Styer Thermo

d3rB
rB
d3rA
rA
Statistical
Mechanics
Daniel F. Styer
November 2016
Statistical Mechanics
Dan Styer
Schiffer Professor of Physics
Oberlin College
Oberlin, Ohio 44074-1088
[email protected]
http://www.oberlin.edu/physics/dstyer
c 22 November 2016

Daniel F. Styer
The copyright holder grants the freedom to copy, modify, convey, adapt, and/or redistribute this work under
the terms of the Creative Commons Attribution Share Alike 4.0 International License. A copy of that license
is available at https://creativecommons.org/licenses/by-sa/4.0/legalcode.
Although, as a matter of history, statistical mechanics owes its origin

to investigations in thermodynamics, it seems eminently worthy of an
independent development, both on account of the elegance and simplicity
of its principles, and because it yields new results and places old truths
in a new light.
J. Willard Gibbs
Elementary Principles in Statistical Mechanics
Contents
0 Preface
1 The Properties of Matter in Bulk
1.1
What is Statistical Mechanics About? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Outline of Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3
Fluid Statics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4
Phase Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5
Additional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Principles of Statistical Mechanics
11
2.1
Microscopic Description of a Classical System . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.2
Macroscopic Description of a Large Equilibrium System . . . . . . . . . . . . . . . . . . . . .
15
2.3
Fundamental Assumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2.4
Statistical Definition of Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
2.5
Entropy of a Monatomic Ideal Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
2.6
Qualitative Features of Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
2.7
Using Entropy to Find (Define) Temperature and Pressure . . . . . . . . . . . . . . . . . . . .
37
2.8
Additional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
3 Thermodynamics
49
3.1
Heat and Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
3.2
Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
50
ii
CONTENTS
3.3
Heat Engines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
3.4
Thermodynamic Quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
3.5
Multivariate Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
3.6
The Thermodynamic Dance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
3.7
Non-fluid Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
3.8
Thermodynamics Applied to Fluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
72
3.9
Thermodynamics Applied to Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . .
80
3.10 Thermodynamics Applied to Chemical Reactions . . . . . . . . . . . . . . . . . . . . . . . . .
81
3.11 Thermodynamics Applied to Light . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
3.12 Additional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
90
4 Ensembles
94
4.1
The Canonical Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
94
4.2
Meaning of the Term Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
96
4.3
Classical Monatomic Ideal Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
96
4.4
Energy Dispersion in the Canonical Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . .
99
4.5
Temperature as a Control Variable for Energy (Canonical Ensemble) . . . . . . . . . . . . . . 102
4.6
The Equivalence of Canonical and Microcanonical Ensembles . . . . . . . . . . . . . . . . . . 103
4.7
The Grand Canonical Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.8
The Grand Canonical Ensemble in the Thermodynamic Limit . . . . . . . . . . . . . . . . . . 104
4.9
Summary of Major Ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.10 Quantal Statistical Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

4.11 Ensemble Problems I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.12 Ensemble Problems II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5 Classical Ideal Gases
118
5.1
Classical Monatomic Ideal Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.2
Classical Diatomic Ideal Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.3
Heat Capacity of an Ideal Gas
5.4
Specific Heat of a Hetero-nuclear Diatomic Ideal Gas . . . . . . . . . . . . . . . . . . . . . . . 126
5.5
Chemical Reactions Between Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.6
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
CONTENTS
6 Quantal Ideal Gases
iii
133
6.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.2
The Interchange Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.3
Quantum Mechanics of Independent Identical Particles . . . . . . . . . . . . . . . . . . . . . . 134
6.4
Statistical Mechanics of Independent Identical Particles . . . . . . . . . . . . . . . . . . . . . 141
6.5
Quantum Mechanics of Free Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.6
Fermi-Dirac Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.7
Bose-Einstein Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.8
Specific Heat of the Ideal Fermion Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.9
Additional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7 Harmonic Lattice Vibrations
165
7.1
The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.2
Statistical Mechanics of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.3
Normal Modes for a One-dimensional Chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.4
Normal Modes in Three Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.5
Low-temperature Heat Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.6
More Realistic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.7
What is a Phonon? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.8
8 Interacting Classical Fluids
170
8.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
8.2
Perturbation Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.3
Variational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
8.4
Distribution Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.5
Correlations and Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
8.6
The Hard Sphere Fluid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
iv
CONTENTS
9 Strongly Interacting Systems and Phase Transitions
183
9.1
Introduction to Magnetic Systems and Models
. . . . . . . . . . . . . . . . . . . . . . . . . . 183
9.2
Free Energy of the One-Dimensional Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . 183
9.3
The Mean-Field Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
9.4
Correlation Functions in the Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
9.5
Computer Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.6
A Series and Integrals
209
B Evaluating the Gaussian Integral
210
C Clinic on the Gamma Function
211
D Volume of a Sphere in d Dimensions
213
E Stirlings Approximation
216
F The Euler-MacLaurin Formula and Asymptotic Series
218
G Ramblings on the Riemann Zeta Function
219
H Tutorial on Matrix Diagonalization
225
H.1 Whats in a name? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

H.2 Vectors in two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
H.3 Tensors in two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
H.4 Tensors in three dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
H.5 Tensors in d dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
H.6 Linear transformations in two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
H.7 What does eigen mean? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
H.8 How to diagonalize a symmetric matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
H.9 A glance at computer algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
H.10 A glance at non-symmetric matrices and the Jordan form . . . . . . . . . . . . . . . . . . . . 242
CONTENTS
I
Catalog of Misconceptions
v
246
J Thermodynamic Master Equations
248
K Useful Formulas
249
Index
250
Chapter 0
Preface
This is a book about statistical mechanics at the advanced undergraduate level. It assumes a background in
classical mechanics through the concept of phase space, in quantum mechanics through the Pauli exclusion
principle, and in mathematics through multivariate calculus. (Section 9.2 also assumes that you can can
diagonalize a 2 2 matrix.)
The book is in draft form and I would appreciate your comments and suggestions for improvement. In
particular, if you have an idea for a new problem, or for a new class of problems, I would appreciate hearing
about it. If you are studying statistical mechanics and find that you are asking yourself Why is. . . true or
What would happen if. . . were changed? or I wish I were more familiar with. . . , then you have the germ
of a new problem. . . so please tell me!
The specific goals of this treatment include:
To demonstrate the extraordinary range of applicability of the ideas of statistical mechanics. These
ideas are applicable to crystals and magnets, superconductors and solutions, surfaces and even bottles
of light. I am always irritated by books that apply statistical mechanics only to fluids, or worse, only
to the ideal gas.
To introduce some modern concepts and tools such as correlation functions and computer simulation.
While a full treatment awaits the final two chapters, the need for these tools is presented throughout
the book.
To develop qualitative insight as well as analytic and technical skills. Serving this goal are particular
sections of text (e.g. section 2.6) and particular problems (e.g. problem 2.4) as well as overall attention
to conceptual issues. Also in support of this goal is an attempt to eliminate misconceptions as well
as present correct ideas (see appendix I, Catalog of Misconceptions). I particularly aim to foster a
qualitative understanding of those central concepts, entropy and chemical potential, emphasizing that
the latter is no more difficult to understand (and no less!) than temperature.
1
CHAPTER 0. PREFACE
To review classical and quantum mechanics, and mathematics, with a special emphasis on difficult
points. It is my experience that some topics are so subtle that they are never learned on first exposure,
no matter how well they are taught. A good example is the interchange rule for identical particles
in quantum mechanics. The rule is so simple and easy to state, yet its consequences are so dramatic
and far reaching, that any student will profit from seeing it treated independently in both a quantum
mechanics course and a statistical mechanics course.
To develop problem solving skills. In the sample problems I attempt to teach both strategy and
tacticsrather than just show how to solve a problem I point out why the steps I use are profitable.
(See the index entry on problem solving tips.) Throughout I emphasize that doing a problem involves
more than simply reaching an algebraic result: it also involves trying to understand that result and
appreciate its significance.
The problems are a central part of this book. (Indeed, I wrote the problems first and designed the book
around them.) In any physics course, the problems play a number of roles: they check to make sure you
understand the material, they force you to be an active (and thus more effective) participant in your own
learning, they expose you to new topics and applications. For this reason, the problems spread over a wide
range of difficulty, and I have identified them as such.
Easy problems (marked with an E) might better be called questions. You should be able to answer
them after just a few moments thought, usually without the aid of pencil and paper. Their main
purpose is to check your understanding or to reinforce conceptual points already made in the text.
You should do all the easy problems as you come to them, but they are not usually appropriate for
assigned problem sets. The level of these problems is similar to that of problems found on the Graduate
Record Examination in physics. . . some of them are even set up in multiple-choice format.
Intermediate problems (marked with an I) are straightforward applications of the text material. They
might better be called exercises because they exercise the skills you learned in the text without
calling them into stress. They are useful both for checking your understanding and for forcing your
participation. Sometimes the detailed derivation of results described in the text is left for such an
exercise, with appropriate direction and clues.
Difficult problems (marked with one to three Ds, to indicate the level of difficulty) are more ambitious
and usually require the student to pull together knowledge and skills from several different pieces of
this book, and from other physics courses as well. Sometimes these problems introduce and develop
topics that are not mentioned in the text at all.
In addition, some easy or intermediate problems are marked with a star (e.g. I*) to indicate that they
are essential problems. These problems are so important to the development of the course material
that you must do them to have any hope of being able to understand the course material. Essential
problems are either relied upon in later text material (e.g. problem 1.2) or else cover important topics
that are not covered in the text itself (e.g. problem 3.6). In the latter case, I covered that important
3
topic through a problem rather than trough text material because I thought it was easier to understand
in that form.
I have attempted to make all of the problems at all of the levels interesting in their own right, rather than
burdens to complete out of necessity. For that reason, I recommend that you read all of the problems, even
the ones that you dont do.
The bibliographies at the end of each chapter are titled Resources rather than Additional Reading
because they include references to computer programs and world wide web pages as well as to books and
articles. The former will no doubt go out of date quickly, but to ignore them would be criminal.
Ideas. Perhaps book should come with disk of some public domain programs. . . or maybe I should set up
a web page for the book.
Appendices J and K should appear as rear and front endpapers, respectively.
Chapter 1
The Properties of Matter in Bulk

1.1
What is Statistical Mechanics About?
Statistical mechanics treats matter in bulk. While most branches of physics. . . classical mechanics, atomic
physics, quantum mechanics, nuclear physics. . . deal with one or two or a few dozen particles, statistical
mechanics deals with, typically, about a mole of particles at one time. A mole is 6.02 1023 , considerably
larger than a few dozen. Lets compare this to a number often considered large, namely the U.S. national
debt. This debt is (2014) about 18 trillion dollars, so the national debt is about thirty trillionth of a mole
of dollars.1 Even so, a mole of water molecules occupies only 18 ml or about half a fluid ounce. . . its just a
sip.
The huge number of particles present in the systems studied by statistical mechanics means that the
traditional questions of physics are impossible to answer. For example, the traditional question of classical
mechanics is the time-development problem: Given the positions and velocities of all the particles now, find
out what they will be at some future time. This problem has has not been completely solved for three
gravitating bodies. . . clearly we will get nowhere asking the same question for 6.02 1023 bodies! But in fact,
a solution of the time-development problem for a mole of water molecules would be useless even if it could be
obtained. Who cares where each molecule is located? No experiment will ever be able to find out. To make
progress, we have to ask different questions, question like How does the pressure change with volume?,
How does the temperature change upon adding particles?, What is the mean distance between atoms?,
or What is the probability for finding two atoms separated by a given distance?. Thus the challenge of
statistical mechanics is two-fold: first find the questions, and only then find the answers.
1 In
contrast, the Milky Way galaxy contains about 0.3 or 0.6 trillionth of a mole of stars. The entire universe probably
contains fewer than a mole of stars.
1.2. OUTLINE OF BOOK
1.2
Outline of Book
This book begins with a chapter, the properties of matter in bulk , that introduces statistical mechanics and
shows why it is so fascinating.
It proceeds to discuss the principles of statistical mechanics. The goal of this chapter is to motivate and
then produce a conceptual definition for that quantity of central importance: entropy. In contrast to, say,
quantum mechanics, it is not useful to cast the foundations of statistical mechanics into a mathematically
rigorous postulate, theorem, proof mold. Our arguments in this chapter are often heuristic and suggestive;
plausibility arguments rather than proofs.
Once we have defined entropy and know a few of its properties, what can we do with it? The subject of
thermodynamics asks what can be discovered about substance by just knowing that entropy exists, without
knowing a formula for it. It is one of the most fascinating fields in all of science, because it produces a large
number of dramatic and unexpected results based on this single modest assumption. This books chapter on
thermodynamics begins begins by developing a concrete operational definition for entropy, in terms of heat
and work, to complement the conceptual definition produced in the previous chapter. It goes on to apply
entropy to situations as diverse as fluids, phase transitions, and light.
The chapter on ensembles returns to issues of principle, and it produces formulas for the entropy that
are considerably easier to apply than the one produced in chapter 2. Armed with these easier formulas, the
rest of the book uses them in various applications.
The first three applications are to the classic topics of classical ideal gases, quantal ideal gases, including
Fermi-Dirac and Bose-Einstein statistics, and harmonic lattice vibrations or phonons.
The subject of ideal gases (i.e. gases of non-interacting particles) is interesting and often useful, but it
clearly does not tell the full story. . . for example, the classical ideal gas can never condense into a liquid, so
it cannot show any of the fascinating and practical phenomena of phase transitions. The next chapter treats
weakly interacting fluids, using the tools of perturbation theory and the variational method. The correlation
function is introduced as a valuable tool. This is the first time in the book that we ask questions more
detailed than the questions of thermodynamics.
Finally we treat strongly interacting systems and phase transitions. Here our emphasis is on magnetic
systems. Tools include mean field theory, transfer matrices, correlation functions, and computer simulations.
Under this heading fall some of the most interesting questions in all of science. . . some answered, many still
open.
The first five chapters (up to and including the chapter on classical ideal gases) are essential background
to the rest of the book, and they must be treated in the sequence presented. The last four chapters are
independent and can be treated in any order.
1.3
CHAPTER 1. THE PROPERTIES OF MATTER IN BULK
Fluid Statics
I mentioned above that statistical mechanics asks questions like How does the pressure change with volume?. But what is pressure? Most people will answer by saying that pressure is force per area:
pressure =
force
.
area
(1.1)
But force is a vector and pressure is a scalar, so how can this formula be correct? The aim of this section is
to investigate what this formula means and find out when it is correct.2
1.3.1
Problems
1.1 (I) The rotating water glass

A cylinder containing a fluid of mass density is placed on the center of a phonograph turntable and
rotated with constant angular velocity . After some initial sloshing of the fluid, everything calms
down to a steady state.
yp p p p
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p6
6
r
6
h
hc
a. The pressure is a function of height h and distance from the axis r. Show that the variation of
pressure with radial distance is
p(r, h)
= 2 r,
(1.2)
r
while the variation with vertical distance is
p(r, h)
= g.
h
(1.3)
(Where g is the acceleration of gravity.)

2 As such, the aim of this section is quite modest. If you want to learn more about the interesting subject of fluid flow, see
the Resources section of this chapter.
1.4. PHASE DIAGRAMS
b. The pressure at the surface of the fluid at the center of the cylinder (r = 0, h = hc ) is of course
atmospheric pressure pa . Integrate the differential equations of part (a.) to show that, at any
point in the fluid,
(1.4)
p(r, h) = pa + 12 2 r2 g(h hc ).
c. Show that the profile of the fluid surface is given by
y(r) =
1.4
2 2
r .
2g
(1.5)
Phase Diagrams
Too often, books such as this one degenerate into a study of gases. . . or even into a study of the ideal gas!
Statistical mechanics in fact applies to all sorts of materials: fluids, crystals, magnets, metals, polymers,
starstuff, even light. I want to show you some of the enormous variety of behaviors exhibited by matter in
bulk, and that can (at least in principle) be explained through statistical mechanics.
Because the axes of a phase diagram are pressure and temperature, the misconception arises that phase
diagrams plot pressure as a function of temperature. No. Pressure and temperature are independent variables. For example, volume is a function of pressure and temperature, V (T, p). Instead, the lines on a phase
diagram mark the places where there are cliffs in the function V (T, p).
End with the high Tc phase diagram of Amnon Aharony discussed by MEF at Gibbs Symposium.
Birgeneau.
Resources
The problems of fluid flow are neglected in the typical American undergraduate physics curriculum. An
introduction to these fascinating problems can be found in the chapters on elasticity and fluids in any
introductory physics book, such as
F.W. Sears, M.W. Zemansky, and H.D. Young, University Physics, fifth edition (Addison-Wesley,
Reading, Massachusetts, 1976), chapters 10, 12, and 13, or
D. Halliday, R. Resnick, and J. Walker, Fundamentals of Physics, fourth edition (John Wiley,
New York, 1993), sections 161 to 167.
More idiosyncratic treatments are given by
R.P. Feynman, R.B. Leighton, and M. Sands, The Feynman Lectures on Physics (Addison-Wesley,
Reading, Massachusetts, 1964), chapters II-40 and II-41, and
Jearl Walker The Flying Circus of Physics (John Wiley, New York, 1975), chapter 4.

Hansen and McDonald
An excellent description of various states of matter (including liquid crystals, antiferromagnets, superfluids, spatially modulated phases, and more) extending our section on Phase Diagrams is
Michael E. Fisher, The States of MatterA Theoretical Perspective in W.O. Milligan,
ed., Modern Structural Methods (The Robert A. Welch Foundation, Houston, Texas, 1980) pp.
74175.
1.5
Additional Problems
1.2 (I*) Compressibility, expansion coefficient

The isothermal compressibility of a substance is defined as
T (p, T ) =
1 V (p, T )
,
V
p
(1.6)
where the volume V (p, T ) is treated as a function of pressure and temperature.

a. Justify the name compressibility. If a substance has a large T is it hard or soft? Since squeeze
is a synonym for compress, is squeezeabilty a synonym for compressibility? Why were the
negative sign and the factor of 1/V included in the definition?
b. The expansion coefficient is
1 V (p, T )
.
V
T
In most situations is positive, but it can be negative. Give one such circumstance.
(p, T ) =
(1.7)
c. What are T and in a region of two-phase coexistence (for example, a liquid in equilibrium with
its vapor)?
d. Find T and for an ideal gas, V (p, T ) = N kB T /p.
e. Show that
T (p, T )
(p, T )
=
T
p
(1.8)
for all substances.

f. Verify this relation for the ideal gas. (Clue: The two expressions are not both equal to N kB /p2 V .)
1.3 (I*) Heat capacity as a susceptibility
Later in this book we will find that the heat capacity of a fluid3 is
CV (T, V, N ) =
E(T, V, N )
,
T
(1.9)
where the energy E(T, V, N ) is considered as a function of temperature, volume, and number of particles. The heat capacity is easy to measure experimentally and is often the first quantity observed
3 Technically,
the heat capacity at constant volume.
1.5. ADDITIONAL PROBLEMS
when new regimes of temperature or pressure are explored. (For example, the first sign of superfluid
He3 was an anomalous dip in the measured heat capacity of that substance.)
a. Explain how to measure the heat capacity of a gas given a strong, insulated bottle, a thermometer,
a resistor, a voltmeter, an ammeter, and a clock.
b. Near the superfluid transition temperature Tc , the heat capacity of Helium is given by
CV (T ) = A ln(|T Tc |/Tc ).
(1.10)
Sketch the heat capacity and the energy as a function of temperature in this region.
c. The heat capacity is one member of a class of thermodynamic quantities called susceptibilities.
Why does it have that name? (Clues: A change in temperature causes a change in energy, but
how much of a change? If the heat capacity is relatively high, is the system relatively sensitive or
insensitive (i.e. susceptible or insusceptible) to such temperature changes?)
d. Interpret the isothermal compressibility (1.6) as a susceptibility. (Clue: A change in pressure
causes a change in volume.)
1.4 (I) The meaning of never
(This problem is modified from Kittel and Kroemer, Thermal Physics, second edition, page 53.)
It has been said4 that six monkeys, set to strum unintelligently on typewriters for millions of years,
would be bound in time to write all the books in the British Museum. This assertion gives a misleading
impression concerning very large numbers.5 Consider the following situation:
The quote considers six monkeys. Lets be generous and allow 1010 monkeys to work. (About
twice the present human population.)
The quote vaguely mentions millions of years. The age of the universe is about 14 billion years,
so lets allow all of that time, about 1018 seconds.
The quote wants to write out all the books in a very large library. Lets be modest and demand
only the production of Shakespeares Hamlet, a work of about 105 characters.
Finally, assume that a monkey can type ten characters per second, and for definiteness, assume a
keyboard of 29 characters (letters, comma, period, space. . . ignore caPitALIzaTion).
a. Show that the probability that a given sequence of 105 characters comes out through a random
striking of 105 keys is
1
10146 240 .
(1.11)
100
29 000
How did you perform the arithmetic?
4 J.
Jeans, Mysterious Universe (Cambridge University Press, Cambridge, 1930) p. 4.

insightful discussion of the monkeys at typewriters problem, and its implications for biological evolution, is given by
Richard Dawkins in his book The Blind Watchmaker (Norton, New York, 1987) pp. 4349.
5 An
10

b. Show that the probability of producing Hamlet through the unintelligent strumming of 1010
monkeys over 1018 seconds is about 10146 211 , which is small enough to be considered zero for
most purposes.
1.5 Human genetics

There are about 21,000 genes in the human genome. Suppose that each gene could have any of three
possible states (called alleles). (For example, if there were a single gene for hair color, the three
alleles might be black, brown, and blond.) Then how many genetically distinct possible people would
there be? Compare to the current worldwide human population. Estimate how long it would take for
every possible genetically distinct individual to be realized. Compare to the age of the universe.
Chapter 2
Principles of Statistical Mechanics

In a book on classical or quantum mechanics, the chapter corresponding to this one would be titled Foundations of Classical (or Quantum) Mechanics. Here, I am careful to use the term principles rather than
foundations. The term foundations suggests rock solid, logically rigorous, hard and fast rules, such
as the experimental evidence that undergirds quantum theory. Statistical mechanics lacks such a rigorous
undergirding. Remember that our first job in statistical mechanics is to find the questions. Of course, we
can ask any question we wish, but we need to find profitable questions. Such a task is not and cannot be
one of rigor.
Our treatment here is based on classical mechanics, not quantum mechanics. This approach is easier
and more straightforward than the approach through quantum mechanics, andas you will soon seethere
are quite enough difficulties and subtleties in the classical approach! After we investigate the principles
of statistical mechanics from a classical perspective, we will outline (in section 4.10) the generalization to
quantum mechanics.
2.1
Microscopic Description of a Classical System
This section deals with classical mechanics, not statistical mechanics. But any work that attempts to build
macroscopic knowledge from microscopic knowledgeas statistical mechanics doesmust begin with a clear
and precise statement of what that microscopic knowledge is.
The microscopic description of a physical system has two components: First, What are the parts of
the system? How do they affect each other?, second, How are those parts arranged?. The first question
is answered by giving the mechanical parameters 1 of the system. The second is answered by giving its
dynamical variables. Rather than give formal definitions of these terms, we give two examples.
1 The
mechanical parameters are sometimes called parameters in the Hamiltonian, or external parameters, or fields.
11
12
CHAPTER 2. PRINCIPLES OF STATISTICAL MECHANICS
The earth-moon system. In this system, the mechanical parameters are the mass of the earth, the mass
of the moon, and (because the earth and moon interact gravitationally), the gravitational constant G. The
dynamical variables are the position and velocity (or momentum) of each body. (Alternative dynamical
variables are the position of the center of mass and the separation between the bodies, plus the total
momentum of the system and the angular momentum of the two bodies about the center of mass.) You can
see from this example that the mechanical parameters give you the knowledge to write down the Hamiltonian
for the system, while the dynamical variables are the quantities that change according to the laws governed
through that Hamiltonian. The mechanical parameters do not depend upon the initial condition of the
systemthe dynamical variables do. Often (although not always) the mechanical parameters are timeconstant while the dynamical variables are time-varying.
Helium atoms in a box. It is natural to begin the description by saying that there are N atoms, each of
mass m. But this innocent beginning is neither obvious nor precisely correct. By saying that the only thing
we need to know about each atom is its mass, we are modeling the atoms as point particles. A more precise
model would describe the system as N nuclei and 2N electrons, but then our treatment would necessarily
involve quantum mechanics rather than classical mechanics. Furthermore we would not gain anything by
this more precise and more difficult description. . . we know from experience that under ordinary conditions
the nuclei and electrons do bind themselves together as atoms. Even the more precise description would
not result in unassailable rigor, because the nuclei themselves are made up of nucleons and the nucleons of
quarks. In fact, a model-building process similar to this one went on unmentioned even in our treatment of
the earth-moon system: When we said that we needed to know only the masses of the two bodies, we were
assuming (a good but not perfect assumption) that the distribution of matter through the earth and moon
was irrelevant to their motion. We will adopt the model that replaces helium atoms by point particles, but
you should keep in mind that it is a model.
To continue in our microscopic description, we need to know how the atoms interact with each other.
A common model is that atoms interact in a pairwise fashion through some sort of atom-atom interaction
potential such as the Lennard-Jones 612 potential:
atom-atom potential energy =
b
a
+ 12 .
r6
r
(2.1)
Here the quantities a and b are mechanical parameters.

An important part of the microscopic description of helium atoms in a box is a description of the box. We
must know where the walls are located and how the walls interact with the atoms. There are many different
models for the wall-atom interaction: common models include the hard wall and soft wall potential
energy functions sketched in figure 2.1. But even by assuming that the walls and atoms interact through a
potential at all we are making a dramatic and easily overlooked assumption: the assumption of a smooth,
non-atomic wall. Real box walls are themselves made up of atoms, which can of course move, so there is no
fixed zero-distance point as implied by the graphs in the figure. The assumption of fixed, smooth walls
implies that when an atom collides with the wall its energy is unchanged, while it is quite possible for the
atom to gain or loose some energy while colliding with a real, atomic wall. The assumption of a smooth wall
is particularly suspicious, and we will find in chapter 4 that it can be relaxed very profitably.
2.1. MICROSCOPIC DESCRIPTION OF A CLASSICAL SYSTEM
13
atom-wall
potential
energy
hard wall
soft wall
distance from wall
Figure 2.1:
The final step is to recognize that there might be externally electric, magnetic, or gravitational fields
that affect the system. If these external fields are present and relevant they will have to be added to the list
of mechanical parameters.
At last we are ready to turn to the dynamical variables for the helium-atoms-in-a-box system. These
are (mercifully) easy to describe: they are just the position and momentum of each particle, a total of 2N
3-dimensional vectors. A shorthand name for this information is point in phase space. Phase space is
an abstract space with one coordinate corresponding to each dynamical variable. Thus in our example the
coordinates are the positions and momenta for particle number 1, number 2,. . . number 9,. . . and number N :
(x1 , y1 , z1 , px,1 , py,1 , pz,1 , x2 , y2 , z2 , px,2 , py,2 , pz,2 , . . . x9 , y9 , z9 , px,9 , py,9 , pz,9 , . . . xN , yN , zN , px,N , py,N , pz,N ).
(2.2)
For a system of N particles, phase space is 6N -dimensional. A single point in phase space gives the position
and momentum of every particle in the system.
2.1.1
Time development of a classical system
Given a microscopic description such as either of the two above, what can we do with it? The time development of the system is represented by the motion of a point in phase space. That point will snake through
many points in the phase space but all of the points visited will have the same energy.2 Because many points
of the given energy will be visited, it is natural to ask whether, in fact, all phase space points corresponding
to a given energy will eventually be visited by a system started at any one of those points.
It is easy to find systems for which this statement is false, but all such examples seem to be in one way or
another atypical. For example, consider two or three or even many millions of non-interacting particles in a
2 For the earth-moon model, all the points visited will also have the same total momentum and angular momentum, but this
is not the case for the helium-in-a-smooth-box model.
14
hard-walled box, and start them all traveling straight up and down. They will travel straight up and down
forever. The points in phase space with identical energy, but with the particles traveling left and right, will
never be visited. This example is atypical because if the particles interacted, even slightly, then they would
fall out of the straight up and down regions of phase space.
Other stuff. Use name ergodic hypothesis. Recent progress through model systems and the name
chaos. Reference. Computer programs.
Problems
2.1 (Q) Three-body interactions: microscopic
Make up a problem involving two charged point particles and a polarizable atom.
2.2 Mechanical parameters and dynamical variables
Here is a classical mechanics problem: A pendulum bob of mass m swings at the end of a cord which
runs through a small hole in the ceiling. The cord is being pulled up through the hole so that its length
is `(t) = `0 t. At time t = 0 the bob is at rest and at angle 0 . Find the subsequent motion of the
bob. Do not solve this problem. Instead, list the mechanical parameters and the dynamical variables
that appear in it.
2.3 A Hamiltonian
Write down the total Hamiltonian of a neutral plasma of N protons and N electrons moving in a
rectangular box with interior dimensions of Lx Ly Lz , assuming that i) any proton or electron
interacts with the wall material through a potential energy function

1
1
W
2
for d < a
0
d2
a
W (d) =
(2.3)
0
otherwise,
where d is the (perpendicular) distance of the particle in question from the wall, and ii) the system is
subject to a uniform electric field in the x
direction of magnitude E. List the mechanical parameters
that appear in this Hamiltonian, and distinguish them from the dynamical variables.
2.4 (Q) For discussion: Mechanical parameters, dynamical variables, and modeling
List the mechanical parameters and dynamical variables of these systems:
a. Hydrogen molecules enclosed in a sphere.
b. Water molecules in a box.
c. A mixture of hydrogen molecules and helium atoms in a box.
To what extent are you making models as you generate descriptions? To what extent are you making
assumptions? (For example, by using non-relativistic classical mechanics.)
2.2. MACROSCOPIC DESCRIPTION OF A LARGE EQUILIBRIUM SYSTEM
2.2
15
Macroscopic Description of a Large Equilibrium System
The title of this section introduces two new terms. . . large and equilibrium. A system is large if it contains
many particles, but just how large is large enough? A few decades ago, physicists dealt with systems of
one or two or a dozen particles, or else they dealt with systems with about 1023 particles, and the distinction
was clear enough. Now that computer chips have such small lines, systems of intermediate size are coming
under investigation. (This field carries the romantic name of mesoscopic physics.) In practical terms, it is
usually easy enough to see when the term large applies, but there is no rigorous criterion and there have
been some surprises in this regard.
A similar situation holds for the term equilibrium. A system is said to be in equilibrium if its macroscopic
properties are not changing with time. Thus a cup of tea, recently stirred and with churning, swirling flow,
is not at equilibrium, whereas some time later the same cup, sedate and calm, is at equilibrium. But some
of the properties of the latter cup are changing with time: for example, the height of water molecule number
173 changes rapidly. Of course this is not a macroscopic property, but then there is no rigorous definition
of macroscopic vs. microscopic properties. Once again there is little difficulty in practice, but a rigorous
criterion is wanting and some excellent physicists have been fooled. (For example, a mixture of hydrogen
gas and oxygen gas can behave like a gas in equilibrium. But if a spark is introduced, these chemicals will
react to form water. The gas mixture is in equilibrium as far as its physical properties are concerned, but
not as far as its chemical properties are concerned.)
With these warnings past, we move on to the macroscopic description. As with the microscopic description, it has two parts: the first saying what the system is and the second saying what condition the system
is in. For definiteness, we consider the helium-in-a-smooth-box model already introduced.
To say what this system is, we again list mechanical parameters. At first these are the same as for the
microscopic description: the number of particles N , the mass of each particle m, and a description of the
atom-atom interaction potential (e.g. the Lennard-Jones parameters a and b). But we usually dont need
to specify the atom-wall interaction or even the location of the walls. . . instead we specify only the volume
of the container V . The reason for this is not hard to see. If we deal with a large system with typical
wall-atom interactions (short range) and typical walls (without intricate projections and recesses) then very
few particles will be interacting with the walls at any one instant, so we expect that the details of wallatom interaction and container shape will be affect only a tiny minority of atoms and hence be irrelevant to
macroscopic properties. There are of course exceptions: for example, in a layered substance such as graphite
or mica, or one of the high-temperature superconductors, the shape of the container might well be relevant.
But in this book we will not often deal with such materials.
Finally, what condition is the system in? (In other words, what corresponds to the dynamical variables
in a microscopic description?) Clearly the point in phase space, giving the positions and momenta of
each and every particle in the system, is far too precise to be an acceptable macroscopic description. But
at the same time it is not acceptable to say we dont care anything about microscopic quantities, because
the energy of this system is conserved, so energy will be a feature of both the microscopic and macroscopic
16
descriptions. In fact, for the helium-in-a-smooth-box model, energy is the only quantity that is conserved,3
so it is the only item in the list of macroscopic descriptors. (Other systems, such as the earth-moon system,
will have additional conserved quantities, and hence additional items in that list.)
What equilibrium is not

Theres a common misconception that at equilibrium, the sample is uniform. This might not be true: A
mixture of ice and water at atmospheric pressure and temperature 0 C is at equilibrium but is not uniform.
Or, consider a container of air 50 kilometers tall with its base on Earth at sea level: most of the molecules
huddle at the bottom of the container, and the top is near-vacuum.
2.5 (Q) Three-body interactions: macroscopic
Even though three-body interactions do exist, they can usually be ignored in a macroscopic description.
(Just as the wall-atom interaction does exist, but it can usually be ignored in a macroscopic description.)
Why?
2.6 (Q,E) Lost in space
A collection of N asteroids floats in space far from other gravitating bodies. Model each asteroid as
a hard sphere of radius R and mass m. What quantities are required for a microscopic description of
this system? For a macroscopic description?
2.3
Fundamental Assumption
Define microstate and macrostate.

There are many microstates corresponding to any given macrostate. The collection of all such microstates
is called an ensemble. (Just as a musical ensemble is a collection of performers.)
Note: An ensemble is a (conceptual) collection of macroscopic systems. It is not the collection of atoms
that makes up a macroscopic system.
(Terminology: A microstate is also called a configuration or a complexion. . . both poor terms. A
macrostate is also called a thermodynamic state. A corresponding microstate is sometimes called an
accessible or a consistent microstate.)
A system is said to be isolated if no energy goes in or out, and if the mechanical parameters (N ,
V , etc.) are also fixed. Most of the systems you deal with in classical mechanics classes, for example, are
isolated.
The fundamental assumption of statistical mechanics is:
3 The
box.
momentum and angular momentum of the helium atoms is not conserved, because there are external forces due to the
2.3. FUNDAMENTAL ASSUMPTION
17
An isolated system in an equilibrium macrostate has equal probability of being in any of the
microstates corresponding to that macrostate.
Conceptual difficulties:
1. What is equilibrium?
2. What is probability? (Any experiment is done on one system.)
3. May be false! (Can be motivated by the ergodic hypothesis, but this is just suggestive, and the ergodic
hypothesis itself has not been proven and may be false. Cellular automata may violate this assumption. . . but do physical/biological systems? See for example Andrew I. Adamatzky, Identification of
Cellular Automata (Taylor and Francis, 1995) and the essay review by Normand Mousseau, in Contemporary Physics, 37 (1996) 321323; Stephen Wolfram, A New Kind of Science (Wolfram Media,
2002).)
Practical difficulties.
1. Gives way to find average values of microscopic quantities, such as
hkinetic energyi, hpotential energyi, hheighti,
but not things like temperature and pressure.
2. Based on model. (In particular, infinitely hard smooth walls.)
3. For point particle model, requires infinitely thin sheet in phase space. Work instead with the volume
of phase space corresponding to energies from E to E + E, and at the end of the calculation take the
limit E 0.
To examine these practical difficulties in more detail, consider again the helium-in-a-smooth-box model.
Suppose we want to find the number of microstates with energy ranging from E to E + E, in a system
with N identical particles in a box of volume V .
Your first answer might be that there are an infinite number of points in phase space satisfying this
criterion. This is true but it misses the point: A thimble and a Mack truck both contain an infinite number
of points, but the Mack truck carries more because it has more volume. Clearly the microstate count we
desire is some sort of measure of phase-space volume. Ill call the region of phase space with energy ranging
from E to E + E by the name (E, E), and call its volume
Z
d.
(Volume in phase space.)
(2.4)
(E,E)
18
However, this volume isnt precisely what we want. Permutation argument for N identical particles.
(For a system with N identical particles, there are N ! points in phase space corresponding to the same
microstate.) A better measure of the microstate count is thus
Z
1
d.
(Delabeled volume in phase space.)
(2.5)
N ! (E,E)
There remains one more problem. We desire a count, which is dimensionless, and the above expression
gives a phase-space volume, which has the dimensions of
(angular momentum)3N .
The solution to this problem is straightforward if clunky. Pick a quantity with the dimensions of angular
momentum. Any quantity will do, so make an arbitrary choice. Call the quantity h0 . (Clearly, by the time
we produce a measurable physical result at the end of a calculation, it had better be independent of our
choice of h0 , just as the value of any measurable result in a classical mechanics problem has to be independent
of choice of the potential energy zero.) Then define the microstate count as the dimensionless quantity
Z
1 1
(E, E, V, N ) = 3N
d.
(Dimensionless, delabeled volume in phase space.)
(2.6)
h0 N ! (E,E)
In poker, there are 52 cards in a deck, so the probability of drawing any given card from a shuffled deck
is 1/52. The number 52 plays a fundamental role in answering any question concerning probability in poker.
In statistical mechanics, there are microstates corresponding to a macrostate, so the probability of
encountering any given microstate is (if the fundamental assumption is correct) 1/. The number plays
the same fundamental role in statistical mechanics that 52 does in poker.
2.7 Microstate count for a mixture
What expression corresponds to (2.6) for a collection of NHe Helium atoms and NAr Argon atoms?
2.4. STATISTICAL DEFINITION OF ENTROPY
2.4
19
Statistical Definition of Entropy
Logarithms and dimensions

You cannot take the logarithm of a number with dimensions.
Perhaps you have heard this rule phrased as you cant take the
logarithm of 3.5 meters or you cant take the logarithm of five
oranges. Why not? A simple argument is Well, what would be
the units of ln(3.5 meters)? A more elaborate argument follows.
The logarithm function is the inverse of the exponential function:
y = ln(x)
means the same as
x = ey .
But remember that

x = ey = 1 + y +
1
1 2
y + y3 + .
2!
3!
If y had the dimensions of length, then the expression above

would be a meaningless sum of 1 plus a length plus an area plus a
volume plus so forth.
You cannot exponentiate a number with dimensions, and you
cannot take the logarithm of a number with dimensions.
Additivity (see problem 2.10) . . . dont physically mix.
S(E, E, V, N ) = kB ln (E, E, V, N )
(2.7)
The constant kB in this definition is called the Boltzmann constant. It is clear from the argument that
the Boltzmann constant could have been chosen to have any value: 1, , whatever. For historical reasons, it
was chosen to have the value
kB = 1.38 1023 joule/kelvin.
(2.8)
There is no particular physical significance to this number: its logical role is analogous to 2.54 cm/inch or
4.186 joule/calorie. In other words, the origin of this number is not to be sought in nature, but in the history
of the definition of the joule and the kelvin. It is a conversion factor.
Problems
2.8 (E) Accessible regions of phase space
Suppose that N non-interacting particles, each of mass m, move freely in a one-dimensional box (i.e. an
infinite square well). Denote the position coordinates by x1 , x2 , . . . , xN and the momentum coordinates
by p1 , p2 , . . . , pN . The box restricts all the positions to fall between xi = 0 and xi = L. The energy of
the system lies between E and E + E.
20

a. If only one particle is present, draw the systems phase space and shade the regions of phase space
that are accessible.
b. If two particles are present then phase space is four dimensional, which makes it difficult to draw.
Draw separately the part of phase space involving positions and the part involving momenta.
Shade the accessible regions of phase space.
c. Suppose two particles are present, and consider the slice of phase space for which x1 = (2/3)L
and p2 equals some constant called p2 . Draw a (carefully labeled) sketch of this slice with the
accessible regions shaded.
d. Describe the accessible regions of phase space if N particles are present.
2.9 Accessible configurations of a spin system

Consider an isolated system of N spin- 12 atoms in a magnetic field H. The atoms are fixed at their
lattice sites and the spins do not interact. Each atom has a magnetic moment m that can point either
up (parallel to the field H) or down (antiparallel to H). A microstate (or configuration) of this
system is specified by giving the direction of every spin. An up spin has energy mH, a down spin
has energy +mH, so a configuration with n up spins and n down spins has energy
E = (n n )mH.
(2.9)
This system is called the ideal paramagnet.

a. Not every energy is possible for this model. What is the maximum possible energy? The minimum? What is the minimum possible non-zero energy difference between configurations?
b. Suppose we know that the system has n up spins and n down spins, but we do not know how
these spins are arranged. How many microstates are consistent with this knowledge?
c. Consider the energy range from E to E + E where E is small compared to N mH but large
compared to mH. What is the approximate number of states (E, E, H, N ) lying in this energy
range? Express your answer in a form that does not include the quantities n or n .
2.10 Microstates for a combined system
System #1 is in a macrostate with three corresponding microstates, labeled A, B, and C. System #2 is
in a macrostate with four corresponding microstates, labeled , , , and . How many microstates are
accessible to the combined system consisting of system #1 and system #2? List all such microstates.
2.11 (E) The logarithm
Suppose that a differentiable function satisfies
f (xy) = f (x) + f (y)
(2.10)
f (x) = k ln(x).
(2.11)
for all positive x and y. Show that

[Clues: 1) Take derivatives with respect to x and with respect to y. 2) If G(x) = G(y), x and y
independent variables, then the function G(x) must be constant. 3) Set y = 1 in equation (2.10).]
2.5. ENTROPY OF A MONATOMIC IDEAL GAS
2.5
21
Entropy of a Monatomic Ideal Gas
So far in this chapter, we have been dealing very abstractly with a very general class of physical systems.
We have made a number of assumptions that are reasonable but that we have not tested in practice. It
is time to put some flesh on these formal bones. We do so by using our statistical definition of entropy to
calculate the entropy of a monatomic ideal gas. (Here monatomic means that we approximate the atoms
by point particles, and ideal means that those particles do not interact with each other. In addition, we
assume that the gas contains only one chemical species and that classical mechanics provides an adequate
description. Thus a more precise name for our system would be the pure classical monatomic idea gas,
but in this case we wisely prefer brevity to precision.) Working with this concrete example will show us that
what we have said is sensible (at least for this system), and guide us in further general developments.
The previous pages have been remarkably free of equations for a physics book. Now is the time to remedy
that situation. Before studying this section, you need to know that the volume of a d-dimensional sphere is
Vd (r) =
d/2 d
r .
(d/2)!
(2.12)
If you dont already know this, then read appendix D, Volume of a Sphere in d Dimensions, before reading
this section. And if you dont know the meaning of x!, where x is a half-integer, then you should read
appendix C, Clinic on the Gamma Function, before reading appendix D. Finally, if you dont know
Stirlings approximation for the factorial function, namely
ln n! n ln n n
for n 1,
(2.13)
then you should also read appendix E, Stirlings Approximation, before reading further. (Do not be
discouraged by this long list of prerequisites. This mathematical material is quite interesting in its own right
and will be valuable throughout this book.)
We consider a system of N identical, classical, non-interacting point particles, each of mass m. The
kinetic energy of this system is
1 2
(p + p22 + + p2N ),
(2.14)
2m 1
and the potential energy is
(
0
if all particles are inside container
(2.15)
otherwise.
(One sometimes hears that the ideal gas has no potential energy. It is true that there is no potential
energy due to atom-atom interaction, but, as the above expression makes clear, there is indeed a potential
energy term due to atom-wall interaction. Because of the character we assume for that term, however, the
numerical value of the potential energy is always zero. Note also that the ideal gas is not the same as the
hard-sphere gas. In the hard-sphere model two atoms have infinite potential energy if they are separated
by a distance of twice the hard-sphere radius or less. In the ideal gas model two atoms do not interact. It
is permissible even for two atoms to occupy the same location. . . only in the model, of course!)
22
Now that the system is completely specified, it is time to begin the problem. We wish to calculate the
entropy
W (E, E, V, N )
S(E, E, V, N ) = kB ln
,
(2.16)
N !h3N
0
where the function W represents the volume in phase space corresponding to energies from E to E + E
(i.e. the volume of the region (E, E)). Before jumping into this (or any other) problem, it is a good idea
to list a few properties that we expect the solution will have. . . this list might guide us in performing the
calculation; it will certainly allow us to check the answer against the list to see if either our mathematics or
our expectations need revision. We expect that:
We will be able to take the limit as E 0 and get sensible results.
The entropy S will depend on only the volume of the container and not on its shape.
If we double the size of the system, by doubling E, V , and N , then we will double S. (Additivity.)
S will depend on h0 in a trivial, sea-level fashion.
The formal expression for the volume of the accessible region of phase space is
W (E, E, V, N )
accessible volume in phase space

Z
=
d
(E,E)
Z
Z
Z
Z
Z
Z
=
dx1 dy1 dz1 dxN dyN dzN
Z
Z
Z
Z
Z
Z
dpx,1 dpy,1 dpz,1 dpx,N dpy,N dpz,N .
(2.17)
(2.18)
The complexity of this integral rests entirely in the complexity of the shape of (E, E) rather than in the
complexity of the integrand, which is just 1. Fortunately the integral factorizes easily into a position part
and a momentum part, and the position part factorizes into a product of integrals for each particle (see
problem 2.8). For, say, particle number 5, if the particle is inside the container it contributes 0 to the energy,
so the total energy might fall between E and E + E (depending on other factors). But if it is outside the
container, then it contributes to the energy, which always exceeds the limit E + E. Thus the integral
is just the volume of the container:
Z
Z
Z
dx5
dy5
dz5 = V.
(2.19)
This integral depends on the volume V but is independent of the shape of the container. We will soon
see that, as a consequence, the entropy depends on volume but not shape (which is in accord with our
expectations).
The integrals over momentum space do not factorize, so we must consider the entire 3N -dimensional
momentum space rather than N separate 3-dimensional spaces. We know that the total potential energy is
zero (unless it is infinite), so the energy restriction is taken up entirely by the kinetic energy. Equation (2.14)
tells us that the momentum space points with energy E fall on the surface of a sphere of radius 2mE.
23
Thus the accessible region in momentum space is a shell with inner radius 2mE and with outer radius
p
2m(E + E). (Notice that we are counting all the microstates within the accessible region of phase space,
not just typical microstates there. For example, one microstate to be counted has all of the particles at
rest, except for one particle that has all the energy of the system and is heading due west. This is to be
counted just as seriously as is the microstate in which the energy is divided up with precise equality among
the several particles, and they are traveling in diverse specified directions. Indeed, the system has exactly
the same probability of being in either of these two microstates.) Using equation (2.12) for the volume of a
3N -dimensional sphere, the volume of that shell is
i
3N/2 h
(2m(E + E))3N/2 (2mE)3N/2 .
(3N/2)!
(2.20)
I prefer to write this result in a form with all the dimensionfull quantities lumped together, namely as
"
#
3N/2
3N/2
E
3N/2
(2mE)
1+
1 .
(2.21)
(3N/2)!
E
The quantity in square brackets is dimensionless.
To find the accessible volume of the entire phase space, we multiply the above result by V N , the result
of performing N separate position integrals. Thus
"
#
3N/2
(2mEV 2/3 )3N/2
E
W (E, E, V, N ) =
1+
1 .
(2.22)
(3N/2)!
E
As promised, W depends upon the variables E, E, V , and N . It also depends upon the unmentioned
mechanical parameter m. The arguments on page 15 have been vindicated. . . the phase space volume W
depends only upon V and not upon the detailed shape of the container.
At last we can find the entropy! It is
"
#)
(
3N/2
3N/2
W
1
E
2mEV 2/3
S = kB ln
1+
1
= kB ln
h20
N !(3N/2)!
E
N !h3N
0
or
S
= 32 N ln
kB
2mEV 2/3
h20

ln N !
ln( 23 N )!
"
+ ln
E
1+
E
3N/2
(2.23)
#
1 .
(2.24)
How does this expression compare to our list of expectations on page 22?
If we take the limit E 0, the entropy approaches , contrary to expectations.
The entropy S depends on the volume of the container but not on its shape, in accord with expectations.
If we double E, V , and N , then S will not exactly double, contrary to expectations.
The entropy S does depend on h0 in a sea-level fashion, in accord with expectations.
24
Only two of our four expectations have been satisfied. (And it was obvious even from equation (2.16) that
the fourth expectation would be correct.) How could we have gone so far wrong?
The trouble with expression (2.24) for the entropy of a monatomic ideal gas is that it attempts to hold
for systems of any size. In justifying the definition of entropy (2.7) (and in writing the list of expectations
on page 22) we relied upon the assumption of a large system, but in deriving expression (2.24) we never
made use of that assumption. On the strength of this revised analysis we realize that our expectations will
hold only approximately for finite systems: they will hold to higher and higher accuracy for larger and larger
systems, but they will hold exactly only for infinite systems.
There is, of course, a real problem in examining an infinite system. The number of particles is infinite, as
is the volume, the energy, and of course the entropy too. Why do we need an equation for the entropy when
we already know that its infinite? Once the problem is stated, the solution is clear: We need an expression
not for the total entropy S, but for the entropy per particle s = S/N . More formally, we want to examine
the system in the thermodynamic limit, in which
V
E
E
e,
v, and
e.
N
N
N
In this limit we expect that the entropy will grow linearly with system size, i.e. that
N in such a way that
S(E, E, V, N ) N s(e, v, e).
(2.25)
(2.26)
The quantities written in lower case, such as e, the energy per particle, and v, the volume per particle, play
the same role in statistical mechanics as per capita quantities do in demographics. (The gross national
product of the United States is much larger than the gross national product of Kuwait, but that is just
because the United States is much larger than Kuwait. The GNP per capita is higher in Kuwait than in the
United States.)
Lets take the thermodynamic limit of expression (2.24) (the entropy of a finite system) to find the
entropy per particle of an infinite monatomic ideal gas. The first thing to do, in preparing to take the
thermodynamic limit, is to write V as vN , E as eN , and E as e N so that the only size-dependent
variable is N . This results in
"
#

3N/2
2/3 5/3
S
2mev
N
e
ln N ! ln( 23 N )! + ln 1 +
= 32 N ln
1 .
(2.27)
kB
h20
e
Next we use Stirlings approximation,
ln n! n ln n n
for n 1,
(2.28)
to simplify the expressions like ln( 23 N )! above. Thus for large values of N we have approximately (an
approximation that becomes exact as N )
"
#

3N/2
S
2mev 2/3 N 5/3
e
3
3
3
3
2 N ln
N ln N + N ( 2 N ) ln( 2 N ) + 2 N + ln 1 +
1
kB
h20
e
"
#

3N/2
2/3
2mev
e
= 23 N ln
+ 32 N ln N 5/3 N ln N + N 23 N ln( 32 N ) + 32 N + ln 1 +
1 .
h20
e
25
The first term on the right increases linearly with N . The next bunch of terms is
=
=
3
2N
5
2N
ln N 5/3 N ln N + N 32 N ln( 32 N ) + 32 N
ln N N ln N + N 32 N ln( 32 ) 23 N ln(N ) + 32 N

N 52 32 ln( 23 ) ,
which again increases linearly with N . The final term is, as N grows,
"
#
3N/2
3N/2

e
e
ln 1 +
1
ln 1 +
e
e

e
3
.
= 2 N ln 1 +
e
(2.29)
(2.30)
This term not only increases linearly with N in the thermodynamic limit, it also vanishes as e 0! (This
is a general principle: One must first take the thermodynamic limit N , and only then take the thin
phase space limit e E/N 0.)
Our expectation (2.26) that in the thermodynamic limit the entropy would be proportional to the system
size N has been fully vindicated. So has the expectation that we could let E 0, although we have seen
that we must do so carefully. The end result is that the entropy per particle of the pure classical monatomic
ideal gas is

4mev 2/3
3
5
s(e, v) = kB 2 ln
(2.31)
+2 .
3h20
This is called the Sackur-Tetrode formula.4 It is often written as

4mEV 2/3
3
5
S(E, V, N ) = kB N 2 ln
+2 ,
3h20 N 5/3
(2.32)
with the understanding that it should be applied only to very large systems, i.e. to systems effectively at the
thermodynamic limit.
Problems
2.12 Entropy of a spin system
Consider again the ideal paramagnet of problem 2.9.
a. Write down an expression for ln (E, E, H, N ) as a function of E. Simplify it using Stirlings
approximation for large values of N . (Clue: Be careful to never take the logarithm of a number
with dimensions.)
b. Find an expression for the entropy per spin s(e, H) as a function of the energy per spin e and
magnetic field H in the thermodynamic limit.
4 Otto Sackur (18801914) was a German physical chemist. Hugo Tetrode (18951931) was a Dutch theoretical physicist.
Each independently uncovered this equation in 1912.
26

c. Sketch the resulting entropy as a function of the dimensionless quantity u e/mH. Does it take
on the proper limits as e mH? (In your sketch pay special attention to endpoints of the
domain, and to any places where the function or its derivative suffers a discontinuity, or a kink, or
goes to zero or infinity. In general, it is important that the sketch convey a correct impression of
the qualitative character of the function, and less important that it be quantitatively accurate.)
2.13 The approach to the thermodynamic limit

For the classical monatomic ideal gas, plot entropy as a function of particle number using both the
finite size form (2.24) and the Sackur-Tetrode form (2.32). We will see in problem 4.11 that for a
gas at room temperature and atmospheric pressure, it is appropriate to use
EV 2/3 /h20 = (1.66 1029 kg1 )N 5/3 .
(2.33)
Use the masses of argon and krypton. All other things being equal, is the thermodynamic limit
approached more rapidly for atoms of high mass or for atoms of low mass?
2.14 Other energy conventions
In the text we found the entropy of a monatomic ideal gas by assuming that the potential energy of an
atom was zero if the atom were inside the box and infinite if it were outside the box. What happens
if we choose a different conventional zero of potential energy so that the potential energy is U for an
atom inside the box and infinite for an atom outside the box?
2.15 Other worlds
Find the entropy as a function of E, V , and N in the thermodynamic limit for a monatomic ideal gas
in a world with arbitrary5 spatial dimensionality d.
2.16 Ideal gas mixtures
Consider a large sample of classical monatomic ideal gas that is a mixture of two components: NA
particles of mass mA and NB particles of mass mB . If N NA + NB , show that the entropy is

4mA EV 2/3
5
+
S(E, V, NA , NB ) = +kB NA 23 ln
2
3h20 N 5/3

4mB EV 2/3
5
+kB NB 32 ln
(2.34)
+
2
3h20 N 5/3

NA
NA
NB
NB
kB N
ln
+
ln
.
N
N
N
N
(Clue: Use the result of problem D.2.)
5 Why, you wonder, should anyone care about a world that is not three-dimensional? For three reasons: (1) There are important physical approximations of two-dimensional worlds (namely surfaces) and of one-dimensional worlds (namely polymers).
(2) The more general formulation might help you in unexpected ways. For example, Ken Wilson and Michael Fisher were trying
to understand an important problem concerning critical points. They found that their technique could not solve the problem in
three dimensions, but it could solve the problem in four dimensions. Then they figured out how to use perturbation theory to
slide carefully from four dimensions to three dimensions, thus making their solution relevant to real physical problems. Wilson
was awarded the Nobel Prize for this work. This illustrates the third reason, namely you never can tell what will be important
and hence: (3) Knowledge is better than ignorance.
2.6. QUALITATIVE FEATURES OF ENTROPY
2.6
27
Qualitative Features of Entropy
The concept of entropy is notoriously difficult to grasp. Even the consummate mathematician and physicist
Johnny von Neumann claimed that nobody really knows what entropy is anyway. Although we have an
exact and remarkably simple formula for the entropy of a macrostate in terms of the number of corresponding
microstates, this simplicity merely hides the subtle characterization needed for a real understanding of the
entropy concept. To gain that understanding, we must examine the truly wonderful (in the original meaning
of that overused word) surprises that this simple formula presents when applied to real physical systems.
2.6.1
Surprises
The monatomic ideal gas

Let us examine the Sackur-Tetrode formula (2.32) qualitatively to see whether it agrees with our understanding of entropy as proportional to the number of microstates corresponding to a given macrostate. If
the volume V is increased, then the formula states that the entropy S increases, which certainly seems
reasonable: If the volume goes up, then each particle has more places where it can be, so the entropy ought
to increase. If the energy E is increased, then S increases, which again seems reasonable: If there is more
energy around, then there will be more different ways to split it up and share it among the particles, so
we expect the entropy to increase. (Just as there are many more ways to distribute a large booty among a
certain number of pirates than there are to distribute a small booty.) But what if the mass m of each particle increases? (Experimentally, one could compare the entropy of, say, argon and krypton under identical
conditions. See problem 2.27.) Our formula shows that entropy increases with mass, but is there any way
to understand this qualitatively?
In fact, I can produce not just one but two qualitative arguments concerning the dependence of S on m.
Unfortunately the two arguments give opposite results! The first relies upon the fact that
E=
1 X 2
p ,
2m i i
(2.35)
so for a given energy E, any individual particle may have a momentum ranging from 0 to 2mE. A larger
mass implies a wider range of possible momenta, which suggests more microstates and a greater entropy.
The second argument relies upon the fact that
mX 2
E=
v ,
(2.36)
2 i i
p
so for a given energy E, any individual particle may have a speed ranging from 0 to 2E/m. A larger mass
implies a narrowed range of possible speeds, which suggests fewer microstates and a smaller entropy. The
moral is simple: Qualitative arguments can backfire!
28

Thats the moral of the paradox. The resolution of the paradox is both deeper and more subtle:
It hinges on the fact that the proper home of statistical mechanics is phase space, not configuration space, because Liouvilles theorem implies conservation of volume in phase space, not in
configuration space. This issue deeply worried the founders of statistical mechanics. See Ludwig
Boltzmann, Vorlesungen u
ber Gastheorie (J.A. Barth, Leipzig, 189698), part II, chapters III
and VII [translated into English by Stephen G. Brush: Lectures on Gas Theory (University of
California Press, Berkeley, 1964)]; J. Willard Gibbs, Elementary Principles in Statistical Mechanics (C. Scribners Sons, New York, 1902), page 3; and Richard C. Tolman, The Principles
of Statistical Mechanics (Oxford University Press, Oxford, U.K., 1938), pages 45, 5152.
Freezing water
It is common to hear entropy associated with disorder, smoothness, or homogeneity. How do these
associations stand up to the simple situation of a bowl of liquid water placed into a freezer? Initially the
water is smooth and homogeneous. As its temperature falls, the sample remains homogeneous until the
freezing point is reached. At the freezing temperature the sample is an inhomogeneous mixture of ice and
liquid water until all the liquid freezes. Then the sample is homogeneous again as the temperature continues
to fall. Thus the sample has passed from homogeneous to inhomogeneous to homogeneous, yet all the while
its entropy has decreased. (We will see later that the entropy of a sample always decreases as its temperature
falls.)
Suppose the ice is then cracked out of its bowl to make slivers, which are placed back into the bowl
and allowed to rest at room temperature until they melt. The jumble of irregular ice slivers certainly
seems disordered relative to the homogeneous bowl of meltwater, yet it is the ice slivers that have the
lower entropy. The moral here is that the huge number of microscopic degrees of freedom in the meltwater
completely overshadow the minute number of macroscopic degrees of freedom in the jumbled ice slivers. But
the analogies of entropy to disorder or smoothness invite us to ignore this moral and concentrate on the
systems gross appearance and nearly irrelevant macroscopic features.
Reentrant phases
When the temperature falls at constant pressure, most pure materials pass from gas to liquid to solid. But
the unusual materials called liquid crystals, which consist of rod-like molecules, display a larger number of
phases. For typical liquid crystals, the high-temperature liquid phase is isotropic, meaning that the positions
and the orientations of the molecules are scattered about nearly at random. At lower temperatures, the
substance undergoes a transition into the so-called nematic phase, in which the molecules tend to orient
in the same direction but in which positions are still scattered. At still lower temperatures it passes into
the smectic phase, in which the molecules orient in the same direction and their positions tend to fall into
planes. Finally, at even lower temperatures, the molecules freeze into a conventional solid. The story told
so far reinforces the picture of entropy as disorder, with lower-temperature (hence lower entropy) phases
showing more and more qualitative order.
29
Figure 2.2: Phase diagram of a liquid crystal mixture. The variable composition refers to the molecular
weight ratio of 6OCB to 8OCB.
But not all liquid crystals behave in exactly this fashion. One material called hexyloxy-cyanobiphenyl or
6OCB passes from isotropic liquid to nematic to smectic and then back to nematic again as the temperature
is lowered. The first transition suggests that the nematic phase is less orderly than the smectic phase,
while the second transition suggests the opposite!
One might argue that the lower-temperature nematic phase the so-called reentrant nematic
is somehow qualitatively different in character from the higher-temperature nematic, but the experiments
summarized in figure 2.2 demonstrate that this is not the case. These experiments involve a similar liquid
crystal material called octyloxy-cyanobiphenyl or 8OCB which has no smectic phase at all. Adding a
bit of 8OCB into a sample of 6OCB reduces the temperature range over which the smectic phase exists.
Adding a bit more reduces that range further. Finally, addition of enough 8OCB makes the smectic phase
disappear altogether. The implication of figure 2.2 is clear: there is no qualitative difference between the
usual nematic and the reentrant nematic phases you can move continuously from one to the other in the
temperaturecomposition phase diagram.
The implication of reentrant phases for entropy is profound: Under some conditions the nematic phase
has more entropy than the smectic phase and under other conditions less, while in all cases the nematic is
qualitatively less ordered.
Another example of reentrant behavior appears in the phase diagram of the mixture of water and nicotine (see figure 2.3). For a wide range of mixing ratios, this mixture is a homogeneous solution at high
30
Figure 2.3: Phase diagram for a mixture of water and nicotine.
temperatures, segregates into water-rich and nicotine-rich phases at moderate temperatures, yet becomes
homogeneous again at low temperatures. At mixing ratios closer to pure water or pure nicotine, the mixture
is homogeneous at all temperatures. Thus the high-temperature and reentrant homogeneous phases are in
fact the same phase. (Reentrant phases are also encountered in type-II superconductors.)
Rust
According to a common misconception, The entropy law describes the tendency for all objects to rust,
break, fall apart, wear out, and otherwise move to a less ordered state.6 Or Entropy tends to make our
eyes grow weaker as we age . . . entropy causes iron and other metals to rust.7 Or Entropy imposes itself in
the form of limitation and in the natural tendency toward confusion and chaos, toward decay and destruction.
Entropy is why ice cream melts and people die. It is why cars rust and forests burn. Preponderantly, the
second law describes effects that we dont want to see.8
At least one error is held in common by these three quotes. Does rusting lead to increased entropy?
The rust reaction is
4 Fe + 3 O2 2 Fe2 O3 .
9
According to standard tables the entropy (at room temperature 298.15 K and at pressure 105 Pa) of one
mole of Fe is 27.280 J/K, of one mole of O2 is 205.147 J/K, and of one mole of Fe2 O3 is 87.404 J/K. The
6 Cutler J. Cleveland and Robert Kaufmann Fundamental principles of energy, in Encyclopedia of Earth, last updated 29
August 2008, http://www.eoearth.org/article/Fundamental principles of energy.
7 Louis M. Savary, Teilhard de Chardin: The Divine Milieu Explained (Paulist Press, Mahwah, NJ, 2007) page 163.
8 Gilbert L. Wedekind, Spiritual Entropy (Xulon Press, Fairfax, VA, 2003) page 68.
9 Ihsan Barin, Thermochemical Data of Pure Substances (VCH Publishers, Weinheim, Germany, 1995) pages 675, 1239, and
702.
31
entropy of the reactants is 724.561 J/K, the entropy of the products is 174.808 J/K, so the reaction results
in an entropy decrease of 549.753 J/K.
It is easy to understand why this should be so: gases typically have much greater entropies that solids.
And of course this doesnt mean that during rusting, the entropy of the universe decreases: although the entropy of the iron plus oxygen decreases, the entropy of the surroundings increases by even more. Nevertheless,
it is clear that rusting itself involves a decrease in entropy, not an increase.
2.6.2
Entropy and the lattice gas
The four examples above should caution us about relying on qualitative arguments concerning entropy. Here
is another situation10 to challenge your intuition: The next page shows two configurations of 132 = 169
squares tossed down on an area that has 35 35 = 1225 empty spaces, each of which could hold a square.
(This system is called the lattice gas model.) The two configurations were produced by two different
computer programs (Toss1 and Toss2 ) that used different rules to position the squares. (The rules will be
presented in due course; for the moment I shall reveal only that both rules employ random numbers.) Which
configuration do you think has the greater entropy? Be sure look at the configurations, ponder, and make a
guess (no matter how ill-informed) before reading on.
10 The
argument of this section was invented by Edward M. Purcell and is summarized in Stephen Jay Gould, Bully for
Brontosaurus (W.W. Norton, New York, 1991), pages 265268, 260261. The computer programs mentioned, which work
under MS-DOS, are available for free downloading through http://www.oberlin.edu/physics/dstyer/.
32
Figure 2.4: A lattice gas configuration generated by the program Toss1.
Figure 2.5: A lattice gas configuration generated by the program Toss2.

Before analyzing these pictures, I must first confess that my question was very misleading. I asked Which
configuration has the greater entropy?, but entropy is not defined in terms of a single configuration (a single
microstate). Instead, entropy is defined for a macrostate, and is related to the number of microstates that
the system can take on and still be classified in that same macrostate. Instead of asking the question I
did, I should have pointed out that I had two classes, two pools, of microstates, and that I used my two
computer programs to select one member from each of those two pools. The selection was done at random,
so the selected configuration should be considered typical. Thus my question should have been Which
configuration was drawn from the larger pool?, or Which configuration is typical of the larger class?.
Given this corrected question, you might want to go back and look at the configurations again.
33
I have asked these questions of a number of people (both students and professionals) and most of them
guess that the lower configuration (the output of program Toss2), is typical of the class with larger entropy.
The lower configuration is smoother, less clumpy. They look at the upper configuration and see patterns,
which suggests some orderly process for producing that configuration, while the smoothness of the lower
configuration suggests a random, disorderly construction process.
Let me end the suspense and tell you how my computer programs produced the two configurations.
The top configuration was produced by tossing 169 = (13)2 squares down at random onto an area with
35 35 = 1225 locations, subject only to the rule that two squares could not fall on the same location. The
bottom configuration was produced in exactly the same manner except that there was an additional rule,
namely that two squares could not fall on adjacent locations either. Thus the top configuration was drawn
from the pool of all patterns with 169 squares, while the bottom configuration was drawn from the much
smaller pool of patterns with 169 squares and with no two squares adjacent. The top configuration is typical
of the class with more configurations and hence greater entropy.
Look again at the bottom configuration. You will notice that there are no squares immediately adjacent
to any given square. This nearest neighbor exclusion rule acts to spread out the squares, giving rise to
the smooth appearance that tricks so many into guessing that the bottom configuration is typical of a class
with high entropy.
Now look again at the top configuration. You will notice holes and clumps of squares, the inhomogeneities
that lead many to guess that it is typical of a small class. But in fact one should expect a random configuration
to have holes only a very exceptional configuration is perfectly smooth.11 This involves the distinction
between a typical configuration and an average configuration. Typical configurations have holes: some have
holes in the upper right, some in the middle left, some in the very center. Because the holes fall in various
locations, the average configuration the one produced by adding all the configurations and dividing by the
number of configurations is smooth. The average configuration is actually atypical. (Analogy: A typical
person is not of average height. A typical person is somewhat taller or somewhat shorter than average,
and very few people are exactly of average height. Any clothing manufacturer that produced only shirts
of average size would quickly go bankrupt.) The presence of holes or clumps, therefore, need not be an
indication of a pattern or of a design. However, we humans tend to find patterns wherever we look, even
when no design is present. In just this way the ancient Greeks looked into the nighttime sky, with stars
sprinkled about at random, and saw the animals, gods, and heroes that became our constellations.
11 In
his book The Second Law (Scientific American Books, New York, 1984), P.W. Atkins promotes the idea that entropy is
a measure of homogeneity. (This despite the everyday observation of two-phase coexistence.) To buttress this argument, the
book presents six illustrations (on pages 54, 72, 74, 75, and 77) of equilibrium lattice gas configurations. Each configuration
has 100 occupied sites on a 40 40 grid. If the occupied sites had been selected at random, then the probability of any site
being occupied would be 100/1600, and the probability of any given pair of sites both being occupied would be 1/(16)2 . The
array contains 2 39 39 adjacent site pairs, so the expected number of occupied adjacent pairs would be 2(39/16)2 = 11.88.
The actual numbers of occupied nearest-neighbor pairs in the six illustrations are 0, 7, 3, 7, 4, and 3. A similar calculation
shows that the expected number of empty rows or columns in a randomly occupied array is (15/16)10 2 40 = 41.96. The
actual numbers for the six illustrations are 28, 5, 5, 4, 4, and 0. I am confident that the sites in these illustrations were not
occupied at random, but rather to give the impression of uniformity.
34
2.6.3
Entropy and poker
An excellent illustration of the nature of entropy is given by the card game poker. There are many possible
hands in poker, some valuable and most less so. For example, the hand
A, K, Q, J, 10
(2.37)
is an example of a royal flush, the most powerful hand in poker. There are only four royal flushes (the royal
flush of hearts, of diamonds, of spades, and of clubs) and any poker player who has ever been dealt a royal
flush will remember it for the rest of his life.
By contrast, no one can remember whether he has been dealt the hand
4, 3, J, 2, 7
(2.38)
because this hand is a member of an enormous class of not-particularly-valuable poker hands. But the
probability of being dealt this hand is exactly the same as the probability of being dealt the royal flush of
hearts. The reason that one hand is memorable and the other is not has nothing to do with the rarity of
that particular hand, it has everything to do with the size of the class of which the hand is a member.
This illustration of the importance of class rather than individual configuration is so powerful and so
graphic that I shall call that distinction the poker paradox.
2.6.4
Conclusion
It is often said that entropy is a measure of the disorder of a system. This qualitative concept has at least
three failings: First, it is vague. There is no precise definition of disorder. Some find the abstract paintings
of Jackson Pollock to be disorderly; others find them pregnant with structure. Second, it uses an emotionally
charged word. Most of us have feelings about disorder (either for it or against it), and the analogy encourages
us to transfer that like or dislike from disorder, where our feelings are appropriate, to entropy, where they
are not. The most important failing, however, is that the analogy between entropy and disorder invites us
to think about a single configuration rather than a class of configurations. In the lattice gas model there
are many orderly configurations (such as the checkerboard pattern of figure 2.6) that are members of both
classes. There are many other orderly configurations (such as the solid block pattern of figure 2.7) that
are members only of the larger (higher entropy!) class.12 The poker hand
2, 4, 6, 8, 10
is very orderly, but a member of a very large class of nearly worthless poker hands.
12 Someone
might raise the objection: Yes, but how many configurations would you have to draw from the pool, on average,
before you obtained exactly the special configuration of figure 2.7? The answer is, Precisely the same number that you would
need to draw, on average, before you obtained exactly the special configuration of figure 2.4. These two configurations are
equally special and equally rare.
35
Figure 2.6: An orderly lattice gas configuration that is a member of both the large class and the small class
of configurations.
Figure 2.7: An orderly lattice gas configuration that is a member of only the large (high entropy!) class of
configurations.
Given the clear need for an intuition concerning entropy, and the appealing but unsatisfactory character
of the simile entropy as disorder, what is to be done? I suggest an additional simile, namely entropy as
freedom, which should be used not by itself but in conjunction with entropy as disorder.
Freedom means a range of possible actions, while entropy means a range of possible microstates.
If only one microstate corresponds to a certain macrostate, then the system has no freedom to choose its
microstate and it has zero entropy. If you are free to manage your own bedroom, then you may keep it
36
either neat or messy, just as high entropy macrostates encompass both orderly and disorderly microstates.
The entropy gives the number of ways that the constituents of a system can be arranged and still be a
member of the club (or class). If the class entropy is high, then there are a number of different ways to
satisfy the class membership criteria. If the class entropy is low, then that class is very demanding very
restrictive about which microstates it will admit as members. In short, the advantage of the entropy
as freedom analogy is that it focuses attention on the variety of microstates corresponding to a macrostate
whereas the entropy as disorder analogy invites focus on a single microstate.
While entropy as freedom has these benefits, it also has two of the drawbacks of entropy as disorder.
First, the term freedom is laden with even more emotional baggage than the term disorder. Second,
it is even more vague: political movements from the far right through the center to the extreme left all
characterize themselves as freedom fighters. Is there any way to reap the benefits of this analogy without
sinking into the mire of drawbacks?
For maximum advantage, I suggest using both of these analogies. The emotions and vaguenesses attached
to freedom are very different from those attached to disorder, so using them together tends to cancel
out the emotion. A simple sentence like For macrostates of high entropy, the system has the freedom to
chose one of a large number of microstates, and the bulk of such microstates are microscopically disordered
directs attention away from glib emotional baggage and toward the perspective of more entropy means more
microstates.
But perhaps it is even better to avoid analogies all together. Just think of entropy as describing the
number of microstates consistent with the prescribed macrostate. And then you will remember always that
entropy applies to a macrostate, i.e. a class of microstates, rather than to an individual microstate.
Problems
2.17 Entropy of the classical monatomic ideal gas: limits
a. Every macrostate has at least one corresponding microstate. Use this to show that in all cases
S 0.
b. Use the Sackur-Tetrode formula (2.31) to find the entropy of the classical monatomic ideal gas in
the limit of high density (i.e. volume per particle approaches zero). [This absurd result indicates
a breakdown in the approximation of ideal non-interacting particles. At high densities, as the
particles crowd together, it is no longer possible to ignore the repulsive interactions which are
responsible for the size of each atom.]
c. What is the entropy of the classical monatomic ideal gas when the energy equals the ground-state
energy (i.e. E 0)? [This absurd result indicates a breakdown in the classical approximation.
As we will see in chapter 6, quantum mechanics becomes a significant effect at low energies.
Historically, this absurdity was one of the first indications that classical mechanics could not be
universally true.]
2.7. USING ENTROPY TO FIND (DEFINE) TEMPERATURE AND PRESSURE
37
2.18 (Q,E) The coin toss

If you flip a coin ten times, you expect on average to get five heads and five tails.
a. The pattern HHHHHHHHHH violates this expectation dramatically. What is the probability of obtaining this pattern?
b. The pattern HTHTHTHTHT matches this expectation exactly. What is the probability of obtaining
this pattern?
c. What is the probability of obtaining the pattern HTTTHHTTHT?
d. What is the probability of obtaining a pattern with one tail and nine heads?
2.19 (Q) Random walks
You are presented with two pictures that supposedly sketch the progress of a random walker as it
moves through space. One was drawn by a computer using a random number generator, the other by
a human being. One picture is more or less uniform, the other has some spaces dark with lines and
other spaces hardly touched. Which picture was drawn in which way?
2.20 Poker
In the game of poker, a hand consists of five cards drawn from a deck of 52 cards. The cards are evenly
divided into four suits (hearts, diamonds, clubs, spades).
a. How many hands are there in poker? Explain your reasoning.
b. How many hands are flushes? (That is, all five cards of the same suit.)
c. Harry plays poker every evening, and each evening he is dealt 100 hands. How long must Harry
play, on average, before being dealt the hand
4, 3, J, 2, 7
2.7
(2.39)
Using Entropy to Find (Define) Temperature and Pressure
References: Schroeder sections 3.1, 3.4, and 3.5. Reif section 3.3. Kittel and Kroemer pages 3041.
You will remember that one of the problems with our ensemble approach to statistical mechanics is
that, while one readily sees how to calculate ensemble average values of microscopic (mechanical) quantities
like the kinetic energy, the potential energy, or the mean particle height, it is hard to see how to calculate
the sort of macroscopic (thermodynamic) quantity that were really more interested in anyway, such as the
temperature or pressure. In fact, it is not even clear how we should define these important quantities. Our
task in this lecture is to motivate such definitions. Because our arguments are motivations rather than
deductions, they will be suggestive rather than definitive. The arguments can be made substantially more
solid, but only at the cost of dramatic increase in mathematical complexity. (See David Ruelles Statistical
Mechanics: Rigorous Results [Physics QC174.8.R84 1989].) At this stage in your education, you should just
accept the arguments and the definitions, realizing that they are not definitive, and go on to learn how to
38
use them. If youre still worried about them a year from now, then that would be a good time to read the
more complex arguments in Ruelles book.
An analogy will help explain the mathematical level of this lecture. Why is the period of the simple
harmonic oscillator independent of amplitude? I can give two answers: 1) At larger amplitudes, the particle
has more distance to move, but it also moves faster, and these two effects exactly cancel. 2) Solve the
p
problem mathematically, and youll see that = k/m which is clearly unaffected by amplitude. The first
answer gives more insight into what is going on physically, but it is not definitive. (It does not explain why
the two effects cancel exactly rather that have one dominate the other, which is the case for non-parabolic
oscillators.) The second argument is bulletproof, but it gives little insight. It is best to use both types of
argument in tandem. Unfortunately, the bulletproof arguments for this section are really too elaborate and
formal to present here. (In a few weeks I will give you a sample of a rigorous statistical mechanical argument:
The Grand Canonical Ensemble in the Thermodynamic Limit, section 4.8. It is the easiest such argument
that I know of, and you will see that its not very easy at all!)
2.7.1
Rapidly increasing character of the (E) function
For a monatomic ideal gas (in the thermodynamic limit) the entropy is

3
4mEV 2/3
5
S(E, V, N ) kB ln (E, V, N ) = kB N
ln
.
+
2
2
3h20 N 5/3
(2.40)
Thus (E, V, N ), the number of microstates consistent with the macrostate (or thermodynamic state) specified by E, V , and N , is

(3/2)N
4mEV 2/3
(5/2)N
(E, V, N ) = e
.
(2.41)
3h20 N 5/3
In particular, as a function of energy,
(E) = (constant)E (3/2)N .
(2.42)
This is a very rapidly increasing function of E. (The function f (x) = x2 increases more rapidly than
23
f (x) = x. The function f (x) = x3 increases more rapidly than f (x) = x2 . The function f (x) = x10
increases very rapidly indeed!) Although we have proven this result only for the ideal gas, it is plausible13
for interacting systems as well: If you throw more energy into a system, there are more ways for the energy
to be arranged. (There is an argument to this effect in Reifs section 2.5. The result is not, however, always
true. . . check out the ideal paramagnet.)
As the particle number N grows larger and larger (approaching the thermodynamic limit) this increase
of with E becomes more and more rapid.
13 Through
the booty argument of page 27.
2.7.2
39
Two isolated systems
Consider two systems, A and B, enclosed in our perfectly reflecting no-skid walls. System A has energy EA ,
volume VA , and particle number NA ; similarly for system B.
A
EA, VA, NA
B
EB, VB, NB
I have drawn the two systems adjacent, but they might as well be miles apart because they cant affect
each other through the no-skid walls. No energy can flow from one system to the other because no energy
can move from the atoms in the system into the walls. The formal name for such walls is insulating or
adiabatic. As we have discussed previously, such perfectly insulating walls do not exist in nature (if nothing
else, system A affects system B through the gravitational interaction), but good experimental realizations
exist and they are certainly handy conceptually.
The total system consisting of A plus B is characterized by the macroscopic quantities EA , EB , VA , VB ,
NA , and NB . The number of microstates corresponding to this macrostate is
T (EA , EB , VA , VB , NA , NB ) = A (EA , VA , NA )B (EB , VB , NB ).
2.7.3
(2.43)
Two systems in thermal contact
Consider the same two systems, A and B, as before, but now allow the wall between them to be an ordinary,
energy-permeable wall. Such a wall is called diathermal. (The walls between the systems and the outside
world are still insulating.)
A
EA, VA, NA
B
EB, VB, NB
Now the mechanical parameters VA , VB , NA , and NB still remain constant when the two systems are brought
together, but the energies EA and EB change. (Of course the total energy ET EA +EB remains constant as
individual energies change.) Such contact between two systems, where energy can flow between the systems
40
but where the mechanical parameters dont change, is called thermal contact. Once they are in contact, I
dont know what the distribution of energy between A and B is: It could be that A is in its ground state and
all the excess energy is in B, or vice versa, and it could change rapidly from one such extreme to the other.
But while I no longer have knowledge of the distribution, I do have an expectation. My expectation, drawn
from daily experience, is that energy will flow from one system to another until EA and EB take on some
pretty well-defined final equilibrium values. The exact energies will then fluctuate about those equilibrium
values, but those fluctuations will be small. And what characterizes this final equilibrium situation? Your
first guess might be that the energy on each side should be equal. But no, if a big system were in contact
with a small system, then youd expect more energy to end up in the big system. Then you might think
that the energy per particle on each side should be equal, but thats not right either. (Ice and liquid water
are in equilibrium at the freezing point, despite the fact that the energy per particle of liquid water is much
greater than than of ice.) The correct expectation (known to any cook) is that:
When two systems A and B are brought into thermal contact, EA and EB will change until they
(e)
(e)
reach equilibrium values EA and EB , with negligible fluctuations, and at equilibrium both
systems A and B will have the same temperature.
This is the qualitative property of temperature that will allow us, in a few moments, to produce a rigorous, quantitative definition. To do so we now look at the process microscopically instead of through our
macroscopic expectations.
Once the two systems are in thermal contact the energy of system A could be anything from 0 to the
total energy ET .14 For any given EA , the number of accessible microstates of the combined system, A plus
B, is A (EA )B (ET EA ). Thus the total number of microstates of the combined system is
T (ET ) =
ET
X
A (EA )B (ET EA ).
(2.44)
EA =0
(Please dont tell me the obvious. . . that EA is a continuous quantity so you have to integrate rather than
sum. This is a seat-of-the-pants argument. If you want attention to fine points, look them up in Ruelles
book.)
Let us examine the character of the summand, A (EA )B (ET EA ), as a function of EA . The functions
A (EA ) and B (EB ) are both rapidly increasing functions of their arguments, so B (ET EA ) is a rapidly
decreasing function of EA . Thus the product A (EA )B (ET EA ) will start out near zero, then jump up
for a narrow range of EA values and then fall quickly to near zero again.
14 Here
and in the rest of this section, we assume that the ground state energy of both system A and system B has been set
by convention to zero.
41
A(EA)B(ETEA)
A(EA)
B(ETEA)
EA
EA
EA
In conclusion, for most values of EA the summand nearly vanishes, so most of the contribution to the sum
A :
comes from its largest term, which is located at an energy we will call E
A )B (ET E
A ) + small change.
T (ET ) = A (E
(2.45)
The sharply peaked character of the product A (EA )B (ET EA ) becomes more and more pronounced as
the thermodynamic limit is approached, so in this limit the small change can be ignored. You can see
A should be interpreted as the
that for macroscopic systems the microscopically most probable value E
(e)
thermodynamic equilibrium value EA .
We locate this value by maximizing
A (EA )B (ET EA ).
(2.46)
To make life somewhat easier, we will not maximize this function directly but, what is the same thing,
maximize its logarithm
f (EA )
=
ln[A (EA )B (ET EA )]
(2.47)
ln A (EA ) + ln B (ET EA ).
(2.48)
Recognizing the entropy terms we have

kB f (EA ) = SA (EA ) + SB (ET EA ),
(2.49)
(kB f (EA ))
SA
SB EB
=
+
.
EA
EA
EB EA
| {z }
(2.50)
and differentiation produces
We locate the maximum by setting the above derivative to zero, which tells us that at equilibrium
SA
SB
=
.
EA
EB
(2.51)
42
Our expectation is that the temperature will be the same for system A and system B at equilibrium, while
our microscopic argument shows that S/E will be the same for system A and system B at equilibrium.
May we conclude from this that
S
= Temperature ?
(2.52)
E
No we cannot! All we can conclude is that
S
= invertible function (Temperature).
E
(2.53)
To find out which invertible function to choose, let us test these ideas for the ideal gas. Using
S(E) = kB ln C + 23 kB N ln E + 52 kB N
(2.54)
we find
S
3 N
= kB ,
(2.55)
E
2
E
so a large value of E results in a small value of S/E. Because we are used to thinking of high temperature
as related to high energy, it makes sense to define (absolute) temperature through
S(E, V, N )
1
.
T (E, V, N )
E
(2.56)
Applying this definition to the ideal gas gives the famous result
E
3
= kB T.
N
2
(2.57)
In fact, this result is too famous, because people often forget that it applies only to the ideal gas. It is a common
misconception that temperature is defined through equation (2.57) rather than through equation (2.56), so that
temperature is always proportional to the average energy per particle or to the average kinetic energy per particle.
In truth, these proportionalities hold exactly only for noninteracting classical point particles.
2.7.4
Two systems in thermal and mechanical contact
Consider the same two systems, A and B, as before, but now allow them to exchange both energy and
volume. (The wall between A and B is a sliding, diathermal wall. The walls between the systems and the
outside world remains insulating and rigid.)
A
EA, VA, NA
B
EB, VB, NB
43
Now the numbers NA and NB remain constant, but the energies EA and EB , and the volumes VA and VB ,
do change. Both the total energy ET EA + EB and the total volume VT VA + VB remain constant as
the individual energies and volumes change. Because of the type of contact, I dont know the distribution
of energy between A and B, nor do I know the distribution of volume between A and B. But just as with
energy, I do have an expectation.
When two systems A and B are brought into thermal and mechanical contact, EA , EB , VA , and
(e)
(e)
(e)
(e)
VB will change until they reach equilibrium values EA , EB , VA , and VB , with negligible
fluctuations, and at equilibrium systems A and B will have the same temperature and the same
pressure.
Just as with energy and temperature, this qualitative expectation will lead us, after a microscopic argument,
to a rigorous, quantitative definition of pressure.
In most cases the system will first reach a preliminary mechanical equilibrium as the wall slides quickly to make the
pressure equal on both sides. Then more slowly, as energy seeps from one side to the other, the two temperatures
will equilibrate. As the temperatures change the pressure on each side changes, and the wall again slides, more
slowly now, to reflect these changes. The argument that follows could be carried out using a sliding insulated wall,
instead of a sliding diathermal wall, but in this case energy is still transferred from A to B through mechanical
work, so the argument is in fact more complicated that way.
The total number of microstates of the combined system is

XX
T (ET , VT ) =
A (EA , VA )B (ET EA , VT VA ).
(2.58)
EA V A
For an ideal gas, (V ) = CV N , but in general we expect (V ) to be a rapidly increasing function of V .

(In analogy to the pirate booty argument, we might call this the sardine argument.) Just as before,
this leads us to expect that the product A (EA , VA )B (ET EA , VT VA ) will nearly vanish except in the
(e)
(e)
vicinity of the equilibrium values EA = EA and VA = VA , and that the sum can be approximated by its
largest value
(e)
(e)
(e)
(e)
T (ET , VT ) A (EA , VA )B (ET EA , VT VA ).
(2.59)
We locate the equilibrium values by maximizing the product A (EA , VA )B (ET EA , VT VA ) with respect
to both EA and VA . This leads directly to the requirements
SB
SA
=
,
EA
EB
(2.60)
SA
SB
=
.
VA
VB
(2.61)
S
= invertible function (Pressure, Temperature).
V
(2.62)
as before, and
We conclude that
44
The dimensions of S/V are

S
V

=
while those for pressure are

[p] =
J/K
m3
(2.63)
J
.
m3
(2.64)
This motivates us to define pressure through

p(E, V, N )
S(E, V, N )
.
T (E, V, N )
V
(2.65)
For an ideal gas,

so
S(V ) = kB N [C + ln V ].
(2.66)
S
kB N
p
=
= ,
V
V
T
(2.67)
pV = N kB T,
(2.68)
leading to
the ideal gas equation of state! (Which is again too famous for its own good. Too many statistical mechanics
courses degenerate into an intensive study of the ideal gas, and it is a shame to waste such a beautiful, powerful,
and diverse subject on the study of only one substance, especially since that substance doesnt even exist in
nature!)
2.7.5
Two systems that can exchange energy, volume, and particles
We return to the same situation as before, but now the wall between A and B is not only diathermal and
sliding, but also allows particles to flow. (Imagine piercing a few tiny holes in the wall.)
A
EA, VA, NA
B
EB, VB, NB
The reasoning in this case exactly parallels the reasoning already given two above, so I will not belabor it.
I will merely point out that we can now define a new quantity, the chemical potential , which governs the
flow of particles just as temperature governs the flow of energy or pressure governs the flow of volume. It is
S(E, V, N )
(E, V, N )
.
T (E, V, N )
N
(2.69)
2.7.6
45
The meaning of chemical potential
The upshot of all our definitions is that

dS =
1
p
dE + dV dN.
T
T
T
(2.70)
What do these definitions mean physically? The first term says that if two systems are brought together so
that they can exchange energy but not volume or particle number, then in the system with high temperature
the energy will decrease while in the system with low temperature the energy will increase. The second
term says that if two systems with the same temperature are brought together so that they can exchange
volume but not particle number, then in the system with high pressure the volume will increase while in
the system with low pressure the volume will decrease. The third term says that if two systems with the
same temperature and pressure are brought together so that they can exchange particles, then in the system
with high chemical potential the number will decrease while in the system with low chemical potential the
number will increase. In summary:
Exchange of energy?
Yes.
Permitted, but T = 0.
Permitted, but T = 0.
Exchange of volume?
Prohibited.
Yes.
Permitted, but p = 0.
Exchange of particles?
Prohibited.
Prohibited.
Yes.
Result:
System with high T decreases in E.
System with high p increases in V .
System with high decreases in N .
Chemists have a great name for chemical potential. . . they call it escaping tendency, because particles
tend to move from a region of high to a region of low just as energy tend to move from a region of high
T to a region of low T . (G.W. Castellan, Physical Chemistry, 2nd edition, Addison-Wesley, Reading, Mass.,
1971, section 112.) You can see that chemical potential is closely related to density, and indeed for an ideal
gas
= kB T ln(3 (T )),
(2.71)
where = N/V is the number density and (T ) is a certain length, dependent upon the temperature, that
we will see again (in equation (5.4)).
Once again, remember that this relation holds only for the ideal gas. At the freezing point of water, when ice
and liquid water are in equilibrium, the chemical potential is the same for both the solid and the liquid, but the
densities clearly are not!
Problems
(e)
2.21 (Q) The location of EA

(e)
In the figure on page 41, does the equilibrium value EA fall where the two curves A (EA ) and
B (ET EA ) cross? If not, is there a geometrical interpretation, in terms of that figure, that does
characterize the crossing?
2.22 (Q) Meaning of temperature and chemical potential
The energy of a microscopic system can be dissected into its components: the energy due to motion,
46

the energy due to gravitational interaction, the energy due to electromagnetic interaction. Can the
temperature be dissected into components in this way? The chemical potential? (Clue: Because the
temperature and chemical potential are derivatives of the entropy, ask yourself first whether the entropy
can be dissected into components.)
2.23 (E) Chemical potential of an ideal gas

a. Show that the chemical potential of a pure classical monatomic ideal gas is

3
4mEV 2/3
= kB T ln
2
3h20 N 5/3
"
3/2 #
2mkB T
V
= kB T ln
.
h20
N
(2.72)
(2.73)
b. Show that when the temperature is sufficiently high and the density sufficiently low (and these are,
after all, the conditions under which the ideal gas approximation is valid) the chemical potential
is negative.
c. Show that
(T, V, N )
(T, V, N ) 3
=
kB .
T
T
2
At a fixed, low density, does increase or decrease with temperature?
(2.74)
2.24 Ideal paramagnet, take three

Find the chemical potential (E, H, N ) of the ideal paramagnet. (Your answer must not be a function
of T .) (To work this problem you must have already worked problems 2.9 and 2.12.)
Resources
Percival as CM for chaos and SM. Hilborn on chaos.
PAS chaos programs, public domain stadium and wedge. CUPS stadium.
Tolman, Ruelle.
2.8
47
Additional Problems
2.25 Accessible configurations in another spin system

The Ising model for a ferromagnetic material such as nickel is different from the ideal paramagnet
discussed in problem 2.9. In this model the spins reside at lattice sites and may point either up or
down, but in contrast to the ideal paramagnet model, two adjacent spins do interact. Specifically, if two
adjacent spins point in the same direction, the pair has energy J. If they point in different directions,
the pair has energy +J. (In the ferromagnetic situation considered here, J > 0.) We consider here a
one-dimensional Ising model of N sites with zero external magnetic field.
b
6
b
?
b
?
b6
b
?
i
b
?
i+1
b6
b6
b
?
N
a. Not every energy is possible for this model. What is the maximum possible energy? The minimum? What is the smallest possible energy spacing between configurations?
b. Consider the energy range from E to E + E where E is small compared to N J but large
compared to J. What is the approximate number of states (E, E, J, N ) lying within this
energy range?
c. Write down an expression for the entropy S(E, E, J, N ) as a function of E. (Clue: Be careful
to never take the logarithm of a number with dimensions.)
d. Take the thermodynamic limit to find the entropy per atom as a function of the energy per atom:
s(e, J). (Clue: Use Stirlings approximation.)
e. Check your work by comparing the Ising entropy per spin s(e, J) when the interactions vanish
(J = 0) to the ideal paramagnet entropy per spin s(e, H) in zero field. (See problem 2.12. Clue:
If J = 0, then e = 0 also. Why?)
General clues: It helps to concentrate not on the status of the spins (up or down) but on the status of
the pairs (alike or different). How many nearest neighbor pairs are there? For each given configuration
of pairs, how many corresponding configurations of spins are there?
2.26 Shake and spin
Make up a problem involving an ideal paramagnet with atoms free to move around.
2.27 Entropy as a function of mass
Use the Sackur-Tetrode formula to find the difference between the entropy of a sample of argon gas
and a sample krypton gas under identical conditions. Use the data in Ihsan Barin, Thermochemical
Data of Pure Substances, third edition (VCH Publishers, New York, 1995) to test this prediction at
the temperatures 300 K and 2000 K.
48
2.28 (Q,D) For discussion: The icy cold of outer space

Is outer space at a high temperature or a low temperature? Before answering, consider these points:
a. The environment of outer space consists of protons and photons streaming away from the sun.
The two systems are virtually uncoupled (do not interact with each other). Im asking for the
temperature of the protons.
b. It was implicitly (perhaps illicitly) assumed in section 2.7.3 that the centers of mass of the two
systems in thermal contact were in the same reference frame. This is a general principle: When
referring to the temperature of a moving body, we mean the temperature as measured by a
thermometer moving along with that body.
c. Remember that in the arguments of section 2.7.3 the two systems could exchange energy but not
particles.
d. What if we approximated the protons by an ideal gas?
e. Look at the Handbook of Chemistry and Physics under index entry Temperature, Earths atmosphere.
2.29 (Q,D) Errors in Reif
Scan Reif pages 4761, but dont read them in detail as they are rife with errors. Point out the major
error in quantum mechanics made in each of the following passages. (Use a maximum of three sentences
per passage.)
a. The examples on pages 48 and 49.
b. The second paragraph on page 57.
c. The first paragraph on page 60.
2.30 Polymers
Id like to think up some problem concerning polymers of fixed length vs. polymers of fixed activity
for length. The activity is analogous to chemical potential in that it is a control knob for polymer
length. It might render the whole concept more concrete.
Chapter 3
Thermodynamics
3.1
Heat and Work
Weve seen how to calculateat least in principle, although our techniques were primitive and difficult to
usethe entropy S(E, V, N ). Weve also argued (the booty argument, section 2.7.1) that for most systems
entropy will increase monotonically with energy when V and N are fixed. Thus we (can)/(will always be
able to) invert the function to find energy as a function of entropy, volume, and number:
E(S, V, N ).
(3.1)
Recall that V and N are just stand-ins as representative mechanical parameters.

But how can we measure entropy? Go into any laboratory and try to count the number of microstates
consistent with a macrostate: you will find no equipment to aid your endeavor. Although we have produced
a conceptual definition of entropy, we need also an operational definition of entropy.
In an operational definition we define a target word, not in terms of other words, but in terms of
a procedure that produces a decision or a numerical value concerning that target word. For example, an
operational definition for the word tuberculosis is a disease marked by wasting, cough, fever, and formation
of small cheesy lumps in the lungs. In contrast, a conceptual definition is a disease caused by the bacillus
bacterium Mycobacterium tuberculosis. The operational definition tells you how to diagnose the disease;
the conceptual definition tells you whats going on and suggests a cure. We need to produce an operational
definition to go along with our conceptual definition of entropy, and thats the starting point of this chapter.
Heat.
Configuration work.
Dissipative work. Always increases the energy of the system. For example, stirring and shaking. Joule
and the waterfall story.
49
50
CHAPTER 3. THERMODYNAMICS
Caution: Heat and work refer to mechanisms for energy change, not to types of energy. Do not think
that heat goes into a change of kinetic energy (or into mixed up energy) whereas work goes into a change
of potential energy (or into ordered energy).
Problems
3.1 What is heat?
Below is an excerpt from a sixth-grade science textbook. Find at least six errors in the excerpt.
What is heat?
You have learned that all matter is made up of atoms. Most of these atoms combine to
form molecules. These molecules are always movingthey have kinetic energy. Heat is the
energy of motion (kinetic energy) of the particles that make up any piece of matter.
The amount of heat a material has depends on how many molecules it has and how fast
the molecules are moving. The greater the number of molecules and the faster they move,
the greater the number of collisions between them. These collisions produce a large amount
of heat.
How is heat measured? Scientists measure heat by using a unit called a calorie. A calorie
is the amount of heat needed to raise the temperature of 1 gram of 1 water 1 degree centigrade
(Celsius).
A gram is a unit used for measuring mass. There are about 454 grams in 1 pound.
What is temperature?
The amount of hotness in an object is called its temperature. A thermometer is used to
measure temperature in units called degrees. Most thermometers contain a liquid.
3.2
Entropy
The operational definition of entropy is that, for a quasistatic change,

dS =
dQ
.
T
(3.2)
Configurational entropy vs. thermal entropy

The operational, thermodynamic definition for entropy, equation (3.2), was in fact discovered before the
more fundamental, conceptual, statistical definition (2.7) was. Because of this dual definition some people
get the misconception that there are two kinds of entropy: the statistical kind (configurational entropy)
and the operational kind (thermal entropy). Thats false. These are two different ways of looking at the
same thing.
3.2. ENTROPY
51
3.2 (I) Ideal gas with variables S and V .

a. Use equation (2.32) to show that for a sample of monatomic ideal gas with a fixed number of
particles, the energy function is
2/3
V0
E(S, V ) = E0
eS/S0 ,
(3.3)
V
where E0 , V0 , and S0 are constants with the dimensions of energy, volume, and entropy.
b. Use the formulas
E(S, V )
E(S, V )
and p(S, V ) =
S
V
to find the temperature and pressure functions for this sample.
T (S, V ) =
(3.4)
3.3 (I) Dependence of heat and work on path

Suppose the ideal gas sample mentioned in the previous problem undergoes a change from an initial
state with volume V0 and entropy S0 to a final state with volume 8V0 and entropy (1 + 5 ln 2)S0 . There
are an infinite number of quasistatic paths that execute this change. For example:
A First increase the entropy from S = S0 to S = (1 + 5 ln 2)S0 , keeping the volume fixed at V = V0 ,
then expand from V = V0 to V = 8V0 , keeping the entropy fixed at S = (1 + 5 ln 2)S0 .
B Increase the volume and entropy simultaneously along the path

V
5
ln
+1 .
S(V ) = S0
3
V0
(3.5)
C First expand from V = V0 to V = 8V0 , keeping the entropy fixed at S = S0 , then increase the
entropy from S = S0 to S = (1 + 5 ln 2)S0 , keeping the volume fixed at V = 8V0 .
Each path has the same change in volume, entropy, and energy, but each will have a different heat and
work. (Remember that each path is quasistatic.)
a. Calculate the heat absorbed by the sample and the work done by the sample for each of these
three paths.
b. Find the difference Q W for each of these three paths.
c. A fourth path from the initial to the final state involves a non-quasistatic change and the absorption of heat Q = 3E0 . What work is done by the sample during this change?
(In making up this problem I choose somewhat peculiar initial and final states, and connecting paths,
so that the above quantities could be calculated analytically. Despite these peculiarities the principle
illustrated by the problem should be clear: heat and work depend upon path, but their difference the
energy change does not.)
3.4 (I*) Work in terms of p and V
We have so far considered work and heat in a fluid system specified by the variables S, V , and N ,
and have often assumed that N is constant. Another way of specifying such a system is through the
variables p (pressure) and V . This problem considers quasistatic processes described as paths in the
(p, V ) plane.
52
a. Show that the work performed by the system in changing quasistatically from the initial state
(Vi , pi ) to the final state (Vf , pf ) along the path shown in the figure below is equal to the area
under the curve. What is the work done if the path is traversed in the opposite direction?
p
(Vf , pf)
(Vi , pi)
0
b. What is the work done through the same change along the path below?
p
(Vf , pf)
(Vi , pi)
0
c. Finally, what is the work done if the system comes back to its starting point as in the cycle shown
below? What if the cycle is instead executed in a counterclockwise fashion?
3.3. HEAT ENGINES
53
3.5 (I) Fluid work

In a quasistatic, adiabatic process, a sample of gas expands from Vi = 1 m3 to Vf = 8 m3 , while the
pressure drops from pi = 32 Pa to pf = 1 Pa. It is observed that during this process the pressure and
volume are related by
pV = constant,
(3.6)
where the constant is 35 . Of course, this is not the only process leading from this particular initial
state to that particular final state. Find the total work done and heat absorbed in each of the following
quasistatic processes, all of which connect the same two thermodynamic states.
a. The adiabatic process described above.
b. At constant pressure, the gas expands from Vi to Vf . (Heat must be added to the system during
this expansion in order to keep the pressure constant.) Then at constant volume, the the pressure
is reduced from pi to pf . (Heat must be extracted from the system during this stage of the
process.)
c. The volume is increased and heat is supplied at just the right rate to make the pressure decrease
linearly with volume.
d. As in part (b.) but with the two steps performed in opposite order.
3.3
Heat Engines
The emphasis of this course is on the properties of matter. This section is independent of the properties of
matter! It is included because:
1. historically important
2. expected coverage (e.g. for GRE)
3. important from an applications and engineering standpoint (See, for example, Hurricane heat engines
by H.E. Willoughby, Nature 401 (14 October 1999) 649650. Or A thermoacoustic Stirling heat
engine by S. Backhaus and G.W. Swift, Nature 399 (27 May 1999) 335338, and Traveling-wave
thermoacoustic electric generator, by S. Backhaus, E. Tward, and M. Petach, Applied Physics Letters
85 (9 August 2004) 10851087, concerning an engine that might replace nuclear thermal power sources
in deep space missions.)
4. fascinating
This section follows Reif sections 3.1 and 5.11.
Heat engines do not include electrical and chemical engines (such as muscles). The Carnot theorems
dont apply to them. The name Carnot is French, so it is pronounced Car-no rather than Car-not.
54
Problems
3.6 (I*) The Carnot cycle
Describe the Carnot cycle. Always means quasistatic. Take this order:
1. Expand at constant T = T1 from A to B, decreasing pressure. Absorb heat q1 from hightemperature heat bath.
2. Expand even more, adibatically, from B to C, decreasing pressure still more.
3. Contract at constant T = T2 from C to D, increasing pressure. Expel heat q2 into low-temperature
heat bath.
4. Contract even more, adibatically, increasing pressure still more, until system comes back to original
state A.
3.7 (E) The Carnot cycle
(This problem is stolen from a GRE Physics test.)
The p-V diagram for a quasistatic Carnot cycle is shown in Reif page 189. Legs bc and da represent
isotherms, while ab and cd represent adiabats. A system is carried through the cycle abcd, taking
in heat q1 from the hot reservoir at temperature T1 and releasing heat q2 to the cold reservoir at
temperature T2 . Which of the following statements is false?
a. q1 /T1 = q2 /T2 .
b. The entropy of the hot reservoir decreases.
c. The entropy of the system increases.
d. The work w done is equal to the net heat absorbed, q1 q2 .
e. The efficiency of the cycle is independent of the working substance.
3.8 (D) The liquifying Carnot cycle
Suppose a quasistatic Carnot cycle is executed with a working substance of steam rather than ideal
gas. Furthermore, suppose that the cycle straddles the liquid-vapor coexistence curve, so that when
the working substance is at a high volume and low pressure it is steam, but when it is at a low volume
and high pressure it is liquid water. Then the (p, V ) diagram of the cycle resembles the following:
A
p
B
D
C
3.4. THERMODYNAMIC QUANTITIES
55
a. Explain the significance of the flat portions of the isotherms.

b. Of the four points A, B, C, and D, which represent liquid and which represent gas?
c. Below is a list of properties of the Carnot cycle executed with an ideal gas. Which of these
properties remain true for our liquifying Carnot cycle?
i. The efficiency is 1 T1 /T2 .
ii. The adiabats BC and DA are described by pV = constant.
iii. The work done is the area enclosed by the cycle.
iv. More.
3.4
Thermodynamic Quantities
By a thermodynamic quantity I mean either a variable or a function. We have already seen that for fluids
we may regard the entropy as a function of energy, volume, and number, S(E, V, N ), or the energy as a
function of entropy, volume, and number, E(S, V, N ): using the term quantity avoids prejudging the issue
of whether entropy, for example, is to be treated as a variable or a function.
It is traditional to divide thermodynamic quantities into two classes: extensive and intensive. Extensive quantities increase linearly with system size while intensive quantities are independent of system size.
(You might ask, Why cant a quantity depend on system size in some other fashion, such as quadratically?
The answer is that this is not impossiblefor example, the radius of a spherical sample increses as the cube
root of the system sizebut it rarely happens in practice for quantities that concern ussee problem 3.13.)
Thus extensive quantities are characteristic of the sample while intensive quantities are characteristic of the
substance.
Examples of extensive quantities are energy, volume, number, entropy, and the magnetization M (i.e. the
magnetic dipole moment of the entire sample):
E,
V,
N,
S,
M.
(3.7)
Examples of intensive quantities are temperature, pressure, chemical potential, and the applied magnetic
field H:
T, p, , H.
(3.8)
If one extensive quantities is divided by another extensive quantity, the quotent is an intensive quantity. The
most frequently used quantities of this type are the number densities, such as
e=
E
,
N
s=
S
,
N
and
v=
V
.
N
(3.9)
u=
E
,
V
S=
S
,
V
and
N
.
V
(3.10)
There are also volume densities like
56
Quantities such as E/S are also intensive, but they are rarely used and dont have special names.
Note that in defining densities such as above, we are not making any assumptions about how the matter
in the sample is arranged. For liquid water in equilibrium with ice (e.g. at atmospheric pressure and a
temperature of 273 K) the water has a greater number density (and a greater entropy density) than the ice,
but the number density of the system is still just the total number divided by the total volume.
Here is a valuable problem solving tip. You know from previous courses that it is important to keep track
of the dimensions of physical quantities: if you derive an equation for which the left side has the dimensions
of meters and the right side has the dimensions of meters/second, your derivation must be wrong. In
thermodynamics there is the equally important check of extensivity: if the left side is extensive and the right
side is intensive, then the equation must be wrong. You should keep a running check of these properties
as you solve problems. The first time you produce an equation that is incorrect in its dimensions or its
extensivity, go back immediately and correct your error.
Problem 1.2 has already introduced the isothermal compressibility
T (p, T ) =
1 V (p, T )
V
p
(3.11)
and the expansion coefficient

1 V (p, T )
.
V
T
Both of these quantities are defined so as to be intensive.
(p, T ) =
(3.12)
Another quantity of interest is the heat capacity, defined informally as the amount of heat required
to quasistatically raise the temperature of a sample by one Kelvin. (The heat capacity was introduced
briefly in problem 1.3.) Because this definition refers to a sample, heat capacity is an extensive quantity.
The definition is incomplete because it doesnt say what path is taken while the heat is added, and we know
that heat is a path-dependent quantity.1 For fluids, the two most frequently used heat capacities are the
heat capacity at constant volume and at constant pressure, denoted by
CV
and
Cp .
(3.13)
Thus Cp (which will in general be a function of temperature, pressure, and number) is the amount of heat
required to increase the temperature of a sample by one Kelvin while the sample is under a constant pressure
(often, in earth-bound laboratories, a pressure of one atmosphere). In contrast CV (which will in general be
a function of temperature, volume, and number) is the amount of heat required to increase the temperature
of a sample by one Kelvin while holding it in a strong box to prevent it from expanding. Because liquids
expand little upon heating, CV and Cp are nearly equal for liquids. But the two quantities differ dramatically
for gases.
For small quasistatic changes, we know that
S =
1 In
Q
T
(3.14)
truth, its worse than this, because it doesnt even specify the end point of the process. If we are describing the fluid
sample by V and T , for example, and the starting point were (Vi , Ti ), then the end point must have Tf = Ti + 1 K, but as far
as our informal definition goes, the final volume could be anything.
3.4. THERMODYNAMIC QUANTITIES
57
whence
T S
.
T
C = lim
T 0
(3.15)
This gives us the formal definitions of heat capacity,

CV = T
S(T, V, N )
T
and
Cp = T
S(T, p, N )
,
T
(3.16)
and from these equations it is again clear that heat capacity is an extensive quantity.
The intensive quantities analogous to heat capacities are called specific heats:
cV =
CV
N
and
cp =
Cp
.
N
(3.17)
The specific heat is readily measured experimentally and we will come back to it again and again.
Problems
3.9 (E) Which one was it?
For a system with a fixed number of particles, the reciprocal of the absolute (Kelvin) temperature T
is given by which of the following derivatives? Explain your reasoning. Do not look up the answer or,
if you have memorized it, dont use that piece of your memory. (Clue: Dimensional analysis will work,
but theres an even easier way.)
p(S, V )
V
p(S, V )
S
S(E, p)
p
V (E, p)
p
S(E, V )
E
(3.18)
3.10 (I) Acceptable forms for the entropy

Below are seven equations purporting to be equations for S(E, V, N ) for various thermodynamic systems. However, four of them are not physically acceptable. Find the four inpermissible equations and
for each indicate why it cannot be correct. The quantities T0 and v0 are in all cases positive, intensive
quantities with the dimensions of temperature and volume, respectively.

S(E, V, N )
=

S(E, V, N )
S(E, V, N )
2
kB
v0 T0
1/3
kB v02
T02
(EV N )1/3
1/3
EN
V
(3.19)
2/3
(3.20)
S(E, V, N )
1/3
kB v02
EN
T02
V
1/2
1/2
kB
kB T0 V 2
=
NE +
T0
v02
2
= kB N ln(EV /N kB T0 v0 )
(3.23)
S(E, V, N )
= kB N exp(EV /N 2 kB v0 )
(3.24)
S(E, V, N )
S(E, V, N )
= kB N exp(EV /N
kB T0 v02 )
(3.21)
(3.22)
(3.25)
58
3.11 (I) Heat capacities for the ideal gas

Use equation (2.32) to find the heat capacities

S
CV (T, V, N ) = T
T V,N
and
Cp (T, p, N ) = T
S
T

p,N
for the pure classical monatomic ideal gas.

3.12 (I) Heat capacity at a phase transition
What is Cp for a fluid when the liquid and gas coexist?
3.13 (D) Gravitational potential energy
Consider a uniform sphere of mass M and radius R.
a. Use dimensional analysis to show that the gravitational potential energy of the sphere is of the
form
M2
UG = cG
,
(3.26)
R
where c is a dimensionless constant independent of M and R.
b. Is UG intensive, extensive, or neither?
c. Is the total energy E (which is equal to UG plus contributions from other forms of energy) intensive,
extensive, or neither?
d. Estimate the value of UG for a glass of water.
You should conclude from this problem that the total energy E is in principle non-extensive, but that
the deviations from being extensive are negligible for everyday thermodynamic systems.
e. (Optional and quite time consuming.) Use ideas from classical mechanics to show that the dimensionless constant in equation (3.26) is c = 3/5.
3.5
Multivariate Calculus
What is a section on multivariate calculus doing in a physics book. . . particularly a physics book which
assumes that you already know multivariate calculus? The answer is that you went through your multivariate
calculus course learning one topic after another, and there are some subtle topics that you covered early in
the course that really couldnt be properly understood until you had covered other topics that came later
in the course. (This is not the fault of your teacher in multivariate calculus, because the later topics could
not be understood at all without an exposure to the earlier topics.) This section goes back and investigates
five subtle points from multivariate calculus to make sure they dont trip you when they are applied to
thermodynamics.
3.5. MULTIVARIATE CALCULUS
3.5.1
59
What is a partial derivative?
Given a function f (x, y, z), what is the meaning of f /y? Many will answer that
f
is the change of f with y while everything else is held constant.
y
(3.27)
This answer is WRONG! If f changes, then f 2 changes, and sin f changes, and so forth, so it cant be that
everything else is held constant. The proper answer is that
f
is the change of f with y while all other variables are held constant.
y
(3.28)
Thus it becomes essential to keep clear which quantities are variables and which are functions. This is not
usually hard in the context of mathematics: the functions are f , g, and h while the variables are x, y, and
z. But in the context of physics we use symbols like E, V , p, T , and N which suggest the quantities they
represent, and it is easy to mix up the functions and the variables.
An illustration from geometry makes this point very well. Consider the set of all right circular cylinders.
Each cylinder can be specified uniquely by the variables r, radius, and h, height. If you know r and h for
a cylinder you can readily calculate any quantity of interestsuch as the area of the top, T (r, h), the area
of the side S(r, h), and the volume V (r, h)as shown on the left side of the table below. But this is not
the only way to specify each cylinder uniquely. For example, if you know the height and the side area of
the cylinder, you may readily calculate the radius and hence find our previous specification. Indeed, the
specification through the variables S and h is just as good as the one through the variables r and h, as is
shown on the right side of the table below. [There are many other possible specifications (e.g. r and S, or T
and V ) but these two sets will be sufficient to make our point.]
Describing a Cylinder
r
variables:
radius
height
functions:
top area
side area
volume
r
h
variables:
side area
height
S
h
T (r, h) = r2
S(r, h) = 2rh
V (r, h) = r2 h
functions:
radius
top area
volume
r(S, h) = S/2h
T (S, h) = S 2 /4h2
V (S, h) = S 2 /4h
All of this is quite straightforward and ordinary. But now we ask one more question concerning the
geometry of cylinders, namely How does the volume change with height? The last line of the table
60
presents two formulas for volume, so taking appropriate derivative gives us either
V
= r2 = T
h
or
V
= S 2 /4h2 = T.
h
(3.29)
What? Is V /h equal to T or to T ? It cant be equal to both!

The problem with equation (3.29) is that we were careless about specifying the variables. The two
expressions for volume,
V (r, h) = r2 h and V (S, h) = S 2 /4h,
(3.30)
are in fact two completely different functions, with completely different variables, so we should expect completely different derivatives. If we increase the height h keeping the radius r fixed, the figure on the left
below makes it clear that the volume increases. But if we increase the height and keep the side area S fixed,
the radius will have to decrease as shown on the right below. The change on the right adds to the volume
at the top of the cylinder but subtracts from the volume all along the sides. It would be most remarkable if
the two volume changes were equal and, as our derivatives have shown, they are not.
A mathematician would say that we got ourselves in trouble in equation (3.29) because we gave two
different functions the same name. A physicist would reply that they both represent the volume, so they
deserve the same name. Rather than get into an argument, it is best to write out the variables of all functions
explicitly, thus rewriting (3.29) as
V (r, h)
= r2 = T (r, h)
h
or
V (S, h)
= S 2 /4h2 = T (S, h).
h
(3.31)
(Physicists often neglect to write out the full list of variables, which saves some time and some ink but which
invites error. R.H. Price and J.D. Romano (Am. J. Phys. 66 (1998) 114) expose a situation in which a
physicist published a deep error, which he made by neglecting to write out an explicit variable list.)
It becomes tiresome to write out the entire argument list for every function, so a shorthand notation
has been developed. A right hand parenthesis is written after the partial derivative, and the functional
arguments that are not being differentiated are listed as subscripts to that parenthesis. Thus the derivatives
above are written as

V (r, h)
V
V (S, h)
V
and
.
(3.32)
=
=
h
h r
h
h S
The expression on the left is read the partial derivative of V with respect to h while r is held constant.
3.5. MULTIVARIATE CALCULUS
3.5.2
61
The Legendre transformation
Let us return to a description of cylinders in terms of the variables r and h. Clearly, one of the functions of
interest is the volume
V (r, h).
(3.33)
A glance at the table on page 59 (or a moments thought about geometry) shows that the total differential
of V as a function of r and h is
dV = S(r, h) dr + T (r, h) dh,
(3.34)
whence

V
V
and T (r, h) =
.
(3.35)
r h
h r
Thus knowledge of the function V (r, h) gives us a bonus. . . if we know V (r, h), then we can take simple
derivatives to find the other quantities of interest concerning cylinders, namely S(r, h) and T (r, h). Because
of the central importance of V (r, h), it is called a master function and the total differential (3.34) is called
a master equation.
S(r, h) =
Is there any way to find a similarly convenient master description in terms of the variables S and h?
Indeed there is, and it is given by the Legendre transformation. In the Legendre transformation from the
variables r and h to the variables S and h, we change the focus of our attention from the master function
V (r, h) to the function
(S, h) = V (r(S, h), h) Sr(S, h).
(3.36)
(The above equation is written out in full with all arguments shown. It is more usually seen as
= V Sr,
(3.37)
although this form raises the possibility that variables and functions will become mixed up.) The total
differential of is
d
= dV S dr r dS
(3.38)
= S dr + T dh S dr r dS
(3.39)
= r dS + T dh.
(3.40)
We have found a new master function! It is

(S, h),
(3.41)
d = r(S, h) dS + T (S, h) dh,
(3.42)
and the new master equation is

giving rise immediately to
r(S, h) =

and
h
T (S, h) =

.
(3.43)
This description has all the characteristics of a master description: once the master function is known, all
the other interesting functions can be found by taking straightforward derivatives.
62
3.5.3
Multivariate chain rule
Begin with exact differentials and divide.
3.5.4
Implicit function theorem
3.5.5
Maxwell relations
Problems
3.14 Partial derivatives in space
A point on the plane can be specified either by the variables (x, y) or by the variables (x0 , y 0 ) where
y6
OC
C
y0 C
C
C
C
C
C
C
C
:

0

x
C

C

C

x
C
x0
y
+ cos() x + sin() y,
= sin() x + cos() y.
If f (x, y) is some function of location on the plane, then write expressions for

f
f
f
,
, and
x0 y0
y 0 x0
x y0
in terms of
f
x

and
y
f
y
(3.44)
(3.45)

.
(3.46)
Interpret f /x)y0 geometrically as a directional derivative. (That is, f /x)y0 is the slope of f along
which curve in the plane?) Given this interpretation, does it have the expected limits as 0 and as
/2?
3.6. THE THERMODYNAMIC DANCE
63
3.15 Maxwell relations for a three-variable system

Suppose (x, y, z) satisfies
d = A(x, y, z) dx + B(x, y, z) dy + C(x, y, z) dz.
(3.47)
State three Maxwell relations relating various first derivatives of A(x, y, z), B(x, y, z), and C(x, y, z),
and a fourth Maxwell relation relating various second derivatives of these functions.
3.16 The cylinder model with three variables
In the three-variable cylinder model the cylinders are described by height h, radius r, and density
. The master function is mass M (h, r, ), and the master equation is
dM (h, r, ) = S(h, r) dr + T (r) dh + V (h, r) d,
(3.48)
where S(h, r) is the side area, T (r) is the top area, and V (h, r) is the volume. Perform a Legendre
transformation to a description in terms of the variables h, S, and using the new master function
(h, S, ) = M Sr.
(3.49)
a. Write down the new master equation.

b. Write down the three first-order Maxwell relations and confirm their correctness using explicit
formulas such as M (h, S, ) = S 2 /(4h).
c. Interpret (h, S, ) physically.
3.17 Multivariate chain rule
Invent a problem going through the chain rule argument of section 3.8.1 with the cylinder model.
3.18 Contours of constant side area
Find contours of constant side area preparing for/using the technique of section 3.8.3.
3.6
The Thermodynamic Dance
A more conventional name for this section would be Changes of Variable in Thermodynamics, but the
changes of variable involved always remind me of an elaborate dance where partners are exchanged and
where patterns seem to dissolve and then reappear, but in fact are always present, even if hidden.
3.6.1
Description in terms of variables (S, V, N )
We know that in many circumstances (e.g. pure, non-magnetic fluids) the thermodynamic state of a system
is uniquely specified by giving the entropy S, the volume V , and the particle number N . The master function
for this description is the energy
E(S, V, N )
(3.50)
64
whose total differential is

dE = T dS p dV + dN.
(3.51)
Remember that a master function is one from which all functions of interest can be obtained by simple
differentiation. . . the equation above, for example, shows directly that

E
p(S, V, N ) =
.
(3.52)
V S,N
Remember also that the variables V and N are just representative mechanical parameters. Other mechanical parameters such as particle mass m, particle radius rHS , or the numbers of two different species,
NH2 and NHe , will be relevant in other circumstances.
3.6.2
Description in terms of variables (T, V, N )
For many purposes the description in terms of S, V , and N is awkward. (Entropy, for example, can be difficult
to measure and to understand.) In these situations a more natural set of variables might be temperature
T , volume V , and number N . We can change2 to this set through a Legendre transformation (a process
sometimes called trading in an S for a T ) obtaining the new master function
F (T, V, N ) = E T S,
(3.53)
which is called the Helmholtz free energy3 and which has the associated master equation
dF = S dT p dV + dN.
(3.54)
(In some books, the Helmholtz free energy is denoted by A rather than F .)
Lots of information can be read directly from the master equation. For example, the entropy can be
found through

F
.
(3.55)
S(T, V, N ) =
T V,N
It is less obvious how to find the energy, but a little thought shows that

F
(F/T )
E(T, V, N ) = F + T S = F T
=
.
T V,N
(1/T ) V,N
(3.56)
This equation is frequently used and carries the name Gibbs-Helmholtz equation.
2 When we changed variable from (S, V, N ) to (E, V, N ) (equation 3.1), I argued that this inversion was mathematically
legitimate. Here I will not be so formal and just assume that everything will come out all right. (See, however, problem 3.27.)
Clifford Truesdell has called this the principle of thoughtless invertibility.
3 It is sometimes called the Helmholtz potential or the Helmholtz function. The term free in this name is of only
historical significance. Its origin lies in the fact that F is related to the maximum amount of energy that is available (free)
for conversion into work through an isothermal process.
65
One sometimes hears that energy must be regarded as a function of (S, V, N ) while Helmholtz free energy must
be regarded as a function of (T, V, N ). This is false. The energy E(T, V, N ) above is a perfectly good function
but it is not a master function. The energy is the master function for the description in terms of (S, V, N ) but
not for the description in terms of (T, V, N ). Similarly, one could find the Helmholtz free energy as a function of
(S, V, N ), but it would not be a master function.
Another immediate consequence of the master equation is that

F
p(T, V, N ) =
.
V T,N
(3.57)
This equation is extremely well know. . . it is just the equation of state, relating pressure, temperature,
volume, and number. Examples are the famous ideal gas equation of state
p(T, V, N ) =
N kB T
V
(3.58)
and the only slightly less famous van der Waals equation of state for non-ideal gases
p(T, V, N ) =
e 0 v0
kB T
,
V /N v0
(V /N )2
(3.59)
where v0 and e0 are positive empirical constants with the dimensions of volume and energy respectively.
Notice that S(T, V, N ), E(T, V, N ), and p(T, V, N ) are all interesting quantities, but none of them are master
functions, that is, none of them contain all the thermodynamic information. It is a common misconception that
the equation of state contains everything one could want to know about a substance. In fact, for example, helium
and nitrogen are both nearly ideal gases at room temperature (so they obey nearly the same equation of state)
but the specific heat of nitrogen is about 67% greater than that of helium (so they differ dramatically in some
thermodynamic properties).
Yet another immediate consequence of the master equation is the Maxwell relation

p
S
=
.
V T,N
T V,N
(3.60)
Each of the derivatives above is not only a mathematical expression, but also an invitation to perform an
experiment. The derivative on the left is measured through an experiment like the following: A sample in
a container of variable volume (such as a piston) is placed within a thermostatically controlled bath (so
that the temperature doesnt change) and is heated in a slow and carefully monitored way (so that the heat
absorbed quasistatically can be divided by the temperature to find the entropy change). As the substance
is heated at constant temperature, the volume of the piston must change. Dividing the heat absorbed by
the temperature and the volume change gives (for small volume changes) the derivative on the left. This
experiment is not impossible, but clearly it is difficult and expensive.
Consider in turn the experiment on the right. The sample is in a strong box container of fixed volume
and its pressure and temperature are measured as the temperature is changed. The change doesnt need to be
66
controlled carefully and the heat absorbed doesnt need to be monitored: you can just blast your sample with
a propane torch. Dividing the measured change in pressure by the measured change in temperature gives (for
small temperature changes) the derivative on the right. It is remarkable that the results of these two very
different experiments is always the same. . . and it shows how thermodynamics can save a lot of experimental
labor! But it is still more remarkablenearly unbelievablethat we know these two experiments give the
same results because we know that the entropy function S(E, V, N ) exists, even though we might not know
what that function is.4 (This Maxwell relation is discussed further in problem 3.23.)
3.6.3
Description in terms of variables (S, p, N )
I dont need to stop with this second description. Both of the descriptions above used volume as a variable.
Most experiments are actually performed with a constant pressure (namely one atmosphere) rather than
with a constant volume, suggesting that pressure should be one of our variables. We may start with the
description in terms of (S, V, N ) and then trade in a V for a p, obtaining a master function
H(S, p, N ) = E + pV,
(3.61)
called the enthalpy, and the related master equation

dH = T dS + V dp + dN.
(3.62)
We could pause here to write down derivatives of H(S, p, N ), Maxwell relations, and so forth, but this
would merely be repetitive of what we did in the last section. Instead, we ask for yet another description
with the advantages of using both temperature instead of entropy and pressure instead of volume.
3.6.4
Description in terms of variables (T, p, N )
In this description, the master function

G(T, p, N ) = F + pV
(3.63)
is called the Gibbs free energy (or the Gibbs potential or the Gibbs function)5 and the master equation
is
dG = S dT + V dp + dN.
(3.64)
4 If
a material existed for which these two experiments did not give identical results, then we could use that substance to
build a perpetual motion machine.
5 I prefer the name Gibbs potential because it is like electrostatic potential the quantity itself is not of intrinsic interest,
but one can find many interesting things by taking its derivatives. The name free energy is particularly inappropriate: What
do you buy from a power company? Youve been told Turn off the lamps to conserve energy, but this is silly: energy is
conserved whether the lamps are on or off. Theres plenty of thermal energy in any room, but you cant get work out of that
energy because the room has uniform temperature. So you dont buy energy from a power company, you in fact buy Gibbs
potential. For this reason it should be called Gibbs expensive energy. I am trying to get the federal government to change
the name of the Department of Energy to the Department of Gibbs Potential, but so far Ive had no success.
67
(It is most unfortunate that in some books the Gibbs free energy is denoted F , the symbol that most books
reserve for the Helmholtz free energy.)
It is immediately obvious that
G
(T, p, N ) =
N

.
(3.65)
T,p
It is also true that G is an extensive quantity, and that it is a function of only one extensive quantity, namely
N . It follows that G must increase linearly with N , whence
(T, p) =
G(T, p, N )
.
N
(3.66)
Notice that , which appeared from equation (3.65) to depend upon N , is actually independent of N .
The chemical potential has so far been a wallflower in the thermodynamic dance. But there is no reason
why we cannot trade in an N for a in the same way that we traded in a V for a p.
3.6.5
Description in terms of variables (T, V, )
Here the master function is

(T, V, ) = F N
(3.67)
d = S dT p dV N d.
(3.68)
with master equation
You might wonder why doesnt have a name. This is because

= F N = F G = pV
(3.69)
(T, V, ) = p(T, )V,
(3.70)
or, to put it more formally,
so already has the name of negative pressure times volume.
3.6.6
Intensive description in terms of variables (T, )
Putting equations (3.68) and (3.70) together gives

p dV V dp = S dT p dV N d
(3.71)
whence
S
N
dT +
d.
V
V
In other words, we have produced a new master function, the pressure
dp =
p(T, )
(3.72)
(3.73)
68
with master equation

dp = S dT + d
(3.74)
where we have used the volume densities

S = S/V
and
= N/V.
(3.75)
The master function p(T, ) differs in important ways from the others that we have seen. First, it is intensive
rather than extensive. Second, it is a function of two rather than three variables, and the two variables are
both intensive. It is clear that we cannot get any information about the system size out of p(T, ). . . thus it
fails the usual test of a master function, namely that it must provide all thermodynamic information. On
the other hand, p(T, ) provides all the information about intensive quantities. The problems in this chapter
demonstrate that p(T, ) is a surprisingly useful master function.
This entire section is summarized in appendix J.
Problems
3.19 Stumbling in the thermodynamic dance
a. From the thermodynamic assembly with variables T , p, and N , with master equation
dG = S dT + V dp + dN
(3.76)
one is tempted to produce an assembly with variables T , p, and , using master function
= G N.
(3.77)
Why does this fail?

b. Show that what is really desired is an intensive-only description in terms of the variables T and
p, for which the master equation is
d = s dT + v dp,
(3.78)
where s = S/N and v = V /N . (This result is called the Gibbs-Duhem equation.)

3.20 Not a master function
The function E(S, V, N ) is a master function: all thermodynamic quantities of interest can be obtained
by taking suitable derivatives of E(S, V, N ). For the variables (T, V, N ) the master function is the
Helmholtz free energy F (T, V, N ). While the energy E(T, V, N ) of course remains a function, it is no
longer a master function. Prove this to yourself by considering two different substances, A and B, with
free energies FA (T, V, N ) and FB (T, V, N ) = FA (T, V, N ) + aT V 2 /N . Show that these two substances
have identical energies E(T, V, N ) but different equations of state p(T, V, N ).
69
3.21 Thermodynamics of a new substance

The entropy of a newly discovered gas is determined to be

S(E, V, N ) =
kB
T0
1/2
kB T0 V 2
NE +
v02
1/2
,
where the constants T0 and v0 are positive, intensive quantities with the dimensions of temperature
and volume, respectively.
a. Verify that this equation is acceptable in its dimensions and in its extensivity.
b. Find T (S, V, N ) and p(S, V, N ) for this gas.
c. Find CV (T, V, N ) for this gas.
3.22 Energy from the Gibbs free energy
The text claims that any thermodynamic function can be obtained from a master function by taking
derivatives. To back up this claim, show that

G
G
E(T, p, N ) = G(T, p, N ) T
p
(3.79)
T p,N
p T,N

(G/p)
G
=
T
(3.80)
(1/p) T,N
T p,N

(G/T )
G
p
.
(3.81)
=
(1/T ) p,N
p T,N
Find an expression for F (T, p, N ) in terms of G(T, p, N ).
3.23 A Maxwell relation: special cases
a. Apply the Maxwell relation (3.60) to the the special case of an ideal gas (equation of state
pV = N kB T ) to show that
S(T, V, N ) = N kB ln(V /V0 (T )),
(3.82)
where V0 (T ) is an undetermined function of integration that differs from one ideal gas to another.
(The remarkable character of our Maxwell relation comes into sharp focus when applied to this
special case: The mechanical-type experiments which uncover the equation of state enable us to
determine much about the entropy function even in the absence of any heat-type experiments.)
b. Another special case worth examination is two-phase coexistence of liquid and gas. Verify the
Maxwell relation both for typical two-phase coexistencein which the low-temperature phase
has higher density than the high-temperature phaseand for the unusual casessuch as the
coexistence of water and icewhere the low-temperature phase has a higher density.
3.24 Chemical potential for mixtures
If a system contains a mixture of two chemical substances, say NA molecules of substance A and NB
70
molecules of substance B, then the list of mechanical parameters must be expanded, and the entropy
is a function
S(E, V, NA , NB ).
(3.83)
In this case the chemical potential of substance A is
S
A (E, V, NA , NB ) = T (E, V, NA , NB )
NA

(3.84)
E,V,NB
and similarly for B. Notice that the chemical potential of substance A depends upon the number of
molecules of B present.
a. Perform the thermodynamic dance down to the function F (T, V, NA , NB ) to show that

F
A (T, V, NA , NB ) =
NA T,V,NB
(3.85)
and similarly for B.

b. Argue that, because A and B are intensive, their functional dependence on V , NA , and NB ,
must be through the number densities A = NA /V and B = NB /V , i.e. that
A (T, V, NA , NB ) = A (T, A , B ).
(3.86)
c. Consider the Gibbs free energy G(T, p, NA , NB ) and modify the extensivity argument that produced equation (3.66) to show that
G(T, p, NA , NB ) = A (T, p, NA /NB )NA + B (T, p, NA /NB )NB ,
(3.87)
where the chemical potentials are independent of the overall numbers NA and NB but may depend
on their ratio.
3.25 More parameters
What is the generalization of
dE = T dS p dV + dN
(3.88)
for systems that can be lifted and tossed? (That is, for systems in which height h and velocity v are
mechanical parameters.)
3.26 Another mechanical parameter
Recall expression (2.32) for the entropy of a monatomic ideal gas, and recall that E = 23 N kB T . Find
expressions for the change in entropy and in volume if the masses of the particles are varied at constant
p, T , and N . Suggest experiments that could verify these relations.
3.27 Which variables to exchange?
In carrying out the thermodynamic dance, we have exchanged the variable S for the function T (S, V, N ),
and we have exchanged the variable V for the function p(T, V, N ), and more. Why would it never be
appropriate to exchange the volume for the temperature? (Clue: Consider the volume of water at
temperatures just above and below freezing.)
3.7. NON-FLUID SYSTEMS
3.7
71
Non-fluid Systems
All the results of the previous section relied on the underlying assumption that the system was a pure fluid
and thus could be described by specifying, for example, the temperature T , the volume V , and the particle
number N . Recall that the latter two were simply examples of many possible mechanical parameters that
could be listed, such as the molecular mass, the molecular moment of inertia, or pair interaction parameters
such as the hard-sphere radius or the Lennard-Jones parameters.
But of course nature presents us with many materials that are not pure fluids! One obvious example is
a fluid mixture, the specification of which requires the number of molecules of each constituent. Another
example is a crystal of a layered material such as graphite. To find the energy, for example, it is not sufficient
to specify only the volume. You must know the area of the layers and the height to which these layers are
stacked.
In this book we will focus instead on a third example, namely magnets. The fundamental thermodynamic
relation for magnetic systems is
dE = T dS M dH,
(3.89)
where H, the applied magnetic field, is an intensive mechanical parameter and M , the magnetization (total
magnetic dipole moment of the sample), is extensive. Just as the thermodynamic equations for fluids
presented in the previous section implicitly assume that the magnetic properties of the sample can be
ignored (either because the substance is non-magnetic or because the magnetic field does not change), so the
equation above implicitly assumes that the volume and number specification of the sample can be ignored.
In another course, you may have learned a mnemonic for remembering the thermodynamic differentials
and Maxwell relations of a pure fluid systems with a constant particle number. Such mnemonics encourage
the very worst problem solving strategy, namely poke around until you discover an equation that fits.
Anyone who uses this strategy finds it impossible to investigate mixtures, crystals, magnets, or any other
member of the rich array of materials that nature has so generously spread before us. Instead of memorizing
equations and hoping that the right one will present itself, you should think about what sort of equation
you will need to solve a problem and then derive it. Appendix J will remind you of the strategy of the
thermodynamic dance and will help you keep your signs straight.
3.28 Magnetic systems
Show that for magnetic systems (see equation (3.89)),

M
S
=
T H
H T
and
H
T

=
M
S
M
(3.90)

.
T
(3.91)
72
3.8
Thermodynamics Applied to Fluids
In the rest of this chapter we apply the general results of section 3.6, The Thermodynamic Dance, to
particular concrete situations. We begin with fluids, that is, systems whose thermodynamic states are
adequately described by giving the variables temperature T , volume V , and particle number N . Furthermore,
in this section we will not allow particles to enter or leave our system, so only two variables, T and V , are
needed.
3.8.1
Heat capacities
Is there any relation between

S
Cp (T, p) = T
T

,
(3.92)
the heat capacity at constant pressure, and

CV (T, V ) = T
S
T

,
(3.93)
the heat capacity at constant volume? Remembering that entirely different experiments are used to measure
the two quantities, you might guess that there is not. But mathematically, the difference between Cp and
CV is related to a change in variable from (T, V ) to (T, p), so you might begin to suspect a relation. [You
might want to prepare for this section by working problem 3.17.]
Begin with the mathematical definition of the total differential of entropy,

S
S
dS =
dT +
dV.
T V
V T
(3.94)
The above holds for any infinitesimal change. We restrict ourselves to infinitesimal changes at constant
pressure, and divide by dT , to find

S
S
S
V
=
+
.
(3.95)
T p
T V
V T T p
Multiplying both sides by T gives a heat capacity relation

S
V
.
Cp = CV + T
V T T p
(3.96)
This relationship is correct but not yet written in its most convenient form. For example, reference books
provide tabulations of Cp and CV , but not of S/V )T or V /T )p .
The first step in changing the heat capacity relation into standard form is an easy one. Recall from
problem 1.2 that the expansion coefficient (an intensive tabulated quantity6 ) is defined by

1 V
(T, p) =
(3.97)
V T p
6 It is easy to see why V /T ) itself is not tabulated. Instead of requiring a tabulation for iron, you would need a tabulation
p
for a five gram sample of iron, for a six gram sample of iron, for a seven gram sample of iron, etc.
3.8. THERMODYNAMICS APPLIED TO FLUIDS

whence
73
S
Cp = CV + T V
V

.
(3.98)
The second step is less obvious. At equation (3.60) I discussed the Maxwell relation

p
S
=
,
V T
T V
(3.99)
and mentioned that the right hand side was much easier to measure. Thus we have the preferable expression

p
Cp = CV + T V
.
(3.100)
T V
This expression is still not very convenient, however, because the quantity p/T )V is neither named nor
tabulated. Although easier to measure than S/V )T is, its measurement still requires a constant-volume
strong box. Measurements at constant pressure are easier to perform and hence more frequently tabulated.
We now write p/T )V in terms of such quantities.
The total differential of V (T, p) is
dV =
V
T
V
p
dT +
p

dp
(3.101)
or, in terms of named quantities,

dV = V dT V T dp.
(3.102)
(The isothermal compressibility

T (T, V )
1 V
V p

(3.103)
T
was defined in problem 1.2.) Restricting the above total differential to changes at constant V gives
V dT = V T dp
or
p
=
T
T
(3.104)

.
(3.105)
This immediately gives us the final form for our relationship between heat capacities,
Cp = CV + T V
2
.
T
(3.106)
This is a far from obvious result that is true for all fluids and, I emphasize, was derived assuming only that
entropy exists and without the use of any explicit formula for the entropy.
Note that T , V , and T are all positive quantities. The expansion coefficient may be either positive or
negative, but it enters the relationship as 2 . Thus
Cp CV .
(3.107)
74
Can we understand these results physically? Heat capacity is the amount of heat required to raise the
temperature of a sample by one Kelvin. If the sample is heated at constant volume, all of the heat absorbed
goes into increasing the temperature. But if the sample is heated at constant pressure, then generally the
substance will expand as well as increase its temperature, so some of the heat absorbed goes into increasing
the temperature and some is converted into expansion work. Thus we expect that more heat will be needed
for a given change of temperature at constant pressure than at constant volume, i.e. that Cp CV . We
also expect that the relation between Cp and CV will depend upon the expansion coefficient . It is hard,
however, to turn these qualitative observations into the precise formula (3.106) without recourse to the
formal mathematics that we have used. It is surprising, for example, that enters into the formula as 2 ,
so that materials which contract with increasing temperature still have Cp CV .
For solids and liquids, there is little expansion and thus little expansion work, whence Cp CV . But for
gases, Cp CV . What happens at two-phase coexistence?
So far we have discussed only the difference between the two heat capacities. It is conventional to also
define the heat capacity ratio
Cp
,
(3.108)
CV
which can be regarded as a function of T and V or as a function of T and p. But the result above shows
that for all values of T and V ,
(T, V ) 1.
(3.109)
We can apply these results to the ideal gas with equation of state
pV = N kB T.
(3.110)
Problem 1.2 used this equation to show that for an ideal gas,
=
1
T
and
T =
1
.
p
(3.111)
Thus the heat capacity relation (3.106) becomes

Cp = CV + T V
p
pV
= CV +
= CV + kB N,
T2
T
and we have
=
3.8.2
Cp
kB N
=1+
.
CV
CV
(3.112)
(3.113)
Energy as a function of temperature and volume
You know that E(S, V ) is a master function with the famous master equation
dE = T dS p dV.
(3.114)
For the variables T and V the master function is the Helmholtz free energy F (T, V ) with master equation
dF = S dT p dV.
(3.115)
75
Thus in terms of the variables T and V the energy in no longer a master function, but that doesnt mean
that its not a function at all. It is possible to consider the energy as a function of T and V , and in this
subsection we will find the total differential of E(T, V ).
We begin by finding the total differential of S(E, V ) and finish off by substituting that expression for dS
into the master equation (3.114). The mathematical expression for that total differential is

S
S
dS =
dT +
dV,
(3.116)
T V
V T
but we showed in the previous subsection that this is more conveniently written as
CV
dT +
dV.
T
T
dS =
Thus
(3.117)

CV
dE = T dS p dV = T
dT +
dV p dV
T
T
(3.118)

p dV.
dE = CV dT + T
T
(3.119)
and, finally,
The most interesting consequence of this exercise is that the heat capacity CV , which was defined in
terms of an entropy derivative, is also equal to an energy derivative:

E
S
=
.
(3.120)
CV (T, V ) T
T V
T V
For an ideal gas
p = 0,
T
(3.121)
whence
dE = CV dT + 0 dV.
(3.122)
Thus the energy, which for most fluids depends upon both temperature and volume, depends in the ideal gas only
upon temperature:
E(T, V ) = E(T ).
(3.123)
The same is true of the heat capacity
CV (T, V ) =
E
T
= CV (T ),
(3.124)
and the heat capacity ratio

(T ) = 1 +
kB N
.
CV (T )
(3.125)
Thermodynamics proves that for an ideal gas CV depends only upon T and not V , but experiment shows that
in fact CV is almost independent of T as well. From the point of view of thermodynamics this is just an
unexpected experimental result and nothing deeper. From the point of view of statistical mechanics this is not
just a coincidence but is something that can be proved. . . we will do so in chapter 5.
76
3.8.3
Quasistatic adiabatic processes
This topic, titled by a string of polysyllabic words, sounds like an arcane and abstract one. In fact it has
direct applications to the speed of sound, the manufacture of refrigerators, to the inflation of soccer balls, and
to the temperature of mountain tops. [You might want to prepare for this section by working problem 3.18.]
If a fluid is changed in a quasistatic, adiabatic manner, its entropy will remain constant during the change.
So, for example, if the state of the system is described by the two variables temperature and volume, then
during a quasistatic adiabatic process the system will move along contour lines7 of constant entropy in the
(T, V ) plane, such as those below.
Our aim in this section is to find an equation V(T) for such constant-entropy contours.
We begin with the now-familiar total differential for the entropy as a function of temperature and volume,
dS =
CV
dT +
dV.
T
T
(3.126)
This equation holds for any infinitesimal change, but we will apply it now to changes along an entropy
contour, i.e. changes for which dS = 0. For such changes
CV
dV =
dT
T
T
(3.127)
whence the differential equation for an entropy contour is

dV (T )
CV (T, V )T (T, V )
=
.
dT
T (T, V )
(3.128)
The quantities on the right are usually (not always) positive, so the entropy contours usually (not always)
slope downward. This is as much as we can say for a general fluid.
7 Such
contours are sometimes called isoentropic curves or, even worse, adiabats. The latter name is not just a linguistic
horror, it also omits the essential qualifier that only quasistatic adiabatic processes move along curves of constant entropy.
77
But for an ideal gas we can fill in the functions on the right to find
dV (T )
CV (T )
CV (T )V
=
=
.
dT
p(T, V )
N kB T
(3.129)
Using equation (3.125), this result is often written in the form

dV (T )
1
V
=
.
dT
1 (T ) T
(3.130)
This is as far as we can go for an arbitrary ideal gas.

But for an ideal gas with constant heat capacity, and hence with constant , this differential equation is readily
solved using separation of variables. We write
1
dV
dT
=
V
T
(3.131)
with the immediate solution

(1 ) ln V = ln T + constant,
(3.132)
giving
V 1 = constant T
or
T V 1 = constant.
(3.133)
Remembering that 1, this confirms the downward slope of the entropy contours on the (V, T ) plane.
This result is most frequently used with the variables p and T , rather than V and T :
pV = constant.
(3.134)
The above equation holds only for quasistatic, adiabatic processes in ideal gases that have heat capacities independent of temperature. You might think that it is such a restricted result that it holds little interest. This
equation is used frequently in applications, in the graduate record examination, and in physics oral examinations!
You should memorize it and know how to work with it.
Problems
3.29 Intensive vs. extensive variables
Equation (3.123), E(T, V ) = E(T ), states that for an ideal gas the energy is a function of temperature
alone. How is it possible for E, an extensive quantity, to be a function of only T , an intensive quantity?
3.30 Heat capacity at constant pressure
Equation (3.120) shows that the heat capacity at constant volume, which is defined in terms of an
entropy derivative, is also equal to an energy derivative. You might suspect a similar relation between
Cp (T, p, N ) and

E(T, p, N )
.
T
p,N
Show that such a suspicion is not correct, and instead find an expression for Cp in terms of a derivative
of enthalpy.
78
3.31 Qualitative heat capacity relations

For an ideal gas, Cp is greater than CV because:
a. The gas does work on its surroundings when its pressure remains constant while its temperature
is increased.
b. The heat input per degree increase in temperature is the same for processes at constant pressure
and at constant volume.
c. The pressure of the gas remains constant when its temperature remains constant.
d. The increase in the gass internal energy is greater when the pressure remains constant than when
the volume remains constant.
e. The heat needed is greater when the volume remains constant than when the pressure remains
constant.
3.32 Heat capacities in a magnetic system
For a magnetic system (see equation (3.89)), show that

S
M
S
, CM = T
, =
,
CH = T
T H
T M
T H
and T =
M
H

(3.135)
T
are related through

CM = CH T 2 /T .
(3.136)
3.33 Isothermal compressibility

a. Show that the isothermal compressibility, defined in problem 1.2 as

1 V
T =
,
V p T,N
is also given by
1
T =
p

T
1
= 2

T
1 2p
= 2
2
where is the number density N/V . Clue:

=
.
p T
T p T
(3.137)

,
(3.138)
(3.139)
b. What does this result tell you about the relation between density and chemical potential?
c. In part (a.) we began with a description in terms of the three variables T , p, N and then reduced
it to an intensive-only description, which requires just two variables, such as and T . Reverse
this process to show that

1 V
V N
T = 2
=
.
(3.140)
N T,V
N T,N
79
3.34 Pressure differential

By regarding pressure as a function of temperature T and number density , show that
1
dT +
d.
T
T
dp =
(3.141)
3.35 Isothermal vs. adiabatic compressibility

In class we derived a remarkable relation between the heat capacities Cp and CV . This problem
uncovers a similar relation between the isothermal and adiabatic compressibilities,

1 V
1 V
and S =
.
(3.142)
T =
V p T
V p S
The adiabatic compressibility S is the compressibility measured when the fluid is thermally insulated.
(It is related to the speed of sound: see problem 3.41.)
a. Use
S
dS =
T
T
p
to show that
S
dT +
p
=
S

dp
(3.143)
T
.
Cp /V
(3.144)
Sketch an experiment to measure this quantity directly.

b. From the mathematical relation
V
dV =
p
V
dp +
T

dT
(3.145)
derive the multivariate chain rule

V
p

S
V
=
p

T
V
+
T
whence
S = T
c. Finally, show that

p
T
p

,
2T
.
Cp /V
Cp
T
=
.
CV
S
3.36 Change of chemical potential with temperature

Prove that

S
= ,
T p,N
N
and that
Hows that for weird?

=
V,N
S
N

=
T,V
(3.146)
(3.147)
(3.148)
(3.149)
S
+
.
N
T
(3.150)
80
3.9
Thermodynamics Applied to Phase Transitions
The Clausius-Clapeyron equation relates the slope of a phase coexistence line px (T ) with the volume and
entropy changes when the sample crosses that line:
dpx
S
=
.
dT
V
(3.151)
Problems
3.37 Thermodynamics applied to the triple point
A triple point has the usual arrangement of a rising gas-solid coexistence line splitting into two rising
coexistence lines: a liquid-solid line and a gas-liquid line.
p
pls(T)
liquid
solid
pgl(T)
gas
pgs(T)
T
a. Relate the slopes of the three coexistence lines that meet at the triple point to the entropy changes
Sgl = Sgas Sliquid and Sls = Sliquid Ssolid .
b. Show that the gas-solid coexistence line, when extended beyond the triple point, must extend into
the liquid phase as show below left. It must not extend into the gas phase as shown below right.
p
p
liquid
liquid
solid
solid
gas
gas
T
Possible
T
Impossible
3.10. THERMODYNAMICS APPLIED TO CHEMICAL REACTIONS
81
c. Show that in most cases, at the triple point the gas-solid coexistence line has nearly the same
slope as the gas-liquid coexistence line.
d. The diagram below illustrates the Wikipedia article Triple point (15 December 2014). The
caption is A typical phase diagram. The solid green line applies to most substances; the dotted
green line gives the anomalous behaviour of water. Critique this diagram.
3.10
Thermodynamics Applied to Chemical Reactions
3.10.1
Thermodynamics of mixtures
So far in this book we have considered mostly pure substances. What happens if we have a mixture of
several chemical species, say the four substances A, B, C, and D? In this case the mechanical parameters
will include the numbers NA , NB , NC , and ND , and the entropy
S(E, V, NA , NB , NC , ND )
will have the total differential
(3.152)
dS =
X i
1
p
dE + dV
dNi .
T
T
T
(3.153)
i=A
The differential above defines the quantity

A (E, V, NA , NB , NC , ND ) = T (E, V, NA , NB , NC , ND )
S
NA

.
(3.154)
E,V,NB ,NC ,ND
Notice that the chemical potential of substance A can depend upon the number of molecules of B, C, and
D present, so it is not necessarily the same as the chemical potential of pure A. (For example, the chemical
82
potential of sugar dissolved in a cup of water is not equal to the chemical potential of pure sugar, unless the
solution is saturated and there is undissolved sugar at the bottom of the cup.)
It is convenient to define the total number of molecules
N = NA + NB + NC + ND
(3.155)
and the (intensive) fractional amounts

fA =
NA
, etc.
N
(3.156)
The thermodynamic dance can be followed for mixtures just as well as it can be for pure substances. One
noteworthy result is
G(T, p, NA , NB , NC , ND ) E T S + pV
D
X
=
i (T, p, fA , fB , fC , fD )Ni
(3.157)
i=A
3.10.2
Equilibrium condition for a chemical reaction
Suppose that molecules of substance A can combine with molecules of substance B to produce molecules
of substance C and molecules of substance D. If this can happen, then of course the reverse can also occur:
molecules of substance C may combine with molecules of substance D to produce molecules of substance
A and molecules of substance B. The first action is called a forward step of the chemical reaction, and
the second is called a backward step. This chemical reaction is symbolized by
A + B C + D.
(3.158)
Thus if we mix together a certain number of molecules each of A, B, C, and D, there can be a reaction and
we might not end up with the same distribution of molecules that we started off with. It is possible that
there will be more forward than backwards steps, in which case we will end up with more Cs and Ds and
fewer As and Bs than we had at the start, or it could go predominantly the other way. It is allowable for
the system to exist in a wide range of conditions, varying from mostly reactants to mostly products, but
experience teaches us that the overwhelmingly likely condition is one with some of each are present. In this
condition reactions still proceed, in both the forward and backward directions, but steps in both directions
proceed at the same rate so there is no net change in the number of molecules of any species. This condition
is called chemical equilibrium.
The above paragraph precisely parallels the discussion in section 2.7, Using entropy to find (define)
temperature and pressure. In that section we saw that if, for example, two systems could exchange volume,
it was possible for all the volume to go to one system or for all of it to go to the other system, but that the
overwhelmingly likely possibility was the equilibrium condition in which the two systems shared volume
in such a way that each had the same pressure. The equilibrium condition was the one with a maximum
number of microstates (or, equivalently, a maximum entropy) with respect to the distribution of volume.
83
Precisely the same effect is at work in chemical equilibrium. The equilibrium state is characterized by a
maximum in the entropy with respect to the number of number of reaction steps taken. At this equilibrium
point the entropy doesnt change when the chemical reaction proceeds by one step in either direction. The
change in entropy through a forward step is of course
S
S
S
S
+
+
,
NA
NB
NC
ND
(3.159)
and the change in entropy through a backward step is the negative of this. In either case, the equilibrium
condition is that this change in entropy vanish, i.e. that
A + B = C + D .
(3.160)
Interpretation in terms of escaping tendency. Specialization to the case of liquid-vapor phase equilibrium.
We cannot go further without formulas for the chemical potentials . Complete formulas are not available
within the realm of thermodynamics, but the next section shows that thermodynamics can put severe
constraints on the permissible functional forms for .
3.10.3
Chemical potential of an ideal gas
I emphasized on page 65 that the equation of state did not contain full thermodynamic information. Thus,
for example, knowledge of the equation of state V (T, p, N ) is not sufficient to uncover the master function
G(T, p, N ). On the other hand, that knowledge is sufficient to restrict the functional form of the Gibbs free
energy, which in turn restricts the possible functional forms of the chemical potential. In this subsection we
uncover that restriction.
Our strategy is to recognize that volume is the derivative of G(T, p, N ) with respect to p, and chemical
potential is the derivative of G(T, p, N ) with respect to N . Given V (T, p, N ), we can integrate with respect
to p to find G(T, p, N ), and then differentiate with respect to N to find (T, p). This process does not
completely determine the free energy or the chemical potential, because a constant of integration is involved.
But it does restrict the permissible functional forms of G(T, p, N ) and (T, p).
The master thermodynamic equation for variables (T, p, N ) is
dG = S dT + V dp + dN,
whence
G
V (T, p, N ) =
p
(3.161)

.
(3.162)
T,N
Thus for a substance obeying the ideal gas equation of state, we have

N kB T
G
=
.
p
p T,N
(3.163)
84
Integrating both sides with respect to p gives

Z
N kB T
or
dp
=
p
G
p

dp
(3.164)
T,N

p
= G(T, p, N ).
(3.165)
p(T, N )
The role of constant of integration is played by p(T, N ), which must be constant with respect to p but
which may vary with T and N .
N kB T ln
In fact, careful examination of the above equation shows what the dependence of p(T, N ) on N must be.
The quantities G and N above are extensive, while T and p are intensive, so p must be intensive so that
both sides will depend linearly on system size. Thus p is independent of N , and we have

p
G(T, p, N ) = N kB T ln
,
(3.166)
p(T )
whence

p
G(T, p, N )
= kB T ln
.
(3.167)
(T, p) =
N
p(T )
This is the form of the Gibbs free energy and of the chemical potential for any ideal gas. The function
p(T ) is undetermined, and it is through differences in p(T ) that one ideal gas differs from another. But the
dependence on p and on N are completely pinned down by the above results.
We have seen that the only expressions for (p, T ) consistent with the ideal gas equation of state for
a pure substance are those of the form (3.167). But what if the ideal gas is a mixture of, say, the four
chemical species A, B, C, and D? In this case, the chemical potential may be a function of p, of T , and of
the composition fractions fA , fB , fC , and fD . I know the result and will show it to you below. I am looking
for a purely thermodynamic argument that will prove it, but until I find one youll just have to trust me.

pfA
.
(3.168)
A (T, p, fA ) = kB T ln
pA (T )
3.10.4
The equilibrium condition for idea gas reactions
[Before plunging in, I want to point out that theres a contradiction embodied in the term ideal gas chemical
reaction. In an ideal gas, molecules do not interact. . . when two molecules approach, they just pass right
through each other. But a chemical reaction is a form of interaction! Clearly, in this section we are thinking
of molecules that interact weakly unless they are very close to each other, but once they approach they
interact strongly and react quickly.]
Putting together the general equilibrium condition for chemical reactions
A + B = C + D
with expression (3.168) for the chemical potential of an ideal gas gives the equilibrium condition

pfA
pfB
pfC
pfD
kB T ln
+ kB T ln
= kB T ln
+ kB T ln
.
pA (T )
pB (T )
pC (T )
pD (T )
(3.169)
(3.170)
85
The repetition of expressions like

pfA
pA (T )
in the equation above suggests that we should examine such expressions carefully. Notice that
N k B T NA
= A kB T,
pfA =
V
N
where A NA /V . This encourages us to define a characteristic density function
A (T ) =
pA (T )
kB T
Using the properties of logarithms, this can be rewritten as

B
C
D
A
=
,
A (T )
B (T )
C (T )
D (T )
C D
A B
C (T )
D (T )
B (T )
A (T )
(3.172)
(3.173)
and to write the equilibrium condition (3.170) as

A
B
C
D
ln
+ ln
= ln
+ ln
.
A (T )
B (T )
C (T )
D (T )
or, after minor rearrangement, as
(3.171)
(3.174)
(3.175)
(3.176)
Chemists like to write this equation in terms of the concentration of A, written as [A], namely as
[C] [D]
= K(T ),
[A] [B]
(3.177)
where K(T ) is called the equilibrium constant (despite the fact that it is a function of temperature). This
result is called the law of mass action despite the fact that it is not a fundamental law (we have just
derived it) and even despite the fact that (because it holds only for ideal gases) its not precisely true for
any real substance! (And also, I might add, despite the fact that the law has nothing to do with mass
or with action.) Even though the result is poorly named, it shows just how valuable thermodynamics is
in connecting disparate areas of research. Who would have guessed that the law of mass action follows from
only thermodynamics and the ideal gas equation of state? We have proven the result without once doing a
chemical experiment!
Problems
3.38 Chemical potential for ideal gas mixtures
The entropy S(E, V, NA , NB ) of a classical monatomic ideal gas mixture was uncovered using statistical
mechanics in problem 2.16. Invert the result of that problem to find E(S, V, NA , NB ), then show that
the temperature and pressure of an ideal gas mixture obey familiar results for pure ideal gases. Show
that
"
#
3/2
h20
A (T, p, fA , fB ) = kB T ln
pfA ,
(3.178)
2mA (kB T )5/3
and relate this expression to the form (3.168). Verify relation (3.158).
86
3.11
Thermodynamics Applied to Light
Another non-fluid system (see section 3.7).

See Robert E. Kelly, Thermodynamics of blackbody radiation, Am. J. Phys. 49 (1981) 714719, and
Max Planck, The Theory of Heat Radiation, part II.
How do you get a box of light? This is blackbody (cavity) radiation: meaning radiation in thermal
equilibrium. (Usually its also in equilibrium with the cavity walls, but you could imagine a container with
mirror walls. . . )
3.11.1
Fundamentals
The fundamental thermodynamic equation for this system involves the master function E(S, V ). It is
dE = T dS p dV.
(3.179)
This equation differs from previously encountered master equations in that there is no term for dN . From
the classical perspective, this is because radiation is made up of fields, not particles. From the quantal
perspective, this is because photon number is not conserved.
Thermodynamics binds one quantity to another, but it must use experiment (or statistical mechanics)
to find the values being bound. For example, thermodynamics tells us that
Cp = CV + T V
2
,
T
but it cannot calculate either Cp , or CV , or , or T : these quantities must be found by experiment (or
through a statistical mechanical calculation). The empirical result that we will employ is that for blackbody
radiation,
1
1E
= u.
(3.180)
p=
3V
3
This equation of state is the parallel for blackbody radiation to pV = N kB T for ideal gases. It was discovered
by experiment, but it can be derived from electrodynamics as well.
3.11.2
Energy density as a function of temperature
Consider the energy as a function of volume and temperature. Because E(V, T ) is extensive, but in the
argument list only V is extensive (recall that N doesnt appear in the argument list), we must have
E(V, T ) = V u(T ).
That is, the energy density u depends only upon temperature. What is this dependence?
(3.181)
3.11. THERMODYNAMICS APPLIED TO LIGHT
87
We seek a differential equation for u(T ). Compare

dE = d(V u) = V du + u dV = V
du
dT + u dV
dT
(3.182)
with
dE = T dS p dV = T dS 31 u dV.
(3.183)
Together these give us a formula for dS:

dS =
V du
T dT
4u
dT +
3T

dV.
The Maxwell relation associated with this differential is

[ ]
( )
=
V
T V
T
or
(3.184)
(3.185)

4 d[u/T ]
4 1 du
1
1 du
=
=
2u .
T dT
3 dT
3 T dT
T
(3.186)
u
du
=4
dT
T
(3.187)
dT
du
=
.
T
u
(3.188)
A rearrangement gives
or
4
The solution is
4 ln T = ln u + const
(3.189)
u(T ) = T 4 .
(3.190)
whence
The Stefan-Boltzmann law!
3.11.3
Quasistatic adiabatic expansion of radiation
Consider a sample of radiation undergoing quasistatic adiabatic change of volume. The entropy is a constant
during this process, although the value of that constant will of course depend on the particular sample thats
expanding.
dE = T dS p dV
(3.191)
but dS = 0, E = uV , and p = u/3 so
u dV + V du = 13 u dV
V du = 43 u dV
ln u = 34 ln V + const
u = KV 4/3
88
Recognizing that this constant will depend upon which adiabat is taken, i.e. that it will depend on the
entropy, we write
u(S, V ) = K(S)V 4/3 .
(3.192)
Using the Stefan-Boltzmann result u = T 4 we find that
T (S, V ) =
C(S)
.
V 1/3
(3.193)
This explains the cooling of the universe as it expands from the initial hot big bang to the current 3 K
microwave background.
3.11.4
Thermodynamics of the energy spectrum
What if we consider not just the energy density per volume, but the energy density per volume and wavelength? Let
u
(T, ) d
(3.194)
represent the energy per volume due to that radiation with wavelength between and + d. As you
know, quantum mechanics was discovered through Plancks efforts to find a theoretical explanation for the
measured function u
(T, ). This is not the place to describe Plancks work. Instead I want to focus on a
purely thermodynamic result that was known long before Planck started his investigations.
This is Weins law, which states that the function u
(T, ), which youd think could have any old form,
must be of the form
u
(T, ) = T 5 f (T ).
(3.195)
which
An immediate consequence of Weins law is the Wein displacement theorem: The wavelength
maximizes u
(T, ) is inversely proportional to temperature:
) = constant ,
(T
T
(3.196)
where the constant is the value of x that maximizes f (x). The consequences of the Wein displacement theorem
are familiar from daily life: low temperature objects (such as people) radiate largely in the infrared, moderate
temperature objects (such as horseshoes in the forge) radiate largely in the red, while high temperature
objects (such as the star Sirius) radiate largely in the blue.
I stated earlier that Weins law is a purely thermodynamic result. Thats almost true, but it also relies
on one more fact from electrodynamics, a result called no mode hoping:
If the volume makes a quasistatic, adiabatic change from V1 to V2 , then the light of wavelength
in the range 1 to 1 + d1 shifts into the range 2 to 2 + d2 where
1
1/3
V1
2
1/3
V2
(3.197)
3.11. THERMODYNAMICS APPLIED TO LIGHT
89
(This result may be derived rigorously from Maxwells equations, but its reasonable through this analogy:
A string of length L vibrating in, say, its third mode, has wavelength = 23 L. If the length of the string is
slowly changed, then the wave remains in its third mode, so /L is constant.)
Now were ready to begin the derivation. Consider the quasistatic adiabatic expansion of light, and while
that expansion is going on focus your attention on the light in wavelength range to + d. According to
the no mode hopping result, during this expansion the quantity
/V 1/3
(3.198)
remains constant during the expansion. Furthermore, according to equation (3.193) the quantity
T V 1/3
(3.199)
also remains constant. Multiplying these two equations, we find that the volume-independent quantity
T
(3.200)
remains constant during the expansion: this number characterizes the expansion.
Another such volume-independent constant can be found by repeating the reasoning of subsection 3.11.3,
Quasistatic adiabatic expansion of radiation, but considering not the energy of all the radiation, but the
energy of the radiation with wavelengths from to + . This energy is E = u
V , and the pressure due
1
. During a quasistatic adiabatic expansion, dE = p dV , so
to this segment of the radiation is p = 3 u
(V ) d
u + (
u) dV + (
uV ) d[] = 31 u
dV.
(3.201)
During the expansion the volume and wavelengths are changing through (see equation 3.197)
d
d[]
= cV 1/3
= c 13 V 2/3 dV =
=
1
dV
3V
1
dV
3 V
(3.202)
so we have
(V ) d
u = 35 u
dV.
(3.203)
Thus
V d
u =
ln u
=
u
=
35 u
dV
53 ln V + const
KV 5/3
(3.204)
But equation (3.193) shows how a larger volume is related to a lower temperature, so the quantity
u
(T, )
T5
(3.205)
90
remains constant during the expansion.

Thus we have two volume-independent ways to characterize the particular curve taken by this expansion.
In the thermodynamics of light, a state is specified by two variables, so a curve is specified by only one
parameter. Hence these two characterizations cannot be independent: one must be a function of the other.
Thus
u
(T, )
= f (T )
T5
or
u
(T, ) = T 5 f (T ).
(3.206)
Weins law.
Problems
3.39 Heat capacity of light
Show that, for blackbody radiation, CV = 4E/T .
Resources
Thermodynamic tables. (G.N. Lewis and M. Randall) Zemansky. Practical heat engines. Callen. Fermi.
Math book. e.g. Taylor and Mann?
Picture of Smithsonian crystal on www?
Thermodynamic data (e.g. steam tables) on www?
3.12
Additional Problems
3.40 Cool mountain air

Model the earths atmosphere as an ideal gas (nitrogen) in a uniform gravitational field. Ignore all
winds. Let m denote the mass of a gas molecule, g the acceleration of gravity, and z the height above
sea level.
a. Use ideas from Newtonian mechanics to show that the change of atmospheric pressure p with
height z is
mg
dp
=
p(z).
(3.207)
dz
kB T (z)
91
b. If the atmosphere is a poor conductor of heat, then the decrease in pressure with height is due to
an adiabatic expansion. (Clue: In other words: At the bottom of the mountain, fill an insulated
balloon with air at the local density, pressure, and temperature. Transport that balloon to the
top of the mountain. During the journey the balloon will expand, so the density, pressure, and
temperature will change. According to our assumption, during this journey the balloons density,
pressure, and temperature will match those of the atmosphere outside.) Show that under this
assumption
dp
p(T )
=
(3.208)
dT
1 T
and hence that
dT
1 mg
=
.
(3.209)
dz
kB
Evaluate this expression in degrees per kilometer for nitrogen, which has = 1.4.
c. In contrast, if the atmosphere were a good conductor of heat, then temperature would be uniform.
Find p(z) under such circumstances. Denote the sea-level pressure and temperature by p0 and
T0 .
d. Similarly find p(z) for an adiabatic atmosphere.
3.41 The speed of sound
When a sound wave passes through a fluid (liquid or gas), the period of vibration is short compared
to the time necessary for significant heat flow, so the compressions may be considered adiabatic.
Analyze the compressions and rarefactions of fluid in a tube. The equilibrium mass density is 0 .
Apply F = ma to a slug of fluid of thickness x, and show that if the variations in pressure p(x, t) are
small then pressure satisfies the wave equation
2
2p
2 p
=
c
t2
x2
(3.210)
where c, the speed of sound, is given by
1
.
0 S
Optional: Use the results of problems 1.2 and 3.35 to show that, for an ideal gas,
r
kB T
.
c=
m
c=
(3.211)
(3.212)
3.42 Thermodynamics of a plastic rod

(This problem is based on Reif problem 5.14.)
For a restricted range of lengths L and temperatures T , the tension force in a stretched plastic rod is
F (T, L) = aT 2 (L L0 ),
(3.213)
where a is a positive constant and L0 is the relaxed (unstretched) length of the rod. When L = L0 ,
the heat capacity CL of the rod (measured at constant length) is given by CL (T, L0 ) = bT , where b is
independent of temperature.
92
a. Write down the fundamental thermodynamic relation for this rod, expressing dE in terms of dS
and dL.
b. Compute (S/L)T . (Clue: Derive an appropriate Maxwell relation for the assembly with variables T and L.)
c. Knowing S(T0 , L0 ), integrate along an appropriate path to find S(T, L) at any temperature and
length within the range of applicability of the equation for F (T, L).
d. If you start at T = Ti and L = Li and then quasi-statically stretch a thermally insulated rod until
it reaches length Lf , what is the final temperature Tf ? Show than when L0 Li < Lf , the rod
is cooled by this process.
e. Find the heat capacity CL (L, T ) of the rod when its length is not necessarily L0 .
f. Find (T /L)S for arbitrary T and L. Can insulated stretches warm as well as cool the rod?
3.43 Magnetic cooling

At low temperatures, paramagnetic salts obey the Curie law
M =c
H
,
T
(3.214)
where c is a positive constant (see equation (3.89)). Assume that the heat capacity CH is a constant
independent of temperature and field. Suppose a sample at magnetic field Hi and temperature Ti is
wrapped in insulation, and then the magnetic field is slowly reduced to zero. Find the final temperature, and show that it is less than Ti . This technique, known as adiabatic demagnetization is the
refrigeration method used to produce the lowest temperatures that have yet been attained.
3.44 Thermodynamics of an electrochemical cell
Reif 5.16.
3.45 Thermodynamics and evolution
Read the essay Thermodynamics and Evolution by John W. Patterson, in Scientists Confront Creationism, Laurie R. Godfrey, ed. (Norton, New York, 1983), pp. 99116, on reserve in the science
library.
a. When a snowflake forms, its surroundings increase in entropy (become more disordered). What
is the name of the heat flow associated with this entropy change?
b. Patterson argues that S < 0 on Earth, due to biological evolution, and that S > 0 somewhere
else in the universe in order to make up for it. Where is that entropy increase taking place?
c. Patterson feels the need to invoke self-organization and Prigogine (pp. 110111) to explain how
his ram pumps could be made. Is this necessary? List two or more situations from nature in
which water does flow uphill.
93
3.46 Entropy and evolution

Creationists often claim that the second law of thermodynamics prohibits biological evolution.
a. The surface of the Sun (mean temperature 5778 K) heats the surface of the Earth (mean temperature 288 K) through visible and near-infrared radiation. The solar energy absorbed by the
Earth each second is 1.732 1017 J. What is the entropy change per second (due to this process)
of the Sun? The Earth? Does the entropy of Sun plus Earth increase or decrease?
b. Yet the mean temperature of the Earth changes slowly, if at all. This is because almost all of the
solar energy absorbed by the Earth is then emitted through far-infrared radiation which in turn
heats outer space the cosmic microwave background (CMB; temperature 2.728 K). What is
the entropy change per second (due to this process) of the Earth? The CMB? Does the entropy
of Earth plus CMB increase or decrease?
c. Now refine the model by supposing that, due to evolution, the entropy of the Earth is not exactly
constant, but is decreasing. (In this case the entropy of the CMB would have to be increasing
faster than rate predicted in part (b).) Suppose that, due to evolution, each individual organism
is 1000 times more improbable than the corresponding individual was 100 years ago. In other
words, if i is the number of microstates consistent with the specification of an organism 100 years
ago, and if f is the number of microstates consistent with the specification of todays improved
and less probable organism, then f = 103 i . What is the corresponding change in entropy
per organism?
d. The population of Earth is about 1018 eukaryotic individuals and 1032 prokaryotic individuals. If
the estimate of part (c) holds for each one of them, what is the change in entropy due to evolution
each second?
e. How accurately would you have to measure the entropy flux of part (b) in order to notice the
diversion of entropy flux calculated in part (d)? Has any scientific quantity ever been measured
to this accuracy?
f. It is generally agreed that the greatest rate of evolution fell during the Cambrian period, from
542 million years ago to 488 million years ago. During this so-called Cambrian explosion
multicellular organisms first formed and then radiated into remarkable variety. Suppose that
during the Cambrian period entropy was diverted into the evolution of living things at the rate
calculated in part (d). And suppose that at the end of the Cambrian there were 1018 multicellular
individuals. How much improved and less probable would each organism be, relative to its
single-celled ancestor at the beginning of the Cambrian period?
The moral of the story? Theres plenty of entropy to go around.
Chapter 4
Ensembles
After the last chapters dip into thermodynamics, we return to the discussion of the principles of statistical
mechanics begun in chapter 2.
4.1
The Canonical Ensemble
Examples of systems and baths:

Bottle of wine/a swimming pool.
Sample of gas/laboratory temperature control apparatus.
One ideal gas molecule/sample of gas.
But not one interacting gas molecule/sample of gas. (The Boltzmann distribution is for systems within
ensembles, not for molecules within systems.)
We denote a microstate by x. Exactly what is meant by a microstate will depend upon circumstances.
For a system of point particles it means specifying all the positions and momenta. For a spin system it
means specifying the up or down state of each spin. Well see other examples later.
Ill try to use different letters for microscopic and macroscopic (thermodynamic) quantities. For example
H(x) versus E (the thermodynamic energy is E = hH(x)i; M(x) versus M = hM(x)i.
94
4.1. THE CANONICAL ENSEMBLE
95
Summary
The canonical ensemble in general:
The probability that the system is in the microstate x is proportional to the Boltzmann factor
eH(x)/kB T .
(4.1)
The normalization factor is called the partition function or sum over all states (German Zustandsumme):
X
Z(T, V, N ) =
eH(x)/kB T .
(4.2)
microstates x
(Note that Z is independent of x.) Thus the probability that the system is in microstate x is
eH(x)/kB T
.
Z(T, V, N )
(4.3)
The connection to thermodynamics is that the Helmholtz free energy is

F (T, V, N ) = kB T ln Z(T, V, N ).
(4.4)
Note that in finding Z we sum over all microstates: the low-energy ones, the high-energy ones, the
orderly ones (e.g. all atoms heading west, or all atoms heading east), the disorderly ones (e.g. atoms
heading in scattered directions).
The canonical ensemble for a pure classical monatomic fluid:
The probability that the system is in the microstate is proportional to the Boltzmann factor
eH()/kB T .
(4.5)
Writing out all the normalizations correctly gives: the probability that the system is in some microstate
within the phase space volume element d about is
eH()/kB T
d,
N !h3N
0 Z(T, V, N )
(4.6)
where the partition function is

1
Z(T, V, N ) =
N !h3N
0
eH()/kB T d.
(4.7)
(The integral runs over all of phase space.) This is an example of partition function, namely the partition
function for a pure classical monatomic fluid. It does not apply to mixtures, to crystals, to the ideal
paramagnet. In contrast, the definition of partition function is equation (4.2), the sum over all states
of the Boltzmann factor.
96
CHAPTER 4. ENSEMBLES
4.2
Meaning of the Term Ensemble
An ensemble is just a collection. For example, a musical group is an ensemble. Think of a bunch of systems
filling up a huge gymnasium.
Microcanonical ensemble: A collection of many systems, all with the same volume, particle number,
particle mass, chemical composition, etc., and each with energy E < H(x) E + E.
Canonical ensemble: A collection of many systems, all with the same volume, particle number, particle
mass, chemical composition, etc., and each connected to the same heat bath (whence each with the
same temperature T ).
The names (selected by J. Willard Gibbs) are pretty poor.
4.3
Classical Monatomic Ideal Gas
Ive told you that its easier to do calculations in the canonical ensemble than in the microcanonical ensemble.
Today Im going to demonstrate the truth of this assertion.
Remember how we found the entropy S(E, V, N ) for a classical monatomic ideal gas? The hard part
involved finding the volume of a shell in 3N -dimensional space. It took us three or four hours to derive the
Sackur-Tetrode formula (equation (2.32)), namely

4mEV 2/3
5
(4.8)
S(E, V, N ) = kB N 32 ln
+
2 .
3h20 N 5/3
It took us another hour or so to show that the energy of this system is (equation (2.57))
E = 23 N kB T
whence
(4.9)

2mkB T V 2/3
5
+
(4.10)
2 .
h20 N 2/3
In this section, well derive this same equation using the canonical ensemble. While the derivation is no
stroll in the park, most people find it considerably easier than the microcanonical derivation.

S(T, V, N ) = kB N
3
2
ln
Our strategy will be: (1) Integrate the Boltzmann factor over all phase space to find the partition function
Z(T, V, N ). (2) Find the Helmholtz free energy using
F (T, V, N ) = kB T ln(Z(T, V, N )).
(4.11)
And finally (3) use the thermodynamic result

F
S(T, V, N ) =
T

(4.12)
V,N
4.3. CLASSICAL MONATOMIC IDEAL GAS
97
to find the entropy.

The Hamiltonian for this model system is
1 2
H() =
(p + p2y,1 + p2z,1 + p2x,2 + + p2z,N ) + U (x1 ) + U (x2 ) + + U (xN ),
2m x,1
where the potential energy function for each individual atom as a function of its location x is
(
0
if x is inside the container
U (x) =
if outside.
(4.13)
(4.14)
This has been background. Now were ready to roll up our sleeves and plunge into the work. Any questions?
Okay, lets go.
The canonical partition function is
Z
1
Z(T, V, N ) =
d eH()
(4.15)
N !h3N
0
Z +
Z +
Z +
Z +
Z +
Z +
1
=
dx
dy
dz
dx
dy
dzN
1
1
1
N
N
N !h3N
0
Z +
Z +
Z +
Z +
Z +
Z +
dpx,1
dpy,1
dpz,1
dpx,N
dpy,N
dpz,N eH() .
There are N triplets of integrals over positions and N triplets of integrals over momenta. . . in total, a 6N dimensional integral. Because the Hamiltonian consists of a sum of terms, the Boltzmann factor consists of
a product of quantities:
2
eH() = e(px,1 /2m) e(py,1 /2m) e(pz,N /2m) eU (x1 ) eU (x2 ) eU (xN ) .
(4.16)
More importantly, however, each of these quantities depends on a different set of variables. Therefore the
partition function also breaks up into a product of quantities:
Z +

Z +

Z +
Z +
Z +
Z +
1
U (x1 )
U (xN )
dx1
dy1
dz1 e
Z(T, V, N ) =

dxN
dyN
dzN e
N !h3N
0
Z +

Z +

2
2
dpx,1 e(px,1 /2m)

dpz,N e(pz,N /2m) .
The partition function, which in general is a 6N -dimensional integral, has been reduced in this case to a
product of N three-dimensional integrals over position and 3N one-dimensional integrals over momentum.
(This near-miraculous simplification comes about because all of the limits of integration are constants. . . if
the limits of integration of py,1 were functions of px,1 as they are in the microcanonical ensemblethen it
wouldnt happen.)
Moreover, the N position integrals differ only in the dummy variable of integration: they all have the
same value. The same holds for the 3N integrals over momentum. The partition function is nothing more
than
Z + Z + Z +
N
1
U (x)
dx
dy
dz e
Z(T, V, N ) =
N !h3N
0
Z +
3N
2
dp e(p /2m)
.
(4.17)
98
It is not difficult to evaluate the two integrals remaining. The position integral ranges over all space, but
for any point in space outside of the container, its integrand is
eU (x) = e() = 0.
(4.18)
Meanwhile, for any point in space within the container, the integrand is
eU (x) = e(0) = 1.
Thus
U (x)
dz e
dy
dx
(4.19)
Z Z Z
=
dx1 dy1 dz1 1 = V.
(4.20)
container
Meanwhile, the momentum integral is a clear cousin to the Gaussian integral (equation B.5)
Z +
2
eu du = .
(4.21)
p
Use of the substitution u = /2m p in the momentum integral gives
Z +
Z +
p
p
2
2
dp e(p /2m) = 2m/
eu du = 2mkB T .
(4.22)
Putting all this together gives us

Z(T, V, N ) =

3N/2
3N
1 2mkB T V 2/3
V N p
=
2mk
T
.
B
N!
h20
N !h3N
0
Weve completed the first step of our three-step plan. Now we must find the free energy
"
3N/2 #

1 2mkB T V 2/3
F (T, V, N ) = kB T ln Z = kB T ln
.
N!
h20
(4.23)
(4.24)
Somethings a bit funny here. If we double V and N , we should double F , but this doesnt exactly happen:
that factor of N ! is sure to mess things up. Whats gone wrong? We ran into the same problem in when
we used the microcanonical ensemble (at equation (2.24)) and it has the same solution: we must take the
thermodynamic limit. For large values of N ,
"

3N/2 #
1 2mkB T V 2/3
F (T, V, N ) = kB T ln
N!
h20
"
#
3N/2
2mkB T V 2/3
ln N !
= kB T ln
h20

2mkB T V 2/3
3
kB T 2 N ln
N ln N + N .
h20
But
N ln N = 23 N ( 23 ln N ) = 23 N ln(N 2/3 ),
4.4. ENERGY DISPERSION IN THE CANONICAL ENSEMBLE

so

F (T, V, N ) = kB T N
3
2

ln
99
2mkB T V 2/3
h20 N 2/3

+1 .
(4.25)
This feels a lot healthier: If we double both V and N , then the ratio (V 2/3 /N 2/3 ) is unchanged, so the free
energy exactly doubles.
Now the final step: Take the derivative with respect to temperature to find the entropy. To do this,
recognize that within the square brackets we have
3

2 ln(T ) + a lot of things independent of T .
With this realization the derivative becomes straightforward:

F
2mkB T V 2/3
3 1
3
+ 1 + kB T N 2
S(T, V, N ) =
= kB N 2 ln
T V,N
T
h20 N 2/3
or

S(T, V, N ) = kB N
3
2

ln
2mkB T V 2/3
h20 N 2/3

+
5
2
(4.26)

.
(4.27)
Wow. This is exactly the same as expression (4.10), which came from a (considerably longer) microcanonical
derivation.
Weve come a long way and it is certainly appropriate to take a moment here to bask in our triumph.
But while doing so, a small but troubling question might arise: Why should we get the same answer from
a canonical and a microcanonical calculation? After all, the ensembles in question are quite distinct: In a
microcanonical ensemble each system has the same energy. In a canonical ensemble the individual systems
can have any energy theyd like, as reflected by states that range from crystalline to plasma. Why should
we get identical entropies from such very different collections?
Loose remark: One often hears that for a gas at temperature T , the probability of an atom having energy
E is proportional to the Boltzman factor eE/kB T . This is true only for non-interacting atoms. If the
atoms interact then the kinetic energy belongs to atoms but the potential energy belongs to atom pairs, so
its impossible to define the energy of an atom. What is always true is that in the canonical ensemble at
temperature T , the probability of a system having energy E is proportional to eE/kB T .
4.4
Energy Dispersion in the Canonical Ensemble
The systems in the canonical ensemble are not restricted to having just one particular energy or falling
within a given range of energies. Instead, systems with any energy from the ground state energy to infinity
are present in the ensemble, but systems with higher energies are less probable. In this circumstance, it is
important to ask not only for the mean energy, but also for the dispersion (uncertainty, fluctuation, spread,
standard deviation) in energy.
Terminology: Uncertainty suggests that theres one correct value, but measurement errors prevent
your knowing it. (For example: You are 183.3 0.3 cm tall.) Dispersion suggests that there are several
100
values, each one correct. (For example: The mean height of people in this room is 172 cm, with a dispersion
(as measured by the standard deviation) of 8 cm.) Fluctuation is similar to dispersion, but suggests
that the value changes with time. (For example: My height fluctuates between when I slouch and when I
stretch.) This book will use the term dispersion or, when tradition dictates, fluctuation. Other books
use the term uncertainty.
The energy of an individual member of the ensemble we call H(x), whereas the average energy of the
ensemble members we call E:
H(x)
E
microscopic energy of an individual system
(4.28)
= hH(x)i = thermodynamic energy for the ensemble.
(4.29)
The dispersion in energy E is given through
E 2
= h(H(x) E)2 i
(4.30)
= hH (x) 2H(x)E + E i
= hH 2 (x)i 2E 2 + E 2
= hH 2 (x)i E 2
2
= hH 2 (x)i hH(x)i
(4.31)
This relation, which holds for the dispersion of any quantity under any type of average, is worth memorizing.
2
Furthermore its easy to memorize: The only thing that might trip you up is whether the result is hH 2 ihHi
2
or hHi hH 2 i, but the result must be positive (it is equal to E 2 ) and its easy to see that the average
of the squares must exceed the square of the averages (consider a list of data containing both positive and
negative numbers).
Now it remains to find hH 2 (x)i. Recall that we evaluated hH(x)i through a slick trick (parametric
differentiation) involving the derivative

ln Z
,
(4.32)
parameters
namely
X
1 Z
1
ln Z
=
=
Z
Z
!
X
eH(x)
H(x)eH(x)
eH(x)
= E
(4.33)
The essential part of the trick was that the derivative with respect to pulls down an H(x) from the exponent
in the Boltzmann factor. In order to pull down two factors of H(x), we will need to take two derivatives.
Thus the average hH 2 (x)i must be related to the second-order derivative

2 ln Z
.
(4.34)
2
parameters
4.4. ENERGY DISPERSION IN THE CANONICAL ENSEMBLE
101
To see precisely how this works out, we take

!
!
!
!
X
X
X
X
H(x)
2
H(x)
H(x)
H(x)
e
H (x)e
H(x)e
H(x)e
2 ln Z
x
x
x
x
=
!2
2
X
H(x)
e
x
X
=
H 2 (x)eH(x)
H(x)
H(x)eH(x)
H(x)
x
2
= hH 2 (x)i hH(x)i .
For our purposes, this result is better than we could ever possibly have hoped. It tells us that
2 ln Z
E T
E
=
=
.
2
T
To simplify the rightmost expression, note that
E 2 =
E
= CV
T
and
(4.35)
(1/kB T )
1
=
=
,
T
T
kB T 2
whence
E 2 = kB T 2 CV
or
E = T
kB CV .
(4.36)
This result is called a fluctuation-susceptibility theorem.

Analogy: susceptibility means herd instinct. A herd of goats is highly susceptible to external influence, because a small influence (such as a waving handkerchief) will cause the whole herd to stampede.
On the other hand a herd of cows is insusceptible to external influence. . . indeed, cowherds often talk about
how hard it is to get the herd into motion. A politician is susceptible if he or she is readily swayed by the
winds of public opinion. The opposite of susceptible is steadfast or stalwart. If a herd (or politician)
is highly susceptible, you expect to see large fluctuations. A herd of goats runs all over its pasture, whereas
a herd of cows stays pretty much in the same place.
How does E behave for large systems, that is in the thermodynamic limit? Because CV is intensive,
E goes up like N as the thermodynamic limit is approached. But of course, we expect things to go to
infinity in the thermodynamic limit! There will be infinite number, volume, energy, entropy, free energy, etc.
The question is what happens to the relative dispersion in energy, E/E. This quantity goes to zero in the
thermodynamic limit.
This resolves the question raised at the end of the previous section. A system in a canonical ensemble is
allowed to have any energy from ground state to infinity. But most of the systems will not make use of that
102
option: they will in fact have energies falling within a very narrow band about the mean energy. For larger
and larger systems, this energy band becomes more and more narrow. This is why the canonical entropy is
the same as the microcanonical entropy.
4.5
Temperature as a Control Variable for Energy (Canonical Ensemble)
Note distinction!
P(x) is the probability of falling in a particular microstate x. This happens to be a function only of
the energy of the microstate, whence this function is often called P(H(x)). (For a continuous
system, P(x) corresponds to p()d. . . the probability of the system falling in some phase
space point within the volume d about point .)
P (H) is the probability of having a particular energy. It is equal to P(H())(H). ((H): the
number of microstates with energy H.)
Temperature as persuasion.
Temperature as an energy control parameter.
Analogy: The Jim Smith effect. A man asks a statistician What is the probability that a man named
Milton Abramowitz will win the lottery? The questioner, whose name is Milton Abramowitz, is disappointed
when the statisticians answer is so tiny. So he tries again: What is the probability that a man named Jim
Smith will win the lottery? The statistician replies with a number that is still small, but not quite so small
as the first reply. In response, the man changes his name to Jim Smith.1
Because P(H()) decreases rapidly with H, and (H) increases rapidly with H, we expect P (H) to
be pretty sharply peaked near some hHi (recall the arguments of section 2.7.3). We also expect this peak
to become sharper and sharper as the system becomes larger and larger (approaching the thermodynamic
limit).
Even at high temperature, the most probable microstate is the ground state. However the most probable
energy increases with temperature.
In the canonical ensemble (where all microstates are accessible) the microstate most likely to be occupied is the ground state, and this is true at any positive temperature, no matter how high. The ground state
energy is not the most probable energy, nor is the ground state typical, yet the ground state is the most
probable microstate. In specific, even at a temperature of 1 000 000 K, a sample of helium is more likely to
be in a particular crystalline microstate than in any particular plasma microstate. However, there are so
many more plasma than crystalline microstates that (in the thermodynamic limit) the sample occupies a
plasma macrostate with probability 1.
1 Only
the most devoted students of American history know that Jim Smith signed the Declaration of Independence.
4.6. THE EQUIVALENCE OF CANONICAL AND MICROCANONICAL ENSEMBLES
103
Economic analogy for temperature as an energy control parameter: strict regulation versus incentives. Market-based approaches work more efficiently than clumsy command-and-control techniquesAlan
S. Blinder, Needed: Planet Insurance, New York Times, 22 October 1997, page A23.
4.1 Probability of microstate vs. probability of energy
P(H) is the probability of being in a particular microstate with energy H, whereas P (H) dH is the
probability of being in any microstate whose energy falls within the range H to H + dH. The number
of microstates with energy from H to H + dH is (H) dH.
Following the arguments of section 2.7.1 (Rapidly increasing character of the (H) function), assume
that (H) = cH . (For a classical monatomic ideal gas of N atoms, = (3/2)N .) Show that under
this assumption
Z(T, V, N ) = c(kB T )+1 ( + 1).
(4.37)
Does Z have the correct dimensions? Show that
P(H) =
eH/kB T
c(kB T )+1 ( + 1)
while
P (H) =
H eH/kB T
.
(kB T )+1 ( + 1)
(4.38)
Sketch (H), P(H), and P (H).
4.6
The Equivalence of Canonical and Microcanonical Ensembles
In a microcanonical ensemble the individual systems are all restricted, by definition, to have a given energy.
In a canonical ensemble the individual systems are allowed to have any energy, from that of the ground state
to that of an ionized plasma, but we have seen that (for large systems) they tend not to use this permission
and instead they cluster within a narrow band of energies. We might expect, therefore, that a calculation
performed in the microcanonical ensemble for systems at a given energy value, say 5 Joules, would give
the same results as a calculation performed in the canonical ensemble for systems at whatever temperature
corresponds to an average energy of 5 Joules.
If the canonical and microcanonical ensembles give the same results, why bother with the canonical
ensemble? We will see soon (section 5.1) that calculations are much easier to perform in the canonical
ensemble.
4.7
The Grand Canonical Ensemble
Grand as in the French big rather than as in the English magnificent. This ensemble is bigger than
the canonical ensemble in that there are more possible microstates.
104
Summary
The grand canonical ensemble for a pure classical monatomic fluid:
The probability that the system has N particles and is in the microstate N is proportional to
e(H(N )N ) ,
(4.39)
where
1
.
(4.40)
kB T
Writing out all the normalizations correctly gives: the probability that the system has N particles and is in
some microstate within the phase space volume element dN about N is
=
e(H(N )N )
dN ,
N !h3N
0 (, V, )
(4.41)
where the grand canonical partition function is

(, V, )
=
=
1
N !h3N
0
N =0
e(H(N )N ) dN
eN Z(, V, N ).
(4.42)
(4.43)
N =0
This sum is expected to converge when is negative.

The connection to thermodynamics is that
(T, V, ) = p(T, )V = kB T ln (T, V, ).
4.8
(4.44)
The Grand Canonical Ensemble in the Thermodynamic Limit
This section uncovers results that are of interest in their own right, but it also serves as an example of a
mathematically rigorous argument in statistical mechanics.
Definitions. The grand partition function is
eN Z(T, V, N ),
(4.45)
F (T, V, N ) = kB T ln Z(T, V, N ),
(4.46)
(T, V, ) =
N =0
the Helmholtz free energy is
and the chemical potential is

F (T, V, N )
(T, V, N ) =
N

.
T,V
(4.47)
4.8. THE GRAND CANONICAL ENSEMBLE IN THE THERMODYNAMIC LIMIT
105
Theorem. Consider a grand canonical ensemble of hard-core particles at equilibrium with a bath of
(a
temperature T and number control parameter . There is a most probable value of N , called N
function of T , V , and ), satisfying
)
(T, V, N
=
,
(4.48)
kB T
such that, in the thermodynamic limit,
)N
+ F (T, V, N
).
kB T ln (T, V, ) (T, V, N
(4.49)
Strategy. The theorem assumes hard-core particles so that there will be a maximum number of particles
in any given volume, whence the sum (4.45) is finite rather than infinite, and convergence questions do not
arise. (Less restrictive conditionse.g. that the repulsion between two nearby particles increases rapidly
enough as they are brought togethermay be used instead of the hard-core condition, but then the proof
becomes more technical and less insightful.) The proof works by establishing both upper and lower bounds
on sum (4.45), and then showing that, in the thermodynamic limit, these two bounds are equal.
Proof. Assume that each hard-core particle has volume v0 . Then the maximum number of particles
that can fit into a container of volume V is N V /v0 . Thus the sum in equation (4.45) does not go to ,
but stops at N .
We will need a trivial mathematical result, called the method of the maximum term: If fN is a sequence
of positive terms, N = 0, 1, . . . N , and if the maximum element of the sequence is max{fN }, then
max{fN }
N
X
fN (N + 1) max{fN }.
(4.50)
N =0
Applying this result to the sum (4.45), we obtain

max{eN ZN } (N + 1) max{eN ZN },
(4.51)
where ZN is shorthand for Z(T, V, N ). Taking the logarithm of each side, and realizing that ln max{f } =
max{ln f }, gives
max{N + ln ZN } ln ln(N + 1) + max{N + ln ZN }.
(4.52)
To prepare for taking the thermodynamic limit, we divide both sides by V and employ the definition FN =
kB T ln ZN , resulting in

FN
ln
ln(N + 1)
N
FN
N
+ max
.
(4.53)
max
V
kB T V
V
V
V
kB T V
Consider the difference between these upper and lower bounds. It is clear that
0
ln(N + 1)
ln(V /v0 + 1)
.
V
V
(4.54)
106
In the limit as V , the right hand side above approaches 0, so

ln(N + 1)
0.
V
(4.55)
Thus in the thermodynamic limit the lower bound approaches the upper bound, and

N
FN
ln
max
.
V
V
kB T V
(4.56)
Now we need to find the maximum of the quantity in curly brackets above. The maximum is located at
, which we find by taking the partial derivative with respect to N and setting it equal to zero:
N =N

1
FN
=
.
(4.57)
V
kB T V N N =N
becomes
Using the definition (4.47), the equation for N
=
)
(T, V, N
.
kB T
(4.58)
implies that, for sufficiently large values of V ,

Returning to equation (4.56), this location of N
) N
FN
(T, V, N
ln
=
,
V
kB T
V
kB T V
(4.59)
from which the final result (4.49) follows immediately.

Resume. The system under study can exchange energy and particles with a heat bath that has energy
control parameter = 1/kB T and number control parameter . The probability that the system contains
exactly N particles is
eN Z(, V, N )
PN =
.
(4.60)
(, V, )
, for which
The proof shows that, in the thermodynamic limit, there is one particular value of N , namely N
PN approaches 1. All the other PN s, of course, approach 0. The condition for locating N is just that it
is located, the result
gives the maximum value of PN . . . this condition gives rise to equation (4.48). Once N
PN 1 becomes equation (4.49).
4.9
Summary of Major Ensembles

boundary
variables
microcanonical
adiabatic
(no-skid)
E, V, N
canonical
heat bath
T, V, N
grand canonical
heat bath,
with holes
T, V,
probability of microstate
p.f.
master function
S(E, V, N ) = kB ln
deH() 1
Z
N !h3N
0
F (T, V, N ) = kB T ln Z
dN eH(N )N 1
N !h3N
0
(T, V, ) = kB T ln
d 1
N !h3N
0
or
4.10. QUANTAL STATISTICAL MECHANICS
107
In all cases, the partition function (p.f. in the above table) is the normalization factor
X
p.f. =
unnormalized probability.
microstates
4.10
Quantal Statistical Mechanics
How shall we define a partition function for quantal systems? A reasonable first guess is that
X
Z(T, parameters) =
eE .
[[first guess]]
all quantal states
This guess runs into problems immediately. Most quantal states are not energy eigenstates, so its not clear
how to interpret E for such states. Another difficultyharder to seearises because there are many
more quantal states than classical states, so the guess above is in some sense counting too many states. Its
hard to see how it could have the correct classical limit. (Indeed, a mathematical investigation of the above
guess would reveal that it doesnt even have the correct thermodynamic limitif you were to calculate the
Helmholtz free energy from the above guess, the resulting F (T, V, N ) would increase far faster than linearly
with system size.)
A second guess might be
Z(T, parameters) =
eE .
[[second guess]]
all energy eigenstates
This avoids the problem of how to interpret E for non-energy eigenstates, but there still seems to be an
over-counting of states in the case of an energy degeneracy: If two energy eigenstates are separate in energy
by any amount, no matter how small, those two energy eigenstates will contribute two terms to the above
sum. But if the energy separation vanishes, then all linear combinations of the two states are also energy
eigenstates, and each of this infinite number of combinations will enter into the sum.
Finally we arrive at a workable definition for quantal partition function, namely
X
Z(T, parameters) =
eEn .
(4.61)
energy eigenbasis n
Note that this is a sum over an energy eigenbasis, not over energy eigenvalues. These two sums differ through
degeneracy: If E6 = E5 , then two equal terms enter into the sum. More formally: if N (m) is the degeneracy
of eigenvalue m, that is, the number of linearly independent states with energy Em , then
X
Z(T, parameters) =
N (m)eEm .
(4.62)
energy eigenvalues m
The above argument has a strange, ex post facto character: We define the quantal partition function
not from first principles, but to insure that it has properties that we consider desirable (namely, that it
avoids subtle interpretation issues, it posses the expected classical limit, and it has desirable behavior at
108
level crossings). Linear combinations of energy eigenstates are quantal states just as legitimate as energy
eigenstates, so why do we ignore them in the sum? The argument reminds me of politics at its worst: Im
going to pass this law because I want it, even though I cant support it with any rationale. The resulting
definition of partition function is strange in that it seems to select the energy eigenbasis as some sort of
God-given basis, better than any other basis. In fact, the result is not as bad as it seems. It turns out
(see problem 4.2) that our definition can be rephrased in a basis-independent manner as
X
X
Z(T, parameters) =
eEn =
hj|e H |ji.
(4.63)
energy eigenbasis n
any basis j
dont worry. For this book

If you dont know what it means to exponentiate the Hamiltonian operator H,
and for most practical calculationsthe most effective definition is the form on the left. The point is simply
that the definition of partition function can be cast into the basis-independent form on the right. The energy
eigenbasis is not special.
In this chapter we have often found, first, results about probabilities, and second, connections to thermodynamics. This section has so far dealt with only the second element. What do these results imply for
probabilities? I will often say things like The probability of having energy Em is
N (m)eEm
.
Z
On the face of it, this statement is absurd, because most of the microstates arent energy eigenstates so they
dont have any energy at all. What I mean is If the energy is measured, the probability of finding energy
Em is
N (m)eEm
.
Z
I will often use the first statement rather than the long and formal second statement. I do so, however,
fully realizing that the first statement is wrong and that I use it just as convenient shorthand for the second
statement. You may use this shorthand too, as long as you dont mean it.
The ultimate source of such problems is that the English language was invented by people who did not understand
quantum mechanics, hence they never produced concise, accurate phrases to describe quantal phenomena. In the
same way, the ancient phrase Search the four corners of the Earth is still colorful and practical, and is used
today even by those who know that the Earth isnt flat.
Similarly, I will often sayThe probability of being in the energy eigenstate n (x1 , . . . , xN ) with energy
En is
eEn
.
Z
But what I really mean is If the microstate is projected onto a basis which includes the energy eigenstate
n (x1 , . . . , xN ), then the probability of projecting onto that state is
eEn
.
Z
4.11. ENSEMBLE PROBLEMS I
109
4.2 The trace

the trace of A,
defined as
Show that for any operator A,
X

trace{A}
hj|A|ji,
(4.64)
any basis j
is independent of basis. What is the trace in terms of the matrix representation of the operator?
4.3 Entropy in the canonical ensemble
From the partition function (4.61), one can find the free energy F and hence the entropy S. Show that
the entropy obtained in this manner equals
X
S(T ) = kB
pn ln pn ,
(4.65)
n
where
pn =
eEn
Z(T )
(4.66)
is the probability that the system is in its nth energy eigenstate (using the shorthand mentioned above).
This expression for the entropy is particularly useful in information theory. (Clue: Work forward from
S = F/T and backwards from equation (4.65) to meet at the middle.)
4.4 Entropy in the microcanonical ensemble
[You must work the previous problem before working this one.] What are the state probabilities
pn , analogous to (4.66), for an system in the microcanonical ensemble? Devise a suitable quantal
replacement for the microcanonical definition of entropy (2.7). Cast your replacement into the form
of a function of the microcanonical probabilities pn . Is formula (4.65) correct in the microcanonical
ensemble?
4.11
Ensemble Problems I
4.5 Classical monatomic ideal gas in the canonical ensemble

In section 5.1 we will show that the canonical partition function of a classical monatomic ideal gas is

N
1
V
Z(T, V, N ) =
,
(4.67)
N ! 3 (T )
where
(T )
h0
.
2mkB T
a. Show that in the thermodynamic limit, the Helmholtz free energy per particle is

V /N
F (T, V, N )
= kB T ln
+
1
.
N
3 (T )
b. Differentiate with respect to T to find S(T, V, N ).
(4.68)
(4.69)
110
c. Change variables to S(E, V, N ) using E = 23 N kB T , and compare the resulting expression to the
entropy derived in class through the microcanonical ensemble.
4.6 Number fluctuations in the grand canonical ensemble

Calculate N , the root mean square fluctuations in the particle number, for a system in the grand
canonical ensemble. (Clue: Take derivatives with respect to .) Show that this result is related to the
isothermal compressibility T through
r
kB T T
N
=
.
(4.70)
N
V
(Clue: Use the results of problem 3.33.)
4.7 Classical monatomic ideal gas in the grand canonical ensemble
a. Use the fact that
Z(T, V, N ) =

N
1
V
,
N ! 3 (T )
(4.71)
to find (T, V, ) for the classical monatomic ideal gas.

b. Use the connection to thermodynamics for the grand canonical ensemble, namely
p(T, )V = kB T ln (T, V, ),
(4.72)
to show that for any substance

N (T, V, ) = kB T
ln

.
(4.73)
T,V
c. Derive the ideal gas equation of state in the grand canonical ensemble.
4.8 Isobaric ensemble
This problem encourages you to think about an ensemble with a specified temperature, pressure, and
particle number. A physical model to keep in mind is a collection of systems each with the same number
of particles and all connected to the same heat bath but enclosed in balloons rather than in rigid boxes.
I want you to guess the answers to the following questions rather than produce mathematical or verbal
arguments. If you are uncomfortable making guesses in a physics problem assignment then do only
the first two parts.
a. Given that T , p, and N are fixed, what quantities fluctuate?
b. In going from the canonical to the isobaric ensemble, what control parameter (comparable to
= 1/kB T or = /kB T ) must be introduced, and how is it related to the pressure?
c. What is the (unnormalized) probability of finding the system in microstate x with energy H(x)
and volume V (x)?
d. What is the relevant state sum (comparable to Z or ), and how is it related to the thermodynamic
master function G(T, p, N )?
111
e. The relevant fluctuation-susceptibility relation connects the fluctuations in volume to what susceptibility?
f. Write out the fluctuation-susceptibility relation, using B to represent a product of several unguessable intensive quantities but writing out the extensive quantities explicitly.
g. What are the dimensions of B? Can you use this knowledge to guess at the unguessable?
4.9 Fluctuation-susceptibility relation for a magnetic system
In the Ising model for a magnet, the spins are located at fixed sites (labeled by i), and can point either
up or down (si = 1). The Ising Hamiltonian is
H = H0 (s1 , . . . , sN ) mH
N
X
si ,
(4.74)
i=1
where m is the magnetic moment of each spin, H is the applied magnetic field, and H0 represents some
spin-spin interaction energy, the details of which are not important here except for the fact that it is
independent of H. The microscopic magnetization, which varies from configuration to configuration,
is
X
M(s1 , . . . , sN ) = m
si ,
(4.75)
i
while the macroscopic (or thermodynamic) magnetization is the canonical

average over all such microscopic magnetizations:
M (T, H) = hMi.
(4.76)
The magnetic susceptibilityeasily measured experimentallyis

M
.
T (T, H) =
H T
(4.77)
Show that in the canonical ensemble the fluctuations in magnetization are related to the susceptibility
by
p
(4.78)
M = kB T T .
4.10 Two definitions of magnetization

The previous problem gave a statistical mechanical definition of magnetization:
M (T, H) = hMi,
(4.79)
while equation (3.89) gave a thermodynamic definition:

E
M (S, H) =
H
Show that these two definitions give equivalent results.

.
S
(4.80)
112
4.11 Values for the approach to the thermodynamic limit

In problem 2.13, The approach to the thermodynamic limit, you were given a value of EV 2/3 /h20
and told that it was appropriate for a gas at room temperature and atmospheric pressure. Justify
this value as
EV 2/3 /h20 = [ 32 (kB T )5/3 /(p2/3 h2 )]N 5/3
(4.81)
4.12 Integrals by parametric differentiation

To find the mean energy or energy dispersion in the canonical ensemble, we introduced the slick trick
of differentiating with respect to . This trick, called parametric differentiation, is in fact useful in
a number of circumstances. For example, in a graduate electrodynamics class I was once assigned a
problem (Jackson, Classical Electrodynamics, problem 14.5(b)) for which I needed to evaluate
Z 2
cos2
d
(4.82)
(1 + b sin )5
0
where b is a constant with |b| < 1. My classmates and I tried everything: substitutions, partial
fractions, conversion to a contour integral in the complex plane, and, most popular of all, look it
up in reference books. I puzzled over it for a dozen hours before giving up. When the solutions
were handed out, I expected to find many pages devoted to evaluating the integral. Instead I found a
two-line argument that began with the known integral (see Dwight 858.546 or Gradshteyn and Ryzhik
3.644.4)
Z 2
2
cos2
d =
(4.83)
a
+
b
sin
a
+
a2 b2
0
and then took a fourth derivative. What was the argument, and what is the value of the integral?
4.13 Parametric differentiation in quantum mechanics
The previous problem showed how parametric differentiation could be useful in electrodynamics. This
one shows how it can be useful in quantum mechanics. In quantum mechanics, an infinite square well
of width has energy eigenfunctions
r
2
sin(nx) n = 1, 2, 3, . . . .
(4.84)
n (x) =
One frequently needs to evaluate matrix elements such as

Z
2
hn|x2 |mi =
sin(nx)x2 sin(mx) dx.
0
(4.85)
Show how this integralas well as a host of other useful matrix elementscan be obtained easily from
the well known result

Z
1 sin(a b) sin(a + b)
sin(ax) sin(bx) dx =
a 6= b.
(4.86)
2
ab
a+b
0
113
4.14 Polymers
A primitive model for a polymer is a random walk on a simple cubic lattice. A random walk consists of
n steps (or links) starting at (anchored to) the origin. (In this model a polymer unit can step back
onto a lattice site already occupied by a different polymer unit. This unrealistic feature is corrected in
a more sophisticated model, the so-called self avoiding walk.)
a. Show that the number of distinct walks consisting of n links is Nn = 6n . Does this formula hold
when n = 0?
For many purposes it is valuable to consider the ensemble of all random walks, regardless of their
size. In this ensemble there is a size control parameter such that the probability of finding a
walk x consisting of n(x) links is proportional to en(x) . (Thus longer walks are less probable in this
ensemble, but there are more of them.) The partition function associated with this model is
X
() =
en(x) .
(4.87)
walks x
b. Show that the mean walk size in this ensemble is a function of given through
hni =
ln ()
.
(4.88)
c. Show that
1
1 6e
Clue: The geometric series sums to
() =
and that
1 + x + x2 + x3 + =
1
1x
hni =
when
6
.
6
(4.89)
|x| < 1.
d. What is the smallest possible value of the control parameter ? Does large correspond to long
polymers or short polymers?
e. Show that the dispersion in n is given through
(n)2 =
2 ln
,
2
whence
n
=
hni
e
=
6
1
+ 1.
hni
(4.90)
(4.91)
Thus the relative dispersion decreases for longer polymers.

(See also loops: S. Leibler, R.R.P. Singh, and M.E. Fisher, Thermodynamic behavior of two-dimensional
vesicles, Phys. Rev. Lett. 59 (1987) 19891992; C. Richard, A.J. Guttmann, and I. Jensen, Scaling
function and universal amplitude combinations for self-avoiding polygons, J. Phys. A 34 (2001) L495
501.)
114
4.15 A class of new ensembles

Any mechanical parameter in the Hamiltonian can be traded in for a control parameter. This
problem details the procedure if the mechanical parameter is intensive and an intrinsically positive
quantity. Examples to keep in mind are 1) the particle mass, 2) the spacing between planes of a high
temperature superconductor, or 3) the natural frequency of vibration for a diatomic molecule modeled
as a simple harmonic oscillator. In honor of this last example I will call the parameter . In the
micro- ensemble, all of the systems in the ensemble have a particular value of the parameter. In
the grand- ensemble, the systems in the ensemble have a variety of values for the parameter, but
larger values are less probable. In fact, the probability is proportional to eN , where is a control
parameter which can be adjusted to set the mean value of to any desired positive value. We suspect
that, in the thermodynamic limit, the results of the two ensembles will be identical when the mean
value of in the grand- ensemble is the same as the fixed value of in the micro- ensemble.
a. The thermodynamic assembly corresponding to the micro- ensemble has as master function the
Helmholtz free energy F (T, V, N, ). Using the definition

F
B(T, V, N, ) =
,
(4.92)
T,V,N
write out the differential equation for dF .
b. Execute a Legendre transformation to an assembly with master function
K(T, V, N, B) = F + B,
and show that in this assembly, the function (not variable!) is

K
.
(T, V, N, B) =
B T,V,N
(4.93)
(4.94)
In the grand- ensemble, the probability that a system has a parameter falling within d about is
eN Z(T, V, N, ) d
R
.
eN 0 Z(T, V, N, 0 ) d 0
0
(4.95)
Define the partition function

Y (T, V, N, ) =
1
0
eN Z(T, V, N, ) d,
(4.96)
where 0 is an arbitrary constant with the dimensions of whose only purpose is to make Y dimensionless. Presumably, the relationship between thermodynamics and statistical mechanics in the grand-
ensemble is
K = kB T ln Y,
(4.97)
by analogy to the similar relation in the canonical ensemble. We will assume this relationship and use
it, first to find the relationship between the statistical mechanical and the thermodynamic B and
second to show that the statistical mechanical equation (4.96) is consistent with the thermodynamic
equation (4.93).
115
c. Show that, in the grand- ensemble,

1 ln Y
N
(4.98)
1 2Y
.
N 2 2
(4.99)
hi =
and
2 =
Note that hi and are functions of T , V , N , and .

d. Compare equations (4.94) and (4.98) and conclude that, if (4.97) is to hold, then
B = N kB T .
e. Show that
s
=
kB T
B
(4.100)

.
(4.101)
T,V,N
f. We suspect that, in the thermodynamic limit N, V , ln Y N . If true, show that hi is
intensive (independent of N ) whereas falls like 1/ N .

g. To evaluate Y () in the thermodynamic limit, write
Z
1
Y (T, V, N, ) =
eN Z(T, V, N, ) d eN hi Z(T, V, N, hi)
.
0 0
0
(4.102)
Take logarithms to show that

ln Y N hi + ln Z + ln(/0 ),
(4.103)
and argue that in the thermodynamic limit the last term becomes negligible. Finally conclude
that
K = kB T N hi + F
(4.104)
or, in light of relationship (4.100),
K = F + Bhi,
(4.105)
which should be compared to equation (4.93).

We have shown only that this ensemble scheme is not inconsistent. It is not obviously wrong, but of
course this does not prove that it is right. A flawless general proof would require detailed mathematical
reasoning, but proofs for particular situations are outlined in the next problem and in problem 4.18.
4.16 The grand- ensemble for the classical monatomic ideal gas
Apply the concepts of the previous problem to the classical monatomic ideal gas where the mechanical
parameter is the particle mass. In particular, find expressions
a. for B(T, V, N, m),
b. for m(T, V, N, B),
116
c. and for hmi and m as functions of T , V , N , and .
4.17 An ensemble of your own

Any mechanical parameter in the Hamiltonian can be traded in for a control parameter. Can you
think up a good problem using this fact where the parameter is, say, an applied magnetic field? Or
the spring constant (or relaxed length) between two atoms in a diatomic molecule (or in the Einstein
model)? Or the moment of inertia of a diatomic molecule modeled as a dumbbell? Or a parameter in
the Lennard-Jones potential? Or the size of hard-sphere atoms? Or the spacing between atoms in a
crystal?
4.12
Ensemble Problems II
The remaining problems in this chapter deal with the same principles issues that the others do, but they
assume some familiarity with physical and mathematical topics that we have not yet treated. I place them
here because of their character, but I do not expect you to do them at this moment. Instead I list their
prerequisites and I hope you will come back to them later.
4.18 The grand- ensemble for the simple harmonic oscillator
(Prerequisite: Problems 4.15 and 5.7.) Problem 4.15 introduced a class of ensembles in which mechanical parameters in the Hamiltonian were not fixed but rather allowed to fluctuate under the control of
some parameters . Apply this concept to the diatomic molecules of problem 5.7, using the natural
frequency of vibration as the mechanical parameter. Develop the suggestive arguments of problem 4.15
into a mathematically rigorous proof.
4.19 Particles in a bowl
(Prerequisite: Chapter 5.) I cant count how many times Ive waved my hands and told you that
the precise character of the boundaries should be irrelevant to the bulk properties found by statistical
mechanics. Now you have a chance to prove it (for a restricted situation of course).
a. Consider N non-interacting, classical point particles moving subject to the potential energy function U (r) = Kr2 /2. Evaluate the classical partition function Zb (T, K, N ) and compare it to the
partition function Zh (T, V, N ) for N particles in a hard-walled container of volume V .
b. For what V (as a function of K) will the two partition functions in part (a.) be the same?
c. The bowl container has no stiff walls, so it is possible for a particle to be located at any distance
from the origin. But is it not likely that a particle will get too far away. Calculate the rms radius
p
hr2 i. (Use equipartition?)
d. Suppose the hard-walled container is spherical with radius R. Place the origin at the center of
p
the sphere and find the rms radius hr2 i.
e. For the V corresponding to K in the sense of part (b.), compare the two rms radii.
4.12. ENSEMBLE PROBLEMS II
117
Extra: Do this problem quantum mechanically. (Clue: Notice the mathematical similarity between
this problem and the Einstein model of lattice vibrations.)
4.20 Quantal monatomic ideal gas
(Prerequisite: Chapter 6.) We have many times considered the problem of N non-interacting classical
point particles in a box of edge length L = V 1/3 . Now we will solve the problem for quantum mechanical
particles.
P
The partition function is Z = eE , where the sum is taken over the many-particle energy eigenstates
(r1 , . . . , rN ). Ignore symmetry-under-interchange requirements so that such states are specified by
a simple listing of the component one-particle levels (or orbitals) k (r). Thus can be specified by
simply listing the relevant k values of its component levels. Show then that
!N
1 X Ek
Z=
e
(4.106)
N!
k
where Ek = h
2 k 2 /2m and the sum is over all k permitted by periodic boundary conditions. Evaluate
the sum (in the thermodynamic limit) by turning it into an integral. Compare the resulting partition
function with that of a classical monatomic ideal gas.
4.21 Quantal monatomic ideal gas in the microcanonical ensemble
Consider the situation of the above problem but calculate the microcanonical partition function
(E, V, N ). Compare with the classical result.
Chapter 5
Classical Ideal Gases

End of principles, beginning of applications. There is still more that could be done with principles, but the
proof of the pudding is in the eating. And in this case the pudding is excellent eating indeed. . . calculations
made from statistical mechanics theory give results that are actually more accurate than experiment.
This chapter considers ideal (i.e. non-interacting) gases made up of atoms or molecules that may have
internal structure (i.e. not point particles). The internal degrees of freedom will be treated either classically
or quantally, but the translational degrees of freedom will always be treated classically.
5.1
Classical Monatomic Ideal Gases
We have already found the Helmholtz free energy of the classical monatomic ideal gas (section 4.3). I think
you will agree that the canonical calculation is considerably easier than the corresponding microcanonical
calculation. Here we will review the calculation, then go back and investigate just what caused the solution
to be so much easier, and this investigation will lead to an important theorem.
5.1.1
Solving the problem
The energy
H() =
1 2
(p + p2y,1 + p2z,1 + p2x,2 + + p2z,N ) + U (x1 ) + U (x2 ) + + U (xN )
2m x,1
(5.1)
is a sum of terms, each involving different variables, so the Boltzmann factor

2
eH() = e(px,1 /2m) e(py,1 /2m) e(pz,N /2m) eU (x1 ) eU (x2 ) eU (xN )
118
(5.2)
5.1. CLASSICAL MONATOMIC IDEAL GASES
119
is a product of factors, each involving different variables. Thus the partition function, which in general is an
integral in a 6N -dimensional space, factors into 3- and 1-dimensional integrals

Z +
Z +
Z +
Z +
Z +
Z +
1
U (xN )
U (x1 )
dx
dy
dz
e
dx
dy
dz
e
Z(T, V, N ) =
N
N
N
1
1
1
N !h3N

Z +
Z +
2
2
dpz,N e(pz,N /2m) .
dpx,1 e(px,1 /2m)
This doesnt happen in the microcanonical calculation!

Moreover, we found it easy to evaluate the N three-dimensional position integrals and the 3N onedimensional momentum integrals. The result was
Z(T, V, N ) =
i3N
V N hp
2mk
T
.
B
N !h3N
(5.3)
The result above is correct, but it is not the most convenient form for the partition function. There
are lots of variables and constants floating around, many of them with hard-to-remember dimensions. The
quantity Z is in fact dimensionless, but thats not obvious from the expression. For this reason is customary
to define the quantity
h
,
(5.4)
(T ) =
2mkB T
called the thermal de Broglie wavelength. Like any wavelength, it has the dimensions of length, and the
notation (T ) suggests this. It is, of course, an intensive quantity. In terms of the thermal de Broglie
wavelength, the partition function for the classical monatomic ideal gas is

N
1
V
Z(T, V, N ) =
.
(5.5)
N ! 3 (T )
If a particle of mass m has kinetic energy kB T , then it has momentum 2mkB T and thus a quantum-mechanical
de Broglie wavelength of
h
h
=
=
.
(5.6)
p
2mkB T
The definition (5.4) above slips in an extra dimensionless factor of = 1.77 just to make the result (5.5) easier
to remember and to work with. Physically, the thermal de Broglie wavelength represents the quantal smearing of
each individual point particle. Thus if the mean separation between atoms is much greater than (T ), we expect
that the classical approximation will be a good one. If the mean separation is less than or about equal to (T ),
we expect that a quantal treatment will be required.
After doing any calculation is it useful to perform a debriefing to figure out whyin general rather than
technical termsthe calculation went the way it did. Almost everyone agrees that this canonical calculation is
a lot easier than the corresponding microcanonical calculation to find (E, V, N ) for the classical monatomic
ideal gas. But what was the particular feature that made it so much easier? Surely it was the factorization of
the Boltzmann factor and the subsequent factorization of the partition function. This enabled us to perform
one- and three-dimensional integrals rather than the 3N -dimensional integral required (see appendix D) for
the microcanonical calculation. Is there anything that can be said more generally about this feature?
120
5.1.2
CHAPTER 5. CLASSICAL IDEAL GASES
Theorem on decoupling Hamiltonians
Indeed there is. Suppose a Hamiltonian is a sum of two pieces

H() = H1 (1 ) + H2 (2 ),
(5.7)
where 1 and 2 are exclusive. (That is, the phase space variables in the list 1 and the phase space variables
in the list 2 together make up the whole list , and no variable appears in both lists.) The Hamiltonian
H() is then said to decouple into the two parts H1 (1 ) and H2 (2 ).1 Then it is easy to see that the
partition function will factor into two parts:
Z
Z
Z
1
1
H()
H(1 )
d e
=
d1 e
d2 eH(2 ) .
(5.8)
Z(T, V, N ) =
N !h3N
N !h3N
If the Hamiltonian decouples into three parts, then this reasoning can be used to show that the partition
function breaks into three factors. And so forth.
For example, the Hamiltonian of the classical monatomic ideal gas that we have just investigated decouples
into 4N pieces,
p2z,N
p2x,1
p2y,1
H() =
+
+ +
+ U (x1 ) + U (x2 ) + + U (xN ).
(5.9)
2m
2m
2m
2
Each term of the form p /2m in the Hamiltonian leads to a factor of h/(T ) in the partition function. . . each
term of the form U (x) leads to a factor of V . Thus our theorem on decoupling Hamiltonians shows that the
partition function is
3N

h
1
V N.
(5.10)
Z(T, V, N ) =
N !h3N (T )
If the particles do interact, then the potential energy part of the Hamiltonian does not break up into
exclusive pieces, but the kinetic energy part still does. Thus for an interacting gas with potential energy of
interaction
UN (x1 , x2 , . . . , xN ),
(5.11)
the partition function is
Z(T, V, N ) =
1
N !3N (T )
dx1
dzN eUN .
(5.12)
Although the potential energy part of the problem is now excruciatingly difficult, our theorem on decoupling
Hamiltonians assures us that the kinetic energy part is just as easy to handle for an interacting as for a
non-interacting gas.
5.1 Gas with pair interactions
For a pure classical monatomic gas that interacts only through pair interactions (i.e. for which there
are no three-body or higher interactions) the Hamiltonian is
N
N
1 X
N
N
X
X
X
p2i
+
UA (xi ) +
UB (xi , xj ),
H() =
2m i=1
i=1 j=i+1
i=1
(5.13)
1 Alternatively, the two parts of the Hamiltonian are called uncoupled. These terms derive from the ordinary English word
to couple, meaning to join or to connect.
5.2. CLASSICAL DIATOMIC IDEAL GASES
121
where UA (x) is the potential energy due to the interaction between a particle at point x and the
container, and UB (x, y) is the potential energy due to the interaction between a particle at point x
and a particle at point y. The potential energy thus breaks up into a sum of N + N (N 1)/2 pieces.
Does the decoupling theorem apply?
5.2
Classical Diatomic Ideal Gases
Rigid linear model. The partition function is

N

3
1
V
4 IkB T
Z(T, V, N ) =
N ! 3 (T )
h2
(5.14)
General results for separation of Z into translational (center of mass) and internal degrees of freedom.
Term internal specific heat.
5.3
5.3.1
Heat Capacity of an Ideal Gas

The equipartition theorem
Equipartition theorem. (For classical systems.) Suppose the Hamiltonian H() decouples into one piece
involving a single phase space variablecall it H1 , plus another piece which involves all the other phase space
variablescall it H2 (2 ). Suppose further that the energy depends quadratically upon this single phase space
variable, and that this variable may take on values from to +. Then, in classical statistical mechanics,
the mean contribution to the energy due to that single variable is
hH1 i = 12 kB T.
(5.15)
Notice how general this theorem is. The remaining piece of the Hamiltonian, H2 (2 ), might decouple
further or it might not. The phase space variable entering H1 might be a momentum,
H1 (p) =
p2
,
2m
(5.16)
or an angular momentum,
`2
,
2I
or even a position coordinate, as in the simple harmonic oscillator energy
H1 (`) =
H1 (x) = 21 kx2 .
(5.17)
(5.18)
Furthermore, in all these circumstances, the mean energy is independent of the particular parameters m or
I or k. . . it depends only upon the temperature. This explains the origin of the name equipartition: the
122
mean translational energy due to motion in the x direction is equal to the mean rotational energy due to the
change of , and this holds true even if the gas is a mixture of molecules with different masses and different
moments of inertia. The energy is equally partitioned among all these different ways of holding energy.
Proof. We will write
H1 (p) = ap2 ,
although the variable might not be a linear momentum. Then the average of H1 is
R +
R
R
dp H1 eH1 d2 eH2 (2 )
d H1 eH()
.
= R +
hH1 i = R
R
d eH()
dp eH1 d2 eH2 (2 )
(5.19)
(5.20)
Clearly, the integrals over 2 cancel in this last expression. (This explains why the form of H2 is irrelevant
to the theorem.) We are left with
R +
2
dp ap2 eap
hH1 i =
.
(5.21)
R +
dp eap2
These two integrals could be evaluated in terms of Gamma functions (see appendix C), but they dont need
to be evaluated yet. Think for a moment about our slick trick of parametric differentiation. . . using it we
can write
Z +

d
ap2
dp e
.
(5.22)
ln
hH1 i =
d
The integral that remains is of Gaussian character and we could evaluate it using the results of Appendix B.
But before rushing in to integrate, lets employ the substitution u = a p to find

Z +
Z +
d
1
1
1
d
u2
u2
hH1 i =
ln
du e
ln + ln
=
du e
.
(5.23)
d
d
a
a
This last expression shows that theres no need to evaluate the integral. Whatever its value is, it is some
number, not a function of , so when we take the derivative with respect to the term involving that
number will differentiate to zero. Similarly for the constant a, which explains why the equipartition result
is independent of that prefactor. We are left with
hH1 i =
1
1 d
11
d
ln =
ln =
d
2 d
2
(5.24)
or, finally, the desired equipartition result

hH1 i = 12 kB T.
5.3.2
Applications of equipartition; Comparison with experiment
5.3.3
Crossover between classical and quantal behavior; Freeze out
(5.25)
At high temperatures, typical thermal energies are much greater than level spacings. Transitions from one
level to another are very easy to do and the granular character of the quantized energies can be ignored.
This is the classical limit, and equipartition holds!
5.3. HEAT CAPACITY OF AN IDEAL GAS
123
At low temperatures, typical thermal energies are less that the level spacing between the ground state
and the first excited state. There is so little thermal energy around that the molecule cannot even be excited
out of its ground state. Virtually all the molecules are in their ground states, and the excited states might
as well just not exist.
In classical mechanics, a diatomic molecule offered a small amount of rotational energy will accept that
energy and rotate slowly. But in quantum mechanics, a diatomic molecule offered a small amount of rotational
energy will reject that energy and remain in the ground state, because the energy offered is not enough to
lift it into the first excited state.2 The quantal diatomic molecule does not rotate at all at low temperatures,
so it behaves exactly like a monatomic molecule with only center-of-mass degrees of freedom.
In short, we explain the high-temperature rotational specific heat (crot
V = kB ) through equipartition. We
rot
explain the low-temperature rotational specific heat (cV vanishes) through difficulty of promotion to the
first quantal excited state. This fall-off of specific heat as the temperature is reduced is called freeze out.
The crossover between the high-temperature and low-temperature regimes occurs in the vicinity of a
characteristic temperature at which the typical thermal energy is equal to energy separation between the
ground state and the first excited state. If the energies of these two states are 0 and 1 respectively, then
we define the characteristic crossover temperature through
k B 1 0 .
(5.26)
5.2 Generalized equipartition theorem and the ultra-relativistic gas

a. Suppose the Hamiltonian H() decouples into two pieces
H() = a|p|n + H2 (2 )
(5.27)
where p is some phase space variable that may take on values from to +, and where 2
represents all the phase space variables except for p. (Note that the absolute value |p| is needed
in order to avoid, for example, taking the square root of a negative number in the case n = 1/2.)
Show that, in classical statistical mechanics, the mean contribution to the energy due to that
single variable is
ha|p|n i = n1 kB T.
(5.28)
b. In special relativity, the energy of a free (i.e. non-interacting) particle is given by
p
(mc2 )2 + (pc)2 ,
(5.29)
where c is the speed of light. As you know, when v c this gives the non-relativistic kinetic
energy KE mc2 + p2 /2m. In the ultra-relativistic limit, where v is close to c, the energy is
approximately pc. What is the heat capacity of a gas of non-interacting ultra-relativistic particles?
2 This
paragraph is written in the shorthand language discussed on page 108, as if energy eigenstates were the only allowed
quantal states.
124

c. Estimate the crossover temperature between the non-relativistic and ultra-relativistic regimes.
5.3 Another generalization of equipartition

Consider the same situation as the equipartition theorem in the text, but now suppose the single phase
space variable takes on values from 0 to +. What is the corresponding result for hH1 i?
5.4 Equipartition and the virial theorem
Look up the term virial theorem in a classical mechanics textbook. Is there any relation between the
virial theorem of classical mechanics and the equipartition theorem of classical statistical mechanics?
Box: The O Notation

Approximations are an important part of physics, and an important part of approximation is to ensure
their reliability and consistency. The O notation (pronounced the big-oh notation) is an important and
practical tool for making approximations reliable and consistent.
The technique is best illustrated through an example. Suppose you desire an approximation for
f (x) =
ex
1x
(5.30)
valid for small values of x, that is, x 1. You know that

ex = 1 x + 21 x2 16 x3 +
(5.31)
and that
1
= 1 + x + x2 + x3 + ,
1x
so it seems that reasonable approximations are
and
(5.32)
ex 1 x
(5.33)
1
1 + x,
1x
(5.34)
whence
ex
(1 x)(1 + x) = 1 x2 .
1x
Lets try out this approximation at x0 = 0.01. A calculator shows that
(5.35)
ex0
= 1.0000503 . . .
1 x0
(5.36)
1 x20 = 0.9999000.
(5.37)
while the value for the approximation is
5.3. HEAT CAPACITY OF AN IDEAL GAS
125
This is a very poor approximation indeed. . . the deviation from f (0) = 1 is even of the wrong sign!
Lets do the problem over again, but this time keeping track of exactly how much weve thrown away
while making each approximation. We write
ex = 1 x + 21 x2 16 x3 +
(5.38)
ex = 1 x + 12 x2 + O(x3 ),
(5.39)
as
where the notation O(x3 ) stands for the small terms that we havent bothered to write out explicitly. The
symbol O(x3 ) means terms that are about the magnitude of x3 , or smaller and is pronounced terms of
order x3 . The O notation will allow us to make controlled approximations in which we keep track of exactly
how good the approximation is.
Similarly, we write
1
= 1 + x + x2 + O(x3 ),
1x
(5.40)
and find the product

f (x)

1 x + 21 x2 + O(x3 ) 1 + x + x2 + O(x3 )

=
1 x + 21 x2 + O(x3 )

+ 1 x + 21 x2 + O(x3 ) x

+ 1 x + 12 x2 + O(x3 ) x2

+ 1 x + 21 x2 + O(x3 ) O(x3 ).
=
(5.41)
(5.42)
(5.43)
(5.44)
(5.45)
Note, however, that x 21 x2 = O(x3 ), and that x2 O(x3 ) = O(x3 ), and so forth, whence
f (x)

1 x + 21 x2 + O(x3 )

+ x x2 + O(x3 )

+ x2 + O(x3 )
(5.46)
+O(x3 )
(5.49)
(5.48)
(5.50)
f (x) 1 + 21 x2 .
(5.51)
1+
1 2
2x
(5.47)
+ O(x ).
Thus we have the approximation

Furthermore, we know that this approximation is accurate to terms of order O(x2 ) (i.e. that the first neglected
terms are of order O(x3 )). Evaluating this approximation at x0 = 0.01 gives
1 + 21 x20 = 1.0000500,
far superior to our old approximation (5.35).
(5.52)
126
What went wrong on our first try? The x2 in approximation (5.35) is the same as the x2 on line (5.47).
However, lines (5.46) and (5.48) demonstrate that there were other terms of about the same size (that is,
other terms of order x2 ) that we neglected in our first attempt.
The O notation is superior to the dot notation (such as ) in that dots stand for a bunch of small
terms, but the dots dont tell you just how small they are. The symbol O(x3 ) also stands for a bunch of
small terms, but in addition it tells you precisely how small those terms are. The O notation allows us to
approximate in a consistent manner, unlike the uncontrolled approximations where we ignore a small term
without knowing whether we have already retained terms that are even smaller.
End of Box: The O Notation
5.4
Specific Heat of a Hetero-nuclear Diatomic Ideal Gas
The classical partition function is

N
Z Z 2
Z +
Z +
1
V
1
(`2 +`2 )/2I
Z=
d
d
d`
d`
e
.
N ! 3 (T ) h2 0
0
(5.53)
Lets make sure we have the quantum mechanics of one molecule straight before we jump into treating
6.02 1023 molecules: First of all, we consider hetero-nuclear rather than homo-nuclear diatomic molecules
(such as CO or HCl rather than H2 or N2 ) because in the latter there are additional issues due to the Pauli
principle. Lets just avoid these issues on our first encounter. The quantal energy eigenvalues are
rot = `(` + 1)
2
h
2I
` = 0, 1, 2, . . . ,
(5.54)
with
degeneracy` = 2` + 1.
(5.55)
Thus the quantal characteristic temperature (or rot ) is

k B 1 0 =
2
h
.
I
(5.56)
And, finally, the quantal partition function is

"
#N
1
V X
`(`+1)
h2 /2I
Z=
(2` + 1)e
.
N ! 3 (T )
(5.57)
`=0
It saves a lot of writing to define the rotational partition function

(T ) rot (T )
X
`=0
(2` + 1)e`(`+1)h
/2I
X
`=0
(2` + 1)e`(`+1)/2T .
(5.58)
5.4. SPECIFIC HEAT OF A HETERO-NUCLEAR DIATOMIC IDEAL GAS
127
The expression on the far right is the one that I like to use. . . instead of having a lot of variables like T , h
,
and I floating around just waiting to get misplaced, everything is neatly packaged into the ratio /T , which
is obviously dimensionless.
Now, for any partition function of the form

N
1
V
= Zmono N (T ).
Z=
(T
)
N ! 3 (T )
(5.59)
We have
F = kB T ln Z = kB T ln Zmono kB T N ln
|
{z
}|
{z
}
Fmono
N f rot
and
erot =
(f rot /T )
ln
=
(1/T )
crot
V =
erot
.
T
(5.60)
(5.61)
Because the rotational partition function depends on temperature only through the combination /T , it
makes sense to use that variable for derivatives:
erot =
(kB T ln /T )
ln
ln
(f rot /T )
=
= kB
= kB
(1/T )
(1/T )
(1/T )
(/T )
(5.62)
erot
erot (/T )
erot
=
= 2
.
T
(/T ) T
T (/T )
(5.63)
and
crot
V =
Now back to the evaluation of (T ). It is quite easy to prove that the infinite sum for (T ) converges.
You will then be tempted to find an expression for (T ) in terms of well-known functions like polynomials or
exponentials. If you try this, you will not find any simple expression. Instead of manipulating (T ) into some
combination of familiar functions, we will have to become familiar with it on our own. It can be evaluated
and plotted on a computer just by adding up a finite number of terms in the defining infinite sum. But doing
so is not the most productive way to approach an understanding of this function. . . given a plot of (T ), how
would you find the specific heat due to rotation? A better way to approach the problem is to find how the
partition function (T ) (and hence the specific heat crot
V (T )) behaves at low and high temperatures.
At low temperatures, we expand in terms of the small variable e/T to find
(T )
=
=
1 + 3e/T + 5e3/T + 7e6/T +

1 + 3e
/T
+ 5e
3/T
+ O(e
6/T
(5.64)
).
The notation O(xN ) is read terms of order xN . (See the box on page 124.) Now for small variables the
logarithm is
ln(1 + ) = 21 2 + 13 3 14 4 + O(5 ),
(5.65)
whence
ln (T )
3e/T + 5e3/T + O(e6/T )
(5.66)
128

12 [3e/T + 5e3/T + O(e6/T )]2
+ 31 [3e/T + 5e3/T + O(e6/T )]3
+O(e4/T )
3e/T + 5e3/T + O(e6/T )
(5.67)
21 [9e2/T + 30e4/T + O(e6/T )]

+ 31 [27e3/T + 135e5/T + O(e6/T )]
4/T
+O(e
3e
/T
9 2/T
2e
+ 14e3/T + O(e4/T ).
(5.68)
Note that in going from equation (5.67) to equation (5.68) it would have been impossible to make sure that
all of the several expansions were kept to equivalent orders of accuracy had we used + rather than the O
notation.
Now it is easy to use equations (5.62) and (5.63) to find
erot = kB [3e/T 9e2/T 42e3/T + O(e4/T )]
and
crot
V
(5.69)
2
[3e/T 18e2/T + 126e3/T + O(e4/T )].

= kB
T
(5.70)
(Note that if we had used equations (5.61) instead, it would have been a lot more work.) The specific heat
vanishes at zero temperature, and it increases exponentially slowly with increasing temperature. This is a
very slow growth indeed. . . not only is the slope zero, but the second derivative is zero, the third derivative
is zero, indeed all orders of derivative vanish at the origin.
What about high temperatures? We have discussed the fact that at high temperatures the rotational
specific heat will approach the classical equipartition value of kB , but how will it approach the limiting value?
In other words, what are the deviations from the high temperature limit? We can answer these questions by
approximating the sum by an integral. The Euler-MacLaurin formula states that if f (x) 0, f 0 (x) 0,
f 00 (x) 0, etc. as x , then
X
`=0
Z
f (`)
0
f (x) dx + 21 f (0)
1 0
12 f (0)
000
1
720 f (0)
(v)
1
(0)
30240 f
+ .
(5.71)
In our case,
f (x) = (2x + 1)ex(x+1)/2T ,
so
Z
f (x) dx
f (0)
f 0 (0)
2T
(5.72)
5.5. CHEMICAL REACTIONS BETWEEN GASES
129
2
f (0) = 6 + O
T
T
2
(v)
f (0) = O
.
T
000
Thus we have
2
T
2
3
1
1
ln (T ) ln
+
+
+O
2T
6T
360 T
T
"
2 #
1
1
T
+O
erot kB + +
6 180 T
T
"
3 #
2
1
rot
cV
kB 1 +
+O
.
180 T
T
(T )
T
1
1
+ +
+O
3 30 T
(5.73)
(5.74)
(5.75)
(5.76)
For high temperatures the specific heat is nearly equal to its classical value, and the quantity grows slightly
larger as the temperature decreases.
Note: The sign in the above formulas represents asymptotic equality. The infinite series on the right
does not necessarily converge, and even if it does then it might not converge to the quantity on the left.
However a finite truncation of the sum can be a good approximation for the quantity on the left, and that
approximation grows better and better with increasing temperature (i.e. smaller values of /T ).
5.5
Chemical Reactions Between Gases
5.6
Problems
5.5 Decoupling quantal Hamiltonians

Prove the decoupling Hamilton implies factoring partition function theorem of section 5.1.2 for
quantal systems.
5.6 Schottky anomaly
A molecule can be accurately modeled by a quantal two-state system with ground state energy 0 and
excited state energy . Show that the internal specific heat is

2

e/kB T
int
cV (T ) = kB
.
(5.77)
kB T
(1 + e/kB T )2
Sketch this specific heat as a function of kB T /. How does the function behave when kB T and
kB T ?
5.7 Simple harmonic oscillator
Suppose a molecule can be accurately modeled as a harmonic oscillator of natural frequency .
130

a. Find the expected internal energy of one such molecule, written as a sum of the ground state
energy plus a temperature-dependent part.
b. Show that the internal specific heat is
cint
V (T ) = kB
h
kB T
2
eh/kB T
.
(1 eh/kB T )2
(5.78)
c. Show that at low temperatures (kB T h),

2
h
cint
(T
)
k
eh/kB T ,
B
V
kB T
(5.79)
whereas at high temperatures (kB T h),

cint
V (T ) kB .
(5.80)
d. (Optional.) Show that the leading quantal correction to the high-temperature specific heat is
cint
V (T ) = kB [1
1 2
12 x
+ O(x3 )],
where
x=
h
.
kB T
(5.81)
e. Sketch the internal specific heat as a function of kB T /h.

5.8 Simple harmonic oscillator entropy
I was speaking with someone who claimed that thermodynamics didnt apply to systems of interacting
atoms, or to living things, or to anything other than the ideal gas. He challenged me Whats the
entropy of a pendulum? The answer is that entropy doesnt apply to a pendulum! Answer his question for a pendulum in the simple harmonic oscillator approximation, with natural frequency and
temperature T .
5.9 Kinetic energy in air
What is the kinetic energy is a cubic liter of air (mostly nitrogen and oxygen) at one atmosphere
pressure and room temperature? At one atmosphere pressure and temperature 600 K? Use classical
mechanics, and express your answer both in Joules and in the commercial unit kilowatt-hours. (Clue:
If you calculate the average kinetic energy of one molecule, youre working way too hard.) Why doesnt
anyone exploit all this valuable energy lying around in every cubic meter of air?
5.10 Conceptual comparison
a. Explain qualitatively why the results of the two previous problems are parallel at low temperatures.
b. (Harder.) Explain qualitatively both high temperature results. (Clue: At high temperatures, the
average energy per particle in the Schottky case approaches /2. Why?)
5.11 Compressibility of a diatomic gas
Find the isothermal compressibility T for an ideal diatomic gas, where each molecule is modeled as a
dumbbell with moment of inertia I.
5.6. PROBLEMS
131
5.12 Systems with a small number of states

(This problem requires no calculation! All the answers can be found in your head.) A collection of
non-interacting particles is in thermal equilibrium. Each particle has only three energy eigenvalues,
namely 0, , and 4.
a. What is the criterion for high temperature in this situation?
b. Suppose there are three non-degenerate energy eigenstates. At high temperatures, what is the
average energy of each particle? (Clue: The answer is not 4.)
c. Now suppose that the lower two energy eigenstates are non-degenerate, but that there are two
independent states with energy 4. What is the average energy per particle at high temperatures
in this case?
5.13 Anharmonic oscillator
The energy eigenvalues of a simple harmonic oscillator are equally spaced, and we have explored the
consequences of this for the heat capacity of a collection of harmonic oscillators. Suppose an anharmonic
oscillator is approximately harmonic (with natural frequency 0 ) for small energies, but that for large
energies (greater than, say, Ex ) the eigenvalues become more closely spaced as energy increases. At
temperatures greater than Ex /kB , will the heat capacity of a collection of such anharmonic oscillators
be greater than or less than that of a collection of harmonic oscillators with the same natural frequency
0 ? Why?
5.14 Descriptive features of models
Two possible models for a diatomic ideal gas are the rigid dumbbell (model R; two point particles
connected by a rigid rod) and the springy dumbbell (model S; two point particles connected by a
spring). In classical statistical mechanics, which of the following statements is true?
a. Model R has a specific heat cV = 23 kB .
b. Model S has a smaller specific heat than model R.
c. Model S is always correct.
d. Model R is always correct.
e. The choice between models R and S depends on the temperature.
5.15 An n-state system, qualitatively
A model molecule has n equally-spaced energy levels, all of them non-degenerate, with energy spacing
. Thus as n varies from 2 to this model interpolates between the Schottky system of problem 5.6
and the simple harmonic oscillator of problem 5.7.
a. Find a low-temperature approximation for the specific heat that is independent of n.
b. At high temperatures, the specific heat approaches zero. What is the criterion for high temperature?
132

c. At high temperatures, what is the expected energy of this model?
d. There is a theorem stating that at any fixed positive temperature, the specific heat must increase
with increasing n. Assume this theorem and use it to prove that as n increases, the maximum in
the specific heat versus temperature curve becomes higher.
5.16 An n-state system, quantitatively

Show that the system of the previous problem has internal specific heat
#
"
2

2
e/kB T
n
en/kB T

int
.
cV (T ) = kB
kB T
kB T
(1 e/kB T )2
(1 en/kB T )2
Does this expression have the proper limits when n = 2 and when n ?
(5.82)
Chapter 6
Quantal Ideal Gases

6.1
Introduction
In the previous chapter, we found that at high temperatures, an ideal gas of diatomic molecules with spring
interactions has a heat capacity of 72 kB per molecule: 32 kB from the translational degrees of freedom, kB from
the rotational degrees of freedom, and kB from the spring degrees of freedom. If the temperature is decreased,
the spring degrees of freedom become governed by quantum mechanics rather than classical mechanics, the
equipartition theorem no longer holds, and eventually these degrees of freedom freeze out and contribute
nothing to the heat capacity: the total heat capacity per molecule becomes 52 kB . If the temperature is
decreased still further, the story is repeated for the rotational degrees of freedom and eventually they freeze
out. What happens if the temperature is decreased yet again? Do the translational degrees of freedom then
freeze out as well? The answer is sort of, but the crossover from the classical to the quantal regime is
complicated in this case by the quantal requirement of interchange symmetry. This requirement gives rise
to a much richer and more interesting crossover behavior than is provided by simple freeze out.
6.2
The Interchange Rule
Before turning to statistical mechanics, let us review the quantal interchange rule. The wavefunction for
a system of three particles is a function of three variables: (x1 , x2 , x3 ). (In this section we ignore spin.) If
the three particles are identical, then the wavefunction must be either symmetric under the interchange of
any pair of variables,
(x1 , x2 , x3 ) = +(x2 , x1 , x3 ) = +(x2 , x3 , x1 ),
etc.,
(6.1)
or else antisymmetric under the interchange of any pair of variables,
(x1 , x2 , x3 ) = (x2 , x1 , x3 ) = +(x2 , x3 , x1 ),
133
etc.
(6.2)
134
CHAPTER 6. QUANTAL IDEAL GASES
The foregoing assertion is an empirical rule that cannot be derived from any of the other principles of
quantum mechanics. (Indeed there is currently considerable interest in anyons, hypothetical particles that
obey all the principles of quantum mechanics except the interchange rule.) The rule holds for all quantal
states, not just energy eigenstates. It holds for interacting as well as for non-interacting particles. And the
rule has a number of surprising consequences, both within the domain of quantum mechanics and atomic
physics, and, as well shall soon see in detail, within the domain of statistical mechanics.
The sign of the interchange symmetry, either + or , is governed only by the type of particle involved:
for pions it is always +, for electrons it is always . Particles for which the sign is always + are called
bosons, and those for which it is always are called fermions. It is an experimental fact that particles with
integral spin s are bosons and those with half integral spin s are fermions.
6.2.1
Problems
6.1 The interchange rule in another representation

Full information about the state of three identical particles is contained not only in the configurational
wave function (x1 , x2 , x3 ) but also in the momentum space wave function
Z
Z
Z
1
3
3
1 , p2 , p3 ) =
(p
d
x
d
x
d3 x3 (x1 , x2 , x3 )ei(p1 x1 +p2 x2 +p3 x3 )/h .
(6.3)
1
2
(2
h)9/2
Show that if one representation is symmetric (or antisymmetric), then the other one is as well.
6.2 Symmetrization
Show that for any function f (x, y, z), the function
fS (x, y, z) = f (x, y, z) + f (y, x, z) + f (z, x, y) + f (x, z, y) + f (y, z, x) + f (z, y, x)
(6.4)
is symmetric under the interchange of any pair of variables. How many such pairs are there? Is there
a similar algorithm for building up an antisymmetric function from any garden-variety function?
6.3
Quantum Mechanics of Independent Identical Particles
This chapter considers collections of independent (i.e. non-interacting) identical monatomic particles. It
does not treat mixtures or diatomic molecules. Notice that independent means only that the particles
do not interact with each other. In contrast, each particle individually may interact with some background
potential, such as a square well or harmonic oscillator potential. Later on (in section 6.5) we will restrict
our consideration to particles that interact only with the walls of their container (free particles), but for
now we treat the more general case.
Doing quantal statistical mechanics requires a basis of states to sum over. In this section we consider
the only the quantum mechanics of our situation in order to produce an energy eigenbasis. . . we defer all
questions of statistical mechanics to the next section.
6.3. QUANTUM MECHANICS OF INDEPENDENT IDENTICAL PARTICLES
135
Most of our discussion in this section is phrased in terms of the three-particle problem rather than the
N -particle problem. The use of a specific number lends the advantage of concreteness, and three is the
smallest number that generalizes readily to the N -particle case.
6.3.1
A basis for non-identical particles
Consider first a single particle moving in the background potential. Elementary quantum mechanics deals
with this situation, and it tells us that there will be a certain number M of (possibly degenerate) energy
eigenstates. (Usually M will be infinite, but there are advantages to calling it M and maintaining the ability
to take the limit M .) The rth energy eigenstate has energy r and is represented by the wavefunction
r (x), where x denotes the arguments of the wavefunction: Thus for a spinless particle, x could stand for
x, y, z or px , py , pz or even px , y, pz . For a particle with spin s, x could stand for expressions like x, y, z, sz or
px , py , pz , sx .
We will soon use these one-particle energy eigenstates as building blocks to construct energy eigenstates
for the N -particle situation, i.e. energy eigenstates for the entire system. Thus we will frequently be talking
about both one-particle energy eigenstates and N -particle energy eigenstates in the same sentence, and
both phrases are mouthfulls. To make our sentences shorter and clearerand to emphasize the dramatic
distinction between these two entitieswe will adopt the common convention of referring to the one-particle
energy eigenstates as levels and to the N -particle energy eigenstates as states. (Some books use the
term orbitals rather than levels.) Thus we have for the one-particle energy eigenproblem:
levels:
energies:
1 (x)
1
2 (x)
2
r (x)
r
M (x)
M
Now consider more than one particlelets say three particlesmoving independently and all subject to
the same background potential. If the particles are not identical (say one helium, one argon, and one neon
atom), then the three-body energy eigenbasis will contain elements (states) such as
r (x1 )s (x2 )t (x3 )
with energy
r + s + t .
(6.5)
(If the particles interacted, then this would still be a state of the system, but it would not be an energy
eigenstate.) We will represent this state by the notation |r, s, ti:
r (x1 )s (x2 )t (x3 )
|r, s, ti.
(6.6)
Of course, degeneracies now must exist: the energy of |r, s, ti is the same as the energy of |s, r, ti. The entire
basis consists of M 3 such states (more generally, for the case of N independent particles, M N ), namely:
136

|1,1, 1
|2,1, 1
|1,2, 1
|1,1, 2
|3,1, 1
..
.
|3,2, 1
|3,1, 2
..
.
|4,2, 1
..
.
|M,M,M
i
i
i
i
i
i
i
i
i
Out of these M 3 states we can build up (by addition and scalar multiplication) any wavefunction, i.e.
any normalized function of (x1 , x2 , x3 ). This is exactly what we want for non-identical particles. But for
identical particles we dont want to be able to build any wavefunction. Because of the interchange rule, the
only relevant wavefunctions are those symmetric (or antisymmetric) under the interchange of any pair of
coordinates, for example
(x1 , x2 , x3 ) = (x3 , x2 , x1 ),
(6.7)
and hence these are the only kind of wavefunctions we want to be able to build. In other words, we need an
energy eigenbasis consisting only of symmetric (or antisymmetric) wavefunctions.
6.3.2
A basis for identical particles
Fortunately, there exists a general procedure for constructing symmetric (or antisymmetric) functions out of
garden-variety (or non-symmetrized) functions, and this procedure will enable us to build the two desired
basis sets. For functions of three variables the procedure (which is readily generalized to N variables) works
like this: Given a function f (x1 , x2 , x3 ), the new function
f (x1 , x2 , x3 ) + f (x1 , x3 , x2 ) + f (x3 , x1 , x2 ) + f (x3 , x2 , x1 ) + f (x2 , x3 , x1 ) + f (x2 , x1 , x3 )
(6.8)
is surely symmetric under the interchange of any pair of variables, and the new function
f (x1 , x2 , x3 ) f (x1 , x3 , x2 ) + f (x3 , x1 , x2 ) f (x3 , x2 , x1 ) + f (x2 , x3 , x1 ) f (x2 , x1 , x3 )
(6.9)
is surely antisymmetric. These two procedures are called symmetrization and antisymmetrization (or
sometimes alternation) respectively. It is of course possible that the resulting function vanishes, but this
does not invalidate the procedure, because zero functions are both symmetric and antisymmetric!
When applied to a quantal wavefunction (x1 , x2 , x3 ), these processes result in the symmetric wavefunction
S(x
1 , x2 , x3 ) As [(x1 , x2 , x3 )+(x1 , x3 , x2 )+(x3 , x1 , x2 )+(x3 , x2 , x1 )+(x2 , x3 , x1 )+(x2 , x1 , x3 )] (6.10)
137
and in the antisymmetric wavefunction
A(x
1 , x2 , x3 ) Aa [(x1 , x2 , x3 )(x1 , x3 , x2 )+(x3 , x1 , x2 )(x3 , x2 , x1 )+(x2 , x3 , x1 )(x2 , x1 , x3 )] (6.11)
where the constants As and Aa are inserted simply to insure normalization. Note that the operators S and
A defined above are linear, whence a basis for symmetric wavefunctions may be obtained by symmetrizing
every element of an ordinary, non-symmetrized, basis, and similarly for antisymmetric wavefunctions.
For functions of three variables that factorize into three functions of one variable, i.e. functions of the
form (6.5), these definitions can be expressed even more simply because, for example, swapping s and t is
the same as swapping x2 and x3 . Thus, if |r, s, ti is an energy eigenstate, then
s, ti As [|r, s, ti + |r, t, si + |t, r, si + |t, s, ri + |s, t, ri + |s, r, ti]
S|r,
(6.12)
is a symmetric state with the same energy, while

s, ti Aa [|r, s, ti |r, t, si + |t, r, si |t, s, ri + |s, t, ri |s, r, ti]
A|r,
(6.13)
in an antisymmetric state with the same energy.

It is easy to see that the symmetrization process is unaffected by the order of the letters, i.e. that
s, ti = S|r,
t, si etc.,
S|r,
(6.14)
whence, for example, the states |1, 1, 2i, |1, 2, 1i, and |2, 1, 1i all symmetrize to the same state. Similarly in
antisymmetrization the order of the letters affects only the sign of the result,
s, ti = A|r,
t, si = A|t,
r, si etc.,
A|r,
(6.15)
2, 3i is the same
but this result is considerably more profound: It shows not only that, for example, A|1,
2, 1i, but also that
state as A|3,
1, 1i = A|3,
1, 1i
A|3,
(6.16)
1, 1i = 0, and, in general, that |r, s, ti antisymmetrizes to zero unless r, s, and t are all distinct.
whence A|3,
Now we construct a basis for symmetric wavefunctions by applying the operator S to every element of
our non-symmetrized basis on page 136. We obtain
138

1,1,1i
S|
2,1,1i
S|
1,2,1i
S|
1,1,2i
S|
3,1,1i
S|
..
.
S| 3 , 2 , 1 i
3,1,2i
S|
..
.
S| 4 , 2 , 1 i
..
.
S|M,M,Mi
where the crossed out elements are those that have already appeared earlier in the list. It is clear that there
are fewer elements in this basis than there are in the non-symmetrized basis, although it is a challenge to
count them exactly. The number of elements turns out to be
(M )(M + 1)(M + 2)
,
3!
(6.17)
or, for the N -particle rather than the three-particle case,

M +N 1
N
!
.
where the binomial coefficient symbol is defined through

!
m
m!
.
n!(m n)!
n
(6.18)
(6.19)
139
We construct a basis for antisymmetric wavefunctions in the same way. The result is
1,1,1i
A|
2,1,1i
A|
1,2,1i
A|
1,1,2i
A|
3,1,1i
A|
..
.
A| 3 , 2 , 1 i
3,1,2i
A|
..
.
A| 4 , 2 , 1 i
..
.
A|M,M,Mi
where now the crossed out elements have either appeared earlier in the list or else are equal to zero. There
are even fewer elements in this basis than there were in the symmetrized basis. In fact there are exactly
(M )(M 1)(M 2)
,
3!
(6.20)
such elements or, for the case of N particles,

M
N
M!
N !(M N )!
(6.21)
elements.
For the case N = 2, there are M (M + 1)/2 elements the symmetric basis and M (M 1)/2 elements in
the antisymmetric basis, so the sum is M 2 , the number of elements in the non-symmetrized basis. In fact,
for N = 2, the set resulting from the conjunction of the symmetric and antisymmetric bases is a basis for
the entire set of any function of two variables. This is related to the fact that any two-variable function can
be written as the sum of a symmetric function and an antisymmetric function. I point out these results to
emphasize that they apply for the two-variable case only, and are not general properties of symmetric and
antisymmetric functions. For N 3, the conjunction of the symmetric and antisymmetric bases does not
span the set of all N -variable functions.
6.3.3
The occupation number representation
We have seen that in order to specify an element of the symmetric or the antisymmetric basis that we have
just produced, it is not necessary to specify the order of the one-particle level building blocks. For example
9, 7i is the same state as A|4,
7, 9i, so theres no need to pay attention to the order in which the 4, 7,
A|4,
and 9 appear. This observation permits the occupation number representation of such states, in which
140
we specify the basis state simply by listing the one-particle levels that are used as building blocks to make
up that state. Or, equivalently but more commonly, we specify the basis state by listing the number nr
of one-body levels of each type r that are used as building blocks. (And, of course, we must also specify
whether were considering the symmetric or the antisymmetric basis.) Thus, for example:
level r:
S|3, 3, 4i has nr :
3, 4i has nr :
A|1,
1
0
1
2
0
0
3
2
1
4
1
1
5
0
0
6
0
0
M
0
0
3, 4i is built by starting with the three levels

The second line in this table means that the state S|3,
3 (x1 ), 3 (x2 ), and 4 (x1 ), multiplying them together, and then symmetrizing. Sometimes you will hear this
state described by the phrase there are two particles in level 3 and one particle in level 4, but that cant be
literally true. . . the three particles are identical, and if they could be assigned to distinct levels they would
not be identical! Phrases such as the one above1 invoke the balls in buckets picture of N -particle quantal
3, 4i is pictured as two balls in bucket number 3 and one ball in bucket number
wavefunctions: The state S|3,
4. It is all right to use this picture and this phraseology, as long as you dont believe it. Always keep in
mind that it is a shorthand for a more elaborate process of building up states from levels by multiplication
and symmetrization.
The very term occupation number for nr is a poor one, because it so strongly suggests the balls-inbuckets picture. A somewhat better name for nr is occupancy, and I will sometimes use it. If you can
think of a better name, please tell me.
To summarize the occupation number representation: an element of the symmetric basis is specified by
the list
nr , for r = 1, 2, . . . M,
where nr is 0, 1, 2, . . . ,
(6.22)
and an element of the antisymmetric basis is specified by the list
nr ,
for
r = 1, 2, . . . M,
where
nr is 0 or 1.
(6.23)
The total number of particles in such a state is

N=
M
X
nr ,
(6.24)
n r r .
(6.25)
r=1
and the energy of the state is

E=
M
X
r=1
Finally, since we have been devoting so much attention to energy eigenstates, I remind you that there
do exist states other than energy states. (Expand on this paragraph and insert a reference to the caveat
concerning meaning of eE for arbitrary states on page 108.)
1 For
example, phrases like the level is filled or the level is empty or the level is half-filled.
6.4. STATISTICAL MECHANICS OF INDEPENDENT IDENTICAL PARTICLES
6.3.4
141
Problems
6.3 Bases in quantum mechanics

We have just produced an energy eigenbasis for independent non-identical particles, one for independent
bosons, and one for independent fermions. In each case did we produce the only possible energy
eigenbasis or just one of several possible energy eigenbases? If the particles interact, the states in
question will no longer constitute an energy eigenbasis. But will they constitute a basis?
6.4 Symmetrization in the N -variable case
Equations (6.8) and (6.9) contain algorithms for constructing (anti)symmetric functions of three variables by summing up six terms. How do these equations generalize to functions of N variables and, in
particular, how many terms appear in each sum?
6.5 Symmetrizing the symmetric
The non-symmetrized state |r, s, ti is already symmetric in the case that r = s = t. What happens if
you attempt to symmetrize this already symmetric state through the procedure (6.12)? What if you
attempt to antisymmetrize it through the procedure (6.13)?
6.6 Normalization constants
Find the normalization constants As and Aa in equations (6.12) and (6.13). Be sure your formula
3, 5i and S|4,
4, 4i. Generalize your result to the N -particle case, where the
works for the case of S|3,
answer is a function of n1 , n2 , . . . , nM as well as N .
6.7 Algorithms for permutations
Invent some problem concerning the Heap algorithm for generating permutations. Better just give a
reference to Sedgwicks paper.
6.8 Algorithms to list basis states
Come up with some algorithm for producing the three sets of basis states that we have discussed,
preferably in some useful order (where useful is to be defined by you). Your algorithm should
generalize readily to the N -particle case.
6.9 Number of basis elements
Perform the counts (6.18) and (6.21). (Clue: Do it first for N = 3, then for N arbitrary.)
6.4
6.4.1
Statistical Mechanics of Independent Identical Particles

Partition function
Now that we have an energy eigenbasis, the obvious thing to do is to calculate the canonical partition function
X
Z() =
eE ,
(6.26)
states
where for fermions and bosons, respectively, the term state implies the occupation number lists:
142

P
(n1 , n2 , , nM ), nr = 0, 1, subject to r nr = N
P
(n1 , n2 , , nM ), subject to r nr = N
fermions
bosons
As we have seen, it is difficult to even count these lists, much less enumerate them and perform the relevant
sum! It can be done, but there is a trick that renders it unnecessary. (Dont be ashamed if you dont see
the trick. . . neither did Einstein or Fermi. They both did it the hard, canonical way.)
The trick here, as in so many places in statistical mechanics, is to use the grand canonical ensemble. In
this ensemble, the partition function is
(, ) =
eE+N =
states
P
r
(nr r nr )
states
M
X Y
enr (r )
(6.27)
states r=1
where the term state now implies the occupation number lists without any restriction on total particle
number:
fermions
bosons
(n1 , n2 , , nM ), nr = 0, 1
(n1 , n2 , , nM )
Writing out the sum over states explicitly, we have for fermions
(, )
1
1
X
X
1
X
#"
e
enr (r )
(6.28)
nM =0 r=1
n1 =0 n2 =0
"
M
1
Y
X
n1 (1 )
n1 =0
1
X
n2 (1 )
"
n2 =0
1
X
#
e
nM (1 )
(6.29)
nM =0
A typical factor in the product is

"
1
X
#
enr (r ) = 1 + e(r ) ,
(6.30)
nr =0
so for fermions
M
Y
(, ) =
1 + e(r ) .
(6.31)
r=1
Meanwhile, for bosons, the explicit state sum is

(, )
M
Y
X
enr (r )
n =0 n =0
nM =0 r=1
"1 2
#"
X
X
n1 (1 )
n1 =0
#
en2 (1 )
n2 =0
and a typical factor in the product is

"
#
X
nr (r )
e
=
(6.32)
"
#
enM (1 ) ,
(6.33)
nM =0
1 + [e(r ) ] + [e(r ) ]2 + [e(r ) ]3 +
(6.34)
nr =0
1
1
e(r )
(6.35)
143
where in the last line we have summed the geometric series under the assumption that r > . Thus for
bosons
M
Y
1
.
(6.36)
(, ) =
(
r )
1e
r=1
The two results are compactly written together as
(, ) =
M
Y
[1 e(r ) ]1 ,
(6.37)
r=1
where the + sign refers to fermions and the sign to bosons.
6.4.2
Mean occupation numbers
In our previous work, we have always found the partition function and worked from there. Surprisingly,
however, for the situation of quantal ideal gases it is more useful to find the mean occupation numbers, such
as
P
P
P
(nr r nr )
(EN )
r
n
e
n
e
5
5
states
states
hn5 i = P
.
(6.38)
=
(EN )
(, )
states e
Note that the averages hnr i are functions of and (as well as of r) but it is notationally clumsy to show
that dependence.
How can such averages be evaluated? A slick trick would be helpful here! Consider the derivative
P
ln
1
1 X
=
=
(n5 )e r (nr r nr ) = hn5 i.
(6.39)
5
5
states
Using the explicit expression (6.37) for (where the + sign refers to fermions and the sign to bosons),
this gives
1 ln
5
(
)
M
X
1
(r ) 1
ln[1 e
]
5 r=1

1
(5 ) 1
ln[1 e
]
5

1
()e(5 )
[1 e(5 ) ]
1
,
e(5 ) 1
hn5 i =
=
=
=
=
(6.40)
(6.41)
(6.42)
(6.43)
(6.44)
leaving us with the final result
1
.
e(r ) 1
As before, the + sign refers to fermions and the sign to bosons.
hnr i =
(6.45)
144
The mean occupation numbers play such an important role that it is easy to forget that they are only
averages, that there will be fluctuations, that for a given T and not all states will have exactly hn5 i building
blocks of level 5 (see problem 6.13). Keep this in mind if you ever find yourself saying occupation number
rather than mean occupation number.
In practice, these results from the grand canonical ensemble are used as follows: One uses these results to
find quantities of interest as functions of temperature, volume, and chemical potential, such as the pressure
p(T, V, ). But most experiments are done with a fixed number of particles N , so at the very end of your
calculation you will want to find (T, V, N ) in order to express your final answer as p(T, V, N ). You can find
(T, V, N ) by demanding that
M
M
X
X
1
N=
.
(6.46)
hnr i =
(r ) 1
e
r=1
r=1
In other words, the quantity serves as a parameter to insure normalization, very much as the quantity Z
serves to insure normalization in the canonical ensemble through
1=
X eEn
n
(6.47)
You might wonder, in fact, about the relation between the canonical probability
eEn
,
Z
(6.48)
which we have seen many times before, and the recently derived occupancy probability
1
1
.
(
)
r
Ne
1
(6.49)
The first result applies to both interacting and non-interacting systems, both classical and quantal. The
second applies only to non-interacting quantal systems. Why do we need a new probability? What was
wrong with our derivation (in section 4.1) of the canonical probability that requires us to replace it with an
occupancy probability? The answer is that nothing was wrong and that the occupancy probability doesnt
replace the canonical probability. The canonical probability and the occupancy probability answer different
questions. The first finds the probability that the entire system is in the many-body state n. The second
finds the probability that the one-body level r is used as a building block in constructing the many-body
state. Indeed, although we derived the occupancy probability result through a grand canonical argument,
it is also possible to derive the occupancy probabilities from strict canonically arguments, proof that these
two probabilities can coexist peacefully.
6.4.3
The Boltzmann limit
This is the limit where particles are far enough apart that overlap of wavefunction is minimal, so we neednt
worry about symmetrization or antisymmetrization. Equivalently, it is the limit where hnr i 1 for all r.
6.4.4
145
Problems
6.10 Evaluation of the grand canonical partition function

Can you find a simple expression for (, ) for non-interacting particles in a one-dimensional harmonic
well? For non-interacting particles in a one-dimensional infinite square well? For any other potential?
Can you do anything valuable with such an expression once youve found it?
6.11 Entropy of quantal ideal gases
This problem derives an expression for the entropy of a quantal ideal gas in terms of the mean occupation numbers hnr i. (Compare problem 4.3.) Throughout the problem, in the symbols and , the
top sign refers to fermions and the bottom sign refers to bosons.
a. Use the connection between thermodynamics and statistical mechanics to show that, for any
system,
S(T, V, )
ln
= ln
.
(6.50)
kB
b. Show that for the quantal ideal gas,

ln (T, V, ) =
ln(1 hnr i).
(6.51)
c. The mean occupation numbers hnr i are functions of T , V , and (although it is notationally
clumsy to show this dependence). Show that

hnr i
= (r )hnr i(1 hnr i)
V,
= [ln(1 hnr i) lnhnr i]hnr i(1 hnr i).
(6.52)
X
[hnr i lnhnr i (1 hnr i) ln(1 hnr i)].
(6.53)
d. Finally, show that

S(T, V, ) = kB
e. Find a good approximation for this expression in the Boltzmann limit, hnr i 1.
f. (Optional.) Find an expression for CV in terms of the quantities hnr i.
6.12 Isothermal compressibility of quantal ideal gases
a. Show that in a quantal ideal gas, the isothermal compressibility is
P

hnr i2
1
T =
1 Pr
,
kB T
r hnr i
(6.54)
where as usual the top sign refers to fermions and the bottom sign to bosons. (Clue: Choose the
most appropriate expression for T from those uncovered in problem 3.33.)
b. Compare this expression to that for a classical (Maxwell-Boltzmann) ideal gas.
146

c. The negative sign in the expression for fermions opens the possibility that T could be negative.
Prove that this potential horror never happens.
d. Do the relative sizes of the three compressibilites (fermion, classical, boson) adhere to your qualitative expectations? (Compare problem 6.29.)
6.13 Dispersion in occupation number

Find an expression analogous to (6.45) giving the dispersion in the occupation numbers. (Clue: A slick
trick would be helpful here.) Answer:
nr =
6.5
e(r )/2
p
1
= hnr i(1 hnr i)
(
)/2
r
e
(6.55)
Quantum Mechanics of Free Particles
Particle in a box. Periodic boundary conditions. k-space. In the thermodynamic limit, the dots in k-space
become densely packed, and it seems appropriate to replace sums over levels with integrals over k-space
volumes. (In fact, there is at least one situation (see equation 6.83) in which this replacement is not correct.)
Density of levels in k-space:
V
8 3
(worth memorizing).
(6.56)
Energy density of levels:
number of one-body levels with r from E to E + dE G(E) dE = V
2m3
E dE.
2 2 h3
(6.57)
How to use energy density of levels:

X
r
Z
f (r )
G(E)f (E) dE
(6.58)
and this approximation (usually) becomes exact in the thermodynamic limit.
6.5.1
Problems
6.14 Free particles in a box

We argued that, for a big box, periodic boundary conditions would give the same results as clamped
boundary conditions. Demonstrate this by finding the density of levels for three-dimensional particle
in a box problem.
6.15 Density of levels for bound particles
What is the density of levels G(E) for a one-dimensional harmonic oscillator with spring constant K?
For a three-dimensional isotropic harmonic oscillator?
6.6. FERMI-DIRAC STATISTICS
147
6.16 Density of levels in d dimensions

What is the density of levels G(E) for free particles subject to periodic boundary conditions in a world
of d dimensions?
6.6
Fermi-Dirac Statistics
In three dimensions, the chemical potential decreases with temperature. Why? It is clear that at T = 0,
= EF > 0. But as the temperature rises the gas approaches the classical limit, for which < 0 (see
problem 2.23). This is not proof, but it makes sense that should decrease with increasing temperature. A
proof is available but surprisingly difficult.
6.6.1
Problems
6.17 Qualitative origin of EF

The mean kinetic energy of electrons in metals at room temperature is usually many times the thermal
energy kB T . Which of the following can best be used to explain this fact?
a. The time-energy uncertainty relation.
b. The Pauli exclusion principle.
c. The degeneracy of the energy levels.
d. The Born approximation.
e. The wave-particle duality.
6.18 Fermion gas in two dimensions
(This problem is based on one in Ashcroft and Mermin, page 53.)
Consider a gas of free, independent, spin- 21 fermions in two dimensions. The gas is contained within
an area (or two dimensional volume) of A.
a. What is the density of one-particle levels in k-space?
b. How does the Fermi energy EF depend upon the density N/A?
P
c. Use r hnr i = N to show that
+ kB T ln(1 + e/kB T ) = EF .
(6.59)
Notice that the chemical potential decreases with temperature.

6.19 Dependence of chemical potential on temperature
Show that for independent (not necessarily free) fermions, the (T ) curve has slope (when N and V
are constant)
R
2
d
1 0 G(E)sech ((E )/2)(E ) dE
R
=
.
(6.60)
dT
T
G(E)sech2 ((E )/2) dE
0
148

Can you use this result to show that the chemical potential must decrease with temperature? I cant.
6.20 Thermodynamics of the fermion gas

Consider a collection of free and independent spin- 12 fermions. Do not assume that the temperature
vanishes.
a. Use the fundamental grand canonical result = kB T ln to show that
Z
p(T, )V = kB T
G(E) ln(1 + e(E) ) dE.
(6.61)
b. Use the expression for G(E) and the change of variable x = E to find
"
#Z
2m3
5/2
p(T, ) = (kB T )
x ln(1 + e ex ) dx.
2 h3
0
(6.62)
c. Integrate by parts to obtain

2
p(T, ) = (kB T )5/2
3
"
2m3
2 h3
#Z
x3/2
ex e + 1
dx.
(6.63)
(Do not attempt to evaluate the integral that remains.)

d. Meanwhile, show that the total energy
Z
E(T, V, ) =
G(E)f (E)E dE
(6.64)
is given by
"
E(T, V, ) = V (kB T )5/2
2m3
2 h3
#Z
0
x3/2
ex e + 1
dx.
(6.65)
Thus the pressure and energy are related by

pV = 32 E,
(6.66)
which is exactly the same relationship they have in the classical monatomic ideal gas! (See problem 6.28.)
6.21 Pressure at zero temperature
Use the result of the previous problem to show that at absolute zero the pressure of a gas of free
and independent spin- 12 fermions does not vanish (as it does for a classical ideal gas). Instead, the
zero-temperature pressure is 25 EF .
6.22 Mass-radius relation for white dwarf stars
(This problem is modified from Kittel and Kroemer, Thermal Physics, second edition, page 216.)
A white dwarf star (see Kittel and Kroemer, pp. 196198) consists of highly compressed hydrogen in
which the atoms have ionized into independent protons and electrons. In most cases it is a good model
to assume that the electrons are free and independent non-relativistic fermions at zero temperature.
Consider a white dwarf of mass M and radius R, containing N electrons.
6.7. BOSE-EINSTEIN STATISTICS
149
a. Show that to a very good approximation, N = M/mp , where mp is the mass of a proton.
b. Show that the gravitational potential energy of a uniform sphere of mass M and radius R is
cGM 2 /R, where G is the gravitational constant and c is a dimensionless constant. (In fact
c = 53 but this is tedious to show.)
c. Show that the kinetic energy of the electrons is

1/3
h2 M 5/3
9 3 2
,
KE =
5/3 R2
20
2
me mp
(6.67)
where me is the mass of an electron.

d. If the potential and kinetic energies satisfy
KE = 12 PE,
(6.68)
as required by the virial theorem of mechanics, show that

RM 1/3 = a constant of approximate value 1017 m kg1/3 .
(6.69)
Evaluate the constant assuming that c = 35 . Note that the radius decreases as the mass increases.
e. If the white dwarf has the mass of the sun (2 1030 kg), what is its radius (in km)? Compare
this to the radius of our sun.
f. Neutron stars (observed as pulsars) are also zero temperature fermion gases, but in this case the
fermions are neutrons rather than electrons. Derive the mass-radius relation for a neutron star,
and use it to find the radius of a neutron star with the mass of the sun.
6.7
Bose-Einstein Statistics
Perhaps you thought the Fermi-Dirac results were strange: Non-interacting particles forming a collection as
hard as steel. . . room temperature being effectively zero. Wait until you see the Bose-Einstein results.
6.7.1
Theory
For independent bosons, whether free or subject to an external potential, the mean occupation number
function is
1
.
(6.70)
b(E) = (E)
e
1
To begin to understand this function, note that
when:
E <
E =
E >
we have:
b(E) negative
b(E) =
b(E) positive
150
Thus this function has the general character sketched below:
b(E)
1 2 3
...
Although we have written b(E) as a function of the continuous variable E, we will in fact have occasion to
evaluate it only at the energy eigenvalues 1 , 2 , 3 ,. . . , r ,. . . , eigenvalues which of course differ for different
external potentials.
It seems bizarre that b(E) can be negative, and indeed this is only a mathematical artifact: Recall that
in our derivation of the Bose function we needed to assume that < 1 in order to insure convergence
(see equation 6.35). Evaluating b(E) at any eigenvalue r will always result in a positive mean occupation
number.
The character of the Bose function is dominated by the singularity at E = , so in trying to understand the
function and its physical implications one must first locate the chemical potential. This section will provide a
tour of Bose-Einstein behavior with decreasing temperature, throughout which the chemical potential shifts
to the right. (When we investigated Fermi-Dirac behavior, we started at T = 0 and toured with increasing
temperature, so the chemical potential shifted left.) This rightward shift presents a potential problem,
because as the temperature decreases might shift right all the way to 1 , and we know that < 1 . We
will just have to go ahead and take the tour, being wary because we know that a = 1 road block might
pop up right in front of us as we view the countryside. With any luck will not yet have reached 1 when
our tour halts at T = 0.
For the case of free and independent bosons (subject to periodic boundary conditions), the ground level
energy is 1 = 0. The natural first step is to find (T, V, N ) by demanding that
Z
N=
G(E)b(E) dE.
(6.71)
0
Natural though this may be, caution is in order. Remember that the integral above is an approximation to
the sum over discrete energy levels
X
N=
b(r ).
(6.72)
r
151
It is legitimate to replace the sum with the integral when the value of b(E) changes little from one energy
level to the next. We saw in section 6.5 that in the thermodynamic limit, the free-particle level spacings
approach zero, so for the most part this approximation is excellent. . . even exact in the thermodynamic
limit. But there is one exception: At E = , the Bose function b(E) is infinite, so b() = is very different
from b( + ), no matter how small the positive number is. Usually we can ignore this caution, because
< 1 = 0. But if approaches 0 then we expect the approximation (6.71) to fail.
In summary, the integral (6.71) is a good approximation for the sum (6.72) as long as the integrand varies
slowly. Now for any value of E greater than , you can make b(E + ) very close to b(E) simply by choosing
> 0 small enough. This is what happens in the thermodynamic limit. But for E = , there is always a
large difference between b(E) = and b(E + ), which is finite. Thus the integral approximation will be a
good one as long as we avoid E = .
In situations where the sum can be legitimately replaced with the integral, we have
Z
G(E)b(E) dE
N =
0"
#Z
2m3
1
= V
E (E)
dE.
3
2
e
1
2 h
0
Use of the obvious substitution x = E gives
"
#
Z
x1/2
2m3
3/2
dx,
N =V
(k
T
)
B
3
ex e 1
2 2 h
0
(6.73)
(6.74)
(6.75)
and remembering the definition (5.4) of thermal de Broglie wavelength results in a more convenient expression
Z
2
V
x1/2
N= 3
dx.
(6.76)
(T ) 0 ex e 1
Note that the definite integral above is not just a number. . . it is a function of the product . There
is no closed-form expression for the integral, although it is readily evaluated numerically and can be found
in tabulations of functions. However it is easy to find an analytic upper bound: Because < 0, we have
e > 1 whence
Z
Z 1/2
x1/2
x
dx
<
dx.
(6.77)
x e 1
x1
e
e
0
0
The expression on the right is just a number, and a little bit of work (see problem 6.26) shows that it is the
number
( 23 )
,
(6.78)
2
where the Riemann zeta function is defined by
(s) =
X
1
ns
n=1
and
( 23 ) = 2.612375348 . . . .
So, how does this upper bound help us? It shows that
Z
2
x1/2
V
V
N= 3
dx < 3
( 3 ),
x
1
(T ) 2
(T ) 0 e e
(6.79)
(6.80)
152
but what is the significance of this result? Remember that the upper bound is just the value of the function
when = 0, which is exactly where we expect a road block due to the breakdown of the integral approximation (6.71). Our hopes that we could avoid the issue have been dashed. The breakdown occurs at the
temperature T0 that satisfies
N=
V
3 (T0 )
( 32 )
or
T0 () =
3 (T0 ) = ( 32 )/
or
h2
2/3 .
2mkB 2/3 ( 32 )
(6.81)
(6.82)
For temperatures above T0 , the chemical potential shifts right as the temperature falls, and the integral
approximation (6.71) is legitimate (indeed, exact in the thermodynamic limit). But as the temperature
decreases below T0 , the chemical potential sticks at = 0 and the integral approximation (6.71) must be
modified.
What is the proper modification? The function b(E) is slowly varying for all values of E except E = , so
the integral approximation is legitimate for all the energy levels except the ground level at 1 = 0 = . Only
for the ground level is some other result needed, so we just add in the ground level occupancy by hand:
Z
N = hn1 i +
G(E)b(E) dE.
(6.83)
0
Now, we have already seen that when = 0the condition for validity of this equationthe integral can
be evaluated exactly and we have
V
( 3 ).
(6.84)
N = hn1 i + 3
(T ) 2
Note that hn1 i here is not given by its traditional formula (6.45), because
1
e(E)
e0
1
= .
1
(6.85)
Instead, equation (6.84) is the formula for hn1 i when T < T0 . The mean ground level occupancy hn1 i is an
intensive quantity when T > T0 but an extensive quantity when T < T0 .
In summary, the correct normalization equation breaks into two parts, namely
Z
2
V
x1/2
dx for T > T0 ()
3 (T )
x
e e
1
0
N=
( 3 )
for T < T0 ().
hn1 i + 3
(T ) 2
(6.86)
We should expect that each part will behave quite differently, i.e. we expect a sudden change of behavior as
the temperature drops through T0 .
What does all this mean physically? A naive reading of equation (6.80) suggests an upper bound on
the number of particles that can be placed into the volume V . This would be sensible if the particles were
marbles with hard-core repulsions. But these are non-interacting particles! Surely we can add more particles
just by throwing them into the container. Indeed we can do so, and the associated excess mean occupation
153
number is due to the level that particles like best at low temperatures, namely the ground level. The ground
level thus has a much higher mean occupancy than the first excited level, and this rapid variation of mean
occupancy with energy renders the approximation of sum by integral invalid. The inequality (6.80) does not
limit the number of particles in the system: instead it shows the domain within which it is legitimate to
approximate the sum (6.72) by the integral (6.71).
The abrupt transition at T0 () is called Bose-Einstein condensation and the material at temperatures
below T0 () is called the Bose condensate. These terms are unfortunate: they conjure images of a gas
condensing into a liquid, in which circumstance the atoms separate into two different classes: those in the
liquid and those remaining in the vapor. This suggests that in Bose-Einstein condensation too there are two
classes of particles: those in the ground level and those in the excited levels. This picture is totally false.
It is incorrect to say one particle is in the ground level, another is in the fourth excited level. In truth
the individual particles are not in individual levels at all: instead the whole system is in a state produced
by multiplying together the individual level wavefunctions (building blocks) and then symmetrizing them.
The literature of Bose-Einstein statistics is full of statements like at temperatures below T0 , any particle
added to the system goes into the ground level. Such statements are wrong. They should be replaced with
statements like at temperatures below T0 , any increase in particle number occurs through an increase in
hn1 i, the mean occupancy of the ground level. Or alternatively, at temperatures below T0 , it is very likely
that many of the building blocks from which the system wavefunction is built are the ground level. Or
again, to be absolutely precise, at temperatures below T0 , if the energy is measured then it is very likely
that many of the building blocks from which the resulting energy eigenfunction is built are the ground level.
Read again the previous paragraphthe one that begins What does all this mean physically?. Notice that
I never need to say that a particle is in or goes into a given level.
6.7.2
Experiment
References: M.H. Anderson, J.R. Ensher, M.R. Matthews, C.E. Wieman, E.A. Cornell, Observation of BoseEinstein condensation in a dilute atomic vapor, Science, 269 (14 July 1995) 198201; Randall G. Hullet at
Rice; Malcolm W. Browne, Two groups of physicists produce matter that Einstein postulated, New York
Times, 14 July 1995, page 1.
6.7.3
Problems
6.23 Character of the Bose function

What are the limits of the Bose function b(E) (equation 6.70) as E ? Is the curvature of the
function greater when the temperature is high or when it is low?
6.24 Thermodynamics of the Bose condensate
For temperatures less than the Bose condensation temperature T0 , find the energy, heat capacity, and
entropy of an ideal gas of spin-zero bosons confined to a volume V . Write your answers in terms of the
154

dimensionless integral
Z
I=
0
x3/2
dx
ex 1
(6.87)
but dont bother to evaluate it. Show that

CV =
5E
2T
and S =
5E
.
3T
(6.88)
2E
.
3V
(6.89)
6.25 More thermodynamics of the Bose condensate

For the system of the previous problem, show that
2
F = E
3
and p =
From this show that at low temperatures, the pressure of a collection of free and independent bosons
goes like p T 5/2 . (This is always less than the classical pressure p T .)
6.26 An integral important for Bose condensation
Show that
Z 1/2
x
3
dx = ( 2 )
x1
e
2
0
where
(s) =
Clue:
6.8
(6.90)
X
1
.
s
n
n=1
X
1
1
x
=
=
e
enx
ex 1
ex (1 ex )
n=0
(6.91)
for x > 0.
(6.92)
Specific Heat of the Ideal Fermion Gas
There is a set of the pants heuristic argument showing that at low temperatures,
CV (T ) kB G(EF )(kB T ).
(6.93)
This argument is not convincing in detail as far as the magnitude is concerned and, indeed, we will soon find
that it is wrong by a factor of 2 /3. On the other hand it is quite clear that the specific heat will increase
linearly with T at low temperatures. From equipartition, the classical result is
CVclassical = 23 N kB ,
(6.94)
and we expect this result to hold at high temperature. Thus our overall expectation is that the specific heat
will behave as sketched below.
6.8. SPECIFIC HEAT OF THE IDEAL FERMION GAS
155
CV(T)/N
(3/2)kB
EF/kB
How can we be more precise and rigorous? That is the burden of this section.
Reminders:
f (E)
EF
1
e(E)/kB T
(6.95)
+1
2/3

2
h
N
3 2
2m
V
2m
3 1
= V 2 3 E =N
E
2 E 3/2
h
F
G(E)
(6.96)
(6.97)
The expressions for N and E are

Z
N (T, V, )
G(E)f (E) dE
Z
E(T, V, )
EG(E)f (E) dE
Z
2m3 1/2
V 2 3
E f (E) dE
h
0
Z
2m3 3/2
E f (E) dE.
V 2 3
h
0
(6.98)
Remember how to use these expressions: The first we will invert to find (T, V, N ), then we will plug this
result into the second to find E(T, V, N ). As a preliminary, recognize that we can clean up some messy
expressions in terms of the physically significant constant EF by writing these two expressions as
Z
3 1
E 1/2 f (E) dE
(6.99)
1 =
2 E 3/2 0
F
Z
E
3 1
=
E 3/2 f (E) dE.
(6.100)
N
2 E 3/2 0
F
In short, we must evaluate integrals like

Z
0
E 1
e(E)/kB T
+1
dE
(6.101)
with the dimensions of [energy] . To make these more mathematical and less physical, convert to the
dimensionless quantities
x =
x0
E/kB T
(6.102)
/kB T
(6.103)
156
to find
E 1
e(E)/kB T + 1
dE = (kB T )
Z
0
Define the dimensionless integral

Z
A (x0 ) =
0
x1
dx.
exx0 + 1
x1
dx.
exx0 + 1
(6.104)
(6.105)
so that
1
3 (kB T )3/2
A3/2 (/kB T )
2 E 3/2
(6.106)
3 (kB T )5/2
A5/2 (/kB T ).
2 E 3/2
(6.107)
E
N
Let us pause before rushing forward. You might think that we should turn to Mathematica (or to a table
of integrals) and try to evaluate these integrals in terms of well-known functions like exponentials and Bessel
functions. Even if we could do this, it wouldnt help: The result would just be an incomprehensible morass
of functions, and to try to understand it we would need to, among other things, look at the low-temperature
limit and the high-temperature limit. Since we are particularly interested in the low-temperature behavior,
lets just set out to find the low-temperature series in the first place. The meaning of low temperature in
this context is kB T or x0 = /kB T 1, so we suspect a series like
E(T ) = E(T = 0) +
E1
E2
+ 2 + .
x0
x0
(6.108)
Whenever we have previously searched for such a series (for example, when we investigated internal
specific heat of the rigid dumbbell, section 5.4, or the simple harmonic oscillator, problem 5.7) we had good
luck using a Taylor series expansion about the origin. In this case, any such technique is doomed to failure,
for the following reason: We need a good approximation for the function
g(x) =
1
,
exx0 + 1
(6.109)
a function that is very flat near the origin.
g(x)
1
x0
In fact, this function is so flat at the origin that, even if we kept dozens of terms in its Taylor series expansion
about the origin, the Taylor approximation would be quite poor.
157
first 50 terms of Taylor series
g(x)
1
x0
Indeed, all of the action associated with this function happens near x = x0 . (We saw earlier, in our heuristic
discussion, that the low-temperature specific heat was due entirely to promotion of a few electrons near the
Fermi energy.) Thus an accurate low-temperature approximation will necessarily involve expansions around
x = x0 , not around x = 0.
Arnold Sommerfeld was aware this situation and had the genius to recognize that a change in focus
would solve the problem. Specifically, the focus needs to change from the function g(x) to its derivative
g 0 (x), because the derivative is very small for all x except near x = x0 , where the action is. In specific, he
noted that
g 0 (x)
=
=
=
exx0
(exx0 + 1)2
1
(xx )/2
0
(e
+ e(xx0 )/2 )2
1
.
4 cosh2 ((x x0 )/2)
The last form makes clear the remarkable fact that g 0 (x) is even under reflections about x0 .
How can we use these facts about g 0 (x) to solve integrals involving g(x)? Through integration by parts.
Z
A (x0 ) =
x1 g(x) dx
0
Z

x
x 0
=
g(x)
g (x) dx
0
Z 0
1
x
=
dx
(6.110)
4 0 cosh2 ((x x0 )/2)
Its now easy to plot the pieces of the integrand, and to see that the integrand nearly vanishes except near
x = x0 .
158
x (for = 1/2)
g(x)
1
cosh2((xx0)/2)
x0
approx
4e|xx0|
Because x varies slowly near x = x0 , it makes sense to expand it in a Taylor series about x = x0 :
f (x) = f (x0 ) + f 0 (x0 )(x x0 ) +
x =
x
+ x1
(x x0 ) +
0
0
1 00
2
2 f (x0 )(x x0 )
2
1
(x x0 )2
2 ( 1)x0
+
+
So
=
=
x
dx
cosh ((x x0 )/2)
0
Z
Z
(x x0 )
1
1
dx
+
x
dx
x
0
0
2
2
cosh
((x
x
)/2)
cosh
((x x0 )/2)
0
0
0
Z
(x x0 )2
+ 21 ( 1)x02
dx + .
cosh2 ((x x0 )/2)
0
Z
4A (x0 )
Lets think about how to evaluate integrals like

Z
(x x0 )n
dx.
cosh2 ((x x0 )/2)
0
(6.111)
The denominator is even in the variable (x x0 ), while the numerator is either even or odd in the variable
(x x0 ). If the lower limit were instead of 0, half of these integrals would vanish. . . an immense labor
savings! Now, of course, the lower limit isnt , but on the other hand if we extended the lower limit from
0 to , we wouldnt pick up much error, because the integrand nearly vanishes when x is negative. In
fact, the error introduced is
Z 0
Z
Z
(x x0 )n
(x x0 )n
(x x0 )n
dx
=
dx
dx,
(6.112)
2
2
cosh2 ((x x0 )/2)
cosh ((x x0 )/2)
cosh ((x x0 )/2)
0
and the scale of this error is about the value of the integrand at x = 0, namely
(x0 )n
(x0 )n (4ex0 ).
cosh2 (x0 /2)
159
We could find more accurate approximations for this error, but theres no need to. We know the error is of
order ex0 and were interested in case that x0 1. This error is utterly negligible. Remember that we
expect an answer of the form
E1
E2
(6.113)
E(T ) = E(T = 0) +
+ 2 + ,
x0
x0
and that (in the limit as x0 ) the error ex0 is smaller than any of these terms. . . even smaller than
E492 /x492
0 . From now on we will change the lower limit of integration from 0 to , with the understanding
that this replacement will introduce a negligible error. Thus, most of the remaining formulas in this section
will show not equality, but rather asymptotic equality, denoted by the symbol .
(Perhaps Im being too glib, because while this error is indeed small, I introduce it an infinite number
of times. This is the kind of problem you might want to take to your friendly neighborhood applied mathematician. These folks can usually solve the problem more easily than you can and they often appreciate
your bringing neat problems to them.)
Using these ideas, we write
Z
4A (x0 ) =
=
x
dx
cosh ((x x0 )/2)
0
Z
Z
(x x0 )
1
1
dx
+
x
dx
x
0
0
2
2
cosh ((x x0 )/2)
cosh ((x x0 )/2)
0
0
Z
(x x0 )2
+ 21 ( 1)x02
dx +
cosh2 ((x x0 )/2)
0
Z
Z
1
(x x0 )
1
x
dx
+
x
dx
0
0
2
2
cosh ((x x0 )/2)
cosh ((x x0 )/2)
Z
(x x0 )2
dx +
+ 21 ( 1)x02
2
cosh ((x x0 )/2)
2
Because the integrals extend from to +, it is easy to change the variable of integration to x x0 , and
then symmetry dictates that
Z
(x x0 )n
d(x x0 ) = 0 for all odd n.
2
cosh ((x x0 )/2)
Using the substitution u = (x x0 )/2, we have
Z

Z
1
4( 1)
u2
4A (x0 ) x0 2
du +
du +
2
2
x20
cosh (u)
cosh (u)
The first integral can be found in Dwight and is
Z
The second is Gradshteyn 3.527.5, namely

Z
1
du = 2.
cosh2 (u)
u2
2
du =
.
2
6
cosh (u)
(6.114)
160
Consequently
A (x0 )

x
1
2 ( 1)
0
+
O
.
1+
2
6
x0
x40
(6.115)
This is the desired series in powers of 1/x0 . . . and surprisingly, all the odd powers vanish! This result is
known as the Sommerfeld expansion.
Now we apply this formula to equation (6.106) to obtain the function (T ):
"

2
4 #
2 kB T
kB T
3/2
1+
+O
1
3/2
8
EF
"

4 #1
2
2
k
T
k
T
B
B
3/2
3/2 EF
+O
1+
8
"

2

4 #2/3
2 kB T
kB T
EF 1 +
+O
8
"

4 #
2
kB T
2 kB T
+O
= EF 1
12
This equation tells us a lot: It confirms that at T = 0, we have = EF . It confirms that at T increases from
zero, the chemical potential decreases. However it doesnt directly give us an expression for (T ) because
appears on both the left and right sides. On the other hand, it does show us that deviates from EF by
an amount proportional to (kB T /)2 , which is small. How small?

2
kB T
(kB T )2
"

2

4 #2
2 kB T
kB T
2
EF 1
+O
12
(kB T )2
2

4 #
2 kB T
kB T
2
EF 1
+O
6

4 #

2 "
2
kB T
kB T
2 kB T
+O
=
1+
EF
6
"
In other words, it is small in the sense that

2
2

4
kB T
kB T
kB T
+O
EF
EF
and that

O
kB T
4

O
kB T
EF
(6.116)
4
This gives us the desired formula for as a function of T :

"

2

4 #
kB T
2 kB T
+O
.
(T ) EF 1
12
EF
EF
(6.117)
(6.118)
161
This result verifies our previous heuristic arguments: (T ) declines as the temperature increases, and it does
so slowly (i.e. to second order in kB T /EF ).
Now apply expansion (6.115) for A to equation (6.107) to find the energy E(T, V, ):
"

2

4 #
E
3 5/2
5 2 kB T
kB T
1+
+O
.
N
5 E 3/2
8
(6.119)
Using equations (6.116) and (6.117):

"

2

4 #
5 2 kB T
3 5/2
kB T
E
1+
+O
.
N
5 E 3/2
8
EF
EF
(6.120)
Now its time to plug (T, V, N ) into E(T, V, ) to find E(T, V, N ): Use equation (6.118) for (T, V, N ), and
find
"

2

4 #5/2
2
k
T
k
T
B
B
5/2
+O
(T )5/2 EF
1
12
EF
EF
"

4 #
2
5 2 kB T
kB T
5/2
= EF
1
+O
.
(6.121)
24
EF
EF
162
Using (6.121) in (6.120) gives

"

2
4 # "

2
4 #
kB T
5 2 kB T
kB T
E
3
5 2 kB T
+O
1+
+O
EF 1
N
5
24
EF
EF
8
EF
EF
whence
"

2
4 #
3
kB T
5 2 kB T
E N EF 1 +
+O
.
5
12
EF
EF
(6.122)
(6.123)
Wow! At long last were able to say

CV =
whence
E
T

=
N,V
E
(kB T /EF )

N,V
kB
EF

(6.124)
"

3 #
3
kB T
5 2 kB T
,
CV N kB
+O
5
6
EF
EF
or, for low temperatures.
2
CV
N kB
2
kB T
EF
(6.125)

.
(6.126)
Was all this worth it, just to get a factor of 2 /3? Perhaps not, but our analysis gives us more than the factor.
It gives us a better understanding of the ideal fermion gas and a deeper appreciation for the significance of
the thin action region near E = EF .
6.9
Additional Problems
6.27 What if there were no interchange rule?

Suppose that the interchange rule did not apply, so that the basis on page 136 were the correct one for
three identical particles. (Alternatively, consider a gas of N non-identical particles.) Find and sketch
the heat capacity as a function of temperature.
6.28 Pressure and energy density
(This problem was inspired by Reif problem 9.5.)
Any non-relativistic monatomic ideal gas, whether classical or quantal, satisfies
p=
2E
.
3V
(6.127)
This remarkable fact can be demonstrated most easily in the canonical ensemble. (Throughout this
problem the particle number is fixed so the N dependence of functions is never mentioned explicitly.)
a. Use E = 32 N kB T to demonstrate the relation for the classical monatomic ideal gas.
b. From thermodynamics, show that
ln Z
E(T, V ) =

V
1 ln Z
and p(T, V ) =
V

.
T
(6.128)
163
c. Argue that temperature and energy eigenvalues enter into the canonical partition function as
quotients:
ln Z(T, V ) = F(1 , 2 , . . . , M ).
(6.129)
d. Show that in non-relativistic quantum mechanics the free particle energy eigenvalues depend on
box volume through
2 r
r
=
.
(6.130)
V
3V
e. Use the last three items together to prove the pressure-energy density relation.
f. How are the pressure and energy density related for blackbody radiation? At what stage does the
above proof break down in this case?
6.29 Pressure comparison
(This problem is modified from one in a GRE Physics test.)
Consider three systems of non-interacting identical particles, each with the same T , V , and N . In one
system the particles are fermions, in another they are bosons, and in a third they behave classically.
Which system has the greatest pressure? Which has the smallest?
6.30 Challenge
For many years I suspected that the chemical potential of an ideal gas would have to decrease or at
least remain constant when the temperature increased (with V and N constant). I tried proving this
in a number of ways, and in my attempts I came across several interesting facts (such as the results
of problem 3.36 and part (f.) of problem 6.11) but I was never been able to prove the desired result.
Thats because the result is false! The chemical potential increases with temperature for ideal fermions
in one dimension. Can you show this?
164
Appendix
Even though we havent been able to derive everything on this graph, I cant resist showing you the equation
of state (that is, pressure as a function of density and temperature), for ideal fermions, ideal classical particles,
and ideal bosons.
classical p = kBT
p(,T)
fermions
T2
(2/5)EF
bosons
T 5/2
0
T
0
T0
TF
Chapter 7
Harmonic Lattice Vibrations

7.1
The Problem
7.1 The harmonic Hamiltonian

The Hamiltonian for lattice vibrations, in the harmonic approximation, is
H=
1
2
3N
X
mi x 2i +
i=1
1
2
3N X
3N
X
xi Aij xj .
(7.1)
i=1 j=1
Notice that this Hamiltonian allows the possibility that atoms at different lattice sites might have
different masses. Accept the fact that any real symmetric matrix S can be diagonalized through an
orthogonal transformation, i.e. that for any such S there exists a matrix B whose inverse is its transpose
and such that
BSB1
(7.2)
is diagonal. Show that the Hamiltonian can be cast into the form
H=
1
2
3N
X
(qr2 + Dr qr2 )
(7.3)
r=1
by a linear change of variables. (Clue: As a first step, introduce the change of variable zi =
7.2
mi xi .)
Statistical Mechanics of the Problem
7.2 Thermal energy versus kinetic energy

According to the Fall 2001 Review Draft of Ohios Academic Content Standards for Science (page 48),
eight-grade students should understand that thermal energy is in the disorderly motion of molecules.
Explain how position as well as motion (i.e. potential energy as well as kinetic energy) contribute to
thermal energy.
165
166
CHAPTER 7. HARMONIC LATTICE VIBRATIONS
7.3
Normal Modes for a One-dimensional Chain
The matrix A is all zeros except for 2 on the diagonal and 2 on the superdiagonal. But this doesnt really
help us solve the problem. The solution comes from physical insight, not mathematical trickery!
Dispersion relation:
r
(k) = 2
K
| sin( 12 ka)|
m
(7.4)
Meaning of term dispersion relation:

Start with an arbitrary wave packet, break it up into Fourier components.
Each such component moves at a particular speed.
After some time, find how all the components have moved, then sew them back together.
The wave packet will have changed shape (usually broadened. . . dispersed).
Remember that we havent done any statistical mechanics in this section, nor even quantum mechanics.
This has been classical mechanics!
7.4
Normal Modes in Three Dimensions
If you wanted to study heat flow, or how sound warms up a crystal, or how light warms up a crystal, youd
have to study these time-dependent dispersion effects. This is hard: the classical mechanics is hard (normal
modes) and the quantum mechanics is hard. For the purposes of this book we need only examine static
effects, and in particular we need only find the heat capacity at low temperatures.
7.5
Low-temperature Heat Capacity
If
G() d = number of normal modes with frequencies from to + d
then
E crystal =
G()eSHO () d
and
CVcrystal =
(7.5)
G()cSHO
() d
V
(7.6)
and so forth.
Density of modes:
G() d
[vol. of shell in k-space](density of modes in k-space)
branches
X
branches
[4(kb ())2 dkb ]
V
8 3

(7.7)
7.5. LOW-TEMPERATURE HEAT CAPACITY
167
This formula holds for any isotropic dispersion relation kb (). For small values of the dispersion relation
for each branch is linear (with sound speed cb ) so
kb =
cb
and
dkb =
d
,
cb
(7.8)
whence
"
G() d
#

2
d
V
=
4
cb
cb
8 3
branches
!
3
X
V
1
=
2 d.
2 2
c3b

(7.9)
b=1
If we define the average sound speed cs through the so-called harmonic cubed average,
3
1
1X 1
,
c3s
3
c3b
(7.10)
b=1
then we have the small- density of modes

G() d
3V 2
d.
2 2 c3s
(7.11)
At any temperature,
CVcrystal
G()cSHO
() d.
V
(7.12)
Recall from equation (5.78) that

cSHO
() = kB
V
h
kB T
2
eh/kB T
,
(1 eh/kB T )2
and using the small- result (7.11), we have the low-temperature result
2

Z
eh/kB T
3V 1
h
2
k
.
CVcrystal =
d
B
2
3
2 cs
kB T
(1 eh/kB T )2
0
(7.13)
(7.14)
For our first step, avoid despair instead convert to the dimensionless variable
x=
h
kB T
and find
CVcrystal
3V 1
kB
2 2 c3s
kB T
h
3 Z
0
x4 ex
dx.
(1 ex )2
(7.15)
The integral is rather hard to do, but we dont need to do it the integral is just a number. We have
achieved our aim, namely to show that at low temperatures, CV T 3 .
168
CHAPTER 7. HARMONIC LATTICE VIBRATIONS

However, if you want to chase down the right numbers, after some fiddling youll find that
Z
x4 ex
4 4
dx = 4(4)(4) =
.
x )2
(1
e
15
0
Thus, the low-temperature specific-heat of a solid due to lattice vibration is

3
2 2 kB T
.
CVcrystal = kB V
5
hcs
(7.16)
7.3 How far do the atoms vibrate?

Consider a simplified classical Einstein model in which N atoms, each of mass m, move classically on
a simple cubic lattice with nearest neighbor separation of a. Each atom is bound to its lattice site by
a spring of spring constant K, and all the values of K are the same. At temperature T , what is the
root mean square average distance of each atom from its equilibrium site? (Note: I am asking for an
ensemble average, not a time average.)
7.6
More Realistic Models
7.4 Debye frequency

In the Debye model, a solid with average sound speed cs has density of normal-mode frequencies
3V 2 for <
D
2 2 c3s
G() =
(7.17)
0
for > D .
Find a formula for D (N, V, cs ), and write G() in terms of D .
7.5 Debye model energy and heat capacity
Find E(T, V, N ) and CV (T, V, N ) for a harmonic solid in the Debye model, in terms of D and the
function
Z x
3
t3
D(x) = 3
dt.
(7.18)
t
x 0 e 1
7.7
What is a Phonon?
7.8
Additional Problems
7.6 Spin waves

For harmonic lattice vibrations at low frequencies, = cs k. There are analogous excitations of ferromagnets called spin waves which, at low frequencies, satisfy = Ak 2 . Find the temperature
dependence of the heat capacity of a ferromagnet at low temperatures. (Do not bother to evaluate
constants. . . I am only looking for the functional form of the temperature dependence.)
169
7.7 Comparison of models

One feature common to both the Debye and Einstein models for the specific heat of a crystal composed
of N identical atoms is that
a. The average energy of each atom is 3kB T .
b. The vibrational energy of the crystal is equivalent to the energy of 3N independent harmonic
oscillators.
c. The crystal is assumed to be continuous for all elastic waves.
d. The speed of longitudinal elastic waves is less than the speed of transverse elastic waves.
e. The upper cutoff frequency of elastic waves is the same.
Chapter 8
Interacting Classical Fluids

8.1
Introduction
The subject of this chapter is also called real gases, or dense gases, or non-ideal gases, or imperfect
gases, or liquids and dense gases. The many names are a clue that the same problem has been approached
by many different scientists from many different points of view, which in turn is a hint that the problem
is one of enormous importance. And in this case the hints are correct. We have treated non-interacting
classical atoms for long enough in this book. We know that this treatment leads to results in which all
substances have the same equation of state (although they can sometimes be distinguished through their
heat capacities) and in which there are no phase transitions, no crystals, no life, and very little of interest.
It galls me that anyone ever called such a situation ideal. The problem of interacting atoms is important
for industrial and military applications, but it is just as important from the point of view of pure science.
We are now going to see what happens when we allow the atoms to interact. Some call the methods and
results messy. I call them fascinating. I would not want to live in the ideal world.
To be specific, we will consider the model problem of:
Monatomic atoms with no internal structure, interacting through spherically symmetric pair potentials.
Classical mechanics.
Many undergraduate books state that this is an important outstanding problem. That is no longer correct.
In fact, the problem was solved (except near phase transitions) in the years 198085 by Lado, Ashcroft, Foiles,
and Rosenfeld (see Talbot, Lebowitz, et al. A comparison of MHNC and MC at high pressures, J. Chem.
Phys. 85 (1986) 21872192, Y. Rosenfeld, Variational MHNC, J. Stat. Phys. 42 (1986) 437457). The
solution is a technique called modified hyper-netted chain. Unfortunately, I will not have time to tell you
what the solution is or even what the words in the name mean. But I can tell you that it is an ingenious
170
8.1. INTRODUCTION
171
combination of perturbation and variational methods, and I will be telling you about these two techniques in
this chapter. (The best references for this chapter are the books by Donald McQuarrie, Statistical Mechanics,
and by J.P. Hansen and I.R. McDonald, Theory of Simple Liquids.)
Problems
8.1 The Lennard-Jones potential
The general Lennard-Jones potential is
u(r) =
a
b
+ n.
r6
r
(8.1)
a. Given a and b, find m , the minimum energy, and r0 , the distance of separation at that minimum.
Write u(r) in terms of the parameters m and r0 rather than the parameters a and b.
b. If is the separation at which u() = 0, find in terms of r0 . Why is m irrelevant here?
c. Note the simplifications in the above results if n = 12. For this case, write u(r) using the
parameters m and .
8.2 Negative compressibilities
Look at figure 8.6.1 on page 307 of Reif. Notice that for temperatures below the critical temperature
(T4 in the figure), the van der Waals fluid can have negative values of T . Suppose a fluid with negative
compressibility existed, and I had a bottle of it on the table in front of me. What would happen if
the volume of the bottle decreased slightly. . . say if I rapped my fingers on it? (This impossibility is
a warning that the van der Waals equation is simply not applicable at temperatures in the vicinity of
the critical temperature or below.)
8.3 Entropy of a real gas
A real gas obeys the equation of state

N
pV = N kB T 1 + B
,
V
(8.2)
where B is a positive constant.

a. In one sentence, describe the behavior of this gas in the limit that T and N are constant but
V . (Answers containing mathematical symbols are automatically disqualified.)
b. Using an appropriate Maxwell relation, calculate the difference in entropy between an ideal gas
and this real gas when both have the same T , V , and N .
c. At a given T , V , and N , which entropy is larger? Is this consistent with the entropy inequality?
Can you justify it with a qualitative argument? Is it possible for the constant B to be negative
or zero?
172
8.2
8.2.1
CHAPTER 8. INTERACTING CLASSICAL FLUIDS
Perturbation Theory
Fluids in the canonical ensemble
The partition function for a fluid is

Z(T, V, N ) =
where
1
h3N N !
deH()
(8.3)
H() =
1 X 2
p + UN (r1 , . . . , rN ).
2m i=1 i
The momentum integrals can be performed precisely as they were for the ideal gas, giving
Z
1
dr1 drN eUN (r1 ,...,rN ) ,
Z(T, V, N ) = 3N
(T )N !
(8.4)
(8.5)
where (T ) is the thermal de Broglie wavelength. It is useful to define the configurational partition function
Z
1
QN (T, V ) =
dr1 drN eUN (r1 ,...,rN ) .
(8.6)
N!
For example, in a three-particle system with spherically symmetric pair potential u(r), the configurational
partition function is
Z
Z
Z
1
Q3 (T, V ) =
dr1
dr2
dr3 e[u(|r1 r2 |)+u(|r1 r3 |)+u(|r2 r3 |)] .
(8.7)
6 V
V
V
Every other time in this book that we have encountered a complicated partition function, like the one
above, it has factorized into the product of many similar terms. (This happened even in our treatment of
harmonic lattice vibrations, through the use of normal coordinates.) The expression above does not factorize
in this neat way. With considerable diligence and patience, it is possible to evaluate QN (T, V ) for N = 3.
But you can see that doing it for cases of interest, such as N = 6.02 1023 , is simply out of the question.
Instead we attempt to use perturbation theory by expanding about an already solved problem, namely the
ideal gas. Such an approximation will be a good one at low densities, where the gas is nearly ideal. The
obvious thing to do is to expand QN (T, V ) in the small parameter = N/V , the number density. But how
can we do that? The quantity doesnt appear in these formulas!
8.2.2
The search for a small parameter
We cannot expand QN (T, V ) in terms of , but we do know that at low densities all fluids behave like an
ideal gas, and that for an ideal gas the chemical potential is

,
(8.8)
= kB T ln
1/3 (T )
so that as 0, and e 0. So here is a second small parameter. You might object that e
doesnt appear in the partition function any more than does, but it does appear in the grand canonical
partition function!
8.2. PERTURBATION THEORY
8.2.3
173
Fluids in the grand canonical ensemble
The grand canonical partition function is
X
(T, V, ) =
eN Z(T, V, N )
=
=
N =0

X
N =0
e
3 (T )
(8.9)
N
QN (T, V )
z N QN (T, V )
(8.10)
(8.11)
N =0
1 + Q1 (T, V )z + Q2 (T, V )z 2 + Q3 (T, V )z 3 + O(z 4 ),
(8.12)
where we have defined the activity

e
.
(8.13)
3 (T )
The activity is the small parameter upon which perturbation theory relies. (For an ideal gas, it is equal to
the number density .)
z
Note that in this approach we still have to calculate QN coefficients, but we will probably stop at N = 4
or 5 rather than N = 6.02 1023 . I will write down the first two coefficients:
Z
Q1 (T, V ) =
dr1 e0 = V
VZ
Z
1
dr1
dr2 eu(|r1 r2 |) .
(8.14)
Q2 (T, V ) =
2 V
V
Because the interactions are short-ranged, most of the time u(|r1 r2 |) nearly vanishes, so Q2 is pretty close
to V 2 /2. It is neither intensive nor extensive.
8.2.4
The Mayer expansion
Now we go from the partition function to an experimentally measurable master function, namely pressure:
p(T, z)
kB T
=
=
1
ln
V
1
ln[1 + Q1 z + Q2 z 2 + Q3 z 3 + O(z 4 )].
V
(8.15)
(8.16)
Employing
ln(1 + ) = 21 2 + 13 3 + O(4 )
(8.17)
we obtain
p(T, z)
kB T
1
[Q1 z + Q2 z 2 + Q3 z 3 21 (Q1 z + Q2 z 2 )2 + 13 (Q1 z)3 + O(z 4 )]
V
1
1
1
=
Q1 z + (Q2 21 Q21 )z 2 + (Q3 Q1 Q2 + 13 Q31 )z 3 + O(z 4 )
V
V
V
X
`
=
b` z .
`=1
(8.18)
(8.19)
(8.20)
174
The last line above defines the Mayer expansion. The expansion coefficients b` = b` (T ) are intensive
functions of temperature (in contrast to the coefficients QN (T, V )).
Using equations (8.14) it is easy to show that b1 = 1 and that
Z

Z
Z
Z
1
b2 =
dr1
dr2 eu(|r1 r2 |)
dr1
dr2 1
2V
V
V
Z V Z V
1
u(|r1 r2 |)
dr1
dr2 (e
1).
=
2V V
V
(8.21)
In contrast to the situation with equation (8.14), the integrand above is nearly always zero, which makes it
much easier to work with. (The same is true for all the the Mayer coefficients.) Indeed, a little more work
will show that
Z
b2 = 2
dr r2 (eu(r)/kB T 1).
(8.22)
0
Specialists in the statistical mechanics of fluids spend a lot of time calculating b` coefficients for a given
potential energy of interaction u(r). We will instead emphasize what to do with the b` s once they are
calculated.
8.2.5
Expansion for the density
To obtain the number density = N/V from the Mayer expansion, use the thermodynamic connection
between and p:

z
p
(p)
p
=
=z
.
(8.23)
=
T
T z T
z
T
Thus
(T, z) =
`b` z ` .
(8.24)
`=1
8.2.6
Eliminating the activity: the virial expansion
Now we have two expansions, (8.20) and (8.24), in terms of the activity:
p(z)
(z)
= z + b2 z 2 + b3 z 3 + O(z 4 )
=
z + 2b2 z 2 + 3b3 z 3 + O(z 4 ).
(8.25)
(8.26)
(The temperature dependence of p, , and the b` s is not shown explicitly.) What we would really like,
however, would be the traditional (and experimentally accessible) equation of state, p(). In other words,
we want to eliminate the zs between the two expansions above. To this end we will invert the expansion for
(z) to find z(), and plug that expression into the expansion (8.25) for p(z). In the appendix we perform
the inversion to find, correct to third order,
z() = 2b2 2 + (8b22 3b3 )3 + O(4 ).
(8.27)
8.2. PERTURBATION THEORY
175
Plugging into (8.25) gives

p()
= z + b2 z 2 + b3 z 3 + O(z 4 )
2
(8b22
(8.28)
3
2 2
[ 2b2 +
3b3 ) ] + b2 [ 2b2 ] + b3 [] + O( )
(8.29)
[ 2b2 2 + (8b22 3b3 )3 ] + b2 [2 4b2 3 ] + b3 []3 + O(4 )
(8.30)
= b2 +
(4b22
2b3 ) + O( ).
(8.31)
This last expression is called the virial expansion, and it is usually written (with the temperature dependence put back in) as
p(T, )
= + B2 (T )2 + B3 (T )3 + O(4 ).
(8.32)
kB T
176
8.2.7
Appendix: Inverting (z) to find z()
Begin with
(z) = z + 2b2 z 2 + 3b3 z 3 + O(z 4 ),
(8.33)
then use
z
= + O(z 2 )
z2
= ( + O(z 2 ))( + O(z 2 ))

= 2 + 2O(z 2 ) + O(z 4 )
= 2 + O(z 3 )
= 2b2 z 2 + O(z 3 )
= 2b2 2 + O(z 3 )
z2
= ( 2b2 2 + O(z 3 ))( 2b2 2 + O(z 3 ))

= 2 4b2 3 + 4b22 4 + O(z 3 )
= 2 4b2 3 + O(z 4 )
= 2b2 z 2 3b3 z 3 + O(z 4 )
= 2b2 [2 4b2 3 + O(z 4 )] 3b3 [3 + O(z 4 )] + O(z 4 ),
z3
= (2 + O(z 3 ))( + O(z 2 ))

= 3 + O(z 4 )
to find
z() = 2b2 2 + (8b22 3b3 )3 + O(4 ).
(8.34)
Problems
8.4 Virial expansion for the ideal gas
Find Q1 , Q2 , Q3 , etc. for an ideal gas. Hence show that the virial equation of state for an ideal gas is
pV = N kB T .
8.5 General properties of expansion coefficients
a. Show that Q1 = V , whence b1 = 1.
b. Show (informally) that the b` Mayer coefficients, as well as the Bn (T ) virial coefficients, are
intensive.
8.3
Variational Methods
8.6 Bogoliubovs contribution to the Gibbs-Bogoliubov inequality

Show that the Gibbs-Bogoliubov inequality, which we derived in class for classical systems, also holds
for quantal systems.
8.4. DISTRIBUTION FUNCTIONS
8.4
8.4.1
177
Distribution Functions
One-particle distribution functions
What is the mean number of particles in the box of volume d3 rA about rA ?
d3rA
rA
origin
The probability that particle 1 is in d3 rA about rA is

R
R
R
R
R
d3 rA d3 r2 d3 r3 d3 rN d3 p1 d3 pN eH(rA ,r2 ,r3 ,...,rN ,p1 ,...,pN )
R
R
R
R
R
R
.
d3 r1 d3 r2 d3 r3 d3 rN d3 p1 d3 pN eH(r1 ,r2 ,r3 ,...,rN ,p1 ,...,pN )
The probability that particle 2 is in d3 rA about rA is
R 3
R
R
R
R
d r1 d3 rA d3 r3 d3 rN d3 p1 d3 pN eH(r1 ,rA ,r3 ,...,rN ,p1 ,...,pN )
R
R
R
R
R
R
.
(8.35)
(8.36)
And so forth. I could write down N different integrals, but all of them would be equal.
Thus the mean number of particles in d3 rA about rA is
n1 (rA ) d3 rA
R
R
R
R
R
R
R
R
=N R 3 R 3 R 3
d r1 d r2 d r3 d3 rN d3 p1 d3 pN eH(r1 ,r2 ,r3 ,...,rN ,p1 ,...,pN )
R
R
R
R
R
=N
h3N N !Z(T, V, N )
R
R
R
d3 rA d3 r2 d3 r3 d3 rN eU (rA ,r2 ,r3 ,...,rN )
1
.
=
(N 1)!
Q(T, V, N )
(8.37)
(8.38)
(8.39)
178
8.4.2
Two-particle distribution functions
What is the mean number of pairs of particles, such that one member of the pair in a box of volume d3 rA
about rA and the other member is in a box of volume d3 rB about rB ?
d3rB
rB
d3rA
rA
origin
The probability that particle 1 is in d3 rA about rA and particle 2 is in d3 rB about rB is

R 3
R
R
R
d3 r d3 r
d r3 d3 rN d3 p1 d3 pN eH(rA ,rB ,r3 ,...,rN ,p1 ,...,pN )
R AR B R
R
R
R
.
R 3
R
R
R
d3 r d3 r
d r3 d3 rN d3 p1 d3 pN eH(rB ,rA ,r3 ,...,rN ,p1 ,...,pN )
R BR A R
R
R
R
.
R 3
R
R
R
d3 r
d r2 d3 rA d3 rN d3 p1 d3 pN eH(rB ,r2 ,rA ,...,rN ,p1 ,...,pN )
R BR
R
R
R
R
.
(8.40)
(8.41)
(8.42)
And so forth. I could write down N (N 1) different integrals, but all of them would be equal.
Thus the mean number of pairs with one particle in d3 rA about rA and the other in d3 rB about rB is
n2 (rA , rB ) d3 rA d3 rB
R
R
R
R
d3 rA d3 rB d3 r3 d3 rN d3 p1 d3 pN eH(rA ,rB ,r3 ,...,rN ,p1 ,...,pN )
R
R
R
= N (N 1) R 3 R 3 R 3
d r1 d r2 d r3 d3 rN d3 p1 d3 pN eH(r1 ,r2 ,r3 ,...,rN ,p1 ,...,pN )
R
R
d3 rA d3 rB d3 r3 d3 rN eU (rA ,rB ,r3 ,...,rN )
1
.
=
(N 2)!
Q(T, V, N )
(8.43)
(8.44)
8.4. DISTRIBUTION FUNCTIONS
179
Problems
8.7 Correlations of nearby particles
Suppose that (as is usual) at small distances the interatomic potential u(r) is highly repulsive. Argue
that at small r,
g2 (r) constant eu(r)/kB T .
(8.45)
Do not write down a long or elaborate derivation. . . Im looking for a simple qualitative argument.
8.8 Correlations between non-interacting identical quantal particles
Guess the form of the pair correlation function g2 (r) for ideal (non-interacting) fermions and bosons.
Sketch your conjectures, and then compare them to the graphs presented by G. Baym in Lectures on
Quantum Mechanics (W.A. Benjamin, Inc., Reading, Mass., 1969) pages 428 and 431.
8.9 Correlation functions and structure factors
A typical isotropic fluid, at temperatures above the critical temperature, has correlation functions
that are complicated at short distances, but that fall off exponentially at long distances. In fact, the
long-distance behavior is
Aer/
(8.46)
g2 (r) = 1 +
r
where , the so-called correlation length, depends on temperature and density. In contrast, at the
critical temperature the correlation function falls off much more slowly, as
g2 (r) = 1 +
Find the structure factor
Z
S(k) =
A
.
r1+
d3 r [g2 (r) 1]eikr
(8.47)
(8.48)
associated with each of these correlation functions. Will your results match those of experiments at
small values of k or at large values (i.e. at long or short wavelengths)?
8.10 Long wavelength structure factor
dS(k)
Show that, for an isotropic fluid,
vanishes at k = 0. Here S(k) is the structure factor
dk
Z
S(k) = 1 + d3 r [g2 (r) 1]eikr .
(8.49)
8.11 Correlations in a magnetic system

In the Ising model for a magnet, described in problem 4.9, the (net) correlation function is defined by
Gi hs0 si i hs0 i2 ,
(8.50)
where the site j = 0 is some arbitrary central spin. Using the results of problem 4.9, show that for
a lattice of N sites,
m2 X
T (T, H) = N
Gi .
(8.51)
kB T i
180
8.5
Correlations and Scattering
8.6
The Hard Sphere Fluid
Consider a fluid in which each atom is modeled by a hard sphere of volume v0 . In this system the potential
energy vanishes unless two spheres overlap, while if they do overlap it is infinite. The criterion for overlap is
that the centers of the two spheres are separated by a distance of 2r0 or less, where v0 = 34 r03 . This model
is certainly oversimplified in that it ignores the attractive part of the interatomic force, but it is an excellent
starting point for perturbation theory, and it also has certain features of interest in its own right, as the
problems below demonstrate. (The first problem below is required background for all of the others.)
8.12 Scaling in the hard sphere fluid
The canonical partition function for a fluid is
Z(T, V, N ) =
1
3N (T )N !
d3 r1
d3 rN eUN (r1 ,...,rN ) .
(8.52)
For an ideal gas, the integral over configuration space gives just V N , so it makes sense to write
N

Z
Z
1
V /N !
3
3
UN (r1 ,...,rN )
d r1 d rN e
Z(T, V, N ) =
3N (T )
VN
Zideal (T, V, N )W (T, V, N ),
(8.53)
where the last line has defined W , the ratio of interacting to non-interacting (ideal) configurational
partition functions.
a. Argue that in the case of hard spheres W is independent of temperature. It does, however, depend
on the mechanical parameter v0 , so we will write the ratio as W (V, N, v0 ).
b. Argue that W (V, N, v0 ) 1, where the equality holds only when v0 = 0.
c. Show that the Helmholtz free energy for any fluid satisfies
FHS (T, V, N, v0 )
= Fideal (T, V, N ) kB T ln W (V, N, v0 )

Fideal + N fex ,
(8.54)
where we have defined fex , the excess free energy per atom. Is the free energy of a hard sphere
fluid greater than or less than the free energy of an ideal gas at the same T , V , and N ?
d. Because fex is intensive, it cannot depend upon V and N separately, but only through their
intensive quotient, the number density = N/V . Use a similar argument to show that
W (V, N, v0 ) = w(, v0 )N ,
where w is a dimensionless, intensive quantity.
(8.55)
8.6. THE HARD SPHERE FLUID
181
e. Use dimensional analysis to show that the hard sphere w (and hence fex ) cannot depend upon
and v0 separately, but only through their product: w(v0 ).
This last result is of enormous importance. It implies that there is no need to perform elaborate
computer simulations for a variety of densities and a variety of radii. It suffices to either simulate
for a variety of densities at one fixed radius or vice versa. It also demonstrates that the quantity of
importance in the statistical mechanics of hard spheres is the scaled density v0 .
f. Use thermodynamics to show that

fex ()
pHS (T, , v0 ) = pideal (T, ) +
.
=v0
(8.56)
8.13 The hard sphere phase transition

a. Interpret the scaled density v0 geometrically. Why is is sometimes called the packing
fraction?
b. Is it possible to have a fluid with 1?
c. Show that if hard spheres are packed into a face-centered cubic lattice (which is probably the
densest possible packing spheres, although no one has been able to prove it), = 2/6 = 0.7405.
In fact, at scaled densities well below the limits mentioned above, the system undergoes a phase
transition from a fluid to a solid. Computer simulations show that the transition occurs when the
scaled density is t 0.496. The conjecture t = 12 is tempting, but apparently not correct. (On
the other hand it is hard to do computer simulations near phase transitions, so the question is not
completely closed!)
8.14 The hard sphere fluid equation of state
Many long and difficult hours of analytic calculation and computer simulation have gone into studying
the hard sphere fluid. In 1964 Ree and Hoover1 codified much of this work into the following empirical
formula which is a good approximation for the hard sphere fluid equation of state:

1 + 0.25403 + 0.27726 2
pHS (T, , v0 ) = kB T 1 + 4
where = v0 .
(8.57)
1 2.2460 + 1.3010 2
Only five years later, Carnahan and Starling2 hit upon a much simpler yet still remarkably accurate
equation of state, namely
pHS (T, , v0 ) = kB T
1 + + 2 3
(1 )3
where
= v0 .
(8.58)
Both of these formulas apply only to the fluid phase (see the previous problem) and neither one is
exact. Answer the following questions for the Carnahan-Starling formula.
1 F.H.
2 N.F.
Ree and W.G. Hoover, J. Chem. Phys., 40 (1964) 939950.

Carnahan and K.E. Starling, J. Chem. Phys., 51 (1969) 635636.
182

a. As the density increases from zero, at what point does the formula become nonsense?
b. On which regions of the (T, ) plane does the hard sphere fluid exert a greater pressure than the
ideal gas? Can you explain your answer through a simple physical argument?
c. Integrate equation (8.56) to show that for the hard sphere fluid,
fex (T, ) = kB T
4 3
.
(1 )2
(8.59)
Does this formula satisfy the inequality established in part (c.) of problem 8.12?
d. Find the difference between the entropy of the hard sphere fluid and that of an ideal gas. At
a given T , V , and N , which entropy is larger? Can you justify this result with a qualitative
argument?
8.15 The hard sphere free energy
Use the Gibbs-Bogoliubov inequality to show that
FHS (T, V, N ) Fideal (T, V, N ).
(8.60)
Reason that EHS (T, V, N ) = Eideal (T, V, N ), and conclude with a relation between entropies.
8.16 Virial expansion for hard spheres
Using the notation of problem 8.12, show that for a hard sphere fluid:
a. Q2 = 21 V (V 8v0 ).
b. b2 = 4v0 .
c. B2 (T ) = 4v0 .
d. Show that both the Ree-Hover and the Carnahan-Starling equations of state, introduced in problem 8.14, expand to give the correct first virial coefficient B2 (T ).
Chapter 9
Strongly Interacting Systems and

Phase Transitions
9.1
Introduction to Magnetic Systems and Models
9.2
Free Energy of the One-Dimensional Ising Model
The N -spin one-dimensional Ising model consists of a horizontal chain of spins, s1 , s2 , . . . , sN , where si = 1.
6
H
b6
b
?
b
?
b6
b
?
i
b
?
b
6
i+1
b6
b
?
N
A vertical magnetic field H is applied, and only nearest neighbor spins interact, so the Hamiltonian is
HN = J
N
1
X
si si+1 mH
i=1
N
X
si .
For this system the partition function is

X
X X
X K PN 1 s s +L PN s
i i+1
i
i=1
i=1
ZN =
eHN =
e
,
s1 =1 s2 =1
sN =1
states
where
K
J
kB T
(9.1)
i=1
and L
183
mH
.
kB T
(9.2)
(9.3)
184
CHAPTER 9. STRONGLY INTERACTING SYSTEMS AND PHASE TRANSITIONS
If J = 0, (the ideal paramagnet) the partition function factorizes, and the problem is easily solved using the
summing Hamiltonian yields factorizing partition function theorem. If H = 0, the partition function nearly
factorizes, and the problem is not too difficult. (See problem 9.3.) But in general, there is no factorization.
We will solve the problem using induction on the size of the system. If we add one more spin (spin
number N + 1), then the change in the systems energy depends only upon the state of the new spin and of
the previous spin (spin number N ). Define ZN

as, not the sum over all states, but the sum over all states in
which the last (i.e. N th) spin is up, and define ZN

as the sum over all states in which the last spin is down,
so that
ZN = ZN
+ ZN
.
(9.4)
Now, if one more spin is added, the extra term in eH results in a factor of
eKsN sN +1 +LsN +1 .
(9.5)
From this, it is very easy to see that
ZN
+1
ZN
+1
K+L
K+L
ZN
e
+ ZN
e
(9.6)
KL
ZN
e
(9.7)
KL
ZN
e
.
This is really the end of the physics of this derivation. The rest is mathematics.
So put on your mathematical hats and look at the pair of equations above. What do you see? A matrix
equation!
!
!
!
ZN
eK+L eK+L
ZN
+1
=
.
(9.8)
ZN
eKL eKL
ZN
+1
We introduce the notation
wN +1 = TwN
(9.9)
for the matrix equation. The 2 2 matrix T, which acts to add one more spin to the chain, is called the
transfer matrix. Of course, the entire chain can be built by applying T repeatedly to an initial chain of one
site, i.e. that
wN +1 = TN w1 ,
(9.10)
where
w1 =
eL
eL
!
.
(9.11)
The fact that we are raising a matrix to a power suggests that we should diagonalize it. The transfer
matrix T has eigenvalues A and B (labeled so that |A | > |B |) and corresponding eigenvectors xA and
xB . Like any other vector, w1 can be expanded in terms of the eigenvectors
w1 = c A x A + c B x B
(9.12)
9.2. FREE ENERGY OF THE ONE-DIMENSIONAL ISING MODEL
185
and in this form it is very easy to see what happens when w1 is multiplied by T N times:
wN +1 = TN w1
= cA TN xA + cB TN xB
N
= cA N
A xA + cB B xB .
(9.13)
So the partition function is
N
N
ZN +1 = ZN
+1 + ZN +1 = cA A (xA + xA ) + cB B (xB + xB ).
(9.14)
By diagonalizing matrix T (that is, by finding both its eigenvalues and its eigenvectors) we could find
every element in the right hand side of the above equation, and hence we could find the partition function
ZN for any N . But of course we are really interested only in the thermodynamic limit N . Because
N
|A | > |B |, N
A dominates B in the thermodynamic limit, and
ZN +1 cA N
A (xA + xA ),
(9.15)
FN +1 = kB T ln ZN +1 kB T N ln A kB T ln[cA (xA + xA )],
(9.16)
provided that cA (xA + xA ) 6= 0. Now,
and this approximation becomes exact in the thermodynamic limit. Thus the free energy per spin is
f (K, L) = lim
FN +1 (K, L)
= kB T ln A .
N +1
(9.17)
So to find the free energy we only need to find the larger eigenvalue of T: we dont need to find the smaller
eigenvalue, and we dont need to find the eigenvectors!
It is a simple matter to find the eigenvalues of our transfer matrix T. They are the two roots of
!
eK+L
eK+L
det
= 0
(9.18)
eKL
eKL
( eK+L )( eKL ) e2K
(9.19)
2K
0,
(9.20)
2K
2e cosh L + e
which are
i
h
p
= eK cosh L cosh2 L 1 + e4K .
It is clear that both eigenvalues are real, and that the larger one is positive, so
h
i
p
A = eK cosh L + sinh2 L + e4K .
Finally, using equation (9.17), we find the free energy per spin
"
#
r
mH
2 mH
f (T, H) = J kB T ln cosh
+ sinh
+ e4J/kB T .
kB T
kB T
(9.21)
(9.22)
(9.23)
186
Results
Knowing the free energy, we can take derivatives to find any thermodynamic quantity (see problem 9.2).
Here Ill sketch and discuss the results obtained through those derivatives.
The heat capacity at constant, vanishing, magnetic field is sketched here:
CH(T,H=0)
1/T2
T
0
kBT about J
Do you think an experimentalist needs a magnet to probe magnetic phenomena? This graph shows that a
magnet is not required: The magnetic effects result in a bump in the heat capacity near kB T = J, even
when no magnetic field is applied.
9.2. FREE ENERGY OF THE ONE-DIMENSIONAL ISING MODEL
187
The magnetic susceptibility is sketched here:
T(T,H=0)
ferromagnetic J>0
paramagnetic J=0 (1/T)
antiferromagnetic J<0
0
T
0
kBT about J
We have already seen that for independent spins (paramagnet, J = 0) the susceptibility falls like 1/T with
temperature (the Curie law). Interacting spins at high temperature (kB T J) behave approximately
the same way. But at low temperatures, the susceptibility for a ferromagnet exceeds the susceptibility for a
paramagnet, while the susceptibility for a antiferromagnet undershoots the susceptibility for a paramagnet.
This makes sense: For a paramagnet, the external magnetic field is inducing the spins to align. For a
ferromagnet, both the external magnetic field and the tendency of neighboring spins to align are inducing
the spins to align. For an antiferromagnet, the external magnetic field is inducing the spins to align, but the
tendency of neighboring spins to antialign is opposing that inducement.
Aligning the spins in a paramagnet is like herding cats: the individual spins are independent and dont
naturally take to pointing all in the same direction. Aligning the spins in a ferromagnet is like herding cows:
the individual spins want to all go in the same direction and dont care which direction it is. Aligning the
spins in an antiferromagnet is like herding siblings in a dysfunctional family, where each sibling says I want
to go to the opposite of wherever my brother/sister wants to go.
188
9.3
The Mean-Field Approximation
Many times, in this book, I have had occasion to make an approximation, but then I argued (sometimes rigorously and sometimes less rigorously) that this approximation would become exact in the thermodynamic
limit. So let me emphasize that the mean-field approximation is not exact in the thermodynamic limit. It
is usually accurate at high temperatures. It is rarely accurate near phase transitions.
9.4
9.4.1
Correlation Functions in the Ising Model

Motivation and definition
How does a spin at one site influence a spin at another site? This is not a question of thermodynamics,
but its an interesting and useful question in statistical mechanics. The answer is given through correlation
functions.
Consider an Ising model with spins si = 1 on lattice sites i. The figure below shows part of a square
lattice, although the discussion holds for any Bravais lattice in any dimension.
b
6
b
?
site 0
site i
Choose a site at the center of the lattice and call it the origin, site 0. Your choice is quite arbitrary, because
the lattice is infinite so any site is at the center. Now choose another site and call it site i. The product of
these two spins is
(
+1 if the spins point in the same direction
s0 si =
(9.24)
1 if the spins point in opposite directions.
What is the average of this product over all the configurations in the ensemble? In general, this will be
a difficult quantity to find, but in certain special cases we can hold reasonable expectations. For example
if the system is ferromagnetic, then at low temperatures we expect that nearest neighbor spins will mostly
be pointing in the same direction, whereas the opposite holds if the system is antiferromagnetic. Thus we
9.4. CORRELATION FUNCTIONS IN THE ISING MODEL
189
expect that, if site i is adjacent to site 0,

(
+1 for ferromagnetic systems at low temperatures
hs0 si i
1 for antiferromagnetic systems at low temperatures.
(9.25)
Can you think of other expectations? (High temperatures? Next-nearest-neighbors at low temperatures?)
Does the average of the product ever exceed 1 in any circumstances? Let me also point out a frequently-held
expectation that is not correct. One might expect that hs0 si i would vanish for sites very far apart. This
is true under some circumstances, but not if an external magnetic field is applied, nor at zero field in a
low-temperature ferromagnet, because in these situations all the spins in the sample (even those very far
apart) tend to point in the same direction.
We have seen that the average of the product hs0 si i is not trivial. In contrast the product of the averages
hs0 ihsi i is very easy, because all sites are equivalent whence hsi i = hs0 i and
2
hs0 ihsi i = hs0 i .
(9.26)
In fact, hs0 si i would be equal to hs0 i if there were no correlations between spins. . . this is essentially the
definition of correlation. This motivates the definition of the correlation function
2
Gi (T, H) = hs0 si i hs0 i .
(9.27)
The correlation function is essentially a measure of peer pressure: How is the spin at site i influenced by
the state of the spin at site 0? If there is a large magnetic field, for example, and if sites i and 0 are far apart,
then both spins will tend to point up, but this is not because of peer pressure, it is because of external
pressure. (For example, two children in different cities might dress the same way, not because one imitates
2
the other, but because both listen to the same advertisements.) The term hs0 i measures external pressure,
and we subtract it from hs0 si i so that Gi (T, H) will measure only peer pressure and not external pressure.
This definition doesnt say how the influence travels from site 0 to site i: If site 2 is two sites to the
right of site 0, for example, then most of the influence will be due to site 0 influencing site 1 and then site 1
influencing site 2. But a little influence will travel up one site, over two sites, then down one site, and still
smaller amounts will take even more circuitous routes.
Note that
if i = 0 then
if i is far from 0 then
Gi (T, H) = 1 hs0 i
Gi (T, H) = 0.
(9.28)
I have three comments to make concerning the correlation function. First, realize that the correlation
function gives information about the system beyond thermodynamic information. For most of this book
we have been using microscopic information to calculate a partition function and then to find macroscopic
(thermodynamic) information about the system. We did not use the partition function to ask any question
that involved the distance between two spins (or two atoms). The correlation function enables us to probe
more deeply into statistical mechanics and ask such important microscopic questions.
190
Second, the correlation function is a measurable quantity. I have produced the definition through conceptual motivations1 that make it seem impossible to find except by examining individual spins. . . a conceptual
nicety but experimental impossibility. But in fact the correlation function can be found experimentally
through neutron scattering. This technique relies upon interference in the de Broglie waves of neutron waves
scattered from nearby spins. It is not trivial: indeed two of the inventors of the technique were awarded
the Nobel prize in physics for 1994. (It is also worth noting that the work of one of these two was based
upon theoretical work done by Bob Weinstock in his Ph.D. thesis.) I will not have enough time to treat this
interesting topic, but you should realize that what follows in this section is not mere theoretical fluff. . . it
has been subjected to, and passed, the cold harsh probe of experimental test.
Third, although the definition given here applies only to Ising models on Bravais lattices, analogous
definitions can be produced for other magnetic systems, for fluids, and for solids.
9.4.2
Susceptibility from the correlation function
I have emphasized above that the correlation function gives information beyond that given by thermodynamics. But the correlation function contains thermodynamic information as well. In this subsection we will
prove that if the correlation function Gi (T, H) is known for all sites in the lattice, then the susceptibility, a
purely thermodynamic quantity, can be found through
(T, H) = N
m2 X
Gi (T, H),
kB T i
(9.29)
where the sum is over all sites in the lattice (including the origin).
We begin with a few reminders. The microscopic magnetization for a single configuration is
X
M(s1 , . . . , sN ) = m
si
(9.30)
and the macroscopic (thermodynamic) magnetization is its ensemble average

M = hMi = mN hs0 i.
(9.31)
The Ising Hamiltonian breaks into two pieces,

H
= H0 (s1 , . . . , sN ) mH
si
(9.32)
= H0 (s1 , . . . , sN ) HM(s1 , . . . , sN ),
(9.33)
where H0 is independent of magnetic field. In the nearest-neighbor Ising model the field-independent part
of the Hamiltonian is
X
H0 (s1 , . . . , sN ) = J
si sj ,
(9.34)
<i,j>
1 Namely,
by asking Wouldnt it be nice to know how one spin is influenced by its neighbors?
191
but we will not use this formula. . . the results of this section are true for spin-spin interactions of arbitrary
complexity. Finally, the Boltzmann factor is
eH = eH0 +HM .
(9.35)
Now we can start the argument by examining the sum of the correlation function over all sites. In what
P
follows, the notation s means the sum over states.
X
X
2
Gi =
hs0 si i hs0 i
(9.36)
i
i
"
#
P
H 2
X P s0 si eH
s
e
0
s
s
=
Z
Z2
i
2
P
P
s0 eH
s0 (M/m)eH
s
s
=
N
2
Z
P
P Z H P
H
H
Z s s0 (M/m)e
N s s0 e
s s0 e
=
2
Z
(9.37)
(9.38)
(9.39)
Look at this last equation carefully. What do you see? The quotient rule! Do you remember our slick
trick for finding dispersions? The M appearing in the leftmost summand can be gotten there by taking a
derivative of equation (9.35) with respect to H:
eH
= MeH .
H
(9.40)
Watch carefully:
P

H
s s0 e
=
Z
P
P
P
H
s MeH s s0 eH
s s0 Me
Z2
P
P
P

Z s s0 MeH mN s s0 eH s s0 eH
=
Z2
Z
(9.41)
(9.42)
(9.43)
But the quantity in square brackets is nothing more than m times equation (9.39), so
P

H
X
s s0 e
= m
Gi .
H
Z
i
But we also know that
P
s s0 e
Z

=
whence, finally
(T, H) = N
(9.44)
hs0 i
M/mN
1
=
=
H
H
mN
(9.45)
m2 X
Gi (T, H).
kB T i
(9.46)
Whew!
Now we can stop running and do our cool-down stretches. Does the above equation, procured at such
expense, make sense? The left hand side is extensive. The right hand side has a factor of N and also a
192
sum over all sites, and usually a sum over all sites is extensive. So at first glance the left hand side scales
linearly with N while the right hand side scales with N 2 , a sure sign of error! In fact, there is no error. The
correlation function Gi vanishes except for sites very close to the origin. (Distant sites are uncorrelated,
see equation (9.28).) Thus the sum over all sites above has a summand that vanishes for most sites, and it
is intensive rather than extensive.
9.4.3
All of thermodynamics from the susceptibility
The previous subsection shows us how to find one thermodynamic quantity, (T, H) at some particular
values for temperature and field, by knowing the correlation function Gi (T, H) for all sites i, at those same
fixed values for temperature and field. In this subsection we will find how to calculate any thermodynamic
quantity at all, at any value of T and H at all, by knowing (T, H) for all values of T and H. Thus
knowing the correlation Gi (T, H) as a function of all its argumentsi, T , and Henables us to find any
thermodynamic quantity at all.
To find any thermodynamic quantity it suffices to find the free energy F (T, H), so the claim above
reduces to saying that we can find F (T, H) given (T, H). Now we can find (T, H) from F (T, H) by simple
differentiation, but going backwards requires some very careful integration. We will do so by looking at the
difference between quantities for the Ising model and for the ideal paramagnet.
Remember that the thermodynamic state of an Ising magnet is specified by the two variables T and H,
as indicated in the figure.
H
H0
Over much of this diagram, namely the part outside the wavy line, the Ising magnet behaves very much as
an ideal paramagnet does, because the typical thermal energy kB T or the typical magnetic energy m|H|
overwhelms the typical spin-spin interaction energy |J|. Thus outside of the dashed line (e.g. at the ) the
Ising thermodynamic quantities are well-approximated by the paramagnetic thermodynamic quantities

m2
M
mH
=N
sech2
(9.47)
P (T, H) =
H T
kB T
kB T

MP (T, H)
FP (T, H)
mH
= mN tanh2
kB T
T

mH
= kB T N ln 2 cosh
,
kB T
=
F
H
193
(9.48)
(9.49)
and this approximation becomes better and better as T or H .

The difference between the susceptibility for our Ising model and for the ideal paramagnet is

(M MP )
(T, H) P (T, H) =
.
H
T
Integrate this equation along the path indicated in the figure to find

Z H
Z H
0
0
0
0 (M MP )
dH [(T, H ) P (T, H )] =
dH
H 0
H0
H0
T
= (M MP )|T,H (M MP )|T,H0 .
Then take the limit H0 so that M (T, H0 ) MP (T, H0 ) and
Z H
M (T, H) = MP (T, H) +
dH 0 [(T, H 0 ) P (T, H 0 )],
(9.50)
(9.51)
(9.52)
(9.53)
where, remember, MP (T, H) and P (T, H) are known functions.

We dont have to stop here. We can consider the difference
(F FP )
M (T, H) MP (T, H) =
H

(9.54)
T
and then perform exactly the same manipulations that we performed above to find
Z H
F (T, H) = FP (T, H)
dH 0 [M (T, H 0 ) MP (T, H 0 )].
(9.55)
Then we can combine equations (9.55) and (9.53) to obtain

Z H
Z H0
F (T, H) = FP (T, H)
dH 0
dH 00 [(T, H 00 ) P (T, H 00 )].
(9.56)
Using the known results for FP (T, H) and P (T, H), as well as the correlation-susceptibility relation (9.46),
this becomes
!

Z H
Z H0
00
X
m2
mH
2 mH
0
00
00
N
dH
. (9.57)
F (T, H) = kB T N ln 2 cosh
dH
Gi (T, H ) sech
kB T
kB T
kB T
i
Let me stress how remarkable this result is. You would think that to find the thermodynamic energy
E(T, H) of the Ising model you would need to know the spin-spin coupling constant J. But you dont: You
can find E(T, H) without knowing J, indeed without knowing the exact form of the spin-spin interaction
Hamiltonian at all, if you do know the purely geometrical information contained within the correlation
function.2 This is one of the most amazing and useful facts in all of physics.
2 This does not mean that the thermodynamic energy is independent of the spin-spin Hamiltonian, because the Hamiltonian
influences the correlation function.
194
9.4.4
Parallel results for fluids
I cannot resist telling you how the results of this section change when applied to fluid rather than magnetic
systems. In that case the correlation function g2 (r; T, V, N ) depends upon the position r rather than the
site index i. Just as the sum of Gi over all sites is related to the susceptibility , so the integral of g2 (r)
over all space is related to the compressibility T . And just as the susceptibility can be integrated twice
(carefully, through comparison to the ideal paramagnet) to give the magnetic free energy F (T, H), so the
compressibility can be integrated twice (carefully, through comparison to the ideal gas) to give the fluid free
energy F (T, V, N ). And just as the magnetic thermodynamic quantities can be found from the correlation
function without knowing the details of the spin-spin interaction Hamiltonian, so the fluid thermodynamic
quantities can be found from the correlation function without knowing the details of the particle interaction
forces. A question for you: How is it possible to get all this information knowing only the correlation function
in space rather than the correlation function in phase space? In other words, why dont you need information
about the velocities?
I should put in some references to experimental tests and to say who thought this up. Are these available
in Joels review article?
9.5
9.5.1
Computer Simulation
Basic strategy
If were going to use a computer to solve problems in statistical mechanics, there are three basic strategies
that we could take. I illustrate them here by showing how they would be applied to the Ising model.
1. Exhaustive enumeration. Just do it! List all the states (microstates, configurations), their associated
probabilities, and average over them! Admittedly, there are a lot of states, but computers are fast, right?
Lets see how many states there are. Consider a pretty small problem: a three dimensional Ising model
on an 8 8 8 cube. This system has 512 spins that can be oriented either up or down, so there are
2512 10154 configurations. Suppose that our computer could generate and examine a million configurations
per second. . . this is about as fast as is currently possible. Then an exhaustive enumeration would require
10148 seconds to do the job. By contrast, the universe is about 1018 seconds old. So for one computer to
do the problem it would take 10130 times the age of the universe. Now this is impractical, but maybe we
could do it by arranging computers in parallel. Well, there are about 1080 protons in the universe, and if
every one were to turn into a spiffy Unix workstation, it would still require 1050 times the age of the universe
to complete the list. And this is for a small problem! Sorry, computers arent fast, and they never will be
fast enough to solve the complete enumeration problem. It is not for nothing that this technique is called
exhaustive enumeration.
2. Random sampling. Instead of trying to list the entire pool of possible configurations, we will dip into
it at random and sample the configurations. This can be implemented by scanning through the lattice and
9.5. COMPUTER SIMULATION
195
orienting spins up with probability 21 and down with probability 21 . The problem in this strategy is that all
configurations are equally likely to be sampled, but because there are many more high energy configurations
than low energy configurations, it is very unlikely to sample a low energy configuration (the poker paradox
again). But in the Boltzmann distribution eenergy/kB T the low energy configurations are in fact more likely
to occur. So you could sample for a long time before encountering even one configuration with non-negligible
probability. This strategy is considerably better than the first one, but it would still take about the age of
the universe to implement.
3. Importance sampling. This strategy is to sample the pool of possible configurations not completely at
random, but in such a way that the most likely configurations are most likely to be sampled. An enormously
successful algorithm to perform importance sampling was developed by Nick Metropolis and his coworkers
Arianna Rosenbluth, Marshall Rosenbluth, Augusta Teller, and Edward Teller (of hydrogen bomb fame) in
1953. (Equation of state calculations by fast computing machines, J. Chem. Phys. 21 10871092.) It
is called the Metropolis algorithm, or the M(RT)2 algorithm, or simply the Monte Carlo algorithm.
The next section treats this algorithm in detail.
An analogy to political polling is worth making here. Exhaustive enumeration corresponds to an election,
random sampling corresponds to a random poll, and importance sampling corresponds to a poll which selects
respondents according to the likelihood that they will vote.
9.5.2
The Metropolis algorithm
This algorithm builds a chain of configurations, each one modified (usually only slightly) from the one before.
For example, if we number the configurations (say, for the 8 8 8 Ising model, from 1 to 2512 ) such a chain
of configurations might be
C171 C49 C1294 C1294 C171 C190 .
Note that it is possible for two successive configurations in the chain to be identical.
To build any such chain, we need some transition probability rule W (Ca Cb ) giving the probability
that configuration Cb will follow configuration Ca . And to be useful for importance sampling, the rule will
have to build the chain in such a way that the probability of Cn appearing in the chain is eEn /Z.
So we need to produce a result of the form: if rule then Boltzmann probability distribution. It
is, however, very difficult to come up with such results. Instead we go the other way to produce a pool of
plausible transition rules. . . plausible in that they are not obviously inconsistent with a Boltzmann probability
distribution.
Consider a long chain of M configurations, in which the probability of a configuration appearing is given
by the Boltzmann distribution. The chain must be at equilibrium in that
number of transitions(Ca Cb ) = number of transitions(Cb Ca ).
(9.58)
196
But the number of Ca s in the chain is M eEa /Z, and similarly for the number of Cb s, so the equation
above is
eEa
eEb
M
W (Ca Cb ) = M
W (Cb Ca ),
(9.59)
Z
Z
whence
W (Ca Cb )
= e(Eb Ea ) .
(9.60)
W (Cb Ca )
This condition, called detailed balance, defines our pool of plausible transition probability rules. It is possible that a rule could satisfy detailed balance and still not sample according to the Boltzmann distribution,
but any rule that does not satisfy detailed balance certainly cannot sample according to the Boltzmann
distribution. In practice, all the rules that satisfy detailed balance seem to work.
Two transitions probability rules that do satisfy detailed balance, and the two rules most commonly used
in practice, are
E
e
(9.61)
W (Ca Cb ) = [normalization constant]
1 + eE
(
)
1
if E 0
W (Ca Cb ) = [normalization constant]
(9.62)
eE if E > 0
where we have taken
E = Eb Ea
and where the normalization constant is fixed so that
X
W (Ca Cb ) = 1.
(9.63)
(9.64)
Cb
The factor for the transition probability aside from the normalization constant, i.e. the part in curly brackets
in equations (9.61) and (9.62), is called wab . Note that it is a positive number less than or equal to one:
0<{
} wab 1.
(9.65)
Notice that for any rule satisfying detailed balance the transition probability must increase as the change
in energy decreases, suggesting a simple physical interpretation: The chain of configurations is like a walker
stepping from configuration to configuration on a configuration landscape. If a step would decrease the
walkers energy, he is likely to take it, whereas if it would increase his energy, he is likely to reject it. Thus
the walker tends to go downhill on the configuration landscape, but it is not impossible for him to go uphill.
This seems like a recipe for a constantly decreasing energy, but it is not, because more uphill steps than
downhill steps are available to be taken.3
From these ingredients, Metropolis brews his algorithm:
3 This paragraph constitutes the most densely packed observation in this book. It looks backward to the poker paradox,
to the definition of temperature, and to the cash incentives interpretation of the canonical ensemble. It looks forward to
equilibration and to optimization by Monte Carlo simulated annealing. Why not read it again?
197
Generate initial configuration [at random or otherwise]

Gather data concerning configuration [e.g. find M = n n ; Msum :=M ]
DO Iconfig = 2, Nconfigs
Generate candidate configuration [e.g. select a spin to flip]
Compute wab for transition to candidate
With probability wab , make the transition
Gather data concerning configuration [e.g. find M = n n ; Msum :=Msum + M ]
END DO
Summarize and print data [e.g. Mave = Msum /Nconfigs]
Ill make three comments concerning this algorithm. First of all, note that the step With probability
wab , make the transition implies that sometimes the transition is not made, in which case two configurations adjacent in the chain will be identical. It is a common misconception that in this case the repeated
configuration should be counted only once, but thats not correct: you must execute the Gather data
concerning configuration step whether the previous candidate was accepted or rejected.
Secondly, I wish to detail how the step
With probability wab , make the transition
is implemented. It is done by expanding the step into the two substeps
Produce a random number z [0 z < 1]
IF z < wab THEN switch to candidate
[i.e. configuration := candidate configuration]
Finally, I need to point out that this algorithm does not precisely implement either of the transition
probability rules (9.61) or (9.62). For example, if the step Generate candidate configuration is done
by selecting a single spin to flip, then the algorithm will never step from one configuration to a configuration
three spin-flips away, regardless of the value of E. Indeed, under such circumstances (and if there are N
spins in the system), the transition probability rule is
if Ca and Cb differ by a single spin flip
N wab
W (Ca Cb ) =
0
if Ca and Cb differ by more than a single spin flip (9.66)
1 Cb 6=Ca W (Ca Cb ) if Ca = Cb .
It is easy to see that this transition probability rule satisfies detailed balance.
9.5.3
Implementing the Metropolis algorithm
It is not atypical to run a Monte Carlo program for about 10,000 Monte Carlo steps per site. As such,
the program might run for hours or even days. This is considerably less than the age of the universe, but
198
probably far longer than you are used to running programs. The following tips are useful for speeding up or
otherwise improving programs implementing the Metropolis algorithm.
1. Use scaled quantities. The parameters J, m, H, and T do not enter in any possible combination, but
only through two independent products. These are usually taken to be
kB T
T =
J
= mH .
and H
J
Thus the Boltzmann exponent for a given configuration is
X
E
1 X
=
si sj + H
si ,
kB T
T hi,ji
i
(9.67)
(9.68)
where si is +1 if the spin at site i is up, 1 if it is down, and where hi, ji denotes a nearest neighbor pair.
2. Dont find total energies. To calculate wab you must first know E, and the obvious way to find
E is to find Eb and Ea (through equation (9.68)) and subtract. This way is obvious but terribly inefficient.
Because the change in configuration is small (usually a single spin flip) the change in energy can be found from
purely local considerations without finding the total energy of the entire system being simulated. Similarly,
if you are finding the average (scaled) magnetization
X
M = n n =
si
(9.69)
i
(as suggested by the square brackets in the algorithm on page 197) you dont need to scan the entire lattice
to find it. Instead, just realize that it changes by 2 with each spin flip.
3. Precompute Boltzmann factors. It is computationally expensive to find evaluate a exponential, yet we
must know the value of eE . However, usually there are only a few possible values of E. (For example
in the square lattice Ising model with nearest neighbor interactions and a field, a single spin flip gives rise to
one of only ten possible values of the energy change.) It saves considerable computational time (and often
makes the program clearer) to precalculate the corresponding values of wab just once at the beginning of
the program, and to store those values in an array for ready reference when they are needed.
4. Average on the fly. The algorithm on page 197 finds the average (scaled) magnetization by summing
the magnetization of each configuration in the chain and then dividing by the number of configurations.
Because the chain is so long this raises the very real possibility of overflow in the value of Msum . It is often
better to keep a running tally of the average by tracking the average so far through
Mave :=Mave (Iconfig 1)/Iconfig + M/Iconfig
(9.70)
or (identical mathematically, but preferable for numerical work)

Mave :=Mave + (M Mave )/Iconfig.
5. Finite size effects. Use periodic or skew-periodic boundary conditions.
(9.71)
199
6. Lattice data structures. Suppose we wish to simulate a two-dimensional Ising model on a 4 3 square
grid. An obvious data structure to hold the configuration in the computers memory is an integer-valued
two-dimensional array declared through the code
INTEGER, PARAMETER :: Nx = 4, Ny = 3
INTEGER :: Spin (1:Nx, 1:Ny)
(I use the notation of Fortran 90. If you are familiar with some other computer language, the intent should
nevertheless be clear.) If the spin at site (3,2) is up, then Spin(3,2) = +1, and if that spin is down, then
Spin(3,2) = 1. The lattice sites are labeled as shown in this figure:
a
(1,3)
a
(2,3)
a
(3,3)
a
(4,3)
a
(1,2)
a
(2,2)
a
(3,2)
a
(4,2)
a
(1,1)
a
(2,1)
a
(3,1)
a
(4,1)
Although this representation is obvious, it suffers from a number of defects. First, it is difficult to
generalize to other lattices, such as the triangular lattice in two dimensions or the face-centered cubic lattice
in three dimensions. Second, finding the nearest neighbors of boundary sites using periodic or skew-periodic
boundary conditions is complicated. And finally because the array is two-dimensional, any reference to an
array element involves a multiplication,4 which slows down the finding of data considerably.
All of these defects are absent in the folded array representation of the lattice sites. In this representation
the sites are stored as an integer-valued one-dimensional array declared through
INTEGER, PARAMETER :: Nx = 4, Ny = 3, NSites = Nx*Ny
INTEGER :: Spin (0:NSites-1)
Now the lattice sites are labeled as:
4 In
Fortran, the datum Site(i,j) is stored at memory location number i + j*(Nx-1), whence the data are stored in the
sequence (1,1), (2,1), (3,1), (4,1), (1,2), (2,2),. . . . In other computer languages the formula for finding the memory location is
different, but in all languages it involves an integer multiply.
200
a
8
a
9
a
10
a
11
a
4
a
5
a
6
a
7
a
0
a
1
a
2
a
3
It is not hard to show that, with skew-periodic boundary conditions, the four neighbors of site number l (l
for location) are
(l + 1) mod NSites, (l 1) mod NSites, (l + Nx) mod NSites, (l Nx) mod NSites.
Unfortunately, the Fortran Mod function was inanely chosen to differ from the mathematical mod function,
so this arithmetic must be implemented through
Mod(l
Mod(l
Mod(l
Mod(l
+
+
+
+
1, NSites)
(NSites - 1), NSites)
Nx, NSites)
(NSites - Nx), NSites)
7. Random number generators. It is hard to find an easy source of high-quality random numbers: this is
a problem for physics, also for government and industry.
D.E. Knuth, Seminumerical Algorithms (AddisonWesley, Reading, Massachusetts, 1981) chapter
3.
W.H. Press, S.A. Teukolsky, W.T. Vettering, and B.P. Flannery, Numerical Recipes chapter
7. (For a confession and a prize announcement, see also W.H. Press and S.A. Teukolsky,
Portable random number generators, Computers in Physics 6 (1992) 522524.)
T.-W. Chiu and T.-S. Guu, A shift-register sequence random number generator, Computer
Physics Communications 47 (1987) 129137. (See particularly figures 1 and 2.)
A.M. Ferrenberg, D.P. Landau, and Y.J. Wong, Monte Carlo simulations: Hidden errors from
good random number generators, Phys. Rev. Lett. 69 (1992) 3382.
S.K. Park and K.W. Miller, Random number generators: Good ones are hard to find, Communications of the ACM 31 (1988) 11921201.
8. Initial configurations and equilibration. The Metropolis algorithm does not specify how to come up
with the first configuration in the chain. Indeed, selecting this configuration is something of an art. If you are
201
simulating at high temperatures, it is usually appropriate to begin with a configuration chosen at random,
so that about half the spins will be up and half down. But if you are simulating at low temperatures (or
at high fields) it might be better to start at the configuration with all spins up. Other possibilities are also
possible. But however you select the initial configuration, it is highly unlikely that the one you pick will be
typical of the configurations for the temperature and magnetic field at which you are simulating.
9. Low temperatures. At low temperatures, most candidate transitions are rejected. BLK algorithm.
(A.B. Bortz, J.L. Lebowitz, and M.H. Kalos, A new algorithm for Monte Carlo simulation of Ising spin
systems, J. Comput. Phys. 17 (1975) 1018.)
10. Critical temperatures. Critical slowing down, response is cluster flipping. (U. Wolff, Collective
Monte Carlo updating for spin systems, Phys. Rev. Lett. 62 (1989) 361364. See also Jian-Sheng Wang
and R.H. Swendsen, Cluster Monte Carlo algorithms, Physica A 167 (1990) 565579.)
11. First-order transitions.
12. Molecular dynamics.
9.5.4
The Wolff algorithm
In 1989 Ulli Wolff proposed an algorithm for Monte Carlo simulation that is particularly effective near
critical points. (U. Wolff, Phys. Rev. Lett., 62 (1989) 361364.) The next page presents the Wolff algorithm
as applied to the zero-field ferromagnetic nearest-neighbor Ising model,
X
H = J
si sj with si = 1,
(9.72)
on a square lattice or a simple cubic lattice.
202

Generate initial configuration
Gather data concerning this configuration
DO Iconfig = 2, Nconfigs
Select site j at random
Flip spin at j
Put j into cluster
NFlippedThisGeneration = 1
DO
IF (NFlippedThisGeneration = 0) EXIT
NFlippedPreviousGeneration = NFlippedThisGeneration
NFlippedThisGeneration = 0
FOR each previous generation member i of the cluster DO
FOR each neighbor j of i DO
IF (spin at j 6= spin at i) THEN
with probability P = 1 exp(2J/kB T )
Flip spin at j
Put j into cluster
NFlippedThisGeneration = NFlippedThisGeneration + 1
END IF
END DO
END DO
END DO
Gather data concerning this configuration
Empty the cluster
END DO
Summarize and print out results
203
Template for example
+
+
configuration a
+ + +
+ + + +
+ +
+ +
+ + + +
+
+ + + +
+ +
wab = N8 PJ7 (1 P J )5
+
+
+
+
configuration b
+ + +
+

+
+ +
+
+ + + +
+ +
wba = N8 PJ7 (1 P J )11
+
+
Demonstration of detailed balance. In these formulas,

PJ = 1 e2J/kB T .
(9.73)
The integer 8 is the number of sites in the cluster, 7 is the number of internal bonds in the skeleton, 5 is the
number of external bonds leading to a + spin, and 11 is the number of external bonds leading to a spin.
The energy difference between configurations a and b depends only upon these last two integers. It is
E = +2J(5) 2J(11) = 2J(11 5).
(9.74)
Thus
wab
= (1 P J )511 = e2J(511)/kB T = eE/kB T
wba
and detailed balance is insured.
9.6
(9.75)
Additional Problems
9.1 Denaturizaton of DNA

(This problem is modified from one posed by M.E. Fisher. It deals with a research question born in
the 1960s that is still of interest today: Douglas Poland and Harold A. Scheraga, Occurrence of a
phase transition in nucleic acid models, J. Chem. Phys. 45 (1966) 14641469; M.E. Fisher, Effect
of excluded volume on phase transitions in biopolymers, J. Chem. Phys. 45 (1966) 14691473; Yariv
Kafri, David Mukamel, and Luca Peliti, Why is the DNA denaturation transition first order? Phys.
Rev. Lett. 85 (2000) 49884991.)
The biological polymer DNA consists of two polymeric strands which, in their usual state, are twisted
together to form a rather rigid double helix. However as the temperature is raised regions of the
molecule may melt, breaking the weak bonds that hold the two strands together. In this case the
two strands separate and coil randomly in the solution. If the entire molecule melts then the DNA is
said to be denatured. A simple model to aid in understanding this phenomenon is the following.
204

Successive segments of the molecule are regarded as being in either the helical state or in the melted
state. A segment in the helical state has a lower energy than one in the melted state ( vs. 0) but
a segment in the melted state has q different possible orientations (due to coiling) whereas a segment
in the helical state has only one. However, the first or last segment of a run of melted segments has
a smaller number of configurations, rq, where r can be as small as 103 . (A melted segment that is
both the first and last segment of a run is doubly restricted and has only r2 q configurations.) Sample
DNA configurations and their associated Boltzmann factors are shown below.
DNA configuration
"Boltzmann factor"
e11r2q7
e8r3q10
e17r2q
a. Introduce a variable hi for the ith segment such that

(
1 if the ith segment is helical
hi =
0 if the ith segment is melted.
(9.76)
Show that the canonical partition function for a molecule of N segments is

ZN (T, , r) =
1
1
X
X
h1 =0 h2 =0
N
1 Y
X
ehi q (1hi ) r(hi hi1 ) .
(9.77)
hN =0 i=1
b. If F (T, , r) is the Helmholtz free energy, show that the mean fraction of helical segments is
(T ) =
1 F
N
(9.78)
while the mean number of junctions between melted and helical runs is
J(T ) =
r F
.
kB T r
(9.79)
h
m
c. By considering partial partition functions, ZN
and ZN
, for a molecule of N segments in which
the last segment is in a helical or melted state, respectively, construct recursion relations from
which the total partition function, ZN , can be found.
205
d. Find the eigenvalues of the corresponding 2 2 matrix and conclude that the free energy per
segment of a very long DNA molecule is
h
i
p
f (T, , r) = kB T ln 21 1 + w + (1 w)2 + 4wr2 ,
(9.80)
where w = qe . Does this result take on the correct value at zero temperature?
e. In the limit that r vanishes show that the molecule undergoes a sharp melting transition at
temperature

.
(9.81)
Tm =
kB ln q
Sketch the energy, entropy, and helical fraction (T ) as a function of temperature and discuss the
character of the transition.
f. If r is small but nonzero, sketch the behavior you expect for (T ).
g. (Optional.) Show that when r is small but nonzero the sharp transition is smeared out over a
temperature range of approximate width
T
4r
.
kB ln2 q
(9.82)
h. (Optional.) Show that the number of junctions J(T ) defined in (b.) varies as Ar both at T = Tm
and at low temperatures. Find the values of A and in each case.
9.2 Thermodynamics of the one-dimensional Ising model
The Helmholtz free energy of the one-dimensional, nearest-neighbor Ising magnet is
io
n
h
p
F (T, H) = N J + kB T ln cosh L + sinh2 L + e4K ,
where
L=
mH
kB T
and K =
J
kB T
(9.83)
(9.84)
Use this result to show that

= JN tanh K,
K2
CH (T, H = 0) = kB N
,
cosh2 K
sinh L
M (T, H) = mN p
,
sinh2 L + e4K
m2 2K
e .
T (T, H = 0) = N
kB T
E(T, H = 0)
(9.85)
(9.86)
(9.87)
(9.88)
9.3 Zero-field one-dimensional Ising model

In class we found the free energy of a one-dimensional, nearest-neighbor Ising magnet with arbitrary
magnetic field using the transfer matrix method. This problem finds the free energy in a much simpler
way, which unfortunately works only in zero magnetic field.
206

a. Show that the partition function can be written as
X X
X
X
ZN =
eKs1 s2
eKs2 s3
s1 =1 s2 =1
s3 =1
eKsN 2 sN 1
sN 1 =1
eKsN 1 sN .
(9.89)
sN =1
b. Show that the last sum is

X
eKsN 1 sN = 2 cosh K,
(9.90)
sN =1
regardless of the value of sN 1 .

c. Conclude that
ZN = 2N coshN 1 K
(9.91)
and that the zero-field Helmholtz free energy per site is

J
f (T ) = kB T ln 2 cosh
.
kB T
(9.92)
d. Find the zero-field heat capacity, and show that it gives the expected behavior at very low temperatures.
9.4 Magnetization near the critical point: mean field approximation
Analyze the Ising model in the mean field approximation to show that the zero-field magnetization
near the critical temperature is

1/2
mN a Tc T T
for T < Tc
M (T ) =
(9.93)
c
0
for T > T .
c
2 5
x + .)
Be sure to specify the value of the numerical constant a. (Hint: tanh x = x 31 x3 + 15
(Remark: Experiment shows that the magnetization is not precisely of the character predicted by
mean field theory: While it does approach zero like (Tc T ) , the exponent is not 1/2 for
two-dimensional systems = 1/8, while for three-dimensional systems 0.326.)
9.5 Correlation functions for paramagnets

In a paramagnet (see problem 2.9) the spins are uncorrelated, i.e. the correlation function is
Gi (T, H) = 0
i 6= 0.
(9.94)
m2 X
Gi ,
kB T i
(9.95)
for
Use this fact to verify the correlation sum rule

T (T, H) = N
derived in problem 8.11.
9.6 Lattice gasIsing magnet
In class we showed that for a lattice gas
(T, ) =
1
1
X
X
n1 =0 n2 =0

N
1
P
P
X
v0
ni +
ni nj
i
nn
e
,
3 (T )
n =0
N
(9.96)
207
whereas for an Ising magnet

Z(T, H) =
s1 =1 s2 =1
emH
P
i
si + J
P
nn
si sj
(9.97)
sN =1
and we concluded that the lattice gas and Ising magnet were more-or-less equivalent, with chemical
potential and magnetic field playing sort of analogous roles. These conclusions are indeed correct, but
this problem renders them precise.
a. Show that an up spin is associated with an occupied site, and a down spin with a vacant site,
through
ni = 21 (si + 1).
(9.98)
b. Write out the sum for (T, ) in terms of the variables si , and show that the correspondence

4J
(9.99)
2mH kB T ln(v0 /3 (T )) 4qJ
(9.100)
(where q is the number of nearest neighbors for each site on the lattice) leads to
p(T, )v0 = mH 23 qJ f (T, H),
(9.101)
where p(T, ) is the pressure of the gas and f (T, H) is the free energy per spin of the magnet.
c. Interpret equation (9.100) using the phrase escaping tendency. (Is this a sensible question?)
d. (Hard.) Take derivatives of equation (9.101) to show that
(T, )v0
2
4 T (T, )v0
Sgas kB N [ln(v0 /3 (T )) + 32 ]
CV
3
2 kB N
1
2 [1
+ M (T, H)/mN ],
2
(9.102)
= (T, H)/m N ,
(9.103)
= Smag ,
(9.104)
= CM .
(9.105)
9.7 Estimates concerning exhaustive enumeration

Consider an Ising model with 6.02 1023 sites. How many distinct configurations can this system take
on? Estimate the mass of paper and ink required to write out a single character. What mass would
be required to write out this number of configurations in decimal notation? Compare your estimate to
the mass of the earth.
9.8 Fluctuations in Monte Carlo simulation
When the mean energy of a system is determined through Monte Carlo simulation, the result will
necessarily have an uncertainty due to the statistical character of the simulation. (The same is true
of any other measured quantity.) So if mean energy is plotted as a function of temperature, the data
points will not fall on a sharp line but instead within a broad band. Using a fluctuation-susceptibility
relation, show that the width of the band depends on its slope. Will steep bands be wide or narrow?
208
9.9 The folded array representation

Implement the folded array representation of a triangular lattice subject to skew-periodic boundary
conditions. Clue: The piece of lattice under study will be a parallelogram.
9.10 Monte Carlo project
Perform a Monte Carlo simulation on the Ising model. The precise form and direction of the simulation
is up to you: here are some suggestions that you are free to use, combine, extend, or ignore.
a. Draw movies showing the spin configuration of a two-dimensional Ising model as the simulation
proceeds. Discuss the qualitative character of the model at high temperatures, low temperatures,
and near the critical temperature (which is known, from Onsagers exact solution, to be kB Tc /J =
2/ ln( 2 1)).
b. Simulate the one-dimensional Ising model and compare your results to the analytic solution obtained in class. Consider using the BKL algorithm5 at low temperatures.
c. Simulate the DNA denaturization model of problem 9.1, and compare your results to the analytic
results of that problem. You will need to modify the Metropolis algorithm.
d. Simulate a two- or three-dimensional Ising ferromagnet at zero field, using an initial configuration
of all spins up. Monitor the magnetization for a number of temperatures ranging from kB T 5J
down to kB T = 0. Record both the equilibrium magnetization and the number of steps required
to reach equilibrium, and plot these as a function of temperature. Can you locate the critical
temperature?
e. Monitor both the mean energy (or magnetization) and the fluctuations in energy (or magnetization). Use a fluctuation-susceptibility relation to find the heat capacity (or susceptibility). In
particular, show how a simulation at zero magnetic field can still give information about the
magnetic susceptibility.
f. Verify the relation discussed in problem 9.8.
g. Implement the Wolff algorithm6 for improved simulation near the critical point.
h. Use finite size scaling7 to obtain accurate results near the critical point.
i. Compare the BKL and Wolff algorithms at low temperature.
j. Find the correlation function
Gi = hs0 si i hs0 i2
(9.106)
for a few values of T and H. How does it behave as the critical point is approached?
k. (Very ambitious.) Verify the correlation-susceptibility relation (9.46).
5 A.B.
Bortz, J.L. Lebowitz, and M.H. Kalos, J. Comput. Phys., 17 (1975) 1018.
Wolff, Phys. Rev. Lett. 62 (1989) 361364. J.-S. Wang and R.H. Swendsen, Physica A 167 (1990) 565579.
7 H. Gould and J. Tobochnik, An Introduction to Computer Simulation Methods: Applications to Physical Systems, part I
(Addison-Wesley, Reading, Mass., 1988) sections 12.4 and 16.5
6 U.
Appendix A
Series and Integrals

1
1x
1 + x + x2 + x3 + x4 +
for |x| < 1
(the geometric series)
1 2
1
1
x + x3 + x4 +
2!
3!
4!
1
1
1
ln(1 + x) = x x2 + x3 x4 +
for |x| < 1
3
4
r 2
Z +
2
b2 /4a
e
for a > 0 (the Gaussian integral)
eax +bx dx =
a
ex
1+x+
209
(A.1)
(A.2)
(A.3)
(A.4)
Appendix B
Evaluating the Gaussian Integral

The integral
ex dx,
(B.1)
called the Gaussian integral, does not fall to any of the methods of attack that you learned in elementary
calculus. But it can be evaluated quite simply using the following trick.
Define the value of the integral to be A. Then
Z +
Z +
Z
2
x2
y 2
A =
e
dx
e
dy =
dx dy e(x
+y 2 )
(B.2)
At the last step we have written A2 as a two-variable integral over the entire plane. This seems perverse,
because most of the times we work hard to reduce two-dimensional integrals to one-dimensional integrals,
whereas here we are going in reverse. But look at the integrand again. When regarded as an integral on the
plane, it is clear that we can regard x2 + y 2 as just r2 , and this suggests we should convert the integral from
Cartesian (x, y) to polar (r, ) coordinates:
Z Z 2
Z
2
2
r 2
A =
dr
r d e
= 2
rer dr.
(B.3)
0
The last integral immediately suggests the substitution u = r2 , giving

Z

A2 =
eu du = eu 0 = .
(B.4)
We conclude that
ex dx =
(B.5)
B.1 (I) Problem: Another integral

Show that
(Clue: Use the substitution y =
Z
0
ex
dx = .
x
x.)
210
(B.6)
Appendix C
Clinic on the Gamma Function

The gamma function (s) is defined, for s > 0, by
Z
(s) =
xs1 ex dx.
(C.1)
Upon seeing any integral, your first thought is to evaluate it. Stay calm. . . first make sure that the integral
exists. A quick check shows that the integral above converges when s > 0.
There is no simple formula for the gamma function for arbitrary s. But for s = 1,

Z

x
x
(1) =
e dx = e = 1.
0
(C.2)
For s > 1 we may integrate by parts,

Z
Z

s1 x
s1 x
x e dx = x e + (s 1)
0
xs2 ex dx,
(C.3)
giving
(s) = (s 1)(s 1)
for s > 1.
(C.4)
Apply equation (C.4) repeatedly for n a positive integer,

(n) = (n 1)(n 1) = (n 1)(n 2)(n 2) = (n 1)(n 2) 2 1 (1) = (n 1)!,
(C.5)
to find a relation between the gamma function and the factorial function. Thus the gamma function generalizes the factorial function to non-integer values, and can be used to define the factorial function through
x! = (x + 1)
for any x > 1.
(C.6)
In particular,
0! = (1) = 1.
211
(C.7)
212
APPENDIX C. CLINIC ON THE GAMMA FUNCTION
(It is a deep and non-obvious result that the gamma function is in fact the simplest generalization of the
factorial function.)
The gamma function can be simplified for half-integral arguments. For example
Z
Z
Z
2
2
ey dy =
y 1 ey (2y dy) =
( 21 ) =
x1/2 ex dx =
where we used the substitution y =
(C.8)
x. Thus
= ( 21 )!,
2
( 52 ) = 32 ( 32 ) = 34 ,
( 32 )
and so forth.
1
1
2 ( 2 )
(C.9)
(C.10)
Appendix D
Volume of a Sphere in d Dimensions

I will call the volume of a d-dimensional sphere, as a function of radius, Vd (r). You know, of course, that
V2 (r) = r2
(D.1)
(two-dimensional volume is commonly called area) and that

V3 (r) = 43 r3 .
(D.2)
But what is the formula for arbitrary d? There are a number of ways to find it. I will use induction on
dimensionality d. That is, I will use the formula for d = 2 to find the formula for d = 3, the formula for
d = 3 to find the formula for d = 4, and in general use the formula for d to find the formula for d + 1.
This is not the most rigorous formal method to derive the formula, but it is very appealing and has much
to recommend it.
To illustrate the process, I will begin with a well-known and easily visualized stage, namely deriving
V3 (r) from V2 (r). Think of a 3-dimensional sphere (of radius r) as a stack of pancakes of various radii, but
each with infinitesimal thickness dz. The pancake on the very bottom of the stack (z = r) has zero radius.
The one above it is slightly broader. They get broader and broader until we get to the middle of the stack
(z = 0), where the pancake has radius r. The pancakes stacked still higher become smaller and smaller, until
they vanish again at the top of the stack (z = +r). Because the equation for the sphere is
x2 + y 2 + z 2 = r 2 ,
the radius of the pancake at height z0 is
This whole process shows that

Z
r2 z02 .
+r
V3 (r) =
p
dz V2 ( r2 z 2 ).
213
(D.3)
(D.4)
(D.5)
214
APPENDIX D. VOLUME OF A SPHERE IN D DIMENSIONS
It is easy to check this integral against the known result for V3 (r):
Z +r
V3 (r) =
dz (r2 z 2 )
(D.6)
+r

= r2 z 31 z 3 r

= 2r3 32 r3
3
4
3 r .
(D.7)
(D.8)
(D.9)
So we havent gone wrong yet.

Now, how to derive V4 (r) from V3 (r)? This requires a more vivid imagination. Last time we started
with a two-dimensional disk of radius r0 in (x, y) space and thickened it a bit into the third dimension (z)
to form a pancake of three-dimensional volume dz V2 (r0 ). Stacking an infinite number of such pancakes in
the z direction, from z = r to z = +r, gave us a three-dimensional sphere. Now we begin with a threedimensional sphere of radius r0 in (w, x, y) space and thicken it a bit into the fourth dimension (z) to form
a thin four-dimensional pancake of four-dimensional volume dz V3 (r0 ). Stacking an infinite number of such
pancakes in the z direction, from z = r to z = +r, gives a four-dimensional sphere. Because the equation
for the four-sphere is
w 2 + x2 + y 2 + z 2 = r 2 ,
(D.10)
the radius of the three-dimensional sphere at height z0 is
q
r2 z02 ,
(D.11)
and the volume of the four-sphere is

Z
+r
V4 (r) =
p
dz V3 ( r2 z 2 ).
(D.12)
In general, the volume of a (d + 1)-sphere is

Z
+r
Vd+1 (r) =
p
dz Vd ( r2 z 2 ).
(D.13)
If we guess that the formula for Vd (r) takes the form

Vd (r) = Cd rd
(D.14)
(which is certainly true for two and three dimensions, and which is reasonable from dimensional analysis),
then
Z +r
Vd+1 (r) =
dz Cd (r2 z 2 )d/2
(D.15)
r
Z +1
=
r du Cd (r2 r2 u2 )d/2
(D.16)
1
Z +1
= rd+1 Cd
du (1 u2 )d/2 .
(D.17)
1
215
This proves our guess and gives us a recursive formula for Cd :
Z +1
Cd+1 = Cd
du (1 u2 )d/2 .
(D.18)
The problem below shows how to build this recursive chain up from C2 = to
Cd =
d/2
d/2
=
.
(d/2)!
( d2 + 1)
(D.19)
Thus the volume of a d-dimensional sphere of radius r is

Vd (r) =
d/2 d
r .
(d/2)!
(D.20)
D.1 (I) Problem: Volume of a d-dimensional sphere

Before attempting this problem, you should read the material concerning beta functions in an applied
mathematics textbook, such as George Arfkens Mathematical Methods for Physicists or Mary Boass
Mathematical Methods in the Physical Sciences.
a. Show that
+1
(1 u2 )d/2 du = B( 21 , d2 + 1).
b. Use
B(p, q) =
and C2 = to conclude that
Cd =
(p)(q)
(p + q)
d/2
.
( d2 + 1)
D.2 (I) Problem: Volume of a d-dimensional ellipse

Show that the volume of the d-dimensional ellipse described by the equation
2 2 2
2
x1
x2
x3
xd
+
+
+ +
=1
a1
a2
a3
ad
is
Vd (r) =
d/2
a1 a2 a3 ad .
(d/2)!
(D.21)
(D.22)
(D.23)
(D.24)
(D.25)
Appendix E
Stirlings Approximation
The Stirling formula is an approximation for n! that is good at large values of n.
n!
ln(n!)
1 2 3 (n 1) n
(E.1)
= |{z}
ln 1 + ln 2 + ln 3 + + ln(n 1) + ln(n)
0
(E.2)
ln x
x
1
Note that the function ln x is nearly flat for large values of x. For example, ln 1023 is about equal to 23.
From the figure
Z
ln(6!) = area under the staircase >
ln x dx,
(E.3)
and in general
Z
ln(n!) >

n
ln x dx = x ln x x = n ln n n + 1.
216
(E.4)
217
For large values of n, where the ln n function is nearly flat, the two expressions above become quite close.
Also, the 1 becomes negligible. We conclude that
ln(n!) n ln n n
for n 1.
(E.5)
This is Stirlings formula. For corrections to the formula, see M. Boas, Mathematical Methods in the Physical
Sciences, sections 9-10 and 9-11.
You know that
An
increases rapidly with n for positive A, but
n!
(E.6)
n n
e
(E.7)
increases a bit more rapidly still.

E.1 Problem: An upper bound for the factorial function
Stirlings approximation gives a rigorous lower bound for n!.
a. Use the general ideas presented in the derivation of that lower bound to show that
Z n
ln(x + 1) dx > ln n!.
(E.8)
b. Conclude that
(n + 1) ln(n + 1) n + 1 2 ln 2 > ln n! > n ln n n + 1.
(E.9)
Appendix F
The Euler-MacLaurin Formula and

Asymptotic Series
You know that a sum can be approximated by an integral. How accurate is that approximation? The
Euler-MacLaurin formula gives the corrections.
n1
X
Z
f (k)
k=0
f (x) dx
0
1
1
1 000
1
[f (n) f (0)] + [f 0 (n) f 0 (0)]
[f (n) f 000 (0)] +
[f (v) (n) f (v) (0)] + .
2
12
720
30240
This series is asymptotic. If the series is truncated at any point, it can give a highly accurate approximation. But the series may be either convergent or divergent, so adding additional terms to the truncated
series might give rise to a poorer approximation. The Stirling approximation is a truncation of an asymptotic
series.
References
C.M. Bender and S.A. Orszag, Advanced Mathematical Methods for Scientists and Engineers, (McGraw-Hill,
New York, 1978).
M. Abramowitz and I. Stegum, Handbook of Mathematical Functions, (National Bureau of Standards, Washington, D.C., 1964).
218
Appendix G
Ramblings on the Riemann Zeta

Function
The function
(s) =
X
1
ns
n=1
for s > 1,
called the Riemann zeta function, has applications to number theory, statistical mechanics, and quantal
chaos.
History
From Simmons:
No great mind of the past has exerted a deeper influence on the mathematics of the twentieth
century than Bernhard Riemann (18261866), the son of a poor country minister in northern
Germany. He studied the works of Euler and Legendre while he was still in secondary school,
and it is said that he mastered Legendres treatise on the theory of numbers in less than a week.
But he was shy and modest, with little awareness of his own extraordinary abilities, so at the
age of nineteen he went to the University of Gottingen with the aim of pleasing his father by
studying theology and becoming a minister himself.
Instead he became a mathematician. He made contributions to the theory of complex variables (CauchyRiemann equations, Riemann sheet, Riemann sphere), to Fourier series, to analysis (Riemann integral), to hypergeometric functions, to convergence of infinite sums (Riemann rearrangement theorem),
to geometry (Riemannian curved space), and to classical mechanics, the theory of fields and other areas
of physics. In his thirties he became interested in the prime number theorem.
219
220
APPENDIX G. RAMBLINGS ON THE RIEMANN ZETA FUNCTION

The first few prime numbers are 2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43,. . . . It is clear
that the primes are distributed among all the positive integers in a rather irregular way; for as
we move out, they seem to occur less and less frequently, and yet there are many adjoining pairs
separated by a single even number [twin primes]. . . . Many attempts have been made to find
simple formulas for the nth prime and for the exact number of primes among the first n positive
integers. All such efforts have failed, and real progress was achieved only when mathematicians
started instead to look for information about the average distribution of the primes among the
positive integers. It is customary to denote by (x) the number of primes less than or equal to
a positive number x. Thus (1) = 0, (2) = 1, (3) = 2, () = 2, (4) = 2, and so on. In
his early youth Gauss studied (x) empirically, with the aim of finding a simple function that
seems to approximate it with a small relative error for large x. On the basis of his observations
he conjectured (perhaps at the age of fourteen or fifteen) that x/ log x is a good approximating
function, in the sense that
(x)
lim
= 1.
x x/ log x
This statement is the famous prime number theorem; and as far as anyone knows, Gauss was
never able to support his guess with even a fragment of proof.
Chebyshev, unaware of Gausss conjecture, was the first mathematician to establish any firm
conclusions about this question. In 1848 and 1850 he proved that
0.9213 . . . <
(x)
< 1.1055 . . .
x/ log x
for all sufficiently large x, and also that if the limit exists, then its value must be 1. (The numbers
in the inequality are A = log 21/2 31/3 51/5 301/30 on the left, and 65 A on the right.) . . .
In 1859 Riemann published his only work on the theory of numbers, a brief but exceedingly
profound paper of less than 10 pages devoted to the prime number theorem. . . . His starting
point was a remarkable identity discovered by Euler over a century earlier: if s is a real number
greater than 1, then
X
Y
1
1
=
,
s
s)
n
1
(1/p
p
n=1
where the expression on the right denotes the product of the numbers (1 ps )1 for all primes
p. To understand how this identity arises, we note that 1/(1 x) = 1 + x + x2 + for |x| < 1,
so for each p we have
1
1
1
= 1 + s + 2s + .
s
1 (1/p )
p
p
On multiplying these series for all primes p and recalling that each integer n > 1 is uniquely
expressible as a product of powers of different primes, we see that

Y
Y
1
1
1
=
1 + s + 2s +
1 (1/ps )
p
p
p
p
=
1+
1
1
1
+ s + + s +
s
2
3
n
221
X
1
,
=
s
n
n=1
which is the desired identity. The sum of this series is evidently a function of the real variable
s > 1, and the identity establishes a connection between the behavior of this function and
properties of the primes. Euler himself exploited this connection in several ways, but Riemann
perceived that access to the deeper features of the distribution of primes can only be gained by
allowing s to be a complex variable. . . . In his paper he proved several important properties of
this function, and in a sovereign way simply stated a number of others without proof. During
the century since his death, many of the finest mathematicians in the world have exerted their
strongest efforts and created rich new branches of analysis in attempts to prove these statements.
The first success was achieved in 1893 by J. Hadamard, and with one exception every statement
has since been settled in the sense Riemann expected. (Hadamards work led him to his 1896
proof of the prime number theorem.) The one exception is the famous Riemann hypothesis:
that all the zeros of (s) in the strip 0 <e{s} 1 lie on the central line <e{s} = 12 . It
stands today as the most important unsolved problem in mathematics, and is probably the most
difficult problem that the mind of man has ever conceived. In a fragmentary note found among
his posthumous papers, Riemann wrote that these theorems follow from an expression for the
function (s) which I have not yet simplified enough to publish. . . .
At the age of thirty-nine Riemann died of tuberculosis in Italy, on the last of several trips he
undertook in order to avoid the cold, wet climate of northern Germany.
General character of the zeta function

It is easy to see that (s) diverges at s = 1 and that it decreases monotonically to 1 as s increases from 1.
As the historical note suggests, its behavior for complex values of s is much more elaborate.
222
223
These graphs were made using Mathematica commands like

Plot[Abs[Zeta[I*y]], {y, 100, 140}].
Its fun to try them with other values of the range.
Exact values for the zeta function

The zeta function can be evaluated exactly for positive even integer arguments:
(m) =
(2)m |Bm |
2m!
for m = 2, 4, 6, . . . .
Here the Bm are the Bernoulli numbers, defined through
X
Bm m
x
=
x .
x
e 1 m=0 m!
The first few Bernoulli numbers are

B0 = 1,
B1 = 12 ,
B2 = 16 ,
whence
(2) =
2
,
6
1
B4 = 30
,
(4) =
4
,
90
B6 =
1
42 ,
(6) =
1
B8 = 30
,
6
,
945
(8) =
B10 =
8
.
9450
5
66 ,
691
B12 = 2730
224
The zeta function near s = 1

The zeta function diverges with a simple pole at s = 1. In the vicinity of the divergence it is approximately
equal to
1
(s) C +
s1
where C is Eulers constant, 0.5772. . . . (The rest of this section is stolen from Bender and Orszag.)
When s is near 1, the defining series
(s) =
X
1
ns
n=1
is very slowly converging. About 1020 terms are needed to compute (1.1) accurate to 1 percent!
Fortunately, there is an expression for the difference between (s) and the N th partial sum:
(s) =
Z s1 N u
N
X
1
u e
1
+
du.
n
s
(s)
eu 1
0
n=0
The integral cannot be evaluated analytically but it can be expanded in an asymptotic series, giving
(s)
N
X
1
1
1
B2 s
B4 s(s + 1)(s + 2)
+
+ s+1 +
+
n
s1
s
s
(s 1)N
2N
N
2N s+3
n=0
where the Bm are again the Bernoulli numbers.

Truncating this formula can give extremely accurate results. For example, using N = 9, only 38 terms
of the series are needed to find (1.1) accurate to 26 decimal places.
References
G.F. Simmons, Differential Equations with Applications and Historical Notes, (McGraw-Hill, New York,
1972) pp. 210211, 214218.
C.M. Bender and S.A. Orszag, Advanced Mathematical Methods for Scientists and Engineers, (McGraw-Hill,
New York, 1978) p. 379.
Jahnke, Emde, and L
osch, Tables of Higher Functions, 6th ed. (McGraw-Hill, New York, 1960) chap. IV.
(Note particularly the beautiful relief plot of |(s)| for complex values of s.)
J. Spanier and K. Oldham, An Atlas of Functions, (Hemisphere Publishing, Washington, D.C., 1987) chap. 3.
(Presents an algorithm purporting to find (s) accurately for any real s.)
M. Abramowitz and I. Stegum, Handbook of Mathematical Functions, (National Bureau of Standards, Washington, D.C., 1964) chap. 23.
M.C. Gutzwiller, Chaos in Classical and Quantum Mechanics, (Springer-Verlag, New York, 1990) section 17.9.
Appendix H
Tutorial on Matrix Diagonalization

You know from as far back as your introductory mechanics course that some problems are difficult given
one choice of coordinate axes and easy or even trivial given another. (For example, the famous monkey
and hunter problem is difficult using a horizontal axis, but easy using an axis stretching from the hunter
to the monkey.) The mathematical field of linear algebra is devoted, in large part, to systematic techniques
for finding coordinate systems that make problems easy. This tutorial introduces the most valuable of these
techniques. It assumes that you are familiar with matrix multiplication and with the ideas of the inverse,
the transpose, and the determinant of a square matrix. It is also useful to have a nodding acquaintance with
the inertia tensor.
This presentation is intentionally non-rigorous. A rigorous, formal treatment of matrix diagonalization
can be found in any linear algebra textbook,1 and there is no need to duplicate that function here. What
is provided here instead is a heuristic picture of whats going on in matrix diagonalization, how it works,
and why anyone would want to do such a thing anyway. Thus this presentation complements, rather than
replaces, the logically impeccable (bulletproof) arguments of the mathematics texts.
Essential problems in this tutorial are marked by asterisks ( ).
Warning: This tutorial is still in draft form.
H.1
Whats in a name?
There is a difference between an entity and its name. For example, a tree is made of wood, whereas its
name tree made of ink. One way to see this is to note that in German, the name for a tree is Baum, so
the name changes upon translation, but the tree itself does not change. (Throughout this tutorial, the term
translate is used as in translate from one language to another rather than as in translate by moving in
a straight line.)
225
226
APPENDIX H. TUTORIAL ON MATRIX DIAGONALIZATION
The same holds for mathematical entities. Suppose a length is represented by the number 2 because
it is two feet long. Then the same length is represented by the number 24 because it is twenty-four inches
long. The same length is represented by two different numbers, just as the same tree has two different names.
The representation of a length as a number depends not only upon the length, but also upon the coordinate
system used to measure the length.
H.2
Vectors in two dimensions
One way of describing a two-dimensional vector V is by giving its x and y components in the form of a 2 1
column matrix
!
Vx
.
(H.1)
Vy
Indeed it is sometimes said that the vector V is equal to the column matrix (H.1). This is not precisely
correctit is better to say that the vector is described by the column matrix or represented by the column
matrix or that its name is the column matrix. This is because if you describe the vector using a different
set of coordinate axes you will come up with a different column matrix to describe the same vector. For
example, in the situation shown below the descriptions in terms of the two different coordinate systems are
related through the matrix equation
!
!
!
Vx0
cos sin
Vx
=
.
(H.2)
Vy 0
sin cos
Vy
y6
OC
C
y0 C
V

C
C
C
C
C
C

:

C
0

x
C
C

C

x
C
The 2 2 matrix above is called the rotation matrix and is usually denoted by R():
!
cos sin
R()
.
sin cos
(H.3)
H.2. VECTORS IN TWO DIMENSIONS
227
One interesting property of the rotation matrix is that it is always invertible, and that its inverse is equal to
its transpose. Such matrices are called orthogonal.1 You could prove this by working a matrix multiplication,
but it is easier to simply realize that the inverse of a rotation by is simply a rotation by , and noting
that
R1 () = R() = R ().
(H.4)
(The dagger represents matrix transposition.)
There are, of course, an infinite number of column matrix representations for any vector, corresponding
to the infinite number of coordinate axis rotations with from 0 to 2. But one of these representations is
special: It is the one in which the x0 -axis lines up with the vector, so the column matrix representation is
just
!
V
,
(H.5)
0
q
where V = |V| = Vx2 + Vy2 is the magnitude of the vector. This set of coordinates is the preferred (or
canonical) set for dealing with this vector: one of the two components is zero, the easiest number to
deal with, and the other component is a physically important number. You might wonder how I can claim
that this representation has full information about the vector: The initial representation (H.1) contains two
independent numbers, whereas the preferred representation (H.5) contains only one. The answer is that the
preferred representation contains one number (the magnitude of the vector) explicitly while another number
(the polar angle of the vector relative to the initial x-axis) is contained implicitly in the rotation needed to
produce the preferred coordinate system.
H.1 Problem: Right angle rotations
Verify equation (H.2) in the special cases = 90 , = 180 , = 270 , and = 360 .
H.2 Problem: The rotation matrix
a. Derive equation (H.2) through purely geometrical arguments.
b. Express i0 and j0 , the unit vectors of the (x0 , y 0 ) coordinate system, as linear combinations of i
and j. Then use
Vx0 = Vi0 and Vy0 = Vj0
(H.6)
to derive equation (H.2).
c. Which derivation do you find easier?
H.3 Problem: Rotation to the preferred coordinate system
In the preferred coordinate system, Vy0 = 0. Use this requirement to show that the preferred system
is rotated from the initial system by an angle with
tan =
1 Although
Vy
.
Vx
(H.7)
all rotation matrices are orthogonal, there are orthogonal matrices that are not rotation matrices: see problem H.4.
228

For any value of Vy /Vx , there are two angles that satisfy this equation. What is the representation of
V in each of these two coordinate systems?
H.4 Problem: A non-rotation orthogonal transformation

In one coordinate system the y-axis is vertical and the x-axis points to the right. In another the y 0 -axis
is vertical and the x0 -axis points to the left. Find the matrix that translates vector coordinates from
one system to the other. Show that this matrix is orthogonal but not a rotation matrix.
H.5 Problem: Other changes of coordinate
Suppose vertical distances (distances in the y direction) are measured in feet while horizontal distances
(distances in the x direction) are measured in miles. (This system is not perverse. It is used in nearly
all American road maps.) Find the matrix that changes the representation of a vector in this coordinate
system to the representation of a vector in a system where all distances are measured in feet. Find the
matrix that translates back. Are these matrices orthogonal?
H.6 Problem: Other special representations
At equation (H.5) we mentioned one special (or canonical) representation of a vector. There are
three others, namely
!
!
!
0
V
0
,
,
.
(H.8)
V
0
V
If coordinate-system rotation angle brings the vector representation into the form (H.5), then what
rotation angle will result in these three representations?
H.3
Tensors in two dimensions
A tensor, like a vector, is a geometrical entity that may be described (named) through components, but
a d-dimensional tensor requires d2 rather than d components. Tensors are less familiar and more difficult
to visualize than vectors, but they are neither less important nor less physical. We will introduce tensors
through the concrete example of the inertia tensor of classical mechanics (see, for example, reference [2]),
but the results we present will be perfectly general.
Just as the two components of a two-dimensional vector are most easily kept track of through a 2 1
matrix, so the four components of two-dimensional tensor are most conveniently written in the form of a 22
matrix. For example, the inertia tensor T of a point particle with mass m located2 at (x, y) has components
!
my 2
mxy
T=
.
(H.9)
mxy mx2
(Note the distinction between the tensor T and its matrix of components, its name, T.) As with vector
components, the tensor components are different in different coordinate systems, although the tensor itself
2 Or,
to be absolutely precise, the particle located at the point represented by the vector with components (x, y).
H.3. TENSORS IN TWO DIMENSIONS
229
does not change. For example, in the primed coordinate system of the figure on page 226, the tensor
components are of course
!
my 02
mx0 y 0
0
T =
.
(H.10)
mx0 y 0
mx02
A little calculation shows that the components of the inertia tensor in two different coordinate systems are
related through
T0 = R()TR1 ().
(H.11)
This relation holds for any tensor, not just the inertia tensor. (In fact, one way to define tensor is as an
entity with four components that satisfy the above relation under rotation.) If the matrix representing a
tensor is symmetric (i.e. the matrix is equal to its transpose) in one coordinate system, then it is symmetric
in all coordinate systems (see problem H.7). Therefore the symmetry is a property of the tensor, not of its
matrix representation, and we may speak of a symmetric tensor rather than just a tensor represented by
a symmetric matrix.
As with vectors, one of the many matrix representations of a given tensor is considered special (or
canonical): It is the one in which the lower left component is zero. Furthermore if the tensor is symmetric
(as the inertia tensor is) then in this preferred coordinate system the upper right component will be zero also,
so the matrix will be all zeros except for the diagonal elements. Such a matrix is called a diagonal matrix
and the process of finding the rotation that renders the matrix representation of a symmetric tensor diagonal
is called diagonalization.3 We may do an accounting of information for this preferred coordinate system
just as we did with vectors. In the initial coordinate system, the symmetric tensor had three independent
components. In the preferred system, it has two independent components manifestly visible in the diagonal
matrix representation, and one number hidden through the specification of the rotation.
H.7 Problem: Representations of symmetric tensors
Show that if the matrix S representing a tensor is symmetric, and if B is any orthogonal matrix, then
all of the representations
BSB
(H.12)
are symmetric. (Clue: If you try to solve this problem for rotations in two dimensions using the explicit
rotation matrix (H.3), you will find it solvable but messy. The clue is that this problem asks you do
prove the result in any number of dimensions, and for any orthogonal matrix B, not just rotation
matrices. This more general problem is considerably easier to solve.)
H.8 Problem: Diagonal inertia tensor
The matrix (H.9) represents the inertia tensor of a point particle with mass m located a distance r
from the origin. Show that the matrix is diagonal in four different coordinate systems: one in which
the x0 -axis points directly toward the particle, one in which the y 0 -axis points directly away from the
particle, one in which the x0 -axis points directly away from the particle, and one in which the y 0 -axis
3 An
efficient algorithm for diagonalization is discussed in section H.8. For the moment, we are more interested in knowing
that a diagonal matrix representation must exist than in knowing how to most easily find that preferred coordinate system.
230

points directly toward the particle. Find the matrix representation in each of these four coordinate
systems.
H.9 Problem: Representations of a certain tensor

Show that a tensor represented in one coordinate system by a diagonal matrix with equal elements,
namely
!
d0 0
,
(H.13)
0 d0
has the same representation in all orthogonal coordinate systems.
H.10 Problem: Rotation to the preferred coordinate system
A tensor is represented in the initial coordinate system by
!
a b
.
b c
(H.14)
Show that the tensor is diagonal in a preferred coordinate system which is rotated from the initial
system by an angle with
2b
.
(H.15)
tan(2) =
ac
This equation has four solutions. Find the rotation matrix for = 90 , then show how the four different
diagonal representations are related. You do not need to find any of the diagonal representations in
terms of a, b and c. . . just show what the other three are given that one of them is
!
d1 0
.
(H.16)
0 d2
H.11 Problem: Inertia tensor in outer product notation

The discussion in this section has emphasized the tensors matrix representation (name) T rather
than the tensor T itself.
a. Define the identity tensor 1 as the tensor represented in some coordinate system by
!
1 0
.
1=
0 1
(H.17)
Show that this tensor has the same representation in any coordinate system.
b. Show that the inner product between two vectors results in a scalar: Namely
!
ax
if vector bf a is represented by
and vector bf b is represented by
ay
then the inner product a b is given through

ax
ay
bx
by
!
= ax bx + ay by ,
bx
by
H.4. TENSORS IN THREE DIMENSIONS
231
and this inner product is a scalar. (A 1 2 matrix times a 2 1 matrix is a 1 1 matrix.) That is,
the vector a is represented by different coordinates in different coordinate systems, and the vector
b is represented by different coordinates in different coordinate systems, but the inner product
a b is the same in all coordinate systems.
c. In contrast, show that the outer product of two vectors is a tensor: Namely
!
!

ax bx ax by
ax
.
.
ab =
bx by =
ay bx ay by
ay
(A 2 1 matrix times a 1 2 matrix is a 2 2 matrix.) That is, show that the representation of
ab transforms from one coordinate system to another as specified through (H.11).
d. Show that the inertia tensor for a single particle of mass m located at position r can be written
in coordinate-independent fashion as
T = m1r2 mrr.
H.4
(H.18)
Tensors in three dimensions
A three-dimensional tensor is represented in component form by a 3 3 matrix with nine entries. If the
tensor is symmetric, there are six independent elements. . . three on the diagonal and three off-diagonal. The
components of a tensor in three dimensions change with coordinate system according to
T0 = RTR ,
(H.19)
where R is the 3 3 rotation matrix.

A rotation in two dimension is described completely by giving a single angle. In three dimensions more
information is required. Specifically, we need not only the amount of the rotation, but we must also know the
plane in which the rotation takes place. We can specify the plane by giving the unit vector perpendicular to
that plane. Specifying an arbitrary vector in three dimensions requires three numbers, but specifying a unit
vector in three dimensions requires only two numbers because the magnitude is already fixed at unity. Thus
three numbers are required to specify a rotation in three dimensions: two to specify the rotations plane, one
to specify the rotations size. (One particularly convenient way to specify a three-dimensional rotation is
through the three Euler angles. Reference [3] defines these angles and shows how to write the 3 3 rotation
matrix in terms of these variables. For the purposes of this tutorial, however, we will not need an explicit
rotation matrix. . . all we need is to know is the number of angles required to specify a rotation.)
In two dimensions, any symmetric tensor (which has three independent elements), could be represented
by a diagonal tensor (with two independent elements) plus a rotation (one angle). We were able to back up
this claim with an explicit expression for the angle.
In three dimensions it seems reasonable that any symmetric tensor (six independent elements) can be
represented by a diagonal tensor (three independent elements) plus a rotation (three angles). The three
232
angles just have to be selected carefully enough to make sure that they cause the off-diagonal elements
to vanish. This supposition is indeed correct, although we will not pause for long enough to prove it by
producing explicit formulas for the three angles.
H.5
Tensors in d dimensions
A d-dimensional tensor is represented by a dd matrix with d2 entries. If the tensor is symmetric, there are d
independent on-diagonal elements and d(d 1)/2 independent off-diagonal elements. The tensor components
will change with coordinate system in the now-familiar form
T0 = RTR ,
(H.20)
where R is the d d rotation matrix.

How many angles does it take to specify a rotation in d dimensions? Remember how we went from
two dimensions to three: The three dimensional rotation took place in a plane, i.e. in a two-dimensional
subspace. It required two (i.e. d 1) angles to specify the orientation of the plane plus one to specify the
rotation within the plane. . . a total of three angles.
A rotation in four dimensions takes place within a three-dimensional subspace. It requires 3 = d 1
angles to specify the orientation of the three-dimensional subspace, plus, as we found above, three angles to
specify the rotation within the three-dimensional subspace. . . a total of six angles.
A rotation in five dimensions requires 4 = d 1 angles to specify the four-dimensional subspace in which
the rotation occurs, plus the six angles that we have just found specify a rotation within that subspace. . . a
total of ten angles.
In general, the number of angles needed to specify a rotation in d dimensions is
Ad = d 1 + Ad1 = d(d 1)/2.
(H.21)
This is exactly the number of independent off-diagonal elements in a symmetric tensor. It seems reasonable
that we can choose the angles to ensure that, in the resulting coordinate system, all the off-diagonal elements
vanish. The proof of this result is difficult and proceeds in a very different manner from the plausibility
argument sketched here. (The proof involves concepts like eigenvectors and eigenvalues, and it gives an
explicit recipe for constructing the rotation matrix. It has the advantage of rigor and the disadvantage of
being so technical that its easy to lose track of the fact that that all youre doing is choosing a coordinate
system.)
H.12 Problem: Non-symmetric tensors
Argue that a non-symmetric tensor can be brought into a triangular representation in which all the
elements below the diagonal are equal to zero and all the elements on and above the diagonal are
independent. (This is indeed the case, although in general some of the non-zero elements remaining
will be complex-valued, and some of the angles will involve rotations into complex-valued vectors.)
H.6. LINEAR TRANSFORMATIONS IN TWO DIMENSIONS
H.6
233
Linear transformations in two dimensions
Section H.3 considered 2 2 matrices as representations of tensors. This section gains additional insight by
considering 2 2 matrices as representations of linear transformations. It demonstrates how diagonalization
can be useful and gives a clue to an efficient algorithm for diagonalization.
A linear transformation is a function from vectors to vectors that can be represented in any given
coordinate system as
!
!
!
a11 a12
x
u
=
.
(H.22)
a21 a22
y
v
If the equation above represents (names) the transformation in one coordinate system, what is its representation in some other coordinate system? We assume that the two coordinate systems are related through
an orthogonal matrix B such that
!
!
!
!
u0
u
x0
x
=B
and
=B
.
(H.23)
v0
v
y0
y
(For example, if the new coordinate system is the
the matrix B that translates from the original to
this translation dictionary, we have
!
u0
=B
v0
primed coordinate system of the figure on page 226, then

the new coordinates is the rotation matrix R().) Given
a11
a21
a12
a22
x
y
!
.
(H.24)
But B is invertible, so
x
y
whence
u0
v0
!
=B
!
1
=B
a11
a21
a12
a22
x0
y0
(H.25)
!
1
x0
y0
!
.
Thus the representation of the transformation in the primed coordinate system is

!
a11 a12
B
B1
a21 a22
(H.26)
(H.27)
(compare equation H.11). This equation has a very direct physical meaning. Remember that the matrix B
translates from the old (x, y) coordinates to the new (x0 , y 0 ) coordinates, while the matrix B1 translates in
the opposite direction. Thus the equation above says that the representation of a transformation in the new
coordinates is given by translating from new to old coordinates (through the matrix B1 ), then applying the
old representation (the a matrix) to those old coordinates, and finally translating back from old to new
coordinates (through the matrix B).
The rest of this section considers only transformations represented by symmetric matrices, which we will
denote by
!
!
!
u
a b
x
=
.
(H.28)
v
b c
y
234
Lets try to understand this transformation as something more than a jumble of symbols awaiting a plunge
into the calculator. First of all, suppose the vector V maps to the vector W. Then the vector 5V will be
mapped to vector 5W. In short, if we know how the transformation acts on vectors with magnitude unity,
we will be able to see immediately how it acts on vectors with other magnitudes. Thus we focus our attention
on vectors on the unit circle:
x2 + y 2 = 1.
(H.29)
A brief calculation shows that the length of the output vector is then
p
p
u2 + v 2 = a2 x2 + b2 + c2 y 2 + 2b(a + c)xy,
(H.30)
which isnt very helpful. Another brief calculation shows that if the input vector has polar angle , then the
output vector has polar angle with
b + c tan
tan =
,
(H.31)
a + b tan
which is similarly opaque and messy.
Instead of trying to understand the transformation in its initial coordinate system, lets instead convert
(rotate) to the special coordinate system in which the transformation is represented by a diagonal matrix.
In this system,
!
!
!
!
u0
d1 0
x0
d1 x0
=
=
.
(H.32)
v0
0 d2
y0
d2 y 0
The unit circle is still
x02 + y 02 = 1,
(H.33)
so the image of the unit circle is
2 0 2
v
u0
+
= 1,
(H.34)
d1
d2
namely an ellipse! This result is transparent in the special coordinate system, but almost impossible to see
in the original one.

Note particularly what happens to a vector pointing along the x0 coordinate axis. For example, the unit
vector in this direction transforms to
!
!
!
d1
d1 0
1
=
.
(H.35)
0
0 d2
0
In other words, the when the vector is transformed it changes in magnitude, but not in direction. Vectors
with this property are called eigenvectors. It is easy to see that any vector on either the x0 or y 0 coordinate
axes are eigenvectors.
H.7
What does eigen mean?
If a vector x is acted upon by a linear transformation B, then the output vector

x0 = Bx
(H.36)
H.8. HOW TO DIAGONALIZE A SYMMETRIC MATRIX
235
will usually be skew to the original vector x. However, for some very special vectors it might just happen
that x0 is parallel to x. Such vectors are called eigenvectors. (This is a terrible name because (1) it gives
no idea of what eigenvectors are or why theyre so important and (2) it sounds gross. However, thats what
theyre called.) We have already seen, in the previous section, that eigenvectors are related to coordinate
systems in which the transformation is particularly easy to understand.
If x is an eigenvector, then
Bx = x,
(H.37)
where is a scalar called the eigenvalue associated with eigenvector x. If x is an eigenvector, then any
vector parallel to x is also an eigenvector with the same eigenvalue. (That is, any vector of the form cx,
where c is any scalar, is also an eigenvector with the same eigenvalue.) Sometimes we speak of a line of
eigenvectors.
The vector x = 0 is never considered an eigenvector, because
B0 = 0,
(H.38)
for any value of for any linear transformation. On the other hand, if
Bx = 0x = 0
(H.39)
for some non-zero vector x, then x is an eigenvector with eigenvalue = 0.

H.13 Problem: Plane of eigenvectors
Suppose x and y are two non-parallel vectors with the same eigenvalue. (In this case the eigenvalue
is said to be degenerate, which sounds like an aspersion cast upon the morals of the eigenvalue but
which is really just poor choice of terminology again.) Show that any vector of the form c1 x + c2 y is
an eigenvector with the same eigenvalue.
H.8
How to diagonalize a symmetric matrix
We saw in section H.3 that for any 2 2 symmetric matrix, represented in its initial basis by, say,
!
a b
,
b c
(H.40)
a simple rotation of axes would produce a new coordinate system in which the matrix representation is
diagonal:
!
d1 0
.
(H.41)
0 d2
These two matrices are related through
d1
0
0
d2
!
= R()
a
b
b
c
!
R1 (),
(H.42)
236
where R() is the rotation matrix (H.3). Problem H.10 gave a direct way to find the desired rotation.
However this direct technique is cumbersome and doesnt generalize readily to higher dimensions. This
section presents a different technique, which relies on eigenvalues and eigenvectors, that is more efficient and
that generalizes readily to complex-valued matrices and to matrices in any dimension, but that is somewhat
sneaky and conceptually roundabout.
We begin by noting that any vector lying along the x0 -axis (of the preferred coordinate system) is an
eigenvector. For example, the vector 5i0 is represented (in the preferred coordinate system) by
!
5
.
(H.43)
0
Multiplying this vector by the matrix in question gives
!
!
d1 0
5
= d1
0 d2
0
5
0
(H.44)
so 5i0 is an eigenvector with eigenvalue d1 . The same holds for any scalar multiple of i0 , whether positive or
negative. Similarly, any scalar multiple of j0 is an eigenvector with eigenvalue d2 . In short, the two elements
on the diagonal in the preferred (diagonal) representation are the two eigenvalues, and the two unit vectors
i0 and j0 of the preferred coordinate system are two of the eigenvectors.
Thus finding the eigenvectors and eigenvalues of a matrix gives you the information needed to diagonalize
that matrix. The unit vectors i0 and j0 constitute an orthonormal basis of eigenvectors. The eigenvectors
even give the rotation matrix directly, as described in the next paragraph.
Lets call the rotation matrix
B=
b11
b21
b12
b22
!
,
(H.45)
so that the inverse (transpose) matrix is

b11
b12
B1 = B =
b21
b22
!
.
(H.46)
The representation of i0 in the preferred basis is

1
0
!
,
(H.47)
so its representation in the initial basis is (see equation H.2)

!
!
!
1
b11 b21
1
B
=
=
0
b12 b22
0
b11
b12
Similarly, the representation of j0 in the initial basis is

!
!
0
b
b
11
21
B
=
1
b12 b22
b21
b22
0
1
!
=
(H.48)
(H.49)

Thus the rotation matrix is
B=
237
initial rep. of i0 , on its side

initial rep. of j0 , on its side
(H.50)
Example
Suppose we need to find a diagonal representation for the matrix
!
7 3
T=
.
3 7
(H.51)
First we search for the special vectorsthe eigenvectorssuch that

!
!
!
7 3
x
x
=
.
3 7
y
y
(H.52)
At the moment, we dont know either the eigenvalue or the associated eigenvector (x, y). Thus it seems
that (bad news) we are trying to solve two equations for three unknowns:
7x + 3y
3x + 7y
(H.53)
Remember, however, that there is not one single eigenvector: any multiple of an eigenvector is also an
eigenvector. (Alternatively, any vector on the line that extends the eigenvector is another eigenvector.) We
only need one of these eigenvectors, so lets take the one that has x = 1 (i.e. the vector on the extension line
where it intersects the vertical line x = 1). (This technique will fail if we have the bad luck that our actual
eigenvector is vertical and hence never passes through the line x = 1.) So we really have two equations in
two unknowns:
7 + 3y
3 + 7y
but note that they are not linear equations. . . the damnable product y in the lower right corner means that
all our techniques for solving linear equations go right out the window. We can solve these two equations
for and y, but theres an easier, if somewhat roundabout, approach.
Finding eigenvalues
Lets go back to equation (H.52) and write it as
!
!
7 3
x
3 7
y
x
y
!
=
0
0
!
.
(H.54)
238
Then
7
3
3
7
x
y
or
7
3
1
0
3
7
0
1
x
y
x
y
=
0
0
=
0
0
!
(H.55)
!
.
Lets think about this. It says that for some matrix M = T 1, we have
!
!
x
0
M
=
.
y
0
(H.56)
(H.57)
You know right away one vector (x, y) that satisfies this equation, namely (x, y) = (0, 0). And most of the
time, this is the only vector that satisfies the equation, because
!
!
!
0
0
x
1
=M
=
.
(H.58)
0
0
y
We appear to have reached a dead end. The solution is (x, y) = (0, 0), but the zero vector is not, by
definition, considered an eigenvector of any transformation. (Because it always gives eigenvalue zero for any
transformation.)
However, if the matrix M is not invertible, then there will be other solutions to
!
!
x
0
M
=
.
y
0
(H.59)
in addition to the trivial solution (x, y) = (0, 0). Thus we must look for those special values of such that
the so-called characteristic matrix M is not invertible. These values come if and only if the determinant of
M vanishes. For this example, we have to find values of such that
!
7
3
det
= 0.
(H.60)
3
7
This is a quadratic equation in
(7 )2 32 = 0
(H.61)
called the characteristic equation. Its two solutions are

7 = 3
(H.62)
= 7 3 = 10 or 4.
(H.63)
or
We have found the two eigenvalues of our matrix!
239
Finding eigenvectors
Lets look now for the eigenvector associated with = 4. Equation (H.53)
7x + 3y
3x + 7y
still holds, but no longer does it look like two equations in three unknowns, because we are now interested
in the case = 4:
7x + 3y
4x
3x + 7y
4y
3x + 3y
3x + 3y
Following our nose gives
and when we see this our heart skips a beat or two. . . a degenerate system of equations! Relax and rest
your heart. This system has an infinite number of solutions and its supposed to have an infinite number
of solutions, because any multiple of an eigenvector is also an eigenvector. The eigenvectors associated with
= 4 are any multiple of
!
1
.
(H.64)
1
An entirely analogous search for the eigenvectors associated with = 10 finds any multiple of
!
1
.
1
(H.65)
Tidying up
We have the two sets of eigenvectors, but which shall we call i0 and which j0 ? This is a matter of individual
choice, but my choice is usually to make the transformation be a rotation (without reflection) through a
small positive angle. Our new, preferred coordinate system is related to the original coordinates by a simple
rotation of 45 if we choose
!
!
1
1
0
1
i0 = 1
and j = 2
.
(H.66)
2
1
1
(Note that we have also normalized the basis, i.e. selected the basis vectors to have magnitude unity.)
Given this choice, the orthogonal rotation matrix that changes coordinates from the original to the preferred
system is (see equation H.50)
!
1
1
B = 12
(H.67)
1 1
240
and the diagonalized matrix (or, more properly, the representation of the matrix in the preferred coordinate
system) is
!
10 0
.
(H.68)
0 4
You dont believe me? Then multiply out
B
7
3
3
7
!
B
(H.69)
and see for yourself.
Problems
H.14 Problem: Diagonalize a 2 2 matrix
Diagonalize the matrix
26
12
12
19
!
.
(H.70)
a. Find its eigenvalues.

b. Find its eigenvectors, and verify that they are orthogonal.
c. Sketch the eigenvectors, and determine the signs and sequence most convenient for assigning axes.
(That is, should the first eigenvector you found be called i0 , i0 , or j0 ?)
d. Find the matrix that translates from the initial basis to the basis of eigenvectors produced in part
(c.).
e. Verify that the matrix produced in part (d.) is orthogonal.
f. Verify that the representation of the matrix above in the basis of eigenvectors is diagonal.
g. (Optional.) What is the rotation angle?
H.15 Problem: Eigenvalues of a 2 2 matrix
Show that the eigenvalues of
a
b
b
c
(a + c)
!
(H.71)
are
=
1
2
i
(a c)2 + 4b2 .
(H.72)
Under what circumstances is an eigenvalue complex valued? Under what circumstances are the two
eigenvalues the same?
H.16 Problem: Diagonalize a 3 3 matrix
Diagonalize the matrix
1182
1
924
625
540
924
643
720
540
720 .
575
(H.73)
H.9. A GLANCE AT COMPUTER ALGORITHMS
241
a. Find its eigenvalues by showing that the characteristic equation is

3 22 5 + 6 = ( 3)( + 2)( 1) = 0.
(H.74)
b. Find its eigenvectors, and verify that they are orthogonal.

c. Show that the translation matrix can be chosen
20
1
B=
9
25
12
to be
15
0
12 20 .
16
15
(H.75)
Why did I use the phrase the translation matrix can be chosen to be rather then the translation
matrix is?
H.17 Problem: A 3 3 matrix eigenproblem
Find the eigenvalues and associated eigenvectors
2
3
H.9
for the matrix
2 3
3 4 .
4 5
(H.76)
A glance at computer algorithms
Anyone who has worked even one of the problems in section H.8 knows that diagonalizing a matrix is no
picnic: theres a lot of mundane arithmetic involved and its very easy to make mistakes. This is a problem
ripe for computer solution. Ones first thought is to program a computer to solve the problem using the same
technique that we used to solve it on paper: first find the eigenvalues through the characteristic equation,
then find the eigenvectors through a degenerate set of linear equations.
This turns out to be a very poor algorithm for automatic computation. The effective algorithm is to
choose a matrix B such that the off-diagonal elements of
BAB1
(H.77)
are smaller than the off-diagonal elements of A. Then choose another, and another. Go through this process
again and again until the off-diagonal elements have been ground down to machine zero. There are many
strategies for choosing the series of B matrices. These are well-described in any edition of Numerical Recipes.4
When you need to diagonalize matrices numerically, I urge you to look at Numerical Recipes to see whats
going on, but I urge you not to code these algorithms yourself. These algorithms rely in an essential way
on the fact that computer arithmetic is approximate rather than exact, and hence they are quite tricky
to implement. Instead of coding the algorithms yourself, I recommend that you use the implementations
in either LAPACK5 (the Linear Algebra PACKage) or EISPACK.6 These packages are probably the finest
computer software ever written, and they are free. They can be obtained through the Guide to Available
Mathematical Software (GAMS) at http://gams.nist.gov.
242
H.10
A glance at non-symmetric matrices and the Jordan form
Many of the matrices that arise in applications are symmetric and hence the results of the previous sections
are the only ones needed. But every once in a while you do encounter a non-symmetric matrix and this
section gives you a guide to treating them. It is just an introduction and treats only 2 2 matrices.
Given a non-symmetric matrix, the first thing to do is rotate the axes to make the matrix representation
triangular, as discussed in problem H.12:
!
a b
.
(H.78)
0 c
Note that b 6= 0 because otherwise the matrix would be symmetric and we would already be done. In this
case vectors on the x-axis are eigenvectors because
!
!
!
a b
1
1
=a
.
(H.79)
0 c
0
0
Are there any other eigenvectors? The equation
!
a b
0 c
x
y
!
=
x
y
!
(H.80)
tells us that
ax + by
= x
cy
= y
whence = c and the eigenvector has polar angle where

tan =
ca
.
b
(H.81)
Note that if c = a (the degenerate case: both eigenvalues are the same) then = 0 or = . In this case
all of the eigenvectors are on the x-axis.
Diagonal form
We already know that that a rotation of orthogonal (Cartesian) coordinates will not diagonalize this matrix.
We must instead transform to a skew coordinate system in which the axes are not perpendicular.
H.10. A GLANCE AT NON-SYMMETRIC MATRICES AND THE JORDAN FORM
y6

y0

243
Vx0
*

Vy 0

V

-0
x, x

Note that in with oblique axes, the coordinates are given by

V = Vx0i0 + Vy0j0
(H.82)
but, because i0 and j0 are not perpendicular, it is not true that

Vx0 = V i0 .
NO!
(H.83)
A little bit of geometry will convince you that the name of the vector V changes according to
!
!
Vx0
Vx
=B
,
Vy0
Vy
where
1
B=
sin
sin cos
0
1
(H.84)
!
.
(H.85)
This matrix is not orthogonal. In fact its inverse is

1
1
0
cos
sin
!
.
(H.86)
Finally, note that we cannot have = 0 or = , because then both Vx0 and Vy0 would give information about
the horizontal component of the vector, and there would be no information about the vertical component of
the vector.
What does this say about the representations of tensors (or, equivalently, of linear transformations)? The
name translation argument of equation (H.27) still applies, so
T0 = BTB1 .
(H.87)
244
Using the explicit matrices already given, this says

!
!
sin cos
a b
1
1
0
T =
sin
0
1
0 c
0
cos
sin
!
=
a
0
(a c) cos + b sin
c
!
.
(H.88)
To make this diagonal, we need only choose a skew coordinate system where the angle gives
(a c) cos + b sin = 0,
(H.89)
that is, one with
ca
.
(H.90)
b
Comparison with equation (H.81) shows that this simply means that the skew coordinate system should
have its axes pointing along two eigenvectors. We have once again found an intimate connection between
diagonal representations and eigenvectors, a connection which is exploited fully in abstract mathematical
treatments of matrix diagonalization.
tan =
Once again we can do an accounting of information. In the initial coordinate system, the four elements
of the matrix contain four independent pieces of information. In the diagonalizing coordinate system, two
of those pieces are explicit in the matrix, and two are implicit in the two axis rotation angles needed to
implement the diagonalization.
This procedure works almost all the time. But, if a = c, then it would involve = 0 or = , and we
have already seen that this is not an acceptable change of coordinates.
Degenerate case
Suppose our matrix has equal eigenvalues, a = c, so that it reads
!
a b
.
0 a
(H.91)
If b = 0, then the matrix is already diagonal. (Indeed, in this case all vectors are eigenvectors with eigenvalue
a, and the linear transformation is simply multiplication of each vector by a).
But if b 6= 0, then, as we have seen, the only eigenvectors are on the x-axis, and it is impossible to make a
basis of eigenvectors. Only one thing can be done to make the matrix representation simpler than it stands
in equation (H.91), and that is a shift in the scale used to measure the y-axis.
For example, suppose that in the (x, y) coordinate system, the y-axis is calibrated in inches. We wish to
switch to the (x0 , y 0 ) system in which the y 0 -axis is calibrated in feet. There is no change in axis orientation
or in the x-axis. It is easy to see that the two sets of coordinates are related through
!
!
!
!
!
!
x0
1
0
x
x
1 0
x0
=
and
=
(H.92)
y0
y0
0 1/12
y
y
0 12
This process is sometimes called a stretching or a scaling of the y-axis.
H.10. A GLANCE AT NON-SYMMETRIC MATRICES AND THE JORDAN FORM
245
The transformation represented by matrix (H.91) in the initial coordinate system is represented in the
new coordinate system by
!
!
!
!
1
0
a b
1 0
a 12b
=
.
(H.93)
0 1/12
0 a
0 12
0 a
The choice of what to do now is clear. Instead of scaling the y-axis by a factor of 12, we can scale it by a
factor of 1/b, and produce a new matrix representation of the form
!
a 1
.
(H.94)
0 a
Where is the information in this case? In the initial coordinate system, the four elements of the matrix
contain four independent pieces of information. In the new coordinate system, two of those pieces are explicit
in the matrix, one is implicit in the rotation angle needed to implement the initial triangularization, and one
is implicit in the y-axis scale transformation.
The Jordan form

Remarkably, the situation discussed above for 2 2 matrices covers all the possible cases for n n matrices.
That is, in n dimensional space, the proper combination of rotations, skews, and stretches of coordinate axes
will bring the matrix representation (the name) of any tensor or linear transformation into a form where
every element is zero except on the diagonal and on the superdiagonal. The elements on the diagonal are
eigenvalues, and each element on the superdiagonal is either zero or one: zero if the two adjacent eigenvalues
differ, either zero or one if they are the same. The warning of problem H.12 applies here as well: The
eigenvalues on the diagonal may well be complex valued, and the same applies for the elements of the new
basis vectors.
References
1
For example, Kenneth Hoffman and Ray Kunze, Linear Algebra, second edition (Prentice-Hall, Englewood Cliffs, New Jersey, 1971).
2
For example, Jerry Marion and Stephen Thorton, Classical Dynamics of Particles and Systems, fourth
edition (Saunders College Publishing, Fort Worth, Texas, 1995) section 11.2.
3
For example, Jerry Marion and Stephen Thorton, Classical Dynamics of Particles and Systems, fourth
edition (Saunders College Publishing, Fort Worth, Texas, 1995) section 11.7.
4
W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flannery, Numerical Recipes (Cambridge University
Press, Cambridge, U.K., 1992).
5
E. Anderson, et al., LAPACK Users Guide (SIAM, Philadelphia, 1992).
B.T. Smith, et al., Matrix Eigensystem RoutinesEISPACK Guide (Springer-Verlag, Berlin, 1976).
Appendix I
Catalog of Misconceptions
Effective teaching does not simply teach students what is correctit also insures that students do not believe
what is incorrect. There are a number of prevalent misconceptions in statistical mechanics. For example, an
excellent history of statistical mechanics is titled The Kind of Motion We Call Heat. This title is wonderfully
memorable and historically justifiable, but it embodies no fewer than three different physics misconceptions:
heat is confused with thermal energy; thermal energy is confused with kinetic energy; and kinetic energy is
confused with motion. Below is a list of misconceptions that are explicitly mentioned in this book, together
with the page number where that misconception is pointed out and corrected.
a microstate (a configuration) has an entropy, 32
all gases are ideal, 44
at high temperature, the most probable microstate has high energy, 102
balls-in-buckets picture of quantal states, 140, 153
Boltzmann distribution is for molecules within systems, 94, 99
Bose-Einstein condensation, 153
canonical probability for a quantal state that doesnt have a definite energy, 108, 123
chemical potential increases with temperature, 46, 147, 148, 150, 163
computers are fast, 194
concerning (anti)symmetric functions, 139
concerning Monte Carlo simulation, 197
concerning multivariate calculus, 59
concerning quantum mechanics, 48
configurational entropy differs from thermal entropy, 50
Cp is E/T )p , 77
definition of partition function, 95
density is uniform at equilibrium, 7, 16, 45, 56
energy must be a function of (S, V, N ), 65, 75
246
247
ensemble is a collection of atoms, 16
entropy as (moral?) decay, 30
entropy as disorder, 34
entropy doesnt apply to a pendulum, 130
equation of state contains all thermodynamic information, 65, 83
heat is thermal energy, 50
highest temperature implies maximum energy, 130, 131
ideal gas is hard-sphere gas, 21
ideal gas potential energy, 21, 97
in Reif, 48
occupancy probability vs. canonical probability, 144
one molecule is an ensemble element, 94, 99
phase diagram plots pressure as a function of temperature, 7
probability of atom having energy E is proportional to eE/kB T , 99
sum over typical microstates, not over all microstates, 23, 95
temperature is energy per particle, 42
the most probable energy is the energy of the most probable microstate, 102
thermal energy is kinetic energy, 50, 165
Appendix J
Thermodynamic Master Equations

E(S, V, N )
dE = T dS p dV + dN
F (T, V, N ) = E T S
dF = S dT p dV + dN
H(S, p, N ) = E + pV
dH = T dS + V dp + dN
G(T, p, N ) = F + pV
dG = S dT + V dp + dN
(T, V, ) = F N = pV
d = S dT p dV N d
p(T, ) [intensive quantities only]

dp = S dT + d
[S = S/V,
= N/V ]
248
Appendix K
Useful Formulas
Isothermal compressibility:
T =
1 V
V p
1
p
=
T,N

=
T
1 2p
2 2

T
Master thermodynamic equation:

dF = S dT p dV M dH +
i dNi
Gibbs-Helmholtz equation:
E(T, V, N ) =
(F/T )
(1/T )

=
V,N
ln Z

V,N
Free energy from partition function:

F (T, V, N ) = kB T ln Z(T, V, N )
Classical pure point-particle partition function:
Z(T, V, N ) =
Z(T, V, N ) =
1
3N
(T )
3N
h3N N !
Z
p
d3N r eU (r)/kB T
d3N r eH(r,p)/kB T
where
Quantal ideal gases (+ for fermions, for bosons):

(, ) =
M
Y
[1 e(r ) ]1
r=1
hnr i =
1
e(r ) 1
249
(T ) =
h
2mkB T
Index
adiabatic compressibility, 79, 91
analogy
booty argument, 27, 49
correlations and peer pressure, 189
entropy as disorder, 34
entropy as freedom, 35
fluctuation-susceptibility as cows vs. goats, 101
Jim Smith effect, 102
poker paradox, 34, 196
susceptibility as herd instinct, 101
temperature and cash incentives, 103, 196
temperature as persuasion, 103
typical vs. average, 33
approximation
controlled, 126
uncontrolled, 126
asteroids, 16
conceptual definition, 49
configuration work, 49
controlled approximation, 126
couple, 120
cows
compared to goats, 101
critical point, 26
crossover temperature, 123
baum, 225
blackbody radiation, 86
Boltzmann constant, 19
Boltzmann factor, 95
boson, 134
de Broglie wavelength, 119

decouple, 120
definition
conceptual, 49
operational, 49
degrees of freedom, 121
densities, 55
density of levels, 146
density of levels in k-space, 146
derivative
as an experiment, 65
dispersion, 100
dispersion relation, 166
dissipative work, 49
Carnot cycle, 54
cavity radiation, 86
characteristic equation, 238
characteristic matrix, 238
chemical potential, 4446, 69, 78, 79, 104, 147, 163
Clausius-Clapeyron equation, 80
compressibility
adiabatic, 79, 91
isothermal, 8, 56, 7879, 110, 130, 145, 171
eigenvalues, 234, 236

eigenvectors, 234, 236
energy density of levels, 146
ensemble, 16
ensembles, summary of, 106
enthalpy, 66
entropy, 1946, 57, 109, 145
of monatomic ideal gas, 25
equilibrium constant, 85
250
INDEX
equipartition theorem, 121
generalized, 123
escaping tendency, 45
Euler-MacLaurin formula, 128, 218
evolution, 9, 92, 93
expansion and compression, 8
extensive quantities, 55, 72
fermion, 134
flat Earth, 108
fluctuation-susceptibility theorem, 101
freedom, degrees of, 121
freeze out, 123
gamma, 74, 79
Gibbs free energy, 66
Gibbs-Duhem equation, 68
Gibbs-Helmholtz equation, 64
goats
compared to cows, 101
gross national product, 24
heat, 49
heat capacity, 8, 56, 58, 7778, 121129, 153
internal, 121
heat capacity ratio, 74, 79
ideal, 21
ideal gas, 8, 2126, 46, 58, 70, 74, 75, 77, 78, 91, 109,
110, 118164, 176, 179
ideal paramagnet, 20, 25, 46, 47
identical particles, 18
information theory, 109
intensive quantities, 24, 55, 72
interchange rule, 133, 162
internal heat capacity, 121
internal specific heat, 121
Ising model, 20, 25, 46, 47, 111, 183208
isothermal compressibility, 8, 7879, 110, 130, 145,
171
Jim Smith effect, 102
251
knowledge, 26
law of mass action, 85
Legendre transformation, 61, 64
level (vs. state), 135
light, thermodynamics of, 8690
macroscopic, 11
magnetic systems, 20, 25, 46, 47, 71, 78, 92, 111, 168,
183208
master equation, 61
master function, 61
mathematical functions, 20, 210224
mathematical rigor, 38, 104
matrix mathematics, 225245
microscopic, 11
misconceptions, catalog of, 246
mixtures, 69
monatomic, 21
monkeys, 9
most probable value, 41, 105
mountain climbing, 90
never, 9
number densities, 55
O notation, 124127
operational definition, 49
orbital, 135
orthogonal matrix, 227
outer product, 231
parametric differentiation, 100, 112, 143, 191
partial derivative
as an experiment, 65
partition function, 95
quantal, 107
perturbation theory, 26
poker, 18
politics, 108
polymer, 113
pressure, 6, 43, 162
252
problem solving tips
check dimensions, 19, 23, 56, 127
check extensivity, 56
check limiting behavior, 128130, 144
check special cases, 171
debrief, 24, 119
distinguish variables and functions, 55, 60
list expectations, 22
poor strategy, 71
preview, 22
use scaled quantities, 180
write variable list, 60
INDEX
thermal de Broglie wavelength, 45, 119
thermodynamic limit, 24, 98, 101, 102
thoughtless invertibility, principle of, 64
trace, 109
trading in variables, 64
tree, 225
Truesdell, Clifford, 64
uncontrolled approximation, 126
uncoupled, 120
universe
age of, 9, 194
size of, 194
quantities
extensive, 55
intensive, 55
virial theorem, 124

volume densities, 55
radiation, thermodynamics of, 8690

random walk, 113
relativistic gas, 123
rotation matrix, 226
Weins law, 88
work
configuration, 49
dissipative, 49
Zustandsumme, 95
Sackur-Tetrode formula, 25, 96
say it, dont believe it, 108, 140
shorthand, 108, 140
slick trick, 100, 112, 113, 122, 143, 146, 191
Smith effect, 102
sound, 91
specific heat, 57
internal, 121
sphere, volume of, 213
spin, 134
spin systems, 20, 25, 46, 47, 71, 78, 92, 111, 168,
183208
stars, 148
state (vs. level), 135
Stefan-Boltzmann law, 87
sum over all states, 95
susceptibility, 8, 111
temperature, 40, 48

Styer Thermo

Uploaded by

Copyright:

Available Formats

Styer Thermo

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Styer Thermo

Uploaded by

Copyright:

Available Formats

d3rB

Although, as a matter of history, statistical mechanics owes its origin

1 The Properties of Matter in Bulk

What is Statistical Mechanics About? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Principles of Statistical Mechanics

Microscopic Description of a Classical System . . . . . . . . . . . . . . . . . . . . . . . . . . .

Macroscopic Description of a Large Equilibrium System . . . . . . . . . . . . . . . . . . . . .

Statistical Definition of Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Entropy of a Monatomic Ideal Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Qualitative Features of Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Using Entropy to Find (Define) Temperature and Pressure . . . . . . . . . . . . . . . . . . . .

Heat and Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

The Thermodynamic Dance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Thermodynamics Applied to Fluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Thermodynamics Applied to Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . .

3.10 Thermodynamics Applied to Chemical Reactions . . . . . . . . . . . . . . . . . . . . . . . . .

3.11 Thermodynamics Applied to Light . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.12 Additional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

The Canonical Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Meaning of the Term Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Classical Monatomic Ideal Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Energy Dispersion in the Canonical Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . .

Temperature as a Control Variable for Energy (Canonical Ensemble) . . . . . . . . . . . . . . 102

The Equivalence of Canonical and Microcanonical Ensembles . . . . . . . . . . . . . . . . . . 103

The Grand Canonical Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

The Grand Canonical Ensemble in the Thermodynamic Limit . . . . . . . . . . . . . . . . . . 104

Summary of Major Ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.10 Quantal Statistical Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

Classical Monatomic Ideal Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

Classical Diatomic Ideal Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

Heat Capacity of an Ideal Gas

Specific Heat of a Hetero-nuclear Diatomic Ideal Gas . . . . . . . . . . . . . . . . . . . . . . . 126

Chemical Reactions Between Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

The Interchange Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Quantum Mechanics of Independent Identical Particles . . . . . . . . . . . . . . . . . . . . . . 134

Statistical Mechanics of Independent Identical Particles . . . . . . . . . . . . . . . . . . . . . 141

Quantum Mechanics of Free Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

Fermi-Dirac Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

Bose-Einstein Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

Specific Heat of the Ideal Fermion Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

Additional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

7 Harmonic Lattice Vibrations

The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

Statistical Mechanics of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

Normal Modes for a One-dimensional Chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

Normal Modes in Three Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

Low-temperature Heat Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

More Realistic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

What is a Phonon? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

Additional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

8 Interacting Classical Fluids

Perturbation Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

Variational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176

Distribution Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Correlations and Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

The Hard Sphere Fluid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

9 Strongly Interacting Systems and Phase Transitions

Introduction to Magnetic Systems and Models

Free Energy of the One-Dimensional Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . 183

The Mean-Field Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188