parameterize the amplitude and spatial constant of the initial state.
An example of a pulse-emitting core is shown in Fig. 7.18, which oscillates at a
frequency of roughly 3 Hz. Pulses are emitted each cycle and travel at a speed of
roughly 30 cm/s, which is determined by the period of the oscillations; the latter is
set by the time constant of synaptic depression. The initial emission of spreading activity appears as a traveling front which propagates from the region activated by the
input current into the surrounding region of zero activity; it travels at the same speed
as the subsequent target waves. The front converts each region of the network into an
oscillatory state that is phase-shifted relative to the core, resulting in the appearance
of a radially symmetric target pattern. Spiral waves can also be induced by breaking
the rotational symmetry of pulse emitter solutions [330]. More specifically, if the
target pattern produced by the emitter has the top and bottom halves of its domain
phase-shifted, then the dynamics evolves into two counterrotating spirals on the left
and right halves of the domain. Closer inspection of one of these spirals reveals that
it has a fixed center about which activity rotates indefinitely as shown in Fig. 7.19.
A very different mechanism for generating periodic waves in a 1D or 2D neural field model is through the combination of adaptation and a spatially localized
input [196, 197]. Recall from Sect. 7.2 that a 1D excitatory neural field with adaptation supports the propagation of solitary traveling pulse, which can be induced by
perturbing the system with a transient localized pulse. (In contrast to the previous
example, we are assuming that the neural field operates in an excitable regime.)
318
In the case of a 2D network with radially symmetric weights, such a pulse will produce a single expanding circular wave. Now suppose that a 2D localized pulse per2
2
sists in the form of a radially symmetric Gaussian input I(r) = I0 er / this could
either represent an external stimulus or a localized region of depolarization. As one
might expect, for sufficiently large input amplitude I0 , the neural field supports a radially symmetric stationary pulse or bump centered about the input. Such a bump is
not self-sustaining, however, since if the input is removed, then the bump disappears
as well. This then raises the question as to what happens to the stability of the bump
as the input amplitude is slowly decreased. One finds that the bump first undergoes
a Hopf instability as I0 is decreased, leading to the formation of a spatially localized
oscillating pulse or breather [196]. Interestingly, as the input amplitude is further
reduced, the breather can undergo a secondary instability such that it now acts as an
oscillating core that emits circular target waves. Thus, a spatially localized stationary input provides a mechanism for the formation of a network pacemaker oscillator.
A linear stability analysis establishes that the primary instability is due to the growth
of radially symmetric eigenmodes. A similar bifurcation scenario also occurs in a
neural field with lateral inhibition, except that now the Hopf bifurcation typically
involves the growth of nonradially symmetric eigenmodes, resulting in asymmetric
breathers and rotating waves [197].
Chapter 8
Neural Field Model of Binocular Rivalry Waves
Binocular rivalry is the phenomenon whereby perception switches back and forth
between different images presented to the two eyes. The resulting fluctuations in
perceptual dominance and suppression provide a basis for noninvasive studies of
the human visual system and the identification of possible neural mechanisms underlying conscious visual awareness [46, 47]. Various psychophysical experiments
have demonstrated that the switch between a dominant and suppressed visual percept propagates as a traveling front for each eye [312, 369, 678]. In this chapter
we show how the neural field theory presented in Chap. 7 can be used to model the
phenomenon of binocular rivalry waves [83, 663].
8.1 Binocular Rivalry Waves and the Structure of Primary

Visual Cortex
We begin by summarizing several important properties of neurons and their functional organization within primary visual cortex (V1), and then review some of the
experimental findings regarding binocular rivalry.
8.1.1 Functional Architecture of V1

V1 is the first cortical area to receive visual information from the retina (see
Fig. 8.1). The output from the retina is conveyed by ganglion cells whose axons
form the optic nerve. The optic nerve conducts the output spike trains of the retinal
ganglion cells to the lateral geniculate nucleus (LGN) of the thalamus, which acts
as a relay station between retina and primary visual cortex (V1). Prior to arriving
at the LGN, some ganglion cell axons cross the midline at the optic chiasm. This
allows the left and right sides of the visual fields from both eyes to be represented on
the right and left sides of the brain, respectively. Note that signals from the left and
319
320
8 Neural Field Model of Binocular Rivalry Waves
right eyes are segregated in the LGN and in input layers of V1. This means that the
corresponding LGN and cortical neurons are monocular, in the sense that they only
respond to stimuli presented to one of the eyes but not the other (ocular dominance).
cortex
retina
/2
LGN
right
left
BR
AR
B
A
0
y
-/2
RETINA
x
VISUAL CORTEX
Fig. 8.1 (a) Visual pathways from the retina through the lateral geniculate nucleus (LGN) of the
thalamus to the primary visual cortex (V1). (b) Schematic illustration of the complex logarithmic
mapping from retina to V1. Foveal region in retina is indicated by gray disk. Regions AR and BR
in the visual field are mapped to regions A and B in cortex
One of the striking features of the early visual system is that the visual world is
mapped onto the cortical surface in a topographic manner. This means that neighboring points in a visual image evoke activity in neighboring regions of visual cortex. Moreover, one finds that the central region of the visual field has a larger representation in V1 than the periphery, partly due to a nonuniform distribution of
retinal ganglion cells. The retinotopic map is defined as the coordinate transformation from points in the visual world to locations on the cortical surface and can be
approximated by a complex logarithm [562]. Superimposed upon the retinotopic
map are additional maps reflecting the fact that neurons respond preferentially to
stimuli with particular features [613]. Neurons in the retina, LGN, and primary visual cortex respond to light stimuli in restricted regions of the visual field called their
classical receptive fields (RFs). Patterns of illumination outside the RF of a given
neuron cannot generate a response directly, although they can significantly modulate
responses to stimuli within the RF via long-range cortical interactions (see below).
The RF is divided into distinct ON and OFF regions. In an ON (OFF) region illumination that is higher (lower) than the background, light intensity enhances firing.
The spatial arrangement of these regions determines the selectivity of the neuron to
different stimuli. For example, one finds that the RFs of most V1 cells are elongated
so that the cells respond preferentially to stimuli with certain preferred orientations
(see Fig. 8.2). Similarly, the width of the ON and OFF regions within the RF determines the optimal spacing of alternating light and dark bars to elicit a response, that
is, the cells spatial frequency preference.
In recent years much information has accumulated about the spatial distribution
of orientation selective cells in V1 [216]. One finds that orientation preferences rotate smoothly over the surface of V1, so that approximately every 300 m the same
preference reappears, i.e., the distribution is -periodic in the orientation preference angle. One also finds that cells with similar feature preferences tend to arrange
8.1 Binocular Rivalry Waves and the Structure of Primary Visual Cortex
321
themselves in vertical columns so that to a first approximation the layered structure

of cortex can be ignored. A more complete picture of the two-dimensional distribution of both orientation preference and ocular dominance in layers 2/3 has been
obtained using optical imaging techniques [50, 51, 53]. The basic experimental procedure involves shining light directly on to the surface of the cortex. The degree of
light absorption within each patch of cortex depends on the local level of activity.
Thus, when an oriented image is presented across a large part of the visual field,
the regions of cortex that are particularly sensitive to that stimulus will be differentiated. The topography revealed by these methods has a number of characteristic
features [469]; see Fig. 8.2b: (i) Orientation preference changes continuously as a
function of cortical location, except at singularities or pinwheels. (ii) There exist
linear zones, approximately 750 750 m2 in area (in primates), bounded by pinwheels, within which iso-orientation regions form parallel slabs. (iii) Linear zones
tend to cross the borders of ocular dominance stripes at right angles; pinwheels tend
to align with the centers of ocular dominance stripes. These experimental findings
suggest that there is an underlying periodicity in the microstructure of V1 with a period of approximately 1 mm (in cats and primates). The fundamental domain of this
approximate periodic (or quasiperiodic) tiling of the cortical plane is the hypercolumn [293, 294, 377], which contains two sets of orientation preferences [0, )
per eye, organized around a pair of singularities; see Fig. 8.2b.
b
1
firing rate
(normalized)
0.8
oriented
bar
RF
0.6
0.4
0.2
0
-90
-45
0
orientation (deg)
45
90
Fig. 8.2 (a) Schematic illustration of an orientation tuning curve of a V1 neuron. Average firing
rate is plotted as a function of the orientation of a bar stimulus that is moved back and forth within
the receptive field (RF) of the neuron. The peak of the orientation tuning curve corresponds to the
orientation preference of the cell. (b) Schematic illustration of iso-orientation (light) and ocular
dominance (dark) contours in a region of primate V1. A cortical hypercolumn consists of two
orientation singularities or pinwheels per ocular dominance column
Within each hypercolumn, neurons with sufficiently similar orientations tend

to excite each other, whereas those with sufficiently different orientations inhibit
each other, and this serves to sharpen a particular neurons orientation preference
[34, 188]. There are also longer-range horizontal or lateral connections that link
cells in different hypercolumns [218, 531]. Optical imaging combined with labeling techniques has generated considerable information concerning the pattern of
322
local connections
lateral connections
hypercolumn
Fig. 8.3 Schematic illustration of anisotropic horizontal connections. Orientation selective cells
within a hypercolumn tend to connect to all neighbors in a roughly isotropic fashion. On the other
hand, longer-range horizontal connections link cells between hypercolumns with similar orientation preferences along a particular visuotopic axis
these connections in superficial layers of V1 [56, 401, 691]. In particular, one finds
that the horizontal connections are patchy and tend to link cells with similar feature preferences. Moreover, in certain animals such as tree shrew and cat, there is
a pronounced anisotropy in the distribution of patchy connections, with differing
iso-orientation patches preferentially connecting to neighboring patches in such a
way as to form continuous contours following the topography of the retinotopic
map [56]. This is illustrated schematically Fig. 8.3. That is, the major axis of the
horizontal connections tends to run parallel to the visuotopic axis of the connected
cells common orientation preference. There is also a clear anisotropy in the patchy
connections of primates [11, 583]. However, in these cases most of the anisotropy
can be accounted for by the fact that V1 is expanded in the direction orthogonal
to ocular dominance columns [11]. Nevertheless, it is possible that when this expansion is factored out, there remains a weak anisotropy correlated with orientation
selectivity. Moreover, patchy feedback connections from higher-order visual areas
in primates are strongly anisotropic [11]. Stimulation of a hypercolumn via lateral
connections modulates rather than initiates spiking activity [278], suggesting that
the long-range interactions provide local cortical processes with contextual information about the global nature of stimuli. As a consequence horizontal and feedback
8.1 Binocular Rivalry Waves and the Structure of Primary Visual Cortex
323
connections have been invoked to explain a wide variety of context-dependent visual

processing phenomena [11, 217]. In summary, the functional relationship between
stimulus feature preferences and synaptic connections within V1 suggests that V1
is a likely substrate of many simple examples of binocular rivalry, such as those
involving sinusoidal grating stimuli.
8.1.2 Binocular Rivalry

The perception of ambiguous figures is a well-known phenomenon where, in response to a single fixed image, a subject switches back and forth between two rival
percepts; see Fig. 8.4. In the case of binocular rivalry, fixed but different images are
presented to the two eyes. Rather than perceiving a superposition of the two images,
a subject typically sees the left-eye image, say, which then switches to the righteye image in an alternating fashion. Although binocular rivalry has been studied
for hundreds of years, only recently have experimentalists clarified some of its specific statistical properties using a combination of psychophysics and brain imaging
[47, 48]:
Fig. 8.4 Ambiguous figures. (a) Necker cube. (b) Faces/vase. (c) Monocular rivalry between a
horizontal orange grating and a vertical green grating. [Reproduced from Wikimedia Commons
(public domain)]
324
1. Increasing the strength of one of the rival figures (brighter, moving rather than
stationary, more densely contoured) increases the percentage of time that it is
dominant. However, this tends to occur due to a decrease in the periods of suppression of the enhanced figure rather than an increase in its dominance periods
[199, 390]. The stimulus dependence of rivalry was summarized in a series of
postulates by Levelt [378].
2. The switching from one percept to the next during binocular rivalry occurs
stochastically, with the switching times distributed according to a gamma distribution [390]. A number of studies suggest that there are little or no correlations
between one dominance time and the next [59, 372].
3. Attention to a particular stimulus may prolong the dominance time of the attended stimulus [270]. Experimental evidence suggests that this correlation may
hold true when the subject practices repeatedly [114]. However, the non-attended
stimulus will eventually appear in the observers perception, no matter how
strongly attention is focused on the other. The fact that attention can bias the
outcomes of binocular rivalry suggests that higher level visual areas may play
a modulatory role in the phenomenon [674]. Attention may increase the apparent contrast of the attended stimulus or simply boost the activity of one of two
competing neural populations [114].
4. The tilt aftereffect is still observed when vertical and tilted lines are rivalrous
stimuli [306]; the tilt aftereffect is the phenomenon by which staring at a slightly
tilted line will make a vertical line appear to be tilted in the opposite direction
if viewed immediately after. Since the neural substrate of this phenomenon is
known to reside in primary visual cortex (V1), this suggests that binocular rivalry
can involve V1.
5. Perceptual dominance transitions are not instantaneous; rather they emerge in a
wavelike fashion, originating at one region of a figure and spreading from there
throughout the rest of the figure [312, 369, 678].
Several different methods of recording neural activity in subjects during binocular rivalry have also been employed in an effort to isolate the specific sites of its encoding. In monkeys, single-electrode recordings have been employed to track electrical activity during binocular rivalry tasks [47]. Evidence of an elevation in some
cells firing rates has been found in the V1, V2, and V4 regions of visual cortex, and
these correspond well with the monkeys reported perception of a stimulus [376].
Thus, it appears that several areas of visual cortex may be involved. However, singleunit recordings have yet to reveal changes in the firing rate of LGN neurons that
correspond to the perceptual changes of binocular rivalry [373]. In humans, less invasive techniques such as scalp recordings and functional magnetic resonance imaging (fMRI) have helped to localize brain regions whose activity reflects the experience of binocular rivalry. Visually evoked potentials measured on the scalp during
a binocular rivalry task reveal that potential waveforms associated with each stimulus possess a time course very closely linked to the subjects perceptual observation
[85]. A number of fMRI studies have verified there are changes in the blood-oxygenlevel-dependent (BOLD) signals in V1 that correspond to the perceived switching in
stimulus dominance [369, 510, 626]. It is likely that the site of binocular rivalry de-
8.2 Competitive Network Models of Binocular Rivalry
325
pends on the nature of the rivalrous images. Indeed, some theories propose that there
may be a hierarchy of visual areas involved so that there is a competition between
either eyes input as well as between differing patterns [627, 674].
One way to observe and measure the speed of perceptual waves in psychophysical experiments of binocular rivalry [369, 678] is to take the rival images to be a
low-contrast radial grating presented to one eye and a high-contrast spiral grating
presented to the other eye. Each image is restricted to an annular region of the visual
field centered on the fixation point of the observer, thus effectively restricting wave
propagation to the one dimension around the annulus. Switches in perceptual dominance can then be triggered using a brief rapid increase in stimulus contrast within a
small region of the suppressed low-contrast grating. This induces a perceptual traveling wave in which the observer perceives the local dominance of the low-contrast
image spreading around the annulus. The observer presses a key when the perceptual
wave reaches a target area at a fixed distance from the trigger zone, and this determines the wave speed [369, 678]. Since the rival images consist of oriented gratings,
one might expect that primary visual cortex (V1) plays some role in the generation
of binocular rivalry waves. Indeed, it has been shown using fMRI that there is a
systematic correspondence between the spatiotemporal dynamics of activity in V1
and the time course of perceptual waves [369]. However, it has not been established
whether the waves originate in V1 or are evoked by feedback from extrastriate cortex. Recently Kang et al. [312, 313] have developed a new psychophysical method
for studying binocular rivalry waves that involves periodic perturbations of the rival images consisting of linear gratings. An observer tracks rivalry within a small,
central region of spatially extended rectangular grating patterns, while alternating
contrast triggers are presented repetitively in the periphery of the rival patterns. The
basic experimental setup is illustrated in Fig. 8.5. A number of interesting results
have been obtained from these studies. First, over a range of trigger frequencies, the
switching between rival percepts within the central regions is entrained to the triggering events. Moreover, the optimal triggering frequency depends on the natural
frequency of spontaneous switching (in the absence of triggers). Second, the latency
between triggering event and perceptual switching increases approximately linearly
with the distance between the triggering site and the central region being tracked by
the observer, consistent with the propagation of a traveling front. Third, the speed of
the traveling wave across observers covaries with the spontaneous switching rate.

The above psychophysical experiments suggest that binocular rivalry waves consist
of two basic components: the switching between rivalrous left-/right-eye states and
the propagation of the switched state across a region of cortex. Let us first focus
on the switching mechanism by neglecting spatial effects. Suppose, for the sake of
illustration, that a horizontally oriented grating is presented to the left eye and a vertically oriented grating is presented to the right eye. This triggers rivalry due to the
combination of orientation and eye-specific cross-inhibition in V1. During left-eye
326
Fig. 8.5 Schematic diagram illustrating experimental protocol used to study binocular rivalry
waves [312]. High-contrast triggers are presented periodically in anti-phase within the upper extended region of one grating pattern and within the lower region of the rival pattern. Subject simply
reports perceptual alternations in rival dominance within the central monitoring region indicated by
the horizontal black lines on each pattern. The monitoring region is a distance d from the trigger
region, which can be adjusted. If t is the latency between the triggering event and the subsequent
observation of a perceptual switch, then the speed c of the wave is given by the slope of the plot
d = c t
stimulus dominance, it is assumed that a group of the left-eye neurons that respond
to horizontal orientations are firing persistently, while right-eye neurons are suppressed by cross-inhibitory connections. Of course, there may still be some low-rate
firing of the right-eye neurons, but it will be less than the firing rate of the left-eye
horizontally tuned neurons [47]. Following this, some slow adaptive process causes
a switch so that right-eye vertical orientation neurons fire persistently, suppressing
327
the left-eye neurons resulting in a repetitive cycle of perceptual dominance between

the left- and right-eye stimuli. The competitive network architecture of reciprocal
inhibition paired with slow adaptation (Fig. 8.6) has been used extensively to model
oscillations in binocular rivalry [199, 328, 355, 577, 578, 586, 619, 678]. (In some
versions of the model, recurrent excitation is omitted.) In most cases a firing rate
model appears sufficient to capture the elevation in neuronal spiking associated with
the dominant stimulus.
recurrent excitation
fast activity u(t)

slow adaptation qu(t)
left eye
cross
inhibition
fast activity v(t)

slow adaptation qv(t)
right eye
Fig. 8.6 Schematic diagram of a competitive network architecture for rivalry oscillations [355,
577, 678], consisting of two homogeneous populations of cells: one driven by left-eye stimuli
and the other by right-eye stimuli. Recurrent connections within each population are assumed to
be excitatory, whereas connections between the two populations are inhibitory (cross-inhibition).
Each network is described by two variables: a fast activity variable and a slow adaptation variable
It remains an open question as to which slow adaptive process is most responsible for the eventual switching of one stimulus dominance to the other [577]. The
mechanism of spike frequency adaptation has been suggested, since it can curtail
excitatory activity in a single neuron [355, 678]. Spike frequency adaptation is the
process by which a hyperpolarizing current is switched on due to a buildup of a
certain ion, like calcium, within the cell due to repetitive firing [604]; see Sect. 4.1.
The maximal firing rate of a neuron is lowered as a result. In the case of binocular rivalry, this may cause the dominant population to eventually drop its firing
rate so that cross-inhibition suppressing the other population is then low enough
for the suppressed populations to rekindle its firing rate into dominance. Since the
recently released population is not adapted, it can then remain in dominance and
suppress the other population for a period of time roughly equal to the time constant
of spike frequency adaptation [355, 392, 678]. Another proposed switching mechanism is that the inhibitory synapses from one eyes neurons to the others undergo
synaptic depression; see Sect. 1.3. (More precisely, synaptic depression tends to be
associated only with excitatory synapses, so that inclusion of depressing inhibitory
connections would have to be mediated by excitatory connections innervating local interneurons, for example.) If inhibitory synapses remain repeatedly active, due
to one eyes neurons suppressing the others, eventually most of those synapses resources will be used up, the effect of inhibition will be weakened and the suppressed
population will escape [328, 355, 577]. Synaptic depression could also occur in the
excitatory recurrent connections. For concreteness, we will take slow adaptation to
328
arise from synaptic depression in all synaptic weights; however, the specific choice
of adaptation does not affect the main results.
Let u(t) and v(t) denote the activity variables of the left- and right-eye populations at time t. The rate-based equations for a competitive network model with
synaptic depression can be constructed as follows [328]:
du(t)
= u(t) + Iu(t) + we qu (t)F(u(t)) wi qv (t)F(v(t)),
dt
dqu (t)
= 1 qu(t) qu(t)F(u(t)),
s
dt
(8.1)
(8.2)
and
dv(t)
= v(t) + Iv(t) + we qv (t)F(v(t)) wi qu (t)F(u(t)),
dt
dqv (t)
= 1 qv(t) qv (t)F(v(t)),
s
dt
(8.3)
(8.4)
where the positive constants we and wi denote the strengths of recurrent excitatory
and cross-inhibitory connections, Iu , Iv denote the input stimuli from the left and
right eyes, and F is a firing rate function that is taken to be the sigmoid (1.14). The
depression variables qu , qv satisfy a rate-based version of (1.45); see also (6.15).
In anticipation of the analysis of traveling front solutions of a spatially extended
version of the model, we will take the high gain limit of F to obtain the Heaviside
function F(u) = H(u ). In this limit, the existence and stability of fixed-point
solutions of (8.1) and (8.4) can be analyzed explicitly [328].
Suppose that Iu = Iv = I with I constant. It turns out that there are four possible
homogeneous fixed points and all are stable. First, there is the off state U = V = I,
which occurs when I < , that is, the input is not strong enough to activate either
population. Second there is the on state or fusion state, where both populations are
simultaneously active:

1
we wi
we wi
1
(U ,V ) =
+ I,
+ I , (Qu , Qv ) =
,
,
1+
1+
1+ 1+
and occurs when I > (we wi )/(1 + ). This case is more likely for very strong
depression ( large), since cross-inhibition will be weak, or when the local connections are strong and excitation-dominated. Finally there are two winner-takes-all
(WTA) states in which one population dominates the other: the left-eye dominant
state

1
we
wi
(U ,V ) =
+ I, I
,1 ,
, (Qu , Qv ) =
1+
1+
1+
and the right-eye dominant state

wi
we
,
+I ,
(U ,V ) = I
1+ 1+
(Qu , Qv )

= 1,
1
1+

.
329
The WTA states exist provided that

I>
we
,
1+
I<+
wi
.
1+
This will occur in the presence of weak depression ( small) and strong crossinhibition such that depression cannot exhaust the dominant hold one population
has on the other. It can also be shown that (8.1)(8.4) also support homogeneous
limit cycle oscillations in which there is periodic switching between left- and righteye dominance consistent with binocular rivalry [328]. Since all the fixed points
are stable, it follows that such oscillations cannot arise via a standard Hopf bifurcation. Indeed, one finds bistable regimes where a rivalry state coexists with a fusion state as illustrated in Fig. 8.7. (Such behavior persists in the case of smooth
sigmoid firing rate functions, at least for sufficiently high gain [328].) For a detailed discussion of the various bifurcation scenarios in competitive neural networks
see [435, 563, 577].
b
0.4
0.4
WTA
fusion
0 Off
rivalry
0.1
0.2
qv
0.1
-0.2
-0.4
0.3
0.2
qu
0.2
0.3
-0.1
200
400
600
time t
-0.2
input I
Fig. 8.7 (a) Bifurcation diagram showing homogeneous solutions for the left population activity u
as a function of the input amplitude I. Solid lines represent stable states, whereas circles represent
the maximum and minimum of rivalry oscillations. It can be seen that there are regions of off/WTA
bistability, WTA/fusion bistability, and fusion/rivalry bistability. Parameters are s = 500, = 0.05,
= 5, we = 0.4, and wi = 1. (b) Homogeneous oscillatory solution in which there is spontaneous
periodic switching between left- and right-eye dominance. Plot against time of the left-eye neural
field u (solid gray) and the right-eye neural activity v (solid black) together with the corresponding
depression variables qu (dashed gray) and qv (dashed black)
In order to take into account the propagation of activity seen in binocular rivalry waves, it is necessary to introduce a spatially extended network model. Therefore, suppose that the neuronal populations responding to stimuli from the left eye,
say, are distributed on a one-dimensional (1D) lattice and are labeled according
to the integer n; a second 1D network responds to stimuli from the right eye; see
Fig. 8.8. In terms of the functional architecture of V1, one can interpret the nth
left-/right-eye populations as consisting of neurons in a given hypercolumn that respond maximally to the distinct stimuli presented to the two eyes at a particular
coarse-grained location in space. In the case of the orientated grating stimuli used
by Kang et al. [312, 313] (see Fig. 8.5), this would mean neurons whose orientation
330
recurrent excitation we
um
un
cross inhibition wi
vm
vn
Fig. 8.8 Schematic diagram of a competitive neural network consisting of two sets of populations
distributed along a 1D lattice. Recurrent connections within each 1D network are assumed to be
excitatory, whereas connections between the two networks are inhibitory (cross-inhibition). Slow
adaptation is incorporated into the model by taking the network connections to exhibit synaptic
depression
preference coincides with the stimulus orientation presented to a given eye. Since
the orientation does not change along the length of the grating, all neurons receive
the same external drive. Letting un , vn , qu,n , qv,n denote the activity and depression
variables of the nth left- and right-eye networks, we have
dum
= um + Iu + [[we ]mn qu,n F(un ) [wi ]mn qv,n F(vn )]
dt
n
dqu,m
= 1 qu,m qu,m F(um )
s
dt
(8.5)
(8.6)
and
dvm
= vm + Iv + [[we ]mn qv,n F(vn ) [wi ]mn qu,n F(un )]
dt
n
dqv,m
= 1 qv,m qv,m F(vm ).
s
dt
(8.7)
(8.8)
Here [we ]mn is the strength of excitation from the nth to the mth population with the
same eye preference, and [wi ]mn is the strength of cross-inhibition between populations with opposite eye preferences. The weights are typically assumed to decrease
with distance of separation |m n| according to an exponential or Gaussian distribution. Note that a similar network architecture was previously considered in a computational model of binocular rivalry waves [313, 678], in which cross-inhibition was
mediated explicitly by interneurons and, rather than including depressing synapses,
the excitatory neurons were taken to exhibit spike frequency adaptation. Numerical
simulations of the model showed that the network supported traveling waves consistent with those observed numerically under physiologically reasonable parameter
regimes.
8.3 Neural Field Model
331

Although the discrete lattice model presented in Sect. 8.2 is directly amenable to
numerical simulations, it is difficult to derive any analytical results for the model
that generate, for example, explicit expressions for how the wave speed depends
on network parameters. Therefore, in this section we turn to a neural field model
that can be constructed by taking an appropriate continuum limit of (8.5) and (8.8)
[83, 663]. Introduce a lattice spacing a, and write [we ]nm = awe (na, ma), un (t) =
u(na,t), etc. Substituting into (8.5)(8.8) and taking the continuum limit a 0
such that na x and ma y, the discrete sums can be replaced by integrals to give
the neural field model
du(x,t)
= u(x,t) + Iu +
dt
we (x x
)qu (x
,t)F(u(x
,t))dx
wi (x x
)qv (x
,t)F(v(x
,t))dx
(8.9a)
dqu (x,t)
= 1 qu(x,t) qu(x,t)F(u(x,t))
dt
(8.9b)
and
dv(x,t)
= v(x,t) + Iv +
dt
we (x x
)qv (x
,t)F(v(x
,t))dx
wi (x x
)qu (x
,t)F(u(x
,t))dx
(8.10a)
dqv (x,t)
= 1 qv(x,t) qv (x,t)F(v(x,t)),
dt
(8.10b)
assuming that the weights only depend on the distance between interacting populations. For concreteness, the distributions we and wi are both taken to be Gaussians:
2
r
we
we (r) =
e 2e2 ,
2e2
r2
wi
wi (r) =
e 2i .
2i2
(8.11)
Following [83], assume that e > i (longer-range excitation) and fix the length
scale by setting e = 2, i = 1. Assuming that excitation spans a single hypercolumn, e should be the same approximate size as a hypercolumn, that is, of the order
200 m. We also take = 1 in units of the membrane time constant, which is typically of the order 10 ms.
8.3.1 Rivalry Waves in a Deterministic Neural Field Model

The next step is to interpret the binocular rivalry wave seen in the experiments of
Kang et al. [312, 313] as a traveling wave-front solution of the neural field equations (8.9) and (8.10), in which a high activity state invades the suppressed left-eye
332
network, say, while retreating from the dominant right-eye network; see Fig. 8.9.
Such a wave is defined as
u(x,t) = U(x ct),
v(x,t) = V (x ct),
(8.12)
where c is the wave speed and = x ct is a traveling wave coordinate together

with the asymptotic conditions
(U( ),V ( )) XL as ,
(U( ),V ( )) XR as
(8.13)
with U( ) a monotonically decreasing function of and V ( ) a monotonically increasing function of . Here XL (XR ) represents a homogeneous left-eye (right-eye)
dominant state. Given the asymptotic behavior of the solution and the requirements
of monotonicity, we see that U( ) and V ( ) each cross threshold at a single location, which may be different for the two eyes. Exploiting translation invariance, we
take U(0) = and V (d) = .
U()
V( )
Fig. 8.9 Sketch of a right-moving traveling wave solution in which a high activity state invades the
suppressed left-eye network while retreating from the dominant right-eye network. Wave is shown
in a moving frame = x ct
It is possible to construct an exact traveling wave solution of equations (8.9)

and (8.10) by taking f to be the Heaviside function (1.15) and making the adiabatic approximation that qu (x,t) = Qu , qv (x,t) = Qv with Qu , Qv constants [83].
The latter is based on the assumption that adaptation is sufficiently slow so that the
wave traverses cortex in a time Tl s . Substituting the traveling wave solution
(8.12) into (8.9a) and (8.10a) with fixed Qu , Qv and F(u) = H(u ) leads to the
equations
0

dU
+ U = Qu
we (
)d
Qv d wi (
)d
+ I (8.14a)
d

0
dV
c
+ V = Qv d we (
)d
Qu
wi (
)d
+ I. (8.14b)
d
333
where, for simplicity, we have set Iu = Iv = I. Multiplying both sides of (8.14a) and
(8.14b) by e /c , integrating with respect to and imposing the threshold conditions
gives

1 z/c
U( ) = e /c
e
d (z)dz I(1 e /c ) ,
(8.15a)
c 0

1 d z/c
V ( ) = e( d)/c
e
d (z)dz I(1 e( d)/c ) (8.15b)
c 0
for > 0, with and defined by

d (z) = Qu
d (z) = Qv

z
we (y)dy Qv
we (y)dy Qu
zd
zd
wi (y)dy.
(8.16)
wi (y)dy.
(8.17)
Finally, requiring that the wave solution remain bounded as (assuming c > 0)
yields the pair of threshold conditions

0
e d (cs)ds + I,

0
es d (cs)ds + I.
(8.18)
In the particular case Qu = Qv = 1 (no synaptic depression), the above equations

have no solution for c = 0, that is, there does not exist a traveling wave solution
[83]. This follows from setting Qu = Qv = 1 in (8.18) and subtracting the pair of
equations to obtain

0
es [d (cs) d (cs)] ds = 0.
The latter equation has no solution for c = 0, since
d (cs) d (cs) =
cs
cs
we (y)dy
csd
csd
wi (y)dy < 0
for all s [0, ). On the other hand, slow synaptic depression with Qu = Qv breaks
the symmetry of the threshold crossing conditions, leading to a unique solution for
c, d as a function of the network parameters. Moreover, one can establish that the
traveling front is linearly stable by constructing the corresponding Evans function
along similar lines to Sect. 7.1.2 [83]. The threshold conditions can also be used to
simplify the expressions for the wave profile, namely,
1 z/c
U( ) =
e
d (z + )dz + I
c 0

1
V ( ) =
ez/c d (z + d)dz + I.
c 0
(8.19a)
(8.19b)
334
Example plots of the wave speed are shown in Fig. 8.10. Baseline parameter values are chosen so that spontaneous oscillations and traveling fronts coexist as found
experimentally [312, 313]. The model wave speed is of the order c = 1 in dimensionless units, that is, c = e /2 where e is the range of excitation and is the
membrane time constant. Taking e to be of the order 200 m and to be of the
order 10 ms yields a wave speed of around 10 mm/s, which is consistent with the
speeds observed experimentally. (In the psychophysical experiments of Kang et al.,
binocular rivalry waves took approximately 0.8 s to traverse 2 degrees of the visual
field. The magnification factor in humans throughout the foveal region is approximately 0.4 cm/deg, which corresponds to 0.8 cm of cortex.) Figure 8.10 shows that
the speed of the wave is a decreasing function of the threshold and an increasing function of the input amplitude I; the latter is consistent with what is found
experimentally when the stimulus contrast is increased [312, 313]. Yet another experimental result that emerges from the model is that the wave speed covaries with
the frequency of spontaneous rivalry oscillations [83].
wavespeed c
b
5
2
2
1
0.05
0.10
threshold
0.15
0.20
0.24
input I
0.28
Fig. 8.10 Plot of wave speed c in units of e /(2 ) as a function of (a) the threshold and (b)
the external input strength I. The default parameters are taken to be wi = 1, we = 0.4, i = 1, e =
2, = 5, = 0.05, I = 0.24, Qu = 0.42, Qv = 0.25 and the corresponding wave speed is c = 1.2.
For this set of parameters, the network operates in a regime that supports both traveling waves
and homogeneous oscillatory solutions, that is, spontaneous switching occurs in the absence of
traveling waves
8.3.2 Effects of Extrinsic Noise in the Fast Activity Variables

As highlighted in Sect. 8.1, the switching between rivalrous images has a stochastic
component. This motivates studying a stochastic version of the neural field model
(8.9), which, in the adiabatic limit of fixed depression variables, takes the form of
an SPDE for the stochastic activity variables U(x,t) and V (x,t) [663]:

we (x y)F[U(y,t)]dy Qv
wi (x y)F[V (y,t)]dy + Iu dt
dU = U + Qu
1
2
+ g(U) dWu
(8.20a)

dV = V + qv
335
we (x y)F[V (y,t)]dy Qu
wi (x y)F[U(y,t)]dy + Iv dt
+ 2 g(V ) dWv .
(8.20b)
with Qu , Qv fixed. Here Wu , Wv represent independent Wiener processes such that

[see (7.158)]
dW{u,v} (x,t) = 0,
dWi (x,t)dW j (x
,t
) = 2i, jC([x x
] / ) (t t
)dtdt
,
(8.21)
where i, j = u, v and denotes averaging with respect to the Wiener processes. The
effects of extrinsic noise on binocular rivalry waves can be analyzed using the multiple time-scale method described in Sect. 2.6 for PDEs and Sect. 7.4 for neural fields
[663]. First, (8.20) and (8.20b) are rewritten so that the fluctuating terms have zero
mean according to Novikovs theorem. Then (U,V ) is expressed as a fixed wave
profile (U0 ,V0 ) that is displaced by an amount (t) from its uniformly translating
position = x c t and a time-dependent fluctuation (U1 ,V1 ) in the wave shape
about its instantaneous position:
U(x,t) = U0 ( (t)) + 1/2U1 ( (t),t),
V (x,t) = V0 ( (t)) + 1/2V1 ( (t),t).
(8.22a)
(8.22b)
The wave profile (U0 ,V0 ) and associated wave speed c are obtained by solving
the deterministic equations (8.14) with the decay terms U and V replaced by h(U)
and h(V ), where h(U) = U + C(0)g
(U)g(U). By carrying out a perturbation
expansion in and imposing a solvability condition, one finds that (t) is given by
a Wiener process with
(t) = 0,
(t)2 = 2D( )t,
(8.23)
The diffusion coefficient D( ) is

V1 ( )2 g(U0 ( ))2 + V2 ( )2 g(V0 ( ))2 d
,
D( ) =
2

V1 ( )U0 ( ) + V2( )V0 ( ) d
(8.24)
where V = (V1 , V2 )T is the null vector of the adjoint linear operator L = (Lu , Lv )T
Lu (B1 , B2 )
dB1
= c
+ h
(U0 )B1 + F
(U0 )Qu
d
F
(V0 )Qv
and
wi (
)B2 (
)d
we (
)B1 (
)d
(8.25)
336
Lv (B1 , B2 ) = c
dB2
+ h
(V0 )B2 + F
(V0 )Qv
d
F
(U0 )Qu
we (
)B2 (
)d
wi (
)B1 (
)d
.
(8.26)
It is important to point out that, in contrast to traveling front solutions of scalar

neural field equations (see Sect. 7.4), we are now considering composite wave solutions consisting of an invading front in the left-eye network, say, co-moving with a
retreating front in the right-eye network. Thus in addition to the center-of-mass of
the composite wave, which moves with speed c in the absence of noise, there is an
additional degree of freedom corresponding to the width of the composite wave.
[In the case of a Heaviside rate function, the width is determined by the threshold
crossing point d; see (8.18).] For simplicity, we assume that the width of the composite wave is only weakly affected by the noise; this is consistent with what is
found numerically [663].
8.3.3 Explicit Results for a Heaviside Rate Function

In order to illustrate the above analysis, we consider a particular example where the
mean speed c and diffusion coefficient D( ) can be calculated explicitly. That is,
set g(U) = g0U for the multiplicative noise term and take F(U) = H(u ). The
deterministic equations for U0 ,V0 can be solved along identical lines to (8.14a) and
(8.14b) and lead to the modified threshold conditions
1
( )

0
esd (cs/ ( ))ds + I,
1
( )

0
es d (cs/ ( ))ds + I,
(8.27)
where ( ) = 1 C(0)g20. The corresponding wave profiles are

1 z /c
e
d (z + )dz + I
c 0

1
V ( ) =
ez /c d (z + d)dz + I.
c 0
U( ) =
(8.28a)
(8.28b)
It immediately follows that both the speed c and displacement d depend on the
noise strength . In order to calculate the diffusion coefficient, it is first necessary
to determine the null vector V ( ) of the adjoint linear operator L . Substituting
F(U) = H(U ) and g(U) = g0U in (8.25) and (8.26) shows that the components
of V satisfy the simultaneous equations
( )
dV1
Qu
+ ( )V1 =
we (z)V1 (z)dz
d
|U0 (0)|

( d)
Qv

wi (z d)V2 (z)dz
|V0 (d)|
(8.29)
337
( )
dV2
Qu
+ ( )V2 =
wi (z)V1 (z)dz
d
|U0 (0)|

( d)
Qv
+
we (z d)V2 (z)dz.
|V0 (d)|
(8.30)
Proceeding along similar lines to Sect. 7.4, we make the ansatz that
V1 ( ) = A1 e /c H( ) B1e [ d]/c H( d),
V2 ( ) = A2 e /c H( ) + B2e [ d]/c H( d).
Substituting back into the adjoint equations yields algebraic conditions for the constant coefficients A j , B j :
Qu
(A1 e [0] B1e [d])
|U0
(0)|
Qv
(B2 i [0] A2i [d])
B1 =
|V0 (d)|
Qu
(A1 i [0] B1 i [d])
A2 =
|U0 (0)|
Qv
(B2 e [0] A2e [d]) ,
B2 =
|V0 (d)|
A1 =
where
j [x] =

0
ez /c w j (x + z)dz,
j = e, i.
(8.31)
(8.32)
(8.33)
(8.34)
(8.35)
Differentiating (8.28a) and (8.28b) with respect to and using (8.16) and (8.17)
show that
(8.36)
U0
(0) = Qu e [0] Qv i [d] < 0,
and
V0
(d) = Qv e [0] + Qui [d] > 0.
(8.37)
We have also used the fact that the weight distributions we (x), wi (x) are even functions of x. Substituting these derivatives into (8.31) and (8.34) gives
B2 =
Qv e [d]
A2 ,
Qu i [d]
B1 =
Qv i [d]
A1 .
Qu e [d]
(8.38)
It follows from (8.33) and (8.38) that A2 , B1 , B2 can all be expressed as constant
multiples of A1 , with the latter determined by normalizing the null vector.
It can be verified that the various assumptions used in the above analysis are reasonable by carrying out numerical simulations of the full system given by (8.20a),
(8.20b), (8.9a) and (8.9b) with Qu,v qu,v [663]. In particular it can be checked that
fluctuations in the separation between the left- and right-eye fronts are negligible
in the long time limit. An illustration of the progression of a stochastic composite front is shown in Fig. 8.11. The corresponding mean X(t) and variance X2 (t)
of wave position as a function of time are plotted in Fig. 8.12 using level sets as
338
0.4
0.3
0.4
0.3
0.2
0.1
0.2
0.1
- 0.1
10
15
20
x
- 0.1
10
15
x
20
10
15
20
0.4
0.3
0.2
0.1
0.4
0.3
0.2
0.1
5
- 0.1
10
15
20
x
- 0.1
Fig. 8.11 Effects of multiplicative noise in the activity variables. Simulation of the activity variables u(x,t) and v(x,t) for (a) t=0, (b) t=1, (c) t=2, (d) t=3. The continuous lines are the functions
U0 (x ct) and V0 (x ct) which were found by solving (2.97). The functions U0 and V0 were also
used for the initial condition. Parameter values were ai = 1, ae = 0.4, i = 1, e = 2, = 5, =
0.05, I = 0.24, = 0.006 with spatial and temporal grid sizes both being 0.01
-4
b x 10
14
13
_ 12
X(t)
11
10
9
2 6
X(t)
4
2
1
2
time t
2
time t
Fig. 8.12 (a) Mean and (b) variance of the wave position for an ensemble of 128 stochastic simulations. Parameters were the same as Fig. 8.11
explained in Sect. 7.4.1. It can be seen that X(t) varies linearly with t, consistent
with the assumption that there is constant speed wave, X(t) c t. The variance initially increases rapidly with respect to t, but then follows linear behavior consistent
with a diffusive-like displacement of the wave from its uniformly translating position at long time scales, X2 (t) 2D( )t. The initial sharp increase in the variance
results from the fact that the left and right fronts do move slightly with respect to
each other, resulting in higher-order behavior. However, one finds numerically that
this component of the variance is small and bounded so that it becomes negligible
as time increases.
339
a U(x,0)
0.4
f(T)
40
0.3
30
0.2
20
0.1
10
10
20
30
40
x
0.60
0.62
0.64
0.66
Fig. 8.13 (a) Snapshot of a stochastic traveling wave (only left-eye activity is shown). (b) First
passage time distribution for the wave to travel a distance L = 1, starting from the wave profile in
(a). Parameter values were ai = 1, ae = 0.4, i = 1, e = 2, = 5, = 0.05, I = 0.24, = 0.006
with spatial and temporal grid sizes both being 0.01.The best fit inverse Gaussian F (T ; , ) for
the histogram gives the parameters = 0.62, = 2200 which is in very good agreement with the
theoretical predictions of = L/c = 0.6, = L2 /D = 2100
The above analysis predicts that the wave position will follow a Brownian motion, which implies that the time taken for a wave to travel a distance L > 0 has a
distribution given by the standard first passage time formula for Brownian motion
with drift c. That is, let TL denote the first passage time for the wave to travel a
distance L: cTL + (TL ) = L given (0) = 0. Then the first passage time density is
given by an inverse Gaussian or Wald distribution:
L L2
f (TL ) = F (TL ; , ),
c D
where

F (T ; , ) =
2 T 3
1/2

(T )2
exp
.
2 2 T
(8.39)
(8.40)
Figure 8.13 shows the first passage time distribution generated over a large number
of simulations for an initial condition given by a fully developed stochastic traveling wave. Consistent with the analysis, the distribution can be fitted by an inverse
Gaussian. However, it has been assumed that the predominant source of noise is in
the activity variables. Now suppose that the main source of noise is in the depression variables. Once again assume that a traveling wave crosses cortex significantly
faster than the relaxation time of synaptic depression, so that qu and qv can be taken
to be constant with respect to time during passage of the wave. However, they are no
longer constant with respect to space nor with respect to trials, that is, qu = Qu (x)
and qv = Qv (x) with Qu , Qv random variables of x. So although the wave itself will
travel deterministically in a given trial, the functions Qu (x) and Qv (x) will be different across trials. In several experimental studies [312, 313] of binocular rivalry
waves, a dominance switch was induced about halfway through a normal dominance
cycle by locally changing the contrast in the depressed stimulus. The latter can be
represented by increasing the input strength I over a small region of the model. This
suggests taking qu and qv to evolve according to
340
s dqu (x,t) = [1 qu(x,t)] dt + dWqu (x,t),

s dqv (x,t) = [1 qv qv ] dt + dWqv (x,t)
(8.41a)
(8.41b)
over the time interval [t0 , T ]. It is assumed that a switch from left- to right-eye dominance occurs at t = t0 so that u(x,t) < and v(x,t) > for t (t0 , T ). The time T is
then chosen so that the system is about 2/3 of the way through a rivalry oscillation,
such that Qu (x) = qu (x, T ) and Qv (x) = qv (x, T ). (Similar results were obtained for
different cycle lengths and choices of T away from the beginning or end of a cycle.) Thus, the quenched random variables Qu (x) and Qv (x) are obtained by taking
a snapshot of two lines of independent OrnsteinUhlenbeck processes. Averaging
over many trials, it is found numerically that the first passage time distribution is
still well approximated by an inverse Gaussian distribution; see Fig. 8.14 and [663].
a
u(x,0)
0.4
0.3
0.2
10
20
30
40
f(T)
0.6
0.5
0.4
0.3
0.2
0.1
5
T
10
Fig. 8.14 Effects of quenched noise in the depression variables. (a) Initial condition given by
a right-eye dominant homogeneous steady state perturbed by a narrow square pulse of left-eye
activity. (b) First passage time distribution for a wave to travel a distance L = 6 starting from the
initial condition shown in (a). Parameter values were ai = 1, ae = 0.4, i = 1, e = 2, = 5, =
0.05, I = 0.24, = 0.037Q, where Q is the spatial average of the quenched depression variables
(to make the results comparable with multiplicative noise). The spatial grid size x = 0.1 and the
temporal grid step is t = 0.01. The solid curve in (b) is the best fit inverse Gaussian
In summary, we have illustrated one application of neural field theory, namely,

to the phenomenon of binocular rivalry waves. Formulating the problem in terms of
continuum neural field equations allows one to study the short time behavior associated with the propagation of eye dominance from an analytical perspective, both
for deterministic and stochastic waves. In the deterministic case, one finds that some
form of slow adaptation such as synaptic depression is needed in order to provide a
symmetry breaking mechanism that allows propagation of a binocular rivalry wave
[83]. That is, the equations for the left- and right-eye networks have to be different on the time scales during which traveling waves propagate. One implication of
this is that purely noise-driven switching between rivalrous states in the absence of
adaptation could not by itself generate rivalry waves, since it would not produce
bias in the appropriate direction. However, noise does have a significant effect on
binocular rivalry waves in the presence of slow adaptation. In particular, the analysis
predicts that motion of the wave in the presence of noise in the activity variables has
a diffusive component that leads to a distribution of first passage times given by an
inverse Gaussian. A similar distribution of first passage times occurs for quenched
noise in the depression variables [83].
8.4 Modeling Primary Visual Cortex Using 2D Neural Fields
341

The neural field model of binocular rivalry waves introduced in Sect. 8.3 is based
on a very simple model of V1. In particular, it is a 1D model that ignores any correlations between long-range horizontal connections and stimulus feature maps. Such
simplifications are reasonable as a first step, particularly when modeling the psychophysical experiments of Kang et al. [312, 313]. However, in order to account
for the dependence of wave velocity on the orientation patterns of annular stimuli
in the experiments of Wilson et al. [678], for example, it is necessary to consider a
more detailed, 2D model of V1 that takes into account such correlations. In the final
section of this chapter, we briefly describe one possible approach, which is based on
a neural field model of coupled hypercolumns [71, 79].
8.4.1 2D Planar Model

For the sake of illustration, consider a 2D scalar neural field (without any slow
adaptation) of the form
u(r,t)
= u(r,t) +
t
w(r, r
)F(u(r
,t))dr
.
(8.42)
Almost all 2D neural field models take w(r, r

) = w(|r r
|) so that w is invariant
with respect to the Euclidean group E(2) of rigid body transformations in the plane.
That is,
w(r, r
) = w( 1 r, 1 r
) = w(r, r
)
for all E(2). The Euclidean group is composed of the (semi-direct) product of
O(2), the group of planar rotations r R r and reflections (x, y) (x, y), with
R2 , the group of planar translations r r + s. Here

cos sin
(8.43)
, [0, 2 ).
R =
sin cos
However, one of the immediate implications of the existence of regularly repeating
feature maps and patchy horizontal (or feedback) connections (Sect. 8.1.1) is that
one can no longer treat the weight distribution w in the neural field equation (8.42) as
Euclidean invariant. That is, one has to consider a more general weight distribution
of the form
(8.44)
w(r, r
) = w(|r r
|) + w (F (r), F (r
)),
where F (r) denotes a cortical feature map, w represents the dependence of excitatory horizontal connections on the feature preferences of the presynaptic and
postsynaptic neuron populations, and is a positive coupling parameter. The local
connections span a single hypercolumn, whereas the patchy horizontal connections
link cells with similar feature preferences in distinct hypercolumns.
342
In the absence of long-range connections ( = 0), the resulting weight distribution is invariant under the action of the Euclidean group. However, the long-range
connections break Euclidean symmetry due to correlations with the feature map
F (r). A certain degree of symmetry still remains under the approximation that the
feature map is periodic. For example, Fig. 8.2b suggests partitioning V1 into a set of
hypercolumns organized around a lattice of orientation pinwheels. Therefore, suppose we treat the distribution of pinwheels as a regular planar lattice L. The resulting
weight distribution for = 0 is then doubly periodic with respect to L:
w(r + , r
+ ) = w(r, r
)
(8.45)
for all L. Additional symmetries may also exist depending on the particular form
of w . There are number of distinct ways in which w may depend on the underlying
feature map F . The first reflects the patchiness of the horizontal connections that
link cells with similar feature preferences. In the case of a periodic feature map, this
may be implemented by taking [65, 66, 526]
w (F (r), F (r
)) =
L
J (r r
),
(8.46)
where (r) is some localized unimodal function that is maximal when r = 0, thus
ensuring that presynaptic and postsynaptic cells with similar feature preferences
are connected. The width of determines the size of the patches and J , = 0
is a monotonically decreasing function of . In this particular example, the patchy
horizontal connections break continuous rotation symmetry down to the discrete
rotation symmetry of the lattice. On the other hand, continuous translation symmetry
(homogeneity) still holds, since w only depends on the relative separation r r
in
cortex.
However, if the anisotropy of horizontal connections is taken into account
(Fig. 8.3), then continuous translation symmetry is broken as well. That is, the
direction of anisotropy is correlated with the orientation preference map and thus
rotates periodically across cortex [56]. Anisotropy can be incorporated into the
model by modifying the weight distribution w along the following lines [75, 79]:
w (F (r), F (r
)) =
L
J AF (r
) () (r r
),
(8.47)
with
1
(H[ (r) | arg (r)|] + H[ (r) | arg (r) |]),
4 (r)
(8.48)
where (r) denotes the orientation preference map. The second term takes account
of the fact that [0, ) whereas arg [0, 2 ). The parameter (r) determines the
degree of anisotropy, that is, the angular spread of the horizontal connections around
the axis joining cells with similar orientation preferences. The degree of anisotropy
is also likely to depend on position r relative to pinwheels, since populations of
AF (r) () =
343
cells around pinwheels have zero average orientation preference so that one expects
the corresponding distribution of weights to be isotropic, in contrast to cells in the
linear zones of the orientation preference map; see Fig. 8.2. In conclusion, at the
submillimeter length scale, there is an approximately periodic modulation of the
synaptic connections that can lead to a periodically varying heterogeneous medium.
We have already explored one implication of this in a simpler 1D model of traveling
waves in a heterogeneous neural field (see Sect. 7.3), namely, that it can cause wave
propagation failure [79, 132, 332]. Another consequence of such an inhomogeneity
is that it can lead to the pinning of a spatially periodic pattern to the underlying
lattice of pinwheels [65, 66].
8.4.2 Coupled Hypercolumn Model

Treating the distribution of pinwheels as a regular lattice does not take into account
the considerable degree of disorder in the distribution of feature preferences across
cortex. One way to avoid such complexity is to collapse each hypercolumn into a
single point (through some form of spatial coarse graining) and to treat V1 as a
continuum of hypercolumns [78, 79]. Thus cortical position r is replaced by the
pair {r, F } with r R2 now labeling the hypercolumn at (coarse-grained) position
r and F labeling the feature preferences of neurons within the hypercolumn. Let
u(r, F ,t) denote the activity of a neuronal population at (r, F ), and suppose that u
evolves according to the neural field equation
u(r, F ,t)
= u(r, F ,t) +
t

R2
w(r, F |r
, F
)F(u(r
, F
,t))DF
dr
,
(8.49)
with DF
an appropriately defined measure on feature space. We decompose w into
local and long-range parts by assuming that the local connections mediate interactions within a hypercolumn, whereas the patchy horizontal connections mediate
interactions between hypercolumns:
w(r, F |r
, F
) = (r r
)w(F , F
) + J(|r r
|)AF
(r r
)w (F , F
),
(8.50)
where w(F , F
) and w (F , F
) represent the dependence of the local and longrange interactions on the feature preferences of the pre- and postsynaptic cell populations, and J(r) with J(0) = 0 is a positive function that determines the variation
in the strength of the long-range interactions with cortical distance. We have also
included the anisotropy factor AF of (8.48). The advantage of collapsing each hypercolumn to a single point in the cortical plane is that a simpler representation of
the internal structure of a hypercolumn can be developed that captures the essential tuning properties of the cells as well as incorporating the modulatory effects of
long-range connections.
344
For the sake of illustration, suppose that we identify F in (8.49) and (8.50) with
the orientation preference [0, ) of cells within a hypercolumn. The weight
distribution (8.50) is taken to have the form [79]
w(r, |r
,
) = (r r
)w(
)
(8.51)
+ J(|r r
|)P(arg(r r
) )w (
)
with
P( ) =
1
[H( | |) + H( | )|].
4
(8.52)
(Note that the direction arg(r r

) can be taken to be correlated with either or
,
since w is a sharply peaked function.) The functions w( ) and w ( ) are assumed
to be even, -periodic functions of , with corresponding Fourier expansions
w( ) = w0 + 2 wn cos2n
n1
w ( ) =
w0
+ 2 wn cos 2n .
(8.53)
n1
The distribution w ( ) is taken to be a positive, narrowly tuned distribution with

w ( ) = 0 for all | | > c and c /2; the long-range connections thus link cells
with similar orientation preferences. Equation (8.49) then describes a continuum
of coupled ring networks, each of which corresponds to a version of the so-called
ring model of orientation tuning [34, 74, 595]. If there is no orientation-dependent
anisotropy, then the weight distribution (8.51) is invariant with respect to the symmetry group E(2) O(2) where O(2) is the group of rotations and reflections on the
ring S1 and E(2) is the Euclidean group acting on R2 . The associated group action is
(r, ) = ( r, ),
(r, ) = (r, + )
(r, ) = (r, ).
E(2)
(8.54)
Invariance of the weight distribution can be expressed as
w(r, |r
, ) = w( 1 (r, )| 1 (r
,
)) = w(r, |r
,
)
for all where = E(2) O(2). Anisotropy reduces the symmetry group to
E(2) with the following shifttwist action on R2 S1 [64, 78]:
s (r, ) = (r + s, )
(r, ) = (R r, + )
(r, ) = (R r, ),
(8.55)
where R denotes the planar rotation through an angle and R denotes the reflection (x1 , x2 ) (x1 , x2 ). It can be seen that the discrete rotation operation comprises
345
a translation or shift of the orientation preference label to + , together with a

rotation or twist of the position vector r by the angle . The fact that the weight
distribution is invariant with respect to this shifttwist action has important consequences for the global dynamics of V1 in the presence of anisotropic horizontal
connections [79].
Finally, one can write down a 2D neural field model for binocular rivalry waves
based on a continuum of coupled hypercolumns for left- and right-eye preferences.
Ignoring the effects of anisotropy and taking an adiabatic approximation for the
depression variables, such a model takes the form
u
= u(r, ,t) +
t
+Qu
w(
)F(u(r,
,t)d
Je (|r r
|)F(u(r
, ,t)dr
Qv
v
= v(r, ,t) +
t
+Qv
2
2
0
Ji (|r r
|)F(v(r
, ,t)dr
w(
)F(v(r,
,t)d
Je (|r r
|)F(v(r
, ,t)dr
Qu
Ji (|r r
|)F(u(r
, ,t)dr
.
An interesting issue is then how to define a traveling wave solution for such a model
that combines orientation tuning within a hypercolumn due to the effects of the local
weight distribution w with a variation of activity across hypercolumns in the form
of a traveling front.
Part III
Development and Disease
Chapter 9
Waves in the Developing and the Diseased Brain
In this final chapter we discuss various wave phenomena in the developing and
diseased brain. We begin by describing two examples of intercellular developmental
waves: Ca2+ waves in radial glial cells, which coordinate cell division and cell
migration during early development, and spontaneous retinal waves, which are
thought to contribute to the activity-dependent development of circuits in visual
cortex. We then consider the possible role of traveling waves in the maturation of
neurons from sphere-like cells to polarized cells consisting of a single axon and
several shorter dendrites. In particular, we consider some new wave phenomena including wave-pinning and cytoskeletal actin waves. Switching to the case of pathological waves in the diseased brain, we focus on two examples: cortical spreading
depression, which is often associated with migraine auras, and epilepsy. In the former case, we review reactiondiffusion models of spreading depression that track
changes in the concentration of extracellular ions such as K+ , and show how kinematic wave theory can be used to model the progression of a migraine aura across
the visual field. It is also necessary to determine the effective diffusion of ions in the
extracellular space, which can be carried out using the theory of porous media, for
example. Epilepsy is a vast field in its own right, so we will focus on how changes in
network structure can result in abnormal cortical states such as seizure-like activity
or population bursting. We also consider recent work on characterizing normal and
abnormal brain states according to how close the cortex is to a critical state. This
leads to another approach to studying the propagation of activity in cortex, based
on the theory of branching processes. We finish by discussing the spread of protein
aggregates in neurodegenerative diseases such as Alzheimers disease, Parkinsons
disease, and Huntingtons disease. There is growing experimental evidence that
such spread has parallels with the neuropathogenesis of prion diseases, in which the
infected agent is a misfolded protein aggregate.
349
350
9 Waves in the Developing and the Diseased Brain
9.1 Intercellular Waves in Neural Development

9.1.1 Calcium Waves in Early Development
There is growing experimental evidence that intercellular Ca2+ waves play an important role in early neural development [146, 368, 498, 666, 697]. In order to
describe such findings, it is first necessary to give a short overview of the early stages
of development. During embryogenesis a strip of specialized cells called the notchcord induces epithelial cells of the ectoderm directly above it to transform into the
primitive nervous system known as the neuroepithelium; see Fig. 9.1. The latter then
wrinkles and folds over to form a hollow tube (the neural tube), which is the precursor of the brain and spinal cord. Meanwhile, the ectoderm and endoderm continue
to curve around and fuse beneath the embryo to create the body cavity, completing the transformation of the embryo from a flattened disk to a three-dimensional
body. Cells originating from the fused tips of the neuroectoderm (neural crest cells)
migrate to various locations throughout the embryo, where they initiate the development of the peripheral nervous system including the retina. As neurogenesis
a
Precursors of
neural crest
Ectoderm
Neural crest
Precursor of
neural tube
Ectoderm
Notochord
Mesoderm
Endoderm
b
Neural folds
Neural
crest
cells
Neural
tube
Body cavity
Fig. 9.1 Formation of the neural tube (cross view). See text for details. (a) Induction of neuroepithelium. (b, c) Folding of the neuroepithelium to form the neural tube. (d) Formation of 3D
body structure and migration of neural crest cells [Reproduced from Wikimedia Commons (public
domain)]
351
Fig. 9.2 Transformation of neuroepithelial cells in the neural tube into radial glial cells, which are
the progenitors of neurons and glial cells. Their radial projections also guide cell migration from
the ventricular zone (VZ) or inner surface of the neural tube to the outer surface. Cortical cells
are also produced in the subventricular zone (SVZ). During early neurogenesis, intercellular ATP
signaling is essential for the migration of SVZ progenitors from the VZ to the SVZ. During late
neurogenesis, ATP released from cortical radial glia mediates Ca2+ waves that are essential for VZ
proliferation (Adapted from [697])
proceeds, neuroepithelial cells of the neural tube transform into radial glial cells,
also known as progenitor cells, which ultimately give rise to neurons and glial cells;
see Fig. 9.2 and [514]. The cell body of radial glial cells is found in the ventricular zone (VZ), or inner surface of the neural tube, and has a radial extension that
projects near the border of the marginal zone, or outer surface. After asymmetrical
division, their daughter cells migrate along the radial projections and differentiate
into neurons. At the onset of cortical neurogenesis, a second proliferating population of progenitor cells appears in a new layer known as the subventricular zone
(SVZ), which generates both neurons and glial cells. Finally, towards the end of
cortical neurogenesis, radial glial cells retract their extensions to become astrocytes.
In summary, radial glial cells function as stem cells for neurogenesis and as scaffolding that guides the migration of neurons out of the ventricular zone. Moreover,
the process of radial migration occurs in an inside-out manner. In other words,
cells that arrive the earliest (the oldest ones) form the deepest layer of the cortex,
whereas the late-arriving (the youngest) neurons form the outermost layer.
An experimental study of embryonic rat cortical slices provides strong evidence
that ATP-mediated calcium waves, which propagate between radial glial cells in the
ventricular zone, are involved in the modulation of cell proliferation [666]. The predominant mechanism for the intracellular Ca2+ waves appears to be paracrine signaling, in which gliotransmission leads to the release of ATP via gap hemichannels,
which diffuses and binds to metabotropic P2Y receptors. This activates a G-proteinmediated IP3 signaling pathway, which leads to an increase in intracellular Ca2+ ;
see Sect. 4.5. As suggested in [666], one possible function of ATP-mediated Ca2+
waves during neurogenesis is that it could coordinate the cell cycle. Recall that the
cell cycle consists of four distinct phases or stages G1SG2M; see Fig. 9.3. Here
G1 and G2 are gap phases, while S is the synthesis stage where DNA replication
352
calcium
waves
G0
G2
G1
M
Fig. 9.3 Phases of the cell cycle: mitosis (M), Gap1 (G1), Gap2 (G2), and synthesis (S). G0 represents the quiescent phase where cells exit the cell cycle either temporarily or permanently. Calcium
waves can coordinate cells to enter the S phase
occurs. The M phase consists of two tightly coupled processes: mitosis, in which
the cells chromosomes are divided into two, and cytokinesis, in which the cells
cytoplasm divides in half forming distinct cells. Activation of each phase is dependent on the proper progression and completion of the previous one. Cells that have
temporarily or reversibly stopped dividing are said to have entered a state of quiescence called the G0 phase. It is known that increases in intracellular Ca2+ are
required for a cell to enter the S and M phases [42]. For example, during the G1
phase Ca2+ may drive a cell to reenter the S phase rather than exit the cell cycle.
Experimentally blocking intercellular Ca2+ waves in the ventricular zone using ATP
receptor antagonists appears to decrease cell proliferation without blocking Ca2+
elevations in individual cells [666]. Hence, one function of the intercellular waves
could be to generate Ca2+ increases that help a nearby cluster of cells to enter the
S phase in a coordinated manner. This hypothesis is also supported by the experimental observation that ventricular zone waves appear to be triggered by cells in the
S phase [666]. In addition to S phase coordination, it is possible that intercellular
Ca2+ waves also regulate differential gene expression [246], as suggested by the
observation that the frequency of Ca2+ waves increases as neurogenesis proceeds.
Intercellular Ca2+ waves are also found in early developmental stages of structures in the peripheral nervous system such as the retina [146, 498, 499]. The
developing retina is a layered structure whose outermost layer, the retinal pigment
epithelium (RPE), is essential for the development of the underlying neural layers. The layer closest to the RPE is called the ventricular zone where, similar to
the developing cortex, progenitor cells divide and differentiate to form the neurons and glia of the fully developed retina. ATP-mediated Ca2+ waves propagate
in the RPE, possibly triggered by a single epithelial cell, and arise via the same
paracrine signaling mechanism as Ca2+ waves in developing cortex; see Fig. 9.4.
Diffusing ATP molecules released from the RPE also bind to P2Y receptors on the
progenitor cells within the ventricular zone, which enhances cell proliferation and
speeds up mitosis. It also appears that ATP from the RPE, possibly in combination with the action of neurotransmitters from differentiated neurons, can induce
353
trigger cell
Ca2+
Ca2+
retinal pigment
epithelium
IP3
ATP
P2R
hemichannels
P2R
ATP
P2R
ventricular
zone
progenitor
cell
Fig. 9.4 ATP-mediated Ca2+ wave in radial pigment epithelium. ATP released from hemichannels
diffuses and binds to metabotropic receptors (P2Rs) in neighboring epithelial cells leading to the
release of intracellular Ca2+ from internal stores. The ATP also binds to P2Rs of progenitor cells
in the ventricular zone, thus promoting cell proliferation
Ca2+ waves in the ventricular zone [498, 614]. The propagation of ventricular Ca2+
waves, which is probably mediated by gap-junction coupling between progenitors
rather than paracrine signaling, can also modulate cell proliferation within a cluster
of progenitor cells.
9.1.2 Retinal Waves in Later Development

The fully developed retina located at the back of the eye converts patterns of light
activity from a visual image into neural spiking activity. The first stage is the conversion of light into graded neural activity by photoreceptors. Such activity is then
modulated by both vertical and horizontal neuronal pathways via synaptic and gapjunction coupling, before reaching the retinal ganglion cells (RGCs). These form the
output layer of the retina, which transmits the resulting pattern of spiking activity
along the optic nerve to the brain; see also Sect. 8.1.1. Following the very early
stages of development discussed at the end of the previous section, but before
photoreceptors are visually responsive, the retina has already formed a layered
354
b Stages I/II
a Stages I/II
ONL
OPL
INL
AcH
IPL
AcH
A
GCL
optic nerve
glutamate
AcH
optic nerve
Fig. 9.5 Simplified schematic of neural circuits that mediate retinal waves during (a) stages I/II
and (b) stage III. The retina is partitioned into three main layers: outer nuclear layer (ONL), inner
nuclear layer (INL), and ganglion cell layer (GCL). There are also two layers of interconnecting processes: inner plexiform layer (IPL) and outer plexiform layer (OPL). During stages I/II,
spontaneous retinal waves between retinal ganglion cells (G) are driven mainly by the interactions
between ganglion cells and cholinergic amacrine cells (A). In stage III, waves are predominantly
driven by glutamatergic bipolar cells (B), which are connected to photoreceptors in the ONL
structure with distinct classes of differentiated neurons, including RGCs. Although

the RGCs are not yet driven by visual stimuli, they are far from silent, firing spontaneous bursts of action potentials that travel across the retinal surface as retinal
waves; see the reviews [190, 222, 682]. Interestingly, such spontaneous rhythmic
activity occurs in many other developing neural circuits [49, 471], including the
cochlea [633], spinal cord [260, 472], cerebellum [661], hippocampus [32], and cortex [326]. There is growing evidence that the resulting spontaneous depolarizations,
and corresponding increases in intracellular Ca2+ (via voltage-gated ion channels
rather than release from intracellular stores), play an important role in driving many
aspects of development, including axon guidance, cell differentiation, local circuit
formation, and the emergence of sensory maps. These phenomena have been extensively studied in the retina, where upregulating or blocking retinal waves has
a major affect on the development of visual circuits [190, 315, 416, 682] and has
led to the suggestion that retinal waves are crucial for the normal development of
the retinotopic map and the segregation of eye dominance in the cortex. However,
their precise role still remains controversial. We briefly summarize some of the main
properties of retinal waves at various stages of development and describe a number
of theoretical models; see also [222].
355
Retinal waves can initiate at any retinal location due to spontaneous activation of
a local cluster of cells and form non-repeating propagation boundaries that are determined, in part, by refractory effects. The waves can be categorized into three distinct
developmental stages [615]. During stages I and II, the retina has only a few functional cell types and the waves arise via a coupled network of RGCs and amacrine
cells forming the inner retina; see Fig. 9.5a. Stage I waves are mediated by the neuromodulator adenosine and by gap junctions, since very few synapses have formed.
Stage II waves first appear at the onset of synaptogenesis and are primarily mediated
by the neurotransmitter acetylcholine; they have a characteristic refractory period so
that following passage of a retinal wave over a section of retina, another wave can
only propagate across the same section after tens of seconds have elapsed. Stage I
waves tend to be faster and occur more frequently than stage II waves. The transition to stage III occurs when vertical bipolar cells form synaptic connections with
RGCs and the neurotransmitter glutamate becomes involved in mediating retinal
waves; see Fig. 9.5b. Although acetylcholine continues to be involved as well, the
nicotinic receptors found in stage II are replaced by muscarinic receptors. One finds
that stage III waves are comparable in speed to stage II waves, but are smaller and
less frequent. In certain species (e.g., the mouse), the retina first becomes sensitive to light near the onset of stage III, whereas in other species (e.g., the rabbit)
light sensitivity occurs later. Spontaneous retinal waves generally disappear after
eye opening, although in the turtle they can persist in diminished form for several
weeks. Finally, retinal waves tend to have speeds in the range 100300 m/s, which
are much slower than waves found in disinhibited cortical slices, for example; see
Chap. 7. On the other hand, the speeds are much faster than processes based on the
extracellular diffusion of signaling molecules.
Most computational studies of retinal waves have focused on developmental
stage II and are based on networks of cholinergically coupled amacrine cells.
The first such model also included a layer of RGCs that acted as a low-pass filter of
the spontaneous activity generated in the amacrine layer [94]. The basic mechanism
of the model was that amacrine cells would spontaneously depolarize and excite
neighboring cells, after which they would enter a refractory state of fixed duration.
A propagating wave was initiated whenever a sufficiently large local cluster of
amacrine cells spontaneously became active and the surrounding medium contained
a sufficient number of non-refractory cells. In spite of the models simplicity, containing only two free parameters (an activation threshold and the probability of
spontaneous activation), it could generate patterns of non-repeating waves that were
statistically similar to experimental data. However, the model was not robust to
changes in parameters, that is, it had to be finely tuned. An alternative model of
retinal waves was proposed by Godfrey and Swindale [223], who took the refractory period of amacrine cells to be activity dependent. This provided a more robust
model and did not require an RGC layer to generate biologically plausible waves.
A more biophysically detailed model of the roles of cholinergic coupling, noise, and
refractory period in the generation of retinal waves in amacrine circuits has subsequently been developed by Hennig et al. [271]. An individual amacrine cell is now
described in terms of a conductance-based model, which consists of a voltage-gated
Ca2+ current that supports spontaneous bursting, and a slow after-hyperpolarization
356
(AHP) current that prevents further bursting. Moreover, the duration of the AHP
current is taken to increase with the level of depolarization. Thus, during passage
of a wave, strong synaptic coupling between adjacent amacrine cells generates a
stronger depolarization than intrinsic bursts in an isolated cell, which in turn causes
a longer-lasting refractory period due to AHP currents. Thus, the effective network
connectedness is dynamically regulated in such a way that the network can operate
close to a percolation threshold, which means that the distribution of propagation
events (frequency and duration) exhibits power-law behavior consistent with some
experimental observations [271]; see also Sect. 9.5.
9.2 Traveling Waves, the Cytoskeleton, and Cell Polarization

9.2.1 Neuronal Polarization and the Cytoskeleton
Cell polarization refers to the breaking of cell symmetry and is an essential process in various cell types, including migration of fibroblasts during wound healing,
asymmetric cell growth or division, and the functional maturation of neurons. Neurons are among the most polarized cells due to their compartmentalization into a
thin, long axon and several shorter, tapered dendrites (Sect. 1.1). Experimental studies of neuronal polarization have mainly been performed on dissociated, embryonic cortical and hippocampal neurons or on postnatal cerebellar granule neurons.
Such studies have identified three basic stages of polarization [13, 452, 509, 581];
see Fig. 9.6. Cultured neurons initially attach to their substrate as round spheres
Stage 1
Stage 2
Stage 3
neurites
axon
growth cone
Fig. 9.6 Stages of neuronal polarization. A neuron attaches to the substrate as a round sphere
surrounded by actin-rich lamellipodia and filopodia (stage 1). Growth cones formation by the consolidation of lamellipodia leads to the establishment of several neurites (stage 2). One neurite starts
to elongate rapidly and forms the axon (stage 3)
357
filopodium
central domain
lamellipodium
actin bundle
microtubule
actin network
Fig. 9.7 Schematic diagram of growth cone showing cytoskeletal structures
surrounded by actin-rich structures such as lamellipodia (mesh-like actin sheets)

and filopodia (packed actin bundles) (stage 1). Lamellipodia then coalesce to form
growth cones, followed by the establishment of several short processes, called neurites (stage 2). Eventually one of the neurites starts to grow more rapidly to become
the axon (stage 3), while the other neurites remain short and develop into dendrites
at later stages of maturation. During the development of a neuron, the cytoskeleton
plays a fundamental role (as does intracellular transport [289]). Since neurite formation is a key event in neuronal polarization, it is instructive to describe the role of
the cytoskeleton during the elongation of neurites.
The growth cone at the mobile tip of an elongating neurite or axon contains
microtubules within a central domain (C-domain) and actin filaments within the
peripheral domain (P-domain); see Fig. 9.7. The microtubules provide the structural
backbone of the shaft and a substrate for intracellular transport to the growth cone.
They polymerize with their growing ends pointed towards the leading edge of the
growth cone. Actin filaments within the P-domain form filopodia and lamellipodia
that shape and direct the motility of the growth cone. In both structures, the actin filaments face with their barbed (growing) ends towards the plasma membrane. Polymerization of actin filaments towards the leading edge causes the extension and protrusion of the growth cone. This creates a force that pushes the actin network and the
tightly linked plasma membrane backwards (retrograde flow) and hinders the invasion of the microtubules into the P-domain. The retrograde flow is also enhanced
by the action of myosin molecular motors, which drag the actin cytoskeleton back
towards the C-domain where actin filaments depolymerize at their pointed ends.
If there is a balance between actin polymerization in the P-domain and retrograde
flow, then there is no elongation. However, signals from surface adhesion receptors
bound to a substrate can suppress the retrograde flow of actin filaments, shifting the
balance towards polymerization-driven forward motion that involves both actin filaments and microtubules. (This signaling mechanism can in turn be modulated by
external chemoattractant cues during axon guidance [394, 470].) The growth of a
neurite proceeds in three stages [225, 432]. First, enhanced polymerization in the
P-domain pushes the growth cone forward (protrusion phase). Second, an actin-free
358
upstream signals
RhoA
GTP
Rac
GTP
GTP
Cdc42
GDP
Rho
RhoGaps
ROCK
GTPase
cycle
PAC
WAVE
WASP
RhoGefs
LIM-K
Rho
GTP
F-actin
profilin
cofilin
Arp 2/3
downstream effectors
actin cytoskeleton
Fig. 9.8 (a) The Rho GTPase cycle. (b) Simplified signaling pathways from GTPases RhoA, Rac,
and Cdc42 to actin accessory proteins cofilin and Arp 2/3 that regulate actin polymerization in the
growth cone (Redrawn from Luo [394])
zone is created by a combination of cytoskeletal restructuring and actin disassembly so that the retrograde flow no longer hinders the advancement of microtubules
into this zone (engorgement phase). Third, vesicles and organelles enter the newly
created zone via intracellular transport, resulting in an elongation of the neurite shaft
(consolidation phase).
One class of signaling molecule that appears to play an important role in the
polarization and migration of many different cell types is Rho GTPases [235, 254].
These act as intracellular molecular switches that cycle between an active GTPbound form and an inactive GDP-bound form; see Fig. 9.8a. Guanine nucleotide
exchange factors (RhoGEFs) facilitate the conversion from GDP-bound to GTPbound form, whereas GTPase-activating proteins (RhoGAPs) enhance GTP hydrolysis and are thus negative regulators. RhoGEFs and RhoGAPs are both regulated
by upstream signals. A major downstream target of the Rho GTPase signaling pathways is the actin cytoskeleton within the growth cone [254, 394, 452]. (There is
growing evidence that the GTPases also regulate microtubules [13], although we
will not discuss this here.) A variety of actin accessory proteins mediate the different components of actin dynamics within the growth cone and can be activated by
the various signaling pathways associated with neuronal polarization. We describe
a few of the major players; see [452, 483] for more details.
First, the actin-related proteins 2 and 3 (Arp2/3) complex stimulates actin polymerization by creating new nucleation cores. The Arp2/3 complex is activated by
members of the WiskottAldrich syndrome protein (WASP) family protein (WAVE)
complex, which localizes to lamellipodia where it facilitates actin polymerization.
The WAVE complex is activated by the small GTPase Rac1, which modulates the
actin cytoskeleton dynamics by controlling the formation of lamellipodia. Filopodia
formation, on the other hand, is regulated by another member of the small GTPases,
called Cdc42. A second important accessory protein is cofilin, a member of the
359
actin-depolymerizing factor (ADF)/cofilin family, which modifies actin dynamics

by increased severing and depolymerization of actin filaments via its binding
to the non-barbed (pointed) ends. Cofilin is inhibited when phosphorylated by
LIM-kinase, which is itself activated by Pak-kinase. Since the latter is a downstream target of Rac1 and Cdc42, it follows that Rac1 and Cdc42 inhibit actin
depolymerization by downregulating cofilin. At first sight, one would expect the
reduction of cofilin to counter the affects of Arp2/3. However, the extent to which
cofilin enhances or reduces cell protrusion depends on the spatial and temporal
scale over which it operates [150]. It turns out that the overall effect of Rac1 and
Cdc42 is to increase actin dynamics thus promoting neurite growth. This is opposed
by the action of a third type of GTPase known as RhoA, which tends to stabilize
the actin network by activating ROCK-kinase, which promotes profilin (an actinbinding protein that catalyzes polymerization) and suppresses cofilin. A summary
of the basic signaling pathways is given in Fig. 9.8b.
9.2.2 Wave-Pinning and Cell Polarity in a Bistable

ReactionDiffusion Model
We now turn to a reactiondiffusion model of cell polarity developed and analyzed
by Keshet et al. [305, 438, 439]; see also [478, 482]. Although the model was originally constructed in order to understand the mechanisms of cell polarization in
motile cells such as fibroblasts, it is possible that similar mechanisms occur prior
to the sprouting of a new neurite during stage 2 (neuritogenesis). The basic idea is
to use reactiondiffusion equations to describe the spatiotemporal dynamics of Rho
GTPases such as Rac1, Cdc42, and RhoA, which are known to cycle between active
membrane-bound forms and inactive cytosolic forms. One finds that a transient and
localized stimulus can be spatially amplified, leading to a robust partitioning of the
cell into a front and back, where the concentration of the active forms is high and
low, respectively. Mathematically speaking, the emergence of cell polarization can
be understood in terms of front propagation in a bistable reactiondiffusion system
(Sect. 2.2) with the following additional features [438, 439]: (1) the inactive and
active states have unequal rates of diffusion; (2) the total amount of each GTPase is
conserved. Consequently, a local stimulus induces a propagating front that decelerates as it propagates across the cell so that it becomes stationary, a process known
as wave-pinning; the stationary front persists in the absence of the stimulus and represents a polarized cell. Note that the wave-pinning mechanism is distinct from the
diffusion-driven Turing mechanism found in other reactiondiffusion models of cell
polarization [417, 482, 609]. One of the basic differences is that in the latter class
of models, a homogeneous state becomes unstable to arbitrarily small fluctuations
and the nonlinear reaction terms support the growth of a spatially varying pattern
via a Turing stability. For a recent application of the Turing mechanism to neurite
formation, see [420].
360

membrane
Rho-GTP
+
Rho-GDP
cytosol
x=0
x=L
Fig. 9.9 Schematic diagram of wave-pinning model showing spatial distribution of membranebound active and cytosolic inactive forms of Rho GTPase
The simplest version of the model considers a single Rho GTPase that can transition between inactive and active forms diffusing in a bounded 1D domain of length L
[438]; see Fig. 9.9. Let a(x,t) and b(x,t) be the concentrations of the active/inactive
states. Then
a
2a
= Da 2 + f (a, b),
t
x
b
2b
= Db 2 f (a, b),
t
x
(9.1a)
(9.1b)
Since the rate of diffusion of the membrane-bound (active) state is significantly

slower than that of the cytosolic (inactive) state, Da Db . The nonlinear function
f (a, b) represents the difference between the rates of activation and inactivation of
the Rho GTPase. Assuming there is cooperative positive feedback in the activation
of the protein, which is modeled as a Hill function of index 2, then

a2
(9.2)
ka.
f (a, b) = b k0 + 2
K + a2
It can be checked that for a range of uniform concentrations of the inactive state,
bmin < b < bmax , the space-clamped version of the model exhibits bistability with
two stable fixed points a (b) separated by an unstable fixed point a0 (b). Equation
(9.1) is supplemented by no-flux boundary conditions at the ends of the domain:

a
b
= 0,
= 0.
(9.3)
x x=0,L
x x=0,L
It follows that there is mass conservation of the total amount of Rho GTPase, that is,
L
0
(a + b)dx = C.
(9.4)
The mathematical explanation of wave-pinning proceeds as follows [438].

First, since Db Da and there are no-flux boundary conditions, one can assume
that b rapidly diffuses to establish a uniform concentration within the bounded
domain [0, L]; b then changes on a slower time scale as the a dynamics evolves
361
(quasi-steady-state approximation). Thus, on short time scales b can be treated as a

fixed global parameter of a scalar equation for a(x,t) given by (9.1a). Suppose that
initially bmin < b < bmax , so (9.1a) is bistable. On an infinite domain, the bistable
equation supports the propagation of a traveling front linking the stable fixed point
a+ (b), a (b) (see Sect. 2.2). That is, for < x < there exists a monotonically
decreasing solution a(x,t) = A( ), = x ct with lim A( ) = a+ (b) and
lim A( ) = a (b). Moreover the wave speed satisfies c = c(b) with [see also
(2.19)]

a+
c(b) =
f (a, b)da
( A/ )2 d
(9.5)
Note that the wave speed depends on the global parameter b. Since the denominator of
(9.5) is always positive, the sign of c(b) will depend on the sign of
I(b) aa+ f (a, b)da, which has a geometrical interpretation in terms of the difference between the area of the curve y = f (a, b) above the straight line y = k a
and the area below; see Fig. 9.10. In the case of a sufficiently sharp front that is
away from the boundaries, these results carry over to the bounded domain [0, L].
Now suppose that a transient stimulus near the edge of the cell at x = 0 triggers at
time t = 0 a traveling front as described above. This implies that bmin < b(0) < bmax
and I(b(0)) > 0. As the front starts to propagate into the interior of the cell, it converts a greater fraction of the domain from a a (b) to a a+ (b). From the
conservation condition (9.4), it follows that the approximately uniform concentration b(t) of the inactive state decreases, eventually reaching a critical value bc ,
bmin < bc < bmax , for which
I(bc )
a+
a
f (a, bc )da = 0,
(9.6)
and wave-pinning occurs. The basic steps are illustrated in Fig. 9.10.
Asymptotic Analysis of Wave-Pinning. As shown by Mori et al. [439], it is possible to analyze wave-pinning in more detail using a multi-time-scale analysis. The first step is to
nondimensionalize (9.1):
a
2a
= 2 2 + f (a, b),
t
x
2b
b
= D 2 f (a, b),
t
x
(9.7a)
(9.7b)
with x [0, 1] and 0 < 1. In order to look at the dynamics on short time scales, set
= t/ and introduce the asymptotic expansions a a0 + a1 + . . . and b b0 + b1 + . . ..
This yields the following pair of equations for a0 and b0 :
a0
= f (a0 , b0 ),
b0
2 b0
= D 2 f (a0 , b0 ).
(9.8a)
(9.8b)
reaction rate f
a
b
a
b
concentration
x=0
x=0
x=L
reaction rate f
concentration
x=0
x=L
reaction rate f
concentration
362
bc
x=L
Fig. 9.10 Schematic diagram explaining the mechanism of wave-pinning developed in [438].
A sequence of snapshots of the traveling front (left column) showing that as the front advances
into the domain, the background concentration b of the inactive state decreases so that the front
decelerates until it becomes stationary. The corresponding geometric construction of I(b) (right
column), which is given by the difference of the shaded areas, shows that I(b) is initially positive
but vanishes at the critical value bc
Suppose that bmin < b0 < bmax so that f (a0 , b0 ) is bistable with respect to a0 . Equation
(9.8a) implies that a0 a+ (b0 ) or a0 a (b0 ) at each point x [0, 1]. Assume initial
conditions in which there exists a single sharp transition layer linking a+ to a (a traveling
front).
Let (t) be the position of the transition layer with respect to the slower time scale t. Divide
the domain [0, 1] into inner and outer regions, with the latter given by [0, (t) O( ))
( (t) + O( ), 1]. Taking the limit 0 in (9.1), with a a0 + a1 + . . ., b b0 + b1 + . . .,
and restricting x to the outer layer, we have
0 = f (a0 , b0 ),
(9.9a)
2 b0
0 = D 2 f (a0 , b0 ).
x
(9.9b)
Adding these two equations immediately implies that 2 b0 / x2 = 0. Combining this with
the no-flux boundary conditions, it follows that

bL (t) 0 x < (t) O( ),
b0 (x,t) =
(9.10)
bR (t) (t) + O( ) < x 1.
363
with bL,R (t) independent of x. The corresponding solution for a0 is

a+ (bL ) 0 x < (t) O( ),
a0 (x,t) =
a (bR ) (t) + O( ) < x 1.
(9.11)
In order to determine the solution in the inner layer near the front, introduce the stretched
coordinate = (x (t))/ , and set
A( ,t) = a([x (t)]/ ,t),
B( ,t) = b([x (t)]/ ,t)
(9.12)
Introduce the asymptotic expansion A A0 + A1 + . . . and similarly for B and . Substitution into (9.7) and expansion to lowest order in then gives
2 A0 d 0 A0
+
+ f (A0 , B0 ) = 0,
2
dt
(9.13a)
2 B0
= 0.
2
(9.13b)
The solution of (9.13b) is of the form

B0 = 1 (t) + 2 (t),
(9.14)
where 1,2 (t) are determined by matching the inner (B0 ) and outer (b0 ) solutions according to
(9.15)
lim B0 ( ) = bL , lim B0 ( ) = bR .

The matching conditions can only be satisfied if B0 is constant in the inner layer, which
means that the inactive state is uniform throughout the whole domain, B0 = b0 . The next
step is to determine the inner solution for A0 , given that B0 is -independent. That is, we
have to solve the boundary value problem (9.13a) with matching conditions:
lim A0 ( ) = a+ (b0 ),
lim A0 ( ) = a (b0 ).
(9.16)
The inner solution thus corresponds to the standard front solution of a bistable equation
with
a+ (b0 )
f (a, b0 )da
d 0
a (b )
c(b0 ) = 0
,
dt
( A0 / )2 d
(9.17)
see (9.5).
The final step is to incorporate the conservation condition (9.4), which to lowest order in
becomes
1
0
a0 dx + b0 = C.
(9.18)
The integral term can be approximated by substituting for a0 using the outer solution:
1
0
a0 dx =
(t)O( )
0
a0 dx +
1
(t)+O( )
a0 dx + O( )
= a+ (b0 )0 (t) + a (b0 )(1 0 (t)) + O( ).
364
Combining the various results, the analysis of a traveling front solution of (9.7) reduces to
the problem of solving the ODE system [439]
d 0
= c(b0 ),
dt
b0 = C a+ (b0 )0 (t) a (b0 )(1 0 (t)).
(9.19)
These equations can be used to establish that the front slows down and eventually stops to
form a stable stationary front. In particular, differentiating the second relation in (9.19) with
respect to t shows that

1+

da+ (b0 )
da (b0 )
d 0
db0
(1 0 )
= [a+ (b0 ) a (b0 )]
.
0 +
db
db
dt
dt
(9.20)
Differentiating the condition f (a (b), b) = 0 with respect to b and imposing the condition
that f (a, b) is bistable, it can be established that 1 + da /db > 0. Since 0 < 0 < 1 and
a+ (b0 ) > a (b0 ), it follows from (9.20) that db0 /dt and d 0 /dt have opposite signs. Hence,
as the front advances, b0 decreases until the front stalls at a critical value bc for which
c(bc ) = 0. The corresponding front position is c with
bc = C a+ (bc )c a (bc )(1 c ).
(9.21)
Explicit conditions for wave-pinning can be obtained if the reaction term f is taken to be a
cubic
f (a, b) = a(1 a)(a 1 b),
(9.22)
rather than the Hill function (9.2). In this case, the speed of the front solution in the inner
layer can be calculated explicitly (see Sect. 2.2), so that (9.19) becomes
b0 1
d 0
= ,
dt
2
b0 = C (1 + b0 )0 .
(9.23)
It follows that the wave stops when b0 = 1 bc and the stall position is c = (C 1)/2.
Finally, note that Mori et al. [439] also extend the asymptotic analysis to the case of multiple
transition layers and carry out a bifurcation analysis to determine parameter regimes for
which wave-pinning occurs.
One interesting issue is to what extent the wave-pinning mechanism differs from
the phenomenon of wave propagation failure due to spatial discretization, as occurs
in myelinated axons, for example (Sect. 2.5). This is particularly important given
that any numerical simulation of the wave-pinning model involves the introduction
of a spatial grid, and the wave becomes more sensitive to discretization effects as
it slows down. A careful numerical study has shown that wave-pinning and propagation failure are distinct effects [655]. In the same study, a stochastic version of
the wave-pinning model was also considered, which takes into account fluctuations
in the number of active and inactive molecules at low concentrations. It was found
that when the total number of molecules is lowered, wave-pinning behavior is lost
due to a broadening of the transition layer as well as increasing fluctuations in the
pinning position.

wave-pinning model
actin wave model
Rho-GTP
+
Rho-GDP
365
Rho-GTP
F-actin
Rho-GDP
Fig. 9.11 Schematic diagram showing the basic pathways involved in the wave-pinning model
[438] and actin wave model [286]
9.2.3 Actin Waves

There are a number of intracellular wavelike phenomena associated with the maturation of single neurons. First, localized Ca2+ transients occur in the growth cone
of developing neurons [233, 234, 245]. Although they spread passively by diffusion
rather than actively, they are often referred to as growth cone Ca2+ waves. These
waves regulate the rate of neurite extension, which is inversely proportional to their
frequency, since elevations in intracellular Ca2+ concentration above baseline slow
any growth. Second, a neuron-wide Ca2+ wave propagating from the growth cone of
an axon to the soma mediates reversal of neuronal migration in response to a repulsive extracellular cue (Slit-2) [247]. It is hypothesized that Slit-2 receptor activation
at the leading edge of the growth cone triggers the Ca2+ wave, which propagates
to the soma, resulting in a front-to-rear gradient of Ca2+ concentration. This gradient in turn causes (by some yet-to-be-identified mechanism) a redistribution of
the GTPase RhoA towards the rear of soma, a process necessary for establishing
the reversed direction of migration. (One possibility is that the calcium wave acts
as a trigger for a polarization wave similar to the one discussed in Sect. 9.2.2.) We
will focus on another type of wave, namely, an actin wave propagating along the
axon shaft of neurites during cell maturation. In developing hippocampal neurons,
neurites undergo periodic extensions and retractions, which correlate with the propagation of growth cone-like structures that originate in the soma and propagate down
the neurites [194, 542, 543]. The waves travel at a speed of around 3 m/min, transporting actin and associated proteins such as cofilin down the neurite shaft to the
growth cone. Waves occur in all neurites during maturation stage 2, but become
more frequent in the emerging axon during the transition from stage 2 to stage 3
[194]. This coincides with an engorgement and increased dynamics of the axons
growth cone. Interestingly, growth cone-like actin waves have also been observed
during the regeneration of injured axons in cultured hippocampal neurons [155].
Actin waves are found in a number of other cell types including Dictyostelium
[643], neutrophils [665], and fibroblasts (see the review [7]). Indeed, most current
models of actin waves are based on studies of nonneuronal cells. They also tend to be
rather complicated, containing details regarding the distribution of actin lengths or
orientations, for example [103, 157, 643]. Here we will consider a simpler model of
actin waves that extends the wave-pinning model of Mori et al. by coupling the Rho
366
GTPase signaling module to a module representing the assembly and disassembly of

actin polymers (F-actin) [286]; see Fig. 9.11. More specifically, active Rho GTPase
promotes the polymerization of F-actin which feeds back to enhance the inactivation
of Rho GTPase (negative feedback). In order to incorporate the effects of negative
feedback, the kinetic rate function (9.2) is modified according to

a3
u
+
s
k
s
f (a, b, u) = b k0 + 3
a,
(9.24)
1
2
K + a3
u0 + u
where u is the concentration of F-actin and s1 , s2 , u0 are additional parameters.
If s2 = 0 and s1 = 1, then we recover the previous wave-pinning model with f
having a hill coefficient of n = 3 rather than n = 2. (Although n = 2 also works, the
choice n = 3 gives more robust waves). It remains to specify the F-actin dynamics.
It is assumed that F-actin is nucleated from an infinite reservoir of actin monomers
with concentration M, at a rate that is proportional to the concentration of active
Rho GTPase. Thus, neglecting the diffusion of F-actin,
u
= [kn M ks u],
t
(9.25)
where kn is the nucleation rate and ks is the disassembly rate.

Extensive numerical simulations show that the actin wave model supports a wide
range of spatiotemporal dynamics not found in the wave-pinning model, including
solitary traveling pulses, periodic wave trains, persistent reflecting waves, and more
exotic localized patterns [286]. In order to characterize the parameter dependence
of these various patterns, Holmes et al. [286] considered the nonlinear stability of
homogeneous steady states with respect to non-infinitesimal localized perturbations.
This analysis identified two distinct stability regimes: an excitable regime where
a pattern can only be triggered by a perturbation whose amplitude exceeds some
threshold and an unstable regime where arbitrarily small noise can induce a pattern.
Traveling pulses occur in the excitable regime, where the negative feedback due to
F-actin plays a role analogous to the refractory variable in the FitzHughNagumo
equations (see Chap. 2). On the other hand, wave trains occur in the unstable regime,
and reflecting waves occur in both regimes. As with other models of actin waves,
the motivating examples are from nonneuronal cells. However, it is possible that the
solitary traveling pulse could represent the growth cone-like wave seen in developing neurites [194, 542, 543].
9.2.4 Diffusion-Based Model of Neurite Symmetry Breaking

A major unresolved issue is how a single axon is specified among equally potential
neurites, that is, what is the underlying symmetry breaking mechanism for neurite differentiation? During stage 3, the cytoskeleton of the emerging axonal growth
cone becomes more dynamic and less dense (due to the action of Rac1 and Cdc42,
367
say), whereas the actin cytoskeleton of the other growth cones remains more stable
and maintains a denser network (due to the action of RhoA, say). Consequently,
microtubules are hindered from entering the growth cone of the inhibited neurites,
whereas they can advance into the P-domain of the axon thus stimulating elongation.
An emerging hypothesis [9, 13, 193, 629] is that neurite extension and retraction are
controlled by positive and negative signaling molecules (morphogens) respectively,
which diffuse within neurites to regulate the behavior of actin filaments and microtubules or to influence cargo transport. In particular, it is hypothesized that prior
to neurite differentiation, positive and negative feedback mechanisms are balanced;
this balance can be broken by an external cue or by spontaneous intrinsic fluctuations, resulting in the enhancement of positive feedback within a single neurite that
elongates to form an axon; simultaneously, the activated neurite generates a strong
negative feedback signal that prevents other neurites from forming a second axon. A
discussion of various candidate signaling cascades that could support autocatalytic
(positive feedback) processes in growth cones can be found elsewhere [9, 13]. Here
we will describe a particular model of neurite differentiation due to Toriyama et
al. [629], which is based on the transport of a protein called Shootin1. Shootin1 is
thought to be an upstream regulator of various signaling molecules responsible for
neuronal polarization [576]. The model is motivated by the experimental observation that Shootin1 is actively transported from the soma to growth cones as discrete
boluses within actin waves [628]. The active anterograde transport combines with
diffusive retrograde transport in such a way that Shootin1 accumulates in the growth
cones in a neurite length-dependent manner [629]. Moreover, during the stage 2/3
transition, Shootin1 predominantly accumulates in the nascent axon. The observation that Shootin1 can induce neurite outgrowth by regulating actin retrograde flow
[576] suggests that the asymmetric accumulation of Shootin1 is one of the critical
events in neurite differentiation.
Consider a developing neuron that consists of N neurites labeled i = 1, . . . , N,
each having length Li ; see Fig. 9.12a. Let C0 denote the concentration of Shootin1
in the soma and Ci the concentration in the ith growth cone. The diffusion coefficient
of Shootin is around 350 m2 /min so that the diffusion time constant = L2 /D is
around 1 min for a typical neurite length of L = 20 m. Given that the volumes of
the cell body (VS ) and growth cone (V ) are large compared to the volume of a thin
neurite shaft, it follows that changes in Shootin1 concentration within the growth
cone and soma occur on a slower time scale than the formation of a concentration
gradient along a neurite. Hence, one can assume as a first approximation that the
diffusive retrograde flow is in quasi-equilibrium. This means that the diffusive flux
Ji along the jth neurite is approximately uniform so that the net flux from the growth
cone is
D
(9.26)
Ji (t) = (Ci (t) C0 (t)).
Li
For simplicity, the propagation of actin waves is not modeled explicitly; rather,
the waves are treated as sequences of discrete stochastic events corresponding to
368
anterograde transport
retrograde transport
neurite growth
growth cone
soma
b
koff
Fs(C)
F1(L)
kon
microtubule
growth cone
Fig. 9.12 Model of diffusion-based neurite growth due to Toriyama et al. [629]. (a) A concentration gradient of Shootin1 is set up along each dendrite due to the combination of active anterograde
flow and length-dependent retrograde diffusion. Accumulation of Shootin1 in a growth cone sets
up a positive feedback loop for neurite growth. Neurite competition for resources leads to symmetry breaking where one neurite grows at the expense of all the others. (b) Neurite growth is
regulated by the balance between a Shootin1-induced traction force FS (C) and a length-dependent
neurite tensional force F1 (L)
Gaussian-like packets of Shootin1 delivered to a neurites growth cone at random

times T jn and amplitudes anj . Thus, the active transport in neurite j is represented as
a sum of discrete events
j (t) =
VS
anj g(t T jn ).
V
n
(9.27)
with g(t) a normalized Gaussian of width . The amplitudes and delivery times are
modeled by gamma distributions fitted to the experimental data [629]. Combining
(9.26) and (9.27), the Shootin1 concentration in the jth growth cone evolves as
AD
dCi
=
(Ci C0 ) + j (t).
dt
V Li
(9.28)
(For simplicity, the growth cone volume V and cross-sectional shaft area A is taken
to be the same for each neurite.)
The next component of the model specifies how the accumulation of Shootin1
in the growth cone regulates neurite growth; see Fig. 9.12b. It is assumed that
369
growth is determined primarily by the balance between two mechanical forces: a

Shootin1-dependent traction force FS (C) and a neurite length-dependent tensional
force F1 (L). The former is taken to be a Hill function
FS (C)
Ch
Ksh + Ch
(9.29)
for constants Ks , h, while the latter is approximated as a logarithmic Hill-like

function,
ln(L/L0 )
(9.30)
F1 (L)
ln(L/L0 ) + ln(K1 /L0 )
for constants K1 , L0 . The growth velocity of the ith neurite is determined as
dLi
= (kion M kioff ),
dt
(9.31)
where is the size of a tubulin monomer (the basic building block of microtubules),
M is the intracellular concentration of tubulin, and kion and kioff are the rates of polymerization and depolymerization, respectively. From the thermodynamics analysis
of a simple ratchet model of polymerization [434], it can be shown that the ratio of
the on and off rates depends on the net force F = F1 (L) Fs (C) according to
koff (F) koff (0) F /kB T
,
= on e
kon (F)
k (0)
(9.32)
where kB is the Boltzmann constant and T is absolute temperature. Fitting to experimental data suggests that the on rate is approximately independent of F. Hence,
setting Fi = F1 (Li ) Fs(Ci ), kion = kon and kioff = koff (Fi ) = koff eFi /kB T , we have

dLi
= kon M koff e[Fs (Ci )F1 (Li )] /kB T .
dt
(9.33)
As it stands, if C0 and M were fixed (reflecting an unlimited supply of tubulin and

Shootin1), then each neurite would independently undergo a positive feedback loop
in which an increase in neurite length Li would slow the rate of retrograde diffusion,
the growth cone concentration Ci of Shootin1 would increase, and this would in
turn speed up the rate of growth. However, the neurites actually compete for a finite
amount of resources. If S is the total amount of Shootin1 in the neuron and V0 is the
volume of the soma, then
S = V0C0 + V
Cj.
(9.34)
j=1
Hence, as a neurite grows, it depletes the level of Shootin1, which inhibits the
growth of the other neurites. There is also competition for tubulin monomers, which
can be expressed as
T = Vtot M +
L j,
j=1
(9.35)
370
where Vtot = V0 + NV , T is the total amount of cellular tubulin and is a parameter

that determines how much tubulin is needed to build a neurite of a given length.
Finally, experimental data indicates that both T and S are upregulated, that is, they
are slowly increasing functions of time that can be fitted by a time-dependent function of the form S(t) s + s /(1 + st 2.5 ) for constants s , s , s , and similarly
for T (t). Numerical simulations of the full model, with parameter values based on
studies of cultured hippocampal neurons, robustly exhibits spontaneous symmetry
breaking and the formation of an axon [629]. The model also reproduces various
atypical behaviors observed in cultured neurons. For example, applying mechanical
tension to a single neurite during stage 2 results in its axonal specification, overexpression of Shootin1 induces the formation of multiple axons, and repression of
Shootin1 inhibits polarization.
Finally, note that the Toriyama et al. model can be viewed as one of a class
of models for neurite differentiation that involve competition among neurites for
a pool of proteins involved in neuronal polarization. The first models of this type
[479, 548, 640] simply took the rate of elongation to be proportional to the intracellular concentration of the protein (possibly tubulin) at the neurite tip. One prediction
of these models is that the rate at which a neuron polarizes to form an axon should
increase with the number of neurites. However, a recent experiment establishes that
the time to polarize is independent of the number of neurites [680], which is consistent with another recent model of neurite differentiation based on the transport
of the protein HRas rather than Shootin1 [193]. It turns out that all of the models
exhibit independence of the number of neurites N, provided that the expression of
the relevant protein increases with N [680], which was one of the assumptions in
[193].
9.3 Spreading Depression

Cortical spreading depression (SD) is a complex wave of transient depolarization
of neurons and glial cells that propagates slowly across cortical and subcortical gray
matter at speeds of 15 mm/min [93, 365, 407, 461, 596]. It occurs as a response to
various forms of brain damage or pathology [237], including focal ischemia, traumatic brain injury, and seizure activity. There is also strong evidence that SD is a
correlate or cause of migraine with aura, where a light scintillation in the visual field
propagates before disappearing to be replaced by a sustained headache [358, 639].
Although SD does not usually damage brain tissue, if it arises repeatedly during
stroke or head trauma, then it may promote neuronal damage. One of the characteristic features of SD is a large increase in extracellular K+ and a dramatic decrease
in extracellular Ca2+ and other ions. Simultaneously, the DC surface potential of
electroencephalogram (EEG) undergoes a negative shift of around 19 mV that lasts
30 s, followed by a smaller but longer positive voltage shift. At the cellular level, the
slow shift in potential coincides with a period of membrane depolarization followed
by a longer period of hyperpolarization. A comprehensive picture of the precise
371
physiological and chemical processes underlying the generation of SD is still

lacking. However, a number of mechanisms are thought to contribute, including diffusion and membrane transport of ions, chemical synapses and neurotransmitters,
spatial buffering, gap junctions, and osmotic effects [433].
Although SD waves are associated with populations of cells, they have several
properties in common with propagating action potentials in the axons of single
neurons (Chap. 2): (1) they are essentially all-or-none phenomena; (2) they travel
with approximately constant speed and wave form; (3) they involve membrane
depolarization; (4) they exhibit refractoriness; (5) multiple waves can be generated
from a sustained stimulus and colliding waves can annihilate. Phenomenologically
speaking, the general characteristics of a solitary SD wave can be reproduced by
appropriately parameterized FitzhughNagumo equations, for example, with the
principal dependent variable being extracellular K+ concentration rather than membrane potential. However, such a model fails to give insights into the underlying
physiological mechanisms and is thus not useful for developing clinical treatments
for associated pathologies such as migraine. Indeed, there are a number of significant differences between SD waves and action potentials: (1) space and time
scales are much longer for SD waves; (2) SD wave propagation mainly involves
ionic fluxes through synaptic rather than non-synaptic membrane; (3) there are substantial changes in extracellular ion concentrations during passage of an SD wave;
(4) the recovery phase of SD involves active transport mechanisms such as ion
pumps.
9.3.1 TuckwellMiura Model

The first physiologically based mathematical model of SD waves was introduced
by Tuckwell and Miura [638]. It centers on the idea that an initial large increase
in extracellular K+ would depolarize presynaptic terminals, resulting in the release
of neurotransmitter. This would then open K+ channels in the postsynaptic membrane of a neuron, leading to the flux of additional K+ into the extracellular space.
The resulting regenerative process would support the propagation of an SD wave.
The model of Tuckwell and Miura is a bidomain continuum model (see Sect. 4.2),
in which the concentration of ions in the extracellular and intracellular domains are
taken to coexist at every point in space. This is a reasonable approximation given
that the spatial scale of an SD wave is much larger than the size of individual neurons. Ions are free to diffuse over significant distances within the extracellular space,
whereas intracellular ions can only diffuse within a cellular domain that is negligible compared to the scale of SD waves. Thus the effective diffusivity of intracellular
ions is taken to be zero. Another assumption of the model is that the equilibrium
membrane potentials for neighboring presynaptic and postsynaptic membranes are
approximately the same and are thus taken to have the common value V (x,t). This
can be justified by first noting that during passage of an SD wave, the changes in
intracellular ion concentrations are negligible, since the ratio of the intracellular to
372
Fig. 9.13 Electron micrograph of a small region of the cerebral cortex of a rat illustrating the small
volume fraction of the extracellular space, which is indicated by the black areas between cells.
(Note, however, that the extracellular space may have been reduced in size as a consequence of
the histological processing.) The asterisk indicates a dendritic spine, which is joined by a synaptic
cleft to a presynaptic terminal containing several small round vesicles filled with neurotransmitter.
Some of the round profiles adjoining the dendrite are likely to be unmyelinated axons cut in cross
section. The scale bar represents a distance of approximately 1 m (Adapted from [461])
extracellular volumes is large; see Fig. 9.13. Second, the local equilibrium membrane potential of a neuron is given by the HodgkinKatz formula
out
K + pNa Naout + pClCl in

kB T
ln
V=
.
(9.36)
q
K in + pNa Nain + pClCl out
Here (Kout , Naout , Clout ) and (Kin , Nain , Clin ) denote the extracellular and intracellular concentrations of potassium, sodium, and chloride, respectively, q is the charge
of a proton, and pNa , pCl are the ratios of the sodium and chloride membrane permeabilities to the potassium permeability. It follows that the membrane potential V (x,t)
simply tracks the extracellular ion concentrations, which are common to presynaptic
and postsynaptic neurons.
in
Let cout
j and c j denote the extracellular and intracellular concentration of ion
species j, with j = 1, 2, 3, 4 corresponding to sodium, chloride, potassium, and calcium, respectively. The basic model equations are then of the general form (in one
space dimension) [638]
cout
2 cout
j
j
+ g j (V V j ) + Pj ,
= Dj
t
x2
cinj
=
[g j (V V j ) + Pj ] .
t
1
(9.37a)
(9.37b)
Here D j is the extracellular diffusion coefficient of the jth ion species, V j is the
corresponding Nernst potential for an ion with charge modulus q j ,
373
Vj =
kB T
ln
qj
cout
j
cinj

,
(9.38)
g j is an effective conductance, and Pj represents the contribution of active transport

mechanisms (pumps). Conservation of ions implies that the non-diffusive rate of
change of each extracellular concentration is the negative of the rate of change of
the corresponding intracellular concentration, scaled by a factor /(1 ), where
is the volume fraction of the extracellular space. In principle, the contribution
to the conductance g j could come from channels at both synaptic and non-synaptic
sites. However, it is assumed that exchange of ions between the extracellular and intracellular domains during passage of an SD wave occurs predominantly at synaptic
sites. First, the Ca2+ conductance is taken to be presynaptic and a sigmoidal function of voltage
(9.39)
g4 = g4 (V ) g0 (1 + tanh[ (V + )]),
for constants g0 , , . A large increase in the extracellular K+ concentration generates a positive shift in the equilibrium potential V given by (9.36). The resulting
depolarization leads to an influx of presynaptic Ca2+ , which causes the release of
neurotransmitter. (For simplicity, only one type of neurotransmitter is considered
here.) The rate of release is taken to be proportional to the calcium current. However,
since such an event lasts a small time interval of size r , the effective concentration
of extracellular neurotransmitter is approximated by
T (x,t) = k4 r g4 (V (x,t))(V (x,t) V4 ),
(9.40)
where k4 is a constant. On the other hand, the conductance associated with the other
ions is assumed to arise from the binding of neurotransmitter to postsynaptic receptors according to
(9.41)
g j (x,t) = k j T (x,t), j = 1, 2, 3,
with constants k j . Finally, the pump rates are given by exponential functions of the
ion concentrations:
out
(9.42)
Pj = f j (1 er j (c j c j ) ),
for constants f j , r j , c j with c j , the resting extracellular concentration. Note that
f1 < 0 for potassium and f j > 0 for all the other ions, since the potassium pump
maintains a higher intracellular concentration. Equations (9.36)(9.42) determine
the full version of the TuckwellMiura model [638]. It is also possible to consider
a simplified version of the model, in which the sodium and chloride concentrations
are kept fixed.
Using physiologically based parameter values, the TuckwellMiura model
produces a wave speed of around 1 mm/min and a wave form consistent with
experimental findings. However, it is necessary to take into account the effects of
tortuosity, that is, the effective increase in path length for diffusion of ions around
cells. This introduces a scaling factor of around 1.5 for the diffusion coefficient
of ions in aqueous solution. One way to estimate the tortuosity of brain tissue as
well as the volume fraction of the extracellular space is to use the theory of porous
374
media [461, 616]; see Sect. 9.3.2. The heterogeneous and anisotropic nature of
brain tissue means that extending the TuckwellMiura model to higher dimensions
is nontrivial, since one would need to determine an effective diffusion tensor. Nevertheless, the model has the advantage of being relatively simple and amenable to
mathematical analysis. For example, one of the interesting mathematical features
of the model is that there exists a family of rest states that depend continuously on
a set of parameters. This degeneracy implies that traveling waves have a number of
properties that are not present in the case of isolated rest points [686]. For example,
a localized disturbance can lead to a pair of outward propagating waves that slow
down and reverse direction before ultimately colliding and annihilating each other.
Since the introduction of the TuckwellMiura model, which focused on extracellular diffusion and membrane ion currents, a number of other mechanisms have been
identified as possible contributors to the initiation and propagation of SD waves:
1. Gap junctions. Experiments have shown that pharmacological blockade of gap
junctions slows or even prevents the propagation of retinal SD waves [359,
406, 448]. As we described in Sect. 4.5, there is extensive gap-junction coupling between astrocytes. However, it would appear that SD waves still occur
in the absence of astrocytes [360]. More specifically, destroying many astrocytes
using the toxin fluorocitrate only reduces the effectiveness of SD by 50%. This
motivated Shapiro [568] to develop a detailed biophysical model of SD waves, in
which gap-junction coupling between neurons played a major role in the transport of K+ . Unfortunately, there is no evidence for such widespread gap-junction
coupling between neurons. In addition, it is possible that after application of
toxins, there are still a sufficient number of intact astrocytes to support some
level of SD.
2. Neuronastrocyte networks. There is growing experimental evidence that SD
waves involve both neurons and astrocytes acting together via the release of ATP
and glutamate. First, intercellular Ca2+ waves precede SD but blocking these
waves by gap-junction inhibitors does not prevent SD [501]. On the other hand,
SD is blocked by antagonists to NMDA receptors suggesting a role for glutamate
signaling. Second, a large amount of ATP is released from astrocytes during SD
[118, 558]. This is consistent with the observation that the ATP antagonist heptanol completely prevents SD in hippocampus [360], presumably by blocking
paracrine signaling. It is also known that glutamate can stimulate the release
of ATP from astrocytes [667] and, conversely, activation of purinergic receptors
by ATP can induce glutamate (and possibly ATP) release from astrocytes [187].
These experimental results have led to a model of SD waves [38], in which a
wave of ATP propagates with the astrocyte Ca2+ wave, resulting in an accompanying wave of glutamate release from the astrocytes. The glutamate acts on
NMDA receptors in the postsynaptic membrane of neurons, which triggers the
large depolarization characteristic of SD together with the presynaptic release of
additional glutamate from neurons. The latter acts on neighboring astrocytes to
initiate the release of additional ATP. Hence, one has a regenerative mechanism
for the propagation of ATP and glutamate, resulting in an SD wave. Note that one
375
major assumption of the model is that glutamate and ATP can be released from
astrocytes via Ca2+ -independent pathways [250, 659].
3. Spatial buffering. Another well-known role of astrocytes is in the regulation
of extracellular K+ by several mechanisms including the action of Na+ /K+
ATPase pumps and K+ ion channels [109, 211, 601]. The spatial buffering of
K+ by astrocytes has only been studied within the context of small perturbations
away from the resting state rather than the large excursions that occur during
the passage of an SD wave. Nevertheless, it is possible that spatial buffering can
account for the range of wave speeds observed during SD. The central mechanism for spatial K+ buffering via ion channels is that a local increase in extracellular K+ causes a local depolarization of glial cell membranes, which then
spreads electrotonically to other glial cells via gap junctions. There is thus an influx of extracellular K+ into glia at the original site of enhanced extracellular K+
and an efflux of K+ into the extracellular space at distal sites where extracellular
K+ is close to the resting level. In other words, there is a rapid net transport
of K+ from sites of high extracellular concentration to sites of low extracellular
concentration, which is mediated by ion channels and gap-junction coupling of
astrocytes.
4. Cell swelling. A number of models of SD waves have taken into account the
effects of cell swelling that results from ions crossing neuronal membranes
leading to osmotic stresses [568, 596]. This reduces the volume fraction of the
extracellular space.
9.3.2 Brain Tissue as a Porous Medium

Diffusion in the extracellular space occurs in the narrow gaps between cells (neurons and glia) and is analogous to diffusion in a porous medium consisting of two
phases such as water in the sandy bed of a river [503]. The connection with porous
media can be exploited to estimate the effective diffusivity of ions moving in the
extracellular space; see the reviews [461, 616]. One immediate observation is that
if one were to draw a line through a section of brain tissue (see Fig. 9.13), one
would find that the extracellular concentration is a discontinuous function of position along the line. In order to construct continuum reactiondiffusion models such
as TuckwellMiura, it is necessary to carry out some form of spatial averaging using
either homogenization theory (see Sect. 4.6) or volume averaging. Although the former approach is more rigorous, we will describe the latter approach here following
[461], since it is more intuitive. The first step is to partition brain tissue into a set
of representative elementary volumes (REVs) that are sufficiently large to include
a sufficient range of cellular elements for averaging, yet sufficiently small to allow
a local description of averaged quantities. Consider an arbitrary REV with center x
relative to a fixed origin and local coordinates about the center; see Fig. 9.14. (One
can view and x as microscopic and macroscopic coordinates, respectively.) Suppose that each REV is composed of two phases , corresponding, respectively, to
376
the extracellular and intracellular domains and occupying the volumes V ,V . Let
V = V +V be the total volume of the REV and let = V /V be the volume fraction of the extracellular space. Derivation of an effective diffusion equation in the
extracellular space now proceeds by applying a spatial averaging theorem to Ficks
law and the particle conservation equation.
Spatial averaging theorem. This relates the average of the divergence of a microscopic
quantity to the divergence of a macroscopic (averaged) quantity [240, 503]. Let F denote
a microscopic state vector that is a function of time t and position x + within the REV
centered at x. Applying the divergence theorem to the volume occupied by extracellular
space V , we have

V
FdV =
F n dS +

S
F n dS ,
(9.43)
where n is the unit normal vector on the boundary of the phase pointing outwards.
Introduce the index function such that (x + ) = 1 if x + lies in the phase and is
zero otherwise. We can then write the first term on the right-hand side as V (F ) n dS
where V is the exterior boundary of the REV. Applying the divergence theorem to the
whole volume V then gives

V
FdV =

V
(F )dV +

S
F n dS .
(9.44)
Since F is a function of x + , we can replace by x and take the divergence operator

outside the volume integral with respect to :

V
FdV = x

V
(F )dV +

S
F n dS .
(9.45)
Finally, noting that the index function restricts the first integral on the right-hand side to
the domain V , we have (after dividing through by V ), the spatial averaging theorem [240]
F = F +
1
V

S
F n dS ,
(9.46)
O
S
Fig. 9.14 A representative elementary volume (REV) composed of and phases and centered at
x. The interface between the phases is denoted by S and the surfaces that make up the boundary
of the REV are denoted by S and S , respectively
377
where
f =
1
V
f dV
(9.47)
for any function f . The spatial averaging theorem for a scalar quantity may then be
obtained by setting F(x + ) = (x + )a with a an arbitrary constant vector:
= +
1
V

S
n dS .
(9.48)
Ficks law for the diffusive flux J in a homogeneous medium takes the form
J = Dc,
(9.49)
where D is the diffusion coefficient and c(x,t) is the concentration of the diffusing
particles at position x and time t. Combining this with particle conservation,
c
+ J = 0,
t
(9.50)
then recovers the standard diffusion equation. Now suppose that Ficks law only
applies to particles located within the extracellular space, which is a multiply connected 3D domain. One would then need to solve the associated diffusion equation
within each simply connected component of extracellular space supplemented by
surface boundary conditions. An alternative approach is to apply the spatial averaging theorem to Ficks law and the conservation equation to obtain a macroscopic
version of the diffusion equation. First, we have
J = Dc = Dc +
1
V
cndS,
(9.51)
where the area integral is over the interfaces A between the extracellular and intracellular spaces. The nontrivial part of the analysis is determining the surface integrals,
which corresponds to the problem of determining the effective diffusion tensor by
solving the cell equations in homogenization theory; see Sect. 4.6. Nevertheless, the
surface integral reduces to the general form [374]
1
V

A
cndS = (K I)c,
(9.52)
where K is a second-order tensor and I is the identity. It follows that Ficks law for
a porous medium takes the form
J = D c,
(9.53)
where D = DK is an effective diffusion tensor. In the special case of an isotropic

medium, this reduces to scalar diffusion coefficient D = D/ 2 , where is
the tortuosity. Similarly, volume averaging the particle conservation equation
gives

c
+ J = 0.
(9.54)
t
378
Applying the spatial averaging theorem to the divergence of the flux implies that
J = J +
1
V

A
J ndS.
In the case of impermeable cell membranes, J n = 0, whereas for permeable membrane

J n = g(c cin )
(9.55)
where g is an effective permeability and cin is the local mean intracellular concentration. For simplicity, we focus on the impermeable case. Assuming that the interfaces
A are fixed, that is, ignoring osmotic effects and cell swelling, then

c
c
.
=
t
t
Collecting the various results finally gives in the isotropic case
where =
c
= D 2 c k(c cin),
t
A dS/V
(9.56)
is a surface-to-volume ratio.
9.3.3 Migraine Auras

In this section we describe the relationship between SD and migraine aura. See also
the review by Dahlem and Chronicle [144]. A characteristic feature of migraine is
an episodic headache of moderate to severe intensity, which is usually unilateral
(one side of the brain) and pulsating in nature. In the West, the lifetime prevalence
of migraine is around 33 % in women and 13 % in men. In the case of migraine
with aura, the migraine attack is preceded by a sensory hallucination called aura
that usually lasts for less than half an hour. Consider, in particular, the classical
visual aura that occurs in approximately 2530 % of patients. Such an aura typically
starts with a scotoma forming near the central region of the visual field; a scotoma
(Greek for darkness) is an area of visual field in which there is partially or severely
reduced visual acuity that is surrounded by a field of normal vision. The scotoma,
which initially has indistinct boundaries, expands into a sickle or C-shaped region
whose advancing edge contains a zigzag pattern of bright lines and shapes oriented
at 60 to each other. The patterns are often called fortification patterns, since they
are reminiscent of baroque city ramparts. An illustration of a fortification pattern
is shown in Fig. 9.15, which has been superimposed upon a photograph of moose
taken in the authors backyard. The fortification pattern becomes more distinct as it
expands into the visual periphery, before disappearing at the edge of the visual field.
The aura rarely returns during a single migraine event and leaves no noticeable permanent disruption of vision. Note that the fortification pattern at the leading edge of
379
Fig. 9.15 Illustration of a fortification pattern within the scotoma of a classical migraine aura
a scotoma reflects a short first phase of intense neuronal excitation, which precedes
the prolonged phase of depressed neural activity that is associated with the scotoma.
The existence of two phases is consistent with early studies of SD waves [93].
One of the important features of the migraine aura is that it clearly corresponds
to a physiological wave propagating across the visual cortex. Moreover, the detailed
structure of the fortification pattern appears to reflect the functional organization of
the cortex, including its topographic organization and finer-scale feature maps; see
Sect. 8.1.1. Indeed, it is now generally accepted that the fortification pattern provides
a direct view of the layout of orientation columns in human visual cortex. Interestingly, many of these observations were anticipated by the psychologist Lashley in
1941 [363]. The first link between migraine aura and SD waves was made in 1945
[366], and there is now considerable experimental evidence supporting such a link
[237]. There is also growing evidence that patients susceptible to SD waves have
hyperexcitable visual cortices [441].
The progression of the scotoma across the visual cortex can be modeled using
a kinematic theory of wave propagation in two-dimensional excitable media [145].
Such a model describes the propagation of a wave front in terms of the motion
of curves with free ends, without detailed knowledge of the underlying reaction
diffusion equations [426]. Wave propagation is analyzed by assuming that the normal velocity c of the wave depends solely on the front curvature and by taking into
account an additional tangential velocity that describes the growth rate of the front
at its open ends. To motivate the curvature dependence of the normal velocity, consider a 2D version of the FitzHughNagumo equations (Sect. 2.3)
v
= f (v, w) + D2 v,
t
w
= w + v.
1
t
(9.57a)
(9.57b)
380
In particular, consider a circular target pattern in which the leading edge of a traveling pulse forms a circle of radius R at time t, with R much larger than the width
of the pulse. K = 1/R is then the curvature of the circular front. Introducing polar
coordinates (r, ) with the origin at the center of the target pattern, we have
D v
v
2v
= f (v, w) +
+D 2,
t
r r
r
w
= w + v.
1
t
(9.58a)
(9.58b)
Since v/ r is negligible outside the transition layers of the pulse, we can replace
D/r by D/R on the right-hand side of (9.58a). Writing the circular wave solution as
v = V ( ), w = W ( ) with = r ct, yields the pair of equations
(c + DK)V
= f (V,W ) + DV
,
c
W
= W + V,
(9.59a)
(9.59b)
with V
= dV /d , etc. These equations are equivalent to the 1D FN equations in the
moving frame of a pulse traveling with speed c0 = c + DK, under the rescaling
(c + DK) /c. Approximating c0 by the singular solution of the 1D FN equations
(see Sect. 2.3), we deduce that the speed of the circular front is a linear function of
the curvature [426]
(9.60)
c(K) = c0 DK.
This result extends locally to a more general curved front, by partitioning the curve
into small segments approximated by circular arcs. Assuming that there are no selfintersections, the curve can be represented in polar coordinates as r = r( ) and there
is now a phase-dependent curvature K( ). Unfortunately, there is no general theory
for the form of the tangential velocity . Following [145], we will take to be a
linearly decreasing function of the curvature at the end point (Fig. 9.16):
(K) = 0 (1 K/Kc ),
(9.61)
up to some critical curvature Kc .

If both c(K) and (K) are known, then one can determine the evolution of an
arbitrary initial excitation pattern, at least numerically. An analytical solution is
only possible when c and are linear function of the curvature and the initial front
is circular symmetric. The basic kinematic theory proceeds as follows [145, 426].
Suppose that the curvature of a point a on the front at time t is Ka , and introduce a
local polar coordinate system (r, ) with the origin at the center of curvature from
point a. Note that r
= r
= 0 at a where r
= r/ . The curve in a neighborhood
of a is given by a function r( ,t) with r(0,t) = ra = 1/Ka . Over a short time interval
dt, the neighborhood of point a moves to a neighborhood of point b with the new
curve segment
r( ,t + dt) = r( ,t) + c( ,t)dt,
c( ,t) = c(K( ,t)).
(9.62)
381
Given a local polar representation of a curve, its local curvature is given by the
formula
r2 + r
2 rr
K= 2
.
(9.63)
[r + r
2 ]3/2
The curvature Kb can then be determined by substituting for r( ,t +dt) using (9.62),
expanding to first order in dt, and setting = 0:

2
2
2 c
dt.
(9.64)
Kb = Ka c(Ka )Ka + Ka
2 =0
In order to integrate this equation, it is necessary to replace by the intrinsic arclength coordinate l using dl = ra d . Assuming that a is an intrinsic distance l from
one free end and taking Ka = K(l,t), Kb Ka = dK(l,t) gives

2c
(9.65)
dK = K 2 c(K) + 2 dt.
l
Note that the points a, b have different values for their intrinsic coordinate, since
a circular wave propagating outwards increases its arc length and the arc length is
measured with respect to an end point of a curve that may be growing or shrinking.
It follows that the difference in arc length between the points a and b is
l
K(s,t)c(K(s,t))ds dt + (0,t)dt.
(9.66)
dl =
0
b
ct
a
ra
O
t
Fig. 9.16 Front position at two successive times t,t + t. Normal velocity at a given point on the
front at time t is c and the tangential velocity at an end point is
382
b
15
30
45 60
Fig. 9.17 (a) Propagation of a front across one half of primary visual cortex. Away from the focus,
horizontal and vertical lines in cortex transform under the inverse retinotopic map to radii and
concentric circles in the corresponding visual hemifield. (b) Progression of scotoma across the
visual hemifield, which is obtained by applying the inverse retinotopic map to the propagating
cortical front in (a)
Since K = K(l,t), we have the alternative expression for dK,

dK =
K
K
dl +
dt.
l
t
(9.67)
with dl given by (9.66). Comparing (9.65) with (9.66) and (9.67) finally yields the
integrodifferential equation [145, 426]
l
K
K
2c
+
= c(K)K 2 2 .
K(s,t)c(K(s,t))ds + (0,t)
(9.68)
t
l
l
0
Dahlem and Muller applied the above kinematic theory to a circular arc propagating in the visual cortex under the assumptions that c c0 and are given by (9.61).
The initial front is taken to be an arc of small radius r0 , which means that K > Kc and
< 0, that is, the end points, shrink as a function of time. Since the normal velocity
is approximately independent of curvature, the front at time t is still a circular arc,
but its arc length l(t) has shrunk. Setting K(t) = 1/r(t) with r(t) = r0 + c0t and
c(K) = c0 , (9.66) implies that

dl V0 l
1
=
+ G0 1
.
dt
r
rkc
Let = l/r be the angle subtended by the arc. Then

1
1
d
= G0
,
dr
r r 2 kc
9.4 Epilepsy
383
so that integrating with respect to r gives

1
(r) = (1/kc ) + G0 ln(rkc ) +
1 .
rkc
(9.69)
Using the analytical expression for the propagation of a circular arc across visual
cortex, it is possible to determine the progression of a corresponding scotoma across
the visual hemifield by applying the inverse retino-cortical map; see Sect. 8.1.1.
This construction was carried out by Dahlem and Muller [145], and an illustrative sketch of their results is shown in Fig. 9.17. Finally, note that it is also possible
to reproduce fortification patterns by introducing a region of enhanced neural excitability in the leading edge of the propagating front and superimposing this onto
an orientation preference map. The resulting activity pattern of orientation-selective
neurons was used to generate a set of population vectors that determined the distribution of oriented edges in the visual field [144].
9.4 Epilepsy
Epilepsy is the worlds most common serious brain disorder, affecting around 50
million people [348, 430, 594]. Although there have been major advances in the
medical and surgical treatment of epileptic seizures, almost a third of patients are
untreatable and, even in successful cases, there can be serious side effects. Traditionally, epilepsy has been divided into two main categories: primary generalized
seizures that arise simultaneously across large parts of the brain and focal or partial seizures that initially occur within a localized area of the brain but may subsequently propagate to other brain regions. If the latter does not cause a disruption
of consciousness or cognitive abilities, then it is referred to as a simple seizure,
otherwise it is called complex. In reality, there is probably a spectrum of different
types of seizure with overlapping or related mechanisms for the initiation, propagation, and termination of the seizure [507]. Most studies of epilepsy have focused
on the molecular, anatomical, and cellular processes underlying the development
of epilepsy (epileptogenesis) and the initiation of seizures [125, 415]. Much of this
work has been based on in vitro and in vivo animal models.
In the cortex and hippocampus, a stereotypical signature of focal epileptiform
activity is the interictal spike that occurs between seizures (ictal activity). Interictal
spikes are brief bursts of synchronized activity that last around 100 ms and which
show up as sharp spikes in EEG recordings and as large amplitude depolarizations
at the intracellular level. The spike appears to be initiated in deep layer V of the
cortex, which contains a subset of pyramidal neurons that can intrinsically generate
high-frequency bursts of up to five action potentials at 200350 Hz and resemble
in some ways the up states of the slow (< 1Hz) oscillations seen in cortical slices
[549]; see Sect. 5.1. Recurrent excitation mediated by glutamatergic synapses then
leads to a rapid spread of activity both vertically and horizontally through a large
384
local network of neurons. Activity is finally terminated by a combination of local

recurrent GABA inhibition, Ca2+ -dependent AHP currents, and possibly synaptic
depression [415]. In the hippocampus, pyramidal neurons in the CA3 region appear
to be the substrate for intrinsic bursters. There are two competing theories for the
primary cause of the large depolarization that initiates an interictal spike. The first
assumes that there is an excessive synchronous synaptic activation of essentially
normal pyramidal neurons due to rewiring of excitatory connections or removal of
inhibition, for example, whereas the second proposes that the excitability properties of the neurons exhibiting strong depolarizations are altered such that normal
synaptic drive results in an abnormal response. Until recently, the first theory was
predominant, but now there is growing experimental evidence that several forms
of epilepsy involve genetically determined ion channel defects that alter intrinsic
excitability [41]. A third possible cause of epileptiform activity has also recently
emerged, namely, neuronastrocyte interactions [311, 564, 647, 668]. As discussed
in Sect. 4.5, astrocytes are in a position to induce local synchronization of neurons
via gliotransmission of glutamate and gap-junction coupling, which could be a precursor of larger-scale epileptiform activity. Alternatively, excessive swelling of astrocytes through the intake of water could reduce the volume of the extracellular
space, thus increasing the concentration of extracellular K+ and resulting in a depolarization of surrounding neurons, see also our discussion of spreading depression
in Sect. 9.3.
Another ongoing debate concerns the precise role of interictal spikes in epileptogenesis? At the cellular level, the transition to seizure-like activity is typically associated with a prolonged depolarization and high-frequency repetitive firing around
1020 Hz [415]. This tonic phase of a seizure is often followed by a period of
irregular periodic bursts known as the clonic phase, in which there is very little
action potential activity between bursts. Finally, the clonic phase is succeeded by
a relatively quite postictal depression phase, during which the membrane potential
is hyperpolarized. One hypothesis is that interictal spikes promote epileptogenesis, which is motivated by the observation that in certain animal models interictal spikes arise in a latent period prior to spontaneous seizures [599]. The main
idea is that persistent abnormal interictal activity over a period of time leads to
the formation of new excitatory connections and a strengthening of existing excitatory circuits via Hebbian plasticity, both of which ultimately reinforce the epileptic
state. A completely opposite proposal is that although interictal spikes resemble
patterns of spiking activity during seizures, they are independent events such that,
in certain cases, the occurrence of interictal activity can actually reduce the likelihood of seizures; the latter effect has been observed in some brain slice experiments [18]. Yet another hypothesis is that deafferentation following brain trauma,
for example, engages homeostatic synaptic plasticity mechanisms, which strengthen
connections between intact neurons ultimately leading to seizures; interictal spikes
could then be an intermediate stage in the progression to full seizure-like activity
[290, 463]. In order to help resolve these issues, a number of computational models have explored how epileptogenesis depends on changes in network structure
[159, 437, 451, 646].
9.4 Epilepsy
385
Recent advances in human brain imaging such as functional magnetic resonance

imaging (fMRI) and diffusion-tensor imaging, together with more traditional electrophysiological methods such as electroencephalogram recordings and chronic
multielectrode implants, have led to a rapid increase in large-scale brain models
of epilepsy [522, 594]. Broadly speaking, such models focus either on the global
temporal dynamics of primary generalized seizures [60, 529, 582] or on characterizing the types of large-scale brain networks that support generalized seizures
and the propagation of focal seizures [348, 375]. Two major types of generalized
seizure are tonicclonic (grand mal) seizures (see above) and absence (petit mal)
seizures. The latter are generated in the thalamocortical loop, where reticular thalamic neurons are recruited by the neocortex and then inhibit thalamocortical neurons; see Sect. 5.4.3. The subsequent activation of various ionic currents generates
post-inhibitory rebound bursting in the thalamocortical neurons, which project onto
and excite neocortical neurons, resulting in a repeating cycle of spiking activity at
around 3 Hz [415]. Since primary generalized seizures occur across large parts of
the brain, it is reasonable as a first approximation to consider a nonspatial (meanfield) model of interacting neural populations (Chap. 6). The resulting model can
be described by a system of ODEs, and bifurcation theory can be used to analyze
the onset of seizures and to identify electrophysiological differences between tonic
clonic and absence seizures [60, 529, 582]. At least two different bifurcation scenarios have been identified [522]: (1) a change in some control parameter transitions the
dynamical system from a normal state to a seizure state or (2) a change in a control
parameter moves the brain from a normal state to a bistable state, in which both normal and seizure states coexist. In the latter case, the switch to the seizure state would
require changes in some intrinsic physiological process such as the sleepwake cycle or an external drive such as a stroboscope. One potential clinical application of
these ODE models is that they generate time series that can be directly compared to
EEG recordings and, in principle, could be used to predict the onset of seizures.
In this section, we focus on network aspects of epilepsy. (For a comprehensive
review of computational models of epilepsy, see the book by Soltesz and Staley
[594].) We begin by summarizing some basic network theory that is particularly relevant to brain networks. We then review recent work on identifying structural and
functional brain networks that support epileptiform bursts and seizures. Ultimately,
one would like to incorporate such findings in neural field models of wave propagation (Chap. 7), in order to better understand the spread of epileptiform activity in
the human brain.
9.4.1 Basics of Network Theory

Much of the modern theory of networks was stimulated by the discovery of smallworld networks in 1998 [662] and scale-free networks in 1999 [23], both of which
will be defined below. There are a number of excellent reviews of modern network
theory [52, 459], including several that are tailored towards brain networks [541,
386
A
E
C
D
B
unweighted graph
weighted graph
directed graph
Fig. 9.18 Graph representation of a network in terms of vertices and edges. The vertex B in the
unweighted graph has degree k = 3 with neighbors C, D, E. The clustering coefficient of B is 2/3,
since there is a maximum of three edges between the neighbors but only CE and ED exist, whereas
CD is missing. The shortest path length from A to C is d = 5
(a)
p=0
(b)
rewiring probability p
(c)
p=1
Fig. 9.19 Three basic network types illustrated by the ring network of Watts and Strogatz [662].
(a) The ordered network has N = 16 vertices and each vertex is connected to k = 4 neighbors. (b)
and (c) The small-world and random networks are constructed by randomly reconnecting edges
with rewiring probability p. For small values of p one obtains a small-world network, whereas a
completely random network is created when p = 1
600]; we will follow [600]. An abstract representation of a network, whether it is

a neural network, a social network, or a communications network, is a graph. The
latter consists of a set or vertices (or nodes) and set of edges or connections. The
connectivity structure of the graph is specified by an adjacency matrix A with Ai j =
1 if the vertices i and j are linked by an edge; otherwise Ai j = 0. The number of
edges ending at given vertex is called the degree k of the vertex. The probability P(k)
that a randomly chosen vertex has degree k is called the degree distribution and is an
important factor in determining network properties. If the direction of information
flow along each edge is unimportant, then the graph is said to be undirected and A is
a symmetric matrix; otherwise it is directed; in the latter case the degree distribution
of incoming and outgoing edges may be distinct. Graphs where all edges are treated
equally are said to be unweighted, whereas if each edge is assigned a weight to
signify its strength or efficacy, then the graph is weighted; see Fig. 9.18.
There are two measures that are frequently used to characterize the structure of an
unweighted graph, namely, the clustering coefficient C and the characteristic path
9.4 Epilepsy
387
scale-free network
Fig. 9.20 Scale-free network with two hubs shaded gray
length L. The clustering coefficient Ci of a vertex i with degree ki is typically defined

as the ratio of the number of existing edges between neighbors of i (those directly
connected to i) and the maximum possible number of such edges. Thus,
Ci =
j,kN Ai j A jk Aki
.
ki (ki 1)
(9.70)
with 0 Ci 1. An equivalent definition of Ci is that it determines the number of

completed triangles with i as one of the vertices. The clustering coefficient of the
graph is then obtained by averaging Ci over all vertices:
C=
1 N
Ci .
N i=1
(9.71)
The clustering coefficient is a local measure of network structure that characterizes

the degree of resilience to random errors, since if C 1, then removal of a single
vertex has a minimal affect on the connectivity pattern of the other vertices. The
second important measure of an unweighted graph is the characteristic path length.
The path length or distance di j between two vertices i and j is the minimal number
of edges that have to be traversed in order to go from i to j. The characteristic path
length of a graph is then defined to be the mean of the path lengths between all
vertex pairs:
1
(9.72)
L=
di j .
N(N 1) i, j,i
=j
Using the degree distribution, clustering coefficient, and path length, one can
distinguish between four different types of graph: ordered, small world, random,
and scale-free. The first three cases can be illustrated using a 1D network on a
ring, which was introduced by Watts and Strogatz [662], see Fig. 9.19. Initially each
vertex is connected to k nearest neighbors (k/2 on either side) to form an ordered
network. (Higher-dimensional versions of an ordered network could be generated
using a regular lattice.) Ordered networks have a high C and a high L; in the ring
network of Fig. 9.19a with k = 4 and N = 16, we have C = 3/4 and L = N/2k.
Now suppose that with a probability p, edges are chosen at random, detached at
one end, and then reconnected to another, randomly chosen vertex. For small p, one
thus obtains a small-world network, which has a C close to an ordered network but
a very small path length close to that of a random network. As p is increased, more
388
and more edges are randomly reconnected until finally all connections are random
when p = 1. A random network has small C and small L. The fourth type of network is a scale-free network, which has a power-law degree distribution P(k). This
was originally formulated in terms of model for network growth by Barabasi and
Albert [23]. They assumed that the probability that a newly added edge will connect
to a vertex depends upon the current degree of the edge. In other words, vertices
with a large number of edges (large degree) are more likely to be assigned even
more edges. A network with this form of growth are characterized by a degree distribution of the form P(k)) k1/ ; the original scale-free network of Barabasi and
Albert has = 3. It turns out that many real-world networks appear to be scale-free,
including the World Wide Web and collaboration networks of researchers. Scalefree networks not only can have very small path lengths [L ln(ln(N))], but may
also have a smaller clustering coefficient than small-world networks. Another interesting feature of scale-free networks is that they typically have a few vertices with
extremely high degrees, which act as network hubs; see Fig. 9.20.
9.4.2 Epilepsy and Network Structure

One major research topic in modern network theory is the relationship between
topological network characteristics and network dynamics, particularly synchronization [52]. A number of studies have investigated how the synchronization
of coupled conductance-based or integrate-and-fire (IF) neurons depends on the
rewiring probability of a small-world network [353, 409, 535]. Here the neural
oscillators are placed at the vertices of the graph and the edges correspond to synaptic connections. Roughly speaking, ordered networks exhibit coherent oscillations
but slow system response, whereas random networks have fast response properties
but cannot support coherent oscillations. On the other hand, small-world networks
support coherent oscillations and a fast system response. The dependence of synchronization on network topology has also been investigated within the context of
epilepsy [159, 437, 451, 646]. In particular, Netoff et al. [451] based their study on
results from hippocampal slice experiments indicating that short synchronous bursts
tend to be seen in the CA3 region, where there is a relatively high degree of recurrent excitation, whereas seizure-like activity is found in the more sparsely connected
CA1 region [450]. They modeled a small-world ring network (with N = 3000;
k = 30 for CA1 and k = 90 for CA3) consisting of various types of stochastic model
neurons such as IF and HodgkinHuxley. The neurons were assumed to be initially
in a quiescent state, so that network activity was kindled by spontaneous action
potentials that propagated as waves around the ring; the pattern of activity that
emerged then depended on the number of long-range connections, as well as the
interplay between noise and refractoriness. It was found that as the rewiring probability was increased, the model network displayed normal spiking behavior (low,
sustained population firing rate), then seizure-like activity (significantly higher, sustained firing rates with a relatively low level of coherence), and finally bursting
(waxing and waning of synchronous activity). Moreover, the transition from seizures
9.4 Epilepsy
389
p2
1-p1
1-p1
p2
p1
Fig. 9.21 Two mechanisms for spontaneous wave generation in the model of Netoff et al. [451].
Excitable (active) neurons are indicated by gray (black) disks. (a) An excitable neuron fires spontaneously at a rate and triggers at least two of its neighbors to fire with probability p2 : the
probability that a neighbor does not fire is 1 p1 . (b) An active neuron with a random long-range
connection induces an excitable neuron to fire with probability p1 , which in turn triggers at least
two of its neighbors to fire with probability p2
to synchronized bursting occurred at a lower value of p for higher degree k, that is,
for the CA3 model.
Netoff et al. [451] also introduced a reduced model of propagating activity, under
the assumption that the number of neurons N is much larger than the number of
refractory or active neurons. This model takes the form of a discrete birthdeath
process for the number of spontaneous waves Xn at the nth time step:
Xn+1 = f (Xn ) = Xn + + (Xn ) (Xn ),
(9.73)
where are the birth and death rates. Suppose that each neuron in the network
has k neighbors and if a given presynaptic neuron fires then there is a probability p1
that this induces a given postsynaptic neuron to fire. A local spontaneous wave is
generated if at least two neighboring postsynaptic neurons fire (see Fig. 9.21), with
associated probability
p2 = 1 (1 p1)k kp1 (1 p1)k1 .
It follows that the rate at which new waves are generated in one time step by local
interactions is
n = En p2 ,
390
where En is the number of excitable neurons and is the spontaneous rate at which
each neuron fires. The latter is taken to be
En = N Xn (1 + R).
(9.74)
Here = k/2 1 is the mean number of active neurons in a wave front and R is
the number of time steps each neuron remains refractory in the wake of a wave. A
second contribution to the birth rate comes from the long-range connections in the
ring network, which takes the form

p 1 p 2 En
n = 2 Xn kp
.
N
The first factor is the number of long-range connections with active presynaptic neurons for a given rewiring probability p and the second factor is the probability that
such a connection results in the formation of a wave emerging from the postsynaptic
neuron. Putting these results together, we have

2 Xn kpp1 p2
+ p2 .
+ (Xn ) = n + n = En
(9.75)
N
Since En is an affine function of Xn , the total birth rate is a quadratic function of Xn .
Finally, the death rate is assumed to be
(Xn ) =
2 Xn
,
En
(9.76)
and is based on an estimate of how often waves collide and annihilate.

Sketches of the function f for various rewiring probabilities are shown in
Fig. 9.22. For sufficiently small rewiring probability p there exists a unique stable
(a)
p = 0.001
p = 0.01
p = 0.05
6
Xn+1
4
Xn
4
Xn
4
Xn
Fig. 9.22 Birthdeath process for activity in a ring network [451]. The map f : Xn Xn+1
describing the evolution of the number Xn of spontaneous waves is sketched for several values of
the rewiring probability p. Other parameter values are = 0.0315, p1 = 0.025, k = 90, N = 3000.
There is a unique fixed point, which is weakly attracting for small p, strongly attracting for intermediate p, and unstable for large p
9.4 Epilepsy
391
synapses per neuron (k)
150
bursting
120
90
60
30
CA3
se
izi
ng
normal
CA1
10-3
10-2
10-1
rewiring probability (p)
Fig. 9.23 Sketch of phase diagram for the reduced model of Netoff et al. [451]. The phase boundaries separating normal, seizing, and bursting states are plotted in the (p, k)-plane. The black (gray)
boundary curve is obtained from the 1D map [(R + 1)-dimensional map]. For a range of rewiring
probabilities the CA3 network with k = 90 is in a bursting state, whereas the CA1 network with
k = 30 is in a seizing state
fixed point X = f (X ), which is only weakly attracting. At each time step the
number of waves remains approximately constant, that is, waves are rarely born
or destroyed. This implies that there is a low rate of spontaneous wave generation consistent with normal brain activity. For larger p, the fixed point becomes
strongly attracting and the number of waves increases, which is interpreted as
a seizure state. Finally, for even larger values of p, the fixed point becomes
unstable via a flip bifurcation, and new waves are generated rapidly, suggestive
of a bursting state. (True population bursting cannot be realized in the reduced
model under the restriction that the number of refractory neurons is much smaller
than N, since termination of a population burst requires most of the network to
enter a refractory state.)The one-dimensional map captures quite well the transition
boundary between the seizing and bursting states. However, in order to capture
the transition between the normal and oscillatory seizing state, it is necessary to
consider a more accurate model of refractory dynamics by replacing (9.74) with
En = N Xn
Xnm .
(9.77)
m=1
This leads to an (R + 1)th-order map that can undergo a Hopf bifurcation resulting
in stable oscillations. A typical phase diagram for the model is shown in Fig. 9.23.
Changes in network structure are not necessary, however, for abnormal rhythms
to occur. Motivated by the observation that epileptiform bursting activity often arises
following severe head trauma, Volman et al. [646] studied a computational model
of traumatized cortex, in which a significant fraction of neurons had their afferent
392
inputs reduced (deafferented neurons). Hence, the model could be viewed as a population of intact neurons embedded in a sea of traumatized neurons. Homeostatic
synaptic plasticity was incorporated into the model so that following deafferentation, the remaining intact neurons strengthened their connections within the intact
subnetwork, as well as their efferents to traumatized neurons. The authors found
that this resulted in the emergence of interictal bursts propagating across the cortical
network, provided that the density of intact neurons was relatively high so that they
could also trigger activity in deafferented neurons. Thus the transition to bursting
involved a strengthening of the intact subnetwork rather than a topological restructuring. Indeed, the level of bursting was found to be weakly dependent on properties
of the subnetwork such as its clustering coefficient and path length [646]; a much
more critical factor was the density of intact neurons.
The above examples correspond to structural or anatomical network models, in
which the vertices of a given network are taken to be individual neurons and the
edges correspond to physical synaptic connections. At this microscopic scale, the
only complete structural network that has been mapped is for the 302 neurons of the
nematode worm C. elegans [669]. Such a feat remains intractable for the human
brain, which consists of more than 1010 neurons and 1012 synapses, although noninvasive neuroimaging techniques such as magnetic resonance imaging (MRI) can
be used to construct coarser-grained macroscopic structural networks in humans.
An alternative representation of brain networks has been gaining increasing attention, namely, one involving functional networks [5, 598, 600]. A functional network
represents the coupling between dynamical activity recorded from separate brain
areas. It is based on the assumption that statistical correlations between the time
series of electrophysiological or metabolic activity reflect functional interactions
between neurons or populations of neurons in different brain regions. Clearly functional connectivity depends on the underlying structural connectivity, but they need
not be identical. A variety of neuroimaging methods have been used to construct
functional networks at different spatial and temporal scales, ranging from single unit
recordings to fMRI. One of the ongoing challenges in establishing a functional brain
network is choosing the most appropriate coupling measure (two-point correlations,
Granger causality) and determining the threshold value above which the level of
coupling constitutes an edge [500]. Roughly speaking, experimental observations
of focal seizures indicate that functional networks acquire larger path lengths and
clustering coefficients at seizure onset and become more small world during seizure
propagation and more random at seizure termination [348]. However, contradictory
results have also been obtained, probably reflecting the fact that epilepsy is a complex, heterogeneous phenomenon.
9.5 Neuronal Avalanches, Criticality, and Branching Processes

Another characterization of large-scale brain activity that has gained prominence in
recent years is the extent to which the cortex operates close to criticality [29, 31,
113, 508, 573]. Roughly speaking, criticality at the network level means that the
393
network is in a balanced state with regard to the propagation of activity induced

by the spontaneous spiking of individual neurons. As a simple gedanken experiment [573], imagine a fully connected network of N neurons and let p denote the
probability that each neuron in the network fires in response to a single spike. The
parameter p represents a variety of neurobiological factors that could contribute to
the efficacy of a single action potential, including the relative number of excitatory and inhibitory neurons, the strength of excitatory versus inhibitory synapses,
the intrinsic excitability of neurons, and the strength of neuromodulators. Clearly,
if p 1/N, then activity in the network quickly dies out, and it is said to be in
the subcritical regime. On the other hand, if p 1/N, then a single spike almost
surely ignites a runaway chain of spiking events that never terminates, and we have
a supercritical regime. Finally, if p = 1/N, then one spike is likely to cause one
subsequent spike, so that there is reverberating activity in the network that does not
overwhelm the network with a hypersynchronous statethe critical regime. Within
the context of epilepsy, the basic idea is that the normal brain operates close to a critical state, whereas the epileptic brain deviates significantly from criticality during
seizures, due to a failure in the underlying homeostatic mechanisms that maintain
the critical state [291, 418].
The concept of criticality has its origins in the statistical physics of equilibrium
phase transitions. A classical example is the phase transition from paramagnetism to
ferromagnetism in an iron bar magnet as the temperature T crosses a critical threshold from above, known as the Curie temperature Tc . The Ising model treats the bar
magnet as a lattice of electrons, each of which can be in one of two spin states
represented by an arrow pointing up (north) or pointing down (south)the arrows
can be viewed as microscopic bar magnets. Nearest-neighbor interactions between
the spins favor spin alignment (order), whereas thermal fluctuations favor disalignment (disorder). Both processes occur in order to minimize the free energy of the
system, the first by reducing the interaction energy and the second by increasing the
entropy. At low temperatures (T < Tc ), the ordering effects of coupling dominate
and the iron bar exhibits macroscopic magnetization due to most spins pointing in
the same direction; see Fig. 9.24. On the other hand, at high temperatures (T > Tc )
the disordering effects of thermal fluctuations win out and the spins point in random
directions so there is no net magnetization. At the critical point T = Tc there is a
T < Tc
T > Tc
Fig. 9.24 Diagram illustrating an ordered (T < Tc ) and a disordered (T > Tc ) alignment of spins
in the 2D Ising model, where Tc is the Curie temperature at criticality
394
balance between the two effects. (Technically speaking, this so-called second-order
phase transition occurs on a 2D or 3D lattice but not a 1D lattice.) Although there
is now no global order, there will be local domains where all the spins point up and
other domains where all the spins point down. At the critical temperature there is
a broad distribution of domain sizes, and the domains are themselves dynamically
shifting and changing size. An important signature of criticality in equilibrium phase
transitions is that various physical quantities exhibit power-law behavior. In the case
of the Ising model, this includes the distribution of domain sizes and the two-point
correlation matrix, which is defined according to
Ci j = (si si )(s j s j ).
(9.78)
Here si = 1 denotes the up/down state of the spin at lattice site i, and the angled
brackets indicate time averaging. Thus (si si ) represents the size of fluctuations
of the ith spin about its average at a given time. It follows that Ci j will only be large
if si and s j both fluctuate (some disorder) and these fluctuations are coordinated in
the same direction (some order). As one might expect, Ci j only becomes large in a
domain where these two effects are balanced, that is, close to criticality. One finds
that Ci j e|i j|/ , where is known as the correlation length and
1
.
|T Tc |
(9.79)
That is, the correlation length diverges as the system approaches the critical temperature Tc according to a power law with critical exponent . (In fact, the critical
exponent differs on either side of the critical point.) An important feature of power
laws is that they show no characteristic scalethey are scale-free. Moreover, they
are a signature of universality, in the sense that a large class of systems that differ at
the microscopic level exhibit the same macroscopic behavior close to criticality
they belong to the same universality class.
Extending the notion of criticality to biological networks such as the cortex is
nontrivial, not only given the fact that these networks are complex and heterogeneous but also because they operate far from thermodynamic equilibrium. Nevertheless, criticality can be understood in a dynamical sense, as suggested by the previous
Z0 = 1
Z1 = 2
Z2 = 4
Z3 = 6
Fig. 9.25 Illustration of a branching process with three generations
395
gedanken experiment concerning spike avalanches, and evidence (but not proof) of
criticality can be obtained by observing power laws in experimental data. The first
substantive study of power-law distributions at the network level arose from multielectrode array recordings in cell cultures and acute brain slices [29]. Bursts of
spontaneous activity were observed such that if the number of electrodes activated
in each distinct burst were counted, then the burst sizes were distributed according
to a power law. That is, denoting the burst size by l,
P(l) l
(9.80)
with 1.5. The index value of 1.5 can be understood using the theory of
branching processes [264, 364]; see below. Since these initial findings, power-law
distributions of these so-called neuronal avalanches have been observed in a variety of animal models, both in vitro and in vivo, using both multielectrode recordings and imaging methods such as fMRI; see the reviews [31, 573]. In spite of
these results, there is still some controversy regarding their interpretation, both in
terms of whether or not the statistical data is truly given by power laws and in
terms of the causes of power-law-like behavior [28, 631]. Nevertheless, neuronal
avalanches and the theory of branching processes provide another interesting framework for understanding the propagation of activity in networks. In the following we
sketch how to obtain the observed power law from a classical branching process,
following [364].
A simple branching process is a discrete Markov model for the evolution of a
population, which in the neural context would correspond to the progression of spiking events in a neuronal avalanche. Let Zn denote the number of members (events)
of the nth generation (time step). Each member of the nth generation gives birth to
a family, which could be empty, of members of the (n + 1)th generation with the
following assumption: the family sizes of the individuals of the branching process
are independent identically distributed random variables [264]. An example realization of a branching process is shown in Fig. 9.25. A useful method for analyzing a
branching process is to use generating functions. Let Gn (s) = E(sZn ) be the generating function of the random variable Zn with probability distribution P(Zn = m).
A number of results can now be established.
1. Recursive structure of generating functions. Each member of the (n + 1)th generation has a unique ancestor in the nth generation such that
Zn+1 = X1 + X2 + . . . + XZn ,
where Xi is the size of the family produced by the ith member of the mth generation. It follows that
Gn+1 (s) = E(sZn+1 ) = E(sX1 +...+XZn )
=
E(sX1 +...+Xm |Zn = m)P(Zn = m)
n=0
396
E(sX1 sX2 . . .Xm |Zn = m)P(Zn = m)
n=0
=
=
E(sX j )
n=0
P(Zn = m)
j=1
[G1 (s)]m P(Zn = m) = Gn (G1 (s)).
n=0
Iterating this resulting and dropping the subscript on G1 , we have the recurrence
relation
(9.81)
Gn (s) = Gn1 (G(s)) = G(G(. . . (G(s)) . . .)).
2. Mean and variance. Let = E(Z1 ) and 2 = var(Z1 ). In order to determine
the mean and variance of Zn , we use the recursive structure of the generating
functions. First,

d
E(Zn ) = Gn (1) =
G(Gn1 (s))
= G
(1)G
n1 (s) = G
n1 (s).
ds
s=1
Iterating this result shows that
E(Zn ) = n .
(9.82)
Similarly,
E(Zn (Zn 1)) = G
n (1) = G
(1)G
n1 (1)2 + G
(1)G
n1 (1).
This gives the iterative result
var(Zn ) = 2 2n2 + var(Zn1 ),
from which one finds that
n 2
var(Zn ) = ( 1) n1
1
2
if = 1
if = 1.
(9.83)
3. Total number of individuals. Let Tn be the total number of individuals up to and

including the nth generation. Then
E(Tn ) = E(Z0 + Z1 + Z2 + . . . + Zn )
= 1 + E(Z1) + E(Z2 ) + . . . + E(Zn )
= 1 + + 2 + ... + n
( n+1
1
= 1,
1 ,
=
n + 1,
= 1.
397
It follows that

lim E(Tn ) =
1
1 ,
1,
< 1.
Let Hn (s) = E(sTn ) be the generating function for the random variable Tn . The
generating functions satisfy the recurrence relation
Hn+1 (s) = sG(Hn (s)).
(9.84)
4. Probability of extinction. One important property of a branching process is

whether or not it eventually becomes extinct, that is, Zn = 0 for some finite n.
We can define the probability of extinction according to
= lim P(Zn = 0).
(9.85)
It can be shown that is the smallest nonnegative root of the equation s = G(s).
Moreover, = 1 if < 1 (extinction occurs almost surely), whereas < 1 if
> 1 (there is a nonzero probability that there is persistent growth). Moreover,
if = 1, then = 1 provided that the variance of the family size distribution
has a strictly positive variance. In light of our discussion of criticality, = 1 is a
critical point separating a subcritical regime ( < 1) from a supercritical regime
( > 1).
For the sake of illustration, consider a geometric branching process where
the distribution of family sizes is given by P(Z1 = k) f (k) = qpk with q =
1 p. In this case, one can calculate the generating function and other quantities
explicitly:
p
p2 p
(9.86)
G(s) = q(1 ps)1, = , 2 = 2 + .
q
q
q
Moreover, it can be shown by induction that
n (n 1)s
n + 1 ns
Gn (s) =
q[pn qn ps(pn1 qn1)]
n+1
p
qn+1 ps(pn qn)
It follows that
P(Zn = 0) = Gn (0) =
n
n+1
q(pn qn)
n+1
p
qn+1
if p = q = 12 ,
(9.87)
if p = q.
if p = q,
(9.88)
if p = q,
and, hence, = 1 if p q and = q/p if p > q. We conclude that for a geometric branching process, extinction occurs almost surely if E(Z1 ) = = p/q 1,
398

1-p
p
p
p
1-p
Fig. 9.26 Branching process for the propagation of activity during an avalanche. Activity propagates to a pair of descendants with probability p, whereas it fails to propagate with probability
1 p. Black nodes are active and white nodes are inactive
otherwise there is a finite probability of persistent growth. In light of our discussions of criticality and avalanches, we can identify the regime p < q as subcritical,
the regime p > q as supercritical, and the point p = q = 1/2 as critical.
In the case of neuronal avalanches, a more appropriate branching process is one
where each node has two descendants and activity can propagate to the descendants
with probability p [29, 364]; see Fig. 9.26. Thus Zn denotes the number of active
members of the nth generation (burst size) with 0 Zn 2n . For this process f (k) =
(1 p)k,0 + pk,2 . It follows that
G(s) = 1 p + ps2,
= 2p,
2 = 4p(1 p),
(9.89)
and the critical point is at p = 1/2. The total number of events Tn is the burst size,
whose distribution can be determined from the generating function Hn (s). For the
given branching process, Hn (s) satisfies the recurrence relation (see (9.84))
Hn+1 (s) = s[(1 p) + pHn(s)2 ].
(9.90)
The solution of this recurrence relation for n 1 is approximately [264, 364]

1 1 4s2 p(1 p)
.
(9.91)
Hn (s) =
2sp
Expanding as a power series in s and comparing with the definition of Hn (s), one
finds that for 1 l < n, the distribution of avalanche sizes l is given by

2(1 p)/p l/lc (p)
3/2 e
,
(9.92)
P(l) P(Tn = l) =
l
with
lc (p) =
2
.
ln[4p(1 p)]
(9.93)
9.6 Protein Aggregate Spreading in Neurodegenerative Diseases
399
For l > n one finds that P(l) decays exponentially. This then establish the 3/2
power-law behavior at the critical point p = 1/2, since 1/lc (1/2) = 0 and

2 3/2
l
(9.94)
P(l)
for 1 l < n and n is large. Also note that at criticality = 1, which is consistent
with the experimental observation that the ratio of the number of active descendant
electrodes to the number of active ancestor electrodes is around one [29].

There is growing experimental evidence suggesting that there are parallels between
prion disease pathogenesis and other, more common neurodegenerative diseases that
are typically not thought to be infectious, including Alzheimers disease, Parkinsons disease, frontal-lobe dementia, and Huntingtons disease [284, 300, 371].
Prion diseases are fatal neurodegenerative diseases that rose to prominence during
the BSE epidemic in England in the 1980s, where there was thought to be a possible
transmission pathway from infected animals to humans. Unlike classical infective
agents such as bacteria or viruses, prion diseases involve an agent consisting solely
of a misfolded protein or prion [513]. Pathologically folded prion protein (PrP*) corrupts normally folded PrP via a fixed conformational change, which usually leads to
the formation of protein aggregates, and this process propagates across brain cells
at a speed of around 0.11mm per day. The initial cause could be an external infective agent or, more commonly, the spontaneous formation of PrP* aggregates.
Recent molecular, cellular, and animal studies suggest that the intercellular propagation of protein misfolding also occurs for a variety of aggregate-prone proteins that
are linked to noninfective neurodegenerative diseases [262, 533, 565, 566]. These
include amyloid- and tau (Alzheimers disease), -synuclein (Parkinsons diseases), and huntingtin (Huntingtons disease).
The conformational change of a normally folded protein tends to occur via direct
contact with a misfolded protein aggregate. This process is commonly called aggregate seeding or seeded polymerization [300]. If the resulting protein aggregate is
small, then it remains soluble, whereas larger aggregates precipitate out of solution
under physiological conditions. Two distinct morphological types of aggregate can
be identified: amorphous aggregates and amyloid fibrils. The former have an irregular granular appearance when viewed under an electron microscope, whereas the
latter consist of highly ordered and repetitive structures in which all the polypeptides adopt a common fold. Amyloid fibrillogenesis consists of multiple stages,
involving nucleation, polymerization, elongation, and aggregate seeding. It used to
be thought that mature amyloid fibrils were ultimately responsible for the toxic
effects of protein aggregates. However, there is a widening consensus that it is
the non-fibrillar assemblies in early stages of amyloid formation that are particularly toxic. The monomers of proteins such as tau and -synuclein are found
within the cytosol of cells. This means that the spread of such protein aggregates
400
requires the internalization of the corresponding fibrils in order to seed the formation of new aggregates. The newly formed aggregates must then be released
back into the extracellular space. The precise mechanisms of internalization and
externalization are currently unknown. Such processes are not required for prions
and amyloid- within cells, since they are both exposed to the extracellular space.
Hence, the latter proteins can diffuse intracellularly and then infect a neighboring
cell directly.
So far there has been very little modeling work on the spread of protein aggregates. Here we will review a reactiondiffusion model for the propagation of prion
disease in the brain [411]. The simplest version of the model uses a heterodimer
model for prionprion interactions, in which one molecule of normal PrP (denoted
by A) combines with one molecule of PrP* (denoted by B) to form two molecules
of B:
A + B AB BB B + B.
(9.95)
For simplicity, this is represented as a single-step reaction with rate kAB :
A + B 2B.
kAB
(9.96)
The molecules also degrade at rates kA and kB , respectively, with kA > kB since PrP*
is more resistant to the action of proteases. If the reaction kinetics are combined with
classical diffusion on a 1D domain x [0, L], then we obtain the reactiondiffusion
system
a
2a
= 0 kA a kABab + DA 2
t
x
b
2b
= kAB ab kBb + DB 2 ,
t
x
(9.97a)
(9.97b)
where a(x,t) and b(x,t) denote the concentration of normal and misfolded PrP at
time t > 0 and location x. The corresponding diffusivities are DA , DB , and 0 represents a constant rate of production of PrP. Equations (9.97) are supplemented by
no-flux boundary conditions at the ends of the domain. The resulting model belongs
to the same class of PDE as the FisherKPP equation (3.44) and the CaMKII model
(3.43) of Sect. 3.2. In particular, it supports a spreading wave in the form of a pulled
front. Suppose that the system is initially in a stable, homogeneous steady-state
a(x, 0) = a = 0 /kA , b(x, 0) = 0. A local transient increase in b at one end of the
domain will induce a propagating front whose speed can be determined by linearizing about the steady state in the leading edge of the wave where a a (Sects. 3.2
and 3.3). The resulting linear equation for b is
b
2b
= Kb + DB 2
t
x
(9.98)
with K = kAB a kB . Taking b(x,t) = B(x ct) = B( ) gives the ODE cB

=
Kb + Db
. The latter has a solution of the form (neglecting boundary effects) b( ) =
401
Hetrodimer model
PrP
PrP*
+
Nucleated polymerization model
polymerization
degradation
splitting
Fig. 9.27 Schematic illustration of the heterodimer and nucleated polymerization models of prion
reaction kinetics
Ce with
c = c( )
K
+ DB .
Finally, minimizing c( ) with respect to yields the minimum speed

cmin = DB K = DB (kAB 0 /kA kB).
(9.99)
(9.100)
This is the asymptotic speed of the front induced by a localized perturbation.

The above model suffers from at least two major simplifications, as highlighted
by Matthaus [411]. First, the heterodimer model is not a good description of
aggregate seeding, which involves the interaction between PrP* polymers of various
lengths and normal PrP monomers. Second, as in the case of cortical spreading
depression (Sect. 9.3), the brain should really be treated as a highly heterogeneous
medium. Focusing on the first issue, one can replace the heterodimer reaction
scheme by a nucleated polymerization model [263, 408], in which the infective
agent is not a single molecule of PrP* but a PrP* oligomer (short polymer of misfolded PrP). Consider, in particular, the model of Masel et al. [408], which is illustrated in Fig. 9.27. There are a number of steps in the formulation of the kinetic
model:
1. Denote a PrP* polymer of size m (m monomer subunits) by Bm . Polymers below a
critical size n (the nucleation length) are unstable and are assumed to disintegrate
instantaneously into m monomers (m < n). The nucleation of new stable PrP*
polymers is neglected.
2. When a PrP molecule A comes in contact with a PrP* polymer of size m, m n,
it forms a polymer of size m + 1:
Bm + A Bm+1
kAB
with the rate kAB independent of m.
402
3. A polymer of size m, m n, can split into two pieces of sizes m k and k:

Bm Bmk + Bk
with the rate independent of m and k. If the resulting pieces are smaller than
the nucleation length, then they immediately disintegrate.
4. PrP is produced at a rate 0 and degrades at a rate kA . All PrP* polymers degrade
at a rate kB .
Let a be the concentration of PrP monomers and bm denote the concentration of
PrP* polymers of size m. Using the law of mass action, the above kinetic scheme
can be expressed in terms of the following system of ODES [408]:
n1
da
= 0 kAa kABa bm + 2 k bm ,
dt
mn
k=1 mk+1
dbm
= kAB a(bm1 bm ) kBbm (m 1)bm + 2 bl
dt
lm+1
(9.101a)
(9.101b)
for m n, and bm 0 for m < n. It turns out that this infinite hierarchy of equations
can be reduced to a closed system of three ODES by summation. Defining
b=
bm ,
mn
z=
mbm ,
(9.102)
mn
one finds that

da
= 0 kAa kABab + n(n 1) b,
dt
db
= kB b z (2n 1) b,
dt
dz
= kAB ab kBz n(n 1) b.
dt
(9.103a)
(9.103b)
(9.103c)
The reaction kinetics given by the system of (9.101) can be incorporated into
a reactiondiffusion model for the spread of prion disease by adding the terms
DA 2 a/ x2 and Dm 2 bm / x2 to the right-hand sides of (9.101a) and (9.101b), respectively [411]. Numerical simulations confirm that the extended model also supports propagating fronts. (Note that since the cytosolic diffusion coefficient Dm will
depend on polymer size m, it is not possible to use summation to reduce (9.103) to
a small closed system of coupled RD equations.)
Finally, we briefly consider the issue of diffusion. The 1D model can capture the
intracellular diffusion of prions along a single nerve. However, applying a homogeneous and isotropic 3D diffusion equation to the spread on larger spatial scales
fails to take into account the fact that neurons form complex networks, which
include both local connections to neighbors and long-range connections. That is,
one has to consider epidemic diseases on complex networks such as the scale-free
403
and small-world networks defined in Sect. 9.4.1. Some insights into the spread of
diseases on networks have been obtained by considering standard models of epidemics such as the susceptibleinfected (SI) model [26, 411, 458]. In the latter
model, each individual is represented by a vertex of the network, and the edges represent the connections between individuals along which the infection may spread.
(In a prion model each vertex would be a neuron and each edge a nerve fiber, say.)
The total population N (number of vertices) is kept fixed so that N = S(t) + I(t),
where S(t) and I(t) are the number of susceptible and infected individuals, respectively.
A first approach to analyzing the SI model is to use a mean-field description
[26]. That is, all the vertices are treated equivalently, which means that the average
density i(t) = I(t)/N of infected individuals satisfies
di(t)
= ki(t)[1 i(t)].
(9.104)
dt
The above equation can be interpreted as follows: the growth rate of infected
individuals is proportional to a constant spreading rate, , the density of susceptible vertices, s(t) = 1 i(t), and the number of infected individuals in contact with
any susceptible vertex. The mean-field approximation takes the latter to be the product of the mean number of neighbors k and the average density i(t). Note that if
the initial density i0 of infected individuals is small, then one can drop the quadratic
term in (9.104) and deduce that the leading behavior at small times is i(t) i0 et/H
with H = 1/( k). The time constant H determines the initial rate at which the
disease spreads. The mean-field approximation does not take into account fluctuations arising from fluctuations in the degree k of vertices. In order to proceed further,
it is necessary to consider the evolution equation for the average density of infected
vertices of degree k, ik (t) = Ik (t)/Nk , where Nk and Ik are the total number and
infected number of k-vertices. Thus
dik (t)
= k[1 ik (t)]k (t),
(9.105)
dt
where k (t) is the density of infected individuals that are neighbors of a k-vertex.
As a further simplification, suppose that k = for all k (a so-called uncorrelated
network). Let P(k) be the degree distribution of the network (Sect. 9.4.1). The probability that a susceptible vertex is connected to an infected vertex of degree k
is
proportional to the number of edges emanating from the latter. Since at least one
edge is likely to be connected to another infected vertex, it follows that this number
is k
1. One thus obtains the result [26]
(t) =
1
(k
1)P(k
)ik
(t).
k
(9.106)
Differentiate this equation with respect to t and use (9.105). Dropping quadratic
terms in the early stages of the epidemic yields the pair of equations

2
k
dik
d
= k (t),
=
1 (t).
(9.107)
dt
dt
k
404
Finally, assume uniform initial conditions ik (0) = i0 for all k. The pair of (9.107)
can then be solved in terms of i(t) = k P(k)ik (t) according to

k2 k t/
i(t) = i0 1 + 2
(9.108)
e 1 ,
k k
with new time constant
k
(k2 k)
(9.109)
In the case of a Poisson degree distribution for which var[k] = k, we recover the
mean-field result = 1/( k). On the other hand, for a scale-free network with
P(k) k , 2 < < 3, we find that k2 as N . In other words, for large
scale-free networks, there is approximately an instantaneous rise in the number of
infected.
References
1. Abbott, L.F., Marder, E.: Modelling small networks. In: Koch, C., Segev, I. (eds.) Methods
in Neuronal Modelling, pp. 361410, 2nd edn. MIT Press, Cambridge (1998)
2. Abbott, L.F., van Vresswijk, C.: Asynchronous states in networks of pulsecoupled
oscillators. Phys. Rev. E 48(2), 14831490 (1993)
3. Abbott, L.F., Fahri, E., Gutmann, S.: The path integral for dendritic trees. Biol. Cybern. 66,
4960 (1991)
4. Abbott, L.F., Varela, J.A., Sen, K., Nelson, S.B.: Synaptic depression and cortical gain
control. Science 275, 220224 (1997)
5. Aertsen, A.M.H.J., Gerstein, G.L., Habib, M.K., Palm, G.: Dynamics of neuronal firing
correlation: modulation of effective connectivity. J. Neurophysiol. 61, 900917 (1989)
6. Alexander, G., Crutcher, M.: Functional architecture of basal ganglia circuits: neural substrates of parallel processing. Trends. Neurosci. 13, 266271 (1990)
7. Allard, J., Moginler, A.: Traveling waves in actin dynamics and cell motility. Curr. Opin.
Cell Biol. 25, 107115 (2013)
8. Amari, S.: Dynamics of pattern formation in lateral inhibition type neural fields. Biol.
Cybern. 27, 7787 (1977)
9. Andersen, S.S., Bi, G.Q.: Axon formation: a molecular model for the generation of neuronal
polarity. Bioessays 22, 172179 (2000)
10. Andersen, P., Morris, R., Amaral, D., Bliss, T., OKeefe, J.: The Hippocampus Book. Oxford
University Press, Oxford (2008)
11. Angelucci, A., Levitt, J.B., Walton, E.J.S., Hupe, J.M., Bullier, J., Lund, J.S.: Circuits for
local and global signal integration in primary visual cortex. J. Neurosci. 22, 86338646
(2002)
12. Antic, S.D., Zhou, W.L., Moore, A.R., Short, S.M., Ikonomu, K.D.: The decade of the
dendritic NMDA spike. J. Neurosci. Res. 88, 29913001 (2010)
13. Arimura, N., Kaibuchi, K.: Neuronal polarity: from extracellular signals to intracellular
mechanisms. Nat. Rev. Neurosci 8, 194205 (2007)
14. Armero, J., Casademunt, J., Ramirez-Piscina, L., Sancho, J.M.: Ballistic and diffusive corrections to front propagation in the presence of multiplicative noise. Phys. Rev. E 58, 54945500
(1998)
15. Aronson, D.G., Weinberger, H.F.: Multidimensional nonlinear diffusion arising in population
genetics. Adv. Math. 30, 3376 (1978)
16. Atay, F.M., Hutt, A.: Stability and bifurcations in neural fields with finite propagation speed
and general connectivity. SIAM J. Appl. Math. 65, 644646 (2005)
17. Austin, T.D.: The emergence of the deterministic Hodgkin-Huxley equations as a limit from
the underlying stochastic ion-channel mechanism. Ann. Appl. Prob. 18, 12791325 (2006)
18. Avoli, M., de Curtis, G.B.M.: Do interictal spikes sustain seizures and epileptogenesis.
Epilepsy Currents 6, 203207 (2006)
DOI 10.1007/978-1-4614-8866-8, Springer Science+Business Media New York 2014
405
406
References
19. Baer, S.M., Rinzel, J.: Propagation of dendritic spikes mediated by excitable spines: a
continuum theory. J. Neurophysiol. 65, 874890 (1991)
20. Bair, W., Koch, C., Newsome, W.T., Britten, K.H.: Power spectrum analysis of bursting cells
in area MT in the behaving monkey. J. Neurosci. 14, 28702893 (1994)
21. Baladron, J., Fasoli, D., Faugeras, O., Touboul, J.: Mean field description of and propagation
of chaos in recurrent multipopulation networks of Hodgkin-Huxley and FitzHugh-Nagumo
neurons. J. Math. Neuro. 2, 10 (2012)
22. Ballesteros-Yanez, I., Benavides-Piccione, R., Elston, G.N., Yuste, R., DeFelipe, J.: Density
and morphology of dendritic spines in mouse neocortex. Neurosci. 138, 403409 (2006)
23. Barabasi, A.L., Albert, R.: Emergence of scaling in random networks. Science 286, 509512
(1999)
24. Baras, F., Mansour, M., Malek, M., Pearson, J.E.: Microscopic simulation of chemical
bistability in homogeneous systems. J. Chem. Phys. 105, 82578261 (1996)
25. Bart, E., Bao, S., Holcman, D.: Modeling the spontaneous activity of the auditory cortex.
J. Comput. Neurosci. 19, 357378 (2005)
26. Barthelemy, M., Barrat, A., Pastor-Satorras, R., Vespignani2, A.: Velocity and hierarchical
spread of epidemic outbreaks in scale-free networks. Phys. Rev. Lett. 92, 178701 (2004)
27. Bazhenov, M., Timofeev, I., Steriade, M., Sejnowski, T.J.: Model of thalamocortical slowwave sleep oscillations and transitions to activated states. J. Neurosci. 22, 86918704 (2002)
28. Bedard, C., Destexhe, A.: Macroscopic models of local field potentials the apparent 1/f noise
in brain activity. Biophys. J. 96, 25892603 (2009)
29. Beggs, J.M., Plenz, D.: Neuronal avalanches in neocortical circuits. J. Neurosci. 23,
1116711177 (2003)
30. Beggs, J.M., Plenz, D.: Neuronal avalanches are diverse and precise activity patterns that are
stable for many hours in cortical slice cultures. J. Neurosci. 24, 52165229 (2004)
31. Beggs, J.M., Timme, N.: Being critical of criticality in the brain. Front. Physiol. 3 (163),
114 (2012)
32. Ben-Ari, Y.: Developing networks play a similar melody. Trend. Neurosci. 24, 353360
(2001)
33. Ben-Jacob, E., Brand, H., Dee, G., Kramer, L., Langer, J.S.: Pattern propagation in nonlinear
dissipative systems. Physica D 14, 348364 (1985)
34. Ben-Yishai, R., Bar-Or, R.L., Sompolinsky, H.: Theory of orientation tuning in visual cortex.
Proc. Nat. Acad. Sci. 92, 38443848 (1995)
35. Benda, J., Herz, A.V.M.: A universal model for spike-frequency adaptation. Neural Comput.
15, 25232564 (2003)
36. Benguria, R., Depassier, M.: Variational characterization of the speed of propagation of fronts
for the nonlinear diffusion equation. Comm. Math. Phys. 175, 221227 (1996)
37. Bennett, M.R., Farnell, L., Gibson, W.G.: A quantitative model of purinergic junctional
transmission of calcium waves in astrocyte networks. Biophys. J. 89, 22352250 (2005)
38. Bennett, M.R., Farnell, L., Gibson, W.G.: A quantitative model of cortical spreading depression due to puringenic and gap-junction transmission in astrocyte networks. Biophys. J. 95,
56485660 (2008)
39. Benucci, A., Frazor, R., Carandini, M.: Standing waves and traveling waves distinguish two
circuits in visual cortex. Neuron 55, 10317 (2007)
40. Berestycki, H., Hamel, F., Roques, L.: Analysis of the periodically fragmented environment
model: II-biological invasions and pulsating travelling fronts. J. Math. Biol. 51, 75113
(2005)
41. Bernard, C., Anderson, A., Becker, A., Poolos, N.P., Beck, H., Johnston, D.: Acquired
dendritic channelopathy in temporal lobe epilepsy. Science 305, 532535 (2004)
42. Berridge, M.J.: Calcium signalling and cell proliferation. Bioessays 17, 491500 (1995)
43. Berridge, M.J.: Neuronal calcium signaling. Neuron 21, 1326 (1998)
44. Berridge, M.J., Lipp, P., Bootman, M.D.: The versatility and universality of calcium
signalling. Nat. Rev. Mol. Cell Biol. 1, 1121 (2000)
References
407
45. Bertram, R., Smith, G.D., Sherman, A.: Modeling study of the effects of overlapping calcium
microdomains on neurotransmitter release. Biophys. J. 76, 735750 (1999)
46. Blake, R.: A primer on binocular rivalry, including current controversies. Brain Mind 2, 538
(2001)
47. Blake, R., Logothetis, N.: Visual competition. Nat. Rev. Neurosci. 3, 111 (2002)
48. Blake, R., Wilson, H.R.: Binocular vision. Vis. Res. 51, 754770 (2011)
49. Blakenship, A.G., Feller, M.B.: Mechanisms underlying spontaneous patterned activity in
developing neural circuits. Nat. Rev. Neurosci. 11, 1829 (2010)
50. Blasdel, G.G.: Orientation selectivity, preference, and continuity in monkey striate cortex.
J. Neurosci. 12, 31393161 (1992)
51. Blasdel, G.G., Salama, G.: Voltage-sensitive dyes reveal a modular organization in monkey
striate cortex. Nature 321, 579585 (1986)
52. Boccaletti, S., Latora, V., Moreno, Y., Chavez, M., Hwang, D.U.: Complex networks:
Structure and dynamics. Phys. Rep. 424, 175308 (2006)
53. Bonhoeffer, T., Grinvald, A.: Orientation columns in cat are organized in pinwheel like
patterns. Nature 364, 166146 (1991)
54. Boonstra, T., Daffertshofer, A., Breakspear, M., Beek, P.: Multivariate time-frequency analysis of electromagnetic brain activity during bimanual motor learning. Neuroimage 36,
370377 (2007)
55. Borisyuk, R., Kirillov, A.B.: Bifurcation analysis of a neural network model. Biol. Cybern.
66, 319325 (1992)
56. Bosking, W.H., Zhang, Y., Schofield, B., Fitzpatrick, D.: Orientation selectivity and the arrangement of horizontal connections in tree shrew striate cortex. J. Neurosci. 17, 21122127
(1997)
57. Boustani, S.E., Destexhe, A.: A master equation formalism for macroscopic modeling of
asynchronous irregular activity states. Neural Comput. 21, 46100 (2009)
58. Brackley, C.A., Turner, M.S.: Random fluctuations of the firing rate function in a continuum
neural field model. Phys. Rev. E 75, 041913 (2007)
59. Brascamp, J.W., van Ee, R., Noest, A.J., Jacobs, R.H., van den Berg, A.V.: The time course
of binocular rivalry reveals a fundamental role of noise. J. Vision 6, 12441256 (2006)
60. Breakspear, M., Roberts, J.A., Terry, J.R., Rodrigues, S., Mahant, N., Robinson, P.A.:
A unifying explanation of primary generalized seizures through nonlinear brain modeling
and bifurcation analysis. Cereb. Cortex 16, 12961313 (2006)
61. Bressloff, P.C.: Resonantlike synchronization and bursting in a model of pulse-coupled neurons with active dendrites. J. Comput. Neurosci. 6, 237249 (1999)
62. Bressloff, P.C.: Synaptically generated wave propagation in excitable neural media. Phys.
Rev. Lett. 82, 29792982 (1999)
63. Bressloff, P.C.: Traveling waves and pulses in a one-dimensional network of integrate-andfire neurons. J. Math. Biol. 40, 169183 (2000)
64. Bressloff, P.C.: Traveling fronts and wave propagation failure in an inhomogeneous neural
network. Physica D 155, 83100 (2001)
65. Bressloff, P.C.: Bloch waves, periodic feature maps and cortical pattern formation. Phys.
Rev. Lett. 89, 088101 (2002)
66. Bressloff, P.C.: Spatially periodic modulation of cortical patterns by long-range horizontal
connections. Physica D 185, 131157 (2003)
67. Bressloff, P.C.: Pattern formation in visual cortex. In: Chow, C.C., Gutkin, B., Hansel, D.,
Meunier, C., Dalibard, J. (eds.) Les Houches 2003: Methods and Models in Neurophysics,
pp. 477574. Elsevier, Amsterdam (2005)
68. Bressloff, P.C.: Stochastic neural field theory and the system-size expansion. SIAM J. Appl.
Math. 70, 14881521 (2009)
69. Bressloff, P.C.: Metastable states and quasicycles in a stochastic Wilson-Cowan model of
neuronal population dynamics. Phys. Rev. E 85, 051903 (2010)
70. Bressloff, P.C.: From invasion to extinction in heterogeneous neural fields. J. Math. Neurosci.
2(6), 127 (2012)
71. Bressloff, P.C.: Spatiotemporal dynamics of continuum neural fields. J. Phys. A 45, 033001
(109 pp.) (2012)
408
References
72. Bressloff, P.C.: Propagation of CaMKII translocation waves in heterogeneous spiny

dendrites. J. Math. Biol. 66, 14991525 (2013)
73. Bressloff, P.C., Coombes, S.: Dynamics of strongly coupled spiking neurons. Neural
Comput. 12, 91129 (2000)
74. Bressloff, P.C., Cowan, J.D.: Amplitude equation approach to contextual effects in visual
cortex. Neural Comput. 14, 493525 (2002)
75. Bressloff, P.C., Cowan, J.D.: The visual cortex as a crystal. Physica D 173, 226258 (2002)
76. Bressloff, P.C., Folias, S.E.: Front bifurcations in an excitatory neural network. SIAM J.
Appl. Math. 65, 131151 (2005)
77. Bressloff, P.C., Dwyer, V.M., Kearney, M.J.: Sum-over-paths approach to diffusion on trees.
J. Phys. A 29, 18811896 (1996)
78. Bressloff, P.C., Cowan, J.D., Golubitsky, M., Thomas, P.J.: Scalar and pseudoscalar
bifurcations: pattern formation on the visual cortex. Nonlinearity 14, 739775 (2001)
79. Bressloff, P.C., Cowan, J.D., Golubitsky, M., Thomas, P.J., Wiener, M.: Geometric Visual
Hallucinations, Euclidean Symmetry and the Functional Architecture of Striate Cortex. Phil.
Trans. Roy. Soc. Lond. B 356, 299330 (2001)
80. Bressloff, P.C., Folias, S.E., Prat, A., Li, Y.X.: Oscillatory waves in inhomogeneous neural
media. Phys. Rev. Lett. 91, 178101 (2003)
81. Bressloff, P.C., Newby, J.: Quasi-steady state analysis of motor-driven transport on a twodimensional microtubular network. Phys. Rev. E 83, 061139 (2011)
82. Bressloff, P.C., Newby, J.M.: Metastability in a stochastic neural network modeled as a
velocity jump Markov process. SIAM J. Appl. Dyn. Sys. 12, 13941435 (2013)
83. Bressloff, P.C., Webber, M.: Neural field model of binocular rivalry waves. J. Comput.
Neurosci. 32, 233252 (2012)
84. Bressloff, P.C., Webber, M.A.: Front propagation in stochastic neural fields. SIAM J. Appl.
Dyn. Syst. 11, 708740 (2012)
85. Brown, R.J., Norcia, A.M.: A method for investigating binocular rivalry in real-time with the
steady state VEP. Vision Res. 37, 24012408 (1997)
86. Brunel, N.: Dynamics of sparsely connected networks of excitatory and inhibitory spiking
neurons. J. Comput. Neurosci 8, 183208 (2000)
87. Brunel, N., Hakim, V.: Fast global oscillations in networks of integrateandfire neurons
with low firing rates. Neural Comput. 11, 16211671 (1999)
88. Buckwar, E., Riedler, M.G.: An exact stochastic hybrid model of excitable membranes
including spatio-temporal evolution. J. Math. Biol. 63, 10511093 (2011)
89. Buckwar, E., Riedler, M.G.: Laws of large numbers and langevin approximations for stochastic neural field equations. J. Math. Neurosci. 3(1), 154 (2013)
90. Buice, M.A., Chow, C.C.: Effective stochastic behavior in dynamical systems with
incomplete information. Phys. Rev. E 84, 051120 (2011)
91. Buice, M., Cowan, J.D.: Field-theoretic approach to fluctuation effects in neural networks.
Phys. Rev. E 75, 051919 (2007)
92. Buice, M., Cowan, J.D., Chow, C.C.: Systematic fluctuation expansion for neural network
activity equations. Neural Comp. 22, 377426 (2010)
93. Bures, J., Buresova, O., Krivanek, J.: The Mechanisms and Applications of Leaos Spreading
Depression of Electrical Activity. Academia, Prague (1974)
94. Butts, D.A., Feller, M.B., Shatz, C.J., Rokshar, D.S.: Retinal waves are governed by collective network properties. J. Neurosci. 19, 35803593 (1999)
95. Butz, E.G., Cowan, J.D.: Transient potentials in dendritic systems of arbitrary geometry.
Biophys. J. 14, 661689 (1974)
96. Buzsaki, G.: Hippocampal sharp waves: their origin and significance. Brain Res. 398,
242252 (1986)
97. Buzsaki, G.: Theta oscillations in the hippocampus. Neuron 33, 32540 (2002)
98. Buzsaki, G.: Rhythms of the Brain. Oxford University Press, Oxford (2006)
99. Byrne, J.H., Roberts, J.L.: From Molecules to Networks: An Introduction to Cellular and
Molecular Neuroscience. Elsevier, Amsterdam (2004)
References
409
100. Cai, D., Tao, L., Shelley, M., McLaughlin, D.W.: An effective kinetic representation of
fluctuationdriven neuronal networks with application to simple and complex cells in visual
cortex. Proc. Natl. Acad. Sci. USA 101, 77577562 (2004)
101. Campbell, K., Gotz., M.: Radial glia: multi-purpose cells for vertebrate brain development.
Trend. Neurosci. 25, 235238 (2002)
102. Cantrell, R.S., Cosner, C.: Spatial Ecology Via Reaction-Diffusion Equations. Wiley, Chichester (2003)
103. Carlsson, A.E.: Dendritic actin filament nucleation causes traveling waves and patches. Phys.
Rev. Lett. 104, 228102 (2010)
104. Carpenter, G.: A geometric approach to singular perturbation problems with applications to
nerve impulse equations. J. Diff. Eqns. 23, 335367 (1977)
105. Casti, A.R.R., Ormutag, A., Sornborger, A., Kaplan, E., Knight, B., Victor, J., Sirovich, L.:
A population study of integrateandfireor burst neurons. Neural Comput. 14, 957986
(2002)
106. Charles, A.C., Merrill, J.E., Dickson, E.R., Sanderson, M.J.: Intercellular signaling in glial
cells: calcium waves and oscillations in response to mechanical stimulation and glutamate.
Neuron 6, 983992 (1991)
107. Charles, A.C., Naus, C.C., Zhu, D., Kidder, G.M., Dirksen, E.R., Sanderson, M.J.: Intercellular calcium signaling via gap junctions in glioma cells. J. Cell. Biol. 118, 195201 (1992)
108. Chen, X.: Existence, uniqueness, and asymptotic stability of traveling waves in nonlocal
evolution equations. Adv. Diff. Eqns. 2, 125 (1997)
109. Chen, K.C., Nicholson, C.: Spatial buffering of potassium ions in brain extracellular space.
Biophys. J. 78, 27762797 (2000)
110. Cheng, H., Lederer, W.J.: Calcium sparks. Physiol. Rev. 88, 14911545 (2008)
111. Cheng, H., Lederer, M.R., Lederer, W.J., Cannell, M.B.: Calcium sparks and waves in cardiac
myocytes. Am. J. Physiol. 270, C148C159 (1996)
112. Chervin, R.D., Pierce, P.A., Connors, B.W.: Periodicity and directionality in the propagation
of epileptiform discharges across neocortex. J. Neurophysiol. 60, 16951713 (1988)
113. Chialvo, D.: Emergent complex neural dynamics. Nat. Phys. 6, 744750 (2010)
114. Chong, S.C., Tadin, D., Blake, R.: Endogenous attention prolongs dominance durations in
binocular rivalry. J. Vision 5, 10441012 (2005)
115. Chow, C.C., Buice, M.: Path integral methods for stochastic differential equations. arXiv
nlin/105966v1 (2011)
116. Chow, C.C., White, J.A.: Spontaneous action potentials due to channel fluctuations. Biophys.
J. 71, 30133021 (1996)
117. Chu, P.H., Milton, J.G., Cowan, J.D.: Connectivity and the dynamics of integrate-and-fire
neural networks. Int. J. Bif. Chaos 4, 237243 (1994)
118. Chuquet, J., Hollander, L., Nimchinsky., E.A.: High-resolution in vivo imaging of the
neurovascular unit during spreading depression. J. Neurosci. 27, 40364044 (2007)
119. Churchland, M.M., et al.: Stimulus onset quenches neural variability: a widespread cortical
phenomenon. Nat. Neurosci. 13, 369378 (2010)
120. Citri, A., Malenka, R.C.: Synaptic plasticity: multiple forms, functions, and mechanisms.
Neuropsychopharmacol. 33, 1841 (2008)
121. Cohen, A.H., Ermentrout, G.B., Kiermel, T., Kopell, N., Sigvardt, K.A., Williams, T.L.:
Modeling of intersegmental coordination in the lamprey central pattern generator for motion.
Trend. Neurosci. 15, 434438 (1992)
122. Compte, A., Sanchez-Vives, M.V., McCormick, D.A., Wang, X.J.: Cellular and network
mechanisms of slow oscillatory activity (< 1 Hz) and wave propagations in a cortical network
model. J. Neurophysiol. 89, 27072725 (2003)
123. Collingridge, G.L., Peineau, S., Howland, J.G., Wang, Y.T.: Long-term depression in the
CNS. Nat. Rev. Neurosci. 11, 459473 (2010)
124. Compte, A., Reig, R., Descalzo, V.F., Harvey, M.A., Puccini, G.D., Sanchez-Vives, M.V.:
Spontaneous high-frequency (1080 Hz) oscillations during up states in the cerebral cortex
in vitro. J. Neurosci. 28, 1382813844 (2008)
410
References
125. Connors, B.W., Amitai, Y.: Epilespy: Models, mechanisms and concepts. In: Schwartkroin,
P.A. (ed.) Generation of Epileptiform Discharge by Local Circuits of Neocortex, pp.
388423. Cambridge University Press, Cambridge (1993)
126. Connors, B.W., Long, M.A.: Electrical synapses in the mammalian brain. Annu. Rev. Neurosci. 27, 393418 (2004)
127. Coombes, S.: The effects of ion pumps on the speed of traveling waves in the fire-diffuse-fire
model of Ca2+ release. Bull. Math. Biol. 63, 120 (2001)
128. Coombes, S.: Dynamics of synaptically coupled integrate-and-fire-or-burst neuron. Phys.
Rev. E 67, 041910 (2003)
129. Coombes, S., Bressloff, P.C.: Solitary waves in a model of dendritic cable with active spines.
SIAM J. Appl. Math. 61, 432453 (2000)
130. Coombes, S., Bressloff, P.C.: Saltatory waves in the spike-diffuse-spike model of active dendritic spines. Phys. Rev. Lett. 91, 028102 (2003)
131. Coombes, S., Bressloff, P.C. (eds.): Bursting: The Genesis of Rhythm in the Nervous System.
World Scientific Press, London (2005)
132. Coombes, S., Laing, C.R.: Pulsating fronts in periodically modulated neural field models.
Phys. Rev. E 83, 011912 (2011)
133. Coombes, S., Lord, G.J.: Intrinsic modulation of pulse-coupled integrate-and-fire neurons.
Phys. Rev. E 56(5), 58095818 (1997)
134. Coombes, S., Owen, M.R.: Evans functions for integral neural field equations with Heaviside
firing rate function. SIAM J. Appl. Dyn. Syst. 4, 574600 (2004)
135. Coombes, S., Owen, M.R.: Bumps, breathers, and waves in a neural network with spike
frequency adaptation. Phys. Rev. Lett. 94, 148102 (2005)
136. Coombes, S., Schmidt, H.: Neural fields with sigmoidal firing rates: approximate solutions.
Discret. Contin. Dyn. Syst. Ser. A 28, 13691379 (2010)
137. Coombes, S., Timofeeva, Y.: Sparks and waves in a stochastic fire-diffue-fire model of Ca2+
release. Phy. Rev. E 68, 021915 (2003)
138. Coombes, S., Owen, M.R., Smith, G.D.: Mode-locking in a periodically forced integrateand-fire-or-burst neuron model. Phys. Rev. E 64, 041914 (2001)
139. Coombes, S., Lord, G.J., Owen, M.R.: Waves and bumps in neuronal networks with
axo-dendritic synaptic interactions. SIAM J. Appl. Dyn. Sys. 5, 552574 (2003)
140. Coombes, S., Hinch, R., Timofeeva, Y.: Receptors, sparks and waves in a fire-diffuse-fire
framework for Ca2+ release. Prog. Biophys. Mol. Biol. 85, 197216 (2004)
141. Coombes, S., Timofeeva, Y., Svensson, C.M., Lord, G.J., Josic, K., Cox, S.J., Colbert, C.M.:
Branching dendrites with resonant membrane: a sum-over-trips approach. Biol. Cybern.
97, 137149 (2007)
142. Coombes, S., Venkov, N.A., Bojak, I., Liley, D.T.J., Laing, C.R.: Modeling electrocortical
activity through improved local approximations of integral neural field equations. Phys. Rev.
E 76, 051901 (2007)
143. Cornell-Bell, A.H., Finkbeiner, S.M., Cooper, M.S., Smith, S.J.: Glutamate induces calcium
waves in cultured astrocytes: long-range glial signaling. Science 247, 470473 (1990)
144. Dahlem, M.A., Chronicle, E.P.: A computational perspective on migraine aura. Prog.
Neurobiol. 74, 351361 (2004)
145. Dahlem, M.A., Muller, S.C.: Migraine aura dynamics after reverse retinotopic mapping of
weak excitation waves in the primary visual cortex. Biol. Cybern. 88, 419424 (2003)
146. Dale, N.: Dynamic ATP signalling and neural development. J. Physiol. 586, 24292436
(2008)
147. Deco, G., Jirsa, V., Robinson, P.A., Breakspear, M., Friston, K.: The dynamic brain: from
spiking neurons to neural masses and cortical fields. PLoS Comput. Biol. 4, e1000092 (2008)
148. Deco, G., Jirsa, V.K., McIntosh., A.R.: Emerging concepts for the dynamical organization of
resting-state activity in the brain. Nat. Rev. Neurosci. 12, 4356 (2011)
149. Delaney, K.R., Galperin, A., Fee, M.S., Flores, J.A., Gervais, R., Tank, D.W., Kleinfeld, D.:
Waves and stimulus-modulated dynamics in an oscillating olfactory network. Proc. Natl.
Acad. Sci. USA 91, 669673 (1994)
References
411
150. Delorme, V., Machacek, M., DerMardirossian, C., Anderson, K.L., Wittmann, T., Hanein,
D., Waterman-Storer, C., Danuser, G., Bokoch, G.M.: Cofilin activity downstream of Pak1
regulates cell protrusion efficiency by organizing lamellipodium and lamella actin networks.
Dev. Cell. 13, 646662 (2007)
151. Derkach, V., Barria, A., Soderling, T.R.: Ca2+ /calmodulin-kinase II enhances channel conductance of -amino-3-hydroxyl-5-methyl-4-isoxazoleproprionate type glutamate receptors.
Proc. Nat. Acad. Sci. (USA) 96, 32693274 (1999)
152. Destexhe, A., Contreras., D.: Neuronal computations with stochastic network states. Science
314, 8590 (2006)
153. Destexhe, A., Sejnowski, T.J.: Thalamocortical Assemblies: How Ion Channels, Single Neurons and Large-Scale Networks Organize Sleep Oscillations. Oxford University Press,
Oxford (2001)
154. Destexhe, A., Mainen, Z.F., Sejnowski, T.J.: Synthesis of models for excitable membranes
synaptic transmission and neuromodulation using a common kinetic formalism. J. Comput.
Neurosci. 1, 195231 (1994)
155. Difato, F., Tsushima, H., Pesce, M., Benfenati, F., Blau, A., Chieregatti, E.: The formation of
actin waves during regeneration after axonal lesion is enhanced by BDNF. Sci. Rep. 1 (183),
118 (2011)
156. Dockery, J.D., Keener, J.P.: Diffusive effects on dispersion in excitable media. SIAM J. Appl.
Dyn. 49, 539566 (1989)
157. Doubrovinski, K., Kruse, K.: Cytoskeletal waves in the absence of molecular motors. Europhys. Lett. 83, 18003 (2008)
158. van Drongelen, W., Lee, H.C., Stevens, R.L., Hereld, M.: Propagation of seizure-like activity
in a model of neocortex. J. Clin. Neurophysiol. 24, 182188 (2007)
159. Dyhrfjeld-Johnsen, J., Santhakumar, V., Morgan, R.J., Huerta, R., Tsimring, L., Soltesz, I.:
Topological determinants of epileptogenesis in large-scale structural and functional models
of the dentate gyrus derived from experimental data. J. Neurophysiol. 97, 15661587 (2007)
160. Dykman, M.I., Mori, E., Ross, J., Hunt, P.M.: Large fluctuations and optimal paths in chemical kinetics. J. Chem. Phys. A 100, 57355750 (1994)
161. Earnshaw, B.A., Bressloff, P.C.: Diffusion-activation model of CaMKII translocation waves
in dendrites. J. Comput. Neurosci. 28, 7789 (2010)
162. Ebert, U., van Saarloos, W.: Front propagation into unstable states: universal algebraic convergence towards uniformly translating pulled fronts. Physica D 146, 199 (2000)
163. Edwards, J.R., Gibson, W.G.: A model for Ca2+ waves in networks of glial cells incorporating both intracellular and extracellular communication pathways. J. Theor. Biol. 263, 4558
(2010)
164. Elgart, V., Kamenev, A.: Rare event statistics in reactiondiffusion systems. Phys. Rev. E 70,
041106 (2004)
165. Elmer, C.E., Vleck, E.S.V.: Analysis and computation of traveling waves solutions of bistable
differential-difference equations. Nonlinearity 12, 771798 (1999)
166. Ermentrout, G.B.: The behaviour of rings of coupled oscillators. Journal of Mathematical
Biology 23, 5574 (1985)
167. Ermentrout, G.B.: The analysis of synaptically generated travelling waves. J. Comput. Neurosci. 5, 191208 (1998)
168. Ermentrout, G.B.: Neural networks as spatio-temporal pattern-forming systems. Rep. Prog.
Phy. 61, 353430 (1998)
169. Ermentrout, G.B., Cowan, J.: A mathematical theory of visual hallucination patterns. Bio.
Cybern. 34, 137150 (1979)
170. Ermentrout, G.B., Kleinfeld, D.: Traveling electrical waves in cortex: insights from phase
dynamics and speculation on a computational role. Neuron 29, 3344 (2001)
412
References
171. Ermentrout, G.B., Kopell, N.: Frequency plateaus in a chain of weakly coupled oscillators.
SIAM J. Appl. Math. 15, 215237 (1984)
172. Ermentrout, G.B., McLeod, J.B.: Existence and uniqueness of travelling waves for a neural
network. Proc. Roy. Soc. Edin. A 123, 461478 (1993)
173. Ermentrout, G.B., Terman, D.: Mathematical Foundations of Neuroscience. Springer, Berlin
(2010)
174. Ermentrout, G.B., Jalics, J.Z., Rubin, J.E.: Stimulus-driven traveling solutions in continuum neuronal models with a general smooth firing rate function. SIAM J. Appl. Math.
70, 30393064 (2010)
175. Escudero, C., Kamanev, A.: Switching rates of multistep reactions. Phys. Rev. E 79, 041149
(2009)
176. Escudero, C., Rodriguez, J.A.: Persistence of instanton connections in chemical reactions
with timedependent rates. Phys. Rev. E 77, 011130 (2008)
177. Evans, J.: Nerve axon equations IV: The stable and unstable impulse. Ind. Univ. Math. J. 24,
11691190 (1975)
178. Evans, L.C., Sougandis, P.E.: A PDE approach to geometric optics for certain semilinear
parabolic equations. Ind. Univ. Math. J. 38, 141172 (1989)
179. Faisal, A.A., Selen, L.P.J., Wolpert, D.M.: Noise in the nervous system. Nat. Rev. Neurosci.
9, 292 (2008)
180. Falcke, M.: On the role of stochastic channel behavior in intracellular Ca2+ dynamics. Biophys. J. 84, 4256 (2003)
181. Falcke, M.: Reading the patterns in living cells - the physics of Ca2+ signaling. Adv. Phys.
53, 255440 (2004)
182. Falcke, M., Tsimiring, L., Levine, H.: Stochastic spreading of intracellular Ca2+ release.
Phys. Rev. E 62, 26362643 (2000)
183. Fath, G.: Propagation failure of traveling waves in a discrete bistable medium. Physica D
116, 176190 (1998)
184. Faugeras, O., Touboul, J., Cessac, B.: A constructive meanfield analysis of multipopulation
neural networks with random synaptic weights and stochastic inputs. Frontiers in Comp.
Neurosci. 3, 128 (2009)
185. Faugeras, O., Veltz, R., Grimbert, F.: Persistent neural states: Stationary localized activity
patterns in nonlinear continuous n-population, q-dimensional neural networks. Neural Comput. 21, 147187 (2009)
186. Fellin, T., Pascual, O., Gobbo, S., Pozzan, T., Haydon, P.G., Carmignoto, G.: Neuronal
synchrony mediated by astrocytic glutamate through activation of extrasynaptic NMDA
receptors. Neuron 43, 729743 (2004)
187. Fellin, T., Pozzan, T., Carmignoto., G.: Purinergic receptors mediate two distinct glutamate
release pathways in hippocampal astrocytes. J. Biol. Chem. 281, 42744284 (2006)
188. Ferster, D., Miller, K.: Neural mechanisms of orientation selectivity in the visual cortex.
Annual review of neuroscience 23(1), 441471 (2000)
189. Fianco, T.A., McCarthy, K.D.: Intracellular astrocyte calcium waves in situ increase the frequency of spontaneous AMPA receptor currents in CA1 pyramidal cells. J. Neurosci. 24,
722732 (2004)
190. Firth, S.I., Wang, C.T., Feller, M.B.: Retinal waves: mechanisms and function in visual system development. Cell Calcium 37, 425432 (2005)
191. Fisher, R.A.: The wave of advance of advantageous genes. Ann. Eugenics 7, 353369 (1937)
192. FitzHugh, R.: Impulses and physiological states in theoretical models of nerve membrane.
Biophys. J. 1, 445466 (1960)
193. Fivaz, M., Bandara, S., Inoue, T., Meyer, T.: Robust neuronal symmetry-breaking by
Ras-triggered local positive feedback. Curr. Biol. 18, 4450 (2008)
194. Flynn, K.C., Pak, C.W., Shaw, A.E., Bradke, F., Bamburg, J.R.: Growth cone-like waves
transport actin and promote axonogenesis and neurite branching. Dev. Neurobiol. 69,
761779 (2009)
References
413
195. Fogelson, A., Zucker, R.S.: Presynaptic calcium diffusion from various arrays of single
channels. implications for transmitter release and synaptic facilitation. Biophys. J. 48,
10031017 (1985)
196. Folias, S.E., Bressloff, P.C.: Breathing pulses in an excitatory neural network. SIAM J. Appl.
Dyn. Syst. 3, 378407 (2004)
197. Folias, S.E., Bressloff, P.C.: Breathers in twodimensional neural media. Phys. Rev. Lett. 95,
208107 (2005)
198. Folias, S.E., Bressloff, P.C.: Stimuluslocked traveling pulses and breathers in an excitatory
neural network. SIAM J. Appl. Math. 65, 20672092 (2005)
199. Fox, R., Rasche, F.: Binocular rivalry and reciprocal inhibition. Attention Percept.
Psychophys. 5(4), 215217 (1969)
200. Fox, R.F., Lu, Y.N.: Emergent collective behavior in large numbers of globally coupled independent stochastic ion channels. Phys. Rev. E 49, 34213431 (1994)
201. Franci, A., Drion, G., Sepulchre, R.: An organizing center in a planar model of neuronal
excitability. SIAM J. Appl. Dyn. Sys. 11, 16981722 (2013)
202. Franks, K.M., Sejnowski, T.J.: Complexity of calcium signaling in synaptic spines. Bioessays
24, 11301144 (2002)
203. Freidlin, M.I.: Limit theorems for large deviations and reaction-diffusion equations. Ann.
Prob. 13, 639675 (1985)
204. Freidlin, M.I.: Geometric optics approach to reactiondiffusion equations. SIAM J. Appl.
Math. 46, 222232 (1986)
205. Freidlin, M.I., Wentzell, A.D.: Random Perturbations of Dynamical Systems. Springer, New
York (1984)
206. Friedman, A., Craciun, G.: A model of intracellular transport of particles in an axon. J. Math.
Biol. 51, 217246 (2005)
207. Friedman, A., Hu, B.: Uniform convergence for approximate traveling waves in linear
reaction-hyperbolic systems. Ind. Univ. Math. J. 56, 21332158 (2007)
208. Fusi, S., Drew, P.J., Abbott, L.F.: Cascade models of synaptically stored memories. Neuron
45, 599611 (2005)
209. Garcia-Ojalvo, J., Sancho, J.M.: External fluctuations in a pattern-forming instability. Phys.
Rev. E 53, 56805689 (1996)
210. Gardiner, C.W.: Handbook of Stochastic Methods, 4th edn. Springer, Berlin (2009)
211. Gardner-Medwin, A.R.: Analysis of potassium dynamics in mammalian brain tissue. J. Physiol. (Lond) 335, 393462 (1983)
212. Gartner, J., Freidlin, M.I.: On the propagation of concentration waves in periodic and random
media. Soviet Math. Dokl. 20, 12821286 (1979)
213. Gerstner, W., van Hemmen, J.L.: Coherence and incoherence in a globally coupled ensemble
of pulseemitting units. Phys. Rev. Lett. 71(3), 312315 (1993)
214. Gerstner, W., Kistler, W.: Spiking Neuron Models. Cambridge University Press, Cambridge
(2002)
215. Gerstner, W., van Hemmen, J.L., Cowan, J.D.: What matters in neuronal locking? Neural
Comput. 8, 16891712 (1996)
216. Gilbert, C.D.: Horizontal integration and cortical dynamics. Neuron 9, 113 (1992)
217. Gilbert, C.D., Das, A., Ito, M., Kapadia, M., Westheimer, G.: Spatial integration and cortical
dynamics. Proc. Nat. Acad. Sci. USA 93, 615622 (1996)
218. Gilbert, C.D., Wiesel, T.N.: Clustered intrinsic connections in cat visual cortex. J. Neurosci.
3, 11161133 (1983)
219. Gilboa, G., Chen, R., Brenner, N.: Historydependent multiple-time-scale dynamics in a
single-neuron model. J. Neurosci. 25, 64796489 (2005)
220. Ginzburg, I., Sompolinsky, H.: Theory of correlations in stochastic neural networks. Phys.
Rev. E 50, 31713191 (1994)
221. Glass, L., Mackey, M.C.: From Clocks to Chaos. Princeton University Press, Princeton
(1988)
222. Godfrey, K.B., Eglen, S.J.: Theoretical models of spontaneous activity generation and
propagation in the developing retina. Mol. Biosyst. 5, 15271535 (2009)
414
References
223. Godfrey, K.B., Swindale, N.V.: Retinal wave behavior through activity-dependent refractory
periods. PLoS Comp. Biol. 3 (11), 24082420 (2007)
224. Goel, P., Sneyd, J., Friedman, A.: Homogenization of the cell cytoplasm: the calcium bidomain equations. SIAM Nultiscal. Model Simul. 5, 10451062 (2006)
225. Goldberg, J.L.: How does an axon grow? Genes Dev. 17, 941958 (2003)
226. Golding, N.L., Staff, N.P., Spruston, N.: Dendritic spikes as a mechanism for cooperative
long-term potentiation. Nature 418, 326331 (2002)
227. Goldwyn, J.H., Shea-Brown, E.: The what and where of adding channel noise to the
Hodgkin-Huxley equations. PLoS Comp. Biol. 7 (11), e1002247 (2011)
228. Golomb, D., Amitai, Y.: Propagating neuronal discharges in neocortical slices: Computational and experimental study. J. Neurophysiol. 78, 11991211 (1997)
229. Golomb, D., Ermentrout, G.: Bistability in pulse propagation in networks of excitatory and
inhibitory populations. Phys. Rev. Lett. 86, 41794182 (2001)
230. Golomb, D., Ermentrout, G.B.: Continuous and lurching traveling waves in neuronal networks with spatially-decaying connectivity and delay. Proc. Natl. Acad Sci. USA 96,
1348013485 (1999)
231. Golomb, D., Rinzel, J.: Clustering in globally coupled inhibitory neurons. Physica D 72,
259282 (1994)
232. Golomb, D., Wang, X.J., Rinzel, J.: Propagation of spindle waves in a thalamic slice model.
J. Neurophysiol. 75, 750769 (1996)
233. Gomez, T.M., Spitzer, N.C.: Regulation of growth cone behavior by calcium: New dynamics
to earlier perspectives. J. Neurobiol. 44, 174183 (1999)
234. Gomez, T.M., Zheng, J.Q.: The molecular basis for calcium-dependent axon pathfinding.
Nat. Rev. Neurosci. 7, 115125 (2006)
235. Gonzalez-Billault, C., Munoz-Llancao, P., Henriquez, D.R., Wojnacki, J., Conde, C., Caceres, A.: The role of small GTPases in neuronal morphogenesis and polarity. Cytoskeleton
69, 464485 (2012)
236. Gordon, G.R., Baimoukhametova, D.V., Hewitt, S.A., Kosala, W.R.A., Rajapaksha, J.S.,
Fisher, T.E., Bains, J.S.: Noropinephrine triggers release of glial ATP to increase postsynaptic efficacy. Nat. Neurosci. 8, 10781086 (2005)
237. Gorji, A.: Spreading depression: A review of the clinical relevance. Brain Res. Revs. 38,
3360 (2001)
238. Gourley, S.A.: Travelling front solutions of a nonlocal fisher equation. J. Math Biol. 41,
272284 (2000)
239. Grannan, E.R., Kleinfeld, D., Sompolinsky, H.: Stimulus dependent synchronization of neuronal assemblies. Neural Comput. 5, 550569 (1993)
240. Gray, W.G., Lee, P.C.Y.: On the theorems for local volume averaging of multiphase systems.
Int. J. Multiphase Flow 3, 333340 (1977)
241. Greenstein, J.L., Hinch, R., Winslow, R.L.: Mechanisms of excitation-contraction coupling
in an integrative model of the cardiac ventricular myocyte. Biophys. J. 90, 7791 (2006)
242. Grindrod, P.: Patterns and Waves: The Theory and Application of Reaction-Diffusion Equations. Clarendon Press, Oxford (1991)
243. Grinvald, A., Lieke, E.E., Frostig, R.D., Hildesheim, R.: Cortical point-spread function and
long-range lateral interactions revealed by real-time optical imaging of macaque monkey
primary visual cortex. J. Neurosci. 14, 25452568 (1994)
244. Groff, J.R., DeRemigio, H., Smith, G.D.: Stochastic methods in neuroscience, chap. 2. In:
Markov Chain Models of Ion Channels and Calcium Release Sites, pp. 2964. Oxford University Press, Oxford (2009)
245. Gu, X., Olson, E.C., Spitzer, N.: Spontaneous neuronal calcium spikes and waves during
early differentiation. . J. Neurosci. 14, 63256335 (1994)
246. Gu, X., Spitzer, N.C.: Breaking the code: regulation of neuronal differentiation by
spontaneous calcium transients. Dev. Neurosci. 19, 3341 (1997)
247. Guan, C.B., Xu, H.T., Jin, M., Yuan, X.B., Poo, M.M.: Long-range Ca2+ signaling from
growth cone to soma mediates reversal of neuronal migration induced by Slit-2. Cell 129,
385395 (2007)
References
415
248. Guckenheimer, J., Holmes, P.J.: Nonlinear Oscillations, Dynamical Systems and Bifurcations
of Vector Fields. Springer, New York (1983)
249. Guthrie, P.B., Knappenberger, J., Segal, M., Bennett, M.V., , Charles, A.C., Kater, S.B.: ATP
released from astrocytes mediates glial calcium waves. J. Neurosci. 19, 520528 (1999)
250. Haas, B., Schipke, C.G., Peters, O., Sohl, G., Willecke, K., Kettenmann, H.: Activitydependent ATP-waves in the mouse neocortex are independent from astrocytic calcium
waves. Cereb. Cortex 16, 237246 (2006)
251. Hagberg, A., Meron, E.: Pattern formation in nongradient reactiondiffusion systems: the
effects of front bifurcations. Nonlinearity 7, 805835 (1994)
252. Hagberg, A., Meron, E., Rubinstein, I., Zaltzman., B.: Controlling domain patterns far from
equilibrium. Phys. Rev. Lett. 76, 427430 (1996)
253. Halassa, M.M., Fellin, T., Haydon, P.G.: The tripartite synapse: roles for gliotransmission in
health and disease. Trend. Mol. Med. 13, 5463 (2007)
254. Hall, A.: Rho GTPases and the actin cytoskeleton. Science 279, 509514 (1998)
255. Han, F., Caporale, N., Dan, Y.: Reverberation of recent visual experience in spontaneous
cortical waves. Neuron 60, 321327 (2008)
256. Hanggi, P., Grabert, H., Talkner, P., Thomas, H.: Bistable systems: master equation versus
FokkerPlanck modeling. Z. Physik B 28, 135 (1984)
257. Hanggi, P., Talkner, P., Borkovec, M.: Reaction-rate theory: fifty years after Kramers. Rev.
Mod. Phys. 62, 251341 (1990)
258. Hansel, D., Mato, G., Meunier, C.: Synchrony in excitatory neural networks. Neural Comput.
7, 23072337 (1995)
259. Hansen, M., Boltano, S., Dirksen, E.R., Sanderson, M.J.: Intercellular calcium signaling
induced by extracellular adenosine 5-triphosphate and mechanical stimulation in airway
epithelial cells. J. Cell. Sci. 106, 9951004 (1993)
260. Hanson, M.G., Landmesser, L.T.: Chracaterization of the circuits that generate spontaneous
episodes of activity in the early embryonic mouse spinal cord. J. Neurosci. 23, 587600
(2003)
261. Hanson, P.I., Meyer, T., Stryer, L., Schulman, H.: Dual role of calmodulin in autophosphorylation of multifunctional cam kinase may underlie decoding of calcium signal. Neuron 12,
943956 (1994)
262. Hardy, J.: A hundred years of Alzheimers disease research. Neuron 52, 313 (2006)
263. Harper, J.D., Lansbury, P.T. Jr.: Models of amyloid seeding in Alzheimers disease and
scrapie: mechanistic truths and physiological consequences of the time-dependent solubility of amyloid proteins. Annu. Rev. Biochem. 66, 385407 (1997)
264. Harris, T.E.: The theory of branching processes. Dover, New York (1989)
265. Harris-White, M.E., Zanotti, S.A., Frautschy, S.A., Charles, A.C.: Spiral intercellular calcium waves in hippocampal slice cultures. J. Neurophysiol. 79, 10451052 (1998)
266. Hassinger, T.D., Guthrie, P.B., Atkinson, P.B., Bennett, M.V., Kater, S.B.: An extracellular
signaling component in propagation of astrocytic calcium waves. Proc. Natl. Acad. Sci. USA
93, 1326813273 (1996)
267. Hastings, S.P.: The existence of progressive wave solutions to the Hodgkin-Huxley equations.
Arc. Rat. Mech. Anal. 60, 229257 (1975)
268. Haydon, P.G.: Glia: listening and talking to the synapse. Nat. Rev. Neurosci. 2, 185193
(2000)
269. Heitmann, S., Gong, P., Breakspear, M.: A computational role for bistability and traveling
waves in motor cortex. Front. Comp. Neurosci. 6 (67), 115 (2012)
270. von Helmholtz, H.: Treatise on physiological optics. Dover, New York (1866)
271. Hennig, M.H., Adams, C., Willshaw, D., Sernagor, E.: Early-stage waves in the retinal network emerge close to a critical state transition between local and global functional connectivity. J. Neurosci. 29, 10771086 (2009)
272. Hill, S., Tononoi, G.: Modeling sleep and wakefulness in the thalamocortical system. J.
Neurophysiol. 93, 16711698 (2005)
416
References
273. Hill, A.A.V., Masino, M.A., Calabrese, R.L.: Intersegmental coordination of rhythmic motor
patterns. J. Neurophysiol. 90, 531538 (2003)
274. Hillen, T., Othmer, H.: The diffusion limit of transport equations derived from velocity-jump
processes. SIAM J. Appl. Math. 61, 751775 (2000)
275. Hinch, E.J.: Perturbation Methods. Cambridge University Press, Cambridge (1991)
276. Hinch, R.: A mathematical analysis of the generation and termination of calcium sparks.
Biophys. J. 86, 12931307 (2004)
277. Hinch, R., Chapman, S.J.: Exponentially slow transitions on a Markov chain: the frequency
of calcium sparks. Eur. J. Appl. Math. 16, 427446 (2005)
278. Hirsch, J.D., Gilbert, C.D.: Synaptic physiology of horizontal connections in the cats visual
cortex. J. Physiol. Lond. 160, 106154 (1991)
279. Hodgkin, A.L., Huxley, A.F.: A quantitative description of membrane and its application to
conduction and excitation in nerve. J. Physiol. 117, 500544 (1952)
280. Hofer, T., Politi, A., Heinrich, R.: Intercellular Ca2+ wave propagation through gapjunctional Ca2+ diffusion: a theoretical study. Biophys. J. 80, 7587 (2001)
281. Hofer, T., Venance, L., Giaume, C.: Control and plasticity of intercellular calcium waves in
astrocytes: a modeling approach. J. Neurosci. 22, 48504859 (2002)
282. Holcman, D., Tsodyks, M.: The emergence of up and down states in cortical networks. PLoS
Comp. Biol. 2(3), 01740181 (2006)
283. Holcman, D., Schuss, Z., Korkotian, E.: Calcium dynamics in dendritic spines and spine
motility. Biophys. J. 87, 8191 (2004)
284. Holmes, B.B., Diamond, M.I.: Cellular mechanisms of protein aggregate propagation. Curr.
Opin. Neuro. 25, 721726 (2012)
285. Holmes, E.E., Lewis, M.A., Banks, J.E., Veit, R.R.: Partial differential equations in ecology:
spatial interactions and population dynamics. Ecology 75, 1729 (1994)
286. Holmes, W.R., Carlsson, A.E., Edelstein-Keshet, L.: Regimes of wave type patterning driven
by refractory actin feedback: transition from static polarization to dynamic wave behaviour.
Phys. Biol. 9, 046005 (2012)
287. Hoogland, T.M., Kuhn, B., Gobel, W., Huang, W., Nakai, J., Helmchen, F., Flint, J., Wang,
S.S.: Radially expanding transglial calcium waves in the intact cerebellum. Proc. Natl. Acad.
Sci. USA 106, 34963501 (2009)
288. Hopfield, J.J.: Neurons with graded response have collective computational properties like
those of twostate neurons. Proc. Natl. Acad. Sci. USA 81, 30883092 (1984)
289. Horton, A.C., Ehlers, M.D.: Neuronal polarity and trafficking. Neuron 40, 277295 (2003)
290. Houweling, A.R., Bazhenov, M., Timofeev, I., Steriade, M., Sejnowski, T.J.: Homeostatic
synaptic plasticity can explain post-traumatic epileptogenesis in chronically isolated neocortex. Cereb. Cortex 15, 834845 (2005)
291. Hsu, D., Tang, A., Hsu, M., Beggs, J.M.: Simple spontaneously active Hebbian learning
model: homeostatsis of activity and connectivity, and consequences for learining and epileptogenesis. Phys. Rev. E 76, 041909 (2007)
292. Huang, X., Troy, W.C., Yang, Q., Ma, H., Laing, C.R., Schiff, S.J., Wu, J.: Spiral waves in
disinhibited mammalian neocortex. J. Neurosci. 24, 98979902 (2004)
293. Hubel, D.H., Wiesel, T.N.: Sequence regularity and geometry of orientation columns in the
monkey striate cortex. J. Comp. Neurol. 158, 267294 (1974)
294. Hubel, D.H., Wiesel, T.N.: Uniformity of monkey striate cortex: A parallel relationship between field size, scatter, and magnification factor. J. Comp. Neurol. 158, 295306 (1974)
295. Hudson, A., Schulman, H.: Neuronal Ca2+ /calmodulin-dependent protein kinase II: the role
of structure and autoregulation in cellular function. Annu. Rev. Biochem. 71, 473510 (2002)
296. Hutt, A., Bestehorn, M., Wennekers, T.: Pattern formation in intracortical neuronal fields.
Network 14, 351368 (2003)
297. Hutt, A., Longtin, A., Schimansky-Geier, L.: Additive noise-induces Turing transitions in
spatial systems with application to neural fields and the Swift-Hohenberg equation. Physica
D 237, 755773 (2008)
298. Iacobas, D.A., Suadicani, S.O., Spray, D.C., Scemes, E.: A stochastic two-dimensional model
of intercellular Ca2+ wave spread in glia. Biophys. J. 90, 2441 (2006)
References
417
299. Idan, S., Rinzel, J., Gordon, M.S. (eds.): The Theoretical Foundations of Dendritic Function:
Selected Papers of Wilfrid Rall with Commentaries. MIT Press, Cambridge (1995)
300. Invernizzi, G., Papaleo, E., Sabate, R., Ventura, S.: Protein aggregation: Mechanisms and
functional consequences. Int. J. Biochem. Cell Biol. 44, 15411554 (2012)
301. Izhikevich, E.: Dynamical Systems in Neuroscience: The Geometry of Excitability and
Bursting. MIT Press, Cambridge (2006)
302. Jacobs, K.: Stochastic Processes for Physicists. Cambridge University Press, Cambridge
(2010)
303. Jaffe, D.B., Brown, T.H.: Metabotropic glutamate receptor activation induces calcium waves
within hippocampal dendrites. J. Neurophysiol. 72, 471474 (1994)
304. Jahr, C., Stevens, C.: A quantitative description of NMDA receptor-channel kinetic behavior.
J. Neurosci. 10, 18301837 (1990)
305. Jilkine, A., Maree, A.F.M., Edelstein-Keshet, L.: Mathematical model for spatial segregation
of the Rho-family GTPases based on inhibitory crosstalk. Bull. Math. Biol. 68, 11691211
(2007)
306. Jin, D.Z., Dragoi, V., Sur, M., Seung, H.S.: Tilt aftereffect and adaptation-induced changes
in orientation tuning and visual cortex. J. Neurosci. 94, 40384050 (2005)
307. Jirsa, V.K., Haken, H.: A derivation of a macroscopic field theory of the brain from the quasimicroscopic neural dynamics. Physica D 99, 503526 (1997)
308. Jirsa, V.K., Kelso, J.A.S.: Spatiotemporal pattern formation in neural systems with heterogeneous topologies. Phys. Rev. E 62, 84628465 (2000)
309. van Kampen, N.G.: Stochastic Processes in Physics and Chemistry. North-Holland, Amsterdam (1992)
310. Kang, M., Othmer, H.G.: Spatiotemporal characteristics of calcium dynamics in astrocytes.
Chaos 19, 037116 (2009)
311. Kang, N., Xu, J., Xu, Q., Nedergaard, M., Kang, J.S.: Astrocytic glutamate release-induced
transient depolarization and epileptiform discharges in hippocampal CA1 pyramidal neuron.
J. Neurophysiol. 94, 41214130 (2005)
312. Kang, M., Heeger, D.J., Blake, R.: Periodic perturbations producing phase-locked fluctuations in visual perception. J. Vision 9(2):8, 112 (2009)
313. Kang, M., Lee, S.H., Kim, J., Heeger, D.J., Blake, R.: Modulation of spatiotemporal dynamics of binocular rivalry by collinear facilitation and pattern-dependent adaptation. J. Vision
10(11):3, 115 (2010)
314. Kapitula, T., Kutz, N., Sandstede, B.: The Evans function for nonlocal equations. Ind. Univ.
Math. J. 53, 10951126 (2004)
315. Katz, L.C., Shatz, C.J.: Synaptic activity and the construction of cortical circuits. Science
274, 11331138 (1996)
316. Keener, J.P.: Waves in excitable media. SIAM J. Appl. Math. 39, 528548 (1981)
317. Keener, J.P.: Propagation and its failure in coupled systems of discrete excitable cells. SIAM
J. Appl. Math 31, 269276 (1987)
318. Keener, J.P.: Homogenization and propagation in the bistable equation. Physica D 136, 117
(2000)
319. Keener, J.P.: Propagation of waves in an excitable medium with discrete release sites. SIAM
J. Appl. Math. 61, 317314 (2000)
320. Keener, J.P.: Stochastic calcium oscillations. Math. Med. Biol. 23, 125 (2006)
321. Keener, J.P., Newby, J.M.: Perturbation analysis of spontaneous action potential initiation by
stochastic ion channels. Phy. Rev. E 84, 011918 (2011)
322. Keener, J., Sneyd, J.: Mathematical Physiology I: Cellular Physiology, 2nd edn. Springer,
New York (2009)
323. Keener, J.P., Hoppensteadt, F.C., Rinzel, J.: Integrateandfire models of nerve membrane
response to oscillatory input. SIAM J. Appl. Math. 41(3), 503517 (1981)
324. Keizer, J., Smith, G.D.: Spark-to-wave transition: saltatory transmission of calcium waves in
cardiac myocytes. Biophys. Chem. 72, 87100 (1998)
325. Keizer, J., Smith, G.D., Ponce-Dawson, S., Pearson, J.E.: Saltatory propagation of Ca2+
waves by Ca2+ sparks. Biophys. J. 75, 595600 (1998)
418
References
326. Khazipov, R., Luhmann, H.J.: Early patterns of electrical activity in the developing cerebral
cortex of humans and rodents. Trend. Neurosci. 29, 414419 (2006)
327. Khovanov, I.A., Polovinkin, A.V., Luchinsky, D.G., McClintock, P.V.E.: Noise-induced
escape in an excitable system. Phys. Rev. E 87, 032116 (2013)
328. Kilpatrick, Z.P., Bressloff, P.C.: Binocular rivalry in a competitive neural network with
synaptic depression. SIAM J. Appl. Dyn. Syst. 9, 13031347 (2010)
329. Kilpatrick, Z.P., Bressloff, P.C.: Effects of synaptic depression and adaptation on spatiotemporal dynamics of an excitatory neuronal network. Physica D 239, 547560 (2010)
330. Kilpatrick, Z.P., Bressloff, P.C.: Spatially structured oscillations in a two-dimensional neuronal network with synaptic depression. J. Comp. Neurosci. 28, 193209 (2010)
331. Kilpatrick, Z.P., Bressloff, P.C.: Stability of bumps in piecewise smooth neural fields with
nonlinear adaptation. Physica D 239, 10481060 (2010)
332. Kilpatrick, Z.P., Folias, S.E., Bressloff, P.C.: Traveling pulses and wave propagation failure
in inhomogeneous neural media. SIAM J. Appl. Dyn. Syst. 7, 161185 (2008)
333. Kim, H.G., Connors, B.W.: Apical dendrites of the neocortex: correlation between sodium
and calcium-dependent spiking and pyramidal cell morphology. J. Neurosci. 13, 53015311
(1993)
334. Kim, U., Bal, T., McCormick, D.A.: Spindle waves are propagating synchronized oscillations
in the ferret LGN in vitro. J. Neurophysiol. 74, 13011323 (1995)
335. Kinezaki, N., Kawasaki, K., Takasu, F., Shigesada, N.: Modeling biological invasions into
periodically fragmented environments. Theor. Popul. Biol. 64, 291302 (2003)
336. Kishimoto, K., Amari, S.: Existence and stability of local excitations in homogeneous neural
fields. J. Math. Biol. 7, 303318 (1979)
337. Kistler, W.M., Seitz, R., van Hemmen, J.L.: Modeling collective excitations in cortical tissue.
Physica D 114, 273295 (1998)
338. Kleinfeld, D., Delaney, K.R., Fee, M.S., Flores, J.A., Tank, D.W., Gelperin, A.: Dynamics of
propagating waves in the olfactory network of a terrestrial mollusc: an electrical and optical
study. J. Neurophysiol. 72, 14021419 (1994)
339. Knessl, C., Mangel, M., Matkowsky, B.J., Schuss, Z., Tier, C.: An asymptotic solution of the
Kramers-Moyal equation and first passage time problems for Markov jump processes. Phys.
Rev. E 29, 33593369 (1984)
340. Knessl, C., Matkowsky, B.J., Schuss, Z., Tier, C.: An asymptotic theory of large deviations
for Markov jump processes. SIAM J. Appl. Math. 46, 10061028 (1985)
341. Koch, C.: Cable theory in neurons with active, linearized membranes. Biol. Cybern. 50,
1533 (1984)
342. Koch, C.: Biophysics of Computation. Oxford University Press, New York (1999)
343. Koch, C., Zador, A.M.: The function of dendritic dendritic spines. devices subserving biochemical rather than electrical compartmentalization. J. Neurosci. 13, 413422 (1993)
344. Koch, C., Poggio, T., Torre, V.: Nonlinear interactions in a dendritic tree: localization, timing
and role in information processing. Proc. Nat. Acad. Sci. (USA) 80, 27992802 (1983)
345. Kolmogorff, A., Petrovsky, I., Piscounoff, N.: Etude

de lequation de la diffusion avec croissance de la quantite de mati`ere et son application a` un probl`eme biologique. Moscow Univ.
Bull. Math. 1, 125 (1937)
346. Konur, S., Rabinowitz, D., Fenstermaker, V.L., Yuste, R.: Systematic regulation of spine sizes
and densities in pyramidal neurons. J. Neurobiol. 56, 95112 (2003)
347. Kopell, N., Ermentrout, G.B.: Symmetry and phase-locking in chains of weakly coupled
oscillators. Comm. Pure Appl. Math. 39, 623660 (1986)
348. Kramer, M.A., Cash, S.S.: Epilepsy as a disorder of cortical network organization. Neuroscientist 18, 360372 (2012)
349. von Krosigk, M., Bal, T., McCormick, D.: Cellular mechanisms of a synchronized oscillation
in the thalamus. Science 261, 361364 (1993)
350. Kuramoto, Y.: Chemical Oscillations, Waves and Turbulence. Springer, New-York (1984)
351. Kurtz, T.G.: Limit theorems for a sequence of jump Markov processes approximating ordinary differential equations. J. Appl. Prob. 8, 344356 (1971)
References
419
352. Kurtz, T.G.: Limit theorems and diffusion approximations for density dependent Markov
chains. Math. Prog. Stud. 5, 67 (1976)
353. Lago-Fernandez, L.F., Huerta, R., Corbacho, F., Siguenza, J.A.: Fast response and temporal
coherent oscillations in small-world networks. Phys. Rev. Lett. 84, 27582761 (2000)
354. Laing, C.R.: Spiral waves in nonlocal equations. SIAM J. Appl. Dyn. 4, 588606 (2005)
355. Laing, C.R., Chow, C.C.: A spiking neuron model for binocular rivalry. J. Comput. Neurosci.
12, 3953 (2002)
356. Laing, C.R., Troy, W.C.: PDE methods for nonlocal models. SIAM J. Appl. Dyn. Syst. 2,
487516 (2003)
357. Lam, Y.W., Cohen, L.B., Wachowiak, M., Zochowski, M.R.: Odors elicit three different
oscillations in the turtle olfactory bulb. J. Neurosci. 20, 749762 (2000)
358. Lance, J.W.: Current concepts of migraine pathogenesis. Neurology 43, 1115 (1993)
359. Largo, C., Cuevas, P., Somjen, G., Rio, R.M., Herreras, O.: The effect of depressing glial
function in rat brain in situ on ion homeostasis, synaptic transmission, and neuronal survival.
J. Neurosci. 16, 12191229 (1996)
360. Largo, C., Ibarz, J.M., Herreras, O.: Effects of the gliotoxin fluorocitrate on spreading depression and glial membrane potential in rat brain in situ. J. Neurophysiol. 78, 295307
(1997)
361. Larkum, M.E., Zhu, J.J., Sakmann, B.: A new cellular mechanism for coupling inputs arriving at different cortical layers. Nature 398, 338341 (1999)
362. Larkum, M.E., Nevian, T., Sandler, M., Polsky, A., Schiller, J.: Synaptic integration in tuft
dendrites of layer 5 pyramidal neurons: a new unifying principle. Science 325, 756760
(2009)
363. Lashley, K.: Patterns of cerebral integration indicated by scotomas of migraine. Arch. Neurol.
Pyschiatry 46, 331339 (1941)
364. Lauritsen, K.B., Zapperi, S., Stanley, H.E.: Self-organized branching processes: avalanche
models with dissipation. Phys. Rev. E 54, 24832488 (1996)
365. Leao, A.A.P.: Spreading depression of activity in the cerebral cortex. J. Neurophysiol. 10,
409414 (1944)
366. Leao, A.A.P., Morrisson, R.S.: Propagation of spreading cortical depression. J. Neurophysiol. 8, 3345 (1945)
367. Lechleiter, J., Girard, S., Peralta, E., Clapham, D.: Spiral calcium wave propagation and
annihilation in xenopus laevis oocytes. Science 252, 123126 (1991)
368. Leclerc, C., Neant, I., Moreau, M.: The calcium: an early signal that initiates the formation of
the nervous system during early embryogenesis. Front. Mol. Neurosci. 5 (64), 112 (2012)
369. Lee, S.H., Blake, R., Heeger, D.J.: Traveling waves of activity in primary visual cortex during
binocular rivalry. Nat. Neurosci. 8, 2223 (2005)
370. Lee, S.J., Escobedo-Lozoya, Y., Szatmari, E.M., Yasuda, R.: Activation of CaMKII in single
dendritic spines during long-term potentiation. Nature 458, 299304 (2009)
371. Lee, S.J., Lim, H.S., Masliah, E., Lee, H.J.: Protein aggregate spreading in neurodegerative
diseases: problems and perspectives. Neuro. Res. 70, 339348 (2011)
372. Lehky, S.R.: Binocular rivalry is not chaotic. Proc. Roy. Soc. Lond. Biol Sci. 259, 7176
(1995)
373. Lehky, S.R.: No binocular rivalry in the LGN of alert macaque monkeys. Vis. Res. 36,
12251234 (1996)
374. Lehner, F.K.: On the validity of Ficks law for transient diffusion through a porous medium.
Chem. Eng. Sci. 34, 821825 (1977)
375. Lemieux, L., Daunizeau, J., Walker, M.C.: Concepts of connectivity and human epileptic
activity. Front. Syst. Neurosci. 5 (12), 113 (2011)
376. Leopold, D.A., Logothetis, N.K.: Activity changes in early visual cortex reflect monkeys
percepts during binocular rivalry. Nature 379, 549553 (1996)
377. LeVay, S., Nelson, S.B.: Columnar organization of the visual cortex. In: Leventhal, A.G.
(ed.) The Neural Basis of Visual Function, pp. 266315. CRC Press, Boca Raton (1991)
378. Levelt, W.J.M.: On Binocular Rivalry. Institute for Perception RVOTNO, Soesterberg, The
Netherlands (1965)
379. Leybaert, L., Paemeleire, K., Strahonja, A., Sanderson, M.J.: Inositol-trisphosphatedependent intercellular calcium signaling in and between astrocytes and endothelial cells.
Glia 24, 398407 (1998)
420
References
380. Leybaert, L., Sanderson, M.J.: Intercellular Ca2+ waves: mechanisms and function. Physiol.
Rev. 92, 13591392 (2012)
381. Li, Y., Rinzel, J.: Equations for InsP3 receptor-mediated calcium oscillations derived from
a detailed kinetic model: a Hodgkin-Huxley like formalism. J. Theor. Biol. 166, 461473
(1994)
382. Liley, D.J.T., Cadusch, P.J., Dafilis, M.P.: A spatially continuous mean field theory of electrocortical activity. Network 13, 67113 (2002)
383. Linder, B., Garcia-Ojalvo, J., Neiman, A., Schimansky-Geier, L.: Effects of noise in excitable
systems. Phys. Rep. 392, 321424 (2004)
384. Lisman, J.E.: Relating hippocampal circuitry to function: recall of memory sequences by
reciprocal dentate-CA3 interactions. Neuron 22, 233242 (1999)
385. Lisman, J.: Long-term potentiation: outstanding questions and attempted synthesis. Phil.
Trans. R. Soc. Lond. B 358, 829842 (2003)
386. Lisman, J.E., Schulman, H., Cline, H.: The molecular basis of CaMKII function in synaptic
and behavioral memory. Nat. Rev. Neurosci. 3, 175190 (2002)
387. Litwin-Kumar, A., Doiron, B.: Slow dynamics and high variability in balanced cortical networks with clustered connections. Nat. Neurosci. 15, 14981505 (2012)
388. Liu, Q.S., Xu, Q., Kang, G.A.J., Nedergaard, M.: Astrocyte-mediated activation of neuronal
kainate receptors. Proc. Natl. Acad Sci. USA 101, 31723177 (2004)
389. Llano, I., Gonzalez, J., Caputo, C., Lai, F.A., Blayney, L.M., Tan, Y.P., Marty, A.: Presynaptic
calcium stores underlie large-amplitude miniature IPSCs and spontaneous calcium transients.
Nat. Neurosci. 3, 12561265 (2000)
390. Logothetis, N.K., Leopold, D.A., Sheinberg, D.L.: What is rivalling during binocular rivalry?
Nature 380, 621624 (1996)
391. London, M., Roth, A., Beeren, L., Hausser, M., Latham, P.E.: Sensitivity to perturbations in
vivo implies high noise and suggests rate coding in cortex. Nature 466, 123127 (2010)
392. Loxley, P., Robinson, P.: Soliton model of competitive neural dynamics during binocular
rivalry. Phys. Rev. Lett. 102(25), 258701 (2009)
393. Lubenov, E.V., Siapas, A.G.: Hippocampal theta oscillations are travelling waves. Nature
459, 534539 (2009)
394. Luo, L.: Actin cytoskeleton regulation in neuronalmorphogenesis and structural plasticity.
Ann. Rev. Cell Dev. Biol. 18, 601635 (2002)
395. Ly, C., Tranchina, D.: Critical analysis of a dimension reduction by a moment closure
method in a population density approach to neural network modeling. Neural Comput. 19,
20322092 (2007)
396. Madison, D.V., Nicoll, R.A.: Control of the repetitive discharge of rat CA1 pyramidal neurones in vitro. J. Physiol. 354, 319331 (1984)
397. Magee, J.C., Johnston, D.: Characterization of single voltage-gated Na+ and Ca2+ channels
in apical dendrites of rat CA1 pyramidal neurons. J. Physiol. 487, 6790 (1995)
398. Maier, R.S., Stein, D.L.: Limiting exit location distribution in the stochastic exit problem.
SIAM J. Appl. Math 57, 752790 (1997)
399. Maier, N., Nimmrich, V., Draguhn, A.: Cellular and network mechanisms underlying spontaneous sharp waveripple complexes in mouse hippocampal slices. J. Physiol. 550, 873887
(2003)
400. Major, G., Polsky, A., Denk, W., Schiller, J., Tank, D.W.: Spatiotemporally graded NMDA
spike/plateau potentials in basal dendrites of neocortical pyramidal neurons. J. Neurophysiol.
99, 25842601 (2008)
401. Malach, R., Harel, Y.A.M., Grinvald, A.: Relationship between intrinsic connections and
functional architecture revealed by optical imaging and in vivo targeted biocytin injections
in primate striate cortex. Proc. Natl. Acad. Sci. USA 90, 046910473 (1993)
402. Malenka, R.C., Bear, M.F.: LTP and LTD: an embarrassment of riches. Neuron 44, 521
(2004)
403. Manita, S., Ross, W.N.: Synaptic activation and membrane potential changes modulate the
frequency of spontaneous elementary Ca2+ releases events in the dendrites of pyramidal
neurons. J. Neurosci. 29, 78337845 (2009)
References
421
404. Marder, E., Calabrese, R.L.: Principles of rhythmic motor pattern generation. Physiol. Rev.
76, 687717 (1996)
405. Markram, H., Tsodyks, M.: Redistribution of synaptic efficacy between neocortical pyramidal neurons. Nature 382, 807810 (1996)
406. Martins-Ferreira, H., Ribeiro, I.J.: Biphasic effects of gap junctional uncoupling agents on
the propagation of retinal spreading depression. Braz. J. Med. Biol. Res. 28, 991994 (1995)
407. Martins-Ferreira, H., Nedergaard, M., Nicholson, C.: Perspectives on spreading depression.
Brain Res. Rev. 32, 215234 (2000)
408. Masel, J., Jansen, V.A.A., Nowak, M.A.: Quantifying the kinetic parameters of prion replication. Biophys. Chem. 77, 13915 (1999)
409. Masuda, N., Aihara, K.: Global and local synchrony of coupled neurons in small-world networks. Biol. Cybern. 90, 302309 (2004)
410. Mataga, N., Mizuguchi, Y., Hensch, T.K.: Experience-dependent pruning of dendritic spines
in visual cortex by tissue plasminogen activator. Neuron 44, 10311041 (2004)
411. Matthaus, F.: Diffusion versus network models as descriptions for the spread of prion diseasess in the brain. J. Theor. Biol. 240, 104113 (2006)
412. Mattia, M., Guidice, P.D.: Population dynamics of interacting spiking neurons. Phys. Rev. E
66, 051917 (2002)
413. Matus, A.: Actin plasticity in dendritic spines. Science 290, 754758 (2000)
414. Mayer, M.L., Westbrook, G.L., Guthrie, P.B.: Voltage-dependent block by Mg2+ of NMDA
responses in spinal cord neurone. Nature 309, 261263 (1984)
415. McCormick, D.A., Contreras, D.: On the cellular and network bases of epileptic seizures.
Annu. Rev. Physiol. 63, 815846 (2001)
416. McLaughlin, T., Torborg, C.L., Feller, M.B., OLeary, D.D.M.: Retinotopic map refinement
requires spontaneous retinal waves during a brief critical period of development. Neuron 40,
11471160 (2003)
417. Meinhardt, H.: Orientation of chemotactic cells and growth cones: models and mechanisms.
J. Cell. Sci. 112, 28672874 (1999)
418. Meisel, C., Storch, A., Hallmeyer-Elgner, S., Bullmore, E., Gross, T.: Failure of adaptive
self-organized criticality during epileptic seizure attacks. PLoS Comp. Biol. 8, e1002312
(2012)
419. Melamed, O., Barak, O., Silberberg, G., Markram, H., Tsodyks, M.: Slow oscillations in
neural networks with facilitating synapses. J. Comp. Neurosci. 25, 308316 (2008)
420. Menchon, S.A., Gartner, A., Roman, P., Dotti, C.G.: Neuronal (bi)polarity as a self-organized
process enhanced by growing membrane. PLoS One 6(9), e24190 (2011)
421. Mendez, V., Fort, J., Rotstein, H.G., Fedotov, S.: Speed of reaction-diffusion fronts in spatially heterogeneous media. Phys. Rev. E 68, 041105 (2003)
422. Mendez, V., Fedotov, S., Horsthemke, W.: Reaction-Transport Systems. Springer, Berlin
(2010)
423. Meng, X., Huguet, G., Rinzel, J.: Type III excitability, slope sensitivity and coincidence
detection. Disc. Cont. Dyn. Systs, 32, 27292757 (2012)
424. Meunier, C., dIncamps, B.L.: Extending cable theory to heterogeneous dendrites. Neural
Comput. 20, 17321775 (2008)
425. Meyer, C., van Vreeswijk, C.: Temporal correlations in stochastic networks of spiking neurons. Neural Comput. 14, 369404 (2002)
426. Mikhailov, A.S., Davydov, V.A., Zykov, V.S.: Complex dynamics of spiral waves and motion
of curves. Physica D 70, 139 (1994)
427. Miller, J.P., Rall, W., Rinzel, J.: Synaptic amplification by active membrane in dendritic
spines. Brain Res. 325, 325330 (1985)
428. Miller, L.D., Golarai, J.J.P., Connor, J.: Ca2+ release from intracellular stores induced by
afferent stimulation of CA3 pyramidal neurons in hippocampal slices. J. Neurophysiol. 76,
554562 (1996)
429. Millhauser, G.L., Salpeter, E.E., Oswald, R.E.: Diffusion model of ionchannel gating and
the origin of powerlaw distributions from singlechannel recordings. Proc. Natl. Acad. Sci.
USA 85, 15031507 (1988)
422
References
430. Milton, J., Jung, P.: Epilepsy as a Dynamic Disease. Springer, Berlin (2003)
431. Mirollo, R.E., Strogatz, S.H.: Synchronisation of pulsecoupled biological oscillators. SIAM
J. Appl. Math. 50(6), 16451662 (1990)
432. Mitchison, T., Kirschner., M.: Cytoskeletal dynamics and nerve growth. Neuron 1, 761772
(1988)
433. Miura, R., Huang, H., Wylie, J.: Cortical spreading depression: An enigma. Eur. Phys. J.
Special Topics 147, 287302 (2007)
434. Mogilner, A., Oster, G.: Cell motility driven by actin polymerization. Biophys. J. 71,
30303045 (1996)
435. Moldakarimov, S., Rollenhagen, J.E., Olson, C.R., Chow, C.C.: Competitive dynamics in
cortical responses to visual stimuli. J. Neurophysiol. 94, 33883396 (2005)
436. Moreno-Bote, R., Rinzel, J., Rubin, N.: Noise-induced alternations in an attractor network
model of perceptual bistability. J. Neurophysiol. 98, 11251139 (2007)
437. Morgan, R.J., Soltesz, I.: Nonrandom connectivity of the epileptic dentate gyrus predicts a
major role for neuronal hubs in seizures. Proc. Natl. Acad. Sci. USA 105, 61796184 (2008)
438. Mori, Y., Jilkine, A., Edelstein-Keshet, L.: Wave-pinning and cell polarity from a bistable
reaction-diffusion system. Biophys. J. 94, 36843697 (2008)
439. Mori, Y., Jilkine, A., Edelstein-Keshet, L.: Asymptotic and bifurcation analysis of wavepinning in a reaction-diffusion model for cell polarization. SIAM J. Appl. Math. 71,
14011427 (2011)
440. Morris, C., Lecar, H.: Voltage oscillations in the barnacle giant muscle fiber. J. Biophys. 35,
193213 (1981)
441. Mulleners, W.M., Chronicle, E.P., Palmer, J.E., Koehler, P.J., Vredeveld, J.W.: Visual cortex
excitability in migraine with and without aura. Headache 41, 565572 (2001)
442. Muller, L., Destexhe, A.: Propagating waves in thalamus, cortex and the thalamocortical
system: experiments and models. J. Physiol. (Paris) 106, 222238 (2012)
443. Mullins, O.J., Hackett, J.T., Buchanan, J.T., Friesen, W.O.: Neuronal control of swimming
behavior: Comparison of vertebrate and invertebrate model systems. Prog. Neurobiol. 93,
244269 (2011)
444. Murray, J.D.: Mathematical Biology, vol. I, II. Springer, Berlin (2002)
445. Naeh, T., Klosek, M.M., Matkowsky, B.J., Schuss, Z.: A direct approach to the exit problem.
SIAM J. Appl. Math. 50, 595627 (1990)
446. Nagumo, J., Armoto, S., Yoshizawa, S.: An active pulse transmission line simulating nerve
axon. Proc. Inst. Radio Eng. 50, 20612070 (1964)
447. Nakamura, T., Barbara, J.G., Nakamura, K., Ross, W.N.: Synergistic release of Ca2+ from
IP3 sensitive stores evoked by synaptic activation of mGluRs paired with backpropagating
action potentials. Neuron 24, 727737 (1999)
448. Nedergaard, M., Cooper, A.J., Goldman, S.A.: Gap junctions are required for the propagation
of spreading depression. J. Neurobiol. 28, 433444 (1995)
449. Nedergaard, M., Ransom, B., Goldman, S.A.: New roles for astrocytes: redefining the functional architecture of the brain. Trend. Neurosci. 26, 523530 (2003)
450. Netoff, T.I., Schiff, S.J.: Decreased neuronal synchronization during experimental seizures.
J. Neurosci. 22, 72977307 (2002)
451. Netoff, T.I., Clewely, R., Arno, S., Keck, T., White, J.A.: Epilepsy in small-world networks.
J. Neurosci. 24, 80758083 (2004)
452. Neukirchen, D., Bradke, F.: Neuronal polarization and the cytoskeleton. Sem. Cell. Dev.
Biol. 22, 825833 (2011)
453. Newby, J.M., Bressloff, P.C.: Local synaptic signalling enhances the stochastic transport of
motor-driven cargo in neurons. Phys. Biol. 7, 036004 (2010)
454. Newby, J.M., Bressloff, P.C.: Quasi-steady state reduction of molecular-based models of
directed intermittent search. Bull. Math. Biol. 72, 18401866 (2010)
455. Newby, J., Chapman, J.: Metastable behavior in Markov processes with internal states: breakdown of model reduction techniques. J. Math. Biol. published online, DOI: 10.1007/s00285013-0723-1 (2013)
References
423
456. Newby, J.M., Bressloff, P.C., Keeener, J.P.: The effect of potassium channels on spontaneous
action potential initiation by stochastic ion channels. Phys. Rev. Lett. 111, 128101 (2013)
457. Newman, E.A.: Propagation of intercellular calcium waves in retinal astrocytes and Muller
cells. J. Neurosci. 21, 22152223 (2001)
458. Newman, M.E.J.: Spread of epidemic disease on networks. Phys. Rev. E 66, 016128 (2002)
459. Newman, M.E.J.: The structure and function of complex networks. SIAM Rev. 45, 167256
(2003)
460. Newman, E.A., Zahs, K.R.: Calcium waves in retinal glial cells. Science 275, 844847
(1997)
461. Nicholson, C.: Diffusion and related transport mechanisms in brain tissue. Rep. Prog. Phys.
64, 815884 (2001)
462. Nimchinsky, E.A., Sabatini, B.L., Svoboda, K.: Structure and function of dendritic spines.
Ann. Rev. Physiol. 64, 313353 (2002)
463. Nita, D.A., Cisse, Y., Timofeev, I., Steriade, M.: Increased propensity to seizures after
chronic cortical deafferentation in vivo. J. Neurophysiol. 95, 902913 (2006)
464. Noble, J.V.: Geographic and temporal development of plagues. Nature 250, 726729 (1974)
465. Novikov, E.A.: Functionals and the random-force method in turbulence theory. Sov. Phys.
JETP 20, 1290 (1965)
466. Nowak, L., Bregestovski, P., Ascher, P., Herbet, A., Prochiantz, A.: Magnesium gates
glutamate-activated channels in mouse central neurones. Nature 307, 462465 (1984)
467. Nunez, P.I.: Neocortical dynamics and human EEG rhythms. Oxford University Press, New
York (1995)
468. Nykamp, D., Tranchina, D.: A population density method that facilitates largescale modeling of neural networks: analysis and application to orientation tuning. J. Comp. Neurosci. 8,
1950 (2000)
469. Obermayer, K., Blasdel, G.G.: Geometry of orientation and ocular dominance columns in
monkey striate cortex. J. Neurosci. 13, 41144129 (1993)
470. ODonnell, M., Chance, R.K., Bashaw, G.J.: Axon growth and guidance: Receptor regulation
and signal transduction. Ann. Rev. Neurosci. 32, 383412 (2009)
471. ODonovan, M.J.: The origin of spontaneous activity in developing networks of the
vertebrate nervous system. Curr. Opin. Neuro. 9, 94104 (1999)
472. ODonovan, M.J., Chub, N., Wenner, P.: Mechanisms of spontaneous activity in the developing retina. J. Neurobiol. 37, 131145 (1998)
473. Ohira, T., Cowan, J.D.: Stochastic neurodynamics and the system size expansion. In: Ellacott, S., Anderson, I.J. (eds.) Proceedings of the First International Conference on Mathematics of Neural Networks, pp. 290294. Academic, New York (1997)
474. OKeefe, J., Dostrovsky, J.: The hippocampus as a spatial map. preliminary evidence from
unit activity in the freely moving rat. Brain Res. 34, 171174 (1971)
475. OKeefe, J., Recce, M.L.: Phase relationship between hippocampal place units and the EEG
theta rhythm. Hippocampus 3, 317330 (1993)
476. Oksendal, B.: Stochastic Differential Equations: An Introduction with Applications.
Springer, New York (2007)
477. Omurtag, A., Knight, B.W., Sirovich, L.: On the simulation of large populations of neurons.
J. Comput. Neurosci. 8, 5163 (2000)
478. Onsum, M., Rao, C.V.: A mathematical model for neutrophil gradient sensing and polarization. PLoS Comput. Biol. 3, e36 (2007)
479. Ooyen, A.V., Graham, B.P., Ramakers, G.J.A.: Competition for tubulin between growing
neurites during development. Neurocomputing 38, 7378 (2001)
480. Oray, S., Majewska, A., Sur, M.: Dendritic spine dynamics are regulated by monocular
deprivation and extracellular matrix degredation. Neuron 44, 10211030 (2004)
481. Othmer, H., Dunbar, S., Alt, W.: Models of dispersal in biological systems. J. Math. Biol.
26, 263298 (1988)
482. Otsuji, M., Ishihara, S., Co, C., Kaibuchi, K., Mochizuki, A., Kuroda, S.: A mass conserved
reaction-diffusion system captures properties of cell polarity. PLoS Comput. Biol. 3, e108
(2007)
424
References
483. Pak, C.W., Flynn, K.C., Bamburg, J.R.: Actin-binding proteins take the reins in growth cones.
Nat. Rev. Neurosci. 9, 136147 (2008)
484. Pakdaman, K., Thieullen, M., Wainrib, G.: Fluid limit theorems for stochastic hybrid systems
with application to neuron models. J. Appl. Prob. 24, 1 (2010)
485. Panatier, A., Theodosis, D.T., Mothet, J.P., Touquet, B., Pollegioni, L., Poulain, D.A., Oliet,
S.H.: Glia-derived D-serine controls NMDA receptor activity and synaptic memory. Cell
125, 775784 (2006)
486. Panja, D.: Effects of fluctuations on propagating fronts. Phys. Rep. 393, 87174 (2004)
487. Papanicolaou, G.C.: Asymptotic analysis of transport processes. Bull. Amer. Math. Soc. 81,
330392 (1975)
488. Parga, N., Abbott, L.F.: Network model of spontaneous activity exhibiting synchronous transitions between up and down states. Front. Neurosci. 1, 5766 (2007)
489. Parker, I., Ivorra, I.: Localized all-or-none calcium liberation by inositol triphosphate. Science 250, 977979 (1990)
490. Parker, I., Yao, Y.: Regenerative release of calcium from functionally discrete subcellular
stores by inositol triphosphate. Proc. Roy. Soc. Lond. B 246, 269274 (1991)
491. Parpura, V., Basarsky, T.A., Liu, F., Jeftinija, K., Jeftinija, S., Haydon, P.G.: Glutamatemediated astrocyte-neuron signalling. Nature 369, 744747 (1994)
492. Parri, H.R., Gould, T.M., Crunelli, V.: Spontaneous astrocytic Ca2+ oscillations in situ drive
NMDAR-mediated neuronal excitation. Nat. Neurosci. 4, 80381 (2001)
493. Pascual, O., Casper, K.B., Kubera, C., Zhang, J., Revilla-Sanchez, R., Sul, J.Y., Takano, H.,
Moss, S.J., McCarthy, K., Haydon, P.G.: Astrocytic purinergic signaling coordinates synaptic
networks. Science 310, 113116 (2005)
494. de Pasquale, F., Gorecki, J., Poielawski., J.: On the stochastic correlations in a randomly
perturbed chemical front. J. Phys. A 25, 433 (1992)
495. Paullet, J.E., Ermentrout, G.B.: Stable rotating waves in two-dimensional discrete active media. SIAM J. Appl. Math. 54, 17201744 (1994)
496. Pavliotis, G.A., Stuart, A.M.: Multiscale Methods: Averaging and Homogenization.
Springer, New York (2008)
497. Pearson, J.E., Ponce-Dawson, S.: Crisis on skid row. Physica A 257, 141148 (1998)
498. Pearson, R., Catsicas, M., Becker, D., Mobbs, P.: Purinergic and muscarinic modulation of
the cell cycle and calcium signaling in the chick retinal ventricular zone. J. Neurosci. 22,
75697579 (2002)
499. Pearson, R.A., Dale, N., Llaudet, E., Mobbs, P.: ATP released via gap junction hemichannels
from the pigment epithelium regulates neural retinal progenitor proliferation. Neuron 46,
731744 (2005)
500. Pereda, E., Quiroga, R., Bhattacharya, J.: Nonlinear multivariate analysis of neurophysiological signals. Prog. Neurobiol. 77, 137 (2005)
501. Peters, O., Schipke, C.G., Hashimoto, Y., Kettenmann, H.: Different mechanisms promote
astrocyte Ca2+ waves and spreading depression in the mouse neocortex. J. Neurosci. 23,
98889896 (2003)
502. Pikovsky, A., Rosenblum, M., Kurths, J.: Synchronization: A Universal Concept in Nonlinear
Sciences. Cambridge University Press, Cambridge (2003)
503. Pinder, G.F., Gray, W.G.: Essentials of Multiphase Flow and Transport in Porous Media.
Wiley, New York (2008)
504. Pinto, D., Ermentrout, G.B.: Spatially structured activity in synaptically coupled neuronal
networks I: traveling fronts and pulses. SIAM J. Appl. Math 62, 206225 (2001)
505. Pinto, D., Ermentrout, G.B.: Spatially structured activity in synaptically coupled neuronal
networks II: lateral inhibition and standing pulses. SIAM J. Appl. Math. 62, 226243 (2001)
506. Pinto, D., Jackson, R.K., Wayne, C.E.: Existence and stability of traveling pulses in a continuous neuronal network. SIAM J. Appl. Dyn. Syst. 4, 954984 (2005)
507. Pinto, D., Patrick, S.L., Huang, W.C., Connors, B.W.: Initiation, propagation, and termination
of epileptiform activity in rodent neocortex in vitro involve distinct mechanisms. J. Neurosci.
25, 81318140 (2005)
508. Plenz, D., Thiagarajan, T.C.: The organizing principles of neuronal avalanches: cell
assemblies in the cortex? Trend. Neurosci. 30, 101110 (2007)
References
425
509. Polleux, F., Snider, W.: Initiating and growing an axon. Cold Spring Harb Perspect Biol 2,
a001925 (2010)
510. Polonsky, A., Blake, R., Braun, J., Heeger, D.J.: Neuronal activity in human primary visual
cortex correlates with perception during binocular rivalry. Nat. Neurosci. 3, 11531159
(2000)
511. Ponce, S.J., Keizer, J., Pearson, J.E.: Fire-diffuse-fire model of dynamics of intracellular
calcium waves. Proc. Nat. Acad. Sci. USA 96, 60606063 (1999)
512. Prechtl, J.C., Cohen, L.B., Pesaran, B., Mitra, P.P., Kleinfeld, D.: Visual stimuli induce waves
of electrical activity in turtle cortex. Proc. Natl. Acad. Sci. USA 94, 76217626 (1997)
513. Prusiner, S.B.: Prions. Proc. Nat. Acad. Sci. (USA) 95, 1336313383 (1998)
514. Rakic, P.: Elusive radial glial cells: historical and evolutionary perspective. Glia 43, 1932
(2003)
515. Rall, W., Snir, H.A.: Cable theory for dendritic neurons. In: Koch, C., Segev, I. (eds.)
Methods in Neuronal Modelling, 2nd edn. MIT Press, Cambridge (1998)
516. Ramirez, O.A., Couve, A.: The endoplasmic reticulum and protein trafficking in dendrites
and axons. Trend. Cell. Biol. 21, 219227 (2011)
517. Rangan, A.V., Kovacic, G., Cai, D.: Kinetic theory for neuronal networks with fast and slow
excitatory conductances driven by the same spike train. Phys. Rev. E 77, 041915 (2008)
518. Reed, M.C., Venakides, S., Blum, J.J.: Approximate traveling waves in linear reactionhyperbolic equations. SIAM J. Appl. Math. 50(1), 167180 (1990)
519. Renart, A., de la Rocha, J., Bartho, P., Hollender, L., Parga, N., Reyes, A., Harris, K.D.: The
asynchronous state in cortical circuits. Science 327, 587590 (2010)
520. Rhodes, P.: The properties and implications of NMDA spikes in neocortical pyramidal cells.
J. Neurosci. 26, 67046715 (2006)
521. Richardson, K.A., Schiff, S.J., Gluckman, B.J.: Control of traveling waves in the mammalian
cortex. Phys. Rev. Lett. 94, 028103 (2005)
522. Richardson, M.P.: Large scale brain models of epilepsy: dynamics meets connectomics.
Epilepsy (2012)
523. Rinzel, J., Maginu, K.: Kinematic analysis of wave pattern formation in excitable media.
In: Pacault, A., Vidal, C. (eds.) Non-Equilibrium Dynamics in Chemical Systems. Springer,
Berlin (1984)
524. Rinzel, J., Terman, D.: Propagation phenomena in a bistable reactiondiffusion system.
SIAM J. Appl. Math. 42, 11111137 (1982)
525. Rinzel, J., Terman, D., Wang, X., Ermentrout, B.: Propagating activity patterns in large-scale
inhibitory neuronal networks. Science 279, 13511355 (1998)
526. Robinson, P.A.: Patchy propagator, brain dynamics, and the generation of spatially structured
gamma oscillations. Phys. Rev. E 73, 041904 (2006)
527. Robinson, P.A., Rennie, C.J., Wright, J.J.: Propagation and stability of waves of electrical
activity in the cerebral cortex. Phys. Rev. E 56, 826840 (1997)
528. Robinson, P.A., Rennie, C.J., Wright, J.J., Bahramali, H., Gordon, E., Rowe, D.I.: Prediction
of electroencephalographic spectra from neurophysiology. Phys. Rev. E 63, 021903 (2001)
529. Robinson, P.A., Rennie, C.J., Rowe, D.L.: Dynamics of large-scale brain activity in normal
arousal states and epileptic seizures. Phys. Rev. E 65, 041924 (2002)
530. Rocco, A., Ebert, U., van Saarloos, W.: Subdiffusive fluctuations of pulled fronts with
multiplicative noise. Phys. Rev. E. 65, R13R16 (2000)
531. Rockland, K.S., Lund, J.: Intrinsic laminar lattice connections in primate visual cortex. J.
Comp. Neurol. 216, 303318 (1983)
532. Rose, J., Jin, S.X., Craig, A.M.: Heterosynaptic molecular dynamics: locally induced propagating synaptic accumulation of CaM Kinase II. Neuron 61, 351358 (2009)
533. Ross, C.A., Poirier, M.A.: Protein aggregation and neurodegenerative disease. Nat. Med.
10(Suppl.), S10S17 (2004)
534. Ross, W.N.: Understanding calcium waves and sparks in central neurons. Nat. Rev. Neurosci.
13, 157168 (2012)
535. Roxin, A., Riecke, H., Solla, S.A.: Self-sustained activity in a small-world network of
excitable neurons. Phys. Rev. Lett. 92, 198101 (2004)
426
References
536. Rubin, J.E.: A nonlocal eigenvalue problem for the stability of a traveling wave in a neuronal
medium. Discret. Contin. Dyn. Syst. 10, 925940 (2004)
537. Rubin, J., Terman, D.: Geometric singular perturbation analysis of neuronal dynamics. In:
Fiedler, B. (ed.) Handbook of Dynamical Systems: Towards Applications, vol. 2. Elsevier,
Amsterdam (2002)
538. Rubin, J.E., Terman, D.: Explicit maps to predict activation order in multiphase rhythms of
a coupled cell network. J. Math. Neurosci. 2, 4 (2012)
539. Rubin, J., Wechselberger, M.: Giant squid-hidden canard: the 3d geometry of the
HodgkinHuxley model. Biol. Cybern. 97, 532 (2007)
540. Rubin, J., Wechselberger, W.: The selection of mixed-mode oscillations in a Hodgkin-Huxley
model with multiple timescales. Chaos 18, 015105 (2008)
541. Rubinov, M., Sporns, O.: Complex network measures of brain connectivity: Uses and interpretations. Neuroimage 52, 10591069 (2010)
542. Ruthel, G., Banker, G.: Actin-dependent anterograde movement of growth-cone-like structures along growing hippocampal axons: A novel form of axonal transport? Cell Motil.
Cytoskel. 40, 160173 (1998)
543. Ruthel, G., Banker, G.: Role of moving growth cone-like wave structures in the outgrowth of
cultured hippocampal axons and dendrites. J. Neurobiol. 39, 97106 (1999)
544. van Saarloos, W.: Front propagation into unstable states. Phys. Rep. 386, 29222 (2003)
545. Saez, J.C., Berthoud, V.M., Branes, M.C., Martinez, A.D., Beyer., E.C.: Plasma membrane
channels formed by connexins: their regulation and functions. Physiol. Rev. 83, 13591400
(2003)
546. Sagues, F., Sancho, J.M., Garcia-Ojalvo, J.: Spatiotemporal order out of noise. Rev. Mod.
Phys. 79, 829882 (2007)
547. Saitoh, T., Schwartz, J.H.: Phosphorylation-dependent subcellular translocation of a
Ca2+ /calmodulin-dependent protein kinase produces an autonomous enzyme in aplysia neurons. J. Cell. Biol. 100, 835842 (1985)
548. Samuels, D.C., Hentschel, H.G.E., Fine, A.: The origin of neuronal polarization: a model of
axon formation. Phil. Trans. R. Soc. Lond. B 351, 11471156 (1996)
549. Sanchez-Vives, M.V., McCormick, D.A.: Cellular and network mechanisms of rhythmic recurrent activity in neocortex. Nat. Neurosci. 3, 10271034 (2000)
550. Sanderson, M.J., Charles, A.C., Dirksen, E.R.: Mechanical stimulation and intercellular communication increases intracellular Ca2+ in epithelial cells. Cell. Regul. 1, 585596 (1990)
551. Sandstede, B.: Stability of travelling waves. In: Handbook of Dynamical Systems, vol. 2, pp.
9831055. North-Holland, Amsterdam (2002)
552. Sandstede, B.: Evans functions and nonlinear stability of traveling waves in neuronal network
models. Int. J. Bifur. Chaos Appl. Sci. Eng. 17, 26932704 (2007)
553. Sato, T.K., Nauhaus, I., Carandini, M.: Traveling waves in visual cortex. Neuron 75, 218229
(2012)
554. Scemes, E., Giaume, C.: Astrocyte calcium waves: what they are and what they do. Glia 54,
716725 (2006)
555. Schiller, J., Schiller, Y., Stuart, G., Sakmann, B.: Calcium action potentials restricted to distal
apical dendrites of rat neocortical pyramidal neurons. J. Physiol. 505, 605616 (1997)
556. Schiller, J., Major, G., Koester, H.J., Schiller, Y.: NMDA spikes in basal dendrites of cortical
pyramidal neurons. Nature 404, 285289 (2000)
557. Schimansky-Geier, L., Mikhailov, A.S., Ebeling., W.: Effects of fluctuations on plane front
propagation in bistable nonequilibrium systems. Ann. Phys. 40, 277 (1983)
558. Schock, S.C., Munyao, N., Yakubchyk, Y., Sabourin, L.A., Hakim, A.M., Ventureyra, E.C.G.,
Thompson, C.S.: Cortical spreading depression releases ATP into the extracellular space and
purinergic receptor activation contributes to the induction of ischemic tolerance. Brain Res.
1168, 129138 (2007)
559. Schuss, Z.: Theory and applications of stochastic processes: an analytical approach. In:
Applied Mathematical Sciences, vol. 170. Springer, New York (2010)
560. Schuster, H.G., Wagner, P.: A model for neuronal oscillations in visual cortex. Biol. Cybern.
64, 7782 (1990)
References
427
561. Schutz, P., Bode, M., Purwins, H.G.: Bifurcations of front dynamics in a reaction-diffusion
system with spatial inhomogeneities. Physica D 82, 382397 (1995)
562. Schwartz, E.: Spatial mapping in the primate sensory projection: analytic structure and relevance to projection. Biol. Cybern. 25, 181194 (1977)
563. Seely, J., Chow, C.C.: Role of mutual inhibition in binocular rivalry. J. Neurophysiol. 106,
21362150 (2011)
564. Seifert, G.: Astrocyte dysfunction in neurological disorders; a molecular perspective. Nat.
Rev. Neurosci. 7, 194206 (2006)
565. Selkoe, D.J.: Folding proteins in fatal way. Nature 426, 900904 (2003)
566. Selkoe, D.J.: Cell biology of protein misfolding: the examples of Alzheimers and Parkinsons diseases. Nat. Cell Biol. 6, 10541061 (2004)
567. Shadlen, M.N., Newsome, W.T.: Noise, neural codes and cortical organization. Curr. Opin.
Neurobiol. 4, 569579 (1994)
568. Shapiro, B.E.: Osmotic forces and gap junctions in spreading depression: A computational
model. J. Comp. Neurosci. 10, 99120 (2001)
569. Shen, K., Meyer, T.: Dynamic control of CaMKII translocation and localization in hippocampal neurons by NMDA receptor simulation. Science 284, 162166 (1999)
570. Shen, K., Tereul, M.N., Subramanian, K., Meyer, T.: CaMKII functions as an F-actin targeting module that localizes CaMKII / heterooligomers to dendritic spines. Neuron 21,
593606 (1998)
571. Shepherd, G.M.: The dendritic spine: a multifunctional unit. J. Neurophysiol. 75, 21972210
(1996)
572. Shepherd, G.M., Brayton, R.K., Miller, J.P., Segev, I., Rinzel, J., Rall, W.: Signal enhancement in distal cortical dendrites by means of interactions between active dendritic spines.
Proc. Natl. Acad. Sci. USA 82, 21922195 (1985)
573. Shew, W.L., Plenz, D.: The functional benefits of criticality in the cortex. Neuroscientist 19,
88100 (2012)
574. Shigesada, N., Kawasaki, K., Teramoto, E.: Traveling periodic waves in heterogeneous environments. Theor. Popul. Biol. 30, 143160 (1986)
575. Shigesada, N., Kawasaki, K.: Biological invasions: theory and practice. Oxford University
Press, Oxford (1997)
576. Shimada, T., Toriyama, M., Uemura, K., Kamiguchi, H., Sugiura, T., Watanabe, N., Inagaki,
N.: Shootin1 interacts with actine retrograde flow and L1-CAM to promote axon outgrowth.
J. Cell. Biol. 181, 817829 (2008)
577. Shpiro, A., Moreno-Bote, R., Rubin, N., Rinzel, J.: Dynamical characteristics common to
neuronal competition models. J. Neurophysiol. 97, 3754 (2007)
578. Shpiro, A., Curtu, R., Rinzel, J., Rubin, N.: Balance between noise and adaptation in competition models of perceptual bistability. J Comp. Neurosci. 27, 462473 (2009)
579. Shuai, J.W., Jung, P.: Stochastic properties of Ca2+ release of Inositol 1,4,5-Triphosphate
receptor clusters. Biophys. J. 83, 8797 (2002)
580. Shusterman, V., Troy, W.C.: From baseline to epileptiform activity: a path to synchronized
rhythmicity in large-scale neural networks. Phys. Rev. E 77, 061911 (2008)
581. Silva, J.S.D., Dotti, C.G.: Breaking the neuronal sphere: regulation of the actin cytoskeleton
in neuritogenesis. Nat. Rev. Neurosci. 3, 694704 (2002)
582. Silva, F.H.L.D., Blanes, W., Kalitzin, S.N., Parra, J., Suffczynski, P., Velis, D.N.: Dynamical
diseases of brain systems: different routes to epileptic seizures. IEEE Trans. Biomed. Eng.
50, 540548 (2003)
583. Sincich, L.C., Blasdel, G.G.: Oriented axon projections in primary visual cortex of the monkey. J. Neurosci. 21, 44164426 (2001)
584. Sinha, S.R.: Basic mechanisms of sleep and epilespy. J. Clin. Neurophysiol. 28, 103110
(2011)
585. Sjostrom, P.J., Rancz, E.A., Roth, A., Hausser, M.: Dendritic excitability and synaptic
plasticity. Physiol. Rev. 88, 769840 (2008)
586. Skinner, F., Kopell, N., Marder, E.: Mechanisms for oscillation and frequency control in
reciprocally inhibitory model neural networks. J. Comput. Neurosci. 1, 6987 (1994)
428
References
587. Smaily, M.E., Hamel, F., Roques, L.: Homogenization and influence of fragmentation in a
biological invasion model. Discret. Contin. Dyn. Syst. Ser. A 25, 321342 (2009)
588. Smith, G.D.: Modeling the stochastic gating of ion channels. In: Fall, C., Marland, E.S.,
wagner, J.M., Tyson, J.J. (eds.) Computational Cell Biology, chap. 11. Springer, New York
(2002)
589. Smith, G.D., Cox, C.L., Sherman, S.M., Rinzel, J.: Fourier analysis of sinusoidally driven
thalamocortical relay neurons and a minimal integrate-and-burst model. J. Neurophysiol. 83,
588610 (2000)
590. Sneyd, J., Charles, A.C., Sanderson, M.J.: A model for the propagation of intercellular calcium waves. Amer. J. Physiol. cell physiol. 266, C293C302 (1994)
591. Sneyd, J., Wetton, B.T., Charles, A.C., Sanderson, M.J.: Intercellular calcium waves mediated by diffusion of inositol trisphosphate: a two-dimensional model. Am. J. Physiol. Cell
Physiol. 268, C1537C1545 (1995)
592. Sneyd, J., Dale, P.D., Duffy, A.: Traveling waves in buffered systems: applications to calcium
waves. SIAM J. Appl. Math. 58, 11781192 (1998)
593. Softky, W.R., Koch, C.: Cortical cell should spike regularly but do not. Neural Comput. 4,
643646 (1992)
594. Soltesz, I., Staley, K.: Computational Neuroscience in Epilepsy. Academic, San Diego (2008)
595. Somers, D.C., Nelson, S., Sur, M.: An emergent model of orientation selectivity in cat visual
cortical simple cells. J. Neurosci. 15, 54485465 (1995)
596. Somjen, G.G.: Mechanisms of spreading depression and hypoxic spreading depression-like
depolarization. Physiol. Rev. 81, 10651096 (2001)
597. Soula, H., Chow, C.C.: Stochastic dynamics of a finite-size spiking neural network. Neural
Comput. 19, 32623292 (2007)
598. Sporns, O.: Networks of the Brain. MIT Press, Cambridge (2010)
599. Staley, K.J., Dudek, F.E.: Interictal spikes and epileptogenesis. Epilepsy Currents 6, 199202
(2006)
600. Stam, C.J., Reijeneveld, J.C.: Graph theoretical analysis of complex networks in the brain.
Nonlin. Biomed. Phys. 1 (3), 119 (2007)
601. Steinberg, B., Wang, Y.Q., Huang, H., Miura, R.: Spatial buffering mechanism: Mathematical
model and computer simulations. Math. Biosci. Eng. 2, 675702 (2005)
602. Steriade, M., McCormick, D.A., Sejnowski, T.J.: Thalamocortical oscillations in the sleeping
and aroused brain. Science 262, 679685 (1993)
603. Steyn-Ross, M.L., Sleigh, D.A.S.R.J.W., Whiting, D.R.: Theoretical predictions for spatial
covariance of the electroencephalographic signal during the anesthetic-induced phase transition: Increased correlation length and emergence of spatial self-organization. Phys. Rev. E
68, 021902 (2003)
604. Stocker, M., Krause, M., Pedarzani, P.: An apamin-sensitive Ca2+ -activated K+ current in
hippocampal pyramidal neurons. Proc. Natl. Acad. Sci. USA 96, 46624667 (1999)
605. Strack, S., Choi, S., Lovinger, D.M., Colbran, R.J.: Translocation of autophosphorylated
calcium/calmodulin-dependent protein kinase II to the postsynaptic density. J. Biol. Chem.
272, 1346713470 (1997)
606. Stuart, G.J., Sakmann, B.: Active propagation of somatic action potentials into neocortical
pyramidal cell dendrites. Nature 367, 6972 (1994)
607. Stuart, G., Spruston, N., Hausser, M.: Dendrites. Oxford University Press, New York (1999)
608. Stuart, G., Spruston, N., Hausser, M. (eds.): Dendrites. Oxford University Press, Oxford
(2007)
609. Subramanian, K., Narang, A.: A mechanistic model for eukaryotic gradient sensing: spontaneous and induced phosphoinositide polarization. J. Theor. Biol. 231, 4967 (2004)
610. Sun, W., McConnell, E., Pare, J.F., Xu, Q., Chen, M., Peng, W., Lovatt, D., Han, X., Y.Smith,
Nedergaard, M.: Glutamate-dependent neuroglial calcium signaling differs between young
and adult brain. Science 339, 197200 (2013)
611. Sutherland, G.R., McNaughton, B.: Memory trace reactivation in hippocampal and neocortical neuronal ensembles. Curr. Opin. Neurobiol. 10, 180186 (2000)
References
429
612. Swillens, S., Dupont, G., Combettes, L., Champeil, P.: From calcium blips to calcium puffs:
theoretical analysis of the requirement for interchannel communication. Proc. Nat. Acad.
Sci. (USA) 96, 1375013755 (1999)
613. Swindale, N.V.: The development of topography in the visualcortex: A review of models.
Network 7, 161274 (1996)
614. Syed, M.M., Lee, S., He, S., Zhou, Z.J.: Spontaneous waves in the ventricular zone of developing mammalian retina. J. Neurophysiol. 91, 19992009 (2004)
615. Syed, M.M., Lee, S., Zheng, J., Zhou, Z.J.: Stage-dependent dynamics and modulation of
spontaneous waves in the developing rabbit retina. J. Physiol. 560, 533549 (2004)
616. Sykova, E., Nicholson, C.: Diffusion in brain extracellular space. Physiol. Rev. 88,
12771340 (2008)
617. Tabak, J., Senn, W., ODonovan, M.J., Rinzel, J.: Modeling of spontaneous activity in developing spinal cord using activity-dependent depression in an excitatory network. J. Neurosci.
20, 30413056 (2000)
618. Tauber, U.C.: Field-theory approaches to nonequilibrium dynamics. Lect. Notes Phys. 716,
295348 (2007)
619. Taylor, A., Cottrell, G., Kristan Jr, W.: Analysis of oscillations in a reciprocally inhibitory
network with synaptic depression. Neural comput. 14(3), 561581 (2002)
620. Terman, D.H., Ermentrout, G.B., Yew, A.C.: Propagating activity patterns in thalamic neuronal networks. SIAM J. Appl. Math 61, 15781604 (2003)
621. Thul, R., Coombes, S., Smith, G.D.: Sensitisation waves in a bidomain fire-diffuse-fire model
of intracellular Ca2+ dynamics. Physica D 238, 21422152 (2009)
622. Thul, R., Smith, G.D., Coombes, S.: A bidomain threshold model of propagating calcium
waves. J. Math. Biol. 56, 435463 (2008)
623. Tian, G.F., , Hooman, A., Takano, T., Xu, Q., Peng, W., Lin, J., Oberheim, N., Lou, N.,
Zielke, R., Kang, J., Nedergaard, M.: An astrocytic basis of epilepsy. Nat. Med. 11, 973981
(2005)
624. Timofeeva, Y.: Traveling waves in a model of quasi-active dendrites with active spines. Physica D 239, 494503 (2010)
625. Timofeeva, Y., Lord, G.J., Coombes, S.: Spatio-temporal filtering properies of a dendritic
cable with active spines: a modeling study in the spike-diffuse-spike framework. J. Comput.
Neurosci. 21, 293306 (2006)
626. Tong, F., Engel, S.A.: Interocular rivalry revealed in the human cortical blind-spot representation. Nature 411, 195199 (2001)
627. Tong, F., Meng, M., Blake, R.: Neural bases of binocular rivalry. Trend. Cog. Sci. 10,
502511 (2006)
628. Toriyama, M., Shimada, T., Kim, K.B., Mitsuba, M., Nomura, E., Katsuta, K., Sakumura, Y.,
Roepstorff, P., Inagaki, N.: Shootin1: a protein involved in the organization of an asymmetric
signal for neuronal polarization. J. Cell Biol. 175, 147157 (2006)
629. Toriyama, M., Sakumura, Y., Shimada, T., Ishii, S., Inagaki, N.: A diffusion-based neurite
length-sensing mechanism involved in neuronal symmetry breaking. Mol. Syst. Biol. 6,
116 (2010)
630. Torquato, S.: Random Heterogeneous Materials. Springer, New York (2002)
631. Touboul, J., Destexhe, A.: Can power-law scaling and neuronal avalanches arise from
stochastic dynamics? PLoS One 5, 14 (2010)
632. Touboul, J., Hermann, G., Faugeras, O.: Noiseinduced behaviors in neural mean field dynamics. SIAM J. Appl. Dyn. Sys. 11, 4981 (2012)
633. Tritsch, N.X., Yi, E., Gale, J.E., Glowatzki, E., Bergles, D.E.: The origin of spontaneous
activity in the developing auditory system. Nature 450, 5055 (2007)
634. Troy, W.C., Shusterman, V.: Patterns and features of families of traveling waves in large-scale
neuronal networks. SIAM J. Appl. Dyn. Syst. 6, 263292 (2007)
635. Tsai, J.C., Sneyd, J.: Are buffers boring? uniqueness and asymptotical sttability of traveling
wave fronts in the buffered bistable equation. J. Math. Biol. 54, 513553 (2007)
636. Tsai, J.C., Sneyd, J.: Traveling waves in the buffered FitzHugh-Nagumo model. SIAM J.
Appl. Math. 71, 16061636 (2011)
430
References
637. Tsodyks, M., Pawelzik, K., Markram, H.: Neural networks with dynamic synapses. Neural
Comput. 10, 821835 (1998)
638. Tuckwell, H.C., Miura, R.M.: A mathematical model for spreading cortical depression. Biophys. J. 23, 257276 (1978)
639. Vecchia, D., Pietrobon, D.: Migrain: a disorder of brain excitatory-inhibitory balance? Trend.
Neurosci. 35, 507520 (2012)
640. Veen, M.V., Pelt, J.V.: Neuritic growth rate described by modelling microtubule dynamics.
Bull. Math. Biol. 56, 249273 (1994)
641. Vellela, M., Qian, H.: Stochastic dynamics and non-equilibrium thermodynamics of a
bistable chemical system: the Schlogl model revisited. J. R. Soc. Interface 6, 925940 (2009)
642. Veltz, R., Faugeras, O.: Local/global analysis of the stationary solutions
of some neural field equations.
SIAM J. Appl. Dyn. Syst. 9, 954998
(2010)
643. Vicker, M.G.: Eukaryotic cell locomotion depends on the propagation of self-organized
reaction-diffusion waves and oscillations of actin filament assembly. J. Expt. Cell Res. 275,
5466 (2002)
644. Vladimirski, B.B., Tabak, J., ODonovan, M.J., Rinzel, J.: Episodic activity in a heterogeneous excitatory network, from spiking neurons to mean field. J. Comp. Neurosci. 25, 3963
(2008)
645. Vogels, T.P., Abbott, L.F.: Signal propagation and logic gating in networks of integrate-andfire neurons. J. Neurosci. 25, 1078610795 (2005)
646. Volman, V., Sejnowski, T.J., Bazhenov, M.: Topological basis of epileptogenesis in a model
of severe cortical trauma. J. Neurophysiol. 106, 19331942 (2011)
647. Volman, V., Bazhenov, M., Sejnowski, T.J.: Computational models of neuron-astrocyte interaction in epilepsy. Front. Comp. Neurosci. 6(58), 110 (2012)
648. Volpert, V., Petrovskii, S.: Reaction-diffusion waves in biology. Phys. Life Rev. 6, 267310
(2009)
649. van Vreeswijk, C.: Partial synchronization in populations of pulse-coupled oscillators. Phys.
Rev. E. 54(5), 55225537 (1996)
650. van Vreeswijk, C., Sompolinsky, H.: Chaotic balanced state in a model of cortical circuits. .
Neural Comput. 10, 13211371 (1998)
651. van Vreeswijk, C., Ermentrout, G.B., Abbott, L.F.: When inhibition not excitation synchronizes neural firing. J. Comput. Neurosci. 1, 313321 (1994)
652. Vuyst, E.D., Wang, N., Decrock, E., Bock, M.D., Vinken, M., Moorhem, M.V., Lai, C., Culot,
M., Rogiers, V., Cecchelli, R., Naus, C.C., Evans, W.H., Leybaert, L.: Ca2+ regulation of
connexin 43 hemichannels in C6 glioma and glial cells. Cell Calcium 46, 176187 (2009)
653. Wagner, J., Sneyd, J.: Effects of rapid buffers on Ca2+ diffusion and Ca2+ oscillations. Biophys. J. 67, 447456 (1994)
654. Wainrib, G., Thieullen, M., Pakdaman, K.: Reduction of stochastic conductance-based neuron models with time-scales separation. J. Comput. Neurosci. 32, 32746 (2012)
655. Walther, G.R., Maree, A.F., Edelstein-Keshet, L., Grieneisen, V.A.: Deterministic versus
stochastic cell polarisation through wave-pinning. Bull. Math. Biol. 74, 25702599 (2012)
656. Wang, X.J.: Decision making in recurrent neuronal networks. Neuron 60, 215234 (2008)
657. Wang, X.J., Rinzel, J.: Alternating and synchronous rhythms in reciprocally inhibitory model
neurons. Neural Comput. 4, 8497 (1992)
658. Wang, S.S., Denk, W., Hausser, M.: Coincidence detection in single dendritic spines
mediated by calcium release. Nat. Neurosci. 3, 12661273 (2000)
659. Wang, Z., Haydon, P.G., Yeung, E.S.: Direct observation of calcium-independent intercellular ATP signaling in astrocytes. . Anal. Chem. 72, 20012007 (2000)
660. Ward, M.J., Lee, J.: On the asymptotic and numerical-analyses of exponentially
ill-conditioned singularly perturbed boundary value problems. Stud. Appl. Math. 94,
271326 (1995)
661. Watt, A.J.: Traveling waves in developing cerebellar cortex mediated by asymmetrical purkinje cell connectivity. Nat. Neurosci. 12, 463473 (2009)
References
431
662. Watts, D.J., Strogatz, S.H.: Collective dynamics of small-world networks. Nature 393,
440442 (1998)
663. Webber, M., Bressloff, P.C.: The effects of noise on binocular rivalry waves: a stochastic
neural field model. J. Stat. Mech 3, P03001 (2013)
664. Weinberger, H.F.: On spreading speeds and traveling waves for growth and migration in a
periodic habitat. J. Math. Biol. 45, 511548 (2002)
665. Weiner, O.D., Marganski, W.A., Wu, L.F., Altschuler, S.J., Kirschner, M.W.: An actin-based
wave generator organizes cell motility. PLoS Biol. 5, 20532063 (2007)
666. Weissman, T.A., Riquelme, P.A., Ivic, L., Flint, A.C., Kriegstein, A.R.: Calcium waves propagate through radial glial cells and modulate proliferation in the developing neocortex. Neuron 43, 647661 (2004)
667. Werry, E.L., Liu, G.J., Bennett., M.R.: Glutamate-stimulated ATP release from spinal cord
astrocytes is potentiated by substance P. J. Neurochem. 99, 924936 (2006)
668. Wetherington, J., Serranoa, G., Dingledine, R.: Astrocytes in the epileptic brain. Neuron 58,
168178 (2008)
669. White, J., Southgate, E., Thomson, J., Brenner, S.: The structure of the nervous system of the
nematode Caenorhabditis elegans. Phil. Trans. R. Soc. Lond. B Biol. Sci. 314, 1340 (1986)
670. White, J.A., Rubinstein, J.T., Kay, A.R.: Channel noise in neurons. Trends. Neurosci. 23,
131137 (2000)
671. Wiley, D.A., Strogatz, S.H., Girvan, M.: The size of the sync basin. Chaos 16, 015103 (2006)
672. Williams, G.S.B., Huertas, M.A., Sobie, E.A., Jafri, M.S., Smith, G.D.: A probability density
approach to modeling local control of calcium-induced calcium release in cardiac myocytes.
Biophys. J. 92, 23112328 (2007)
673. Williams, G.S.B., Huertas, M.A., Sobie, E.A., Jafri, M.S., Smith, G.D.: Moment closure
for local control models of calcium-induced calcium release in cardiac myocytes. Biophys.
J. 95, 16891703 (2008)
674. Wilson, H.R.: Computational evidence for a rivalry hierarchy in vision. Proc. Natl. Acad Sci.
USA 100, 14499144503 (2003)
675. Wilson, H.R., Cowan, J.D.: Excitatory and inhibitory interactions in localized populations of
model neurons. Biophys. J. 12, 123 (1972)
676. Wilson, H.R., Cowan, J.D.: A mathematical theory of the functional dynamics of cortical and
thalamic nervous tissue. Kybernetik 13, 5580 (1973)
677. Wilson, M.A., McNaughton, B.L.: Reactivation of hippocampal ensemble memories during
sleep. Science 265, 676679 (1994)
678. Wilson, H.R., Blake, R., Lee, S.H.: Dynamics of traveling waves in visual perception. Nature
412, 907910 (2001)
679. Winfree, A.: The Geometry of Biological Time. Springer, New York (1980)
680. Wissner-Gross, Z.D., Scott, M.A., Steinmeyer, J.D., Yanik, M.F.: Synchronous symmetry
breaking in neurons with different neurite counts. PLoS One 8 (2), e54905 (2013)
681. Wolszon, L.R., Rehder, V., Kater, S.B., Macagno, E.R.: Calcium wave fronts that cross gap
junctions may signal neuronal death during development. J. Neurosci. 14, 34373448 (1994)
682. Wong, R.O.L.: Retinal waves and visual development. Ann. Rev. Neurosci. 22, 2947 (1999)
683. Wu, J.: Propagating waves of activity in the neocortex: what they are, what they do. Neuroscientist 14, 487502 (2008)
684. Wu, J.Y., Guan, L., Tsau, Y.: Propagating activation during oscillations and evoked responses
in neocortical slices. J. Neurosci. 19, 50055015 (1999)
685. Wu, J., Guan, L., Bai, L., Yang, Q.: Spatiotemporal properties of an evoked population
activity in rat sensory cortical slices. J. Neurophysiol. 86, 246174 (2001)
686. Wylie, J.J., Miura, R.M.: Traveling waves in coupled reaction-diffusion models with
degenerate sources. Phys. Rev. E 74, 021909 (2006)
687. Xin, J.: Front propagation in heterogeneous media. SIAM Rev. 42, 161230 (2000)
688. Xu, W., Huang, X., Takagaki, K., Wu, J.Y.: Compression and reflection of visually evoked
cortical waves. Neuron 55, 119129 (2007)
689. Yang, E., Schulman, H.: Structural examination of autoregulation of multifunctional
calcium/calmodulin-dependent protein kinase II. J. Biol. Chem. 274, 2619926208 (1999)
432
References
690. Yao, Y., Choi, J., Parker, I.: Quantal puff of intracellular Ca2+ evoked by inositol triphosphate
in Xenopus oocytes. J. Physiol. 482, 533553 (1995)
691. Yoshioka, T., Blasdel, G.G., Levitt, J.B., Lund, J.S.: Relation between patterns of intrinsic
lateral connectivity, ocular dominance, and cytochrome oxidasereactive regions in macaque
monkey striate cortex. Cereb. Cortex 6, 297310 (1996)
692. Young, G.W.D., Keizer, J.: A single pool IP3 -receptor model for agonist stimulated Ca2+
oscillations. Proc. Nat. Acad. Sci. (USA) 89, 98959899 (1992)
693. Yuste, R., Bonhoeffer, T.: Morphological changes in dendritic spines associated with LTP.
Ann. Rev. Neurosci. 24, 10711089 (2001)
694. Yuste, R., Denk, W.: Dendritic spines as a basic functional units of neuronal integration.
Nature 375, 682684 (1995)
695. Zahradnikova, A., Zahradnik, I.: A minimal gating model for the cardiac Ca2+ release channel. Biophys. J. 71, 29963012 (1996)
696. Zhang, L.: On the stability of traveling wave solutions in synaptically coupled neuronal networks. Diff. Int. Eq. 16, 513536 (2003)
697. Zimmermann, H.: Purinergic signaling in neural development. Sem. Cell Dev. Biol. 22,
194204 (2011)
698. Zinn-Justin, J.: Quantum Field Theory and Critical Phenomena, 4th edn. Oxford University
Press, Oxford (2002)
699. Zinner, B.: Existence of traveling wavefront solutions for the discrete Nagumo equation.
J. Diff. Eqns. 96, 127 (1992)
700. Zucker, R.S.: Short term synaptic plasticity. Ann. Rev. Neurosci. 12, 1331 (1989)
Index
D-serine, 171
action potential, 7, 65
active dendrites, 3236
adaptative threshold, 237
adenosine triphosphate (ATP), 171, 175, 351
alpha function, 18
Alzheimers disease, 399
ambiguous figures, 323
AMPA receptor, 18
amyloid fibrils, 400
astrocytes, 169173, 375
asynchronous state, 242
Averaging Theorem, 16
averaging theory
calcium release sites, 148150
discrete bistable equation, 8690
neural field equation, 290294
axon growth, 356370
axonal delays, 237, 268
axons
action potentials, 7
myelinated, 83
propagation delays, 221
balanced networks, 246
bidomain models, 143, 371
binocular rivalry, 323325
competitive network model, 325330
effects of noise, 334340
neural field model, 330334
traveling fronts, 330334
bistability, 44, 157, 257, 261, 278
branching processes, 395399
breathers, 317
buffering, 144146, 375
cable equation, 2429, 103

calcium
bidomain model, 143
buffering, 144146
calcium signaling in neurons, 137141
calcium-induced-calcium release, 137
De Young-Keizer model, 153
endoplasmic reticulum, 137
fire-diffuse-fire model, 150152
intercellular waves, 168177, 352
IP3 receptors, 137
Li-Rinzel model, 153
neural development, 350353
paracrine signaling, 175177, 352
reaction-diffusion models, 141146
ryanodine receptors, 137
sparks, 141, 152160
synaptic plasticity, 140
calcium T -current, 229
CaMKII, 101
CaMKII translocation waves, 113130
cardiac myocytes, 155
cell cycle, 352
cell polarization, 356370
cerebral cortex, 186
characteristic path length, 386
clustering coefficient, 386
conductance-based models, 39
connexins, 173
coupled hypercolumn model of V1, 343345
criticality, 392399
cytoskeleton, 356370
dendrites
active, 3236
cable equation, 2429, 103
dendritic filtering, 29
DOI 10.1007/978-1-4614-8866-8, Springer Science+Business Media New York 2014
433
434
Index
dendritic spines, 101

dendritic tree, 28
NMDA spikes, 32
quasilinear, 35
spike-diffuse-spike model, 103112
development, 350356
discrete bistable equation, 84
interfacial dynamics, 294

ion channels, 5, 3656, 153160
ionotropic receptor, 20
Ising model, 393
isochrones, 10
Ito vs. Stratonovich, 60
Itos formula, 60
E-I networks, 225227, 238, 266

electroencephalogram (EEG), 190
epidemics on networks, 403
epilepsy, 194, 383392
Euclidean shift-twist symmetry, 345
Euclidean symmetry, 341
Evans function, 8083, 274277, 285
excitable neuron, 6466
exponential dichotomies, 80
kinematic theory, 7677, 379383

Kramers-Moyal expansion, 253
Kuramoto model, 204, 208
fictive waves, 194

firing rate function, 9
Fisher-KPP equation, 115, 130, 277
FitzHugh-Nagumo equations, 6469
flip bifurcation, 390
fMRI, 324
Fokker-Planck equation, 38, 43, 45, 61
Fredholm Alternative, 85, 97
functional networks, 392
GABA inhibition, 19
gap junctions, 173, 374
generating functions, 395397
gliotransmission, 171
glutamate, 18, 169, 171
Greens function, 26, 30, 111
growth cone, 356370
half-center oscillator, 201
Hamilton-Jacobi equations, 126130,
162163, 262, 297299
hemichannels, 173
hippocampus, 192
Hodgkin-Huxley model, 5, 7
homeostatic synaptic plasticity, 392
homogenization theory, 177181
dendritic spines, 121126
spreading depression, 375378
Hopf bifurcation, 11, 218, 318
Hopfield network, 235
horizontal connections in V1, 321, 343
hypercolumn, 320
hysteresis, 210
integrate-and-fire model, 103, 212227
integrate-and-fire-or-burst model, 229231
Langevin equation, 39, 42, 252, 268

limbic system, 186
locomotion, 200204
long term potentiation, 19
lurching waves, 223
matched asymptotics, 52
mean first passage time, 4356, 158, 258, 261,
338
metabotropic receptor, 20
migraine aura, 378383
Morris-Lecar model, 14, 66
Multi-pulse wave, 293
multiplicative noise, 60, 91
myelinated axons, 83
Nernst equation, 4
network theory, 385388
characteristic path length, 386
clustering coefficient, 386
scale-free, 387
small-world, 387
neural fields
adaptive, 281289
and synaptic depression, 328334
construction, 264269
heterogeneous, 289303
oscillatory, 315318
stimulus-locking, 309313
stochastic, 264269, 303315
traveling fronts, 272281
traveling pulses, 281289
two-dimensional, 340345
neural master equation, 250264
neural variability, 246
neuritogenesis, 356370
neurodegenerative diseases, 399404
neuronal avalanches, 394
NMDA receptor, 18
nodes of Ranvier, 83
Novikovs theorem, 91
Index
nucleated polymerization model of prions
spreading, 401
nullclines, 65, 69
olfactory bulb, 189
optical imaging, 321
orientation preference map, 320
Ornstein-Uhlenbeck equation, 40, 310
phase coherence, 210
phase diagram, 112, 227, 239, 391
phase oscillators, 1017, 195212, 239242
phase resetting curves, 1014, 197, 213,
239242
phase-locking, 1417, 198200
phase-plane analysis, 6476
Poisson processes, 247
population-based network models, 233238
porous media, 375378
power laws, 394, 398
primary visual cortex, 319323
prions, 399404
protein aggregates, 400
pulled fronts, 130136
quasi-steady-state approximation, 47, 255
quasilinear membrane, 35
radial glial cells, 350353
reaction-diffusion models, 113116, 141146,
371373
retina, 352356
retinotopic map, 320, 382
Rho GTPase, 358
ring network, 204208
ripple waves, 192
saddle-node bifurcation, 227
scale-free networks, 387, 403
sharp waves, 192
shooting method, 71
SI model of infection, 115, 403
singular perturbation theory, 65, 7274, 287
sleep-wake cycle, 190
slow waves, 190, 227, 260
slow-fast analysis, 65, 287
small-world networks, 387
and epilepsy, 388391
spectral projection method, 48
spectral theory, 80, 97
spike response model, 225
spike times, 7
spindle waves, 190, 228
spiral waves, 318
435
spontaneous action potentials, 6769
spreading depression, 194, 370375
steepest descents, 132133, 299300
stochastic calculus, 5762
stochastic hybrid system, 42, 4656, 252
stochastic hybrid systems, 255
stochastic integrate-and-fire model, 242
stochastic ion channels
calcium release, 153160
diffusion approximation, 38
master equation, 37, 154, 157
voltage fluctuations, 4156
voltage-gated channels, 3656
stochastic Morris-Lecar equation, 6769
stochastic neural fields, 264269, 303315,
334340
stochastic phase-plane analysis, 165
stochastic population models, 246264
stochastic traveling waves, 9195, 303315,
334340
subdiffusion, 313
synapses
conductance, 17
kinetics, 2124
tripartite synapse, 170
synaptic depression, 20, 236, 315, 328334
synaptic plasticity, 19, 101, 140
sync basin, 204208
synchronization, 16, 204208, 213
thalamic reticular nucleus, 190, 228
thalamus, 190, 228
theta waves, 192
tripartite synapse, 170
Tuckwell-Miura model, 371373
two-dimensional waves, 208, 315
Type I and Type II phase oscillators, 13
up and down states, 190
variational method, 135
Volterra integral equation, 30
wave-pinning, 359
waves
actin, 364366
CaMKII translocation waves, 113130
cortical waves, 185195, 272
Evans function, 8083
in binocular rivalry, 330340
in fire-diffuse-fire model, 150152
in Fisher-KPP equation, 130136
in heterogeneous media, 119130, 289303
in hippocampus, 192
436
in integrate-and-fire networks, 219227
in neural development, 349356
in neural fields, 272318, 330340
in oscillator networks, 200212
in spike-diffuse-spike model, 103112
in the bistable equation, 6971
in the discrete bistable equation, 8490
in the FitzHugh-Nagumo equations, 7276
intercellular calcium waves, 168177, 352
intracellular calcium waves, 146152
kinematics, 7677, 379383
linear spreading velocity, 130
lurching waves, 223
migraine aura, 378383
periodic waves, 74
propagation failure, 8690, 146150
pulled fronts, 130136, 277281, 313
Index
pulsating waves, 119126, 296
retinal waves, 353356
saltatory waves, 112, 150
spiral waves, 318
spreading depression, 370375
stability, 7783, 107, 221223, 274277,
285
stimulus-locked fronts, 309313
stochastic traveling waves, 9195, 303315,
334340
swimming locomotion, 200204
thalamic waves, 228231
Wiener process, 58
Wilson-Cowan equations, 266
WKB method, 5152, 6769, 160165, 255,
261, 297

Waves in Neural Media

Uploaded by

Copyright:

Available Formats

Waves in Neural Media

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Waves in Neural Media

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What mathematical models are discussed in the book?

What mathematical models are discussed in the book?

Lecture Notes on Mathematical Modelling

in the Life Sciences

Lecture Notes on Mathematical Modelling

For further volumes:

Waves in Neural Media

This is a book on mathematical neuroscience, in which the unifying theme is

molecules to macroscopic descriptions of states of matter. The latter also supervene

Single Neuron Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Traveling Waves in One-Dimensional Excitable Media . . . . . . . . . . . . .

Wave Propagation Along Spiny Dendrites . . . . . . . . . . . . . . . . . . . . . . . . . 101

Calcium Waves and Sparks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

4.4.3 WKB Method and Quasistationary Approximation . . . . . . . . 160

Waves in Synaptically Coupled Spiking Networks . . . . . . . . . . . . . . . . . . 185

Population Models and Neural Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

Waves in Excitable Neural Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

7.3 Wave Propagation in Heterogeneous Neural Fields . . . . . . . . . . . . . . . 289

Neural Field Model of Binocular Rivalry Waves . . . . . . . . . . . . . . . . . . . 319

Part III Development and Disease

Waves in the Developing and the Diseased Brain . . . . . . . . . . . . . . . . . . . 349

9.5 Neuronal Avalanches, Criticality, and Branching Processes . . . . . . . . 392

Single Neuron Modeling

In this introductory chapter we describe conductance-based models of single

1.1 Conductance-Based Model of a Neuron

1 Single Neuron Modeling

where gs is the conductance due to channels of type s and Vs is the corresponding

1.1 Conductance-Based Model of a Neuron

terminating an action potential by repolarizing a neuron. One finds that opening of

1 Single Neuron Modeling

h = 0.07 exp[0.05(v + 65)],

1.1 Conductance-Based Model of a Neuron

Once a neuron has generated an action potential, it propagates as a traveling pulse

1 Single Neuron Modeling

1.1 Conductance-Based Model of a Neuron

forms of excitability as detailed elsewhere [201, 423].) If one includes stochastic

F(I) = (u ), < I < + 1 ,

1 Single Neuron Modeling

1.2 Periodically Forced Neural Oscillator

where p(x,t) = p(x,t + ) and = 2 / is the forcing frequency. If the amplitude

1.2.1 Isochrones and Phase-Resetting Curves

1.2 Periodically Forced Neural Oscillator

Fig. 1.5 Isochrones in the neighborhood of a stable limit cycle

is a 2 -periodic function of and a -periodic function of t.

1 Single Neuron Modeling

In polar coordinates A = Rei ,

The solution for arbitrary initial data R(0) = R0 , (0) = 0 is

(t) = 0 + 0t log(R20 + (1 R20)e2t ),

Hence, we can define a phase on the whole plane

It follows that the isochrones are logarithmic spirals with ln R = constant.

1.2 Periodically Forced Neural Oscillator

1 Single Neuron Modeling

(v) = cosh[(v v3 )/(2v4 )].

1.2.2 Phase-Locking and Synchronization

1.2 Periodically Forced Neural Oscillator

Q( ,t) = al,k eik +il t .

Substitute for using the zero-order approximation = 0t + 0 :

The phase equation then becomes