Unconventional Computation

Unconventional Computation
Bruce MacLennan
Department of Electrical Engineering and Computer Science
University of Tennessee, Knoxville
Copyright 2018
c
Version of
November 17, 2018

2
Contents
I Introduction 3
A What is unconventional computation? . . . . . . . . . . . . . . 3
B Post-Moore’s Law computing . . . . . . . . . . . . . . . . . . 5
C Super-Turing vs. non-Turing . . . . . . . . . . . . . . . . . . . 7
C.1 The limits of Turing computation . . . . . . . . . . . . 8
C.2 New computational models . . . . . . . . . . . . . . . . 11
C.3 Computation in general . . . . . . . . . . . . . . . . . 16
C.4 Expanding the range of computing technologies . . . . 22
C.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 26
II Physics of Computation 29
A Energy dissipation . . . . . . . . . . . . . . . . . . . . . . . . 29
B Thermodynamics of computation . . . . . . . . . . . . . . . . 35
B.1 Von Neumann-Landaur Principle . . . . . . . . . . . . 35
B.2 Mechanical and thermal modes . . . . . . . . . . . . . 43
C Reversible computing . . . . . . . . . . . . . . . . . . . . . . . 47
C.1 Reversible computing as a solution . . . . . . . . . . . 47
C.2 Foundations of Conservative Computation . . . . . . . 51
C.3 Conservative logic circuits . . . . . . . . . . . . . . . . 54
C.4 Universality . . . . . . . . . . . . . . . . . . . . . . . . 55
C.5 Constants and garbage . . . . . . . . . . . . . . . . . . 57
C.6 Garbageless conservative logic . . . . . . . . . . . . . . 57
C.7 Ballistic computation . . . . . . . . . . . . . . . . . . . 60
D Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
III Quantum Computation 67

A Mathematical preliminaries . . . . . . . . . . . . . . . . . . . 67
A.1 Complex numbers . . . . . . . . . . . . . . . . . . . . . 67
3
4 CONTENTS
A.2 Linear algebra review . . . . . . . . . . . . . . . . . . . 67

B Basic concepts from quantum theory . . . . . . . . . . . . . . 77
B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 77
B.2 Postulates of QM . . . . . . . . . . . . . . . . . . . . . 78
B.3 Wave-particle duality (supplementary) . . . . . . . . . 85
B.4 Superposition . . . . . . . . . . . . . . . . . . . . . . . 86
B.5 No-cloning theorem . . . . . . . . . . . . . . . . . . . . 89
B.6 Entanglement . . . . . . . . . . . . . . . . . . . . . . . 89
B.7 Uncertainty principle (supplementary) . . . . . . . . . 93
C Quantum information . . . . . . . . . . . . . . . . . . . . . . . 99
C.1 Qubits . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
C.2 Quantum gates . . . . . . . . . . . . . . . . . . . . . . 104
C.3 Quantum circuits . . . . . . . . . . . . . . . . . . . . . 110
C.4 Quantum gate arrays . . . . . . . . . . . . . . . . . . . 112
C.5 Quantum parallelism . . . . . . . . . . . . . . . . . . . 115
C.6 Applications . . . . . . . . . . . . . . . . . . . . . . . . 117
C.7 Universal quantum gates . . . . . . . . . . . . . . . . . 122
D Quantum algorithms . . . . . . . . . . . . . . . . . . . . . . . 124
D.1 Deutsch-Jozsa algorithm . . . . . . . . . . . . . . . . . 124
D.2 Simon’s algorithm . . . . . . . . . . . . . . . . . . . . . 131
D.3 Shor’s algorithm . . . . . . . . . . . . . . . . . . . . . 135
D.4 Search problems . . . . . . . . . . . . . . . . . . . . . . 146
D.5 Quantum error correction . . . . . . . . . . . . . . . . 158
E Abrams-Lloyd theorem . . . . . . . . . . . . . . . . . . . . . . 166
E.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . 166
E.2 Basic algorithm . . . . . . . . . . . . . . . . . . . . . . 166
E.3 Discussion of further results . . . . . . . . . . . . . . . 169
F Universal quantum computers . . . . . . . . . . . . . . . . . . 170
F.1 Feynman on quantum computation . . . . . . . . . . . 170
F.2 Benioff’s quantum Turing machine . . . . . . . . . . . 176
F.3 Deutsch’s universal quantum computer . . . . . . . . . 176
G Quantum probability in cognition . . . . . . . . . . . . . . . . 178
G.1 Theories of decision making . . . . . . . . . . . . . . . 178
G.2 Framework . . . . . . . . . . . . . . . . . . . . . . . . 179
G.3 Experimental evidence . . . . . . . . . . . . . . . . . . 184
G.4 Cognition in Hilbert space . . . . . . . . . . . . . . . . 188
G.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 191
H Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
CONTENTS 5
IV Molecular Computation 199

A Basic concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
A.1 Introduction to molecular computation . . . . . . . . . 199
A.2 DNA basics . . . . . . . . . . . . . . . . . . . . . . . . 199
A.3 DNA manipulation . . . . . . . . . . . . . . . . . . . . 204
B Filtering models . . . . . . . . . . . . . . . . . . . . . . . . . . 208
B.1 Adleman: HPP . . . . . . . . . . . . . . . . . . . . . . 208
B.2 Lipton: SAT . . . . . . . . . . . . . . . . . . . . . . . . 213
B.3 Test tube programming language . . . . . . . . . . . . 217
B.4 Parallel filtering model . . . . . . . . . . . . . . . . . . 220
C Formal models . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
C.1 Sticker systems . . . . . . . . . . . . . . . . . . . . . . 223
C.2 Splicing systems . . . . . . . . . . . . . . . . . . . . . . 226
D Enzymatic computation . . . . . . . . . . . . . . . . . . . . . 227
V Analog Computation 231

A Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
B Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
B.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
B.2 Chapter roadmap . . . . . . . . . . . . . . . . . . . . . 251
C Fundamentals of analog computing . . . . . . . . . . . . . . . 252
C.1 Continuous state space . . . . . . . . . . . . . . . . . . 252
C.2 Computational process . . . . . . . . . . . . . . . . . . 254
C.3 Analog computer programs . . . . . . . . . . . . . . . . 261
C.4 Characteristics of analog computation . . . . . . . . . . 262
D Analog Computation in Nature . . . . . . . . . . . . . . . . . 266
D.1 Neural computation . . . . . . . . . . . . . . . . . . . . 267
D.2 Adaptive self-organization in social insects . . . . . . . 268
D.3 Genetic circuits . . . . . . . . . . . . . . . . . . . . . . 269
D.4 Is everything a computer? . . . . . . . . . . . . . . . . 270
E General-purpose analog computation . . . . . . . . . . . . . . 271
E.1 The importance of general-purpose computers . . . . . 271
E.2 General-purpose electronic analog computers . . . . . . 271
E.3 Shannon’s analysis . . . . . . . . . . . . . . . . . . . . 272
E.4 Rubel’s Extended Analog Computer . . . . . . . . . . 273
F Analog computation and the Turing limit . . . . . . . . . . . . 274
F.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 274
F.2 A sampling of theoretical results . . . . . . . . . . . . . 275
CONTENTS 1
F.3 Real-valued inputs, output, and constants . . . . . . . 278

F.4 The issue of simulation by Turing machines and digital
computers . . . . . . . . . . . . . . . . . . . . . . . . . 279
F.5 The problem of models of computation . . . . . . . . . 280
F.6 Relevant issues for analog computation . . . . . . . . . 281
F.7 Transcending Turing computability . . . . . . . . . . . 282
G Analog thinking . . . . . . . . . . . . . . . . . . . . . . . . . . 283
H Future directions . . . . . . . . . . . . . . . . . . . . . . . . . 284
H.1 Post-Moore’s Law computing . . . . . . . . . . . . . . 284
2 CONTENTS
Chapter I
Introduction
These lecture notes are exclusively for the use of students in Prof. MacLen-
nan’s Unconventional Computation course. 2018,
c B. J. MacLennan, EECS,
University of Tennessee, Knoxville. Version of November 17, 2018.
A What is unconventional computation?

Unconventional computation and its synonym, nonstandard computation, are
negative terms; they refer to computation that is not “conventional” or “stan-
dard.” Conventional computation is the kind we are all familiar with, which
has dominated computer science since the 1950s. We can enumerate its com-
mon characteristics:
• Information representation and processing is digital and in fact binary.
• Computers are organized according to the von Neumann architecture,
which separates memory from a processor that fetches data from mem-
ory for processing and then returns it to memory.
• Memory is organized as one or more arrays of units (bytes or words)
indexed by natural numbers.
• The processor executes instructions from memory, sequentially, in dis-
crete steps, and one at a time; therefore instructions are encoded in
binary.
• The sequential binary logic of the processor is implemented by means
of electronically-controlled electronic switches.
3
4 CHAPTER I. INTRODUCTION
• Programs are written as hierarchically organized sequences of instruc-

tions, which perform computation, input, output, and control of pro-
gram execution.
• Programs can decide between alternative sequences of instructions.
• Programs can repeat sequences of instructions until some condition is
satisfied.
• Programs can be hierarchically organized into subprograms, subpro-
grams of subprograms, etc.
Unconventional computers, then, are different in at least one of these charac-
teristics. For example, analog computers represent information by continu-
ous physical quantities rather than bits, and some analog computers operate
in continuous time (obeying differential equations) rather than in discrete
sequential steps. Nonelectronic computers may use light, fluids, microorgan-
isms, or DNA and other molecules as a computational medium. In molecular
computers, vast numbers of operations take place asynchronously in parallel.
Ballistic computers implement reversible computation, so they may dissipate
arbitrarily small amounts of energy, but cannot use common programming
operations, such as storing into a memory. Functional programming does not
use rewritable variables, and computers designed for functional programming,
such as reduction machines, do not have a programmer-accessible storable
memory. Quantum computers do not do one thing at a time, but execute an
exponentially large number of operations simultaneously in quantum super-
position. Alternative computational paths can be executed in superposition
and otherwise iterative processes can be executed all at once. Neuromor-
phic computers, which are inspired by the brain, don’t have an addressible
memory in the usual sense (they learn by analog changes in vast numbers
of connection strengths, so that individual memories are distributed over
large overlapping sets of connections). They are also massively parallel and
use analog computation. Programs for unconventional computers can take
unconventional forms, such as recursive function definitions, ordinary or par-
tial differential equations, quantum operations, neural network architectures,
sequences of chemical reactions, genetic regulatory networks, or even contin-
uous images. These are just a few of the ways in which computation can be
“unconventional” or “nonstandard.” You may be asking, however, why we
would want to compute in these unconventional ways, given the outstanding
success of conventional computation. That is our next topic.
Differences in Spatial Scale
B. POST-MOORE’S LAW COMPUTING 5
2.71828
0010111001100010100
… …
Fall 2012 Unconventional Computation 7
Figure I.1: Hierarchy of spatial scales in conventional computing.
(Images from Wikipedia)‫‏‬
B Post-Moore’s Law computing

Although estimates differ, it is clear that the end of Moore’s Law is in sight;
there are physical limits to the density of binary logic devices and to their
speed of operation.1 This will require us to approach computation in new
ways, which will present significant challenges, but can also broaden and
deepen our concept of computation in natural and artificial systems.
In the past there has been a significant difference in scales between com-
putational processes and the physical processes by which they are realized.
For example, there are differences in spatial scale: the data with which pro-
grams operate (integers, floating point numbers, characters, pointers, etc.)
are represented by large numbers of physical devices comprising even larger
numbers of physical particles (Fig. I.1). Also, there are differences in time
scale: elementary computational operations (arithmetic, instruction sequenc-
ing, memory access, etc.), are the result of large numbers of state changes
at the device level (typically involving a device moving from one saturated
state to another) (Fig. I.2). However, increasing the density and speed of
1
This section is adapted from MacLennan (2008). Moore’s Law was presented in Moore
(1965).
X := Y / Z
P[0] := N
i := 0
while i < n do
if P[i] >= 0 then
q[n-(i+1)] := 1
P[i+1] := 2*P[i] - D
else
q[n-(i+1)] := -1
P[i+1] := 2*P[i] + D
end if
i := i + 1
end while
Fall 2012 Unconventional Computation 8

(Images from Wikipedia)‫‏‬
Figure I.2: Hierarchy of temporal scales in conventional computing.
C. SUPER-TURING VS. NON-TURING 7
computation will force it to take place on the scale (spatial and temporal)
near that of the underlying physical processes. With fewer hierarchical levels
between computations and their physical realizations, and less time for im-
plementing computational processes, computation will have to become more
like the underlying physical processes. That is, post-Moore’s Law computing
will depend on a greater assimilation of computation to physics.
In discussing the role of physical embodiment in the “grand challenge” of
unconventional computing, Stepney (2004, p 29) writes,
Computation is physical; it is necessarily embodied in a device
whose behaviour is guided by the laws of physics and cannot be
completely captured by a closed mathematical model. This fact
of embodiment is becoming ever more apparent as we push the
bounds of those physical laws.
Traditionally, a sort of Cartesian dualism has reigned in computer science;
programs and algorithms have been conceived as idealized mathematical ob-
jects; software has been developed, explained, and analyzed independently
of hardware; the focus has been on the formal rather than the material.
Post-Moore’s Law computing, in contrast, because of its greater assimilation
to physics, will be less idealized, less independent of its physical realization.
On one hand, this will increase the difficulty of programming since it will
be dependent on (or, some might say, contaminated by) physical concerns.
On the other hand, as will be explored in this book, it also presents many
opportunities that will contribute to our understanding and application of
information processing in the future.
C Super-Turing vs. non-Turing

In addition to the practical issues of producing more powerful computers, un-
conventional computation poses the question of its theoretical power.2 Might
some forms of unconventional computation provide a means of super-Turing
computation (also called hypercomputation), that is, of computation beyond
what is possible on a Turing machine? This is an important theoretical ques-
tion, but obsession with it can divert attention from more significant issues
in unconventional computing. Therefore it is worth addressing it before we
get into specific unconventional computing paradigms.
2
This section is adapted from MacLennan (2009).
C.1 The limits of Turing computation
C.1.a Frames of Relevance
It is important to remember that Church-Turing (CT) computation is a model

of computation, and that computation is a physical process taking place
in certain physical objects (such as computers). Models are intended to
help us understand some class of phenomena, and they accomplish this by
making simplifying assumptions (typically idealizing assumptions, which omit
physical details taken to be of secondary importance). For example, we might
use a linear mathematical model of a physical process even though we know
that its dynamics is only approximately linear; or a fluid might be modeled
as infinitely divisible, although we know it is composed of discrete molecules.
We are familiar also with the fact that several models may be used for a
single system, each model suited to understanding certain aspects of the
system but not others. For example, a circuit diagram shows the electrical
interconnections among components (qua electrical devices), and a layout
diagram shows the components’ sizes, shapes, spatial relationships, etc.
As a consequence of its simplifying assumptions, each model comes with
a (generally unstated) frame of relevance, which delimits (often vaguely) the
questions that the model can answer accurately. For example, it would be a
mistake to draw conclusions from a circuit diagram about the size, shape, or
physical placement of circuit components. Conversely, little can be inferred
about the electrical properties of a circuit from a layout diagram.
Within a (useful) model’s frame of relevance, its simplifying assumptions
are sensible (e.g., they are good approximations); outside of it they may not
be. That is, within its frame of relevance a model will give us good answers
(not necessarily 100% correct) and help us to understand the characteristics
of the system that are most relevant in that frame. Outside of its intended
frame, a model might give good answers (showing that its actual frame can
be larger than its intended frame), but we cannot assume that to be so. Out-
side of its frame, the answers provided by a model may reflect the simplifying
assumptions of the model more than the system being modeled. For exam-
ple, in the frame of relevance of macroscopic volumes, fluids are commonly
modeled as infinitely divisible continua (an idealizing assumption), but if we
apply such a model to microscopic (i.e., molecular scale) volumes, we will get
misleading answers, which are a consequence of the simplifying assumptions.
C.1.b The frame of relevance of Church-Turing computation
It is important to explicate the frame of relevance of Church-Turing (CT)

computation, by which I mean not just Turing machines, but also equivalent
models of computation, such as the lambda calculus and Post productions,
as well as other more or less powerful models based on similar assumptions
(discussed below). (Note however that the familiar notions of equivalence and
power are themselves dependent on the frame of relevance of these models, as
will be discussed.) The CT frame of relevance becomes apparent if we recall
the original questions the model was intended to answer, namely questions
of effective calculability and formal derivability. As is well known, the CT
model arises from an idealized description of what a mathematician could do
with pencil and paper. Although a full analysis of the CT frame of relevance
is beyond the scope of this chapter (but see MacLennan, 1994b, 2003, 2004),
I will mention a few of the idealizing assumptions.
Within the CT frame of relevance, something is computable if it can be
computed with finite but unbounded resources (e.g., time, memory). This
is a reasonable idealizing assumption for answering questions about formal
derivability, since we don’t want our notion of a proof to be limited in length
or “width” (size of the formal propositions). It is also a reasonable simplifying
assumption for investigating the limits of effective calculability, which is a
idealized model of arithmetic with paper and pencil. Again, in the context
of the formalist research program in mathematics, there was no reason to
place an a priori limit on the number of steps or the amount of paper (or
pencil lead!) required. Note that these are idealizing assumptions: so far
as we know, physical resources are not unbounded, but these bounds were
not considered relevant to the questions that the CT model was originally
intended to address; in this frame of relevance “finite but unbounded” is a
good idealization of “too large to be worth worrying about.”
Both formal derivation and effective calculation make use of finite for-
mulas composed of discrete tokens, of a finite number of types, arranged
in definite structures (e.g., strings) built up according to a finite number of
primitive structural relationships (e.g., left-right adjacency). It is further
assumed that the types of the tokens are positively determinable, as are the
primitive interrelationships among them. Thus, in particular, we assume that
there is no uncertainty in determining whether a token is present, whether a
configuration is one token or more than one, what is a token’s type, or how
the tokens are arranged, and we assume that they can be rearranged with
perfect accuracy according to the rules of derivation. These are reasonable

assumptions in the study of formal mathematics and effective calculability,
but it is important to realize that they are idealizing assumptions, for even
mathematicians can make mistakes in reading and copying formulas and in
applying formal rules!
Many of these assumptions are captured by the idea of a calculus, but
a phenomenological analysis of this concept is necessary to reveal its back-
ground of assumptions (MacLennan, 1994b). Briefly, we may state that both
information representation and information processing are assumed to be
formal, finite, and definite (MacLennan, 2003, 2004). These and other as-
sumptions are taken for granted by the CT model because they are reasonable
in the context of formal mathematics and effective calculability. Although
the originators of the model discussed some of these simplifying assumptions
(e.g., Markov, 1961), many people today do not think of them as assumptions
at all, or consider that they might not be appropriate in some other frames
of relevance.
It is important to mention the concept of time presupposed in the CT
model, for it is not discrete time in the familiar sense in which each unit of
time has the same duration; it is more accurate to call it sequential time (van
Gelder, 1997). This is because the CT model does not take into consideration
the time required by an individual step in a derivation or calculation, so long
as it is finite. Therefore, while we can count the number of steps, we cannot
translate that count into real time, since the individual steps have no definite
duration. As a consequence, the only reasonable way to compare the time
required by computational processes is in terms of their asymptotic behavior.
Again, sequential time is reasonable in a model of formal derivability or
effective calculability, since the time required for individual operations was
not relevant to the research programme of formalist mathematics (that is, the
timing was irrelevant in that frame of relevance), but it can be very relevant
in other contexts, as will be discussed.
Finally I will mention a simplifying assumption of the CT model that is
especially relevant to hypercomputation, namely, the assumption that com-
putation is equivalent to evaluating a well-defined function on an argument.
Certainly, the mathematical function, in the full generality of its definition,
is a powerful and versatile mathematical concept. Almost any mathemat-
ical object can be treated as a function, and functions are essential to the
description of processes and change in the physical sciences. Therefore, it
was natural, in the context of the formalist program, to focus on functions in
the investigation of effective calculation and derivation. Furthermore, many

early applications of computers amounted to function evaluations: you put
in a deck of punched cards or mounted a paper or magnetic tape, started
the program, it computed for awhile, and when it stopped you had an out-
put in the form of cards, tape, or paper. Input — compute — output, that
was all there was to it. If you ran the program again with a different input,
that amounted to an independent function evaluation. The only relevant as-
pect of a program’s behavior was the input-output correspondence (i.e., the
mathematical function).
This view can be contrasted with newer ones in which, for example, a
computation involves continuous, non-terminating interaction with its envi-
ronment, such as might be found in control systems and autonomous robotics.
Some new models of computation have moved away from the idea of compu-
tation as the evaluation of a fixed function (Eberbach et al., 2003; Milner,
1993; Milner et al., 1992; Wegner, 1997, 1998; Wegner & Goldin, 2003). In
the CT frame of relevance, however, the natural way to compare the “power”
of models of computation was in terms of the classes of functions they could
compute, a single dimension of power now generalized into a partial order of
set inclusions (but still based on a single conception of power: computing a
class of functions).3
C.2 New computational models

A reasonable position, which many people take explicitly or implicitly, is
that the CT model is a perfectly adequate model of everything we mean
by “computation,” and therefore that any answers that it affords us are
definitive. However, as we have seen, the CT model exists in a frame of
relevance, which delimits the kinds of questions that it can answer accurately,
and, as I will show, there are important computational questions that fall
outside this frame of relevance.
C.2.a Natural computation

Natural computation may be defined as computation occurring in nature or
inspired by computation in nature. The information processing and control
3
I note in passing that this approach raises all sorts of knotty cardinality questions,
which are inevitable when we deal with such “large” classes; therefore in some cases results
depend on a particular axiomatization or philosophy of mathematics.
that occurs in the brain is perhaps the most familiar example of computa-
tion in nature, but there are many others, such as the distributed and self-
organized computation by which social insects solve complicated optimiza-
tion problems and construct sophisticated, highly structured nests. Also,
the DNA of multicellular organisms defines a developmental program that
creates the detailed and complex structure of the adult organism. For exam-
ples of computation inspired by that in nature, we may cite artificial neural
networks, genetic algorithms, artificial immune systems, and ant swarm op-
timization, to name just a few. Next I will consider a few of the issues that
are important in natural computation, but outside the frame of relevance of
the CT model.
One of the most obvious issues is that, because computation in nature
serves an adaptive purpose, it must satisfy stringent real-time constraints.
For example, an animal’s nervous system must respond to a stimulus —
fight or flight, for example — in a fraction of a second. Also, in order to
control coordinated sensorimotor behavior, the nervous system has to be
able to process sensory and proprioceptive inputs quickly enough to generate
effector control signals at a rate appropriate to the behavior. And an ant
colony must be able to allocate workers appropriately to various tasks in real
time in order to maintain the health of the colony.
In nature, asymptotic complexity is generally irrelevant; the constants
matter and input size is generally fixed or varies over a relatively limited range
(e.g., numbers of sensory receptors, colony size). Whether the algorithm
is linear, quadratic, or exponential is not so important as whether it can
deliver useful results in required real-time bounds in the cases that actually
occur. The same applies to other computational resources. For example, it is
not so important whether the number of neurons required varies linearly or
quadratically with the number of inputs to the network; what matters is the
absolute number of neurons required for the actual number of inputs, and
how well the system will perform with the number of inputs and neurons it
actually has.
Therefore, in natural computation, what does matter is how the real-time
response rate of the system is related to the real-time rates of its components
(e.g., neurons, ants) and to the actual number of components. This means
that it is not adequate to treat basic computational processes as having an
indeterminate duration or speed, as is commonly done in the CT model.
In the natural-computation frame of relevance, knowing that a computation
will eventually produce a correct result using finite but unbounded resources
is largely irrelevant. The question is whether it will produce a good-enough

result using available resources subject to real-time constraints.
Many of the inputs and outputs to natural computation are continuous
in magnitude and vary continuously in real time (e.g., intensities, concen-
trations, forces, spatial relations). Many of the computational processes are
also continuous, operating in continuous real time on continuous quantities
(e.g., neural firing frequencies and phases, dendritic electrical signals, pro-
tein synthesis rates, metabolic rates). Obviously these real variables can be
approximated arbitrarily closely by discrete quantities, but that is largely
irrelevant in the natural-computation frame of relevance. The most natural
way to model these systems is in terms of continuous quantities and processes.
If the answers to questions in natural computation seem to depend on
“metaphysical issues,” such as whether only Turing-computable reals exist,
or whether all the reals of standard analysis exist, or whether non-standard
reals exist, then I think that is a sign that we are out of the model’s frame
of relevance, and that the answers are more indicative of the model itself
than of the modeled natural-computation system. For models of natural
computation, naive real analysis, like that commonly used in science and
engineering, should be more than adequate; it seems unlikely that disputes
in the foundations of mathematics will be relevant to our understanding how
brains coordinate animal behavior, how ants and wasps organize their nests,
how embryos self-organize, and so forth.
C.2.b Cross-frame comparisons

This example illustrates the more general pitfalls that arise from cross-frame
comparisons. If two models have different frames of relevance, then they will
make different simplifying and idealizing assumptions; for example objects
whose existence is assumed in one frame (such as standard real numbers)
may not exist in the other (where all objects are computable). Therefore,
a comparison requires that one of the models be translated from its own
frame to the other (or that both be translated to a third), and, in doing this
translation, assumptions compatible with the new frame will have to be made.
For example, if we want to investigate the computational power of neural nets
in the CT frame (i.e., in terms of classes of functions of the integers), then
we will have to decide how to translate the naive continuous variables of
the neural net model into objects that exist in the CT frame. For instance,
we might choose fixed-point numbers, computable reals (represented in some
way by finite programs), or arbitrary reals (represented by infinite discrete

structures). We then discover (as reported in the literature: e.g., Maass &
Sontag, 1999a; Siegelmann & Sontag, 1994a), that our conclusions depend
on the choice of numerical representation (which is largely irrelevant in the
natural-computation frame). That is, our conclusions are more a function of
the specifics of the cross-frame translation than of the modeled systems.
Such results tell us nothing about, for example, why brains do some
things so much better than do contemporary computers, which are made of
much faster components. That is, in the frame of natural computation, the
issue of the representation of continuous quantities does not arise, for it is
irrelevant to the questions addressed by this frame, but this issue is crucial
in the CT frame. Conversely, from within the frame of the CT model, much
of what is interesting about neural net models (parallelism, robustness, real-
time response) becomes irrelevant. Similar issues arise when the CT model
is taken as a benchmark against which to compare unconventional models of
computation, such as quantum and molecular computation.
C.2.c Relevant issues for natural computation

We have seen that important issues in the CT frame of relevance, such as
asymptotic complexity and the computability of classes of functions, are not
so important in natural computation. What, then, are the relevant issues?
One important issue in natural computation is robustness, by which I
mean effective operation in the presence of noise, uncertainty, imprecision,
error, and damage, all of which may affect the computational process as
well as its inputs. In the CT model, we assume that a computation should
produce an output exactly corresponding to the evaluation of a well-defined
function on a precisely specified input; we can, of course, deal with error
and uncertainty, but it’s generally added as an afterthought. Natural com-
putation is better served by models that incorporate this indefiniteness a
priori.
In the CT model, the basic standard of correctness is that a program
correctly compute the same outputs as a well-defined function evaluated on
inputs in that function’s domain. In natural computation, however, we are
often concerned with generality and flexibility, for example: How well does
a natural computation system (such as a neural network) respond to inputs
that are not in its intended domain (the domain over which it was trained or
for which it was designed)? How well does a neural control system respond
to unanticipated inputs or damage to its sensors or effectors? A related

issue is adaptability: How well does a natural computation system change its
behavior (which therefore does not correspond to a fixed function)?
Finally, many natural computation systems are not usefully viewed as
computing a function at all. As previously remarked, with a little cleverness
anything can be viewed as a function, but this is not the simplest way to
treat many natural systems, which often are in open and continuous interac-
tion with their environments and are effectively nonterminating. In natural
computation we need to take a more biological view of a computational sys-
tem’s “correctness” (better: effectiveness). It will be apparent that the CT
model is not particularly well suited to addressing many of these issues, and
in a number of cases begs the questions or makes assumptions incompatible
with addressing them. Nevertheless, real-time response, generality, flexibility,
adaptability, and robustness in the presence of noise, error, and uncertainty
are important issues in the frame of relevance of natural computation.
C.2.d Nanocomputation
Nanocomputation is another domain of computation that seems to be out-
side the frame of relevance of the CT model. By nanocomputation I mean
any computational process involving sub-micron devices and arrangements of
information; it includes quantum computation (Ch. III) and molecular com-
putation (e.g., DNA computation), in which computation proceeds through
molecular interactions and conformational changes (Ch. IV).
Due to thermal noise, quantum effects, etc., error and instability are un-
avoidable characteristics of nanostructures. Therefore they must be taken
as givens in nanocomputational devices and in their interrelationships (e.g.,
interconnections), and also in the structures constructed by nanocomputa-
tional processes (e.g., in algorithmic self-assembly: Winfree, 1998). There-
fore, a “perfect” structure is an over-idealized assumption in the context
of nanocomputation; defects are unavoidable. In many cases structures are
not fixed, but are stationary states occurring in a system in constant flux.
Similarly, unlike in the CT model, nanocomputational operations cannot
be assumed to proceed correctly, for the probability of error is always non-
negligible. Error cannot be considered a second-order detail added to an
assumed perfect computational system, but should be built into a model
of nanocomputation from the beginning. Indeed, operation cannot even be
assumed to proceed uniformly forward. For example, chemical reactions al-
ways have a non-zero probability of moving backwards, and therefore molec-

ular computation systems must be designed so that they accomplish their
purposes in spite of such reversals. This is a fundamental characteristic of
molecular computation and should be an essential part of any model of it.
C.2.e Summary of issues

In summary, the notion of super-Turing computation, stricto sensu, exists
only in the frame of relevance of the Church-Turing model of computation,
for the notion of being able to compute “more” than a Turing machine pre-
supposes a particular notion of “power.” Although it is interesting and im-
portant to investigate where unconventional models of computation fall in
this computational hierarchy, it is also important to explore non-Turing com-
putation, that is, models of computation with different frames of relevance
from the CT model. Several issues arise in the investigation of non-Turing
computation: (1) What is computation in the broad sense? (2) What frames
of relevance are appropriate to unconventional conceptions of computation
(such as natural computation and nanocomputation), and what sorts of mod-
els do we need for them? (3) How can we fundamentally incorporate error,
uncertainty, imperfection, and reversibility into computational models? (4)
How can we systematically exploit new physical processes (quantum, molec-
ular, biological, optical) for computation? The remainder of this chapter
addresses issues (1) and (4).
C.3 Computation in general

C.3.a Kinds of computation
Historically, there have been many kinds of computation, and the existence of
alternative frames of relevance shows us the importance of non-Turing models
of computation. How, then, can we define “computation” in sufficiently
broad terms? Prior to the twentieth century computation involved operations
on mathematical objects by means of physical manipulation. The familiar
examples are arithmetic operations on numbers, but we are also familiar with
the geometric operations on spatial objects of Euclidean geometry, and with
logical operations on formal propositions. Modern computers operate on a
much wider variety of objects, including character strings, images, sounds,
and much else. Therefore, the observation that computation uses physical
processes to accomplish mathematical operations on mathematical objects

must be understood in the broadest sense, that is, as abstract operations on
abstract objects. In terms of the traditional distinction between form and
matter, we may say that computation uses material states and processes to
realize (implement) formal operations on abstract forms. But what sorts of
physical processes?
C.3.b Effectiveness and mechanism

The concepts of effectiveness and mechanism, familiar from CT computa-
tion, are also relevant to computation in a broader sense, but they must be
similarly broadened. To do this, we may consider the two primary uses to
which models of computation are put: understanding computation in nature
and designing computing devices. In both cases the model relates informa-
tion representation and processing to underlying physical processes that are
considered unproblematic within the frame of relevance of the model.
For example, the CT model sought to understand effective calculability
and formal derivability in terms of simple processes of symbol recognition and
manipulation, such as are routinely performed by mathematicians. Although
these are complex processes from a cognitive science standpoint, they were
considered unproblematic in the context of metamathematics. Similarly, in
the context of natural computation, we may expect a model of computation
to explain intelligent information processing in the brain in terms of electro-
chemical processes in neurons (considered unproblematic in the context of
neural network models). Or we may expect a different model to explain the
efficient organization of an ant colony in term of pheromone emission and
detection, simple stimulus-response rules, etc. In all these cases the explana-
tion is mechanistic, in the sense that it refers to primary qualities, which can
be objectively measured or positively determined, as opposed to secondary
qualities, which are subjective or depend on human judgment, feeling, etc.
(all, of course, in the context to the intended purpose of the model); mea-
surements and determinations of primary qualities are effective in that their
outcomes are reliable and dependable.
A mechanistic physical realization is also essential if a model of compu-
tation is to be applied to the design of computing devices. We want to use
physical processes that are effective in the broad sense that they result re-
liably in the intended computations. In this regard, electronic binary logic
has proved to be an extraordinarily effective mechanism for computation. (In
Sec. C.4.b I will discuss some general effectiveness criteria.)
C.3.c Multiple realizability

Although the forms operated upon by a computation must be materially
realized in some way, a characteristic of computation that distinguishes it
from other physical processes is that it is independent of specific material
realization. That is, although a computation must be materially realized
in some way, it can be realized in any physical system having the required
formal structure. (Of course, there will be practical differences between dif-
ferent physical realizations, but I will defer consideration of them until later.)
Therefore, when we consider computation qua computation, we must, on the
one hand, restrict our attention to formal structures that are mechanistically
realizable, but, on the other, consider the processes independently of any
particular mechanistic realization.
These observations provide a basis for determining whether or not a par-
ticular physical system (in the brain, for example) is computational (MacLen-
nan, 1994c, 2004). If the system could, in principle at least, be replaced by
another physical system having the same formal properties and still accom-
plish its purpose, then it is reasonable to consider the system computational
(because its formal structure is sufficient to fulfill its purpose). On the other
hand, if a system can fulfill its purpose only by control of particular sub-
stances or particular forms of energy (i.e., it is not independent of a specific
material realization), then it cannot be purely computational. (Nevertheless,
a computational system will not be able to accomplish its purpose unless it
can interface properly with its physical environment; this is a topic I will
consider in Sec. C.3.f.)
C.3.d Defining computation

Based on the foregoing considerations, we have the following definition of
computation (MacLennan, 1994c, 2004, 2009):
Definition 1 Computation is a mechanistic process, the purpose of which
is to perform abstract operations on abstract objects.
Alternately, we may say that computation accomplishes the formal trans-
formation of formal objects by means of mechanistic processes operating on
the objects’ material embodiment. The next definition specifies the relation
between the physical and abstract processes:
Definition 2 A mechanistic physical system realizes a computation if, at the

level of abstraction appropriate to its purpose, the abstract transformation of
the abstract objects is a sufficiently accurate model of the physical process.
Such a physical system is called a realization of the computation.
That is, the physical system realizes the computation if we can see the ma-
terial process as a sufficiently accurate embodiment of the formal structure,
where the sufficiency of the accuracy must be evaluated in the context of
the system’s purpose. Mathematically, we may say that there is a homomor-
phism from the physical system to the abstract system, because the abstract
system has some, but not all, of the formal properties of the physical system
(MacLennan, 1994a, 2004). The next definition classifies various systems,
both natural and artificial, as computational:
Definition 3 A physical system is computational if its function (purpose)
is to realize a computation.
Definition 4 A computer is an artificial computational system.
Thus the term “computer” is restricted to intentionally manufactured com-

putational devices; to call the brain a computer is a metaphor. These def-
initions raise a number of issues, which I will discuss briefly; no doubt the
definitions can be improved.
C.3.e Purpose
First, these definitions make reference to the function or purpose of a system,
but philosophers and scientists are justifiably wary of appeals to purpose, es-
pecially in a biological context. However, the use of purpose in the definition
of computation is unproblematic, for in most cases of practical interest, pur-
pose is easy to establish. (There are, of course, borderline cases, but that
fact does not invalidate the definition.) On the one hand, in a technological
context, we can look to the stated purpose for which an artificial system was
designed. On the other, in a biological context, scientists routinely inves-
tigate the purposes (functions) of biological systems, such as the digestive
system and immune system, and make empirically testable hypotheses about
their purposes. Ultimately such claims of biological purpose may be reduced
to a system’s selective advantage to a particular species in that species’ en-
vironment of evolutionary adaptedness, but in most cases we can appeal to
more immediate ideas of purpose.
On this basis we may identify many natural computational systems. For

example, the function of the brain is primarily computational (in the sense
used here), which is easiest to see in sensory areas. For example, there is
considerable evidence that an important function of primary visual cortex is
to perform a Gabor wavelet transform on visual data (Daugman, 1993); this
is an abstract operation that could, in principal, be realized by a non-neural
physical system (such as a computer chip). Also, pheromone-mediated in-
teractions among insects in colonies often realize computational ends such as
allocation of workers to tasks and optimization of trails to food sources. Like-
wise DNA transcription, translation, replication, repair, etc., are primarily
computational processes.
However, there is a complication that arises in biology and can be ex-
pected to arise in our biologically-inspired robots and more generally in
post-Moore’s Law computing. That is, while the distinction between com-
putational and non-computational systems is significant to us, it does not
seem to be especially significant to biology. The reason may be that we are
concerned with the multiple realizability of computations, that is, with the
fact that they have alternative realizations, for this property allows us to
consider the implementation of a computation in a different technology, for
example in electronics rather than in neurons. In nature, typically, the real-
ization is given, since natural life is built upon a limited range of substances
and processes. On the other hand, there is often selective pressure in favor of
exploiting a biological system for as many purposes as possible. Therefore, in
a biological context, we expect physical systems to serve multiple functions,
and therefore many such systems will not be purely computational; they
will fulfill other functions besides computation. From this perspective, it is
remarkable how free nervous systems are of non-computational functions.
C.3.f Transduction
The purpose of computation is the abstract transformation of abstract ob-
jects, but obviously these formal operations will be pointless unless the com-
putational system interfaces with its environment in some way. Certainly
our computers need input and output interfaces in order to be useful. So
also computational systems in the brain must interface with sensory recep-
tors, muscles, and many other noncomputational systems to fulfill their func-
tions. In addition to these practical issues, the computational interface to the
physical world is relevant to the symbol grounding problem, the philosophi-
cal question of how abstract symbols can have real-world content (Harnad,
1990, 1993; MacLennan, 1993). Therefore we need to consider the interface
between a computational system and its environment, which comprises input
and output transducers.
The relation of transduction to computation is easiest to see in the case
of analog computers. The inputs and outputs of the computational system
have some physical dimensions (light intensity, air pressure, mechanical force,
etc.), because they must have a specific physical realization for the system
to accomplish its purpose. On the other hand, the computation itself is
essentially dimensionless, since it manipulates pure numbers. Of course,
these internal numbers must be represented by some physical quantities, but
they can be represented in any appropriate physical medium. In other words,
computation is generically realized, that is, realized by any physical system
with an appropriate formal structure, whereas the inputs and outputs are
specifically realized, that is, constrained by the environment with which they
interface to accomplish the computational system’s purpose.
Therefore we can think of (pure) transduction as changing matter (or
energy) while leaving form unchanged, and of computation as transforming
form independently of matter (or energy). In fact, most transduction is not
pure, for it modifies the form as well as the material substrate, for example, by
filtering. Likewise, transductions between digital and analog representations
transform the signal between discrete and continuous spaces.
C.3.g Classification of computational dynamics

The preceding definition of computation has been framed quite broadly, to
make it topology-neutral, so that it encompasses all the forms of computation
found in natural and artificial systems. It includes, of course, the familiar
computational processes operating in discrete steps and on discrete state
spaces, such as in ordinary digital computers. It also includes continuous-
time processes operating on continuous state spaces, such as found in conven-
tional analog computers and field computers (Adamatzky, 2001; Adamatzky
et al., 2005; MacLennan, 1987, 1999). However, it also includes hybrid pro-
cesses, incorporating both discrete and continuous computation, so long as
they are mathematically consistent (MacLennan, 2010). As we expand our
computational technologies outside of the binary electronic realm, we will
have to consider these other topologies of computation. This is not so much
a problem as an opportunity, for many important applications, especially in
natural computation, are better matched to these alternative topologies.

In connection with the classification of computational processes in terms
of their topologies, it is necessary to say a few words about the relation be-
tween computations and their realizations. A little thought will show that
a computation and its realizations do not have to have the same topology,
for example, discrete or continuous. For instance, the discrete computations
performed on our digital computers are in fact realized by continuous phys-
ical systems obeying Maxwell’s equations. The realization is approximate,
but exact enough for practical purposes. Conversely a discrete system can
approximately realize a continuous system, analogously to numerical integra-
tion on a digital computer. In comparing the topologies of the computation
and its realization, we must describe the physical process at the relevant
level of analysis, for a physical system that is discrete on one level may be
continuous on another. (The classification of computations and realizations
is discussed in MacLennan, 2004.)
C.4 Expanding the range of computing technologies

C.4.a A vicious cycle
A powerful feedback loop has amplified the success of digital VLSI technol-
ogy to the exclusion of all other computational technologies. The success
of digital VLSI encourages and finances investment in improved tools, tech-
nologies, and manufacturing methods, which further promote the success of
digital VLSI. Unfortunately this feedback loop threatens to become a vi-
cious cycle. We know that there are limits to digital VLSI technology, and,
although estimates differ, we will reach them soon (see Ch. II). We have as-
sumed there will always be more bits and more MIPS, but that assumption
is false. Unfortunately, alternative technologies and models of computation
remain undeveloped and largely uninvestigated, because the rapid advance
of digital VLSI has surpassed them before they could be adequately refined.
Investigation of alternative computational technologies is further constrained
by the assumption that they must support binary logic, because that is the
only way we know how to compute, or because our investment in this model
of computation is so large. Nevertheless, we must break out of this vicious
cycle or we will be technologically unprepared when digital VLSI finally, and
inevitably, reaches its limits.
C.4.b General guidelines
Therefore, as a means of breaking out of this vicious cycle, let us step back
and look at computation and computational technologies in the broadest
sense. What sorts of physical processes can we reasonably expect to use for
computation? Based on the preceding discussion, we can see that any math-
ematical process, that is, any abstract transformation of abstract objects, is
a potential computation. Therefore, in principle, any reasonably controllable,
mathematically described, physical process can be used for computation. Of
course, there are practical limitations on the physical processes usable for
computation, but the range of possible technologies is much broader than
might be suggested by a narrow conception of computation. Considering
some of the requirements for computational technologies will reveal some of
the possibilities as well as the limitations.
One obvious issue is speed. The rate of the physical process may be either
too slow or too fast for a particular computational application. That it might
be too slow is obvious, for the development of conventional computing tech-
nology has been driven by speed. Nevertheless, there are many applications
that have limited speed requirements, for example, if they are interacting
with an environment with its own limited rates. Conversely, these applica-
tions may benefit from other characteristics of a slower technology, such as
energy efficiency; insensitivity to uncertainty, error, and damage; and the
ability to be reconfigured or to adapt or repair itself. Sometimes we simply
want to slow a simulated process down so we can observe it. Another con-
sideration that may supersede speed is whether the computational medium
is suited to the application: Is it organic or inorganic? Living or nonliving?
Chemical, optical, or electrical?
A second requirement is the ability to implement the transducers required
for the application. Although computation is theoretically independent of its
physical embodiment, its inputs and outputs are not, and some conversions
to and from a computational medium may be easier than others. For exam-
ple, if the inputs and outputs to a computation are chemical, then chemical
or molecular computation may permit simpler transducers than electronic
computation. Also, if the system to be controlled is biological, then some
form of biological computation may suit it best.
Finally, a physical realization should have the accuracy, stability, control-
lability, etc. required for the application. Fortunately, natural computation
provides many examples of useful computations that are accomplished by
realizations that are not very accurate, for example, neuronal signals have at
most about one digit of precision. Also, nature shows us how systems that
are subject to many sources of noise and error may be stabilized and thereby
accomplish their purposes.
C.4.c Learning to use new technologies

A key component of the vicious cycle is our extensive knowledge about de-
signing and programming digital computers. We are naturally reluctant to
abandon this investment, which pays off so well, but as long as we restrict
our attention to existing methods, we will be blind to the opportunities of
new technologies. On the other hand, no one is going to invest much time
or money in technologies that we don’t know how to use. How can we break
the cycle?
In many respects natural computation provides the best opportunity, for
nature offers many examples of useful computations based on different mod-
els from digital logic. When we understand these processes in computational
terms, that is, as abstractions independent of their physical realizations in
nature, we can begin to see how to apply them to our own computational
needs and how to realize them in alternative physical processes. As examples
we may take information processing and control in the brain, and emergent
self-organization in animal societies, both of which have been applied already
to a variety of computational problems (e.g., artificial neural networks, ge-
netic algorithms, ant colony optimization, etc.). But there is much more
that we can learn from these and other natural computation systems, and
we have not made much progress in developing computers better suited to
them. More generally we need to increase our understanding of computation
in nature and keep our eyes open for physical processes with useful math-
ematical structure (Calude et al., 1998; Calude & Paun, 2001). Therefore,
one important step toward a more broadly based computer technology will
be a knowledge-base of well-matched computational methods and physical
realizations.
Computation in nature gives us many examples of the matching of physi-
cal processes to the needs of natural computation, and so we may learn valu-
able lessons from nature. First, we may apply the actual natural processes as
realizations of our artificial systems, for example using biological neurons or
populations of microorganisms for computation. Second, by understanding
the formal structure of these computational systems in nature, we may realize
them in alternative physical systems with the same abstract structure. For
example, neural computation or insect colony-like self-organization might be
realized in an optical system.
C.4.d General-purpose computation

An important lesson learned from digital computer technology is the value
of programmable general-purpose computers, both for prototyping special-
purpose computers as well as for use in production systems. Therefore to
make better use of an expanded range of computational methodologies and
technologies, it will useful to have general-purpose computers in which the
computational process is controlled by easily modifiable parameters. That
is, we will want generic computers capable of a wide range of specific compu-
tations under the control of an easily modifiable representation. As has been
the case for digital computers, the availability of such general-purpose com-
puters will accelerate the development and application of new computational
models and technologies.
We must be careful, however, lest we fall into the “Turing Trap,” which is
to assume that the notion of universal computation found in Turing machine
theory is the appropriate notion in all frames of relevance. The criteria of uni-
versal computation defined by Turing and his contemporaries was appropriate
for their purposes, that is, studying effective calculability and derivability in
formal mathematics. For them, all that mattered was whether a result was
obtainable in a finite number of atomic operations and using a finite number
of discrete units of space. Two machines, for example a particular Turing
machine and a programmed universal Turing machine, were considered to
be of the same power if they computed the same function by these criteria.
Notions of equivalence and reducibility in contemporary complexity theory
are not much different.
It is obvious that there are many important uses of computers, such as
real-time control applications, for which this notion of universality is irrel-
evant. In some of these applications, one computer can be said to emulate
another only if it does so at the same speed. In other cases, a general-purpose
computer may be required to emulate a particular computer with at most a
fixed extra amount of a computational resource, such as storage space. The
point is that in the full range of computer applications, in particular in nat-
ural computation, there may be considerably different criteria of equivalence
than computing the same mathematical function. Therefore, in any partic-
ular application area, we must consider in what respects the programmed

general-purpose computer must behave the same as the computer it is em-
ulating, and in what respects it may behave differently, and by how much.
That is, each notion of universality comes with a frame of relevance, and
we must uncover and explicate the frame of relevance appropriate to our
application area.
There has been limited work on general-purpose computers in the non-
Turing context. For example, theoretical analysis of general-purpose analog
computation goes back to Claude Shannon (1941), with more recent work
by Pour-El, Lipshitz, and Rubel (Pour-El, 1974b; Lipshitz & Rubel, 1987;
Rubel, 1993; Shannon, 1941, 1993). In the area of neural networks we have
several theorems based on Sprecher’s improvement of the Kolmogorov super-
position theorem (Sprecher, 1965), which defines one notion of universality
for feed-forward neural networks, although perhaps not a very useful one, and
there are several “universal approximation theorems” for neural networks and
related computational models (Haykin, 2008, pp. 166–168, 219–220, 236–239,
323–326). Also, there are some CT-relative universality results for molecular
computation (Calude & Paun, 2001) and for computation in nonlinear me-
dia (Adamatzky, 2001). Finally, we have done some work on general-purpose
field computers (MacLennan, 1987, 1990, 1999, 2015) and on general-purpose
computation over second-countable metric spaces (which includes both ana-
log and digital computation) (MacLennan, 2010). In any case, much more
work needs to be done, especially towards articulating the relation between
notions of universality and their frames of relevance.
It is worth remarking that these new types of general-purpose computers
might not be programmed with anything that looks like an ordinary program,
that is, a textual description of rules of operation. For example, a guiding
image, such as a potential surface, might be used to govern a gradient descent
process or even a nondeterministic continuous process (MacLennan, 1995,
2004). We are, indeed, quite far from universal Turing machines and the
associated notions of programs and computation, but non-Turing models are
often more relevant in natural computation and other kinds of unconventional
computation.
C.5 Conclusions
The historical roots of Church-Turing computation remind us that the the-
ory exists in a frame of relevance, which is not well suited to post-Moore’s
Law unconventional computation. Therefore we need to supplement it with

new models based on different assumptions and suited to answering differ-
ent questions. Central issues include real-time response, generality, flexibil-
ity, adaptability, and robustness in the presence of noise, uncertainty, error,
and damage. Once we understand computation in a broader sense than the
Church-Turing model, we begin to see new possibilities for using physical pro-
cesses to achieve our computational goals. These possibilities will increase
in importance as we approach the limits of electronic binary logic as a ba-
sis for computation, and they will also help us to understand computational
processes in nature.
Chapter II
Physics of Computation
A Energy dissipation
As an introduction to the physics of computation, and further motivation
for unconventional computation, we will discuss Michael P. Frank’s analysis
of energy dissipation in conventional computing technologies (Frank, 2005b).
The performance R of a computer system can meeasured by the number of
computational operations executed per unit time. This ratio is the product
of the number operations per unit of dissipated energy times the energy
dissipation per unit time:
Nops Nops Ediss

R= = × = FE × Pdiss . (II.1)
t Ediss t
Here we have defined Pdiss to be the power dissipated by the computation
and the energy efficiency FE to be to be the number of low-level bit opera-
tions performed per unit of energy. The key parameter is FE , which is the
reciprocal of the energy dissipated per bit operation.
This energy can be estimated as follows. Contemporary digital electronics
uses CMOS technology, which represents a bit as the charge on a capacitor.
The energy to set or reset the bit is (approximately) the energy to charge
the capacitor or the energy dissipated when it discharges. Voltage is energy
29
30 CHAPTER II. PHYSICS OF COMPUTATION
per unit charge, so the work to move an infinitesimal charge dq from one
plate to the other is V dq, where V is the voltage between the plates. But
V is proportional to the charge already on the capacitor, V = q/C. So the
change in energy is dE = V dq = Cq dq. Hence the energy to reach a charge
Q is Z Q
q 1 Q2
E= dq = .
0 C 2C
Therefore, E = 12 (CV )2 /C = 12 CV 2 and FE ≈ (1 op)/( 12 CV 2 ).
Frank observes that Moore’s law in the 1985–2005 period was a result of
an exponential decrease in C resulting from decreasing feature sizes (since
capacitance is proportional to area) and a decrease in logic voltage V from
5V to about 1V (further improving E by a factor of 25). The clock rate also
went up with smaller feature sizes. (See Fig. II.1.)
Unfortunately, neither the transistor lengths nor the voltage can be re-
duced much more, for if the signal is too small in comparison with thermal
energy, then thermal noise will lead to unreliable operation, because the
thermal fluctuations will be of the same order as the signals (Fig. II.2). The
thermal energy is ET = kB T , where kB is Boltzmann’s constant and T is the
absolute temperature. Since kB ≈ 8.6 × 10−5 eV/K = 1.38 × 10−23 J/K, and
room temperature T ≈ 300K, room-temperature thermal energy is
ET = kB T ≈ 26 meV ≈ 4.14 × 10−21 J ≈ 4 zJ.
(Fig. II.1 shows ET .)
We have seen that Esig = 21 CV 2 , but for reliable operation, how big
should it be in comparison to ET ? Frank estimates Esig ≥ kB T ln R, where
the reliability R = 1/perr , for a desired probability of error perr .1 For example,
for a reasonable reliability R = 2 × 1017 , Esig ≥ 40kB T ≈ 1 eV, which is the
energy to move one electron with 1V logic levels. This implies a maximum
energy efficiency of
1 op
FE = 1 op/eV ≈ −19
= 6.25 × 1018 op/J. (II.2)
1.6 × 10 J
A round 100kB T corresponds to an error probability of perr = e−100 = 3.72 ×
10−44 (at room temperature). Therefore, a reasonable target for reliable
operation is
Esig & 100kB T ≈ 2.6 eV = 414 zJ.
1
Frank (2005b, slide 7).
A. ENERGY DISSIPATION 31
Trend of ITRS
Min.'97-'03
Transistor Switching
Gate Energy Trends Energy
Based on ITRS ’97-03 roadmaps
1.E-14
250 LP min gate energy, aJ
1.E-15 180 HP min gate energy, aJ
130 Node numbers fJ
100 k(300 K)
90 (nm DRAM hp) ln(2) k(300 K)
1.E-16 65
1 eV
45
32 k(300 K)
CVV/2 energy, J
1.E-17 22
Practical limit for CMOS?
1.E-18 Room-temperature 100 kT reliability limit aJ
One electron volt
1.E-19
1.E-20 Room-temperature kT thermal energy
Room-temperature von Neumann - Landauer limit

1.E-21 zJ
1.E-22
1995 2000 2005 2010 2015 2020 2025 2030 2035 2040 2045
Year
2005/05/05 M. Frank, "Introduction to Reversible Computing" 6
Figure II.1: Historical and extrapolated switching energy. Figure from Frank
(2005b, slide 9).
0.75
0.50
0.25
0.00
0.25
0.50
0.75
0 1 2 3 4 5
Figure II.2: Depiction of 0-1-0-1-1 pulses in the presence of high thermal

noise.
This, therefore, is an estimate of the minimum energy dissipation per op-

eration for reliable operation using conventional technology. Nevertheless,
these conclusions are independent of technology (electronic, optical, carbon
nanotube, etc.), since they depend only on relative energy levels for reliable
operation.2
One apparent solution is to operate at a lower temperature T , but it does
not help much, since the effective T has to reflect the environment into which
the energy is eventually dissipated (i.e., the energy dissipation has to include
the refrigeration to operate below ambient temperature). Another possible
solution, operating closer to kB T and compensating for low reliability with
error-correcting codes, does not help, because we need to consider the total
energy for encoding a bit. That is, we have to include the additional bits
required for error detection and correction.
Frank observed in 2005 that the smallest logic signals were about 104 kB T ,
and therefore that there were only about two orders of magnitude improve-
ment in reliable operation. “A factor of 100 means only around 10 years re-
main of further performance improvements, given the historical performance
doubling period of about 1.5 years. Thus, by about 2015, the performance
2
Frank presentation, “Reversible Computing: A Cross-Disciplinary Introduction” (Be-
yond Moore), Mar. 10, 2014. put in bib
A. ENERGY DISSIPATION 33
of conventional computing will stop improving, at least at the device level”

(Frank, 2005b).
In fact, these limitations are becoming apparent. By 2011 computer en-
gineers were worrying about “the 3 GHz wall,” since computer clock speeds
had been stalled at about that rate for five years.3 Recent processors have
gone a little beyond the barrier, but a “power wall” remains, for although in-
dividual transistors can be operated at higher speeds, the millions or billions
of transistors on a chip dissipate excessive amounts of energy. This presents
an obstacle for future supercomputers.
As of August 2017 the fastest computer was the Sunway TaihuLight.4 It is
rated at 93 petaflops on the LINPACK benchmark and has 10,649,600 cores
with 1.31 PB of memory. It consumes 16 MW, which is about the power
consumed by 1400 homes, and is quite efficient, with FE = 6 Gflops/W,
that is 166 pJ/flop (the fourth most efficient supercomputer at the time).
To convert floating-point operations to basic logic operations, including all
the overhead etc., one conversion estimate is 107 to 108 ops/flop.5 There-
fore, we can compare the theoretical best energy efficiency (Eq. II.2), FE−1 =
1.6 × 10−7 pJ/op ≈ 1.6 to 16 pJ/flop, with the 144pJ/flop of the Sunway Tai-
huLight. The gap is only about one order of magnitude. Indeed, it has been
estimated that scaling up current technology to 1 exaflops would consume
1.5 GW, more than 0.1% of US power grid.6 This is impractical.
It might be possible to get energy consumption down to 5 to 10 pJ/flop,
but “the energy to perform an arithmetic operation is trivial in comparison
with the energy needed to shuffle the data around, from one chip to another,
from one board to another, and even from rack to rack.”7 Indeed, due to the
difficulty of programming parallel computers, and due to delays in internal
data transmission, it is difficult to use more than 5% to 10% of a supercom-
3
Spectrum (Feb. 2011) spectrum.ieee.org/computing/hardware/nextgeneration-
supercomputers/0 (accessed Aug. 20, 2012).
4
https://en.wikipedia.org/wiki/Sunway TaihuLight (accessed Aug. 11, 2017). I will
use “flop” for “floating point operations” and “flops” for “floating point operations per
second” = flop/s. Note that “flop” is a count and “flops” is a rate. Also note that since
W = J/s, flops/W = flop/J.
5
And so this is one estimate of the difference in time scale between computational
abstractions and the logic that implements them, which was discussed in Ch. I (p. 5).
6
7
puter’s capacity for any extended period; most of the processors are idling.8
So with those 10.6 × 106 cores, most of the time about ten million of them
are idle! There has to be a better way.
8
supercomputers/0 (accessed 2012-08-20).
B. THERMODYNAMICS OF COMPUTATION 35
B Thermodynamics of computation
“Computers may be thought of as engines for transforming free energy into
waste heat and mathematical work.” — Bennett (1982)
B.1 Von Neumann-Landaur Principle

B.1.a Information and entropy
We will begin with a quick introduction or review of the entropy concept;
we will look at it in more detail soon (Sec. B.2). The information content
of a signal (message) measures our “surprise,” that is, how unlikely it is.
I(s) = − logb P{s}, where P{s} is the probability of signal s. We take logs
so that the information content of independent signals is additive. We can
use any base, and for b = 2, e, and 10, the corresponding units are bits (or
shannons), nats, and dits (also, hartleys, bans). Therefore, if a signal has a
50% probability, then it conveys one bit of information. The entropy of a
distribution of signals is their average information content:
X X
H(S) = E{I(s) | s ∈ S} = P{s}I(s) = − P{s} log P{s}.
s∈S s∈S
P
Or more briefly, H = − k pk log pk . For example, if P{1} = 1/16 and
P{0} = 15/16, we will receive, on the average, H = 0.3 bits of information.
According to a well-known story, Shannon was trying to decide what to
call this quantity and had considered both “information” and “uncertainty.”
Because it has the same mathematical form as statistical entropy in physics,
von Neumann suggested he call it “entropy,” because “nobody knows what
entropy really is, so in a debate you will always have the advantage.”9 (This
is one version of the quote.) I will call it information entropy when I need to
distinguish it from the thermodynamical concept.
An important special case is the entropy of a uniform distribution. If
there are N signals that are all equally likely, then H = log N . Therefore, if
we have eight equally likely possibilities, the entropy is H = lg 8 = 3 bits.10
9
https://en.wikipedia.org/wiki/History of entropy (accessed 2012-08-24).
Ralph Hartley laid the foundations of information theory in 1928, on which Claude
Shannon built his information theory in 1948.
10
I use the notations lg x = log2 x and ln x = loge x.
A uniform distribution maximizes the entropy (and minimizes the ability to

guess).
In computing, we are often concerned with the state of the computation,
which is realized by the state of a physical system. Consider a physical
system with three degrees of freedom (DoF), each with 1024 possible values.
There are N = 10243 = 230 possible states, each describable by three 10-bit
integers. If we don’t care about the distance between states (i.e., distance
on each axis), then states can be specified equally well by six 5-bit numbers
or one 30-bit number, etc. (or ten digits, since 30 log10 2 ≈ 9.03 digits). Any
scheme that allows us to identify all 230 states will do. We can say that there
are 30 binary degrees of freedom.
In computing we often have to deal with things that grow exponentially
or are exponentially large (due to combinatorial explosion), such as solution
spaces. (For example, NP problems are characterized by the apparent neces-
sity to search a space that grows exponentially with problem size.) In such
cases, we are often most concerned with the exponents and how they relate.
Therefore it is convenient to deal with their logarithms (i.e., with logarith-
mic quantities). The logarithm represents, in a scale-independent way, the
degrees of freedom generating the space.
Different logarithm bases amount to different units of measurement for
logarithmic quantities (such as information and entropy). As with other
quantities, we can leave the units unspecified, so long as we do so consistently.
I will use the notation “log x” for an unspecific logarithm, that is, a logarithm
with an unspecified base.11 When I mean a specific base, I will write ln x,
lg x, log10 x, etc. Logarithms in specific bases can be defined in terms of
unspecific logarithms as follows: lg x = log x/ log 2, ln x = log x/ log e, etc.
(The units can be defined bit = log 2, nat = log e, dit = log 10, etc.)
B.1.b The von Neumann-Landauer bound

Thermodynamic entropy is unknown information residing in the physical
state. The macroscopic thermodynamic entropy S is related to microscopic
information entropy H by Boltzmann’s constant, which expresses the entropy
in thermodynamical units (energy over temperature). If H is measured in
nats, then S = kB H = kB ln N , for N equally likely states. When using
11
Frank (2005a) provides a formal definition for the indefinite logarithm. I am using
the idea less formally, an “unspecific logarithm,” whose base is not mentioned. This is a
compromise between Frank’s concept and familiar notation; we’ll see how well it works!
Figure II.3: Physical microstates representing logical states. Setting the

binary device decreases the entropy: ∆H = lg N − lg(2N ) = −1 bit. That
is, we have one bit of information about its microstate.
unspecific logarithms, I will drop the “B” subscript: S = kH = k log N . The

physical dimensions of entropy are usually expressed as energy over temper-
ature (e.g., joules per kelvin), but the dimensions of temperature are energy
per degree of freedom (measured logarithmically), so the fundamental di-
mension of entropy is degrees of freedom, as we would expect. (There are
technical details that I am skipping.)
Consider a macroscopic system composed of many microscopic parts (e.g.,
a fluid composed of many molecules). In general a very large number of
microstates (or microconfigurations) — such as positions and momentums of
molecules — will correspond to a given macrostate (or macroconfiguration)
— such as a combination of pressure and termperature. For example, with
m = 1020 particles we have 6m degrees of freedom, and a 6m-dimensional
phase space.
Now suppose we partition the microstates of a system into two macro-
scopically distinguishable macrostates, one representing 0 and the other rep-
resenting 1. For example, whether the electrons are on one plate of a ca-
pacitor or the other could determine whether a 0 or 1 bit is stored on it.
Next suppose N microconfigurations correspond to each macroconfiguration
(Fig. II.3). This could be all the positions, velocities, and spins of the many
electrons, which we don’t care about and cannot control individually. If we
confine the system to one half of its microstate space in order to represent a
0 or a 1, then the entropy (average uncertainty in identifying the microstate)
will decrease by one bit. We don’t know the exact microstate, but at least
additional
DoF
logical 0
logical 0
logical 1
process
t t+1
Figure II.4: Thermodynamics of erasing a bit. On the left is the initial state
(time t), which may be logical 0 or logical 1; on the right (time t + 1) the
binary device has been set to logical 0. In each case there are N microstates
representing each prior state, so a total of 2N microstates. However, at time
t + 1 the system must be in one of N posterior microstates. Therefore N
of the microstate trajectories must exit the defined region of phase space
by expanding into additional, uncontrolled degrees of freedom. Therefore
entropy of the environment must increase by at least ∆S = k log(2N ) −
k log N = k log 2. We lose track of this information because it passes into
uncontrolled degrees of freedom.
we know which half of the state-space it is in.

In general, in physically realizing a computation we distinguish information-
bearing degrees of freedom (IBDF), which we control and use for computation,
from non-information-bearing degrees of freedom (NIBDF), which we do not
control and are irrelevant to the computation (Bennett, 2003).
Consider the process of erasing or clearing a bit (i.e., setting it to 0, no
matter what its previous state): we are losing one bit of physical information.
The physical information still exists, but we have lost track of it.
Suppose we have N physical microstates per logical macrostate (logical
0 or logical 1). Before the bit is erased it can be in one of 2N possible
microstates, but there are only N microstates representing its final state.
The laws of physics are reversible,12 so they cannot lose any information.
Since physical information can’t be destroyed, it must go into NIBDF (e.g.,
the environment or thermal motion of the atoms) (Fig. II.4). The trajectories
have to expand into other degrees of freedom (NIBDF) to maintain the phase
12
This is true in both classical and quantum physics. In the latter case, we cannot have
2N quantum states mapping reversibly into only N quantum states.
space volume.
The information lost, or dissipated into NIBDF (typically as heat), is
∆S = k log(2N )−k log N = k log 2. (In physical units this is ∆S = kB ln 2 ≈
10−23 J/K.) Therefore the increase of energy in the device’s environment is
∆Q = ∆S × Tenv = kB Tenv ln 2 ≈ 0.7kTenv . At Tenv = 300K, kB Tenv ln 2 ≈
18 meV ≈ 3 × 10−9 pJ = 3 zJ. We will see that this is the minimum energy
dissipation for any irreversible operation (such as erasing a bit); it is called the
von Neumann–Landauer (VNL) bound (or sometimes simply the Landauer
bound). Von Neumann suggested the idea in 1949, but it was published first
by Rolf Landauer (IBM) in 1961.13 Recall that for reliable operation we need
minimum logic levels around 40kB Tenv to 100kB Tenv , which is two orders of
magnitude above the von Neumann–Landauer limit of 0.7kB Tenv . “From a
technological perspective, energy dissipation per logic operation in present-
day silicon-based digital circuits is about a factor of 1,000 greater than the
ultimate Landauer limit, but is predicted to quickly attain it within the next
couple of decades” (Berut et al., 2012). That is, current circuits have signal
levels of about 18 eV, which we may compare to the VNL, 18 meV.
In research reported in March 2012 Berut et al. (2012) confirmed ex-
perimentally the Landauer Principle and showed that it is the erasure that
dissipates energy. They trapped a 2µ silica ball in either of two laser traps,
representing logical 0 and logical 1. For storage, the potential barrier was
greater than 8kB T , and for erasure, the barrier was lowered to 2.2kB T by
decreasing the power of the lasers and by tilting the device to put it into
the logical 0 state (see Fig. II.5). At these small sizes, heat is a stochastic
property, so the dissipated heat was computed by averaging the trajectory
of the particle over multiple trials:
Z τ
∂U (x, t)
hQi = − ẋ(t) dt .
0 ∂x x
(The angle brackets means “average value.”) Complete erasure results in the
ball being in the logical 0 state; incomplete erasure results in it being in the
logical 0 state with probability p. They established a generalized Landauer
bound in which the dissipated heat depends on the completeness of erasure:
hQipLandauer = kT [log 2 + p log p + (1 − p) log(1 − p)] = kT [log 2 − H(p, 1 − p)].

13
See Landauer (1961), reprinted in Leff & Rex (1990) and Leff & Rex (2003), which
include a number of other papers analyzing the VNL principle.
Figure II.5: Erasing a bit by changing potential barrier. (Figure from Berut
et al. (2012).)
Therefore, for p = 1, heat dissipation is kT ln 2, but for p = 1/2 (ineffective

erasure) no heat needs to be dissipated. Notice how the dissipated energy
depends on the entropy H(p, 1 − p) of the final macrostate.
B.1.c Irreversible operations

Suppose the phase space is divided into M macrostates of size N1 , N2 , . . . , NM ,
where N = N1 + N2 + · · · + NM . Let pij be the probability the device is in
microstate i of macrostate j. The total entropy is
X
S = −k pij log pij . (II.3)
ij
We can separate this into the macroscopic entropy associated with the macro-
states (IBDF) and the microscopic entropy associated with the microstates
PNj
(NIBDF). Now let Pj = i=1 pij be the probability of being in macrostate
j. Then Eq. II.3 can be rearranged (Exer. II.1):
Nj
X X X pij pij
S = −k Pj log Pj − k Pj log = Si + Sh . (II.4)
j j i=1
Pj Pj
The first term is the macrostate entropy (IBDF):

X
Si = −k Pj log Pj .
j
The second is the microstate entropy (NIBDF):

Nj
X X pij pij
Sh = −k Pj log .
j i=1
Pj Pj
Note that the ratios in the inner summation are essentially conditional prob-
abilities, and that the inner summation is the conditional entropy given that
you are in macrostate j.
When we erase a bit, we go from a maximum Si of 1 bit (if 0 and 1 are
equally likely), to 0 bits (since there is no uncertainty). Thus we lose one bit
of information, and the macrostate entropy decreases ∆Si = −k log 2. (The
actual entropy decrease can be less than 1 bit if the 0 and 1 are not equally
likely initial states.) Since according to the Second Law of Thermodynamics
∆S ≥ 0, we have a corresponding minimum increase in microstate entropy,

∆Sh ≥ k log 2. Typically this is dissipated as heat, ∆Q ≥ kT log 2. The
information becomes inaccessible and unusable.
The standard logic gates (And, Or, Xor, Nand, etc.) have two input
bits and one output bit. Therefore the output will have lower entropy than
the input, and so these gates must dissipate at least 1 bit of entropy, kT log 2
energy. Consider, for example, And. If the four inputs 00, 01, 10, 11, are
equally likely, then the input entropy is Hi = 2 bits. However the output
entropy will be Ho = −(1/4) lg(1/4) − (3/4) lg(3/4) = 0.811, so the entropy
lost is 1.189 bits. To compute the dissipated energy in Joules, multiply by
ln 2 to convert bits to nats (or shannons to hartleys):
∆Q = T ∆S ≥ −T kB (Ho − Hi ) ln 2 ≈ 1.189kB T ln 2 ≈ 0.83kB T.
For each gate, we can express Ho in terms of the probabilities of the inputs
and compute the decrease from Hi (exercise). If the inputs are not equally
likely, then the input entropy will be less than 2 bits, but we will still have
Hi > Ho and energy will be dissipated. (Except in a trivial, uninteresting
case. What is it?)
More generally, any irreversible operation (non-invertible function) will
lose information, which has to be dissipated into the environment. That is,
it is irreversible because it loses information, and every time we lose informa-
tion, we lose energy. If the function is not one-to-one (injective), then at least
two inputs map to the same output, and so information about the inputs is
lost. For example, changing a bit, that is, overwriting a bit with another bit,
is a fundamental irreversible operation, subject to the VNL limit. Therefore,
the assignment operation is bound by it. Also, when two control paths join,
we forget where we came from, and so again we must dissipate at least a
bit’s worth of entropy (Bennett, 2003). These considerations suggest that
reversible operations might not be subject to the VNL limit, and this is in
fact the case, as we will see.
The preceding observations have important connections with the problem
of Maxwell’s Demon and its resolution. Briefly, the demon has to reset its
mechanism after each measurement in preparation for the next measurement,
and this dissipates at least kT log 2 energy into the heat bath for each decision
that it makes. That is, the demon must “pay” for any information that it
acquires. Therefore, the demon cannot do useful work. Further discussion
is outside the scope of this book, so if you are interested, please see Leff &
Rex (2003) and Leff & Rex (1990) (which have a large intersection), in which
many of the papers on the topic are collected.
B.2 Mechanical and thermal modes

We need to understand in more detail the reason for the increase of entropy
and its relation to reversibility and irreversibility.14 We can classify systems
according to their size and completeness of specification:
specification: complete incomplete
size: ∼1 ∼ 100 ∼ 1023
laws: dynamical statistical thermodynamical
reversible: yes no no
Dynamical systems have a relatively small number of particles or degrees
of freedom and can be completely specified. For example, we may have 6
degrees of freedom for each particle (x, y, z, px , py , pz ). We can prepare an
individual dynamical system in an initial state and expect that it will behave
according to the dynamical laws that describe it. Think of billiard balls or
pucks on a frictionless surface, or electrons moving through an electric or
magnetic field. So far as we know, the laws of physics at this level (either
classical or quantum) are reversible.
If there are a large number of particles with many degrees of freedom
(several orders of magnitude), then it is impractical to specify the system
completely. Moreover, small errors in the initial state will have a larger
effect, due to complex interaction of the particles. Also, there are small
effects from interaction with the environment. Therefore we must resort
to statistical laws and we have a statistical system. If we can’t manage all
the degrees of freedom, then we average over those that we can’t manage.
Statistical laws don’t tell us how an individual system will behave (there are
too many sources of variability), but they tell us how ensembles of similar
systems (or preparations) behave. We can talk about the average behavior
of such systems, but we also have to consider the variance, because unlikely
outcomes are not impossible. For example, tossing 10 coins has a probability
of 1/1024 of turning up all heads; this is small, but not negligible. Statistical
laws are in general irreversible (because there are many ways to get to the
same state).
14
This section is based primarily on Edward Fredkin and Tommaso Toffoli’s “Conser-
vative logic” (Fredkin & Toffoli, 1982).
Finally we consider thermodynamical systems. Macroscopic systems have

a very large number of particles (∼ 1023 ) and a correspondingly large num-
ber of degrees of freedom. We call these “Avogadro scale” numbers. It is
important to grasp how truly enormous these numbers are; in comparison
(Tong, 2012, p. 37): the number of grains of sand on all beaches ≈ 1018 , the
number of stars in our galaxy ≈ 1011 , and the number of stars in the visible
universe ≈ 1022 , but the number of water molecules in a cup of tea ≈ 1023 .
Obviously such systems cannot be completely specified (we cannot describe
the initial state and trajectory of every atom). Indeed, because information
is physical, it is physically impossible to “know” (i.e., to represent physically)
the physical state of a macroscopic system (i.e., to use the macrostates of
one system to represent the microstates of another macroscopic system).
We can derive statistical laws for thermodynamical systems, but in these
cases most macrostates become so improbable that they are virtually im-
possible (for example, the cream unmixing from your coffee). The central
limit theorem shows that the variance decreases with n: By the law of large
numbers (specifically, Bernoulli’s Theorem), variance in the relative number
of successes is σ 2 = p(1 − p)/n, that is, σ 2 = 1/4n for p = 1/2. There-
fore the standard deviation is σ = 1/(2n1/2 ). For example, for n = 1022 ,
σ = 5×10−10 . Also, 99.99% of the probability density is within 4σ = 2×10−9
(see Thomas, Intro. Appl. Prob. & Rand. Proc., p. 111). The probability
of deviating more than from the mean decreases exponentially with n:
1
6
exp(−2 n/2) + 21 exp(−22 n/3).
In the thermodynamic limit, the likely is inevitable, and the unlikely is
impossible. In these cases, thermodynamical laws describe the virtually de-
terministic (but irreversible) dynamics of the system.
Sometimes in a macroscopic system we can separate a small number of
mechanical modes (DoF) from the thermal modes. “Mechanical” here in-
cludes “electric, magnetic, chemical, etc. degrees of freedom.” The mechani-
cal modes are strongly coupled to each other but weakly coupled to the ther-
mal modes. For example, in a rigid body (e.g., a bullet, a billiard ball) the
mechanical modes are the positions and momentums of the particles in the
body. Thus the mechanical modes can be treated exactly or approximately
independently of the thermal modes. In the ideal case the mechanical modes
are completely decoupled from the thermal modes, and so the mechanical
modes can be treated as an isolated (and reversible) dynamical system. The
energy of the mechanical modes (once initialized) is independent of the en-
ergy (∼ kT ) of the thermal modes. The mechanical modes are conservative;
Figure II.6: Complementary relation of damping and fluctuations. The

macroscopic ball has few mechanical degrees of (e.g., 6), but during the
collision there is an interaction of the enormously many degrees of freedom
of the microstates of the ball and those of the wall.
they don’t dissipate any energy. (This is what we have with elastic collisions.)
This is the approach of reversible computing.
Suppose we want irreversible mechanical modes, e.g., for implementing
irreversible logic. The underlying physics is reversible, and so the informa-
tion lost by the mechanical modes cannot simply disappear; it must be trans-
ferred to the thermal modes. This is damping: Information in the mechanical
modes, where it is accessible and usable, is transferred to the thermal modes,
where it is inaccessible and unusable. This is the thermalization of informa-
tion, the transfer of physical information from accessible DoF to inaccessible
DoF. But the interaction is bidirectional, so noise (uncontrolled DoF) will
flow from the thermal modes back to the mechanical modes, making the
system nondeterministic.
As Feynman said, “If we know where the damping comes from, it turns
out that that is also the source of the fluctuations” [Feynman, 1963]. Think
of a bullet ricocheting off a flexible wall filled with sand. It dissipates energy
into the sand and also acquires noise in its trajectory (see Fig. II.6). To avoid
nondeterminacy, the information may be encoded redundantly so that the
noise can be filtered out. For example, the signal can be encoded in multiple
mechanical modes, over which we take a majority vote or an average. Or
the signal can be encoded with energy much greater than any one of the
thermal modes, E kT , to bias the energy flow from mechanical to ther-
mal (preferring dissipation over noise). Free energy must be used to refresh
the mechanical modes, and heat must be flushed from the thermal modes.
“[I]mperfect knowledge of the dynamical laws leads to uncertainties in the
behavior of a system comparable to those arising from imperfect knowledge

of its initial conditions. . . Thus, the same regenerative processes which help
overcome thermal noise also permit reliable operation in spite of substantial
fabrication tolerances.” (Fredkin & Toffoli, 1982)
Damped mechanisms have proved to be very successful, but they are
inherently inefficient.
In a damped circuit, the rate of heat generation is proportional

to the number of computing elements, and thus approximately
to the useful volume; on the other hand, the rate of heat re-
moval is only proportional to the free surface of the circuit. As a
consequence, computing circuits using damped mechanisms can
grow arbitrarily large in two dimensions only, thus precluding the
much tighter packing that would be possible in three dimensions.
(Fredkin & Toffoli, 1982)
In an extreme case (force of impact > binding forces), a signal’s interact-

ing with the environment might cause it to lose its coherence (the correlation
of its constituent DoFs, such as the correlation between the positions and
momenta of its particles). The information implicit in the mechanical modes
is lost into the thermal modes.
C. REVERSIBLE COMPUTING 47
C Reversible computing
C.1 Reversible computing as a solution
Notice that the key quantity FE in Eqn. II.1 depends on the energy dissipated
as heat.15 The 100kB T limit depends on the energy in the signal (necessary
to resist thermal fluctuation causing a bit flip). Although most common
computing operations must dissipate a minimum amount of energy (as given
by VNL), there is nothing that says that information processing has to be
done this way. There is no need to dissipate energy, and an arbitrarily large
amount of it can be recovered for future operations (“arbitrary” in the sense
that there is no inherent physical lower bound on the energy that must be
dissipated and cannot be recovered). Accomplishing this becomes a mat-
ter of precise energy management: moving it around in different patterns,
with as little dissipation as possible. Indeed, Esig can be increased to im-
prove reliability, provided we minimize dissipation of energy. This goal can
be accomplished by making the computation logically reversible (i.e., each
successor state has only one predecessor state).
All fundamental physical theories are Hamiltonian dynamical systems,
and all such systems are time-reversible. That is, if ψ(t) is a solution, then
so is ψ(−t). That is, in general, physics is reversible. Therefore, physical
information cannot be lost, but we can lose track of it. This is entropy:
“unknown information residing in the physical state.” Note how this is fun-
damentally a matter of information and knowledge: processes are irreversible
because information becomes inaccessible. Entropy is ignorance.
To avoid dissipation, don’t erase information. The problem is to keep
track of information that would otherwise be dissipated, to avoid squeezing
information out of logical space (IBDF) into thermal space (NIBDF). This is
accomplished by making computation logically reversible (it is already phys-
ically reversible). In effect, computational information is rearranged and
recombined in place. (We will see lots of examples of how to do this.)
C.1.a Information Mechanics
In 1970s, Ed Fredkin, Tommaso Toffoli, and others at MIT formed the Infor-
mation Mechanics group to the study the physics of information. As we will
15
This section is based on Frank (2005b).
see, Fredkin and Toffoli described computation with idealized, perfectly elas-
tic balls reflecting off barriers. The balls have minimum dissipation and are
propelled by (conserved) momentum. The model is unrealistic but illustrates
many ideas of reversible computing. Later we will look at it briefly (Sec. C.7).
They also suggested a more realistic implementation involving “charge pack-
ets bouncing around along inductive paths between capacitors.” Richard
Feynman (Caltech) had been interacting with Information Mechanics group,
and developed “a full quantum model of a serial reversible computer” (Feyn-
man, 1986).
Charles Bennett (1973) (IBM) first showed how any computation could be
embedded in an equivalent reversible computation. Rather than discarding
information (and hence dissipating energy), it keeps it around so it can later
“decompute” it back to its initial state. This was a theoretical proof based on
Turing machines, and did not address the issue of physical implementation.
Bennett (1982) suggested Brownian computers (or Brownian motion ma-
chines) as a possible physical implementation of reversible computation. The
idea is that rather than trying to avoid randomization of kinetic energy
(transfer from mechanical modes to thermal modes), perhaps it can be ex-
ploited. This is an example of respecting the medium in embodied computa-
tion. A Brownian computer makes logical transitions as a result of thermal
agitation. However, because it is operating at thermal equilibrium, it is
about as likely to go backward as forward; it is essentially conducting a ran-
dom walk, and therefore can be expected to take Θ(n2 ) time to advance n
steps from its initial state. A small energy input — a very weak external
driving force — biases the process in the forward direction, so that it pre-
cedes linearly, but still very slowly. This means we will need to look at the
relation between energy and computation speed (Sec. C.1.b). DNA polymer-
ization provides an example. We can compare its energy dissipation, about
40kB T (∼ 1 eV) per nucleotide, with its rate, about a thousand nucleotides
per second.
Bennett (1973) also described a chemical Turing machine, in which the
tape is a large macromolecule analogous to RNA. An added group encodes
the state and head location, and for each transition rule there is a hypothet-
ical enzyme that catalyzes the state transition. We will look at molecular
computation in much more detail later in the class.
As in ballistic computing, Brownian computing needs logical reversibility.
With no driving force, it is equally likely to move forward or backward, but
any driving force will ensure forward movement. Brownian computing can
Figure II.7: Different degrees of logical reversibility. [from Bennett (1982)]
accommodate a small degree of irreversibility (see Fig. II.7). In these cases,

there are a few backward detours (e.g., it might be equally likely to go in
one forard path or two backward paths), but the process can still be biased
forward. For forward computation on such a tree “the dissipation per step
must exceed kT times the log of the mean number of immediate predeces-
sors to each state” (Bennett, 1982, p. 923). Brownian computation cannot
accommodate exponentially backward branching trees, since the computer
will spend much more of its time going backward than going forward, and
since one backward step is likely to lead to even more backward steps. These
undesired states may outnumber desired states by factors of 2100 , requiring
driving forces on the order of 100kT . Why 2100 ? Think of the number of pos-
sible predecessors to a state that does something like x := 0; assigning to a
64-bit variable has 264 possible predecessor states. Consider also the number
of ways of getting to the next statement after a loop. Next we consider the
relation between energy dissipation and computation speed.
C.1.b Energy coefficient

We have seen that a greater driving force can lead to faster computation,
therefore we define an energy coefficient that relates energy dissipation to
computation speed (Frank, 2005b). Let Ediss be the energy dissipated per
operation and fop be the frequency of operations; then the energy coefficient
is defined:
def
cE = Ediss /fop .
For example, for DNA, cE ≈ (40kT )/(1kHz) = 40×26 meV/kHz ≈ 1 eV/kHz

(since at room temperature, kB T ≈ 26 meV: see Sec. A, p. 30). If our goal,
however, is to operate at GHz frequencies (fop ≈ 109 ) and energy dissipation
below kB T (which is below VNL, but possible for reversible logic), then we
need energy coefficients vastly lower than DNA. This is an issue, of course,
for molecular computation.
C.1.c Adiabatic circuits
Since the 1980s, and especially in the 1990s there has been work in adiabatic
circuits. An adiabatic process takes place without input or dissipation of
energy, and adiabatic circuits minimize energy use by obeying certain circuit
design rules. For example: (1) Never turn on a transistor when there is a
voltage potential between the source and drain. (2) Never turn off a transistor
when current is flowing through it. “[A]rbitrary, pipelined, sequential logic
could be implemented in a fully-reversible fashion, limited only by the energy
coefficients and leakage currents of the underlying transistors.” As of 2004,
about cE = 3 meV/kHz was achievable, which is about 250× less than DNA.
“It is difficult to tell for certain, but a wide variety of post-

transistor device technologies have been proposed . . . that have
energy coefficients ranging from 105 to 1012 times lower than
present-day CMOS! This translates to logic circuits that could
run at GHz to THz frequencies, with dissipation per op that is
still less (in some cases orders of magnitude less) than the VNL
bound of kB T ln 2 . . . that applies to all irreversible logic technolo-
gies. Some of these new device ideas have even been prototyped
in laboratory experiments [2001].” (Frank, 2005b, p. 388)
Frank (2005b, p. 388) notes, “fully-reversible processor architectures [1998]

and instruction sets [1999] have been designed and implemented in silicon.”
Reversible circuit design is, however, outside of the scope of this book.
C.2 Foundations of Conservative Computation

If we want to avoid the von Neumann-Landauer limit, then we have to do re-
versible computation (we cannot throw logical information away). Moreover,
if we want to do fast, reliable computation, we need to use driving forces
and signal levels well above this limit, but this energy cannot be dissipated
into the environment. Therefore, we need to investigate a conservative logic
by which energy and other resources are conserved.16 What this means is
that the mechanical modes (computation) must be separated from the ther-
mal modes (heat) to minimize damping and fluctuation and the consequent
thermalization of information (recall Sec. B.2).
According to Fredkin & Toffoli (1982), “Computation is based on the stor-
age, transmission, and processing of discrete signals.” They outline several
physical principles implicit in the axioms of conventional dissipative logic:
P1 “The speed of propagation of information is bounded.” That is, there

is no action at a distance.
P2 “The amount of information which can be encoded in the state of a

finite system is bounded.” This is ultimately a consequence of thermo-
dynamics and quantum theory.
P3 “It is possible to construct macroscopic, dissipative physical devices

which perform in a recognizable and reliable way the logical functions
AND, NOT, and FAN-OUT.” This is a simple empirical fact (i.e., we
build these things).
Since only macroscopic systems are irreversible, as we go to the microscopic

level, we need to understand reversible logic. This leads to new physical
principles of computing:
P4 “Identity of transmission and storage.” From a relativistic perspective,

information storage in one reference frame may be information trans-
mission in another. For an example, consider leaving a note on a table
in an airplane. In the reference frame of the airplane, it is information
storage. If the airplane travels from one place to another, then in the
reference plane of the earth it is information transmission.
16
This section is based primarily on Fredkin & Toffoli (1982).
(and at its output at time t + 1) is called the state of the wire at time t.
From the unit wire one obtains by composition more general wires of arbitrary length. Thus, a wire of length
i (i ! 1) represents a space-time signal path whose ends are separated by an interval of i time units. For the
moment we shall not concern ourselves with the specific spatial layout of such a path (cf. constraint P8).
Observe that the unit wire is invertible, conservative (i.e., it conserves in the output the number of 0's and l's
that are present at the input), and is mapped into its inverse by the transformation t -t.
2.4. Conservative-Logic
52
Gates; The Fredkin Gate. Having introduced a primitive whose role
CHAPTER II. PHYSICS OF COMPUTATION
is to represent signals, we now need primitives to represent in a stylized way physical computing events.
Figure II.8: SymbolFigure 1. The

for unit unit(Fredkin
wire. wire. & Toffoli, 1982)
A conservative-logic gate is any Boolean function that is invertible and conservative (cf. Assumptions P5
and P7 above). It is well known that, under the ordinary rules of function composition (where fan-out is
allowed), the P5
two-input NAND gate This
“Reversibility.” constitutes a universal
is because primitivephysics
microscopic for the is
setreversible.
of all Boolean functions. In
There-
conservative logic, an analogous role is played by a single signal-processing primitive, namely, the Fredkin
fore,
gate, defined by the our computational primitives will need to be invertible.
table
u x1 x2 v y1 y2
000 00
P6 “One-to-one composition.” 0
Physically, fan-out is not trivial (even in
0 0 1 0 1 0
conventional logic), so we cannot assume that one function output can
0 1 0number
be substituted for any 0 0 1 of input variables. Copying a signal
0 1 1 0 1 1 (2)
can be complicated1(and,
00 100
in some cases, impossible, as in quantum
computing). We have 1 0 1 1 0 1 fan-out as a specific signal-processing
to treat
element. 110 110
111 111
and graphically represented as in Figure 2a. This computing element can be visualized as a device that
P7 “Conservation of additive quantities.” It can be shown that in a re-
performs conditional crossover of two data signals according to the value of a control signal (Figure 2b).
When this value isversible systems
1 the two there follow
data signals are a parallel
numberpaths;
of independent
when 0, theyconserved
cross over.quanti-
Observe that the
ties, and in many systems they are additive
Fredkin gate is nonlinear and coincides with its own inverse. over the subsystems, which
makes them more useful. Emmy Noether (1882–1935) proved a famous
theorem: that any symmetry has a corresponding conservation law,
and vice versa; that is, there is a one-to-one correspondence between
physical invariances and conserved quantities. In particular, time in-
variance corresponds to conservation of energy, translational invariance
corresponds to conservation of linear momentum, and rotational invari-
ance corresponds to conservation of angular momentum. Conservative
Figure
logic has to 2. (a) Symbol
obey at leastand
one(b)additive
operationconservation
of the Fredkinlaw.
gate.
In conservative logic, all signal processing is ultimately reduced to conditional routing of signals. Roughly
speaking, signals are treated
P8 “The as unalterable
topology objects that
of space-time is can be moved
locally around in“Intuitively,
Euclidean. the course of athe
computation
but never created amount
or destroyed. For the physical significance of this approach, see Section
of ‘room’ available as one moves away from a certain point 6.
in space increases
2.5. Conservative-Logic as a Finally,
Circuits. power (rather
we shall than as ana scheme
introduce exponential) of the
for connecting signals,
distance
represented by unit from
wires, with thatrepresented
events, point, thus severely limitinggates.
by conservative-logic the connectivity of a
circuit.”
We will see that two primitive operations are sufficient for conservative logic:
the unit wire and the Fredkin gate.
c c 0 0 1 1
a a' a a a b
b b' b b b a
(a) (b)
Figure II.9: Fredkin gate or CSWAP (conditional swap): (a) symbol and (b)
operation.
C.2.a Unit wire

The basic operation of information storage/transmission is the unit wire,
which moves one bit of information between two space-time points separated
by one unit of time (Fig. II.8). The input value at time t, which is considered
the wire’s state at time t, becomes the output value at time t + 1. The unit
wire is reversible and conservative (since it conserves the number of 0s and 1s
in its input). (Note that there are mathematically reversible functions that
are not conservative, e.g., Not.)
C.2.b Fredkin gate

A conservative logic gate is a Boolean function that is both invertible and
conservative (preserves the number of 0s and 1s). Since the number of 1s and
0s is conserved, conservative computing is essentially conditional rerouting,
that is, the initial supply of 0s and 1s is rearranged. Conventional models
of computation are based on rewriting (e.g., Turing machines, the lambda
calculus, register machines, term-rewriting systems, Post and Markov pro-
ductions), but we have seen that overwriting dissipates energy (and is thus
non-conservative). In conservative logic we rearrange bits without creating
or destroying them. There is no infinite “bit supply” and no “bit bucket.”
In the context of the physics of computation, these are physically real, not
metaphors!
A swap is the simplest operation on two bits, and the Fredkin gate, which
is a conditional swap operation (also called CSWAP), is an example of a
conservative logic operation on three bits. It is defined:
(0, a, b) 7→ (0, a, b),

c c c c c c a a'
a a' a a' a a' b b'
b b' b b' b b' c c
Figure
A conservative-logic circuit is a II.10:
directedAlternative
graph whosenotations for Fredkin gate.gates and whose arcs are
nodes are conservative-logic
wires of any length (cf. Figure 3).
Figure 3. (a) closed and (b) open conservative-logic circuits.

Figure II.11: “(a) closed and (b) open conservative-logic circuits.” (Fredkin
Any output&ofToffoli,
a gate can be connected only to the input of a wire, and sirnilarly any input of a gate only to
1982)
the output of a wire. The interpretation of such a circuit in terms of conventional sequential computation is
immediate, as the gate plays the role of an "instantaneous" combinational element and the wire that of a
delay element embedded in an interconnection line. In a closed conservative-logic circuit, all inputs and
outputs of any elements are connected within (1, a,the 7→ (1,
b) circuit b, a).3a). Such a circuit corresponds to what in
(Figure
physics is called a a closed (or isolated) system. An open conservative-logic circuit possesses a number of
Theand
external input first input
output is a(Figure
ports control3b).signal and the
In isolation, other
such twomight
a circuit are data or controlled
be thought of as a transducer
signals. Here, 1 signals a swap, but Fredkin’s original definition
(typically, with memory) which, depending on its initial state, will respond with a particular used 0 to Output
signal a swap. See Fig. II.9 and Fig. II.12(a) for the operation of the Fredkin
sequence to any particular input sequence. However, usually such a circuit will be thought of as a portion of
a larger circuit; thence
gate; Fig. the shows
II.10 notation for input notations.
alternative and output Check
ports (Figure
that the3b),Fredkin
which gate
is suggestive
is of,
respectively, the trailing and the leading edge of a wire. Observe that in
reversible and conservative. As we will see, the Fredkin gate is a universal conservative-logic circuits the
number of Boolean
output ports always equals
primitive that of input ones.
for conservative logic.
The junction between two adjacent unit wires can be formally treated as a node consisting of a trivial
conservative-logic gate, namely, the identity gate. Inwhat follows, whenever we speak of the realizability of
C.3
a function in Conservative
terms of logic circuits
a certain set of conservative-logic primitives, the unit wire and the identity gate will be
tacitly assumed to be included in this set.
A conservative-logic circuit is a directed graph constructed from conservative
logic gatescircuit
A conservative-logic connected by wires (see
is a time-discrete Fig. II.11
dynamical for examples).
system. We can
The unit wires think the
represent of system’s
individual the
stategates
variables, while the gates (including, of course, any occurrence
as instantaneous and the unit wire as being a unit delay, of which we of the identity gate)
collectivelycan
represent the system’s transition function. The number N of unit
make a sequence (or imagine intervening identity gates). A closed circuitwires that are present in the
circuit mayis be thought of as the number of degrees of freedom of the system.
a closed (or isolated) physical system, whereas an open circuit has external Of these N wires, at any
moment N1 will be in state 1, and the remaining N0 (= N - N1) will be in state 0. The quantity N1 is an
inputs
additive function of andtheoutputs.
system’s The
state,number of outputs
i.e., is defined must
for any equalofthe
portion thenumber of inputs,
circuit and its value for the
or is
whole circuit thethecircuit
sum ofwill not be reversible.
the individual Anfrom
contributions openallcircuit may
portions. be partsince
Moreover, of aboth
larger
the unit wire
and the gates return at their outputs as many l’s as are present at their inputs, the quantity N1 is an integral of
the motion of the system, i.e., is constant along any trajectory. (Analogous considerations apply to the
quantity N0, but, of course, N0 and N1 are not independent integrals of the motion.) It is from this
"conservation principle" for the quantities in which signals are encoded that conservative logic derives its
name.
It must be noted that reversibility (in the sense of mathematical invertibility) and conservation are
independent properties, that is, there exist computing circuits that are reversible but not "bit-conserving,"
(Toffoli, 1980) and vice versa (Kinoshita, 1976).
0
u u a
a a' = ūa+ub b
b b' = ua+ūb ab
āb a
(a) (b)
Figure II.12: (a) Logical behavior of Fredkin gate. (b) Implementation of

AND gate by Fredkin gate by constraining one input to 0 and discarding two
“garbage” outputs.
conservative circuit, or connected to the environment. A conservative-logic

circuit is a discrete-time dynamical system, that is, it computes in discrete
steps in each of which bits move through the gates and unit wires. The
number N of unit wires in the circuit is its number of degrees of freedom
(specifically, IBDF). The numbers of 0s and 1s at any time is conserved,
N = N0 + N1 .
C.4 Universality
Fig. II.12(b) illustrates how the Fredkin gate can be used to implement AND;
other conventional gates, such as NOT, OR, and FAN-OUT can be imple-
mented similarly. (You will show this in Exercises II.4 to II.5). Notice that
the implementation of AND requires that we provide a constant 0 on the
second data line, and the Fredkin gate produces two “garbage bits” whose
values (ab and a) we might not need. Fig. II.14 shows a more complicated
example, a 1-line to 4-line demultiplexer. Depending on the value of the
address bits A1 A0 it will direct the input X to Y0 , Y1 , Y2 , or Y3 . The circuit
requires three constant 0 inputs and produces two garbage bits in addition
the the desired outputs.
As a consequence, you can convert conventional logic circuits (constructed
from AND, OR, NOT, etc.) into conservative circuits, but the process is not
very efficient, because of the need for many ancillary (constant input) and
garbage bits. It’s better to design the conservative circuit from scratch. Nev-
Figure 4. Behavior of the Fredkin gate (a) with unconstrained inputs, and (b) with x2 constrained to the
value 0, thus realizing the AND function.
Figure 5.Figure
Realization by !using source
II.13:of f“Realization of fand
bysink. The function
φ using source and: (c, x)
sink.(y, The
g) is chosen so that, for a
function
particular value of c, y = f(x).
φ : (c, x) 7→ (y, g) is chosen so that, for a particular value of c, y = f (x).”
(Fredkin & Toffoli, 1982)
Terminology: source, sink, constants, garbage. Given any finite function , one obtains a new function f
"embedded" in it by assigning specified values to certain distinguished input lines (collectively called the
source) and disregarding certain distinguished output lines (collectively called the sink). The remaining
input lines will constitute the argument, and the remaining output lines, the result. This construction (Figure
5) is called a realization of f by means of !using source and sink. In realizing f by means of , the source
lines will be fed with constant values, i.e., with values that do not depend on the argument. On the other
hand, the sink lines in general will yield values that depend on the argument, and thus cannot be used as
input constants for a new computation. Such values will be termed garbage. (Much as in ordinary life, this
0 0 0
A0
A1 Y3
X Y2
Y1
Y0
A1 A0
Figure II.14: 1-line-to 4-line demultiplexer. The address bits A1 A0 =

00, 01, 10, 11 direct the data bit X into Y0 , Y1 , Y2 or Y3 , respectively. Note
that each Fredkin gate uses an address bit to route X into either of two wires.
(Adapted from circuit in Fredkin & Toffoli (1982).)
ertheless, this shows that any conventional sequential circuit can be converted
into a conservative logic circuit, provided there is a source for constants and
a sink for garbage. As a consequence, the unit wire and Fredkin gate are a
universal set of conservative operators, since they can be used to implement
(for example) AND, NOT, and FAN-OUT, which are universal.
C.5 Constants and garbage

You have seen that the Fredkin gate can be used to compute non-invertible
functions such as AND, if we are willing to provide appropriate constants
(called “ancillary values”) and to accept unwanted outputs. In general, an
irreversible function can be embedded in a reversible by providing appropriate
constants from a source and ignoring some of the outputs, the sink, which are
considered garbage (Fig. II.13). That is, if we want to compute f : x 7→ y, we
provide appropriate constants c so that it can be embedded in a conservative
computation φ : (c, x) 7→ (y, g), which produces the desired output y along
with garbage g. However, this garbage cannot be thrown away (which would
dissipate energy), so it must be recycled in some way.
C.6 Garbageless conservative logic

To reuse the apparatus for a new computation, we would have to throw away
the garbage and provide fresh constants, both of which would dissipate en-
ergy. This is a significant problem if dissipative circuits are naively translated
to conservative circuits because: (1) the amount of garbage tends to increase
with the number of gates, and (2) with the naive translation, the number of
gates tends to increase exponentially with the number of input lines. How-
ever there is a way to make the garbage about the same size as the input,
and thereby limit the dissipated energy.
First observe that a combinational conservative-logic network (one with
no feedback loops) can be composed with its inverse to consume all the
garbage (Fig. II.15). That is, if φ converts (c, x) into (y, g), then φ−1 , its in-
verse, will convert (y, g) back to (c, x). We can always implement φ−1 because
the unit wire and the Fredkin gate are invertible (in fact, their own inverses).
This in itself would not be useful, since in “decomputing” (y, g) back to (c, x)
we have lost the result y of the computation. Therefore, observe that the
desired output can be extracted by a “spy circuit” (Fig. II.16) interposed
on a wire. It works because the Fredkin gate satisfies (a, 0, 1) 7→ (a, a, ā).
-1
computation is "undone" by the inverse network,
Figure 20. (a) Computation of y = f(x) by means of a combinational conservative-logic network . (b) This
58 -1
computation
Figure 21. The network obtained isCHAPTER
"undone"
by combining byII.
and the PHYSICS
-1 inverse OF COMPUTATION
network,
'looks from the outside like a bundle of parallel
wires. The value y(=f(x)) is buried in the middle.
The remarkable achievements of this construction are discussed below with the help of the schematic
representation of Figure 24. In this figure, it will be convenient to visualize the input registers as "magnetic
bulletin boards," in which identical, undestroyable magnetic tokens can be moved on the board surface. A
token at a given position on the board represents a 1, while the absence of a token at that position represents
a 0. The capacity of a board is the maximum number of tokens that can be placed on it. Three such registers
Figure 21. The network
Figure obtained by combining
II.15: box"
Composition and -1 'looks
of combinational from the outside like
conservative-logic a bundle of parallel
are sent through a "black F, which represents the conservative-logic circuit of network
Figure 23,with
and when they
its inverse wires. The value y(=f(x)) is buried in the middle.
reappear some of the to consume
tokens the garbage.
may have been moved, (Fredkin
but none& Toffoli,
taken away1982)
or added. Let us follow this
process, register by register.
The remarkable achievements of this construction are discussed below with the help of the schematic
representation of Figure 24. In this figure, it will be convenient to visualize the input registers as "magnetic
bulletin boards," in which identical, undestroyable magnetic tokens can be moved on the board surface. A
token at a given position on the board represents a 1, while the absence of a token at that position represents
a 0. The capacity of a board is the maximum number of tokens that can be placed on it. Three such registers
are sent through a "black box" F, which represents the conservative-logic circuit of Figure 23, and when they
reappear some of the tokens may have been moved, but none taken away or added. Let us follow this
process, register by register.
Figure 22. Figure
The value a carried
II.16: Theby“spy
an arbitrary
circuit”linefor(a)tapping
can be inspected
into the in aoutput.
nondestructive way by
(Fredkin & the "spy"
device in (b).
Toffoli, 1982)
We use this circuit on the bits of y between the computation φ and the de-
computation φ−1 . Notice that the spy circuit requires two ancillary bits for
Figure 22.each
The value a carried
bit that by an arbitrary
it extracts; line (a)the
it outputs can be inspected
desired in aand
value nondestructive way by the "spy"
its complement
(presumably garbage). device in (b).
We can use these ideas to design a general approach to garbageless compu-
tation (Fig. II.17). The desired computation has m input bits, x1 , x2 , . . . , xm ,
and n output bits y1 , . . . , yn . To do it reversibly requires (we suppose) h con-
stants c1 , . . . , ch and generates (necessarily) h + m − n garbage bits (for the
number of outputs of a reversible computation has to equal the number of
inputs). Extracting the output requires the provision of 2n new constants
and generates the n output bits and their n complements (which can be con-
sidered garbage). Initializing the machine for a new computation requires
putting in the new input, which will dissipate energy, and restoring the out-
put registers y ȳ to their 00 · · · 0011 · · · 11 state, which also dissipates energy.
Therefore the energy dissipated will be proportional to the size of the input
and output (specifically, m + n). The ancillary constants are automatically
restored by decomputation.
Consider the more schematic diagram in Fig. II.18. Think of arranging
tokens (representing 1-bits) in the input registers, both to represent the input
Figure 23. A "garbageless" circuit for computing the function y = f(x). Inputs C ,..., C and X ,…., X are
Figure 23. A "garbageless" circuit for computing the function y = f(x). Inputs C11,..., Chhand X11,…., Xmmare
Figure
returned unchanged, while II.17: Garbageless
the constants circuit.
0,...,0 and 1,..., 1 (Fredkin & part
in the lower Toffoli,
of the1982)
circuits are replaced by the
returned unchanged, while the constants 0,...,0 and 1,..., 1 in the lower part of the circuits are replaced by the
result, y1,...., yn and its complement, ¬y1,...., ¬yn
result, y1,...., yn and its complement, ¬y1,...., ¬yn
Figure 24. The conservative-logic scheme for garbageless computation. Three data registers are "shot"
Figure Figure
24. The II.18:
conservative-logic scheme for garbageless
“The conservative-logic schemecomputation. Three data
for garbageless registers are "shot"
computation.
through a conservative-logic black-box F. The register with the argument, x, is returned unchanged; the
throughThree
a conservative-logic black-box F. The register with the argument, x, is returnedFunchanged; the
clean registerdata
on topregisters are ‘shot’
of the figure, through
representing anaappropriate
conservative-logic black-box
supply of input . The
constants, is used as a
clean register
register on top of the figure, representing an appropriate supply of input constants, is used as a
scratchpad during with the argument,
the computation (cf. thex,c is
andreturned
g lines in unchanged;
Figure 23) butthe clean register
is returned on end of
clean at the the
scratchpad during the computation (cf. the c and g lines in Figure 23) but is returned clean at the end of the
top
computation. of the
Finally, figure, representing
the tokens on the registeran atappropriate
the bottom ofsupply of
the figure input
are constants,
rearranged so as to
computation. Finally, the tokens on the register at the bottom of the figure are rearranged so as to encode the
is encode the
used as a scratchpad during result
theyyand
and its complement
computation ¬y c and g lines in Figure
(cf. ¬y
the
result its complement
[II.17]) but is returned clean at the end of the computation. Finally, the
(a) The "argument" register, containing a given arrangement of tokens x, is returned unchanged. The
(a) The "argument"
tokens on theregister,
registercontaining a given arrangement
at the bottom of are
of the figure tokens x, is returned
rearranged unchanged.
so as to The
capacity of this register is m, i.e., the number of bits in x.
capacity of this register is m, i.e., the number of bits in x.
encode the result y and its complement ¬y” (Fredkin & Toffoli, 1982)
(b) A clean "scratchpad register" with a capacity of h tokens is supplied, and will be returned clean.
(b) A clean "scratchpad register" with a capacity of h tokens is supplied, and will be returned clean.
(This is the main supply of constants-namely, c , . . . , c in Figure 23.) Note that a clean register means one
(This is the main supply of constants-namely, c11, . . . , chhin Figure 23.) Note that a clean register means one
with all 0's (i.e., no tokens), while we used both 0's and l's as constants, as needed, in the construction of
with all 0's (i.e., no tokens), while we used both 0's and l's as constants, as needed, in the construction of
Figure 10. However, a proof due to N. Margolus shows that all 0's can be used in this register without loss of
Figure 10. However, a proof due to N. Margolus shows that all 0's can be used in this register without loss of
generality. In other words, the essential function of this register is to provide the computation with spare
generality. In other words, the essential function of this register is to provide the computation with spare
room rather than tokens.
room rather than tokens.
(c) Finally, we supply a clean "result" register of capacity 2n (where n is the number of bits in y). For
(c) Finally, we supply a clean "result" register of capacity 2n (where n is the number of bits in y). For
this register, clean means that the top half is empty and the bottom half completely filled with tokens. The
this register, clean means that the top half is empty and the bottom half completely filled with tokens. The
x, but also to provide a supply of n of them in the black lower square. Next,
run the computation (including both the forward and backward passes). The
backward pass restores the input argument tokens to their initial positions.
The 2n-bit string 00 · · · 0011 · · · 11 in the lower register has been rearranged
to yield the result and its complement, y ȳ. Restoring the 0 · · · 01 · · · 1 inputs
for another computation dissipates energy, but the amount of energy depends
on the size of the output (number of bits), not the amount of computation.17
C.7 Ballistic computation

“Consider a spherical cow moving in a vacuum. . . ”
To illustrate how conservative computation could dissipate arbitrarily small

amounts of energy, Fredkin and Toffoli developed an idealized model of dis-
sipationless ballistic computation, often called billiard ball computation. It
is based on the same assumptions as the classical kinetic theory of gasses:
perfectly elastic spheres and surfaces. In this case we can think of pucks on
frictionless table.
Fig. II.19 shows the general structure of a billiard ball computer. 1-bits
are represented by the presence of a ball at a location, and 0-bits by its
absence. Input is provided by simultaneously firing balls into the input ports
for the 1s in the argument. Inside the box the balls ricochet off each other and
off of fixed reflectors, which performs the computation. After a fixed time
delay, the balls emerging (or not) from the output ports define the output.
Obviously the number of 1s (balls) is conserved, and the computation is
reversible because the laws of motion are reversible.
Since in this idealized model collisions are perfectly elastic, and there is no
friction, no energy is dissipated, and the tiniest initial velocities are sufficient
for a computation of arbitrary length. Therefore, there is no lower bound
on the energy required for the computation. (Computation will go faster, of
course, if the balls are shot in at higher velocity.) Since the laws of classi-
cal dynamics are reversible, the computation will be reversible, assuming of
course that billiard balls can be made to compute at all!
In fact, they can, and Fig. II.20 shows the realization of the computational
primitive, the interaction gate. If balls representing 1-bits are shot in from
the left at p and q, then the balls emerging (or not) on the right will represent
17
Finite loops can be unrolled, which shows that they can be done without dissipation.
(Cf. also that billiard balls can circulate in a frictionless system.)
Figure II.19: Overall structure of ballistic computer. (Bennett, 1982)
Figure
Figure II.20: 14 Billiard
“Billiard ball ball model
model realizationofofthe
realization theinteraction
interaction gate.
gate.” (Fred-
kin & Toffoli, 1982)
All of the above requirements are met by introducing, in addition to collisions between two balls, collisions
between a ball and a fixed plane mirror. In this way, one can easily deflect the trajectory of a ball (Figure
15a), shift it sideways (Figure 15b), introduce a delay of an arbitrary number of time steps (Figure 1 Sc), and
guarantee correct signal crossover (Figure 15d). Of course, no special precautions need be taken for trivial
crossover, where the logic or the timing are such that two balls cannot possibly be present at the same
moment at the crossover point (cf. Figure 18 or 12a). Thus, in the billiard ball model a conservative-logic
wire is realized as a potential ball path, as determined by the mirrors.
Note that, since balls have finite diameter, both gates and wires require a certain clearance in order to
function properly. As a consequence, the metric of the space in which the circuit is embedded (here, we are
considering the Euclidean plane) is reflected in certain circuit-layout constraints (cf. P8, Section 2).
Essentially, with polynomial packing (corresponding to the Abelian-group connectivity of Euclidean space)
some wires may have to be made longer than with exponential packing (corresponding to an abstract space
Figure 12.62(a) Balls of radius l/sqrt(2) traveling
CHAPTERon a unit
II.grid. (b) Right-angle
PHYSICS elastic collision between two
OF COMPUTATION
balls.
Figure 13. (a) The interaction gate and (b) its inverse.
Figure II.21: “(a) The interaction gate and (b) its inverse.” (Fredkin &
Toffoli, 1982)
6.2. The Interaction Gate. The interaction gate is the conservative-logic primitive defined by Figure
13a, which also assigns its graphical representation.7
In the billiard ball model, the interaction gate is realized simply as the potential locus of collision of two
balls. Withfour logical
reference possibilities,
to Figure 14, let p,pqq,bep̄the
q, values
p q̄, andat ap̄certain
q̄ (theinstant
latterofrepresented by no associated
the binary variables
balls emerging from the gate). (Of course, the gate is conservative:
with the two points P, Q, and consider the values-four time steps later in this particular example-of the the
variables associated
number of with the four
balls pointsitA,has
entering B, C,toD.equal
It is clear
thethat these values
number are, inNotice
exiting.) the order shown in the
that
figure, pq,the
¬pq,interaction
p¬q; and pq.gate
In other words, there
is invertible, will beifa we
because ballput
at Ainif on
andthe
onlyright
if there
onewas
of athe
ball at P and
one at Q; similarly, there will be a ball at B if and only if there was a ball at
four possible outputs (11, 10, 01, 00), we will get the corresponding input onQ and none at P; etc.
the left (p q, p̄ q, Timing
6.3. Interconnection; p q̄, p̄ q̄, respectively).
and Crossover;CheckThe
to make sure Owing
Mirror. you seetothis (Ex. and NOT
its AND
II.10).
capabilities, Fig. II.21
the interaction is isa clearly
gate more abstract
a universalsymbol for the (as
logic primitive interaction
explained ingate and 5,
Section itswe assume
the availability of input constants). To verify that these capabilities are retained in the billiard ball model,
inverse.
one must make sureinteraction
The that one cangate
realize the appropriate
is universal interconnections,
because it can compute i.e., both
that one
ANDcanand
suitably route
balls fromNOT.
one collision
However, we must make provisions for arbitrary interconnections in aconsidering
locus to another and maintain proper timing. In particular, since we are
a planar grid, one must provide a way of performing signal crossover.
planar grid. Therefore, we need to implement signal crossover and to control
timing so that balls arrive at the correct time in order to interact. In fact,
it’s only necessary to deal with non-trivial crossover, for trivial crossover
is when two balls cannot possibly be at the same place at the same time.
Fig. II.22 shows mechanisms for realizing nontrivial crossover, delays, and
7
Note thatdirection
the interaction gate has
changes. four that
Notice outputthe
lines“wires”
but onlyinfour
this(rather than 24are
computer ) output states-in other
virtual,
words, therepresented
output variables
by the possible trajectories of the balls, and not physical objects. the same
are constrained. When one considers its inverse (Figure 13b),
constraintsFor
appear
reversible input
on the variables.the
computing, In Fredkin
composinggate
functions
is moreof this kind, one
relevant, andmust
Fig.exercise
II.23 due care
that the constraints are satisfied.
shows its realization in terms of multiple interaction gates. (The “bridge”
indicates non-trivial crossover.) Since the Fredkin gate is universal, any
reversible computation can be implemented on the billiard ball computer.
Of course, the billiard ball computer is an idealized model of computation,
and like other abstract models, such as the Turing machine, it has practical
limitations (Bennett, 1982). For example, minuscule errors of any sort (po-
sition, velocity, alignment) will accumulate rapidly (by about a factor of 2
between a ball and a fixed plane mirror. In this way, one can easily deflect the trajectory of a ball (Figure
15a), shift it sideways (Figure 15b), introduce a delay of an arbitrary number of time steps (Figure 1 Sc), and
guarantee correct signal crossover (Figure 15d). Of course, no special precautions need be taken for trivial
crossover, where the logic or the timing are such that two balls cannot possibly be present at the same
moment at the crossover point (cf. Figure 18 or 12a). Thus, in the billiard ball model a conservative-logic
wire is realized as a potential ball path, as determined by the mirrors.
Note that, since balls have finite diameter, both gates and wires require a certain clearance in order to
C. REVERSIBLE
function properly. COMPUTING
As a consequence, the metric of the space in which the circuit is embedded 63
(here, we are
considering the Euclidean plane) is reflected in certain circuit-layout constraints (cf. P8, Section 2).
Essentially, with polynomial packing (corresponding to the Abelian-group connectivity of Euclidean space)
some wires may have to be made longer than with exponential packing (corresponding to an abstract space
with free-group connectivity) (Toffoli, 1977).
Figure 15. The mirror (indicated by a solid dash) can be used to deflect a ball’s path (a), introduce a
Figure II.22: “The
sideways mirror
shift (b), (indicated
introduce by and
a delay (c), a solid dash)
realize can be
nontrivial used to(d).
crossover deflect
a ball’s path (a), introduce a sideways shift (b), introduce a delay (c), and
realize nontrivial crossover (d).” (Fredkin & Toffoli, 1982)
156 Introduction to computer science
Figure 16. The switch gate and its inverse. Input signal x is routed to one of two output paths depending on
the value of the control signal, C.
!
! !$
" "$
# #$
Figure 3.14. A simple billiard ball computer, with three input bits and three output bits, shown entering on the left
Figure II.23:
and leaving on theRealization
right, respectively.ofThe
the Fredkin
presence gate
or absence of a in terms
billiard of multiple
ball indicates a 1 or a 0, interaction
respectively.
gates.
Empty [NC]
circles illustrate potential paths due to collisions. This particular computer implements the Fredkin classical
reversible logic gate, discussed in the text.
we will ignore the effects of noise on the billiard ball computer, and concentrate on
understanding the essential elements of reversible computation.
The billiard ball computer provides an elegant means for implementing a reversible
universal logic gate known as the Fredkin gate. Indeed, the properties of the Fredkin gate
provide an informative overview of the general principles of reversible logic gates and
circuits. The Fredkin gate has three input bits and three output bits, which we refer to
as a, b, c and a! , b! , c! , respectively. The bit c is a control bit, whose value is not changed
by the action of the Fredkin gate, that is, c! = c. The reason c is called the control bit
is because it controls what happens to the other two bits, a and b. If c is set to 0 then a
and b are left alone, a! = a, b! = b. If c is set to 1, a and b are swapped, a! = b, b! = a.
The explicit truth table for the Fredkin gate is shown in Figure 3.15. It is easy to see
at each collision). Therefore, an initial random error of 1/1015 in position or

velocity (about what would be expected from Heisenberg uncertainty prin-
ciple) would lead to a completely unpredictable trajectory after a few dozen
collisions, leading to a Maxwell distribution of velocities, as in a gas. That is,
errors grow exponentially in the length of a computation. “Even if classical
balls could be shot with perfect accuracy into a perfect apparatus, fluctuat-
ing tidal forces from turbulence in the atmosphere of nearby stars would be
enough to randomize their motion within a few hundred collisions” (Bennett,
1982, p. 910). Various solutions to these problems have been considered, but
they all have limitations. Bennett (1982, p. 911) concludes, “In summary,
although ballistic computation is consistent with the laws of classical and
quantum mechanics, there is no evident way to prevent the signals’ kinetic
energy from spreading into the computer’s other degrees of freedom.” Of
course, signals can be restored, but this introduces dissipation, and we are
back where we began. Nevertheless, ballistic computation, as found in the
billiard ball computer, illustrates some of the principles of reversible comput-
ing that are used in quantum computation, the topic of the next chapter.
D. EXERCISES 65
D Exercises
Exercise II.1 Show that Eq. II.3 (p. 41) can be rearranged to Eq. II.4 (p.
41).
Exercise II.2 Show that the Fredkin gate is reversible.
Exercise II.3 Show that the Fredkin gate implements (a, 0, 1) 7→ (a, a, ā)
(the “spy circuit”).
Exercise II.4 Show how to use a single Fredkin gate to implement each of
the NOT, OR, and FAN-OUT gates. (FAN-OUT copies its input value to
two output wires.) Note! You cannot use FAN-IN or FAN-OUT in your
implementations, only Fredkin gates.
Exercise II.5 Use the Fredkin gate to implement XOR. Minimize the num-
ber of Fredkin gates you use.
Exercise II.6 Give a truth table for a reversible gate that is not conservative
(does not preserve the total number of 1s and 0s). It should have the same
number of outputs as inputs.
Exercise II.7 Give a truth table for a 2-input / 2-output gate that is con-
servative but not reversible.
Exercise II.8 Show for the eight possible inputs that Fig. II.14 is a correct
implementation of a 1-line to 4-line demultiplexer. That is, show in each
of the four cases A1 A0 = 00, 01, 10, 11 the bit X = 0 or 1 gets routed to
Y0 , Y1 , Y2 , Y3 , respectively. You can use a Boolean algebra proof, if you prefer.
Exercise II.9 Show that implementation of a J-K̄ flip-flop with Fredkin

gates in Fig. II.24 is correct. A J-K̄ flip-flop has the following behavior:
J K̄ behavior
0 0 reset, Q → 0
0 1 hold, Q doesn’t change
1 0 toggle, Q → Q̄
1 1 set, Q → 1
Q
J
K
?
Figure II.24: Implementation of J-K̄ flip-flop. [adapted from Fredkin & Tof-
foli (1982)]
Exercise II.10 Show that the inverse of the interaction gate (Fig. II.20)
works correctly. Hint: It only needs to work correctly for outputs that actu-
ally occur. Therefore, to invert a pq output, balls must be shot into outputs
A and D simultaneously.
Exercise II.11 Use interaction gates (and constant inputs and garbage out-
puts, as necessary) to implement NOT, OR (inclusive or), and XOR (exclu-
sive or).
Exercise II.12 Show that the realization of the Fredkin gate in terms of
interaction gates (Fig. II.23) is correct, by labeling the inputs and outputs
of the interaction gates with Boolean expressions of a, b, and c.
Chapter III
Quantum Computation
A Mathematical preliminaries
“[I]nformation is physical, and surprising physical theories such as quantum
mechanics may predict surprising information processing abilities.” (Nielsen
& Chuang, 2010, p. 98)
A.1 Complex numbers

If you go to the course webpage, and look under Quantum Computation in
the Topics section, you will see a link to “complex number review [FFC-
ch4].” Depending on how familiar you are with complex numbers, read or
skim it through section 4.4.2.1 (pp. 41–53). This should tell you all you need
to know (and a little more).
A.2 Linear algebra review

A.2.a Dirac bracket notation
Much of the math of quantum computation is just elementary linear algebra,
but the notation is different (and of course there is a physical interpretation).
The Dirac bracket notation will seem peculiar if you are not used to it, but
67
68 CHAPTER III. QUANTUM COMPUTATION
it is elegant and powerful, as are all good notations. Think of it like a new
programming language.
First, the notation |ψi represents an n-dimensional vector, which we
can write in a particular basis as a n × 1 complex column vector, |ψi =
(v1 , . . . , vn )T . We pronounce |ψi “ket psi” or “psi ket.” Normally the vectors
are finite-dimensional, but they can be infinite-dimensional if thePvectors have
a finite magnitude (their components are square-summable), k |vk |2 < ∞.
The Dirac notation has the advantage that we can use arbitrary names
for vectors, for example:
|excitedi, |zeroi, |onei, | ↑i, | %i, |1i, |101i, |5i, |f (x)i, |1⊗g(1)i.
It may be easier to remember if you notice that it looks kind of like an arrow;
compare |vi and ~v .
The notation hφ| represents a 1×n complex row vector, hφ| = (u1 , . . . , un )
in a particular basis. We pronounce hψ| “bra psi” or “psi bra.” If |ψi =
(v1 , . . . , vn )T , then hψ| = (v1 , . . . , vn ), where vk is the complex conjugate of
vk . Recall that
x + iy = x − iy and reiφ = re−iφ .
We define the adjoint (conjugate transpose, Hermitian transpose) M † of
a matrix M by
(M † )jk = Mkj .
We pronounce it “M dagger,” “M adjoint,” etc. Note that corresponding
bras and kets are adjoints: hψ| = |ψi† and |ψi = hψ|† .
A.2.b Inner product

Suppose |φi = (u1 , . . . , un )T and |ψi = (v1 , . . . , vn )TPin the same basis. Then
the complex inner product of the vectors is defined k uk vk = hφ| |ψi. Thus,
the inner product of two vectors is the conjugate transpose of the first times
the second. This is the convention in physics, which we will follow; math-
ematicians usually put the complex conjugate on the second argument. It
doesn’t make any difference so long as you are consistent. Since the inner
product multiplies a 1 × n matrix by an n × 1 matrix, the result is a 1 × 1
matrix, or scalar. This product of a bra and a ket is usually abbreviated
hφ | ψi = hφ| |ψi, which can be pronounced “φ-bra ket-ψ” or “φ bra-ket ψ.”
The complex inner product satisfies several important properties:
A. MATHEMATICAL PRELIMINARIES 69
1. It is positive definite:
hψ | ψi > 0, if |ψi =
6 0,
hψ | ψi = 0, if |ψi = 0.
2. It has conjugate symmetry: hφ | ψi = hψ | φi.
3. It is linear in its second argument (the one that’s not conjugated):
hφ | cψi = chφ | ψi, for c ∈ C,

hφ | ψ + χi = hφ | ψi + hφ | χi.
Note that conjugate symmetry and linearity in the second argument together
imply that hcφ | ψi = chφ | ψi (antilinearity in the first argument). The
complex inner product is called sesquilinear, which means “one-and-a-half
linear” (in contrast to the inner product of real vectors, which is linear in
both arguments, i.e., bilinear). p
The norm or magnitude of a vector is defined k|ψik = hψ | ψi. That
is, k|ψik2 = |v1 |2 + · · · |vn |2 . A vector is normalized if k|ψik = 1. Note that
normalized vectors fall on the surface of an n-dimensional hypersphere.
A.2.c Bases and Generalized Fourier Series

Vectors |φi and |ψi are orthogonal if hφ | ψi = 0. A set of vectors is or-
thogonal if each vector is orthogonal to all the others. An orthonormal (ON)
set of vectors is an orthogonal set of normalized vectors. A set of vectors
|φ1 i, |φ2 i, . . . spans a vector space if for every vector
P |ψi in the space there
are complex coefficients c1 , c2 , . . . such that |ψi = k ck |φk i. A basis for a
vector space is a linearly independent set of vectors that spans the space.
Equivalently, a basis is a minimal generating set for the space; that is, all
of the vectors in the space can be generated by linear combinations of the
basis vectors. An (orthonormal) basis for a vector space is an (orthonormal)
set of vectors that spans the space. In general, when I write “basis” I mean
“orthonormal basis.” Any vector in the space has a unique representation as
a linear combination of the basis vectors.
A Hilbert space is a complete inner-product space. “Complete” means
that all Cauchy sequences of vectors (or functions) have a limit in the space.
(In a Cauchy sequence, kxm − xn k → 0 as m, n → ∞.) Hilbert spaces may
be finite- or infinite-dimensional. Suppose |1i, |2i, . . . is an ON basis for a

Hilbert space H. Note, that these are just the names of the basis vectors (we
could have used |e1 i, |e2 i, . . . or something similar), and have nothing to do
with the integers 1, 2, etc. Given such a basis, any |ψi in H can be expanded
in a generalized Fourier series:
X
|ψi = ck |ki.
k
The generalized Fourier coefficients ck can be determined by taking inner

products with the corresponding basis vectors:
X X
hk | ψi = hk| cj |ji = cj hk | ji = ck .
j j
Therefore, ck = hk | ψi. Hence,

X X X
|ψi = ck |ki = hk | ψi |ki = |kihk | ψi.
k k k
This is just the vector’s representation inP a particular basis. (Note that this
equation implies the identity matrix I = k |kihk|.)
A linear operator L : H → Ĥ satisfies L(c|φi + d|ψi) = cL(|φi) + dL(|ψi)
for all |φi, |ψi ∈ H and c, d ∈ C. For example, differentiation is a linear
operator.
A linear operator L : H → Ĥ can be represented by a (possibly infinite-
dimensional) matrix relative to bases for H and Ĥ. To see this, suppose |1i,
|2i, . . . is a basis for H and |1̂i, |2̂i, . . . is a basis for Ĥ. Consider |φi = L|ψi
and represent the vectors in these bases by their Fourier coefficients: bj = h̂ |
φi and ck = hk | ψi. Hence |φi is represented by the vector b = (b1 , b2 , . . .)T
and |ψi by the vector c = (c1 , c2 , . . .)T . Apply the linearity of L:
bj = h̂ | φi
= h̂ | L | ψi
!
X
= h̂|L ck |ki
k
!
X
= h̂| ck L|ki
k
X
= h̂ | L | kick .
k
def
Therefore, define the matrix Mjk = h̂ | L | ki and we see b = Mc. For
this reason, an expression of the form h̂ | L | ki is sometimes called a matrix
element of the operator L. Note that the matrix depends on the bases we
choose.
A.2.d Outer product or dyad

We can form the product of a ket and a bra, which is called a dyad or outer
product. Consider first the finite dimensional case. If |φi is an m × 1 column
vector, and |ψi is an n × 1 column vector (so that hψ| is a 1 × n row vector),
then the outer product |φihψ| is an m × n matrix. In most cases of interest
m = n. Since matrix multiplication is associative, (|φihψ|) |χi = |φi hψ | χi.
More generally, we can form outer products of infinite-dimensional vectors
in Hilbert spaces. If |φi ∈ H0 and |ψi ∈ H, then |φihψ| is the linear operator
L : H → H0 defined, for any |χi ∈ H:
L|χi = (|φihψ|)|χi = |φi hψ | χi.
That is, |φihψ| is the operator that returns |φi scaled by the inner product of
|ψi and its argument. To the extent that the inner product hψ | χi measures
the similarity of |ψi and |χi, the result |φi is weighted by this similarity. The
product of a ket and a bra |φihψ| can be pronounced “φ-ket bra-ψ” or “φ
ket-bra ψ,” and abbreviated |φihψ|. This product is also called a dyad.
The special case of |φihφ| in which |φi is normalized is called a projector
onto |φi. This is because |φihφ| |ψi = |φi hφ | ψi, that is, |φi scaled by the
projection
P of |ψi on |φi. More generally, if |η1 i, . . . , |ηm i are orthonormal,
then m k=1 k ihηk | projects into the m-dimensional subspace spanned by these
|η
vectors.
Any linear operator can be represented as a weighted sum of outer prod-
ucts. To see this, suppose L : H → Ĥ, |̂i is a basis for Ĥ, and |ki is a basis
for H. Consider |φi = L|ψi. We know from Sec. A.2.c that
X
h̂ | φi = Mjk ck , where Mjk = h̂ | L | ki, and ck = hk | ψi.
k
Hence,
X
|φi = |̂i h̂ | φi
j
!
X X
= |̂i Mjk hk | ψi
j k
!
X X
= |̂i Mjk hk| |ψi
j k
!
X
= Mjk |̂ihk| |ψi.
jk
Hence, we have a sum-of-outer-products representation of the operator in

terms of its matrix elements:
X
L= Mjk |̂ihk|, where Mjk = h̂ | L | ki.
jk
A.2.e Tensor products

In this section we define the tensor product, which is of central importance in
quantum computation. To see where we are headed, suppose |φi and |ψi are
the quantum states of two objects (e.g., the states of two qubits, or quantum
bits). Then the state of the composite system (e.g., the pair of qubits) will
be represented by the tensor product |φi ⊗ |ψi, which we can think of as a
sort of concatenation or structure formed of |φi and |ψi. If H and H0 are the
Hilbert spaces from which these states are drawn, then the tensor product
space H ⊗ H0 is the space of all possible states of the two-qubit system (i.e.,
all such pairs or structures). We can apply the tensor product to operators
as well: L ⊗ M is the operator that applies, in parallel, L to the first qubit
of the pair and M to the second qubit:
(L ⊗ M )(|φi ⊗ |ψi) = (L|φi) ⊗ (M |ψi).
For vectors, operators, and spaces, we pronounce L⊗M as “L tensor M .” As

we will see, the tensor product is essential to much of the power of quantum
computation. Next we develop these ideas more formally.
Suppose that |ηj i is an ON basis for H and |ηk0 i is an ON basis for H0 .
For every pair of basis vectors, define the tensor product |ηj i ⊗ |ηk0 i as a
sort of couple or pair of the two basis vectors; that is, there is a one-to-one
correspondence between the |ηj i ⊗ |ηk0 i and the pairs in {|η0 i, |η1 i, . . .} ×
{|η00 i, |η10 i, . . .}. Define the tensor product space H ⊗ H0 as the space spanned
0
by all linear combinations of the basis vectors |ηP
j i ⊗ |ηk i. Therefore each
element of H ⊗ H0 is represented by a unique sum jk cjk |ηj i ⊗ |ηk0 i. In order
to make H ⊗ H0 a Hilbert space, we need an inner product:
hφ1 ⊗ φ2 | ψ1 ⊗ ψ2 i = hφ1 | ψ1 i hφ2 | ψ2 i.
That is, we multiply the inner products of the corresponding elements of the
tensor product pairs.
Usually, we are dealing with finite-dimensional spaces, in which case the
tensor products can be defined less abstractly in terms of matrices. Suppose
in a given basis |φi = (u1 , . . . , um )T and |ψi = (v1 , . . . , vn )T , then their tensor
product in that basis can be defined by the Kronecker product:
 
u1 |ψi
 .. 
|φi ⊗ |ψi =  . 
um |ψi
T
= u1 |ψiT , . . . , um |ψiT
= (u1 v1 , . . . , u1 vn , . . . , um v1 . . . , um vn )T .
Note that this is an mn × 1 column vector and that
(|φi ⊗ |ψi)(j−1)n+k = uj vk .
This combinatorial explosion of dimension is what gives quantum computa-

tion its power.
The following abbreviations are frequent: |φψi = |φ, ψi = |φi|ψi = |φi ⊗
|ψi. Note that |φi|ψi can only be a tensor product because it would not be a
legal matrix product. These are some useful properties of the tensor product
(which is bilinear):
(c|φi) ⊗ |ψi = c(|φi ⊗ |ψi) = |φi ⊗ (c|ψi),

(|φi + |ψi) ⊗ |χi = (|φi|χi) + (|ψi|χi),
|φi ⊗ (|ψi + |χi) = (|φi ⊗ |ψi) + (|φi ⊗ |χi).
The tensor product of linear operators is defined
(L ⊗ M ) (|φi ⊗ |ψi) = L|φi ⊗ M |ψi.

P
Using the fact that |ψi = jk cjk |ηj i ⊗ |ηk0 i you can compute (L ⊗ M )|ψi
for an arbitrary |ψi ∈ H ⊗ H0 (exercise). If M is a k × m matrix and N is a
l × n matrix, then their Kronecker product is a kl × mn matrix:
 
M11 N M12 N · · · M1m N
 M21 N M22 N · · · M2m N 
 
M⊗N= .. .. . . .. .
 . . . . 
Mk1 N Mk2 N · · · Mkm N
Matrices, of course, are linear operators and satisfy (M ⊗ N) (|φi ⊗ |ψi) =

M|φi ⊗ N|ψi.
For a vector, operator, or space M , we define the tensor power M ⊗n to
be M tensored with itself n times:
n
z }| {
⊗n
M = M ⊗ M ⊗ ··· ⊗ M .
A.2.f Properies of operators and matrices

Several properties of operators and matrices are important in quantum com-
putation. An operator L : H → H is normal if L† L = LL† . The same applies
to square matrices. That is, normal operators commute with their adjoints.
An operator L : H → H is Hermitian or self-adjoint if L† = L. The same
applies to square matrices. (Hermitian matrices are the complex analogues
of symmetric matrices.) Note that Hermitian operators are normal. It is
easy to see that L is Hermitian if and only if hφ | L | ψi = hψ | L | φi for all
|φi, |ψi (since hψ | L | φi = hφ | L† | ψi = hφ | L | ψi). A normal matrix is
Hermitian if and only if it has real eigenvalues (exercise). This is important
in quantum mechanics, since measurement results are usually assumed to be
real.
An operator U is unitary if U † U = U U † = I. That is, a unitary operator
is invertible and its inverse is its adjoint: if U is unitary, then U −1 = U † .
Obviously every unitary operator is also normal. (A normal matrix is unitary
if and only if its spectrum — the multiset of its eigenvalues — is contained
in the unit circle in the complex plane.)
Unitary operators are like rotations of a complex vector space (analo-
gous to orthogonal operators, which are rotations of a real vector space).
Just as orthogonal transformations preserve the angles between vectors, uni-
tary operators preserve their inner products (which are analogous to angles).
Consider the inner product between U |φi and U |ψi:
(U |φi)† U |ψi = hφ | U † U | ψi = hφ | ψi.
Hence, the inner product of U |φi and U |ψi is the same as the inner product
of |φi and |ψi. Therefore, unitary operators are isometric, that is, they
preserve norms:
kU |ψik2 = hψ | U † U | ψi = hψ | ψi = k|ψik2 .
Unitary operators are important in quantum computation because the evo-

lution of quantum systems is unitary.
A.2.g Spectral decomposition and operator functions (sup-

plementary)
For any normal operator on a finite-dimensional Hilbert space, there is an
ON basis that diagonalizes the operator, and conversely, any diagonalizable
operator is normal. This is called a spectral decomposition of the opera-
tor. The ON basis comprises its set of eigenvectors, which we can write
|0i, |1i, . . . , |ni and the correspondingPeigenvalues λk are the diagonal ele-
n
ments (cf. Sec. A.2.d, p. 72): L = k=1 λk |kihk|. Therefore, a matrix is
normal if and only if it can be diagonalized by a unitary transform (see
A.2.f, above). That is, it is normal if and only if there is a unitary U such
that L = U ΛU † , where Λ = diag(λ1 , . . . λn ). If |0i, |1i, . . . , |ni is the basis,
then U = (|0i, |1i, . . . , |ni) and
 
h0|
 h1| 
 
U † =  ..  .
 . 
hn|
(More generally, this property holds for compact normal operators.)

√
It is often convenient to extend various complex functions (e.g., ln, exp, )
to normal matrices and operators. If f : C → C and L : H → H, then we
define:
n
X
def
f (L) = f (λk )|kihk|,
k=1
P
where L = nk=1 λk |kihk| is the spectral decomposition
√ of L. LTherefore, for
a normal linear operator or matrix L we can write L, ln L, e , etc. This is
not an arbitrary extension, but follows from the power series expansion for
f.
B. BASIC CONCEPTS FROM QUANTUM THEORY 77
B Basic concepts from quantum theory

B.1 Introduction
B.1.a Bases
In quantum mechanics certain physical quantities are quantized, such as the
energy of an electron in an atom. Therefore an atom might be in certain
distinct energy states |groundi, |first excitedi, |second excitedi, . . . . Other
particles might have distinct states such as spin-up |↑i and spin-down |↓i. In
each case these alternative states correspond to orthonormal vectors:
h↑|↓i = 0,
hground | first excitedi = 0,
hground | second excitedi = 0,
hfirst excited | second excitedi = 0.
In general we may express the same state with respect to different bases,
such as vertical or horizontal polarization |→i, |↑i; or orthogonal diagonal
polarizations | %i, | &i.
B.1.b Superpositions of Basis States

One of the unique characteristics of quantum mechanics is that a physical
system can be in a superposition of basis states, for example,
|ψi = c0 |groundi + c1 |first excitedi + c2 |second excitedi,
where the cj are complex numbers, called (probability) amplitudes. With

respect to a given basis, a state |ψi is interchangeable with its vector of
coefficients, c = (c0 , c1 , . . . , cn )T . When the basis is understood, we can use
|ψi as a name for this vector. This ability of a quantum system to be in
many states simultaneously is the foundation of quantum parallelism.
As we will see, when we measure the quantum state
c0 |E0 i + c1 |E1 i + . . . + cn |En i
with respect to the |E0 i, . . . , |En i basis, we will get the result |Ej i with
probability |cj |2 and the state will “collapse” into state |Ej i. Since the prob-
abilities must add to 1, |c0 |2 + |c1 |2 + · · · + |cn |2 = 1, we know k|ψik = 1, that
is, the vector is normalized.
Figure III.1: Probability density of first six hydrogen orbitals. The main
quantum number (n = 1, 2, 3) and the angular momentum quantum number
(` = 0, 1, 2 = s, p, d) are shown. (The magnetic quantum number m = 0 in
these plots.) [fig. from wikipedia commons]
For the purposes of quantum computation, we usually pick two basis

states and use them to represent the bits 1 and 0, for example, |1i = |groundi
and |0i = |excitedi. We call this the computational basis. I’ve picked the
opposite of the “obvious” assignment (|0i = |groundi) just to show that the
assignment is arbitrary (just as for classical bits). Note that |0i =6 0, the
zero element of the vector space, since k|0ik = 1 but k0k = 0. (Thus 0 does
not represent a physical state, since it is not normalized.)
B.2 Postulates of QM
In this section you will learn the four fundamental postulates of quantum
mechanics.1
1
Quotes are from Nielsen & Chuang (2010) unless otherwise specified.
B.2.a Postulate 1: state space

Postulate 1: Associated with any isolated physical system is a state space,
which is a Hilbert space. The state of the system “is completely defined by
its state vector, which is a unit vector in the system’s state space” (Nielsen
& Chuang, 2010). The state vector has to be normalized so that the total
probability is 1; it is equivalent to the probability axiom that states that the
maximum probability (probability of the whole sample space) = 1.
In previous examples, the state vectors have been finite dimensional, but
Hilbert spaces can be infinite dimensional as well. For example, a quantum
system might have an unlimited number of energy P∞levels, |0i, |1i, |2i, . . . . If
the state of the system is a superposition,
P∞ |ψi = k=0 ck |ki, then the squared
2
amplitudes must sum to 1, k=0 |ck | = 1.
A quantum state |ψi is often a wavefunction, which defines the probability
amplitude distribution (actually, the probability density function) of some
continuous quantity. For example, |ψi may define the complex amplitude
ψ(r) associated with each location r in space, and |Ψi may define the complex
amplitude of Ψ(p) associated with each momentum p (see Fig. III.1). Infinite
dimensional Hilbert spaces also include spaces of wavefunctions such as these.
The inner product of wavefunctions is defined:
Z
hφ | ψi = φ(r)ψ(r)dr.
R3
(For this example we are assuming the R domain2 is 3D space.) Wavefunctions

2
are also normalized, 1 = k|ψik = R3 |ψ(r)| dr. For our purposes, finite
dimensional spaces are usually adequate.
In quantum mechanics, global phase has no physical meaning; all that
matters is relative phase. In other words, if you consider all the angles
around the circle, there is no distinguished 0◦ (see Fig. III.2). Likewise, in
a continuous wave (such as a sine wave), there is no distinguished starting
point (see Fig. III.3).
To say all quantum states are normalized is equivalent to saying that their
absolute length has no physical meaning. That is, only their form (shape)
matters, not their absolute size. This is a characteristic of information.
Another way of looking at quantum states is as rays in a projective Hilbert
space. A ray is an equivalence class of nonzero vectors under the relation,
φ∼= ψ iff ∃z 6= 0 ∈ C : φ = zψ, where φ, ψ 6= 0. That is, global magnitude
and phase (r and φ in z = reiφ ) are irrelevant (i.e., have no physical meaning).
Figure III.2: Relative phase vs. global phase. What matters in quantum
mechanics is the relative phase between state vectors (e.g., θ in the figure).
Global phase “has no physical meaning”; i.e., we can choose to put the 0◦
point anywhere we like.
Figure III.3: Relative phase vs. global phase of sine waves. There is no
privileged point from which to start measuring absolute phase, but there is
a definite relative phase between the two waves.
This is another way of expressing the fact that the form is significant, but
not the size. However, it is more convenient to use normalized vectors in
ordinary Hilbert spaces and to ignore global phase.
B.2.b Postulate 2: evolution

Postulate 2: “The evolution of a closed quantum system is described by
a unitary transformation” (Nielsen & Chuang, 2010). Therefore a closed
quantum system evolves by “complex rotation” of a Hilbert space. More
precisely, the state |ψi of the system at time t is related to the state |ψ 0 i of
the system at time t0 by a unitary operator U which depends only on the
times t and t0 ,
|ψ 0 i = U (t, t0 )|ψi = U |ψi.
This postulate describes the evolution of systems that don’t interact with
the rest of the world. That is, the quantum system is a dynamical system of
relatively low dimension, whereas the environment, including any measure-
ment apparatus, is a thermodynamical system (recall Ch. II, Sec. B).
Dynamics (supplementary) The laws of quantum mechanics, like the

laws of classical mechanics, are expressed in differential equations. However,
in quantum computation we usually deal with quantum gates operating in
discrete time, so it is worth mentioning their relation.
The continuous-time evolution of a closed quantum mechanical system is
given by the Schrödinger equation:
d
i~ |ψ(t)i = H|ψ(t)i,
dt
or more compactly, i~|ψ̇i = H|ψi. H is the Hamiltonian of the system
(a fixed Hermitian operator), and ~ is the reduced Planck constant (often
absorbed into H). P
Since H is Hermitian, it has a spectral decomposition, H = E E|EihE|,
where the normalized |Ei are energy eigenstates (or stationary states) with
corresponding energies E. The lowest energy is the ground state energy.
In quantum computing, we are generally interested in the discrete-time
dynamics of quantum systems. Stone’s theorem shows that the solution to
the Schrödinger equation is:
|ψ(t + s)i = e−iHt/~ |ψ(s)i.
def
Therefore define U (t) = exp(−iHt/~); then |ψ(t + s)i = U (t)|ψ(s)i. It turns
out that U is unitary (Exer. III.6). Hence the evolution of a closed quantum
mechanical system from a state |ψi at time t to a state |ψ 0 i at time t0 can
be described by a unitary operator, |ψ 0 i = U |ψi. Conversely, for any unitary
operator U there is a Hermitian K such that U = exp(iK) (Exer. III.7).
B.2.c Postulate 3: quantum measurement

What happens if the system is no longer closed, that is, if it interacts with
the larger environment? In particular, what happens if a quantum system
interacts with a much larger measurement apparatus, the purpose of which
is to translate a microscopic state into a macroscopic, observable effect? For
example, suppose we have a quantum system that can be in two distinct
states, for example, an atom that can be in a ground state |0i and an ex-
cited state |1i. Since they are distinct states, they correspond to orthogonal
vectors, h0 | 1i = 0. Suppose further that we have a measurement appara-
tus that turns on one light if it measures state |0i and a different light if it
measures state |1i. √
Now consider an atom in a quantum state |ψi = 21 |0i + 23 |1i, a superpo-
sition of the states |0i and |1i. When we measure
1 2 1 |ψi in the computational
basis, we will measure |0i with probability 2 = 4 , and we will measure |1i
√ 2

with probability 23 = 34 . After measurement, the system is in the state we
measured (|0i or |1i, respectively); this is the “collapse” of the wavefunction.
We depict the possibilities as follows:
1/4
|ψi −→ |0i,
3/4
|ψi −→ |1i.
Now consider a more complicated example, a quantum system that can be

in three distinct states, say an atom that can be in a ground state |0i or two
excited states, |1i and |2i. Note that h0 | 1i = h1 | 2i = h0 | 2i = 0. Suppose
the quantum system is in state |ψi = √12 |0i+ 12 |1i+ 21 |2i. Further, suppose we
have a measurement apparatus that turns on a light if it measures state |0i
2does not turn it on otherwise. When we measure |ψi, with probability
and
√1
2 = 21 we will measure |0i and after measurement it will collapse to state
2 2
|0i. With probability 21 + 12 = 21 it will not measure state |0i and the
light won’t go on. In this case, it will collapse to state √1 |1i + √1 |2i, which
2 2
we get by renormalizing the state measured:
1
2
+ 21 |2i
|1i 1 1
q = √ |1i + √ |2i.
1 2 + 1 2 2 2
2 2
We can depict the possible outcomes as follows:


1 1 1  −→1/2
|0i
|ψi = √ |0i + |1i + |2i 1/2 .
2 2 2  −→ √1 |1i + √1 |2i
2 2
In other words, we zero out the coefficients of the states we didn’t measure
and renormalize (because quantum states are always normalized). Now we
develop these ideas more formally.
A measurement can be characterized by a set of projectors Pm , for each
possible measurement outcome m. In the first example above, the measure-
ment operators are P1 = |0ih0| and P2 = |1ih1|. In the second example, the
operators are P1 = |0ih0| and P2 = |1ih1|+|2ih2|. In the latter case, P1 projects
the quantum state into the subspace spanned by {|0i}, and P2 projects the
quantum state into the subspace spanned by {|1i, |2i}. These are orthogonal
subspaces of the original space (spanned by {|0i, |1i, |2i}).
Since a measurement must measure some definite state, a projective mea-
surement is a set of projectors P1 , . . . , PN satisfying: (1) They project into
orthogonal subspaces, so for m 6= n we have Pm PP n = 0, the identically zero
operator. (2) They are complete, that is, I = N m=1 Pm , so measurement
always produces a result. Projectors are also idempotent, Pm Pm = Pm , since
if a vector is already projected into the m subspace, projecting it again has
no effect. Finally, projectors are Hermitian (self-adjoint), as we can see:
!†
X X X
Pm† = |ηj ihηj | = (|ηj ihηj |)† = |ηj ihηj | = Pm .
j j j
Now we can state Postulate 3.
Postulate 3: Quantum measurements are described by a complete set of

orthogonal projectors, Pm , for each possible measurement outcome m.
Measurement projects the state into a subspace with a probability given

by the squared magnitude of the projection. Therefore, the probability of
measurement m of state |ψi is given by:
p(m) = kPm |ψik2 = hψ | Pm† Pm | ψi = hψ | Pm Pm | ψi = hψ | Pm | ψi.

(III.1)
This is Born’s Rule, which gives the probability of a measurement outcome.
The measurement probabilities must sum to 1, which we can check:
!
X X X
p(m) = hψ | Pm | ψi = hψ| Pm |ψi = hψ | I | ψi = hψ | ψi = 1.
m m m
P
This follows from the completeness if the projectors, m Pm = I.
For an example, suppose Pm P = |mihm|, and write the quantum state in
the measurement basis: |ψi = m cm |mi. Then the probability p(m) of
measuring m is:
p(m) = hψ | Pm | ψi
= hψ|(|mihm|)|ψi
= hψ | mihm | ψi
= hm | ψihm | ψi
= |hm | ψi|2
= |cm |2 .
P
P if P2m projects into a subspace, Pm = k |kihk|;
More generally, the same holds
the probability is p(m) = k |ck | . Alternatively, we can “zero out” the cj
for the orthogonal subspace, that is, for the |jihj| omitted by Pm . To maintain
a total probability of 1, the normalized state vector after measurement is
P |ψi Pm |ψi
pm = .
p(m) kPm |ψik
B.2.d Postulate 4: composite systems

Postulate 4: “The state space of a composite physical system is the tensor
product of the state spaces of the component physical systems” (Nielsen &
Chuang, 2010). If there are n subsystems, and subsystem j is prepared in
state |ψj i, then the composite system is in state

n
O
|ψ1 i ⊗ |ψ2 i ⊗ · · · ⊗ |ψn i = |ψj i.
j=1
B.3 Wave-particle duality (supplementary)

Some of the capabilities of quantum computation depend on the fact that
microscopic objects behave as both particles and waves. To see why, imagine
performing the double-slit experiment with three different kinds of objects.
Imagine a stream of classical particles impinging on the two slits and
consider the probability of their arriving on a screen. Define Pj (x) to be the
probability of a particle arriving at x with just slit j open, and P12 (x) to
be the probability of a particle arriving at x with both open. We observe
P12 = P1 + P2 , as expected.
Now consider classical waves, such as water waves, passing through the
two slits. The energy I of a water wave depends on the square of its height
H, which may be positive or negative. Hence,
2
I12 = H12 = (H1 + H2 )2 = H12 + 2H1 H2 + H22 = I1 + 2H1 H2 + I2 .
The 2H1 H2 term may be positive or negative, which leads to constructive

and destructive interference.
Finally, consider quantum particles. The probability of observing a par-
ticle is given by the rule for waves. In particular, the probability P is given
by the square of a complex amplitude A:
P12 = |A1 + A2 |2 = A1 A1 + A1 A2 + A2 A1 + A2 A2 ,
= P1 + A1 A2 + A1 A2 + P2 .
Again, the interference terms A1 A2 + A1 A2 can be positive or negative lead-

ing to constructive and destructive interference. How does a particle going
through one slit “know” whether or not the other slit is open?
6 · E. Rieffel and W. Polak
A C

Finally, after filter B is inserted between A and C, a small amount of light will be visible
on the screen, exactly one eighth of the original amount of light.
A B C
Here we have a nonintuitive effect. Classical experience suggests that adding a filter should
only be able toFigure
decreaseIII.4: Fig.offrom
the number Rieffel
photons getting&through.
Polak How
(2000).
can it increase it?
2.1.2 The Explanation. A photon’s polarization state can be modelled by a unit vector
pointing in the appropriate direction. Any arbitrary polarization can be expressed as a
B.4 Superposition
linear combination a|↑" + b|→" of the two basis vectors2 |→" (horizontal polarization) and
|↑" (vertical polarization).
A simpleSinceexperiment demonstrates
we are only interested quantum
in the direction of theeffects that (the
polarization cannotion
not be of explained
“magni-
2
by tude”
classical
is notphysics (seethe
meaningful), Fig. III.4).
state vectorSuppose
will be a unitwe vector,
have three + |b|2 = 1.filters,
i.e., |a|polarizing In
◦
A, general,
B, andthe C,polarization
polarized ofhorizontally,
a photon can be45 expressed
, and as a|↑" + b|→"
vertically, where a and b Place
respectively. are
numbers3 such that |a|2 + |b|2 = 1. Note, the choice of basis for this representa-
thecomplex
horizontal filter A between a strong light source, such as a laser, and
tion is completely arbitrary: any two orthogonal unit vectors will do (e.g. {|$", |%"}).
a screen. The lightpostulate
The measurement intensity is reduced
of quantum mechanics by states
one that
halfany and themeasuring
device light isa hori-
2-
zontally
dimensional system has an associated orthonormal basis with respect to which thei.e.,
polarized. (Note: Since the light source is unpolarized, it has
quantum
all measurement
polarizations, takesthe
place. Measurement
resulting of a state
intensity would transforms
be much the less
state than
into oneoneof half
the if
themeasuring device’sonly
filter allowed associated
exactly basishorizontally
vectors. The probability
polarizedthat the state
light is measured
through, as
as would
basis vector |u" is the square of the norm of the amplitude of the component of the original
be implied by a sieve model of polarization.) Next insert filter C, polarized
state in the direction of the basis vector |u". For example, given a device for measuring
vertically, and the
the polarization intensity
of photons drops tobasis
with associated zero.{|↑",
This
|to"},isthe
notstatesurprising,
ψ = a|↑" + since
b|→" isthe
filters are cross-polarized.
measured as |↑" with probabilityFinally, insert
|a|2 and as |→" withfilter B, polarized
probability diagonally,
|b|2 (see Figure 1). Notebe-
that A
tween different
and C, measuring devices with some
and surprisingly have different associated
light (about 1/8basis, and measurements
intensity) will return!
using these devices will have different outcomes.
This can’t be explained by the sieve model. How can putting As measurements are always made with
in more filters
respect to an orthonormal basis, throughout the rest of this paper all bases will be assumed
increase the light intensity?
to be orthonormal.
Quantum
Furthermore,mechanics
measurement provides a simple
of the quantum explanation
state will change the state of the
to thethis effect;
result of the in
fact, it’s exactly
measurement. Thatwhat
is, if we should expect.
measurement of ψ = a|↑" A+ photon’s
b|→" results polarization state
in |↑", then the statecan
ψ changes to |↑" and a second measurement with respect to
be represented by a unit vector pointing in appropriate direction. Therefore, the same basis will return |↑"
with probability 1. Thus, unless the original state happened
arbitrary polarization can be expressed by a|0i+b|1i for any two basis vectors to be one of the basis vectors,
measurement will change that state, and it is not possible to determine what the original
|0i state
andwas.|1i, where |a|2 + |b|2 = 1.
A polarizing filter measures a state with respect to a basis that includes
2 The notation |→" is explained in section 2.2.
a vector parallel to its polarization and one orthogonal to it. The effect of
3 Imaginary coefficients correspond to circular polarization.
filter A is the projector PA = | →ih→ |. To get the probability amplitude,
def
apply it to |ψi = a|→i + b| ↑i:
p(A) = | h→ |ψ i |2 = |h→ |(a|→i + b|↑i)|2 = |ah→|→i + bh→|↑i|2 = |a|2 .
So with probability |a|2 we get | →i (recall Eqn. III.1, p. 84). So if the
polarizations are randomly distributed from the source, half will get through
|↑>
|diag up>
|→>
|diag down>
Figure III.5: Alternative polarization bases for measuring photons (black =

rectilinear basis, red = diagonal basis). Note | %i = √12 (| ↑i + | →i) and
|→i = √12 (| %i + | &i).
and all of them will be in state |→i. Why one half? Note that a = cos θ,
where θ is the angle between |ψi and |→i, and that
Z 2π
2 1 1
ha i = cos2 θ dθ = .
2π 0 2
When we insert filter C we are measuring with the projector PC = | ↑ih↑ |

and the result is 0, as expected:
p(AC) = |h↑|→i|2 = 0.
Now insert the diagonal filter B between the horizontal and vertical filters
A and C. Filter B measures with respect to the projector {| %i, | &i} basis
(see Fig. III.5). Transmitted light is given by the projector PB = | %ih% |.
To find the result of applying filter B to the horizontally polarized light
emerging from filter A, we must express |→i in the diagonal basis:
1
|→i = √ (| %i + | &i).
2
So if filter B is | %ih% | we get | %i photons passing through filter B with

probability 1/2:
2 2
1 1 1 1
2
p(B) = |h%|→i| = h% | √ (| %i + | &i = √ h%|%i = √ h%|&i = .

2 2 2 2
Hence, the probability of source photons passing though filters A and B is

p(AB) = p(A)p(B) = 1/4.
The effect of filter C, then, is to measure | %i by projecting against |↑i.
Note that
1
| %i = √ (|↑i + |→i).
2
The probability of these photons getting through filter C is
2
1 1 2 1
|h→|%i| = h→| √ (|↑i + |→i) = √ = .
2
2 2 2
Therefore we get |→i with another 1/2 decrease in intensity (so 1/8 overall).
B.5 No-cloning theorem

Copying and erasing are two of the fundamental (blackboard-inspired) op-
erations of conventional computing. However, the No-cloning Theorem of
quantum mechanics states that it is impossible to copy the state of a qubit.
To see this, assume on the contrary that we have a unitary transformation
U that does the copying, so that U (|ψi ⊗ |ci) = |ψi ⊗ |ψi, where |ci is an
arbitrary constant qubit (actually, |ci can be any quantum state). That is,
U |ψci = |ψψi. Next suppose that |ψi = a|0i + b|1i. By the linearity of U :
U |ψi|ci = U (a|0i + b|1i)|ci

= U (a|0i|ci + b|1i|ci) distrib. of tensor prod.
= U (a|0ci + b|1ci)
= a(U |0ci) + b(U |1ci) linearity
= a|00i + b|11i copying property.
On the other hand, by expanding |ψψi we have:
U |ψci = |ψψi
= (a|0i + b|1i) ⊗ (a|0i + b|1i)
= a2 |00i + ba|10i + ab|01i + b2 |11i.
Note that these two expansions cannot be made equal in general, so no such
unitary transformation exists. Cloning is possible only in the special cases
a = 0, b = 1 or a = 1, b = 0, that is, only where we know that we are
cloning a determinate (classical) basis state. The inability to simply copy
a quantum state is one of the characteristics of quantum computation that
makes it significantly different from classical computation.
B.6 Entanglement
B.6.a Entangled and decomposable states
The possibility of entangled quantum states is one of the most remarkable
characteristics distinguishing quantum from classical systems. Suppose that
H0 and H00 are the state spaces of two quantum systems. Then H = H0 ⊗ H00
is the state space of the composite system (Postulate 4). For simplicity,
suppose that both spaces have the basis {|0i, |1i}. Then H0 ⊗ H00 has the
basis {|00i, |01i, |10i, |11i}. (Recall that |01i = |0i ⊗ |1i, etc.) Arbitrary
elements of H0 ⊗ H00 can be written in the form
X X
cjk |jki = cjk |j 0 i ⊗ |k 00 i.
j,k=0,1 j,k=0,1
Sometimes the state of the composite systems can be written as the tensor
product of the states of the subsystems, |ψi = |ψ 0 i ⊗ |ψ 00 i. Such a state is
called a separable, decomposable or product state. In other cases the state
cannot be decomposed, in which case it is called an entangled state
For an example of an entangled state, consider the Bell state |β01 i, which
might arise from a process that produced two particles with opposite spin
(but without determining which is which):
def 1 def
|β01 i = √ (|01i + |10i) = |Φ+ i. (III.2)
2
(The notations |β01 i and |Φ+ i are both used.) Note that the states |01i and
|10i both have probability 1/2. Such a state might arise, for example, from
a process that emits two particles with opposite spin angular momentum in
order to preserve conservation of spin angular momentum.
To show that |β01 i is entangled, we need to show that it cannot be de-
composed, that is, that we cannot write |β01 i = |ψ 0 i ⊗ |ψ 00 i, for two state
vectors |ψ 0 i = a0 |0i + a1 |1i and |ψ 00 i = b0 |0i + b1 |1i. Let’s try a separation
or decomposition:
?
|β01 i = (a0 |0i + a1 |1i) ⊗ (b0 |0i + b1 |1i).
Multiplying out the RHS yields:
a0 b0 |00i + a0 b1 |01i + a1 b0 |10i + a1 b1 |11i.
Therefore we must have a0 b0 = 0 and a1 b1 √ = 0. But this implies that

either a0 b1 = 0 or a1 b0 = 0 (as opposed to 1/ 2), so the decomposition is
impossible.
For an example of a decomposable state, consider 12 (|00i + |01i + |10i +
|11i). Writing out the product (a0 |0i + a1 |1i) ⊗ (b0 |0i + b1 |1i) as before, we
require a0 b0 = a0 b1 = a1 b0 = a1 b1 = 21 . This is satisfied by a0 = a1 = b0 =
b1 = √12 , therefore the state is decomposable.
In addition to Eq. III.2, the other three Bell states are defined:
def 1 def
|β00 i = √ (|00i + |11i) = |Ψ+ i, (III.3)
2
def 1 def
|β10 i = √ (|00i − |11i) = |Ψ− i, (III.4)
2
def 1 def
|β11 i = √ (|01i − |10i) = |Φ− i. (III.5)
2
The Ψ states have two identical qubits, the Φ states have opposite qubits.
The + superscript indicates they are added, the − that they are subtracted.
The general definition is:
1
|βxy i = √ (|0, yi + (−1)x |1, ¬yi).
2
Remember this useful formula! The Bell states are orthogonal and in fact
constitute a basis for H0 ⊗ H00 (exercise).
B.6.b EPR paradox

The EPR Paradox was proposed by Einstein, Podolsky, and Rosen in 1935 to
show problems in quantum mechanics. Our discussion here will be informal.
Suppose a source produces an entangled EPR pair (or Bell state) |Ψ+ i =
|β00 i = √12 (|00i + |11i), and the entangled particles are sent to Alice and
Bob. If Alice measures her particle and gets |0i, then that collapses the
state to |00i, and so Bob will have to get |0i if he measures his particle.
Likewise, if Alice happens to get |1i, Bob is also required to get |1i if he
measures. This happens instantaneously (but it does not permit faster-than-
light communication, as explained below).
One explanation is that there is some internal state in the particles that
will determine the result of the measurement. Both particles have the same
internal state. Such hidden-variable theories of quantum mechanics assume
that particles are “really” in some definite state and that superposition re-
flects our ignorance of its state. However, they cannot explain the results of
measurements in different bases. In 1964 John Bell showed that any local
hidden variable theory would lead to measurements satisfying a certain in-
equality (Bell’s inequality). Actual experiments, which have been conducted
over tens of kilometers, violate Bell’s inequality. Thus local hidden variable
theories cannot be correct.
Another explanation is that Alice’s measurement affects Bob’s (or vice

versa, if Bob measures first). These are called causal theories. According to
relativity theory, however, in some frames of reference Alice’s measurement
comes first, and in other frames, Bob’s comes first. Therefore there is no con-
sistent cause-effect relation. This is why Alice and Bob cannot use entangled
pairs to communicate.
B.7 Uncertainty principle (supplementary)

You might be surprised that the famous Heisenberg uncertainty principle
is not among the postulates of quantum mechanics. That is because it is
not a postulate, but a theorem, which can be proved from the postulates.
This section is optional, since the uncertainty principle is not required for
quantum computation.
B.7.a Informally
The uncertainty principle states a lower bound on the precision with which
certain pairs of variables, called conjugate variables, can be measured. These
are such pairs as position and momentum, and energy and time. For example,
the same state can be represented by the wave function ψ(x) as a function
of space and by φ(p) as a function of momentum. The most familiar version
of the Heisenberg principle, limits the precision with which location and mo-
mentum can be measured simultaneously: ∆x ∆p ≥ ~/2, where the reduced
Plank constant ~ = h/2π, where h is Planck’s constant.
It is often supposed that the uncertainty principle is a manifestation of
the observer effect, the inevitable effect that measuring a system has on it,
but this is not the case. “While it is true that measurements in quantum
mechanics cause disturbance to the system being measured, this is most em-
phatically not the content of the uncertainty principle.”(Nielsen & Chuang,
2010, p. 89)
Often the uncertainty principle is a result of the variables representing
measurements in two bases that are Fourier transforms of each other. Con-
sider an audio signal ψ(t) and its Fourier transform Ψ(ω) (its spectrum).
Note that ψ is a function of time, with dimension t, and its spectrum Ψ is a
function of frequency, with dimension t−1 . They are reciprocals of each other,
and that is always the case with Fourier transforms. Simultaneous mea-
surement in the time and frequency domains obeys the uncertainty relation
∆t∆ω ≥ 1/2. (For more details on this, including an intuitive explanation,
see MacLennan (prep, ch. 6).)
Time and energy are also conjugate, as a result of the de Broglie relation,
according to which energy is proportional to frequency: E = hν (ν in Hertz,
or cycles per second) or E = ~ω (ω in radians per second). Therefore simul-
taneous measurement in the time and energy domains obeys the uncertainty
principle ∆t∆E ≥ ~/2.
More generally, the observables are represented by Hermitian operators

P, Q that do not commute. That is, to the extent they do not commute,
to that extent you cannot measure them both (because you would have to
do either P Q or QP , but they do not give the same result). The best
interpretation of the uncertainty principle is that if you set up the experiment
multiple times, and measure the outcomes, you will find
2 ∆P ∆Q ≥ |h[P, Q]i|,
where P and Q are conjugate observables. (The commutator [P, Q] is defined

below, Def. B.2, p. 96.)
Note that this is a purely mathematical result (proved in Sec. B.7.b). Any
system obeying the QM postulates will have uncertainty principles for every
pair of non-commuting observables.
B.7.b Formally
In this section we’ll derive the uncertainty principle more formally. Since
it deals with the variances of measurements, we begin with their definition.
To understand the motivation for these definitions, suppose we have a quan-
tum system (such as an atom) that can be in three distinct states |groundi,
|first excitedi, |second excitedi with energies e0 , e1 , e2 , respectively. Then the
energy observable is the operator
E = e0 |groundihground| + e1 |first excitedihfirst excited|

+ e2 |second excitedihsecond excited|,
P
or more briefly, 2m=0 em |mihm|.
Definition B.1 (observable) An observable M is a Hermitian operator on

the state space.
An observable M has a spectral decomposition (Sec. A.2.g):

N
X
M= e m Pm ,
m=1
where the Pm are projectors onto the eigenspaces of M , and the eigenvalues
em are the corresponding measurement results. The projector Pm projects
into the eigenspace corresponding to eigenvalue em . (For projectors, see Sec.

A.2.d.) Since an observable is described by a Hermitian PN operator M , it has
a spectral decomposition with real eigenvalues, M = m=1 em |mihm|, where
|mi is the measurement basis. Therefore we can write M = U EU † , where
E = diag(e1 , e2 , . . . , eN ), U = (|1i, |2i, . . . , |N i), and
 
h1|
 h2| 
† †  
U = (|1i, |2i, . . . , |N i) =  ..  .
 . 
hN |
U † expresses the state in the measurement basis and U translates back. In

the measurement basis, the matrix for an observable is a diagonal matrix:
E = diag(e1 , . . . , eN ). The probability of measuring em is
p(m) = hψ | Pm† Pm | ψi = hψ | Pm Pm | ψi = hψ | Pm | ψi.
We can derive the mean or expectation value of an energy measurement
for a given quantum state |ψi:
def def
hEi = µE = E{E}
X
= em p(m)
m
X
= em hψ | mihm | ψi
m
X
= hψ| em |mihm| |ψi
m
!
X
= hψ| em |mihm| |ψi
m
= hψ | E | ψi.
This formula can be used to derive the standard deviation σE and variance
σE2 , which are important in the uncertainty principle:
def def
σE2 = (∆E)2 = Var{E}
= E{(E − hEi)2 }
= hE 2 i − hEi2
= hψ | E 2 | ψi − (hψ | E | ψi)2 .
Note that E 2 , the matrix E multiplied by Pitself, is also the operator that
2 2
measures the square of the energy, E = j em |mihm|. (This is because E
is diagonal in this basis; alternately, E 2 can be interpreted as an operator
function.)
We now proceed to the derivation of the uncertainty principle.2
Definition B.2 (commutator) If L, M : H → H are linear operators,

then their commutator is defined:
[L, M ] = LM − M L. (III.6)
Remark B.1 In effect, [L, M ] distills out the non-commutative part of the
product of L and M . If the operators commute, then [L, M ] = 0, the iden-
tically zero operator. Constant-valued operators always commute (cL = Lc),
and so [c, L] = 0.
Definition B.3 (anti-commutator) If L, M : H → H are linear opera-

tors, then their anti-commutator is defined:
{L, M } = LM + M L. (III.7)
If {L, M } = 0, we say that L and M anti-commute, LM = −M L.
See B.2.c (p. 82) for the justification of the following definitions.
Definition B.4 (mean of measurement) If M is a Hermitian operator

representing an observable, then the mean value of the measurement of a
state |ψi is
hM i = hψ | M | ψi.
Definition B.5 (variance and standard deviation of measurement)

If M is a Hermitian operator representing an observable, then the variance
in the measurement of a state |ψi is
Var{M } = h(M − hM i2 )i = hM 2 i − hM i2 .
As usual, the standard deviation ∆M of the measurement is defined

p
∆M = Var{M }.
2
The following derivation is from MacLennan (prep, ch. 5).
Proposition B.1 If L and M are Hermitian operators on H and |ψi ∈ H,

then
4hψ | L2 | ψi hψ | M 2 | ψi ≥ |hψ | [L, M ] | ψi|2 + |hψ | {L, M } | ψi|2 .
More briefly, in terms of average measurements,
4hL2 ihM 2 i ≥ |h[L, M ]i|2 + |h{L, M }i|2 .
Proof: Let x + iy = hψ | LM | ψi. Then,
2x = hψ | LM | ψi + (hψ | LM | ψi)∗
= hψ | LM | ψi + hψ | M † L† | ψi
= hψ | LM | ψi + hψ | M L | ψi since L, M are Hermitian
= hψ | {L, M } | ψi.
Likewise,
2iy = hψ | LM | ψi − (hψ | LM | ψi)∗

= hψ | LM | ψi − hψ | M L | ψi
= hψ | [L, M ] | ψi.
Hence,
|hψ | LM | ψi|2 = 4(x2 + y 2 )

= |hψ | [L, M ] | ψi|2 + |hψ | {L, M } | ψi|2 .
Let |λi = L|ψi and |µi = M |ψi. By the Cauchy-Schwarz inequality, k|λik k|µik ≥
|hλ | µi| and so hλ | λi hµ | µi ≥ |hλ | µi|2 . Hence,
hψ | L2 | ψi hψ | M 2 | ψi ≥ |hψ | LM | ψi|2 .
The result follows.

Proposition B.2 Prop. B.1 can be weakened into a more useful form:
4hψ | L2 | ψi hψ | M 2 | ψi ≥ |hψ | [L, M ] | ψi|2 ,
or 4hL2 ihM 2 i ≥ |h[L, M ]i|2

Proposition B.3 (uncertainty principle) If Hermitian operators P and

Q are measurements (observables), then
1
∆P ∆Q ≥ |hψ | [P, Q] | ψi|.
2
That is, ∆P ∆Q ≥ |h[P, Q]i|/2. So the product of the variances is bounded
below by the degree to which the operators do not commute.
Proof: Let L = P − hP i and M = Q − hQi. By Prop. B.2 we have
4 Var{P } Var{Q} = 4hL2 ihM 2 i

≥ |h[L, M ]i|2
= | h[P − hP i, Q − hQi]i |2
= |h[P, Q]i|2 .
Hence,
2 ∆P ∆Q ≥ |h[P, Q]i|

C. QUANTUM INFORMATION 99
C Quantum information
C.1 Qubits
C.1.a Single qubits
Just as the bits 0 and 1 are represented by distinct physical states in a conven-
tional computer, so the quantum bits (or qubits) |0i and |1i are represented
by distinct quantum states. We call |0i and |1i the computational or stan-
dard measurement basis. What distinguishes qubits from classical bits is that
they can be in a superposition of states, a0 |0i + a1 |1i, for a0 , a1 ∈ C, where
|a0 |2 + |a1 |2 = 1. If we measure this state in the computational basis, we will
observe |0i with probability |a0 |2 and likewise for |1i; after measurement the
qubit is in the observed state. This applies, of course, to measurement in
any basis. I will depict the measurement possibilities this way:
|a0 |2
a0 |0i + a1 |1i −→ |0i,
|a1 |2
a0 |0i + a1 |1i −→ |1i.
The following sign basis is often useful:
def 1
|+i = √ (|0i + |1i), (III.8)
2
def 1
|−i = √ (|0i − |1i). (III.9)
2
Notice that |+i is “halfway” between |0i and |1i, and likewise |−i is halfway
between |0i and −|1i. Draw them to be sure you see this. As a consequence
(Exer. III.33):
1
|0i = √ (|+i + |−i),
2
1
|1i = √ (|+i − |−i).
2
To remember this, think (+x) + (−x) = 0 and (+x) − (−x) = (+2x), which
is nonzero (this is just a mnemonic).
original.
3.1 Quantum Key Distribution

Sequences of single qubits can be used to transmit private keys on insecure channels. In
1984 Bennett and Brassard described the first quantum key distribution scheme [Bennett
and Brassard 1987; Bennett et al. 1992]. Classically, public key encryption techniques,
e.g. RSA, are used for key distribution.
Consider the situation in which Alice and Bob want to agree on a secret key so that they
100can communicate privately. TheyCHAPTER
are connected by
III.an ordinary
QUANTUM bi-directional open channel
COMPUTATION
and a uni-directional quantum channel both of which can be observed by Eve, who wishes
to eavesdrop on their conversation. This situation is illustrated in the figure below. The
quantum channel allows Alice to send individual particles (e.g. photons) to Bob who can
measure their quantum state. Eve can attempt to measure the state of these particles and
can resend the particles to Bob.
classical channel
Alice Bob
quantum channel
Eve
To begin the process of establishing a secret key, Alice sends a sequence of bits to Bob
Figure III.6:
by encoding Quantum
each key distribution
bit in the quantum [fromasRieffel
state of a photon follows.& For
Polak
each (2000)].
bit, Alice
randomly uses one of the following two bases for encoding each bit:
0 → |↑#
1 → |→#
Figure III.7: Example if QKD without interference. [fig. from wikipedia]

Figure III.8: Example if QKD with eavesdropping. [fig. from wikipedia]
C.1.b Quantum key distribution

In 1984 Bennett and Brassard showed how sequences of qubits could be used
to distribute an encryption key securely.3 This is called the “BB84 protocol.”
Ironically, the idea was proposed initially by Stephen Wiesner in the 1970s,
but he couldn’t get it published.
We are supposing that Alice is transmitting a key to Bob over two chan-
nels, one classical and one quantum. Eve may eavesdrop on both channels
and even replace the signals in them. Over the quantum channel Alice will
send the photons to Bob that encode the key bits in two different bases, ei-
ther {|↑i, |→i}, which I’ll call the “+ basis,” or {| %i, | &i} (the “× basis”)
(respectively 0, 1 in each basis). Alice chooses randomly the basis in which
to encode her bits (see Fig. III.7). Bob will measure the photons according
to these two bases, also chosen randomly and independently of Alice. After
the transmission, Alice and Bob will communicate over the classical channel
and compare their random choices; where they picked the same basis, they
will keep the bit, otherwise they will discard it. (They will have agreed on
about 50% of the choices.)
Suppose Eve is eavesdropping on the quantum channel, measuring the
qubits and retransmitting them to Bob (see Fig. III.8). About 50% of the
time, she will guess the wrong basis, and will also resend it in this same
incorrect basis. If this is one of the times Alice and Bob chose the same
basis, the bit will nevertheless be incorrect about half of the time (the times
3
This section is based on Rieffel & Polak (2000), which is also the source for otherwise
unattributed quotes.
Eve chose the wrong basis). That is, about 50% of the time Eve picks the
same basis as Alice, so she reads the bit correctly and transmits it to Bob
correctly. About 50% of the time Eve guesses the wrong basis. She will
know this, if she is listening in on the classical channel, but she has already
transmitted it to Bob in the wrong basis. If this is a case in which Alice and
Bob used the same basis (and so Bob should get it correct), he will get it
incorrect 50% of the time, since Eve transmitted it in the other basis. So
25% of the bits that should be correct will be wrong. This high error rate
will be apparent to Alice and Bob if they have been using an error-detecting
code for the key. (In effect Eve is introducing significant, detectable noise
into the channel.) Furthermore, Eve’s version of the key will be about 25%
incorrect. Therefore Bob knows that the key was not transmitted securely
and Eve gets an incorrect key.
This is only the most basic technique, and it has some vulnerabilities, and
so other techniques have been proposed, but they are outside the scope of this
book. “The highest bit rate system currently demonstrated exchanges secure
keys at 1 Mbit/s (over 20 km of optical fibre) and 10 kbit/s (over 100 km
of fibre)”4 “As of March 2007 the longest distance over which quantum key
distribution has been demonstrated using optic fibre is 148.7 km, achieved
by Los Alamos National Laboratory/NIST using the BB84 protocol.” In
Aug. 2015 keys were distributed over a 307 km optical cable, with 12.7 kbps
key generation rate. “The distance record for free space QCD [quantum
key distribution] is 144 km between two of the Canary Islands, achieved
by a European collaboration using entangled photons (the Ekert scheme)
in 2006,[7] and using BB84 enhanced with decoy states[8] in 2007.[9] The
experiments suggest transmission to satellites is possible, due to the lower
atmospheric density at higher altitudes.” At least three companies offer
commercial QKD. “Quantum encryption technology provided by the Swiss
company Id Quantique was used in the Swiss canton (state) of Geneva to
transmit ballot results to the capitol in the national election occurring on
October 21, 2007.” Four QKD networks have been in operation since mid-
late 2000s. Among them,
[t]he world’s first computer network protected by quantum key
distribution was implemented in October 2008, at a scientific
conference in Vienna. The name of this network is SECOQC
(Secure Communication Based on Quantum Cryptography) and
4
https://en.wikipedia.org/wiki/Quantum key distribution (accessed 12-09-18).
EU funded this project. The network used 200 km of standard

fibre optic cable to interconnect six locations across Vienna and
the town of St Poelten located 69 km to the west.
C.1.c Multiple qubits

We can combine multiple qubits into a quantum register. By Postulate 4, if
H is the state space of one qubit, then the tensor power H⊗n will be the state
space of an n-qubit quantum register. The computational basis of this space
def
is the set of all vectors |b1 b2 · · · bn i with bk ∈ 2. (I define 2 = {0, 1} to be
the set of bits, and in general I use a boldface integer N for the set integers
{0, 1, . . . , N − 1}.) Therefore the dimension of the space H⊗n is 2n , and the
n
set of states is the set of normalized vectors in C2 . For 10 qubits we are
dealing with 1024-dimensional complex vectors (because each of the 210 basis
vectors has its own complex amplitude). This is a huge space, exponentially
larger than the 2n classical n-bit strings. This is part of the origin of quantum
parallelism, because we can compute on all of these qubit strings in parallel.
Consider a quantum computer with 500 qubits; it could be very small (e.g.,
500 atoms), but it is computing in a space of 2500 complex numbers. Note
that 2500 is more than the number of particles in the universe times the age
of the universe in femtoseconds! That is, a 500-qubit quantum computer is
equivalent to a universe-sized computer working at high speed since the Big
Bang.
Whereas an ordinary direct product has dimension dim(S ×T ) = dim S +
dim T , a tensor product has dimension dim(S ⊗ T ) = dim S × dim T . Hence
if dim S = 2, dim S ⊗n = 2n .
Measuring some of the qubits in a register causes partial collapse of the
quantum state. Suppose we have a composite state
|ψi = a00 |00i + a01 |01i + a10 |10i + a11 |11i,
and we measure just the first bit. We will get 0 with probability |a00 |2 +|a01 |2
and it will collapse into the state a00 |00i + a01 |01i, but we must renormalize
it:
a00 |00i + a01 |01i
|ψ 0 i = p .
|a00 |2 + |a01 |2
Do this by striking out all terms in |ψi that have 1 in the first qubit.
|a00 |2 +|a01 |2 a00 |00i + a01 |01i
|ψi −→ a00 |00i + a01 |01i ∼
= p .
|a00 |2 + |a01 |2
Quantum computation 21
!
! !"' ! ! $!% "
"
"#$ "%$
! ! "# " !
! &"# "
" "
"&$ "'$
!
"($ ! !$!% " !
"
!
")$ ! !"# " !
"
Figure 1.6. On the left are some standard single and multiple bit gates, while on the right is the prototypical
Figure
multipleIII.9: Left:
qubit gate, classical gates.
the controlled- Right:
. The matrix controlled-Not
representation of the controlled- gate.
, UCN[from Nielsen
, is written with
& respect
Chuang to the(2010,
amplitudesFig.
for |00!, |01!, |10!, and |11!, in that order.
1.6)]
qubit. The action of the gate may be described as follows. If the control qubit is set to
C.2
0, thenQuantum
the target qubit gates
is left alone. If the control qubit is set to 1, then the target qubit
is flipped. In equations:
Quantum gates are analogous to ordinary logic gates (the fundamental build-
ing blocks of circuits), but |01!
|00! → |00!; they
→ must be →
|01!; |10! unitary transformations
|11!; |11! → |10!. (see Fig.
(1.18)
III.9, left, for ordinarty logic gates). Fortunately, Bennett, Fredkin, and
Another
Toffoli wayalready
have of describing the how is
shown allas the
a generalization
usual logic of the classical cangate,
operations be since
done
the action In
reversibly. of the
thisgate may beyou
section summarized
will learnas |A,
the B ⊕ A!, where
→ |A,important
B!most ⊕ is addition
quantum gates.
modulo two, which is exactly what the gate does. That is, the control qubit and the
target qubit are ed and stored in the target qubit.
C.2.a Single-qubit gates
Yet another way of describing the action of the is to give a matrix represen-
Thetation,
NOT as shown
gate isinsimple
the bottom right of
because it Figure 1.6. You can
is reversible: easily verify
NOT|0i = |1i,that the first=
NOT|1i
column of
|0i. Its desired U describes the transformation
CNbehavior can be represented: that occurs to |00!, and similarly for the
other computational basis states, |01!, |10!, and |11!. As for the single qubit case, the
requirement that probability be conserved
NOT : |0i 7→ |1iin the fact that UCN is a unitary
is expressed
†
matrix, that is, UCN UCN = I. |1i 7→ |0i.
We noticed that the can be regarded as a type of generalized- gate. Can
other
Note classical
that gates such
defining it onasathe or the regular
basis defines gate be understood
it on all quantum as unitary
states. Therefore
gatesbe
it can in awritten
sense similar to theof
as a sum way the quantum
dyads gate represents the classical
(outer products):
gate? It turns out that this is not possible. The reason is because the and gates
NOT = |1ih0|For
are essentially irreversible or non-invertible. |0ih1|. given the output A ⊕ B from
+ example,
an gate, it is not possible to determine what the inputs A and B were; there is an
You can read this, “return |1i if the input is |0i, and return
irretrievable loss of information associated with the irreversible T
|0i if the input
action of the gate.
|1i.”
is On the Recall thatunitary
other hand, in the standard
quantum gates basis |0i =
are always (1 0) since
invertible, and the = (0 of1)aT .
|1iinverse
unitary matrix is also a unitary matrix, and thus a quantum gate can always be inverted
by another quantum gate. Understanding how to do classical logic in this reversible or
invertible sense will be a crucial step in understanding how to harness the power of
Therefore NOT can be represented in the standard basis by computing the

outer products:

0 1 0 0 0 1 0 1
NOT = (1 0) + (0 1) = + = .
1 0 1 0 0 0 1 0
The first column represents the result for |0i, which is |1i, and the second
represents the result for |1i, which is |0i.
Although NOT is defined in terms of the computational basis vectors, it
applies to any qubit, in particular to superpositions of |0i and |1i:
NOT(a|0i + b|1i) = aNOT|0i + bNOT|1i = a|1i + b|0i = b|0i + a|1i.
Therefore, NOT exchanges the amplitudes of |0i and |1i.

In quantum mechanics, the NOT transformation is usually called X. It
is one of four useful unitary operations, called the Pauli matrices, which are
worth remembering. In the standard basis:

def def 1 0
I = σ0 = (III.10)
0 1

def def def 0 1
X = σx = σ1 = (III.11)
1 0

def def def 0 i
Y = σy = σ2 = (III.12)
−i 0

def def def 1 0
Z = σz = σ3 = (III.13)
0 −1
We have seen that X is NOT, and I is obviously the identity gate. Z leaves |0i
unchanged and maps |1i to −|1i. It is called the phase-flip operator because
it flips the phase of the |1i component by π relative to the |0i component.
(Recall that global/absolute phase doesn’t matter.) The Pauli matrices span
the space of 2 × 2 complex matrices (Exer. III.20).
Note that Z|+i = |−i and Z|−i = |+i. It is thus the analog in the sign
basis of X (NOT) in the computational basis. What is the effect of Y on the
computational basis vectors? (Exer. III.14)
Note that there is an alternative definition of Y that differs only in global
phase:
def 0 1
Y = .
−1 0
This is a 90◦ = π/2 counterclockwise rotation: Y (a|0i + b|1i) = b|0i − a|1i.

Draw a diagram to make sure you see this.
Note that the Pauli operations apply to any state, not just basis states.
The X, Y , and Z operators get their names from the fact that they reflect
state vectors along the x, y, z axes of the Bloch-sphere representation of a
qubit, which we will not use in this book. Since they are reflections, they are
Hermitian (their own inverses).
C.2.b Multiple-qubit gates

We know that any logic circuit can be built up from NAND gates. Can we
do the same for quantum logic, that is, is there a universal quantum logic
gate? We can’t use NAND, because it’s not reversible, but we will see that
there are universal sets of quantum gates.
The controlled-NOT or CNOT gate has two inputs: the first determines
what it does to the second (negate it or not).
CNOT : |00i 7→ |00i

|01i 7→ |01i
|10i 7→ |11i
|11i 7→ |10i.
Its first argument is called the control and its second is called the target,
controlled, or data qubit. It is a simple example of conditional quantum
computation. CNOT can be translated into a sum-of-dyads representation
(Sec. A.2.d), which can be written in matrix form (Ex. III.23, p. 195):
CNOT = |00ih00|
+ |01ih01|
+ |11ih10|
+ |10ih11|
We can also define it (for x, y ∈ 2), CNOT|xyi = |xzi, where z = x ⊕ y,

the exclusive OR of x and y. That is, CNOT|x, yi = |x, x ⊕ yi CNOT is
the only non-trivial 2-qubit reversible logic gate. Note that CNOT is unitary
since obviously CNOT = CNOT† (which you can show using its dyadic
representation or its matrix representation, Ex. III.23, p. 195). See the right
t. In general there can be multiple control bits. Some authors use a solid circle
negative control, in which the subject bit is toggled when the control bit is 0.
rly, the controlled-controlled-NOT, which negates the last bit of three if and o
t two are both 1, has the following graphical representation.
!
×
Figure III.10: Diagram for CCNOT or Toffoli gate [fig. from Nielsen &
bit operations
Chuangare graphically
(2010)]. represented
Sometimes the × by appropriately
is replaced by ⊕ because CCNOT|xyzi = labelled boxes
|x, y, xy ⊕ zi.
panel of Fig. III.9 (p. 104) for the matrix and note the diagram notation for
CNOT.
Y
CNOT can be used to produce an entangled state:

1 1 1
2 2
Z
CNOT √ (|0i + |1i) |0i = CNOT √ (|00i+|10i) = √ (|00i+|11i) = |β00 i.
2
Note also that CNOT|x, 0i = |x, xi, that is, FAN-OUT, which would seem
to violate the No-cloning Theorem, but it works as expected only for x ∈ 2.
In general CNOT|ψi|0i = 6 |ψi|ψi (Exer. III.24).
Another useful gate is the three-input/output Toffoli gate or controlled-
controlled-NOT. It negates the third qubit if and only if the first two qubits
are both 1. For x, y, z ∈ 2,
def
CCNOT|1, 1, zi = |1, 1, ¬zi,
def
CCNOT|x, y, zi = |x, y, zi, otherwise.
That is, CCNOT|x, y, zi = |x, y, xy ⊕ zi. All the Boolean operations can be
implemented (reversibly!) by using Toffoli gates (Exer. III.29). For example,
CCNOT|x, y, 0i = |x, y, x ∧ yi. Thus it is a universal gate for quantum logic.
In Jan. 2009 CCNOT was implemented successfully using trapped ions.5
5
Monz, T.; Kim, K.; Hänsel, W.; Riebe, M.; Villar, A. S.; Schindler, P.; Chwalla, M.;
Hennrich, M. et al. (Jan 2009). “Realization of the Quantum Toffoli Gate with Trapped
Ions.” Phys. Rev. Lett. 102 (4): 040501. arXiv:0804.0082.
C.2.c Walsh-Hadamard transformation

def def
Recall that the sign basis is defined |+i = √12 (|0i + |1i) and |−i = √1 (|0i −
2
|1i). The Hadamard transformation or gate is defined:
def
H|0i = |+i, (III.14)
def
H|1i = |−i. (III.15)
def
In sum-of-dyads form: H = |+ih0| + |−ih1|. In matrix form (with respect to
the standard basis):
def 1 1 1
H = √ . (III.16)
2 1 −1
Note that H is self-adjoint, H 2 = I (since H † √
= H). H can be defined also
in terms of the Pauli matrices: H = (X + Z)/ 2 (Exer. III.37).
The H transform can be used to transform the computational basis into
the sign basis and back (Exer. III.36):
H(a|0i + b|1i) = a|+i + b|−i,
H(a|+i + b|−i) = a|0i + b|1i.
Alice and Bob could use this in quantum key distribution.
When applied to a |0i, H generates an (equal-amplitude) superposition of
the two bit-values, H|0i = √12 |0i + √12 |1i. This is a useful way of generating a
superposition of both possible input bits, and the Walsh transform, a tensor
power of H, can be applied to a quantum register to generate a superposition
of all possible register values. Consider the n = 2 case:
H ⊗2 |ψ, φi = (H ⊗ H) (|ψi ⊗ |φi)
= (H|ψi) ⊗ (H|φi)
In particular,
H ⊗2 |00i = (H|0i) ⊗ (H|0i)
= |+i⊗2
⊗2
1
= √ (|0i + |1i)
2
2
1
= √ (|0i + |1i)(|0i + |1i)
2
1
= √ (|00i + |01i + |10i + |11i).
22
Notice that this is an equal superposition of all possible values

√ of the 2-qubit
register. (I wrote the amplitude in a complicated way, 1/ 22 , to help you
see the general case.) In general,
1
H ⊗n |0i⊗n = √ (|0i + |1i)⊗n
2n
n
1 z }| {
= √ (|0i + |1i) ⊗ (|0i + |1i) ⊗ · · · ⊗ (|0i + |1i)
2n
1
= √ (|00 . . . 00i + |00 . . . 01i + · · · + |11 . . . 11i)
2m
1 X
= √ |xi
2n x∈2n
2 −1n
1 X
= √ |xi.
2n x=0
Note that “2n −1” represents a string of n 1-bits, and that 2 = {0, 1}. Hence,
H ⊗n |0i⊗n generates an equal superposition of all the 2n possible values of the
n-qubit register. We often write Wn = H ⊗n for the Walsh transformation.
An linear operation applied to such a superposition state in effect applies
the operation simultaneously to all 2n possible input values. This is expo-
nential quantum parallelism and suggests that quantum computation might
be able to solve exponential problems much more efficiently than classical
computers. To see this, suppose U |xi = |f (x)i. Then:
" n −1
2X
# n −1
2X n −1
2X
1 1 1
U (H ⊗n |0i⊗n ) = U √ |xi = √ U |xi = √ |f (x)i
2n x=0 2n x=0 2n x=0
This is a superposition of the function values f (x) for all of the 2n possible
values of x; it is computed by one pass through the operator U .
sequence of gates has the following sequence of effects on a computational basis state
|a, b!,
|a, b! −→ |a, a ⊕ b!
−→ |a ⊕ (a ⊕ b), a ⊕ b! = |b, a ⊕ b!
−→ |b, (a ⊕ b) ⊕ b! = |b, a! , (1.20)
where all additions are done modulo 2. The effect of the circuit, therefore, is to inter-
change the state of the two qubits.
Figure 1.7. Circuit swapping two qubits, and an equivalent schematic symbol notation for this common and useful
circuit.
Figure III.11: Diagram for swap [from Nielsen & Chuang (2010)].
There are a few features allowed in classical circuits that are not usually present in
quantum circuits. First of all, we don’t allow ‘loops’, that is, feedback from one part of the
C.3quantum Quantum
circuit to another;circuits
we say the circuit is acyclic. Second, classical circuits allow
wires to be ‘joined’ together, an operation known as , with the resulting single wire
Acontaining
quantumthecircuit
bitwise is a sequential series of quantum
of the inputs. Obviously this operation transformations
is not reversible and on a
therefore not
quantum unitary,The
register. so weinputs
don’t allow
are usually in our quantum circuits.
computational Third,
basis (all |0i
the inverse
states
operation,
unless , wherebyQuantum
stated otherwise). several copies of a bit
circuit are produced
diagrams is also not
are drawn with allowed
timeingo-
ingquantum circuits.
from left In fact,with
to right, it turns
theout that quantum
quantum gates mechanics
crossingforbids
one or themore
copying of a
“wires”
qubit, making
(qubits) the
as appropriate. operation
The circuitimpossible! We’ll see
represents an example of
a sequence of this in theopera-
unitary next
section when we attempt to design a circuit
tions on a quantum register rather than physical wires. to copy a qubit.
As we proceed we’ll introduce new quantum gates as needed. It’s convenient to in-
These “circuits” are different in several respects from ordinary sequential
troduce another convention about quantum circuits at this point. This convention is
logic circuits.
illustrated First,
in Figure 1.8.loops
Suppose (feedback) are not
U is any unitary allowed,
matrix acting onbutsomeyou can napply
number of
transforms repeatedly. Second, Fan-In (equivalent to OR)
qubits, so U can be regarded as a quantum gate on those qubits. Then we can define a is not allowed,
controlled-U
since it it notgate which is aornatural
reversible unitary. Fan-Out
extension of the controlled- gate. Suchbecause
is also not allowed, a gate
it has a single
would controlthe
violate qubit, indicated by
No-cloning the line with
Theorem. the black
(N.B.: dot,does
This and nnottarget qubits,
contradict
theindicated by the boxed
universality . If the or
of theUToffoli control qubitgates,
Fredkin is set towhich
0 thenarenothing happens
universal onlyto the
with
target qubits. If the control qubit
respect to logical or classical states.) is set to 1 then the gate U is applied to the target qubits.
The prototypical example of the controlled-U gate is the controlled- gate, which is
Fig. III.9 (right) on page 104 shows the symbol for CNOT and its effect.
a controlled-U gate with U = X, as illustrated in Figure 1.9.
Another important operation is measurement, which we represent by a ‘meter’ symbol,
The swap operation is defined |xyi 7→ |yxi, or explicitly
X
SWAP = |yxihxy|.
x,y∈2
We can put three CNOTs in series to swap two qubits (Exer. III.39). Swap
has a special symbol as shown in Fig. III.11.
In general, any unitary operator U (on any number of qubits) can be
conditionally controlled (see Fig. III.12); this is the quantum analogue of
an if-then statement. If the control bit is 0, this operation does nothing,
otherwise it does U . This is implemented by |0ih0| ⊗ I + |1ih1| ⊗ U . Effectively,
the operators are entangled.
Suppose the control bit is in superposition, |χi = a|0i + b|1i. The effect
C.24 QUANTUM INFORMATION
Introduction and overview 111
Figure 1.8. Controlled-U gate.
Figure III.12: Diagram for controlled-U [from Nielsen & Chuang (2010)].
of the conditional operation is:

Figure 1.9. Two different representations for the controlled- .
(|0ih0| ⊗ I + |1ih1| ⊗ U )|χ, ψi
as shown in= (|0ih1.10.
Figure 0| ⊗AsI+ |1ih1| ⊗ described,
previously U )(a|0i + b|1i)
this ⊗ |ψiconverts a single qubit
operation
state |ψ! = α|0!
= |0i + β|1! into a+probabilistic
h0|(a|0i b|1i) ⊗ I|ψi classical
+ |1ibit M (distinguished
h1|(a|0i + b|1i) ⊗ from
U |ψia qubit by
2
drawing it as=a double-line
a|0i ⊗ |ψiwire), which
+ b|1i ⊗ Uis|ψi0 with probability |α| , or 1 with probability
|β|2 .
= a|0, ψi + b|1, U ψi.
"## " " " " " ! " " ##
## !! ##
The result is a superposition of entangled outputs. Notice that CNOT is a
## !!
#" " " " "! ! " " " ##
#
special case of this construction, a controlled X.

Figure 1.10. Quantum circuit symbol for measurement.
We also have a quantum analogue for an if-then-else construction. If U0
and We
U1 shall
are find
unitary operators,
quantum then aswemodels
circuits useful can make the choice
of all quantum between
processes, them
including
conditional on a control bit as follows:
but not limited to computation, communication, and even quantum noise. Several simple
examples illustrate this below.
|0ih0| ⊗ U0 + |1ih1| ⊗ U1 .
1.3.5 Qubit copying circuit?
For
Theexample,
gate is useful for demonstrating one particularly fundamental property of
quantum information. Consider
CNOTthe = |0i
h0|of⊗copying
task I + |1iha1|classical
⊗ X. bit. This may be(III.17)
done
using a classical gate, which takes in the bit to copy (in some unknown state x)
Inaquantum
and ‘scratchpad’circuit diagrams,
bit initialized theas symbol
to zero, illustratedfor
in the CCNOT
Figure 1.11. Thegate
outputis show
is two in
Fig.
bits,III.10,
both ofor withare•infor
which thetop
sametwo
stateconnections
x. and ⊕ for bottom, suggesting
CCNOT|x, y, zi = |x, y, xy ⊕ zi. Alternately,
Suppose we try to copy a qubit in the unknown state |ψ! put “CCNot”
= a |0! + bin|1!a in
box. Other
the same
manner by may
operations using be
a shown gate.byThe input state
putting of theortwo
a letter qubits in
symbol may
a be
box,written as
for example
“H” for the Hadamard gate. ! "
a |0! + b |1! |0! = a |00! + b |10!, (1.21)
The Hadamard gate can be used to generate Bell states (Exer. III.38):
The function of is to negate the second qubit when the first qubit is 1, and thus
the output is simply a |00! CNOT(H ⊗ we
+ b |11!. Have I)|xyi = |βxycopied
successfully i. (III.18)
|ψ!? That is, have we
created the state |ψ!|ψ!? In the case where |ψ! = |0! or |ψ! = |1! that is indeed what this
circuit does; it is possible to use quantum circuits to copy classical information encoded
as a |0! or a |1!. However, for a general state |ψ! we see that
|ψ!|ψ! = a2 |00! + ab|01! + ab|10! + b2 |11!. (1.22)
26 Introduction and overview
understood via the equations

|0, y! + (−1)x |1, ȳ!
|βxy ! ≡ √ , (1.27)
2
112ȳ is the negation of y.
where CHAPTER III. QUANTUM COMPUTATION
Figure 1.8. Controlled-U gate.
In Out
√
|00! (|00! + |11!)/ 2 ≡ |β00 !
√
|01! (|01! + |10!)/ 2 ≡ |β01 !
√
|10! (|00! − |11!)/ 2 ≡ |β10 !
√
Figure 1.9. Two different representations for the controlled- .
|11! (|01! − |10!)/ 2 ≡ |β11 !
as shown in Figure 1.10. As previously described, this operation converts a single qubit
Figure 1.12. Quantum circuit to create Bell states, and its input–ouput quantum ‘truth table’.
state |ψ! =
Figure III.13: + β|1! into acircuit
α|0! Quantum probabilistic classical bit MBell
for generating (distinguished
states. [from from aNielsen
qubit by &
2
drawing it as a double-line
Chuang (2010, fig. 1.12)] wire), which is 0 with probability |α| , or 1 with probability
|β|2 . 1.3.7 Example: quantum teleportation
We will now apply the techniques of the last few"# " " pages to understand something non-
## " " !" ! " " ###
trivial, surprising, and a lot of fun – quantum teleportation! ## !
###
Quantum teleportation is a
!!
technique for moving quantum states around, even " " " in
" "! " the " " absence of a quantum commu-
## ! #
nications channel linking theFigure sender

1.10.of the quantum
Quantum circuit symbol statefortomeasurement.
the recipient.
Here’s how quantum teleportation
Figure III.14: Symbol for measurement of a quantum works. Alice and Bob met longstate ago but(fromnow Nielsen
live &
far Chuang
apart. While
We shall together they generated an EPR pair, each
find quantum circuits useful as models of all quantum processes, including
(2010)). taking one qubit of the EPR
pair when
but not they separated.
limited Many yearscommunication,
to computation, later, Bob is in and hiding, evenand Alice’s noise.
quantum mission, should
Several simple
she choose
examples to accept
illustrate it, this
is tobelow.
deliver a qubit |ψ! to Bob. She does not know the state of
the qubit, and moreover can only send classical information to Bob. Should Alice accept
theThe circuit for generating
mission? BellQubit
1.3.5 statescopying
(Eq. III.18) circuit?is shown in Fig. III.13.
Intuitively,
TheIt’s also things
gate look
convenient pretty
is useful tobad
for for Alice.
have a symbol
demonstrating She onedoesn’t for know fundamental
quantum
particularly the state
state |ψ! of the of
measurement,
property
qubit quantum information. Consider the task of copying a classical bit. This mayfrom
suchshe ashas to
Fig. send
III.14.to Bob, and the laws of quantum mechanics prevent her be done
determining the state when gate,
using a classical she only
whichhastakes
a single
in the copybitofto|ψ! copyin her possession.
(in some unknown What’sstate x)
worse,andeven if she did know
a ‘scratchpad’ the state |ψ!,
bit initialized describing
to zero, as illustrated it precisely takes1.11.
in Figure an infinite amountis two
The output
C.4
of classical Quantum
information
bits, both of which since
gate
are in|ψ!thetakes
arrays
state x.in a continuous space. So even if she did
samevalues
know |ψ!, it
Suppose wouldwe take forever
try toacopy for Alice
a qubit in to describe
the unknown the state to=Bob. |0!It’s
+ bnot|1! looking
Fig.
goodmanner
III.15 Fortunately
for Alice.
shows quantum
for
circuit for a state 1-bit|ψ! full aadder. inAsthewe same will
by using a gate. The input state of the two qubits may be writtenthe
Alice, quantum teleportation is a way of utilizing as
entangled EPR pair in order to! send |ψ! to"Bob, with only a small overhead of classical any
see (Sec. C.7), it is possible to construct reversible quantum gates for
classically computable function.
communication. a |0! + b |1! In|0!particular
= a |00! + bthe |10!,Fredkin and Toffoli gates
(1.21)
In
areoutline,
universal.the steps of the solution are as follows: Alice interacts the qubit |ψ! with
The
her half offunction
the EPR
Because ofpair, andis then
quantum to negate the second
measures
computation theistwo aqubit qubits
unitary when the
in her first qubititobtaining
possession,
operator, ismust
1, and be thusre-
the output is simply + Have we successfully copied
oneversible.
of four possibleYou know a |00!
classicalthat b
results,|11!.
an 00, 01, 10, andcomputation
irreversible 11. She sends this |ψ!?
f (x) cantobewe
That
x 7→information is, have
em-
Bob. created
Depending the state |ψ!|ψ!? In the message,
case where |ψ!performs = |0! or one |ψ! of= four
|1! that is indeedonwhat this
bedded in aonreversible
Alice’s classical
computation Bob (x, c) 7→ (g(x), operations
f (x)), where c his are suit-
circuit
halfable
of the does;pair.
EPR it is Amazingly,
possible to use quantum
by doing circuits to copy the classical information encoded
as a ancillary
|0! or a |1!. constants
However, and
for a g(x)this
general
he can recover
represents
state |ψ! we thethat
see
original
garbage state |ψ!!Note
qubits. that
The quantum circuit shown in Figure 1.13 gives a more precise description of quantum
throwing away the garbage qubits (dumping them into the environment) will
teleportation. The state to |ψ!|ψ!
be teleported is |ψ!
= a2 |00! = α|0!+β|1!,
+ ab|01! + ab|10!where α and β are unknown(1.22)
+ b2 |11!.
collapseThe
amplitudes. thestate
quantum state
input into the (equivalent
circuit |ψ0 ! isto measurement) by entangling them
in the many degrees of freedom of the environment. Typically these garbage
qubits will be ! = |ψ!|β00 ! with other qubits in the computation,(1.28)
|ψ0entangled collapsing
them as well, and interfering with the computation. Therefore the garbage
stance of this conditional definition:
T = |0!"0| ⊗ I ⊗ I + |1!"1| ⊗ Cnot .
The Toffoli gate T can be used to construct complete set of boolean connectives, as can
be seen from the fact that it can be used to construct the AND and NOT operators in the
following way:
T |1, 1, x! = |1, 1, ¬x!
T |x, y, 0! = |x, y, x ∧ y!
The T gate is sufficient to construct arbitrary combinatorial circuits.

The following quantum circuit, for example, implements a 1 bit full adder using Toffoli
and controlled-NOT gates:
|c! ! ! ! |c!
|x! ! ! ! |x!
|y! ! ! ! |y!
|0! × × × |s!
|0! × × × |c! !
Figure
whereIII.15: Quantum
x and y are circuit
the data bits, forsum
s is their 1-bit full2),adder
(modulo c is the[from
incomingRieffel & and
carry bit, Polak
!
c is the new carry bit. Vedral, Barenco and Ekert [Vedral et al.
(2000)]. “x and y are the data bits, s is their sum (modulo 2), c is the 1996] define more complex
circuits that
incoming include
carry bit,in-place
and c0addition
is the and
new modular
carryaddition.
bit.”
The Fredkin gate is a “controlled swap” and can be defined as
F = |0!"0| ⊗ I ⊗ I + |1!"1| ⊗ S
where S is the swap operation
S = |00!"00| + |01!"10| + |10!"01| + |11!"11|.
The reader can verify that F , like T , is complete for combinatorial circuits.
x
x x
g(x) g(x)
0 0
Φ Φ-1
0 0
f(x) CNOT f(x)
y y⊕f(x)
y y⊕f(x)
Uf
Figure III.16: Quantum gate array for reversible quantum computation.

must be produced in a standard state independent of x. This is accomplished

by uncomputing, as we did in classical reversible computing (Ch. II, Sec. C.6,
p. 57).
Since NOT is reversible, each 1 bit in c can be replaced by a 0 bit followed
by a NOT, so we need only consider computations of the form (x, 0) 7→
(g(x), f (x)); that is, all the constant bits can be zero.
Therefore, we begin by embedding our irreversible computation of f in a
reversible computation Φ, which we get by providing 0 constants and gen-
erating garbage g(x); see Fig. III.16. That is, Φ will perform the following
computation on four registers (data, workspace, result, target):
(x, 0, 0, y) 7→ (x, g(x), f (x), y).
The result f (x) is in the result register and the garbage g(x) is in the
workspace register. Notice that x and y (data and target) are passed through.
Now use CNOTs between corresponding places in the result and target reg-
isters to compute y ⊕ f (x), where ⊕ represents bitwise exclusive-or, in the
target register. Thus we have computed:
(x, 0, 0, y) 7→ (x, g(x), f (x), y ⊕ f (x)).
Now we uncompute with Φ−1 , but since the data and target registers are
passed through, we get (x, 0, 0, y ⊕ f (x)) in the registers. We have restored
the data, workspace, and result registers to their initial values and have
y ⊕ f (x) in the target register. Ignoring the result and workspace registers,
we write
(x, y) 7→ (x, y ⊕ f (x)).
This is the standard approach we will use for embedding a classical compu-
tation in a quantum computation.
Therefore, for any computable f : 2m → 2n , there is a reversible quantum
gate array Uf : Hm+n → Hm+n such that for x ∈ 2m and y ∈ 2n ,
Uf |x, yi = |x, y ⊕ f (x)i,
See Fig. III.17. In particular, Uf |x, 0i = |x, f (x)i. The first m qubits are
called the data register and the last n are called the target register.
gates for any classically computable function. In fact, it is possible to conceive of a univer-
sal quantum Turing machine [Bernstein and Vazirani 1997]. In this construction we must
assume a sufficient supply of bits that correspond to the tape of a Turing machine.
Knowing that an arbitrary classical function f with m input and k output bits can be im-
plemented on quantum computer, we assume the existence of a quantum gatearray Uf that
implements f . Uf is a m + k bit transformation of the form Uf : |x, y! → |x, y ⊕ f (x)!
where ⊕ denotes the bitwise exclusive-OR6. Quantum gate arrays Uf , defined in this way,
are unitary for any function f . To compute f (x) we apply Uf to |x! tensored with k
C. QUANTUM
zores INFORMATION
|x, 0!. Since f (x) ⊕ f (x) = 0 we have Uf Uf = I. Graphically the transformation 115
Uf : |x, y! → |x, y ⊕ f (x)! is depicted as
|x! |x!
Uf
|y! |y ⊕ f (x)!.
While the T and F gates are complete for combinatorial circuits, they cannot achieve ar-
Figurequantum
bitrary III.17:state
Computation
transformations.of function
In order by arbitrary
to realize quantum gate
unitary array (Rieffel
transformations 7
, &
single
Polak,bit2000).
rotations need to be included. Barenco et. al. [Barenco et al. 1995] show that
Cnot together with all 1-bit quantum gates is a universal gate set. It suffices to include the
following one-bit transformations
! " ! iα "
C.5 Quantum parallelism cos α sin α
,
e 0
− sin α cos α 0 e−iα
Since Uf is linear, if it is applied to a superposition of bit strings, it will
for all 0 ≤ α ≤ 2π together with the Cnot to obtain a universal set of gates. As we shall
produce a superposition of the results of applying f to them in parallel (i.e.,
see, such non-classical transformations are crucial for exploiting the power of quantum
in the same time it takes to compute it on one input):
computers.
(c1 |x1 i +Parallelism
UfQuantum
5.2 c2 |x2 i + · · · + ck |xk i) = c1 Uf |x1 i + c2 Uf |x2 i + · · · + ck Uf |xk i.
What happens if U
For example, iffweis applied
have atosuperposition
input which is inof a superposition?
the inputs xThe answer is easy
1 and x2 ,
but powerful: since Uf is a linear transformation,
! it is applied to all basis vectors in the
√
superposition simultaneously and will generate a
√
superposition of the results. In this way,
3 1 3 1
it is possible |x1 if (x)
Uf to compute + for|xn2 ivalues
⊗ |0i
of x= |x1application
in a single , f (x1 )i +of U|xf .2This
, f (xeffect
2 )i. is
2
called quantum parallelism. 2 2 2
The power of quantum algorithms comes from taking advantage of quantum parallelism
Theentanglement.
and amplitude So ofmost
a result
quantumy will be the
algorithms sum
begin of the amplitudes
by computing a function ofof all x
interest such
that
on y = f (x). of all values as follows. Start with an n-qubit state |00 . . . 0!. Apply the
a superposition
If we apply Uf to a superposition of all possible 2m inputs, it will compute
a superposition of all the corresponding outputs in parallel (i.e., in the same
6 ⊕ is not the direct sum of vectors.
7 More precisely, we mean arbitrary unitary transformations up to a constant phase factor. A constant phase shift
time as required for one function evaluation)! The Walsh-Hadamard trans-
of the state has no physical, and therefore no computational, significance.
formation can be used to produce this superposition of all possible inputs:
1
Wm |00 . . . 0i = √ (|00 . . . 0i + |00 . . . 1i + · · · + |11 . . . 1i)
2m
1 X
= √ |xi
2m x∈2m
2 −1 m
1 X
= √ |xi.
2m x=0
In the last line we are obviously interpreting the bit strings as natural num-
bers. Hence,
2m −1
! 2m −1 2m −1
1 X 1 X 1 X
Uf Wm |0i = Uf √ |x, 0i = √ Uf |x, 0i = √ |x, f (x)i.
2m x=0 2m x=0 2m x=0
A single circuit does all 2m computations simultaneously! “Note that since

n qubits enable working simultaneously with 2n states, quantum parallelism
circumvents the time/space trade-off of classical parallelism through its abil-
ity to provide an exponential amount of computational space in a linear
amount of physical space.” (Rieffel & Polak, 2000)
This is amazing, but not immediately useful. If we measure the input
bits, we will get a random value, and the state will be projected into a
superposition of the outputs for the inputs we measured. If we measure an
output bit, we will get a value probabilistically, and a superposition of all
the inputs that can produce the measured output. Neither of the above is
especially useful, so most quantum algorithms transform the state in such a
way that the values of interest have a high probability of being measured.
The other thing we can do is to extract common properties of all values of
f (x). Both of these require different programming techniques than classical
computing.
The key to both dense coding and teleportation is the use of entangled particles. The
initial set up is the same for both processes. Alice and Bob wish to communicate. Each is
sent one of the entangled particles making up an EPR pair,
1
ψ0 = √ (|00" + |11").
2
Say Alice is sent the first particle, and Bob the second. So until a particle is transmit-
ted, only Alice can perform transformations on her particle, and only Bob can perform
C. transformations
QUANTUMonINFORMATION
his. 117
4.2.1 Dense Coding
Alice Bob
Encoder Decoder
EPR
source
Alice. Alice receives two classical bits, encoding the numbers 0 through 3. Depending
FigureAlice
on this number III.18: Superdense
performs one of the coding. (Rieffel
transformations & Y,
{I, X, Polak,
Z} on 2000)
her qubit of the
entangled pair ψ0 . Transforming just one bit of an entangled pair means performing the
identity transformation on the other bit. The resulting state is shown in the table.
C.6 Value
Applications Transformation New state
0 ψ0 = (I ⊗ I)ψ0 √12 (|00" + |11")
C.6.a Superdense 1coding ψ1 = (X ⊗ I)ψ0 √12 (|10" + |01")
2 ψ2 = (Y ⊗ I)ψ0 √12 (−|10" + |01")
We will consider a couple
3 simple applications
ψ3 = (Z of these
⊗ I)ψ0 √12 (|00" ideas.
− |11") The first is called
superdense coding or (more modestly) dense coding, since it is a method by
whichAlice
onethen sends her qubit to Bob.
quantum particle can be used to transmit two classical bits of
Bob. Bob applies
information. It wasa controlled-
described to the
NOTby two qubits
Bennett and of the entangledinpair.
Wiesner 1992, and was
partially validated experimentally by 1998.
Here is the idea. Alice and Bob share an entangled pair of qubits. To
transmit two bits of information, Alice applies one of four transformations
to her qubit. She then sends her qubit to Bob, who can apply an operation
to the entangled pair to determine which of the four transformations she
applied, and hence recover the two bits of information.
Now let’s work it through more carefully. Suppose Alice and Bob share
the entangled pair |β00 i = √12 (|00i + |11i). Since the four Bell states are a
basis for the quantum state of the pair of qubits, Alice’s two bits of infor-
mation can be encoded as one of the four Bell states. For example, Alice
can use the state |βzx i to encode the bits z, x (the correspondence is ar-
bitrary so long as we are consistent, but this one is easy to remember).
Recall the circuit for generating Bell states (Fig. III.13, p. 112). Its effect
is CNOT(H ⊗ I)|zxi = |βzx i. This cannot be used by Alice for generating
the Bell states, because she doesn’t have access to Bob’s qubit. However,
the Bell states differ from each other only in the relative parity and phase
of their component qubits (i.e., whether they have the same or opposite bit
values and the same or opposite signs). Therefore, Alice can alter the par-
ity and phase of just her qubit to transform the entangled pair into any of
the Bell states. In particular, if she uses zx to select I, X, Z, or ZX = Y
(corresponding to zx = 00, 01, 10, 11 respectively) and applies it to just her
qubit, she can generate the corresponding Bell state |βzx i. I’ve picked this
correspondence because of the simple relation between the bits z, x and the
application of the operators Z, X, but this is not necessary; any other 1-1
correspondence between the two bits and the four operators could be used.
When Alice applies this transformation to her qubit, Bob’s qubit is unaf-
fected, and so the transformation on the entangled pair is I ⊗ I, X ⊗ I,
Z ⊗ I, or ZX ⊗ I. We can check the results as follows:
bits transformation result

00 I ⊗I √1 (|00i + |11i) = |β00 i
2
01 X ⊗I √1 (|10i + |01i) = |β01 i
2
10 Z ⊗I √1 (|00i − |11i) = |β10 i
2
11 ZX ⊗ I √1 (−|10i + |01i) = |β11 i
2
For example,hin the second-to-last

i case, since Z|0i = |0i and Z|1i = −|1i,
1 1
we see Z ⊗ I √2 (|00i + |11i) = √2 (|00i − |11i). Make sure you can explain
the results in the other cases (Exer. III.43).
When Alice wants to send her information, she applies the appropriate
operator to her qubit and sends her single transformed qubit to Bob, which
he uses with his qubit to recover the information by measuring the pair
of qubits in the Bell basis. This can be done by inverting the Bell state
generator, which, since the CNOT and H are self-adjoint, is simply:
(H ⊗ I)CNOT|βzx i = |zxi.
This translates the Bell basis into the computational basis, so Bob can mea-
sure the bits exactly.
C.6.b Quantum teleportation

Quantum teleportation is not quite as exciting as it sounds! Its goal is to
transfer the exact quantum state of a particle from Alice to Bob by means
a classical channel (Figs. III.19, III.20). Of course, the No Cloning Theorem
says we cannot copy a quantum state, but we can “teleport” it by destroying
! "
ψ0 √1 (|0! + |1!) √1 √1 (|0! + |1!) + √1 (|0! − |1!) = |0!
2 2! 2 2 "
ψ1 √1 (|1! + |0!) √1 √1 (|0! − |1!) + √1 (|0! + |1!) = |0!
2 2! 2 2 "
ψ2 √1 (−|1! + |0!) √1 − √1 (|0! − |1!) + √1 (|0! + |1!) = |1!
2 2 ! 2 2 "
ψ3 √1 (|0! − |1!) √1 √1 (|0! + |1!) − √1 (|0! − |1!) = |1!
2 2 2 2
Finally, Bob measures the resulting bit which allows him to distinguish between 0 and
3, and 1 and 2.
C. QUANTUM INFORMATION
4.2.2 Teleportation. The objective is to transmit the quantum state of a particle using119
classical bits and reconstruct the exact quantum state at the receiver. Since quantum state
cannot be copied, the quantum state of the given particle will necessarily be destroyed. Sin-
gle bit teleportation has been realized experimentally [Bouwmeester et al. 1997; Nielsen
et al. 1998; Boschi et al. 1998].
Alice Bob
Decoder Encoder
EPR
source
Alice. Alice has a qubit whose state she doesn’t know. She wants to send the state of ths
qubitFigure III.19: Quantum teleportation. (Rieffel & Polak, 2000)
φ = a|0! + b|1!
to Bob through classical channels. As with dense coding, Alice and Bob each possess one
qubit of an entangled pair
1
ψ0 = √ (|00! + |11!).
2
Figure III.20: Possible setup for quantum teleportation. [from wikipedia

commons]
the original and recreating it elsewhere. Single-qubit quantum teleportation

was described by Bennett in 1993 and first demonstrated experimentally in
the late 1990s.
This is how it works. Alice and Bob begin by sharing the halves of an
entangled pair, |β00 i = √12 (|00i + |11i). Suppose that the quantum state that
Alice wants to share is |ψi = a|0i + b|1i. The composite system comprising
the unknown state and the Bell state is
def
|ψ0 i = |ψ, β00 i
1
= (a|0i + b|1i) √ (|00i + |11i)
2
1
= √ [a|0i(|00i + |11i) + b|1i(|00i + |11i)]
2
1
= √ (a|0, 00i + a|0, 11i + b|1, 00i + b|1, 11i).
2
Alice applies the decoding circuit used for superdense coding to the unknown
state and her qubit from the entangled pair. This function is (H ⊗ I)CNOT;
it measures her two qubits in the Bell basis. When Alice applies CNOT to
her two qubits (leaving Bob’s qubit alone) the resulting composite state is:
def
|ψ1 i = (CNOT ⊗ I)|ψ0 i

1
= (CNOT ⊗ I) √ (a|00, 0i + a|01, 1i + b|10, 0i + b|11, 1i)
2
1
= √ (a|00, 0i + a|01, 1i + b|11, 0i + b|10, 1i).
2
Notice that the amplitude a of |ψi has been transferred to the components of
the shared pair having the same parity (|00i and |11i), whereas the amplitude
b has been transferred to the components having the opposite parity (|10i
and |01i). When Alice applies H ⊗ I to her qubits the result is:
def
|ψ2 i = (H ⊗ I ⊗ I)|ψ1 i
1
= (H ⊗ I ⊗ I) √ (a|0, 00i + a|0, 11i + b|1, 10i + b|1, 01i)
2
1
= [a(|0, 00i + |1, 00i + |0, 11i + |1, 11i)
2
+b(|0, 10i − |1, 10i + |0, 01i − |1, 01i)] .
This is because H|0i = |+i = √12 (|0i + |1i) and H|1i = |−i = √12 (|0i − |1i).
Rearranging and factoring to separate Alice’s qubits from Bob’s, we have:
1
|ψ2 i = [|00i(a|0i + b|1i) + |01i(a|1i + b|0i)
2
+|10i(a|0i − b|1i) + |11i(a|1i − b|0i)] .
Thus the unknown amplitudes have been transferred from the first qubit
(Alice’s) to the third (Bob’s), which now incorporates the amplitudes a and
b, but in different ways depending on the first two qubits. In fact you can
see that the amplitudes are transformed by the Pauli matrices, and Bob can
restore the quantum state by applying the correct Pauli matrix. Therefore
Alice measures the first two bits (completing measurement in the Bell basis)
and sends them to Bob over the classical channel. This measurement par-
tially collapses the state, which includes Bob’s qubit, but in a way that is
determined by the first two qubits.
When Bob receives the two classical bits from Alice, he uses them to
select a transformation for his qubit, which restores the amplitudes to the
correct basis vectors. These transformations are the Pauli matrices (which
are their own inverses):
bits gate input

00 I a|0i + b|1i (identity)
01 X a|1i + b|0i (exchange)
10 Z a|0i − b|1i (flip)
11 ZX a|1i − b|0i (exchange–flip)
In each case, applying the specified gate to its input yields |ψi = a|0i + b|1i,
Alice’s original quantum state. This is obvious in the 00 case, but you should
verify the others (Exer. III.44). Notice that since Alice had to measure her
qubits, the original quantum state of her particle has collapsed. Thus it has
been “teleported,” not copied.
The quantum circuit in Fig. III.21 is slightly different from what we’ve
described, since it uses the fact that the appropriate transformations can be
expressed in the form Z M1 X M2 , where M1 and M2 are the two classical bits.
You should verify that ZX = Y (Exer. III.45).
Both superdense coding and teleportation indicate that with an entangled
pair, two bits can be interchanged with one qubit. This is one example of
a method of interchanging resources. However, quantum teleportation does
Quantum computation 27
"## " " " " " ! " " ##

## !! ##
## !! # #
#" " " " " ! ! " " " #
"## " " " " " ! " " ##
## !! ###
## !! #
#" " " " " ! ! " " " #
Figure 1.13. Quantum circuit for teleporting a qubit. The two top lines represent Alice’s system, while the bottom
line is Bob’s system. The meters represent measurement, and the double lines coming out of them carry classical
Figure III.21: Circuit for quantum teleportation. [from Nielsen & Chuang
bits (recall that single lines denote qubits).
(2010)]
1 ! "
= √ α|0"(|00" + |11") + β|1"(|00" + |11") , (1.29)
2
notwhere
allow faster-than-light communication, since Alice has to transmit her
we use the convention that the first two qubits (on the left) belong to Alice, and
twotheclassical bits
third qubit to Bob.
to Bob. As we explained previously, Alice’s second qubit and Bob’s qubit
Entangled states can be teleported
start out in an EPR state. Alice in a through
sends her qubits similaraway. Free-space quantum
gate, obtaining
teleportation has been demonstrated
1 !
over 143 km between" two of the Canary
Islands (Nature, 13 = √ 2012).
|ψ1 "Sept. 6 + |11") + β|1"(|10" + |01") .
α|0"(|00"In Sept. 2015 teleportation was achieved (1.30)
2
over 101 km through supercooled nanowire. For teleporting material systems,
theShe then sends the first qubit through a Hadamard gate, obtaining
current record is 21 m.
1! "
|ψ2 " = α(|0" + |1")(|00" + |11") + β(|0" − |1")(|10" + |01") .
2
C.7 Universal quantum gates (1.31)
WeThis
havestate may
seen be re-written
several in the following
interesting examplesway, simply by regrouping
of quantum computing terms:
using gates
such as CNOT and the1 Hadamard ! # 7
and$ Pauli #operators.$ Since the imple-
|ψ " = |00" α|0" + β|1" + |01" α|1" + β|0"
mentation of each 2of these 2 is a technical challenge, it raises the important
# $ # $"
question: What gates are sufficient
+ |10" for implementing
α|0" − β|1" any quantum
+ |11" α|1" − β|0" . compu-
(1.32)
tation?
This expression naturally breaks down into four terms. The first term has Alice’s qubits
Both the Fredkin (controlled swap) and Toffoli (controlled-controlled-
in the state |00", and Bob’s qubit in the state α|0" + β|1" – which is the original state
NOT) gates are sufficient for classical logic circuits. In fact, they can op-
|ψ". If Alice performs a measurement and obtains the result 00 then Bob’s system will
erate
be inasthewell
stateon
|ψ".qubits in from
Similarly, superposition. But whatweabout
the previous expression can readother quantum
off Bob’s post-
operators?
measurement state, given the result of Alice’s measurement:
It can be proved that single-qubit unitary ! "
operators can be approximated
00 −
$ → |ψ 3 (00)" ≡ α|0" + β|1"
arbitrarily closely by the Hadamard gate and the T (π/8) gate, which (1.33)is
! "
6 01 $−→ |ψ3 (01)" ≡ α|1" + β|0" (1.34)
http://www.nature.com/nature/journal/v489/n7415/full/nature11472.html
! "
(accessed 12-09-18). 10 $−→ |ψ3 (10)" ≡ α|0" − β|1" (1.35)
7
This lecture follows Nielsen & Chuang (2010,! §4.5). "
11 $−→ |ψ3 (11)" ≡ α|1" − β|0" . (1.36)
Depending on Alice’s measurement outcome, Bob’s qubit will end up in one of these
four possible states. Of course, to know which state it is in, Bob must be told the result of
Alice’s measurement – we will show later that it is this fact which prevents teleportation
defined:
1 0 ∼ e−iπ/8 0
T = iπ/4 = iπ/8 (III.19)
0 e 0 e
(ignoring global phase). To approximate within any single-qubit unitary
operation, you need O(logc (1/)) gates, where c ≈ 2. For an m-gate circuit
(of CNOTs and single-qubit unitaries) and an accuracy of , O(m logc (m/)),
where c ≈ 2, gates are needed (Solovay-Kitaev theorem).
A two-level operation is a unitary operator on a d-dimensional Hilbert
space that non-trivially affects only two qubits out of n (where d = 2n ).
It can be proved that any two-level unitary operation can be computed by
a combination of CNOTs and single-qubit operations. This requires O(n2 )
single-qubit and CNOT gates.
It also can be proved that an arbitrary d-dimensional unitary matrix
can be decomposed into a product of two-level unitary matrices. At most
d(d − 1)/2 of them are required. Therefore a unitary operator on an n-qubit
system requires at most 2n−1 (2n − 1) two-level matrices.
In conclusion, the H (Hadamard), CNOT, and π/8 gates are sufficient
for quantum computation. For fault-tolerance, either the standard set —
H (Hadamard), CNOT, π/8, and S (phase) — can be used, or H, CNOT,
Toffoli, and S. The latter phase gate is defined:

2 1 0
S=T = . (III.20)
0 i
Quantum algorithms 33
Figure 1.19. Quantum circuit implementing Deutsch’s algorithm.

Figure III.22: Quantum circuit for Deutsch algorithm. [fig. from Nielsen &
Chuang (2010)]
is sent through two Hadamard gates to give
! "! "
|0! + |1! |0! − |1!
|ψ1 ! = √ √ . (1.42)
D Quantum algorithms 2 2
√
A little thought shows that if we√apply Uf to the state |x!(|0! − |1!)/ 2 then we obtain
D.1 (−1)f (x) |x!(|0! − |1!)/ algorithm
the stateDeutsch-Jozsa 2. Applying Uf to |ψ1 ! therefore leaves us with one of
two possibilities:
D.1.a Deutsch’s 
!
algorithm "!
|0! + |1! |0! − |1!
"

 ± √ √ if f (0) = f (1)

 2 2
In this section |ψ you2! =
will encounter your first example of a quantum algorithm (1.43)
 ! "! "
that can compute faster 
 |0!
than − |1!
a |0! −
classical |1!
algorithm for the same problem.

 ± √ √ if f (0) $= f (1).
2
This is a simplified version of Deutsch’s 2
original algorithm, which shows how
it The
is possible to extract
final Hadamard gate global information
on the first qubit thus about
gives usa function by using quantum

parallelism and interference (Fig.|0!III.22).! " 8

 − |1!
Suppose we have a function 


±|0! f √ : 22 → 2, ifasf (0) in =Sec.
f (1) C.5. The goal is to
determine whether|ψf3 !(0) = = f (1) ! with a" single function evaluation. This (1.44) is

 |0! − |1!
not a very interesting problem 
 ±|1!(since
 √ there are if fonly
(0) $= four
f (1). such functions), but
2
it is a warmup for the Deutsch-Jozsa algorithm. Simple as it is, it could be
Realizing that
expensive f (0) ⊕ fon
to decide (1) a
is 0classical
if f (0) = computer.
f (1) and 1 otherwise, we can rewrite
For example, supposethis fresult
(0) =
concisely as
the billionth bit of π and f (1) = the billionth bit of e. Then the problem is
! "
to decide if the billionth bits of π and e are |0!the same. It is mathematically
− |1!
|ψ3 ! = ±|f (0) ⊕ f (1)! √ , (1.45)
simple, but computationally complex. 2
soTo
by see how we
measuring themight solvewethis
first qubit mayproblem,
determine suppose weThis
f (0) ⊕ f (1). haveisavery
quantum gate
interesting
indeed:
array Uf the
forquantum
f ; thatcircuit
is, Ufhas|xi|yi
given=us|xi|y ⊕ f (x)i.
the ability to determine
In particular, Uf |xi|0iof =
a global property
f (x), namely f (0) ⊕ f (1), using only one evaluation of f (x)! This is faster than is possible
8
with a classical
This apparatus,
is the 1998 which would
improvement require
by Cleve et al.attoleast two evaluations.
Deutsch’s 1985 algorithm (Nielsen &
This example highlights the difference between quantum parallelism and classical
Chuang, 2010, p. 59).
randomized algorithms. Naively, one might think that the state |0!|f (0)! + |1!|f (1)!
corresponds rather closely to a probabilistic classical computer that evaluates f (0) with
probability one-half, or f (1) with probability one-half. The difference is that in a classical
computer these two alternatives forever exclude one another; in a quantum computer it is
D. QUANTUM ALGORITHMS 125
|xi|f (x)i and Uf |xi|1i = |xi|¬f (x)i. Usually we set y = 0 to get the result
|f (x)i, but here you will see an application in which we want y = 1.
Now consider the result of applying Uf to |xi in the data register and to
the superposition |−i = √12 (|0i − |1i) in the target register.
1 1 1
Uf |xi|−i = √ |xi|f (x)i − √ |xi|¬f (x)i = √ |xi[|f (x)i − |¬f (x)i].
2 2 2
Now the rightmost square bracket is |0i − |1i if f (x) = 0 or |1i − |0i if
f (x) = 1. Therefore, we can write
1
Uf |xi|−i = √ |xi(−)f (x) (|0i − |1i) = (−)f (x) |xi|−i. (III.21)
2
[Here, (−)x is an abbreviation for (−1)x when we want to emphasize that
the sign is all that matters.] Since Uf |xi|−i = (−)f (x) |xi|−i, the result of
applying it to an equal superposition of x = 0 and x = 1 is:
1 X 1 X
√ Uf |xi|−i = √ (−)f (x) |xi|−i.
2 x∈2 2 x∈2
If f is a constant function, then f (0) = f (1), and the summation is ± √12 (|0i+
|1i)|−i = ±|+i|−i because both components have the same sign.. On the
other hand, if f (0) 6= f (1), then the summation is ± √12 (|0i − |1i)|−i =
±|−i|−i because the components have opposite signs. That is, a constant
function gives the |0i and |1i components of the data qubit the same phase,
and otherwise gives them the opposite phase. Therefore, we can determine
whether the function is constant or not by measuring the first qubit in the sign
basis; we get |+i if f (0) = f (1) and |−i otherwise. With this background,
we can state Deutsch’s algorithm.
algorithm Deutsch:
def
Initial state: Begin with the qubits |ψ0 i = |01i.
Superposition: Transform it to a pair of superpositions

def 1 1
|ψ1 i = √ (|0i + |1i) ⊗ √ (|0i − |1i) = | + −i. (III.22)
2 2
by a pair of Hadamard gates. Recall that H|0i = √1 (|0i + |1i) = |+i and
2
H|1i = √12 (|0i − |1i) = |−i.
Function application: Next apply Uf to |ψ1 i = | + −i. As we’ve seen,

Uf |xi|0i = |xi|0 ⊕ f (x)i = |xi|f (x)i, and Uf |xi|1i = |xi|1 ⊕ f (x)i =
|xi|¬f (x)i. Therefore, expand Eq. III.22 and apply Uf :
def
|ψ2 i = Uf |ψ1 i

1 1
= Uf √ (|0i + |1i) ⊗ √ (|0i − |1i)
2 2
1
= [Uf |00i − Uf |01i + Uf |10i − Uf |11i]
2
1
= [|0, f (0)i − |0, ¬f (0)i + |1, f (1)i − |1, ¬f (1)i]
2
There are two cases: f (0) = f (1) and f (0) 6= f (1).
Equal (constant function): If f (0) = f (1), then
1
|ψ2 i = [|0, f (0)i − |0, ¬f (0)i + |1, f (0)i − |1, ¬f (0)i]
2
1
= [|0i(|f (0)i − |¬f (0)i) + |1i(|f (0)i − |¬f (0)i)]
2
1
= (|0i + |1i)(|f (0)i − |¬f (0)i)
2
1
= ± (|0i + |1i)(|0i − |1i)
2
1
= ± √ (|0i + |1i)|−i
2
= | + −i.
The last line applies because global phase (including ±) doesn’t matter.
Unequal (balanced function): If f (0) 6= f (1), then
1
|ψ2 i = [|0, f (0)i − |0, ¬f (0)i + |1, ¬f (0)i − |1, f (0)i]
2
1
= [|0i(|f (0)i − |¬f (0)i) + |1i(|¬f (0)i − |f (0)i)]
2
1
= [|0i(|f (0)i − |¬f (0)i) − |1i(|f (0)i − |¬f (0)i)]
2
1
= (|0i − |1i)(|f (0)i − |¬f (0)i)
2
1
= ± (|0i − |1i)(|0i − |1i)
2
1
= ± √ (|0i − |1i)|−i
2
= | − −i
Clearly we can discriminate between the two cases by measuring the first
qubit in the sign basis. In particular, note that in the unequal case, the |1i
component has the opposite phase from the |0i component.
Measurement: Therefore we can determine whether f (0) = f (1) or not by

measuring the first bit of |ψ2 i in the sign basis, which we can do with the
Hadamard gate (recall H|+i = |0i and H|−i = |1i):
def
|ψ3 i = (H ⊗ I)|ψ2 i

±|0i|−i, if f (0) = f (1)
=
±|1i|−i, if f (0) 6= f (1)
= ±|f (0) ⊕ f (1)i|−i.
Notice that the information we need is in the data register, not the target
register. This technique is called phase kick-back (i.e., kicked back into the
phase of the data register).
In conclusion, we can determine whether or not f (0) = f (1) with a single
evaluation of f , which is quite remarkable. In effect, we are evaluating f on
a superposition of |0i and |1i and determining how the results interfere with
each other. As a result we get a definite (not probabilistic) determination of
a global property with a single evaluation.
This is a clear example where a quantum computer can do something
faster than a classical computer. However, note that Uf has to uncompute
Quantum algorithms 35
Figure 1.20. Quantum circuit implementing the general Deutsch–Jozsa algorithm. The wire with a ‘/’ through it
Figure III.23:
represents a set of nQuantum
qubits, similar circuit for engineering
to the common Deutsch-Jozsa
notation. algorithm. [fig. from NC]
evenly weighted superposition of 0 and 1. Next, the function f is evaluated (by Bob)
f , using
which : |x, y! as
Uf takes → |x, (x)!, as
y ⊕ ftime
much giving
computing it, but we will see other cases
(Deutsch-Jozsa) where the speedup ! (−1)is much
f (x) " more than# 2×.
|x! |0! − |1!
|ψ2 ! = √ √ . (1.48)
x 2n 2
D.1.b The Deutsch-Jozsa algorithm
Alice now has a set of qubits in which the result of Bob’s function evaluation is stored
in the amplitude of the qubit superposition state. She now interferes terms in the super-
The Deutsch-Jozsa algorithm is a generalization of the Deutsch algorithm to
position using a Hadamard transform on the query register. To determine the result of
n bits, which they published it in 1992; here we present the improved version
the Hadamard transform it helps to first calculate the effect of the Hadamard transform
of on
Nielsen & Chuang
a state |x!. (2010,
By checking the p. 59).
cases x = 0 and x = 1 separately we see that for a single
$ √
xz Suppose we are given an unknown function f : 2n →
This is the problem:
qubit H|x! = z (−1) |z!/ 2. Thus
n+1
2 in the form of a unitary transform $ Uf ∈x1L(Hz1 +·· +xn zn
, H) (Fig. III.23). We are
z ,...,z (−1) |z1 , . . . , zn !
told only thatHf⊗nis either constant
|x1 , . . . , xn ! = 1
or n
balanced,
√ which means .that it is 0 on
(1.49)
2
half its domain and 1 on the other half. Our task is to determine into which
n
Thisthe
class cangiven
be summarized
f falls. more succinctly in the very useful equation
Consider first the classical situation. $ Wex·z can
⊗n z (−1) |z! try different input bit strings
H
x. We might (if we’re lucky) discover after |x! = √ , (1.50)
2n the second query of f that it is
n
not constant. But we might require as many as 2 /2 + 1 queries to answer
where x · z is the bitwise inner product of x and z, modulo 2. Using this equation
the question. So we’re facing O(2n−1 ) function evaluations.
and (1.48) we can now evaluate |ψ3 !,
! ! (−1)x·z+f (x) |z! " |0! − |1! #
|ψ3 ! = √ . (1.51)
z x
2n 2
algorithm Deutsch-Jozsa:
Alice now observes the query register. Note that the amplitude for the state |0!⊗n is
$ f (x)
x (−1) /2n . Let’s look at the two possible cases – f constant and f balanced – to
def
discern what happens.
Initial state: In the
As in the case where
Deutsch f is constant
algorithm, the amplitude
prepare for |0!
the initial state
⊗n
is |ψ
+10 ior=
⊗n depending on the constant value f (x) takes. Because |ψ ! is of unit length it follows
|0i−1, |1i. 3
that all the other amplitudes must be zero, and an observation will yield 0s for all qubits
in the query register. If f is balanced then the positive and negative contributions to the
amplitude for |0!⊗n cancel, leaving an amplitude of zero, and a measurement must yield
a result other than 0 on at least one qubit in the query register. Summarizing, if Alice
Superposition: Use the Walsh-Hadamard transformation to create a su-

perposition of all possible inputs:
def
X 1
|ψ1 i = (H ⊗n ⊗ H)|ψ0 i = √ |x, −i.
x∈2 n 2n
Claim: Similarly to the single qubit case (Eq. III.21), we can see that
Uf |x, −i = (−)f (x) |xi|−i, where (−)n is an abbreviation for (−1)n . From
the definition of |−i and Uf , Uf |x, −i = |xi √12 (|f (x)i − |¬f (x)i). Since
f (x) ∈ 2, √12 (|f (x)i − |¬f (x)i) = |−i if f (x) = 0, and it = −|−i if f (x) = 1.
This establishes the claim.
Function application: Therefore, you can see that:

def
X 1
|ψ2 i = Uf |ψ1 i = √ (−)f (x) |xi|−i. (III.23)
x∈2n
2n
In the case of a constant function, all the components of the data state have
the same phase, otherwise they do not.
To see how we can make use of this information, let’s consider the state
in more detail. For a single bit you can show (Exer. III.46):
1 1 X X 1
H|xi = √ (|0i + (−)x |1i) = √ (−)xz |zi = √ (−)xz |zi.
2 2 z∈2 z∈2
2
(This is just another way of writing H|0i = √12 (|0i+|1i) and H|1i = √12 (|0i−
|1i).) Therefore, the general formula for the Walsh transform of n bits is:
1 X
H ⊗n |x1 , x2 , . . . , xn i = √ (−)x1 z1 +···+xn zn |z1 , z2 , . . . , zn i
2n z1 ,...,zn ∈2
1 X
= √ (−)x·z |zi, (III.24)
2n z∈2n
where x · z is the bitwise inner product. (It doesn’t matter if you do addition
or ⊕ since only the parity of the result is significant.) Remember this formula!
Combining this and the result in Eq. III.23,
def
X X 1
|ψ3 i = (H ⊗n ⊗ I)|ψ2 i = n
(−)x·z+f (x) |zi|−i.
z∈2n x∈2n
2
Measurement: Consider the first n qubits and the amplitude of one par-
ticular basis state, z = |0i = |0i⊗n , which we separate from the rest of the
summation:
X 1 X X 1
f (x)
|ψ3 i = n
(−) |0i|−i + n
(−)x·z+f (x) |zi|−i
x∈2n
2 n x∈2n
2
z∈2 −{0}
P
Hence, the amplitude of |0i⊗n , the all-|0i qubit string, is x∈2n 21n (−)f (x) .
Recall how in the basic Deutsch algorithm we got a sum of signs (either all
the same or not) for all the function evaluations. The result is similar here,
but we have 2n values rather than just two. We now have two cases:
Constant function: If the function is constant, then all the exponents of

−1 will be the same (either all 0 or all 1), and so the amplitude will be ±1.
Therefore all the other amplitudes are 0 and any measurement must yield 0
for all the qubits (since only |0i⊗n has nonzero amplitude).
Balanced function: If the function is not constant then (ex hypothesi) it

is balanced, but more specifically, if it is balanced, then there must be an
equal number of +1 and −1 contributions to the amplitude of |0i⊗n , so its
amplitude is 0. Therefore, when we measure the state, at least one qubit
must be nonzero (since the all-0s state has amplitude 0).
The good news about the Deutsch-Jozsa algorithm is that with one quan-
tum function evaluation we have got a result that would require between 2
and O(2n−1 ) classical function evaluations (exponential speedup!). The bad
news is that the algorithm has no known applications! Even if it were useful,
however, the problem could be solved probabilistically on a classical com-
puter with only a few evaluations of f : for an error probability of , it takes
O(log −1 ) function evaluations. However, it illustrates principles of quantum
computing that can be used in more useful algorithms.
D.2 Simon’s algorithm

Simon’s algorithm was first presented in 1994 and can be found in Simon, D.
(1997), “On the power of quantum computation,” SIAM Journ. Computing,
26 (5), pp. 1474–83.9 For breaking RSA we will see that its useful to know
the period of a function: that r such that f (x + r) = f (x). Simon’s problem
is a warmup for this.
Simon’s Problem: Suppose we are given an unknown function f :
2 → 2n and we are told that it is two-to-one. This means f (x) = f (y) iff
n
x ⊕ y = r for some fixed r ∈ 2n . The vector r can be considered the period

of f , since f (x ⊕ r) = f (x). The problem is to determine the period r of a
given unknown f .
Consider first the classical solution. Since we don’t know anything about
f , the best we can do is evaluate it on random inputs. If we are ever lucky
enough to find x and x0 such that f (x) = f (x0 ), then we have our answer,
r = x⊕x0 . After testing m values, you will have eliminated about m(m−1)/2
possible r vectors (namely, x ⊕ x0 for every pair of these m vectors). You
will be done when m2 ≈ 2n . Therefore, on the average you need to do
2n/2 function evaluations, which is exponential in the size of the input. For
n = 100, it would require about 250 ≈ 1015 evaluations. “At 10 million
calls per second it would take about three years” (Mermin, 2007, p. 55).
We will see that a quantum computer can determine r with high probability
(> 1 − 10−6 ) in about 120 evaluations. At 10 million calls per second, this
would take about 12 microseconds!
algorithm Simon:
Input superposition: As before, start by using the Walsh-Hadamard trans-

form to create a superposition of all possible inputs:
def 1 X
|ψ1 i = H ⊗n |0i⊗n = n/2 |xi.
2 x∈2n
9
The following presentation follows Mermin’s Quantum Computer Science (Mermin,
2007, §2.5, pp. 55–8).
Function evaluation: Suppose that Uf is the quantum gate array imple-

menting f and recall Uf |xi|yi = |xi|y ⊕ f (x)i. Therefore:
def 1 X
|ψ2 i = Uf |ψ1 i|0i⊗n = n/2 |xi|f (x)i.
2 x∈2n
Therefore we have an equal superposition of corresponding input-output val-
ues.
Output measurement: Measure the output register (in the computational

basis) to obtain some |zi. Since the function is two-to-one, the projection
will have a superposition of two inputs:
1
√ (|x0 i + |x0 ⊕ ri)|zi,
2
where f (x0 ) = z = f (x0 ⊕ r). The information we need is contained in the
input register,
def 1
|ψ3 i = √ (|x0 i + |x0 ⊕ ri),
2
but it cannot be extracted directly. If we measure it, we will get either x0
or x0 ⊕ r, but not both, and we need both to get r. (We cannot make two
copies, due to the no-cloning theorem.)
Suppose we apply the Walsh-Hadamard transform to this superposition:
1
H ⊗n |ψ3 i = H ⊗n √ (|x0 i + |x0 ⊕ ri)
2
1
= √ (H ⊗n |x0 i + H ⊗n |x0 ⊕ ri).
2
Now, recall (D.1.b, p. 129) that
1 X
H ⊗n |xi = n/2 (−1)x·y |yi.
2 y∈2n
(This is the general expression for the Walsh transform of a bit string. The
phase depends on the number of common 1-bits.) Therefore,
" #
1 1 X 1 X
H ⊗n |ψ3 i = √ n/2
(−1)x0 ·y |yi + n/2 (−1)(x0 +r)·y |yi
2 2 y∈2n
2 y∈2n
1 X
= (n+1)/2 (−1)x0 ·y + (−1)(x0 +r)·y |yi.
2 y∈2n
Note that
(−1)(x0 +r)·y = (−1)x0 ·y (−1)r·y .
Therefore, if r · y = 1, then the bracketed expression is 0 (since the terms
have opposite sign and cancel). However, if r · y = 0, then the bracketed
expression is 2(−1)x0 ·y (since they don’t cancel). Hence the result of the
Walsh-Hadamard transform is
1 X
|ψ4 i = H ⊗n |ψ3 i = (−1)x0 ·y |yi.
2(n−1)/2 y s.t. r·y=0
Measurement: Measuring the input register (in the computational basis)

will collapse it with equal probability into a state |y(1) i such that r·y(1) = 0.
First equation: Since we know y(1) , this gives us some information about
r, expressed in the equation:
(1) (1)
y1 r1 + y2 r2 + · · · + yn(1) rn = 0 (mod 2).
Iteration: The quantum computation can be repeated, producing a series

of bit strings y(1) , y(2) , . . . such that y(k) · r = 0. From them we can build
up a system of n linearly-independent equations and solve for r. (If you
get a linearly-dependent equation, you have to try again.) Note that each
quantum step (involving one evaluation of f ) produces an equation (except
in the unlikely case y(k) = 0 or that it’s linearly dependent), and therefore
determines one of the bits in terms of the other bits. That is, each iteration
reduced the candidates for r by approximately one-half.
A mathematical analysis (Mermin, 2007, App. G) shows that with n + m

iterations the probability of having enough information to determine r is
1
> 1 − 2m+1 . “Thus the odds are more than a million to one that with
n + 20 invocations of Uf we will learn [r], no matter how large n may be.”
(Mermin, 2007, p. 57) Note that the “extra” evaluations are independent of
n. Therefore Simon’s problem can be solved in linear time on a quantum
computer, but requires exponential time on a classical computer.
D.3 Shor’s algorithm

If computers that you build are quantum,
Then spies everywhere will all want ’em.
Our codes will all fail,
And they’ll read our email,
Till we get crypto that’s quantum, and daunt ’em.
— Jennifer and Peter Shor (Nielsen & Chuang, 2010, p. 216)
The widely used RSA public-key cryptography system is based on the diffi-
culty of factoring large numbers.10 The best classical algorithms are nearly
exponential in the size of the input, m = ln M . Specifically, the best
current (2006) algorithm (the number field sieve algorithm) runs in time
1/3 2/3
eO(m ln m) . This is subexponential but very inefficient. Shor’s quantum
algorithm is bounded error-probability quantum polynomial time (BQP),
specifically, O(m3 ). Shor’s algorithm was invented in 1994, inspired by Si-
mon’s algorithm.
Shor’s algorithm reduces factoring to finding the period of a function.
The connection between factoring and period finding can be understood as
follows. Assume you are trying to factor M . Suppose you can find x such
that x2 = 1 (mod M ). Then x2 − 1 = 0 (mod M ). Therefore (x + 1)(x − 1) =
0 (mod M ). Therefore both x + 1 and x − 1 have common factors with M
(except in the trivial case x = 1, and so long as neither is a multiple of M ).
Now pick an a that is coprime (relatively prime) to M . If ar = 1 (mod M )
and r happens to be even, we’re done (since we can find a factor of M as
explained above). (The smallest such r is called the order of a.) This r is
the period of ax (mod M ), since ax+r = ax ar = ax (mod M ).
In summary, if we can find the order of an appropriate a and it is even,
then we can probably factor the number. To accomplish this, we need to
find the period of ax (mod M ), which can be determined through a Fourier
transform.
Like the classical Fourier transform, the quantum Fourier transform puts
all the amplitude of the function into multiples of the frequency (reciprocal
period). Therefore, measuring the state yields the period with high proba-
bility.
10
These section is based primarily on Rieffel & Polak (2000).
D.3.a Quantum Fourier transform

Before explaining Shor’s algorithm, it’s necessary to explain the quantum
Fourier transform, and to do so it’s helpful to begin with a review of the
classical Fourier transform.
Let f be a function defined on [0, 2π). We know it can be represented as
a Fourier series,
∞ ∞
a0 X A0 X
f (x) = + (ak cos kx + bk sin kx) = + Ak cos(kx + φk ),
2 k=1
2 k=1
where k = 0, 1, 2, . . . represents the overtone series (natural number multiples

of the fundamental frequency). You know also that the Fourier transform
can be represented in the ciscoid (cosine + i sine) basis, where we define
def
uk (x) = cis(−kx) = e−ikx . (The “−” sign is irrelevant, but will be convenient
later.) The R 2πuk are orthogonal but not normalized, so we divide them by
2π, since 0 cos x + sin2 x dx = 2π. The Fourier series in this basis is
2
P
f (x) = ∞ ˆ ˆ 1
k=−∞ fk cis(−kx). The Fourier coefficients are given by fk = 2π huk |
R 2π
f i = 2π1 0 eikx f (x)dx. They give the amplitude and phase of the component
signals uk .
For the discrete Fourier transform (DFT) we suppose that f is repre-
def def
sented by N samples, fj = f (xj ), where xj = 2π Nj , with j ∈ N =
{0, 1, . . . , N − 1}. Let f = (f0 , . . . , fN −1 )T . Note that the xj are the 1/N
segments of a circle. (Realistically, N is big.)
Likewise each of the basis functions is represented by a vector of N sam-
def
ples: uk = (uk,0 , . . . , uk,N −1 )T . Thus we have a matrix of all the basis sam-
ples:
def
ukj = cis(−kxj ) = e−2πikj/N , j ∈ N.
In e−2πikj/N , note that 2πi represents a full cycle, k is the overtone, and j/N
represents the fraction of a full cycle.
Recall that every complex number has N principal N th -roots, and in
particular the number 1 (unity) has N principal N th -roots. Notice that N
samples of the fundamental period correspond to the N principal N th -roots
of unity, that is, ω j where (for a particular N ) ω = e2πi/N . Hence, ukj = ω −kj .
That is, uk = (ω −k·0 , ω −k·1 . . . , ω −k·(N −1) )√T . It is easy to show that the vectors
uk are orthogonal, and in fact that uk / N are ON (exercise). Therefore, f
can be represented by a Fourier series,

1 X ˆ 1 X †
f=√ fk uk = (u f )uk .
N k∈N N k∈N k
Define the discrete Fourier transform √ of the vector f , f̂ = Ff , to be the

vector of Fourier coefficients, fˆk = u†k f / N . Determine F as follows:
     
fˆ0 u†0 f u†0
 fˆ1  1  u†1 f  1  u†1 
     
f̂ =  .. = √  .. = √  ..  f.
 .  N .  N . 
ˆ
fN −1 u†N −1 f u†N −1
Therefore let
 
  ω 0·0 ω 0·1 ··· ω 0·(N −1)
u†0  
  ω 1·0 ω 1·1 ··· ω 1·(N −1)
1def  u†1  1 
 ω 2·0 ω 2·1 ··· ω 2·(N −1)


F= √  .. = √  .
N  .  N
.. .. .. .. 

. . . .
u†N −1
ω (N −1)·0 ω (N −1)·1 · · · ω (N −1)·(N −1)
√ √ (III.25)
kj
That is, Fkj = ukj / N = ω / N for k, j ∈ N. Note that the “−” signs in
the complex exponentials were eliminated by the conjugate transpose. F is
unitary transformation (exercise).
The fast Fourier transform (FFT) reduces the number of operations re-
quired from O(N 2 ) to O(N log N ).11 It does this with a recursive algorithm
that avoids recomputing values. However, it is restricted to powers of two,
N = 2n .
The quantum Fourier transform (QFT) is even faster, O(log2 N ), that is,
O(n2 ). However, because the spectrum is encoded in the amplitudes of the
quantum state, we cannot get them all. Like the FFT, the QFT is restricted
to powers of two, N = 2n . The QFT transforms the amplitudes of a quantum
state: X X
UQFT fj |ji = fˆk |ki,
j∈N k∈N
11
The FFT is O(N log N ), but N > M 2 = e2m . Therefore the FFT is O(M 2 log M 2 ) =
O(M 2 log M ) = O(me2m )
def
where f̂ = Ff .
Suppose f has period r, and suppose that the period evenly divides the
number of samples, r | N . Then all the amplitude of fˆ should be at multiples
of its fundamental frequency, N/r. If r 6 | N , then the amplitude will be
concentrated near multiples of N/r. The approximation is improved by using
larger n.
The FFT (and QFT) are implemented in terms of additions and multi-
plications by various roots of unity (powers of ω). In QFT, these are phase
shifts. In fact, the QFT can be implemented with n(n + 1)/2 gates of two
types: (1) One is Hj , the Hadamard transformation of the jth qubit. (2) The
other is a controlled phase-shift Sj,k , which uses qubit xj to control whether
it does a particular phase shift on the |1i component of qubit xk . That is,
Sj,k |xj xk i 7→ |xj x0k i is defined by
def
Sj,k = |00ih00| + |01ih01| + |10ih10| + eiθk−j |11ih11|,
where θk−j = π/2k−j . That is, the phase shift depends on the indices j and
k.
It can be shown that the QFT can be defined:12
n−1
Y n−1
Y
UQFT = Hj Sj,k .
j=0 k=j+1
This is O(n2 ) gates.
D.3.b Shor’s algorithm step by step

Shor’s algorithm depends on many results from number theory, which are
outside of the scope of this course. Since this is not a course in cryptography
or number theory, I will just illustrate the ideas.
algorithm Shor:
12
See Rieffel & Polak (2000) for this, with a detailed explanation in Nielsen & Chuang
(2010, §5.1, pp. 517–21).
Input: Suppose we are factoring M (and M = 21 will be used for con-

crete examples, but of course the goal is to factor very large numbers). Let
def
m = dlg M e = 5 in the case M = 21.
Step 1: Pick a random number a < M . If a and M are not coprime (rela-
tively prime), we are done. (Euclid’s algorithm is O(m2 ) = O(log2 M ).) For
our example, suppose we pick a = 11, which is relatively prime with 21.
def def
Modular exponentiation: Let g(x) = ax (mod M ), for x ∈ M =
{0, 1, . . . , M − 1}. This takes O(m3 ) gates and is the most complex part
of the algorithm! (Reversible circuits typically use m3 gates for m qubits.)
In our example case, g(x) = 11x (mod 21), so
g(x) = 1, 11, 16, 8, 4, 2, 1, 11, 16, 8, 4, . . .
| {z }
period
In order to get a good QFT approximation, pick n such that M 2 ≤ 2n < 2M 2 .
def
Let N = 2n . Equivalently, pick n such that 2 lg M ≤ n < 2 lg M + 1, that
is, roughly twice as many qubits as in M . Note that although the number
of samples is N = 2n , we need only n qubits (thanks to the tensor product).
This is where quantum computation gets its speedup over classical compu-
tation; M is very large, so N > M 2 is extremely large. The QFT computes
all these in parallel. For our example M = 21, and so we pick n = 9 for
N = 512 since 441 ≤ 512 < 882. Therefore m = 5.
Step 2 (quantum parallelism): Apply Ug to the superposition

def 1 X
|ψ0 i = H ⊗n |0i⊗n = √ |xi
N x∈N
to get
def 1 X
|ψ1 i = Ug |ψ0 i|0i⊗m = √ |x, g(x)i.
N x∈N
For our example problem, 14 qubits are required [n = 9 for x and m = 5 for
g(x)]. The quantum state looks like this (note the periodicity):
1
|ψ1 i = √ ( |0, 1i + |1, 11i + |2, 16i + |3, 8i + |4, 4i + |5, 2i +
512
|6, 1i + |7, 11i + |8, 16i + |9, 8i + |10, 4i + |11, 2i + · · ·)
Step 3 (measurement): The function g has a period r, which we want

to transfer to the amplitudes of the state so that we can apply the QFT.
This is accomplished by measuring (and discarding) the result register (as
in Simon’s algorithm).13 Suppose the result register collapses into state g ∗
(e.g., g ∗ = 8). The input register will collapse into a superposition of all x
such that g(x) = g ∗ . We can write it:
" #
def 1 X 1 X 1 X
|ψ2 i = |x, g ∗ i = fx |x, g ∗ i = fx |xi |g ∗ i,
Z ∗
Z x∈N
Z x∈N
x∈N s.t. g(x)=g
where
def 1, if g(x) = g ∗
fx = ,
0, otherwise
def p
and Z = |{x | g(x) = g ∗ }| is a normalization factor. For example,
1
|ψ2 i = (|3, 8i + |9, 8i + |15, 8i + · · ·)
Z
1
= (|3i + |9i + |15i + · · ·)|8i
Z
Note that the values x for which fx 6= 0 differ from each other by the period;
This produces a function f of very strong periodicity. As in Simon’s algo-
rithm, if we could measure two such x, we would have useful information,
but we can’t. Suppose we measure the result register and get g ∗ = 8. Fig.
III.24 shows the corresponding f = (0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, . . .).
Step 4 (QFT): Apply the QFT to obtain,

!
def 1 X
|ψ3 i = UQFT fx |xi
Z x∈N
1 X ˆ
= fx̂ |x̂i.
Z x̂∈N
13
As it turns out, this measurement of the result register can be avoided. This is in
general true for “internal” measurement processes in quantum algorithms (Bernstein &
Vazirani 1997).
0.012
0.0108
0.0096
0.0084
0.012
0.0072
0.0108
0.006
0.0096
0.0048
0.0084
0.0036
0.0072
0.0024
0.006
0.0012
0.0048
0.0
0.0036
0 64 128 192 256 320 384 448 512
0.0024
0.0012
!
Fig. 2. Probabilities
FigureP III.24: for measuring
x mod 21 = 8}}
Example x when probability
measuring the state distribution
C x∈X |fx |in2 Stepfor
|x, 8! obtained 2, where
state
X
−1 = {x|211
0.0
Z x∈N0fx |x, 8i. 64 In128this example
192 256the period
320 is r =448 6 (e.g.,
384 512 at
x = 3, 9, 15, . . .). [fig. source: Rieffel & Polak (2000)]
0.17 !
Fig. 2. Probabilities for measuring x when measuring the state C x∈X
|x, 8! obtained in Step 2, where
0.153 x mod 21 = 8}}
X = {x|211
0.136
0.119
0.17
0.102
0.153
0.085
0.136
0.068
0.119
0.051
0.102
0.034
0.085
0.017
0.068
0.0
0.051
0 64 128 192 256 320 384 448 512
0.034
0.017 Fig. 3. Probability distribution of the quantum state after Fourier Transformation.
0.0
0 64 128 192 256 320 384 448 512
where the amplitude is 0 except at multiples of 2m /r. When the period r does not divide
2m , the transform approximates the of
exact case so most of the amplitude is attached to
Figure
integers
Fig. 3. Probabilityprobability
III.25:
close toExample
distribution
multiples of 2r .
m thedistribution fˆx̂ |2 of
quantum state after|Fourier Transformation.
the quantum Fourier
transform
Example.ofFigure
f . The3 showsspectrum
the result of is applying
concentrated
the quantum near multiples
Fourier Transform of toN/6
the =
m
where
512/6 the amplitude
= 85 1/3,
state obtained is
in Stepthat 0 except
is that
2. Note at multiples
85 Figure
1/3, 170 of 2
2/3,
3 is the /r.
256,
graph When the period
etc.fast [fig.
of the r does
source:
Fourier not divide
transformRieffel
of the &
m
2function
Polak , the transform approximates
shown in Figure
(2000)] 2. In2mthisthe exact case
particular so most
example the of the amplitude
period of f does notis attached
divide 2mto.
integers close to multiples of r .
Step 4. Extracting the period. Measure the state in the standard basis for quantum com-
Example.
putation, andFigure
call the3 shows theInresult
result v. of applying
the case where thetheperiod
quantum Fourier
happens to Transform
be a powertoofthe 2,
state
so thatobtained in StepFourier
the quantum 2. Notetransform
that Figure 3 isexactly
gives the graph of the fast
multiples of 2Fourier
m transform
/r, the period isofeasy
the
function
to extract.shown in Figure
In this case, v 2.=Inj 2this
m
particular example the period of f does not divide 2m .
r for some j. Most of the time j and r will be relatively
Step 4. Extracting the period. Measure the state in the standard basis for quantum com-
putation, and call the result v. In the case where the period happens to be a power of 2,
so that the quantum Fourier transform gives exactly multiples of 2m /r, the period is easy
m
to extract. In this case, v = j 2r for some j. Most of the time j and r will be relatively
(The collapsed result register |g ∗ i has been omitted.)

If the period r divides N = 2n , then f̂ will be nonzero only at multiples
of the fundamental frequency N/r. That is, the nonzero components will be
|kN/ri. If it doesn’t divide evenly, then the amplitude will be concentrated
around these |kN/ri. See Fig. III.24 and Fig. III.25 for examples of the
probability distributions |fx |2 and |fˆx̂ |2 .
Step 5 (period extraction): Measure the state in the computational basis.
def
Period a power of 2: If r | N , then the resulting state will be v = |kN/ri
for some k ∈ N. Therefore k/r = v/N . If k and r are relatively prime, as is
likely, then reducing the fraction v/N to lowest terms will produce r in the
denominator. In this case the period is discovered.
Period not a power of 2: In the case r does not divide N , it’s often pos-
sible to guess the period from a continued fraction expansion of v/N .14 If
v/N is sufficiently close to p/q, then a continued fraction expansion of v/N
will contain the continued fraction expansion of p/q. For example, suppose
the measurement returns v = 427, which is not a power of two. This is the
result of the continued fraction expansion of v/N (in this case, 427/512) (see
IQC):
i ai pi qi i
0 0 0 1 0.8339844
1 1 1 1 0.1990632
2 5 5 6 0.02352941
3 42 211 253 0.5
“which terminates with 6 = q2 < M ≤ q3 . Thus, q = 6 is likely to be the

period of f .” [IQC]
Step 6 (finding a factor): The following computation applies however we

14
See Rieffel & Polak (2000, App. B) for an explanation of this procedure and citations
for why it works.
got the period q in Step 5. If the guess q is even, then aq/2 + 1 and aq/2 − 1
are likely to have common factors with M . Use the Euclidean algorithm to
check this. The reason is as follows. If q is the period of g(x) = ax (mod M ),
then aq = 1 (mod M ). This is because, if q is the period, then for all x,
g(x + q) = g(x), that is, aq+x = aq ax = ax (mod M ) for all x. Therefore
aq − 1 = 0 (mod M ). Hence,
(aq/2 + 1)(aq/2 − 1) = 0 (mod M ).
Therefore, unless one of the factors is a multiple of M (and hence = 0 mod

M ), one of them has a nontrivial common factor with M .
In the case of our example, the continued fraction gave us a guess q = 6, so
with a = 11 we should consider 113 + 1 = 1332 and 113 − 1 = 1330. For M =
21 the Euclidean algorithm yields gcd(21, 1332) = 3 and gcd(21, 1330) = 7.
We’ve factored 21!
Iteration: There are several reasons that the preceding steps might not have
succeeded: (1) The value v projected from the spectrum might not be close
enough to a multiple of N/r (D.3.b). (2) In D.3.b, k and r might not be
relatively prime, so that the denominator is only a factor of the period, but
not the period itself. (3) In D.3.b, one of the two factors turns out to be a
multiple of M . (4) In D.3.b, q was odd. In these cases, a few repetitions of
the preceding steps yields a factor of M .
D.3.c Recent progress

To read our E-mail, how mean
of the spies and their quantum machine;
be comforted though,
they do not yet know
how to factorize twelve or fifteen.
— Volker Strassen (Nielsen & Chuang, 2010, p. 216)
In this section we review recent progress in hardware implementation of

Figure III.26: Hardware implementation of Shor’s algorithm developed at

UCSB (2012). The Mj are quantum memory elements, B is a quantum
“bus,” and the Qj are phase qubits that can be used to implement qubit
operations between the bus and memory elements. [source: CPF]
Figure III.27: Circuit of hardware implementation of Shor’s algorithm devel-

oped at UCSB. [source: CPF]
Shor’s algorithm.15 In Aug. 2012 a group at UC Santa Barbara described

a quantum implementation of Shor’s algorithm that correctly factored 15
about 48% of the time (50% being the theoretical success rate). (There have
been NMR hardware factorizations of 15 since 2001, but there is some doubt
if entanglement was involved.) This is a 3-qubit compiled version of Shor’s
algorithm, where “compiled” means that the implementation of modular ex-
ponentiation is for fixed M and a. This compiled version used a = 4 as the
coprime to M = 15. In this case the correct period r = 2. The device (Fig.
III.26) has nine quantum devices, including four phase qubits and five su-
perconducing co-planar waveguide (CPW) microwave resonators. The four
CPWs (Mj ) can be used as memory elements and fifth (B) can be used as a
“bus” to mediate entangling operations. In effect the qubits Qj can be read
and written. Radio frequency pulses in the bias coil can be used to adjust the
qubit’s frequency, and gigahertz pulses can be used to manipulate and mea-
sure the qubit’s state. SQUIDs are used for one-shot readout of the qubits.
The qubits Qj can be tuned into resonance with the bus B or memory ele-
ments Mj . The quantum processor can be used to implement the single-qubit
gates X, Y, Z, H, and the two-qubit swap (iSWAP) and controlled-phase (Cφ )
gates. The entanglement protocol can be scaled to an arbitrary number of
qubits. The relaxation and dephasing times are about 200ns.
Another group has reported the quantum factoring of 21.16 Their pro-
cedure operates by using one qubit instead of the n qubits in the (upper)
control qubits. It does this by doing all the unitaries associated with the
lowest-order control qubit, then for the next control qubit, updating the
work register after each step, for n interations.
15
This section is based primarily on Erik Lucero, R. Barends, Y. Chen, J. Kelly, M.
Mariantoni, A. Megrant, P. O’Malley, D. Sank, A. Vainsencher, J. Wenner, T. White, Y.
Yin, A. N. Cleland & John M. Martinis, “Computing prime factors with a Josephson phase
qubit quantum processor.” Nature Physics 8, 719–723 (2012) doi:10.1038/nphys2385
[CPF].
16
See Martin-Lópex et al., “Experimental realization of Shor’s quantum factoring algo-
rithm using qubit recycling,” Nature Photonics 6, 773–6 (2012).
D.4 Search problems

D.4.a Overview
Many problems can be formulated as search problems over a solution space
S.17 That is, find the x ∈ S such that some predicate P (x) is true. For
example, hard problems such as the Hamiltonian path problem and Boolean
satisfiability can be formulated this way. An unstructured search problem
is a problem that makes no assumptions about the structure of the search
space, or for which there is no known way to make use of it (also called a
needle in a haystack problem). That is, information about a particular value
P (x0 ) does not give us usable information about another value P (x1 ). In
contrast, a structured search problem is a problem in which the structure of
the solution space can be used to guide the search, for example, searching an
alphabetized array. In general, unstructured search takes O(M ) evaluations,
where M = |S| is the size of the solution space (which is often exponential in
the size of the problem). On the average it will be M/2 (think of searching
an unordered array) to find a solution with 50% probability.
We will see that Grover’s algorithm can do unstructured
√ search on a quan-
tum computer with bounded probability in O( M ) time, that is, quadratic
speedup. This is provably more efficient than any algorithm on a classical
computer, which is good (but not great). Unfortunately, it has been proved
that Grover’s algorithm is optimal for unstructured search. Therefore, to
do better requires exploiting the structure of the solution space. Quantum
computers do not exempt us from understanding the problems we are trying
to solve! Shor’s algorithm is an excellent example of exploiting the structure
of a problem domain. Later we will take a look at heuristic quantum search
algorithms that do make use of problem structure.
D.4.b Grover’s Algorithm
algorithm Grover:
Input: Let M be the size of the solution space and pick n such that 2n ≥ M .
17
This section is based primarily on Rieffel & Polak (2000).
measurement has projected the state of Eq. 2 onto the subspace !ki=1 |xi , 1! where
measurement has projected the state of Eq. 2 onto the subspace √12kk i=1 |xi , 1! where
k is the number of solutions. Further measurement of the remaining2 bits will provide one
kofisthese
the number of solutions. Further measurement of the remaining bits
solutions. If the measurement of qubit P (x) yields 0, then the whole will provide one
process is
of theseover
started solutions.
and theIfsuperposition
the measurement
of Eq.of2 qubit P (x)
must be yields 0,
computed then the whole process is
again.
started over algorithm
Grover’s and the superposition
then consistsofofEq.
the2following
must be computed
steps: again.
Grover’s algorithm then consists of the following steps:
(1) Prepare a register containing a superposition of all possible values xi ∈ [0 . . . 2n − 1].
(1) Prepare a register containing a superposition of all possible values xi ∈ [0 . . . 2n − 1].
(2) Compute P (xi ) on this register.
(2) Compute P (xi ) on this register.
(3) Change amplitude aj to −aj for xj such that P (xj ) = 1. An efficient algorithm for
(3) QUANTUM
D. Change
changingamplitude aj to −a
ALGORITHMS
selected signs j for xj such
is described that P 7.1.2.
in section (xj ) =A1.plot
Anofefficient algorithmafter
the amplitudes for 147
changing selected signs
this step is shown here. is described in section 7.1.2. A plot of the amplitudes after
this step is shown here.
average
average
0
0
(4) Apply inversion about the average to increase amplitude of xj with P (xj ) = 1. The
(4) Apply
Figure inversion
III.28:
quantum about
Depiction
algorithm the of
average to increase
theperform
to efficiently result of amplitude
phase
inversion of
thexjaverage
rotation
about with P is
(xgiven
j ) = 1.
(changing The
inthe
sec-sign)
quantum
of solutions algorithm
tion 7.1.1.inThe
Grover’sto efficiently perform
algorithm.
resulting amplitudes look inversion
[source: about
as shown,Rieffel the average
& amplitude
where the is given
Polak (2000)] in sec-
of all the xi ’s
tion
with7.1.1.
P (xi )The
= 0resulting
have beenamplitudes look
diminished as shown, where the amplitude of all the xi ’s
imperceptibly.
with P (xi ) = 0 have been diminished imperceptibly.
average
average
0
0 √
(5) Repeat steps 2 through 4 ππ4 √ 2n times.
(5) Repeat steps 2 through 4 4 2n times.
(6) Read
Figure the result.
III.29: Depiction of result of inversion about the mean in Grover’s
(6) Read the result.
algorithm.
Boyer et.al.[source:
[Boyer et Rieffel
al. 1996] & Polak
provide (2000)]
a detailed analysis of the performance of Grover’s
Boyer et.al.
algorithm. They[Boyer
proveetthat
al. 1996] provide
Grover’s a detailed
algorithm analysis
is optimal upoftothe performance
a constant of no
factor; Grover’s
quan-
algorithm. They prove that Grover’s algorithm
tum algorithm can perform an unstructured search √ is optimal up to a constant factor;
faster. They also show that if there no quan-is
tum
onlyalgorithm
a single x0can perform
such that P (xan )unstructured searchπ √ faster. They also
n iterations show 2that if there is
0 is true, then after π 2 of steps through 4 the
def nx such that P (x
onlyNa single def ) is √ then after 8 2n iterations of steps
Let = 2is 0.5.
failure rate andAfter
0 let N =0 2nππ4true,
iterating =2{0,
√ 1, . .the
n times N − 1},
. ,8failure ratethedropssetto of−n
2−n
2 through
n-bit 4 the We
strings.
. Interestingly,
failure rate is 0.5. After iterating 2 n times the failure rate drops to 2√ . Interestingly,
are given iterations
additional a computable predicate
will increase P rate.
4the failure : N For → example,
2. Suppose after ππ2we
√ 2nhave
n
a quantum
iterations the
additional
failure iterations
rate isUclose to will
1. increase the failure rate. For example, after 2 2 iterations the
gate array P (an oracle) that computes the predicate:
failure
Thererate
areismany
close to 1. algorithms in which a procedure is repeated over and over again
classical
There are many classical
for ever better results. Repeatingalgorithms
quantumin which a procedure
procedures may is repeated
improve overfor
results anda over again
while, but
for ever better results. RepeatingUquantumP |x, yi procedures
= |x, y ⊕may P (x)i.
improve results for a while, but
Application: Consider what happens if, as usual, we apply the function to

a superposition of all possible inputs |ψ0 i:
" #
1 X 1 X
UP |ψ0 i|0i = UP √ |x, 0i = √ |x, P (x)i.
N x∈N N x∈N
Notice that the components we want, |x, 1i, and the components we don’t
want, |x, 0i, all have the same amplitude, √1N . So if we measure the state,
the chances of getting a hit are very small, O(2−n ). The trick, therefore, is
to amplify the components that we want at the expense of the ones we don’t
want; this is what Grover’s algorithm accomplishes.
Sign-change: To do this, first we changePthe sign of every solution (a phase

rotation of π). That is, if the state is j aj |xj , P (xj )i, then we want to
change aj to −aj whenever P (xj ) = 1. See Fig. III.28. I’ll get to the
technique in a moment.
Inversion about mean: Next, we invert all the components around their
mean amplitude (which is a little less than the amplitudes of the non-
solutions); the result is shown in Fig. III.29. As a result of this operation,
amplitudes of non-solutions go from a little above the mean to a little below
it, but amplitudes of solutions go from far below the mean to far above it.
This amplifies the solutions.
Iteration: This Grover iteration (the sign change and inversion about the
√
π N
√
mean) is repeated 4 times. Thus the algorithm is O( N ).
Measurement: Measurement yields an x0 for which P (x0 ) = 1 with high √

π N
probability. Specifically, if there is exactly one solution x 0 ∈ S, then 8
√
iterations will yield it with probability 1/2. With π 4N iterations, the prob-
ability of failure drops to 1/N = 2−n . Unlike with most classical algorithms,
additional iterations will give a worse result! This is because Grover iter-
ations are unitary rotations, and so excessive rotations can rotate past the
solution. Therefore it is critical to know when to stop. Fortunately there
is a quantum technique (Brassard et al. 1998) for determining the optimal
stopping point. Grover’s iteration can be used for a wide variety of problems
as a part of other quantum algorithms.
In the following geometric analysis, I will suppose that there is just one
answer α such that P (α) = 1; then |αi is the desired answer vector. Let
|ωi be a uniform superposition of all the other (non-answer) states, and
observe that |αiqand |ωi are orthonormal. Therefore, initially the state is
√1 |αi N −1
|ψ0 i = N
+ N
|ωi. In general, after k iterations the state is |ψk i =
2â
Figure III.30: Process of inversion about the mean in Grover’s algorithm.

The black lines represent the original amplitudes aj . The red lines represent
2ā − aj , with the arrow heads indicating the new amplitudes a0j .
a|αi + w|ωi, for some a, w with |a|2 + |w|2 = 1.

The sign change operation transforms the state as follows:
|ψk i = a|αi + w|ωi 7→ −a|αi + w|ωi = |ψk0 i,
where I’ve called the result |ψk0 i. This is a reflection across the |ωi vector,
which means that it will be useful to look at reflections more generally.
Suppose that |φi and |φ⊥ i are orthonormal vectors and that |ψi = a|φ⊥ i+
b|φi is an arbitrary vector in the space they span (Fig. III.31). The reflection
of |ψi across |φi is |ψ 0 i = −a|φ⊥ i + b|φi. Since a = hφ⊥ | ψi and b =
hφ | ψi, we know |ψi = |φihφ | ψi + |φ⊥ ihφ⊥ | ψi, and you can see that
|ψ 0 i = |φihφ | ψi − |φ⊥ ihφ⊥ | ψi. Hence the operator to reflect across |φi is
def
Rφ = |φihφ| − |φ⊥ ihφ⊥ |. Alternate forms of this operator are 2|φihφ| − I and
I − 2|φ⊥ ihφ⊥ |, that is, subtract twice the perpendicular component.
The sign change can be expressed as a reflection:
Rω = |ωihω| − |αihα| = I − 2|αihα|,
which expresses the sign-change of the answer vector clearly. Of course we
don’t know |αi, which is why we will have to use a different process to
accomplish this reflection (see p. 151). We also will see that the inversion
about the mean is equivalent to reflecting that state vector across |ψ0 i.
But first, taking this for granted, let’s see the effect of the Grover iteration
(Fig. III.32). Let θ be the angle between |ψ0 i and |ωi. It’s given by the inner
∣𝜑⊥⟩
∣𝜓⟩
b
∣𝜑⟩
–a
∣𝜓′⟩
Figure III.31: Reflection across arbitrary vector. Reflection of |ψi across |φi
in plane with |φ⊥ i. |ψi = a|φ⊥ i + b|φi becomes |ψ 0 i = −a|φ⊥ i + b|φi.
|α〉
|ψ1〉
2θ
|ψ0〉
θ
θ |ω〉
|ψ'0〉
Figure III.32: Geometry of first Grover iteration.
q
product cos θ = hψ0 | ωi = NN−1 . Therefore the sign change reflects |ψ0 i
from θ above |ωi into |ψ00 i, which is θ below it. Inversion about the mean
reflects |ψ00 i from 2θ below |ψ0 i into a state we call |ψ1 i, which is 2θ above it.
Therefore, in going from |ψ0 i to |ψ1 i the state vector has rotated 2θ closer
to |αi.
You can see that after k iterations, the state vector |ψk i will be (2k + 1)θ
above |ωi. We can solve (2k + 1)θ = π/2 to get the required number of
iterations to bring |ψk i to |αi. Note that for small θ, θ ≈ sin θ = √1N (which
√ √
is certainly small).√ Hence, we want (2k +1)/ N ≈ π/2, or 2k +1 ≈ π N /2.
That
√ is, k ≈ π N /4 is the required number of iterations. Note that after
π N /8 iterations, we are about halfway there (i.e., π/4), so the probability
of success is 50%. In general, the probability of success is about sin2 2k+1 √ .
N
Now for the techniques for changing the sign and inversion about the
mean. Let |ψk i be the state after k iterations (k ≥ 0). To change the sign,
simply apply UP to |ψk i|−i. To see the result, let X0 = {x | P (x) = 0} and
X1 = {x | P (x) = 1}, the solution set. Then:
UP |ψk i|−i
" #
X
= UP ax |x, −i
x∈N
" #
1 X
= UP √ ax |x, 0i − ax |x, 1i
2 x∈N
" #
1 X X X X
= √ UP ax |x, 0i + ax |x, 0i − ax |x, 1i − ax |x, 1i
2 x∈X0 x∈X1 x∈X0 x∈X1
"
1 X X
= √ ax UP |x, 0i + ax UP |x, 0i
2 x∈X0 x∈X1
#
X X
− ax UP |x, 1i − ax UP |x, 1i
x∈X0 x∈X1
"
1 X X
= √ ax |x, 0i + ax |x, 1i
2 x∈X0 x∈X1
#
X X
− ax |x, 1 ⊕ 0i − ax |x, 1 ⊕ 1i
x∈X0 x∈X1
" #
1 X X X X
= √ ax |xi|0i + ax |xi|1i − ax |xi|1i − ax |xi|0i
2 x∈X0 x∈X1 x∈X0 x∈X1
!
1 X X
= √ ax |xi − ax |xi (|0i − |1i)
2 x∈X0 x∈X1
!
X X
= ax |xi − ax |xi |−i.
x∈X0 x∈X1
Therefore the signs of the solutions have been reversed (they have been ro-
tated by π). Notice how |−i in the target register can be used to separate
the 0 and 1 results by rotation. This is a useful idea!
It remains to show the connection between inversion about the mean and
reflection across |ψ0 i. This reflection is given by Rψ0 = 2|ψ0 ihψ0 | − I. Note
that:
! !
1 X 1 X 1 XX
|ψ0 ihψ0 | = √ |xi √ hy| = |xihy|.
N x∈N N y∈N N x∈N y∈N
This is the diffusion matrix:

 1 1 1

N N
··· N
 1 1
··· 1 
 N N N 
 .. .. .. ..  ,
 . . . . 
1 1 1
N N
··· N
which, as we will see, does the averaging.

To perform inversion about the mean, let ā be the average of the aj (see
Fig. III.30). Inversion about the mean is accomplished by the transformation:
X X
aj |xj i 7→ (2ā − aj )|xj i.
j∈N j∈N
To see this, write aj = ā ± δj , that is, as a difference from the mean. Then
2ā − aj = 2ā − (ā ± δj ) = ā ∓ δj . Therefore an amplitude δj below the
mean will be transformed to δj above, and vice verse. But an amplitude
that is negative, and thus very far below the mean, will be transformed to
an amplitude much above the mean. This is exactly what we want in order
to amplify the negative components, which correspond to solutions.
Inversion about the mean is accomplished by a “Grover diffusion trans-

formation” D. To derive the matrix D, consider the new amplitude a0j as a
function of all the others:
N −1
!
1 X X 2 2
0 def
aj = 2ā − aj = 2 ak − aj = ak + − 1 aj .
N k=0 k6=j
N N
2 2
This matrix has N
− 1 on the main diagonal and N
in the off-diagonal
elements:  2 
2 2
N
−1 N
··· N
 2 2
− 1 ··· 2 
 
D =  N.. N
.. ...
N
.. .
 . . . 
2 2 2
N N
··· N
−1
Note that D = 2|ψ0 ihψ0 |−I = Rψ0 . It is easy to confirm that DD† = I (Exer.
III.49), so the matrix is unitary and therefore a possible quantum operation,
but it remains to be seen if it can be implemented efficiently.
We claim D = W RW , where W = H ⊗n is the n-qubit Walsh-Hadamard
transform and R is the phase rotation matrix:
 
1 0 ··· 0
 0 
def  0 −1 · · · 
R =  .. .. . . ..  .
 . . . . 
0 0 · · · −1
To see this, let  

2 0 ··· 0
 0 0 ··· 0 
def  
R0 = R + I =  .... . . .. .
 . . . . 
0 0 ··· 0
Then W RW = W (R0 − I)W = W R0 W − W W = W R0 W − I. It is easy to
show (Exer. III.50) that:
 2 2 2

N N
··· N
 2 2 ··· 2 
 
W R0 W =  N.. N.. . . N
..  .
 . . . . 
2 2 2
N N
··· N
Figure III.33: Circuit for√ Grover’s algorithm. The Grover iteration in the
dashed box is repeated π 4N times.
It therefore follows that D = W R0 W − I = W RW . See Fig. III.33 for a

diagram of Grover’s algorithm.
It remains to consider the possibility that there may be several solutions
√
π N/s
to the problem. If there are s solutions, then run the Grover iteration 4
p
times, which is optimal (Exer. III.51). It can be done in N/s iterations
even if s is unknown.
D.4.c Hogg’s heuristic search algorithms

Many important problems can be formulated as constraint satisfaction prob-
lems, in which we try to find a set of assignments to variables that satisfy
def
specified constraints. More specifically let V = {v1 , . . . , vn } be a set of vari-
def
ables, and let X = {x1 , . . . , xn } be a set of values that can be assigned to
the variables, and let C1 , . . . , Cl be the constraints. The set of all possible
assignments of values to variables is V × X. Subsets of this set correspond
to full or partial assignments, including inconsistent assignments. The set of
all such assignments is P(V × X).
The sets of assignments form a lattice under the ⊆ partial order (Fig.
III.34). By assigning bits to the elements of V × X, elements of P(V × X)
can be represented by mn-element bit strings (i.e., integers in the set MN =
{0, . . . , 2mn − 1}); see Fig. III.35. Hogg’s algorithms are based on the ob-
servation that if an assignment violates the constraints, then so do all those
assignments above it in the lattice.
 Introduction to Quantum Computing

 · 33

 v1 = 0 

v1 = 1
 v2 = 0 
 
v2 = 1
$$!!
$ $$ #" !!!
$$ # " !!
$$ # " !!
# $ # $ # $ # $
v1 = 1 v1 = 0 v1 = 0 v1 = 0
v2 = 0 v2 = 0 v1 = 1 v1 = 1
v2 = 1 v2 = 1 v2 = 1 v2 = 0
&''''((('''( (&((
&'''' &&%%
&&(((( '' (( '
(' &' '' &''' %
&&
( (( ( (( '
&'&' '&&'
' ''%
'%
(( " ! (( " ! & " '!
! & '& " '! ' " '!
%
' "
v2 = 0 v1 = 1 v1 = 0 v1 = 1 v1 = 0 v1 = 0
v2 = 1 v2 = 1 v2 = 1 v2 = 0 v2 = 0 v1 = 1
'
%' ''' ' ( ((
%''' ' && ''' && (((( ((& &
%% ''' &&'''' &&'(' (
('' (( ((
&&
%%&&''''( &'
&(('(' '(
'((('''&&
{v1 = 0} {v1 = 1} {v2 = 0} {v2 = 1}
!! $$
!! " # $$
!! " #$$$
!!
#$
!$
"
∅
Fig. 4. Lattice of variable assignments in a CSP

Figure III.34: Lattice of variable assignments. [source: Rieffel & Polak
(2000)] 1 rn−1 r0
= √ (|0" + (−1) |1") ⊗ . . . ⊗ (|0" + (−1) |1")
2n
2n −1
1 ,
= √ (−1)sn−1 rn−1 |sn−1 " ⊗ . . . ⊗ (−1)s0 r0 |s0 "
2n s=0
n
2 −1
1 ,
= √ (−1)s·r |s".
2n s=0
7.2.2 Overview of Hogg’s algorithms. A constraint satisfaction problem (CSP) has n

variables V = {v1 , . . . , vn } which can take m different values X = {x1 , . . . , xm } subject
to certain constraints C1 , . . . , Cl . Solutions to a constraint satisfaction problem lie in the
space of assignments of xi ’s to vj ’s, V ×X. There is a natural lattice structure on this space
given by set containment. Figure 4 shows the assignment space and its lattice structure for
n = 2, m = 2, x1 = 0, and x2 = 1. Note that the lattice includes both incomplete and
inconsistent assignments.
Using the standard correspondence between sets of enumerated elements and binary
sequences, in which a 1 in the nth place corresponds to inclusion of the nth element and a
0 corresponds to exclusion, standard basis vectors for a quantum state space can be put in
one to one correspondence with the sets. For example, Figure 5 shows the lattice of Figure
4 rewritten in ket notation where the elements v1 = 0, v1 = 1, v2 = 0 and v2 = 1 have
156
CHAPTER III. QUANTUM COMPUTATION
|1111!
$$!!
$$$ #" !!!
$$ # " !!
$$ # " !!
|1110! |1101! |1011! |0111!
& &'''( '(((''' (((
(' & &'''' &&%%
& ((( '' (' (( &''' &''' %
&&
(((( (((( ' &'
&' '
' &'
&' ''
''%
'%
'%
&
( ( & '& '
|1100! |1010! |1001! |0110! |0101! |0011!
'
%' ''' ' ( ((
%''' && ''''
'' && (((( ((& &
%% ''' && ''' &&(' (' (' ( (((&&
%%&&''''( &( &'
('(' '(
'((('''&&
|1000! |0100! |0010! |0001!
!! $$
!! " # $$
!! " #$$$
!!
"#$
!$
|0000!
Fig. 5. Lattice of variable assignments in ket form

Figure III.35: Lattice of binary strings corresponding to all subsets of a 4-
element Hogg
set. takes
[source: Rieffelquantum
in designing & Polak (2000)]for constraint satisfaction problems is to be-
algorithms
gin with all the amplitude concentrated in the |0 . . . 0! state and to iteratively move ampli-
tude up the lattice from sets to supersets and away from sets that violate the constraints.
Note that this algorithm begins differently than Shor’s algorithm and Grover’s algorithm,
which both begin by computing a function on a superposition of all the input values at
once.
algorithm Hogg:
Hogg gives two ways [Hogg 1996; Hogg 1998] of constructing a unitary matrix for
moving amplitude up the lattice. We will describe both methods, and then describe how he
moves amplitude away from bad sets.
Initialization:
Moving The algorithm
amplitude begins1. with
up: Method Thereall theobvious
is an amplitude concentrated
transformation that moves
amplitude
in the bottom offrom
the sets to supersets.
lattice, |0 · · · 0iAny amplitude
(i.e., the emptyassociated to assignments).
set of the empty set is evenly
distributed among all sets with a single element. Any amplitude associated to a set with a
single element is evenly distributed among all two element sets which contain that element
and so on. For the lattice of a three element set
Movement: The algorithm proceeds by moving amplitude up the lattice,
while avoiding assignments that violate&|111!
the%constraints; that is, we want
& % For example, we want to
&& %%
to move amplitude from a set to its supersets.
redistribute the amplitude from |1010i
|011! to |1110i
|101! |110!|1011i. Hogg has de-
and
veloped several methods. One method% & %
is based &
on the assumption that the
%
&&% %
&&% ⊗mn
transformation has the form W DW , where W = H% , the mn-dimensional
& % &
Walsh-Hadamard transformation, |001!and D|010! |100! The elements of D
is diagonal.
%% &&
depend on the size of the sets. Recall%(D.1.b,
% && p. 129) that
|000!
We want to transform 1 X
W |xi = √ √
(−)x·z |zi.
2mn
|000! → 1/ z∈MN
3(|001! + |010! + |100!
As shown in Sec. A.2.c (p. 70), we can derive a matrix representation for W :
Wjk = hj | W | ki
1 X
= hj| √ (−)k·z |zi
2mn z∈MN
1 X
= √ (−)k·z hj | zi
mn
2 z∈MN
1
= √ (−1)k·j .
2 mn
Note that k · j = |k ∩ j|, where on the right-hand side we interpret the bit
strings as sets.
The general approach is to try to steer amplitude away from sets that
violate the constraints, but the best technique depends on the particular
problem. One technique is to invert the phase on bad subsets so that they
tend to cancel the contribution of good subsets to supersets. This could be
done by a process like Grover’s algorithm using a predicate that tests for
violation of constraints. Another approach is to assign random phases to
bad sets.
It is difficult to analyze the probability that an iteration of a heuristic
algorithm will produce a solution, and so its efficiency is usually evaluated
empirically, but empirical tests will be difficult to apply to quantum heuristic
search until larger quantum computers are available, since classical computers
require exponential time to simulate quantum systems. Small simulations,
however, indicate that Hogg’s algorithms may provide polynomial speedup
over Grover’s algorithm.
Figure III.36: Effects of decoherence on a qubit. On the left is a qubit |yi that
is mostly isoloated from its environment |Ωi. On the right, a weak interaction
between the qubit and the environment has led to a possibly altered qubit
|xi and a correspondingly (slightly) altered environment |Ωxy i.
D.5 Quantum error correction

D.5.a Motivation
Quantum coherence is very difficult to maintain for long.18 Even weak inter-
actions with the environment can affect the quantum state, and we’ve seen
that the amplitudes of the quantum state are critical to quantum algorithms.
On classical computers, bits are represented by very large numbers of parti-
cles (but that is changing). On quantum computers, qubits are represented
by atomic-scale states or objects (photons, nuclear spins, electrons, trapped
ions, etc.). They are very likely to become entangled with computationally
irrelevant states of the computer and its environment, which are out of our
control. Quantum error correction is similar to classical error correction in
that additional bits are introduced, creating redundancy that can be used to
correct errors. It is different from classical error correction in that: (a) We
want to restore the entire quantum state (i.e., the continuous amplitudes),
not just 0s and 1s. Further, errors are continuous and can accumulate. (b)
It must obey the no-cloning theorem. (c) Measurement destroys quantum
information.
18
This section follows Rieffel & Polak (2000).
D.5.b Effect of decoherence

Ideally the environment |Ωi, considered as a quantum system, does not inter-
act with the computational state. But if it does, the effect can be categorized
as a unitary transformation on the environment-qubit system. Consider de-
coherence operator D describing a bit flip error in a single qubit (Fig. III.36):

|Ωi|0i =⇒ |Ω00 i|0i + |Ω10 i|1i
D: .
|Ωi|1i =⇒ |Ω01 i|0i + |Ω11 i|1i
In this notation the state vectors |Ωxy i are not normalized, but incorporate
the amplitudes of the various outcomes. In the case of no error, |Ω00 i =
|Ω11 i = |Ωi and |Ω01 i = |Ω10 i = 0. If the entanglement with the environment
is small, then kΩ01 k, kΩ10 k 1 (small exchange of amplitude).
def
Define decoherence operators Dxy |Ωi = |Ωxy i, for x, y ∈ 2, which describe
the effect of the decoherence on the environment. (These are not unitary, but
are the products of scalar amplitudes and unitary operators for the various
outcomes.) Then the evolution of the joint system is defined by the equations:
D|Ωi|0i = (D00 ⊗ I + D10 ⊗ X)|Ωi|0i,

D|Ωi|1i = (D01 ⊗ X + D11 ⊗ I)|Ωi|1i.
Alternately, we can define it:
D = D00 ⊗ |0ih0| + D10 ⊗ |1ih0| + D01 ⊗ |0ih1| + D11 ⊗ |1ih1|.
Now, it’s easy to show (Exer. III.19):

1 1 1 1
|0ih0| = (I + Z), |0ih1| = (X − Y ), |1ih0| = (X + Y ), |1ih1| = (I − Z),
2 2 2 2

0 −1
where Y = . Therefore
1 0
1
D = [D00 ⊗ (I + Z) + D01 ⊗ (X − Y ) +
2
D10 ⊗ (X + Y ) + D11 ⊗ (I − Z)]
1
= [(D00 + D11 ) ⊗ I + (D10 + D01 ) ⊗ X +
2
(D10 − D01 ) ⊗ Y + (D00 − D11 ) ⊗ Z].
Therefore the bit flip error can be described as a linear combination of the
Pauli matrices. It is generally the case that the effect of decoherence on a
single qubit can be described by a linear combination of the Pauli matrices,
which is important, since qubits are subject to various errors beside bit flips.
This is a distinctive feature about quantum errors: they have a finite basis,
and because they are unitary, they are therefore invertible. In other words,
single-qubit errors can be characterized in terms of a linear combination
of the Pauli matrices (which span the space of 2 × 2 self-adjoint unitary
matrices: C.2.a, p. 105): I (no error), X (bit flip error), Y (phase error), and
Z = Y X (bit flip phase error).PTherefore a single qubit error is represented
by e0 σ0 + e1 σ1 + e2 σ2 + e3 σ3 = 3j=0 ej σj , where the σj are the Pauli matrices
(Sec. C.2.a, p. 105).
D.5.c Correcting the quantum state

We consider a basis set of unitary “error operators” Ej , so that the error
def P
transformation is a superposition E = j ej Ej . In the more general case
of quantum registers, the Ej affect the entire register, not just a single qubit.
algorithm quantum error correction:
Encoding: An n-bit register is encoded in n + m bits, where the extra bits

def
are used for error correction. Let y = C(x) ∈ 2m+n be the n + m bit code
for x ∈ 2n . As in classical error correcting codes, we embed the message in
a space of higher dimension.
Error process: Suppose ỹ ∈ 2m+n is the result of error type k, ỹ = Ek (y).
Syndrome: Let k = S(ỹ) be a function that determines the error syndrome,

which identifies the error Ek from the corrupted code. That is, S(Ek (y)) = k.
Correction: Since the errors are unitary, and the syndrome is known, we
−1
can invert the error and thereby correct it: y = ES(ỹ) (ỹ).
Figure III.37: Circuit for quantum error correction. |ψi is the n-qubit quan-
tum state to be encoded by C, which adds m error-correction qubits to yield
the encoded state |φi. E is a unitary superposition of error operators Ej ,
which alter the quantum state to |φ̃i. S is the syndrome extraction operator,
which computes a superposition of codes for the errors E. The syndrome
register is measured, to yield a particular syndrome code j ∗ , which is used to
select a corresponding inverse error transformation Ej−1
∗ to correct the error.
Quantum case: Now consider the quantum case, in which the state |ψi is a
superposition
P of basis vectors, and the error is a superposition of error types,
E = j ej Ej . This is an orthogonal decomposition of E (see Fig. III.37).
def
Encoding: The encoded state is |φi = C|ψi|0i. There are several require-
ments for a useful quantum error correcting code. Obviously, the codes for
orthogonal inputs must be orthogonal; that is, if hψ | ψ 0 i = 0, then C|ψ, 0i
and C|ψ 0 , 0i are orthogonal: hψ, 0|C † C|ψ 0 , 0i = 0. Next, if |φi and |φ0 i are
the codes of distinct inputs, we do not want them to be confused by the error
processes, so hφ|Ej† Ek |φ0 i = 0 for all i, j. Finally, we require that for each
pair of error indices j, k, there is a number mjk such that hφ|Ej† Ek |φi = mjk
for every code |φi. This means that the error syndromes are independent of
the codes, and therefore the syndromes can be measured without collapsing
superpositions in the codes, which would make them useless for quantum
computation.
def
Error process: Let |φ̃i = E|φi be the code corrupted by error.
428 Quantum error-correction
the logical |0! and logical |1! states, not the physical zero and one states. A circuit
162 CHAPTER
performing this encoding is illustrated in FigureIII.
10.2.QUANTUM COMPUTATION
|ψ" • •
|0" ⊕
|0" ⊕
Figure 10.2. Encoding circuit for the three qubit bit flip code. The data to be encoded enters the circuit on the top
line.
Figure III.38: Quantum encoding circuit for triple repetition code. [source:
NC]
Exercise 10.1: Verify that the encoding circuit in Figure 10.2 works as claimed.
Suppose theextraction:
Syndrome initial state a|0!Apply
+ b|1! has
thebeen perfectly encoded
syndrome as a|000!
extraction + b|111!.
operator to the
Each of the three qubits is passed through an independent copy of
encoded state, augmented with enough extra qubits to represent the set ofthe bit flip channel.
Suppose a bit flip occurred on one or fewer of the qubits. There is a simple two stage
syndromes. This yields a superposition of syndromes:
error-correction procedure which ! can be used to recover the correct quantum state in
this case: X X X
S|φ̃, 0i = S ej Ej |φi ⊗ |0i = ej (SEj |φi|0i) = ej (Ej |φi|ji).
(1) Error-detectionj or syndrome diagnosis: Wej perform a measurement j which tells us
what error, if any, occurred on the quantum state. The measurement result is called
the error syndrome. For the bit flip channel there are four error syndromes,
corresponding to the four projection operators:
Measurement: Measure the syndrome register to obtain some j ∗ and the
≡ |000!#000| + |111!#111| no error (10.5)
collapsed state EjP∗0|φi|j ∗ 19
i.
P1 ≡ |100!#100| + |011!#011| bit flip on qubit one (10.6)
P2 ≡ |010!#010| + |101!#101| bit flip on qubit two (10.7)
Correction: Apply Ej−1
∗ to correct the error.
P3 ≡ |001!#001| + |110!#110| bit flip on qubit three. (10.8)

Suppose for example that a bit flip occurs on qubit one, so the corrupted state is
a|100! + b|011!. Notice that #ψ|P1 |ψ! = 1 in this case, so the outcome of the
measurement result (the error
Note the remarkable syndrome)
fact that is certainly
although there1. was
Furthermore, the syndrome
a superposition of er-
measurement does not cause any change to the state: it is a|100! + b|011! both
rors, we only have to correct one of them to get the original state back. This
before and after syndrome measurement. Note that the syndrome contains only
is because measurement
information of the
about what error haserror syndrome
occurred, and doescollapses
not allow into
us to ainfer
state affected
anything
by just that one error.
about the value of a or b, that is, it contains no information about the state being
protected. This is a generic feature of syndrome measurements, since to obtain
D.5.d information
Exampleabout the identity of a quantum state it is in general necessary to
perturb that state.
We’ll work through
(2) Recovery: We use an the example
value of thetoerror
illustrate
syndrome thetoerror
tell uscorrection process.
what procedure to useFor
an example, suppose
to recover the we use
initial state. For aexample,
simpleif triple redundancy
the error syndrome wascode that assigns
1, indicating a
19 bit flip on the first qubit, then we flip that qubit again, recovering the original state
As we mentioned the discussion of in Shor’s algorithm (p. 140), it is not necessary to
a|000!
actually + b|111!
perform the with perfect accuracy.
measurement; the sameThe fourcan
effect possible error syndromes
be obtained by unitaryand the
operations.
recovery procedure in each case are: 0 (no error) – do nothing; 1 (bit flip on first
qubit) – flip the first qubit again; 2 (bit flip on second qubit) – flip the second qubit
|0i 7→ |000i and |1i 7→ |111i. This is accomplished by a simple quantum gate
array:
C|0i|00i = |000i, C|1i|00i = |111i.
This is not a sophisticated code! It’s called a repetition code. The three-qubit
codes are called logical zero and logical one (See Fig. III.38). This code can
correct single bit flips (by majority voting); the errors are represented by the
operators:
E0 = I ⊗I ⊗I
E1 = I ⊗I ⊗X
E2 = I ⊗X ⊗I
E3 = X ⊗ I ⊗ I.
The following works as a syndrome extraction operator:

def
S|x3 , x2 , x1 , 0, 0, 0i = |x3 , x2 , x1 , x1 ⊕ x2 , x1 ⊕ x3 , x2 ⊕ x3 i.
The ⊕s compare each pair of bits, and so the ⊕ will be zero if the two bits
are the same (the majority). The following table shows the bit flipped (if
any), the corresponding syndrome, and the operator to correct it (which is
the same as the operator that caused the error):
bit flipped syndrome error correction

none |000i I ⊗I ⊗I
1 |110i I ⊗I ⊗X
2 |101i I ⊗X ⊗I
3 |011i X ⊗I ⊗I
(Note that the correction operators need not be the same as the error oper-
ators, although they are in this case.)
For an example, suppose we want to encode the state |ψi = √12 (|0i − |1i).
Its code is |φi = √12 (|000i − |111i). Suppose the following error occurs:
E = 54 X ⊗ I ⊗ I + 53 I ⊗ X ⊗ I (that is, the bit 3 flips with probability 16/25,
and bit 2 with probability 9/25). The resulting error state is
|φ̃i = E|φi

4 3 1
= X ⊗ I ⊗ I + I ⊗ X ⊗ I √ (|000i − |111i)
5 5 2
4 3
= √ X ⊗ I ⊗ I(|000i − |111i) + √ I ⊗ X ⊗ I(|000i − |111i)
5 2 5 2
4 3
= √ (|100i − |011i) + √ (|010i − |101i).
5 2 5 2
Applying the syndrome extraction operator yields:

4 3
S|φ̃, 000i = S √ (|100000i − |011000i) + √ (|010000i − |101000i)
5 2 5 2
4 3
= √ (|100011i − |011011i) + √ (|010101i − |101101i)
5 2 5 2
4 3
= √ (|100i − |011i) ⊗ |011i + √ (|010i − |101i) ⊗ |101i
5 2 5 2
Measuring the syndrome register yields either |011i (representing an error in
bit 3) or |101i (representing an error in bit 2). Suppose we get |011i. The
state collapses into:
1
√ (|100i − |011i) ⊗ |011i.
2
Note that we have projected into a subspace for just one of the two bit-flip
errors that occurred (the flip in bit 3). The measured syndrome |011i tells
us to apply X ⊗ I ⊗ I to the first three bits, which restores |φi:
1 1
(X ⊗ I ⊗ I) √ (|100i − |011i) = √ (|000i − |111i) = |φi.
2 2
We can do something similar to correct single phase flip (Z) errors by using
the encoding |0i 7→ | + ++i, |1i 7→ | − −−i (Exer. III.54). To see this, recall
that Z in the sign basis is the analog of X is the computational basis.
D.5.e Discussion
There is a nine-qubit code, called the Shor code, that can correct arbitrary
errors on a single qubit, even replacing the entire qubit by garbage (Nielsen
& Chuang, 2010, §10.2). The essence of this code is that triple redundancy
is used to correct X errors, and triple redundancy again to correct Z errors,
thus requiring nine code qubits for each logical qubut. Since Y = ZX and
the Pauli matrices are a basis, this code is able to correct all errors.
Quantum error correction is remarkable in that an entire continuum of
errors can be corrected by correcting only a discrete set of errors. This
works in quantum computation, but not classical analog computing. The

general goal in syndrome extraction is to separate the syndrome information
from the computational information in such a way that the syndrome can
be measured without collapsing any of the computational information. Since
the syndrome is unitary, it can be inverted to correct the error.
What do we do about noise in the gates that do the encoding and decod-
ing? It is possible to do fault-tolerant quantum computation. “Even more
impressively, fault-tolerance allow us to perform logical operations on en-
coded quantum states, in a manner which tolerates faults in the underlying
gate operations.” (Nielsen & Chuang, 2010, p. 425) Indeed, “provided the
noise in individual quantum gates is below a certain constant threshold it
is possible to efficiently perform an arbitrarily large quantum computation.”
(Nielsen & Chuang, 2010, p. 425)20
20
See Nielsen & Chuang (2010, §10.6.4).
E Abrams-Lloyd theorem
E.1 Overview
All experiments to date confirm the linearity of QM, but what would be
the consequences of slight nonlinearities?21 We will see that nonlinearities
can be exploited to solve NP problems (and in fact harder problems) in
polynomial time. This fact demonstrates clearly that computability and
complexity are not purely mathematical matters. Because computation is
inherently physical, fundamental physics is intertwined with fundamental
computation theory. It also exposes the fact that there are hidden physical
assumptions in the traditional theory of computation.
How could nonlinearities be exploited in quantum computation? Recall
that quantum state vectors lie on the unit sphere and unitary transformations
preserve “angles” (inner products) between vectors. Nonunitary transforma-
tions would, in effect, stretch the sphere, so the angles between vectors could
change. Unitary transformations could be used to position the vectors to the
correct position for subsequent nonunitary transformations. The following
algorithm exploits a nonlinear operator to separate vectors that are initially
close together.
E.2 Basic algorithm

The Lyapunov exponent λ describes the divergence of trajectories in a dy-
namical system. If ∆θ(0) is the initial separation, then the separation after
time t is given by |∆θ(t)| ≈ eλt |∆θ(0)|. If λ > 0, the system is usually
chaotic.
Suppose there is some nonlinear operation N with a positive Lyapunov
exponent over some finite region of the unit sphere. Further suppose we
have an oracle P : 2n → 2. We want to determine if there is an x such
that P (x) = 1. Next suppose we are given a quantum gate array UP as in
21
This section is based on Daniel S. Abrams and Seth Lloyd (1998), “Non-
linear quantum mechanics implies polynomial-time solution for NP-complete and
#P problems.” Phys. Rev. Lett. 81, 3992–3995 (1998), preprint available at
http://arxiv.org/abs/quant-ph/9801041v1. See also Scott Aaronson, “NP-complete
Problems and Physical Reality,” SIGACT News, Complexity Theory Column, March
2005. quant-ph/0502072. http://www.scottaaronson.com/papers/npcomplete.pdf
(accessed 2012-10-27).
E. ABRAMS-LLOYD THEOREM 167
Figure III.39: Quantum circuit for Abrams-Lloyd algorithm. The first mea-
surement produces 0 with probability greater than 1/4, but if it yields a
nonzero state, we try again. The N parallelogram represents a hypothetical
nonlinear quantum state transformation, which may be repeated to yield a
macroscopically observable separation of the solution and no-solution vectors.
Grover’s algorithm. It is defined on a n-qubit data register and a 1-qubit

result register. See Fig. III.39.
algorithm Abrams-Lloyd:
Step 1: As usual, apply the Walsh-Hadamard transform to a zero data

register to get a superposition of all possible inputs:
1 X
|ψ0 i = (Wn |0i)|0i = √ |x, 0i.
2n x∈2n
Step 2 (apply oracle): Apply the oracle to get a superposition of input-

output pairs:
1 X
|ψ1 i = UP |ψ0 i = √ |x, P (x)i.
2n x∈2n
Step 3 (measure data register): First, apply the Walsh transformation

to the data register to get:
|ψ2 i = (Wn ⊗ I)|ψ1 i
1 X
= √ (Wn |xi)|P (x)i
2n x∈2n
" #
1 X 1 X
= √ √ (−)x·z |zi |P (x)i.
n
2 x∈2n n
2 z∈2n
The last step applies by Eq. III.24 (p. 129). That is,
X X 1
|ψ2 i = n
(−)x·z |zi|P (x)i.
x∈2n z∈2n
2
Separate out the state z = 0 in order to see its amplitude:

X 1 X 1
x·0
n
(−) |0i|P (x)i = n
|0i|P (x)i.
x∈2 n
2 x∈2 n
2
At least half of the 2n vectors x must have the same value, a = P (x) (since
P (x) ∈ 2). Therefore the amplitude of |0, ai is at least 1/2, and the proba-
bility of observing |0, ai is at least 1/4. We get a non-zero data register with
probability ≤ 3/4. (If we happen to observe |x0 , 1i, then of course we have
our answer.)
In the case in which we get a zero data register, measurement of the data
register yields the state:

≥ 14 −1 s 2n − s −1 s 2n − s
|ψ2 i −→ Z |0i|1i + |0i|0i = Z |0i n |1i + |0i ,
2n 2n 2 2n
where s is the number of solutions (the number of x such that P (x) = 1)

and Z −1 renormalizes after the state collapse.
The information we want is in the result qubit, but if s is small (as ex-
pected), then measurement will almost always yield |0i. Recall what we did
n
in Grover’s algorithm. For s 2n , the vector Z −1 |0i 2sn |1i + 2 2−sn |0i is
very close to the vector |0, 0i. Therefore, we would like to drive them apart.
Step 4: Applying the nonlinear operator N repeatedly will separate the

vectors at an exponential rate. “[E]ventually, at a time determined by a
polynomial function of the number of qubits n, the number of solutions s,
and the rate of spreading (Lyapunov exponent) λ, the two cases will become
E. ABRAMS-LLOYD THEOREM 169
macroscopically distinguishable.”
Step 5 (measure result register): Measure the result qubit. If the vec-
tors have been sufficiently separated, there will be a significant probability
of observing |1i in the s 6= 0 case.
If η is the angular extent of the nonlinear region, it might take O((π/η)2 )

trials to get |1i with high probability. For large η, just one iteration might
be sufficient.
E.3 Discussion of further results

The preceding algorithm depends on exponential precision, but Abrams and
Lloyd present another algorithm that is robust against small errors. Each
iteration doubles the number of components that have |1i in the result qubit,
and after n iterations it yields a result with probability 1, and so the algo-
rithm is linear.
Scott Aaronson has expressed doubts that the required nonlinear OR gate
can be implemented. Abrams and Lloyd summarize: “We have demonstrated
that nonlinear time evolution can in fact be exploited to allow a quantum
computer to solve NP-complete and #P problems in polynomial time.” (#P
or “number-P” asks how many accepting paths in a NTM running in polyno-
mial time. #P problems are at least as hard as corresponding NP problems.)
Nevertheless, they continue, “we believe that quantum mechanics is in all
likelihood exactly linear, and that the above conclusions might be viewed
most profitably as further evidence that this is indeed the case.”
F Universal quantum computers

Hitherto we have used a practical definition of universality: since conven-
tional digital computers are implemented in terms of binary digital logic,
we have taken the ability to implement binary digital logic as sufficient for
universality. This leave open the question of the relation of quantum com-
puters to the theoretical standard of computational universality: the Turing
machine. Therefore, a natural question is: What is the power of a quantum
computer? Is it super-Turing or sub-Turing? Another question is: What is
its efficiency? Can it solve NP problems efficiently? There are a number
of universal quantum computing models for both theoretical and practical
purposes.
F.1 Feynman on quantum computation

F.1.a Simulating quantum systems
In 1982 Richard Feynman discussed what would be required to simulate a
quantum mechanical system on a digital computer.22 First he considered a
classical probabilistic physical system. Suppose we want to use a conven-
tional computer to calculate the probabilities as the system evolves in time.
Further suppose that the system comprises R particles that are confined to
N locations in space, and that each configuration c has a probability p(c).
There are N R possible configurations, since a configuration assigns a location
N to each of the R particles (i.e., the number of functions R → N ). There-
fore to simulate all the possibilities would require keeping track of a number
of quantities (the probabilities) that grows exponentially with the size of the
system. This is infeasible.
So let’s take a weaker goal: we want a simulator that exhibits the same
probabilistic behavior as the system. Our goal is that if we run both of them
over and over, we will see the same distribution of behaviors. This we can
do. You can implement this by having a nondeterministic computer that has
the same state transition probabilities as the primary system.
Let’s try the same trick with quantum systems, that is, have a conven-
tional computer that exhibits the same probabilities as the quantum system.
If you do the math (which we won’t), it turns out that this is impossible.
The reason is that, in effect, some of the state transitions would have to have
22
This section is based primarily on Feynman (1982).
For example, if the value is zero we cannot tell whether it came from
(a, b) = (0, 0) or from (1, 1) but we keep the other line a' = a to resolve the
ambiguity.
We will represent the C O N T R O L L E D N O T by putting a 0 on the
control wire, connected with a vertical line to an X on the wire which is
controlled.
This element can also supply us with F A N O U T , for if b = 0 we see
that a is copied onto line b'. This C O P Y function will be important later
F. UNIVERSAL QUANTUM
on. It also supplies us withCOMPUTERS
E X C H A N G E , for three of them used 171
a ..... o
b ,,, SUM
o ^ CARRY
Fig. 4. Adder.
Figure III.40: Simple adder using reversible logic. [fig. from Feynman (1986)]
Quantum Mechanical Computers 511
0 ~ C I ...... (3 = 0I
SI
b A s'- --l---b b'
C SUM c'
C1
d=o' CARRY = d'
Fig. 5. Full adder.
Figure III.41: Full adder using reversible logic. [fig. from Feynman (1986)]
successively on a pair of lines, but with alternate choice for control line,
accomplishes an exchange of the information on the lines (Fig. 3b).
It turns out that combinations of just these two elements alone
what are
amount to negative
insufficient probabilities,
to accomplish arbitrary and
logicalwefunctions.
don’t know Somehow to do this
element
classically. We’ve
involving threeseen
lineshow in quantum
is necessary. mechanics,
We have chosen whatprobabilities
we can callcanthein effect
cancelC O
byN Tdestructive
ROLLED CO interference
N T R O L L E D ofNOT.
the Here
wavefunctions.
(see Fig. 3c) weThehave
conclusion
two is
control lines a, b, which appear unchanged in the output
that no conventional computer can efficiently simulate a quantum computer. and which change
the third line c to N O T c only if both lines are activated (a = 1 and b = 1).
Therefore, if we want to (efficiently) simulate any physical system, we need
Otherwise c ' = c. If the third line input c is set to 0, then evidently it
a quantum computer.
becomes 1(c' = 1) only if both a and b are 1 and therefore supplies us with
the AND function (see Table I).
Three combinations
F.1.b Universal quantum for (a, computer
b), namely (0, 0), (0, 1), and (1, 0), all give
the same value, 0, to the AND (a, b) function so the ambiguity requires
twoFeynman
In 1985 bits to resolve it. These
described are kept
several in the lines
possible designsa, b in
forthea universal
output so the
quantum
function
23 can be reversed (by itself, in fact). The AND function is the carry
computer. He observes that NOT, CNOT, and CCNOT are sufficient for
bit for the sum of a and b.
any logic From
gate, these
as well as for COPY and EXCHANGE, and therefore for
elements it is known that any logical circuit can be put
universal computation.
together by using them Heinexhibits circuits
combination, andforin afact,
simple adder science
computer (Fig. III.40)
and a full adder (Fig. III.41).
The goal is to construct a Hamiltonian to govern the operation of a quan-
Table I.
tum computer. Feynman describes quantum logic gates in terms of two prim-
itive operations, which change
a b
thec stateo ' of b'an ¢'“atom” (two-state system or
“wire”). Letters near the beginning of the alphabet (a, b, c, . . .) are used for
0 0 0 0 0 0
23
This section is based primarily
0 on
0 Feynman
1 0 (1986).
0 1
0 1 0 0 1 0
0 1 1 0 1 1
1 0 0 1 0 0
1 0 1 1 0 l
1 1 0 1 1 1
1 1 1 1 1 0
data or register atoms, and those toward the end (p, q, r, . . .) for program
atoms (which are used for sequencing operations). In this simple sequential
computer, only one program atom is set at a time.
For a single line a, the annihilation operator is defined:

0 1
a= = |0ih1|.
0 0
The annihilator changes the state |1i to |0i; typically it lowers the energy of
a quantum system. Applied to |0i, it leaves the state unchanged and returns
the zero vector 0 (which is not a meaningful quantum state). It matches
|1i and resets it to |0i. The operation is not unitary (because not norm
preserving). It is a “partial NOT” operation.
Its conjugate transpose it the creation operation:24

∗ 0 0
a = = |1ih0|.
1 0
The creator transforms |0i to |1i, but leaves |1i alone, returning 0; typically
it raises the energy of a quantum system. It matches |0i and sets it to |1i.
This is the other half of NOT = a + a∗ .
Feynman also defines a number operation or 1-test: Consider25

∗ 0 0
Na = a a = = |1ih1|.
0 1
This has the effect of returning |1i for input |1i, but 0 for |0i:
Na = a∗ a = |1ih0| |0ih1| = |1ih0 | 0ih1| = |1ih1|.
Thus it’s a test for |1i. (This is a partial identity operation.)

Similarly, the 0-test is defined:26

∗ 1 0
1 − Na = aa = = |0ih0|.
0 0
24
Note that Feynman uses a∗ for the adjoint (conjugate transpose) of a.
25
This matrix is not the same as that given in Feynman (1982, 1986), since Feynman
uses the basis |1i = (1, 0)T , |0i = (0, 1)T , whereas we use |0i = (1, 0)T and |1i = (0, 1)T .
26
This matrix is different from that given in Feynman (1982, 1986), as explained in the
previous footnote.
been studied in normal computers. But until we find a specific implemen-
tation for this computer, I do not know how to proceed to analyze these
effects. However, it appears that they would be very important, in practice.
This computer seems to be very delicate and these imperfections may
produce considerable havoc.
The time needed to make a step of calculation depends on the strength
or the energy of the interactions in the terms of the Hamiltonian. If each of
the terms in the Hamiltonian is supposed to be of the order of 0.1 electron
F. UNIVERSAL QUANTUM
volts, then it appears that the COMPUTERS
time for the cursor to make each step, if done 173
<o
in a ballistic fashion, is of the order 6 x 10 -15 sec. This does not represent
C
...... q
IF c = t GO p TO q AND PUT c =O
p ~ IF c =O GO p TO r AND PUT c = I
I
IF c = I GO r TO p AND PUT c : O
H : q* cp + r*c*p IF c =O GO q TO p AND PUT c = I

+ p*c*q + p * c r
Fig. 7. Switch.
Figure III.42: Switch element. 0/1 annotations on the wires show the c
values. [fig. from Feynman (1986)]
(Feynman writes this 1 − Na because he writes 1 = I.) This has the effect
of returning |0i for input |0i, but 0 for |1i. This is test for |0i. (It is the rest
of the identity operation.)
The two operations a and a∗ are sufficient for creating all 2 × 2 matrices,
and therefore all transformations on a single qubit. Note that

w x
= waa∗ + xa + ya∗ + za∗ a.
y z
Feynman writes Aa for the negation (NOT) operation applied to a. Obvi-

ously, Aa = a + a∗ (it annihilates |1i and creates from |0i) and 1 = aa∗ + a∗ a
(it passes |0i and passes |1i. We can prove that Aa Aa = 1 (Exer. III.55).
Feynman writes Aa,b for the CNOT operation applied to lines a and b.
Observe, Aa,b = a∗ a(b + b∗ ) + aa∗ . Notice that this is a tensor product on
the register |a, bi: Aa,b = a∗ a ⊗ (b + b∗ ) + aa∗ ⊗ 1. You can write this formula
Na ⊗ Ab + (1 − Na ) ⊗ 1. That is, if Na detects |1i, then it negates b. If 1 − Na
detects |0i, then it leaves b alone.
Feynman writes Aab,c for the CCNOT operation applied to lines a, b, and
c. Note that Aab,c = 1 + a∗ ab∗ b(c + c∗ − 1) (Exer. III.56). This formula is
more comprehensible in this form:
Aab,c = 1 + Na Nb (Ac − 1).
One of Feynman’s universal computers is based on only two logic gates,

Not and Switch (Fig. III.42). If c = |1i, then the “cursor” (locus of control)
524 Feynman
O a
sMiNoT01tM
s =b+ t
I sN tN I
Fig. 8. CONTROLLED NOT by switches,
Figure III.43: CNOT implemented by switches. 0/1 annotations on the wires
show the a values. [fig. from Feynman (1986)]
In these diagrams, horizontal or vertical lines will represent program
atoms. The switches are represented by diagonal lines and in boxes we'll
put the other matrices that operate on registers such as the N O T b. To be
at p moves to the
specific, q, but if c = |0i
Hamiltonian foritthis
moves to r. Itofalso
little section a C Onegates
N T R O L Lc Ein
D the process.
NOT,
thinking of it as starting at
It’s also reversible (see Fig. III.42). s and ending at t, is given below:
The switch is aHe(s,
tensor product on |c, p, q, ri:
t) = s ' a s + t*a*tM + t*(b + b*) sM + s~va*s
q ∗ cpnt- +
l ' a t ∗u q-
r c∗ pt~rs
+ u[p+∗ cC.C
∗
q + p∗ cr].
(The c.c means to add the complex conjugate of all the previous terms.)
(The bracketed Although expression
there seemis tojustbe the
two complex
routes hereconjugate
which would of the first part,
possibly
required for reversibility.)
produce Read the factors
all kinds of complications in each
characteristic of term
quantumfrommechanics,
right to left:
(1) q ∗ cp:
this ifis pnotand
so. cIf are
the entire computer
set, then unsetsystem
themisand started
setinq.a definite state for
(2) r∗ ca∗ p:
by if thep time
is settheand
cursor
c isreaches s, the
not set, theatom
unseta ispstill
andin set
some definite
c and r. state
(although possibly different from its initial state due to previous computer
Fig.operations
III.43 shows CNOT implemented by switches. This is the controlled-
on it). Thus only one of the two routes is taken. The expression
NOT applied to data by
may be simplified a, bomitting
and sequenced by cursor
the S * I N term atoms
and putting t u = Ss,N.
t (= start, ter-
minate). If One a need
= 1 not thebecursor state
concerned in moves
that case,along theroute
that one top isline, and(two
longer if a = 0
along thecursor sites) than
bottom. If itthe otheralong
moves (one cursor
the top, site)then
for it
again there bis+no
applies b∗ inter-
to negate b
ference.
(otherwise No scattering
leaving is produced
it alone). In eitherin case,
any casethebycursor
the insertion
arrivesinto at athe
chain
reversed
of coupled sites, an extra piece of chain of any number of sites with the
switch,samewhere sets the next cursor atom t. We can write it
mutual coupling between sites (analogous to matching impedances in
transmission lines).
∗
Ha,b (s, t)To
= sstudy, + t∗ athings
M as these
∗
t∗M (b +web∗think
tM +further, ∗ ∗
)sM +ofsputting t∗ atNtogether.
N a s +pieces + t∗N sNA+ c.c,
piece (see Fig. 9) M might be represented as a logical unit of interacting
where parts
“c.c”in means
which weto only
add represent
the complex
the firstconjugates
input cursorofsite
theas preceding
sM and the terms.
final one at the other end as tM. All the
Read the factors in each term from right to left: rest of the program sites that are
∗ between sM and tM are considered internal parts of M, and M contains its
(1) sM as: if s and a are set, then unset them and set sM .
registers. Only sM and t M are sites that may be coupled externally.
(4) s∗N a∗ s: if s is set and a in unset, then unset s and set sN and a.
(6) t∗N sN : if sN is set, then unset it and set tN .
(3) t∗M (b + b∗ )sM : if sM is set, then unset it, negate b and set tM .
The first thing we do is to copy (using C O N T R O L L E D NOT's) our
external input into M. Then M operates, and the cursor goes on the top
line in our drawing. It copies the output out of M into the external output
register. M now contains garbage. Next it changes f to N O T f, comes down
on the other line of the switch, backs out through M clearing the garbage,
and uncopies the input again.
When you copy data and do it again, you reduce one of the registers
to 0, the register into which you coied the first time. After the coying, it
F. UNIVERSAL
goes out (sinceQUANTUM COMPUTERS
f is now changed) on the other line where we restore f to 0 175
f f
coPY I
i ~ NOT f IJ
Fig. 13. Garbage clearer.

Figure III.44: Garbage clearer. 0/1 annotations on the wires show the f
values. [fig. from Feynman (1986)]
(5) t∗ atN : if tN and a are set (as a must be to get here), then unset them
and set t.
(2) t∗ a∗ tM : if tM is set and a is unset (as it must be to get here), then reverse
their states and set t.
(The t∗N sN term can be eliminated by setting tN = sN .)
F.1.c Garbage clearer

Instead of having a separate copy of the machine to clear out the garbage,
it’s possible to run the same machine backwards (Fig. III.44). An external
register In contains the input, and the output register Out and all machine
registers are initialized to 0s. Let s be the starting program atom. The flag
f is initially 0.
The f = 0 flag routes control through the reversed switch (setting f = 1)
to Copy. The Copy box uses CNOTs to copy the external input into M .
Next M operates, generating the result in an internal register. At the end of
the process M contains garbage.
The f = 1 flag directs control into the upper branch (resetting f = 0),
which uses CNOTs to copy the result into the external output register Out.
Control then passes out from the upper branch of the switch down and back
into the lower branch, which negates f , setting it to f = 1. Control passes
back into the machine through the lower switch branch (resetting f = 0), and
backwards through M , clearing out all the garbage, restoring all the registers
to 0s. It passes backwards through the Copy box, copying the input back
from M to the external input register In. This restores the internal registers
to 0s. Control finally passes out through the lower branch of the left switch
(setting f = 1), but it negates f again, so f = 0. It arrives at the terminal
program atom t. At the end of the process, everything is reset to the initial
conditions, except that we have the result in the Out register. Feynman also
discusses how to do subroutines and other useful programming constructs,
but we won’t go into those details.
F.2 Benioff ’s quantum Turing machine

In 1980 Paul Benioff published the first design for a universal quantum com-
puter, which was based on the Turing machine (Benioff, 1980). The tape
is represented by a finite lattice of quantum spin systems with eigenstates
corresponding to the tape symbols. (Therefore, he cannot implement an
open-ended TM tape, but neither can an ordinary digital computer.) The
read/write head is a spinless system that moves along the lattice. The state
of the TM was represented by another spin system. Benioff defined unitary
operators for doing the various operations (e.g., changing the tape). In 1982
he extended his model to erase the tape, as in Bennett’s model (Benioff,
1982). Each step was performed by measuring both the tape state under the
head and the internal state (thus collapsing them) and using this to control
the unitary operator applied to the tape and state. As a consequence, the
computer does not make much use of superposition.
F.3 Deutsch’s universal quantum computer

Benioff’s computer is effectively classical; it can be simulated by a classical
Turing machine. Moreover, Feynman’s construction is not a true univer-
sal computer, since you need to construct it for each computation, and it’s
not obvious how to get the required dynamical behavior. Deutsch sought a
broader definition of quantum computation, and a universal quantum com-
puter Q.27 M binary observables are used to represent the processor, {ňi }
for i ∈ M, where M = {0, . . . , M − 1}. Collectively they are called ň. An
infinite sequence of binary observables implements the memory, {m̌i } (i ∈ Z)
Collectively the sequence is called m̌. An observable x̌, with spectrum Z,
represents the tape position (address) of the head. The computational basis
states have the form:
def
|x; n; mi = |x; n0 , n1 , . . . , nM −1 ; . . . , m−1 , m0 , m1 , . . .i.
27
This section is based on Deutsch (1985).
F. UNIVERSAL QUANTUM COMPUTERS 177
Here the eigenvectors are labeled by their eigenvalues x, n, and m.

The dynamics of computation is described by a unitary operator U , which
advances the computation by one step of duration T :
|ψ(nT )i = U n |ψ(0)i.
Initially, only a finite number of the memory elements are prepared in a

non-zero state.
X X
|ψ(0)i = λm |0; 0; mi, where |λm |2 = 1,
m m
That is, only finitely many λm 6= 0, and in particular λm = 0 when an infinite

number of the m are non-zero. The non-zero entries are the program and its
input. Note that the initial state may be a superposition of initial tapes.
The matrix elements of U (relating the new state to the current state)
have the form:
hx0 ; n0 ; m0 | U | x; n; mi
Y m
= [δxx+1
0 U + (n0 , m0x |n, mx ) + δxx−1
0 U − (n0 , m0x |n, mx )] δm0yy .
y6=x
U + and U − represent moves to the right and left, respectively. The first
two δ functions ensure that the tape position cannot move by more than one
position in a step. The final product of deltas ensures that all the other tape
positions are unchanged; it’s equivalent to: ∀y 6= x : m0y = my . The U +
and U − functions define the actual transitions of the machine in terms of the
processor state and the symbol under the tape head. Each choice defines a
quantum computer Q[U + , U − ].
The machine cannot be observed before it has halted, since that would
generally alter its state. Therefore one of the processor’s bits is chosen as a
halt indicator. It can be observed from time to time without affecting the
computation.
Q can simulate TMs, but also any other quantum computer to arbitrary
precision. In fact, it can simulate any finitely realizable physical system to
arbitrary precision, and it can simulate some physical systems that go beyond
the power of TMs (hypercomputation).
G Quantum probability in cognition

There are interesting connections between the mathematics of quantum me-
chanics and information processing in the brain. This is not so remarkable
when we recall that the foundations of quantum mechanics are matters of
information and knowledge.28
G.1 Theories of decision making

How do people make decisions under uncertainty? There have been three
major phases of models.
(i) Logic. From Aristotle’s time, the most common model of human
thinking has been formal logic, especially deductive logic. For example, the
title of George Boole’s book, in which he introduced Boolean algebra, was
called The Laws of Thought, and that is what he supposed it to be. The first
AI program (1956) was the Logic Theorist, and formal deductive logic
still dominates many AI systems. Since the 1960s, however, there has been
accumulating psychological evidence that classical logic is not a good model
of everyday reasoning.
An additional, more technical problem is that classical logic is monotonic,
that is, the body of derived theorems can only increase. But everyday rea-
soning is nonmonotonic: propositions that were previously taken to be true
can become false (either because the facts have changed or an assumption has
been invalidated. As a consequence, the body of truths can shrink or change
in other ways. Existing truths can be nullified. An additional problem is that
much of our reasoning is inductive rather than deductive, that is, it moves
from more particular premises to more general conclusions, rather than vice
versa, as deductive logic does. But after many years of research, there really
isn’t an adequate inductive logic that accounts for scientific reasoning as well
as everyday generalization.
(ii) Classical probability (CP). The most common models of human
decision making have been based on classical probability theory (CP) and
Bayesian inference. Amos Tversky and Daniel Kahneman were pioneers
(from the 1970s) in the study of how people actually make decisions and
judgments. (In 2002 Kahneman recieved the Nobel Prize in Economics for
28
This chapter is based primarily on Emmanuel M. Pothos and Jerome R. Busemeyer,
“Can quantum probability provide a new direction for cognitive modeling?” Behavioral
and Brain Sciences (Pothos & Busemeyer, 2013), including MacLennan (2013).
G. QUANTUM PROBABILITY IN COGNITION 179
this work; Tversky had already died.) Since then many other psychologists
have confirmed and extended their findings. They concluded that every-
day human reasoning follows the laws of neither classical logic nor classical
probability theory. “Many of these findings relate to order/context effects,
violations of the law of total probability (which is fundamental to Bayesian
modeling), and failures of compositionality.” (Pothos & Busemeyer, 2013)
(iii) Quantum probability (QP). The mathematics of quantum me-
chanics provides an alternative system (axiomatization) of probability which
has the potential to account for these violations of CP, as we will see. Note
that QP is just a probability theory; there is no presumption that physical
quantum phenomena are significant in the brain (although they might be).
In this sense, our brains appear to be using a kind of quantum computation.
G.2 Framework
G.2.a Questions & outcomes
Just as CP begins by defining a sample space, QP begins by defining a Hilbert
space, which defines all possible answers that could be produced for all possi-
ble questions (addressed by the model). Corresponding to the quantum state
is the cognitive state, which you can think of as the indeterminate state of the
brain before there is a decision or determination to act in some way (such as
answering a question). Corresponding to observables in quantum mechanics,
we have questions in QP. More generally, we might refer to quandries, that
is, unsettled dispositions to act. Corresponding to projectors into subspaces
we have decisions. Often the subspaces are one-dimensional, that is, rays.
G.2.b Happiness example

Consider asking a person whether they are happy or not. Before asking the
question, they might be in an indefinite (superposition) state (Fig. III.45(a)):
|ψi = a|happyi + b|unhappyi.
It is not just that we do not know whether the person is happy or not; rather
the person “is in an indefinite state regarding happiness, simultaneously en-
tertaining both possibilities, but being uncommitted to either” (Pothos &
Busemeyer, 2013). More realistically “happy” and “unhappy” are likely to
|| PA PB |ψ 〉 ||2 Prob( B ∧ then A)
Prob( A | B) = = (this is called Lüder’s law).
|| PB |ψ 〉 ||2 Prob( B)
It is clear that the definition of conditional probability in QP theory is analogous
to that in CP theory, but for potential order effects in the sequential projection PAPB,
when A and B are incompatible.
Ha pp y, &e m p lo ye d
lo ye d
Em p
Ha pp y
Ha pp y
Ha pp y,&unem p lo yed
Unha p py Unha p py
Une m Unha p py,&em p loyed

p lo ye
a d b c
Figure 1. An illustration of basic processes in QP theory. In Figure 1b, all vectors are co-
planar, and the figure is a two-dimensional one. In Figure 1c, the three vectors “Happy,
Figure III.45:
employed,” “Happy,“An illustration
unemployed,” of basic
and “Unhappy, processes
employed” in QP theory.
are all orthogonal to each In Figure
other,
[b], allso vectors
that the figure
areis acoplanar,
three-dimensional
and one.
the(The fourthis
figure dimension, “unhappy,
a two-dimensional one. In
unemployed” is not shown).
Figure [c], the three vectors ‘Happy, employed’, ‘Happy, unemployed’, and
‘Unhappy, employed’ are all orthogonal to each other, so that the figure is
a three-dimensional
The magnitude of a one. (The
projection fourth
depends dimension,
upon the angle between‘unhappy, unemployed’ is
the corresponding
not shown).” (Pothos & Busemeyer, 2013)
subspaces. For example, when the angle is large, a lot of amplitude is lost between
successive projections. As can be seen in Figure 1b,

be complex subspaces, not rays, but for the sake of the example, we use a
2D outcome space.
Asking the question is equivalent to measuring the state in the “happiness
basis,” which comprises two projectors Phappy and Punhappy :
Phappy = |happyihhappy|,
Punhappy = |unhappyihunhappy|.
The probability that the person responds “happy” is, as expected:
kPhappy |ψik2 = k|happyihhappy | ψik2 = |a|2 .
Measurement (decision) collapses the indefinite state to a definite basis state,

|happyi, with probability |a|2 . The judgment or decision is not just a “read
out”; it is constructed from the state and the question, which actively disam-
biguates the superposition state.
G.2.c Incompatibility
As in quantum mechanics, questions can be compatible or incompatible. In
fact, Neils Bohr borrowed the notion of incompatible questions from the psy-
chologist William James. Compatible questions can be asked in any order;
they commute; incompatible questions do not commute.
In CP it is always possible to specify a joint probability distribution over
the four possible pairs of answers (the unicity principle). In QP you can do
this for compatible questions, but not for incompatible ones. Psychologically,
in the incompatible case the person cannot form a single thought for all
combinations of possible outcomes (because they are linearly dependent).
In the incompatible case, asking the first question alters the context of the
second question, and thus affects its answer. Therefore, in applying QP in
psychology, we can ask whether one decision is likely to affect the other.
Suppose we are going to ask a person two questions, whether they are
happy or not and whether they are employed or not. It is plausible that
happiness and employment are related, so we postulate a single 2D space
spanned by both bases (Fig. III.45(b)). The angle between the two bases
reflects the fact that happiness is likely to be correlated to employment.
Notice that once we get an answer regarding happiness, we will be in an
indefinite state regarding employment, and vice versa.
Suppose we ask if the subject is employed and then ask if they are happy.
The probability that they answer “yes” to both is given by:29
P{employed && happy} = P{employed} × P{happy | employed}.
The rules of QP give the first probability: P{employed} = kPemployed |ψik2 .

Asking about employment has collapsed the state, which is now
Pemployed |ψi
|ψemployed i = .
kPemployed |ψik
The probability of a happy response is then
P{happy | employed} = kPhappy |ψemployed ik2 .
Hence the probability of the two responses is
P{employed && happy} = kPhappy Pemployed |ψik2 .
From this example, we can see that the law for conditional probability in
QP, called Lüder’s Law, is:
kPA PB |ψik2 P{B && A}
P{A | B} = 2
= .
kPB |ψik P{B}
Look at Fig. III.45(b). You can see that
P{happy} < P{employed && happy},
which cannot happen in CP (since P{A} ≥ P{A ∧ B} always). The psy-

chological interpretation would be that the subject’s consciousness of be-
ing employed makes her more likely to say she is happy. This is because
happiness and employment are correlated, but this correlation does not af-
fect the outcome without the prior question about employment. In general,
P{A && B} = 6 P{B && A}, which cannot happen in CP. That is, conjunc-
tion is not commutative. You can see
P{happy && employed} < P{employed && happy}.
This is because the subject was more uncertain about their happiness than
their employment, and therefore the state vector lost a lot of its amplitude
29
As in C++, “ && ” should be read “and then” (sequential “and”).
via its projection first onto |happyi. “The size of such angles and the relative
dimensionality of the subspaces are the cornerstones of QP cognitive models
and are determined by the known psychology of the problem. These angles
(and the initial state vector) have a role in QP theory analogous to that
of prior and conditional distributions in Bayesian modeling.” (Pothos &
Busemeyer, 2013)
G.2.d Compatible questions

Fig. III.45(c) displays the case where the questions are compatible (only
three of the four basis vectors are shown). In this case we have a tensor
product between the space spanned by {|happyi, |unhappyi} and the space
spanned by {|employedi, |unemployedi}. For compatible questions the states
are composite vectors, e.g.,
|Hi = η|happyi + η 0 |unhappyi,
|Ei = |employedi + 0 |unemployedi,
|Ψi = |Hi ⊗ |Ei
= η|happyi|employedi + η0 |happyi|unemployedi
+η 0 |unhappyi|employedi + η 0 0 |unhappyi|unemployedi.
Then, for example, as in CP the joint probability
P{happy ∧ employed} = |η|2 = P{happy}P{employed}.
G.2.e Structured representations and entanglement

Many concepts seem to have structured representations, that is, components,
properties, or attributes,, which are “aligned” when concepts are compared.30
Structured concepts are naturally represented in QP by tensored spaces rep-
resenting the concept’s components. However QP also permits entangled
(non-product) states, such as
α|happyi|employedi + β|unhappyi|unemployedi.
This represents a state in which happiness and employment are strongly
interdependent. It represents a stronger degree of dependency than can be
expressed in CP. In CP you can construct a complete joint probability out
of pairwise joints, but this is not possible in QP.
30
Think of the variable components (fields) of a C++ class.
Ba nk%Te lle r
t
nis
mi
Fe
~ Ba nk% Te lle r
~F
em
inis
t
Figure 2. An illustration of the QP explanation for the conjunction fallacy.
Figure III.46: Hypothetical basis state space of the “Linda experiment.”

(Pothos & Busemeyer, 2013)3.2. Failures of commutativity in decision making
We next consider failures of commutativity in decision making, whereby asking the same
two questions in different orders can lead to changes in response (Feldman & Lynch
1988; Schuman & Presser 1981; Tourangeau et al. 1991). Consider the questions “Is
G.2.f Time evolution Clinton honest?” and “Is Gore honest?” and the same questions in a reverse order. When
the first two questions were asked in a Gallup poll, the probabilities of answering yes for
Time evolution is CP is defined by “a transition matrix (the solution to Kol-
Clinton and Gore were 50% and 68%, respectively. The corresponding probabilities for
mogorov’s forward equation)” (Pothos & Busemeyer, 2013). It transforms
asking the questions in the reverse order were, by contrast, 57% and 60% (Moore 2002).
the probabilities without violating the law of total probability. In QP am-
Such order effects are puzzling according to CP theory, because, as noted, the probability
plitudes change by a unitaryof transformation.
saying yes to question A and then yes to question B equals
G.3 Experimental evidence

The “Linda experiment.” In 1983 Tversky and Kahneman reported on
experiments in which subjects read a description of a hypothetical person
named Linda that suggested she was a feminist. Subjects were asked to
compare the probability of two statements: “Linda is a bank teller” (ex-
tremely unlikely given Linda’s description), and “Linda is a bank teller and
a feminist.” Most subjects concluded:
P{bank teller} < P{bank teller ∧ feminist},
which violates CP; it is an example of the conjunction fallacy. Many experi-

ments of this sort have shown that everyday reasoning commits this fallacy.
Tversky and Kahneman proposed that people use heuristics rather than for-
mal CP, but it can also be explained by QP.
The actual QP theory model developed for such failures in commutativity was
based on the abovementioned idea, but was more general, so as to provide a parameter
free test of the relevant empirical data (e.g., there are various specific types of order
effects; Wang & Busemeyer, under review).

s
n%Ye
C linto
G ore %Ye s
o
to n%N
C lin
G ore %No
Figure 3. An illustration of order effects in Gallup polls.
Figure III.47: Example of order effects in Gallop polls. [fig. from PB]
A related failure of commutativity concerns the order of assessing different pieces
of evidence for a particular hypothesis. According to CP theory, the order in which
evidence A and B is considered, in relation to a hypothesis H, is irrelevant, as
The QP explanation is as follows. We suppose that the written descrip-

tion makes it a priori likely that Linda
Prob(H|A∧B)= is a feminist and unlikely that she is
Prob (H|B∧A).
a bank teller; these priors are depicted in Fig. III.46. However, notice that
being a feminist is largelyHowever,
independent of being
there have been demonstrations that, ina
fact,bank teller. In making a
judgment like “Linda is a bank teller and a feminist” it is supposed that it is a

sequential conjunction, with the most likely judgment evaluated first, in this
case, “feminist && bank teller.” Look at the figure. The green projection
onto |feministi and then onto |bank telleri is longer than the blue projection
directly onto |bank telleri. Projection can be thought of as an abstraction
process, and so the projection of Linda onto |feministi throws away details
about her (it stereotypes her, we might say), and makes it more likely that
she is a bank teller (since there is not a strong correlation between feminists
and bank tellers). This may be compared to decoherence and loss of informa-
tion in a quantum system. “In general, QP theory does not always predict
an overestimation of conjunction. However, given the details of the Linda
problem, an overestimation of conjunction necessarily follows. Moreover, the
same model was able to account for several related empirical findings, such
as the disjunction fallacy, event dependencies, order effects, and unpacking
effects. . . ” (Pothos & Busemeyer, 2013)
G.3.a Failure of commutativity
The “Clinton-Gore experiment.” A Gallup poll asked “Is Clinton hon-

est?” and “Is Gore honest?” Results depended on the order in which they
were asked:
order Clinton Gore

Clinton — Gore 50% 68%
Gore — Clinton 57% 60%
This is also a common characteristic of everyday judgment; it is also common
in the assessment of evidence for a hypothesis. QP explains this as follows.
The “Yes” basis vectors have a smaller angle reflecting an expected correla-
tion between the answers (since Clinton and Gore ran together). The initial
state vector is a little closer to the |Gore Yesi vector reflecting the assump-
tion that Gore’s honesty is a priori more likely than Clinton’s. You can see
this by looking at the green projection onto |Gore Yesi, which is longer than
its blue projection onto |Clinton Yesi. Note further that the two-step blue
projection onto |Clinton Yesi is longer than the direct projection onto it.
That is, judging Gore to be honest increases the probability of also judging
Clinton to be honest.
G.3.b Violations of the sure-thing principle

“The sure thing principle is the expectation that human behavior ought
to conform to the law of total probability.” (Pothos & Busemeyer, 2013)
In 1992 Shafir and Tversky reported experiments showing violations of the
sure-thing principle in the one-shot prisoner’s dilemma: The subject has to
decide whether to cooperate or defect, as does their opponent. This is a
typical payoff matrix; it shows the payoff for you and your opponent for each
pair of choices:
opponent
↓ you cooperate defect
cooperate 3, 3 0, 5
defect 5, 0 1, 1
If you are told what your opponent is going to do, then you should defect.
This is what subjects usually do. If you don’t know, then the optimal strategy
is still to defect. This is the “sure thing”: you should defect in either case.
However, some subjects decide to cooperate anyway (thus violating the sure-
thing principle). One explanation is “wishful thinking.” If you have a bias
toward cooperation, you might suppose (in the absence of evidence) that
your opponent has a similar bias.
The QP explanation is as follows. Suppose |ψC i and |ψD i are the states of
knowing that your opponent will cooperate and defect, respectively. Suppose
PC and PD are projections representing your decision to cooperate or defect.

Under the condition where you know what your opponent is going to do, the
probability of you deciding to defect in the two cases is:
P{you defect} = kPD |ψC ik2 ,
P{you defect} = kPD |ψD ik2 .
In the unknown condition, we can suppose the state is |ψi = √12 (|ψC i+|ψD i).
Hence, in this case the probability of you deciding to defect is:
2
1
P{you defect} = √2 (PD |ψC i + PD |ψD i)

1
= (hψC | + hψD |)PD† PD (|ψC i + |ψD i)
2
1 1
= kPD |ψC ik2 + kPD |ψD ik2 + hψD | PD† PD | ψC i.
2 2
The interference term hψD | PD† PD | ψC i could be positive or negative, in the
latter case decreasing the probability below unity.
G.3.c Asymmetric similarity

In 1977 Tversky showed that similarity judgments violate metric axioms, in
particular, symmetry.
China-Korea experiment. For example, North Korea was judged more
similar to China, than China was judged to be similar to North Korea:
Sim(North Korea, China) > Sim(China, North Korea).
The QP explanation is that concepts correspond to subspaces of various
dimensions, with the dimension of the subspace roughly corresponding to
the number of known properties of the concept (i.e., how much someone
knows about it). The judgment of the similarity of A to B is modeled by the
projection of the initial state into A and then into B. It’s assumed that the
initial state is neutral with respect to A and B (i.e., the subject hasn’t been
thinking about either). If |ψi is the initial state, then
Sim(A, B) = kPB PA |ψik2 = P{A && B}.
The subjects in this case are assumed to be more familiar with China than
with North Korea, so the China subspace is larger (see Fig. III.48). When
Pothos and Busemeyer (2011), whose results indicate that, as long as one subspace has a
greater dimensionality than another, on average the transition from the lower
dimensionality subspace to the higher dimensionality one would retain more amplitude
than the converse transition (it has not been proved that this is always the case, but note
that participant results with such tasks are not uniform).
(a) Korea to China
a
re
Ko
C hina
(b) China to Korea
a
re
Ko C hina
Figure 4. Figure 4a corresponds to the similarity of Korea to China and 4b to the

similarity of China to Korea. Projecting to a higher dimensionality subspace last (as in
Figure III.48: QP model of China – (North) Korea experiment. [fig. from
PB]
North Korea is compared to China, more of its amplitude is retained by the

final projection into the higher dimensional subspace corresponding to China:
Fig. III.48(a). In the opposite case, the projection into the lower dimensional
North Korea subspace loses more amplitude: Fig. III.48(b). This is not
universally true.
G.4 Cognition in Hilbert space
Pothos & Busemeyer (2013) defend the application of QP in a function-first

or top-down approach to modeling cognition.31 This is done by postulating
vectors in a low-dimensional space. I argue that consideration of the high-
dimensional complex-valued wavefunction underlying the state vector will
expand the value of QP in cognitive science.
31
Material in this section is adapted from MacLennan (2013), my commentary on Pothos
& Busemeyer (2013).
G.4.a QM premises
To this end, application of QP in cognitive science would be aided by im-
porting two premises from quantum mechanics:
The first premise is that the fundamental reality is the wavefunction. In
cognitive science this corresponds to postulating a spatially-distributed pat-
tern of neural activity as the elements of the cognitive state space. Therefore
the basis vectors used in QP are in fact basis functions for an infinite (or
very high) dimensional Hilbert space.
The second important fact is that the wavefunction is complex-valued
and that wavefunctions combine with complex coefficients. This is the main
reason for interference and other non-classical properties. The authors ac-
knowledge this, but do not make explicit use of complex numbers in the
target article.
G.4.b Possible neural substrates

What is the analog of the complex-valued wavefunction in neurophysiology?
There are several possibilities, but perhaps the most obvious is the distri-
bution of neural activity across a region of cortex; even a square millimeter
of which can have hundreds of thousands of neurons. The dynamics will be
defined by a time-varying Hamiltonian, with each eigenstate being a spatial
distribution of neurons firing at a particular rate. The most direct represen-
tation of the magnitude and phase (or argument) of a complex quantity is
frequency and phase of neural impulses.
G.4.c Projection
Possible neural mechanisms: Pothos & Busemeyer (2013) specify that a
judgment or decision corresponds to measurement of a quantum state, which
projects it into a corresponding subspace, but it is informative to consider
possible mechanisms. For example, the need to act definitely (such as coming
to a conclusion in order to answer a question) can lead to mutually compet-
itive mechanisms, such as among the minicolumns in a macrocolumn, which
creates dynamical attractors corresponding to measurement eigenspaces. Ap-
proach to the attractor amplifies certain patterns of activity at the expense
of others. Orthogonal projectors filter the neural activity and win the com-
petition with a probability proportional to the squared amplitude of the
patterns to which they are matched. (In the case where the phases of neu-
ral impulses encode complex phases, matching occurs when the phases are
delayed in such a way that the impulses reinforce.) The winner positively
reinforces its matched signal and the loser negatively reinforces the signal to
which it is matched. Regardless of mechanism, during collapse the energy of
the observed eigenstate of the decision (measurement) operator receives the
energy of the orthogonal eigenstates (this is the effect of renormalization).
The projection switches a jumble of frequencies and phases into a smaller,
more coherent collection, corresponding to the answer (observed) eigenspace.
No inherent bases: The target article suggests that a QP model of a
process begins by postulating basis vectors and qualitative angles between
alternative decision bases (significantly, only real rotations are discussed).
As a consequence, a QP model is treated as a low-dimensional vector space.
This is a reasonable, top-down strategy for defining a QP cognitive model,
but it can be misleading. There is no reason to suppose that particular
decision bases are inherent to a cognitive Hilbert space. There may be a
small number of ”hard-wired” decisions, such as fight-or-flight, but the vast
majority are learned. Certainly this is the case for decisions corresponding
to lexical items such as (un-)happy and (un-)employed.
Creation/modification of observables: Investigation of the dynamics
of cognitive wavefunction collapse would illuminate the mechanisms of deci-
sion making but also of the processes by which observables form. This would
allow modeling changes in the decision bases, either temporary through con-
text effects or longer lasting through learning. Many decision bases are ad
hoc, as when we ask, “Do you admire Telemachus in the Odyssey?” How such
ad hoc projectors are organized requires looking beneath a priori state basis
vectors to the underlying neural wavefunctions and the processes shaping
them.
G.4.d Incompatible decisions

The commutator and anti-commutator: In quantum mechanics the
uncertainty principle is a consequence of non-commuting measurement oper-
ators, and the degree of non-commutativity can be quantified (see Sec. B.7,
p. 93). Two measurement operators P and Q commute if P Q = QP , that
is, if the operator P Q − QP is identically 0. If they do not commute, then
P Q − QP measures the degree of non-commutativity. This is expressed in
quantum mechanics by the “commutator” [P, Q] = P Q − QP . It is relatively
easy to show that this implies an uncertainty relation: ∆P ∆Q ≥ |h[P, Q]i|.

That is, the product of the uncertainties on a state is bounded below by
the absolute mean value of the commutator on the state. Suppose H is
a measurement that returns 1 for |happyi and 0 for |unhappyi, and E is
a measurement that returns 1 for |employedi and 0 for |unemployedi. If
|employedi = a|happyi + b|unhappyi, then the commutator is

0 1
[H, E] = ab .
−1 0
The absolute mean value of this commutator (applied to a state) gives a

minimum joint uncertainty. If we could measure [P, Q] for various pairs of
questions, P and Q, we could make quantitative empirical predictions of the
joint uncertainty in decisions.
Might we design experiments to measure the commutators and so quantify
incompatibility among decisions? Certainly there are difficulties, such as
making independent measurements of both P Q and QP for a single subject,
or accounting for intersubject variability in decision operators. But making
such measurements would put more quantitative teeth into QP as a cognitive
model.
G.4.e Suggestions
Pothos & Busemeyer (2013) do an admirable job of defending QP as a fruitful
top-down model of decision making, but I believe it would be more valuable
if it paid greater attention to the complex-valued wavefunction that under-
lies QP in both quantum mechanics and cognition. This would allow a more
detailed account of the origin of interference effects and of the structure of
both learned and ad hoc decision operators. Finally, the treatment of incom-
patible decisions can be made more rigorous by treating them quantitatively
as noncommuting operators.
G.5 Conclusions
You might wonder why it is so important to understand the less-then-perfect
inferential abilities of humans. There are at least two reasons, scientific and
technological. First, it is important to understand human inference as both
pure and applied science. It reveals much about our human nature, and
specifically provides hints as to how the brain works. From a more applied
perspective, we need to understand how humans determine their actions in

order to predict (and even influence) human behavior. In terms of technology,
it might seem that the last thing we might want to do would be to emulate
in our machine intelligence the “imperfect, fallacious” reasoning of humans.
It might be the case, however, that QP-based reasoning is better than CP
for real-time purposeful action in natural, complex situations, where the
premisses of CP are inaccurate.
H. EXERCISES 193
H Exercises
Exercise III.1 Compute the probability of measuring |0i and |1i for each
of the following quantum states:
1. 0.6|0i + 0.8|1i.
p
2. √13 |0i + 2/3|1i.
√
3
3. 2
|0i − 12 |1i.
1
4. − 25 (24|0i − 7|1i).
iπ/6
5. − √12 |0i + e√
2
|1i.
Exercise III.2 Compute the probability of the four states if the following
are measured in the computational basis:
√ √ √
1. (ei |00i + 2|01i + 3|10i + 2e2i |11i)/ 10.
1
2. 2
(−|0i + |1i) ⊗ (eπi |0i + e−πi |1i).
3.

p p √ eπi/4 eπi/2
( 1/3|0i − 2/3|1i) ⊗ 2 |0i + |1i .
2 2
Exercise III.3 Suppose that a two-qubit register is in the state

√
3 7 eiπ/2 2
|ψi = |00i − |01i + √ |10i − |11i.
5 5 5 5
1. Suppose we measure just the first qubit. Compute the probability of
measuring a |0i or a |1i and the resulting register state in each case.
2. Do the same, but supposing instead that we measure just the second
qubit.
Exercise III.4 Prove that projectors are idempotent, that is, P 2 = P .
Exercise III.5 Prove that a normal matrix is Hermitian if and only if it has
real eigenvalues.
def
Exercise III.6 Prove that U (t) = exp(−iHt/~) is unitary.
Exercise III.7 Use spectral decomposition to show that K = −i log(U ) is

Hermitian for any unitary U , and thus U = exp(iK) for some Hermitian K.
Exercise III.8 Show that the commutators ([L, M ] and {L, M }) are bilin-
ear (linear in both of their arguments).
Exercise III.9 Show that [L, M ] is anticommutative, i.e., [M, L] = −[L, M ],

and that {L, M } is commutative.
[L,M ]+{L,M }
Exercise III.10 Show that LM = 2
.
Exercise III.11 Show that the four Bell states are orthonormal (i.e., both
orthogonal and normalized).
Exercise III.12 Prove that |β11 i is entangled.
Exercise III.13 Prove that √1 (|000i + |111i) is entangled.

2
Exercise III.14 What is the effect of Y (imaginary definition) on the com-

putational basis vectors? What is its effect if you use the real definition
(C.2.a, p. 105)?
Exercise III.15 Prove that I, X, Y , and Z are unitary. Use either the imag-
inary or real definition of Y (C.2.a, p. 105).
Exercise III.16 What is the matrix for H in the sign basis?
Exercise III.17 Show that the X, Y, Z and H gates are Hermitian (their
own inverses) and prove your answers. Use either the imaginary or real
definition of Y (C.2.a, p. 105).
Exercise III.18 Prove the following useful identities:
HXH = Z, HY H = −Y, HZH = X.
Exercise III.19 Show (using the real definition of Y , C.2.a, p. 105):

|0ih0| = 21 (I + Z), |0ih1| = 12 (X − Y ), |1ih0| = 12 (X + Y ), |1ih1| = 21 (I − Z).
H. EXERCISES 195
Exercise III.20 Prove that the Pauli matrices span the space of 2 × 2 ma-
trices.
Exercise III.21 Prove |βxy i = (P ⊗ I)|β00 i, where xy = 00, 01, 11, 10 for
P = I, X, Y, Z, respectively.
Exercise III.22 Suppose that P is one of the Pauli operators, but you don’t
know which one. However, you are able to pick a 2-qubit state |ψ0 i and
operate on it, |ψ1 i = (P ⊗ I)|ψ0 i. Further, you are able to select a unitary
operation U to apply to |ψ1 i, and to measure the 2-qubit result, |ψ2 i = U |ψ1 i,
in the computational basis. Select |ψ0 i and U so that you can determine with
certainty the unknown Pauli operator P .
Exercise III.23 What is the matrix for CNOT in the standard basis? Prove
your answer.
Exercise III.24 Show that CNOT does not violate the No-cloning Theorem
6 |ψi|ψi. Under what conditions
by showing that, in general, CNOT|ψi|0i =
does the equality hold?
Exercise III.25 What quantum state results from

1
CNOT(H ⊗ I) (c00 |00i + c01 |01i + c10 |10i + c11 |11i)?
2
Express the result in the computational basis.
Exercise III.26 Compute (Y ⊗ I)CNOT(H ⊗ I)|00i. Show your work.
Exercise III.27
1. Compute (H ⊗ I ⊗ I)(CNOT ⊗ I)[( 54 |0i + 35 |1i) ⊗ |β00 i.
2. Give the probabilities and resulting states for measuring the first two
qubits in the computational basis.
3. Apply Z to the state resulting from measuring |10i.
Exercise III.28 What is the matrix for CCNOT in the standard basis?
Prove your answer.
Exercise III.29 Use a single Toffoli gate to implement each of NOT, NAND,
and XOR.
Exercise III.30 Use Toffoli gates to implement FAN-OUT. FAN-OUT would

seem to violate the No-cloning Theorem, but it doesn’t. Explain why.
Exercise III.31 Design a quantum circuit to transform |000i into the en-
tangled state √12 (|000i + |111i).
Exercise III.32 Show that |+i, |−i is an ON basis.
Exercise III.33 Prove:

1
|0i = √ (|+i + |−i),
2
1
|1i = √ (|+i − |−i).
2
Exercise III.34 What are the possible outcomes (probabilities and result-
ing states) of measuring a|+i + b|−i in the computational basis (of course,
|a|2 + |b|2 = 1)?
Exercise III.35 Prove that Z|+i = |−i and Z|−i = |+i.
Exercise III.36 Prove:
H(a|0i + b|1i) = a|+i + b|−i,

H(a|+i + b|−i) = a|0i + b|1i.
√
Exercise III.37 Prove H = (X + Z)/ 2.
Exercise III.38 Prove Eq. III.18 (p. 111).
Exercise III.39 Show that three successive CNOTs, connected as in Fig.

III.11 (p. 110), will swap two qubits.
Exercise III.40 Recall the conditional selection between two operators (C.3,
p. 111): |0ih0| ⊗ U0 + |1ih1| ⊗ U1 . Suppose the control bit is a superposition
|χi = a|0i + b|1i. Show that:
(|0ih0| ⊗ U0 + |1ih1| ⊗ U1 )|χ, ψi = a|0, U0 ψi + b|1, U1 ψi.

H. EXERCISES 197
Exercise III.41 Show that the 1-bit full adder (Fig. III.15, p. 113) is cor-
rect.
Exercise III.42 Show that the operator Uf is unitary:

def
Uf |x, yi = |x, y ⊕ f (x)i,
Exercise III.43 Verify the remaining superdense encoding transformations

in Sec. C.6.a (p. 117).
Exercise III.44 Verify the remaining decoding cases for quantum telepor-
tation Sec. C.6.b (p. 121).
Exercise III.45 Confirm the quantum teleportation circuit in Fig. III.21

(p. 122).
Exercise III.46 Complete the following step from the derivation of the
Deutsch-Jozsa algorithm (Sec. D.1, p. 129):
X 1
H|xi = √ (−1)xz |zi.
z∈2
2
Exercise III.47 Show that CNOT(H ⊗ I) = (I ⊗ H)CZ H ⊗2 , where CZ is

the controlled-Z gate.
Exercise III.48 Show that the Fourier transform matrix (Eq. III.25, p. 137,
Sec. D.3.a) is unitary.
Exercise III.49 Prove the claim on page 153 (Sec. D.4.b) that D is unitary.
Exercise III.50 Prove the claim on page 153 (Sec. D.4.b) that
 2 2 2

N N
··· N
 2 2 ··· 2 
 
W R0 W =  N.. N.. . . N
..  .
 . . . . 
2 2 2
N N
··· N
Exercise
√ III.51 Show that if there are s solutions x such that P (x) = 1,
π N/s
then 4
is the optimal number of iterations in Grover’s algorithm.
Exercise III.52 Design a quantum gate array for the following syndrome
extraction operator (Sec. D.5.d, p. 163):
def
S|x3 , x2 , x1 , 0, 0, 0i = |x3 , x2 , x1 , x1 ⊕ x2 , x1 ⊕ x3 , x2 ⊕ x3 i.
Exercise III.53 Design a quantum gate array to apply the appropriate error
correction for the extracted syndrome as given in Sec. D.5.d, p. 163:
bit flipped syndrome error correction

none |000i I ⊗I ⊗I
1 |110i I ⊗I ⊗X
2 |101i I ⊗X ⊗I
3 |011i X ⊗I ⊗I
Exercise III.54 Design encoding, syndrome extraction, and error correc-

tion quantum circuits for the code |0i 7→ | + ++i, |1i 7→ | − −−i to correct
single phase flip (Z) errors.
Exercise III.55 Prove that Aa Aa = 1 (Sec. F.1.b).
Exercise III.56 Prove that Aab,c = 1+a† ab† b(c+c† −1) = 1+Na Nb (Ac −1)
is a correct definition of CCNOT by showing how it transforms the quantum
register |a, b, ci (Sec. F.1.b).
Exercise III.57 Show that the following definition of Feynman’s switch is

unitary (Sec. F.1.b):
q † cp + r† c† p + p† c† q + p† cr.
Chapter IV
Molecular Computation
A Basic concepts
A.1 Introduction to molecular computation
Molecular computation uses molecules to represent information and molec-
ular processes to implement information processing. DNA computation is
the most important (and most explored) variety of molecular computation,
and it is the principal topic of this chapter. The principal motivation is that
molecular computation is massively parallel (molar or Avogadro-scale par-
allelism, that is, on the order of 1026 ). Data representations are also very
dense (on the order of nanometers), and molecular computation can be very
efficient in terms of energy.
A.2 DNA basics

The DNA molecule is a polymer (poly = many, mer = part), a sequence
of nucleotides. A DNA nucleotide has three parts: a deoxyribose sugar, a
phosphate group, and a nucleobase (C = cytosine, G = guanine, A = ade-
nine, T= thymine).1 See Fig. IV.1 for deoxyribose. The DNA backbone is a
1
In RNA, uracil replaces thymine.
199
200 CHAPTER IV. MOLECULAR COMPUTATION
Figure IV.1: Deoxyribose. [source: wikipedia]
Figure IV.2: DNA backbone. [source: wikipedia]

A. BASIC CONCEPTS 201
sequence of deoxyribose sugars alternating with phosphate groups connected

by covalent phosphodiester bonds (Fig. IV.2). The backbone connects the
hydroxyl group on the 30 carbon (the “30 -hydroxyl group”) of one sugar to
the 50 -hydroxyl of the next. In this way we distinguish the 50 and 30 ends of
a polynucleotide: 50 has a terminal phosphate group, 30 a terminal hydroxl
group. We generally write the bases in the 50 to 30 order.
Adenine and thymine each have two hydrogen-bonding sites and can bind
together; similarly, guanine and cytosine each have three hydrogen-bonds and
can bind together. This results in Watson-Crick complementarity, which
means that two complementary polynucleotides can bind together. This can
occur only if the two single strands are antiparallel, that is, run in opposite
directions (30 to 50 versus 50 to 30 ); see Fig. IV.3. Since we write the sense-
strand in the 50 to 30 order, we write the antisense-strand in the 30 to 50
order. Complementary single stranded DNA can hybridize, leaving sticky
ends (unmatched single strands) at either or both ends. They are called
“sticky” because they are available for bonding by complementary sequences.
Fig. IV.4 gives some idea what the actual double helix looks like. It is
worth keeping in mind that it is a very complex organic molecule, not a
string of the letters A, C, T, G. For example, in addition to the double helix
formed by the backbones of the two strands, there are two groves between the
strands (Fig. IV.5). They are adjacent to the bases and therefore provide
binding sites for proteins. Because the backbones are not symmetrically
placed, the grooves have two sizes: minor (1.2nm) and major (2.2nm). The
latter provides the best binding site.
To give some sense of the scale, here are some numbers. The nucleotides
are about 3.3nm long. The chain is 2.2 to 2.6nm wide. The radius of the
coil is about 1nm. The longest human chromosome (chromosome 1) is about
220 × 106 bp (base pairs) long.
DNA data storage: In Aug. 2012 a team at Harvard reported convert-
ing a 53,000 word book to DNA and then reading it out by DNA sequenc-
ing.2 This is the densest consolidation of data in any medium (including
flash memory, but also compared to experimental media, such as quantum
2
Church, George M, Gao, Yuan, and Kosuri, Sriram. “Next-generation digital
information storage in DNA.” Science 337 (28 Sept. 2012): 1628. See also Sci-
ence (Aug. 16, 2012), DOI: 10.1126/science.1226355 (accessed 2012-08-24). See also
http://spectrum.ieee.org/biomedical/imaging/reading-and-writing-a-book-with-dna/
(accessed 2012-08-24).
Figure IV.3: Watcon-Crick complementarity. Complementary H-bonds al-

low guanine to bind to cytosine and adenine to bind to thymine. [source:
wikimedia commons]
Figure IV.4: [source: wikimedia commons]
Figure IV.5: Major and minor grooves. [source: wikimedia commons]

holography). The book included 11 jpegs and a javascript program, for a

total of 5.27 Mbits, and the decoded version had only 10 incorrect bits. The
encoded book could be reproduced by DNA replication. This is a storage
density of 5.5 petabits/mm3 = 5.5 × 1015 bits/mm3 . The authors estimate
that DNA can encode 455 exabytes (4.55 × 1020 bytes) per gram of single-
stranded DNA. (This is based on 2 bits per nucleotide, and doesn’t take
redundancy, error correction, etc., into account, which would decrease the
density by several orders of magnitude.3 ). In Jan. 2013 Nature, it was re-
ported that 5 Mbits had been encoded and perfectly decoded (99.99 to 100%
accuracy with quadruple redundancy. It cost $12,400 for the encoding and
$220 for retrieval. The cost of DNA synthesis has been decreasing about 5×
per year, and the cost of sequencing by about 12× per year, so it could be-
come economical for archival storage. Enclosed in silica glass spheres, DNA
storage could last for a million years.
A.3 DNA manipulation

There are several standard procedures for manipulating DNA that have been
applied to DNA computing, which has benefitted from well-developed labo-
ratory procedures and apparatus for handling DNA. The following sections
review these procedures briefly, omitting details not directly relevant to DNA
computation.
A.3.a Denaturation
Denaturation causes DNA double strands to unwind and separate into its
two single strands. This occurs by breaking the H-bonds between the bases,
which are weak compared to covalent bonds (e.g., those in the backbone).
Thermal denaturing (heating to about 95◦ C) is the most common procedure,
but there are also chemical denaturing agents.
A.3.b Hybridization
Hybridization is the process by which two DNA single strands combine in
a sequence specific way through the formation of non-covalent H-bonds be-
tween the bases. Annealing is used to achieve an energetically favorable result
(a minimal number mismatches).
3
Church, op. cit., Supplementary Material
A.3.c Polymerase extension

Enzymes called polymerases (polymerizing enzymes) add nucleotides one at
a time to the 30 end of a DNA molecule, an operation called polymerase
extension. It can be used to fill in a sticky end, or a short sequence bound
to a complementary sequence can operate as a primer causing the rest of the
sequence to be filled in.
A.3.d Ligation
Hybridization links to strands together by H-bonds, but there is a nick in the
DNA backbone resulting from a missing phosphodiester bond. Certain en-
zymes will ligate the strands by filling in the missing covalent bonds, making
a more stable strand.
A.3.e Nuclease degradation

Nucleases are enzymes that remove nucleotides from a DNA strand. Exonu-
cleases remove nucleotides from the ends (either 30 or 50 depending on the
exonuclease), whereas endonucleases cut DNA strands (single or double) at
specific places. Restriction enzymes are the most common endonucleases.
They operate at specific restriction sites, typically 4 to 8 bases long. The cut
can be blunt or staggered leaving sticky ends.
A.3.f Synthesis
Synthesis is the process of assembling single-stranded oligonucleotides (“oli-
gos”), which are short strands with a desired sequence. Assembly occurs in
the 30 to 50 direction.The process is completely automated nowadays, but due
to the accumulation of errors, the current limit is about 200 nucleotides.
A.3.g Polymerase chain reaction

The Polymerase Chain Reaction (PCR) is a method for exponentially am-
plifying the concentration of a selected DNA sequence in a sample. It was a
breakthrough and remains one of the most important processes in practical
DNA technology. Here is the basic procedure.
The sequence to be replicated is identified by a known short sequence
(about 20 bases) at the 30 ends of both its sense and antisense strands. Short
Figure IV.6: Example of result of gel electrophoresis.
DNA strands called primers, which are complementary to the identifying

sequences, are synthesized. That is, two primers are used to bound or delimit
the sequence of interest (at the 30 end of the sense strand and the 30 end of
the antisense strand). (You can also have primers that bind to the 50 ends
of the target sequence, but 30 is typical.)
Step 1 (add primers): The primers and DNA polymerase are added to
the DNA sample (possibly containing the desired sequence and many others).
Step 2 (denaturation): The mixture is heated to about 95◦ C for about
30 seconds, which causes all the double strands to separate.
Step 3 (annealing): The mixture is cooled to about 68◦ C for a minute,
which allows the primers to hybridize with the other strands.
Step 4 (polymerase extension): The mixture is warmed to about
◦
72 C for 30 seconds or more (depending on the length of the sequence to
be replicated, which goes at about about 1000 bp/min.). This allows the
polymerase to extend the primer on each single strand so that it becomes a
double strand. Pairs of long strands can rehybridize, but the lighter oligonu-
cleotides tend to get there faster, so instead we get two double strands. As
a result, the concentration of the desired DNA sequence has doubled.
Step 5 (repeat): Repeat Steps 2 to 4 until the desired concentration is
reached. The concentration of the desired sequence increases exponentially,
2n , in the number of cycles. There are “PCR machines” that automate the
entire PCR process.
A.3.h Gel electrophoresis

Gel electrophoresis is a method of determining the size of DNA fragments in a
sample. The sample is put in a well at one end of a layer of gel, and a positive
electric field is applied at the other end of the gel. Since DNA is negatively
charged, it migrates through the gel, but smaller fragments migrate faster.
Later the DNA can be stained and the gel imaged to determine the relative
concentration of fragments of different sizes in the sample. The resulting
pattern is called a DNA ladder; see Fig. IV.6 for several “lanes” of DNA
ladders.
A.3.i Filtering
There are several ways to filter DNA samples for specific sequences. One way
to select for an oligonucleotide o in a sample S is to attach the complementary
sequence o to a filter. Then when S goes through the filter, the o in it will
hybridize with the o in the filter, and stay there; what passes through is
S − o. If desired, the oo on the filter can be denatured to separate the
strands containing o from the o on the filter.
A.3.j Modification
Certain enzymes can be used to change subsequences of a DNA sequence.
A.3.k Sequencing
Sequencing is the process of reading out the sequence of a DNA molecule.
This is traditionally accomplished by extension of a primed template se-
quence, but a variety of other techniques have been developed since the 1970s.
The latest techniques do massively parallel sequencing of many fragments.
B Filtering models
Filtering models, an important class of DNA algorithms, operate by filtering
out of the solution molecules that are not part of the desired result. A
chemical solution can be treated mathematically as a finite bag or multi-set
of molecules, and filtering operations can be treated as operations to produce
multi-sets from multi-sets. Typically, for a problem of size n, strings of size
O(n) are required. The chemical solution should contain enough strings
to include many copies all possible answers. Therefore, for an exponential
problem, we will have O(k n ) strings. Filtering is essentially a brute-force
method of solving problems.
B.1 Adleman: HPP

B.1.a Review of HPP
Leonard Adleman’s solution of the Hamiltonian Path Problem was the first
successful demonstration of DNA computing. The Hamiltonian Path Prob-
lem (HPP) is to determine, for a given directed graph G = (V, E) and two
of its vertices vin , vout ∈ V , whether there is a Hamiltonian path from vin
to vout , that is, a path that goes through each vertex exactly once. HPP is
an NP-complete problem, but we will see that for Adleman’s algorithm the
number of algorithm steps is linear in problem size.
Adleman (the “A,” by the way, of “RSA.”) gave a laboratory demonstra-
tion of the procedure in 1994 for n = 7, which is a very small instance of
HPP. (We will use this instance, shown in Fig. IV.7, as an example.) Later
his group applied similar techniques to solving a 20-variable 3-SAT problem,
which has more than a million potential solutions (see p. 216).4
B.1.b Problem Representation

The heart of Adleman’s algorithm is a clever way to encode candidate paths in
DNA. Vertices are represented by single-stranded 20mers, that is, sequences
of 20nt (nucleotides). They were generated at random and assigned to the
vertices, but with the restriction that none of them were too similar or com-
plementary. Each vertex code can be considered a catenation of two 10mers:
vi = ai bi (i.e., ai is the 50 10mer and bi is the 30 10mer). Edges are also
4
https://en.wikipedia.org/wiki/Adleman, Leonard (accessed 2012-11-04).
112 5 Physical Implementations
5.4 Adleman’s Implementation

Adleman utilized the incredible storage capacity of DNA to implement a
brute-force algorithm for the directed Hamiltonian Path Problem (HPP). Re-
call that the HPP involves finding a path through a graph that visits each
vertex exactly once. The instance of the HPP that Adleman solved is depicted
in Fig. 5.2, with the unique Hamiltonian Path (HP) highlighted by a dashed
B. FILTERING
line.
MODELS 209
1 2
7 3
6 4
Fig. 5.2. Instance of the HPP solved by Adleman

Figure IV.7: HPP solved by Adleman. The HP is indicated by the dotted
edges. [source: Amos, Fig. 5.2]
Adleman’s approach was simple:
1. Generate strands encoding random paths such that the Hamiltonian Path
represented
(HP)by
is 20mers. The
represented edge
with highfrom vertex The
probability. i to quantities
vertex j ofis represented
DNA used by
far exceeded those necessary for the small graph under consideration, so
it is likelyethat
i→j = bi ajstrands
many , whereencoding
vi = aithe
bi , and vj =present.
HP were aj b j .
2. Remove all strands that do not encode the HP.
Paths 3.are represented
Check by usingstrands
that the remaining complements of the vertex
encode a solution to the HPP.20mers to stitch
together the edge 20mers. (Of course, using the complements of the edges to
The individual steps were implemented as follows:
stitch together the vertices works as well.) For example, a path 2 → 3 → 4
Stage 1: Each vertex and edge was assigned a distinct 20-mer sequence of
is represented:
DNA (Fig. 5.3a). This implies that strands encoding a HP were of length 140
b.p. Sequences representing edges act as ‘splints’ between strands representing
e e
z 2→3
}|their z 3→4
{ endpoints
}| { (Fig. 5.3b).
b3 a4terms, the sequence associated with an edge i → j is the 3’ 10-
b2 a3 In formal
mer
a3 of bthe sequence representing vi followed by the 5’ 10-mer of the sequence
| {z }3 vj . These oligonucleotides were then combined to form strands
representing
v3
encoding random paths through the graph. An (illegal) example path (v1 →
→ v →
The edges fromv4v)inis and
v depicted
to vin Fig. 5.4.
out have special representations as 30mers:
2 3
Fixed amounts (50 pmol) of each oligonucleotide were mixed together in
a single ligation reaction.
e At
= thev end
a , of this reaction,
where v = a bit, is assumed that a
in→j in j j j j
ei→out = bi vout , where vi = ai bi .
Note that the special representation of the initial and terminal edges results
in blunt ends for complete paths.
Therefore, for the n = 7 problems, candidate solutions were 140bp in
length: There are n − 1 edges, but the first and last edges are 30mers. Hence
2 × 30 + (n − 3) × 20 = 140. Ligation is used to remove the nicks in the
backbone.
B.1.c Adleman’s Algorithm
algorithm Adlemen:
Step 1 (generation of all paths): Generate multiple representations of

all possible paths through the graph. This is done by combining the oligos
for the edges with the oligos for the complements of the vertices in a single
ligation reaction.
Step 2: Amplify the concentration of paths beginning with vin and ending
with vout . This is done by PCR using vin and vout as primers. Remember
that denaturation separates the sense and antisense strands. PCR extends
the sense strand in the 30 direction from vin , and extends the antisense strand
in the 30 direction from vout . At the end of this step we have paths of all
sorts from vin to vout .
Step 3: Only paths with the correct length are retained; for n = 7 this
is 140bp. This operation is accomplished by gel electrophoresis. The band
corresponding to 140bp is determined by comparison with a marker lane,
and the DNA is extracted from this band and amplified by PCR. Gel elec-
trophoresis and PCR are repeated to get a sufficient quantity of the DNA.
We now have paths from vin to vout , but they might not be Hamiltonian.
Step 4 (affinity purification): Select for paths that contain all the ver-
tices (and are thus necessarily Hamiltonian). This is done by first selecting
all those paths that contain v1 , and then, of those, all that contain v2 , and
so forth. To select for vi , first heat the solution to separate the strands.
Then add the vi bound to a magnetic bead. Rehybridize (so the beads are
bound to strands containing vi ), and use a magnet to extract all the paths
containing vi . Repeat this process for each vertex.
Step 5 If there any paths left, they are Hamiltonian. Therefore amplify them
by PCR and inspect the result by gel electrophoresis to see if there are any
B. FILTERING MODELS 211
Figure IV.8: Electrophoresis showing solution to HPP problem.
strands of the correct length. If there are, then there is a Hamiltonian path;
if there aren’t, then there is not. If desired, the precise HP can be determined
by a graduated PCR procedure: Run n − 1 parallel PCR reactions. In the
ith lane, vin is the left primer and vi is the right primer. This will produce
bands with lengths 40, 60, 80, 100, 120, and 140 bp. The lane that has a
band at 40 corresponds to the first vertex after vin in the path, the lane with
a band at 60 corresponds to the next vertex, etc. This final readout process
depends on there being only one Hamiltonian path, and it is error-prone due
to its dependence on PCR.
B.1.d Discussion
Adleman’s algorithm is linear in the number of nodes, since the only iteration
is Step 4, which is repeated for each vertex. Step 5 is also linear if the path
is read out. Thanks to the massive parallelism of molecular computation,
it solves this NP-complete problem in linear time. Adleman’s experiment
took about a week, but with a more automated approach it could be done
in a few hours. On the other hand, the PCR process cannot be significantly
shortened.
In addition to time, we need to consider the molecular resources required.
The number of different oligos required is proportional to n, but the number
of strands is much larger, since there must be multiples instances of each
possible path. If d is the average degree of the graph, then there are about
dn possible paths (exponential in n). For example, if d = 10 and n = 80, then
the required 1080 DNA molecules is more than the estimated number of atoms
in the universe. Hartmanis calculated that for n = 200 the weight of the DNA
would exceed the weight of the earth. So this brute-force approach is still
defeated by exponential explosion. Lipton (1995) estimates that Adleman’s
algorithm is feasible for n ≤ 70, based on an upper limit of 1021 ≈ 270 DNA
strands (Boneh, Dunworth, Lipton & Sgall, 1996), but this is also feasible on
conventional computers.
Nevertheless, Adleman’s algorithm illustrates the massive parallelism of
molecular computation. Step 1 (generation of all possible paths) took about
an hour for n = 7. Adleman estimates that about 1014 ligation operations
were performed, and that it could be scaled up to 1020 operations. There-
fore, speeds of about 1015 to 1016 ops/sec (1–10 peta-operations/s) should be
achievable, which is, digital supercomputer range. Adlemen also estimates
that 2×1019 ligation operations were performed per joule of energy. Contem-
porary supercomputers perform only 109 operations per joule, so molecular
computation is 1010 more energy-efficient. It is near the thermodynamic limit
of 34×1019 operations per joule. Recall (Ch. II, Sec. B.1) kT ln 2 ≈ 3×10−9 pJ
= 3 × 10−21 J, so there can be about 3.3 × 1020 bit changes/J.5
A more pervasive problem is the inherent error in the filtering processes
(due to incorrect hybridization). Some strands we don’t want, get through;
and some that we do want, don’t. With many filtering stages the errors
accumulate to the extent that the algorithms fail. There are some approaches
to error-resistant DNA computing, but this is an open problem.
5
DNA is of course space efficient. One bit of information occupies about 1 cubic nm,
whereas contemporary disks store a bit in about 1010 cubic nm. That is, DNA is a 1010
improvement in density.
Figure IV.9: Graph G2 for Lipton’s algorithm (with two variables, x and y).
[source: Lipton (1995)]
B.2 Lipton: SAT

In this section we will discuss DNA solution of another classic NP-complete
problem, Boolean satisfiability, in fact the first problem proved to be NP-
complete.6
B.2.a Review of SAT problem

In the Boolean satisfiability problem (called “SAT”), we are given a Boolean
expression of n variables. The problem is to determine if the expression is
satisfiable, that is, if there is an assignment of Boolean values to the variables
that makes the expression true.
Without loss of generality, we can restrict our attention to expressions in
conjunctive normal form, for every Boolean expression can be put into this
form. That is, the expression is a conjunction of clauses, each of which is a
disjunction of either positive or negated variables, such as this:
(x1 ∨ x02 ∨ x03 ) ∧ (x3 ∨ x05 ∨ x6 ) ∧ (x3 ∨ x06 ∨ x4 ) ∧ (x4 ∨ x5 ∨ x6 ),
For convenience we use primes for negation, for example, x02 = ¬x2 . In
the above example, we have n = 6 variables m = 4 clauses. The (possibly
negated) variables are called literals.
B.2.b Data representation

To apply DNA computation, we have to find a way to represent potential
solutions to the problem as DNA strands. Potential solutions to SAT are
6
This section is based on Richard J. Lipton (1995), “DNA solution of hard computa-
tional problems,” Science 268: 542–5.
n-bit binary strings, which can be thought of as paths through a particular

graph Gn (see Fig. IV.9). For vertices ak , xk , x0k , k = 1, . . . , n, and an+1 ,
there are edges from ak to xk and x0k , and from xk and x0k to ak+1 . Binary
strings are represented by paths from a1 to an+1 . A path that goes through
xk encodes the assignment xk = 1 and a path through x0k encodes xk = 0.
The DNA encoding of these paths is essentially the same as in Adleman’s
algorithm.
B.2.c Lipton’s Algorithm
algorithm Lipton:
Input: Suppose we have an instance (formula) to be solved: I = C1 ∧ C2 ∧

· · · ∧ Cm . The algorithm will use a series of “test tubes” (reaction vessels)
i i
T0 , T1 , . . . , Tm and T1i , T 1 , . . . , Tmi , T m , for i = 0, . . . , n.
Step 1 (initialization): Create in a test tube T0 a library of all possible

n-bit binary strings, encoded as above as paths through the graph.
Step 2 (clause satisfaction): For each clause Ck , k = 1, . . . , m: we will

extract from Tk−1 only those strings that satisfy Ck , and put them in Tk .
(These successive filtrations in effect do an AND operation.) The goal is
that the DNA in Tk satisfies the first k clauses of the formula. That is,
∀x ∈ Tk ∀ 1 ≤ j ≤ k : Cj (x) = 1. Here are the details.
For k = 0, . . . , m − 1 do the following steps:
Precondition: The strings in Tk satisfy clauses C1 , . . . , Ck .
Let ` = |Ck+1 | (the number of literals in Ck+1 ), and suppose Ck+1 has the
form v1 ∨ · · · ∨ v` , where the vi are literals (positive or negative variables).
Our goal is to find all strings that satisfy at least one of these literals. To
accomplish this we will use an extraction operation E(T, i, a) that extracts

from test tube T all (or most) of the strings whose ith bit is a.
0
Let T k = Tk . Do the following for literals i = 1, . . . , `.
i−1
Positive literal: Suppose vi = xj (some positive literal). Let Tki = E(T k , j, 1)
and let a = 1 (used below). These are the paths that satisfy this positive
literal, since they have 1 in position j.
Negative literal: Suppose vi = x0j (some negative literal). Let Tki =

i−1
E(T k , j, 0) and let a = 0. These are the paths that satisfy this nega-
tive literal, since they have 0 in position j.
i
In either case, Tki are the strings that satisfy literal i of the clause. Let T k =
i−1
E(T k , j, ¬a) be the remaining strings (which do not satisfy this literal).
Continue the process above until all the literals in the clause are processed.
At the end, for each i = 1, . . . , `, Tki will contain the strings that satisfy literal
i of clause k.
Combine Tk1 , . . . , Tk` into Tk+1 . (Combining the test tubes effectively does
OR.) These will be the strings that satisfy at least one of the literals in clause
k + 1.
Postcondition: The strings in Tk+1 satisfy clauses C1 , . . . , Ck+1 .
Continue the above for k = 1, . . . , m.
Step 3 (detection): At this point, the strings in Tm (if any) are those that
satisfy C1 , . . . , Cm , so do a detect operation (for example, with PCR and gel
electrophoresis) to see if there are any strings left. If there are, the formula
is satisfiable; if there aren’t, then it is not.
If the number of literals per clause is fixed (as in the 3-SAT problem),
then performance is linear in m. The main problem with this algorithm is the
effect of errors, but imperfections in extraction are not fatal, so long as there
are enough copies of the desired sequence. In 2002, Adelman’s group solved
a 20-variable 3-SAT problem with 24 clauses, finding the unique satisfying
string.7 In this case the number of possible solutions is 220 ≈ 106 . Since the
degree of the specialized graph used for this problem is small, the number
of possible paths is not excessive (as it might be in the Hamiltonian Path
Problem). They stated, “This computational problem may be the largest
yet solved by nonelectronic means,” and they conjectured that their method
might be extended to 30 variables.
7
Ravinderjit S. Braich, Nickolas Chelyapov, Cliff Johnson, Paul W. K. Rothemund,
Leonard Adleman, “Solution of a 20-Variable 3-SAT Problem on a DNA Computer,”
Science 296 (19 Apr. 2002), 499–502.
B.3 Test tube programming language

Filtering algorithms use a small set of basic DNA operations, which can
be extended to a Test Tube Programming Language (TTPL), such as was
developed in the mid 90s by Lipton and Adleman (Adleman, 1995).
B.3.a Basic Operations

DNA algorithms operate on “test tubes,” which are multi-sets of strings over
Σ = {A, C, T, G}. There are four basic operations (all implementable):
Extract (or separate): There are two complementary extraction (or
separation) operations. Given a test tube t and a string w, +(t, w) returns
all strings in t that have w as a subsequence:
def
+(t, w) = {s ∈ t | ∃u, v ∈ Σ∗ : s = uwv}.
Likewise, −(t, w) returns a test tube of all the remaining strings:
def
−(t, w) = t − +(t, w) (multi-set difference).
Merge: The merge operation combines several test tubes into one test
tube:
def
∪(t1 , t2 , . . . , tn ) = t1 ∪ t2 ∪ · · · ∪ tn .
Detect: The detect operation determines if any DNA strings remain in
a test tube:
def true, if t 6= ∅
detect(t) = .
false, otherwise
Amplify: Given a test tube t, the amplify operation produces two copies
of it: t0 , t00 ← amplify(t). Amplification is a problematic operation, which
depends on the special properties of DNA and RNA, and it may be error
prone. Therefore it is useful to consider a restricted model of DNA computing
that avoids or minimizes the use of amplification.
The following additional operations have been proposed:
Length-separate: This operation produces a test tube containing all
the strands less than a specified length:
def
(t, ≤ n) = {s ∈ t | |s| ≤ n}.
Position-separate: There are two position-separation operations, one

that selects for strings that begin with a given sequence, and one for sequences
that end with it:
def
B(t, w) = {s ∈ t | ∃v ∈ Σ∗ : s = wv},
def
E(t, w) = {s ∈ t | ∃u ∈ Σ∗ : s = uw}.
B.3.b Examples
AllC: The following example algorithm detects if there are any sequences
that contain only C:
procedure [out] = AllC(t, A, T, G)

t ← –(t, A)
t ← –(t, T)
t ← –(t, G)
out ← detect (t)
end procedure
HPP: Adelman’s solution of the HPP can be expressed in TTPL:
procedure [out] = HPP(t, vin, vout)

t ← B(t, vin) //begin with vin
t ← E(t, vout) //end with vout
t ← (t, ≤ 140) //correct length
for i=1 to 5 do //except vin and vout
t ← +(t, s[i]) //contains vertex i
end for
out ← detect(t) //any HP left?
end procedure
SAT: Programming Lipton’s solution to Sat requires another primi-

tive operation, which extracts all sequences for which the jth bit is a ∈ 2:
E(t, j, a). Recall that these are represented by the sequences containing xj
and x0j . Therefore:
E(t, j, 1) = +(t, xj ),
E(t, j, 0) = +(t, x0j ).
procedure [out] = Sat(t)

for k = 1 to m do // for each clause
for i = 1 to n do // for each literal
if C[k][i] = xj // i-th literal in clause k
then t[i] ← E(t,j,1)
else t[i] ← E(t,j,0)
end if
end for
t ← merge(t[1], t[2], . . . , t[n]) // solutions for clauses 1,...,k
end for
out ← detect(t)
end procedure
B.4 Parallel filtering model

The parallel filtering model (PFM) was developed in the mid 90s by Martyn
Amos and colleagues to be a means of describing DNA algorithms for any
NP problem (as opposed to Ableson’s and Lipton’s algorthms, which are
specialized to particular problems). “Our choice is determined by what we
know can be effectively implemented by very precise and complete chemical
reactions within the DNA implementation.”8 All PFM algorithms begin
with a multi-set of all candidate solutions. The PFM differs from other DNA
computation models in that removed strings are discarded and cannot be
used in further operations. Therefore this is a “mark and destroy” approach
to DNA computation.
B.4.a Basic operations

The basic operations are remove, union, copy, and select.
Remove: The operation remove(U, {S1 , . . . , Sn }) removes from U any
strings that contain any of the substrings Si . Remove is implemented by two
primitive operations, mark and destroy:
Mark: mark(U, S) marks all strands that have S as a substring. This is
done by adding S as a primer with polymerase to make it double-stranded.
Destroy: destroy(U ) removes all the marked sequences from U . This is
done by adding a restriction enzyme that cuts up the double-stranded part.
These fragments can be removed by gel electrophoresis, or left in the solution
(since they won’t affect it). Restriction enzymes are much more reliable than
other DNA operations, which is one advantage of the PFM approach.
Union: The operation union({U1 , . . . , Un }, U ) combines in parallel the
multi-sets U1 , . . . , Un into U .
Copy: The operation copy(U, {U1 , . . . , Un }) divides multi-set U into n
equal multi-sets U1 , . . . , Un .
Select: The operation select(U ) returns a random element of U . If U = ∅,
then it returns ∅.
Homogeneous DNA can be detected and sequenced by PCR, and nested
PCR can be used in non-homogeneous cases (multiple solutions). All of
these operations are assumed to be constant-time. Periodic amplification
(especially after copy operations) may be necessary to ensure an adequate
8
Amos, p. 50.
• mark(U, S). This operation marks all strings in the set U which contains
at least one occurrence of the substring S.
• destroy(U ). This operation removes all marked strings from U .
mark(U, S) is implemented by adding to U many copies of a primer corre-

sponding to S (Fig. 5.7b). This primer only anneals to single strands contain-
ing the subsequence S. We then add DNA polymerase to extend the primers
once they have annealed, making only the single strands containing S double
B. FILTERING
stranded (Fig. MODELS
5.7b). 221
(a)
Restriction site Target sequence
(b)
Polymerase extends
Primer block
(c)
Restrict Restrict Restrict
(d)
Fig. 5.7. Implementation of destroy

Figure IV.10: Remove operation implemented by mark and destroy. [source:
Amos]
We may then destroy strands containing S by adding the appropriate restric-
tion enzyme. Double-stranded DNA (i.e. strands marked as containing S) is
cut at the restriction sites embedded within, single strands remaining intact
number of instances. Amos et al. have done a number of experiments to
determine optimum reactions parameters and procedures.
B.4.b Permutations
Amos et al. describe a PFM algorithm for generating all possible permuta-
tions of a set of integers.
algorithm Permutations:
Input: “The input set U consists of all strings of the form p1 i1 p2 i2 · · · pn in

where, for all j, pj uniquely encodes ‘position j’ and each ij is in {1, 2, . . . , n}.
Thus each string consists of n integers with (possibly) many occurrences of
the same integer.”9
9
Amos, p. 51.
Iteration:
for j = 1 to n − 1 do
copy(U, {U1 , U2 , . . . , Un })
for i = 1, 2, . . . , n and all k > j
in parallel do remove(Ui , {pj ij 6= pj i, pk i})
// Ui contains i in jth position and no other is
union({U1 , U2 , . . . , Un }, U )
end for
Pn ← U
In the preceding, remove(Ui , {pj ij 6= pj i, pk i}) means to remove from Ui all

strings that have a pj value not equal to i and all strings containing pk i for any
k > j. For example, if i = 2 and j = n − 1, this remove operation translates
to remove(U2 , {pn−1 1, pn−1 3, pn−1 4, . . . , pn−1 n, pn 2}). That is, it eliminates
all strings except those with 2 in the n − 1 position, and eliminates those
with 2 in the n position. At the end of iteration j we have:
α
z }| {
p1 i1 p2 i2 · · · pj ij pj+1 ij+1 · · · pn in
| {z }
β
where α represents a permutation of j integers from 1, . . . , n, and none of

these integers i1 , . . . , ij are in β.
Amos shows how to do a number of NP-complete problems, including
3-vertex-colorability, HPP, subgraph isomorphism, and maximum clique.
C. FORMAL MODELS 223
C Formal models
C.1 Sticker systems
C.1.a Basic operations
The sticker model was developed by Rosweis et al. in the mid-1990s. It
depends primarily on separation by means of hybridization and makes no
use of strand extension and enzymes. It implements a sort of random-access
binary memory. Each bit position is represented by a substrand of length
m. A memory strand comprises k contiguous substrands, and so has length
n = km and can store k bits. Sticker strands or stickers are strands that are
complementary to substrands representing bits. When a sticker is bound to
a bit, it represents 1, and if no sticker is bound, the bit is 0. Such a strand,
which is partly double and partly single, is called a complex strand.
Computations begin with a prepared library of strings. A (k, l) library
uses the first l ≤ k bits as inputs to the algorithm, and the remaining k −l for
output and working storage. Therefore, the last k − l are initially 0. There
are four basic operations, which act on multi-sets of binary strings:
Merge: Creates the union of two tubes (multi-sets).
Separate: The operation separate(N, i) separates a tube N into two
tubes: +(N, i) contains all strings in which bit i is 1, and −(N, i) contains
all strings in which bit i is 0.
Set: The operation set(N, i) produces a tube in which every string from
N has had its ith bit set to 1.
Clear: The operation clear(N, i) produces a tube in which every string
from N has had its ith bit cleared to 0.
C.1.b Set cover problem

The set cover problem is a classic NP-complete problem. Given a finite
set of p objects S, and a finite collection of subsets of S (C1 , . . . , Cq ⊂ S)
whose union is S, find the smallest collection of these subsets whose union
is S. For an example, consider S = {1, 2, 3, 4, 5} and C1 = {3, 4, 5}, C2 =
{1, 3, 4}, C3 = {1, 2, 5}, C4 = {3, 4}. In this case there are three minimal
solutions: {C1 , C3 }, {C3 , C4 }, {C2 , C3 }.
algorithm Minimum Set Cover:
Data representation: The memory strands are of size k = p + q. Each

strand represents a collection of subsets, and the first q bits encode which
subsets are in the collection; call them subset bits. For example 1011 rep-
resents {C1 , C3 , C4 } and 0010 represents {C3 }. Eventually, the last p bits
will represent the union of the collection, that is, the elements of S that are
contained in at lease one subset in the collection; call them element bits. For
example, 0101 10110 represents {C2 , C4 } {1, 3, 4}.
Library: The algorithm begins with the (p + q, q) library, which must be

initialized to reflect the subsets’ members.
Step 1 (initialization): For all strands, if the i subset bit is set, then set
the bits for all the elements of that subset. Call the result tube N0 . This is
accomplished by the following code:
Initialize (p + q, q) library in N0
for i = 1 to q do
separate(+(N0 , i), −(N0 , i)) //separate those with subset i
for j = 1 to |Ci | do
set(+(N0 , i), q + cji ) //set bit for jth element of set i
end for
N0 ← merge(+(N0 , i), −(N0 , i)) //recombine
end for
Step 2 (retain covers): Retain only the strands that represent collections
that cover the set. To do this, retain in N0 only the strands whose last p bits
are set.
for i = q + 1 to q + p do
N0 ← +(N0 , i) //retain those with element i
end for
C. FORMAL MODELS 225
3.2 Filtering Models 59
N0 N1 N2 N3 N4
(3,4), (2,3),
(2,3,4), (1,3),
(1,3,4), (1,2,3)
(1,2,3,4)
Separate on 1
(3,4), (2,3), (1,3)

(1,3,4), (1,2,3)
(2,3,4)
(1,2,3,4)
Separate on 2
(1,3)
(1,2,3)
(1,3,4) (2,3)
(3,4) (2,3,4) (1,2,3,4)
Separate on 3
(1,3) (1,2,3)
(1,3,4) (2,3)
(3,4) (2,3,4) (1,2,3,4)
Separate on 4
(1,3)
(3,4) (2,3) (1,3,4) (1,2,3)
(2,3,4) (1,2,3,4)
Fig. 3.4. Sorting procedure

Figure IV.11: Sorting of covers by repeated separations. [source: Amos, Fig.
3.4]
and so on until we find a tube that contains a covering. In this case, tube
N2 contains three coverings, each using two bags. The algorithm is formally
expressed within the sticker model as follows.
(1) Initialize (p,q) library in tube N0

(2) for i = 1 to q do begin
(3)
Step 3 (isolate 0 ← separatei(+(N
Nminimum 0 , i), −(N
covers): 0 , i))
Tube N0 now holds all covers, so we
(4) for j = 1 to | C |
have to somehow sort its contents to find thej minimum cover(s). Set up a
i
(5) set(+(N0 , i), q + ci )
row of tubes
(6)
N0 , N1 , . . . end
, Nq .for
We will arrange things so that Ni contains the
covers of size
(7) i; then we N0 ← merge(+(N0 , i), −(N0 , tube
just have to find the first i)) with some DNA in it.
(8) end for
Sorting: For i = 1, . . . , q, “drag” to the right all collections containing Ci ,

Thisfor
that is, section
whichsetsbit
thei object
is set identifying
(see Fig. substrands. Noteisthat
IV.11). This cji denotes the
accomplished by the
jth element
following code:10of set Ci . We now separate out for further use only those memory
complexes where each of the last p substrands is set to on.
for i = 0 (1) q −i 1= do
to for q + 1 to q + p do begin
for j =(2)
i down N0to 0 do0 , i)
← +(N
separate(+(N
(3) end for j , i + 1), −(Nj , i + 1)) //those that do & don’t have i
10
Corrected from Amos p. 60.
Nj+1 ← merge(+(Nj , i + 1), Nj+1 ) //move those that do to Nj+1

Nj ← −(Nj , i + 1) //leave those that don’t in Nj
end for
end for
Detection: Find the minimum i such that Ni contains DNA; Ni contains

the minimum covers.
The algorithm is O(pq).
C.2 Splicing systems

It has been argued that the full power of a TM requires some sort of string
editing operation. Therefore, beginning with Tom Head (1987), a number of
splcing systems have been defined. The splicing operations takes two strings
S = S1 S2 and T = T1 T2 and performs a “crossover” at a specified location,
yielding S1 T2 and T1 S2 . Finite extended splicing systems have been shown
to be computationally universal (1996).
The Parallel Associative Memory (PAM) Model was defined by Reif in
1995. It is based on a restricted splicing operation called parallel associative
matching (PA-Match) operation, which is named Rsplice. Suppose S = S1 S2
and T = T1 T2 . Then,
Rsplice(S, T ) = S1 T2 , if S2 = T1 ,
and is undefined otherwise. The PAM model can simulate nondeterministic

TMs and parallel random access machines.
D. ENZYMATIC COMPUTATION 227
Figure IV.12: The fokI restriction enzyme bound to DNA. [source: wikipedia]
D Enzymatic computation
The molecular computation processes that we have seen so far are externally
controlled by a person or conventional automatic controller sequencing the
chemical operations.11 In autonomous molecular computation the chemical
processes sequence themselves so that they do not require external control.
This is also called “one-pot” molecular computation; that is, you put all the
reactants in one pot, and the molecular processes do the rest. Autonomous
molecular computation is essential for, example, controlling drug delivery in
the body.
Shapiro and his colleagues have demonstrated how to implement finite
state machines (FSMs) by autonomous molecular computation. In addition
to DNA, it uses a restriction enzyme, ligase, and ATP (for fuel).
The implementation is based on the fokI restriction enzyme. “Once the
11
This section is based primarily on: (1) Yaakov Benenson, Tamar Paz-Elizur, Rivka
Adar, Ehud Keinan, Zvi Livneh, and Ehud Shapiro. Programmable and autonomous
computing machine made of biomolecules. Nature, 414:430–434, 2001.
(2) Yaakov Benenson, Rivka Adam, Tamar Paz-Livneh, and Ehud Shapiro. DNA molecule
provides a computing machine with both data and fuel. Proceedings of the National
Academies of Science, 100(5):2191–2196, 2003.
protein is bound to duplex DNA via its DNA-binding domain at the 50 -

GGATG-30 : 50 -CATCC-30 recognition site, the DNA cleavage domain is
activated and cleaves, without further sequence specificity, the first strand
9 nucleotides downstream and the second strand 13 nucleotides upstream of
the nearest nucleotide of the recognition site.”12 It leaves 4-nucleotide sticky
ends. That is, the restriction enzyme cuts the DNA as follows:
GGATGNNNNNNNNN NNNNNNNNNN
CCTACNNNNNNNNNNNNN NNNNNN
The ‘N’s can be any nucleotides (respecting Watson-Crick complementarity,

of course).
Both the current state of the FSM and the input string are represented
by a double DNA strand. fokI operates at the beginning of this string and
leaves a sticky end that encodes both the current state and the next input
symbol (see p. 229 below). The state transitions of the FSM are encoded
in transition molecules, which have sticky ends complementary to the state-
symbol code at the beginning of the string. The rest of a transition molecule
ensures that the string properly encodes the new state, including adding a
new recognition site for the enzyme. A matching transition molecule binds
to the string’s sticky end, providing a new opportunity for fokI to operate,
and so the process continues.
A state transition (q, s1 ) → q 0 can be represented:
[q, s1 ]s2 s3 · · · sn t =⇒ [q 0 , s2 ]s3 · · · sn t
where [q, s] represents a DNA sequence encoding both state q and symbol
s, and t is a terminator for the string. The fokI enzyme cleaves off [q, s1 ] in
such a way that a transition molecule can bind to the sticky end in a way
that encodes [q 0 , s2 ]. A special terminator symbol marks the end of the input
string.
As an example we will consider a two-state FSM on {a, b} that accepts
strings with an even number of ‘b’s. Ignoring the terminator, DNA codes are
assigned to the two symbols ‘a’ and ‘b’ as follows:
a →
7 AAααaa
b → 7 BBββbb
12
wikipedia, s.v. fokI.
D. ENZYMATIC COMPUTATION 229
where A, α, a, B, β, b are unspecified (by me) bases.13 The bases are selected
in such a way that either the first four bases (AAαα, BBββ) or the last four
bases (ααaa, ββbb) encode the symbol. These alternatives represent the two
machine states.
The transition molecules are constructed so that the distance between
the recognition site (for fokI) and the next symbol depends on new state. As
a consequence, when fokI operates it cleaves the next symbol code at a place
that depends on the state. Therefore the sticky end encodes the state in the
way that it represents the next symbol:
[q0 , a] 7→ ααaa
[q1 , a] 7 → AAαα
[q0 , b] 7 → ββbb
[q1 , b] 7 → BBββ
The transition molecules are:
(q0 , a) → q0 GGATGN N N
CCTACN N N ααaa
(q1 , a) → q1 GGATGN N N
CCTACN N N AAαα
(q0 , b) → q1 GGATGN N N N N
CCTACN N N N N ββbb
(q1 , b) → q0 GGATGN
CCTACN BBββ
The N s represent any bases as before. They are used as spacers to adjust
the restriction site to represent the new state.
After transition to the new state the sense strand will look like this (for
convenience assuming the next symbol is ‘a’):
q0 GGATGN XXY Y yyAA.ααaa

q1 GGATGN N N XXY Y yy.AAααaa
This is after attachment of the transition molecule but before restriction.
Here XX represents either spacers or the first two bases of the previous first
symbol, and Y Y yy represents the last four bases of this symbol. The cleavage
site is indicated by a period.
13
Note that repeated letters might refer to different bases.
The longest strings processed in the PNAS experiments were 12.14 Op-
eration required about 20 seconds per step. However, the parallel speed was
about 6.6 × 1010 ops/s/µl. Energy consumption was about 34kT per tran-
sition, which is only about 50× the von Neumann-Landauer limit (kT ln 2).
The authors note, “Reaction rates were surprisingly insensitive to tempera-
ture and remained similar over the range of 2–20◦ C.” This implementation
also handles nondeterministic FSMs (just put in all the transition molecules),
but the yield decreases exponentially (due to following out all the nondeter-
ministic paths, breadth-first). Therefore it doesn’t seem to be practical for
nondeterministic machines.
14
Benenson et al., PNAS 100 (5), March 4, 2003.
Chapter V
Analog Computation
1
A Definition
Although analog computation was eclipsed by digital computation in the
second half of the twentieth century, it is returning as an important alterna-
tive computing technology. Indeed, as explained in this chapter, theoretical
results imply that analog computation can escape from the limitations of
digital computation. Furthermore, analog computation has emerged as an
important theoretical framework for discussing computation in the brain and
other natural systems.
Analog computation gets its name from an analogy, or systematic rela-
tionship, between the physical processes in the computer and those in the
system it is intended to model or simulate (the primary system). For exam-
ple, the electrical quantities voltage, current, and conductance might be used
as analogs of the fluid pressure, flow rate, and pipe diameter of a hydrolic sys-
tem. More specifically, in traditional analog computation, physical quantities
in the computation obey the same mathematical laws as physical quantities
in the primary system. Thus the computational quantities are proportional
1
This chapter is based on an unedited draft for an article that appeared in the Ency-
clopedia of Complexity and System Science (Springer, 2008).
231
232 CHAPTER V. ANALOG COMPUTATION
to the modeled quantities. This is in contrast to digital computation, in which

quantities are represented by strings of symbols (e.g., binary digits) that have
no direct physical relationship to the modeled quantities. According to the
Oxford English Dictionary (2nd ed., s.vv. analogue, digital), these usages
emerged in the 1940s.
However, in a fundamental sense all computing is based on an analogy,
that is, on a systematic relationship between the states and processes in the
computer and those in the primary system. In a digital computer, the rela-
tionship is more abstract and complex than simple proportionality, but even
so simple an analog computer as a slide rule goes beyond strict proportion
(i.e., distance on the rule is proportional to the logarithm of the number).
In both analog and digital computation—indeed in all computation—the
relevant abstract mathematical structure of the problem is realized in the
physical states and processes of the computer, but the realization may be
more or less direct (MacLennan, 1994a,c, 2004).
Therefore, despite the etymologies of the terms “analog” and “digital,”
in modern usage the principal distinction between digital and analog com-
putation is that the former operates on discrete representations in discrete
steps, while the later operated on continuous representations by means of
continuous processes (e.g., MacLennan 2004, Siegelmann 1999, p. 147, Small
2001, p. 30, Weyrick 1969, p. 3). That is, the primary distinction resides in
the topologies of the states and processes, and it would be more accurate to
refer to discrete and continuous computation (Goldstine, 1972, p. 39). (Con-
sider so-called analog and digital clocks. The principal difference resides in
the continuity or discreteness of the representation of time; the motion of the
two (or three) hands of an “analog” clock do not mimic the motion of the
rotating earth or the position of the sun relative to it.)
B Introduction
B.1 History
B.1.a Pre-electronic analog computation
Just like digital calculation, analog computation was originally performed by
hand. Thus we find several analog computational procedures in the “con-
structions” of Euclidean geometry (Euclid, fl. 300 BCE), which derive from
techniques used in ancient surveying and architecture. For example, Problem
B. INTRODUCTION 233
Figure V.1: Euclid Problem VI.13: To find a mean proportional between two
given straight lines. Solution: Take on any line AC parts AB, BC respectively
equal to X, Y. On AC describe a semicircle ADC. Erect BD at right angles to
AC, meeting the semicircle in D. BD will be the mean proportional required.
II.51 is “to divide a given straight line into two parts, so that the rectangle
contained by the whole and one of the parts shall be equal to the square
of the other part.” Also, Problem VI.13 is “to find a mean proportional
between two given straight lines” (Fig. V.1), and VI.30 is “to cut a given
straight line in extreme and mean ratio.” These procedures do not make
use of measurements in terms of any fixed unit or of digital calculation; the
lengths and other continuous quantities are manipulated directly (via com-
pass and straightedge). On the other hand, the techniques involve discrete,
precise operational steps, and so they can be considered algorithms, but over
continuous magnitudes rather than discrete numbers.
It is interesting to note that the ancient Greeks distinguished continuous
magnitudes (Grk., megethoi), which have physical dimensions (e.g., length,
area, rate), from discrete numbers (Grk., arithmoi), which do not (Maziarz &
Greenwood, 1968). Euclid axiomatizes them separately (magnitudes in Book
V, numbers in Book VII), and a mathematical system comprising both dis-
crete and continuous quantities was not achieved until the nineteenth century
in the work of Weierstrass and Dedekind.
The earliest known mechanical analog computer is the “Antikythera mech-
anism,” which was found in 1900 in a shipwreck under the sea near the Greek
island of Antikythera (between Kythera and Crete) (Figs. V.3, V.4). It dates
to the second century BCE and appears to be intended for astronomical cal-
culations. The device is sophisticated (at least 70 gears) and well engineered,
suggesting that it was not the first of its type, and therefore that other analog
Figure V.2: Euclid Problem VI.30 from Byrne’s First

Six Books of Euclid’s Elements (1847). [source:
http://www.math.ubc.ca/˜cass/Euclid/byrne.html]
B. INTRODUCTION 235
Figure V.3: The Antikythera Mechanism. An ancient analog computer

(2nd century BCE) for astronomical calculations including eclipses. [source:
wikipedia]
(a) front (b) back
Figure V.4: Computer reconstruction of Antikythera Mechanism. [source:

wikipedia]
B. INTRODUCTION 237
Figure V.5: Example nomograph.

computing devices may have been used in the ancient Mediterranean world
(Freeth et al., 2006). Indeed, according to Cicero (Rep. 22) and other au-
thors, Archimedes (c. 287–c. 212 BCE) and other ancient scientists also built
analog computers, such as armillary spheres, for astronomical simulation and
computation. Other antique mechanical analog computers include the astro-
labe, which is used for the determination of longitude and a variety of other
astronomical purposes, and the torquetum, which converts astronomical mea-
surements between equatorial, ecliptic, and horizontal coordinates.
A class of special-purpose analog computer, which is simple in conception
but may be used for a wide range of purposes, is the nomograph (also, nomo-
gram, alignment chart) (Fig. V.5). In its most common form, it permits the
solution of quite arbitrary equations in three real variables, f (u, v, w) = 0.
The nomograph is a chart or graph with scales for each of the variables;
typically these scales are curved and have non-uniform numerical markings.
Given values for any two of the variables, a straightedge is laid across their
positions on their scales, and the value of the third variable is read off where
the straightedge crosses the third scale. Nomographs were used to solve many
problems in engineering and applied mathematics. They improve intuitive
understanding by allowing the relationships among the variables to be visu-
alized, and facilitate exploring their variation by moving the straightedge.
Lipka (1918) is an example of a course in graphical and mechanical methods
of analog computation, including nomographs and slide rules.
Until the introduction of portable electronic calculators in the early 1970s,
the slide rule was the most familiar analog computing device. Slide rules use
logarithms for multiplication and division, and they were invented in the early
seventeenth century shortly after John Napier’s description of logarithms.
The mid-nineteenth century saw the development of the field analogy
method by G. Kirchhoff (1824–87) and others (Kirchhoff, 1845). In this ap-
proach an electrical field in an electrolytic tank or conductive paper was
used to solve two-dimensional boundary problems for temperature distribu-
tions and magnetic fields (Small, 2001, p. 34). It is an early example of
analog field computation, which operates on continuous spatial distributions
of quantity (i.e., fields).
In the nineteenth century a number of mechanical analog computers were
developed for integration and differentiation (e.g., Lipka 1918, pp. 246–56,
Clymer 1993). For example, the planimeter measures the area under a curve
or within a closed boundary (Fig. V.6). While the operator moves a pointer
along the curve, a rotating wheel accumulates the area. Similarly, the inte-
B. INTRODUCTION 239
Figure V.6: Planimeter for measuring the area inside an arbitrary curve
(1908). [source: wikipedia]
graph is able to draw the integral of a given function as its shape is traced
(Fig. V.7). Other mechanical devices can draw the derivative of a curve or
compute a tangent line at a given point.
In the late nineteenth century William

Thomson, Lord Kelvin, constructed several
analog computers, including a “tide predic-
tor” and a “harmonic analyzer,” which com-
puted the Fourier coefficients of a tidal curve
(Thomson, 1878, 1938) (Fig. V.8). In 1876
he described how the mechanical integrators
invented by his brother could be connected
together in a feedback loop in order to solve
second and higher order differential equa-
tions (Small 2001, pp. 34–5, 42, Thomson
1876). He was unable to construct this dif-
ferential analyzer, which had to await the Figure V.8: Lord Kelvin’s ana-
invention of the torque amplifier in 1927. log tide computer. [source:
The torque amplifier and other technical wikipedia]
advancements permitted Vannevar Bush at
MIT to construct the first practical differen-
tial analyzer in 1930 (Small, 2001, pp. 42–5) (Fig. V.9). It had six integrators
and could also do addition, subtraction, multiplication, and division. Input
Figure V.7: Integraph for drawing the integral of an arbitrary curve (Lipka,
1918).
B. INTRODUCTION 241
Figure V.9: Meccano differential analyzer at Cambridge University, 1938.

The computer was constructed by Douglas Hartree based on Vannevar Bush’s
design. [source: wikipedia]
data were entered in the form of continuous curves, and the machine auto-
matically plotted the output curves continuously as the equations were inte-
grated. Similar differential analyzers were constructed at other laboratories
in the US and the UK.
Setting up a problem on the MIT differential analyzer took a long time;
gears and rods had to be arranged to define the required dependencies among
the variables. Bush later designed a much more sophisticated machine, the
Rockefeller Differential Analyzer, which became operational in 1947. With
18 integrators (out of a planned 30), it provided programmatic control of ma-
chine setup, and permitted several jobs to be run simultaneously. Mechanical
differential analyzers were rapidly supplanted by electronic analog comput-
ers in the mid-1950s, and most were disassembled in the 1960s (Bowles 1996,
Owens 1986, Small 2001, pp. 50–5).
During World War II, and even later wars, an important application
of optical and mechanical analog computation was in “gun directors” and
“bomb sights,” which performed ballistic computations to accurately target
artillery and dropped ordnance.
B.1.b Electronic analog computation in the 20th century

It is commonly supposed that electronic analog computers were superior
to mechanical analog computers, and they were in many respects, including
speed, cost, ease of construction, size, and portability (Small, 2001, pp. 54–6).
On the other hand, mechanical integrators produced higher precision results
(0.1%, vs. 1% for early electronic devices) and had greater mathematical
flexibility (they were able to integrate with respect to any variable, not just
time). However, many important applications did not require high precision
and focused on dynamic systems for which time integration was sufficient;
for these, electronic analog computers were superior.
Analog computers (non-electronic as well as electronic) can be divided
into active-element and passive-element computers; the former involve some
kind of amplification, the latter do not (Truitt & Rogers, 1960, pp. 2-1–4).
Passive-element computers included the network analyzers that were devel-
oped in the 1920s to analyze electric power distribution networks, and which
continued in use through the 1950s (Small, 2001, pp. 35–40). They were
also applied to problems in thermodynamics, aircraft design, and mechan-
ical engineering. In these systems networks or grids of resistive elements
or reactive elements (i.e., involving capacitance and inductance as well as
B. INTRODUCTION 243
resistance) were used to model the spatial distribution of physical quanti-

ties such as voltage, current, and power (in electric distribution networks),
electrical potential in space, stress in solid materials, temperature (in heat
diffusion problems), pressure, fluid flow rate, and wave amplitude (Truitt
& Rogers, 1960, p. 2-2). That is, network analyzers dealt with partial dif-
ferential equations (PDEs), whereas active-element computers, such as the
differential analyzer and its electronic successors, were restricted to ordinary
differential equations (ODEs) in which time was the independent variable.
Large network analyzers are early examples of analog field computers.
Electronic analog computers became feasible after the invention of the
DC operational amplifier (“op amp”) c. 1940 (Small, 2001, pp. 64, 67–72).
Already in the 1930s scientists at Bell Telephone Laboratories (BTL) had
developed the DC-coupled feedback-stabilized amplifier, which is the basis
of the op amp. In 1940, as the USA prepared to enter World War II, D.
L. Parkinson at BTL had a dream in which he saw DC amplifiers being
used to control an anti-aircraft gun. As a consequence, with his colleagues
C. A. Lovell and B. T. Weber, he wrote a series of papers on “electrical
mathematics,” which described electrical circuits to “operationalize” addi-
tion, subtraction, integration, differentiation, etc. The project to produce
an electronic gun-director led to the development and refinement of DC op
amps suitable for analog computation.
The war-time work at BTL was focused primarily on control applications
of analog devices, such as the gun-director. Other researchers, such as E.
Lakatos at BTL, were more interested in applying them to general-purpose
analog computation for science and engineering, which resulted in the de-
sign of the General Purpose Analog Computer (GPAC), also called “Gypsy,”
completed in 1949 (Small, 2001, pp. 69–71). Building on the BTL op amp
design, fundamental work on electronic analog computation was conducted
at Columbia University in the 1940s. In particular, this research showed how
analog computation could be applied to the simulation of dynamic systems
and to the solution of nonlinear equations.
Commercial general-purpose analog computers (GPACs) emerged in the
late 1940s and early 1950s (Small, 2001, pp. 72–3) (Fig. V.10). Typically they
provided several dozen integrators, but several GPACs could be connected
together to solve larger problems. Later, large-scale GPACs might have up
to 500 amplifiers and compute with 0.01%–0.1% precision (Truitt & Rogers,
1960, pp. 2–33).
Besides integrators, typical GPACs provided adders, subtracters, multi-
Figure V.10: Analog computer at Lewis Flight Propulsion Laboratory circa

1949.
B. INTRODUCTION 245
pliers, fixed function generators (e.g., logarithms, exponentials, trigonometric

functions), and variable function generators (for user-defined functions) (Tru-
itt & Rogers, 1960, chs. 1.3, 2.4). A GPAC was programmed by connecting
these components together, often by means of a patch panel. In addition,
parameters could be set by adjusting potentiometers (attenuators), and ar-
bitrary functions could be entered in the form of graphs (Truitt & Rogers,
1960, pp. 1-72–81, 2-154–156). Output devices plotted data continuously or
displayed it numerically (Truitt & Rogers, 1960, pp. 3-1–30).
The most basic way of using a GPAC was in single-shot mode (Weyrick,
1969, pp. 168–70). First, parameters and initial values were entered into the
potentiometers. Next, putting a master switch in “reset” mode controlled
relays to apply the initial values to the integrators. Turning the switch to
“operate” or “compute” mode allowed the computation to take place (i.e., the
integrators to integrate). Finally, placing the switch in “hold” mode stopped
the computation and stabilized the values, allowing them to be read from
the computer (e.g., on voltmeters). Although single-shot operation was also
called “slow operation” (in comparison to “repetitive operation,” discussed
next), it was in practice quite fast. Because all of the devices computed in
parallel and at electronic speeds, analog computers usually solved problems
in real-time but often much faster (Truitt & Rogers 1960, pp. 1-30–32, Small
2001, p. 72).
One common application of GPACs was to explore the effect of one or
more parameters on the behavior of a system. To facilitate this exploration
of the parameter space, some GPACs provided a repetitive operation mode,
which worked as follows (Weyrick 1969, p. 170, Small 2001, p. 72). An
electronic clock switched the computer between reset and compute modes at
an adjustable rate (e.g., 10–1000 cycles per second) (Ashley, 1963, p. 280, n.
1). In effect the simulation was rerun at the clock rate, but if any parameters
were adjusted, the simulation results would vary along with them. Therefore,
within a few seconds, an entire family of related simulations could be run.
More importantly, the operator could acquire an intuitive understanding of
the system’s dependence on its parameters.
B.1.c The eclipse of analog computing

A common view is that electronic analog computers were a primitive pre-
decessor of the digital computer, and that their use was just a historical
episode, or even a digression, in the inevitable triumph of digital technol-
ogy. It is supposed that the current digital hegemony is a simple matter of

technological superiority. However, the history is much more complicated,
and involves a number of social, economic, historical, pedagogical, and also
technical factors, which are outside the scope of this book (see Small 1993
and Small 2001, especially ch. 8, for more information). In any case, begin-
ning after World War II and continuing for twenty-five years, there was lively
debate about the relative merits of analog and digital computation.
Speed was an oft-cited advantage of analog computers (Small, 2001, ch.
8). While early digital computers were much faster than mechanical dif-
ferential analyzers, they were slower (often by several orders of magnitude)
than electronic analog computers. Furthermore, although digital computers
could perform individual arithmetic operations rapidly, complete problems
were solved sequentially, one operation at a time, whereas analog comput-
ers operated in parallel. Thus it was argued that increasingly large problems
required more time to solve on a digital computer, whereas on an analog com-
puter they might require more hardware but not more time. Even as digital
computing speed was improved, analog computing retained its advantage for
several decades, but this advantage eroded steadily.
Another important issue was the comparative precision of digital and ana-
log computation (Small, 2001, ch. 8). Analog computers typically computed
with three or four digits of precision, and it was very expensive to do much
better, due to the difficulty of manufacturing the parts and other factors. In
contrast, digital computers could perform arithmetic operations with many
digits of precision, and the hardware cost was approximately proportional
to the number of digits. Against this, analog computing advocates argued
that many problems did not require such high precision, because the mea-
surements were known to only a few significant figures and the mathematical
models were approximations. Further, they distinguished between precision
and accuracy, which refers to the conformity of the computation to physi-
cal reality, and they argued that digital computation was often less accurate
than analog, due to numerical limitations (e.g., truncation, cumulative error
in numerical integration). Nevertheless, some important applications, such
as the calculation of missile trajectories, required greater precision, and for
these, digital computation had the advantage. Indeed, to some extent pre-
cision was viewed as inherently desirable, even in applications where it was
unimportant, and it was easily mistaken for accuracy. (See Sec. C.4.a for
more on precision and accuracy.)
There was even a social factor involved, in that the written programs,
B. INTRODUCTION 247
precision, and exactness of digital computation were associated with mathe-

matics and science, but the hands-on operation, parameter variation, and ap-
proximate solutions of analog computation were associated with engineering,
and so analog computing inherited “the lower status of engineering vis-à-vis
science” (Small, 2001, p. 251). Thus the status of digital computing was fur-
ther enhanced as engineering became more mathematical and scientific after
World War II (Small, 2001, pp. 247–51).
Already by the mid-1950s the competition between analog and digital
had evolved into the idea that they were complementary technologies. This
resulted in the development of a variety of hybrid analog/digital computing
systems (Small, 2001, pp. 251–3, 263–6). In some cases this involved using a
digital computer to control an analog computer by using digital logic to con-
nect the analog computing elements, to set parameters, and to gather data.
This improved the accessibility and usability of analog computers, but had
the disadvantage of distancing the user from the physical analog system. The
intercontinental ballistic missile program in the USA stimulated the further
development of hybrid computers in the late 1950s and 1960s (Small, 1993).
These applications required the speed of analog computation to simulate the
closed-loop control systems and the precision of digital computation for ac-
curate computation of trajectories. However, by the early 1970s hybrids were
being displaced by all-digital systems. Certainly part of the reason was the
steady improvement in digital technology, driven by a vibrant digital com-
puter industry, but contemporaries also pointed to an inaccurate perception
that analog computing was obsolete and to a lack of education about the
advantages and techniques of analog computing.
Another argument made in favor of digital computers was that they
were general-purpose, since they could be used in business data processing
and other application domains, whereas analog computers were essentially
special-purpose, since they were limited to scientific computation (Small,
2001, pp. 248–50). Against this it was argued that all computing is essen-
tially computing by analogy, and therefore analog computation was general-
purpose because the class of analog computers included digital computers!
(See also Sec. A on computing by analogy.) Be that as it may, analog com-
putation, as normally understood, is restricted to continuous variables, and
so it was not immediately applicable to discrete data, such as that manipu-
lated in business computing and other nonscientific applications. Therefore
business (and eventually consumer) applications motivated the computer in-
dustry’s investment in digital computer technology at the expense of analog
technology.
Although it is commonly believed that analog computers quickly disap-
peared after digital computers became available, this is inaccurate, for both
general-purpose and special-purpose analog computers have continued to be
used in specialized applications to the present time. For example, a general-
purpose electrical (vs. electronic) analog computer, the Anacom, was still
in use in 1991. This is not technological atavism, for “there is no doubt
considerable truth in the fact that Anacom continued to be used because it
effectively met a need in a historically neglected but nevertheless important
computer application area” (Aspray, 1993). As mentioned, the reasons for
the eclipse of analog computing were not simply the technological superiority
of digital computation; the conditions were much more complex. Therefore
a change in conditions has necessitated a reevaluation of analog technology.
B.1.d Analog VLSI

In the mid-1980s, Carver Mead, who already had made important contri-
butions to digital VLSI technology, began to advocate for the development
of analog VLSI (Mead, 1987, 1989). His motivation was that “the nervous
system of even a very simple animal contains computing paradigms that are
orders of magnitude more effective than are those found in systems made
by humans” and that they “can be realized in our most commonly available
technology—silicon integrated circuits” (Mead, 1989, p. xi). However, he
argued, since these natural computation systems are analog and highly non-
linear, progress would require understanding neural information processing
in animals and applying it in a new analog VLSI technology.
Because analog computation is closer to the physical laws by which all
computation is realized (which are continuous), analog circuits often use
fewer devices than corresponding digital circuits. For example, a four-quadrant
adder (capable of adding two signed numbers) can be fabricated from four
transistors (Mead, 1989, pp. 87–8), and a four-quadrant multiplier from nine
to seventeen, depending on the required range of operation (Mead, 1989, pp.
90–6). Intuitions derived from digital logic about what is simple or complex
to compute are often misleading when applied to analog computation. For ex-
ample, two transistors are sufficient to compute the logarithm or exponential,
five for the hyperbolic tangent (which is very useful in neural computation),
and three for the square root (Mead, 1989, pp. 70–1, 97–9). Thus analog
VLSI is an attractive approach to “post-Moore’s Law computing” (see Sec.
B. INTRODUCTION 249
Figure V.11: Mahowald/Mead silicon retina.
H, p. 284 below). Mead and his colleagues demonstrated a number of analog

VLSI devices inspired by the nervous system, including a “silicon retina” and
an “electronic cochlea” (Fig. V.11) (Mead, 1989, chs. 15–16), research that
has lead to a renaissance of interest in electronic analog computing.
B.1.e Field-programmable analog arrays
Field Programmable Analog Arrays (FPAAs) permit the programming of

analog VLSI systems comparable to Field Programmable Gate Arrays (FP-
GAs) for digital systems (Fig. V.12). An FPAA comprises a number of iden-
tical Computational Analog Blocks (CABs), each of which contains a small
number of analog computing elements. Programmable switching matrices
control the interconnections among the elements of a CAB and the intercon-
nections between the CABs. Contemporary FPAAs make use of floating-gate
transistors, in which the gate has no DC connection to other circuit elements
and thus is able to hold a charge indefinitely cite . Therefore the floating
gate can be used to store a continuous value that governs the impedance
of the transistor by several orders of magnitude. The gate charge can be
changed by processes such as electron tunneling, which increases the charge,
and hot-electron injection, which decreases it. Digital decoders allow indi-
vidual floating-gate transistors in the switching matrices to be addressed and
programmed. At the extremes of zero and infinite impedance the transistors
operate as perfect switches, connecting or disconnecting circuit elements.
Programming the connections to these extreme values is time consuming,
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS
Fig. 2. RASP 3.0 functional block diagram illustrating the resulting computational blocks and resulting routing architecture. The infrastructure control
includes a µP developed from an open-source MSP 430 processor [1], as well as on-chip structures include the on-chip DACs, current-to-voltage conversion,
Figure V.12: RASP 3.0 FPAA. “RASP 3.0 integrates divergent concepts from
and voltage measurement, to program each FG device. The FG switches in the connection (C) blocks, the switch (S) blocks, and the local routing are a single
multiple previous FPAA designs . . . along with low-power digital computa-
pFET FG transistor programmed to be a closed switch over the entire fabric signal swing of 0–2.5 V [9]. The CABs and the CLBs are similar to previous
approaches [3]. Eight, four input BLE lookup tables with a latch comprise the CLB blocks. Transconductance amplifiers, transistors, capacitors, switches, and
tion, including a 16-bit microprocessor (µP), interface circuitry, and DACs
other elements comprise the CAB blocks.
+ ADCs.
digital Theapproaches,
computational FPAA SoC die timing,
capacitance, photorapid measures 12 and
as utilization mmaccessibility
7 mm,of fabricated
the resulting computational
reconfigurability of the routing fabric, implementation
in a 350-nm standard CMOS process. The die photo identifiesof data resources for the data flow. The µP,
closestSRAM
high utilization struc-
converters in the mixed-mode fabric, and utilizing the routing ture (i.e., PSoC5 [10]) has nearly a 600 000 factor less in
memory,
fabric as part DACs, and programming (DACs +parameter
of the computation. ADC)densityinfrastructure; the mixed
than this SoC FPAA device.
array of the
This paper FPAA fabric
demonstrates (SectionisV)
composed of interdigitated analog (A) and digital
the first embed-
ded classifier structure (command-word recognition) compiled II. A RCHITECTURE D ESCRIPTION OF THE FPAA SoC IC
(D) configurable blocks on a single routing grid.
onto a single FPAA device, going from sensor input (audio) to
DACs and programming in-
Fig. 2 shows the block diagram for the RASP 3.0 FPAA
frastructure
classified are accessed
word, experimentally through
demonstrated memory-mapped
in analog hard- IC based on registers.” (George
a Manhattan FPAA et al.,including the
architecture,
ware.
2016) This demonstration is a small fraction of the overall IC. array of computation blocks and routing, composed of con-
The SoC FPAA compiled system system power (23 µW) is nection (C) and switch (S) blocks. This configurable fabric
consistent with the ×1000 improvement factor (comparison of effectively integrates analog (A) and digital (D) components
MACs) for physical computation over digital approaches, with in a hardware platform easily mapped toward compiler tools.
future opportunities for improved performance in the same IC. The switchable analog and digital devices are a combination of
Section VI summarizes the SoC FPAA design, as well as the components in the computational analog blocks (CABs),
presents the comparison showing the SoC FPAA as the most in the computational logic blocks (CLBs), and in the devices
sophisticated FPAA device built to date. The presented SoC in the routing architectures that are programmed to nonbinary
FPAA device maximizes both parameter area normalized to levels. The architecture is based on floating-gate (FG) device,
the process node, nearly a factor of 500 improvement in area circuit, and system techniques; we present the particular
efficiency as typical of other analog FPAA devices, as well FG programming approach elsewhere [12].
B. INTRODUCTION 251
however, and so in practice some tradeoff is made between programming time

and switch impedance. Each CAB contains several Operational Transcon-
ductance Amplifiers (OTAs), which are op-amps whose gain is controlled by
a bias current. They are the principal analog computing elements, since they
can be used for operations such as integration, differentiation, and gain am-
plification. Other computing elements may include tunable band-pass filters,
which can be used for Fourier signal processing, and small matrix-vector mul-
tipliers, which can be used to implement linear operators. Current FPAAs
can compute with a resolution of 10 bits (precision of 10−3 ).
B.1.f Non-electronic analog computation
As will be explained later in this chapter, analog computation suggests many

opportunities for future computing technologies. Many physical phenomena
are potential media for analog computation provided they have useful math-
ematical structure (i.e., the mathematical laws describing them are math-
ematical functions useful for general- or special-purpose computation), and
they are sufficiently controllable for practical use.
B.2 Chapter roadmap

The remainder of this chapter will begin by summarizing the fundamentals of
analog computing, starting with the continuous state space and the various
processes by which analog computation can be organized in time. Next it
will discuss analog computation in nature, which provides models and inspi-
ration for many contemporary uses of analog computation, such as neural
networks. Then we consider general-purpose analog computing, both from
a theoretical perspective and in terms of practical general-purpose analog
computers. This leads to a discussion of the theoretical power of analog
computation and in particular to the issue of whether analog computing is
in some sense more powerful than digital computing. We briefly consider the
cognitive aspects of analog computing, and whether it leads to a different
approach to computation than does digital computing. Finally, we conclude
with some observations on the role of analog computation in “post-Moore’s
Law computing.”
C Fundamentals of analog computing

C.1 Continuous state space
As discussed in Sec. B, the fundamental characteristic that distinguishes
analog from digital computation is that the state space is continuous in analog
computation and discrete in digital computation. Therefore it might be
more accurate to call analog and digital computation continuous and discrete
computation, respectively. Furthermore, since the earliest days there have
been hybrid computers that combine continuous and discrete state spaces
and processes. Thus, there are several respects in which the state space may
be continuous.
In the simplest case the state space comprises a finite (generally mod-
est) number of variables, each holding a continuous quantity (e.g., voltage,
current, charge). In a traditional GPAC they correspond to the variables in
the ODEs defining the computational process, each typically having some
independent meaning in the analysis of the problem. Mathematically, the
variables are taken to contain bounded real numbers, although complex-
valued variables are also possible (e.g., in AC electronic analog computers).
In a practical sense, however, their precision is limited by noise, stability,
device tolerance, and other factors (discussed below, Sec. C.4).
In typical analog neural networks the state space is larger in dimension
but more structured than in traditional analog computers. The artificial
neurons are organized into one or more layers, each composed of a (possi-
bly large) number of artificial neurons. Commonly each layer of neurons is
densely connected to the next layer (i.e., each neuron in one layer is connected
to every neuron in the next). In general the layers each have some meaning
in the problem domain, but the individual neurons constituting them do not
(and so, in mathematical descriptions, the neurons are typically numbered
rather than named).
The individual artificial neurons usually perform a simple computation
such as this:
Xn
y = σ(s), where s = b + w i xi ,
i=1
and where y is the activity of the neuron, x1 , . . . , xn are the activities of

the neurons that provide its inputs, b is a bias term, and w1 , . . . , wn are the
weights or strengths of the connections. Often the activation function σ is a
C. FUNDAMENTALS OF ANALOG COMPUTING 253
real-valued sigmoid (“S-shaped”) function, such as the logistic sigmoid,

1
σ(s) = ,
1 + e−s
in which case the neuron activity y is a real number, but some applications
use a discontinuous threshold function, such as the Heaviside function,

+1 , if s ≥ 0
U (s) = ,
0 , if s < 0
in which case the activity is a discrete quantity. The saturated-linear or

piecewise-linear sigmoid is also used occasionally:

 +1 , if s > 1
σ(s) = s , if 0 ≤ s ≤ 1 .

0 , if s < 0
Regardless of whether the activation function is continuous or discrete,

the bias b and connection
P weights w1 , . . . , wn are real numbers, as is the “net
input” s = b + i wi xi to the activation function. Analog computation may
be used to evaluate the linear combination s and the activation function σ(s),
if it is real-valued. If it is discrete, analog computation can approximate
it with a sufficiently sharp sigmoid. The biases and weights are normally
determined by a learning algorithm (e.g., back-propagation), which is also a
good candidate for analog implementation.
In summary, the continuous state space of a neural network includes the
bias values and net inputs of the neurons and the interconnection strengths
between the neurons. It also includes the activity values of the neurons, if
the activation function is a real-valued sigmoid function, as is often the case.
Often large groups (“layers”) of neurons (and the connections between these
groups) have some intuitive meaning in the problem domain, but typically
the individual neuron activities, bias values, and interconnection weights do
not (they are “sub-symbolic”).
If we extrapolate the number of neurons in a layer to the continuum limit,
we get a field, which may be defined as a spatially continuous distribution
of continuous quantity. Treating a group of artificial or biological neurons
as a continuous mass is a reasonable mathematical approximation if their
number is sufficiently large and if their spatial arrangement is significant (as
it generally is in the brain). Fields are especially useful in modeling cortical
maps, in which information is represented by the pattern of activity over a

region of neural cortex.
In field computation the state space in continuous in two ways: it is
continuous in variation but also in space. Therefore, field computation is
especially applicable to solving PDEs and to processing spatially extended
information such as visual images. Some early analog computing devices were
capable of field computation (Truitt & Rogers, 1960, pp. 1-14–17, 2-2–16).
For example, as previously mentioned (Sec. B), large resistor and capacitor
networks could be used for solving PDEs such as diffusion problems. In these
cases a discrete ensemble of resistors and capacitors was used to approximate
a continuous field, while in other cases the computing medium was spatially
continuous. The latter made use of conductive sheets (for two-dimensional
fields) or electrolytic tanks (for two- or three-dimensional fields). When they
were applied to steady-state spatial problems, these analog computers were
called field plotters or potential analyzers.
The ability to fabricate very large arrays of analog computing devices,
combined with the need to exploit massive parallelism in realtime computa-
tion and control applications, creates new opportunities for field computa-
tion (MacLennan, 1987, 1990, 1999). There is also renewed interest in using
physical fields in analog computation. For example, Rubel (1993) defined an
abstract extended analog computer (EAC), which augments Shannon’s (1941)
general purpose analog computer with (unspecified) facilities for field com-
putation, such as PDE solvers (see Secs. E.3–E.4 below). J. W. Mills has
explored the practical application of these ideas in his artificial neural field
networks and VLSI EACs, which use the diffusion of electrons in bulk silicon
or conductive gels and plastics for 2D and 3D field computation (Mills, 1996;
Mills et al., 2006).
C.2 Computational process
We have considered the continuous state space, which is the basis for analog
computing, but there are a variety of ways in which analog computers can
operate on the state. In particular, the state can change continuously in time
or be updated at distinct instants (as in digital computation).
C.2.a Continuous time
Since the laws of physics on which analog computing is based are differential
equations, many analog computations proceed in continuous real time. Also,
as we have seen, an important application of analog computers in the late
19th and early 20th centuries was the integration of ODEs in which time
is the independent variable. A common technique in analog simulation of
physical systems is time scaling, in which the differential equations are altered
systematically so the simulation proceeds either more slowly or more quickly
than the primary system (see Sec. C.4 for more on time scaling). On the
other hand, because analog computations are close to the physical processes
that realize them, analog computing is rapid, which makes it very suitable
for real-time control applications.
In principle, any mathematically describable physical process operating
on time-varying physical quantities can be used for analog computation. In
practice, however, analog computers typically provide familiar operations
that scientists and engineers use in differential equations (Rogers & Con-
nolly, 1960; Truitt & Rogers, 1960). These include basic arithmetic opera-
tions, such as algebraic sum and difference (u(t) = v(t) ± w(t)), constant
multiplication or scaling (u(t) = cv(t)), variable multiplication and division
(u(t) = v(t)w(t), u(t) = v(t)/w(t)), and inversion (u(t) = −v(t)). Transcen-
dental functions may be provided, such as the exponential (u(t) = exp v(t)),
logarithm (u(t) = ln v(t)), trigonometric functions (u(t) = sin v(t), etc.), and
resolvers for converting between polar and rectangular Rcoordinates. Most
t
important, of course, is definite integration (u(t) = v0 + 0 v(τ )dτ ), but dif-
ferentiation may also be provided (u(t) = v̇(t)). Generally, however, direct
differentiation is avoided, since noise tends to have a higher frequency than
the signal, and therefore differentiation amplifies noise; typically problems
are reformulated to avoid direct differentiation (Weyrick, 1969, pp. 26–7).
As previously mentioned, many GPACs include (arbitrary) function genera-
tors, which allow the use of functions defined only by a graph and for which
no mathematical definition might be available; in this way empirically defined
functions can be used (Rogers & Connolly, 1960, pp. 32–42). Thus, given a
graph (x, f (x)), or a sufficient set of samples, (xk , f (xk )), the function gen-
erator approximates u(t) = f (v(t)). Rather less common are generators for
arbitrary functions of two variables, u(t) = f (v(t), w(t)), in which the func-
tion may be defined by a surface, (x, y, f (x, y)), or by sufficient samples from
it.
Although analog computing is primarily continuous, there are situations

in which discontinuous behavior is required. Therefore some analog comput-
ers provide comparators, which produce a discontinuous result depending on
the relative value of two input values. For example,

k , if v ≥ w,
u=
0 , if v < w.
Typically, this would be implemented as a Heaviside (unit step) function

applied to the difference of the inputs, u = kU (v − w). In addition to
allowing the definition of discontinuous functions, comparators provide a
primitive decision making ability, and may be used, for example to terminate
a computation (switching the computer from “operate” to “hold” mode).
Other operations that have proved useful in analog computation are time
delays and noise generators (Howe, 1961, ch. 7). The function of a time delay
is simply to retard the signal by an adjustable delay T > 0: u(t + T ) = v(t).
One common application is to model delays in the primary system (e.g.,
human response time).
Typically a noise generator produces time-invariant Gaussian-distributed
noise with zero mean and a flat power spectrum (over a band compatible with
the analog computing process). The standard deviation can be adjusted by
scaling, the mean can be shifted by addition, and the spectrum altered by
filtering, as required by the application. Historically noise generators were
used to model noise and other random effects in the primary system, to
determine, for example, its sensitivity to effects such as turbulence. However,
noise can make a positive contribution in some analog computing algorithms
(e.g., for symmetry breaking and in simulated annealing, weight perturbation
learning, and stochastic resonance).
As already mentioned, some analog computing devices for the direct so-
lution of PDEs have been developed. In general a PDE solver depends on
an analogous physical process, that is, on a process obeying the same class
of PDEs that it is intended to solve. For example, in Mills’ EAC, diffusion
of electrons in conductive sheets or solids is used to solve diffusion equations
(Mills, 1996; Mills et al., 2006). Historically, PDEs were solved on electronic
GPACs by discretizing all but one of the independent variables, thus replac-
ing the differential equations by difference equations (Rogers & Connolly,
1960, pp. 173–93). That is, computation over a field was approximated by
computation over a finite real array.
Reaction-diffusion computation is an important example of continuous-

time analog computing. The state is represented by a set of time-varying
chemical concentration fields, c1 , . . . , cn . These fields are distributed across
a one-, two-, or three-dimensional space Ω, so that, for x ∈ Ω, ck (x, t) repre-
sents the concentration of chemical k at location x and time t. Computation
proceeds in continuous time according to reaction-diffusion equations, which
have the form:
∂c/∂t = D∇2 c + F(c),
where c = (c1 , . . . , cn )T is the vector of concentrations, D = diag(d1 , . . . , dn )
is a diagonal matrix of positive diffusion rates, and F is nonlinear vector
function that describes how the chemical reactions affect the concentrations.
Some neural net models operate in continuous time and thus are examples
of continuous-time analog computation. For example, Grossberg (Grossberg,
1967, 1973, 1976) defines the activity of a neuron by differential equations
such as this:
n
X n
X
(+) (−)
ẋi = −ai xi + bij wij fj (xj ) − cij wij gj (xj ) + Ii .
j=1 j=1
This describes the continuous change in the activity of neuron i resulting

from passive decay (first term), positive feedback from other neurons (second
term), negative feedback (third term), and input (last term). The fj and
(+) (−)
gj are nonlinear activation functions, and the wij and wij are adaptable
excitatory and inhibitory connection strengths, respectively.
The continuous Hopfield network is another example of continuous-time
analog computation (Hopfield, 1984). The output yi of a neuron is a nonlinear
function of its internal state xi , yi = σ(xi ), where the hyperbolic tangent is
usually used as the activation function, σ(x) = tanh x, because its range is
[−1, 1]. The internal state is defined by a differential equation,
n
X
τi ẋi = −ai xi + bi + wij yj ,
j=1
where τi is a time constant, ai is the decay rate, bi is the bias, and wij is the
connection weight to neuron i from neuron j. In a Hopfield network every
neuron is symmetrically connected to every other (wij = wji ) but not to itself
(wii = 0).
Of course analog VLSI implementations of neural networks also operate

in continuous time (e.g., Mead, 1989; Fakhraie & Smith, 1997)
Concurrent with the resurgence of interest in analog computation have
been innovative reconceptualizations of continuous-time computation. For
example, Brockett (1988) has shown that dynamical systems can perform a
number of problems normally considered to be intrinsically sequential. In
particular, a certain system of ODEs (a nonperiodic finite Toda lattice) can
sort a list of numbers by continuous-time analog computation. The system
is started with the vector x equal to the values to be sorted and a vector
y initialized to small nonzero values; the y vector converges to a sorted
permutation of x.
C.2.b Sequential time

Sequential-time computation refers to computation in which discrete compu-
tational operations take place in succession but at no definite interval (van
Gelder, 1997). Ordinary digital computer programs take place in sequential
time, for the operations occur one after another, but the individual oper-
ations are not required to have any specific duration, so long as they take
finite time.
One of the oldest examples of sequential analog computation is provided
by the compass-and-straightedge constructions of traditional Euclidean ge-
ometry (Sec. B). These computations proceed by a sequence of discrete
operations, but the individual operations involve continuous representations
(e.g., compass settings, straightedge positions) and operate on a continuous
state (the figure under construction). Slide rule calculation might seem to be
an example of sequential analog computation, but if we look at it, we see that
although the operations are performed by an analog device, the intermediate
results are recorded digitally (and so this part of the state space is discrete).
Thus it is a kind of hybrid computation.
The familiar digital computer automates sequential digital computations
that once were performed manually by human “computers.” Sequential ana-
log computation can be similarly automated. That is, just as the control unit
of an ordinary digital computer sequences digital computations, so a digital
control unit can sequence analog computations. In addition to the analog
computation devices (adders, multipliers, etc.), such a computer must pro-
vide variables and registers capable of holding continuous quantities between
the sequential steps of the computation (see also Sec. C.2.c below).
The primitive operations of sequential-time analog computation are typ-

ically similar to those in continuous-time computation (e.g., addition, multi-
plication, transcendental functions), but integration and differentiation with
respect to sequential time do not make sense. However, continuous-time
integration within a single step, and space-domain integration, as in PDE
solvers or field computation devices, are compatible with sequential analog
computation.
In general, any model of digital computation can be converted to a similar
model of sequential analog computation by changing the discrete state space
to a continuum, and making appropriate changes to the rest of the model.
For example, we can make an analog Turing machine by allowing it to write
a bounded real number (rather than a symbol from a finite alphabet) onto a
tape cell. The Turing machine’s finite control can be altered to test for tape
markings in some specified range.
Similarly, in a series of publications Blum, Shub, and Smale developed a
theory of computation over the reals, which is an abstract model of sequential-
time analog computation (Blum et al., 1998, 1988). In this “BSS model”
programs are represented as flowcharts, but they are able to operate on real-
valued variables. Using this model they were able to prove a number of
theorems about the complexity of sequential analog algorithms.
The BSS model, and some other sequential analog computation models,
assume that it is possible to make exact comparisons between real numbers
(analogous to exact comparisons between integers or discrete symbols in dig-
ital computation) and to use the result of the comparison to control the path
of execution. Comparisons of this kind are problematic because they imply
infinite precision in the comparator (which may be defensible in a mathemat-
ical model but is impossible in physical analog devices), and because they
make the execution path a discontinuous function of the state (whereas ana-
log computation is usually continuous). Indeed, it has been argued that this
is not “true” analog computation (Siegelmann, 1999, p. 148).
Many artificial neural network models are examples of sequential-time
analog computation. In a simple feed-forward neural network, an input vector
is processed by the layers in order, as in a pipeline. That is, the output
of layer n becomes the input of layer n + 1. Since the model does not
make any assumptions about the amount of time it takes a vector to be
processed by each layer and to propagate to the next, execution takes place
in sequential time. Most recurrent neural networks, which have feedback, also
operate in sequential time, since the activities of all the neurons are updated
synchronously (that is, signals propagate through the layers, or back to earlier
layers, in lockstep).
Many artificial neural-net learning algorithms are also sequential-time
analog computations. For example, the back-propagation algorithm updates
a network’s weights, moving sequentially backward through the layers.
In summary, the correctness of sequential time computation (analog or
digital) depends on the order of operations, not on their duration, and sim-
ilarly the efficiency of sequential computations is evaluated in terms of the
number of operations, not on their total duration.
C.2.c Discrete time

Discrete-time analog computation has similarities to both continuous-time
and sequential-time analog computation. Like the latter, it proceeds by a
sequence of discrete (analog) computation steps; like the former, these steps
occur at a constant rate in real time (e.g., some “frame rate”). If the real-
time rate is sufficient for the application, then discrete-time computation can
approximate continuous-time computation (including integration and differ-
entiation).
Some electronic GPACs implemented discrete-time analog computation
by a modification of repetitive operation mode, called iterative analog compu-
tation (Ashley, 1963, ch. 9). Recall (Sec. B.1.b) that in repetitive operation
mode a clock rapidly switched the computer between reset and compute
modes, thus repeating the same analog computation, but with different pa-
rameters (set by the operator). However, each repetition was independent of
the others. Iterative operation was different in that analog values computed
by one iteration could be used as initial values in the next. This was accom-
plished by means of an analog memory circuit (based on an op amp) that
sampled an analog value at the end of one compute cycle (effectively during
hold mode) and used it to initialize an integrator during the following reset
cycle. (A modified version of the memory circuit could be used to retain a
value over several iterations.) Iterative computation was used for problems
such as determining, by iterative search or refinement, the initial conditions
that would lead to a desired state at a future time. Since the analog compu-
tations were iterated at a fixed clock rate, iterative operation is an example
of discrete-time analog computation. However, the clock rate is not directly
relevant in some applications (such as the iterative solution of boundary
value problems), in which case iterative operation is better characterized as
sequential analog computation.

The principal contemporary examples of discrete-time analog computing
are in neural network applications to time-series analysis and (discrete-time)
control. In each of these cases the input to the neural net is a sequence
of discrete-time samples, which propagate through the net and generate
discrete-time output signals. Many of these neural nets are recurrent, that
is, values from later layers are fed back into earlier layers, which allows the
net to remember information from one sample to the next.
C.3 Analog computer programs

The concept of a program is central to digital computing, both practically,
for it is the means for programming general-purpose digital computers, and
theoretically, for it defines the limits of what can be computed by a universal
machine, such as a universal Turing machine. Therefore it is important to
discuss means for describing or specifying analog computations.
Traditionally, analog computers were used to solve ODEs (and sometimes
PDEs), and so in one sense a mathematical differential equation is one way
to represent an analog computation. However, since the equations were usu-
ally not suitable for direct solution on an analog computer, the process of
programming involved the translation of the equations into a schematic dia-
gram showing how the analog computing devices (integrators etc.) should be
connected to solve the problem. These diagrams are the closest analogies to
digital computer programs and may be compared to flowcharts, which were
once popular in digital computer programming. It is worth noting, how-
ever, that flowcharts (and ordinary computer programs) represent sequences
among operations, whereas analog computing diagrams represent functional
relationships among variables, and therefore a kind of parallel data flow.
Differential equations and schematic diagrams are suitable for continuous-
time computation, but for sequential analog computation something more
akin to a conventional digital program can be used. Thus, as previously
discussed (Sec. C.2.b), the BSS system uses flowcharts to describe sequen-
tial computations over the reals. Similarly, Moore (1996) defines recursive
functions over the reals by means of a notation similar to a programming
language.
In principle any sort of analog computation might involve constants that
are arbitrary real numbers, which therefore might not be expressible in finite
form (e.g., as a finite string of digits). Although this is of theoretical interest
(see Sec. F.3 below), from a practical standpoint these constants could be
set with about at most four digits of precision (Rogers & Connolly, 1960,
p. 11). Indeed, automatic potentiometer-setting devices were constructed
that read a series of decimal numerals from punched paper tape and used
them to set the potentiometers for the constants (Truitt & Rogers, 1960,
pp. 3-58–60). Nevertheless it is worth observing that analog computers do
allow continuous inputs that need not be expressed in digital notation, for
example, when the parameters of a simulation are continuously varied by
the operator. In principle, therefore, an analog program can incorporate
constants that are represented by a real-valued physical quantity (e.g., an
angle or a distance), which need not be expressed digitally. Further, as we
have seen (Sec. B.1.b), some electronic analog computers could compute a
function by means of an arbitrarily drawn curve, that is, not represented by
an equation or a finite set of digitized points. Therefore, in the context of
analog computing it is natural to expand the concept of a program beyond
discrete symbols to include continuous representations (scalar magnitudes,
vectors, curves, shapes, surfaces, etc.).
Typically such continuous representations would be used as adjuncts to
conventional discrete representations of the analog computational process,
such as equations or diagrams. However, in some cases the most natural static
representation of the process is itself continuous, in which case it is more like
a “guiding image” than a textual prescription (MacLennan, 1995). A simple
example is a potential surface, which defines a continuum of trajectories from
initial states (possible inputs) to fixed-point attractors (the results of the
computations). Such a “program” may define a deterministic computation
(e.g., if the computation proceeds by gradient descent), or it may constrain
a nondeterministic computation (e.g., if the computation may proceed by
any potential-decreasing trajectory). Thus analog computation suggests a
broadened notion of programs and programming.
C.4 Characteristics of analog computation

C.4.a Precision
Analog computation is evaluated in terms of both accuracy and precision,
but the two must be distinguished carefully (Ashley 1963, pp. 25–8, Weyrick
1969, pp. 12–13, Small 2001, pp. 257–61). Accuracy refers primarily to the
relationship between a simulation and the primary system it is simulating
or, more generally, to the relationship between the results of a computation

and the mathematically correct result. Accuracy is a result of many factors,
including the mathematical model chosen, the way it is set up on a computer,
and the precision of the analog computing devices. Precision, therefore, is a
narrower notion, which refers to the quality of a representation or computing
device. In analog computing, precision depends on resolution (fineness of op-
eration) and stability (absence of drift), and may be measured as a fraction
of the represented value. Thus a precision of 0.01% means that the represen-
tation will stay within 0.01% of the represented value for a reasonable period
of time. For purposes of comparing analog devices, the precision is usually
expressed as a fraction of full-scale variation (i.e., the difference between the
maximum and minimum representable values).
It is apparent that the precision of analog computing devices depends
on many factors. One is the choice of physical process and the way it is
utilized in the device. For example a linear mathematical operation can be
realized by using a linear region of a nonlinear physical process, but the
realization will be approximate and have some inherent imprecision. Also,
associated, unavoidable physical effects (e.g., loading, and leakage and other
losses) may prevent precise implementation of an intended mathematical
function. Further, there are fundamental physical limitations to resolution
(e.g., quantum effects, diffraction). Noise is inevitable, both intrinsic (e.g.,
thermal noise) and extrinsic (e.g., ambient radiation). Changes in ambient
physical conditions, such as temperature, can affect the physical processes
and decrease precision. At slower time scales, materials and components
age and their physical characteristics change. In addition, there are always
technical and economic limits to the control of components, materials, and
processes in analog device fabrication.
The precision of analog and digital computing devices depend on very
different factors. The precision of a (binary) digital device depends on the
number of bits, which influences the amount of hardware, but not its quality.
For example, a 64-bit adder is about twice the size of a 32-bit adder, but can
made out of the same components. At worst, the size of a digital device might
increase with the square of the number of bits of precision. This is because
binary digital devices only need to represent two states, and therefore they
can operate in saturation. The fabrication standards sufficient for the first bit
of precision are also sufficient for the 64th bit. Analog devices, in contrast,
need to be able to represent a continuum of states precisely. Therefore, the
fabrication of high-precision analog devices is much more expensive than low-
precision devices, since the quality of components, materials, and processes

must be much more carefully controlled. Doubling the precision of an analog
device may be expensive, whereas the cost of each additional bit of digital
precision is incremental; that is, the cost is proportional to the logarithm of
the precision expressed as a fraction of full range.
The forgoing considerations might seem to be a convincing argument for
the superiority of digital to analog technology, and indeed they were an im-
portant factor in the competition between analog and digital computers in
the middle of the twentieth century (Small, 2001, pp. 257–61). However, as
was argued at that time, many computer applications do not require high pre-
cision. Indeed, in many engineering applications, the input data are known
to only a few digits, and the equations may be approximate or derived from
experiments. In these cases the very high precision of digital computation
is unnecessary and may in fact be misleading (e.g., if one displays all 14
digits of a result that is accurate to only three). Furthermore, many appli-
cations in image processing and control do not require high precision. More
recently, research in artificial neural networks (ANNs) has shown that low-
precision analog computation is sufficient for almost all ANN applications.
Indeed, neural information processing in the brain seems to operate with very
low precision — perhaps less than 10% (McClelland et al., 1986, p. 378) —
for which it compensates with massive parallelism. For example, by coarse
coding a population of low-precision devices can represent information with
relatively high precision (Rumelhart et al. 1986, pp. 91–6, Sanger 1996).
C.4.b Scaling
An important aspect of analog computing is scaling, which is used to adjust a
problem to an analog computer. First is time scaling, which adjusts a problem
to the characteristic time scale at which a computer operates, which is a
consequence of its design and the physical processes by which it is realized
(Peterson 1967, pp. 37–44, Rogers & Connolly 1960, pp. 262–3, Weyrick
1969, pp. 241–3). For example, we might want a simulation to proceed on
a very different time scale from the primary system. Thus a weather or
economic simulation should proceed faster than real time in order to get
useful predictions. Conversely, we might want to slow down a simulation of
protein folding so that we can observe the stages in the process. Also, for
accurate results it is necessary to avoid exceeding the maximum response rate
of the analog devices, which might dictate a slower simulation speed. On the
other hand, too slow a computation might be inaccurate as a consequence of

instability (e.g., drift and leakage in the integrators).
Time scaling affects only time-dependant operations such as integration.
For example, suppose t, time in the primary system or “problem time,” is
related Rto τ , time in the computer, by τ = βt. Therefore, an integration
t
u(t) R= 0 v(t0 )dt0 in the primary system is replaced by the integration u(τ ) =
τ
β −1 0 v(τ 0 )dτ 0 on the computer. Thus time scaling may be accomplished
simply by decreasing the input gain to the integrator by a factor of β.
Fundamental to analog computation is the representation of a continuous
quantity in the primary system by a continuous quantity in the computer. For
example, a displacement x in meters might be represented by a potential V in
volts. The two are related by an amplitude or magnitude scale factor, V = αx,
(with units volts/meter), chosen to meet two criteria (Ashley 1963, pp. 103–6,
Peterson 1967, ch. 4, Rogers & Connolly 1960, pp. 127–8, Weyrick 1969, pp.
233–40). On the one hand, α must be sufficiently small so that the range of
the problem variable is accommodated within the range of values supported
by the computing device. Exceeding the device’s intended operating range
may lead to inaccurate results (e.g., forcing a linear device into nonlinear
behavior). On the other hand, the scale factor should not be too small, or
relevant variation in the problem variable will be less than the resolution of
the device, also leading to inaccuracy. (Recall that precision is specified as a
fraction of full-range variation.)
In addition to the explicit variables of the primary system, there are im-
plicit variables, such as the time derivatives of the explicit variables, and scale
factors must be chosen for them too. For example, in addition to displace-
ment x, a problem might include velocity ẋ and acceleration ẍ. Therefore,
scale factors α, α0 , and α00 must be chosen so that αx, α0 ẋ, and α00 ẍ have an
appropriate range of variation (neither too large nor too small).
Once a scale factor has been chosen, the primary system equations are
adjusted to obtain the analog computing equations. For Rexample, if we have
t
scaled u = αx and v = α0 ẋ, then the integration x(t) = 0 ẋ(t0 )dt0 would be
computed by scaled equation:
Z t
α
u(t) = 0 v(t0 )dt0 .
α 0
This is accomplished by simply setting the input gain of the integrator to

α/α0 .
In practice, time scaling and magnitude scaling are not independent

(Rogers & Connolly, 1960, p. 262). For example, if the derivatives of a
variable can be large, then the variable can change rapidly, and so it may
be necessary to slow down the computation to avoid exceeding the high-
frequency response of the computer. Conversely, small derivatives might
require the computation to be run faster to avoid integrator leakage etc. Ap-
propriate scale factors are determined by considering both the physics and
the mathematics of the problem (Peterson, 1967, pp. 40–4). That is, first,
the physics of the primary system may limit the ranges of the variables and
their derivatives. Second, analysis of the mathematical equations describing
the system can give additional information on the ranges of the variables. For
example, in some cases the natural frequency of a system can be estimated
from the coefficients of the differential equations; the maximum of the nth
derivative is then estimated as the n power of this frequency (Peterson 1967,
p. 42, Weyrick 1969, pp. 238–40). In any case, it is not necessary to have
accurate values for the ranges; rough estimates giving orders of magnitude
are adequate.
It is tempting to think of magnitude scaling as a problem unique to ana-
log computing, but before the invention of floating-point numbers it was also
necessary in digital computer programming. In any case it is an essential as-
pect of analog computing, in which physical processes are more directly used
for computation than they are in digital computing. Although the necessity
of scaling has been a source of criticism, advocates for analog computing
have argued that it is a blessing in disguise, because it leads to improved
understanding of the primary system, which was often the goal of the com-
putation in the first place (Bissell 2004, Small 2001, ch. 8). Practitioners of
analog computing are more likely to have an intuitive understanding of both
the primary system and its mathematical description (see Sec. G).
D Analog Computation in Nature

Computational processes—that is to say, information processing and control—
occur in many living systems, most obviously in nervous systems, but also
in the self-organized behavior of groups of organisms. In most cases natural
computation is analog, either because it makes use of continuous natural pro-
cesses, or because it makes use of discrete but stochastic processes. Several
examples will be considered briefly.
D. ANALOG COMPUTATION IN NATURE 267
D.1 Neural computation

In the past neurons were thought of binary computing devices, something like
digital logic gates. This was a consequence of the “all or nothing” response of
a neuron, which refers to the fact that it does or does not generate an action
potential (voltage spike) depending, respectively, on whether its total input
exceeds a threshold or not (more accurately, it generates an action potential
if the membrane depolarization at the axon hillock exceeds the threshold and
the neuron is not in its refractory period). Certainly some neurons (e.g., so-
called “command neurons”) do act something like logic gates. However, most
neurons are analyzed better as analog devices, because the rate of impulse
generation represents significant information. In particular, an amplitude
code, the membrane potential near the axon hillock (which is a summation
of the electrical influences on the neuron), is translated into a rate code
for more reliable long-distance transmission along the axons. Nevertheless,
the code is low precision (about one digit), since information theory shows
that it takes at least N milliseconds (and probably more like 5N msec.)
to discriminate N values (MacLennan, 1991). The rate code is translated
back to an amplitude code by the synapses, since successive impulses release
neurotransmitter from the axon terminal, which diffuses across the synaptic
cleft to receptors. Thus a synapse acts as a leaky integrator to time-average
the impulses.
As previously discussed (Sec. C.1), many artificial neural net models have
real-valued neural activities, which correspond to rate-encoded axonal signals
of biological neurons. On the other hand, these models typically treat the
input connections as simple real-valued weights, which ignores the analog
signal processing that takes place in the dendritic trees of biological neurons.
The dendritic trees of many neurons are complex structures, which often
have tens of thousands of synaptic inputs. The binding of neurotransmitters
to receptors causes minute voltage fluctuations, which propagate along the
membrane, and ultimately cause voltage fluctuations at the axon hillock,
which influence the impulse rate. Since the dendrites have both resistance
and capacitance, to a first approximation the signal propagation is described
by the “cable equations,” which describe passive signal propagation in cables
of specified diameter, capacitance, and resistance (Anderson, 1995, ch. 1).
Therefore, to a first approximation, a neuron’s dendritic net operates as an
adaptive linear analog filter with thousands of inputs, and so it is capable
of quite complex signal processing. More accurately, however, it must be
treated as a nonlinear analog filter, since voltage-gated ion channels introduce

nonlinear effects. The extent of analog signal processing in dendritic trees is
still poorly understood.
In most cases, then, neural information processing is treated best as
low-precision analog computation. Although individual neurons have quite
broadly tuned responses, accuracy in perception and sensorimotor control is
achieved through coarse coding, as already discussed (Sec. C.4). Further,
one widely used neural representation is the cortical map, in which neurons
are systematically arranged in accord with one or more dimensions of their
stimulus space, so that stimuli are represented by patterns of activity over
the map. (Examples are tonotopic maps, in which pitch is mapped to cortical
location, and retinotopic maps, in which cortical location represents retinal
location.) Since neural density in the cortex is at least 146 000 neurons per
square millimeter (Changeux, 1985, p. 51), even relatively small cortical maps
can be treated as fields and information processing in them as analog field
computation. Overall, the brain demonstrates what can be accomplished
by massively parallel analog computation, even if the individual devices are
comparatively slow and of low precision.
D.2 Adaptive self-organization in social insects

Another example of analog computation in nature is provided by the self-
organizing behavior of social insects, microorganisms, and other populations
(Camazine et al., 2001). Often such organisms respond to concentrations, or
gradients in the concentrations, of chemicals produced by other members of
the population. These chemicals may be deposited and diffuse through the
environment. In other cases, insects and other organisms communicate by
contact, but may maintain estimates of the relative proportions of different
kinds of contacts. Because the quantities are effectively continuous, all these
are examples of analog control and computation.
Self-organizing populations provide many informative examples of the use
of natural processes for analog information processing and control. For ex-
ample, diffusion of pheromones is a common means of self-organizzation in
insect colonies, facilitating the creation of paths to resources, the construction
of nests, and many other functions (Camazine et al., 2001). Real diffusion
(as opposed to sequential simulations of it) executes, in effect, a massively
parallel search of paths from the chemical’s source to its recipients and al-
lows the identification of near-optimal paths. Furthermore, if the chemical
D. ANALOG COMPUTATION IN NATURE 269
degrades, as is generally the case, then the system will be adaptive, in effect
continually searching out the shortest paths, so long as source continues to
function (Camazine et al., 2001). Simulated diffusion has been applied to
robot path planning (Khatib, 1986; Rimon & Koditschek, 1989).
D.3 Genetic circuits

Another example of natural analog computing is provided by the genetic reg-
ulatory networks that control the behavior of cells, in multicellular organisms
as well as single-celled ones (Davidson, 2006). These networks are defined by
the mutually interdependent regulatory genes, promoters, and repressors that
control the internal and external behavior of a cell. The interdependencies
are mediated by proteins, the synthesis of which is governed by genes, and
which in turn regulate the synthesis of other gene products (or themselves).
Since it is the quantities of these substances that is relevant, many of the
regulatory motifs can be described in computational terms as adders, sub-
tracters, integrators, etc. Thus the genetic regulatory network implements
an analog control system for the cell (Reiner, 1968).
It might be argued that the number of intracellular molecules of a par-
ticular protein is a (relatively small) discrete number, and therefore that it
is inaccurate to treat it as a continuous quantity. However, the molecular
processes in the cell are stochastic, and so the relevant quantity is the prob-
ability that a regulatory protein will bind to a regulatory site. Further, the
processes take place in continuous real time, and so the rates are generally the
significant quantities. Finally, although in some cases gene activity is either
on or off (more accurately: very low), in other cases it varies continuously
between these extremes (Hartl, 1994, pp. 388–90).
Embryological development combines the analog control of individual cells
with the sort of self-organization of populations seen in social insects and
other colonial organisms. Locomotion of the cells and the expression of spe-
cific genes is controlled by chemical signals, among other mechanisms (David-
son, 2006; Davies, 2005). Thus PDEs have proved useful in explaining some
aspects of development; for example reaction-diffusion equations have been
used to describe the formation of hair-coat patterns and related phenomena
(Camazine et al., 2001; Maini & Othmer, 2001; Murray, 1977). Therefore
the developmental process is governed by naturally occurring analog compu-
tation.
D.4 Is everything a computer?
It might seem that any continuous physical process could be viewed as analog
computation, which would make the term almost meaningless. As the ques-
tion has been put, is it meaningful (or useful) to say that the solar system
is computing Kepler’s laws? In fact, it is possible and worthwhile to make a
distinction between computation and other physical processes that happen
to be described by mathematical laws (MacLennan, 1994a,c, 2001, 2004).
If we recall the original meaning of analog computation (Sec. A), we see

that the computational system is used to solve some mathematical problem
with respect to a primary system. What makes this possible is that the com-
putational system and the primary system have the same, or systematically
related, abstract (mathematical) structures. Thus the computational system
can inform us about the primary system, or be used to control it, etc. Al-
though from a practical standpoint some analogs are better than others, in
principle any physical system can be used that obeys the same equations as
the primary system.
Based on these considerations we may define computation as a physical

process the purpose of which is the abstract manipulation of abstract objects
(i.e., information processing); this definition applies to analog, digital, and
hybrid computation (MacLennan, 1994a,c, 2001, 2004). Therefore, to deter-
mine if a natural system is computational we need to look to its purpose or
function within the context of the living system of which it is a part. One
test of whether its function is the abstract manipulation of abstract objects is
to ask whether it could still fulfill its function if realized by different physical
processes, a property called multiple realizability. (Similarly, in artificial sys-
tems, a simulation of the economy might be realized equally accurately by a
hydraulic analog computer or an electronic analog computer (Bissell, 2004).)
By this standard, the majority of the nervous system is purely computational;
in principle it could be replaced by electronic devices obeying the same dif-
ferential equations. In the other cases we have considered (self-organization
of living populations, genetic circuits) there are instances of both pure com-
putation and computation mixed with other functions (for example, where
the specific substances used have other—e.g. metabolic—roles in the living
system).
E. GENERAL-PURPOSE ANALOG COMPUTATION 271
E General-purpose analog computation
E.1 The importance of general-purpose computers

Although special-purpose analog and digital computers have been developed,
and continue to be developed, for many purposes, the importance of general-
purpose computers, which can be adapted easily for a wide variety of pur-
poses, has been recognized since at least the nineteenth century. Babbage’s
plans for a general-purpose digital computer, his analytical engine (1835),
are well known, but a general-purpose differential analyzer was advocated
by Kelvin (Thomson, 1876). Practical general-purpose analog and digital
computers were first developed at about the same time: from the early 1930s
through the war years. General-purpose computers of both kinds permit the
prototyping of special-purpose computers and, more importantly, permit the
flexible reuse of computer hardware for different or evolving purposes.
The concept of a general-purpose computer is useful also for determin-
ing the limits of a computing paradigm. If one can design—theoretically
or practically—a universal computer, that is, a general-purpose computer
capable of simulating any computer in a relevant class, then anything un-
computable by the universal computer will also be uncomputable by any
computer in that class. This is, of course, the approach used to show that
certain functions are uncomputable by any Turing machine because they
are uncomputable by a universal Turing machine. For the same reason, the
concept of general-purpose analog computers, and in particular of universal
analog computers are theoretically important for establishing limits to analog
computation.
E.2 General-purpose electronic analog computers

Before taking up these theoretical issues, it is worth recalling that a typ-
ical electronic GPAC would include linear elements, such as adders, sub-
tracters, constant multipliers, integrators, and differentiators; nonlinear ele-
ments, such as variable multipliers and function generators; other computa-
tional elements, such as comparators, noise generators, and delay elements
(Sec. B.1.b). These are, of course, in addition to input/output devices, which
would not affect its computational abilities.
E.3 Shannon’s analysis

Claude Shannon did an important analysis of the computational capabil-
ities of the differential analyzer, which applies to many GPACs (Shannon,
1941, 1993). He considered an abstract differential analyzer equipped with an
unlimited number of integrators, adders, constant multipliers, and function
generators (for functions with only a finite number of finite discontinuities),
with at most one source of drive (which limits possible interconnections be-
tween units). This was based on prior work that had shown that almost
all the generally used elementary functions could be generated with addition
and integration. We will summarize informally a few of Shannon’s results;
for details, please consult the original paper.
First Shannon offers proofs that, by setting up the correct ODEs, a GPAC
with the mentioned facilities can generate any function if and only if is not
hypertranscendental (Theorem II); thus the GPAC can generate any function
that is algebraic transcendental (a very large class), but not, for example,
Euler’s gamma function and Riemann’s zeta function. He also shows that
the GPAC can generate functions derived from generable functions, such as
the integrals, derivatives, inverses, and compositions of generable functions
(Thms. III, IV). These results can be generalized to functions of any number
of variables, and to their compositions, partial derivatives, and inverses with
respect to any one variable (Thms. VI, VII, IX, X).
Next Shannon shows that a function of any number of variables that
is continuous over a closed region of space can be approximated arbitrarily
closely over that region with a finite number of adders and integrators (Thms.
V, VIII).
Shannon then turns from the generation of functions to the solution of
ODEs and shows that the GPAC can solve any system of ODEs defined in
terms of non-hypertranscendental functions (Thm. XI).
Finally, Shannon addresses a question that might seem of limited interest,
but turns out to be relevant to the computational power of analog computers
(see Sec. F below). To understand it we must recall that he was investigating
the differential analyzer—a mechanical analog computer—but similar issues
arise in other analog computing technologies. The question is whether it is
possible to perform an arbitrary constant multiplication, u = kv, by means of
gear ratios. He show that if we have just two gear ratios a and b (a, b 6= 0, 1),
such that b is not a rational power of a, then by combinations of these gears
we can approximate k arbitrarily closely (Thm. XII). That is, to approximate
E. GENERAL-PURPOSE ANALOG COMPUTATION 273
multiplication by arbitrary real numbers, it is sufficient to be able to multiply

by a, b, and their inverses, provided a and b are not related by a rational
power.
Shannon mentions an R valternative method of constant multiplication, which
uses integration, kv = 0 kdv, but this requires setting the integrand to the
constant function k. Therefore, multiplying by an arbitrary real number re-
quires the ability to input an arbitrary real as the integrand. The issue of
real-valued inputs and outputs to analog computers is relevant both to their
theoretical power and to practical matters of their application (see Sec. F.3).
Shannon’s proofs, which were incomplete, were eventually refined by
Pour-El (1974a) and finally corrected by Lipshitz & Rubel (1987). Rubel
(1988) proved that Shannon’s GPAC cannot solve the Dirichlet problem for
Laplace’s equation on the disk; indeed, it is limited to initial-value problems
for algebraic ODEs. Specifically, the Shannon–Pour-El Thesis is that the
outputs of the GPAC are exactly the solutions of the algebraic differential
equations, that is, equations of the form
P [x, y(x), y 0 (x), y 00 (x), . . . , y (n) (x)] = 0,
where P is a polynomial that is not identically vanishing in any of its vari-
ables (these are the differentially algebraic functions) (Rubel, 1985). (For
details please consult the cited papers.) The limitations of Shannon’s GPAC
motivated Rubel’s definition of the Extended Analog Computer.
E.4 Rubel’s Extended Analog Computer

The combination of Rubel’s (1985) conviction that the brain is an analog
computer together with the limitations of Shannon’s GPAC led him to pro-
pose the Extended Analog Computer (EAC) (Rubel, 1993).
Like Shannon’s GPAC (and the Turing machine), the EAC is a concep-
tual computer intended to facilitate theoretical investigation of the limits of
a class of computers. The EAC extends the GPAC in a number of respects.
For example, whereas the GPAC solves equations defined over a single vari-
able (time), the EAC can generate functions over any finite number of real
variables. Further, whereas the GPAC is restricted to initial-value problems
for ODEs, the EAC solves both initial- and boundary-value problems for a
variety of PDEs.
The EAC is structured into a series of levels, each more powerful than the
ones below it, from which it accepts inputs. The inputs to the lowest level
are a finite number of real variables (“settings”). At this level it operates on

real polynomials, from which it is able to generate the differentially algebraic
functions. The computation on each level is accomplished by conceptual
analog devices, which include constant real-number generators, adders, mul-
tipliers, differentiators, “substituters” (for function composition), devices for
analytic continuation, and inverters, which solve systems of equations de-
fined over functions generated by the lower levels. Most characteristic of the
EAC is the “boundary-value-problem box,” which solves systems of PDEs
and ODEs subject to boundary conditions and other constraints. The PDEs
are defined in terms of functions generated by the lower levels. Such PDE
solvers may seem implausible, and so it is important to recall field-computing
devices for this purpose were implemented in some practical analog comput-
ers (see Sec. B.1) and more recently in Mills’ EAC (Mills et al., 2006). As
Rubel observed, PDE solvers could be implemented by physical processes
that obey the same PDEs (heat equation, wave equation, etc.). (See also
Sec. H.1 below.)
Finally, the EAC is required to be “extremely well-posed,” which means
that each level is relatively insensitive to perturbations in its inputs; thus
“all the outputs depend in a strongly deterministic and stable way on the
initial settings of the machine” (Rubel, 1993).
Rubel (1993) proves that the EAC can compute everything that the
GPAC can compute, but also such functions as the gamma and zeta, and
that it can solve the Dirichlet problem for Laplace’s equation on the disk, all
of which are beyond the GPAC’s capabilities. Further, whereas the GPAC
can compute differentially algebraic functions of time, the EAC can compute
differentially algebraic functions of any finite number of real variables. In
fact, Rubel did not find any real-analytic (C ∞ ) function that is not com-
putable on the EAC, but he observes that if the EAC can indeed generate
every real-analytic function, it would be too broad to be useful as a model
of analog computation.
F Analog computation and the Turing limit

F.1 Introduction
The Church-Turing Thesis asserts that anything that is effectively com-
putable is computable by a Turing machine, but the Turing machine (and
F. ANALOG COMPUTATION AND THE TURING LIMIT 275
equivalent models, such as the lambda calculus) are models of discrete com-
putation, and so it is natural to wonder how analog computing compares in
power, and in particular whether it can compute beyond the “Turing limit.”
Superficial answers are easy to obtain, but the issue is subtle because it de-
pends upon choices among definitions, none of which is obviously correct,
it involves the foundations of mathematics and its philosophy, and it raises
epistemological issues about the role of models in scientific theories. This is
an active research area, but many of the results are apparently inconsistent
due to the differing assumptions on which they are based. Therefore this
section will be limited to a mention of a few of the interesting results, but
without attempting a comprehensive, systematic, or detailed survey; Siegel-
mann (1999) can serve as an introduction to the literature.
F.2 A sampling of theoretical results

F.2.a Continuous-time models
Orponen’s (1997) survey of continuous-time computation theory is a good
introduction to the literature as of that time; here we give a sample of these
and more recent results.
There are several results showing that—under various assumptions—
analog computers have at least the power of Turing machines (TMs). For
example, Branicky (1994) showed that a TM could be simulated by ODEs,
but he used non-differentiable functions; Bournez et al. (2006) provide an
alternative construction using only analytic functions. They also prove that
the GPAC computability coincides with (Turing-)computable analysis, which
is surprising, since the gamma function is Turing-computable but, as we have
seen, the GPAC cannot generate it. The paradox is resolved by a distinction
between generating a function and computing it, with the latter, broader no-
tion permitting convergent computation of the function (that is, as t → ∞).
However, the computational power of general ODEs has not been determined
in general (Siegelmann, 1999, p. 149). M. B. Pour-El and I. Richards exhibit
a Turing-computable ODE that does not have a Turing-computable solution
(Pour-El & Richards, 1979, 1982). Stannett (1990) also defined a continuous-
time analog computer that could solve the halting problem.
Moore (1996) defines a class of continuous-time recursive functions over
the reals, which includes a zero-finding operator µ. Functions can be classified
into a hierarchy depending on the number of uses of µ, with the lowest level
(no µs) corresponding approximately to Shannon’s GPAC. Higher levels can

compute non-Turing-computable functions, such as the decision procedure
for the halting problem, but he questions whether this result is relevant in
the physical world, which is constrained by “noise, quantum effects, finite
accuracy, and limited resources.” Bournez & Cosnard (1996) have extended
these results and shown that many dynamical systems have super-Turing
power.
Omohundro (1984) showed that a system of ten coupled nonlinear PDEs
could simulate an arbitrary cellular automaton, which implies that PDEs
have at least Turing power. Further, D. Wolpert and B. J. MacLennan
(Wolpert, 1991; Wolpert & MacLennan, 1993) showed that any TM can be
simulated by a field computer with linear dynamics, but the construction
uses Dirac delta functions. Pour-El and Richards exhibit a wave equation
in three-dimensional space with Turing-computable initial conditions, but
for which the unique solution is Turing-uncomputable (Pour-El & Richards,
1981, 1982).
F.2.b Sequential-time models
We will mention a few of the results that have been obtained concerning the
power of sequential-time analog computation.
Although the BSS model has been investigated extensively, its power
has not been completely determined (Blum et al., 1998, 1988). It is known
to depend on whether just rational numbers or arbitrary real numbers are
allowed in its programs (Siegelmann, 1999, p. 148).
A coupled map lattice (CML) is a cellular automaton with real-valued
states; it is a sequential-time analog computer, which can be considered a
discrete-space approximation to a simple sequential-time field computer. Or-
ponen & Matamala (1996) showed that a finite CML can simulate a universal
Turing machine. However, since a CML can simulate a BSS program or a
recurrent neural network (see Sec. F.2.c below), it actually has super-Turing
power (Siegelmann, 1999, p. 149).
Recurrent neural networks are some of the most important examples of
sequential analog computers, and so the following section is devoted to them.
F.2.c Recurrent neural networks
With the renewed interest in neural networks in the mid-1980s, many in-
vestigators wondered if recurrent neural nets have super-Turing power. M.
Garzon and S. Franklin showed that a sequential-time net with a countable
infinity of neurons could exceed Turing power (Franklin & Garzon, 1990; Gar-
zon & Franklin, 1989, 1990). Indeed, Siegelmann & Sontag (1994b) showed
that finite neural nets with real-valued weights have super-Turing power, but
Maass & Sontag (1999b) showed that recurrent nets with Gaussian or sim-
ilar noise had sub-Turing power, illustrating again the dependence on these
results on assumptions about what is a reasonable mathematical model of
analog computing.
For recent results on recurrent neural networks, we will restrict our at-
tention of the work of Siegelmann (1999), who addresses the computational
power of these network in terms of the classes of languages they can rec-
ognize. Without loss of generality the languages are restricted to sets of
binary strings. A string to be tested is fed to the network one bit at a time,
along with an input that indicates when the end of the input string has been
reached. The network is said to decide whether the string is in the language if
it correctly indicates whether it is in the set or not, after some finite number
of sequential steps since input began.
Siegelmann shows that, if exponential time is allowed for recognition,
finite recurrent neural networks with real-valued weights (and saturated-
linear activation functions) can compute all languages, and thus they are
more powerful than Turing machines. Similarly, stochastic networks with
rational weights also have super-Turing power, although less power than the
deterministic nets with real weights. (Specifically, they compute P/POLY
and BPP/log∗ respectively; see Siegelmann 1999, chs. 4, 9 for details.) She
further argues that these neural networks serve as a “standard model” of
(sequential) analog computation (comparable to Turing machines in Church-
Turing computation), and therefore that the limits and capabilities of these
nets apply to sequential analog computation generally.
Siegelmann (1999, p 156) observes that the super-Turing power of recur-
rent neural networks is a consequence of their use of non-rational real-valued
weights. In effect, a real number can contain an infinite number of bits of
information. This raises the question of how the non-rational weights of a net-
work can ever be set, since it is not possible to define a physical quantity with
infinite precision. However, although non-rational weights may not be able
to be set from outside the network, they can be computed within the network
by learning algorithms, which are analog computations. Thus, Siegelmann
suggests, the fundamental distinction may be between static computational
models, such as the Turing machine and its equivalents, and dynamically
evolving computational models, which can tune continuously variable param-
eters and thereby achieve super-Turing power.
F.2.d Dissipative models

Beyond the issue of the power of analog computing relative to the Tur-
ing limit, there are also questions of its relative efficiency. For example,
could analog computing solve NP-hard problems in polynomial or even lin-
ear time? In traditional computational complexity theory, efficiency issues
are addressed in terms of the asymptotic number of computation steps to
compute a function as the size of the function’s input increases. One way to
address corresponding issues in an analog context is by treating an analog
computation as a dissipative system, which in this context means a system
that decreases some quantity (analogous to energy) so that the system state
converges to an point attractor. From this perspective, the initial state of
the system incorporates the input to the computation, and the attractor
represents its output. Therefore, H. T. Sieglemann, S. Fishman, and A.
Ben-Hur have developed a complexity theory for dissipative systems, in both
sequential and continuous time, which addresses the rate of convergence in
terms of the underlying rates of the system (Ben-Hur et al., 2002; Siegelmann
et al., 1999). The relation between dissipative complexity classes (e.g., Pd ,
NPd ) and corresponding classical complexity classes (P, NP) remains unclear
(Siegelmann, 1999, p. 151).
F.3 Real-valued inputs, output, and constants

A common argument, with relevance to the theoretical power of analog com-
putation, is that an input to an analog computer must be determined by
setting a dial to a number or by typing a number into digital-to-analog con-
version device, and therefore that the input will be a rational number. The
same argument applies to any internal constants in the analog computation.
Similarly, it is argued, any output from an analog computer must be mea-
sured, and the accuracy of measurement is limited, so that the result will
be a rational number. Therefore, it is claimed, real numbers are irrelevant
to analog computing, since any practical analog computer computes a func-

tion from the rationals to the rationals, and can therefore be simulated by a
Turing machine.2
There are a number of interrelated issues here, which may be considered
briefly. First, the argument is couched in terms of the input or output of
digital representations, and the numbers so represented are necessarily ratio-
nal (more generally, computable). This seems natural enough when we think
of an analog computer as a calculating device, and in fact many historical
analog computers were used in this way and had digital inputs and outputs
(since this is our most reliable way of recording and reproducing quantities).
However, in many analog control systems, the inputs and outputs are con-
tinuous physical quantities that vary continuously in time (also a continuous
physical quantity); that is, according to current physical theory, these quan-
tities are real numbers, which vary according to differential equations. It is
worth recalling that physical quantities are neither rational nor irrational;
they can be so classified only in comparison with each other or with respect
to a unit, that is, only if they are measured and digitally represented. Fur-
thermore, physical quantities are neither computable nor uncomputable (in
a Church-Turing sense); these terms apply only to discrete representations
of these quantities (i.e., to numerals or other digital representations).
Therefore, in accord with ordinary mathematical descriptions of physical
processes, analog computations can can be treated as having arbitrary real
numbers (in some range) as inputs, outputs, or internal states; like other
continuous processes, continuous-time analog computations pass through all
the reals in some range, including non-Turing-computable reals. Paradox-
ically, however, these same physical processes can be simulated on digital
computers.
F.4 The issue of simulation by Turing machines and

digital computers
Theoretical results about the computational power, relative to Turing ma-
chines, of neural networks and other analog models of computation raise
difficult issues, some of which are epistemological rather than strictly tech-
nical. On the one hand, we have a series of theoretical results proving the
super-Turing power of analog computation models of various kinds. On the
2
See related arguments by Martin Davis (2004, 2006).
other hand, we have the obvious fact that neural nets are routinely simulated
on ordinary digital computers, which have at most the power of Turing ma-
chines. Furthermore, it is reasonable to suppose that any physical process
that might be used to realize analog computation—and certainly the known
processes—could be simulated on a digital computer, as is done routinely in
computational science. This would seem to be incontrovertible proof that
analog computation is no more powerful than Turing machines. The crux
of the paradox lies, of course, in the non-Turing-computable reals. These
numbers are a familiar, accepted, and necessary part of standard mathe-
matics, in which physical theory is formulated, but from the standpoint of
Church-Turing (CT) computation they do not exist. This suggests that the
the paradox is not a contradiction, but reflects a divergence between the
goals and assumptions of the two models of computation.
F.5 The problem of models of computation

These issues may be put in context by recalling that the Church-Turing (CT)
model of computation is in fact a model, and therefore that it has the limita-
tions of all models. A model is a cognitive tool that improves our ability to
understand some class of phenomena by preserving relevant characteristics
of the phenomena while altering other, irrelevant (or less relevant) charac-
teristics. For example, a scale model alters the size (taken to be irrelevant)
while preserving shape and other characteristics. Often a model achieves
its purposes by making simplifying or idealizing assumptions, which facili-
tate analysis or simulation of the system. For example, we may use a linear
mathematical model of a physical process that is only approximately linear.
For a model to be effective it must preserve characteristics and make sim-
plifying assumptions that are appropriate to the domain of questions it is
intended to answer, its frame of relevance (MacLennan, 2004). If a model
is applied to problems outside of its frame of relevance, then it may give
answers that are misleading or incorrect, because they depend more on the
simplifying assumptions than on the phenomena being modeled. Therefore
we must be especially cautious applying a model outside of its frame of rel-
evance, or even at the limits of its frame, where the simplifying assumptions
become progressively less appropriate. The problem is aggravated by the fact
that often the frame of relevance is not explicitly defined, but resides in a
tacit background of practices and skills within some discipline.
Therefore, to determine the applicability of the CT model of computa-
tion to analog computing, we must consider the frame of relevance of the

CT model. This is easiest if we recall the domain of issues and questions
it was originally developed to address: issues of effective calculability and
derivability in formalized mathematics. This frame of relevance determines
many of the assumptions of the CT model, for example, that information is
represented by finite discrete structures of symbols from a finite alphabet,
that information processing proceeds by the application of definite formal
rules at discrete instants of time, and that a computational or derivational
process must be completed in a finite number of these steps.3 Many of these
assumptions are incompatible with analog computing and with the frames of
relevance of many models of analog computation.
F.6 Relevant issues for analog computation

Analog computation is often used for control. Historically, analog computers
were used in control systems and to simulate control systems, but contempo-
rary analog VLSI is also frequently applied in control. Natural analog com-
putation also frequently serves a control function, for example, sensorimotor
control by the nervous system, genetic regulation in cells, and self-organized
cooperation in insect colonies. Therefore, control systems delimit one frame
of relevance for models of analog computation.
In this frame of relevance real-time response is a critical issue, which mod-
els of analog computation, therefore, ought to be able to address. Thus it
is necessary to be able to relate the speed and frequency response of analog
computation to the rates of the physical processes by which the computa-
tion is realized. Traditional methods of algorithm analysis, which are based
on sequential time and asymptotic behavior, are inadequate in this frame
of relevance. On the one hand, the constants (time scale factors), which
reflect the underlying rate of computation are absolutely critical (but ig-
nored in asymptotic analysis); on the other hand, in control applications the
asymptotic behavior of algorithm is generally irrelevant, since the inputs are
typically fixed in size or of a limited range of sizes.
The CT model of computation is oriented around the idea that the pur-
pose of a computation is to evaluate a mathematical function. Therefore
the basic criterion of adequacy for a computation is correctness, that is, that
3
See MacLennan (2003, 2004) for a more detailed discussion of the frame of relevance
of the CT model.
given a precise representation of an input to the function, it will produce (af-

ter finitely many steps) a precise representation of the corresponding output
of the function. In the context of natural computation and control, however,
other criteria may be equally or even more relevant. For example, robustness
is important: how well does the system respond in the presence of noise,
uncertainty, imprecision, and error, which are unavoidable in physical nat-
ural and artificial control systems, and how well does it respond to defects
and damage, which arise in many natural and artificial contexts. Since the
real world is unpredictable, flexibility is also important: how well does an
artificial system respond to inputs for which it was not designed, and how
well does a natural system behave in situations outside the range of those to
which it is evolutionarily adapted. Therefore, adaptability (through learning
and other means) is another important issue in this frame of relevance.4
F.7 Transcending Turing computability
Thus we see that many applications of analog computation raise different

questions from those addressed by the CT model of computation; the most
useful models of analog computing will have a different frame of relevance.
In order to address traditional questions such as whether analog computers
can compute “beyond the Turing limit,” or whether they can solve NP-hard
problems in polynomial time, it is necessary to construct models of analog
computation within the CT frame of relevance. Unfortunately, constructing
such models requires making commitments about many issues (such as the
representation of reals and the discretization of time), that may affect the
answers to these questions, but are fundamentally unimportant in the frame
of relevance of the most useful applications of the concept of analog compu-
tation. Therefore, being overly focused on traditional problems in the theory
of computation (which was formulated for a different frame of relevance) may
distract us from formulating models of analog computation that can address
important issues in its own frame of relevance.
4
See MacLennan (2003, 2004) for a more detailed discussion of the frames of relevance
of natural computation and control.
G. ANALOG THINKING 283
G Analog thinking
It will be worthwhile to say a few words about the cognitive implications of
analog computing, which are a largely forgotten aspect of analog vs. digital
debates of the late 20th century. For example, it was argued that analog
computing provides a deeper intuitive understanding of a system than the
alternatives do (Bissell 2004, Small 2001, ch. 8). On the one hand, analog
computers afforded a means of understanding analytically intractable sys-
tems by means of “dynamic models.” By setting up an analog simulation, it
was possible to vary the parameters and explore interactively the behavior
of a dynamical system that could not be analyzed mathematically. Digital
simulations, in contrast, were orders of magnitude slower and did not permit
this kind of interactive investigation. (Performance has improved sufficiently
in contemporary digital computers so that in many cases digital simulations
can be used as dynamic models, sometimes with an interface that mimics an
analog computer; see Bissell 2004.)
Analog computing is also relevant to the cognitive distinction between
knowing how (procedural knowledge) and knowing that (declarative knowl-
edge) (Small, 2001, ch. 8). The latter (“know-that”) is more characteristic of
scientific culture, which strives for generality and exactness, often by design-
ing experiments that allow phenomena to be studied in isolation, whereas the
former (“know-how”) is more characteristic of engineering culture; at least
it was so through the first half of the twentieth century, before the develop-
ment of “engineering science” and the widespread use of analytic techniques
in engineering education and practice. Engineers were faced with analyt-
ically intractable systems, with inexact measurements, and with empirical
relationships (characteristic curves, etc.), all of which made analog comput-
ers attractive for solving engineering problems. Furthermore, because ana-
log computing made use of physical phenomena that were mathematically
analogous to those in the primary system, the engineer’s intuition and un-
derstanding of one system could be transferred to the other. Some commen-
tators have mourned the loss of hands-on intuitive understanding resulting
from the increasingly scientific orientation of engineering education and the
disappearance of analog computers (Bissell, 2004; Lang, 2000; Owens, 1986;
Puchta, 1996).
I will mention one last cognitive issue relevant to the differences between
analog and digital computing. As already discussed Sec. C.4, it is generally
agreed that it is less expensive to achieve high precision with digital tech-
nology than with analog technology. Of course, high precision may not be
important, for example when the available data are inexact or in natural
computation. Further, some advocates of analog computing argue that high
precision digital results are often misleading (Small, 2001, p. 261). Precision
does not imply accuracy, and the fact that an answer is displayed with 10
digits does not guarantee that it is accurate to 10 digits; in particular, engi-
neering data may be known to only a few significant figures, and the accuracy
of digital calculation may be limited by numerical problems. Therefore, on
the one hand, users of digital computers might fall into the trap of trusting
their apparently exact results, but users of modest-precision analog comput-
ers were more inclined to healthy skepticism about their computations. Or
so it was claimed.
H Future directions
H.1 Post-Moore’s Law computing
Certainly there are many purposes that are best served by digital technology;
indeed there is a tendency nowadays to think that everything is done better
digitally. Therefore it will be worthwhile to consider whether analog com-
putation should have a role in future computing technologies. I will argue
that the approaching end of Moore’s Law (Moore, 1965), which has predicted
exponential growth in digital logic densities, will encourage the development
of new analog computing technologies.
Two avenues present themselves as ways toward greater computing power:
faster individual computing elements and greater densities of computing el-
ements. Greater density increases power by facilitating parallel computing,
and by enabling greater computing power to be put into smaller packages.
Other things being equal, the fewer the layers of implementation between the
computational operations and the physical processes that realize them, that
is to say, the more directly the physical processes implement the computa-
tions, the more quickly they will be able to proceed. Since most physical pro-
cesses are continuous (defined by differential equations), analog computation
is generally faster than digital. For example, we may compare analog addi-
tion, implemented directly by the additive combination of physical quantities,
with the sequential process of digital addition. Similarly, other things being
equal, the fewer physical devices required to implement a computational ele-
H. FUTURE DIRECTIONS 285
ment, the greater will be the density of these elements. Therefore, in general,
the closer the computational process is to the physical processes that realize
it, the fewer devices will be required, and so the continuity of physical law
suggests that analog computation has the potential for greater density than
digital. For example, four transistors can realize analog addition, whereas
many more are required for digital addition. Both considerations argue for
an increasing role of analog computation in post-Moore’s Law computing.
From this broad perspective, there are many physical phenomena that are
potentially usable for future analog computing technologies. We seek phe-
nomena that can be described by well-known and useful mathematical func-
tions (e.g., addition, multiplication, exponential, logarithm, convolution).
These descriptions do not need to be exact for the phenomena to be useful
in many applications, for which limited range and precision are adequate.
Furthermore, in some applications speed is not an important criterion; for
example, in some control applications, small size, low power, robustness,
etc. may be more important than speed, so long as the computer responds
quickly enough to accomplish the control task. Of course there are many
other considerations in determining whether given physical phenomena can
be used for practical analog computation in a given application (MacLen-
nan, 2009). These include stability, controllability, manufacturability, and
the ease of interfacing with input and output transducers and other devices.
Nevertheless, in the post-Moore’s Law world, we will have to be willing to
consider all physical phenomena as potential computing technologies, and in
many cases we will find that analog computing is the most effective way to
utilize them.
Natural computation provides many examples of effective analog com-
putation realized by relatively slow, low-precision operations, often through
massive parallelism. Therefore, post-Moore’s Law computing has much to
learn from the natural world.
Bibliography
Adamatzky, A. (2001). Computing in Nonlinear Media and Automata Col-

lectives. Bristol: Institute of Physics Publishing.
Adamatzky, A., De Lacy Costello, B., & Asai, T. (2005). Reaction-Diffusion

Computers. Amsterdam: Elsevier.
Adleman, L. M. (1995). On constructing a molecular computer. In R. J.

Lipton & E. B. Baum (Eds.), DNA Based Computers, Proc. of a DIMACS
Workshop (pp. 1–22).
Anderson, J. (1995). An Introduction to Neural Networks. Cambridge, MA:

MIT Press.
Ashley, J. R. (1963). Introduction to Analog Computing. New York: John

Wiley & Sons.
Aspray, W. (1993). Edwin l. harder and the anacom: Analog computing at

westinghouse. IEEE Annals of the History of Computing, 15(2), 35–52.
Ben-Hur, A., Siegelmann, H., & Fishman, S. (2002). A theory of complexity

for continuous time systems. Journal of Complexity, 18, 51–86.
Benioff, P. (1980). The computer as a physical system: A microscopic quan-

tum mechanical Hamiltonian model of computers as represented by Turing
machines. Journal of Statistical Physics, 22(5), 563–591.
Benioff, P. (1982). Quantum mechanical Hamiltonian models of Turing ma-

chines. Journal of Statistical Physics, 29(3), 515–546.
Bennett, C. H. (1973). Logical reversibility of computation. IBM Journal of

Research and Development, 17(6), 525–532.
287
288 BIBLIOGRAPHY
Bennett, C. H. (1982). The thermodynamics of computation — a review.

Int. J. Theo. Phys., 21(12), 905–940.
Bennett, C. H. (2003). Notes on Landauer’s principle, reversible computa-

tion, and Maxwell’s Demon. Studies in History and Philosophy of Modern
Physics, 34, 501–510.
Berut, A., Arakelyan, A., Petrosyan, A., Ciliberto, S., Dillenschneider, R., &
Lutz, E. (2012). Experimental verification of Landauer’s principle linking
information and thermodynamics. Nature, 483, 187–189.
Bissell, C. C. (2004). A great disappearing act: The electronic analogue

computer. In IEEE Conference on the History of Electronics Bletchley,
UK.
Blum, L., Cucker, F., Shub, M., & Smale, S. (1998). Complexity and Real
Computation. Berlin: Springer-Verlag.
Blum, L., Shub, M., & Smale, S. (1988). On a theory of computation and
complexity over the real numbers: Np completeness, recursive functions
and universal machines. The Bulletin of the American Mathematical So-
ciety, 21, 1–46.
Bournez, O., Campagnolo, M., Graça, D., & Hainry, E. (2006). The General
Purpose Analog Computer and computable analysis are two equivalent
paradigms of analog computation. In Theory and Applications of Models
of Computation (TAMC 2006), volume 3959 of Lectures Notes in Computer
Science (pp. 631–43). Berlin: Springer-Verlag.
Bournez, O. & Cosnard, M. (1996). On the computational power of dynam-

ical systems and hybrid systems. Theoretical Computer Science, 168(2),
417–59.
Bowles, M. D. (1996). U.S. technological enthusiasm and British technolog-

ical skepticism in the age of the analog brain. Annals of the History of
Computing, 18(4), 5–15.
Branicky, M. (1994). Analog computation with continuous ODEs. In Proceed-

ings IEEE Workshop on Physics and Computation (pp. 265–74). Dallas,
TX.
BIBLIOGRAPHY 289
Brockett, R. (1988). Dynamical systems that sort lists, diagonalize matri-

ces and solve linear programming problems. In Proc. 27th IEEE Conf.
Decision and Control (pp. 799–803). Austin, TX.
Calude, C., Casti, J., & Dinneen, M., Eds. (1998). Unconventional Models
of Computation. Singapore & New York: Springer.
Calude, C. & Paun, G. (2001). Computing with Cells and Atoms. London &
New York: Taylor & Francis.
Camazine, S., Deneubourg, J., Franks, N. R., Sneyd, J. Theraulaz, G., &
Bonabeau, E. (2001). Self-organization in Biological Systems. Princeton.
Changeux, J. (1985). Neuronal Man: The Biology of Mind. Oxford: Oxford

University Press. tr. by L. Garey.
Clymer, A. B. (1993). The mechanical analog computers of hannibal ford and

william newell. IEEE Annals of the History of Computing, 15(2), 19–34.
Daugman, J. (1993). An information-theoretic view of analog representation

in striate cortex. In E. Schwartz (Ed.), Computational Neuroscience (pp.
403–423). Cambridge: MIT Press.
Davidson, E. H. (2006). The Regulatory Genome: Gene Regulatory Networks

in Development and Evolution. Amsterdam: Academic Press.
Davies, J. A. (2005). Mechanisms of Morphogensis. Amsterdam: Elsevier.
Davis, M. (2004). The myth of hypercomputation. In C. Teuscher (Ed.),

Alan Turing: Life and Legacy of a Great Thinker (pp. 195–212). Berlin:
Springer-Verlag.
Davis, M. (2006). Why there is no such discipline as hypercomputation.

Applied Mathematics and Computation, 178, 4–7.
Deutsch, D. (1985). Quantum theory, the Church-Turing principle, and the

universal quantum computer. Proceedings of the Royal Society London A,
400, 97–119.
Eberbach, E., Goldin, D., & Wegner, P. (2003). Turing’s ideas and models
of computation. In C. Teuscher (Ed.), Alan Turing: Life and Legacy of a
Great Thinker. Berlin, Heidelberg & New York: Springer-Verlag.
290 BIBLIOGRAPHY
Fakhraie, S. M. & Smith, K. C. (1997). VLSI-Compatible Implementation

for Artificial Neural Networks. Boston: Kluwer Academic Publishers.
Feynman, R. (1986). Quantum mechanical computers. Foundations of
Physics, 16(6), 507–531.
Feynman, R. P. (1982). Simulating physics with computers. International
Journal of Theoretical Physics, 21, 467–488.
Frank, M. P. (2005a). The indefinite logarithm, logarithmic units, and the
nature of entropy. Dept. of Electrical & Computer Engineering, FAMU-
FSU College of Engineering.
Frank, M. P. (2005b). Introduction to reversible computing: Motivation,
progress, and challenges. In CF ‘05, May 4–6, 2005, Ischia, Italy.
Franklin, S. & Garzon, M. (1990). Neural computability. In O. M. Omidvar
(Ed.), Progress in Neural Networks, volume 1 (pp. 127–145). Norwood, NJ:
Ablex.
Fredkin, E. F. & Toffoli, T. (1982). Conservative logic. Int. J. Theo. Phys.,
21(3/4), 219–253.
Freeth, T., Bitsakis, Y., Moussas, X., Seiradakis, J., Tselikas, A., Mangou,
H., Zafeiropoulou, M., Hadland, R., Bate, D., Ramsey, A., Allen, M.,
Crawley, A., Hockley, P., Malzbender, T., Gelb, D., Ambrisco, W., &
Edmunds, M. (2006). Decoding the ancient Greek astronomical calculator
known as the Antikythera mechanism. Nature, 444, 587–91.
Garzon, M. & Franklin, S. (1989). Neural computability ii (extended ab-
stract). In Proceedings, IJCNN International Joint Conference on Neural
Networks, volume 1 (pp. 631–637). New York, NJ: Institute of Electrical
and Electronic Engineers.
Garzon, M. & Franklin, S. (1990). Computation on graphs. In O. M. Omidvar
(Ed.), Progress in Neural Networks, volume 2 chapter 13. Norwood, NJ:
Ablex.
George, S., Kim, S., Shah, S., Hasler, J., Collins, M., Adil, F., Wunderlich,
R., Nease, S., & Ramakrishnan, S. (2016). A programmable and config-
urable mixed-mode FPAA SoC. IEEE Transactions on Very Large Scale
Integration (VLSI) Systems, 24(6), 2253–2261.
BIBLIOGRAPHY 291
Goldstine, H. (1972). The Computer from Pascal to von Neumann. Princeton,

NJ: Princeton.
Grossberg, S. (1967). Nonlinear difference-differential equations in prediction

and learning theory. Proceedings of the National Academy of Sciences,
USA, 58(4), 1329–1334.
Grossberg, S. (1973). Contour enhancement, short term memory, and con-

stancies in reverberating neural networks. Studies in Applied Mathematics,
LII, 213–257.
Grossberg, S. (1976). Adaptive pattern classification and universal recoding:

I. parallel development and coding of neural feature detectors. Biological
Cybernetics, 23, 121–134.
Harnad, S. (1990). The symbol grounding problem. Physica D, 42, 335–346.
Harnad, S. (1993). Grounding symbols in the analog world. Think, 2, 12–78.
Hartl, D. L. (1994). Genetics. Boston: Jones & Bartlett, 3rd edition.
Haykin, S. (2008). Neural Networks and Learning Machines. New York:

Pearson Education, third edition.
Hopfield, J. (1984). Neurons with graded response have collective computa-

tional properties like those of two-state neurons. Proceedings of the Na-
tional Academy of Sciences USA, 81, 3088–92.
Howe, R. M. (1961). Design Fundamentals of Analog Computer Components.

Princeton, NJ: Van Nostrand.
Khatib, O. (1986). Real-time obstacle avoidance for manipulators and mobile

robots. International Journal of Robotics Research, 5, 90–9.
Kirchhoff, G. (1845). Ueber den durchgang eines elektrischen stromes durch

eine ebene, insbesondere durch eine kreisförmige. Annalen der Physik und
Chemie, 140/64(4), 497–514.
Landauer, R. (1961). Irreversibility and heat generation in the comput-

ing process. IBM Journal of Research and Development, 5(3), 183–191.
Reprinted, Vol. 44 No. 1/2, Jan./March 2000, pp. 261–269.
292 BIBLIOGRAPHY
Lang, G. F. (2000). Analog was not a computer trademark! Why would

anyone write about analog computers in year 2000? Sound and Vibration,
(pp. 16–24).
Leff, H. S. & Rex, A. F. (1990). Maxwell’s Demon: Entropy, Information,

Computing. Princeton, NJ: Princeton University Press.
Leff, H. S. & Rex, A. F. (2003). Maxwell’s Demon 2: Entropy, Classical and

Quantum Information, Computing. Bristol and Philadelphia: Institute of
Physics Publishing.
Lipka, J. (1918). Graphical and Mechanical Computation. New York: Wiley.
Lipshitz, L. & Rubel, L. A. (1987). A differentially algebraic replacment

theorem. Proceedings of the American Mathematical Society, 99(2), 367–
72.
Maass, W. & Sontag, E. (1999a). Analog neural nets with Gaussian or other
common noise distributions cannot recognize arbitrary regular languages.
Neural Computation, 11, 771–782.
Maass, W. & Sontag, E. (1999b). Analog neural nets with Gaussian or other
common noise distributions cannot recognize arbitrary regular languages.
Neural Computation, 11(3), 771–782.
MacLennan, B. J. (1987). Technology-independent design of neurocomput-

ers: The universal field computer. In M. Caudill & C. Butler (Eds.), Pro-
ceedings of the IEEE First International Conference on Neural Networks,
volume 3 (pp. 39–49).: IEEE Press.
MacLennan, B. J. (1990). Field Computation: A Theoretical Framework

for Massively Parallel Analog Computation, Parts I–IV. Technical Re-
port CS-90-100, Department of Computer Science, University of Tennessee,
Knoxville. Also available from web.eecs.utk.edu/~mclennan.
MacLennan, B. J. (1991). Gabor Representations of Spatiotemporal

Visual Images. Technical Report CS-91-144, Department of Com-
puter Science, University of Tennessee, Knoxville. Also available from
web.eecs.utk.edu/~mclennan.
BIBLIOGRAPHY 293
MacLennan, B. J. (1993). Grounding analog computers. Think, 2,

48–51. Also available from web.eecs.utk.edu/~mclennan and at
cogprints.org/542/.
MacLennan, B. J. (1994a). Continuous computation and the emergence of
the discrete. In K. H. Pribram (Ed.), Origins: Brain & Self-Organization
(pp. 121–151). Hillsdale, NJ: Lawrence Erlbaum. Also available from
web.eecs.utk.edu/~mclennan and at cogprints.org/540/.
MacLennan, B. J. (1994b). Continuous symbol systems: The logic of con-
nectionism. In D. S. Levine & M. Aparicio IV (Eds.), Neural Networks
for Knowledge Representation and Inference (pp. 83–120). Hillsdale, NJ:
Lawrence Erlbaum. Also available from web.eecs.utk.edu/~mclennan.
MacLennan, B. J. (1994c). “Words lie in our way”. Minds and Machines,
4(4), 421–437. Also available from web.eecs.utk.edu/~mclennan and at
cogprints.org/383/.
MacLennan, B. J. (1995). Continuous formal systems: A unifying model
in language and cognition. In Proceedings of the IEEE Workshop on
Architectures for Semiotic Modeling and Situation Analysis in Large
Complex Systems (pp. 161–172). Monterey, CA. Also available from
web.eecs.utk.edu/~mclennan and at cogprints.org/541.
MacLennan, B. J. (1999). Field computation in natural and artificial
intelligence. Information Sciences, 119, 73–89. Also available from
web.eecs.utk.edu/~mclennan.
MacLennan, B. J. (2001). Can Differential Equations Compute? Technical
Report UT-CS-01-459, Department of Computer Science, University of
Tennessee, Knoxville. Also available from web.eecs.utk.edu/~mclennan.
MacLennan, B. J. (2003). Transcending Turing computability. Minds and
Machines, 13, 3–22.
MacLennan, B. J. (2004). Natural computation and non-Turing models of
computation. Theoretical Computer Science, 317, 115–145.
MacLennan, B. J. (2008). Aspects of Embodied Computation: Toward a
Reunification of the Physical and the Formal. Technical Report UT-CS-
08-610, Department of Electrical Engineering and Computer Science, Uni-
versity of Tennessee, Knoxville.
294 BIBLIOGRAPHY
MacLennan, B. J. (2009). Super-Turing or non-Turing? Extending the con-

cept of computation. International Journal of Unconventional Computing,
5(3–4), 369–387.
MacLennan, B. J. (2010). The U-machine: A model of generalized com-

putation. International Journal of Unconventional Computing, 6(3–4),
265–283.
MacLennan, B. J. (2013). Cognition in Hilbert space. Behavioral and Brain

Sciences, 36(3), 296–7.
MacLennan, B. J. (2015). Field computation in natural and artificial intelli-

gence. In R. Meyers et al. (Ed.), Encyclopedia of Complexity and System
Science chapter 6, entry 199, (pp. 3334–3360). Springer, 2nd edition.
MacLennan, B. J. (prep). Foundations of Field Computation. Draft available

from author.
Maini, P. K. & Othmer, H. G., Eds. (2001). Mathematical Models for Bio-
logical Pattern Formation. Springer- Verlag.
Markov, A. (1961). Theory of Algorithms. Jerusalem, Israel: Israel Program

for Scientific Translation. US Dept. of Commerce Office of Technical Ser-
vice OTS 60-51085, transl. by Jacques J. Schorr-Kon & PST Staff. Trans-
lation of Teoriya Algorifmov, Academy of Sciences of the USSR, Moscow,
1954.
Maziarz, E. & Greenwood, T. (1968). Greek Mathematical Philosophy. New

York: Frederick Ungar.
McClelland, J., Rumelhart, D., & the PDP Research Group (1986). Parallel
Distributed Processing: Explorations in the Microstructure of Cognition,
Volume 2: Psychological and Biological Models. Cambridge, MA: MIT
Press.
Mead, C. (1987). Silicon models of neural computation. In M. Caudill & C.

Butler (Eds.), Proceedings, IEEE First International Conference on Neural
Networks, volume I (pp. 91–106). Piscataway NJ: IEEE Press.
Mead, C. (1989). Analog VLSI and Neural Systems. Reading, MA: Addison-
Wesley.
BIBLIOGRAPHY 295
Mermin, N. D. (2007). Quantum Computer Science: An Introduction. Cam-

bridge: Cambridge University Press.
Mills, J. W. (1996). The continuous retina: Image processing with a single-

sensor artificial neural field network. In Proceedings IEEE Conference on
Neural Networks: IEEE Press.
Mills, J. W., Himebaugh, B., Kopecky, B., Parker, M., Shue, C., & Weile-
mann, C. (2006). “Empty space” computes: The evolution of an unconven-
tional supercomputer. In Proceedings of the 3rd Conference on Computing
Frontiers (pp. 115–26). New York: ACM Press.
Milner, R. (1993). Elements of interaction. Communications of the ACM,

36(1), 78–89.
Milner, R., Parrow, J., & Walker, D. (1992). A calculus of mobile processes,
I & II. Information and Computation, 100, 1–77.
Moore, C. (1996). Recursion theory on the reals and continuous-time com-

putation. Theoretical Computer Science, 162, 23–44.
Moore, G. E. (1965). Cramming more components onto integrated circuits.

Electronics, 38(8), 114–117.
Murray, J. D. (1977). Lectures on Nonlinear Differential-Equation Models in

Biology. Oxford: Oxford.
Nielsen, M. A. & Chuang, I. L. (2010). Quantum Computation and Quantum

Information. Cambridge, 10th anniversary edition edition.
Omohundro, S. (1984). Modeling cellular automata with partial differential

equations. Physica D, 10, 128–34.
Orponen, P. (1997). A survey of continous-time computation theory. In

Advances in Algorithms, Languages, and Complexity (pp. 209–224).
Orponen, P. & Matamala, M. (1996). Universal computation by finite two-

dimensional coupled map lattices. In Proceedings, Physics and Computa-
tion 1996 (pp. 243–7). Cambridge, MA: New England Complex Systems
Institute.
296 BIBLIOGRAPHY
Owens, L. (1986). Vannevar bush and the differential analyzer: The text and
context of an early computer. Technology and Culture, 27(1), 63–95.
Peterson, G. R. (1967). Basic Analog Computation. New York: Macmillan.
Pothos, E. M. & Busemeyer, J. R. (2013). Can quantum probability provide

a new direction for cognitive modeling? Behavioral and Brain Sciences,
36, 255–327.
Pour-El, M. (1974a). Abstract computability and its relation to the gen-

eral purpose analog computer (some connections between logic, differential
equations and analog computers). Transactions of the American Mathe-
matical Society, 199, 1–29.
Pour-El, M. & Richards, I. (1979). A computable ordinary differential equa-

tion which possesses no computable solution. Annals of Mathematical
Logic, 17, 61–90.
Pour-El, M. & Richards, I. (1981). The wave equation with computable

initial data such that its unique solution is not computable. Advances in
Mathematics, 39, 215–239.
Pour-El, M. & Richards, I. (1982). Noncomputability in models of physical

phenomena. International Journal of Theoretical Physics, 21, 553–555.
Pour-El, M. B. (1974b). Abstract computability and its relation to the gen-

eral purpose analog computer (some connections between logic, differential
equations and analog computers). Transactions of the American Mathe-
matical Society, 199, 1–29.
Puchta, S. (1996). On the role of mathematics and mathematical knowledge

in the invention of vannevar bush’s early analog computers. IEEE Annals
of the History of Computing, 18(4), 49–59.
Reiner, J. M. (1968). The Organism as an Adaptive Control System. Engle-

wood Cliffs: Prentice-Hall.
Rieffel, E. & Polak, W. (2000). An introduction to quantum computing for

non-physicists.
BIBLIOGRAPHY 297
Rimon, E. & Koditschek, D. (1989). The construction of analytic diffeomor-

phisms for exact robot navigation on star worlds. In Proceedings of the
1989 IEEE International Conference on Robotics and Automation, Scotts-
dale AZ (pp. 21–6). New York: IEEE Press.
Rogers, A. & Connolly, T. (1960). Analog Computation in Engineering DE-

SIGN. New York: McGraw-Hill.
Rubel, L. A. (1985). The brain as an analog computer. Journal of Theoretical

Neurobiology, 4, 73–81.
Rubel, L. A. (1988). Some mathematical limitations of the general-purpose

analog computer. Advances in Applied Mathematics, 9, 22–34.
Rubel, L. A. (1993). The extended analog computer. Advances in Applied

Mathematics, 14, 39–50.
Rumelhart, D., McClelland, J., & the PDP Research Group (1986). Parallel
Distributed Processing: Explorations in the Microstructure of Cognition,
Volume 1: Foundations. Cambridge, MA: MIT Press.
Sanger, T. (1996). Probability density estimation for the interpretation of

neural population codes. Journal of Neurophysiology, 76, 2790–3.
Shannon, C. E. (1941). Mathematical theory of the differential analyzer.

Journal of Mathematics and Physics of the Massachusetts Institute Tech-
nology, 20, 337–354.
Shannon, C. E. (1993). Mathematical theory of the differential analyzer. In

N. J. A. Sloane & A. D. Wyner (Eds.), Claude Elwood Shannon: Collected
Papers (pp. 496–513). New York: IEEE Press.
Siegelmann, H., Ben-Hur, A., & Fishman, S. (1999). Computational complex-

ity for continuous time dynamics. Physical Review Letters, 83(7), 1463–6.
Siegelmann, H. & Sontag, E. (1994a). Analog computation via neural net-

works. Theoretical Computer Science, 131, 331–360.
Siegelmann, H. & Sontag, E. (1994b). Analog computation via neural net-

works. Theoretical Computer Science, 131, 331–360.
298 BIBLIOGRAPHY
Siegelmann, H. T. (1999). Neural Networks and Analog Computation: Be-

yond the Turing Limit. Boston: Birkhäuser.
Small, J. S. (1993). General-purpose electronic analog computing. IEEE

Annals of the History of Computing, 15(2), 8–18.
Small, J. S. (2001). The Analogue Alternative. London & New York: Rout-
ledge.
Sprecher, D. (1965). On the structure of continuous functions of several

variables. Transactions of the American Mathematical Society, 115, 340–
355.
Stannett, M. (1990). X-machines and the halting problem: Building a super-

Turing machine. Formal Aspects of Computing, 2, 331–341.
Stepney, S. (2004). Journeys in non-classical computation. In T. Hoare &

R. Milner (Eds.), Grand Challenges in Computing Research (pp. 29–32).
Swindon: BCS.
Thomson, W. (1876). Mechanical integration of the general linear differential

equation of any order with variable coefficients. Proceedings of the Royal
Society, 24, 271–275.
Thomson, W. (1878). Harmonic analyzer. Proceedings of the Royal Society,

27, 371–373.
Thomson, W. (1938). The tides. In The Harvard Classics, volume 30: Sci-
entific Papers (Lord Kelvin) (pp. 274–307). New York: Collier.
Tong, D. (Lent Term, 2011 and 2012). Statistical Physics: University of

Cambridge Part II Mathematical Tripos.
Truitt, T. D. & Rogers, A. E. (1960). Basics of Analog Computers. New

York: John F. Rider.
van Gelder, T. (1997). Dynamics and cognition. In J. Haugeland (Ed.), Mind

Design II: Philosophy, Psychology and Artificial Intelligence chapter 16,
(pp. 421–450). Cambridge MA: MIT Press, revised & enlarged edition.
Wegner, P. (1997). Why interaction is more powerful than algorithms. Com-

munications of the ACM, 40(5), 81–91.
BIBLIOGRAPHY 299
Wegner, P. (1998). Interactive foundations of computing. Theoretical Com-

puter Science, 192(2), 315–351.
Wegner, P. & Goldin, D. (2003). Computation beyond Turing machines:

Seeking appropriate methods to model computing and human thought.
Communications of the ACM, 46(4), 100–102.
Weyrick, R. C. (1969). Fundamentals of Analog Computers. Englewood

Cliffs: Prentice-Hall.
Winfree, E. (1998). Algorithmic Self-Assembly of DNA. PhD thesis, Califor-

nia Institute of Technology.
Wolpert, D. (1991). A Computationally Universal Field Computer which is

Purely Linear. Technical Report LA-UR-91-2937, Los Alamos National
Laboratory.
Wolpert, D. H. & MacLennan, B. J. (1993). A Computationally Universal

Field Computer that is Purely Linear. Technical Report CS-93-206, Dept.
of Computer Science, University of Tennessee, Knoxville. Also available
from web.eecs.utk.edu/~mclennan.

Unconventional Computation - MacLennan - 2018

Uploaded by

Copyright:

Available Formats

Unconventional Computation - MacLennan - 2018

Uploaded by

Document Information

Original Title

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

Unconventional Computation - MacLennan - 2018

Uploaded by

Copyright:

Available Formats

November 17, 2018

III Quantum Computation 67

A.2 Linear algebra review . . . . . . . . . . . . . . . . . . . 67

IV Molecular Computation 199

V Analog Computation 231

F.3 Real-valued inputs, output, and constants . . . . . . . 278

A What is unconventional computation?

• Programs are written as hierarchically organized sequences of instruc-

B Post-Moore’s Law computing

Fall 2012 Unconventional Computation 8

C Super-Turing vs. non-Turing

C.1 The limits of Turing computation

C.1.a Frames of Relevance

It is important to remember that Church-Turing (CT) computation is a model

C.1.b The frame of relevance of Church-Turing computation

It is important to explicate the frame of relevance of Church-Turing (CT)

perfect accuracy according to the rules of derivation. These are reasonable

the investigation of effective calculation and derivation. Furthermore, many

C.2 New computational models

C.2.a Natural computation

is largely irrelevant. The question is whether it will produce a good-enough

C.2.b Cross-frame comparisons

way by finite programs), or arbitrary reals (represented by infinite discrete

C.2.c Relevant issues for natural computation

to unanticipated inputs or damage to its sensors or effectors? A related

ways have a non-zero probability of moving backwards, and therefore molec-

C.2.e Summary of issues

C.3 Computation in general

processes to accomplish mathematical operations on mathematical objects

C.3.b Effectiveness and mechanism

Sec. C.4.b I will discuss some general effectiveness criteria.)

C.3.c Multiple realizability

C.3.d Defining computation

Definition 2 A mechanistic physical system realizes a computation if, at the

Definition 4 A computer is an artificial computational system.

Thus the term “computer” is restricted to intentionally manufactured com-

On this basis we may identify many natural computational systems. For

C.3.g Classification of computational dynamics

natural computation, are better matched to these alternative topologies.

C.4 Expanding the range of computing technologies

C.4.b General guidelines

C.4.c Learning to use new technologies

C.4.d General-purpose computation

ular application area, we must consider in what respects the programmed

Law unconventional computation. Therefore we need to supplement it with

Nops Nops Ediss

1.E-20 Room-temperature kT thermal energy

Room-temperature von Neumann - Landauer limit

Figure II.2: Depiction of 0-1-0-1-1 pulses in the presence of high thermal

This, therefore, is an estimate of the minimum energy dissipation per op-

of conventional computing will stop improving, at least at the device level”

B.1 Von Neumann-Landaur Principle

A uniform distribution maximizes the entropy (and minimizes the ability to

B.1.b The von Neumann-Landauer bound

Figure II.3: Physical microstates representing logical states. Setting the

unspecific logarithms, I will drop the “B” subscript: S = kH = k log N . The

we know which half of the state-space it is in.

hQipLandauer = kT [log 2 + p log p + (1 − p) log(1 − p)] = kT [log 2 − H(p, 1 − p)].