Lecture Notes on Thermodynamics and Statistical Mechanics

(A Work in Progress)

May 29, 2018

Contents

Contents i

List of Tables xiv

List of Figures xv

0.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

0.2 General references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1 Fundamentals of Probability 3

1.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Statistical Properties of Random Walks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.1 One-dimensional random walk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.2 Thermodynamic limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2.3 Entropy and energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.3 Basic Concepts in Probability Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.3.1 Fundamental definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.3.2 Bayesian statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.3.3 Random variables and their averages . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.4 Entropy and Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.4.1 Entropy and information theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.4.2 Probability distributions from maximum entropy . . . . . . . . . . . . . . . . . . 13

1.4.3 Continuous probability distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 17

i
ii CONTENTS

1.5 General Aspects of Probability Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.5.1 Discrete and continuous distributions . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5.2 Central limit theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5.3 Moments and cumulants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.4 Multidimensional Gaussian integral . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.6 Bayesian Statistical Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.6.1 Frequentists and Bayesians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.6.2 Updating Bayesian priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.6.3 Hyperparameters and conjugate priors . . . . . . . . . . . . . . . . . . . . . . . . 25
1.6.4 The problem with priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2 Thermodynamics 29
2.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 What is Thermodynamics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.1 Thermodynamic systems and state variables . . . . . . . . . . . . . . . . . . . . . 30
2.2.2 Heat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.3 Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.4 Pressure and Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.5 Standard temperature and pressure . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 The Zeroth Law of Thermodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 Mathematical Interlude : Exact and Inexact Differentials . . . . . . . . . . . . . . . . . . 37
2.5 The First Law of Thermodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.1 Conservation of energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.2 Single component systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.3 Ideal gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.5.4 Adiabatic transformations of ideal gases . . . . . . . . . . . . . . . . . . . . . . . 44
2.5.5 Adiabatic free expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.6 Heat Engines and the Second Law of Thermodynamics . . . . . . . . . . . . . . . . . . . 47
2.6.1 There’s no free lunch so quit asking . . . . . . . . . . . . . . . . . . . . . . . . . . 47
CONTENTS iii

2.6.2 Engines and refrigerators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.6.3 Nothing beats a Carnot engine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.6.4 The Carnot cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

2.6.5 The Stirling cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

2.6.6 The Otto and Diesel cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

2.6.7 The Joule-Brayton cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

2.6.8 Carnot engine at maximum power output . . . . . . . . . . . . . . . . . . . . . . . 58

2.7 The Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

2.7.1 Entropy and heat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

2.7.2 The Third Law of Thermodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 60

2.7.3 Entropy changes in cyclic processes . . . . . . . . . . . . . . . . . . . . . . . . . . 61

2.7.4 Gibbs-Duhem relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

2.7.5 Entropy for an ideal gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

2.7.6 Example system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

2.7.7 Measuring the entropy of a substance . . . . . . . . . . . . . . . . . . . . . . . . . 66

2.8 Thermodynamic Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

2.8.1 Energy E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

2.8.2 Helmholtz free energy F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

2.8.3 Enthalpy H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

2.8.4 Gibbs free energy G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

2.8.5 Grand potential Ω . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

2.9 Maxwell Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

2.9.1 Relations deriving from E(S, V, N ) . . . . . . . . . . . . . . . . . . . . . . . . . . 70

2.9.2 Relations deriving from F (T, V, N ) . . . . . . . . . . . . . . . . . . . . . . . . . . 70

2.9.3 Relations deriving from H(S, p, N ) . . . . . . . . . . . . . . . . . . . . . . . . . . 71

2.9.4 Relations deriving from G(T, p, N ) . . . . . . . . . . . . . . . . . . . . . . . . . . 71

2.9.5 Relations deriving from Ω(T, V, µ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

2.9.6 Generalized thermodynamic potentials . . . . . . . . . . . . . . . . . . . . . . . . 73

iv CONTENTS

2.10 Equilibrium and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

2.10.1 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

2.10.2 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

2.11 Applications of Thermodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

2.11.1 Adiabatic free expansion revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

2.11.2 Energy and volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

2.11.3 van der Waals equation of state . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

2.11.4 Thermodynamic response functions . . . . . . . . . . . . . . . . . . . . . . . . . . 80

2.11.5 Joule effect: free expansion of a gas . . . . . . . . . . . . . . . . . . . . . . . . . . 83

2.11.6 Throttling: the Joule-Thompson effect . . . . . . . . . . . . . . . . . . . . . . . . 84

2.12 Phase Transitions and Phase Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

2.12.1 p-v-T surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

2.12.2 The Clausius-Clapeyron relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

2.12.3 Liquid-solid line in H2 O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

2.12.4 Slow melting of ice : a quasistatic but irreversible process . . . . . . . . . . . . . . 92

2.12.5 Gibbs phase rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

2.13 Entropy of Mixing and the Gibbs Paradox . . . . . . . . . . . . . . . . . . . . . . . . . . 96

2.13.1 Computing the entropy of mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

2.13.2 Entropy and combinatorics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

2.13.3 Weak solutions and osmotic pressure . . . . . . . . . . . . . . . . . . . . . . . . . 99

2.13.4 Effect of impurities on boiling and freezing points . . . . . . . . . . . . . . . . . . 101

2.13.5 Binary solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

2.14 Some Concepts in Thermochemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

2.14.1 Chemical reactions and the law of mass action . . . . . . . . . . . . . . . . . . . . 111

2.14.2 Enthalpy of formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

2.14.3 Bond enthalpies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

2.15 Appendix I : Integrating Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

2.16 Appendix II : Legendre Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

CONTENTS v

2.17 Appendix III : Useful Mathematical Relations . . . . . . . . . . . . . . . . . . . . . . . . 122

3 Ergodicity and the Approach to Equilibrium 127

3.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

3.2 Modeling the Approach to Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

3.2.1 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

3.2.2 The Master Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

3.2.3 Equilibrium distribution and detailed balance . . . . . . . . . . . . . . . . . . . . 129

3.2.4 Boltzmann’s H-theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

3.3 Phase Flows in Classical Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

3.3.1 Hamiltonian evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

3.3.2 Dynamical systems and the evolution of phase space volumes . . . . . . . . . . . 132

3.3.3 Liouville’s equation and the microcanonical distribution . . . . . . . . . . . . . . 135

3.4 Irreversibility and Poincaré Recurrence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

3.4.1 Poincaré recurrence theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

3.4.2 Kac ring model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

3.5 Remarks on Ergodic Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

3.5.1 Definition of ergodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

3.5.2 The microcanonical ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

3.5.3 Ergodicity and mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

3.6 Thermalization of Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

3.6.1 Quantum dephasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

3.6.2 Eigenstate thermalization hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . 150

3.6.3 When is the ETH true? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

3.7 Appendix I : Formal Solution of the Master Equation . . . . . . . . . . . . . . . . . . . . 152

3.8 Appendix II : Radioactive Decay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

3.9 Appendix III: Transition to Ergodicity in a Simple Model . . . . . . . . . . . . . . . . . . 155

4 Statistical Ensembles 165

vi CONTENTS

4.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

4.2 Microcanonical Ensemble (µCE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

4.2.1 The microcanonical distribution function . . . . . . . . . . . . . . . . . . . . . . . 166

4.2.2 Density of states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

4.2.3 Arbitrariness in the definition of S(E) . . . . . . . . . . . . . . . . . . . . . . . . 170

4.2.4 Ultra-relativistic ideal gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

4.2.5 Discrete systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

4.3 The Quantum Mechanical Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

4.3.1 The density matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

4.3.2 Averaging the DOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

4.3.3 Coherent states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

4.4 Thermal Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

4.4.1 Two systems in thermal contact . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

4.4.2 Thermal, mechanical and chemical equilibrium . . . . . . . . . . . . . . . . . . . . 177

4.4.3 Gibbs-Duhem relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

4.5 Ordinary Canonical Ensemble (OCE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178

4.5.1 Canonical distribution and partition function . . . . . . . . . . . . . . . . . . . . . 178

4.5.2 The difference between P (En ) and Pn . . . . . . . . . . . . . . . . . . . . . . . . . 179

4.5.3 Additional remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

4.5.4 Averages within the OCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

4.5.5 Entropy and free energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

4.5.6 Fluctuations in the OCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

4.5.7 Thermodynamics revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

4.5.8 Generalized susceptibilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

4.6 Grand Canonical Ensemble (GCE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

4.6.1 Grand canonical distribution and partition function . . . . . . . . . . . . . . . . . 185

4.6.2 Entropy and Gibbs-Duhem relation . . . . . . . . . . . . . . . . . . . . . . . . . . 186

4.6.3 Generalized susceptibilities in the GCE . . . . . . . . . . . . . . . . . . . . . . . . 187

CONTENTS vii

4.6.4 Fluctuations in the GCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

4.6.5 Gibbs ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

4.7 Statistical Ensembles from Maximum Entropy . . . . . . . . . . . . . . . . . . . . . . . . 189

4.7.1 µCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

4.7.2 OCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

4.7.3 GCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

4.8 Ideal Gas Statistical Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

4.8.1 Maxwell velocity distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

4.8.2 Equipartition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

4.9 Selected Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

4.9.1 Spins in an external magnetic field . . . . . . . . . . . . . . . . . . . . . . . . . . 195

4.9.2 Negative temperature (!) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

4.9.3 Adsorption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

4.9.4 Elasticity of wool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

4.9.5 Noninteracting spin dimers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

4.10 Statistical Mechanics of Molecular Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . 202

4.10.1 Separation of translational and internal degrees of freedom . . . . . . . . . . . . . 202

4.10.2 Ideal gas law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

4.10.3 The internal coordinate partition function . . . . . . . . . . . . . . . . . . . . . . 204

4.10.4 Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

4.10.5 Vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207

4.10.6 Two-level systems : Schottky anomaly . . . . . . . . . . . . . . . . . . . . . . . . 207

4.10.7 Electronic and nuclear excitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

4.11 Appendix I : Additional Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

4.11.1 Three state system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

4.11.2 Spins and vacancies on a surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

4.11.3 Fluctuating interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

4.11.4 Dissociation of molecular hydrogen . . . . . . . . . . . . . . . . . . . . . . . . . . 216

viii CONTENTS

5 Noninteracting Quantum Systems 219

5.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

5.2 Statistical Mechanics of Noninteracting Quantum Systems . . . . . . . . . . . . . . . . . . 220

5.2.1 Bose and Fermi systems in the grand canonical ensemble . . . . . . . . . . . . . . 220

5.2.2 Quantum statistics and the Maxwell-Boltzmann limit . . . . . . . . . . . . . . . . 221

5.2.3 Single particle density of states . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

5.3 Quantum Ideal Gases : Low Density Expansions . . . . . . . . . . . . . . . . . . . . . . . 224

5.3.1 Expansion in powers of the fugacity . . . . . . . . . . . . . . . . . . . . . . . . . . 224

5.3.2 Virial expansion of the equation of state . . . . . . . . . . . . . . . . . . . . . . . 225

5.3.3 Ballistic dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

5.4 Entropy and Counting States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

5.5 Photon Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

5.5.1 Thermodynamics of the photon gas . . . . . . . . . . . . . . . . . . . . . . . . . . 229

5.5.2 Classical arguments for the photon gas . . . . . . . . . . . . . . . . . . . . . . . . 231

5.5.3 Surface temperature of the earth . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232

5.5.4 Distribution of blackbody radiation . . . . . . . . . . . . . . . . . . . . . . . . . . 233

5.5.5 What if the sun emitted ferromagnetic spin waves? . . . . . . . . . . . . . . . . . 234

5.6 Lattice Vibrations : Einstein and Debye Models . . . . . . . . . . . . . . . . . . . . . . . 235

5.6.1 One-dimensional chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

5.6.2 General theory of lattice vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . 237

5.6.3 Einstein and Debye models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240

5.6.4 Melting and the Lindemann criterion . . . . . . . . . . . . . . . . . . . . . . . . . 242

5.6.5 Goldstone bosons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

5.7 The Ideal Bose Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

5.7.1 General formulation for noninteracting systems . . . . . . . . . . . . . . . . . . . 247

5.7.2 Ballistic dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

5.7.3 Isotherms for the ideal Bose gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

5.7.4 The λ-transition in Liquid 4 He . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

CONTENTS ix

5.7.5 Fountain effect in superfluid 4 He . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

5.7.6 Bose condensation in optical traps . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
5.8 The Ideal Fermi Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
5.8.1 Grand potential and particle number . . . . . . . . . . . . . . . . . . . . . . . . . 259
5.8.2 The Fermi distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
5.8.3 T = 0 and the Fermi surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
5.8.4 The Sommerfeld expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
5.8.5 Magnetic susceptibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
5.8.6 Moment formation in interacting itinerant electron systems . . . . . . . . . . . . . 270
5.8.7 White dwarf stars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
5.9 Appendix I : Second Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
5.9.1 Basis states and creation/annihilation operators . . . . . . . . . . . . . . . . . . . 281
5.9.2 Second quantized operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
5.10 Appendix II : Ideal Bose Gas Condensation . . . . . . . . . . . . . . . . . . . . . . . . . . 284
5.11 Appendix III : Example Bose Condensation Problem . . . . . . . . . . . . . . . . . . . . . 286
Chapter 1

Fundamentals of Probability

1.1 References

– C. Gardiner, Stochastic Methods (4th edition, Springer-Verlag, 2010)

Very clear and complete text on stochastic methods with many applications.

– J. M. Bernardo and A. F. M. Smith, Bayesian Theory (Wiley, 2000)

A thorough textbook on Bayesian methods.

– D. Williams, Weighing the Odds: A Course in Probability and Statistics (Cambridge, 2001)
A good overall statistics textbook, according to a mathematician colleague.

– E. T. Jaynes, Probability Theory (Cambridge, 2007)

An extensive, descriptive, and highly opinionated presentation, with a strongly Bayesian approach.

– A. N. Kolmogorov, Foundations of the Theory of Probability (Chelsea, 1956)

The Urtext of mathematical probability theory.

3
4 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

1.2 Statistical Properties of Random Walks

1.2.1 One-dimensional random walk

Consider the mechanical system depicted in Fig. 1.1, a version of which is often sold in novelty shops. A
ball is released from the top, which cascades consecutively through N levels. The details of each ball’s
motion are governed by Newton’s laws of motion. However, to predict where any given ball will end up in
the bottom row is difficult, because the ball’s trajectory depends sensitively on its initial conditions, and
may even be influenced by random vibrations of the entire apparatus. We therefore abandon all hope of
integrating the equations of motion and treat the system statistically. That is, we assume, at each level,
that the ball moves to the right with probability p and to the left with probability q = 1 − p. If there is
no bias in the system, then p = q = 12 . The position XN after N steps may be written
N
X
X= σj , (1.1)
j=1

where σj = +1 if the ball moves to the right at level j, and σj = −1 if the ball moves to the left at level
j. At each level, the probability for these two outcomes is given by
(
p if σ = +1
Pσ = p δσ,+1 + q δσ,−1 = (1.2)
q if σ = −1 .

This is a normalized discrete probability distribution of the type discussed in section 1.5 below. The
multivariate distribution for all the steps is then
N
Y
P (σ1 , . . . , σN ) = P (σj ) . (1.3)
j=1

Our system is equivalent to a one-dimensional random walk . Imagine an inebriated pedestrian on a

sidewalk taking steps to the right and left at random. After N steps, the pedestrian’s location is X.
Now let’s compute the average of X:
N

X  X
hXi = σj = N hσi = N σ P (σ) = N (p − q) = N (2p − 1) . (1.4)
j=1 σ=±1

This could be identified as an equation of state for our system, as it relates a measurable quantity X to
the number of steps N and the local bias p. Next, let’s compute the average of X 2 :
N X
X N
2
hX i = hσj σj 0 i = N 2 (p − q)2 + 4N pq . (1.5)
j=1 j 0 =1

Here we have used (

2 1 if j = j 0
hσj σj 0 i = δjj 0 + 1 − δjj 0 (p − q) = (1.6)
(p − q)2 6 j0 .
if j =
1.2. STATISTICAL PROPERTIES OF RANDOM WALKS 5

Figure 1.1: The falling ball system, which mimics a one-dimensional random walk.

Note that hX 2 i ≥ hXi2 , which must be so because

2 
Var(X) = h(∆X)2 i ≡ X − hXi = hX 2 i − hXi2 .

(1.7)
This is called the variance of X. We have Var(X)p = 4N p q. The root mean square deviation, ∆Xrms ,
is the square root of the variance: ∆Xrms = Var(X). Note that the mean value of X is linearly
proportional to N 1 , but the RMS fluctuations ∆Xrms are proportional to N 1/2 . In the limit N → ∞
then, the ratio ∆Xrms /hXi vanishes as N −1/2 . This is a consequence of the central limit theorem (see
§1.5.2 below), and we shall meet up with it again on several occasions.
We can do even better. We can find the complete probability distribution for X. It is given by
 
N
PN,X = pNR q NL , (1.8)
NR
where NR/L are the numbers of steps taken to the right/left, with N = NR + NL , and X = NR − NL .
There are many independent ways to take NR steps to the right. For example, our first NR steps could
all be to the right, and the remaining NL = N − NR steps would then all be to the left. Or our final NR
steps could all be to the right. For each of these independent possibilities, the probability is pNR q NL .
How many possibilities are there? Elementary combinatorics tells us this number is
 
N N!
= . (1.9)
NR NR ! NL !
Note that N ± X = 2NR/L , so we can replace NR/L = 12 (N ± X). Thus,
N!
PN,X = N +X

N −X
p(N +X)/2 q (N −X)/2 . (1.10)
2 ! 2 !
1
The exception is the unbiased case p = q = 12 , where hXi = 0.
6 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

Figure 1.2: Comparison of exact distribution of eqn. 1.10 (red squares) with the Gaussian distribution
of eqn. 1.19 (blue line).

1.2.2 Thermodynamic limit

Consider the limit N → ∞ but with x ≡ X/N finite. This is analogous to what is called the thermody-
namic limit in statistical mechanics. Since N is large, x may be considered a continuous variable. We
evaluate ln PN,X using Stirling’s asymptotic expansion

ln N ! ' N ln N − N + O(ln N ) . (1.11)

We then have
h i
ln PN,X ' N ln N − N − 12 N (1 + x) ln 12 N (1 + x) + 12 N (1 + x)
h i
− 12 N (1 − x) ln 12 N (1 − x) + 21 N (1 − x) + 21 N (1 + x) ln p + 12 N (1 − x) ln q (1.12)
h
i h i
= −N 1+x 1+x 1−x 1−x 1+x 1−x






2 ln 2 + 2 ln 2 + N 2 ln p + 2 ln q .

Notice that the terms proportional to N ln N have all cancelled, leaving us with a quantity which is linear
in N . We may therefore write ln PN,X = −N f (x) + O(ln N ), where
h
i h 1+x
i
f (x) = 1+x 1+x 1−x 1−x 1−x





2 ln 2 + 2 ln 2 − 2 ln p + 2 ln q . (1.13)

We have just shown that in the large N limit we may write

PN,X = C e−N f (X/N ) , (1.14)

where C is a normalization constant2 . Since N is by assumption large, the function PN,X is dominated
by the minimum (or minima) of f (x), where the probability is maximized. To find the minimum of f (x),
2
The origin of C lies in the O(ln N ) and O(N 0 ) terms in the asymptotic expansion of ln N !. We have ignored these terms
here. Accounting for them carefully reproduces the correct value of C in eqn. 1.20.
1.2. STATISTICAL PROPERTIES OF RANDOM WALKS 7

we set f 0 (x) = 0, where  

0 1 q 1+x
f (x) = 2 ln · . (1.15)
p 1−x
Setting f 0 (x) = 0, we obtain
1+x p
= ⇒ x̄ = p − q . (1.16)
1−x q
We also have
1
f 00 (x) = , (1.17)
1 − x2
so invoking Taylor’s theorem,

f (x) = f (x̄) + 21 f 00 (x̄) (x − x̄)2 + . . . . (1.18)

Putting it all together, we have

" # " #
N (x − x̄)2 (X − X̄)2
PN,X ≈ C exp − = C exp − , (1.19)
8pq 8N pq

where X̄ = hXi = N (p − q) = N x̄. The constant C is determined by the normalization condition,

∞ Z∞ " #
X
1 (X − X̄)2 p
PN,X ≈ 2 dX C exp − = 2πN pq C , (1.20)
8N pq
X=−∞ −∞
√
and thus C = 1/ 2πN pq. Why don’t we go beyond second order in the Taylor expansion of f (x)? We
will find out in §1.5.2 below.

1.2.3 Entropy and energy

The function f (x) can be written as a sum of two contributions, f (x) = e(x) − s(x), where

s(x) = − 1+x ln 1+x − 1−x ln 1−x

2 2 2 2
(1.21)
e(x) = − 21 ln(pq) − 12 x ln(p/q) .

The function S(N, x) ≡ N s(x) is analogous to the statistical entropy of our system3 . We have
   
N N
S(N, x) = N s(x) = ln = ln 1 . (1.22)
NR 2 N (1 + x)

Thus, the statistical entropy is the logarithm of the number of ways the system can be configured so as to
yield the same value of X (at fixed N ). The second contribution to f (x) is the energy term. We write

E(N, x) = N e(x) = − 21 N ln(pq) − 12 N x ln(p/q) . (1.23)

The energy term biases the probability PN,X = exp(S − E) so that low energy configurations are more
probable than high energy configurations. For our system, we see that when p < q (i.e. p < 21 ), the energy
3
The function s(x) is the specific entropy.
8 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

is minimized by taking x as small as possible (meaning as negative as possible). The smallest possible
allowed value of x = X/N is x = −1. Conversely, when p > q (i.e. p > 21 ), the energy is minimized by
taking x as large as possible, which means x = 1. The average value of x, as we have computed explicitly,
is x̄ = p − q = 2p − 1, which falls somewhere in between these two extremes.
In actual thermodynamic systems, entropy and energy are not dimensionless. What we have called S
here is really S/kB , which is the entropy in units of Boltzmann’s constant. And what we have called E
here is really E/kB T , which is energy in units of Boltzmann’s constant times temperature.

1.3 Basic Concepts in Probability Theory

Here we recite the basics of probability theory.

1.3.1 Fundamental definitions

The natural mathematical setting is set theory. Sets are generalized collections of objects. The basics:
ω ∈ A is a binary relation which says that the object ω is an element of the set A. Another binary
relation is set inclusion. If all members of A are in B, we write A ⊆ B. The union of sets A and B is
denoted A ∪ B and the intersection of A and B is denoted A ∩ B. The Cartesian product of A and B,
denoted A × B, is the set of all ordered elements (a, b) where a ∈ A and b ∈ B.
Some details: If ω is not in A, we write ω ∈ / A. Sets may also be objects, so we may speak of sets of
sets, but typically the sets which will concern us are simple discrete collections of numbers, such as the
possible rolls of a die {1,2,3,4,5,6}, or the real numbers R, or Cartesian products such as RN . If A ⊆ B
but A 6= B, we say that A is a proper subset of B and write A ⊂ B. Another binary operation is the set
difference A\B, which contains all ω such that ω ∈ A and ω ∈ / B.
In probability theory, each object ω is identified as an event. We denote by Ω the set of all events, and
∅ denotes the set of no events. There are three basic axioms of probability:

i) To each set A is associated a non-negative real number P (A), which is called the probability of A.

ii) P (Ω) = 1.

iii) If {Ai } is a collection of disjoint sets, i.e. if Ai ∩ Aj = ∅ for all i 6= j, then

[  X
P Ai = P (Ai ) . (1.24)
i i

From these axioms follow a number of conclusions. Among them, let ¬A = Ω\A be the complement of
A, i.e. the set of all events not in A. Then since A ∪ ¬A = Ω, we have P (¬A) = 1 − P (A). Taking A = Ω,
we conclude P (∅) = 0.
The meaning of P (A) is that if events ω are chosen from Ω at random, then the relative frequency for
ω ∈ A approaches P (A) as the number of trials tends to infinity. But what do we mean by ’at random’ ?
1.3. BASIC CONCEPTS IN PROBABILITY THEORY 9

One meaning we can impart to the notion of randomness is that a process is random if its outcomes
can be accurately modeled using the axioms of probability. This entails the identification of a probability
space Ω as well as a probability measure P . For example, in the microcanonical ensemble of classical
statistical physics, the space Ω is the Q
collection of phase space points ϕ = {q1 , . . . , qn , p1 , . . . , pn } and the
−1 n

probability measure is dµ = Σ (E) i=1 dqi dpi δ E − H(q, p) , so that for A ∈ Ω the probability of A
is P (A) = dµ χA (ϕ), where χA (ϕ) = 1 if ϕ ∈ A and χA (ϕ)
R
R = 0 if ϕ ∈
/ A is the characteristic function
of A. The quantity Σ(E) is determined by normalization: dµ = 1.

1.3.2 Bayesian statistics

We now introduce two additional probabilities. The joint probability for sets A and B together is written
P (A ∩ B). That is, P (A ∩ B) = Prob[ω ∈ A and ω ∈ B]. For example, A might denote the set of all
politicians, B the set of all American citizens, and C the set of all living humans with an IQ greater than
60. Then A ∩ B would be the set of all politicians who are also American citizens, etc. Exercise: estimate
P (A ∩ B ∩ C).
The conditional probability of B given A is written P (B|A). We can compute the joint probability
P (A ∩ B) = P (B ∩ A) in two ways:

P (A ∩ B) = P (A|B) · P (B) = P (B|A) · P (A) . (1.25)

Thus,
P (B|A) P (A)
P (A|B) = , (1.26)
P (B)
a result known as Bayes’ theorem. Now suppose the ‘event space’ is partitioned as {Ai }. Then
X
P (B) = P (B|Ai ) P (Ai ) . (1.27)
i

We then have
P (B|Ai ) P (Ai )
P (Ai |B) = P , (1.28)
j P (B|Aj ) P (Aj )

a result sometimes known as the extended form of Bayes’ theorem. When the event space is a ‘binary
partition’ {A, ¬A}, we have

P (B|A) P (A)
P (A|B) = . (1.29)
P (B|A) P (A) + P (B|¬A) P (¬A)

Note that P (A|B) + P (¬A|B) = 1 (which follows from ¬¬A = A).

As an example, consider the following problem in epidemiology. Suppose there is a rare but highly
contagious disease A which occurs in 0.01% of the general population. Suppose further that there is
a simple test for the disease which is accurate 99.99% of the time. That is, out of every 10,000 tests,
the correct answer is returned 9,999 times, and the incorrect answer is returned only once. Now let us
administer the test to a large group of people from the general population. Those who test positive are
quarantined. Question: what is the probability that someone chosen at random from the quarantine
10 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

group actually has the disease? We use Bayes’ theorem with the binary partition {A, ¬A}. Let B denote
the event that an individual tests positive. Anyone from the quarantine group has tested positive. Given
this datum, we want to know the probability that that person has the disease. That is, we want P (A|B).
Applying eqn. 1.29 with

P (A) = 0.0001 , P (¬A) = 0.9999 , P (B|A) = 0.9999 , P (B|¬A) = 0.0001 ,

we find P (A|B) = 21 . That is, there is only a 50% chance that someone who tested positive actually has
the disease, despite the test being 99.99% accurate! The reason is that, given the rarity of the disease in
the general population, the number of false positives is statistically equal to the number of true positives.
In the above example, we had P (B|A) + P (B|¬A) = 1, but this is not generally the case. What is true
instead is P (B|A) + P (¬B|A) = 1. Epidemiologists define the sensitivity of a binary classification test as
the fraction of actual positives which are correctly identified, and the specificity as the fraction of actual
negatives that are correctly identified. Thus, se = P (B|A) is the sensitivity and sp = P (¬B|¬A) is the
specificity. We then have P (B|¬A) = 1 − P (¬B|¬A). Therefore,

P (B|A) + P (B|¬A) = 1 + P (B|A) − P (¬B|¬A) = 1 + se − sp . (1.30)

In our previous example, se = sp = 0.9999, in which case the RHS above gives 1. In general, if P (A) ≡ f
is the fraction of the population which is afflicted, then
f · se
P (infected | positive) = . (1.31)
f · se + (1 − f ) · (1 − sp)

For continuous distributions, we speak of a probability density. We then have

Z
P (y) = dx P (y|x) P (x) (1.32)

and
P (y|x) P (x)
P (x|y) = R . (1.33)
dx0 P (y|x0 ) P (x0 )
The range of integration may depend on the specific application.
The quantities P (Ai ) are called the prior distribution. Clearly in order to compute P (B) or P (Ai |B)
we must know the priors, and this is usually the weakest link in the Bayesian chain of reasoning. If
our prior distribution is not accurate, Bayes’ theorem will generate incorrect results. One approach to
approximating prior probabilities P (Ai ) is to derive them from a maximum entropy construction.

1.3.3 Random variables and their averages

Consider an abstract probability space X whose elements (i.e. events) are labeled by x. The average of
any function f (x) is denoted as Ef or hf i, and is defined for discrete sets as
X
Ef = hf i = f (x) P (x) , (1.34)
x∈X
1.4. ENTROPY AND PROBABILITY 11

where P (x) is the probability of x. For continuous sets, we have

Z
Ef = hf i = dx f (x) P (x) . (1.35)
X

Typically for continuous sets we have X = R or X = R≥0 . Gardiner and other authors introduce an
extra symbol, X, to denote a random variable, with X(x) = x being its value. This is formally useful
but notationally confusing, so we’ll avoid it here and speak loosely of x as a random variable.
When there are two random variables x ∈ X and y ∈ Y, we have Ω = X × Y is the product space, and
XX
Ef (x, y) = hf (x, y)i = f (x, y) P (x, y) , (1.36)
x∈X y∈Y

with the obvious generalization to continuous sets. This generalizes to higher rank products, i.e. xi ∈ Xi
with i ∈ {1, . . . , N }. The covariance of xi and xj is defined as




Cij ≡ xi − hxi i xj − hxj i = hxi xj i − hxi ihxj i . (1.37)

If f (x) is a convex function then one has

Ef (x) ≥ f (Ex) . (1.38)

For continuous functions, f (x) is convex if f 00 (x) ≥ 0 everywhere4 . If f (x) is convex on some interval
[a, b] then for x1,2 ∈ [a, b] we must have


f λx1 + (1 − λ)x2 ≤ λf (x1 ) + (1 − λ)f (x2 ) , (1.39)

where λ ∈ [0, 1]. This is easily generalized to

X  X
f p n xn ≤ pn f (xn ) , (1.40)
n n

where pn = P (xn ), a result known as Jensen’s theorem.

1.4 Entropy and Probability

1.4.1 Entropy and information theory

It was shown in the classic 1948 work of Claude Shannon that entropy is in fact a measure of information 5 .
Suppose we observe that a particular event occurs with probability p. We associate with this observation
an amount of information I(p). The information I(p) should satisfy certain desiderata:
4
A function g(x) is concave if −g(x) is convex.
5
See ‘An Introduction to Information Theory and Entropy’ by T. Carter, Santa Fe Complex Systems Summer School,
June 2011. Available online at http://astarte.csustan.edu/$\sim$tom/SFI-CSSS/info-theory/info-lec.pdf.
12 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

1 Information is non-negative, i.e. I(p) ≥ 0.

2 If two events occur independently so their joint probability is p1 p2 , then their information is addi-
tive, i.e. I(p1 p2 ) = I(p1 ) + I(p2 ).

3 I(p) is a continuous function of p.

4 There is no information content to an event which is always observed, i.e. I(1) = 0.

From these four properties, it is easy to show that the only possible function I(p) is

I(p) = −A ln p , (1.41)

where A is an arbitrary constant that can be absorbed into the base of the logarithm, since logb x =
ln x/ ln b. We will take A = 1 and use e as the base, so I(p) = − ln p. Another common choice is to take
the base of the logarithm to be 2, so I(p) = − log2 p. In this latter case, the units of information are
known as bits. Note that I(0) = ∞. This means that the observation of an extremely rare event carries
a great deal of information6
Now suppose we have a set of events labeled by an integer n which occur with probabilities {pn }. What
is the expected amount of information in N observations? Since event n occurs an average of N pn times,
and the information content in pn is − ln pn , we have that the average information per observation is

hIN i X
S= =− pn ln pn , (1.42)
N n

which is known as the entropy of the distribution. Thus, maximizing S is equivalent to maximizing the
information content per observation.
Consider, for example, the information content of course grades. As we shall see, if the only constraint on
the probability distribution is that of overall normalization, then S is maximized when all the probabilities
pn are equal. The binary entropy is then S = log2 Γ , since pn = 1/Γ . Thus, for pass/fail grading, the
maximum average information per grade is − log2 ( 12 ) = log2 2 = 1 bit. If only A, B, C, D, and F grades are
assigned, then the maximum average information per grade is log2 5 = 2.32 bits. If we expand the grade
options to include {A+, A, A-, B+, B, B-, C+, C, C-, D, F}, then the maximum average information
per grade is log2 11 = 3.46 bits.
Equivalently, consider, following the discussion in vol. 1 of Kardar, a random sequence {n1 , n2 , . . . , nN }
where each element nj takes one of K possible values. There are then K N such possible sequences, and
to specify one of them requires log2 (K N ) = N log2 K bits of information. However, if the value n occurs
with probability pn , then on average it will occur Nn = N pn times in a sequence of length N , and the
total number of such sequences will be

N!
g(N ) = QK . (1.43)
n=1 Nn !
6
John McGreevy refers to I(p) as the surprise of observing an event which occurs with probability p. I like this very much.
1.4. ENTROPY AND PROBABILITY 13

In general, this is far less that the total possible number K N , and the number of bits necessary to specify
one from among these g(N ) possibilities is

K
X K
X
log2 g(N ) = log2 (N !) − log2 (Nn !) ≈ −N pn log2 pn , (1.44)
n=1 n=1

up to terms of order unity. Here we have invoked Stirling’s approximation. If the distribution is uniform,
then we have pn = K1 for all n ∈ {1, . . . , K}, and log2 g(N ) = N log2 K.

1.4.2 Probability distributions from maximum entropy

We have shown how one can proceed from a probability distribution and compute various averages. We
now seek to go in the other direction, and determine the full probability distribution based on a knowledge
of certain averages.
At first, this seems impossible. Suppose we want to reproduce the full probability distribution for an
N -step random walk from knowledge of the average hXi = (2p − 1)N , where p is the probability of
moving to the right at each step (see §1.2 above). The problem seems ridiculously underdetermined,
since there are 2N possible configurations for an N -step random walk: σj = ±1 for j = 1, . . . , N . Overall
normalization requires
X
P (σ1 , . . . , σN ) = 1 , (1.45)
{σj }

but this just imposes one constraint on the 2N probabilities P (σ1 , . . . , σN ), leaving 2N − 1 overall param-
eters. What principle allows us to reconstruct the full probability distribution

N
Y N

Y
P (σ1 , . . . , σN ) = p δσj ,1 + q δσj ,−1 = p(1+σj )/2 q (1−σj )/2 , (1.46)
j=1 j=1

corresponding to N independent steps?

The principle of maximum entropy

The entropy of a discrete probability distribution {pn } is defined as

X
S=− pn ln pn , (1.47)
n

where here we take e as the base of the logarithm.

The entropy may therefore be regarded as a function of
the probability distribution: S = S {pn } . One special
 Aproperty of the entropy is the following. Suppose
B

we have two independent normalized distributions pa and pb . The joint probability for events a
14 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

and b is then Pa,b = pA B

a pb . The entropy of the joint distribution is then
XX XX XX
pA B A B
pA B A B



S=− Pa,b ln Pa,b = − a pb ln pa pb = − a pb ln pa + ln pb
a b a b a b
X X X X X X
=− pA A
a ln pa · pB
b − pB B
b ln pb · pA
a =− pA A
a ln pa − pB B
b ln pb
a b b a a b
A B
=S +S .
Thus, the entropy of a joint distribution formed from two independent distributions is additive.
P
Suppose all we knew about {pn } was that it was normalized. Then n pn = 1. This is a constraint on
the values {pn }. Let us now extremize the entropy S with respect to the distribution {pn }, but subject
to the normalization constraint. We do this using Lagrange’s method of undetermined multipliers. We
define X X 
S ∗ {pn }, λ = −


pn ln pn − λ pn − 1 (1.48)
n n
and we freely extremize S∗ over all its arguments. Thus, for all n we have
∂S ∗

0= = − ln pn + 1 + λ
∂pn
∂S ∗ X (1.49)
0= = pn − 1 .
∂λ n

From the first of these equations, we obtain pn = e−(1+λ) , and from the second we obtain
X X
pn = e−(1+λ) · 1 = Γ e−(1+λ) , (1.50)
n n
P
where Γ ≡ n 1 is the total number of possible events. Thus, pn = 1/Γ , which says that all events are
equally probable.
P
Now suppose we know one other piece of information, which is the average value X = n Xn pn of some
quantity. We now extremize S subject to two constraints, and so we define
X X  X 
S ∗ {pn }, λ0 , λ1 = −


pn ln pn − λ0 pn − 1 − λ1 Xn pn − X . (1.51)
n n n

We then have
∂S ∗

= − ln pn + 1 + λ0 + λ1 Xn = 0 , (1.52)
∂pn
which yields the two-parameter distribution
pn = e−(1+λ0 ) e−λ1 Xn . (1.53)
P P
To fully determine the distribution {pn } we need to invoke the two equations n pn = 1 and n Xn pn =
X, which come from extremizing S ∗ with respect to λ0 and λ1 , respectively:
X
1 = e−(1+λ0 ) e−λ1 Xn
n
X (1.54)
−(1+λ0 )
X=e Xn e−λ1 Xn .
n
1.4. ENTROPY AND PROBABILITY 15

General formulation

The generalization to K extra pieces of information (plus normalization) is immediately apparent. We

have X
Xa = Xna pn , (1.55)
n

and therefore we define

X K
X X 
∗
Xna pn − X a ,


S {pn }, {λa } = − pn ln pn − λa (1.56)
n a=0 n

(a=0)
with Xn ≡ X (a=0) = 1. Then the optimal distribution which extremizes S subject to the K + 1
constraints is
K
( )
X
a
pn = exp − 1 − λa Xn
a=0
(1.57)
K
( )
1 X
= exp − λa Xna ,
Z
a=1

where Z = e1+λ0 is determined by normalization:

P
n pn = 1. This is a (K + 1)-parameter distribution,
with {λ0 , λ1 , . . . , λK } determined by the K + 1 constraints in eqn. 1.55.

Example

As an example, consider the random walk problem. We have two pieces of information:
X X
··· P (σ1 , . . . , σN ) = 1
σ1 σN

X X N
X (1.58)
··· P (σ1 , . . . , σN ) σj = X .
σ1 σN j=1

Here the discrete label n from §1.4.2 ranges over 2N possible values, and may be written as an N digit
binary number rN · · · r1 , where rj = 12 (1 + σj ) is 0 or 1. Extremizing S subject to these constraints, we
obtain
N
( )
e−λ σj ,
X Y
P (σ1 , . . . , σN ) = C exp − λ σj = C (1.59)
j j=1

where C ≡ e−(1+λ0 ) and λ ≡ λ1 . Normalization then requires

X
N
Tr P ≡ P (σ1 , . . . , σN ) = C eλ + e−λ , (1.60)
{σj }
16 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

hence C = (cosh λ)−N . We then have

N N
Y e−λσj Y

P (σ1 , . . . , σN ) = λ −λ
= p δσj ,1 + q δσj ,−1 , (1.61)
e +e
j=1 j=1

where
e−λ eλ
,
p= q = 1 − p = . (1.62)
eλ + e−λ eλ + e−λ
We then have X = (2p − 1)N , which determines p = 12 (N + X), and we have recovered the Bernoulli
distribution.
Of course there are no miracles7 , and there are an infinite family of distributions for which X = (2p −
PN −1that are not Bernoulli. For example, we could have imposed another constraint, such as E =
1)N
j=1 σj σj+1 . This would result in the distribution

N N −1
( )
1 X X
P (σ1 , . . . , σN ) = exp − λ1 σj − λ 2 σj σj+1 , (1.63)
Z
j=1 j=1
P
with Z(λ1 , λ2 ) determined by normalization: σ P (σ) = 1. This is the one-dimensional Ising chain
0 0
of classical equilibrium statistical physics. Defining the transfer matrix Rss0 = e−λ1 (s+s )/2 e−λ2 ss with
s, s0 = ±1 ,
 −λ −λ
eλ2

e 1 2
R=
eλ2 eλ1 −λ2 (1.64)
= e−λ2 cosh λ1 I + eλ2 τ x − e−λ2 sinh λ1 τ z ,
where τ x and τ z are Pauli matrices, we have that
Zring = Tr RN , Zchain = Tr RN −1 S ,



(1.65)
0
where Sss0 = e−λ1 (s+s )/2 , i.e.
 −λ 
e 1 1
S=
1 eλ1 (1.66)
x z
= cosh λ1 I + τ − sinh λ1 τ .
The appropriate case here is that of the chain, but in the thermodynamic limit N → ∞ both chain and
ring yield identical results, so we will examine here the results for the ring, which are somewhat easier to
obtain. Clearly Zring = ζ+ N + ζ N , where ζ are the eigenvalues of R:
− ±
q
ζ± = e−λ2 cosh λ1 ± e−2λ2 sinh2 λ1 + e2λ2 . (1.67)
N . We now have
In the thermodynamic limit, the ζ+ eigenvalue dominates, and Zring ' ζ+
N
DX E ∂ ln Z N sinh λ1
X= σj = − = −q . (1.68)
∂λ1 sinh2 λ + e4λ2
j=1 1
7
See §10 of An Enquiry Concerning Human Understanding by David Hume (1748).
1.4. ENTROPY AND PROBABILITY 17

We also have E = −∂ ln Z/∂λ2 . These two equations determine the Lagrange multipliers λ1 (X, E, N ) and
λ2 (X, E, N ). In the thermodynamic limit, we have λi = λi (X/N, E/N ). Thus, if we fix X/N = 2p − 1
alone, there is a continuous one-parameter family of distributions, parametrized ε = E/N , which satisfy
the constraint on X.
So what is it about the maximum entropy approach that is so compelling? Maximum entropy gives us
a calculable distribution which is consistent with maximum ignorance given our known constraints. In
that sense, it is as unbiased as possible, from an information theoretic point of view. As a starting point,
a maximum entropy distribution may be improved upon, using Bayesian methods for example (see §1.6.2
below).

1.4.3 Continuous probability distributions

Suppose we have a continuous probability density P (ϕ) defined over some set Ω. We have observables
Z
X = dµ X a (ϕ) P (ϕ) ,
a
(1.69)
Ω

where dµ is the appropriate integration measure. We assume dµ = D

Q
j=1 dϕj , where D is the dimension
of Ω. Then we extremize the functional
K
Z Z !
X
S ∗ P (ϕ), {λa } = − dµ P (ϕ) ln P (ϕ) − dµ P (ϕ) X a (ϕ) − X a
 
λa (1.70)
Ω a=0 Ω

with respect to P (ϕ) and with respect to {λa }. Again, X 0 (ϕ) ≡ X 0 ≡ 1. This yields the following result:
K
X
ln P (ϕ) = −1 − λa X a (ϕ) . (1.71)
a=0

The K + 1 Lagrange multipliers {λa } are then determined from the K + 1 constraint equations in eqn.
1.69.
As an example, consider a distribution P (x) over the real numbers R. We constrain
Z∞ Z∞ Z∞
dx P (x) = 1 , dx x P (x) = µ , dx x2 P (x) = µ2 + σ 2 . (1.72)
−∞ −∞ −∞

Extremizing the entropy, we then obtain

2
P (x) = C e−λ1 x−λ2 x , (1.73)

where C = e−(1+λ0 ) . We already know the answer:

1 2 /2σ 2
P (x) = √ e−(x−µ) . (1.74)
2πσ 2
In other words, λ1 = −µ/σ 2 and λ2 = 1/2σ 2 , with C = (2πσ 2 )−1/2 exp(−µ2 /2σ 2 ).
18 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

1.5 General Aspects of Probability Distributions

1.5.1 Discrete and continuous distributions

Consider a system whose possible configurations | n i can be labeled by a discrete variable n ∈ C, where
C is the set of possible configurations. The total number of possible configurations, which is to say the
order of the set C, may be finite or infinite. Next, consider an ensemble of such systems, and let Pn
denote the probability that a given random element from that ensemble is in the state (configuration)
| n i. The collection {Pn } forms a discrete probability distribution. We assume that the distribution is
normalized , meaning X
Pn = 1 . (1.75)
n∈C

Now let An be a quantity which takes values depending on n. The average of A is given by
X
hAi = Pn An . (1.76)
n∈C

Typically, C is the set of integers (Z) or some subset thereof, but it could be any countable set. As an
example, consider the throw of a single six-sided die. Then Pn = 61 for each n ∈ {1, . . . , 6}. Let An = 0
if n is even and 1 if n is odd. Then find hAi = 21 , i.e. on average half the throws of the die will result in
an even number.
It may be that the system’s configurations are described by several discrete variables {n1 , n2 ,P
n3 , . . .}. We
can combine these into a vector n and then we write Pn for the discrete distribution, with n Pn = 1.
Another possibility is that the system’s configurations are parameterized by a collection of continuous
variables, ϕ = {ϕ1 , . . . , ϕn }. We write ϕ ∈ Ω, where Ω is the phase space (or configuration space) of the
system. Let dµ be a measure on this space. In general, we can write

dµ = W (ϕ1 , . . . , ϕn ) dϕ1 dϕ2 · · · dϕn . (1.77)

The phase space measure used in classical statistical mechanics gives equal weight W to equal phase
space volumes:
Yr
dµ = C dqσ dpσ , (1.78)
σ=1

where C is a constant we shall discuss later on below8 .

Any continuous probability distribution P (ϕ) is normalized according to
Z
dµ P (ϕ) = 1 . (1.79)
Ω

8
Such a measure is invariant with respect to canonical transformations, which are the broad class of transformations
among coordinates and momenta which leave Hamilton’s equations of motion invariant, and which preserve phase space
volumes under Hamiltonian evolution. For this reason dµ is called an invariant phase space measure.
1.5. GENERAL ASPECTS OF PROBABILITY DISTRIBUTIONS 19

The average of a function A(ϕ) on configuration space is then

Z
hAi = dµ P (ϕ) A(ϕ) . (1.80)
Ω

For example, consider the Gaussian distribution

1 2 /2σ 2
P (x) = √ e−(x−µ) . (1.81)
2πσ 2
From the result9
Z∞ r
−αx2 −βx π β 2 /4α
dx e e = e , (1.82)
α
−∞

we see that P (x) is normalized. One can then compute

hxi = µ
(1.83)
hx i − hxi2 = σ 2 .
2

We call µ the mean and σ the standard deviation of the distribution, eqn. 1.81.
The quantity P (ϕ) is called the distribution or probability density. One has

P (ϕ) dµ = probability that configuration lies within volume dµ centered at ϕ

 
For example, consider the probability density P = 1 normalized on the interval x ∈ 0, 1 . The probability
that some x chosen at random will be exactly 21 , say, is infinitesimal – one
 would have to specify each of
1

the infinitely many digits of x. However, we can say that x ∈ 0.45 , 0.55 with probability 10 .
If x is distributed according to P1 (x), then the probability distribution on the product space (x1 , x2 )
is simply the product of the distributions: P2 (x1 , x2 ) = P1 (x1 ) P1 (x2 ). Suppose we have a function
φ(x1 , . . . , xN ). How is it distributed? Let P (φ) be the distribution for φ. We then have
Z∞ Z∞  
P (φ) = dx1 · · · dxN PN (x1 , . . . , xN ) δ φ(x1 , . . . , xN ) − φ
−∞ −∞
(1.84)
Z∞ Z∞  
= dx1 · · · dxN P1 (x1 ) · · · P1 (xN ) δ φ(x1 , . . . , xN ) − φ ,
−∞ −∞

where the second line is appropriate if the {xj } are themselves distributed independently. Note that

Z∞
dφ P (φ) = 1 , (1.85)
−∞

so P (φ) is itself normalized.

9
Memorize this!
20 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

1.5.2 Central limit theorem

PN
In particular, consider the distribution function of the sum X = i=1 xi . We will be particularly
interested in the case where N is large. For general N , though, we have
Z∞ Z∞


PN (X) = dx1 · · · dxN P1 (x1 ) · · · P1 (xN ) δ x1 + x2 + . . . + xN − X . (1.86)
−∞ −∞

It is convenient to compute the Fourier transform10 of P (X):

Z∞
P̂N (k) = dX PN (X) e−ikX
−∞
(1.87)
Z∞ Z∞ Z∞
N
dX dx1 · · · dxN P1 (x1 ) · · · P1 (xN ) δ x1 + . . . + xN − X) e−ikX = P̂1 (k) ,

=
−∞ −∞ −∞

where
Z∞
P̂1 (k) = dx P1 (x) e−ikx (1.88)
−∞

is the Fourier transform of the single variable distribution P1 (x). The distribution PN (X) is a convo-
lution of the individual P1 (xi ) distributions. We have therefore proven that the Fourier transform of a
convolution is the product of the Fourier transforms.
OK, now we can write for P̂1 (k)

Z∞
P̂1 (k) = dx P1 (x) 1 − ikx − 21 k 2 x2 + 16 i k 3 x3 + . . .


(1.89)
−∞
= 1 − ikhxi − 12 k 2 hx2 i + 16 i k 3 hx3 i + . . . .
10
Jean Baptiste Joseph Fourier (1768-1830) had an illustrious career. The son of a tailor, and orphaned at age eight,
Fourier’s ignoble status rendered him ineligible to receive a commission in the scientific corps of the French army. A
Benedictine minister at the École Royale Militaire of Auxerre remarked, ”Fourier, not being noble, could not enter the
artillery, although he were a second Newton.” Fourier prepared for the priesthood but his affinity for mathematics proved
overwhelming, and so he left the abbey and soon thereafter accepted a military lectureship position. Despite his initial
support for revolution in France, in 1794 Fourier ran afoul of a rival sect while on a trip to Orleans and was arrested
and very nearly guillotined. Fortunately the Reign of Terror ended soon after the death of Robespierre, and Fourier was
released. He went on Napoleon Bonaparte’s 1798 expedition to Egypt, where he was appointed governor of Lower Egypt. His
organizational skills impressed Napoleon, and upon return to France he was appointed to a position of prefect in Grenoble.
It was in Grenoble that Fourier performed his landmark studies of heat, and his famous work on partial differential equations
and Fourier series. It seems that Fourier’s fascination with heat began in Egypt, where he developed an appreciation of
desert climate. His fascination developed into an obsession, and he became convinced that heat could promote a healthy
body. He would cover himself in blankets, like a mummy, in his heated apartment, even during the middle of summer. On
May 4, 1830, Fourier, so arrayed, tripped and fell down a flight of stairs. This aggravated a developing heart condition,
which he refused to treat with anything other than more heat. Two weeks later, he died. Fourier’s is one of the 72 names
of scientists, engineers and other luminaries which are engraved on the Eiffel Tower.
1.5. GENERAL ASPECTS OF PROBABILITY DISTRIBUTIONS 21

Thus,
ln P̂1 (k) = −iµk − 12 σ 2 k 2 + 61 i γ 3 k 3 + . . . , (1.90)
where

µ = hxi
σ 2 = hx2 i − hxi2 (1.91)
3 3 2 3
γ = hx i − 3 hx i hxi + 2 hxi

We can now write

N 2 k 2 /2 3 k 3 /6
= e−iN µk e−N σ eiN γ

P̂1 (k) ··· (1.92)
3 k 3 /6
Now for the inverse transform. In computing PN (X), we will expand the term eiN γ and all subsequent
terms in the above product as a power series in k. We then have
Z∞
dk ik(X−N µ) −N σ2 k2 /2 n o
PN (X) = e e 1 + 61 i N γ 3 k 3 + . . .
2π
−∞

γ3 ∂3
 
1 2 2 (1.93)
= 1− N 3
+ ... √ e−(X−N µ) /2N σ
6 ∂X 2πN σ 2
γ 3 −1/2 ∂ 3
 
1 2 2
= 1− N 3
+ ... √ e−ξ /2σ .
6 ∂ξ 2πN σ 2

√
In going from the second line to the third, we have written X = N µ+ N ξ, in which case ∂X = N −1/2 ∂ξ ,
and the non-Gaussian terms give a subleading contribution which vanishes in the N → ∞ limit. We have
just proven the central limit theorem: in the limit N → ∞, the distribution
√ of a sum of N independent
random variables xi is a Gaussian with mean N µ and standard deviation N σ. Our only assumptions
are that the mean µ and standard deviation σ exist for the distribution P1 (x). Note that P1 (x) itself
need not be a Gaussian – it could be a very peculiar distribution indeed, but so long as itsPfirst and
second moment exist, where the k th moment is simply hxk i, the distribution of the sum X = N i=1 xi is
a Gaussian.

1.5.3 Moments and cumulants

Consider a general multivariate distribution P (x1 , . . . , xN ) and define the multivariate Fourier transform

Z∞ Z∞  N
X 
P̂ (k1 , . . . , kN ) = dx1 · · · dxN P (x1 , . . . , xN ) exp − i kj xj . (1.94)
−∞ −∞ j=1

The inverse relation is

Z∞ Z∞  N 
dk1 dkN X
P (x1 , . . . , xN ) = ··· P̂ (k1 , . . . , kN ) exp + i k j xj . (1.95)
2π 2π
−∞ −∞ j=1
22 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

∂
Acting on P̂ (k), the differential operator i ∂k brings down from the exponential a factor of xi inside the
i
integral. Thus,
" #
∂ m1 ∂ mN
  

m m 
i ··· i P̂ (k) = x1 1 · · · xN N . (1.96)
∂k1 ∂kN
k=0
Similarly, we can reconstruct the distribution from its moments, viz .
∞ ∞
X X (−ik1 )m1 (−ikN )mN
m1 m 
P̂ (k) = ··· ··· x1 · · · xN N . (1.97)
m1 ! mN !
m1 =0 mN =0

m m
The cumulants hhx1 1 · · · xN N ii are defined by the Taylor expansion of ln P̂ (k):
∞ ∞
X X (−ik1 )m1 (−ikN )mN

m1 m 
ln P̂ (k) = ··· ··· x1 · · · xN N . (1.98)
m1 ! mN !
m1 =0 mN =0

There is no general form for the cumulants. It is straightforward to derive the following low order results:

hhxi ii = hxi i
hhxi xj ii = hxi xj i − hxi ihxj i (1.99)
hhxi xj xk ii = hxi xj xk i − hxi xj ihxk i − hxj xk ihxi i − hxk xi ihxj i + 2hxi ihxj ihxk i .

1.5.4 Multidimensional Gaussian integral

Consider the multivariable Gaussian distribution,

det A 1/2
   
1
P (x) ≡ exp − 2 i ij j ,
x A x (1.100)
(2π)n
where A is a positive definite matrix of rank n. A mathematical result which is extremely important
throughout physics is the following:
 1/2 Z∞ Z∞
det A    
Z(b) = dx1 · · · dxn exp − 21 xi Aij xj + bi xi = exp 21 bi A−1
ij j .
b (1.101)
(2π)n
−∞ −∞

Here, the vector b = (b1 , . . . , bn ) is identified as a source. Since Z(0) = 1, we have that the distribution
P (x) is normalized. Now consider averages of the form

∂ nZ(b) 
Z 
n
h xj1· · · xj i = d x P (x) xj1· · · xj =
2k 2k ∂bj · · · ∂bj b=0
X 1 2k (1.102)
−1 −1
= Aj j · · · Aj j .
σ(1) σ(2) σ(2k−1) σ(2k)
contractions

The sum in the last term is over all contractions of the indices {j1 , . . . , j2k }. A contraction is an
arrangement of the 2k indices into k pairs. There are C2k = (2k)!/2k k! possible such contractions. To
1.6. BAYESIAN STATISTICAL INFERENCE 23

obtain this result for Ck , we start with the first index and then find a mate among the remaining 2k − 1
indices. Then we choose the next unpaired index and find a mate among the remaining 2k − 3 indices.
Proceeding in this manner, we have
(2k)!
C2k = (2k − 1) · (2k − 3) · · · 3 · 1 =
. (1.103)
2k k!
Equivalently, we can take all possible permutations of the 2k indices, and then divide by 2k k! since
permutation within a given pair results in the same contraction and permutation among the k pairs
results in the same contraction. For example, for k = 2, we have C4 = 3, and
h xj1 xj2 xj3 xj4 i = A−1 −1 −1 −1 −1 −1
j j Aj j + Aj j Aj j + Aj j Aj j . (1.104)
1 2 3 4 1 3 2 4 1 4 2 3

If we define bi = iki , we have  

P̂ (k) = exp − 21 ki A−1
ij kj , (1.105)
from which we read off the cumulants hhxi xj ii = A−1
ij , with all higher order cumulants vanishing.

1.6 Bayesian Statistical Inference

1.6.1 Frequentists and Bayesians

There field of statistical inference is roughly divided into two schools of practice: frequentism and
Bayesianism. You can find several articles on the web discussing the differences in these two approaches.
In both cases we would like to model observable data x by a distribution. The distribution in general
depends on one or more parameters θ. The basic worldviews of the two approaches are as follows:

Frequentism: Data x are a random sample drawn from an infinite pool at some frequency.
The underlying parameters θ, which are to be estimated, remain fixed during this process.
There is no information prior to the model specification. The experimental conditions under
which the data are collected are presumed to be controlled and repeatable. Results are gen-
erally expressed in terms of confidence intervals and confidence levels, obtained via statistical
hypothesis testing. Probabilities have meaning only for data yet to be collected. Calculations
generally are computationally straightforward.

Bayesianism: The only data x which matter are those which have been observed. The
parameters θ are unknown and described probabilistically using a prior distribution, which
is generally based on some available information but which also may be at least partially
subjective. The priors are then to be updated based on observed data x. Results are expressed
in terms of posterior distributions and credible intervals. Calculations can be computationally
intensive.

In essence, frequentists say the data are random and the parameters are fixed . while Bayesians say the
data are fixed and the parameters are random 11 . Overall, frequentism has dominated over the past several
11
”A frequentist is a person whose long-run ambition is to be wrong 5% of the time. A Bayesian is one who, vaguely
expecting a horse, and catching glimpse of a donkey, strongly believes he has seen a mule.” – Charles Annis.
24 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

hundred years, but Bayesianism has been coming on strong of late, and many physicists seem naturally
drawn to the Bayesian perspective.

1.6.2 Updating Bayesian priors

Given data D and a hypothesis H, Bayes’ theorem tells us

P (D|H) P (H)
P (H|D) = . (1.106)
P (D)

Typically the data is in the form of a set of values x = {x1 , . . . , xN }, and the hypothesis in the form
of a set of parameters θ = {θ1 , . . . , θK }. It is notationally helpful to express distributions of x and
distributions of x conditioned on θ using the symbol f , and distributions of θ and distributions of θ
conditioned on x using the symbol π, rather than using the symbol P everywhere. We then have

f (x|θ) π(θ)
π(θ|x) = R , (1.107)
dθ 0 f (x|θ 0 ) π(θ 0 )
Θ
R
where Θ 3 θ is the space of parameters. Note that Θ dθ π(θ|x) = 1. The denominator of the RHS
is simply f (x), which is independent of θ, hence π(θ|x) ∝ f (x|θ) π(θ). We call π(θ) the prior for θ,
f (x|θ) the likelihood of x given θ, and π(θ|x) the posterior for θ given x. The idea here is that while
our initial guess at the θ distribution is given by the prior π(θ), after taking data, we should update this
distribution to the posterior π(θ|x). The likelihood f (x|θ) is entailed by our model for the phenomenon
which produces the data. We can use the posterior to find the distribution of new data points y, called
the posterior predictive distribution,
Z
f (y|x) = dθ f (y|θ) π(θ|x) . (1.108)
Θ

This is the update of the prior predictive distribution,

Z
f (x) = dθ f (x|θ) π(θ) . (1.109)
Θ

Example: coin flipping

Consider a model of coin flipping based on a standard Bernoulli distribution, where θ ∈ [0, 1] is the
probability for heads (x = 1) and 1 − θ the probability for tails (x = 0). That is,

N h
Y i
f (x1 , . . . , xN |θ) = (1 − θ) δxj ,0 + θ δxj ,1
j=1 (1.110)
= θX (1 − θ)N −X ,
1.6. BAYESIAN STATISTICAL INFERENCE 25

where X = N
P
j=1 xj is the observed total number of heads, and N − X the corresponding number of tails.
We now need a prior π(θ). We choose the Beta distribution,

θα−1 (1 − θ)β−1
π(θ) = , (1.111)
B(α, β)

where B(α, β) =R 1Γ(α) Γ(β)/Γ(α + β) is the Beta function. One can check that π(θ) is normalized on the
unit interval: 0 dθ π(θ) = 1 for all positive α, β. Even if we limit ourselves to this form of the prior,
different Bayesians might bring different assumptions about the values of α and β. Note that if we choose
α = β = 1, the prior distribution for θ is flat, with π(θ) = 1.
We now compute the posterior distribution for θ:

f (x1 , . . . , xN |θ) π(θ) θX+α−1 (1 − θ)N −X+β−1

π(θ|x1 , . . . , xN ) = R 1 = . (1.112)
0 0 0 B(X + α, N − X + β)
0 dθ f (x1 , . . . , xN |θ ) π(θ )

Thus, we retain the form of the Beta distribution, but with updated parameters,

α0 = X + α
(1.113)
β0 = N − X + β .

The fact that the functional form of the prior is retained by the posterior is generally not the case in
Bayesian updating. We can also compute the prior predictive,

Z1
f (x1 , . . . , xN ) = dθ f (x1 , . . . , xN |θ) π(θ)
0
(1.114)
Z1
1 B(X + α, N − X + β)
= dθ θX+α−1 (1 − θ)N −X+β−1 = .
B(α, β) B(α, β)
0

The posterior predictive is then

Z1
f (y1 , . . . , yM |x1 , . . . , xN ) = dθ f (y1 , . . . , yM |θ) π(θ|x1 , . . . , xN )
0
Z1
1 (1.115)
= dθ θX+Y +α−1 (1 − θ)N −X+M −Y +β−1
B(X + α, N − X + β)
0
B(X + Y + α, N − X + M − Y + β)
= .
B(X + α, N − X + β)

1.6.3 Hyperparameters and conjugate priors

In the above example, θ is a parameter of the Bernoulli distribution, i.e. the likelihood, while quantities
α and β are hyperparameters which enter the prior π(θ). Accordingly, we could have written π(θ|α, β)
26 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

for the prior. We then have for the posterior

f (x|θ) π(θ|α)
π(θ|x, α) = R , (1.116)
dθ 0 f (x|θ 0 ) π(θ 0 |α)
Θ

replacing eqn. 1.107, etc., where α ∈ A is the vector of hyperparameters. The hyperparameters can
also be distributed, according to a hyperprior ρ(α), and the hyperpriors can further be parameterized by
hyperhyperparameters, which can have their own distributions, ad nauseum.
What use is all this? We’ve already seen a compelling example: when the posterior is of the same form
as the prior, the Bayesian update can be viewed as an automorphism of the hyperparameter space A, i.e.
one set of hyperparameters α is mapped to a new set of hyperparameters α.e


Definition: A parametric family of distributions P = π(θ|α) | θ ∈ Θ, α ∈ A is called a
conjugate family for a family of distributions f (x|θ) | x ∈ X , θ ∈ Θ if, for all x ∈ X and
α ∈ A,
f (x|θ) π(θ|α)
π(θ|x, α) ≡ R 0 ∈P . (1.117)
dθ f (x|θ 0 ) π(θ 0 |α)
Θ

That is, π(θ|x, α) = π(θ|α) e ∈ A, with α

e for some α e = α(α,
e x).

As an example, consider the conjugate Bayesian analysis of the Gaussian distribution. We assume a
likelihood
N
( )
1 X
f (x|u, s) = (2πs2 )−N/2 exp − 2 (xj − u)2 . (1.118)
2s
j=1

The parameters here are θ = {u, s}. Now consider the prior distribution
( )
(u − µ0 )2
π(u, s|µ0 , σ0 ) = (2πσ02 )−1/2 exp − . (1.119)
2σ02

Note that the prior distribution is independent of the parameter s and only depends on u and the
hyperparameters α = (µ0 , σ0 ). We now compute the posterior:

π(u, s|x, µ0 , σ0 ) ∝ f (x|u, s) π(u, s|µ0 , σ0 )

(   2 )
2i (1.120)
  
1 N µ N hxi µ N hx
= exp − + u2 + 0
+ 2 u− 0
+ ,
2σ02 2s2 σ02 s 2σ02 2s2

with hxi = N1 N 2 1 PN 2
P
j=1 xj and hx i = N j=1 xj . This is also a Gaussian distribution for u, and after
supplying the appropriate normalization one finds
( )
2 −1/2 (u − µ1 )2
π(u, s|x, µ0 , σ0 ) = (2πσ1 ) exp − , (1.121)
2σ12
1.6. BAYESIAN STATISTICAL INFERENCE 27

with
N hxi − µ0 σ02


µ1 = µ0 +
s2 + N σ02
(1.122)
s2 σ02
σ12 = 2 .
s + N σ02

Thus, the posterior is among the same family as the prior, and we have derived the update rule for the
hyperparameters (µ0 , σ0 ) → (µ1 , σ1 ). Note that σ1 < σ0 , so the updated Gaussian prior is sharper than
the original. The updated mean µ1 shifts in the direction of hxi obtained from the data set.

1.6.4 The problem with priors

We might think that the for the coin flipping problem, the flat prior π(θ) = 1 is an appropriate initial
one, since it does not privilege any value of θ. This prior therefore seems ’objective’ or ’unbiased’, also
called ’uninformative’. But supposewe make a change of variables, mapping the interval θ ∈ [0, 1] to the
entire real line according to ζ = ln θ/(1 − θ) . In terms of the new parameter ζ, we write the prior as
π̃(ζ). Clearly π(θ) dθ = π̃(ζ) dζ, so π̃(ζ) = π(θ) dθ/dζ. For our example, find π̃(ζ) = 41 sech2(ζ/2), which
is not flat. Thus what was uninformative in terms of θ has become very informative in terms of the new
parameter ζ. Is there any truly unbiased way of selecting a Bayesian prior?
One approach, advocated by E. T. Jaynes, is to choose the prior distribution π(θ) according to the
principle of maximum entropy. For continuous parameter spaces, we must first define a parameter space
metric so as to be able to ’count’ the number ofR different parameter states. The entropy of a distribution
π(θ) is then dependent on this metric: S = − dµ(θ) π(θ) ln π(θ).
Another approach, due to Jeffreys, is to derive a parameterization-independent prior from the likelihood
f (x|θ) using the so-called Fisher information matrix ,
 2 
∂ lnf (x|θ)
Iij (θ) = −Eθ
∂θi ∂θj
(1.123)
∂ 2 lnf (x|θ)
Z
= − dx f (x|θ) .
∂θi ∂θj

The Jeffreys prior πJ (θ) is defined as p

πJ (θ) ∝ det I(θ) . (1.124)
One can check that the Jeffries prior is invariant under reparameterization. As an
P example, consider the
Bernoulli process, for which ln f (x|θ) = X ln θ + (N − X) ln(1 − θ), where X = N j=1 xj . Then

d2 ln p(x|θ) X N −X
− 2
= 2+ , (1.125)
dθ θ (1 − θ)2
and since Eθ X = N θ, we have

N 1 1
I(θ) = ⇒ πJ (θ) = p , (1.126)
θ(1 − θ) π θ(1 − θ)
28 CHAPTER 1. FUNDAMENTALS OF PROBABILITY

which felicitously corresponds to a Beta distribution with α = β = 21 . In this example the Jeffries prior
turned out to be a conjugate prior, but in general this is not the case.
We can try to implement the Jeffreys procedure for a two-parameter family where each xj is normally
distributed with mean µ and standard deviation σ. Let the parameters be (θ1 , θ2 ) = (µ, σ). Then
N
√ 1 X
− ln f (x|θ) = N ln 2π + N ln σ + 2 (xj − µ)2 , (1.127)
2σ
j=1

and the Fisher information matrix is

N σ −2 σ −3
P
− µ)
 
∂ 2 lnf (x|θ) j (xj
I(θ) = − =  . (1.128)
∂θi ∂θj −3
P −2 −4
P 2
σ j (xj − µ) −N σ + 3σ j (xj − µ)

Taking the expectation value, we have E (xj − µ) = 0 and E (xj − µ)2 = σ 2 , hence

N σ −2
 
0
E I(θ) = (1.129)
0 2N σ −2

and the Jeffries prior is πJ (µ, σ) ∝ σ −2 . This is problematic because if we choose a flat metric on the
(µ, σ) upper half plane, the Jeffries prior is not normalizable. Note also that the Jeffreys prior no longer
resembles a Gaussian, and hence is not a conjugate prior.
Chapter 2

Thermodynamics

2.1 References

– E. Fermi, Thermodynamics (Dover, 1956)

This outstanding and inexpensive little book is a model of clarity.

– A. H. Carter, Classical and Statistical Thermodynamics

(Benjamin Cummings, 2000)
A very relaxed treatment appropriate for undergraduate physics majors.

– H. B. Callen, Thermodynamics and an Introduction to Thermostatistics

(2nd edition, Wiley, 1985)
A comprehensive text appropriate for an extended course on thermodynamics.

– D. V. Schroeder, An Introduction to Thermal Physics (Addison-Wesley, 2000)

An excellent thermodynamics text appropriate for upper division undergraduates. Contains many
illustrative practical applications.

– D. Kondepudi and I. Prigogine, Modern Thermodynamics: From Heat Engines to Dissipative Struc-
tures (Wiley, 1998)
Lively modern text with excellent choice of topics and good historical content. More focus on
chemical and materials applications than in Callen.

– L. E. Reichl, A Modern Course in Statistical Physics (2nd edition, Wiley, 1998)

A graduate level text with an excellent and crisp section on thermodynamics.

29
30 CHAPTER 2. THERMODYNAMICS

2.2 What is Thermodynamics?

Thermodynamics is the study of relations among the state variables describing a thermodynamic system,
and of transformations of heat into work and vice versa.

2.2.1 Thermodynamic systems and state variables

Thermodynamic systems contain large numbers of constituent particles, and are described by a set of
state variables which describe the system’s properties in an average sense. State variables are classified
as being either extensive or intensive.
Extensive variables, such as volume V , particle number N , total internal energy E, magnetization M ,
etc., scale linearly with the system size, i.e. as the first power of the system volume. If we take two
identical thermodynamic systems, place them next to each other, and remove any barriers between them,
then all the extensive variables will double in size.
Intensive variables, such as the pressure p, the temperature T , the chemical potential µ, the electric field
E, etc., are independent of system size, scaling as the zeroth power of the volume. They are the same
throughout the system, if that system is in an appropriate state of equilibrium. The ratio of any two
extensive variables is an intensive variable. For example, we write n = N/V for the number density,
which scales as V 0 . Intensive variables may also be inhomogeneous. For example, n(r) is the number
density at position r, and is defined as the limit of ∆N/∆V of the number of particles ∆N inside a
volume ∆V which contains the point r, in the limit V  ∆V  V /N .
Classically, the full motion of a system of N point particles requires 6N variables to fully describe it (3N
positions and 3N velocities or momenta, in three space dimensions)1 . Since the constituents are very
small, N is typically very large. A typical solid or liquid, for example, has a mass density on the order
of % ∼ 1 g/cm3 ; for gases, % ∼ 10−3 g/cm3 . The constituent atoms have masses of 100 to 102 grams per
mole, where one mole of X contains NA of X, and NA = 6.0221415 × 1023 is Avogadro’s number2 . Thus,
for solids and liquids we roughly expect number densities n of 10−2 − 100 mol/cm3 for solids and liquids,
and 10−5 − 10−3 mol/cm3 for gases. Clearly we are dealing with fantastically large numbers of constituent
particles in a typical thermodynamic system. The underlying theoretical basis for thermodynamics, where
we use a small number of state variables to describe a system, is provided by the microscopic theory of
statistical mechanics, which we shall study in the weeks ahead.
Intensive quantities such as p, T , and n ultimately involve averages over both space and time. Consider for
example the case of a gas enclosed in a container. We can measure the pressure (relative to atmospheric
pressure) by attaching a spring to a moveable wall, as shown in Fig. 2.2. From the displacement of the
spring and the value of its spring constant k we determine the force F . This force is due to the difference
in pressures, so p = p0 + F/A. Microscopically, the gas consists of constituent atoms or molecules, which
are constantly undergoing collisions with each other and with the walls of the container. When a particle
bounces off a wall, it imparts an impulse 2n̂(n̂ · p), where p is the particle’s momentum and n̂ is the unit
1
For a system of N molecules which can freely rotate, we must then specify 3N additional orientational variables – the
Euler angles – and their 3N conjugate momenta. The dimension of phase space is then 12N .
2
Hence, 1 guacamole = 6.0221415 × 1023 guacas.
2.2. WHAT IS THERMODYNAMICS? 31

Figure 2.1: From microscale to macroscale : physical versus social sciences.

vector normal to the wall. (Only particles with p · n̂ > 0 will hit the wall.) Multiply this by the number
of particles colliding with the wall per unit time, and one finds the net force on the wall; dividing by
the area gives the pressure p. Within the gas, each particle travels for a distance `, called the mean free
path, before it undergoes a collision. We can write ` = v̄τ , where v̄ is the average particle speed and τ is
the mean free time. When we study the kinetic theory of gases, we will derive formulas for ` and v̄ (and
hence τ ). For now it is helpful to quote some numbers to get an idea of the relevant distance and time
scales. For O2 gas at standard temperature and pressure (T = 0◦ C, p = 1 atm), the mean free path is
` ≈ 1.1 × 10−5 cm, the average speed is v̄ ≈ 480 m/s, and the mean free time is τ ≈ 2.5 × 10−10 s. Thus,
particles in the gas undergo collisions at a rate τ −1 ≈ 4.0 × 109 s−1 . A measuring device, such as our
spring, or a thermometer, effectively performs time and space averages. If there are Nc collisions with
a particular patch of wall during some time interval on which our measurement device responds, then
−1/2
the root mean square relative fluctuations in the local pressure will be on the order of Nc times the
average. Since Nc is a very large number, the fluctuations are negligible.
If the system is in steady state, the state variables do not change with time. If furthermore there are
no macroscopic currents of energy or particle number flowing through the system, the system is said
to be in equilibrium. A continuous succession of equilibrium states is known as a thermodynamic path,
which can be represented as a smooth curve in a multidimensional space whose axes are labeled by state
variables. A thermodynamic process is any change or succession of changes which results in a change of
the state variables. In a cyclic process, the initial and final states are the same. In a quasistatic process,
the system passes through a continuous succession of equilibria. A reversible process is one where the
external conditions and the thermodynamic path of the system can be reversed; it is both quasi-static and
non-dissipative (i.e. no friction). The slow expansion of a gas against a piston head, whose counter-force
is always infinitesimally less than the force pA exerted by the gas, is reversible. To reverse this process,
we simply add infinitesimally more force to pA and the gas compresses. An example of a quasistatic
process which is not reversible: slowly dragging a block across the floor, or the slow leak of air from a
tire. Irreversible processes, as a rule, are dissipative. Other special processes include isothermal (dT = 0),
isobaric (dp = 0), isochoric (dV = 0), and adiabatic (dQ ¯ = 0, i.e. no heat exchange):

reversible: dQ
¯ = T dS isothermal: dT = 0
spontaneous: dQ
¯ < T dS isochoric: dV = 0
adiabatic: dQ
¯ =0 isobaric: dp = 0 .

We shall discuss later the entropy S and its connection with irreversibility.
32 CHAPTER 2. THERMODYNAMICS

Figure 2.2: The pressure p of a gas is due to an average over space and time of the impulses due to the
constituent particles.

How many state variables are necessary to fully specify the equilibrium state of a thermodynamic system?
For a single component system, such as water which is composed of one constituent molecule, the answer
is three. These can be taken to be T , p, and V . One always must specify at least one extensive variable,
else we cannot determine the overall size of the system. For a multicomponent system with g different
species, we must specify g + 2 state variables, which may be {T, p, N1 , . . . , Ng }, where Na is the number
of particles of species a. Another possibility
Pg is the set (T, p, V, x1 , . . . , xg−1 }, where the concentration
Pg of
species a is xa = Na /N . Here, N = a=1 Na is the total number of particles. Note that a=1 xa = 1.
If then follows that if we specify more than g + 2 state variables, there must exist a relation among them.
Such relations are known as equations of state. The most famous example is the ideal gas law,

pV = N kB T , (2.1)

relating the four state variables T , p, V , and N . Here kB = 1.3806503 × 10−16 erg/K is Boltzmann’s
constant. Another example is the van der Waals equation,
aN 2
 
p + 2 (V − bN ) = N kB T , (2.2)
V
where a and b are constants which depend on the molecule which forms the gas. For a third example,
consider a paramagnet, where
M CH
= , (2.3)
V T
where M is the magnetization, H the magnetic field, and C the Curie constant.
Any quantity which, in equilibrium, depends only on the state variables is called a state function. For
example, the total internal energy E of a thermodynamics system is a state function, and we may write
E = E(T, p, V ). State functions can also serve as state variables, although the most natural state variables
are those which can be directly measured.
2.2. WHAT IS THERMODYNAMICS? 33

2.2.2 Heat

Once thought to be a type of fluid, heat is now understood in terms of the kinetic theory of gases, liquids,
and solids as a form of energy stored in the disordered motion of constituent particles. The units of
heat are therefore units of energy, and it is appropriate to speak of heat energy, which we shall simply
abbreviate as heat:3

1 J = 107 erg = 6.242 × 1018 eV = 2.390 × 10−4 kcal = 9.478 × 10−4 BTU . (2.4)

We will use the symbol Q to denote the amount of heat energy absorbed by a system during some given
thermodynamic process, and dQ ¯ to denote a differential amount of heat energy. The symbol d̄ indicates
an ‘inexact differential’, about which we shall have more to say presently. This means that heat is not a
state function: there is no ‘heat function’ Q(T, p, V ).

2.2.3 Work

In general we will write the differential element of work dW

¯ done by the system as
X
dW
¯ = Fi dXi , (2.5)
i

where Fi is a generalized force and dXi a generalized displacement 4 . The generalized forces and displace-
ments are themselves state variables, and by convention we will take the generalized forces to be intensive
and the generalized displacements to be extensive. As an example, in a simple one-component system,
we have dW
¯ = p dV . More generally, we write
P
− yj dXj
P
j a µa dNa
z }|
{ z }|
{
dW
¯ = p dV − H · dM − E · dP − σ dA + . . . − µ1 dN1 + µ2 dN2 + . . . (2.6)

Here we distinguish between two types of work. The first involves changes in quantities such as volume,
magnetization, electric polarization, area, etc. The conjugate forces yi applied to the system are then
−p, the magnetic field H, the electric field E, the surface tension σ, respectively. The second type of
work involves changes in the number of constituents of a given species. For example, energy is required
in order to dissociate two hydrogen atoms in an H2 molecule. The effect of such a process is dNH = −1
2
and dNH = +2.
As with heat, dW
¯ is an inexact differential, and work W is not a state variable, since it is path-dependent.
There is no ‘work function’ W (T, p, V ).

3
One calorie (cal) is the amount of heat needed to raise 1 g of H2 O from T0 = 14.5◦ C to T1 = 15.5◦ C at a pressure
of p0 = 1 atm. One British Thermal Unit (BTU) is the amount of heat needed to raise 1 lb. of H2 O from T0 = 63◦ F to
T1 = 64◦ F at a pressure of p0 = 1 atm.
4
We use the symbol d¯ in the differential dW
¯ to indicate that this is not an exact differential. More on this in section 2.4
below.
34 CHAPTER 2. THERMODYNAMICS

2.2.4 Pressure and Temperature

The units of pressure (p) are force per unit area. The SI unit is the Pascal (Pa): 1 Pa = 1 N/m2 =
1 kg/m s2 . Other units of pressure we will encounter:

1 bar ≡ 105 Pa
1 atm ≡ 1.01325 × 105 Pa
1 torr ≡ 133.3 Pa .

Temperature (T ) has a very precise definition from the point of view of statistical mechanics, as we shall
see. Many physical properties depend on the temperature – such properties are called thermometric
properties. For example, the resistivity of a metal ρ(T, p) or the number density of a gas n(T, p) are both
thermometric properties, and can be used to define a temperature scale. Consider the device known as
the ‘constant volume gas thermometer’ depicted in Fig. 2.3, in which the volume or pressure of a gas may
be used to measure temperature. The gas is assumed to be in equilibrium at some pressure p, volume
V , and temperature T . An incompressible fluid of density % is used to measure the pressure difference
∆p = p − p0 , where p0 is the ambient pressure at the top of the reservoir:

p − p0 = %g(h2 − h1 ) , (2.7)

where g is the acceleration due to gravity. The height h1 of the left column of fluid in the U-tube provides
a measure of the change in the volume of the gas:

V (h1 ) = V (0) − Ah1 , (2.8)

where A is the (assumed constant) cross-sectional area of the left arm of the U-tube. The device can
operate in two modes:

• Constant pressure mode : The height of the reservoir is adjusted so that the height difference h2 −h1
is held constant. This fixes the pressure p of the gas. The gas volume still varies with temperature
T , and we can define
T V
= , (2.9)
Tref Vref
where Tref and Vref are the reference temperature and volume, respectively.

• Constant volume mode : The height of the reservoir is adjusted so that h1 = 0, hence the volume
of the gas is held fixed, and the pressure varies with temperature. We then define
T p
= , (2.10)
Tref pref

where Tref and pref are the reference temperature and pressure, respectively.

What should we use for a reference? One might think that a pot of boiling water will do, but anyone
who has gone camping in the mountains knows that water boils at lower temperatures at high altitude
2.2. WHAT IS THERMODYNAMICS? 35

Figure 2.3: The constant volume gas thermometer. The gas is placed in thermal contact with an
object of temperature T . An incompressible fluid of density % is used to measure the pressure difference
∆p = pgas − p0 .

(lower pressure). This phenomenon is reflected in the phase diagram for H2 O, depicted in Fig. 2.4.
There are two special points in the phase diagram, however. One is the triple point, where the solid,
liquid, and vapor (gas) phases all coexist. The second is the critical point, which is the terminus of the
curve separating liquid from gas. At the critical point, the latent heat of transition between liquid and
gas phases vanishes (more on this later on). The triple point temperature Tt at thus unique and is by
definition Tt = 273.16 K. The pressure at the triple point is 611.7 Pa = 6.056 × 10−3 atm.
A question remains: are the two modes of the thermometer compatible? E.g. it we boil water at p = p0 =
1 atm, do they yield the same value for T ? And what if we use a different gas in our measurements? In
fact, all these measurements will in general be incompatible, yielding different results for the temperature
T . However, in the limit that we use a very low density gas, all the results converge. This is because all
low density gases behave as ideal gases, and obey the ideal gas equation of state pV = N kB T .

2.2.5 Standard temperature and pressure

It is customary in the physical sciences to define certain standard conditions with respect to which any
arbitrary conditions may be compared. In thermodynamics, there is a notion of standard temperature
and pressure, abbreviated STP. Unfortunately, there are two different definitions of STP currently in use,
one from the International Union of Pure and Applied Chemistry (IUPAC), and the other from the U.S.
National Institute of Standards and Technology (NIST). The two standards are:

IUPAC : T0 = 0◦ C = 273.15 K , p0 = 105 Pa

NIST : T0 = 20◦ C = 293.15 K , p0 = 1 atm = 1.01325 × 105 Pa
36 CHAPTER 2. THERMODYNAMICS

Figure 2.4: A sketch of the phase diagram of H2 O (water). Two special points are identified: the triple
point (Tt , pt ) at which there is three phase coexistence, and the critical point (Tc , pc ), where the latent
heat of transformation from liquid to gas vanishes. Not shown are transitions between several different
solid phases.

To make matters worse, in the past it was customary to define STP as T0 = 0◦ C and p0 = 1 atm. We will
use the NIST definition in this course. Unless I slip and use the IUPAC definition. Figuring out what I
mean by STP will keep you on your toes.
The volume of one mole of ideal gas at STP is then
(
NA kB T0 22.711 ` (IUPAC)
V = = (2.11)
p0 24.219 ` (NIST) ,
where 1 ` = 106 cm3 = 10−3 m3 is one liter. Under the old definition of STP as T0 = 0◦ C and p0 = 1 atm,
the volume of one mole of gas at STP is 22.414 `, which is a figure I remember from my 10th grade
chemistry class with Mr. Lawrence.

2.3 The Zeroth Law of Thermodynamics

Equilibrium is established by the exchange of energy, volume, or particle number between different systems
or subsystems:
energy exchange =⇒ T = constant =⇒ thermal equilibrium
p
volume exchange =⇒ = constant =⇒ mechanical equilibrium
T
µ
particle exchange =⇒ = constant =⇒ chemical equilibrium
T
2.4. MATHEMATICAL INTERLUDE : EXACT AND INEXACT DIFFERENTIALS 37

Figure 2.5: As the gas density tends to zero, the readings of the constant volume gas thermometer
converge.

Equilibrium is transitive, so

If A is in equilibrium with B, and B is in equilibrium with C, then A is in equilibrium with

C.

This known as the Zeroth Law of Thermodynamics5 .

2.4 Mathematical Interlude : Exact and Inexact Differentials

The differential
k
X
dF = Ai dxi (2.12)
i=1
is called exact if there is a function F (x1 , . . . , xk ) whose differential gives the right hand side of eqn.
2.188. In this case, we have
∂F ∂Ai ∂Aj
Ai = ⇐⇒ = ∀ i, j . (2.13)
∂xi ∂xj ∂xi
For exact differentials, the integral between fixed endpoints is path-independent:
ZB
dF = F (xB1 , . . . , xBk ) − F (xA1 , . . . , xAk ) , (2.14)
A

from which it follows that the integral of dF around any closed path must vanish:
I
dF = 0 . (2.15)

5
As we shall see further below, thermomechanical equilibrium in fact leads to constant p/T , and thermochemical equi-
librium to constant µ/T . If there is thermal equilibrium, then T is already constant, and so thermomechanical and thermo-
chemical equilibria then guarantee the constancy of p and µ.
38 CHAPTER 2. THERMODYNAMICS

Figure 2.6: Two distinct paths with identical endpoints.

When the cross derivatives are not identical, i.e. when ∂Ai /∂xj 6= ∂Aj /∂xi , the differential is inexact.
In this case, the integral of dF is path dependent, and does not depend solely on the endpoints.
As an example, consider the differential
dF = K1 y dx + K2 x dy . (2.16)
Let’s evaluate the integral of dF , which is the work done, along each of the two paths in Fig. 2.6:

ZxB ZyB
W (I) = K1 dx yA + K2 dy xB = K1 yA (xB − xA ) + K2 xB (yB − yA ) (2.17)
xA yA

ZxB ZyB
W (II) = K1 dx yB + K2 dy xA = K1 yB (xB − xA ) + K2 xA (yB − yA ) . (2.18)
xA yA

Note that in general W (I) 6= W (II) . Thus, if we start at point A, the kinetic energy at point B will depend
on the path taken, since the work done is path-dependent.
The difference between the work done along the two paths is
I
(I) (II)
W − W = dF = (K2 − K1 ) (xB − xA ) (yB − yA ) . (2.19)

Thus, we see that if K1 = K2 , the work is the same for the two paths. In fact, if K1 = K2 , the work
would be path-independent, and would depend only on the endpoints. This is true for any path, and
not just piecewise linear paths of the type depicted in Fig. 2.6. Thus, if K1 = K2 , we are justified in
using the notation dF for the differential in eqn. 2.16; explicitly, we then have F = K1 xy. However, if
K1 6= K2 , the differential is inexact, and we will henceforth write dF
¯ in such cases.
2.5. THE FIRST LAW OF THERMODYNAMICS 39

Figure 2.7: The first law of thermodynamics is a statement of energy conservation.

2.5 The First Law of Thermodynamics

2.5.1 Conservation of energy

The first law is a statement of energy conservation, and is depicted in Fig. 2.7. It says, quite simply, that
during a thermodynamic process, the change in a system’s internal energy E is given by the heat energy
Q added to the system, minus the work W done by the system:

∆E = Q − W . (2.20)

The differential form of this, the First Law of Thermodynamics, is

dE = d̄Q − dW
¯ . (2.21)

We use the symbol d¯ in the differentials dQ

¯ and dW¯ to remind us that these are inexact differentials.
The energy E, however, is a state function, hence dE is an exact differential.
Consider a volume V of fluid held in a flask, initially at temperature T0 , and held at atmospheric pressure.
The internal energy is then E0 = E(T0 , p, V ). Now let us contemplate changing the temperature in two
different ways. The first method (A) is to place the flask on a hot plate until the temperature of the fluid
rises to a value T1 . The second method (B) is to stir the fluid vigorously. In the first case, we add heat
QA > 0 but no work is done, so WA = 0. In the second case, if we thermally insulate the flask and use
a stirrer of very low thermal conductivity, then no heat is added, i.e. QB = 0. However, the stirrer does
work −WB > 0 on the fluid (remember W is the work done by the system). If we end up at the same
temperature T1 , then the final energy is E1 = E(T1 , p, V ) in both cases. We then have

∆E = E1 − E0 = QA = −WB . (2.22)

It also follows that for any cyclic transformation, where the state variables are the same at the beginning
and the end, we have
∆Ecyclic = Q − W = 0 =⇒ Q = W (cyclic) . (2.23)

2.5.2 Single component systems

A single component system is specified by three state variables. In many applications, the total number
of particles N is conserved, so it is useful to take N as one of the state variables. The remaining two can
40 CHAPTER 2. THERMODYNAMICS

be (T, V ) or (T, p) or (p, V ). The differential form of the first law says

dE = d̄Q − dW
¯
= d̄Q − p dV + µ dN . (2.24)

The quantity µ is called the chemical potential . We ask: how much heat is required in order to make an
infinitesimal change in temperature, pressure, volume, or particle number? We start by rewriting eqn.
2.24 as
dQ
¯ = dE + p dV − µ dN . (2.25)
We now must roll up our sleeves and do some work with partial derivatives.
• (T, V, N ) systems : If the state variables are (T, V, N ), we write
     
∂E ∂E ∂E
dE = dT + dV + dN . (2.26)
∂T V,N ∂V T,N ∂N T,V

Then   "  # "  #

∂E ∂E ∂E
dQ
¯ = dT + + p dV + − µ dN . (2.27)
∂T V,N ∂V T,N ∂N T,V

• (T, p, N ) systems : If the state variables are (T, p, N ), we write

     
∂E ∂E ∂E
dE = dT + dp + dN . (2.28)
∂T p,N ∂p T,N ∂N T,p

We also write      
∂V ∂V ∂V
dV = dT + dp + dN . (2.29)
∂T p,N ∂p T,N ∂N T,p
Then
"   #  "    #
∂E ∂V ∂E ∂V
dQ
¯ = +p dT + +p dp
∂T p,N ∂T p,N ∂p T,N ∂p T,N
"    # (2.30)
∂E ∂V
+ +p − µ dN .
∂N T,p ∂N T,p

• (p, V, N ) systems : If the state variables are (p, V, N ), we write

     
∂E ∂E ∂E
dE = dp + dV + dN . (2.31)
∂p V,N ∂V p,N ∂N p,V

Then   "  # "  #

∂E ∂E ∂E
dQ
¯ = dp + + p dV + − µ dN . (2.32)
∂p V,N ∂V p,N ∂N p,V

The heat capacity of a body, C, is by definition the ratio dQ/dT

¯ of the amount of heat absorbed by
the body to the associated infinitesimal change in temperature dT . The heat capacity will in general be
2.5. THE FIRST LAW OF THERMODYNAMICS 41

cp c̃p cp c̃p
SUBSTANCE (J/mol K) (J/g K) SUBSTANCE (J/mol K) (J/g K)
Air 29.07 1.01 H2 O (25◦ C) 75.34 4.181
Aluminum 24.2 0.897 H2 O (100◦+ C) 37.47 2.08
Copper 24.47 0.385 Iron 25.1 0.450
CO2 36.94 0.839 Lead 26.4 0.127
Diamond 6.115 0.509 Lithium 24.8 3.58
Ethanol 112 2.44 Neon 20.786 1.03
Gold 25.42 0.129 Oxygen 29.38 0.918
Helium 20.786 5.193 Paraffin (wax) 900 2.5
Hydrogen 28.82 5.19 Uranium 27.7 0.116
H2 O (−10◦ C) 38.09 2.05 Zinc 25.3 0.387

Table 2.1: Specific heat (at 25◦ C, unless otherwise noted) of some common substances. (Source:
Wikipedia.)

different if the body is heated at constant volume or at constant pressure. Setting dV = 0 gives, from
eqn. 2.27,
   
dQ
¯ ∂E
CV,N = = . (2.33)
dT V,N ∂T V,N

Similarly, if we set dp = 0, then eqn. 2.30 yields

     
dQ
¯ ∂E ∂V
Cp,N = = +p . (2.34)
dT p,N ∂T p,N ∂T p,N

Unless explicitly stated as otherwise, we shall assume that N is fixed, and will write CV for CV,N and Cp
for Cp,N .
The units of heat capacity are energy divided by temperature, e.g. J/K. The heat capacity is an extensive
quantity, scaling with the size of the system. If we divide by the number of moles N/NA , we obtain the
molar heat capacity, sometimes called the molar specific heat: c = C/ν, where ν = N/NA is the number
of moles of substance. Specific heat is also sometimes quoted in units of heat capacity per gram of
substance. We shall define
C c heat capacity per mole
c̃ = = = . (2.35)
mN M mass per mole

Here m is the mass per particle and M is the mass per mole: M = NA m.
Suppose we raise the temperature of a body from T = TA to T = TB . How much heat is required? We
have
ZTB
Q = dT C(T ) , (2.36)
TA
42 CHAPTER 2. THERMODYNAMICS

Figure 2.8: Heat capacity CV for one mole of hydrogen (H2 ) gas. At the lowest temperatures, only
translational degrees of freedom are relevant, and f = 3. At around 200 K, two rotational modes are
excitable and f = 5. Above 1000 K, the vibrational excitations begin to contribute. Note the logarithmic
temperature scale. (Data from H. W. Wooley et al., Jour. Natl. Bureau of Standards, 41, 379 (1948).)

where C = CV or C = Cp depending on whether volume or pressure is held constant. For ideal gases, as
we shall discuss below, C(T ) is constant, and thus

Q
Q = C(TB − TA ) =⇒ TB = TA + . (2.37)
C

In metals at very low temperatures one finds C = γT , where γ is a constant6 . We then have

ZTB
Q = dT C(T ) = 12 γ TB2 − TA2


(2.38)
TA
p
TB = TA2 + 2γ −1 Q . (2.39)

2.5.3 Ideal gases

The ideal gas equation of state is pV = N kB T . In order to invoke the formulae in eqns. 2.27, 2.30, and
2.32, we need to know the state function E(T, V, N ). A landmark experiment by Joule in the mid-19th
century established that the energy of a low density gas is independent of its volume7 . Essentially, a gas
at temperature T was allowed to freely expand from one volume V to a larger volume V 0 > V , with no
added heat Q and no work W done. Therefore the energy cannot change. What Joule found was that
the temperature also did not change. This means that E(T, V, N ) = E(T, N ) cannot be a function of the
volume.
6
In most metals, the difference between CV and Cp is negligible.
7
See the description in E. Fermi, Thermodynamics, pp. 22-23.
2.5. THE FIRST LAW OF THERMODYNAMICS 43

Since E is extensive, we conclude that

E(T, V, N ) = ν ε(T ) , (2.40)

where ν = N/NA is the number of moles of substance. Note that ν is an extensive variable. From eqns.
2.33 and 2.34, we conclude

CV (T ) = ν ε0 (T ) , Cp (T ) = CV (T ) + νR , (2.41)

where we invoke the ideal gas law to obtain the second of these. Empirically it is found that CV (T )
is temperature independent over a wide range of T , far enough from boiling point. We can then write
CV = ν cV , where ν ≡ N/NA is the number of moles, and where cV is the molar heat capacity. We then
have
cp = cV + R , (2.42)
where R = NA kB = 8.31457 J/mol K is the gas constant. We denote by γ = cp /cV the ratio of specific
heat at constant pressure and at constant volume.
From the kinetic theory of gases, one can show that

monatomic gases: cV = 23 R , cp = 52 R , γ= 5
3
diatomic gases: cV = 52 R , cp = 72 R , γ= 7
5
4
polyatomic gases: cV = 3R , cp = 4R , γ= 3 .

Digression : kinetic theory of gases

We will conclude in general from noninteracting classical statistical mechanics that the specific heat of a
substance is cv = 12 f R, where f is the number of phase space coordinates, per particle, for which there
is a quadratic kinetic or potential energy function. For example, a point particle has three translational
degrees of freedom, and the kinetic energy is a quadratic function of their conjugate momenta: H0 =
(p2x +p2y +p2z )/2m. Thus, f = 3. Diatomic molecules have two additional rotational degrees of freedom – we
don’t count rotations about the symmetry axis – and their conjugate momenta also appear quadratically
in the kinetic energy, leading to f = 5. For polyatomic molecules, all three Euler angles and their
conjugate momenta are in play, and f = 6.
The reason that f = 5 for diatomic molecules rather than f = 6 is due to quantum mechanics. While
translational eigenstates form a continuum, or are quantized in a box with ∆kα = 2π/Lα being very
small, since the dimensions Lα are macroscopic, angular momentum, and hence rotational kinetic energy,
is quantized. For rotations about a principal axis with very low moment of inertia I, the corresponding
energy scale ~2 /2I is very large, and a high temperature is required in order to thermally populate these
states. Thus, degrees of freedom with a quantization energy on the order or greater than ε0 are ‘frozen
out’ for temperatures T < ∼ ε0 /kB .
In solids, each atom is effectively connected to its neighbors by springs; such a potential arises from
quantum mechanical and electrostatic consideration of the interacting atoms. Thus, each degree of
freedom contributes to the potential energy, and its conjugate momentum contributes to the kinetic
energy. This results in f = 6. Assuming only lattice vibrations, then, the high temperature limit for
44 CHAPTER 2. THERMODYNAMICS

Figure 2.9: Molar heat capacities cV for three solids. The solid curves correspond to the predictions of
the Debye model, which we shall discuss later.

cV (T ) for any solid is predicted to be 3R = 24.944 J/mol K. This is called the Dulong-Petit law . The
high temperature limit is reached above the so-called Debye temperature, which is roughly proportional
to the melting temperature of the solid.
In table 2.1, we list cp and c̃p for some common substances at T = 25◦ C (unless otherwise noted).
Note that cp for the monatomic gases He and Ne is to high accuracy given by the value from kinetic
theory, cp = 25 R = 20.7864 J/mol K. For the diatomic gases oxygen (O2 ) and air (mostly N2 and O2 ),
kinetic theory predicts cp = 72 R = 29.10, which is close to the measured values. Kinetic theory predicts
cp = 4R = 33.258 for polyatomic gases; the measured values for CO2 and H2 O are both about 10%
higher.

2.5.4 Adiabatic transformations of ideal gases

Assuming dN = 0 and E = ν ε(T ), eqn. 2.27 tells us that

dQ
¯ = CV dT + p dV . (2.43)

Invoking the ideal gas law to write p = νRT /V , and remembering CV = ν cV , we have, setting dQ
¯ = 0,
dT R dV
+ =0. (2.44)
T cV V
We can immediately integrate to obtain

γ−1 = constant
T V

dQ
¯ =0 =⇒ pV γ = constant (2.45)

 γ 1−γ
T p = constant

where the second two equations are obtained from the first by invoking the ideal gas law. These are all
adiabatic equations of state. Note the difference between the adiabatic equation of state d(pV γ ) = 0 and
2.5. THE FIRST LAW OF THERMODYNAMICS 45

the isothermal equation of state d(pV ) = 0. Equivalently, we can write these three conditions as

V 2 T f = V02 T0f , pf V f +2 = pf0 V0f +2 , T f +2 p−2 = T0f +2 p−2

0 . (2.46)

It turns out that air is a rather poor conductor of heat. This suggests the following model for an adiabatic
atmosphere. The hydrostatic pressure decrease associated with an increase dz in height is dp = −%g dz,
where % is the density and g the acceleration due to gravity. Assuming the gas is ideal, the density can
be written as % = M p/RT , where M is the molar mass. Thus,
dp Mg
=− dz . (2.47)
p RT

If the height changes are adiabatic, then, from d(T γ p1−γ ) = 0, we have
γ − 1 T dp γ − 1 Mg
dT = =− dz , (2.48)
γ p γ R
with the solution  
γ − 1 Mg γ−1 z
T (z) = T0 − z= 1− T0 , (2.49)
γ R γ λ
where T0 = T (0) is the temperature at the earth’s surface, and

RT0
λ= . (2.50)
Mg

With M = 28.88 g and γ = 75 for air, and assuming T0 = 293 K, we find λ = 8.6 km, and dT /dz =
−(1 − γ −1 ) T0 /λ = −9.7 K/km. Note that in this model the atmosphere ends at a height zmax =
γλ/(γ − 1) = 30 km.
Again invoking the adiabatic equation of state, we can find p(z):
  γ   γ
p(z) T γ−1 γ − 1 z γ−1
= = 1− (2.51)
p0 T0 γ λ

Recall that  x k
ex = lim 1+ . (2.52)
k→∞ k
Thus, in the limit γ → 1, where k = γ/(γ − 1) → ∞, we have p(z) = p0 exp(−z/λ). Finally, since
% ∝ p/T from the ideal gas law, we have
  1
%(z) γ−1 z γ−1
= 1− . (2.53)
%0 γ λ

2.5.5 Adiabatic free expansion

Consider the situation depicted in Fig. 2.10. A quantity (ν moles) of gas in equilibrium at temperature
T and volume V1 is allowed to expand freely into an evacuated chamber of volume V2 by the removal of
46 CHAPTER 2. THERMODYNAMICS

a barrier. Clearly no work is done on or by the gas during this process, hence W = 0. If the walls are
everywhere insulating, so that no heat can pass through them, then Q = 0 as well. The First Law then
gives ∆E = Q − W = 0, and there is no change in energy.
If the gas is ideal, then since E(T, V, N ) = N cV T , then ∆E = 0 gives ∆T = 0, and there is no change in
temperature. (If the walls are insulating against the passage of heat, they must also prevent the passage
of particles, so ∆N = 0.) There is of course a change in volume: ∆V = V2 , hence there is a change in
pressure. The initial pressure is p = N kB T /V1 and the final pressure is p0 = N kB T /(V1 + V2 ).
If the gas is nonideal, then the temperature will in general change. Suppose E(T, V, N ) = α V x N 1−x T y ,
where α, x, and y are constants. This form is properly extensive: if V and N double, then E doubles. If
the volume changes from V to V 0 under an adiabatic free expansion, then we must have, from ∆E = 0,
x y x/y
T0
  
V 0 V
= =⇒ T =T · . (2.54)
V0 T V0

If x/y > 0, the temperature decreases upon the expansion. If x/y < 0, the temperature increases.
Without an equation of state, we can’t say precisely what happens to the pressure, although we know on
general grounds that it must decrease because, as we shall see, thermodynamic stability entails a positive
isothermal compressibility: κT = − V1 ∂V


∂p T,N > 0.
Adiabatic free expansion of a gas is a spontaneous process, arising due to the natural internal dynamics
of the system. It is also irreversible. If we wish to take the gas back to its original state, we must do
work on it to compress it. If the gas is ideal, then the initial and final temperatures are identical, so we
can place the system in thermal contact with a reservoir at temperature T and follow a thermodynamic
path along an isotherm. The work done on the gas during compression is then

ZVi    
dV Vf V2
W = −N kB T = N kB T ln = N kB T ln 1 + (2.55)
V Vi V1
Vf

Figure 2.10: In the adiabatic free expansion of a gas, there is volume expansion with no work or heat
exchange with the environment: ∆E = Q = W = 0.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 47

Figure 2.11: A perfect engine would extract heat Q from a thermal reservoir at some temperature T and
convert it into useful mechanical work W . This process is alas impossible, according to the Second Law
of thermodynamics. The inverse process, where work W is converted into heat Q, is always possible.

R
The work done by the gas is W = p dV = −W. During the compression, heat energy Q = W < 0 is
transferred to the gas from the reservoir. Thus, Q = W > 0 is given off by the gas to its environment.

2.6 Heat Engines and the Second Law of Thermodynamics

2.6.1 There’s no free lunch so quit asking

A heat engine is a device which takes a thermodynamic system through a repeated cycle which can be
represented as a succession of equilibrium states: A → B → C · · · → A. The net result of such a cyclic
process is to convert heat into mechanical work, or vice versa.
For a system in equilibrium at temperature T , there is a thermodynamically large amount of internal
energy stored in the random internal motion of its constituent particles. Later, when we study statistical
mechanics, we will see how each ‘quadratic’ degree of freedom in the Hamiltonian contributes 12 kB T to the
total internal energy. An immense body in equilibrium at temperature T has an enormous heat capacity
C, hence extracting a finite quantity of heat Q from it results in a temperature change ∆T = −Q/C which
is utterly negligible. Such a body is called a heat bath, or thermal reservoir . A perfect engine would, in
each cycle, extract an amount of heat Q from the bath and convert it into work. Since ∆E = 0 for a
cyclic process, the First Law then gives W = Q. This situation is depicted schematically in Fig. 2.11.
One could imagine running this process virtually indefinitely, slowly sucking energy out of an immense
heat bath, converting the random thermal motion of its constituent molecules into useful mechanical
work. Sadly, this is not possible:

A transformation whose only final result is to extract heat from a source at fixed temperature
and transform that heat into work is impossible.

This is known as the Postulate of Lord Kelvin. It is equivalent to the postulate of Clausius,

A transformation whose only result is to transfer heat from a body at a given temperature to
a body at higher temperature is impossible.
48 CHAPTER 2. THERMODYNAMICS

These postulates which have been repeatedly validated by empirical observations, constitute the Second
Law of Thermodynamics.

2.6.2 Engines and refrigerators

While it is not possible to convert heat into work with 100% efficiency, it is possible to transfer heat
from one thermal reservoir to another one, at lower temperature, and to convert some of that heat into
work. This is what an engine does. The energy accounting for one cycle of the engine is depicted in the
left hand panel of Fig. 2.12. An amount of heat Q2 > 0 is extracted- from the reservoir at temperature
T2 . Since the reservoir is assumed to be enormous, its temperature change ∆T2 = −Q2 /C2 is negligible,
and its temperature remains constant – this is what it means for an object to be a reservoir. A lesser
amount of heat, Q1 , with 0 < Q1 < Q2 , is deposited in a second reservoir at a lower temperature T1 . Its
temperature change ∆T1 = +Q1 /C1 is also negligible. The difference W = Q2 − Q1 is extracted as useful
work. We define the efficiency, η, of the engine as the ratio of the work done to the heat extracted from
the upper reservoir, per cycle:
W Q
η= =1− 1 . (2.56)
Q2 Q2
This is a natural definition of efficiency, since it will cost us fuel to maintain the temperature of the upper
reservoir over many cycles of the engine. Thus, the efficiency is proportional to the ratio of the work
done to the cost of the fuel.
A refrigerator works according to the same principles, but the process runs in reverse. An amount of heat
Q1 is extracted from the lower reservoir – the inside of our refrigerator – and is pumped into the upper
reservoir. As Clausius’ form of the Second Law asserts, it is impossible for this to be the only result of
our cycle. Some amount of work W must be performed on the refrigerator in order for it to extract the
heat Q1 . Since ∆E = 0 for the cycle, a heat Q2 = W + Q1 must be deposited into the upper reservoir
during each cycle. The analog of efficiency here is called the coefficient of refrigeration, κ, defined as
Q1 Q1
κ= = . (2.57)
W Q2 − Q1

Thus, κ is proportional to the ratio of the heat extracted to the cost of electricity, per cycle.
Please note the deliberate notation here. I am using symbols Q and W to denote the heat supplied to
the engine (or refrigerator) and the work done by the engine, respectively, and Q and W to denote the
heat taken from the engine and the work done on the engine.
A perfect engine has Q1 = 0 and η = 1; a perfect refrigerator has Q1 = Q2 and κ = ∞. Both violate
the Second Law. Sadi Carnot8 (1796 – 1832) realized that a reversible cyclic engine operating between
two thermal reservoirs must produce the maximum amount of work W , and that the amount of work
produced is independent of the material properties of the engine. We call any such engine a Carnot
engine.
The efficiency of a Carnot engine may be used to define a temperature scale. We know from Carnot’s
observations that the efficiency ηC can only be a function of the temperatures T1 and T2 : ηC = ηC (T1 , T2 ).
8
Carnot died during cholera epidemic of 1832. His is one of the 72 names engraved on the Eiffel Tower.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 49

Figure 2.12: An engine (left) extracts heat Q2 from a reservoir at temperature T2 and deposits a
smaller amount of heat Q1 into a reservoir at a lower temperature T1 , during each cycle. The difference
W = Q2 − Q1 is transformed into mechanical work. A refrigerator (right) performs the inverse process,
drawing heat Q1 from a low temperature reservoir and depositing heat Q2 = Q1 + W into a high
temperature reservoir, where W is the mechanical (or electrical) work done per cycle.

We can then define

T1
≡ 1 − ηC (T1 , T2 ) . (2.58)
T2
Below, in §2.6.4, we will see that how, using an ideal gas as the ‘working substance’ of the Carnot engine,
this temperature scale coincides precisely with the ideal gas temperature scale from §2.2.4.

2.6.3 Nothing beats a Carnot engine

The Carnot engine is the most efficient engine possible operating between two thermal reservoirs. To see
this, let’s suppose that an amazing wonder engine has an efficiency even greater than that of the Carnot
engine. A key feature of the Carnot engine is its reversibility – we can just go around its cycle in the
opposite direction, creating a Carnot refrigerator. Let’s use our notional wonder engine to drive a Carnot
refrigerator, as depicted in Fig. 2.13.
We assume that
W W0
= ηwonder > ηCarnot = 0 . (2.59)
Q2 Q2
But from the figure, we have W = W 0 , and therefore the heat energy Q02 − Q2 transferred to the upper
reservoir is positive. From
W = Q2 − Q1 = Q02 − Q01 = W 0 , (2.60)

we see that this is equal to the heat energy extracted from the lower reservoir, since no external work is
done on the system:
Q02 − Q2 = Q01 − Q1 > 0 . (2.61)
50 CHAPTER 2. THERMODYNAMICS

Figure 2.13: A wonder engine driving a Carnot refrigerator.

Therefore, the existence of the wonder engine entails a violation of the Second Law. Since the Second
Law is correct – Lord Kelvin articulated it, and who are we to argue with a Lord ? – the wonder engine
cannot exist.
We further conclude that all reversible engines running between two thermal reservoirs have the same
efficiency, which is the efficiency of a Carnot engine. For an irreversible engine, we must have

W Q T
η= = 1 − 1 ≤ 1 − 1 = ηC . (2.62)
Q2 Q2 T2

Thus,
Q2 Q1
− ≤0. (2.63)
T2 T1

2.6.4 The Carnot cycle

Let us now consider a specific cycle, known as the Carnot cycle, depicted in Fig. 2.14. The cycle consists
of two adiabats and two isotherms. The work done per cycle is simply the area inside the curve on our
p − V diagram: I
W = p dV . (2.64)

The gas inside our Carnot engine is called the ‘working substance’. Whatever it may be, the system
obeys the First Law,
dE = d̄Q − dW¯ = d̄Q − p dV . (2.65)

We will now assume that the working material is an ideal gas, and we compute W as well as Q1 and Q2
to find the efficiency of this cycle. In order to do this, we will rely upon the ideal gas equations,

νRT
E= , pV = νRT , (2.66)
γ−1
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 51

Figure 2.14: The Carnot cycle consists of two adiabats (dark red) and two isotherms (blue).

where γ = cp /cv = 1 + f2 , where f is the effective number of molecular degrees of freedom contributing
to the internal energy. Recall f = 3 for monatomic gases, f = 5 for diatomic gases, and f = 6 for
polyatomic gases. The finite difference form of the first law is

∆E = Ef − Ei = Qif − Wif , (2.67)

where i denotes the initial state and f the final state.

AB: This stage is an isothermal expansion at temperature T2 . It is the ‘power stroke’ of the engine. We
have
ZVB  
νRT2 VB
WAB = dV = νRT2 ln (2.68)
V VA
VA
νRT2
EA = EB = , (2.69)
γ−1
hence  
VB
QAB = ∆EAB + WAB = νRT2 ln . (2.70)
VA
BC: This stage is an adiabatic expansion. We have

QBC = 0 (2.71)
νR
∆EBC = EC − EB = (T − T2 ) . (2.72)
γ−1 1
The energy change is negative, and the heat exchange is zero, so the engine still does some work
during this stage:
νR
WBC = QBC − ∆EBC = (T − T1 ) . (2.73)
γ−1 2
52 CHAPTER 2. THERMODYNAMICS

CD: This stage is an isothermal compression, and we may apply the analysis of the isothermal expansion,
mutatis mutandis:
ZVD  
νRT1 VD
WCD = dV = νRT1 ln (2.74)
V VC
VC
νRT1
EC = ED = , (2.75)
γ−1
hence  
VD
QCD = ∆ECD + WCD = νRT1 ln . (2.76)
VC
DA: This last stage is an adiabatic compression, and we may draw on the results from the adiabatic
expansion in BC:
QDA = 0 (2.77)
νR
∆EDA = ED − EA = (T − T1 ) . (2.78)
γ−1 2
The energy change is positive, and the heat exchange is zero, so work is done on the engine:
νR
WDA = QDA − ∆EDA = (T − T2 ) . (2.79)
γ−1 1

We now add up all the work values from the individual stages to get for the cycle
W = WAB + WBC + WCD + WDA
(2.80)
   
VB VD
= νRT2 ln + νRT1 ln .
VA VC
Since we are analyzing a cyclic process, we must have ∆E = 0, we must have Q = W , which can of course
be verified explicitly, by computing Q = QAB + QBC + QCD + QDA . To finish up, recall the adiabatic ideal
gas equation of state, d(T V γ−1 ) = 0. This tells us that
T2 VBγ−1 = T1 VCγ−1 (2.81)
γ−1 γ−1
T2 VA = T1 VD . (2.82)
Dividing these two equations, we find
VB V
= C , (2.83)
VA VD
and therefore
 
VB
W = νR(T2 − T1 ) ln (2.84)
VA
 
VB
QAB = νRT2 ln . (2.85)
VA
Finally, the efficiency is given by the ratio of these two quantities:
W T
η= =1− 1 . (2.86)
QAB T2
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 53

Figure 2.15: A Stirling cycle consists of two isotherms (blue) and two isochores (green).

2.6.5 The Stirling cycle

Many other engine cycles are possible. The Stirling cycle, depicted in Fig. 2.15, consists of two isotherms
and two isochores. Recall the isothermal ideal gas equation of state, d(pV ) = 0. Thus, for an ideal gas
Stirling cycle, we have
pA V1 = pB V2 , pD V1 = pC V2 , (2.87)
which says
pB p V
= C = 1 . (2.88)
pA pD V2

AB: This isothermal expansion is the power stroke. Assuming ν moles of ideal gas throughout, we have
pV = νRT2 = p1 V1 , hence
ZV2  
νRT2 V2
WAB = dV = νRT2 ln . (2.89)
V V1
V1

Since AB is an isotherm, we have EA = EB , and from ∆EAB = 0 we conclude QAB = WAB .

BC: Isochoric cooling. Since dV = 0 we have WBC = 0. The energy change is given by
νR(T1 − T2 )
∆EBC = EC − EB = , (2.90)
γ−1
which is negative. Since WBC = 0, we have QBC = ∆EBC .

CD: Isothermal compression. Clearly

ZV1  
νRT1 V2
WCD = dV = −νRT1 ln . (2.91)
V V1
V2
54 CHAPTER 2. THERMODYNAMICS

Since CD is an isotherm, we have EC = ED , and from ∆ECD = 0 we conclude QCD = WCD .

DA: Isochoric heating. Since dV = 0 we have WDA = 0. The energy change is given by
νR(T2 − T1 )
∆EDA = EA − ED = , (2.92)
γ−1
which is positive, and opposite to ∆EBC . Since WDA = 0, we have QDA = ∆EDA .

We now add up all the work contributions to obtain

W = WAB + WBC + WCD + WDA
(2.93)
 
V2
= νR(T2 − T1 ) ln .
V1
The cycle efficiency is once again
W T
η= =1− 1 . (2.94)
QAB T2

2.6.6 The Otto and Diesel cycles

The Otto cycle is a rough approximation to the physics of a gasoline engine. It consists of two adiabats
and two isochores, and is depicted in Fig. 2.16. Assuming an ideal gas, along the adiabats we have
d(pV γ ) = 0. Thus,
pA V1γ = pB V2γ , pD V1γ = pC V2γ , (2.95)
which says  γ
pB p V1
= C = . (2.96)
pA pD V2

AB: Adiabatic expansion, the power stroke. The heat transfer is QAB = 0, so from the First Law we
have WAB = −∆EAB = EA − EB , thus
"  γ−1 #
pA V1 − pB V2 pA V1 V1
WAB = = 1− . (2.97)
γ−1 γ−1 V2

Note that this result can also be obtained from the adiabatic equation of state pV γ = pA V1γ :

ZV2 ZV2 "  γ−1 #

γ −γ pA V1 V1
WAB = p dV = pA V1 dV V = 1− . (2.98)
γ−1 V2
V1 V1

BC: Isochoric cooling (exhaust); dV = 0 hence WBC = 0. The heat QBC absorbed is then
V2
QBC = EC − EB = (p − pB ) . (2.99)
γ−1 C
In a realistic engine, this is the stage in which the old burned gas is ejected and new gas is inserted.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 55

Figure 2.16: An Otto cycle consists of two adiabats (dark red) and two isochores (green).

Figure 2.17: A Diesel cycle consists of two adiabats (dark red), one isobar (light blue), and one isochore
(green).

CD: Adiabatic compression; QCD = 0 and WCD = EC − ED :

"  γ−1 #
pC V2 − pD V1 pD V1 V1
WCD = =− 1− . (2.100)
γ−1 γ−1 V2

DA: Isochoric heating, i.e. the combustion of the gas. As with BC we have dV = 0, and thus WDA = 0.
The heat QDA absorbed by the gas is then

V1
QDA = EA − ED = (p − pD ) . (2.101)
γ−1 A
56 CHAPTER 2. THERMODYNAMICS

The total work done per cycle is then

W = WAB + WBC + WCD + WDA
"  γ−1 #
(pA − pD )V1 V1 (2.102)
= 1− ,
γ−1 V2

and the efficiency is defined to be

 γ−1
W V1
η≡ =1− . (2.103)
QDA V2
The ratio V2 /V1 is called the compression ratio. We can make our Otto cycle more efficient simply by
increasing the compression ratio. The problem with this scheme is that if the fuel mixture becomes too
hot, it will spontaneously ‘preignite’, and the pressure will jump up before point D in the cycle is reached.
A Diesel engine avoids preignition by compressing the air only, and then later spraying the fuel into the
cylinder when the air temperature is sufficient for fuel ignition. The rate at which fuel is injected is
adjusted so that the ignition process takes place at constant pressure. Thus, in a Diesel engine, step DA
is an isobar. The compression ratio is r ≡ VB /VD , and the cutoff ratio is s ≡ VA /VD . This refinement of
the Otto cycle allows for higher compression ratios (of about 20) in practice, and greater engine efficiency.
For the Diesel cycle, we have, briefly,
pA VA − pB VB pC VC − pD VD
W = pA (VA − VD ) + +
γ−1 γ−1
(2.104)
γ pA (VA − VD ) (pB − pC )VB
= −
γ−1 γ−1
and
γ pA (VA − VD )
QDA = . (2.105)
γ−1
To find the efficiency, we will need to eliminate pB and pC in favor of pA using the adiabatic equation of
state d(pV γ ) = 0. Thus,  γ  γ
VA VD
pB = pA · , pC = pA · , (2.106)
VB VB
where we’ve used pD = pA and VC = VB . Putting it all together, the efficiency of the Diesel cycle is
W 1 r1−γ (sγ − 1)
η= =1− . (2.107)
QDA γ s−1

2.6.7 The Joule-Brayton cycle

Our final example is the Joule-Brayton cycle, depicted in Fig. 2.18, consisting of two adiabats and two
isobars. Along the adiabats we have Thus,
p2 VAγ = p1 VDγ , p2 VBγ = p1 VCγ , (2.108)
which says
 γ −1
VD V p2
= C = . (2.109)
VA VB p1
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 57

Figure 2.18: A Joule-Brayton cycle consists of two adiabats (dark red) and two isobars (light blue).

AB: This isobaric expansion at p = p2 is the power stroke. We have

ZVB
WAB = dV p2 = p2 (VB − VA ) (2.110)
VA

p2 (VB − VA )
∆EAB = EB − EA = (2.111)
γ−1
γ p2 (VB − VA )
QAB = ∆EAB + WAB = . (2.112)
γ−1

BC: Adiabatic expansion; QBC = 0 and WBC = EB − EC . The work done by the gas is
 
p2 VB − p1 VC p2 VB p1 VC
WBC = = 1− ·
γ−1 γ−1 p2 VB
"  1−γ −1 # (2.113)
p2 VB p1
= 1− .
γ−1 p2

CD: Isobaric compression at p = p1 .

ZVD  1−γ −1
p1
WCD = dV p1 = p1 (VD − VC ) = −p2 (VB − VA ) (2.114)
p2
VC

p1 (VD − VC )
∆ECD = ED − EC = (2.115)
γ−1
 1−γ −1
γ p2 p1
QCD = ∆ECD + WCD =− (V − VA ) . (2.116)
γ−1 B p2
58 CHAPTER 2. THERMODYNAMICS

DA: Adiabatic expansion; QDA = 0 and WDA = ED − EA . The work done by the gas is
 
p1 VD − p2 VA p2 VA p1 VD
WDA = =− 1− ·
γ−1 γ−1 p2 VA
"  1−γ −1 # (2.117)
p2 VA p1
=− 1− .
γ−1 p2

The total work done per cycle is then

W = WAB + WBC + WCD + WDA
"  1−γ −1 #
γ p2 (VB − VA ) p1 (2.118)
= 1−
γ−1 p2

and the efficiency is defined to be

 1−γ −1
W p1
η≡ =1− . (2.119)
QAB p2

2.6.8 Carnot engine at maximum power output

While the Carnot engine described above in §2.6.4 has maximum efficiency, it is practically useless,
because the isothermal processes must take place infinitely slowly in order for the working material to
remain in thermal equilibrium with each reservoir. Thus, while the work done per cycle is finite, the cycle
period is infinite, and the engine power is zero.
A modification of the ideal Carnot cycle is necessary to create a practical engine. The idea9 is as follows.
During the isothermal expansion stage, the working material is maintained at a temperature T2w < T2 .
The temperature difference between the working material and the hot reservoir drives a thermal current,
dQ
¯ 2
= κ2 (T2 − T2w ) . (2.120)
dt
Here, κ2 is a transport coefficient which describes the thermal conductivity of the chamber walls, multiplied
by a geometric parameter (which is the ratio of the total wall area to its thickness). Similarly, during the
isothermal compression, the working material is maintained at a temperature T1w > T1 , which drives a
thermal current to the cold reservoir,
dQ
¯ 1
= κ1 (T1w − T1 ) . (2.121)
dt
Now let us assume that the upper isothermal stage requires a duration ∆t2 and the lower isotherm a
duration ∆t1 . Then

Q2 = κ2 ∆t2 (T2 − T2w ) (2.122)

Q1 = κ1 ∆t1 (T1w − T1 ) . (2.123)
9
See F. L. Curzon and B. Ahlborn, Am. J. Phys. 43, 22 (1975). I am grateful to Professor Asle Sudbø for correcting a
typo in one expression and providing a simplified form of another.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 59

Power source T1 (◦ C) T2 (◦ C) ηCarnot η (theor.) η (obs.)

West Thurrock (UK)
Coal Fired Steam Plant ∼ 25 565 0.641 0.40 0.36
CANDU (Canada)
PHW Nuclear Reactor ∼ 25 300 0.480 0.28 0.30
Larderello (Italy)
Geothermal Steam Plant ∼ 80 250 0.323 0.175 0.16

Table 2.2: Observed performances of real heat engines, taken from table 1 from Curzon and Albhorn
(1975).

Since the engine is reversible, we must have

Q1 Q
= 2 , (2.124)
T1w T2w

which says
∆t1 κ T (T − T2w )
= 2 1w 2 . (2.125)
∆t2 κ1 T2w (T1w − T1 )
The power is
Q2 − Q1
P = , (2.126)
(1 + α) (∆t1 + ∆t2 )
where we assume that the adiabatic stages require a combined time of α (∆t1 + ∆t2 ). Thus, we find

κ1 κ2 (T2w − T1w ) (T1w − T1 ) (T2 − T2w )

P = · (2.127)
1 + α κ1 T2w (T1w − T1 ) + κ2 T1w (T2 − T2w )

We optimize the engine by maximizing P with respect to the temperatures T1w and T2w . This yields
p
T2 − T1 T2
T2w = T2 − p (2.128)
1 + κ2 /κ1
p
T1 T2 − T1
T1w = T1 + p . (2.129)
1 + κ1 /κ2

The efficiency at maximum power is then

s
Q − Q1 T T1
η= 2 = 1 − 1w = 1 − . (2.130)
Q2 T2w T2

One also finds at maximum power s

∆t2 κ1
= . (2.131)
∆t1 κ2
60 CHAPTER 2. THERMODYNAMICS

Finally, the maximized power is

p p !2
κ κ T2 − T1
Pmax = 1 2 p p . (2.132)
1+α κ1 + κ2

Table 2.2, taken from the article of Curzon and Albhorn (1975), shows how the efficiency of this practical
Carnot cycle, given by eqn. 2.130, rather accurately predicts the efficiencies of functioning power plants.

2.7 The Entropy

2.7.1 Entropy and heat

The Second Law guarantees us that an engine operating between two heat baths at temperatures T1 and
T2 must satisfy
Q1 Q2
+ ≤0, (2.133)
T1 T2
with the equality holding for reversible processes. This is a restatement of eqn. 2.63, after writing
Q1 = −Q1 for the heat transferred to the engine from reservoir #1. Consider now an arbitrary curve in
the p − V plane. We can describe such a curve, to arbitrary accuracy, as a combination of Carnot cycles,
as shown in Fig. 2.19. Each little Carnot cycle consists of two adiabats and two isotherms. We then
conclude
XQ I
dQ
¯
i
−→ ≤0, (2.134)
Ti T
i C
with equality holding if all the cycles are reversible. Rudolf Clausius, in 1865, realized that one could
then define a new state function, which he called the entropy, S, that depended only on the initial and
final states of a reversible process:

ZB
dQ
¯ dQ
¯
dS = =⇒ SB − SA = . (2.135)
T T
A

Since Q is extensive, so is S; the units of entropy are [S] = J/K.

2.7.2 The Third Law of Thermodynamics

Eqn. 2.135 determines the entropy up to a constant. By choosing a standard state Υ, we can define
SΥ = 0, and then by taking A = Υ in the above equation, we can define the absolute entropy S for
any state. However, it turns out that this seemingly arbitrary constant SΥ in the entropy does have
consequences, for example in the theory of gaseous equilibrium. The proper definition of entropy, from
the point of view of statistical mechanics, will lead us to understand how the zero temperature entropy
of a system is related to its quantum mechanical ground state degeneracy. Walther Nernst, in 1906,
articulated a principle which is sometimes called the Third Law of Thermodynamics,
2.7. THE ENTROPY 61

The entropy of every system at absolute zero temperature always vanishes.

Again, this is not quite correct, and quantum mechanics tells us that S(T = 0) = kB ln g, where g is the
ground state degeneracy. Nernst’s law holds when g = 1.
We can combine the First and Second laws to write

dE + dW
¯ = d̄Q ≤ T dS , (2.136)

where the equality holds for reversible processes.

2.7.3 Entropy changes in cyclic processes

For a cyclic process, whether reversible or not, the change in entropy around a cycle is zero: ∆SCYC = 0.
This is because the entropy S is a state function, with a unique value for every equilibrium state. A
cyclical process returns to the same equilibrium state, hence S must return as well to its corresponding
value from the previous cycle.
Consider now a general engine, as in Fig. 2.12. Let us compute the total entropy change in the entire
Universe over one cycle. We have

(∆S)TOTAL = (∆S)ENGINE + (∆S)HOT + (∆S)COLD , (2.137)

written as a sum over entropy changes of the engine itself, the hot reservoir, and the cold reservoir10 .
10
We neglect any interfacial contributions to the entropy change, which will be small compared with the bulk entropy
change in the thermodynamic limit of large system size.

Figure 2.19: An arbitrarily shaped cycle in the p − V plane can be decomposed into a number of smaller
Carnot cycles. Red curves indicate isotherms and blue curves adiabats, with γ = 35 .
62 CHAPTER 2. THERMODYNAMICS

Clearly (∆S)ENGINE = 0. The changes in the reservoir entropies are

dQ Q
Z
¯ HOT
(∆S)HOT = =− 2 < 0 (2.138)
T T2
T =T2
dQ Q Q
Z
¯ COLD
(∆S)COLD = = 1 =− 1 > 0, (2.139)
T T1 T1
T =T1

because the hot reservoir loses heat Q2 > 0 to the engine, and the cold reservoir gains heat Q1 = −Q1 > 0
from the engine. Therefore,  
Q1 Q2
(∆S)TOTAL = − + ≥0. (2.140)
T1 T2
Thus, for a reversible cycle, the net change in the total entropy of the engine plus reservoirs is zero. For
an irreversible cycle, there is an increase in total entropy, due to spontaneous processes.

2.7.4 Gibbs-Duhem relation

Recall eqn. 2.6: X X

dW
¯ =− yj dXj − µa dNa . (2.141)
j a

For reversible systems, we can therefore write

X X
dE = T dS + yj dXj + µa dNa . (2.142)
j a

This says that the energy E is a function of the entropy S, the generalized displacements {Xj }, and the
particle numbers {Na }:

E = E S, {Xj }, {Na } . (2.143)
Furthermore, we have
  !  
∂E ∂E ∂E
T = , yj = , µa = (2.144)
∂S {Xj ,Na } ∂Xj ∂Na S,{Xj ,Nb(6=a) }
S,{Xi(6=j) ,Na }

Since E and all its arguments are extensive, we have

λE = E λS, {λXj }, {λNa } . (2.145)

We now differentiate the LHS and RHS above with respect to λ, setting λ = 1 afterward. The result is

∂E X ∂E X ∂E
E=S + Xj + Na
∂S ∂Xj a
∂Na
j
X X (2.146)
= TS + yj Xj + µa Na .
j a
2.7. THE ENTROPY 63

Mathematically astute readers will recognize this result as an example of Euler’s theorem for homogeneous
functions. Taking the differential of eqn. 2.146, and then subtracting eqn. 2.142, we obtain
X X
S dT + Xj dyj + Na dµa = 0 . (2.147)
j a

This is called the Gibbs-Duhem relation. It says that there is one equation of state which may be written
in terms of all the intensive quantities alone. For example, for a single component system, we must have
p = p(T, µ), which follows from
S dT − V dp + N dµ = 0 . (2.148)

2.7.5 Entropy for an ideal gas

For an ideal gas, we have E = 21 f N kB T , and

1 p µ
dS = dE + dV − dN
T T T   (2.149)
1 dT p 1 µ
= 2 f N kB + dV + 2 f kB − dN .
T T T

Invoking the ideal gas equation of state pV = N kB T , we have

dS N = 12 f N kB d ln T + N kB d ln V .

(2.150)

Integrating, we obtain
S(T, V, N ) = 21 f N kB ln T + N kB ln V + ϕ(N ) , (2.151)

where ϕ(N ) is an arbitrary function. Extensivity of S places restrictions on ϕ(N ), so that the most
general case is
 
1 V
S(T, V, N ) = 2 f N kB ln T + N kB ln + Na , (2.152)
N
where a is a constant. Equivalently, we could write
   
1 E V
S(E, V, N ) = 2 f N kB ln + N kB ln + Nb , (2.153)
N N

where b = a − 21 f kB ln( 12 f kB ) is another constant. When we study statistical mechanics, we will find that
for the monatomic ideal gas the entropy is
"  #
5 V
S(T, V, N ) = N kB 2 + ln , (2.154)
N λ3T
p
where λT = 2π~2 /mkB T is the thermal wavelength, which involved Planck’s constant. Let’s now
contrast two illustrative cases.
64 CHAPTER 2. THERMODYNAMICS

• Adiabatic free expansion – Suppose the volume freely expands from Vi to Vf = r Vi , with r > 1. Such
an expansion can be effected by a removal of a partition between two chambers that are otherwise
thermally insulated (see Fig. 2.10). We have already seen how this process entails

∆E = Q = W = 0 . (2.155)

But the entropy changes! According to eqn. 2.153, we have

∆S = Sf − Si = N kB ln r . (2.156)

• Reversible adiabatic expansion – If the gas expands quasistatically and reversibly, then S =
S(E, V, N ) holds everywhere along the thermodynamic path. We then have, assuming dN = 0,
dE dV
0 = dS = 12 f N kB + N kB
E V (2.157)
= N kB d ln V E f /2 .

Integrating, we find
 2/f
E V0
= . (2.158)
E0 V
Thus,
Ef = r−2/f Ei ⇐⇒ Tf = r−2/f Ti . (2.159)

2.7.6 Example system

Consider a model thermodynamic system for which

aS 3
E(S, V, N ) = , (2.160)
NV
where a is a constant. We have
dE = T dS − p dV + µ dN , (2.161)
and therefore
3aS 2
 
∂E
T = = (2.162)
∂S V,N NV
aS 3
 
∂E
p=− = (2.163)
∂V S,N NV 2
aS 3
 
∂E
µ= =− 2 . (2.164)
∂N S,V N V

Choosing any two of these equations, we can eliminate S, which is inconvenient for experimental purposes.
This yields three equations of state,

T3 V T3 N p N
= 27a , = 27a , =− , (2.165)
p2 N µ2 V µ V
2.7. THE ENTROPY 65

only two of which are independent.

What about CV and Cp ? To find CV , we recast eqn. 2.162 as
 1/2
NV T
S= . (2.166)
3a
We then have 1/2
N T2
  
∂S 1 NV T
CV = T = = , (2.167)
∂T V,N 2 3a 18a p
where the last equality on the RHS follows upon invoking the first of the equations of state in eqn. 2.165.
To find Cp , we eliminate V from eqns. 2.162 and 2.163, obtaining T 2 /p = 9aS/N . From this we obtain

2N T 2
 
∂S
Cp = T = . (2.168)
∂T p,N 9a p

Thus, Cp /CV = 4.

We can derive still more. To find the isothermal compressibility κT = − V1 ∂V

, use the first of the
∂p T,N
equations of state in eqn. 2.165. To derive the adiabatic compressibility κS = − V1 ∂V


∂p S,N , use eqn.
2.163, and then eliminate the inconvenient variable S.
Suppose we use this system as the working substance for a Carnot engine. Let’s compute the work done
and the engine efficiency. To do this, it is helpful to eliminate S in the expression for the energy, and to
rewrite the equation of state:
r r
N 1/2 3/2 N T 3/2
E = pV = V T , p= . (2.169)
27a 27a V 1/2
We assume dN = 0 throughout. We now see that for isotherms,
E
dT = 0 : √ = constant (2.170)
V
Furthermore, since r
 N 3/2 dV 
dW
¯ T = T 1/2
= 2 dE T , (2.171)
27a V
we conclude that

dT = 0 : Wif = 2(Ef − Ei ) , Qif = Ef − Ei + Wif = 3(Ef − Ei ) . (2.172)

For adiabats, eqn. 2.162 says d(T V ) = 0, and therefore

E
dQ
¯ = 0 : T V = constant , = constant , EV = constant (2.173)
T
as well as Wif = Ei − Ef . We can use these relations to derive the following:
s s
VB T1 VB T
EB = E , EC = E , ED = 1 EA . (2.174)
VA A T2 VA A T2
66 CHAPTER 2. THERMODYNAMICS

Now we can write

s !
VB
WAB = 2(EB − EA ) = 2 − 1 EA (2.175)
VA
s !
VB T1
WBC = (EB − EC ) = 1− EA (2.176)
VA T2
s !
T1 VB
WCD = 2(ED − EC ) = 2 1− EA (2.177)
T2 VA
!
T1
WDA = (ED − EA ) = − 1 EA (2.178)
T2

Adding up all the work, we obtain

W = WAB + WBC + WCD + WDA

s ! ! (2.179)
VB T1
=3 −1 1− EA .
VA T2

Since s !
3 VB
QAB = 3(EB − EA ) = 2 WAB =3 − 1 EA , (2.180)
VA
we find once again
W T
η= =1− 1 . (2.181)
QAB T2

2.7.7 Measuring the entropy of a substance

If we can measure the heat capacity CV (T ) or Cp (T ) of a substance as a function of temperature down to

the lowest temperatures, then we can measure the entropy. At constant pressure, for example, we have
T dS = Cp dT , hence
ZT
Cp (T 0 )
S(p, T ) = S(p, T = 0) + dT 0 . (2.182)
T0
0

The zero temperature entropy is S(p, T = 0) = kB ln g where g is the quantum ground state degeneracy
at pressure p. In all but highly unusual cases, g = 1 and S(p, T = 0) = 0.

2.8 Thermodynamic Potentials

Thermodynamic systems may do work on their environments. Under certain constraints, the work done
may be bounded from above by the change in an appropriately defined thermodynamic potential .
2.8. THERMODYNAMIC POTENTIALS 67

2.8.1 Energy E

Suppose we wish to create a thermodynamic system from scratch. Let’s imagine that we create it from
scratch in a thermally insulated box of volume V . The work we must to to assemble the system is
then W = E . After we bring all the constituent particles together, pulling them in from infinity (say),
the system will have total energy E. After we finish, the system may not be in thermal equilibrium.
Spontaneous processes will then occur so as to maximize the system’s entropy, but the internal energy
remains at E.
We have, from the First Law, dE = dQ
¯ − dW
¯ and combining this with the Second Law in the form
dQ
¯ ≤ T dS yields
dE ≤ T dS − dW
¯ . (2.183)
Rearranging terms, we have dW¯ ≤ T dS − dE . Hence, the work done by a thermodynamic system under
conditions of constant entropy is bounded above by −dE, and the maximum dW
¯ is achieved for a reversible
process. It is sometimes useful to define the quantity

dW
¯ free = d̄W − p dV , (2.184)

which is the differential work done by the system other than that required to change its volume. Then
we have
dW
¯ free ≤ T dS − p dV − dE , (2.185)
and we conclude for systems at fixed (S, V ) that dW
¯ free ≤ −dE.
In equilibrium, the equality in Eqn. 2.183 holds, and for single component systems where dW
¯ = p dV −
µ dN we have E = E(S, V, N ) with
     
∂E ∂E ∂E
T = , −p = , µ= . (2.186)
∂S V,N ∂V S,N ∂N S,V

These expressions are easily generalized to multicomponent systems, magnetic systems, etc.
Now consider a single component system at fixed (S, V, N ). We conclude that dE ≤ 0 , which says that
spontaneous processes in a system with dS = dV = dN = 0 always lead to a reduction in the internal
energy E. Therefore, spontaneous processes drive the internal energy E to a minimum in systems at fixed
(S, V, N ).

2.8.2 Helmholtz free energy F

Suppose that when we spontaneously create our system while it is in constant contact with a thermal
reservoir at temperature T . Then as we create our system, it will absorb heat from the reservoir.
Therefore, we don’t have to supply the full internal energy E, but rather only E − Q, since the system
receives heat energy Q from the reservoir. In other words, we must perform work W = E − T S to create
our system, if it is constantly in equilibrium at temperature T . The quantity E − T S is known as the
Helmholtz free energy, F , which is related to the energy E by a Legendre transformation,

F = E − TS . (2.187)
68 CHAPTER 2. THERMODYNAMICS

The general properties of Legendre transformations are discussed in Appendix II, §2.16.
Again invoking the Second Law, we have

dF ≤ −S dT − dW
¯ . (2.188)

Rearranging terms, we have dW ¯ ≤ −S dT − dF , which says that the work done by a thermodynamic
system under conditions of constant temperature is bounded above by −dF , and the maximum dW
¯ is
achieved for a reversible process. We also have the general result

dW
¯ free ≤ −S dT − p dV − dF , (2.189)

and we conclude, for systems at fixed (T, V ), that d̄Wfree ≤ −dF .

Under equilibrium conditions, the equality in Eqn. 2.188 holds, and for single component systems where
dW
¯ = p dV − µ dN we have dF = −S dT − p dV + µ dN . This says that F = F (T, V, N ) with
     
∂F ∂F ∂F
−S = , −p = , µ= . (2.190)
∂T V,N ∂V T,N ∂N T,V

For spontaneous processes, dF ≤ −S dT −p dV +µ dN says that spontaneous processes drive the Helmholtz
free energy F to a minimum in systems at fixed (T, V, N ).

2.8.3 Enthalpy H

Suppose that when we spontaneously create our system while it is thermally insulated, but in constant
mechanical contact with a ‘volume bath’ at pressure p. For example, we could create our system inside
a thermally insulated chamber with one movable wall where the external pressure is fixed at p. Thus,
when creating the system, in addition to the system’s internal energy E, we must also perform work pV
in order to make room for it. In other words, we must perform work W = E + pV . The quantity E + pV
is known as the enthalpy, H. (We use the calligraphic font for H for enthalpy to avoid confusing it with
magnetic field, H.) The enthalpy is obtained from the energy via a different Legendre transformation
than that used to obtain the Helmholtz free energy F , i.e.

H = E + pV . (2.191)

Again invoking the Second Law, we have

dH ≤ T dS − dW
¯ + p dV + V dp , (2.192)

hence with d̄Wfree = d̄W − p dV , we have in general

dW
¯ free ≤ T dS + V dp − dH , (2.193)

and we conclude, for systems at fixed (S, p), that d̄Wfree ≤ −dH.
In equilibrium, for single component systems,

dH = T dS + V dp + µ dN , (2.194)
2.8. THERMODYNAMIC POTENTIALS 69

which says H = H(S, p, N ), with

     
∂H ∂H ∂H
T = , V = , µ= . (2.195)
∂S p,N ∂p S,N ∂N S,p

For spontaneous processes, dH ≤ T dS + V dp + µ dN , which says that spontaneous processes drive the
enthalpy H to a minimum in systems at fixed (S, p, N ).

2.8.4 Gibbs free energy G

If we create a thermodynamic system at conditions of constant temperature T and constant pressure

p, then it absorbs heat energy Q = T S from the reservoir and we must expend work energy pV in
order to make room for it. Thus, the total amount of work we must do in assembling our system is
W = E − T S + pV . This is the Gibbs free energy, G. The Gibbs free energy is obtained from E after
two Legendre transformations, viz.
G = E − T S + pV (2.196)
Note that G = F + pV = H − T S. The Second Law says that

dG ≤ −S dT + V dp + p dV − dW
¯ , (2.197)

which we may rearrange as dW

¯ free ≤ −S dT + V dp − dG . Accordingly, we conclude, for systems at fixed
(T, p), that dW
¯ free ≤ −dG.
For equilibrium one-component systems, the differential of G is

dG = −S dT + V dp + µ dN , (2.198)

therefore G = G(T, p, N ), with

     
∂G ∂G ∂G
−S = , V = , µ= . (2.199)
∂T p,N ∂p T,N ∂N T,p

Recall that Euler’s theorem for single component systems requires E = T S − pV + µN which says
G = µN , Thus, the chemical potential µ is the Gibbs free energy per particle. For spontaneous processes,
dG ≤ −S dT + V dp + µ dN , hence spontaneous processes drive the Gibbs free energy G to a minimum in
systems at fixed (T, p, N ).

2.8.5 Grand potential Ω

The grand potential, sometimes called the Landau free energy, is defined by

Ω = E − T S − µN . (2.200)

Under equilibrium conditions, its differential is

dΩ = −S dT − p dV − N dµ , (2.201)
70 CHAPTER 2. THERMODYNAMICS

hence      
∂Ω ∂Ω ∂Ω
−S = , −p = , −N = . (2.202)
∂T V,µ ∂V T,µ ∂µ T,V

Again invoking eqn. 2.146, we find Ω = −pV , which says that the pressure is the negative of the grand
potential per unit volume.
The Second Law tells us
dΩ ≤ −dW
¯ − S dT − µ dN − N dµ , (2.203)
hence
d¯W
f ≡ dW
free ¯ free + µ dN ≤ −S dT − p dV − N dµ − dΩ . (2.204)
We conclude, for systems at fixed (T, V, µ), that d̄W
ffree ≤ −dΩ.

2.9 Maxwell Relations

Maxwell relations are conditions equating certain derivatives of state variables which follow from the
exactness of the differentials of the various state functions.

2.9.1 Relations deriving from E(S, V, N )

The energy E(S, V, N ) is a state function, with

dE = T dS − p dV + µ dN , (2.205)

and therefore      
∂E ∂E ∂E
T = , −p = , µ= . (2.206)
∂S V,N ∂V S,N ∂N S,V

Taking the mixed second derivatives, we find

∂ 2E
   
∂T ∂p
= =− (2.207)
∂S ∂V ∂V S,N ∂S V,N
∂ 2E
   
∂T ∂µ
= = (2.208)
∂S ∂N ∂N S,V ∂S V,N

∂ 2E
   
∂p ∂µ
=− = . (2.209)
∂V ∂N ∂N S,V ∂V S,N

2.9.2 Relations deriving from F (T, V, N )

The energy F (T, V, N ) is a state function, with

dF = −S dT − p dV + µ dN , (2.210)
2.9. MAXWELL RELATIONS 71

and therefore
     
∂F ∂F ∂F
−S = , −p = , µ= . (2.211)
∂T V,N ∂V T,N ∂N T,V

Taking the mixed second derivatives, we find

∂ 2F
   
∂S ∂p
=− =− (2.212)
∂T ∂V ∂V T,N ∂T V,N
∂ 2F
   
∂S ∂µ
=− = (2.213)
∂T ∂N ∂N T,V ∂T V,N

∂ 2F
   
∂p ∂µ
=− = . (2.214)
∂V ∂N ∂N T,V ∂V T,N

2.9.3 Relations deriving from H(S, p, N )

The enthalpy H(S, p, N ) satisfies

dH = T dS + V dp + µ dN , (2.215)
which says H = H(S, p, N ), with
     
∂H ∂H ∂H
T = , V = , µ= . (2.216)
∂S p,N ∂p S,N ∂N S,p

Taking the mixed second derivatives, we find

∂2H
   
∂T ∂V
= = (2.217)
∂S ∂p ∂p
S,N ∂S p,N
∂2H
   
∂T ∂µ
= = (2.218)
∂S ∂N ∂N S,p ∂S p,N

∂2H
   
∂V ∂µ
= = . (2.219)
∂p ∂N ∂N S,p ∂p S,N

2.9.4 Relations deriving from G(T, p, N )

The Gibbs free energy G(T, p, N ) satisfies

dG = −S dT + V dp + µ dN , (2.220)

therefore G = G(T, p, N ), with

     
∂G ∂G ∂G
−S = , V = , µ= . (2.221)
∂T p,N ∂p T,N ∂N T,p
72 CHAPTER 2. THERMODYNAMICS

Taking the mixed second derivatives, we find

∂2G
   
∂S ∂V
=− = (2.222)
∂T ∂p ∂p T,N ∂T p,N

∂2G
   
∂S ∂µ
=− = (2.223)
∂T ∂N ∂N T,p ∂T p,N

∂2G
   
∂V ∂µ
= = . (2.224)
∂p ∂N ∂N T,p ∂p T,N

2.9.5 Relations deriving from Ω(T, V, µ)

The grand potential Ω(T, V, µ) satisfied

dΩ = −S dT − p dV − N dµ , (2.225)

hence      
∂Ω ∂Ω ∂Ω
−S = , −p = , −N = . (2.226)
∂T V,µ ∂V T,µ ∂µ T,V

Taking the mixed second derivatives, we find

∂ 2Ω
   
∂S ∂p
=− =− (2.227)
∂T ∂V ∂V T,µ ∂T V,µ

∂ 2Ω
   
∂S ∂N
=− =− (2.228)
∂T ∂µ ∂µ T,V ∂T V,µ

∂ 2Ω
   
∂p ∂N
=− =− . (2.229)
∂V ∂µ ∂µ T,V ∂V T,µ

Relations deriving from S(E, V, N )

We can also derive Maxwell relations based on the entropy S(E, V, N ) itself. For example, we have

1 p µ
dS = dE + dV − dN . (2.230)
T T T

Therefore S = S(E, V, N ) and

∂ 2S ∂(T −1 ) ∂(pT −1 )
   
= = , (2.231)
∂E ∂V ∂V E,N ∂E V,N

et cetera.
2.10. EQUILIBRIUM AND STABILITY 73

2.9.6 Generalized thermodynamic potentials

We have up until now assumed a generalized force-displacement pair (y, X) = (−p, V ). But the above
results also generalize to e.g. magnetic systems, where (y, X) = (H, M ). In general, we have

THIS SPACE AVAILABLE dE = T dS + y dX + µ dN (2.232)

F = E − TS dF = −S dT + y dX + µ dN (2.233)

H = E − yX dH = T dS − X dy + µ dN (2.234)

G = E − T S − yX dG = −S dT − X dy + µ dN (2.235)

Ω = E − T S − µN dΩ = −S dT + y dX − N dµ . (2.236)

Generalizing (−p, V ) → (y, X), we also obtain, mutatis mutandis, the following Maxwell relations:
           
∂T ∂y ∂T ∂µ ∂y ∂µ
= = =
∂X S,N ∂S X,N ∂N S,X ∂S X,N ∂N S,X ∂X S,N
           
∂T ∂X ∂T ∂µ ∂X ∂µ
=− = =−
∂y S,N ∂S y,N ∂N S,y ∂S y,N ∂N S,y ∂y S,N
           
∂S ∂y ∂S ∂µ ∂y ∂µ
=− =− =
∂X T,N ∂T X,N ∂N T,X ∂T X,N ∂N T,X ∂X T,N
           
∂S ∂X ∂S ∂µ ∂X ∂µ
= =− =−
∂y T,N ∂T y,N ∂N T,y ∂T y,N ∂N T,y ∂y T,N
           
∂S ∂y ∂S ∂N ∂y ∂N
=− = =− .
∂X T,µ ∂T X,µ ∂µ T,X ∂T X,µ ∂µ T,X ∂X T,µ

2.10 Equilibrium and Stability

2.10.1 Equilibrium

Suppose we have two systems, A and B, which are free to exchange energy, volume, and particle number,
subject to overall conservation rules

EA + EB = E , V A + VB = V , NA + NB = N , (2.237)

where E, V , and N are fixed. Now let us compute the change in the total entropy of the combined
systems when they are allowed to exchange energy, volume, or particle number. We assume that the
74 CHAPTER 2. THERMODYNAMICS

entropy is additive, i.e.

"    # "    #
∂SA ∂SB ∂SA ∂SB
dS = − dEA + − dVA
∂EA VA ,NA ∂EB V ,N ∂VA E ,N ∂VB E ,N
B B A A B B
"    #
∂SA ∂SB
+ − dNA . (2.238)
∂NA E ,V ∂NB E ,V
A A B B

Note that we have used dEB = −dEA , dVB = −dVA , and dNB = −dNA . Now we know from the Second
Law that spontaneous processes result in T dS > 0, which means that S tends to a maximum. If S is a
maximum, it must be that the coefficients of dEA , dVA , and dNA all vanish, else we could increase the total
entropy of the system by a judicious choice of these three differentials. From T dS = dE + p dV − µ, dN ,
we have
     
1 ∂S p ∂S µ ∂S
= , = , =− . (2.239)
T ∂E V,N T ∂V E,N T ∂N E,V

Thus, we conclude that in order for the system to be in equilibrium, so that S is maximized and can
increase no further under spontaneous processes, we must have

TA = TB (thermal equilibrium) (2.240)

pA p
= B (mechanical equilibrium) (2.241)
TA TB
µA µ
= B (chemical equilibrium) (2.242)
TA TB

2.10.2 Stability

Next, consider a uniform system with energy E 0 = 2E, volume V 0 = 2V , and particle number N 0 = 2N .
We wish to check that this system is not unstable with respect to spontaneously becoming inhomogeneous.
To that end, we imagine dividing the system in half. Each half would have energy E, volume V , and
particle number N . But suppose we divided up these quantities differently, so that the left half had
slightly different energy, volume, and particle number than the right, as depicted in Fig. 2.20. Does the
entropy increase or decrease? We have

∆S = S(E + ∆E, V + ∆V, N + ∆N ) + S(E − ∆E, V − ∆V, N − ∆N ) − S(2E, 2V, 2N )

∂ 2S 2 ∂ 2S 2 ∂ 2S
= (∆E) + (∆V ) + (∆N )2 (2.243)
∂E 2 ∂V 2 ∂N 2
∂ 2S ∂ 2S ∂ 2S
+2 ∆E ∆V + 2 ∆E ∆N + 2 ∆V ∆N .
∂E ∂V ∂E ∂N ∂V ∂N

Thus, we can write

X
∆S = Qij Ψi Ψj , (2.244)
i,j
2.10. EQUILIBRIUM AND STABILITY 75

Figure 2.20: To check for an instability, we compare the energy of a system to its total energy when we
reapportion its energy, volume, and particle number slightly unequally.

where
∂ 2S ∂ 2S ∂ 2S
 
∂E 2 ∂E ∂V ∂E ∂N
 
 
 2S ∂ 2S ∂ 2S
Q =  ∂E∂ ∂V (2.245)

 ∂V 2 ∂V ∂N 

 
∂ 2S ∂ 2S ∂ 2S
∂E ∂N ∂V ∂N ∂N 2
is the matrix of second derivatives, known in mathematical parlance as the Hessian, and Ψ = (∆E, ∆V, ∆N ).
Note that Q is a symmetric matrix.
Since S must be a maximum in order for the system to be in equilibrium, we are tempted to conclude
that the homogeneous system is stable if and only if all three eigenvalues of Q are negative. If one or more
of the eigenvalues is positive, then it is possible to choose a set of variations Ψ such that ∆S > 0, which
would contradict the assumption that the homogeneous state is one of maximum entropy. A matrix with
this restriction is said to be negative definite. While it is true that Q can have no positive eigenvalues,
it is clear from homogeneity of S(E, V, N ) that one of the three eigenvalues must be zero, corresponding
to the eigenvector Ψ = (E, V, N ). Homogeneity means S(λE, λV, λN ) = λS(E, V, N ). Now let us take
λ = 1 + η, where η is infinitesimal. Then ∆E = ηE, ∆V = ηV , and ∆N = ηN , and homogeneity says
S(E ± ∆E, V ± ∆V, N ± ∆N ) = (1 ± η) S(E, V, N ) and ∆S = (1 + η)S + (1 − η)S − 2S = 0. We then
have a slightly weaker characterization of Q as negative semidefinite.
However, if we fix one of the components of (∆E, ∆V, ∆N ) to be zero, then Ψ must have some component
orthogonal to the zero eigenvector, in which case ∆S < 0. Suppose we set ∆N = 0 and we just examine
the stability with respect to inhomogeneities in energy and volume. We then restrict our attention to the
upper left 2 × 2 submatrix of Q. A general symmetric 2 × 2 matrix may be written
 
a b
Q= (2.246)
b c
It is easy to solve for the eigenvalues of Q. One finds
 s
a−c 2
 
a+c
λ± = ± + b2 . (2.247)
2 2
In order for Q to be negative definite, we require λ+ < 0 and λ− < 0. Thus, Tr Q = a + c = λ+ + λ− < 0
and det Q = ac − b2 = λ+ λ− > 0. Taken together, these conditions require
a<0 , c<0 , ac > b2 . (2.248)
76 CHAPTER 2. THERMODYNAMICS

Going back to thermodynamic variables, this requires

2
∂ 2S ∂ 2S ∂ 2S ∂ 2S ∂ 2S

<0 , <0 , · > . (2.249)
∂E 2 ∂V 2 ∂E 2 ∂V 2 ∂E ∂V

Thus the entropy is a concave function of E and V at fixed N . Had we set ∆E = 0 and considered
the lower right 2 × 2 submatrix of Q, we’d have concluded that S(V, N ) is concave at fixed E. Since
∂S

−1 , we have ∂ 2S = − 1 ∂T

C
∂E V = T ∂E 2 T 2 ∂E V
= − TV2 < 0 and we conclude CV > 0 for stability.
Many thermodynamic systems are held at fixed (T, p, N ), which suggests we examine the stability criteria
for G(T, p, N ). Suppose our system is in equilibrium with a reservoir at temperature T0 and pressure p0 .
Then, suppressing N (which is assumed constant), we have

G(T0 , p0 ) = E − T0 S + p0 V . (2.250)

Now suppose there is a fluctuation in the entropy and the volume of our system, which is held at fixed
particle number. Going to second order in ∆S and ∆V , we have
"  # "  #
∂E ∂E
∆G = − T0 ∆S + + p0 ∆V
∂S V ∂V S
" # (2.251)
1 ∂ 2E 2 ∂ 2E ∂ 2E 2
+ (∆S) + 2 ∆S ∆V + (∆V ) + . . . .
2 ∂S 2 ∂S ∂V ∂V 2

Equilibrium requires that the coefficients of ∆S and ∆V both vanish, i.e. that T = ∂E


∂S V,N = T0 and
∂E


p = − ∂V S,N
= p0 . The condition for stability is that ∆G > 0 for all (∆S, ∆V ). Stability therefore
requires that the Hessian matrix Q be positive definite, with
 2 2

∂ E ∂ E
∂S 2 ∂S ∂V
Q=  . (2.252)
 
∂ 2E ∂ 2E
∂S ∂V ∂V 2

Thus, we have the following three conditions:

∂ 2E
 
∂T T
2
= = >0 (2.253)
∂S ∂S V CV
∂ 2E
 
∂p 1
2
=− = >0 (2.254)
∂V ∂V S V κS
 2 2
∂ 2E ∂ 2E ∂T 2
 
∂E T
· − = − >0. (2.255)
∂S 2 ∂V 2 ∂S ∂V V κ S CV ∂V S

As we shall discuss below, the quantity αS ≡ V1 ∂V

∂T S,N is the adiabatic thermal expansivity coefficient.
We therefore conclude that stability of any thermodynamic system requires
r
CV κS CV
>0 , κS > 0 , αS > . (2.256)
T VT
2.11. APPLICATIONS OF THERMODYNAMICS 77

Figure 2.21: Adiabatic free expansion via a thermal path. The initial and final states do not lie along
an adabat! Rather, for an ideal gas, the initial and final states lie along an isotherm.

2.11 Applications of Thermodynamics

A discussion of various useful mathematical relations among partial derivatives may be found in the
appendix in §2.17. Some facility with the differential multivariable calculus is extremely useful in the
analysis of thermodynamics problems.

2.11.1 Adiabatic free expansion revisited

Consider once again the adiabatic free expansion of a gas from initial volume Vi to final volume Vf = rVi .
Since the system is not in equilibrium during the free expansion process, the initial and final states do
not lie along an adiabat, i.e. they do not have the same entropy. Rather, as we found, from Q = W = 0,
we have that Ei = Ef , which means they have the same energy, and, in the case of an ideal gas, the
same temperature (assuming N is constant). Thus, the initial and final states lie along an isotherm. The
situation is depicted in Fig. 2.21. Now let us compute the change in entropy ∆S = Sf − Si by integrating
along this isotherm. Note that the actual dynamics are irreversible and do not quasistatically follow
any continuous thermodynamic path. However, we can use what is a fictitious thermodynamic path as a
means of comparing S in the initial and final states.
We have
ZVf  
∂S
∆S = Sf − Si = dV . (2.257)
∂V T,N
Vi

But from a Maxwell equation deriving from F , we have

   
∂S ∂p
= , (2.258)
∂V T,N ∂T V,N
78 CHAPTER 2. THERMODYNAMICS

hence
ZVf  
∂p
∆S = dV . (2.259)
∂T V,N
Vi

For an ideal gas, we can use the equation of state pV = N kB T to obtain

 
∂p N kB
= . (2.260)
∂T V,N V

The integral can now be computed:

ZrVi
N kB
∆S = dV = N kB ln r , (2.261)
V
Vi

as we found before, in eqn. 2.156 What is different about this derivation? Previously, we derived the
entropy change from the explicit formula for S(E, V, N ). Here, we did not need to know this function.
The Maxwell relation allowed us to compute the entropy change using only the equation of state.

2.11.2 Energy and volume

We saw how E(T, V, N ) = 12 f N kB T for an ideal gas, independent of the volume. In general we should
have
V


E(T, V, N ) = N φ T, N . (2.262)
V
= 21 f kB T is a function of T alone and is independent on the other intensive


For the ideal gas, φ T, N
quantity V /N . How does energy vary with volume? At fixed temperature and particle number, we have,
from E = F + T S,
       
∂E ∂F ∂S ∂p
= +T = −p + T , (2.263)
∂V T,N ∂V T,N ∂V T,N ∂T V,N

∂S

∂p

where we have used the Maxwell relation ∂V T.N
= ∂T V,N , derived from the mixed second derivative
∂ 2F
∂T ∂V. Another way to derive this result is as follows. Write dE = T dS − p dV + µ dN and then express
dS in terms of dT , dV , and dN , resulting in
  "   # "   #
∂S ∂S ∂µ
dE = T dT + T − p dV − T + µ dN . (2.264)
∂T V,N ∂V T,N ∂T V,N

∂E


Now read off ∂V V,N
and use the same Maxwell relation as before to recover eqn. 2.263. Applying this
result to the ideal gas law pV = N kB T results in the vanishing of the RHS, hence for any substance
obeying the ideal gas law we must have

E(T, V, N ) = ν ε(T ) = N ε(T )/NA . (2.265)

2.11. APPLICATIONS OF THERMODYNAMICS 79

 
L2 ·bar L


gas a mol2
b mol pc (bar) Tc (K) vc (L/mol)
Acetone 14.09 0.0994 52.82 505.1 0.2982
Argon 1.363 0.03219 48.72 150.9 0.0966
Carbon dioxide 3.640 0.04267 7404 304.0 0.1280
Ethanol 12.18 0.08407 63.83 516.3 0.2522
Freon 10.78 0.0998 40.09 384.9 0.2994
Helium 0.03457 0.0237 2.279 5.198 0.0711
Hydrogen 0.2476 0.02661 12.95 33.16 0.0798
Mercury 8.200 0.01696 1055 1723 0.0509
Methane 2.283 0.04278 46.20 190.2 0.1283
Nitrogen 1.408 0.03913 34.06 128.2 0.1174
Oxygen 1.378 0.03183 50.37 154.3 0.0955
Water 5.536 0.03049 220.6 647.0 0.0915

Table 2.3: Van der Waals parameters for some common gases. (Source: Wikipedia.)

2.11.3 van der Waals equation of state

It is clear that the same conclusion follows for any equation of state of the form p(T, V, N ) = T · f (V /N ),
where f (V /N ) is an arbitrary function of its argument: the ideal gas law remains valid11 . This is not
true, however, for the van der Waals equation of state,
 
a
p + 2 v − b) = RT , (2.266)
v
where v = NA V /N is the molar volume. We then find (always assuming constant N ),
     
∂E ∂ε ∂p a
= =T −p= 2 , (2.267)
∂V T ∂v T ∂T V v
where E(T, V, N ) ≡ ν ε(T, v). We can integrate this to obtain
a
ε(T, v) = ω(T ) − , (2.268)
v
where ω(T ) is arbitrary. From eqn. 2.33, we immediately have
 
∂ε
cV = = ω 0 (T ) . (2.269)
∂T v

What about cp ? This requires a bit of work. We start with eqn. 2.34,
   
∂ε ∂v
cp = +p
∂T p ∂T p
   (2.270)
a ∂v
= ω 0 (T ) + p + 2
v ∂T p
11
Note V /N = v/NA .
80 CHAPTER 2. THERMODYNAMICS

We next take the differential of the equation of state (at constant N ):

   
a
2a
R dT = p + 2 dv + v − b dp − 3 dv
v v
  (2.271)
a 2ab

= p − 2 + 3 dv + v − b dp .
v v
We can now read off the result for the volume expansion coefficient,
 
1 ∂v 1 R
αp = = · . (2.272)
v ∂T p v p − va2 + 2ab v3

We now have for cp ,

a


0 p+ v2
R
cp = ω (T ) + a 2ab
p− v2
+ v3
(2.273)
R2 T v 3
= ω 0 (T ) + .
RT v 3 − 2a(v − b)2

where v = V NA /N is the molar volume.

To fix ω(T ), we consider the v → ∞ limit, where the density of the gas vanishes. In this limit, the gas
must be ideal, hence eqn. 2.268 says that ω(T ) = 21 f RT . Therefore cV (T, v) = 21 f R, just as in the case
of an ideal gas. However, rather than cp = cV + R, which holds for ideal gases, cp (T, v) is given by eqn.
2.273. Thus,

cVDW
V = 21 f R (2.274)
R2 T v 3
cVDW
p = 12 f R + . (2.275)
RT v 3 − 2a(v − b)2
Note that cp (a → 0) = cV + R, which is the ideal gas result.
As we shall see in chapter 7, the van der Waals system in unstable throughout a region of parameters,
where it undergoes phase separation between high density (liquid) and low density (gas) phases. The
above results are valid only in the stable regions of the phase diagram.

2.11.4 Thermodynamic response functions

Consider the entropy S expressed as a function of T , V , and N :

     
∂S ∂S ∂S
dS = dT + dV + dN . (2.276)
∂T V,N ∂V T,N ∂N T,V

Dividing by dT , multiplying by T , and assuming dN = 0 throughout, we have

   
∂S ∂V
Cp − CV = T . (2.277)
∂V T ∂T p
2.11. APPLICATIONS OF THERMODYNAMICS 81

Appealing to a Maxwell relation derived from F (T, V, N ), and then appealing to eqn. 2.479, we have
       
∂S ∂p ∂p ∂V
= =− . (2.278)
∂V T ∂T V ∂V T ∂T p

This allows us to write    2

∂p ∂V
Cp − CV = −T . (2.279)
∂V T ∂T p

We define the response functions,

1 ∂2G
 
1 ∂V
isothermal compressibility: κT = − =− (2.280)
V ∂p T V ∂p2
1 ∂2H
 
1 ∂V
adiabatic compressibility: κS = − =− (2.281)
V ∂p S V ∂p2
 
1 ∂V
thermal expansivity: αp = . (2.282)
V ∂T p

Thus,
T αp2
Cp − CV = V , (2.283)
κT
or, in terms of intensive quantities,
v T αp2
cp − cV = , (2.284)
κT
where, as always, v = V NA /N is the molar volume.
This above relation generalizes to any conjugate force-displacement pair (−p, V ) → (y, X):
   
∂y ∂X
Cy − CX = −T
∂T X ∂T y
(2.285)
∂X 2
   
∂y
=T .
∂X T ∂T y

For example, we could have (y, X) = (H α , M α ).

A similar relationship can be derived between the compressibilities κT and κS . We then clearly must
start with the volume, writing
     
∂V ∂V ∂V
dV = dp + dS + dN . (2.286)
∂p S,N ∂S p,N ∂p S,p

Dividing by dp, multiplying by −V −1 , and keeping N constant, we have

  
1 ∂V ∂S
κT − κS = − . (2.287)
V ∂S p ∂p T
82 CHAPTER 2. THERMODYNAMICS

Again we appeal to a Maxwell relation, writing

   
∂S ∂V
=− , (2.288)
∂p T ∂T p
and after invoking the chain rule,
      
∂V ∂V ∂T T ∂V
= = , (2.289)
∂S p ∂T p ∂S p Cp ∂T p

we obtain
v T αp2
κT − κS = . (2.290)
cp
Comparing eqns. 2.284 and 2.290, we find
(cp − cV ) κT = (κT − κS ) cp = v T αp2 . (2.291)
This result entails
cp κ
= T . (2.292)
cV κS
The corresponding result for magnetic systems is
 2
∂m
(cH − cM ) χT = (χT − χS ) cH =T , (2.293)
∂T H

where m = M/ν is the magnetization per mole of substance, and

∂ 2G
 
∂M 1
isothermal susceptibility: χT = =− (2.294)
∂H T ν ∂H 2
∂2H
 
∂M 1
adiabatic susceptibility: χS = =− . (2.295)
∂H S ν ∂H 2

Here the enthalpy and Gibbs free energy are

H = E − HM dH = T dS − M dH (2.296)
G = E − T S − HM dG = −S dT − M dH . (2.297)

Remark: The previous discussion has assumed an isotropic magnetic system where M and H are
collinear, hence H ·M = HM .
∂M α ∂ 2G
 
1
χαβ
T = = − (2.298)
∂H β T ν ∂H α ∂H β
∂M α ∂2H
 
1
χαβ
S = =− . (2.299)
β
∂H S ν ∂H α ∂H β

In this case, the enthalpy and Gibbs free energy are

H = E − H ·M dH = T dS − M ·dH (2.300)
G = E − T S − H ·M dG = −S dT − M ·dH . (2.301)
2.11. APPLICATIONS OF THERMODYNAMICS 83

2.11.5 Joule effect: free expansion of a gas

Previously we considered the adiabatic free expansion of an ideal gas. We found that Q = W = 0 hence
∆E = 0, which means the process is isothermal, since E = νε(T ) is volume-independent. The entropy
changes, however, since S(E, V, N ) = N kB ln(V /N ) + 21 f N kB ln(E/N ) + N s0 . Thus,
 
V
Sf = Si + N kB ln f . (2.302)
Vi

What happens if the gas is nonideal?

We integrate along a fictitious thermodynamic path connecting initial and final states, where dE = 0
along the path. We have    
∂E ∂E
0 = dE = dV + dT (2.303)
∂V T ∂T V
hence    
∂T (∂E/∂V )T 1 ∂E
=− =− . (2.304)
∂V E (∂E/∂T )V CV ∂V T
We also have      
∂E ∂S ∂p
=T −p=T −p . (2.305)
∂V T ∂V T ∂T V
Thus, "
    #
∂T 1 ∂p
= p−T . (2.306)
∂V E CV ∂T V
Note that the term in square brackets vanishes for any system obeying the ideal gas law. For a nonideal
gas,
ZVf  
∂T
∆T = dV , (2.307)
∂V E
Vi

which is in general nonzero.

Now consider a van der Waals gas, for which
 
a
p + 2 (v − b) = RT .
v

We then have
aν 2
 
∂p a
p−T =− 2
=− 2 . (2.308)
∂T V v V
In §2.11.3 we concluded that CV = 21 f νR for the van der Waals gas, hence

ZVf  
2aν dV 2a 1 1
∆T = − = − . (2.309)
fR V2 f R vf vi
Vi
84 CHAPTER 2. THERMODYNAMICS

Figure 2.22: In a throttle, a gas is pushed through a porous plug separating regions of different pressure.
The change in energy is the work done, hence enthalpy is conserved during the throttling process.

Thus, if Vf > Vi , we have Tf < Ti and the gas cools upon expansion.
Consider O2 gas with an initial specific volume of vi = 22.4 L/mol, which is the STP value for an ideal gas,
freely expanding to a volume vf = ∞ for maximum cooling. According to table 2.3, a = 1.378 L2 ·bar/mol2 ,
and we have ∆T = −2a/f Rvi = −0.296 K, which is a pitifully small amount of cooling. Adiabatic free
expansion is a very inefficient way to cool a gas.

2.11.6 Throttling: the Joule-Thompson effect

In a throttle, depicted in Fig. 2.22, a gas is forced through a porous plug which separates regions of
different pressures. According to the figure, the work done on a given element of gas is

ZVf ZVi
W = dV pf − dV pi = pf Vf − pi Vi . (2.310)
0 0

Now we assume that the system is thermally isolated so that the gas exchanges no heat with its environ-
ment, nor with the plug. Then Q = 0 so ∆E = −W , and

Ei + pi Vi = Ef + pf Vf (2.311)
Hi = Hf , (2.312)

where H is enthalpy. Thus, the throttling process is isenthalpic. We can therefore study it by defining a
fictitious thermodynamic path along which dH = 0. The, choosing T and p as state variables,
   
∂H ∂H
0 = dH = dT + dp (2.313)
∂T p ∂p T

hence  
∂T (∂H/∂p)T
=− . (2.314)
∂p H (∂H/∂T )p
The numerator on the RHS is computed by writing dH = T dS + V dp and then dividing by dp, to obtain
     
∂H ∂S ∂V
=V +T =V −T . (2.315)
∂p T ∂p T ∂T p
2.11. APPLICATIONS OF THERMODYNAMICS 85

The denominator is
    
∂H ∂H ∂S
=
∂T p ∂S p ∂T p
  (2.316)
∂S
=T = Cp .
∂T p
Thus,
  "   #
∂T 1 ∂v
= T −v
∂p H cp ∂T p
(2.317)
v

= T αp − 1 ,
cp
1 ∂V


where αp = V ∂T p is the volume expansion coefficient.
From the van der Waals equation of state, we obtain, from eqn. 2.272,
 
T ∂v RT /v v−b
T αp = = a 2ab
=
2 . (2.318)
v ∂T p p − v2 + v3 v − 2a v−b RT v
a
Assuming v  RT , b, we have
   
∂T 1 2a
= −b . (2.319)
∂p H cp RT
 
Thus, for T > T ∗ = Rb
2a
, we have ∂T
∂p H < 0 and the gas heats up upon an isenthalpic pressure decrease.
∗
For T < T , the gas cools under such conditions.
∗ for the van der Waals gas. To see this, we set T α = 1,
In fact, there are two inversion temperatures T1,2 p
which is the criterion for inversion. From eqn. 2.318 it is easy to derive
r
b bRT
=1− . (2.320)
v 2a
We insert this into the van der Waals equation of state to derive a relationship T = T ∗ (p) at which
T αp = 1 holds. After a little work, we find
r
3RT 8aRT a
p=− + 3
− 2 . (2.321)
2b b b
This is a quadratic equation for T , the solution of which is
r !2
∗ 2a 3b2 p
T (p) = 2± 1− . (2.322)
9 bR a
 
∂T
In Fig. 2.23 we plot pressure versus temperature in scaled units, showing the curve along which ∂p H =
0. The volume, pressure, and temperature scales defined are
a 8a
vc = 3b , pc = , Tc = . (2.323)
27 b2 27 bR
86 CHAPTER 2. THERMODYNAMICS

Figure 2.23: Inversion temperature T ∗ (p) for the van der Waals gas. Pressure and temperature are given
in terms of pc = a/27b2 and Tc = 8a/27bR, respectively.

Values for pc , Tc , and vc are provided in table 2.3. If we define v = v/vc , p = p/pc , and T = T /Tc , then
the van der Waals equation of state may be written in dimensionless form:
 
3
p + 2 3v − 1) = 8T . (2.324)
v
 
∂T
In terms of the scaled parameters, the equation for the inversion curve ∂p H = 0 becomes
 q 2  q 2
p = 9 − 36 1 − 13 T ⇐⇒ T = 3 1 ± 1 − 91 p . (2.325)

Thus, there is no inversion for p > 9 pc . We are usually interested in the upper inversion temperature, T2∗ ,
corresponding to the upper sign in eqn. 2.322. The maximum inversion temperature occurs for p = 0,
∗
where Tmax 2a
= 27 ∗
= bR 4 Tc . For H2 , from the data in table 2.3, we find Tmax (H2 ) = 224 K, which is within
10% of the experimentally measured value of 205 K.
 
What happens when H2 gas leaks from a container with T > T2∗ ? Since ∂T ∂p H < 0 and ∆p < 0, we have
∆T > 0. The gas warms up, and the heat facilitates the reaction 2 H2 + O2 −→ 2 H2 O, which releases
energy, and we have a nice explosion.

2.12 Phase Transitions and Phase Equilibria

A typical phase diagram of a p-V -T system is shown in the Fig. 2.24(a). The solid lines delineate
boundaries between distinct thermodynamic phases. These lines are called coexistence curves. Along
2.12. PHASE TRANSITIONS AND PHASE EQUILIBRIA 87

(a) (b) (c)

generic
substance
ppressure

temperature T 3He 4He

Figure 2.24: (a) Typical thermodynamic phase diagram of a single component p-V -T system, showing
triple point (three phase coexistence) and critical point. (Source: Univ. of Helsinki.) Also shown: phase
diagrams for 3 He (b) and 4 He (c). What a difference a neutron makes! (Source: Brittanica.)

these curves, we can have coexistence of two phases, and the thermodynamic potentials are singular. The
order of the singularity is often taken as a classification of the phase transition. I.e. if the thermodynamic
potentials E, F , G, and H have discontinuous or divergent mth derivatives, the transition between the
respective phases is said to be mth order . Modern theories of phase transitions generally only recognize
two possibilities: first order transitions, where the order parameter changes discontinuously through
the transition, and second order transitions, where the order parameter vanishes continuously at the
boundary from ordered to disordered phases12 . We’ll discuss order parameters during Physics 140B.
For a more interesting phase diagram, see Fig. 2.24(b,c), which displays the phase diagrams for 3 He and
4 He. The only difference between these two atoms is that the former has one fewer neutron: (2p + 1n +
2e) in 3 He versus (2p + 2n + 2e) in 4 He. As we shall learn when we study quantum statistics, this extra
neutron makes all the difference, because 3 He is a fermion while 4 He is a boson.

2.12.1 p-v-T surfaces

The equation of state for a single component system may be written as

f (p, v, T ) = 0 . (2.326)

This may in principle be inverted to yield p = p(v, T ) or v = v(T, p) or T = T (p, v). The single constraint
f (p, v, T ) on the three state variables defines a surface in {p, v, T } space. An example of such a surface
is shown in Fig. 2.25, for the ideal gas.
Real p-v-T surfaces are much richer than that for the ideal gas, because real systems undergo phase
transitions in which thermodynamic properties are singular or discontinuous along certain curves on the
p-v-T surface. An example is shown in Fig. 2.26. The high temperature isotherms resemble those of the
ideal gas, but as one cools below the critical temperature Tc , the isotherms become singular. Precisely
12
Some exotic phase transitions in quantum matter, which do not quite fit the usual classification schemes, have recently
been proposed.
88 CHAPTER 2. THERMODYNAMICS

Figure 2.25: The surface p(v, T ) = RT /v corresponding to the ideal gas equation of state, and its
projections onto the (p, T ), (p, v), and (T, v) planes.

at T = Tc , the isotherm p = p(v, Tc ) becomes perfectly horizontal at v = vc , which is the critical molar
volume. This means that the isothermal compressibility, κT = − v1 ∂v


∂p T diverges at T = Tc . Below Tc ,
the isotherms have a flat portion, as shown in Fig. 2.28, corresponding to a two-phase region where liquid
and vapor coexist. In the (p, T ) plane, sketched for H2 O in Fig. 2.4 and shown for CO2 in Fig. 2.29,
this liquid-vapor phase coexistence occurs along a curve, called the vaporization (or boiling) curve. The
density changes discontinuously across this curve; for H2 O, the liquid is approximately 1000 times denser
than the vapor at atmospheric pressure. The density discontinuity vanishes at the critical point. Note
that one can continuously transform between liquid and vapor phases, without encountering any phase
transitions, by going around the critical point and avoiding the two-phase region.
In addition to liquid-vapor coexistence, solid-liquid and solid-vapor coexistence also occur, as shown in
Fig. 2.26. The triple point (Tt , pt ) lies at the confluence of these three coexistence regions. For H2 O, the
location of the triple point and critical point are given by

Tt = 273.16 K Tc = 647 K
−3
pt = 611.7 Pa = 6.037 × 10 atm pc = 22.06 MPa = 217.7 atm
2.12. PHASE TRANSITIONS AND PHASE EQUILIBRIA 89

Figure 2.26: A p-v-T surface for a substance which contracts upon freezing. The red dot is the critical
point and the red dashed line is the critical isotherm. The yellow dot is the triple point at which there
is three phase coexistence of solid, liquid, and vapor.

2.12.2 The Clausius-Clapeyron relation

Recall that the homogeneity of E(S, V, N ) guaranteed E = T S − pV + µN , from Euler’s theorem. It also
guarantees a relation between the intensive variables T , p, and µ, according to eqn. 2.148. Let us define
g ≡ G/ν = NA µ, the Gibbs free energy per mole. Then

dg = −s dT + v dp , (2.327)

where s = S/ν and v = V /ν are the molar entropy and molar volume, respectively. Along a coexistence
curve between phase #1 and phase #2, we must have g1 = g2 , since the phases are free to exchange
energy and particle number, i.e. they are in thermal and chemical equilibrium. This means

dg1 = −s1 dT + v1 dp = −s2 dT + v2 dp = dg2 . (2.328)

Therefore, along the coexistence curve we must have

 
dp s − s1 `
= 2 = , (2.329)
dT coex v2 − v1 T ∆v
where
` ≡ T ∆s = T (s2 − s1 ) (2.330)
is the molar latent heat of transition. A heat ` must be supplied in order to change from phase #1 to
phase #2, even without changing p or T . If ` is the latent heat per mole, then we write `˜ as the latent
heat per gram: `˜ = `/M , where M is the molar mass.
90 CHAPTER 2. THERMODYNAMICS

Figure 2.27: Equation of state for a substance which expands upon freezing, projected to the (v, T ) and
(v, p) and (T, p) planes.

Along the liquid-gas coexistence curve, we typically have vgas  vliquid , and assuming the vapor is ideal,
we may write ∆v ≈ vgas ≈ RT /p. Thus,
 
dp ` p`
= ≈ . (2.331)
dT liq−gas T ∆v RT 2

If ` remains constant throughout a section of the liquid-gas coexistence curve, we may integrate the above
equation to get
dp ` dT
= =⇒ p(T ) = p(T0 ) e`/RT0 e−`/RT . (2.332)
p R T2

2.12.3 Liquid-solid line in H2 O

Life on planet earth owes much of its existence to a peculiar property of water: the solid is less dense
than the liquid along the coexistence curve. For example at T = 273.1 K and p = 1 atm,

ṽwater = 1.00013 cm3 /g , ṽice = 1.0907 cm3 /g . (2.333)

The latent heat of the transition is `˜ = 333 J/g = 79.5 cal/g. Thus,

`˜
 
dp 333 J/g
= =
dT liq−sol T ∆ṽ (273.1 K) (−9.05 × 10−2 cm3 /g)
(2.334)
dyn
8 atm
= −1.35 × 10 2
= −134 ◦ .
cm K C
2.12. PHASE TRANSITIONS AND PHASE EQUILIBRIA 91

Figure 2.28: Projection of the p-v-T surface of Fig. 2.26 onto the (v, p) plane.

The negative slope of the melting curve is invoked to explain the movement of glaciers: as glaciers slide
down a rocky slope, they generate enormous pressure at obstacles13 Due to this pressure, the story goes,
the melting temperature decreases, and the glacier melts around the obstacle, so it can flow past it, after
which it refreezes. But it is not the case that the bottom of the glacier melts under the pressure, for
consider a glacier of height h = 1 km. The pressure at the bottom is p ∼ gh/ṽ ∼ 107 Pa, which is only
about 100 atmospheres. Such a pressure can produce only a small shift in the melting temperature of
about ∆Tmelt = −0.75◦ C.
Does the Clausius-Clapeyron relation explain how we can skate on ice? When my daughter was seven
years old, she had a mass of about M = 20 kg. Her ice skates had blades of width about 5 mm and length
about 10 cm. Thus, even on one foot, she imparted an additional pressure of only

Mg 20 kg × 9.8 m/s2
∆p = ≈ = 3.9 × 105 Pa = 3.9 atm . (2.335)
A (5 × 10−3 m) × (10−1 m)

The corresponding change in the melting temperature is thus minuscule: ∆Tmelt ≈ −0.03◦ C.
So why could my daughter skate so nicely? The answer isn’t so clear!14 There seem to be two relevant
issues in play. First, friction generates heat which can locally melt the surface of the ice. Second, the
surface of ice, and of many solids, is naturally slippery. Indeed, this is the case for ice even if one is
standing still, generating no frictional forces. Why is this so? It turns out that the Gibbs free energy
of the ice-air interface is larger than the sum of free energies of ice-water and water-air interfaces. That
is to say, ice, as well as many simple solids, prefers to have a thin layer of liquid on its surface, even at
13
The melting curve has a negative slope at relatively low pressures, where the solid has the so-called Ih hexagonal crystal
structure. At pressures above about 2500 atmospheres, the crystal structure changes, and the slope of the melting curve
becomes positive.
14
For a recent discussion, see R. Rosenberg, Physics Today 58, 50 (2005).
92 CHAPTER 2. THERMODYNAMICS

temperatures well below its bulk melting point. If the intermolecular interactions are not short-ranged15 ,
theory predicts a surface melt thickness d ∝ (Tm − T )−1/3 . In Fig. 2.30 we show measurements by Gilpin
(1980) of the surface melt on ice, down to about −50◦ C. Near 0◦ C the melt layer thickness is about
40 nm, but this decreases to ∼ 1 nm at T = −35◦ C. At very low temperatures, skates stick rather than
glide. Of course, the skate material is also important, since that will affect the energetics of the second
interface. The 19th century novel, Hans Brinker, or The Silver Skates by Mary Mapes Dodge tells the
story of the poor but stereotypically decent and hardworking Dutch boy Hans Brinker, who dreams of
winning an upcoming ice skating race, along with the top prize: a pair of silver skates. All he has are some
lousy wooden skates, which won’t do him any good in the race. He has money saved to buy steel skates,
but of course his father desperately needs an operation because – I am not making this up – he fell off a
dike and lost his mind. The family has no other way to pay for the doctor. What a story! At this point,
I imagine the suspense must be too much for you to bear, but this isn’t an American Literature class, so
you can use Google to find out what happens (or rent the 1958 movie, directed by Sidney Lumet). My
point here is that Hans’ crappy wooden skates can’t compare to the metal ones, even though the surface
melt between the ice and the air is the same. The skate blade material also makes a difference, both for
the interface energy and, perhaps more importantly, for the generation of friction as well.

2.12.4 Slow melting of ice : a quasistatic but irreversible process

Suppose we have an ice cube initially at temperature T0 < Θ ≡ 273.15 K (i.e. Θ = 0◦ C) and we toss it
into a pond of water. We regard the pond as a heat bath at some temperature T1 > Θ. Let the mass of
the ice be M . How much heat Q is absorbed by the ice in order to raise its temperature to T1 ? Clearly
Q = M c̃S (Θ − T0 ) + M `˜ + M c̃L (T1 − Θ) , (2.336)
where c̃S and c̃L are the specific heats of ice (solid) and water (liquid), respectively16 , and `˜ is the latent
heat of melting per unit mass. The pond must give up this much heat to the ice, hence the entropy of
the pond, discounting the new water which will come from the melted ice, must decrease:
Q
∆Spond = − . (2.337)
T1

Now we ask what is the entropy change of the H2 O in the ice. We have
ZΘ ZT1
dQ M c̃S M `˜ M c̃L
Z
¯
∆Sice = = dT + + dT
T T Θ T
T0 Θ (2.338)
M `˜
   
Θ T
= M c̃S ln + + M c̃L ln 1 .
T0 Θ Θ
The total entropy change of the system is then
∆Stotal = ∆Spond + ∆Sice
(2.339)
         
Θ Θ − T0 ˜ 1 1 T1 T1 − Θ
= M c̃S ln − M c̃S + M` − + M c̃L ln − M c̃L
T0 T1 Θ T1 Θ T1
15
For example, they could be of the van der Waals form, due to virtual dipole fluctuations, with an attractive 1/r6 tail.
16
We assume c̃S (T ) and c̃L (T ) have no appreciable temperature dependence, and we regard them both as constants.
2.12. PHASE TRANSITIONS AND PHASE EQUILIBRIA 93

Figure 2.29: Phase diagram for CO2 in the (p, T ) plane. (Source: www.scifun.org.)

Now since T0 < Θ < T1 , we have

   
Θ − T0 Θ − T0
M c̃S < M c̃S . (2.340)
T1 Θ

Therefore,  
1 1
∆S > M `˜



− + M c̃S f T0 /Θ + M c̃L f Θ/T1 , (2.341)
Θ T1
where f (x) = x − 1 − ln x. Clearly f 0 (x) = 1 − x−1 is negative on the interval (0, 1), which means that
the maximum of f (x) occurs at x = 0 and the minimum at x = 1. But f (0) = ∞ and f (1) = 0, which
means that f (x) ≥ 0 for x ∈ [0, 1]. Since T0 < Θ < T1 , we conclude ∆Stotal > 0.

2.12.5 Gibbs phase rule

Equilibrium between two phases means that p, T , and µ(p, T ) are identical. From

µ1 (p, T ) = µ2 (p, T ) , (2.342)

we derive an equation for the slope of the coexistence curve, the Clausius-Clapeyron relation. Note that
we have one equation in two unknowns (T, p), so the solution set is a curve. For three phase coexistence,
we have
µ1 (p, T ) = µ2 (p, T ) = µ3 (p, T ) , (2.343)
which gives us two equations in two unknowns. The solution is then a point (or a set of points). A critical
point also is a solution of two simultaneous equations:

critical point =⇒ v1 (p, T ) = v2 (p, T ) , µ1 (p, T ) = µ2 (p, T ) . (2.344)

94 CHAPTER 2. THERMODYNAMICS

Figure 2.30: Left panel: data from R. R. Gilpin, J. Colloid Interface Sci. 77, 435 (1980) showing
measured thickness of the surface melt on ice at temperatures below 0◦ C. The straight line has slope − 31 ,
as predicted by theory. Right panel: phase diagram of H2 O, showing various high pressure solid phases.
(Source : Physics Today, December 2005).

∂µ

Recall v = NA ∂p T . Note that there can be no four phase coexistence for a simple p-V -T system.
Now for the general result. Suppose we have σ species, with particle numbers Na , where a = 1, . . . , σ. It is
useful to briefly recapitulate the derivation of the Gibbs-Duhem relation. The energy E(S, V, N1 , . . . , Nσ )
is a homogeneous function of degree one:

E(λS, λV, λN1 , . . . , λNσ ) = λE(S, V, N1 , . . . , Nσ ) . (2.345)

From Euler’s theorem for homogeneous functions (just differentiate with respect to λ and then set λ = 1),
we have
Xσ
E = TS − pV + µa Na . (2.346)
a=1

Taking the differential, and invoking the First Law,

σ
X
dE = T dS − p dV + µa dNa , (2.347)
a=1

we arrive at the relation

σ
X
S dT − V dp + Na dµa = 0 , (2.348)
a=1

of which eqn. 2.147 is a generalization to additional internal ‘work’ variables. This says that the σ + 2
quantities (T, p, µ1 , . . . , µσ ) are not all independent. We can therefore write


µσ = µσ T, p, µ1 , . . . , µσ−1 . (2.349)
2.12. PHASE TRANSITIONS AND PHASE EQUILIBRIA 95

(j)
If there are ϕ different phases, then in each phase j, with j = 1, . . . , ϕ, there is a chemical potential µa
for each species a. We then have
 
(j) (j)
µ(j)
σ = µσ
(j)
T, p, µ1 , . . . , µσ−1 . (2.350)

(j)
Here µa is the chemical potential of the ath species in the j th phase. Thus, there are ϕ such equations
 (j)

relating the 2 + ϕ σ variables T, p, µa , meaning that only 2 + ϕ (σ − 1) of them may be chosen as
independent. This, then, is the dimension of ‘thermodynamic space’ containing a maximal number of
intensive variables:
dTD (σ, ϕ) = 2 + ϕ (σ − 1) . (2.351)
To completely specify the state of our system, we of course introduce
Pσ a single extensive variable, such
as the total volume V . Note that the total particle number N = a=1 Na may not be conserved in the
presence of chemical reactions!
Now suppose we have equilibrium among ϕ phases. We have implicitly assumed thermal and mechanical
equilibrium among all the phases, meaning that p and T are constant. Chemical equilibrium applies on
a species-by-species basis. This means
(j 0 )
µ(j)
a = µa (2.352)
where j, j 0 ∈ {1, . . . , ϕ}. This gives σ(ϕ − 1) independent equations equations17 . Thus, we can have phase
equilibrium among the ϕ phases of σ species over a region of dimension

dPE (σ, ϕ) = 2 + ϕ (σ − 1) − σ (ϕ − 1)
(2.353)
=2+σ−ϕ .

Since dPE ≥ 0, we must have ϕ ≤ σ + 2. Thus, with two species (σ = 2), we could have at most four
phase coexistence.
If the various species can undergo ρ distinct chemical reactions of the form
(r) (r)
ζ1 A1 + ζ2 A2 + · · · + ζσ(r) Aσ = 0 , (2.354)

(r)
where Aa is the chemical formula for species a, and ζa is the stoichiometric coefficient for the ath species
in the rth reaction, with r = 1, . . . , ρ, then we have an additional ρ constraints of the form
σ
X
ζa(r) µ(j)
a =0 . (2.355)
a=1

Therefore,
dPE (σ, ϕ, ρ) = 2 + σ − ϕ − ρ . (2.356)
One might ask what value of j are we to use in eqn. 2.355, or do we in fact have ϕ such equations for
each r? The answer is that eqn. 2.352 guarantees that the chemical potential of species a is the same in
all the phases, hence it doesn’t matter what value one chooses for j in eqn. 2.355.
17
Set j = 1 and let j 0 range over the ϕ − 1 values 2, . . . , ϕ.
96 CHAPTER 2. THERMODYNAMICS

Pσ
Let us assume that no reactions take place, i.e. ρ = 0, so the total number of particles b=1 Nb is
(j)
conserved. Instead of choosing (T, p, µ1 , . . . , µσ−1 ) as dTD intensive variables, we could have chosen
(j)
(T, p, µ1 , . . . , xσ−1 ), where xa = Na /N is the concentration of species a.
Why do phase diagrams in the (p, v) and (T, v) plane look different than those in the (p, T ) plane?18
For example, Fig. 2.27 shows projections of the p-v-T surface of a typical single component substance
onto the (T, v), (p, v), and (p, T ) planes. Coexistence takes place along curves in the (p, T ) plane, but in
extended two-dimensional regions in the (T, v) and (p, v) planes. The reason that p and T are special is
that temperature, pressure, and chemical potential must be equal throughout an equilibrium phase if it
is truly in thermal, mechanical, and chemical equilibrium. This is not the case for an intensive variable
such as specific volume v = NA V /N or chemical concentration xa = Na /N .

2.13 Entropy of Mixing and the Gibbs Paradox

2.13.1 Computing the entropy of mixing

Entropy is widely understood as a measure of disorder. Of course, such a definition should be supple-
mented by a more precise definition of disorder – after all, one man’s trash is another man’s treasure.
To gain
P some intuition about entropy, let us explore the mixing of a multicomponent ideal gas. Let
N = a Na be the total
P number of particles of all species, and let xa = Na /N be the concentration of
species a. Note that a xa = 1.
For any substance obeying the ideal gas law pV = N kB T , the entropy is

S(T, V, N ) = N kB ln(V /N ) + N φ(T ) , (2.357)

∂S

∂p
Nk
since ∂V T,N
= ∂T V,N
= V B . Note that in eqn. 2.357 we have divided V by N before taking the
logarithm. This is essential if the entropy is to be an extensive function (see §2.7.5). One might think
that the configurational entropy of an ideal gas should scale as ln(V N ) = N ln V , since each particle can
be anywhere in the volume V . However, if the particles are indistinguishable, then permuting the particle
labels does not result in a distinct configuration, and so the configurational entropy is proportional to
ln(V N /N !) ∼ N ln(V /N ) − N . The origin of this indistinguishability factor will become clear when we
discuss the quantum mechanical formulation of statistical mechanics. For now, note that such a correction
is necessary in order that the entropy be an extensive function.
If we did not include this factor and instead wrote S ∗ (T, V, N ) = N kB ln V + N φ(T ), then we would
find S ∗ (T, V, N ) − 2S ∗ (T, 21 V, 12 N ) = N kB ln 2, i.e. the total entropy of two identical systems of particles
separated by a barrier will increase if the barrier is removed and they are allowed to mix. This seems
absurd, though, because we could just as well regard the barriers as invisible. This is known as the Gibbs
paradox . The resolution of the Gibbs paradox is to include the indistinguishability correction, which
renders S extensive, in which case S(T, V, N ) = 2S(T, 12 V, 12 N ).

18
The same can be said for multicomponent systems: the phase diagram in the (T, x) plane at constant p looks different
than the phase diagram in the (T, µ) plane at constant p.
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 97

Figure 2.31: A multicomponent system consisting of isolated gases, each at temperature T and pressure
p. Then system entropy increases when all the walls between the different subsystems are removed.

Consider now the situation in Fig. 2.31, where we have separated the different components into their own
volumes Va . Let the pressure and temperature be the same everywhere, so pVa = Na kB T . The entropy
of the unmixed system is then
X Xh i
Sunmixed = Sa = Na kB ln(Va /Na ) + Na φa (T ) . (2.358)
a a

Now let us imagine removing all the barriers separating the different gases and letting the particles mix
thoroughly. The result is that each component gas occupies the full volume V , so the entropy is
X Xh i
Smixed = Sa = Na kB ln(V /Na ) + Na φa (T ) . (2.359)
a a

Thus, the entropy of mixing is

∆Smix = Smixed − Sunmixed
=
X
Na kB ln(V /Va ) = −N kB
X
xa ln xa , (2.360)
a a
Na Va
where xa = N = V is the fraction of species a. Note that ∆Smix ≥ 0.
What if all the components were initially identical? It seems absurd that the entropy should increase
simply by removing some invisible barriers. This is again the Gibbs paradox. In this case, the resolution
of the paradox is to note that the sum in the expression for Smixed is a sum over distinct species. Hence
if the particles are all identical, we have Smixed = N kB ln(V /N ) + N φ(T ) = Sunmixed , hence ∆Smix = 0.

2.13.2 Entropy and combinatorics

As we shall learn when we study statistical mechanics, the entropy may be interpreted in terms of the
number of ways W (E, V, N ) a system at fixed energy and volume can arrange itself. One has
S(E, V, N ) = kB ln W (E, V, N ) . (2.361)
98 CHAPTER 2. THERMODYNAMICS

Figure 2.32: Mixing among three different species of particles. The mixed configuration has an additional
entropy, ∆Smix .

Consider a system consisting of σ different species of particles. Now let it be that for each species label
a, Na particles of that species are confined among Qa little boxes such that at most one particle can fit
in a box (see Fig. 2.32). How many ways W are there to configure N identical particles among Q boxes?
Clearly  
Q Q!
W = = . (2.362)
N N ! (Q − N )!
Q!
Were the particles distinct, we’d have Wdistinct = (Q−N )! , which is N ! times greater. This is because
permuting distinct particles results in a different configuration, and there are N ! ways to permute N
particles.
 
Qa
The entropy for species a is then Sa = kB ln Wa = kB ln N . We then use Stirling’s approximation,
a

ln(K!) = K ln K − K + 12 ln K + 21 ln(2π) + O(K −1 ) , (2.363)

which is an asymptotic expansion valid for K  1. One then finds for Q, N  1, with x = N/Q ∈ [0, 1],
  
Q    

ln = Q ln Q − Q − xQ ln(xQ) − xQ − (1 − x)Q ln (1 − x)Q − (1 − x)Q
N
h i
= −Q x ln x + (1 − x) ln(1 − x) . (2.364)

This is valid up to terms of order Q in Stirling’s expansion. Since ln Q  Q, the next term is small and
we are safe to stop here. Summing up the contributions from all the species, we get
σ
X σ
X h i
Sunmixed = kB ln Wa = −kB Qa xa ln xa + (1 − xa ) ln(1 − xa ) , (2.365)
a=1 a=1

where xa = Na /Qa is the initial dimensionless density of species a.

Now let’s remove all the partitions between
P the different species so that each of the particles is free to
explore all of the boxes. There are Q = a Qa boxes in all. The total number of ways of placing N1
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 99

particles of species a = 1 through Nσ particles of species σ is

Q!
Wmixed = , (2.366)
N0 ! N1 ! · · · Nσ !
Pσ
where N0 = Q − a=1 Na is the number of vacant boxes. Again using Stirling’s rule, we find
σ
X
Smixed = −kB Q x
ea ln x
ea , (2.367)
a=0

where x
ea = Na /Q is the fraction of all boxes containing a particle of species a, and N0 is the number of
empty boxes. Note that
N N Q
ea = a = a · a = xa fa ,
x (2.368)
Q Qa Q
where fa ≡ Qa /Q. Note that σa=1 fa = 1.
P

Qa Na
Let’s assume all the densities are initially the same, so xa = x∀a, so x
ea = x fa . In this case, fa = Q = N
is the fraction of species a among all the particles. We then have x e0 = 1 − x, and
σ
X
Smixed = −kB Q xfa ln(xfa ) − kB Q x
e0 ln x
e0
a=1
h i σ (2.369)
X
= −kB Q x ln x + (1 − x) ln(1 − x) − kB x Q fa ln fa .
a=1

Thus, the entropy of mixing is

σ
X
∆Smix = −N kB fa ln fa , (2.370)
a=1

where N = σa=1 Na is the total number of particles among all species (excluding vacancies) and fa =
P
Na /(N + N0 ) is the fraction of all boxes occupied by species a.

2.13.3 Weak solutions and osmotic pressure

Suppose one of the species is much more plentiful than all the others, and label it with a = 0. We will
call this the solvent. The entropy of mixing is then
σ
"   X  #
N0 Na
∆Smix = −kB N0 ln + Na ln , (2.371)
N0 + N 0 N0 + N 0
a=1
Pσ
where N 0 = a=1 Na is the total number of solvent molecules, summed over all species. We assume the
solution is weak , which means Na ≤ N 0  N0 . Expanding in powers of N 0 /N0 and Na /N0 , we find
σ
"   #
X Na 2
− Na + O N 0 /N0 .


∆Smix = −kB Na ln (2.372)
N0
a=1
100 CHAPTER 2. THERMODYNAMICS

Consider now a solution consisting of N0 molecules of a solvent and Na molecules of species a of solute,
where a = 1, . . . , σ. We begin by expanding the Gibbs free energy G(T, p, N0 , N1 , . . . , Nσ ), where there
are σ species of solutes, as a power series in the small quantities Na . We have
 

X Na
G T, p, N0 , {Na } = N0 g0 (T, p) + kB T Na ln
a
eN0
(2.373)
X 1 X
+ Na ψa (T, p) + Aab (T, p) Na Nb .
a
2N0
a,b

The first term on the RHS corresponds to the Gibbs free energy of the solvent. The second term is due
to the entropy of mixing. The third term is the contribution to the total free energy from the individual
species. Note the factor of e in the denominator inside the logarithm, which accounts for the second term
in the brackets on the RHS of eqn. 2.372. The last term is due to interactions between the species; it is
truncated at second order in the solute numbers.
The chemical potential for the solvent is
∂G X X
µ0 (T, p) = = g0 (T, p) − kB T xa − 21 Aab (T, p) xa xb , (2.374)
∂N0 a a,b

and the chemical potential for species a is

∂G X
µa (T, p) = = kB T ln xa + ψa (T, p) + Aab (T, p) xb , (2.375)
∂Na
b

where xa = Na /N0 is the concentrations of solute species a. By assumption, the last term on the RHS
of each of these equations is small, since Nsolute  N0 , where Nsolute = σa=1 Na is the total number of
P
solute molecules. To lowest order, then, we have
µ0 (T, p) = g0 (T, p) − x kB T (2.376)
µa (T, p) = kB T ln xa + ψa (T, p) , (2.377)
P
where x = a xa is the total solute concentration.
If we add sugar to a solution confined by a semipermeable membrane19 , the pressure increases! To
see why, consider a situation where a rigid semipermeable membrane separates a solution (solvent plus
solutes) from a pure solvent. There is energy exchange through the membrane, so the temperature is T
throughout. There is no volume exchange, however: dV = dV 0 = 0, hence the pressure need not be the
same. Since the membrane is permeable to the solvent, we have that the chemical potential µ0 is the
same on each side. This means
g0 (T, pR ) − xkB T = g0 (T, pL ) , (2.378)
where pL,R is the pressure on the left and right sides of the membrane, and x = N/N0 is again the
total solute concentration. This equation once again tells us that the pressure p cannot be the same on
both sides of the membrane. If the pressure difference is small, we can expand in powers of the osmotic
pressure, π ≡ pR − pL , and we find  
∂µ0
π = x kB T . (2.379)
∂p T
19
‘Semipermeable’ in this context means permeable to the solvent but not the solute(s).
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 101

Figure 2.33: Osmotic pressure causes the column on the right side of the U-tube to rise higher than the
column on the left by an amount ∆h = π/% g.

But a Maxwell relation (§2.9) guarantees

   
∂µ ∂V
= = v(T, p)/NA , (2.380)
∂p T,N ∂N T,p

where v(T, p) is the molar volume of the solvent.

πv = xRT , (2.381)

which looks very much like the ideal gas law, even though we are talking about dense (but ‘weak’)
solutions! The resulting pressure has a demonstrable effect, as sketched in Fig. 2.33. Consider a solution
containing ν moles of sucrose (C12 H22 O11 ) per kilogram (55.52 mol) of water at 30◦ C. We find π = 2.5 atm
when ν = 0.1.
One might worry about the expansion in powers of π when π is much larger than the ambient pressure.
But in fact the next term in the expansion is smaller than the first term by a factor of πκT , where κT is
the isothermal compressibility. For water one has κT ≈ 4.4 × 10−5 (atm)−1 , hence we can safely ignore
the higher order terms in the Taylor expansion.

2.13.4 Effect of impurities on boiling and freezing points

Along the coexistence curve separating liquid and vapor phases, the chemical potentials of the two phases
are identical:
µ0L (T, p) = µ0V (T, p) . (2.382)
Here we write µ0 for µ to emphasize that we are talking about a phase with no impurities present. This
equation provides a single constraint on the two variables T and p, hence one can, in principle, solve to
obtain T = T0∗ (p), which is the equation of the liquid-vapor coexistence curve in the (T, p) plane. Now
suppose there is a solute present in the liquid. We then have

µL (T, p, x) = µ0L (T, p) − xkB T , (2.383)

102 CHAPTER 2. THERMODYNAMICS

Latent Heat Melting Latent Heat of Boiling

Substance of Fusion `˜f Point Vaporization `˜v Point
J/g ◦C J/g ◦C

C2 H5 OH 108 -114 855 78.3

NH3 339 -75 1369 -33.34
CO2 184 -57 574 -78
He – – 21 -268.93
H 58 -259 455 -253
Pb 24.5 372.3 871 1750
N2 25.7 -210 200 -196
O2 13.9 -219 213 -183
H2 O 334 0 2270 100

Table 2.4: Latent heats of fusion and vaporization at p = 1 atm.

where x is the dimensionless solute concentration, summed over all species. The condition for liquid-vapor
coexistence now becomes
µ0L (T, p) − xkB T = µ0V (T, p) . (2.384)
This will lead to a shift in the
boiling temperature at fixed p. Assuming this shift is small, let us expand
to lowest order in T − T0∗ (p) , writing

∂µ0L ∂µ0V
   
µL (T0∗ , p)
0
T0∗ µV (T0∗ , p)
0
T − T0∗ .



+ T− − xkB T = + (2.385)
∂T p ∂T p

Note that    
∂µ ∂S
=− (2.386)
∂T p,N ∂N T,p

from a Maxwell relation deriving from exactness of dG. Since S is extensive, we can write S =
(N/NA ) s(T, p), where s(T, p) is the molar entropy. Solving for T , we obtain
2
xR T0∗ (p)

∗
T (p, x) = T0∗ (p) + , (2.387)
`v (p)

where `v = T0∗ · (sV − sL ) is the latent heat of the liquid-vapor transition20 . The shift ∆T ∗ = T ∗ − T0∗ is
called the boiling point elevation.
As an example, consider seawater, which contains approximately 35 g of dissolved Na+ Cl− per kilogram of
H2 O. The atomic masses of Na and Cl are 23.0 and 35.4, respectively, hence the total ionic concentration
in seawater (neglecting everything but sodium and chlorine) is given by

2 · 35 1000
x= ≈ 0.022 . (2.388)
23.0 + 35.4 18
20
We shall discuss latent heat again in §2.12.2 below.
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 103

The latent heat of vaporization of H2 O at atmospheric pressure is ` = 40.7 kJ/mol, hence

(0.022)(8.3 J/mol K)(373 K)2

∆T ∗ = ≈ 0.6 K . (2.389)
4.1 × 104 J/mol
Put another way, the boiling point elevation of H2 O at atmospheric pressure is about 0.28◦ C per percent
solute. We can express this as ∆T ∗ = Km, where the molality m is the number of moles of solute per
kilogram of solvent. For H2 O, we find K = 0.51◦ C kg/mol.
Similar considerations apply at the freezing point, when we equate the chemical potential of the solvent
plus solute to that of the pure solid. The latent heat of fusion for H2 O is about `f = Tf0 ·(sLIQUID −sSOLID ) =
 2
6.01 kJ/mol21 We thus predict a freezing point depression of ∆T ∗ = −xR T0∗ /`f = 1.03◦ C · x[%]. This
can be expressed once again as ∆T ∗ = −Km, with K = 1.86◦ C kg/mol22 .

2.13.5 Binary solutions

Consider a binary solution, and write the Gibbs free energy G(T, p, NA , NB ) as
 
0 0 NA
G(T, p, NA , NB ) = NA µA (T, p) + NB µB (T, p) + NA kB T ln
NA + NB
  (2.390)
NB NA NB
+ NB kB T ln +λ .
NA + NB NA + NB

The first four terms on the RHS represent the free energy of the individual component fluids and the
entropy of mixing. The last term is an interaction contribution. With λ > 0, the interaction term prefers
that the system be either fully A or fully B. The entropy contribution prefers a mixture, so there is a
competition. What is the stable thermodynamic state?
It is useful to write the Gibbs free energy per particle, g(T, p, x) = G/(NA + NB ), in terms of T , p, and
the concentration x ≡ xB = NB /(NA + NB ) of species B (hence xA = 1 − x is the concentration of species
A). Then
h i
g(T, p, x) = (1 − x) µ0A + x µ0B + kB T x ln x + (1 − x) ln(1 − x) + λ x (1 − x) . (2.391)

In order for the system to be stable against phase separation into relatively A-rich and B-rich regions, we
must have that g(T, p, x) be a convex function of x. Our first check should be for a local instability, i.e.
spinodal decomposition. We have
 
∂g 0 0 x
= µB − µA + kB T ln + λ (1 − 2x) (2.392)
∂x 1−x
and
∂ 2g k T k T
= B + B − 2λ . (2.393)
∂x2 x 1−x
21
See table 2.4, and recall M = 18 g is the molar mass of H2 O.
It is more customary to write ∆T ∗ = Tpure
22 ∗ ∗
solvent − Tsolution in the case of the freezing point depression, in which case
∗
∆T is positive.
104 CHAPTER 2. THERMODYNAMICS

Figure 2.34: Gibbs free energy per particle for a binary solution as a function of concentration x = xB
of the B species (pure A at the left end x = 0 ; pure B at the right end x = 1), in units of the interaction
parameter λ. Dark red curve: T = 0.65 λ/kB > Tc ; green curve: T = λ/2kB = Tc ; blue curve:
T = 0.40 λ/kB < Tc . We have chosen µ0A = 0.60 λ − 0.50 kB T and µ0B = 0.50 λ − 0. 50 kB T . Note that the
free energy g(T, p, x) is not convex in x for T < Tc , indicating an instability and necessitating a Maxwell
construction.

∂ 2g
The spinodal is given by the solution to the equation ∂x2
= 0, which is
2λ
T ∗ (x) = x (1 − x) . (2.394)
kB
Since x (1 − x) achieves its maximum value of 1
4 at x = 12 , we have T ∗ ≤ kB /2λ.
In Fig. 2.34 we sketch the free energy g(T, p, x) versus x for three representative temperatures. For
T > λ/2kB , the free energy is everywhere convex in λ. When T < λ/2kB , there free energy resembles the
blue curve in Fig. 2.34, and the system is unstable to phase separation. The two phases are said to be
immiscible, or, equivalently, there exists a solubility gap. To determine the coexistence curve, we perform
a Maxwell construction, writing
 
g(x2 ) − g(x1 ) ∂g  ∂g 
= = . (2.395)
x2 − x1 ∂x x ∂x x
1 2

Here, x1 and x2 are the boundaries of the two phase region. These equations admit a symmetry of
x ↔ 1 − x, hence we can set x = x1 and x2 = 1 − x. We find
g(1 − x) − g(x) = (1 − 2x) µ0B − µ0A ,


(2.396)
and invoking eqns. 2.395 and 2.392 we obtain the solution
λ 1 − 2x
Tcoex (x) = ·
. (2.397)
kB ln 1−x
x
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 105

Figure 2.35: Upper panels: chemical potential shifts ∆µ± = ∆µA ± ∆µB versus concentration x = xB .
The dashed black line is the spinodal, and the solid black line the coexistence boundary. Temperatures
range from T = 0 (dark blue) to T = 0.6 λ/kB (red) in units of 0.1 λ/kB . Lower panels: phase diagram in
the (T, ∆µ± ) planes. The black dot is the critical point.

The phase diagram for the binary system is shown in Fig. 2.36. For T < T ∗ (x), the system is unstable,
and spinodal decomposition occurs. For T ∗ (x) < T < Tcoex (x), the system is metastable, just like the
van der Waals gas in its corresponding regime. Real binary solutions behave qualitatively like the model
discussed here, although the coexistence curve is generally not symmetric under x ↔ 1 − x, and the single
phase region extends down to low temperatures for x ≈ 0 and x ≈ 1. If λ itself is temperature-dependent,
there can be multiple solutions to eqns. 2.394 and 2.397. For example, one could take

λ0 T 2
λ(T ) = . (2.398)
T 2 + T02

In this case, kB T > λ at both high and low temperatures, and we expect the single phase region to be
reentrant. Such a phenomenon occurs in water-nicotine mixtures, for example.
106 CHAPTER 2. THERMODYNAMICS

Figure 2.36: Phase diagram for the binary system. The black curve is the coexistence curve, and the
dark red curve is the spinodal. A-rich material is to the left and B-rich to the right.

It is instructive to consider the phase diagram in the (T, µ) plane. We define the chemical potential shifts,
∆µA ≡ µA − µ0A = kB T ln(1 − x) + λ x2 (2.399)
∆µB ≡ µB − µ0B 2
= kB T ln x + λ (1 − x) , (2.400)
and their sum and difference, ∆µ± ≡ ∆µA ± ∆µB . From the Gibbs-Duhem relation, we know that we
can write µB as a function of T , p, and µA . Alternately, we could write ∆µ± in terms of T , p, and
∆µ∓ , so we can choose which among ∆µ+ and ∆µ− we wish to use in our phase diagram. The results
are plotted in Fig. 2.35. It is perhaps easiest to understand the phase diagram in the (T, ∆µ− ) plane.
At low temperatures, below T = Tc = λ/2kB , there is a first order phase transition at ∆µ− = 0. For
T < Tc = λ/2kB and ∆µ− = 0+ , i.e. infinitesimally positive, the system is in the A-rich phase, but for
∆µ− = 0− , i.e. infinitesimally negative, it is B-rich. The concentration x = xB changes discontinuously
across the phase boundary. The critical point lies at (T, ∆µ− ) = (λ/2kB , 0).
If we choose N = NA + NB to be the extensive variable, then fixing N means dNA + dNB = 0. So st fixed
T and p,
 
dGT,p = µA dNA + µB dNB ⇒ dg T,p = −∆µ− dx . (2.401)
Since ∆µ− (x, T ) = ϕ(x, T ) − ϕ(1 − x, T ) = −∆µ− (1 − x, T ), where ϕ(x, T ) = λx − kB T ln x, we have that
1−x
the coexistence boundary in the (x, ∆− ) plane is simply the line ∆µ− = 0, because dx0 ∆µ− (x0 , T ) = 0.
R
x

Note also that there is no two-phase region in the (T, ∆µ) plane; the phase boundary in this plane is
a curve which terminates at a critical point. As we saw in §2.12, the same situation pertains in single
component (p, v, T ) systems. That is, the phase diagram in the (p, v) or (T, v) plane contains two-phase
regions, but in the (p, T ) plane the boundaries between phases are one-dimensional curves. Any two-phase
behavior is confined to these curves, where the thermodynamic potentials are singular.
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 107

Figure 2.37: Gibbs free energy per particle g for an ideal binary solution for temperatures T ∈ [TA∗ , TB∗ ].
The Maxwell construction is shown for the case TA∗ < T < TB∗ . Right: phase diagram, showing two-phase
region and distillation sequence in (x, T ) space.

The phase separation can be seen in a number of systems. A popular example involves mixtures of
water and ouzo or other anise-based liqueurs, such as arak and absinthe. Starting with the pure liqueur
(x = 1), and at a temperature below the coexistence curve maximum, the concentration is diluted by
adding water. Follow along on Fig. 2.36 by starting at the point (x = 1 , kB T /λ = 0.4) and move to
the left. Eventually, one hits the boundary of the two-phase region. At this point, the mixture turns
milky, due to the formation of large droplets of the pure phases on either side of coexistence region which
scatter light, a process known as spontaneous emulsification 23 . As one continues to dilute the solution
with more water, eventually one passes all the way through the coexistence region, at which point the
solution becomes clear once again, and described as a single phase.
What happens if λ < 0 ? In this case, both the entropy and the interaction energy prefer a mixed phase,
and there is no instability to phase separation. The two fluids are said to be completely miscible. An
example would be benzene, C6 H6 , and toluene, C7 H8 (i.e. C6 H5 CH3 ). The phase diagram would be
blank, with no phase boundaries below the boiling transition, because the fluid could exist as a mixture
in any proportion.
Any fluid will eventually boil if the temperature is raised sufficiently high. Let us assume that the boiling
points of our A and B fluids are TA,B∗ , and without loss of generality let us take T ∗ < T ∗ at some given
A B
fixed pressure . This means µA (TA∗ , p) = µVA (TA∗ , p) and µLB (TB∗ , p) = µVB (TB∗ , p). What happens to the
24 L

23
An emulsion is a mixture of two or more immiscible liquids.
24
We assume the boiling temperatures are not exactly equal!
108 CHAPTER 2. THERMODYNAMICS

Figure 2.38: Negative (left) and positive (right) azeotrope phase diagrams. From Wikipedia.

mixture? We begin by writing the free energies of the mixed liquid and mixed vapor phases as
h i
gL (T, p, x) = (1 − x) µLA (T, p) + x µLB (T, p) + kB T x ln x + (1 − x) ln(1 − x) + λL x(1 − x) (2.402)
h i
gV (T, p, x) = (1 − x) µVA (T, p) + x µVB (T, p) + kB T x ln x + (1 − x) ln(1 − x) + λV x(1 − x) . (2.403)

Typically λV ≈ 0. Consider these two free energies as functions of the concentration x, at fixed T and
p. If the curves never cross, and gL (x) < gV (x) for all x ∈ [0, 1], then the liquid is always the state of
lowest free energy. This is the situation in the first panel of Fig. 2.37. Similarly, if gV (x) < gL (x) over
this range, then the mixture is in the vapor phase throughout. What happens if the two curves cross at
some value of x? This situation is depicted in the second panel of Fig. 2.37. In this case, there is always
a Maxwell construction which lowers the free energy throughout some range of concentration, i.e. the
system undergoes phase separation.
In an ideal fluid , we have λL = λV = 0, and setting gL = gV requires

(1 − x) ∆µA (T, p) + x ∆µB (T, p) = 0 , (2.404)

where ∆µA/B (T, p) = µLA/B (T, p) − µVA/B (T, p). Expanding the chemical potential about a given tempera-
ture T ∗ ,
cp (T ∗ , p)
µ(T, p) = µ(T ∗ , p) − s(T ∗ , p) (T − T ∗ ) − (T − T ∗ )2 + . . . , (2.405)
2T
∂µ
∂S

∂s


where we have used ∂T p,N = − ∂N T,p = −s(T, p), the entropy per particle, and ∂T p,N = cp /T .
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 109

Figure 2.39: Free energies before Maxwell constructions for a binary fluid mixture in equilibrium with
a vapor (λV = 0). Panels show (a) λL = 0 (ideal fluid), (b) λL < 0 (miscible fluid; negative azeotrope),
(c) λLAB > 0 (positive azeotrope), (d) λLAB > 0 (heteroazeotrope). Thick blue and red lines correspond
to temperatures TA∗ and TB∗ , respectively, with TA∗ < TB∗ . Thin blue and red curves are for temperatures
outside the range [TA∗ , TB∗ ]. The black curves show the locus of points where g is discontinuous, i.e. where
the liquid and vapor free energy curves cross. The yellow curve in (d) corresponds to the coexistence
temperature for the fluid mixture. In this case the azeotrope forms within the coexistence region.

∗ , we have
Thus, expanding ∆µA/B about TA/B

cVpA − cLpA
L V V L
∆µA ≡ µA − µA = (sA − sA )(T − TA∗ ) + (T − TA∗ )2 + . . .
2TA∗
(2.406)
cpB − cLpB
V
L V V L
∆µB ≡ µB − µB = (sB − sB )(T − TB∗ ) + (T − TB∗ )2 + ...
2TB∗
We assume sVA/B > sLA/B , i.e. the vapor phase has greater entropy per particle. Thus, ∆µA/B (T ) changes
∗ . If we assume that these are the only sign changes
sign from negative to positive as T rises through TA/B
for ∆µA/B (T ) at fixed p, then eqn. 2.404 can only be solved for T ∈ [TA∗ , TB∗ ]. This immediately leads to
the phase diagram in the rightmost panel of Fig. 2.37.
According to the Gibbs phase rule, with σ = 2, two-phase equilibrium (ϕ = 2) occurs along a subspace
110 CHAPTER 2. THERMODYNAMICS

of dimension dPE = 2 + σ − ϕ = 2. Thus, if we fix the pressure p and the concentration x = xB , liquid-gas
equilibrium occurs at a particular temperature T ∗ , known as the boiling point. Since the liquid and the
vapor with which it is in equilibrium at T ∗ may have different composition, i.e. different values of x, one
may distill the mixture to separate the two pure substances, as follows. First, given a liquid mixture of
A and B, we bring it to boiling, as shown in the rightmost panel of Fig. 2.37. The vapor is at a different
concentration x than the liquid (a lower value of x if the boiling point of pure A is less than that of
pure B, as shown). If we collect the vapor, the remaining fluid is at a higher value of x. The collected
vapor is then captured and then condensed, forming a liquid at the lower x value. This is then brought
to a boil, and the resulting vapor is drawn off and condensed, etc The result is a purified A state. The
remaining liquid is then at a higher B concentration. By repeated boiling and condensation, A and B
can be separated. For liquid-vapor transitions, the upper curve, representing the lowest temperature at
a given concentration for which the mixture is a homogeneous vapor, is called the dew point curve. The
lower curve, representing the highest temperature at a given concentration for which the mixture is a
homogeneous liquid, is called the bubble point curve. The same phase diagram applies to liquid-solid
mixtures where both phases are completely miscible. In that case, the upper curve is called the liquidus,
and the lower curve the solidus.
When a homogeneous liquid or vapor at concentration x is heated or cooled to a temperature T such that
(x, T ) lies within the two-phase region, the mixture phase separates into the the two end components
(x∗L , T ) and (x∗V , T ), which lie on opposite sides of the boundary of the two-phase region, at the same
temperature. The locus of points at constant T joining these two points is called the tie line. To
determine how much of each of these two homogeneous phases separates out, we use particle number
conservation. If ηL,V is the fraction of the homogeneous liquid and homogeneous vapor phases present,
then ηL x∗L + ηV x∗V = x, which says ηL = (x − x∗V )/(x∗L − x∗V ) and ηV = (x − x∗L )/(x∗V − x∗L ). This is known
as the lever rule.
For many binary mixtures, the boiling point curve is as shown in Fig. 2.38. Such cases are called
azeotropes. For negative azeotropes, the maximum of the boiling curve lies above both TA,B ∗ . The free
∗ ∗
energy curves for this case are shown in panel (b) of Fig. 2.39. For x < x , where x is the azeotropic
composition, one can distill A but not B. Similarly, for x > x∗ one can distill B but not A. The
situation is different for positive azeotropes, where the minimum of the boiling curve lies below both
∗ , corresponding to the free energy curves in panel (c) of Fig. 2.39. In this case, distillation (i.e.
TA,B
condensing and reboiling the collected vapor) from either side of x∗ results in the azeotrope. One can
of course collect the fluid instead of the vapor. In general, for both positive and negative azeotropes,
starting from a given concentration x, one can only arrive at pure A plus azeotrope (if x < x∗ ) or pure B
plus azeotrope (if x > x∗ ). Ethanol (C2 H5 OH) and water (H2 O) form a positive azeotrope which is 95.6%
ethanol and 4.4% water by weight. The individual boiling points are TC∗ 2 H5 OH = 78.4◦ C , TH∗ 2 O = 100◦ C,
while the azeotrope boils at TAZ∗ = 78.2◦ C. No amount of distillation of this mixture can purify ethanol

beyond the 95.6% level. To go beyond this level of purity, one must resort to azeotropic distillation, which
involves introducing another component, such as benzene (or a less carcinogenic additive), which alters
the molecular interactions.
To model the azeotrope system, we need to take λL 6= 0, in which case one can find two solutions to
the energy crossing condition gV (x) = gL (x). With two such crossings come two Maxwell constructions,
hence the phase diagrams in Fig. 2.38. Generally, negative azeotropes are found in systems with λL < 0 ,
whereas positive azeotropes are found when λL > 0. As we’ve seen, such repulsive interactions between
2.14. SOME CONCEPTS IN THERMOCHEMISTRY 111

Figure 2.40: Phase diagram for a eutectic mixture in which a liquid L is in equilibrium with two solid
phases α and β. The same phase diagram holds for heteroazeotropes, where a vapor is in equilibrium
with two liquid phases.

the A and B components in general lead to a phase separation below a coexistence temperature TCOEX (x)
given by eqn. 2.397. What happens if the minimum boiling point lies within the coexistence region?
This is the situation depicted in panel (d) of Fig. 2.39. The system is then a liquid/vapor version of the
solid/liquid eutectic (see Fig. 2.40), and the minimum boiling point mixture is called a heteroazeotrope.

2.14 Some Concepts in Thermochemistry

2.14.1 Chemical reactions and the law of mass action

Suppose we have a chemical reaction among σ species, written as

ζ1 A1 + ζ2 A2 + · · · + ζσ Aσ = 0 , (2.407)

where

Aa = chemical formula
ζa = stoichiometric coefficient .

For example, we could have

− 3 H2 − N2 + 2 NH3 = 0 (3 H2 + N2
2 NH3 ) (2.408)

for which
ζ(H2 ) = −3 , ζ(N2 ) = −1 , ζ(NH3 ) = 2 . (2.409)
112 CHAPTER 2. THERMODYNAMICS

When ζa > 0, the corresponding Aa is a product; when ζa < 0, the corresponding Aa is a reactant. The
bookkeeping of the coefficients ζa which ensures conservation of each individual species of atom in the
reaction(s) is known as stoichiometry 25
Now we ask: what are the conditions for equilibrium? At
constant T and p, which is typical for many
chemical reactions, the conditions are that G T, p, {Na } be a minimum. Now
X
dG = −S dT + V dp + µa dNa , (2.410)
i

so if we let the reaction go forward, we have dNa = ζa , and if it runs in reverse we have dNa = −ζa .
Thus, setting dT = dp = 0, we have the equilibrium condition
σ
X
ζa µ a = 0 . (2.411)
a=1

Let us investigate the consequences of this relation for ideal gases. The chemical potential of the ath
species is
µa (T, p) = kB T φa (T ) + kB T ln pa . (2.412)
P
Here pa = p xa is the partial pressure of species a, where xa = Na / b Nb the dimensionless concentration
of species a. Chemists sometimes write xa = [Aa ] for the concentration of species a. In equilibrium we
must have X h i
ζa ln p + ln xa + φa (T ) = 0 , (2.413)
a

which says X X X
ζa ln xa = − ζa ln p − ζa φa (T ) . (2.414)
a a a

Exponentiating, we obtain the law of mass action:

Y P
 X 
xaζa = p− a ζa exp − ζa φa (T ) ≡ κ(p, T ) . (2.415)
a a
25
Antoine Lavoisier, the ”father of modern chemistry”, made pioneering contributions in both chemistry and biology. In
particular, he is often credited as the progenitor of stoichiometry. An aristocrat by birth, Lavoisier was an administrator of
the Ferme générale, an organization in pre-revolutionary France which collected taxes on behalf of the king. At the age of
28, Lavoisier married Marie-Anne Pierette Paulze, the 13-year-old daughter of one of his business partners. She would later
join her husband in his research, and she played a role in his disproof of the phlogiston theory of combustion. The phlogiston
theory was superseded by Lavoisier’s work, where, based on contemporary experiments by Joseph Priestley, he correctly
identified the pivotal role played by oxygen in both chemical and biological processes (i.e. respiration). Despite his fame as
a scientist, Lavoisier succumbed to the Reign of Terror. His association with the Ferme générale, which collected taxes from
the poor and the downtrodden, was a significant liability in revolutionary France (think Mitt Romney vis-a-vis Bain Capital).
Furthermore – and let this be a lesson to all of us – Lavoisier had unwisely ridiculed a worthless pseudoscientific pamphlet,
ostensibly on the physics of fire, and its author, Jean-Paul Marat. Marat was a journalist with scientific pretensions, but
apparently little in the way of scientific talent or acumen. Lavoisier effectively blackballed Marat’s candidacy to the French
Academy of Sciences, and the time came when Marat sought revenge. Marat was instrumental in getting Lavoisier and
other members of the Ferme générale arrested on charges of counterrevolutionary activities, and on May 8, 1794, after a
trial lasting less than a day, Lavoisier was guillotined. Along with Fourier and Carnot, Lavoisier’s name is one of the 72
engraved on the Eiffel Tower. Source: http://www.vigyanprasar.gov.in/scientists/ALLavoisier.htm.
2.14. SOME CONCEPTS IN THERMOCHEMISTRY 113

The quantity κ(p, T ) is called the equilibrium constant. When κ is large, the LHS of the above equation
is large. This favors maximal concentration xa for the products (ζa > 0) and minimal concentration xa
for the reactants (ζa < 0). This means that the equation REACTANTS
PRODUCTS is shifted to the
right, i.e. the products are plentiful and the reactants are scarce. When κ is small, the LHS is small and
the reaction is shifted to the left, i.e. the reactants are plentiful and the products are scarce.
P Remember
we are describing equilibrium conditions here. Now we observe that reactions for which a ζa > 0 shift
to the Pleft with increasing pressure and shift to the right with decreasing pressure, while reactions for
which a ζa > 0 the situation is P reversed: they shift to the right with increasing pressure and to the left
with decreasing pressure. When a ζa = 0 there is no shift upon increasing or decreasing pressure.
The rate at which the equilibrium constant changes with temperature is given by
 
∂ ln κ X
=− ζa φ0a (T ) . (2.416)
∂T p a

Now from eqn. 2.412 we have that the enthalpy per particle for species i is
 
∂µa
ha = µa − T , (2.417)
∂T p

since H = G + T S and S = − ∂G


∂T p . We find

ha = −kB T 2 φ0a (T ) , (2.418)

and thus   P
∂ ln κ i ζa ha ∆h
= = , (2.419)
∂T p k T2
B kB T 2
where ∆h is the enthalpy of the reaction, which is the heat absorbed or emitted as a result of the reaction.
When ∆h > 0 the reaction is endothermic and the yield increases with increasing T . When ∆h < 0 the
reaction is exothermic and the yield decreases with increasing T .
As an example, consider the reaction H2 + I2
2 HI. We have
ζ(H2 ) = −1 , ζ(I2 ) = −1 ζ(HI) = 2 . (2.420)
Suppose our initial system consists of ν10 moles of H2 , ν20 = 0 moles of I2 , and ν30 moles of undissociated
0
P
HI . These mole numbers determine the initial concentrations xa , where xa = νa / b νb . Define
x03 − x3
α≡ , (2.421)
x3
in which case we have
x1 = x01 + 21 αx03 , x2 = 12 αx03 , x3 = (1 − α) x03 . (2.422)
Then the law of mass action gives
4 (1 − α)2
=κ. (2.423)
α(α + 2r)
where r ≡ x01 /x03 = ν10 /ν30 . This yields
P a quadratic equation, which can be solved to find α(κ, r). Note
that κ = κ(T ) for this reaction since a ζa = 0. The enthalpy of this reaction is positive: ∆h > 0.
114 CHAPTER 2. THERMODYNAMICS

∆Hf0 ∆Hf0
Formula Name State kJ/mol Formula Name State kJ/mol
Ag Silver crystal 0.0 NiSO4 Nickel sulfate crystal -872.9
C Graphite crystal 0.0 Al2 O3 Aluminum oxide crystal -1657.7
C Diamond crystal 1.9 Ca3 P2 O8 Calcium phosphate gas -4120.8
O3 Ozone gas 142.7 HCN Hydrogen cyanide liquid 108.9
H2 O Water liquid -285.8 SF6 Sulfur hexafluoride gas -1220.5
H3 BO3 Boric acid crystal -1094.3 CaF2 Calcium fluoride crystal -1228.0
ZnSO4 Zinc sulfate crystal -982.8 CaCl2 Calcium chloride crystal -795.4

Table 2.5: Enthalpies of formation of some common substances.

2.14.2 Enthalpy of formation

Most chemical reactions take place under constant pressure. The heat Qif associated with a given isobaric
process is
Zf Zf
Qif = dE + p dV = (Ef − Ei ) + p (Vf − Vi ) = Hf − Hi , (2.424)
i i

where H is the enthalpy,

H = E + pV . (2.425)
Note that the enthalpy H is a state function, since E is a state function and p and V are state variables.
Hence, we can meaningfully speak of changes in enthalpy: ∆H = Hf −Hi . If ∆H < 0 for a given reaction,
we call it exothermic – this is the case when Qif < 0 and thus heat is transferred to the surroundings. Such
reactions can occur spontaneously, and, in really fun cases, can produce explosions. The combustion of
fuels is always exothermic. If ∆H > 0, the reaction is called endothermic. Endothermic reactions require
that heat be supplied in order for the reaction to proceed. Photosynthesis is an example of an endothermic
reaction.
Suppose we have two reactions
(∆H)
1
A + B −−−−→ C (2.426)
and
(∆H)
2
C + D −−−−→ E. (2.427)
Then we may write
(∆H)
3
A + B + D −−−−→ E, (2.428)
with
(∆H)1 + (∆H)2 = (∆H)3 . (2.429)
We can use this additivity of reaction enthalpies to define a standard molar enthalpy of formation. We
first define the standard state of a pure substance at a given temperature to be its state (gas, liquid, or
solid) at a pressure p = 1 bar. The standard reaction enthalpies at a given temperature are then defined
to be the reaction enthalpies when the reactants and products are all in their standard states. Finally,
2.14. SOME CONCEPTS IN THERMOCHEMISTRY 115

Figure 2.41: Left panel: reaction enthalpy and activation energy (exothermic case shown). Right panel:
reaction enthalpy as a difference between enthalpy of formation of reactants and products.

we define the standard molar enthalpy of formation ∆Hf0 (X) of a compound X at temperature T as the
reaction enthalpy for the compound X to be produced by its constituents when they are in their standard
state. For example, if X = SO2 , then we write
f∆H0 [SO2 ]
S + O2 −−−−− −−−→ SO2 . (2.430)

The enthalpy of formation of any substance in its standard state is zero at all temperatures, by definition:
∆H0f [O2 ] = ∆H0f [He] = ∆H0f [K] = ∆H0f [Mn] = 0, etc.
Suppose now we have a reaction
∆H
a A + b B −−−−→ c C + d D . (2.431)
To compute the reaction enthalpy ∆H, we can imagine forming the components A and B from their
standard state constituents. Similarly, we can imagine doing the same for C and D. Since the number
of atoms of a given kind is conserved in the process, the constituents of the reactants must be the same
as those of the products, we have

∆H = −a ∆Hf0 (A) − b ∆Hf0 (B) + c ∆Hf0 (C) + d ∆Hf0 (D) . (2.432)

A list of a few enthalpies of formation is provided in table 2.5. Note that the reaction enthalpy is
independent of the actual reaction path. That is, the difference in enthalpy between A and B is the same
whether the reaction is A −→ B or A −→ X −→ (Y + Z) −→ B. This statement is known as Hess’s
Law .
Note that
dH = dE + p dV + V dp = d̄Q + V dp , (2.433)
hence    
dQ
¯ ∂H
Cp = = . (2.434)
dT p ∂T p
116 CHAPTER 2. THERMODYNAMICS

enthalpy enthalpy enthalpy enthalpy

bond (kJ/mol) bond (kJ/mol) bond (kJ/mol) bond (kJ/mol)
H−H 436 C−C 348 C−S 259 F−F 155
H−C 412 C=C 612 N−N 163 F − Cl 254
H−N 388 C≡C 811 N=N 409 Cl − Br 219
H−O 463 C−N 305 N≡N 945 Cl − I 210
H−F 565 C=N 613 N−O 157 Cl − S 250
H − Cl 431 C≡N 890 N−F 270 Br − Br 193
H − Br 366 C−O 360 N − Cl 200 Br − I 178
H−I 299 C=O 743 N − Si 374 Br − S 212
H−S 338 C−F 484 O−O 146 I−I 151
H−P 322 C − Cl 338 O=O 497 S−S 264
H − Si 318 C − Br 276 O−F 185 P−P 172
C−I 238 O − Cl 203 Si − Si 176

Table 2.6: Average bond enthalpies for some common bonds. (Source: L. Pauling, The Nature of the
Chemical Bond (Cornell Univ. Press, NY, 1960).)

We therefore have
ZT
H(T, p, ν) = H(T0 , p, ν) + ν dT 0 cp (T 0 ) . (2.435)
T0
1
For ideal gases, we have cp (T ) = (1 + 2 f ) R. For real gases, over a range of temperatures, there are small
variations:
cp (T ) = α + β T + γ T 2 . (2.436)
Two examples (300 K < T < 1500 K, p = 1 atm):
J J J
O2 : α = 25.503 , β = 13.612 × 10−3 , γ = −42.553 × 10−7
mol K mol K2 mol K3
J J J
H2 O : α = 30.206 , β = 9.936 × 10−3 , γ = 11.14 × 10−7
mol K mol K2 mol K3

If all the gaseous components in a reaction can be approximated as ideal, then we may write
X
(∆H)rxn = (∆E)rxn + ζa RT , (2.437)
a

where the subscript ‘rxn’ stands for ‘reaction’. Here (∆E)rxn is the change in energy from reactants to
products.

2.14.3 Bond enthalpies

The enthalpy needed to break a chemical bond is called the bond enthalpy, h[ • ]. The bond enthalpy
is the energy required to dissociate one mole of gaseous bonds to form gaseous atoms. A table of bond
2.14. SOME CONCEPTS IN THERMOCHEMISTRY 117

Figure 2.42: Calculation of reaction enthalpy for the hydrogenation of ethene (ethylene), C2 H4 .

enthalpies is given in Tab. 2.6. Bond enthalpies are endothermic, since energy is required to break
chemical bonds. Of course, the actual bond energies can depend on the location of a bond in a given
molecule, and the values listed in the table reflect averages over the possible bond environment.
The bond enthalpies in Tab. 2.6 may be used to compute reaction enthalpies. Consider, for example, the
reaction 2 H2 (g) + O2 (g) −→ 2 H2 O(l). We then have, from the table,

(∆H)rxn = 2 h[H−H] + h[O = O] − 4 h[H−O]

(2.438)
= −483 kJ/mol O2 .

Thus, 483 kJ of heat would be released for every two moles of H2 O produced, if the H2 O were in the
gaseous phase. Since H2 O is liquid at STP, we should also include the condensation energy of the gaseous
water vapor into liquid water. At T = 100◦ C the latent heat of vaporization is `˜ = 2270 J/g, but at
T = 20◦ C, one has `˜ = 2450 J/g, hence with M = 18 we have ` = 44.1 kJ/mol. Therefore, the heat
produced by the reaction 2 H2 (g) + O2 (g) −* )− 2 H2 O(l) is (∆H)rxn = −571.2 kJ / mol O2 . Since the
reaction produces two moles of water, we conclude that the enthalpy of formation of liquid water at STP
is half this value: ∆H0f [H2 O] = 285.6 kJ/mol.
Consider next the hydrogenation of ethene (ethylene): C2 H4 + H2 )− −* C2 H6 . The product is known
as ethane. The energy accounting is shown in Fig. 2.42. To compute the enthalpies of formation of
ethene and ethane from the bond enthalpies, we need one more bit of information, which is the standard
enthalpy of formation of C(g) from C(s), since the solid is the standard state at STP. This value is
118 CHAPTER 2. THERMODYNAMICS

∆H0f [C(g)] = 718 kJ/mol. We may now write

−2260 kJ
2 C(g) + 4 H(g) −−−−−−−−→ C2 H4 (g)
1436 kJ
2 C(s) −−−−−−−−→ 2 C(g)
872 kJ
2 H2 (g) −−−−−−−−→ 4 H(g) .

Thus, using Hess’s law, i.e. adding up these reaction equations, we have
48 kJ
2 C(s) + 2 H2 (g) −−−−−−−−→ C2 H4 (g) .

Thus, the formation of ethene is endothermic. For ethane,

−2820 kJ
2 C(g) + 6 H(g) −−−−−−−−→ C2 H6 (g)
1436 kJ
2 C(s) −−−−−−−−→ 2 C(g)
1306 kJ
3 H2 (g) −−−−−−−−→ 6 H(g)

For ethane,
−76 kJ
2 C(s) + 3 H2 (g) −−−−−−−−→ C2 H6 (g) ,
which is exothermic.

2.15 Appendix I : Integrating Factors

Suppose we have an inexact differential

dW
¯ = Ai dxi . (2.439)
Here I am adopting the ‘Einstein convention’ where we sum over repeated indices unless otherwise ex-
plicitly stated; Ai dxi = i Ai dxi . An integrating factor eL(~x) is a function which, when divided into
P
dF
¯ , yields an exact differential:
∂U
dU = e−L dW¯ = dx . (2.440)
∂xi i
Clearly we must have
∂2U ∂ ∂
e−L Aj = e−L Ai .



= (2.441)
∂xi ∂xj ∂xi ∂xj

Applying the Leibniz rule and then multiplying by eL yields

∂Aj ∂L ∂Ai ∂L
− Aj = − Ai . (2.442)
∂xi ∂xi ∂xj ∂xj

If there are K independent variables {x1 , . . . , xK }, then there are 21 K(K − 1) independent equations of
the above form – one for each distinct (i, j) pair. These equations can be written compactly as

∂L
Ωijk = Fij , (2.443)
∂xk
2.16. APPENDIX II : LEGENDRE TRANSFORMATIONS 119

where

Ωijk = Aj δik − Ai δjk (2.444)

∂Aj ∂Ai
Fij = − . (2.445)
∂xi ∂xj

Note that Fij is antisymmetric, and resembles a field strength tensor, and that Ωijk = −Ωjik is antisym-
metric in the first two indices (but is not totally antisymmetric in all three).
Can we solve these 21 K(K − 1) coupled equations to find an integrating factor L? In general the answer
is no. However, when K = 2 we can always find an integrating factor. To see why, let’s call x ≡ x1 and
y ≡ x2 . Consider now the ODE
dy A (x, y)
=− x . (2.446)
dx Ay (x, y)
This equation can be integrated to yield a one-parameter set of integral curves, indexed by an initial
condition. The equation for these curves may be written as Uc (x, y) = 0, where c labels the curves. Then
along each curve we have

dUc ∂Ux ∂Uc dy

0= = +
dx ∂x ∂y dx
(2.447)
∂Uc Ax ∂Uc
= − .
∂x Ay ∂y

Thus,
∂Uc ∂Uc
Ay = A ≡ e−L Ax Ay . (2.448)
∂x ∂y x
This equation defines the integrating factor L :
   
1 ∂Uc 1 ∂Uc
L = − ln = − ln . (2.449)
Ax ∂x Ay ∂y

We now have that

∂Uc ∂Uc
Ax = eL , Ay = eL , (2.450)
∂x ∂y
and hence
∂Uc ∂Uc
e−L dW
¯ = dx + dy = dUc . (2.451)
∂x ∂y

2.16 Appendix II : Legendre Transformations

A convex function of a single variable f (x) is one for which f 00 (x) > 0 everywhere. The Legendre transform
of a convex function f (x) is a function g(p) defined as follows. Let p be a real number, and consider the
line y = px, as shown in Fig. 2.43. We define the point x(p) as the value of x for which the difference
120 CHAPTER 2. THERMODYNAMICS

Figure 2.43: Construction for the Legendre transformation of a function f (x).

F (x, p) = px − f (x) is greatest. Then define g(p) = F x(p), p .26 The value x(p) is unique if f (x) is

convex, since x(p) is determined by the equation

f 0 x(p) = p .


(2.452)

Note that from p = f 0 x(p) we have, according to the chain rule,

d 0 h
i−1
f x(p) = f 00 x(p) x0 (p) x0 (p) = f 00 x(p)



=⇒ . (2.453)
dp
From this, we can prove that g(p) is itself convex:
dh
i
g 0 (p) = p x(p) − f x(p)
dp (2.454)
= p x0 (p) + x(p) − f 0 x(p) x0 (p) = x(p) ,

hence
00 0
h
00

i−1
g (p) = x (p) = f x(p) >0. (2.455)

In higher dimensions, the generalization of the definition f 00 (x) > 0 is that a function F (x1 , . . . , xn ) is
convex if the matrix of second derivatives, called the Hessian,
∂ 2F
Hij (x) = (2.456)
∂xi ∂xj

is positive definite. That is, all the eigenvalues of Hij (x) must be positive for every x. We then define
the Legendre transform G(p) as
G(p) = p · x − F (x) (2.457)
26
Note that g(p) may be a negative number, if the line y = px lies everywhere below f (x).
2.16. APPENDIX II : LEGENDRE TRANSFORMATIONS 121

where
p = ∇F . (2.458)
Note that
dG = x · dp + p · dx − ∇F · dx = x · dp , (2.459)
which establishes that G is a function of p and that
∂G
= xj . (2.460)
∂pj
Note also that the Legendre transformation is self dual , which is to say that the Legendre transform of
G(p) is F (x): F → G → F under successive Legendre transformations.
We can also define a partial Legendre transformation as follows. Consider a function of q variables F (x, y),
where x = {x1 , . . . , xm } and y = {y1 , . . . , yn }, with q = m + n. Define p = {p1 , . . . , pm }, and
G(p, y) = p · x − F (x, y) , (2.461)
where
∂F
pa = (a = 1, . . . , m) . (2.462)
∂xa
These equations are then to be inverted to yield
∂G
xa = xa (p, y) = . (2.463)
∂pa

Note that
∂F

pa = x(p, y), y . (2.464)
∂xa
Thus, from the chain rule,
∂pa ∂ 2F ∂xc ∂ 2F ∂ 2G
δab = = = , (2.465)
∂pb ∂xa ∂xc ∂pb ∂xa ∂xc ∂pc ∂pb
which says
∂ 2G ∂xa
= = K−1
ab , (2.466)
∂pa ∂pb ∂pb
where the m × m partial Hessian is
∂ 2F ∂p
= a = Kab . (2.467)
∂xa ∂xb ∂xb
Note that Kab = Kba is symmetric. And with respect to the y coordinates,
∂ 2G ∂ 2F
=− = −Lµν , (2.468)
∂yµ ∂yν ∂yµ ∂yν
where
∂ 2F
Lµν = (2.469)
∂yµ ∂yν
is the partial Hessian in the y coordinates. Now it is easy to see that if the full q × q Hessian matrix Hij
is positive definite, then any submatrix such as Kab or Lµν must also be positive definite. In this case,
the partial Legendre transform is convex in {p1 , . . . , pm } and concave in {y1 , . . . , yn }.
122 CHAPTER 2. THERMODYNAMICS

2.17 Appendix III : Useful Mathematical Relations

Consider a set of n independent variables {x1 , . . . , xn }, which can be thought of as a point in n-dimensional
space. Let {y1 , . . . , yn } and {z1 , . . . , zn } be other choices of coordinates. Then
∂xi ∂xi ∂yj
= . (2.470)
∂zk ∂yj ∂zk

Note that this entails a matrix multiplication: Aik = Bij Cjk , where Aik = ∂xi /∂zk , Bij = ∂xi /∂yj , and
Cjk = ∂yj /∂zk . We define the determinant
 
∂xi ∂(x1 , . . . , xn )
det ≡ . (2.471)
∂zk ∂(z1 , . . . , zn )
Such a determinant is called a Jacobian. Now if A = BC, then det(A) = det(B) · det(C). Thus,
∂(x1 , . . . , xn ) ∂(x1 , . . . , xn ) ∂(y1 , . . . , yn )
= · . (2.472)
∂(z1 , . . . , zn ) ∂(y1 , . . . , yn ) ∂(z1 , . . . , zn )
Recall also that
∂xi
= δik . (2.473)
∂xk

Consider the case n = 2. We have

∂x ∂x



∂u v ∂v u       
∂(x, y) ∂x ∂y ∂x ∂y
= det   = − . (2.474)
 
∂(u, v) ∂y

∂y ∂u v ∂v u ∂v u ∂u v
∂u v ∂v u

We also have
∂(x, y) ∂(u, v) ∂(x, y)
· = . (2.475)
∂(u, v) ∂(r, s) ∂(r, s)
From this simple mathematics follows several very useful results.
1) First, write
" #−1
∂(x, y) ∂(u, v)
= . (2.476)
∂(u, v) ∂(x, y)
Now let v = y :  
∂(x, y) ∂x 1
= =
. (2.477)
∂(u, y) ∂u y
∂u
∂x y
Thus,  
∂x .  ∂u 
=1 (2.478)
∂u y ∂x y

2) Second, we have
    
∂(x, y) ∂x ∂(x, y) ∂(x, u) ∂y ∂x
= = · =− ,
∂(u, y) ∂u y ∂(x, u) ∂(u, y) ∂u x ∂y u
2.17. APPENDIX III : USEFUL MATHEMATICAL RELATIONS 123

which is to say      
∂x ∂y ∂x
=− . (2.479)
∂y u ∂u x ∂u y

Invoking eqn. 2.478, we conclude that

    
∂x ∂y ∂u
= −1 . (2.480)
∂y u ∂u x ∂x y

3) Third, we have
∂(x, v) ∂(x, v) ∂(y, v)
= · , (2.481)
∂(u, v) ∂(y, v) ∂(u, v)
which says     
∂x ∂x ∂y
= (2.482)
∂u v ∂y v ∂u v

This is simply the chain rule of partial differentiation.

4) Fourth, we have

∂(x, y) ∂(x, y) ∂(u, v)

= ·
∂(u, y) ∂(u, v) ∂(u, y)
          (2.483)
∂x ∂y ∂v ∂x ∂y ∂v
= − ,
∂u v ∂v u ∂y u ∂v u ∂u v ∂y u

which says        
∂x ∂x ∂x ∂y
= − (2.484)
∂u y ∂u v ∂y u ∂u v

5) Fifth, whenever we differentiate one extensive quantity with respect to another, holding only intensive
quantities constant, the result is simply the ratio of those extensive quantities. For example,
 
∂S S
= . (2.485)
∂V p,T V

The reason should be obvious. In the above example, S(p, V, T ) = V φ(p, T ), where φ is a function of the
two intensive quantities p and T . Hence differentiating S with respect to V holding p and T constant is
the same as dividing S by V . Note that this implies
     
∂S ∂S ∂S S
= = = , (2.486)
∂V p,T ∂V p,µ ∂V n,T V

where n = N/V is the particle density.

6) Sixth, suppose we have a function Φ(y, v) and we write

dΦ = x dy + u dv . (2.487)
124 CHAPTER 2. THERMODYNAMICS

That is,    
∂Φ ∂Φ
x= ≡ Φy , u= ≡ Φv . (2.488)
∂y v ∂v y

Now we may write

dx = Φyy dy + Φyv dv (2.489)

du = Φvy dy + Φvv dv . (2.490)

If we demand du = 0, this yields

Φyy
 
∂x
= . (2.491)
∂u v Φvy
Note that Φvy = Φyv . From the equation du = 0 we also derive
 
∂y Φvv
=− . (2.492)
∂v u Φvy

Next, we use eqn. 2.490 with du = 0 to eliminate dy in favor of dv, and then substitute into eqn. 2.489.
This yields
Φyy Φvv
 
∂x
= Φyv − . (2.493)
∂v u Φvy
Finally, eqn. 2.490 with dv = 0 yields  
∂y 1
= . (2.494)
∂u v Φvy

Combining the results of eqns. 2.491, 2.492, 2.493, and 2.494, we have
      
∂(x, y) ∂x ∂y ∂x ∂y
= −
∂(u, v) ∂u v ∂v u ∂v u ∂u v
(2.495)
Φyy Φyy Φvv
     
Φvv 1
= − − Φyv − = −1 .
Φvy Φvy Φvy Φvy

Thus, if Φ = E(S, V ),̇ then (x, y) = (T, S) and (u, v) = (−p, V ), we have

∂(T, S)
= −1 . (2.496)
∂(−p, V )

Nota bene: It is important to understand what other quantities are kept constant, otherwise we can run
into trouble. For example, it would seem that eqn. 2.495 would also yield
∂(µ, N )
=1. (2.497)
∂(p, V )
But then we should have
∂(T, S) ∂(T, S) ∂(−p, V )
= · = +1 (WRONG!)
∂(µ, N ) ∂(−p, V ) ∂(µ, N )
2.17. APPENDIX III : USEFUL MATHEMATICAL RELATIONS 125

when according to eqn. 2.495 it should be −1. What has gone wrong?
The problem is that we have not properly specified what else is being held constant. In eqn. 2.496 it is
N (or µ) which is being held constant, while in eqn. 2.497 it is S (or T ) which is being held constant.
Therefore a naive application of the chain rule for determinants yields the wrong result, as we have seen.
Let’s be more careful. Applying the same derivation to dE = x dy + u dv + r ds and holding s constant,
we conclude        
∂(x, y, s) ∂x ∂y ∂x ∂y
= − = −1 . (2.498)
∂(u, v, s) ∂u v,s ∂v u,s ∂v u,s ∂u v,s
Thus, if
dE = T dS + y dX + µ dN , (2.499)
where (y, X) = (−p, V ) or (H α , M α ) or (E α , P α ), the appropriate thermodynamic relations are

∂(T, S, N ) ∂(T, S, µ)
= −1 = −1
∂(y, X, N ) ∂(y, X, µ)
∂(µ, N, X) ∂(µ, N, y)
= −1 = −1 (2.500)
∂(T, S, X) ∂(T, S, y)
∂(y, X, S) ∂(y, X, T )
= −1 = −1
∂(µ, N, S) ∂(µ, N, T )

For example,
∂(T, S, N ) ∂(−p, V, S) ∂(µ, N, V )
= = = −1 (2.501)
∂(−p, V, N ) ∂(µ, N, S) ∂(T, S, V )
and
∂(T, S, µ) ∂(−p, V, T ) ∂(µ, N, −p)
= = = −1 . (2.502)
∂(−p, V, µ) ∂(µ, N, T ) ∂(T, S, −p)

If we are careful, then the results in eq. 2.500 can be quite handy, especially when used in conjunction
with eqn. 2.472. For example, we have
=1
  z }| {  
∂S ∂(T, S, N ) ∂(T, S, N ) ∂(p, V, N ) ∂p
= = · = , (2.503)
∂V T,N ∂(T, V, N ) ∂(p, V, N ) ∂(T, V, N ) ∂T V,N

which is one of the Maxwell relations derived from the exactness of dF (T, V, N ). Some other examples
include
=1
  z }| {  
∂V ∂(V, p, N ) ∂(V, p, N ) ∂(S, T, N ) ∂T
= = · = (2.504)
∂S p,N ∂(S, p, N ) ∂(S, T, N ) ∂(S, p, N ) ∂p S,N
=1
  z }| {  
∂S ∂(S, T, p) ∂(S, T, p) ∂(µ, N, p) ∂µ
= = · =− , (2.505)
∂N T,p ∂(N, T, p) ∂(µ, N, p) ∂(N, T, p) ∂T p,N
126 CHAPTER 2. THERMODYNAMICS

which are Maxwell relations deriving from dH(S, p, N ) and dG(T, p, N ), respectively. Note that due to
the alternating nature of the determinant – it is antisymmetric under interchange of any two rows or
columns – we have
∂(x, y, z) ∂(y, x, z) ∂(y, x, z)
=− = = ... . (2.506)
∂(u, v, w) ∂(u, v, w) ∂(w, v, u)

In general, it is usually advisable to eliminate S from a Jacobian. If we have a Jacobian involving T , S,

and N , we can write
∂(T, S, N ) ∂(T, S, N ) ∂(p, V, N ) ∂(p, V, N )
= = , (2.507)
∂( • , • , N ) ∂(p, V, N ) ∂( • , • , N ) ∂( • , • , N )
where each • is a distinct arbitrary state variable other than N .
If our Jacobian involves the S, V , and N , we write

∂(S, V, N ) ∂(S, V, N ) ∂(T, V, N ) C ∂(T, V, N )

= · = V · . (2.508)
∂( • , • , N ) ∂(T, V, N ) ∂( • , • , N ) T ∂( • , • , N )

If our Jacobian involves the S, p, and N , we write

∂(S, p, N ) ∂(S, p, N ) ∂(T, p, N ) Cp ∂(T, p, N )

= · = · . (2.509)
∂( • , • , N ) ∂(T, p, N ) ∂( • , • , N ) T ∂( • , • , N )

For example,
=1
  z }| {  
∂T ∂(T, S, N ) ∂(T, S, N ) ∂(p, V, N ) ∂(p, T, N ) T ∂V
= = · · = (2.510)
∂p S,N ∂(p, S, N ) ∂(p, V, N ) ∂(p, T, N ) ∂(p, S, N ) Cp ∂T p,N
   
∂V ∂(V, S, N ) ∂(V, S, N ) ∂(V, T, N ) ∂(p, T, N ) CV ∂V
= = · · = . (2.511)
∂p S,N ∂(p, S, N ) ∂(V, T, N ) ∂(p, T, N ) ∂(p, S, N ) Cp ∂p T,N

With κ ≡ − V1 ∂V ∂p the compressibility, we see that the second of these equations says κT cV = κS cp ,
relating the isothermal and adiabatic compressibilities and the molar heat capacities at constant volume
and constant pressure. This relation was previously established in eqn. 2.292
Chapter 3

Ergodicity and the Approach to

Equilibrium

3.1 References

– R. Balescu, Equilibrium and Nonequilibrium Statistical Mechanics (Wiley, 1975)

An advanced text with an emphasis on fluids and kinetics.

– R. Balian, From Macrophysics to Microphysics (2 vols., Springer-Verlag, 2006)

A very detailed discussion of the fundamental postulates of statistical mechanics and their implica-
tions.)

127
128 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

3.2 Modeling the Approach to Equilibrium

3.2.1 Equilibrium

A thermodynamic system typically consists of an enormously large number of constituent particles, a

typical ‘large number’ being Avogadro’s number, NA = 6.02 × 1023 . Nevertheless, in equilibrium, such
a system is characterized by a relatively small number of thermodynamic state variables. Thus, while
a complete description of a (classical) system would require us to account for O 1023 evolving degrees

of freedom, with respect to the physical quantities in which we are interested, the details of the initial
conditions are effectively forgotten over some microscopic time scale τ , called the collision time, and over
some microscopic distance scale, `, called the mean free path1 . The equilibrium state is time-independent.

3.2.2 The Master Equation

Relaxation to equilibrium is often modeled with something called the master equation. Let Pi (t) be the
probability that the system is in a quantum or classical state i at time t. Then write

dPi X

= Wij Pj − Wji Pi . (3.1)
dt
j

Here, Wij is the rate at which j makes a transition to i. Note that we can write this equation as

dPi X
=− Γij Pj , (3.2)
dt
j

where (
−W if i 6= j
Γij = P0 ij (3.3)
k Wkj if i = j ,

where the prime on the sum indicates that k = j is to be excluded. The constraints on the Wij are that
Wij ≥ 0 for all i, j, and we may take Wii ≡ 0 (no sum on i). Fermi’s Golden Rule of quantum mechanics
says that
2π  2
Wij = h i | V̂ | j i ρ(Ej ) , (3.4)
~
   
where Ĥ0  i = Ei  i , V̂ is an additional potential which leads to transitions, and ρ(Ei ) is the density
of final states at energy Ei . The fact that Wij ≥ 0 means that if each Pi (t = 0) ≥ 0, then Pi (t) ≥ 0 for all
t ≥ 0. To see this, suppose that at some time t > 0 one of the P probabilities Pi is crossing zero and about
to become negative. But then eqn. 3.1 says that Ṗi (t) = j Wij Pj (t) ≥ 0. So Pi (t) can never become
negative.
1
Exceptions involve quantities which are conserved by collisions, such as overall particle number, momentum, and energy.
These quantities relax to equilibrium in a special way called hydrodynamics.
3.2. MODELING THE APPROACH TO EQUILIBRIUM 129

3.2.3 Equilibrium distribution and detailed balance

If the transition rates Wij are themselves time-independent, then we may formally write

Pi (t) = e−Γ t


ij
Pj (0) . (3.5)

Here we have used the Einstein ‘summation convention’ in which repeated indices are summed over (in
this case, the j index). Note that X
Γij = 0 , (3.6)
i
P
which says that the total probability i Pi is conserved:

d X X X X 
Pi = − Γij Pj = − Pj Γij = 0 . (3.7)
dt
i i,j j i

We conclude that φ ~ = (1, 1, . . . , 1) is a left eigenvector of Γ with eigenvalue λ = 0. The corresponding

right eigenvector, which we write as Pieq , satisfies Γij Pjeq = 0, and is a stationary (i.e. time independent)
solution to the master equation. Generally, there is only one right/left eigenvector pair corresponding to
λ = 0, in which case any initial probability distribution Pi (0) converges to Pieq as t → ∞, as shown in
Appendix I (§3.7).
In equilibrium, the net rate of transitions into a state | i i is equal to the rate of transitions out of | i i.
If, for each state | j i the transition rate from | i i to | j i is equal to the transition rate from | j i to | i i,
we say that the rates satisfy the condition of detailed balance. In other words,

Wij Pjeq = Wji Pieq . (3.8)

Assuming Wij 6= 0 and Pjeq 6= 0, we can divide to obtain

Wji Pjeq
= eq . (3.9)
Wij Pi

Note that detailed balance is a stronger condition than that required for a stationary solution to the
master equation.
If Γ = Γ t is symmetric, then the right eigenvectors and left eigenvectors are transposes of each other,
hence P eq = 1/N , where N is the dimension of Γ . The system then satisfies the conditions of detailed
balance. See Appendix II (§3.8) for an example of this formalism applied to a model of radioactive decay.

3.2.4 Boltzmann’s H-theorem

Suppose for the moment that Γ is a symmetric matrix, i.e. Γij = Γji . Then construct the function
X
H(t) = Pi (t) ln Pi (t) . (3.10)
i
130 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Then
dH X dPi X dP
i
= 1 + ln Pi ) = ln Pi
dt dt dt
i i
X
=− Γij Pj ln Pi (3.11)
i,j
X

= Γij Pj ln Pj − ln Pi ,
i,j
P
where we have used i Γij = 0. Now switch i ↔ j in the above sum and add the terms to get

dH 1X


= Γij Pi − Pj ln Pi − ln Pj . (3.12)
dt 2
i,j

Note that the i = j term does not contribute to the sum. For i 6= j we have Γij = −Wij ≤ 0, and using
the result
(x − y) (ln x − ln y) ≥ 0 , (3.13)
we conclude
dH
≤0. (3.14)
dt
In equilibrium, Pieq is a constant, independent of i. We write

1
Pieq =
X
, Ω= 1 =⇒ H = − ln Ω . (3.15)
Ω
i

If Γij 6= Γji , we can still prove a version of the H-theorem. Define a new symmetric matrix

W ij ≡ Wij Pjeq = Wji Pieq = W ji , (3.16)

and the generalized H-function,  

X P (t)
H(t) ≡ Pi (t) ln i eq . (3.17)
Pi
i

Then "  #
Pj Pj
  
dH 1 X Pi Pi
=− W ij − ln − ln ≤0. (3.18)
dt 2 Pieq Pjeq Pieq Pjeq
i,j

3.3 Phase Flows in Classical Mechanics

3.3.1 Hamiltonian evolution

The master equation provides us with a semi-phenomenological description of a dynamical system’s

relaxation to equilibrium. It explicitly breaks time reversal symmetry. Yet the microscopic laws of Nature
3.3. PHASE FLOWS IN CLASSICAL MECHANICS 131

are (approximately) time-reversal symmetric. How can a system which obeys Hamilton’s equations of
motion come to equilibrium?
Let’s start our investigation by reviewing the basics of Hamiltonian dynamics. Recall the
R Lagrangian
L = L(q, q̇, t) = T − V . The Euler-Lagrange equations of motion for the action S q(t) = dt L are
 
d ∂L ∂L
ṗσ = = , (3.19)
dt ∂ q̇σ ∂qσ
where pσ is the canonical momentum conjugate to the generalized coordinate qσ :
∂L
pσ = . (3.20)
∂ q̇σ

The Hamiltonian, H(q, p) is obtained by a Legendre transformation,

r
X
H(q, p) = pσ q̇σ − L . (3.21)
σ=1

Note that
r  
X ∂L ∂L ∂L
dH = pσ dq̇σ + q̇σ dpσ − dqσ − dq̇σ − dt
∂qσ ∂ q̇σ ∂t
σ=1
r   (3.22)
X ∂L ∂L
= q̇σ dpσ − dqσ − dt .
∂qσ ∂t
σ=1

Thus, we obtain Hamilton’s equations of motion,

∂H ∂H ∂L
= q̇σ , =− = −ṗσ (3.23)
∂pσ ∂qσ ∂qσ
and
dH ∂H ∂L
= =− . (3.24)
dt ∂t ∂t
Define the rank 2r vector ϕ by its components,

qi
 if 1 ≤ i ≤ r
ϕi = (3.25)

pi−r if r ≤ i ≤ 2r .


Then we may write Hamilton’s equations compactly as

∂H
ϕ̇i = Jij , (3.26)
∂ϕj
where !
0r×r 1r×r
J= (3.27)
−1r×r 0r×r

is a rank 2r matrix. Note that J t = −J, i.e. J is antisymmetric, and that J 2 = −12r×2r .
132 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

3.3.2 Dynamical systems and the evolution of phase space volumes

Consider a general dynamical system,

dϕ
= V (ϕ) , (3.28)
dt
where ϕ(t) is a point in an n-dimensional phase space. Consider now a compact2 region R0 in phase
space, and consider its evolution under the dynamics. That is, R0 consists of a set of points ϕ | ϕ ∈ R0 ,
and if we regard each ϕ ∈ R0 as an initial condition, we can define the time-dependent set R(t) as the
set of points ϕ(t) that were in R0 at time t = 0:
 
R(t) = ϕ(t)  ϕ(0) ∈ R0 . (3.29)

Now consider the volume Ω(t) of the set R(t). We have

Z
Ω(t) = dµ (3.30)
R(t)

where
dµ = dϕ1 dϕ2 · · · dϕn , (3.31)
for an n-dimensional phase space. We then have
 
 ∂ϕi (t + dt) 
Z Z
0
Ω(t + dt) = dµ = dµ 
 , (3.32)
∂ϕ (t)  j
R(t+dt) R(t)

where
 ∂ϕi (t + dt)  ∂(ϕ01 , . . . , ϕ0n )
 
 ∂ϕ (t)  ≡ ∂(ϕ , . . . , ϕ ) (3.33)
 
j 1 n

is a determinant, which
 0 is the Jacobean
of the transformation from the set of coordinates ϕi = ϕi (t)
to the coordinates ϕi = ϕi (t + dt) . But according to the dynamics, we have

ϕi (t + dt) = ϕi (t) + Vi ϕ(t) dt + O(dt2 )

(3.34)

and therefore
∂ϕi (t + dt) ∂Vi
= δij + dt + O(dt2 ) . (3.35)
∂ϕj (t) ∂ϕj
We now make use of the equality
ln det M = Tr ln M , (3.36)
for any matrix M , which gives us3 , for small ε,

2 
det 1 + εA = exp Tr ln 1 + εA = 1 + ε Tr A + 21 ε2 Tr A − Tr (A2 ) + . . .



(3.37)
2
‘Compact’ in the parlance of mathematical analysis means ‘closed and bounded’.
3
The equality ln det M = Tr ln M is most easily proven by bringing the matrix to diagonal form via a similarity transfor-
mation, and proving the equality for diagonal matrices.
3.3. PHASE FLOWS IN CLASSICAL MECHANICS 133

Thus, Z
Ω(t + dt) = Ω(t) + dµ ∇·V dt + O(dt2 ) , (3.38)
R(t)

which says
dΩ
Z Z
= dµ ∇·V = dS n̂ · V (3.39)
dt
R(t) ∂R(t)

Here, the divergence is the phase space divergence,

n
X ∂Vi
∇·V = , (3.40)
∂ϕi
i=1

and we have used the divergence theorem to convert the volume integral of the divergence to a surface
integral of n̂ · V , where n̂ is the surface normal and dS is the differential element of surface area, and
∂R denotes the boundary of the region R. We see that if ∇·V = 0 everywhere in phase space, then Ω(t)
is a constant, and phase space volumes are preserved by the evolution of the system.
For an alternative derivation, consider a function %(ϕ, t) which is defined to be the density of some
collection of points in phase space at phase space position ϕ and time t. This must satisfy the continuity
equation,
∂%
+ ∇·(%V ) = 0 . (3.41)
∂t
This is called the continuity equation. It says that ‘nobody gets lost’. If we integrate it over a region of
phase space R, we have
d
Z Z Z
dµ % = − dµ ∇·(%V ) = − dS n̂ · (%V ) . (3.42)
dt
R R ∂R

It is perhaps helpful to think of % as a charge density, in which case J = %V is the current density. The
above equation then says
dQR
Z
= − dS n̂ · J , (3.43)
dt
∂R

where QR is the total charge contained inside the region R. In other words, the rate of increase or
decrease of the charge within the region R is equal to the total integrated current flowing in or out of R
at its boundary.
The Leibniz rule lets us write the continuity equation as

∂%
+ V ·∇% + % ∇·V = 0 . (3.44)
∂t
But now suppose that the phase flow is divergenceless, i.e. ∇·V = 0. Then we have
 
D% ∂
≡ + V ·∇ % = 0 . (3.45)
Dt ∂t
134 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 3.1: Time evolution of two immiscible fluids. The local density remains constant.

The combination inside the brackets above is known as the convective derivative. It tells us the total rate
of change of % for an observer moving with the phase flow . That is

d
∂% dϕi ∂%
% ϕ(t), t = +
dt ∂ϕi dt ∂t
n (3.46)
X ∂ρ ∂% D%
= Vi + = .
∂ϕi ∂t Dt
i=1

If D%/Dt = 0, the local density remains the same during the evolution of the system. If we consider the
‘characteristic function’ (
1 if ϕ ∈ R0
%(ϕ, t = 0) = (3.47)
0 otherwise
then the vanishing of the convective derivative means that the image of the set R0 under time evolution
will always have the same volume.
Hamiltonian evolution in classical mechanics is volume preserving. The equations of motion are

∂H ∂H
q̇i = + , ṗi = − (3.48)
∂pi ∂qi

A point in phase space is specified by r positions qi and r momenta pi , hence the dimension of phase
space is n = 2r:      
q q̇ ∂H/∂p
ϕ= , V = = . (3.49)
p ṗ −∂H/∂q
3.3. PHASE FLOWS IN CLASSICAL MECHANICS 135

Hamilton’s equations of motion guarantee that the phase space flow is divergenceless:
r  
X ∂ ṗi
∂ q̇i
∇·V = +
∂qi ∂pi
i=1
r
(    ) (3.50)
X ∂ ∂H ∂ ∂H
= + − =0.
∂qi ∂pi ∂pi ∂qi
i=1

Thus, we have that the convective derivative vanishes, viz.

D% ∂%
≡ + V ·∇% = 0 , (3.51)
Dt ∂t
for any distribution %(ϕ, t) on phase space. Thus, the value of the density %(ϕ(t), t) is constant, which
tells us that the phase flow is incompressible. In particular, phase space volumes are preserved.

3.3.3 Liouville’s equation and the microcanonical distribution

Let %(ϕ) = %(q, p) be a distribution on phase space. Assuming the evolution is Hamiltonian, we can
write
r  
∂% X ∂ ∂
= −ϕ̇ · ∇% = − q̇k + ṗk % = −iL̂% , (3.52)
∂t ∂qk ∂pk
k=1

where L̂ is a differential operator known as the Liouvillian:

r
( )
X ∂H ∂ ∂H ∂
L̂ = −i − . (3.53)
∂pk ∂qk ∂qk ∂pk
k=1

Eqn. 3.52, known as Liouville’s equation, bears an obvious resemblance to the Schrödinger equation from
quantum mechanics.
Suppose that Λa (ϕ) is conserved by the dynamics of the system. Typical conserved quantities include
the components of the total linear momentum (if there is translational invariance), the components of the
total angular momentum (if there is rotational invariance), and the Hamiltonian itself (if the Lagrangian
is not explicitly time-dependent). Now consider a distribution %(ϕ, t) = %(Λ1 , Λ2 , . . . , Λk ) which is a
function only of these various conserved quantities. Then from the chain rule, we have
X ∂%
ϕ̇ · ∇% = ϕ̇ · ∇Λa = 0 , (3.54)
a
∂Λa

since for each a we have

r  
dΛa X ∂Λa ∂Λa
= q̇ + ṗ = ϕ̇ · ∇Λa = 0 . (3.55)
dt ∂qσ σ ∂pσ σ
σ=1

We conclude that any distribution %(ϕ, t) = %(Λ1 , Λ2 , . . . , Λk ) which is a function solely of conserved
dynamical quantities is a stationary solution to Liouville’s equation.
136 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Clearly the microcanonical distribution,

δ E − H(ϕ) δ E − H(ϕ)
%E (ϕ) = =R
, (3.56)
D(E) dµ δ E − H(ϕ)

is a fixed point solution of Liouville’s equation.

3.4 Irreversibility and Poincaré Recurrence

The dynamics of the master equation describe an approach to equilibrium. These dynamics are irre-
versible: dH/dt ≤ 0, where H is Boltzmann’s H-function. However, the microscopic laws of physics are
(almost) time-reversal invariant4 , so how can we understand the emergence of irreversibility? Further-
more, any dynamics which are deterministic and volume-preserving in a finite phase space exhibits the
phenomenon of Poincaré recurrence, which guarantees that phase space trajectories are arbitrarily close
to periodic if one waits long enough.

3.4.1 Poincaré recurrence theorem

The proof of the recurrence theorem is simple. Let gτ be the ‘τ -advance mapping’ which evolves points
in phase space according to Hamilton’s equations. Assume that gτ is invertible and volume-preserving,
as is the case for Hamiltonian flow. Further assume that phase space volume is finite. Since energy is
preserved in the case of time-independent Hamiltonians, we simply ask that the volume of phase space
at fixed total energy E be finite, i.e.
Z


dµ δ E − H(q, p) < ∞ , (3.57)

where dµ = dq dp is the phase space uniform integration measure.

Theorem: In any finite neighborhood R0 of phase space there exists a point ϕ0 which will return to R0
after m applications of gτ , where m is finite.
Proof: Assume the theorem fails; we will show this assumption results in a contradiction. Consider the
set Υ formed from the union of all sets gτk R for all m:
∞
[
Υ= gτk R0 (3.58)
k=0

We assume that the set {gτk R0 | k ∈ N} is disjoint5 . The volume of a union of disjoint sets is the sum of
4
Actually, the microscopic laws of physics are not time-reversal invariant, but rather are invariant under the product
P CT , where P is parity, C is charge conjugation, and T is time reversal.
5
The natural numbers N is the set of non-negative integers {0, 1, 2, . . .}.
3.4. IRREVERSIBILITY AND POINCARÉ RECURRENCE 137

Figure 3.2: Successive images of a set R0 under the τ -advance mapping gτ , projected onto a two-
dimensional phase plane. The Poincaré recurrence theorem guarantees that if phase space has finite
volume, and gτ is invertible and volume preserving, then for any set R0 there exists an integer m such
that R0 ∩ gτm R0 6= ∅.

the individual volumes. Thus,

∞
X
vol gτk R0


vol(Υ) =
k=0
∞ (3.59)
X
= vol(R0 ) · 1=∞,
k=0

since vol gτk R0 = vol R0 from volume preservation. But clearly Υ is a subset of the entire phase space,

hence we have a contradiction, because by assumption phase space is of finite volume.

Thus, the assumption that the set {gτk R0 | k ∈ Z+ } is disjoint fails. This means that there exists some
pair of integers k and l, with k 6= l, such that gτk R0 ∩ gτl R0 6= ∅. Without loss of generality we may
assume k < l. Apply the inverse gτ−1 to this relation k times to get gτl−k R0 ∩ R0 6= ∅. Now choose any
point ϕ1 ∈ gτm R0 ∩ R0 , where m = l − k, and define ϕ0 = gτ−m ϕ1 . Then by construction both ϕ0 and
gτm ϕ0 lie within R0 and the theorem is proven.
Poincaré recurrence has remarkable implications. Consider a bottle of perfume which is opened in an
otherwise evacuated room, as depicted in fig. 3.3. The perfume molecules evolve according to Hamiltonian
evolution. The positions are bounded because physical space is finite. The momenta are bounded because
the total energy is conserved, hence no single particle can have a momentum such that T (p) > ETOT ,
where T (p) is the single particle kinetic energy function6 . Thus, phase space, however large, is still
bounded. Hamiltonian evolution, as we have seen, is invertible and volume preserving, therefore the
6
In the nonrelativistic limit, T = p2 /2m. For relativistic particles, we have T = (p2 c2 + m2 c4 )1/2 − mc2 .
138 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 3.3: Poincaré recurrence guarantees that if we remove the cap from a bottle of perfume in an
otherwise evacuated room, all the perfume molecules will eventually return to the bottle! (Here H is the
Hubble constant.)

system is recurrent. All the molecules must eventually return to the bottle. What’s more, they all must
return with momenta arbitrarily close to their initial momenta!7 In this case, we could define the region
R0 as
R0 = (q1 , . . . , qr , p1 , . . . , pr )  |qi − qi0 | ≤ ∆q and |pj − p0j | ≤ ∆p ∀ i, j ,
 
(3.60)

which specifies a hypercube in phase space centered about the point (q 0 , p0 ).

Each of the three central assumptions – finite phase space, invertibility, and volume preservation – is
crucial. If any one of these assumptions does not hold, the proof fails. Obviously if phase space is infinite
the flow needn’t be recurrent since it can keep moving off in a particular direction. Consider next a
volume-preserving map which is not invertible. An example might be a mapping f : R → R which takes
any real number to its fractional part. Thus, f (π) = 0.14159265 . . .. Let us restrict our attention to
intervals of width less than unity. Clearly f is then volume preserving. The action of f on the interval
[2, 3) is to map it to the interval [0, 1). But [0, 1) remains fixed under the action of f , so no point within
the interval [2, 3) will ever return under repeated iterations of f . Thus, f does not exhibit Poincaré
recurrence.
Consider next the case of the damped harmonic oscillator. In this case, phase space volumes contract.
For a one-dimensional oscillator obeying ẍ + 2β ẋ + Ω02 x = 0 one has ∇ · V = −2β < 0, since β > 0
for physical damping. Thus the convective derivative is Dt % = −(∇ · V )% = 2β% which says that the
density increases exponentially in the comoving frame, as %(t) = e2βt %(0). Thus, phase space volumes
collapse: Ω(t) = e−2β2 Ω(0), and are not preserved by the dynamics. The proof of recurrence therefore
fails. In this case, it is possible for the set Υ to be of finite volume, even if it is the union of an infinite
number of sets gτk R0 , because the volumes of these component sets themselves decrease exponentially, as
vol(gτn R0 ) = e−2nβτ vol(R0 ). A damped pendulum, released from rest at some small angle θ0 , will not
return arbitrarily close to these initial conditions.

7
Actually, what the recurrence theorem guarantees is that there is a configuration arbitrarily close to the initial one which
recurs, to within the same degree of closeness.
3.4. IRREVERSIBILITY AND POINCARÉ RECURRENCE 139

Figure 3.4: Left: A configuration of the Kac ring with N = 16 sites and F = 4 flippers. The flippers,
which live on the links, are represented by blue dots. Right: The ring system after one time step.
Evolution proceeds by clockwise rotation. Spins passing through flippers are flipped.

3.4.2 Kac ring model

The implications of the Poincaré recurrence theorem are surprising – even shocking. If one takes a bottle
of perfume in a sealed, evacuated room and opens it, the perfume molecules will diffuse throughout the
room. The recurrence theorem guarantees that after some finite time T all the molecules will go back
inside the bottle (and arbitrarily close to their initial velocities as well). The hitch is that this could take
a very long time, e.g. much much longer than the age of the Universe.
On less absurd time scales, we know that most systems come to thermodynamic equilibrium. But
how can a system both exhibit equilibration and Poincaré recurrence? The two concepts seem utterly
incompatible!
A beautifully simple model due to Kac shows how a recurrent system can exhibit the phenomenon of
equilibration. Consider a ring with N sites. On each site, place a ‘spin’ which can be in one of two states:
up or down. Along the N links of the system, F of them contain ‘flippers’. The configuration of the
flippers is set at the outset and never changes. The dynamics of the system are as follows: during each
time step, every spin moves clockwise a distance of one lattice spacing. Spins which pass through flippers
reverse their orientation: up becomes down, and down becomes up.
The ‘phase space’ for this system consists of 2N discrete configurations. Since each configuration maps
onto a unique image under the evolution of the system, phase space ‘volume’ is preserved. The evolution
is invertible; the inverse is obtained simply by rotating the spins counterclockwise. Figure 3.4 depicts an
example configuration for the system, and its first iteration under the dynamics.
Suppose the flippers were not fixed, but moved about randomly. In this case, we could focus on a single
spin and determine its configuration probabilistically. Let pn be the probability that a given spin is in
140 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 3.5: Three simulations of the Kac ring model with N = 2500 sites and three different concen-
trations of flippers. The red line shows the magnetization as a function of time, starting from an initial
configuration in which 100% of the spins are up. The blue line shows the prediction of the Stosszahlansatz ,
which yields an exponentially decaying magnetization with time constant τ .

the up configuration at time n. The probability that it is up at time (n + 1) is then

pn+1 = (1 − x) pn + x (1 − pn ) , (3.61)
where x = F/N is the fraction of flippers in the system. In words: a spin will be up at time (n + 1) if it
was up at time n and did not pass through a flipper, or if it was down at time n and did pass through
a flipper. If the flipper locations are randomized at each time step, then the probability of flipping is
simply x = F/N . Equation 3.61 can be solved immediately:
pn = 1
2 + (1 − 2x)n (p0 − 12 ) , (3.62)
1
which decays exponentially to the equilibrium value of peq = 2 with time scale
1
τ (x) = − . (3.63)
ln |1 − 2x|
3.4. IRREVERSIBILITY AND POINCARÉ RECURRENCE 141

Figure 3.6: Simulations of the Kac ring model. Top: N = 2500 sites with F = 201 flippers. After
2500 iterations, each spin has flipped an odd number of times, so the recurrence time is 2N . Middle:
N = 2500 with F = 2400, resulting in a near-complete reversal of the population with every iteration.
Bottom: N = 25000 with N = 1000, showing long time equilibration and dramatic resurgence of the spin
population.

We identify τ (x) as the microscopic relaxation time over which local equilibrium is established. If we
define the magnetization m ≡ (N↑ − N↓ )/N , then m = 2p − 1, so mn = (1 − 2x)n m0 . The equilibrium
magnetization is meq = 0. Note that for 21 < x < 1 that the magnetization reverses sign each time step,
as well as decreasing exponentially in magnitude.
The assumption that leads to equation 3.61 is called the Stosszahlansatz 8 , a long German word mean-
ing, approximately, ‘assumption on the counting of hits’. The resulting dynamics are irreversible: the
magnetization inexorably decays to zero. However, the Kac ring model is purely deterministic, and the
Stosszahlansatz can at best be an approximation to the true dynamics. Clearly the Stosszahlansatz fails
to account for correlations such as the following: if spin i is flipped at time n, then spin i + 1 will have
8
Unfortunately, many important physicists were German and we have to put up with a legacy of long German words like
Gedankenexperiment, Zitterbewegung, Brehmsstrahlung, Stosszahlansatz , Kartoffelsalat, etc.
142 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

been flipped at time n − 1. Also if spin i is flipped at time n, then it also will be flipped at time n + N .
Indeed, since the dynamics of the Kac ring model are invertible and volume preserving, it must exhibit
Poincaré recurrence. We see this most vividly in figs. 3.5 and 3.6.
The model is trivial to simulate. The results of such a simulation are shown in figure 3.5 for a ring of
N = 1000 sites, with F = 100 and F = 24 flippers. Note how the magnetization decays and fluctuates
about the equilibrium value meq = 0, but that after N iterations m recovers its initial value: mN = m0 .
The recurrence time for this system is simply N if F is even, and 2N if F is odd, since every spin will
then have flipped an even number of times.
In figure 3.6 we plot two other simulations. The top panel shows what happens when x > 21 , so that the
magnetization wants to reverse its sign with every iteration. The bottom panel shows a simulation for a
larger ring, with N = 25000 sites. Note that the fluctuations in m about equilibrium are smaller than in
the cases with N = 1000 sites. Why?

3.5 Remarks on Ergodic Theory

3.5.1 Definition of ergodicity

A mechanical system evolves according to Hamilton’s equations of motion. We have seen how such a
system is recurrent in the sense of Poincaré.
There is a level beyond recurrence called ergodicity. In an ergodic system, time averages over intervals
[0, T ] with T → ∞ may be replaced by phase space averages. The time average of a function f (ϕ) is
defined as
ZT

 1

f (ϕ) t = lim dt f ϕ(t) . (3.64)
T →∞ T
0
For a Hamiltonian system, the phase space average of the same function is defined by
Z Z




f (ϕ) S = dµ f (ϕ) δ E − H(ϕ) dµ δ E − H(ϕ) , (3.65)

where H(ϕ) = H(q, p) is the Hamiltonian, and where δ(x) is the Dirac δ-function. Thus,



ergodicity ⇐⇒ f (ϕ) t = f (ϕ) S , (3.66)


for all smooth functions f (ϕ) for which f (ϕ) S exists and is finite. Note that we do not average over
all of phase space. Rather, we average only over a hypersurface along which H(ϕ) = E is fixed, i.e. over
one of the level sets of the Hamiltonian function. This is because the dynamics preserves the energy.
Ergodicity means that almost all points ϕ will, upon Hamiltonian evolution, move in such a way as to
eventually pass through every finite neighborhood on the energy surface, and will spend equal time in
equal regions of phase space.
Let χR (ϕ) be the characteristic function of a region R:
(
1 if ϕ ∈ R
χR (ϕ) = (3.67)
0 otherwise,
3.5. REMARKS ON ERGODIC THEORY 143

where H(ϕ) = E for all ϕ ∈ R. Then

 
time spent in R
χR (ϕ) = lim


t
. (3.68)
T →∞ T
If the system is ergodic, then
DR (E)
χR (ϕ)


t
= P (R) = , (3.69)
D(E)
where P (R) is the a priori probability to find ϕ ∈ R, based solely on the relative volumes of R and of
the entire phase space. The latter is given by
Z


D(E) = dµ δ E − H(ϕ) , (3.70)

called the density of states, is the surface area of phase space at energy E, and
Z


DR (E) = dµ δ E − H(ϕ) . (3.71)
R

is the density of states for the phase space subset R. Note that

dS
Z Z


D(E) ≡ dµ δ E − H(ϕ) = (3.72)
|∇H|
SE

d
dΩ(E)
Z
= dµ Θ E − H(ϕ) = . (3.73)
dE dE
Here, dS is the differential surface element, SE is the constant H hypersurface H(ϕ) = E, and Ω(E) is
the volume of phase space over which H(ϕ) < E. Note also that we may write
dµ = dE dΣE , (3.74)
where 
dS 
dΣE = (3.75)
|∇H| H(ϕ)=E
is the the invariant surface element.

3.5.2 The microcanonical ensemble

The distribution,


δ E − H(ϕ) δ E − H(ϕ)
%E (ϕ) = =R
, (3.76)
D(E) dµ δ E − H(ϕ)
defines the microcanonical ensemble (µCE) of Gibbs.
We could also write Z

 1
f (ϕ) S = dΣE f (ϕ) , (3.77)
D(E)
SE

integrating over the hypersurface SE rather than the entire phase space.
144 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 3.7: Constant phase space velocity at an irrational angle over a toroidal phase space is ergodic,
but not mixing. A circle remains a circle, and a blob remains a blob.

3.5.3 Ergodicity and mixing

Just because a system is ergodic, it doesn’t necessarily

 mean that %(ϕ, t) → %eq (ϕ), for consider the
following motion on the toroidal space ϕ = (q, p)  0 ≤ q < 1 , 0 ≤ p < 1 , where we identify opposite
edges, i.e. we impose periodic boundary conditions. We also take q and p to be dimensionless, for
simplicity of notation. Let the dynamics be given by

q̇ = 1 , ṗ = α . (3.78)

The solution is
q(t) = q0 + t , p(t) = p0 + αt , (3.79)
hence the phase curves are given by
p = p0 + α(q − q0 ) . (3.80)

Now consider the average of some function f (q, p). We can write f (q, p) in terms of its Fourier transform,
X
f (q, p) = fˆmn e2πi(mq+np) . (3.81)
m,n

We have, then,

X
f q(t), p(t) = fˆmn e2πi(mq0 +np0 ) e2πi(m+αn)t . (3.82)
m,n

We can now perform the time average of f :

1 X0 ˆ e2πi(m+αn)T − 1
f (q, p) t = fˆ00 + lim fmn e2πi(mq0 +np0 )


T →∞ T m,n 2πi(m + αn) (3.83)
= fˆ00 if α irrational.

Clearly,
Z1 Z1
dq dp f (q, p) = fˆ00 = f (q, p) t ,



f (q, p) S = (3.84)
0 0

so the system is ergodic.

3.5. REMARKS ON ERGODIC THEORY 145

Figure 3.8: The baker’s transformation is a successive stretching, cutting, and restacking.

The situation is depicted in fig. 3.7. If we start with the characteristic function of a disc,

%(q, p, t = 0) = Θ a2 − (q − q0 )2 − (p − p0 )2 ,


(3.85)

then it remains the characteristic function of a disc:

%(q, p, t) = Θ a2 − (q − q0 − t)2 − (p − p0 − αt)2 ,

(3.86)

For an example of a transition to ergodicity in a simple dynamical Hamiltonian model, see §3.9.
A stronger condition one could impose is the following. Let A and B be subsets of SE . Define the measure
Z
DA (E)
Z
ν(A) = dΣE χA (ϕ) dΣE = , (3.87)
D(E)

where χA (ϕ) is the characteristic function of A. The measure of a set A is the fraction of the energy
surface SE covered by A. This means ν(SE ) = 1, since SE is the entire phase space at energy E. Now
let g be a volume-preserving map on phase space. Given two measurable sets A and B, we say that a
system is mixing if
 
mixing ⇐⇒ lim ν g nA ∩ B = ν(A) ν(B) . (3.88)
n→∞

In other words, the fraction of B covered by the nth iterate of A, i.e. g nA, is, as n → ∞, simply the
fraction of SE covered by A. The iterated map g n distorts the region A so severely that it eventually
spreads out ‘evenly’ over the entire energy hypersurface. Of course by ‘evenly’ we mean ‘with respect to
any finite length scale’, because at the very smallest scales, the phase space density is still locally constant
as one evolves with the dynamics.
146 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 3.9: The multiply iterated baker’s transformation. The set A covers half the phase space and its
area is preserved under the map. Initially, the fraction of B covered by A is zero. After many iterations,
the fraction of B covered by g nA approaches 12 .

Mixing means that

Z


f (ϕ) = dµ %(ϕ, t) f (ϕ)
Z Z
(3.89)



−−−−→ dµ f (ϕ) δ E − H(ϕ) dµ δ E − H(ϕ)
t→∞
h
i. h
i
≡ Tr f (ϕ) δ E − H(ϕ) Tr δ E − H(ϕ) .

Physically, we can imagine regions of phase space being successively stretched and folded. During the
stretching process, the volume is preserved, so the successive stretch and fold operations map phase space
back onto itself.
An example of a mixing system is the baker’s transformation, depicted in fig. 3.8. The baker map is
defined by

1
if 0 ≤ q < 12


 2q , 2 p

g(q, p) = (3.90)

2q − 1 , 12 p + 12 if 12 ≤ q < 1 .


Note that g is invertible and volume-preserving. The baker’s transformation consists of an initial stretch
in which q is expanded by a factor of two and p is contracted by a factor of two, which preserves the total
volume. The system is then mapped back onto the original area by cutting and restacking, which we can
call a ‘fold’. The inverse transformation is accomplished by stretching first in the vertical (p) direction
3.5. REMARKS ON ERGODIC THEORY 147

Figure 3.10: The Arnold cat map applied to an image of 150 × 150 pixels. After 300 iterations, the image
repeats itself. (Source: Wikipedia)

and squashing in the horizontal (q) direction, followed by a slicing and restacking. Explicitly,

1
if 0 ≤ p < 12


 2 q , 2p

g −1 (q, p) = (3.91)

 1 1
if 12 ≤ p < 1 .


2 q + 2 , 2p − 1

Another example of a mixing system is Arnold’s ‘cat map’9

g(q, p) = [q + p] , [q + 2p] , (3.92)

where [x] denotes the fractional part of x. One can write this in matrix form as
M
 0  z }| {  
q 1 1 q
0 = mod Z2 . (3.93)
p 1 2 p

The matrix M is very special because it has integer entries and its determinant is det M = 1. This means
that the inverse also has integer entries. The inverse transformation is then

M −1
z }| {  
q0
  
q 2 −1
= mod Z2 . (3.94)
p −1 1 p0

Now for something cool. Suppose that our image consists of a set of discrete points located at (n1 /k , n2 /k),
where the denominator k ∈ Z is fixed, and where n1 and n2 range over the set {1, . . . , k}. Clearly g and
its inverse preserve this set, since the entries of M and M −1 are integers. If there are two possibilities
9
The cat map gets its name from its initial application, by Arnold, to the image of a cat’s face.
148 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 3.11: The hierarchy of dynamical systems.

2
for each pixel (say off and on, or black and white), then there are 2(k ) possible images, and the cat map
will map us invertibly from one image to another. Therefore it must exhibit Poincaré recurrence! This
phenomenon is demonstrated vividly in fig. 3.10, which shows a k = 150 pixel (square) image of a cat
subjected to the iterated cat map. The image is stretched and folded with each successive application of
the cat map, but after 300 iterations the image is restored! How can this be if the cat map is mixing? The
point is that only the discrete set of points (n1 /k , n2 /k) is periodic. Points with different denominators
will exhibit a different periodicity, and points with irrational coordinates will in general never return
to their exact initial conditions, although recurrence says they will come arbitrarily close, given enough
iterations. The baker’s transformation is also different in this respect, since the denominator of the p
coordinate is doubled upon each successive iteration.
The student should now contemplate the hierarchy of dynamical systems depicted in fig. 3.11, under-
standing the characteristic features of each successive refinement10 .

3.6 Thermalization of Quantum Systems

3.6.1 Quantum dephasing

Thermalization of quantum systems is fundamentally different from that of classical systems. Whereas
time evolution in classical mechanics is in general a nonlinear dynamical system, the Schrödinger equation
for time evolution in quantum mechanics is linear:
∂Ψ
i~ = ĤΨ , (3.95)
∂t
where Ĥ is a many-body Hamiltonian. In classical mechanics, the thermal state is constructed by time
evolution – this is the content of the ergodic theorem. In quantum mechanics, as we shall see, the thermal
distribution must be encoded in the eigenstates themselves.
10
There is something beyond mixing, called a K-system. A K-system has positive Kolmogorov-Sinai entropy. For such
a system, closed orbits separate exponentially in time, and consequently the Liouvillian L has a Lebesgue spectrum with
denumerably infinite multiplicity.
3.6. THERMALIZATION OF QUANTUM SYSTEMS 149

Let us assume an initial condition at t = 0,

X
|Ψ(0)i = Cα |Ψα i , (3.96)
α

where | Ψα i is an orthonormal
P eigenbasis for Ĥ satisfying Ĥ |Ψα i = Eα |Ψα i. The expansion coefficients
satisfy Cα = hΨα |Ψ(0)i and α |Cα |2 = 1. Normalization requires
X
h Ψ(0) | Ψ(0) i = |Cα |2 = 1 . (3.97)
α

The time evolution of |Ψi is then given by

X
|Ψ(t)i = Cα e−iEα t/~ |Ψα i . (3.98)
α

The energy is distributed according to the time-independent function

X
P (E) = h Ψ(t) | δ(E − Ĥ) | Ψ(t) i = |Cα |2 δ(E − Eα ) . (3.99)
α

Thus, the average energy is time-independent and is given by

Z∞ X
hEi = h Ψ(t) | Ĥ | Ψ(t) i = dE P (E) E = |Cα |2 Eα . (3.100)
−∞ α

The root mean square fluctuations of the energy are given by

s
D
2 E1/2 X X 2
(∆E)rms = E − hEi = |Cα |2 Eα2 − |Cα |2 Eα . (3.101)
α α

Typically we assume that the distribution P (E) is narrowly peaked about hEi, such that (∆E)rms 
E − E0 , where E0 is the ground state energy. Note that P (E) = 0 for E < E0 , i.e. the eigenspectrum of
Ĥ is bounded from below.
Now consider a general quantum observable described by an operator A. We have

Cα∗ Cβ ei(Eα −Eβ )t/~ Aαβ ,

X
hA(t)i = h Ψ(t) | A | Ψ(t) i = (3.102)
α,β

where Aαβ = hΨα |A|Ψβ i. In the limit of large times, we have

ZT
1 X
hAit ≡ lim dt hA(t)i = |Cα |2 Aαα . (3.103)
T →∞ T
0 α

Note that this implies that all coherence between different eigenstates is lost in the long time limit, due
to dephasing.
150 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

3.6.2 Eigenstate thermalization hypothesis

The essential ideas behind the eigenstate thermalization hypothesis (ETH) were described independently
by J. Deutsch (1991) and by M. Srednicki (1994). The argument goes as follows. If the total energy is
the only conserved quantity, and if A is a local, translationally-invariant, few-body operator, then the
time average hAi is given by its microcanonical value,
P
α Aαα Θ(Eα ∈ I)
X
2
hAit = |Cα | Aαα = P ≡ hAiE , (3.104)
α α Θ(Eα ∈ I)
 
where I = E, E + ∆E is an energy interval of width ∆E. So once again, time averages are micro
canonical averages.
But how is it that this is the case? The hypothesis of Deutsch and of Srednicki is that thermalization
in isolated and bounded quantum systems occurs at the level of individual eigenstates. That is, for all
eigenstates |Ψα i with Eα ∈ I, one has
Aαα = hAiEα . (3.105)
This means that thermal information is encoded in each eigenstate. This is called the eigenstate ther-
malization hypothesis (ETH).
An equivalent version of the ETH is the following scenario. Suppose we have an infinite or extremely
large quantum system U (the ‘universe’) fixed in an eigenstate |Ψα i. Then form the projection operator
Pα = |Ψα ihΨα |. Projection operators satisfy P 2 = P and their eigenspectrum consists of one eigenvalue
1 and the rest of the eigenvalues are zero11 . Now consider a partition of U = W ∪ S, where W  S. We
imagine S to be the ‘system’ and W the ‘world’. We can always decompose the state |Ψα i in a complete
product basis for W and S, viz.
NW NS
X X
|Ψα i = Qαpj |ψpW i ⊗ |ψjS i . (3.106)
p=1 j=1

Here NW/S is the size of the basis for W/S. The reduced density matrix for S is defined as

NS NW !
X X
ρS = Tr Pα = Qαpj Qα∗
pj 0 |ψjS ihψjS0 | . (3.107)
W
j,j 0 =1 p=1

The claim is that ρS approximates a thermal density matrix on S, i.e.

1 −β Ĥ
ρS ≈ e S , (3.108)
ZS

where ĤS is some Hamiltonian on S, and ZS = Tr e−β ĤS , so that Tr ρS = 1 and ρS is properly normalized.
A number of issues remain to be clarified:
11
More generally, we could project onto a K-dimensional subspace, in which case there would be K eigenvalues of +1 and
N − K eigenvalues of 0, where N is the dimension of the entire vector space.
3.6. THERMALIZATION OF QUANTUM SYSTEMS 151

(i) What do we mean by “approximates”?

(ii) What do we mean by ĤS ?

(iii) What do we mean by the temperature T ?

We address these in reverse order. The temperature T of an eigenstate |Ψα i of a Hamiltonian Ĥ is defined
by setting its energy density Eα /VU to the thermal energy density, i.e.

Eα 1 Tr Ĥ e−β Ĥ
= . (3.109)
V V Tr e−β Ĥ

Here, Ĥ = ĤU is the full Hamiltonian of the universe U = W ∪ S. Our intuition is that ĤS should reflect
a restriction of the original Hamiltonian ĤU to the system S. What should be done, though, about
the interface parts of ĤU which link S and W ? For lattice Hamiltonians, we can simply but somewhat
arbitrarily cut all the bonds coupling S and W . But we could easily imagine some other prescription, such
as halving the coupling strength along all such interface bonds. Indeed, the definition of HS is somewhat
arbitrary. However, so long as we use ρS to compute averages of local operators which lie sufficiently far
from the boundary of S, the precise details of how we truncate ĤU to ĤS are unimportant. This brings
us to the first issue: the approximation of ρS by its Gibbs form in eqn. 3.108 is only valid when we
consider averages of local operators lying within the bulk of S. This means that we must only examine
operators whose support is confined to regions greater than some distance ξT from ∂S, where ξT is a
thermal correlation length. This, in turn, requires that LS  ξT , i.e. the region S is very large on the
scale of ξT . How do we define ξT ? For a model such as the Ising model, it can be taken to be the usual
correlation length obtained from the spin-spin correlation function hσr σr0 iT . More generally, we may
choose the largest correlation length from among the correlators of all the independent local operators
in our system. Again, the requirement is that exp(−d∂ (r)/ξT )  1, where d∂ (r) is the shortest distance
from the location of our local operator Or to the boundary of S. At criticality, the exponential is replaced
by a power law (d∂ (r)/ξT )−p , where p is a critical exponent. Another implicit assumption here is that
VS  VW .

3.6.3 When is the ETH true?

There is no rigorous proof of the ETH. Deutsch showed that the ETH holds for the case of an integrable
Hamiltonian weakly perturbed by a single Gaussian random matrix. Horoi et al. (1995) showed that
nuclear shell model wavefunctions reproduce thermodynamic predictions. Recent numerical work by M.
Rigol and collaborators has verified the applicability of the ETH in small interacting boson systems.
ETH fails for so-called integrable models, where there are a large number of conserved quantities, which
commute with the Hamiltonian. Integrable models are, however, quite special, and as Deutsch showed,
integrability is spoiled by weak perturbations, in which case ETH then applies.
ETH also fails in the case of noninteracting disordered systems which exhibit Anderson localization.
Single particle energy eigenstates ψj whose energies εj the localized portion of the eigenspectrum decay
exponentially, as |ψj (r)|2 ∼ exp − |r − rj |/ξ(εj ) , where rj is some position in space associated with

ψj and ξ(εj ) is the localization length. Within the localized portion of the spectrum, ξ(ε) is finite.
152 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

As ε approaches a mobility edge, ξ(ε) diverges as a power law. In the delocalized regime, eigenstates
are spatially extended and typically decay at worst as a power law12 . Exponentially localized states
are unable to thermalize with other distantly removed localized states. Of course, all noninteracting
systems will violate ETH, because they are integrable. The interacting version of this phenomenon,
many-body localization (MBL), is a topic of intense current interest in condensed matter and statistical
physics. MBL systems also exhibit a large number of conserved quantities, but in contrast to the case of
integrable systems, where each conserved quantity is in general expressed in terms of an integral of a local
density, in MBL systems the conserved quantities are themselves local, although emergent. The emergent
nature of locally conserved quantities in MBL systems means that they are not simply expressed in terms
of the original local operators of the system, but rather are arrived at via a sequence of local unitary
transformations.
Note again that in contrast to the classical case, time evolution of a quantum state does not create the
thermal state. Rather, it reveals the thermal distribution which is encoded in all eigenstates after sufficient
time for dephasing to occur, so that correlations between all the wavefunction expansion coefficients {Cα }
for α 6= α0 are all lost.

3.7 Appendix I : Formal Solution of the Master Equation

Recall the master equation Ṗi = −Γij Pj . The matrix Γij is real but not necessarily symmetric. For such
a matrix, the left eigenvectors φαi and the right eigenvectors ψjβ are not the same: general different:

φαi Γij = λα φαj

(3.110)
Γij ψjβ = λβ ψiβ .

Note that the eigenvalue equation for the right eigenvectors is Γ ψ = λψ while that for the left eigenvectors
is Γ t φ = λφ. The characteristic polynomial is the same in both cases:

F (λ) ≡ det (λ − Γ ) = det (λ − Γ t ) , (3.111)

∗
which means that the left and right eigenvalues are the same. Note also that F (λ) = F (λ∗ ), hence the


eigenvalues are either real or appear in complex conjugate pairs. Multiplying the eigenvector equation
for φα on the right by ψjβ and summing over j, and multiplying the eigenvector equation for ψ β on the
left by φαi and summing over i, and subtracting the two results yields

λ α − λ β φα  ψ β = 0 ,


 
(3.112)

where the inner product is

  X
φψ = φi ψi . (3.113)
i

We can now demand

φα  ψ β = δαβ ,

 
(3.114)
12
Recall that in systems with no disorder, eigenstates exhibit Bloch periodicity in space.
3.8. APPENDIX II : RADIOACTIVE DECAY 153

in which case we can write

X X
λ α  ψ α φα  λα ψiα φαj .
 

Γ = ⇐⇒ Γij = (3.115)
α α

We have seen that φ ~ = (1, 1, . . . , 1) is a left eigenvector with eigenvalue λ = 0, since P Γ = 0. We do

i ij
not know a priori the corresponding right eigenvector, which depends on other details of Γij . Now let’s
expand Pi (t) in the right eigenvectors of Γ , writing
X
Pi (t) = Cα (t) ψiα . (3.116)
α
Then
dPi X dCα α
= ψ
dt α
dt i
X
= −Γij Pj = − Cα Γij ψjα (3.117)
α
X
=− λα Cα ψiα .
α
This allows us to write
dCα
= −λα Cα =⇒ Cα (t) = Cα (0) e−λα t . (3.118)
dt
Hence, we can write X
Pi (t) = Cα (0) e−λα t ψiα . (3.119)
α

It is now easy to see that Re (λα ) ≥ 0 for all λ, or else the probabilities will become negative. For suppose
Re (λα ) < 0 for some α. Then as t → ∞, the sum in eqn. 3.119 will be dominated by the term for
whichPλα has the largest αnegative real part; all other contributions will be subleading. But we must
α ~ α=0 = (1, 1, . . . , 1). Therefore,

have i ψi = 0 since ψ  must be orthogonal to the left eigenvector φ
α
at least one component of ψi (i.e. for some value of i) must have a negative real part, which means a
negative probability!13 As we have already proven that an initial nonnegative distribution {Pi (t = 0)}
will remain nonnegative under the evolution of the master equation, we conclude that Pi (t) → Pieq as
t → ∞, relaxing to the λ = 0 right eigenvector, with Re (λα ) ≥ 0 for all α.

3.8 Appendix II : Radioactive Decay

Consider a group of atoms, some of which are in an excited state which can undergo nuclear decay. Let
Pn (t) be the probability that n atoms are excited at some time t. We then model the decay dynamics by

0
 if m ≥ n
Wmn = nγ if m = n − 1 (3.120)

0 if m < n − 1 .


13
Since the probability Pi (t) is real, if the eigenvalue with the smallest (i.e. largest negative) real part is complex, there
will be a corresponding complex conjugate eigenvalue, and summing over all eigenvectors will result in a real value for Pi (t).
154 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Here, γ is the decay rate of an individual atom, which can be determined from quantum mechanics. The
master equation then tells us
dPn
= (n + 1) γ Pn+1 − n γ Pn . (3.121)
dt
 
The interpretation here is as follows: let  n denote a state in which n atoms are excited. Then Pn (t) =
h ψ(t) | n i2 . Then Pn (t) will increase due to spontaneous transitions from | n+1 i to | n i, and will
 

decrease due to spontaneous transitions from | n i to | n−1 i.

The average number of particles in the system is
∞
X
N (t) = n Pn (t) . (3.122)
n=0

Note that
∞
dN X h i
= n (n + 1) γ Pn+1 − n γ Pn
dt
n=0
∞ h
X i
=γ n(n − 1) Pn − n2 Pn (3.123)
n=0
X∞
= −γ n Pn = −γ N .
n=0

Thus,
N (t) = N (0) e−γt . (3.124)
The relaxation time is τ = γ −1 , and the equilibrium distribution is

Pneq = δn,0 . (3.125)

Note that this satisfies detailed balance.

We can go a bit farther here. Let us define
∞
X
P (z, t) ≡ z n Pn (t) . (3.126)
n=0

This is sometimes called a generating function. Then

∞
∂P X h i
=γ z n (n + 1) Pn+1 − n Pn
∂t
n=0 (3.127)
∂P ∂P
=γ − γz .
∂z ∂z
Thus,
1 ∂P ∂P
− (1 − z) =0. (3.128)
γ ∂t ∂z
3.9. APPENDIX III: TRANSITION TO ERGODICITY IN A SIMPLE MODEL 155

We now see that any function f (ξ) satisfies the above equation, where ξ = γt − ln(1 − z). Thus, we can
write

P (z, t) = f γt − ln(1 − z) . (3.129)
Setting t = 0 we have P (z, 0) = f −ln(1 − z) , and inverting this result we obtain f (u) = P (1 − e−u , 0),

i.e.
P (z, t) = P 1 + (z − 1) e−γt , 0 .


(3.130)
P∞
The total probability is P (z = 1, t) = n=0 Pn , which clearly is conserved: P (1, t) = P (1, 0). The average
particle number is
∞ 
X ∂P 
N (t) = n Pn (t) = = e−γt P (1, 0) = N (0) e−γt . (3.131)
∂z 
z=1
n=0

3.9 Appendix III: Transition to Ergodicity in a Simple Model

A ball of mass m executes perfect one-dimensional motion along the symmetry axis of a piston. Above
the ball lies a mobile piston head of mass M which slides frictionlessly inside the piston. Both the ball
and piston head execute ballistic motion, with two types of collision possible: (i) the ball may bounce off
the floor, which is assumed to be infinitely massive and fixed in space, and (ii) the ball and piston head
may engage in a one-dimensional elastic collision. The Hamiltonian is

P2 p2
H= + + M gX + mgx ,
2M 2m
where X is the height of the piston head and x the height of the ball. Another quantity is conserved by
the dynamics: Θ(X − x). I.e., the ball always is below the piston head.

(a) Choose
√ scale L, and then energy scale E0 = M gL, momentum scale P0 =
an arbitrary length p
M gL, and time scale τ0 = L/g. Show that the dimensionless Hamiltonian becomes

p̄2
H̄ = 21 P̄ 2 + X̄ + + rx̄ ,
2r
with r = m/M , and with equations of motion dX/dt = ∂ H̄/∂ P̄ , etc. (Here the bar indicates
dimensionless variables: P̄ = P/P0 , t̄ = t/τ0 , etc.) What special dynamical consequences hold for
r = 1?

(b) Compute the microcanonical average piston height hXi. The analogous dynamical average is

ZT
1
hXit = lim dt X(t) .
T →∞ T
0

When computing microcanonical averages, it is helpful to use the Laplace transform, discussed
toward the end of §3.3 of the notes. (It is possible to compute the microcanonical average by more
brute force methods as well.)
156 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

(c) Compute the microcanonical average of the rate of collisions between the ball and the floor. Show
that this is given by

X
δ(t − ti ) = Θ(v) v δ(x − 0+ ) .



i
The analogous dynamical average is

ZT X
1
hγit = lim dt δ(t − ti ) ,
T →∞ T
0 i

where {ti } is the set of times at which the ball hits the floor.

(d) How do your results change if you do not enforce the dynamical constraint X ≥ x?

(e) Write a computer program to simulate this system. The only input should be the mass ratio r
(set Ē = 10 to fix the energy). You also may wish to input the initial conditions, or perhaps to
choose the initial conditions randomly (all satisfying energy conservation, of course!). Have your
program compute the microcanonical as well as dynamical averages in parts (b) and (c). Plot out
the Poincaré section of P vs. X for those times when the ball hits the floor. Investigate this for
several values of r. Just to show you that this is interesting, I’ve plotted some of my own numerical
results in fig. 3.12.

Solution:
√ √
we choose a length scale L (arbitrary), we may define E0 = M gL, P0 = M gL, V0 = gL,
(a) Once p
and τ0 = L/g as energy, momentum, velocity, and time scales, respectively, the result follows directly.
Rather than write P̄ = P/P0 etc., we will drop the bar notation and write

p2
H = 21 P 2 + X + + rx .
2r

(b) What is missing from the Hamiltonian of course is the interaction potential between the ball and
the piston head. We assume that both objects are impenetrable, so the potential energy is infinite when
the two overlap. We further assume that the ball is a point particle (otherwise reset ground level to
minus the diameter of the ball). We can eliminate the interaction potential from H if we enforce that
each time X = x the ball and the piston head undergo an elastic collision. From energy and momentum
conservation, it is easy to derive the elastic collision formulae

1−r 2
P0 = P+ p
1+r 1+r

2r 1−r
p0 = P− p.
1+r 1+r

We can now answer the last question from part (a). When r = 1, we have that P 0 = p and p0 = P ,
i.e. the ball and piston simply exchange momenta. The problem is then equivalent to two identical
3.9. APPENDIX III: TRANSITION TO ERGODICITY IN A SIMPLE MODEL 157

Figure 3.12: Poincaré sections for the ball and piston head problem. Each color corresponds to a different
initial condition. When the mass ratio r = m/M exceeds unity, the system apparently becomes ergodic.

particles elastically bouncing off the bottom of the piston, and moving through each other as if they were
completely transparent. When the trajectories cross, however, the particles exchange identities.
Averages within the microcanonical ensemble are normally performed with respect to the phase space
distribution

δ E − H(ϕ)
%(ϕ) =
,
Tr δ E − H(ϕ)
where ϕ = (P, X, p, x), and

Z∞ Z∞ Z∞ Z∞
Tr F (ϕ) = dP dX dp dx F (P, X, p, x) .
−∞ 0 −∞ 0

Since X ≥ x is a dynamical constraint, we should define an appropriately restricted microcanonical

158 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

average: 

 h
i

F (ϕ) µce ≡ Tr F (ϕ) δ E − H(ϕ)
e e δ E − H(ϕ)
Tr

where
Z∞ Z∞ Z∞ ZX
e F (ϕ) ≡
Tr dP dX dp dx F (P, X, p, x)
−∞ 0 −∞ 0
is the modified trace. Note that the integral over x has an upper limit of X rather than ∞, since the
region of phase space with x > X is dynamically inaccessible.
When computing the traces, we shall make use of the following result from the theory of Laplace trans-
forms. The Laplace transform of a function K(E) is
Z∞
K(β)
b = dE K(E) e−βE .
0

The inverse Laplace transform is given by

c+i∞
dβ b
Z
K(E) = K(β) eβE ,
2πi
c−i∞

where the integration contour, which is a line extending from β = c − i∞ to β = c + i∞, lies to the right
of any singularities of K(β)
b in the complex β-plane. For this problem, all we shall need is the following:
E t−1
K(E) = ⇐⇒ K(β)
b = β −t .
Γ(t)
For a proof, see §4.2.2 of the lecture notes.
We’re now ready to compute the microcanonical average of X. We have
N (E)
hXi = ,
D(E)
where
 
N (E) = Tr
e X δ(E − H)
D(E) = Tr
e δ(E − H) .

Let’s first compute D(E). To do this, we compute the Laplace transform D(β):
b

D(β)
b e e−βH
= Tr
Z∞ Z∞ Z∞ ZX
−βP 2 /2 −βp2 /2r −βX
= dP e dp e dX e dx e−βrx
−∞ −∞ 0 0
√ Z∞ √
1 − e−βrX
 
2π r −βX r 2π
= dX e = · .
β βr 1 + r β3
0
3.9. APPENDIX III: TRANSITION TO ERGODICITY IN A SIMPLE MODEL 159

Similarly for N
b (β) we have

N e X e−βH
b (β) = Tr
Z∞ Z∞ Z∞ ZX
−βP 2 /2 −βp2 /2r −βX
= dP e dp e dX X e dx e−βrx
−∞ −∞ 0 0
√ Z∞
1 − e−βrX (2 + r) r3/2 2π
 
2π r −βX
= dX X e = · 4 .
β βr (1 + r)2 β
0

Taking the inverse Laplace transform, we then have

√ √
r (2 + r) r 1
D(E) = · πE 2 , N (E) = · 3 πE 3 .
1+r (1 + r)2

We then have
 
N (E) 2+r
hXi = = · 13 E .
D(E) 1+r

The ‘brute force’ evaluation of the integrals isn’t so bad either. We have

Z∞ Z∞ Z∞ ZX
1 2 1 2


D(E) = dP dX dp dx δ 2P + 2r p + X + rx − E .
−∞ 0 −∞ 0

√ √ √
To evaluate, define P = 2 ux and p = 2r uy . Then we have dP dp = 2 r dux duy and 21 P 2 + 1
2r p2 =
u2x + u2y . Now convert to 2D polar coordinates with w ≡ u2x + u2y . Thus,

Z∞ Z∞ ZX
√

D(E) = 2π r dw dX dx δ w + X + rx − E
0 0 0
Z∞ Z∞ ZX
2π
=√ dw dX dx Θ(E − w − X) Θ(X + rX − E + w)
r
0 0 0
ZE E−w √ ZE √
2π r r
Z
2π
=√ dw dX = dq q = · πE 2 ,
r 1+r 1+r
0 E−w 0
1+r
160 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Rt
Figure 3.13: Long time running numerical averages Xav (t) ≡ t−1 0 dt0 X(t0 ) for r = 0.3 (top) and r = 1.2
(bottom), each for three different initial conditions, with E = 10 in all cases. Note how in the r = 0.3 case
the long time average is dependent on the initial condition, while the r = 1.2 case is ergodic and hence
independent of initial conditions. The dashed black line shows the restricted microcanonical average,
(2+r) 1
hXiµce = (1+r) · 3 E.

with q = E − w. Similarly,

Z∞ Z∞ ZX
√

N (E) = 2π r dw dX X dx δ w + X + rx − E
0 0 0
Z∞ Z∞ ZX
2π
=√ dw dX X dx Θ(E − w − X) Θ(X + rX − E + w)
r
0 0 0
ZE E−w ZE     √
2+r r 1
Z
2π 2π 1 1 2
=√ dw dX X = √ dq 1 − · 2 q = · · πE 3 .
r r (1 + r)2 1+r 1+r 3
0 E−w 0
1+r

(c) Using the general result

X δ(x − xi )
δ F (x) − A = F 0 (x ) ,
 
i i
3.9. APPENDIX III: TRANSITION TO ERGODICITY IN A SIMPLE MODEL 161

where F (xi ) = A, we recover the desired expression. We should be careful not to double count, so to
avoid this difficulty we can evaluate δ(t − t+ + +
i ), where ti = ti + 0 is infinitesimally later than ti . The
+
point here is that when t = ti we have p = r v > 0 (i.e. just after hitting the bottom). Similarly, at
times t = t−i we have p < 0 (i.e. just prior to hitting the bottom). Note v = p/r. Again we write
γ(E) = N (E)/D(E), this time with

e Θ(p) r−1 p δ(x − 0+ ) δ(E − H) .

 
N (E) = Tr

The Laplace transform is

Z∞ Z∞ Z∞
−βP 2 /2 −1 −βp 2 /2r
N
b (β) = dP e dp r p e dX e−βX
−∞ 0 0
√
r
2π 1 1
= · · = 2π β −5/2 .
β β β

Thus,
√
N (E) = 4 2
3 E 3/2

and
√
 
N (E) 1+r
hγi = = 4 2
3π
√ E −1/2 .
D(E) r

r X(0) hX(t)i hXiµce hγ(t)i hγiµce r X(0) hX(t)i hXiµce hγ(t)i hγiµce
0.3 0.1 6.1743 5.8974 0.5283 0.4505 1.2 0.1 4.8509 4.8545 0.3816 0.3812
0.3 1.0 5.7303 5.8974 0.4170 0.4505 1.2 1.0 4.8479 4.8545 0.3811 0.3812
0.3 3.0 5.7876 5.8974 0.4217 0.4505 1.2 3.0 4.8493 4.8545 0.3813 0.3812
0.3 5.0 5.8231 5.8974 0.4228 0.4505 1.2 5.0 4.8482 4.8545 0.3813 0.3812
0.3 7.0 5.8227 5.8974 0.4228 0.4505 1.2 7.0 4.8472 4.8545 0.3808 0.3812
0.3 9.0 5.8016 5.8974 0.4234 0.4505 1.2 9.0 4.8466 4.8545 0.3808 0.3812
0.3 9.9 6.1539 5.8974 0.5249 0.4505 1.2 9.9 4.8444 4.8545 0.3807 0.3812

Table 3.1: Comparison of time averages and microcanonical ensemble averages for r = 0.3 and r = 0.9.
Initial conditions are P (0) = x(0) = 0, with X(0) given in the table and E = 10. Averages were performed
over a period extending for Nb = 107 bounces.

(d) When the constraint X ≥ x is removed, we integrate over all phase space. We then have

D(β)
b = Tr e−βH
Z∞ Z∞ Z∞ Z∞ √
−βP 2 /2 −βp2 /2r −βX −βrx 2π r
= dP e dp e dX e dx e = .
β3
−∞ −∞ 0 0
162 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

For part (b) we would then have

b (β) = Tr X e−βH
N
Z∞ Z∞ Z∞ Z∞ √
2
−βP /2 2
−βp /2r −βX −βrx 2π r
= dP e dp e dX X e dx e = .
β4
−∞ −∞ 0 0
√ √
The respective inverse Laplace transforms are D(E) = π rE 2 and N (E) = 31 π rE 3 . The microcanonical
average of X would then be
hXi = 31 E .
Using the restricted phase space, we obtained a value which is greater than this by a factor of (2+r)/(1+r).
That the restricted average gives a larger value makes good sense, since X is not allowed to descend below
x in that case. For part (c), we would obtain the same result for N (E) since x = 0 in the average. We
would then obtain √
hγi = 43π2 r−1/2 E −1/2 .
The restricted microcanonical average yields a rate which is larger by a factor 1 + r. Again, it makes
good sense that the restricted average should yield a higher rate, since the ball is not allowed to attain a
height greater than the instantaneous value of X.
(e) It is straightforward to simulate the dynamics. So long as 0 < x(t) < X(t), we have
p
Ẋ = P , Ṗ = −1 , ẋ = , ṗ = −r .
r
Starting at an arbitrary time t0 , these equations are integrated to yield

X(t) = X(t0 ) + P (t0 ) (t − t0 ) − 21 (t − t0 )2

P (t) = P (t0 ) − (t − t0 )
p(t0 )
x(t) = x(t0 ) + (t − t0 ) − 21 (t − t0 )2
r
p(t) = p(t0 ) − r(t − t0 ) .

We must stop the evolution when one of two things happens. The first possibility is a bounce at t = tb ,
−
meaning x(tb ) = 0. The momentum p(t) changes discontinuously at the bounce, with p(t+ b ) = −p(tb ),
−
and where p(tb ) < 0 necessarily. The second possibility is a collision at t = tc , meaning X(tc ) = x(tc ).
Integrating across the collision, we must conserve both energy and momentum. This means

1−r 2
P (t+
c )= P (t−
c )+ p(t−
c )
1+r 1+r

2r 1−r
p(t+
c )= P (t−
c )− p(t−
c ) .
1+r 1+r

In the following tables I report on the results of numerical simulations, comparing dynamical averages with
(restricted) phase space averages within the microcanonical ensemble. For r = 0.3 the microcanonical
3.9. APPENDIX III: TRANSITION TO ERGODICITY IN A SIMPLE MODEL 163

r X(0) Nb hX(t)i hXiµce hγ(t)i hγiµce

1.2 7.0 104 4.8054892 4.8484848 0.37560388 0.38118510
1.2 7.0 105 4.8436969 4.8484848 0.38120356 0.38118510
1.2 7.0 106 4.8479414 4.8484848 0.38122778 0.38118510
1.2 7.0 107 4.8471686 4.8484848 0.38083749 0.38118510
1.2 7.0 108 4.8485825 4.8484848 0.38116282 0.38118510
1.2 7.0 109 4.8486682 4.8484848 0.38120259 0.38118510
1.2 1.0 109 4.8485381 4.8484848 0.38118069 0.38118510
1.2 9.9 109 4.8484886 4.8484848 0.38116295 0.38118510

Table 3.2: Comparison of time averages and microcanonical ensemble averages for r = 1.2, with Nb
ranging from 104 to 109 .

averages poorly approximate the dynamical averages, and the dynamical averages are dependent on
the initial conditions, indicating that the system is not ergodic. For r = 1.2, the agreement between
dynamical and microcanonical averages generally improves with averaging time. Indeed, it has been
shown by N. I. Chernov, Physica D 53, 233 (1991), building on the work of M. P. Wojtkowski, Comm.
Math. Phys. 126, 507 (1990) that this system is ergodic for r > 1. Wojtkowski also showed that this
system is equivalent to the wedge billiard particle of mass m bounces −1
 , in which a single point inside
p ma
two-dimensional wedge-shaped region (x, y) x ≥ 0 , y ≥ x ctn φ for some fixed angle φ = tan

M.
To see this, pass to relative (X ) and center-of-mass (Y) coordinates,

mP − M p
X =X −x Px =
M +m
M X + mx
Y= Py = P + p .
M +m

Then
(M + m) Px2 Py2
H= + + (M + m) gY .
2M m 2(M + m)
There are two constraints. One requires X ≥ x, i.e. X ≥ 0. The second requires x > 0, i.e.
M
x=Y− X ≥0.
M +m
√
M +m Mm
Now define x ≡ X , px ≡ Px , and rescale y ≡ √
Mm
Y and py ≡ M +m Py to obtain

1 2
px + p2y + M g y


H=
2µ

Mm
√ q
M
with µ = M +m the familiar reduced mass and M = M m. The constraints are then x ≥ 0 and y ≥ m x.
164 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
Chapter 4

Statistical Ensembles

4.1 References

– F. Reif, Fundamentals of Statistical and Thermal Physics (McGraw-Hill, 1987)

This has been perhaps the most popular undergraduate text since it first appeared in 1967, and
with good reason.

– A. H. Carter, Classical and Statistical Thermodynamics

(Benjamin Cummings, 2000)
A very relaxed treatment appropriate for undergraduate physics majors.

– D. V. Schroeder, An Introduction to Thermal Physics (Addison-Wesley, 2000)

This is the best undergraduate thermodynamics book I’ve come across, but only 40% of the book
treats statistical mechanics.

– C. Kittel, Elementary Statistical Physics (Dover, 2004)

Remarkably crisp, though dated, this text is organized as a series of brief discussions of key concepts
and examples. Published by Dover, so you can’t beat the price.

– M. Kardar, Statistical Physics of Particles (Cambridge, 2007)

A superb modern text, with many insightful presentations of key concepts.

– M. Plischke and B. Bergersen, Equilibrium Statistical Physics (3rd edition, World Scientific, 2006)
An excellent graduate level text. Less insightful than Kardar but still a good modern treatment of
the subject. Good discussion of mean field theory.

– E. M. Lifshitz and L. P. Pitaevskii, Statistical Physics (part I, 3rd edition, Pergamon, 1980)
This is volume 5 in the famous Landau and Lifshitz Course of Theoretical Physics. Though dated,
it still contains a wealth of information and physical insight.

165
166 CHAPTER 4. STATISTICAL ENSEMBLES

4.2 Microcanonical Ensemble (µCE)

4.2.1 The microcanonical distribution function

We have seen how in an ergodic dynamical system, time averages can be replaced by phase space averages:



ergodicity ⇐⇒ f (ϕ) t = f (ϕ) S , (4.1)
where
ZT

 1

f (ϕ) t = lim dt f ϕ(t) . (4.2)
T →∞ T
0
and Z Z




f (ϕ) S = dµ f (ϕ) δ E − Ĥ(ϕ) dµ δ E − Ĥ(ϕ) . (4.3)

Here Ĥ(ϕ) = Ĥ(q, p) is the Hamiltonian, and where δ(x) is the Dirac δ-function1 . Thus, averages are
taken over a constant energy hypersurface which is a subset of the entire phase space.
We’ve also seen how any phase space distribution %(Λ1 , . . . , Λk ) which is a function of conserved quantitied
Λa (ϕ) is automatically a stationary (time-independent) solution to Liouville’s equation. Note that the
microcanonical distribution,
Z



%E (ϕ) = δ E − Ĥ(ϕ) dµ δ E − Ĥ(ϕ) , (4.4)

is of this form, since Ĥ(ϕ) is conserved by the dynamics. Linear and angular momentum conservation
generally are broken by elastic scattering off the walls of the sample.
So averages in the microcanonical ensemble are computed by evaluating the ratio

 Tr A δ(E − Ĥ)
A = , (4.5)
Tr δ(E − Ĥ)
where Tr means ‘trace’, which entails an integration over all phase space:
N Z
1 Y ddpi ddqi
Tr A(q, p) ≡ A(q, p) . (4.6)
N! (2π~)d
i=1
Here N is the total number of particles and d is the dimension of physical space in which each particle
moves. The factor of 1/N !, which cancels in the ratio between numerator and denominator, is present for
indistinguishable particles 2 . The normalization factor (2π~)−N d renders the trace dimensionless. Again,
this cancels between numerator and denominator. These factors may then seem arbitrary in the definition
of the trace, but we’ll see how they in fact are required from quantum mechanical considerations. So we
now adopt the following metric for classical phase space integration:
N
1 Y ddpi ddqi
dµ = . (4.7)
N! (2π~)d
i=1
1
We write the Hamiltonian as Ĥ (classical or quantum) in order to distinguish it from magnetic field (H) or enthalpy
(H).
2
More on this in chapter 5.
4.2. MICROCANONICAL ENSEMBLE (µCE) 167

4.2.2 Density of states

The denominator,
D(E) = Tr δ(E − Ĥ) , (4.8)
is called the density of states. It has dimensions of inverse energy, such that
E+∆E
Z Z Z
0 0
D(E) ∆E = dE dµ δ(E − Ĥ) = dµ (4.9)
E E<Ĥ<E+∆E
= # of states with energies between E and E + ∆E .

Let us now compute D(E) for the nonrelativistic ideal gas. The Hamiltonian is
N
X p2i
Ĥ(q, p) = . (4.10)
2m
i=1

We assume that the gas is enclosed in a region of volume V , and we’ll do a purely classical calculation,
neglecting discreteness of its quantum spectrum. We must compute
N N
ddpi ddqi p2i
Z Y  
1 X
D(E) = δ E− . (4.11)
N! (2π~)d 2m
i=1 i=1

We shall calculate D(E) in two ways. The first method utilizes the Laplace transform, Z(β):
Z∞
dE e−βE D(E) = Tr e−β Ĥ .
 
Z(β) = L D(E) ≡ (4.12)
0

The inverse Laplace transform is then

c+i∞
dβ βE
Z
−1
 
D(E) = L Z(β) ≡ e Z(β) , (4.13)
2πi
c−i∞

where c is such that the integration contour is to the right of any singularities of Z(β) in the complex
β-plane. We then have
N Z
1 Y ddxi ddpi −βp2 /2m
Z(β) = e i
N! (2π~)d
i=1
 ∞ N d
VN  dp −βp2 /2m 
Z
= e (4.14)
N! 2π~
−∞
N
 N d/2
V m
= β −N d/2 .
N! 2π~2
168 CHAPTER 4. STATISTICAL ENSEMBLES

Figure 4.1: Complex integration contours C for inverse Laplace transform L−1 Z(β) = D(E). When
 

the product dN is odd, there is a branch cut along the negative Re β axis.

The inverse Laplace transform is then

N d/2 I
VN

m dβ βE −N d/2
D(E) = e β
N! 2π~2 2πi
C
1
(4.15)
N d/2 N d−1
N E 2

V m
= ,
N! 2π~2 Γ(N d/2)

exactly as before. The integration contour for the inverse Laplace transform is extended in an infinite
semicircle in the left half β-plane. When N d is even, the function β −N d/2 has a simple pole of order
N d/2 at the origin. When N d is odd, there is a branch cut extending along the negative Re β axis, and
the integration contour must avoid the cut, as shown in fig. 4.1. One can check that this results in the
same expression above, i.e. we may analytically continue from even values of N d to all positive values of
N d.
 
For a general system, the Laplace transform, Z(β) = L D(E) also is called the partition function. We
shall again meet up with Z(β) when we discuss the ordinary canonical ensemble.
Our final result, then, is
1
N d/2 N d−1
VN E2

m
D(E, V, N ) = . (4.16)
N! 2π~2 Γ(N d/2)
Here we have emphasized that the density of states is a function of E, V , and N . Using Stirling’s
approximation,
ln N ! = N ln N − N + 21 ln N + 12 ln(2π) + O N −1 ,


(4.17)
we may define the statistical entropy,
 
E V
S(E, V, N ) ≡ kB ln D(E, V, N ) = N kB φ , + O(ln N ) , (4.18)
N N
4.2. MICROCANONICAL ENSEMBLE (µCE) 169

where        
E V d E V d m 1


φ , = ln + ln + ln + 1 + 2 d . (4.19)
N N 2 N N 2 dπ~2
Recall kB = 1.3806503 × 10−16 erg/K is Boltzmann’s constant.

Second method
√
The second method invokes a mathematical trick. First, let’s rescale pαi ≡ 2mE uαi . We then have
√ !N d
VN
Z
2mE 1
dMu δ u21 + u22 + . . . + u2M − 1 .


D(E) = (4.20)
N! h E

Here we have written u = (u1 , u2 , . . . , uM ) with M = N d as a M -dimensional vector. We’ve also used
the rule δ(Ex) = E −1 δ(x) for δ-functions. We can now write

dMu = uM −1 du dΩM , (4.21)

where dΩM is the M -dimensional differential solid angle. We now have our answer:3
√ !N d
VN 2m 1
N d−1 1
D(E) = E2 · 2 ΩN d . (4.22)
N! h

What remains is for us to compute ΩM , the total solid angle in M dimensions. We do this by a nifty
mathematical trick. Consider the integral
Z Z∞
−u2 2
IM = dMu e = ΩM du uM −1 e−u
0
(4.23)
Z∞ 1
1 M −1 −s
e = 21 ΩM Γ 1


= 2 ΩM ds s 2 2M ,
0

where s = u2 , and where

Z∞
Γ(z) = dt tz−1 e−t (4.24)
0

is the Gamma function, which satisfies z Γ(z) = Γ(z + 1).4 On the other hand, we can compute IM in
Cartesian coordinates, writing
 ∞ M
√
M
Z
2
IM =  du1 e−u1  = π . (4.25)
−∞
3
The factor of 12 preceding ΩM in eqn. 4.22 appears because δ(u2 − 1) = 1
2
δ(u − 1) + 1
2
δ(u + 1). Since u = |u| ≥ 0, the
second term can be dropped.
4
Note that for integer argument, Γ(k) = (k − 1)!
170 CHAPTER 4. STATISTICAL ENSEMBLES

Therefore
2π M/2
ΩM = . (4.26)
Γ(M/2)
We thereby obtain Ω2 = 2π, Ω3 = 4π, Ω4 = 2π 2 , etc., the first two of which are familiar.

4.2.3 Arbitrariness in the definition of S(E)

Note that D(E) has dimensions of inverse energy, so one might ask how we are to take the logarithm
of a dimensionful quantity in eqn. 4.18. We must introduce an energy scale, such as ∆E in eqn. 4.9,
and define D̃(E; ∆E) = D(E) ∆E and S(E; ∆E) ≡ kB ln D̃(E; ∆E). The definition of statistical entropy
then involves the arbitrary parameter ∆E, however this only affects S(E) in an additive way. That is,
 
∆E1
S(E, V, N ; ∆E1 ) = S(E, V, N ; ∆E2 ) + kB ln . (4.27)
∆E2

Note that the difference between the two definitions of S depends only on the ratio ∆E1 /∆E2 , and is
independent of E, V , and N .

4.2.4 Ultra-relativistic ideal gas

Consider an ultrarelativistic ideal gas, with single particle dispersion ε(p) = cp. We then have
N
Z∞

VN ΩdN  dp pd−1 e−βcp 
Z(β) =
N! hN d
0 (4.28)
N
 N
V Γ(d) Ωd
= .
N! cd hd β d
E V


The statistical entropy is S(E, V, N ) = kB ln D(E, V, N ) = N kB φ N , N , with
       
E V E V Ωd Γ(d)
φ , = d ln + ln + ln + (d + 1) (4.29)
N N N N (dhc)d

4.2.5 Discrete systems

For classical systems where the energy levels are discrete, the states of the system | σ i are labeled by a
set of discrete quantities {σ1 , σ2 , . . .}, where each variable σi takes discrete values. The number of ways
of configuring the system at fixed energy E is then
X
Ω(E, N ) = δĤ(σ),E , (4.30)
σ

where the sum is over all possible configurations. Here N labels the total number of particles. For
example, if we have N spin- 21 particles on a lattice which are placed in a magnetic field H, so the
4.3. THE QUANTUM MECHANICAL TRACE 171

individual particle energy is εi = −µ0 Hσ, where σ = ±1, then in a configuration in which N↑ particles
have σi = +1 and N↓ = N − N↑ particles have σi = −1, the energy is E = (N↓ − N↑ )µ0 H. The number
of configurations at fixed energy E is
 
N N!
Ω(E, N ) = = N E

N E

, (4.31)
N↑ 2 − 2µ H ! 2 + 2µ H ! 0 0

N E
since N↑/↓ = 2 ∓ 2µ0 H . The statistical entropy is S(E, N ) = kB ln Ω(E, N ).

4.3 The Quantum Mechanical Trace

Thus far our understanding of ergodicity is rooted in the dynamics of classical mechanics. A Hamiltonian
flow which is ergodic is one in which time averages can be replaced by phase space averages using the
microcanonical ensemble. What happens, though, if our system is quantum mechanical, as all systems
ultimately are?

4.3.1 The density matrix

First, let us consider that our system S willin general

 be in contact with a world W . We call the union
 
of S and W the universe, U = W ∪ S. Let  N denote a quantum mechanical state of W , and let  n
denote a quantum mechanical state of S. Then the most general wavefunction we can write is of the form
  X    
Ψ = ΨN,n  N ⊗  n . (4.32)
N,n

Now let us   0 the expectation value of some operator  which acts as the identity within W ,

compute
meaning N    N =  δN N 0 , where  is the ‘reduced’ operator which acts within S alone. We then
have
XX
Ψ∗N,n ΨN 0 ,n0 δN N 0 n  Â  n0

  
  
Ψ  Â  Ψ =
N,N 0 n,n0 (4.33)


= Tr %̂ Â ,
where XX
Ψ∗N,n ΨN,n0  n0 n 
 

%̂ = (4.34)
N n,n0

is the density matrix . The time-dependence of %̂ is easily found:

XX
Ψ∗N,n ΨN,n0  n0 (t) n(t) 
 

%̂(t) =
N n,n0 (4.35)
= e−iĤt/~ %̂ e+iĤt/~ ,

where Ĥ is the Hamiltonian for the system S. Thus, we find

∂ %̂  
i~ = Ĥ, %̂ . (4.36)
∂t
172 CHAPTER 4. STATISTICAL ENSEMBLES

Figure 4.2: A system S in contact with a ‘world’ W . The union of the two, universe U = W ∪ S, is said
to be the ‘universe’.

Note that the density matrix evolves according to a slightly different equation than an operator in the
Heisenberg picture, for which

∂ Â 
Â(t) = e+iHt/~ A e−iĤt/~
  
=⇒ i~ = Â, Ĥ = − Ĥ, Â . (4.37)
∂t

For Hamiltonian systems, we found that the phase space distribution %(q, p, t) evolved according to the
Liouville equation,
∂%
i = L% , (4.38)
∂t
where the Liouvillian L is the differential operator
Nd
( )
X ∂ Ĥ ∂ ∂ Ĥ ∂
L = −i − . (4.39)
∂pj ∂qj ∂qj ∂pj
j=1

Accordingly, any distribution %(Λ1 , . . . , Λk ) which is a function of constants of the motion Λa (q, p) is
a stationary solution to the Liouville equation: ∂t %(Λ1 , . . . , Λk ) = 0. Similarly, any quantum mechan-
ical density matrix which commutes with the Hamiltonian is a stationary solution to eqn. 4.36. The
corresponding microcanonical distribution is


%̂E = δ E − Ĥ . (4.40)

4.3.2 Averaging the DOS

If our quantum mechanical system is placed in a finite volume, the energy levels will be discrete, rather
than continuous, and the density of states (DOS) will be of the form

X
D(E) = Tr δ E − Ĥ = δ(E − El ) , (4.41)
l
4.3. THE QUANTUM MECHANICAL TRACE 173

Figure 4.3: Averaging the quantum mechanical discrete density of states yields a continuous curve.

where {El } are the eigenvalues of the Hamiltonian Ĥ. In the thermodynamic limit, V → ∞, and the
discrete spectrum of kinetic energies remains discrete for all finite V but must approach the continuum
result. To recover the continuum result, we average the DOS over a window of width ∆E:
E+∆E
Z
1
D(E) = dE 0 D(E 0 ) . (4.42)
∆E
E

If we take the limit ∆E → 0 but with ∆E  δE, where δE is the spacing between successive quantized
levels, we recover a smooth function, as shown in fig. 4.3. We will in general drop the bar and refer to
this function as D(E). Note that δE ∼ 1/D(E) = e−N φ(ε,v) is (typically) exponentially small in the size
of the system, hence if we took ∆E ∝ V −1 which vanishes in the thermodynamic limit, there are still
exponentially many energy levels within an interval of width ∆E.

4.3.3 Coherent states

The quantum-classical correspondence is elucidated with the use of coherent states. Recall that the
one-dimensional harmonic oscillator Hamiltonian may be written
p2
Ĥ0 = + 1 m ω02 q 2
2m 2 (4.43)
= ~ω0 a† a + 21 ,

where a and a† are ladder operators satisfying a, a† = 1, which can be taken to be

 

∂ q ∂ q
a=` + , a† = −` + , (4.44)
∂q 2` ∂q 2`
p
with ` = ~/2mω0 . Note that
~
q = ` a + a† a − a† .



, p= (4.45)
2i`
174 CHAPTER 4. STATISTICAL ENSEMBLES

The ground state satisfies a ψ0 (q) = 0, which yields

2 /4`2
ψ0 (q) = (2π`2 )−1/4 e−q . (4.46)

The normalized coherent state | z i is defined as

∞
1 2 † 1 2 zn
| z i = e− 2 |z| eza | 0 i = e− 2 |z|
X
√ |ni . (4.47)
n=0 n!

The overlap of coherent states is given by

1 2 1 2
h z1 | z2 i = e− 2 |z1 | e− 2 |z2 | ez̄1 z2 , (4.48)

hence different coherent states are not orthogonal. Despite this nonorthogonality, the coherent states
allow a simple resolution of the identity,
Z 2
dz d2z d Rez d Imz
1= | z ih z | ; ≡ (4.49)
2πi 2πi π
which is straightforward to establish.
To gain some physical intuition about the coherent states, define
Q i`P
z≡ + (4.50)
2` ~
and write | z i ≡ | Q, P i. One finds (exercise!)
2 /4`2
ψQ,P (q) = h q | z i = (2π`2 )−1/4 e−iP Q/2~ eiP q/~ e−(q−Q) , (4.51)

hence the coherent state ψQ,P (q) is a wavepacket Gaussianly localized about q = Q, but oscillating with
average momentum P .
For example, we can compute

Q, P  q  Q, P = z  ` (a + a† )  z = 2` Re z = Q

  
  
(4.52)

 ~   ~
(a − a† )  z = Im z = P

 
Q, P  p  Q, P = z  (4.53)
2i` `
as well as

Q, P  q 2  Q, P = z  `2 (a + a† )2  z = Q2 + `2

  
  
(4.54)

 ~2 ~2
Q, P  p2  Q, P = − z  2 (a − a† )2  z = P 2 + 2 .

    
(4.55)
4` 4`
Thus, the root mean square fluctuations in the coherent state | Q, P i are
s r
~ ~ m~ω0
∆q = ` = , ∆p = = , (4.56)
2mω0 2` 2
4.4. THERMAL EQUILIBRIUM 175

and ∆q · ∆p = 21 ~. Thus we learn that the coherent state ψQ,P (q) is localized in phase space, i.e. in both
position and momentum. If we have a general operator Â(q, p), we can then write

  
Q, P  Â(q, p)  Q, P = A(Q, P ) + O(~) , (4.57)

where A(Q, P ) is formed from Â(q, p) by replacing q → Q and p → P .

Since
d2z d Rez d Imz dQ dP
≡ = , (4.58)
2πi π 2π~
we can write the trace using coherent states as
Z∞ Z∞
1
  
Tr  = dQ dP Q, P    Q, P . (4.59)
2π~
−∞ −∞

We now can understand the origin of the factor 2π~ in the denominator of each (qi , pi ) integral over
classical phase space in eqn. 4.6.
Note that ω0 is arbitrary in our discussion. By increasing ω0 , the states become more localized in q
and more plane wave like in p. However, so long as ω0 is finite, the width of the coherent state in each
direction is proportional to ~1/2 , and thus vanishes in the classical limit.

4.4 Thermal Equilibrium

4.4.1 Two systems in thermal contact

Consider two systems in thermal contact, as depicted in fig. 4.4. The two subsystems #1 and #2 are
free to exchange energy, but their respective volumes and particle numbers remain fixed. We assume
the contact is made over a surface, and that the energy associated with that surface is negligible when
compared with the bulk energies E1 and E2 . Let the total energy be E = E1 + E2 . Then the density of
states D(E) for the combined system is
Z
D(E) = dE1 D1 (E1 ) D2 (E − E1 ) . (4.60)

The probability density for system #1 to have energy E1 is then

D1 (E1 ) D2 (E − E1 )
P1 (E1 ) = . (4.61)
D(E)
R
Note that P1 (E1 ) is normalized: dE1 P1 (E1 ) = 1. We now ask: what is the most probable value of E1 ?
We find out by differentiating P1 (E1 ) with respect to E1 and setting the result to zero. This requires
1 dP1 (E1 ) ∂
0= = ln P1 (E1 )
P1 (E1 ) dE1 ∂E1
(4.62)
∂ ∂
= ln D1 (E1 ) + ln D2 (E − E1 ) .
∂E1 ∂E1
176 CHAPTER 4. STATISTICAL ENSEMBLES

Figure 4.4: Two systems in thermal contact.

We conclude that the maximally likely partition of energy between systems #1 and #2 is realized when

∂S1 ∂S2
= . (4.63)
∂E1 ∂E2

This guarantees that

S(E, E1 ) = S1 (E1 ) + S2 (E − E1 ) (4.64)

is a maximum with respect to the energy E1 , at fixed total energy E.

The temperature T is defined as
 
1 ∂S
= , (4.65)
T ∂E V,N

a result familiar from thermodynamics. The difference is now we have a more rigorous definition of the
entropy. When the total entropy S is maximized, we have that T1 = T2 . Once again, two systems in
thermal contact and can exchange energy will in equilibrium have equal temperatures.
According to eqns. 4.19 and 4.29, the entropies of nonrelativistic and ultrarelativistic ideal gases in d
space dimensions are given by
   
1 E V
SNR = 2 N d kB ln + N kB ln + const. (4.66)
N N
   
E V
SUR = N d kB ln + N kB ln + const. . (4.67)
N N

Invoking eqn. 4.65, we then have

ENR = 12 N d kB T , EUR = N d kB T . (4.68)

We saw that the probability distribution P1 (E1 ) is maximized when T1 = T2 , but how sharp is the peak
in the distribution? Let us write E1 = E1∗ + ∆E1 , where E1∗ is the solution to eqn. 4.62. We then have

1 ∂ 2S1  1 ∂ 2S2 
 
ln P1 (E1∗ + ∆E1 ) = ln P1 (E1∗ ) + 2
(∆E1 ) + (∆E1 )2 + . . . , (4.69)
2kB ∂E12 E ∗ 2kB ∂E22 E ∗
1 2
4.4. THERMAL EQUILIBRIUM 177

where E2∗ = E − E1∗ . We must now evaluate

∂ 2S
   
∂ 1 1 ∂T 1
2
= =− 2 =− 2 , (4.70)
∂E ∂E T T ∂E V,N T CV


where CV = ∂E/∂T V,N is the heat capacity. Thus,
2 /2k T 2 C̄
P1 = P1∗ e−(∆E1 ) B V , (4.71)
where
CV,1 CV,2
C̄V = . (4.72)
CV,1 + CV,2
The distribution is therefore a Gaussian, and the fluctuations in ∆E1 can now be computed:
q
(∆E1 )2 = kB T 2 C̄V


=⇒ (∆E1 )RMS = kB T C̄V /kB . (4.73)
The individual heat capacities CV,1 and CV,2 scale with the volumes V1 and V2 , respectively. If V2  V1 ,
then CV,2  CV,1 , in which case C̄V ≈ CV,1 . Therefore the RMS fluctuations in ∆E1 are proportional to
the square root of the system size, whereas E1 itself is extensive. Thus, the ratio (∆E1 )RMS /E1 ∝ V −1/2
scales as the inverse square root of the volume. The distribution P1 (E1 ) is thus extremely sharp.

4.4.2 Thermal, mechanical and chemical equilibrium

1

We have dS V,N = T dE , but in general S = S(E, V, N ). Equivalently, we may write E = E(S, V, N ).
The full differential of E(S, V, N ) is then dE = T dS − p dV + µ dN , with T = ∂E ∂E



∂S V,N and p = − ∂V S,N
∂E


and µ = ∂N S,V
. As we shall discuss in more detail, p is the pressure and µ is the chemical potential .
We may thus write the total differential dS as
1 p µ
dS = dE + dV − dN . (4.74)
T T T
Employing the same reasoning as in the previous section, we conclude that entropy maximization for two
systems in contact requires the following:

• If two systems can exchange energy, then T1 = T2 . This is thermal equilibrium.

• If two systems can exchange volume, then p1 /T1 = p2 /T2 . This is mechanical equilibrium.
• If two systems can exchange particle number, then µ1 /T1 = µ2 /T2 . This is chemical equilibrium.

4.4.3 Gibbs-Duhem relation

The energy E(S, V, N ) is an extensive function of extensive variables, i.e. it is homogeneous of degree
one in its arguments. Therefore E(λS, λV, λN ) = λE, and taking the derivative with respect to λ yields
     
∂E ∂E ∂E
E=S +V +N
∂S V,N ∂V S,N ∂N S,V (4.75)
= T S − pV + µN .
178 CHAPTER 4. STATISTICAL ENSEMBLES

Taking the differential of each side, using the Leibniz rule on the RHS, and plugging in dE = T dS −
p dV + µ dN , we arrive at the Gibbs-Duhem relation 5 ,
S dT − V dp + N dµ = 0 . (4.76)
This, in turn, says that any one of the intensive quantities (T, p, µ) can be written as a function of the
other two, in the case of a single component system.

4.5 Ordinary Canonical Ensemble (OCE)

4.5.1 Canonical distribution and partition function

Consider a system S in contact with a world W , and let their union U = W ∪ S be called the ‘universe’.
The situation is depicted in fig. 4.2. The volume VS and particle number NS of the system are held
fixed, but the energy is allowed to fluctuate by exchange with the world W . We are interested in the
limit NS → ∞, NW → ∞, with NS  NW , with similar relations holding for the respective volumes and
energies. We now ask what is the probability that S is in a state | n i with energy En . This is given by
the ratio
DW (EU − En ) ∆E
Pn = lim
∆E→0 DU (EU ) ∆E
(4.77)
# of states accessible to W given that ES = En
= .
total # of states in U

Then
ln Pn = ln DW (EU − En ) − ln DU (EU )

∂ ln DW (E) 
= ln DW (EU ) − ln DU (EU ) − En + ... (4.78)
∂E 
E=E
U

≡ −α − βEn .
The constant β is given by

∂ ln DW (E)  1
β= = . (4.79)
∂E 
E=E kB T
U

Thus, we find Pn = e−α e−βEn . The constant α is fixed by the requirement that n Pn = 1:
P

1 X
Pn = e−βEn , Z(T, V, N ) = e−βEn = Tr e−β Ĥ . (4.80)
Z n

We’ve already met Z(β) in eqn. 4.12 – it is the Laplace transform of the density of states. It is also
called the partition function of the system S. Quantum mechanically, we can write the ordinary canonical
density matrix as
e−β Ĥ
%̂ = , (4.81)
Tr e−β Ĥ
5
See §2.7.4.
4.5. ORDINARY CANONICAL ENSEMBLE (OCE) 179

 
which is known as the Gibbs distribution. Note that %̂, Ĥ = 0, hence the ordinary canonical distribution
is a stationary solution to the evolution equation for the density matrix. Note that the OCE is specified
by three parameters: T , V , and N .

4.5.2 The difference between P (En ) and Pn

Let the total energy of the Universe be fixed at EU . The joint probability density P (ES , EW ) for the
system to have energy ES and the world to have energy EW is

P (ES , EW ) = DS (ES ) DW (EW ) δ(EU − ES − EW ) DU (EU ) , (4.82)
where
Z∞
DU (EU ) = dES DS (ES ) DW (EU − ES ) , (4.83)
−∞
R R
which ensures that dES dEW P (ES , EW ) = 1. The probability density P (ES ) is defined such that
P (ES ) dES is the (differential) probability for the system to have an energy in the range [ES , ES + dES ].
The units of P (ES ) are E −1 . To obtain P (ES ), we simply integrate the joint probability density P (ES , EW )
over all possible values of EW , obtaining
DS (ES ) DW (EU − ES )
P (ES ) = , (4.84)
DU (EU )
as we have in eqn. 4.77.
Now suppose we wish to know the probability Pn that the system is in a particular state | n i with energy
En . Clearly
probability that ES ∈ [En , En + ∆E] P (En ) ∆E D (E − En )
Pn = lim = = W U . (4.85)
∆E→0 # of S states with ES ∈ [En , En + ∆E] DS (En ) ∆E DU (EU )

4.5.3 Additional remarks

The formula of eqn. 4.77 is quite general and holds in the case where NS /NW = O(1), so long as we are
in the thermodynamic limit, where the energy associated with the interface between S and W may be
neglected. In this case, however, one is not licensed to perform the subsequent Taylor expansion, and the
distribution Pn is no longer of the Gibbs form. It is also valid for quantum systems6 , in which case we
interpret Pn = hn|%S |ni as a diagonal element of the density matrix %S . The density of states functions
may then be replaced by
EU −E
Z n +∆E
SW (EU −En , ∆E)
DW (EU − En ) ∆E → e ≡ Tra dE δ(E − ĤW )
W
EU −En
(4.86)
EUZ+∆E

DU (EU ) ∆E → eSU (EU , ∆E) ≡ Tra dE δ(E − ĤU ) .

U
EU
6
See T.-C. Lu and T. Grover, arXiv 1709.08784.
180 CHAPTER 4. STATISTICAL ENSEMBLES

The off-diagonal matrix elements of %S are negligible in the thermodynamic limit.

4.5.4 Averages within the OCE

To compute averages within the OCE,

Â|ni e−βEn
P
n hn|



 = Tr %̂  = P −βE , (4.87)
ne
n

where we have conveniently taken the trace in a basis of energy eigenstates. In the classical limit, we
have Z
1
%(ϕ) = e−β Ĥ(ϕ) , Z = Tr e−β Ĥ = dµ e−β Ĥ(ϕ) , (4.88)
Z
with dµ = N1 ! N d d d
Q
j=1 (d qj d pj /h ) for identical particles (‘Maxwell-Boltzmann statistics’). Thus,

dµ A(ϕ) e−β Ĥ(ϕ)

R
hAi = Tr (%A) = R . (4.89)
dµ e−β Ĥ(ϕ)

4.5.5 Entropy and free energy

The Boltzmann entropy is defined by

X
S = −kB Tr %̂ ln %̂) = −kB Pn ln Pn . (4.90)
n

The Boltzmann entropy and the statistical entropy S = kB ln D(E) are identical in the thermodynamic
limit. We define the Helmholtz free energy F (T, V, N ) as

F (T, V, N ) = −kB T ln Z(T, V, N ) , (4.91)

hence
Pn = eβF e−βEn , ln Pn = βF − βEn . (4.92)
Therefore the entropy is
X
F h Ĥ i
S = −kB Pn βF − βEn = − + , (4.93)
n
T T
which is to say F = E − T S, where

X Tr Ĥ e−β Ĥ
E= Pn En = (4.94)
n Tr e−β Ĥ

is the average energy. We also see that

−βEn
P
n En e ∂ ∂
X
Z = Tr e−β Ĥ = e−βEn


=⇒ E= P −βEn
= − ln Z = βF . (4.95)
n ne ∂β ∂β
4.5. ORDINARY CANONICAL ENSEMBLE (OCE) 181

Thus, F (T, V, N ) is a Legendre transform of E(S, V, N ), with

dF = −S dT − p dV + µ dN , (4.96)

which means
     
∂F ∂F ∂F
S=− , p=− , µ=+ . (4.97)
∂T V,N ∂V T,N ∂N T,V

4.5.6 Fluctuations in the OCE

In the OCE, the energy is not fixed. It therefore fluctuates about its average value E = hĤi. Note that

∂E ∂E ∂ 2 ln Z
− = kB T 2 =
∂β ∂T ∂β 2
!2
Tr Ĥ 2 e−β Ĥ Tr Ĥ e−β Ĥ (4.98)
= −
Tr e−β Ĥ Tr e−β Ĥ

2

2
= Ĥ − Ĥ .

Thus, the heat capacity is related to the fluctuations in the energy, just as we saw at the end of §4.4:
 
∂E 1 
2 
2 
CV = = Ĥ − Ĥ (4.99)
∂T V,N kB T 2

For the nonrelativistic ideal gas, we found CV = d2 N kB , hence the ratio of RMS fluctuations in the energy
to the energy itself is q

(∆Ĥ)2
p r
k B T 2 CV 2
= d = , (4.100)
hĤi 2 N kB T
Nd
and the ratio of the RMS fluctuations to the mean value vanishes in the thermodynamic limit.
The full distribution function for the energy is

 Tr δ(E − Ĥ) e−β Ĥ 1

= D(E) e−βE .

P (E) = δ(E − Ĥ) = (4.101)

Tr e−β Ĥ Z
Thus,
e−β[E−T S(E)]
P (E) = R , (4.102)
dE 0 e−β[E 0 −T S(E 0 )]
where S(E) = kB ln D(E) is the statistical entropy. Let’s write E = E + δE, where E extremizes the
combination E − T S(E), i.e. the solution to T S 0 (E) = 1, where the energy derivative of S is performed
at fixed volume V and particle number N . We now expand S(E + δE) to second order in δE, obtaining

2
δE δE
S(E + δE) = S(E) + − + ... (4.103)
T 2T 2 CV
182 CHAPTER 4. STATISTICAL ENSEMBLES

Recall that S 00 (E) = ∂ 1

= − T 21C . Thus,


∂E T V

(δE)2
+ O (δE)3 .


E − T S(E) = E − T S(E) + (4.104)
2T CV

Applying this to both numerator and denominator of eqn. 4.102, we obtain7

" #
(δE)2
P (E) = N exp − , (4.105)
2kB T 2 CV

where N = (2πkB T 2 CV )−1/2 is a normalization constant which guarantees dE P (E) = 1. Once again,
R
p
we see that the distribution is a Gaussian centered at hEi = E, and of width (∆E)RMS = kB T 2 CV .
This is a consequence of the Central Limit Theorem.

4.5.7 Thermodynamics revisited

The average energy within the OCE is X

E= En Pn , (4.106)
n
and therefore
X X
dE = En dPn + Pn dEn
n n (4.107)
= d̄Q − dW
¯ ,

where
X
dW
¯ =− Pn dEn (4.108)
n
X
dQ
¯ = En dPn . (4.109)
n

Finally, from Pn = Z −1 e−En /kB T , we can write

En = −kB T ln Z − kB T ln Pn , (4.110)

with which we obtain

X
dQ
¯ = En dPn
n
X X
= −kB T ln Z dPn − kB T ln Pn dPn (4.111)
n n
 X 
= T d − kB Pn ln Pn = T dS .
n
7
In applying eqn. 4.104 to the denominator of eqn. 4.102, we shift E 0 by E and integrate over the difference δE 0 ≡ E 0 − E,
retaining terms up to quadratic order in δE 0 in the argument of the exponent.
4.5. ORDINARY CANONICAL ENSEMBLE (OCE) 183

Figure 4.5: Microscopic, statistical interpretation of the First Law of Thermodynamics.

Note also that

X
dW
¯ =− Pn dEn
n
!
X X ∂E
n
=− Pn dXi (4.112)
n
∂Xi
i
X
 ∂ Ĥ   X
=− Pn n   n dXi ≡ Fi dXi ,
∂Xi
n,i i

so the generalized force Fi conjugate to the generalized displacement dXi is

 
X ∂En ∂ Ĥ
Fi = − Pn =− . (4.113)
n
∂Xi ∂Xi

This is the force acting on the system8 . In the chapter on thermodynamics, we defined the generalized
force conjugate to Xi as yi ≡ −Fi .
Thus we see from eqn. 4.107 that there are two ways that the average energy can change; these are
depicted in the sketch of fig. 4.5. Starting from a set of energy levels {En } and probabilities {Pn }, we
8
In deriving eqn. 4.113, we have used the so-called Feynman-Hellman theorem of quantum mechanics: dhn|Ĥ|ni =
hn| dĤ |ni, if |ni is an energy eigenstate.
184 CHAPTER 4. STATISTICAL ENSEMBLES

can shift the energies to {En0 }. The resulting change in energy (∆E)I = −W is identified with the work
done on the system. We could also modify the probabilities to {Pn0 } without changing the energies. The
energy change in this case is the heat absorbed by the system: (∆E)II = Q. This provides us with a
statistical and microscopic interpretation of the First Law of Thermodynamics.

4.5.8 Generalized susceptibilities

Suppose our Hamiltonian is of the form

Ĥ = Ĥ(λ) = Ĥ0 − λ Q̂ , (4.114)

where λ is an intensive parameter, such as magnetic field. Then

Z(λ) = Tr e−β(Ĥ0 −λQ̂) (4.115)

and
1 ∂Z 1  
= β · Tr Q̂ e−β Ĥ(λ) = β hQ̂i . (4.116)
Z ∂λ Z
But then from Z = e−βF we have
 
∂F
Q(λ, T ) = h Q̂ i = − . (4.117)
∂λ T

Typically we will take Q to be an extensive quantity. We can now define the susceptibility χ as

1 ∂Q 1 ∂ 2F
χ= =− . (4.118)
V ∂λ V ∂λ2
The volume factor in the denominator ensures that χ is intensive.
 
It is important to realize that we have assumed here that Ĥ0 , Q̂ = 0, i.e. the ‘bare’ Hamiltonian Ĥ0
and the operator Q̂ commute. If they do not commute, then the response functions must be computed
within a proper quantum mechanical formalism, which we shall not discuss here.
  
Note also that we can imagine an entire family of observables Q̂k satisfying Q̂k , Q̂k0 = 0 and
Ĥ0 , Q̂k = 0, for all k and k 0 . Then for the Hamiltonian
 

Ĥ (~λ) = Ĥ0 −
X
λk Q̂k , (4.119)
k

we have that  
∂F
Qk (~λ, T ) = h Q̂k i = − (4.120)
∂λk T, Na , λk0 6=k

and we may define an entire matrix of susceptibilities,

1 ∂Qk 1 ∂ 2F
χkl = =− . (4.121)
V ∂λl V ∂λk ∂λl
4.6. GRAND CANONICAL ENSEMBLE (GCE) 185

4.6 Grand Canonical Ensemble (GCE)

4.6.1 Grand canonical distribution and partition function

Consider once again the situation depicted in fig. 4.2, where a system S is in contact with a world W ,
their union U = W ∪ S being called the ‘universe’. We assume that the system’s volume VS is fixed,
but otherwise it is allowed to exchange energy and particle number with W . Hence, the system’s energy
ES and particle number NS will fluctuate. We ask what is the probability that S is in a state | n i with
energy En and particle number Nn . This is given by the ratio
DW (EU − En , NU − Nn ) ∆E
Pn = lim
∆E→0 DU (EU , NU ) ∆E
(4.122)
# of states accessible to W given that ES = En and NS = Nn
= .
total # of states in U

Then
ln Pn = ln DW (EU − En , NU − Nn ) − ln DU (EU , NU )

= ln DW (EU , NU ) − ln DU (EU , NU )
  (4.123)
∂ ln DW (E, N )  ∂ ln DW (E, N ) 
− En  E=EU − Nn  E=EU + . . .
∂E ∂N
N =N N =N
U U

≡ −α − βEn + βµNn .
The constants β and µ are given by

∂ ln DW (E, N )  1
β= = (4.124)
∂E  E=E
U kB T
N =N
U

∂ ln DW (E, N ) 
µ = −kB T  E=EU . (4.125)
∂N
N =N
U

The quantity µ has dimensions of energy and is called the chemical potential . Nota bene: Some texts
define the ‘grand canonical Hamiltonian’ K̂ as

K̂ ≡ Ĥ − µN̂ . (4.126)

Thus, Pn = e−α e−β(En −µNn ) . Once again, the constant α is fixed by the requirement that
P
n Pn = 1:
1 −β(En −µNn ) X
Pn = e , Ξ(β, V, µ) = e−β(En −µNn ) = Tr e−β(Ĥ−µN̂ ) = Tr e−β K̂ . (4.127)
Ξ n

Thus, the quantum mechanical grand canonical density matrix is given by

e−β K̂
%̂ = . (4.128)
Tr e−β K̂
186 CHAPTER 4. STATISTICAL ENSEMBLES

 
Note that %̂, K̂ = 0. The quantity Ξ(T, V, µ) is called the grand partition function. It stands in relation
to a corresponding free energy in the usual way:

Ξ(T, V, µ) ≡ e−βΩ(T,V,µ) ⇐⇒ Ω = −kB T ln Ξ , (4.129)

where Ω(T, V, µ) is the grand potential , also known as the Landau free energy. The dimensionless quantity
z ≡ eβµ is called the fugacity.
 
If Ĥ, N̂ = 0, the grand potential may be expressed as a sum over contributions from each N sector, viz.
X
Ξ(T, V, µ) = eβµN Z(T, V, N ) . (4.130)
N

When there is more than one species, we have several chemical potentials {µa }, and accordingly we define
X
K̂ = Ĥ − µa N̂a , (4.131)
a

with Ξ = Tr e−β K̂ as before.

4.6.2 Entropy and Gibbs-Duhem relation

In the GCE, the Boltzmann entropy is

X
S = −kB Pn ln Pn
n
X  
= −kB Pn βΩ − βEn + βµNn (4.132)
n
Ω hĤi µ hN̂ i
=− + − ,
T T T
which says
Ω = E − T S − µN , (4.133)
where
X

E= En Pn = Tr %̂ Ĥ (4.134)
n
X

N= Nn Pn = Tr %̂ N̂ . (4.135)
n

Therefore, Ω(T, V, µ) is a double Legendre transform of E(S, V, N ), with

dΩ = −S dT − p dV − N dµ , (4.136)

which entails
     
∂Ω ∂Ω ∂Ω
S=− , p=− , N =− . (4.137)
∂T V,µ ∂V T,µ ∂µ T,V
4.6. GRAND CANONICAL ENSEMBLE (GCE) 187

Since Ω(T, V, µ) is an extensive quantity, we must be able to write Ω = V ω(T, µ). We identify the
function ω(T, µ) as the negative of the pressure:
 
∂Ω kB T ∂Ξ 1 X ∂En −β(En −µNn )
=− = e
∂V Ξ ∂V T,µ Ξ n ∂V
  (4.138)
∂E
= = −p(T, µ) .
∂V T,µ
Therefore,
Ω = −pV , p = p(T, µ) (equation of state) . (4.139)
This is consistent with the result from thermodynamics that G = E − T S + pV = µN . Taking the
differential, we recover the Gibbs-Duhem relation,
dΩ = −S dT − p dV − N dµ = −p dV − V dp ⇒ S dT − V dp + N dµ = 0 . (4.140)

4.6.3 Generalized susceptibilities in the GCE

We can appropriate the results §4.5.8 and

 from  apply them,
 mutatis
 mutandis,
 to the
 GCE. Suppose we
have a family of observables Q̂k satisfying Q̂k , Q̂k0 = 0 and Ĥ0 , Q̂k = 0 and N̂a , Q̂k = 0 for all
k, k 0 , and a. Then for the grand canonical Hamiltonian

K̂ (~λ) = Ĥ0 −
X X
µa N̂a − λk Q̂k , (4.141)
a k

we have that  
∂Ω
Qk (~λ, T ) = h Q̂k i = − (4.142)
∂λk T,µa , λk0 6=k

and we may define the matrix of generalized susceptibilities,

1 ∂Qk 1 ∂2Ω
χkl = =− . (4.143)
V ∂λl V ∂λk ∂λl

4.6.4 Fluctuations in the GCE

Both energy and particle number fluctuate in the GCE. Let us compute the fluctuations in particle
number. We have
Tr N̂ e−β(Ĥ−µN̂ ) 1 ∂
N = h N̂ i = = ln Ξ . (4.144)
Tr e−β(Ĥ−µN̂ ) β ∂µ
Therefore,
!2
1 ∂N Tr N̂ 2 e−β(Ĥ−µN̂ ) Tr N̂ e−β(Ĥ−µN̂ )
= −
β ∂µ Tr e−β(Ĥ−µN̂ ) Tr e−β(Ĥ−µN̂ )
(4.145)

2 
2
= N̂ − N̂ .
188 CHAPTER 4. STATISTICAL ENSEMBLES

Note now that 

2
N̂ 2 − N̂

 
kB T ∂N k T

2 = 2
= B κT , (4.146)
N̂ N ∂µ T,V V
where κT is the isothermal compressibility. Note:
 
∂N ∂(N, T, V ) ∂(N, T, V )
= =−
∂µ T,V ∂(µ, T, V ) ∂(V, T, µ)
1
z }| {
∂(N, T, V ) ∂(N, T, p) ∂(V, T, p) ∂(N, T, µ) (4.147)
=− · · ·
∂(N, T, p) ∂(V, T, p) ∂(N, T, µ) ∂(V, T, µ)
N 2 ∂V N2
 
=− 2 = κ .
V ∂p T,N V T

Thus, r
(∆N )RMS kB T κT
= , (4.148)
N V
which again scales as V −1/2 .

4.6.5 Gibbs ensemble

Let the system’s particle number N be fixed, but let it exchange energy and volume with the world W .
Mutatis mutandis, we have
DW (EU − En , VU − Vn ) ∆E ∆V
Pn = lim lim . (4.149)
∆E→0 ∆V →0 DU (EU , VU ) ∆E ∆V
Then
ln Pn = ln DW (EU − En , VU − Vn ) − ln DU (EU , VU )

= ln DW (EU , VU ) − ln DU (EU , VU )
  (4.150)
∂ ln DW (E, V )  ∂ ln DW (E, V ) 
− En  E=EU − Vn  E=EU + . . .
∂E ∂V
V =V V =V
U U

≡ −α − βEn − βp Vn .

The constants β and p are given by


∂ ln DW (E, V )  1
β= = (4.151)
∂E  E=E
U kB T
V =V
U

∂ ln DW (E, V ) 
p = kB T  E=EU . (4.152)
∂V
V =V
U
4.7. STATISTICAL ENSEMBLES FROM MAXIMUM ENTROPY 189

The corresponding partition function is

Z∞
−β(Ĥ+pV ) 1
Y (T, p, N ) = Tr e = dV e−βpV Z(T, V, N ) ≡ e−βG(T,p,N ) , (4.153)
V0
0

where V0 is a constant which has dimensions of volume. The factor V0−1 in front of the integral renders
Y dimensionless. Note that G(V00 ) = G(V0 ) + kB T ln(V00 /V0 ), so the difference is not extensive and can
be neglected in the thermodynamic limit. In other words, it doesn’t matter what constant we choose for
V0 since it contributes subextensively to G. Moreover, in computing averages, the constant V0 divides
out in the ratio of numerator and denominator. Like the Helmholtz free energy, the Gibbs free energy
G(T, p, N ) is also a double Legendre transform of the energy E(S, V, N ), viz.

G = E − T S + pV
(4.154)
dG = −S dT + V dp + µ dN ,

which entails
     
∂G ∂G ∂G
S=− , V =+ , µ=+ . (4.155)
∂T p,N ∂p T,N ∂N T,p

4.7 Statistical Ensembles from Maximum Entropy

The basic principle: maximize the entropy,

X
S = −kB Pn ln Pn . (4.156)
n

4.7.1 µCE

We maximize S subject to the single constraint

X
C= Pn − 1 = 0 . (4.157)
n

We implement the constraint C = 0 with a Lagrange multiplier, λ̄ ≡ kB λ, writing

S ∗ = S − kB λ C , (4.158)

and freely extremizing over the distribution {Pn } and the Lagrange multiplier λ. Thus,

δS ∗ = δS − kB λ δC − kB C δλ
Xh i
= −kB ln Pn + 1 + λ δPn − kB C δλ ≡ 0 . (4.159)
n

We conclude that C = 0 and that

ln Pn = − 1 + λ , (4.160)
190 CHAPTER 4. STATISTICAL ENSEMBLES

P
and we fix λ by the normalization condition n Pn = 1. This gives
1 X
Pn = , Ω= Θ(E + ∆E − En ) Θ(En − E) . (4.161)
Ω n

Note that Ω is the number of states with energies between E and E + ∆E.

4.7.2 OCE

We maximize S subject to the two constraints

X X
C1 = Pn − 1 = 0 , C2 = En Pn − E = 0 . (4.162)
n n

We now have two Lagrange multipliers. We write

2
X
∗
S = S − kB λj Cj , (4.163)
j=1

and we freely extremize over {Pn } and {Cj }. We therefore have

X 2
X
∗


δS = δS − kB λ1 + λ2 En δPn − kB Cj δλj
n j=1
(4.164)
Xh i 2
X
= −kB ln Pn + 1 + λ1 + λ2 En δPn − kB Cj δλj ≡ 0 .
n j=1

Thus, C1 = C2 = 0 and

ln Pn = − 1 + λ1 + λ2 En . (4.165)
We define λ2 ≡ β and we fix λ1 by normalization. This yields
1 −βEn X
Pn = e , Z= e−βEn . (4.166)
Z n

4.7.3 GCE

We maximize S subject to the three constraints

X X X
C1 = Pn − 1 = 0 , C2 = En Pn − E = 0 , C3 = Nn Pn − N = 0 . (4.167)
n n n

We now have three Lagrange multipliers. We write

3
X
S ∗ = S − kB λj Cj , (4.168)
j=1
4.8. IDEAL GAS STATISTICAL MECHANICS 191

and hence
X 3
X
δS ∗ = δS − kB


λ1 + λ2 En + λ3 Nn δPn − kB Cj δλj
n j=1
(4.169)
Xh i 3
X
= −kB ln Pn + 1 + λ1 + λ2 En + λ3 Nn δPn − kB Cj δλj ≡ 0 .
n j=1

Thus, C1 = C2 = C3 = 0 and

ln Pn = − 1 + λ1 + λ2 En + λ3 Nn . (4.170)
We define λ2 ≡ β and λ3 ≡ −βµ, and we fix λ1 by normalization. This yields
1 −β(En −µNn ) X
Pn = e , Ξ= e−β(En −µNn ) . (4.171)
Ξ n

4.8 Ideal Gas Statistical Mechanics

The ordinary canonical partition function for the ideal gas was computed in eqn. 4.14. We found
N Z
1 Y ddxi ddpi −βp2 /2m
Z(T, V, N ) = e i
N! (2π~)d
i=1
 ∞ N d
VN  dp −βp2 /2m 
Z
= e (4.172)
N! 2π~
−∞
 N
1 V
= ,
N ! λdT

where λT is the thermal wavelength: p

λT = 2π~2 /mkB T . (4.173)
The physical interpretation of λT is that it is the de Broglie wavelength for a particle of mass m which
has a kinetic energy of kB T .
In the GCE, we have
∞
X
Ξ(T, V, µ) = eβµN Z(T, V, N )
N =0
∞
!N ! (4.174)
X 1 V eµ/kB T V eµ/kB T
= = exp .
N!
N =1
λdT λdT

From Ξ = e−Ω/kB T , we have the grand potential is

Ω(T, V, µ) = −V kB T eµ/kB T λdT .


(4.175)
192 CHAPTER 4. STATISTICAL ENSEMBLES

Since Ω = −pV (see §4.6.2), we have

p(T, µ) = kB T λ−d
T e
µ/kB T
. (4.176)
The number density can also be calculated:
 
N 1 ∂Ω
n= =− = λ−d
T e
µ/kB T
. (4.177)
V V ∂µ T,V

Combined, the last two equations recapitulate the ideal gas law, pV = N kB T .

4.8.1 Maxwell velocity distribution

The distribution function for momenta is given by

N
D1 X E
g(p) = δ(pi − p) . (4.178)
N
i=1


Note that g(p) = δ(pi − p) is the same for every particle, independent of its label i. We compute the
average hAi = Tr Ae−β Ĥ / Tr e−β Ĥ . Setting i = 1, all the integrals other than that over p1 divide out

between numerator and denominator. We then have

2
d p1 δ(p1 − p) e−βp1 /2m
R 3
g(p) = 2
d3p1 e−βp1 /2m
R
(4.179)
2 /2m
= (2πmkB T )−3/2 e−βp .
Textbooks commonly refer to the velocity distribution f (v), which is related to g(p) by
f (v) d3v = g(p) d3p . (4.180)
Hence,
 3/2
m 2 /2k T
f (v) = e−mv B . (4.181)
2πkB T
This is known as the Maxwell velocity distribution. Note that the distributions are normalized, viz.
Z Z
d3p g(p) = d3v f (v) = 1 . (4.182)

If we are only interested in averaging functions of v = |v| which are isotropic, then we can define the
Maxwell speed distribution, f˜(v), as
 3/2
m 2
˜ 2
f (v) = 4π v f (v) = 4π v 2 e−mv /2kB T . (4.183)
2πkB T

Note that f˜(v) is normalized according to

Z∞
dv f˜(v) = 1 . (4.184)
0
4.8. IDEAL GAS STATISTICAL MECHANICS 193

√
Figure 4.6: Maxwell distribution of speeds ϕ(v/v0 ). The most probable speed is vMAX = 2 v0 . The
q
8
√
average speed is vAVG = π v0 . The RMS speed is vRMS = 3 v0 .

p
It is convenient to represent v in units of v0 = kB T /m, in which case

1
q
2 /2
f˜(v) = ϕ(v/v0 ) , ϕ(s) = 2
π s2 e−s . (4.185)
v0

The distribution ϕ(s) is shown in fig. 4.6. Computing averages, we have

Z∞
2
k
Ck ≡ hs i = ds sk ϕ(s) = 2k/2 · √ Γ 3 k


2 + 2 . (4.186)
π
0
q
8
Thus, C0 = 1, C1 = π, C2 = 3, etc. The speed averages are

 k/2

k kB T
v = Ck . (4.187)
m
p
Note that the average velocity is hvi = 0, but the average speed is hvi = 8kB T /πm. The speed
distribution is plotted in fig. 4.6.

4.8.2 Equipartition

The Hamiltonian for ballistic (i.e. massive nonrelativistic) particles is quadratic in the individual com-
ponents of each momentum pi . There are other cases in which a classical degree of freedom appears
194 CHAPTER 4. STATISTICAL ENSEMBLES

quadratically in Ĥ as well. For example, an individual normal mode ξ of a system of coupled oscillators
has the Lagrangian
L = 21 ξ˙2 − 12 ω02 ξ 2 , (4.188)
where the dimensions of ξ are [ξ] = M 1/2 L by convention. The Hamiltonian for this normal mode is then

p2 1 2 2
Ĥ = + 2 ω0 ξ , (4.189)
2
from which we see that both the kinetic as well as potential energy terms enter quadratically into the
Hamiltonian. The classical rotational kinetic energy is also quadratic in the angular momentum compo-
nents.
Let us compute the contribution of a single quadratic degree of freedom in Ĥ to the partition function.
We’ll call this degree of freedom ζ – it may be a position or momentum or angular momentum – and
we’ll write its contribution to Ĥ as
Ĥζ = 12 Kζ 2 , (4.190)
where K is some constant. Integrating over ζ yields the following factor in the partition function:
Z∞
2π 1/2
 
−βKζ 2 /2
dζ e = . (4.191)
Kβ
−∞

The contribution to the Helmholtz free energy is then

 
1 K
∆Fζ = 2 kB T ln , (4.192)
2πkB T

and therefore the contribution to the internal energy E is

∂ 1
= 12 kB T .


∆Eζ = β ∆Fζ = (4.193)
∂β 2β
We have thus derived what is commonly called the equipartition theorem of classical statistical mechanics:

To each degree of freedom which enters the Hamiltonian quadratically is associated a contri-
bution 21 kB T to the internal energy of the system. This results in a concomitant contribution
of 12 kB to the heat capacity.

We now see why the internal energy of a classical ideal gas with f degrees of freedom per molecule
is E = 21 f N kB T , and CV = 12 N kB . This result also has applications in the theory of solids. The
atoms in a solid possess kinetic energy due to their motion, and potential energy due to the spring-like
interatomic potentials which tend to keep the atoms in their preferred crystalline positions. Thus, for a
three-dimensional crystal, there are six quadratic degrees of freedom (three positions and three momenta)
per atom, and the classical energy should be E = 3N kB T , and the heat capacity CV = 3N kB . As we
shall see, quantum mechanics modifies this result considerably at temperatures below the highest normal
mode (i.e. phonon) frequency, but the high temperature limit is given by the classical value CV = 3νR
(where ν = N/NA is the number of moles) derived here, known as the Dulong-Petit limit.
4.9. SELECTED EXAMPLES 195

4.9 Selected Examples

4.9.1 Spins in an external magnetic field

Consider a system of NS spins , each of which can be either up (σ = +1) or down (σ = −1). The
Hamiltonian for this system is
NS
X
Ĥ = −µ0 H σj , (4.194)
j=1

where now we write Ĥ for the Hamiltonian, to distinguish it from the external magnetic field H, and µ0
is the magnetic moment per particle. We treat this system within the ordinary canonical ensemble. The
partition function is X X
Z= ··· e−β Ĥ = ζ NS , (4.195)
σ1 σN
S

where ζ is the single particle partition function:

 
X
µ0 Hσ/kB T µ0 H
ζ= e = 2 cosh . (4.196)
kB T
σ=±1

The Helmholtz free energy is then

 "#
µ0 H
F (T, H, NS ) = −kB T ln Z = −NS kB T ln 2 cosh . (4.197)
kB T

The magnetization is    
∂F µ0 H
M =− = NS µ0 tanh . (4.198)
∂H T, NS kB T
The energy is  
∂ µ H
βF = −NS µ0 H tanh 0


E= . (4.199)
∂β kB T
Hence, E = −HM , which we already knew, from the form of Ĥ itself.
Each spin here is independent. The probability that a given spin has polarization σ is

eβµ0 Hσ
Pσ = . (4.200)
eβµ0 H + e−βµ0 H
The total probability is unity, and the average polarization is a weighted average of σ = +1 and σ = −1
contributions:  
µ H
P↑ + P↓ = 1 , hσi = P↑ − P↓ = tanh 0 . (4.201)
kB T
At low temperatures T  µ0 H/kB , we have P↑ ≈ 1 − e−2µ0 H/kB T . At high temperatures T > µ0 H/kB ,
 
σµ H
the two polarizations are equally likely, and Pσ ≈ 21 1 + k 0T .
B
196 CHAPTER 4. STATISTICAL ENSEMBLES

The isothermal magnetic susceptibility is defined as

µ20
   
1 ∂M 2 µ0 H
χT = = sech . (4.202)
NS ∂H T kB T kB T

(Typically this is computed per unit volume rather than per particle.) At H = 0, we have χT = µ20 /kB T ,
which is known as the Curie law .

Aside

The energy E = −HM here is not the same quantity we discussed in our study of thermodynamics.
In fact, the thermodynamic energy for this problem vanishes! Here is why. To avoid confusion, we’ll
need to invoke a new symbol for the thermodynamic energy, E. Recall that the thermodynamic energy
E is a function of extensive quantities, meaning E = E(S, M, NS ). It is obtained from the free energy
F (T, H, NS ) by a double Legendre transform:

E(S, M, NS ) = F (T, H, NS ) + T S + HM . (4.203)

Now from eqn. 4.197 we derive the entropy

"  #  
∂F µ0 H µ0 H µ0 H
S=− = NS kB ln 2 cosh − NS tanh . (4.204)
∂T kB T T kB T

Thus, using eqns. 4.197 and 4.198, we obtain E(S, M, NS ) = 0.

The potential confusion here arises from our use of the expression F (T, H, NS ). In thermodynamics, it is
the Gibbs free energy G(T, p, N ) which is a double Legendre transform of the energy: G = E − T S + pV .
By analogy, with magnetic systems we should perhaps write G = E − T S − HM , but in keeping with
many textbooks we shall use the symbol F and refer to it as the Helmholtz free energy. The quantity
we’ve called E in eqn. 4.199 is in fact E = E − HM , which means E = 0. The energy E(S, M, NS )
vanishes here because the spins are noninteracting.

4.9.2 Negative temperature (!)

Consider again a system of NS spins, each of which can be either up (+) or down (−). Let Nσ be the
number of sites with spin σ, where σ = ±1. Clearly N+ + N− = NS . We now treat this system within
the microcanonical ensemble.
The energy of the system is
E = −HM , (4.205)
where H is an external magnetic field, and M = (N+ − N− ) µ0 is the total magnetization. We now
compute S(E) using the ordinary canonical ensemble. The number of ways of arranging the system with
N+ up spins is  
NS
Ω= , (4.206)
N+
4.9. SELECTED EXAMPLES 197

Figure 4.7: When entropy decreases with increasing energy, the temperature is negative. Typically,
kinetic degrees of freedom prevent this peculiarity from manifesting in physical systems.

hence the entropy is n o

S = kB ln Ω = −NS kB x ln x + (1 − x) ln(1 − x) (4.207)

in the thermodynamic limit: NS → ∞, N+ → ∞, x = N+ /NS constant. Now the magnetization is

M = (N+ − N− )µ0 = (2N+ − NS )µ0 , hence if we define the maximum energy E0 ≡ NS µ0 H, then

E M E0 − E
=− = 1 − 2x =⇒ x= . (4.208)
E0 NS µ0 2E0

We therefore have
"       #
E0 − E E0 − E E0 + E E0 + E
S(E, NS ) = −NS kB ln + ln . (4.209)
2E0 2E0 2E0 2E0

We now have    
1 ∂S ∂S ∂x NS kB E0 − E
= = = ln . (4.210)
T ∂E NS ∂x ∂E 2E0 E0 + E
We see that the temperature is positive for −E0 ≤ E < 0 and is negative for 0 < E ≤ E0 .
What has gone wrong? The answer is that nothing has gone wrong – all our calculations are perfectly
correct. This system does exhibit the possibility of negative temperature. It is, however, unphysical in
that we have neglected kinetic degrees of freedom, which result in an entropy function S(E, NS ) which
is an increasing function of energy. In this system, S(E, NS ) achieves a maximum of Smax = NS kB ln 2
at E = 0 (i.e. x = 21 ), and then turns over and starts decreasing. In fact, our results are completely
consistent with eqn. 4.199 : the energy E is an odd function of temperature. Positive energy requires
negative temperature! Another example of this peculiarity is provided in the appendix in §4.11.2.
198 CHAPTER 4. STATISTICAL ENSEMBLES

4.9.3 Adsorption

PROBLEM: A surface containing NS adsorption sites is in equilibrium with a monatomic ideal gas. Atoms
adsorbed on the surface have an energy −∆ and no kinetic energy. Each adsorption site can accommodate
at most one atom. Calculate the fraction f of occupied adsorption sites as a function of the gas density
n, the temperature T , the binding energy ∆, and physical constants.
The grand partition function for the surface is
NS  
−Ωsurf /kB T
X N
Ξsurf = e = S
ej(µ+∆)/kB T
j (4.211)
j=0
T N
= 1 + eµ/kB T e∆/kB


S
.

The fraction of occupied sites is

hN̂surf i 1 ∂Ωsurf eµ/kB T

f= =− = µ/k T . (4.212)
NS NS ∂µ e B + e−∆/kB T
Since the surface is in equilibrium with the gas, its fugacity z = exp(µ/kB T ) and temperature T are the
same as in the gas.
SOLUTION:
p For a monatomic ideal gas, the single particle partition function is ζ = V λ−3
T , where λT =
2π~2 /mkB T is the thermal wavelength. Thus, the grand partition function, for indistinguishable par-
ticles, is  
Ξgas = exp V λ−3T e µ/kB T
. (4.213)
The gas density is
hN̂gas i 1 ∂Ωgas
n= =− = λ−3
T e
µ/kB T
. (4.214)
V V ∂µ
We can now solve for the fugacity: z = eµ/kB T = nλ3T . Thus, the fraction of occupied adsorption sites is

nλ3T
f= . (4.215)
nλ3T + e−∆/kB T
Interestingly, the solution for f involves the constant ~.
It is always advisable to check that the solution makes sense in various limits. First of all, if the gas
density tends to zero at fixed T and ∆, we have f → 0. On the other hand, if n → ∞ we have f → 1,
which also makes sense. At fixed n and T , if the adsorption energy is (−∆) → −∞, then once again
f = 1 since every adsorption site wants to be occupied. Conversely, taking (−∆) → +∞ results in n → 0,
since the energetic cost of adsorption is infinitely high.

4.9.4 Elasticity of wool

Wool consists of interlocking protein molecules which can stretch into an elongated configuration, but
reversibly so. This feature gives wool its very useful elasticity. Let us model a chain of these proteins
4.9. SELECTED EXAMPLES 199

Figure 4.8: The monomers in wool are modeled as existing in one of two states. The low energy unde-
formed state is A, and the higher energy deformed state is B. Applying tension induces more monomers
to enter the B state.

by assuming they can exist in one of two states, which we will call A and B, with energies εA and εB
and lengths `A and `B . The situation is depicted in fig. 4.8. We model these conformational degrees of
freedom by a spin variable σ = ±1 for each molecule, where σ = +1 in the A state and σ = −1 in the B
state. Suppose a chain consisting of N monomers is placed under a tension τ . We then have

N h
X i
Ĥ = εA δσj ,+1 + εB δσj ,−1 . (4.216)
j=1

Similarly, the length is

N h
X i
L̂ = `A δσj ,+1 + `B δσj ,−1 . (4.217)
j=1

The Gibbs partition function is Y = Tr e−K̂/kB T , with K̂ = Ĥ − τ L̂ :

N h
X i
K̂ = ε̃A δσj ,+1 + ε̃B δσj ,−1 , (4.218)
j=1

where ε̃A ≡ εA − τ `A and ε̃B ≡ εB − τ `B . At τ = 0 the A state is preferred for each monomer, but when
τ exceeds τ ∗ , defined by the relation ε̃A = ε̃B , the B state is preferred. One finds

εB − εA
τ∗ = . (4.219)
`B − `A

Once again, we have a set of N noninteracting spins. The partition function is Y = ζ N , where ζ is the
single monomer partition function, ζ = Tr e−β ĥ , where

ĥ = ε̃A δσj ,1 + ε̃B δσj ,−1 (4.220)

is the single “spin” Hamiltonian. Thus,

ζ = Tr e−β ĥ = e−β ε̃A + e−β ε̃B , (4.221)

It is convenient to define the differences

∆ε = εB − εA , ∆` = `B − `A , ∆ε̃ = ε̃B − ε̃A (4.222)

200 CHAPTER 4. STATISTICAL ENSEMBLES

Figure 4.9: Upper panel: length L(τ, T ) for kB T /ε̃ = 0.01 (blue), 0.1 (green), 0.5 (dark red), and 1.0
(red). Bottom panel: dimensionless force constant k/N (∆`)2 versus temperature.

in which case the partition function Y is

h iN
Y (T, τ, N ) = e−N β ε̃A 1 + e−β∆ε̃ (4.223)
h i
G(T, τ, N ) = N ε̃A − N kB T ln 1 + e−∆ε̃/kB T (4.224)

The average length is

 
∂G
L = hL̂i = −
∂τ T,N
(4.225)
N ∆`
= N `A + .
e(∆ε−τ ∆`)/kB T + 1
The polymer behaves as a spring, and for small τ the spring constant is
  
∂τ  4kB T 2 ∆ε
k= = cosh . (4.226)
∂L τ =0 N (∆`)2 2kB T
The results are shown in fig. 4.9. Note that length increases with temperature for τ < τ ∗ and decreases
with temperature for τ > τ ∗ . Note also that k diverges at both low and high temperatures. At low T , the
energy gap ∆ε dominates and L = N `A , while at high temperatures kB T dominates and L = 12 N (`A +`B ).
4.9. SELECTED EXAMPLES 201

4.9.5 Noninteracting spin dimers

Consider a system of noninteracting spin dimers as depicted in fig. 4.10. Each dimer contains two spins,
and is described by the Hamiltonian

Ĥdimer = −J σ1 σ2 − µ0 H (σ1 + σ2 ) . (4.227)

Here, J is an interaction energy between the spins which comprise the dimer. If J > 0 the interaction is
ferromagnetic, which prefers that the spins are aligned. That is, the lowest energy states are | ↑↑ i and
| ↓↓ i. If J < 0 the interaction is antiferromagnetic, which prefers that spins be anti-aligned: | ↑↓ i and
| ↓↑ i.9
Suppose there are Nd dimers. Then the OCE partition function is Z = ζ Nd , where ζ(T, H) is the single
dimer partition function. To obtain ζ(T, H), we sum over the four possible states of the two spins,
obtaining

ζ = Tr e−Ĥdimer /kB T
 
−J/kB T J/kB T 2µ0 H
= 2e + 2e cosh .
kB T
Thus, the free energy is
"  #
2µ0 H
F (T, H, Nd ) = −Nd kB T ln 2 − Nd kB T ln e−J/kB T + eJ/kB T cosh . (4.228)
kB T

The magnetization is
 
2µ H

∂F
 eJ/kB T sinh k 0T
B
M =− = 2Nd µ0 ·   (4.229)
∂H T,Nd e −J/kB T J/k T
+ e B cosh k 0T
2µ H
B

It is instructive to consider the zero field isothermal susceptibility per spin,

µ20 2 eJ/kB T

1 ∂M 
χT = = · . (4.230)
2Nd ∂H H=0 kB T eJ/kB T + e−J/kB T

The quantity µ20 /kB T is simply the Curie susceptibility for noninteracting classical spins. Note that we
correctly recover the Curie result when J = 0, since then the individual spins comprising each dimer are
in fact noninteracting. For the ferromagnetic case, if J  kB T , then we obtain
2µ20
χT (J  kB T ) ≈ . (4.231)
kB T
This has the following simple interpretation. When J  kB T , the spins of each dimer are effectively
locked in parallel. Thus, each dimer has an effective magnetic moment µeff = 2µ0 . On the other hand,
there are only half as many dimers as there are spins, so the resulting Curie susceptibility per spin is
1 2
2 × (2µ0 ) /kB T .
9
Nota bene we are concerned with classical spin configurations only – there is no superposition of states allowed in this
model!
202 CHAPTER 4. STATISTICAL ENSEMBLES

Figure 4.10: A model of noninteracting spin dimers on a lattice. Each red dot represents a classical spin
for which σj = ±1.

When −J  kB T , the spins of each dimer are effectively locked in one of the two antiparallel configura-
tions. We then have
2µ20 −2|J|/k T
χT (−J  kB T ) ≈ e B . (4.232)
kB T
In this case, the individual dimers have essentially zero magnetic moment.

4.10 Statistical Mechanics of Molecular Gases

4.10.1 Separation of translational and internal degrees of freedom

The states of a noninteracting atom or molecule are labeled by its total momentum p and its internal
quantum numbers, which we will simply write with a collective index α, specifying rotational, vibrational,
and electronic degrees of freedom. The single particle Hamiltonian is then
p2
ĥ = + ĥint , (4.233)
2m
with
~2 k2
 
   
ĥ  k , α = + εα  k , α . (4.234)
2m
The partition function is
2 /2m
gj e−βεj .
X X
ζ = Tr e−β ĥ = e−βp (4.235)
p j
Here we have replaced the internal label α with a label j of energy eigenvalues, with gj being the
degeneracy of the internal state with energy εj . To do the p sum, we quantize in a box of dimensions
L1 × L2 × · · · × Ld , using periodic boundary conditions. Then
 
2π~n1 2π~n2 2π~nd
p= , , ... , , (4.236)
L1 L2 Ld
4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 203

where each ni is an integer. Since the differences between neighboring quantized p vectors are very tiny,
we can replace the sum over p by an integral:
ddp
X Z
−→ (4.237)
p
∆p1 · · · ∆pd

where the volume in momentum space of an elementary rectangle is

(2π~)d (2π~)d
∆p1 · · · ∆pd = = . (4.238)
L1 · · · Ld V
Thus,
ddp
Z
−p2 /2mkB T
gj e−εj /kB T = V λ−d
X
ζ=V e T ξ (4.239)
(2π~)d
j

gj e−εj /kB T .
X
ξ(T ) = (4.240)
j

Here, ξ(T ) is the internal coordinate partition function. The full N -particle ordinary canonical partition
function is then
1 V N N
 
ZN = ξ (T ) . (4.241)
N ! λdT
Using Stirling’s approximation, we find the Helmholtz free energy F = −kB T ln Z is
"   #
V
F (T, V, N ) = −N kB T ln + 1 + ln ξ(T )
N λdT
"   # (4.242)
V
= −N kB T ln + 1 + N ϕ(T ) ,
N λdT

where
ϕ(T ) = −kB T ln ξ(T ) (4.243)
is the internal coordinate contribution to the single particle free energy. We could also compute the
partition function in the Gibbs (T, p, N ) ensemble:
Z∞
−βG(T,p,N ) 1
Y (T, p, N ) = e = dV e−βpV Z(T, V, N )
V0
0 (4.244)
  N
kB T kB T
= ξ N (T ) .
pV0 p λdT
Thus, in the thermodynamic limit,
 d
G(T, p, N ) p λT
µ(T, p) = = kB T ln − kB T ln ξ(T )
N kB T
 d (4.245)
p λT
= kB T ln + ϕ(T ) .
kB T
204 CHAPTER 4. STATISTICAL ENSEMBLES

4.10.2 Ideal gas law

Since the internal coordinate contribution to the free energy is volume-independent, we have
 
∂G N kB T
V = = , (4.246)
∂p T,N p

and the ideal gas law applies. The entropy is

  "   #
∂G kB T
S=− = N kB ln + 1 + 21 d − N ϕ0 (T ) , (4.247)
∂T p,N pλdT

and therefore the heat capacity is

 
∂S 1
+ 1 N kB − N T ϕ00 (T )


Cp = T = 2d (4.248)
∂T p,N
 
∂S
CV = T = 21 dN kB − N T ϕ00 (T ) . (4.249)
∂T V,N

Thus, any temperature variation in Cp must be due to the internal degrees of freedom.

4.10.3 The internal coordinate partition function

At energy scales of interest we can separate the internal degrees of freedom into distinct classes, writing

ĥint = ĥrot + ĥvib + ĥelec (4.250)

as a sum over internal Hamiltonians governing rotational, vibrational, and electronic degrees of freedom.
Then
ξint = ξrot · ξvib · ξelec . (4.251)

Associated with each class of excitation is a characteristic temperature Θ. Rotational and vibrational
temperatures of a few common molecules are listed in table tab. 4.1.

4.10.4 Rotations

Consider a class of molecules which can be approximated as an axisymmetric top. The rotational Hamil-
tonian is then

L2a + L2b L2
ĥrot = + c
2I1 2I3
2
  (4.252)
~ L(L + 1) 1 1
= + − L2c ,
2I1 2I3 2I1
4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 205

molecule Θrot (K) Θvib (K)

H2 85.4 6100
N2 2.86 3340
H2 O 13.7 , 21.0 , 39.4 2290 , 5180 , 5400

Table 4.1: Some rotational and vibrational temperatures of common molecules.

where n̂a.b,c (t) are the principal axes, with n̂c the symmetry axis, and La,b,c are the components of the
angular momentum vector L about these instantaneous body-fixed principal axes. The components of L
along space-fixed axes {x, y, z} are written as Lx,y,z . Note that
 µ
L , Lc = nνc Lµ , Lν + Lµ , nνc Lν = iµνλ nνc Lλ + iµνλ nλc Lν = 0 ,
    
(4.253)
which is equivalent to the statement that Lc = n̂c ·L is a rotational scalar. We can therefore simultaneously
specify the eigenvalues of {L2 , Lz , Lc }, which form a complete set of commuting observables (CSCO)10 .
The eigenvalues of Lz are m~ with m ∈ {−L, . . . , L}, while those of Lc are k~ with k ∈ {−L, . . . , L}.
There is a (2L + 1)-fold degeneracy associated with the Lz quantum number.
We assume the molecule is prolate, so that I3 < I1 . We can the define two temperature scales,
~2 ~2
Θ= , Θ
e= . (4.254)
2I1 kB 2I3 kB

Prolateness then means Θ

e > Θ. We conclude that the rotational partition function for an axisymmetric
molecule is given by
∞ L
2 e
X X
−L(L+1) Θ/T
ξrot (T ) = (2L + 1) e e−k (Θ−Θ)/T (4.255)
L=0 k=−L

In diatomic molecules, I3 is extremely small, and Θ

e  k T at all relevant temperatures. Only the k = 0
B
term contributes to the partition sum, and we have
∞
X
ξrot (T ) = (2L + 1) e−L(L+1) Θ/T . (4.256)
L=0

When T  Θ, only the first few terms contribute, and

ξrot (T ) = 1 + 3 e−2Θ/T + 5 e−6Θ/T + . . . (4.257)
In the high temperature limit, we have a slowly varying summand. The Euler-MacLaurin summation
formula may be used to evaluate such a series:
n Zn ∞
X
1
  X B2j h (2j−1) (2j−1)
i
Fk = dk F (k) + 2 F (0) + F (n) + F (n) − F (0) (4.258)
(2j)!
k=0 0 j=1

10
Note that while we cannot simultaneously specify the eigenvalues of two components of L along axes fixed in space,
we can simultaneously specify the components of L along one axis fixed in space and one axis rotating with a body. See
Landau and Lifshitz, Quantum Mechanics, §103.
206 CHAPTER 4. STATISTICAL ENSEMBLES

where Bj is the j th Bernoulli number where

B0 = 1 , B1 = − 21 , B2 = 1
6 , 1
B4 = − 30 , B6 = 1
42 . (4.259)

Thus,
∞ Z∞
X
1 0 1
Fk = dx F (x) + 21 F (0) − 12 F (0) − F 000 (0) + . . . . (4.260)
720
k=0 0
R∞
We have F (x) = (2x + 1) e−x(x+1)Θ/T , for which dx F (x) = T
Θ , hence
0
 2
T 1 1 Θ 4 Θ
ξrot = + + + + ... . (4.261)
Θ 3 15 T 315 T

Recall that ϕ(T ) = −kB T ln ξ(T ). We conclude that ϕrot (T ) ≈ −3kB T e−2Θ/T for T  Θ and ϕrot (T ) ≈
−kB T ln(T /Θ) for T  Θ. We have seen that the internal coordinate contribution to the heat capacity
is ∆CV = −N T ϕ00 (T ). For diatomic molecules, then, this contribution is exponentially suppressed for
T  Θ, while for high temperatures we have ∆CV = N kB . One says that the rotational excitations are
‘frozen out’ at temperatures much below Θ. Including the first few terms, we have
 2
Θ
∆CV (T  Θ) = 12 N kB e−2Θ/T + . . . (4.262)
T
( )
1 Θ 2 16 Θ 3
   
∆CV (T  Θ) = N kB 1 + + + ... . (4.263)
45 T 945 T

Note that CV overshoots its limiting value of N kB and asymptotically approaches it from above.
Special care must be taken in the case of homonuclear diatomic molecules, for then only even or odd L
states are allowed, depending on the total nuclear spin. This is discussed below in §4.10.7.
For polyatomic molecules, the moments of inertia generally are large enough that the molecule’s rotations
can be considered classically. We then have

L2a L2 L2
ε(La , Lb , Lc ) = + b + c . (4.264)
2I1 2I2 2I3

We then have
dLa dLb dLc dφ dθ dψ −ε(La Lb Lc )/k T
Z
1
ξrot (T ) = e B , (4.265)
grot (2π~)3
where (φ, θ ψ) are the Euler angles. Recall φ ∈ [0, 2π], θ ∈ [0, π], and ψ ∈ [0, 2π]. The factor grot
accounts for physically indistinguishable orientations of the molecule brought about by rotations, which
can happen when more than one of the nuclei is the same. We then have
 3/2
2kB T p
ξrot (T ) = πI1 I2 I3 . (4.266)
~2

This leads to ∆CV = 23 N kB .

4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 207

4.10.5 Vibrations

Vibrational frequencies are often given in units of inverse wavelength, such as cm−1 , called a wavenumber .
To convert to a temperature scale T ∗ , we write kB T ∗ = hν = hc/λ, hence T ∗ = (hc/kB ) λ−1 , and we
multiply by
hc
= 1.436 K · cm . (4.267)
kB
For example, infrared absorption (∼ 50 cm−1 to 104 cm−1 ) reveals that the ‘asymmetric stretch’ mode of
the H2 O molecule has a vibrational frequency of ν = 3756 cm−1 . The corresponding temperature scale is
T ∗ = 5394 K.
Vibrations are normal modes of oscillations. A single normal mode Hamiltonian is of the form

p2
+ 21 mω 2 q 2 = ~ω a† a + 12 .


ĥ = (4.268)
2m
In general there are many vibrational modes, hence many normal mode frequencies ωα . We then must
sum over all of them, resulting in Y (α)
ξvib = ξvib . (4.269)
α
For each such normal mode, the contribution is
∞ ∞  n
X 1 X
ξ= e−(n+ 2 )~ω/kB T = e−~ω/2kB T e−~ω/kB T
n=0 n=0 (4.270)
e−~ω/2kB T 1
= = ,
1 − e−~ω/kB T 2 sinh(Θ/2T )
where Θ = ~ω/kB . Then
 
ϕ = kB T ln 2 sinh(Θ/2T )
(4.271)
= 21 kB Θ + kB T ln 1 − e−Θ/T .

The contribution to the heat capacity is

2
eΘ/T

Θ
∆CV = N kB
(eΘ/T − 1)2
T
( (4.272)
N kB (Θ/T )2 exp(−Θ/T ) (T → 0)
=
N kB (T → ∞)

4.10.6 Two-level systems : Schottky anomaly

Consider now a two-level system, with energies ε0 and ε1 . We define ∆ ≡ ε1 − ε0 and assume without
loss of generality that ∆ > 0. The partition function is

ζ = e−βε0 + e−βε1 = e−βε0 1 + e−β∆ .

(4.273)
208 CHAPTER 4. STATISTICAL ENSEMBLES

Figure 4.11: Heat capacity per molecule as a function of temperature for (a) heteronuclear diatomic
gases, (b) a single vibrational mode, and (c) a single two-level system.

The free energy is

f = −kB T ln ζ = ε0 − kB T ln 1 + e−∆/kB T .


(4.274)
The entropy for a given two level system is then
∂f
∆ 1
s=− = kB ln 1 + e−∆/kB T + · ∆/k T (4.275)
∂T T e B +1
and the heat capacity is = T (∂s/∂T ), i.e.

∆2 e∆/kB T
c(T ) = ·
. (4.276)
kB T 2 e∆/kB T + 1 2

Thus,

∆2 −∆/k T
c (T  ∆) = e B (4.277)
kB T 2
∆2
c (T  ∆) = . (4.278)
4kB T 2
We find that c(T ) has a characteristic peak at T ∗ ≈ 0.42 ∆/kB . The heat capacity vanishes in both the
low temperature and high temperature limits. At low temperatures, the gap to the excited state is much
greater than kB T , and it is not possible to populate it and store energy. At high temperatures, both
ground state and excited state are equally populated, and once again there is no way to store energy.
If we have a distribution of independent two-level systems, the heat capacity of such a system is a sum
over the individual Schottky functions:

X Z∞
C(T ) = c (∆i /kB T ) = N d∆ P (∆) e
e c(∆/T ) , (4.279)
i 0
4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 209

where N is the number of two level systems, e c(x) = kB x2 ex /(ex + 1)2 , and where P (∆) is the normalized
distribution function, which satisfies the normalization condition
Z∞
d∆ P (∆) = 1 . (4.280)
0

NS is the total number of two level systems. If P (∆) ∝ ∆r for ∆ → 0, then the low temperature heat
capacity behaves as C(T ) ∝ T 1+r . Many amorphous or glassy systems contain such a distribution of two
level systems, with r ≈ 0 for glasses, leading to a linear low-temperature heat capacity. The origin of
these two-level systems is not always so clear but is generally believed to be associated with local atomic
configurations for which there are two low-lying states which are close in energy. The paradigmatic
example is the mixed crystalline solid (KBr)1−x (KCN)x which over the range 0.1 < ∼x< ∼ 0.6 forms an
‘orientational glass’ at low temperatures. The two level systems are associated with different orientation
of the cyanide (CN) dipoles.

4.10.7 Electronic and nuclear excitations

For a monatomic gas, the internal coordinate partition function arises due to electronic and nuclear
degrees of freedom. Let’s first consider the electronic degrees of freedom. We assume that kB T is small
compared with energy differences between successive electronic shells. The atomic ground state is then
computed by filling up the hydrogenic orbitals until all the electrons are used up. If the atomic number
is a ‘magic number’ (A = 2 (He), 10 (Ne), 18 (Ar), 36 (Kr), 54 (Xe), etc.) then the atom has all shells
filled and L = 0 and S = 0. Otherwise the last shell is partially filled and one or both of L and S will be
nonzero. The atomic ground state configuration 2J+1 LS is then determined by Hund’s rules:

1. The LS multiplet with the largest S has the lowest energy.

2. If the largest value of S is associated with several multiplets, the multiplet with the largest L has
the lowest energy.
3. If an incomplete shell is not more than half-filled, then the lowest energy state has J = |L − S|. If
the shell is more than half-filled, then J = L + S.

The last of Hund’s rules distinguishes between the (2S + 1)(2L + 1) states which result upon fixing S and
L as per rules #1 and #2. It arises due to the atomic spin-orbit coupling, whose effective Hamiltonian
may be written Ĥ = ΛL · S, where Λ is the Russell-Saunders coupling. If the last shell is less than or
equal to half-filled, then Λ > 0 and the ground state has J = |L − S|. If the last shell is more than
half-filled, the coupling is inverted , i.e. Λ < 0, and the ground state has J = L + S.11
The electronic contribution to ξ is then
L+S
X
ξelec = (2J + 1) e−∆ε(L,S,J)/kB T (4.281)
J=|L−S|

11
See e.g. §72 of Landau and Lifshitz, Quantum Mechanics, which, in my humble estimation, is the greatest physics book
ever written.
210 CHAPTER 4. STATISTICAL ENSEMBLES

where h i
∆ε(L, S, J) = 12 Λ J(J + 1) − L(L + 1) − S(S + 1) . (4.282)

At high temperatures, kB T is larger than the energy difference between the different J multiplets, and we
have ξelec ∼ (2L+1)(2S +1) e−βε0 , where ε0 is the ground state energy. At low temperatures, a particular
value of J is selected – that determined by Hund’s third rule – and we have ξelec ∼ (2J + 1) e−βε0 . If, in
addition, there is a nonzero nuclear spin I, then we also must include a factor ξnuc = (2I + 1), neglecting
the small hyperfine splittings due to the coupling of nuclear and electronic angular momenta.
For heteronuclear diatomic molecules, i.e. molecules composed from two different atomic nuclei, the
(1) (2)
internal partition function simply receives a factor of ξelec · ξnuc · ξnuc , where the first term is a sum over
molecular electronic states, and the second two terms arise from the spin degeneracies of the two nuclei.
For homonuclear diatomic molecules, the exchange of nuclear centers is a symmetry operation, and does
not represent a distinct quantum state. To correctly count the electronic states, we first assume that the
total electronic spin is S = 0. This is generally a very safe assumption. Exchange symmetry now puts
restrictions on the possible values of the molecular angular momentum L, depending on the total nuclear
angular momentum Itot . If Itot is even, then the molecular angular momentum L must also be even.
If the total nuclear angular momentum is odd, then L must be odd. This is so because the molecular
ground state configuration is 1Σg+ .12
The total number of nuclear states for the molecule is (2I + 1)2 , of which some are even under nuclear
exchange, and some are odd. The number of even states, corresponding to even total nuclear angular
momentum is written as gg , where the subscript conventionally stands for the (mercifully short) German
word gerade, meaning ‘even’. The number of odd (Ger. ungerade) states is written gu . Table 4.2 gives
the values of gg,u corresponding to half-odd-integer I and integer I.
The final answer for the rotational component of the internal molecular partition function is then

ξrot (T ) = gg ζg + gu ζu , (4.283)

where

X
ζg = (2L + 1) e−L(L+1) Θrot /T
L even
X (4.284)
−L(L+1) Θrot /T
ζu = (2L + 1) e .
L odd

For hydrogen, the molecules with the larger nuclear statistical weight are called orthohydrogen and those
with the smaller statistical weight are called parahydrogen. For H2 , we have I = 21 hence the ortho state
has gu = 3 and the para state has gg = 1. In D2 , we have I = 1 and the ortho state has gg = 6 while the
para state has gu = 3. In equilibrium, the ratio of ortho to para states is then

NHortho g ζ 3ζ NDortho gg ζg 2 ζg
2
= u u = u , 2
= = . (4.285)
NHpara gg ζg ζg NDpara gu ζu ζu
2 2

12
See Landau and Lifshitz, Quantum Mechanics, §86.
4.11. APPENDIX I : ADDITIONAL EXAMPLES 211

2I gg gu
odd I(2I + 1) (I + 1)(2I + 1)
even (I + 1)(2I + 1) I(2I + 1)

Table 4.2: Number of even (gg ) and odd (gu ) total nuclear angular momentum states for a homonuclear
diatomic molecule. I is the ground state nuclear spin.

Incidentally, how do we derive the results in Tab. ?? ? The total nuclear angular momentum Itot is the
quantum mechanical sum of the two individual nuclear angular momenta, each of which are of magnitude
I. From elementary addition of angular momenta, we have

I ⊗ I = 0 ⊕ 1 ⊕ 2 ⊕ · · · ⊕ 2I . (4.286)

The right hand side of the above equation lists all the possible multiplets. Thus, Itot ∈ {0, 1, . . . , 2I}.
Now let us count the total number of states with even Itot . If 2I is even, which is to say if I is an integer,
we have
I n
X o
gg(2I=even) = 2 · (2n) + 1 = (I + 1)(2I + 1) , (4.287)
n=0

because the degeneracy of each multiplet is 2Itot + 1. It follows that

gu(2I=even) = (2I + 1)2 − gg = I(2I + 1) . (4.288)

On the other hand, if 2I is odd, which is to say I is a half odd integer, then

I− 21
Xn o
gg(2I=odd) = 2 · (2n) + 1 = I(2I + 1) . (4.289)
n=0

It follows that
gu(2I=odd) = (2I + 1)2 − gg = (I + 1)(2I + 1) . (4.290)

4.11 Appendix I : Additional Examples

4.11.1 Three state system

Consider a spin-1 particle where σ = −1, 0, +1. We model this with the single particle Hamiltonian

ĥ = −µ0 H σ + ∆(1 − σ 2 ) . (4.291)

We can also interpret this as describing a spin if σ = ±1 and a vacancy if σ = 0. The parameter ∆ then
represents the vacancy formation energy. The single particle partition function is

ζ = Tr e−β ĥ = e−β∆ + 2 cosh(βµ0 H) . (4.292)

212 CHAPTER 4. STATISTICAL ENSEMBLES

With NS distinguishable noninteracting spins (e.g. at different sites in a crystalline lattice), we have
Z = ζ NS and h i
F ≡ NSf = −kB T ln Z = −NS kB T ln e−β∆ + 2 cosh(βµ0 H) , (4.293)

where f = −kB T ln ζ is the free energy of a single particle. Note that

∂ ĥ
n̂V = 1 − σ 2 = (4.294)
∂∆
∂ ĥ
m̂ = µ0 σ = − (4.295)
∂H
are the vacancy number and magnetization, respectively. Thus,

 ∂f e−∆/kB T
nV = n̂V = = −∆/k T (4.296)
∂∆ e B + 2 cosh(µ0 H/kB T )

and
∂f 2µ sinh(µ0 H/kB T )
= −∆/k 0T


m = m̂ = − . (4.297)
∂H e B + 2 cosh(µ0 H/kB T )
At weak fields we can compute

µ20

∂m  2
χT = = · . (4.298)
∂H H=0 kB T 2 + e
 −∆/k BT

We thus obtain a modified Curie law. At temperatures T  ∆/kB , the vacancies are frozen out and we
recover the usual Curie behavior. At high temperatures, where T  ∆/kB , the low temperature result
is reduced by a factor of 32 , which accounts for the fact that one third of the time the particle is in a
nonmagnetic state with σ = 0.

4.11.2 Spins and vacancies on a surface

PROBLEM: A collection of spin- 21 particles is confined to a surface with N sites. For each site, let σ = 0 if
there is a vacancy, σ = +1 if there is particle present with spin up, and σ = −1 if there is a particle present
with spin down. The particles are non-interacting, and the energy for each site is given by ε = −W σ 2 ,
where −W < 0 is the binding energy.

(a) Let Q = N↑ + N↓ be the number of spins, and N0 be the number of vacancies. The surface
magnetization is M = N↑ − N↓ . Compute, in the microcanonical ensemble, the statistical entropy
S(Q, M ).

(b) Let q = Q/N and m = M/N be the dimensionless particle density and magnetization density,
respectively. Assuming that we are in the thermodynamic limit, where N , Q, and M all tend to
infinity, but with q and m finite, Find the temperature T (q, m). Recall Stirling’s formula

ln(N !) = N ln N − N + O(ln N ) .
4.11. APPENDIX I : ADDITIONAL EXAMPLES 213

(c) Show explicitly that T can be negative for this system. What does negative T mean? What physical
degrees of freedom have been left out that would avoid this strange property?

SOLUTION: There is a constraint on N↑ , N0 , and N↓ :

N↑ + N0 + N↓ = Q + N0 = N . (4.299)

The total energy of the system is E = −W Q.

(a) The number of states available to the system is

N!
Ω= . (4.300)
N↑ ! N0 ! N↓ !

Fixing Q and M , along with the above constraint, is enough to completely determine {N↑ , N0 , N↓ }:
1 1
N↑ = 2 (Q + M ) , N0 = N − Q , N↓ = 2 (Q − M ) , (4.301)

whence
N!
Ω(Q, M ) =  1  1  . (4.302)
2 (Q + M ) ! 2 (Q − M ) ! (N − Q)!
The statistical entropy is S = kB ln Ω:

S(Q, M ) = kB ln(N !) − kB ln 21 (Q + M )! − kB ln 12 (Q − M )! − kB ln (N − Q)! .

     
(4.303)

(b) Now we invoke Stirling’s rule,

ln(N !) = N ln N − N + O(ln N ) , (4.304)

to obtain
ln Ω(Q, M ) = N ln N − N − 21 (Q + M ) ln
1
+ M ) + 21 (Q + M )

2 (Q

− 12 (Q − M ) ln
1
− M ) + 12 (Q − M )

2 (Q
(4.305)
− (N − Q) ln(N − Q) + (N − Q)
 
1
h
1 2 2
i
1 Q+M
= N ln N − 2 Q ln 4 (Q − M ) − 2 M ln
Q−M
Combining terms,
 
h p
1
i
1 q+m
ln Ω(Q, M ) = −N q ln 2 q2 − m2 − 2 N m ln − N (1 − q) ln(1 − q) , (4.306)
q−m
where Q = N q and M = N m. Note that the entropy S = kB ln Ω is extensive. The statistical
entropy per site is thus
 
h p
1 2 2
i
1 q+m
s(q, m) = −kB q ln 2 q − m − 2 kB m ln − kB (1 − q) ln(1 − q) . (4.307)
q−m
214 CHAPTER 4. STATISTICAL ENSEMBLES

The temperature is obtained from the relation

   
1 ∂S 1 ∂s
= =
T ∂E M W ∂q m
(4.308)
1 1 h p i
= ln(1 − q) − ln 12 q 2 − m2 .
W W

Thus,
W/kB
T =  p  . (4.309)
ln 2(1 − q)/ q 2 − m2

(c) We have 0 ≤ q ≤ 1 and −q ≤ m ≤ q, so T is real (thank heavens!). But it is easy to choose {q, m}
such that T < 0. For example, when m = 0 we have T = W/kB ln(2q −1 − 2) and T < 0 for all
q ∈ 23 , 1 . The reason for this strange state of affairs is that the entropy S is bounded, and is not an
monotonically increasing function of the energy E (or the dimensionless quantity Q). The entropy
is maximized for N ↑= N0 = N↓ = 31 , which says m = 0 and q = 23 . Increasing q beyond this
point (with m = 0 fixed) starts to reduce the entropy, and hence (∂S/∂E) < 0 in this range, which
immediately gives T < 0. What we’ve left out are kinetic degrees of freedom, such as vibrations
and rotations, whose energies are unbounded, and which result in an increasing S(E) function.

4.11.3 Fluctuating interface

Consider an interface between two dissimilar fluids. In equilibrium, in a uniform gravitational field, the
denser fluid is on the bottom. Let z = z(x, y) be the height the interface between the fluids, relative to
equilibrium. The potential energy is a sum of gravitational and surface tension terms, with
Z Zz
Ugrav = d2x dz 0 ∆ρ g z 0 (4.310)
0
Z
Usurf = d2 x 21 σ (∇z)2 . (4.311)

We won’t need the kinetic energy in our calculations, but we can include it just for completeness. It isn’t
so clear how to model it a priori so we will assume a rather general form
∂z(x, t) ∂z(x0 , t)
Z Z
T = d x d2x0 21 µ(x, x0 )
2
. (4.312)
∂t ∂t
We assume that the (x, y) plane is a rectangle of dimensions Lx × Ly . We also assume µ(x, x0 ) =
µ |x − x0 | . We can then Fourier transform

−1/2 X
z(x) = Lx Ly zk eik·x , (4.313)
k

where the wavevectors k are quantized according to

2πnx 2πny
k= x̂ + ŷ , (4.314)
Lx Ly
4.11. APPENDIX I : ADDITIONAL EXAMPLES 215

with integer nx and ny , if we impose periodic boundary conditions (for calculational convenience). The
Lagrangian is then
1 X h  2
 2 i
L= µk żk − g ∆ρ + σk2 zk  , (4.315)
2
k
where Z
d2x µ |x| e−ik·x .


µk = (4.316)

Since z(x, t) is real, we have the relation z−k = zk∗ , therefore the Fourier coefficients at k and −k are not
independent. The canonical momenta are given by
∂L ∂L
pk = = µk żk , p∗k = = µk żk∗ (4.317)
∂ żk∗ ∂ żk
The Hamiltonian is then
X0 h i
Ĥ = pk zk∗ + p∗k zk − L (4.318)
k
X0  |p |2 
k 2 2


= + g ∆ρ + σk |zk | , (4.319)
µk
k

where the prime on the k sum indicates that only one of the pair {k, −k} is to be included, for each k.
We may now compute the ordinary canonical partition function:
Y0 Z d2p d2z
k k −|pk |2 /µk kB T −(g ∆ρ+σk2 ) |zk |2 /kB T
Z= 2
e e
(2π~)
k
(4.320)
Y0  k T 2  µk

B
= .
2~ g ∆ρ + σk2
k

Thus,  
X kB T
F = −kB T ln , (4.321)
2~Ωk
k
where13 1/2
g ∆ρ + σk2

Ωk = . (4.322)
µk
is the normal mode frequency
 for surface oscillations at wavevector k. For deep water waves, it is
appropriate to take µk = ∆ρ |k|, where ∆ρ = ρL − ρG ≈ ρL is the difference between the densities of
water and air.
It is now easy to compute the thermal average
Z Z
2 −(g ∆ρ+σk2 ) |zk |2 /kB T 2 2
2 2
d2zk e−(g ∆ρ+σk ) |zk | /kB T


|zk | = d zk |zk | e
(4.323)
kB T
= .
g ∆ρ + σk2
13
Note that there is no prime on the k sum for F , as we have divided the logarithm of Z by two and replaced the half
sum by the whole sum.
216 CHAPTER 4. STATISTICAL ENSEMBLES

Note that this result does not depend on µk , i.e. on our choice of kinetic energy. One defines the
correlation function
Z 2  
1 X
2
 ik·x dk kB T
eik·x


C(x) ≡ z(x) z(0) = |zk | e =
Lx Ly (2π)2 g ∆ρ + σk2
k
Z∞ (4.324)
kB T eik|x| kB T

= dq p = K |x|/ξ ,
4πσ q2 + ξ2 4πσ 0
0
p
where ξ = g ∆ρ/σ is the correlation length, and where K0 (z) is the Bessel function of imaginary
argument. The asymptotic behavior of K0 (z) for small z is K0 (z) ∼ ln(2/z), whereas for large z one has
K0 (z) ∼ (π/2z)1/2 e−z . We see that on large length scales the correlations decay exponentially, but on
small length scales they diverge. This divergence is due to the improper energetics we have assigned to
short wavelength fluctuations of the interface. Roughly, it can cured by imposing a cutoff on the integral,
or by insisting that the shortest distance scale is a molecular diameter.

4.11.4 Dissociation of molecular hydrogen

Consider the reaction

)− p+ + e− .
H −* (4.325)
In equilibrium, we have
µH = µ p + µe . (4.326)
What is the relationship between the temperature T and the fraction x of hydrogen which is dissociated?
Let us assume a fraction x of the hydrogen is dissociated. Then the densities of H, p, and e are then

nH = (1 − x) n , np = xn , ne = xn . (4.327)

The single particle partition function for each species is

g N V N −N ε /k T
 
ζ= e int B , (4.328)
N ! λ3T
where g is the degeneracy and εint the internal energy for a given species. We have εint = 0 for p and e,
and εint = −∆ for H, where ∆ = e2 /2aB = 13.6 eV, the binding energy of hydrogen. Neglecting hyperfine
splittings14 , we have gH = 4, while ge = gp = 2 because each has spin S = 21 . Thus, the associated grand
potentials are

ΩH = −gH V kB T λ−3
T,H e
(µH +∆)/kB T
(4.329)

Ωp = −gp V kB T λ−3
T,p e
µp /kB T
(4.330)

Ωe = −ge V kB T λ−3
T,e e
µe /kB T
, (4.331)

The hyperfine splitting in hydrogen is on the order of (me /mp ) α4 me c2 ∼ 10−6 eV, which is on the order of 0.01 K. Here
14

α = e2 /~c is the fine structure constant.

4.11. APPENDIX I : ADDITIONAL EXAMPLES 217

where s
2π~2
λT,a = (4.332)
ma kB T
for species a. The corresponding number densities are
 
1 ∂Ω
n= = g λ−3
T e
(µ−εint )/kB T
, (4.333)
V ∂µ T,V

and the fugacity z = eµ/kB T of a given species is given by

z = g −1 n λ3T eεint /kB T . (4.334)

We now invoke µH = µp + µe , which says zH = zp ze , or

−1
nH λ3T,H e−∆/kB T = gp−1 np λ3T,p ge−1 ne λ3T,e ,



gH (4.335)

which yields
x2
 
nλ̃3T = e−∆/kB T , (4.336)
1−x
2π~2 /m∗ kB T , with m∗ = mp me /mH ≈ me . Note that
p
where λ̃T =
s s
4πmH ∆
λ̃T = aB , (4.337)
mp kB T

where aB = 0.529 Å is the Bohr radius. Thus, we have

3/2
x2
  
T
· (4π) 3/2
ν= e−T0 /T , (4.338)
1−x T0

where T0 = ∆/kB = 1.578 × 105 K and ν = na3B . Consider for example a temperature T = 3000 K, for
which T0 /T = 52.6, and assume that x = 21 . We then find ν = 1.69 × 10−27 , corresponding to a density
of n = 1.14 × 10−2 cm−3 . At this temperature, the fraction of hydrogen molecules in their first excited
(2s) state is x0 ∼ e−T0 /2T = 3.8 × 10−12 . This is quite striking: half the hydrogen atoms are completely
dissociated, which requires an energy of ∆, yet the number in their first excited state, requiring energy
1
2 ∆, is twelve orders of magnitude smaller. The student should reflect on why this can be the case.
218 CHAPTER 4. STATISTICAL ENSEMBLES
Chapter 5

Noninteracting Quantum Systems

5.1 References

– F. Reif, Fundamentals of Statistical and Thermal Physics (McGraw-Hill, 1987)

This has been perhaps the most popular undergraduate text since it first appeared in 1967, and
with good reason.

– A. H. Carter, Classical and Statistical Thermodynamics

(Benjamin Cummings, 2000)
A very relaxed treatment appropriate for undergraduate physics majors.

– D. V. Schroeder, An Introduction to Thermal Physics (Addison-Wesley, 2000)

This is the best undergraduate thermodynamics book I’ve come across, but only 40% of the book
treats statistical mechanics.

– C. Kittel, Elementary Statistical Physics (Dover, 2004)

Remarkably crisp, though dated, this text is organized as a series of brief discussions of key concepts
and examples. Published by Dover, so you can’t beat the price.

– R. K. Pathria, Statistical Mechanics (2nd edition, Butterworth-Heinemann, 1996)

This popular graduate level text contains many detailed derivations which are helpful for the stu-
dent.

– M. Plischke and B. Bergersen, Equilibrium Statistical Physics (3rd edition, World Scientific, 2006)
An excellent graduate level text. Less insightful than Kardar but still a good modern treatment of
the subject. Good discussion of mean field theory.

– E. M. Lifshitz and L. P. Pitaevskii, Statistical Physics (part I, 3rd edition, Pergamon, 1980)
This is volume 5 in the famous Landau and Lifshitz Course of Theoretical Physics. Though dated,
it still contains a wealth of information and physical insight.

219
220 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

5.2 Statistical Mechanics of Noninteracting Quantum Systems

5.2.1 Bose and Fermi systems in the grand canonical ensemble

A noninteracting many-particle quantum Hamiltonian may be written as1

X
Ĥ = εα n̂α , (5.1)
α

where n̂α is the number of particles in the quantum state α with energy εα . This form is called the
second quantized representation of the Hamiltonian. The number eigenbasis is therefore also an energy
eigenbasis. Any eigenstate of Ĥ may be labeled by the integer eigenvalues of the n̂α number operators,
and written as  n1 , n2 , . . . . We then have
   
n̂α  ~n = nα  ~n (5.2)
and   X  
Ĥ  ~n = nα εα  ~n . (5.3)
α
The eigenvalues nα take on different possible values depending on whether the constituent particles are
bosons or fermions, viz.

bosons : nα ∈ 0 , 1 , 2 , 3 , . . .
 (5.4)
fermions : nα ∈ 0 , 1 .
In other words, for bosons, the occupation numbers are nonnegative integers. For fermions, the occupation
numbers are either 0 or 1 due to the Pauli principle, which says that at most one fermion can occupy
any single particle quantum state. There is no Pauli principle for bosons.
The N -particle partition function ZN is then
X P
ZN = e−β α nα εα δN , P nα , (5.5)
α
{nα }

where the sum is over all allowed values of the set {nα }, which depends on the statistics of the particles.
Bosons satisfy Bose-Einstein (BE) statistics, in which nα ∈ {0 , 1 , 2 , . . .}. Fermions satisfy Fermi-Dirac
(FD) statistics, in which nα ∈ {0 , 1}.
P
The OCE partition sum is difficult to perform, owing to the constraint α nα = N on the total number
of particles. This constraint is relaxed in the GCE, where
X
Ξ= eβµN ZN
N
X P P
= e−β α nα εα βµ
e α nα

{nα } (5.6)
!
Y X
= e−β(εα −µ) nα .
α nα
1
For a review of the formalism of second quantization, see the appendix in §5.9.
5.2. STATISTICAL MECHANICS OF NONINTERACTING QUANTUM SYSTEMS 221

Note that the grand partition function Ξ takes the form of a product over contributions from the indi-
vidual single particle states.
We now perform the single particle sums:
∞
X 1
e−β(ε−µ) n = (bosons) (5.7)
n=0
1− e−β(ε−µ)
X1
e−β(ε−µ) n = 1 + e−β(ε−µ) (fermions) . (5.8)
n=0

Therefore we have
Y 1
ΞBE =
1−
α e−(εα −µ)/kB T
X   (5.9)
ΩBE = kB T ln 1 − e−(εα −µ)/kB T
α

and
Y 
ΞFD = 1 + e−(εα −µ)/kB T
α
X   (5.10)
ΩFD = −kB T ln 1 + e−(εα −µ)/kB T .
α

We can combine these expressions into one, writing

X  
Ω(T, V, µ) = ±kB T ln 1 ∓ e−(εα −µ)/kB T , (5.11)
α

where we take the upper sign for Bose-Einstein statistics and the lower sign for Fermi-Dirac statistics.
Note that the average occupancy of single particle state α is
∂Ω 1
hn̂α i = = (ε −µ)/k T , (5.12)
∂εα e α B ∓1
and the total particle number is then
X 1
N (T, V, µ) = . (5.13)
α e(εα −µ)/kB T ∓ 1

We will henceforth write nα (µ, T ) = hn̂α i for the thermodynamic average of this occupancy.

5.2.2 Quantum statistics and the Maxwell-Boltzmann limit

Consider a system composed of N noninteracting particles. The Hamiltonian is

N
X
Ĥ = ĥj . (5.14)
j=1
222 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

The single particle Hamiltonian ĥ has eigenstates |αi with corresponding energy eigenvalues εα . What is
the partition function? Is it


?
X X −β ε + ε + ... + ε
Z = ··· e α1 α2 α
N = ζN , (5.15)
α1 αN

where ζ is the single particle partition function,

X
ζ= e−βεα . (5.16)
α

For systems where the individual particles are distinguishable, such as spins on a lattice which have fixed
positions, this is indeed correct. But for particles free to move in a gas, this equation is wrong. The
reason is that for indistinguishable particles the many particle quantum mechanical states are specified
by a collection of occupation numbers nα , which tell us how many particles are in the single-particle state
| α i. The energy is X
E= n α εα (5.17)
α

and the total number of particles is X

N= nα . (5.18)
α
 
That is, each collection of occupation numbers {nα } labels a unique many particle state  {nα } . In the
product ζ N , the collection {nα } occurs many times. We have therefore overcounted the contribution to
ZN due to this state. By what factor have we overcounted? It is easy to see that the overcounting factor
is
N!
degree of overcounting = Q ,
α nα !

which is the number of ways we can rearrange the labels αj to arrive at the same collection {nα }. This
follows from the multinomial theorem,

K
!N
X XX X N! n n n
xα = ··· x 1 x 2 · · · xKK δN,n1 + ... + n . (5.19)
n1 ! n2 ! · · · nK ! 1 2 K
α=1 n1 n2 nK

Thus, the correct expression for ZN is

X P
ZN = e−β α nα εα
δN,P nα
α
{nα }
X  Q n !  −β(ε + (5.20)
XX
α α α1 εα + ... + εα )
= ··· e 2 N .
N!
α1 α2 αN

In the high temperature limit, almost all the nα are either 0 or 1, hence

ζN
ZN ≈ . (5.21)
N!
5.2. STATISTICAL MECHANICS OF NONINTERACTING QUANTUM SYSTEMS 223

This is the classical Maxwell-Boltzmann limit of quantum statistical mechanics. We now see the origin
of the 1/N ! term which is so important in the thermodynamics of entropy of mixing.
Finally, starting with the expressions for the grand partition function for Bose-Einstein or Fermi-Dirac
particles, and working in the low density limit where nα (µ, T )  1 , we have εα − µ  kB T , and
consequently
X  
ΩBE/FD = ±kB T ln 1 ∓ e−(εα −µ)/kB T
α
X (5.22)
−→ −kB T e−(εα −µ)/kB T ≡ ΩMB .
α

This is the Maxwell-Boltzmann limit of quantum statistical mechanics. The occupation number average
in the Maxwell-Boltzmann limit is then

hn̂α i = e−(εα −µ)/kB T . (5.23)

5.2.3 Single particle density of states

The single particle density of states per unit volume g(ε) is defined as
1 X
g(ε) = δ(ε − εα ) . (5.24)
V α

We can then write

Z∞  
Ω(T, V, µ) = ±V kB T dε g(ε) ln 1 ∓ e−(ε−µ)/kB T . (5.25)
−∞

For particles with a dispersion ε(k), with p = ~k, we have

Z d
dk

g(ε) = g d
δ(ε − ε(k)
(2π)
(5.26)
g Ωd k d−1
= .
(2π)d dε/dk
where g = 2S +1 is the spin degeneracy, and where we assume that ε(k) is both isotropic and a mono-
tonically increasing function of k. Thus, we have
g
dk

 π dε d=1



g Ωd k d−1


g dk
g(ε) = = 2π k dε d=2 (5.27)
(2π)d dε/dk  



 g 2 dk

2π 2
k dε d = 3 .

In order to obtain g(ε) as a function of the energy ε one must invert the dispersion relation ε = ε(k) to
obtain k = k(ε).
224 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Note that we can equivalently write

ddk g Ωd d−1
g(ε) dε = g d
= k dk (5.28)
(2π) (2π)d

to derive g(ε).
For a spin-S particle with ballistic dispersion ε(k) = ~2 k2 /2m, we have

m d/2 d −1
 
2S +1
g(ε) = ε 2 Θ(ε) , (5.29)
Γ(d/2) 2π~2

where Θ(ε) is the step function, which takes the value 0 for ε < 0 and 1 for ε ≥ 0. The appearance of Θ(ε)
simply says that all the single particle energy eigenvalues are nonnegative. Note that we are assuming a
box of volume V but we are ignoring the quantization of kinetic energy, and assuming that the difference
between successive quantized single particle energy eigenvalues is negligible so that g(ε) can be replaced
by the average in the above expression. Note that
1
n(ε, T, µ) = . (5.30)
e(ε−µ)/kB T ∓1
This result holds true independent of the form of g(ε). The average total number of particles is then
Z∞
1
N (T, V, µ) = V dε g(ε) , (5.31)
e(ε−µ)/kB T ∓1
−∞

which does depend on g(ε).

5.3 Quantum Ideal Gases : Low Density Expansions

5.3.1 Expansion in powers of the fugacity

From eqn. 5.31, we have that the number density n = N/V is

Z∞
g(ε)
n(T, z) = dε
z −1 eε/kB T ∓1
−∞ (5.32)
X∞
= (±1)j−1 Cj (T ) z j ,
j=1

where z = exp(µ/kB T ) is the fugacity and

Z∞
Cj (T ) = dε g(ε) e−jε/kB T . (5.33)
−∞
5.3. QUANTUM IDEAL GASES : LOW DENSITY EXPANSIONS 225

From Ω = −pV and our expression above for Ω(T, V, µ), we have

Z∞  
p(T, z) = ∓ kB T dε g(ε) ln 1 ∓ z e−ε/kB T
−∞ (5.34)
∞
X
= kB T (±1)j−1 j −1 Cj (T ) z j .
j=1

5.3.2 Virial expansion of the equation of state

Eqns. 5.32 and 5.34 express n(T, z) and p(T, z) as power series in the fugacity z, with T -dependent
coefficients. In principal, we can eliminate z using eqn. 5.32, writing z = z(T, n) as a power series in the
number density n, and substitute this into eqn. 5.34 to obtain an equation of state p = p(T, n) of the
form  
p(T, n) = n kB T 1 + B2 (T ) n + B3 (T ) n2 + . . . . (5.35)

Note that the low density limit n → 0 yields the ideal gas law independent of the density of states g(ε).
This follows from expanding n(T, z) and p(T, z) to lowest order in z, yielding n = C1 z + O(z 2 ) and
p = kB T C1 z + O(z 2 ). Dividing the second of these equations by the first yields p = n kB T + O(n2 ),
which is the ideal gas law. Note that z = n/C1 + O(n2 ) can formally be written as a power series in n.
Unfortunately, there is no general analytic expression for the virial coefficients Bj (T ) in terms of the
expansion coefficients nj (T ). The only way is to grind things out order by order in our expansions. Let’s
roll up our sleeves and see how this is done. We start by formally writing z(T, n) as a power series in the
density n with T -dependent coefficients Aj (T ):

z = A1 n + A2 n 2 + A3 n 3 + . . . . (5.36)

We then insert this into the series for n(T, z):

n = C1 z ± C2 z 2 + C3 z 3 + . . .

2
= C1 A1 n + A2 n2 + A3 n3 + . . . ± C2 A1 n + A2 n2 + A3 n3 + . . .


(5.37)

3
+ C 3 A1 n + A2 n 2 + A3 n 3 + . . . + . . . .

Let’s expand the RHS to order n3 . Collecting terms, we have

n = C1 A1 n + C1 A2 ± C2 A21 n2 + C1 A3 ± 2C2 A1 A2 + C3 A31 n3 + . . .

. (5.38)

In order for this equation to be true we require that the coefficient of n on the RHS be unity, and that
the coefficients of nj for all j > 1 must vanish. Thus,

C 1 A1 = 1
C1 A2 ± C2 A21 = 0 (5.39)
C1 A3 ± 2C2 A1 A2 + C3 A31 = 0 .
226 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

The first of these yields A1 :

1
A1 = . (5.40)
C1
We now insert this into the second equation to obtain A2 :
C2
A2 = ∓ . (5.41)
C13
Next, insert the expressions for A1 and A2 into the third equation to obtain A3 :

2C22 C
A3 = 5 − 34 . (5.42)
C1 C1
This procedure rapidly gets tedious!
And we’re only half way done. We still must express p in terms of n:
p
2
= C1 A1 n + A2 n2 + A3 n3 + . . . ± 12 C2 A1 n + A2 n2 + A3 n3 + . . .


kB T

3
+ 31 C3 A1 n + A2 n2 + A3 n3 + . . . + . . .
(5.43)
2 2
1 1
C3 A31 3



= C1 A1 n + C1 A2 ± 2 C2 A1 n + C1 A3 ± C2 A1 A2 + 3 n + ...

= n + B2 n2 + B3 n3 + . . .

We can now write

C2
B2 = C1 A2 ± 21 C2 A21 = ∓
2C12
(5.44)
C2 2 C3
B3 = C1 A3 ± C2 A1 A2 + 1
3 C3 A31 = 24 − .
C1 3 C13

It is easy to derive the general result that BjF = (−1)j−1 BjB , where the superscripts denote Fermi (F) or
Bose (B) statistics.
We remark that the equation of state for classical (and quantum) interacting systems also can be expanded
in terms of virial coefficients. Consider, for example, the van der Waals equation of state,

aN 2
 
p+ 2 V − N b) = N kB T . (5.45)
V
This may be recast as
nkB T
p= − an2
1 − bn (5.46)
2 2 3 3 4


= nkB T + b kB T − a n + kB T b n + kB T b n + . . . ,

where n = N/V . Thus, for the van der Waals system, we have B2 = (b kB T − a) and Bk = kB T bk−1 for
all k ≥ 3.
5.4. ENTROPY AND COUNTING STATES 227

5.3.3 Ballistic dispersion

For the ballistic dispersion ε(p) = p2 /2m we computed the density of states in eqn. 5.29. One finds
Z∞
g λ−d d
Cj (T ) = S T dt t 2 −1 e−jt = gS λ−d
T j
−d/2
. (5.47)
Γ(d/2)
0

We then have
d
B2 (T ) = ∓ 2−( 2 +1) · g−1
S λT
d
 d
 (5.48)
B3 (T ) = 2−(d+1) − 3−( 2 +1) · 2 g−2 2d
S λT .

Note that B2 (T ) is negative for bosons and positive for fermions. This is because bosons have a tendency
to bunch and under certain circumstances may exhibit a phenomenon known as Bose-Einstein conden-
sation (BEC). Fermions, on the other hand, obey the Pauli principle, which results in an extra positive
correction to the pressure in the low density limit.
We may also write
n(T, z) = ±gS λ−d
T Li d (±z) (5.49)
2

and
p(T, z) = ±gS kB T λ−d
T Li d +1 (±z) , (5.50)
2

where
∞
X zn
Liq (z) ≡ (5.51)
n=1
nq

is the polylogarithm function 2 . Note that Liq (z) obeys a recursion relation in its index, viz.

∂
z Li (z) = Liq−1 (z) , (5.52)
∂z q
and that
∞
X 1
Liq (1) = = ζ(q) . (5.53)
n=1
nq

5.4 Entropy and Counting States

Suppose we are to partition N particles among J possible distinct single particle states. How many ways
Ω are there of accomplishing this task? The answer
depends on the statistics of the particles. If the
J
particles are fermions, the answer is easy: ΩFD = N . For bosons, the number of possible partitions
can be evaluated via the following argument. Imagine that we line up all the N particles in a row, and
we place J − 1 barriers among the particles, as shown below in Fig. 5.1. The number of partitions is
2
Several texts, such as Pathria and Reichl, write gq (z) for Liq (z). I adopt the latter notation since we are already using
the symbol g for the density of states function g(ε) and for the internal degeneracy g.
228 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

then the total number of ways of placing

N +J−1

the N particles among these N + J − 1 objects (particles
N
plus
barriers), hence we have ΩBE = N . For Maxwell-Boltzmann statistics, we take ΩMB = J /N ! Note
that ΩMB is not necessarily an integer, so Maxwell-Boltzmann statistics does not represent any actual
state counting. Rather, it manifests itself as a common limit of the Bose and Fermi distributions, as we
have seen and shall see again shortly.

Figure 5.1: Partitioning N bosons into J possible states (N = 14 and J = 5 shown). The N black dots
represent bosons, while the J − 1 white dots represent markers separating the different single particle
populations. Here n1 = 3, n2 = 1, n3 = 4, n4 = 2, and n5 = 4.

The entropy in each case is simply S = kB ln Ω. We assume N  1 and J  1, with n ≡ N/J finite.
Then using Stirling’s approximation, ln(K!) = K ln K − K + O(ln K), we have

SMB = −JkB n ln n
 
SBE = −JkB n ln n − (1 + n) ln(1 + n) (5.54)
 
SFD = −JkB n ln n + (1 − n) ln(1 − n) .

In the Maxwell-Boltzmann limit, n  1, and all three expressions agree. Note thatR
 
∂SMB

= −kB 1 + ln n
∂N J
 
∂SBE
= kB ln n−1 + 1


(5.55)
∂N J
 
∂SFD
= kB ln n−1 − 1 .


∂N J

Now let’s imagine grouping the single particle spectrum into intervals of J consecutive energy states.
If J is finite and the spectrum is continuous and we are in the thermodynamic limit, then these states
will all be degenerate. Therefore, using α as a label for the energies, we have that the grand potential
Ω = E − T S − µN is given in each case by
Xh i
ΩMB = J (εα − µ) nα + kB T nα ln nα
α
Xh i
ΩBE = J (εα − µ) nα + kB T nα ln nα − kB T (1 + nα ) ln(1 + nα ) (5.56)
α
Xh i
ΩFD = J (εα − µ) nα + kB T nα ln nα + kB T (1 − nα ) ln(1 − nα ) .
α

Now - lo and behold! - treating Ω as a function of the distribution

P {nα } and extremizing in each case,
subject to the constraint of total particle number N = J α nα , one obtains the Maxwell-Boltzmann,
5.5. PHOTON STATISTICS 229

Bose-Einstein, and Fermi-Dirac distributions, respectively:


nMB (µ−εα )/kB T


 α =e



δ  X   −1
= e(εα −µ)/kB T − 1

Ω − λJ n α0 = 0 ⇒ nBE
α (5.57)
δnα
α0





 −1
 FD  (εα −µ)/kB T
nα = e +1 .

As long as J is finite, so the states in each block all remain at the same energy, the results are independent
of J.

5.5 Photon Statistics

5.5.1 Thermodynamics of the photon gas

There exists a certain class of particles, including photons and certain elementary excitations in solids
such as phonons (i.e. lattice vibrations) and magnons (i.e. spin waves) which obey bosonic statistics
but with zero chemical potential. This is because their overall number is not conserved (under typical
conditions) – photons can be emitted and absorbed by the atoms in the wall of a container, phonon and
magnon number is also not conserved due to various processes, etc. In such cases, the free energy attains
its minimum value with respect to particle number when
 
∂F
µ= =0. (5.58)
∂N T.V
The number distribution, from eqn. 5.12, is then
1
n(ε) = . (5.59)
eβε −1

The grand partition function for a system of particles with µ = 0 is

Z∞
Ω(T, V ) = V kB T dε g(ε) ln 1 − e−ε/kB T ,


(5.60)
−∞

where g(ε) is the density of states per unit volume.

Suppose the particle dispersion is ε(p) = A|p|σ . We can compute the density of states g(ε):
Z d Z∞
dp
gΩd
g(ε) = g δ ε − A|p| = d dp pd−1 δ(ε − Apσ )
σ
hd h
0
 √ d d (5.61)
Z∞
gΩd − σd d
−1 2g π −1
= A dx x σ δ(ε − x) = ε σ Θ(ε) ,
σhd σ Γ(d/2) hA1/σ
0
230 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

where g is the internal degeneracy,

 due, for example, to different polarization states of the photon. We
have used the result Ωd = 2π d/2 Γ(d/2) for the solid angle in d dimensions. The step function Θ(ε) is
perhaps overly formal, but it reminds us that the energy spectrum is bounded from below by ε = 0, i.e.
there are no negative energy states.
For the photon, we have ε(p) = cp, hence σ = 1 and
2g π d/2 εd−1
g(ε) = Θ(ε) . (5.62)
Γ(d/2) (hc)d
In d = 3 dimensions the degeneracy is g = 2, the number of independent polarization states. The pressure
p(T ) is then obtained using Ω = −pV . We have
Z∞
p(T ) = −kB T dε g(ε) ln 1 − e−ε/kB T

−∞
Z∞
2 g π d/2
(hc) kB T dε εd−1 ln 1 − e−ε/kB T
−d


=− (5.63)
Γ(d/2)
0
Z∞
2 g π d/2 (kB T )d+1
dt td−1 ln 1 − e−t .


=− d
Γ(d/2) (hc)
0

We can make some progress with the dimensionless integral:

Z∞
Id ≡ − dt td−1 ln 1 − e−t

0
∞ Z∞
X 1
= dt td−1 e−nt (5.64)
n
n=1 0
∞
X 1
= Γ(d) = Γ(d) ζ(d + 1) .
nd+1
n=1

Finally, we invoke a result from the mathematics of the gamma function known as the doubling formula,
2z−1 z

z+1


Γ(z) = √ Γ 2 Γ 2 . (5.65)
π
Putting it all together, we find
− 12 (d+1) d+1

(kB T )d+1
p(T ) = g π Γ 2 ζ(d + 1) . (5.66)
(~c)d
The number density is found to be
Z∞
g(ε)
n(T ) = dε
eε/kB T −1
−∞ (5.67)
 d
− 12 (d+1) d+1

kB T
= gπ Γ 2 ζ(d) .
~c
5.5. PHOTON STATISTICS 231

For photons in d = 3 dimensions, we have g = 2 and thus

2 ζ(3) kB T 3 2 ζ(4) (kB T )4

 
n(T ) = , p(T ) = . (5.68)
π2 ~c π2 (~c)3
π4
It turns out that ζ(4) = 90 .

Note that ~c/kB = 0.22855 cm · K, so

kB T
= 4.3755 T [K] cm−1 =⇒ n(T ) = 20.405 × T 3 [K3 ] cm−3 . (5.69)
~c
To find the entropy, we use Gibbs-Duhem:
dp
dµ = 0 = −s dT + v dp =⇒ s=v , (5.70)
dT
where s is the entropy per particle and v = n−1 is the volume per particle. We then find
ζ(d+1)
s(T ) = (d+1) kB . (5.71)
ζ(d)
The entropy per particle is constant. The internal energy is
∂ ln Ξ ∂
E=− =− βpV ) = d · p V , (5.72)
∂β ∂β
and hence the energy per particle is
E d · ζ(d+1)
ε= = d · pv = kB T . (5.73)
N ζ(d)

5.5.2 Classical arguments for the photon gas

A number of thermodynamic properties of the photon gas can be determined from purely classical argu-
ments. Here we recapitulate a few important ones.

1. Suppose our photon gas is confined to a rectangular box of dimensions Lx ×Ly ×Lz . Suppose further
that the dimensions are all expanded by a factor λ1/3 , i.e. the volume is isotropically expanded by
a factor of λ. The cavity modes of the electromagnetic radiation have quantized wavevectors, even
within classical electromagnetic theory, given by
2πnx 2πny 2πnz
 
k= , , . (5.74)
Lx Ly Lz

Since the energy for a given mode is ε(k) = ~c|k|, we see that the energy changes by a factor λ−1/3
under an adiabatic volume expansion V → λV , where the distribution of different electromagnetic
mode occupancies remains fixed. Thus,
   
∂E ∂E
V =λ = − 13 E . (5.75)
∂V S ∂λ S
232 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Thus,  
∂E E
p=− = , (5.76)
∂V S 3V
as we found in eqn. 5.72. Since E = E(T, V ) is extensive, we must have p = p(T ) alone.

2. Since p = p(T ) alone, we have

   
∂E ∂E
= = 3p
∂V T ∂V p
  (5.77)
∂p
=T −p ,
∂T V
∂S

∂p

where the second line follows the Maxwell relation ∂V p
= ∂T V
, after invoking the First Law
dE = T dS − p dV . Thus,
dp
T = 4p =⇒ p(T ) = A T 4 , (5.78)
dT
where A is a constant. Thus, we recover the temperature dependence found microscopically in eqn.
5.66.

3. Given an energy density E/V , the differential energy flux emitted in a direction θ relative to a
surface normal is
E dΩ
djε = c · · cos θ · , (5.79)
V 4π
where dΩ is the differential solid angle. Thus, the power emitted per unit area is

Zπ/2 Z2π
dP cE cE
= dθ dφ sin θ · cos θ = = 3
4 c p(T ) ≡ σ T 4 , (5.80)
dA 4πV 4V
0 0

where σ = 34 cA, with p(T ) = A T 4 as we found above. From quantum statistical mechanical
considerations, we have
π 2 kB4 W
σ= 2 3
= 5.67 × 10−8 2 4 (5.81)
60 c ~ m K
is Stefan’s constant.

5.5.3 Surface temperature of the earth

We derived the result P = σT 4 · A where σ = 5.67 × 10−8 W/m2 K4 for the power emitted by an
electromagnetic ‘black body’. Let’s apply this result to the earth-sun system. We’ll need three lengths:
the radius of the sun R = 6.96 × 108 m, the radius of the earth Re = 6.38 × 106 m, and the radius of the
earth’s orbit ae = 1.50 × 1011 m. Let’s assume that the earth has achieved a steady state temperature
of Te . We balance the total power incident upon the earth with the power radiated by the earth. The
power incident upon the earth is

πRe2 4 2 Re2 R2
Pincident = · σT · 4πR = · πσT4 . (5.82)
4πa2e a2e
5.5. PHOTON STATISTICS 233

Figure 5.2: Spectral density ρε (ν, T ) for blackbody radiation at three temperatures.

The power radiated by the earth is

Pradiated = σTe4 · 4πRe2 . (5.83)
Setting Pincident = Pradiated , we obtain
1/2
R

Te = T . (5.84)
2 ae
Thus, we find Te = 0.04817 T , and with T = 5780 K, we obtain Te = 278.4 K. The mean surface
temperature of the earth is T̄e = 287 K, which is only about 10 K higher. The difference is due to the fact
that the earth is not a perfect blackbody, i.e. an object which absorbs all incident radiation upon it and
emits radiation according to Stefan’s law. As you know, the earth’s atmosphere retraps a fraction of the
emitted radiation – a phenomenon known as the greenhouse effect.

5.5.4 Distribution of blackbody radiation

Recall that the frequency of an electromagnetic wave of wavevector k is ν = c/λ = ck/2π. Therefore the
number of photons NT (ν, T ) per unit frequency in thermodynamic equilibrium is (recall there are two
polarization states)
2V d3k V k 2 dk
N (ν, T ) dν = 3 · ~ck/k T = 2 · ~ck/k T . (5.85)
8π e B −1 π e B −1
We therefore have
8πV ν2
N (ν, T ) = 3 · hν/k T . (5.86)
c e B −1
Since a photon of frequency ν carries energy hν, the energy per unit frequency E(ν) is
8πhV ν3
E(ν, T ) = · . (5.87)
c3 ehν/kB T − 1
234 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Note what happens if Planck’s constant h vanishes, as it does in the classical limit. The denominator
can then be written
hν
ehν/kB T − 1 = + O(h2 ) (5.88)
kB T
and
8πkB T 2
ECL (ν, T ) = lim E(ν) = V · ν . (5.89)
h→0 c3
In classical electromagnetic theory, then, the total energy integrated over all frequencies diverges. This
is known as the ultraviolet catastrophe, since the divergence comes from the large ν part of the integral,
which in the optical spectrum is the ultraviolet portion. With quantization, the Bose-Einstein factor
imposes an effective ultraviolet cutoff kB T /h on the frequency integral, and the total energy, as we found
above, is finite:
Z∞
π 2 (kB T )4
E(T ) = dν E(ν) = 3pV = V · . (5.90)
15 (~c)3
0

We can define the spectral density ρε (ν) of the radiation as

E(ν, T ) 15 h (hν/kB T )3
ρε (ν, T ) ≡ = 4 (5.91)
E(T ) π kB T ehν/kB T − 1

so that ρε (ν, T ) dν is the fraction of the electromagnetic energy, under equilibrium conditions, between
R∞
frequencies ν and ν + dν, i.e. dν ρε (ν, T ) = 1. In fig. 5.2 we plot this in fig. 5.2 for three different
0
temperatures. The maximum occurs when s ≡ hν/kB T satisfies
 3 
d s s
=0 =⇒ =3 =⇒ s = 2.82144 . (5.92)
s
ds e − 1 1 − e−s

5.5.5 What if the sun emitted ferromagnetic spin waves?

We saw in eqn. 5.79 that the power emitted per unit surface area by a blackbody is σT 4 . The power law
here follows from the ultrarelativistic dispersion ε = ~ck of the photons. Suppose that we replace this
dispersion with the general form ε = ε(k). Now consider a large box in equilibrium at temperature T .
The energy current incident on a differential area dA of surface normal to ẑ is
Z 3
dk 1 ∂ε(k) 1
dP = dA · 3
Θ(cos θ) · ε(k) · · ε(k)/k T . (5.93)
(2π) ~ ∂kz e B −1
Let us assume an isotropic power law dispersion of the form ε(k) = Ck α . Then after a straightforward
calculation we obtain
dP 2
= σ T 2+ α , (5.94)
dA
where
2+ 2 2
2

2

g kB α C − α
σ =ζ 2+ α Γ 2+ α · . (5.95)
8π 2 ~
One can check that for g = 2, C = ~c, and α = 1 that this result reduces to that of eqn. 5.81.
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 235

5.6 Lattice Vibrations : Einstein and Debye Models

Crystalline solids support propagating waves called phonons, which are quantized vibrations of the lattice.
p2
Recall that the quantum mechanical Hamiltonian for a single harmonic oscillator, Ĥ = 2m + 12 mω02 q 2 ,
† 1 †
may be written
 †
 as Ĥ = ~ω0 (a a + 2 ), where a and a are ‘ladder operators’ satisfying commutation
relations a , a = 1.

5.6.1 One-dimensional chain

Consider the linear chain of masses and springs depicted in fig. 5.3. We assume that our system consists
of N mass points on a large ring of circumference L. In equilibrium, the masses are spaced evenly by a
distance b = L/N . That is, x0n = nb is the equilibrium position of particle n. We define un = xn − x0n to
be the difference between the position of mass n and The Hamiltonian is then
X  p2 
n 1 2
Ĥ = + κ (xn+1 − xn − a)
n
2m 2
X  p2  (5.96)
n 1 2 1 2
= + κ (un+1 − un ) + 2 N κ(b − a) ,
n
2m 2

where a is the unstretched length of each spring, m is the mass of each mass point, κ is the force constant
of each spring, and N is the total number of mass points. If b 6= a the springs are under tension in
equilibrium, but as we see this only leads to an additive constant in the Hamiltonian, and hence does not
enter the equations of motion.
The classical equations of motion are

∂ Ĥ p
u̇n = = n
∂pn m
(5.97)
∂ Ĥ

ṗn = − = κ un+1 + un−1 − 2un .
∂un

Taking the time derivative of the first equation and substituting into the second yields
κ

ün = un+1 + un−1 − 2un . (5.98)
m
We now write
1 X
un = √ ũk eikna , (5.99)
N k

where periodicity uN +n = un requires that the k values are quantized so that eikN a = 1, i.e. k = 2πj/N a
where j ∈ {0, 1, . . . , N −1}. The inverse of this discrete Fourier transform is

1 X
ũk = √ un e−ikna . (5.100)
N n
236 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.3: A linear chain of masses and springs. The black circles represent the equilibrium positions
of the masses. The displacement of mass n relative to its equilibrium value is un .

Note that ũk is in general complex, but that ũ∗k = ũ−k . In terms of the ũk , the equations of motion take
the form
¨k = − 2κ 1 − cos(ka) ũk ≡ −ωk2 ũk .


ũ (5.101)
m
Thus, each ũk is a normal mode, and the normal mode frequencies are
r
κ 
sin 21 ka  .


ωk = 2 (5.102)
m
The density of states for this band of phonon excitations is

Zπ/a
dk
g(ε) = δ(ε − ~ωk )
2π
−π/a (5.103)
2
−1/2
= J 2 − ε2 Θ(ε) Θ(J − ε) ,
πa
p
where J = 2~ κ/m is the phonon bandwidth. The step functions require 0 ≤ ε ≤ J; outside this range
there are no phonon energy levels and the density of states accordingly vanishes.
 
The entire theory can be quantized, taking pn , un0 = −i~δnn0 . We then define
1 X 1 X
pn = √ p̃k eikna , p̃k = √ pn e−ikna , (5.104)
N k N n

in which case p̃k , ũk0 = −i~δkk0 . Note that ũ†k = ũ−k and p̃†k = p̃−k . We then define the ladder operator
 

 1/2  1/2
1 mωk
ak = p̃k − i ũk (5.105)
2m~ωk 2~

and its Hermitean conjugate a†k , in terms of which the Hamiltonian is

~ωk a†k ak + 21 ,
X

Ĥ = (5.106)
k

which is a sum over independent harmonic oscillator modes. Note that the sum over k is restricted to
an interval of width 2π, e.g. k ∈ − πa , πa , which is the first Brillouin zone for the one-dimensional chain


structure. The state at wavevector k + 2π a is identical to that at k, as we see from eqn. 5.100.
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 237

Figure 5.4: A crystal structure with an underlying square Bravais lattice and a three element basis.

5.6.2 General theory of lattice vibrations

The most general model of a harmonic solid is described by a Hamiltonian of the form
X p2 (R) 1 XX X α
Ĥ = i
+ ui (R) Φαβ 0 β 0
ij (R − R ) uj (R ) , (5.107)
2Mi 2 0
R,i i,j α,β R,R

where the dynamical matrix is

∂2U
Φαβ 0
ij (R − R ) = , (5.108)
∂uαi (R) ∂uβj (R0 )

where U is the potential energy of interaction among all the atoms. Here we have simply expanded the
potential to second order in the local displacements uαi (R). The lattice sites R are elements of a Bravais
lattice. The indices i and j specify basis elements with respect to this lattice, and the indices α and
β range over {1, . . . , d}, the number of possible directions in space. The subject of crystallography is
beyond the scope of these notes, but, very briefly, a Bravais lattice in d dimensions is specified by a set of
d linearly independent primitive direct lattice vectors al , such that any point in the Bravais lattice may
be written as a sum over the primitive vectors with integer coefficients: R = dl=1 nl al . The set of all
P
such vectors {R} is called the direct lattice. The direct lattice is closed under the operation of vector
addition: if R and R0 are points in a Bravais lattice, then so is R + R0 .
A crystal is a periodic arrangement of lattice sites. The fundamental repeating unit is called the unit
cell . Not every crystal is a Bravais lattice, however. Indeed, Bravais lattices are special crystals in which
there is only one atom per unit cell. Consider, for example, the structure in fig. 5.4. The blue dots form
238 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

a square Bravais lattice with primitive direct lattice vectors a1 = a x̂ and a2 = a ŷ, where a is the lattice
constant, which is the distance between any neighboring pair of blue dots. The red squares and green
triangles, along with the blue dots, form a basis for the crystal structure which label each sublattice. Our
crystal in fig. 5.4 is formally classified as a square Bravais lattice with a three element basis. To specify
an arbitrary site in the crystal, we must specify both a direct lattice vector R as well as a basis index
j ∈ {1, . . . , r}, so that the location is R + ηj . The vectors {ηj } are the basis vectors for our crystal
structure. We see that a general crystal structure consists of a repeating unit, known as a unit cell . The
centers (or corners, if one prefers) of the unit cells form a Bravais lattice. Within a given unit cell, the
individual sublattice sites are located at positions ηj with respect to the unit cell position R.
Upon diagonalization, the Hamiltonian of eqn. 5.107 takes the form
X  
Ĥ = ~ωa (k) A†a (k) Aa (k) + 12 , (5.109)
k,a

where
Aa (k) , A†b (k0 ) = δab δkk0 .
 
(5.110)
The eigenfrequencies are solutions to the eigenvalue equation
(a) (a)
Φ̃αβ
X
2
ij (k) ejβ (k) = Mi ωa (k) eiα (k) , (5.111)
j,β

where
Φ̃αβ Φαβ
X
−ik·R
ij (k) = ij (R) e . (5.112)
R

Here, k lies within the first Brillouin zone, which is the unit cell of the reciprocal lattice of points G
satisfying eiG·R = 1 for all G and R. The reciprocal lattice is also a Bravais lattice, with Pprimitive
reciprocal lattice vectors bl , such that any point on the reciprocal lattice may be written G = dl=1 ml bl .
One also has that al · bl0 = 2πδll0 . The index a ranges from 1 to d · r and labels the mode of oscillation at
(a)
wavevector k. The vector eiα (k) is the polarization vector for the ath phonon branch. In solids of high
symmetry, phonon modes can be classified as longitudinal or transverse excitations.
For a crystalline lattice with an r-element basis, there are then d · r phonon modes for each wavevector
k lying in the first Brillouin zone. If we impose periodic boundary conditions, then the k points within
the first Brillouin zone are themselves quantized, as in the d = 1 case where we found k = 2πn/N . There
are N distinct k points in the first Brillouin zone – one for every direct lattice site. The total number of
modes is than d · r · N , which is the total number of translational degrees of freedom in our system: rN
total atoms (N unit cells each with an r atom basis) each free to vibrate in d dimensions. Of the d · r
branches of phonon excitations, d of them will be acoustic modes whose frequency vanishes as k → 0.
The remaining d(r − 1) branches are optical modes and oscillate at finite frequencies. Basically, in an
acoustic mode, for k close to the (Brillouin) zone center k = 0, all the atoms in each unit cell move
together in the same direction at any moment of time. In an optical mode, the different basis atoms
move in different directions.
There is no number conservation law for phonons – they may be freely created or destroyed in anharmonic
processes, where two photons with wavevectors k and q can combine into a single phonon with wavevector
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 239

k + q, and vice versa. Therefore the chemical potential for phonons is µ = 0. We define the density of
states ga (ω) for the ath phonon mode as
Z d
1 X
dk

ga (ω) = δ ω − ωa (k) = V0 d
δ ω − ωa (k) , (5.113)
N (2π)
k BZ

where N is the number of unit cells, V0 is the unit cell volume of the direct lattice, and the k sum
and integral are over the first Brillouin zone only. Note that ω here has dimensions of frequency. The
functions ga (ω) is normalized to unity:
Z∞
dω ga (ω) = 1 . (5.114)
0

The total phonon density of states per unit cell is given by3
dr
X
g(ω) = ga (ω) . (5.115)
a=1

The grand potential for the phonon gas is

∞

na (k)+ 12
e−β~ωa (k)
Y X
Ω(T, V ) = −kB T ln
k,a na (k)=0
" # 
X ~ωa (k)
= kB T ln 2 sinh (5.116)
2kB T
k,a
Z∞ "  #
~ω
= N kB T dω g(ω) ln 2 sinh .
2kB T
0

∂Ω


Note that V = N V0 since there are N unit cells, each of volume V0 . The entropy is given by S = − ∂T V
and thus the heat capacity is
Z∞
∂ 2Ω ~ω 2
   
2 ~ω
CV = −T = N kB dω g(ω) csch (5.117)
∂T 2 2kB T 2kB T
0

2kB T
Note that as T → ∞ we have csch ~ω
2kB T → ~ω , and therefore

Z∞
lim CV (T ) = N kB dω g(ω) = rdN kB . (5.118)
T →∞
0

This is the classical Dulong-Petit limit of 12 kB per quadratic degree of freedom; there are rN atoms
moving in d dimensions, hence d · rN positions and an equal number of momenta, resulting in a high
temperature limit of CV = rdN kB .
3
Note the dimensions of g(ω) are (frequency)−1 . By contrast, the dimensions of g(ε) in eqn. 5.29 are (energy)−1 ·
(volume)−1 . The difference lies in the a factor of V0 · ~, where V0 is the unit cell volume.
240 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.5: Upper panel: phonon spectrum in elemental rhodium (Rh) at T = 297 K measured by high
precision inelastic neutron scattering (INS) by A. Eichler et al., Phys. Rev. B 57, 324 (1998). Note the
three acoustic branches and no optical branches, corresponding to d = 3 and r = 1. Lower panel: phonon
spectrum in gallium arsenide (GaAs) at T = 12 K, comparing theoretical lattice-dynamical calculations
with INS results of D. Strauch and B. Dorner, J. Phys.: Condens. Matter 2, 1457 (1990). Note the three
acoustic branches and three optical branches, corresponding to d = 3 and r = 2. The Greek letters along
the x-axis indicate points of high symmetry in the Brillouin zone.

5.6.3 Einstein and Debye models

HIstorically, two models of lattice vibrations have received wide attention. First is the so-called Einstein
model , in which there is no dispersion to the individual phonon modes. We approximate ga (ω) ≈ δ(ω−ωa ),
in which case
X  ~ω 2 
~ωa

a 2
CV (T ) = N kB csch . (5.119)
a
2kB T 2kB T

At low temperatures, the contribution from each branch vanishes exponentially, since csch2 ~ωa

2kB T '
4 e−~ωa /kB T → 0. Real solids don’t behave this way.
A more realistic model. due to Debye, accounts for the low-lying acoustic phonon branches. Since the
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 241

acoustic phonon dispersion vanishes linearly with |k| as k → 0, there is no temperature at which the
acoustic phonons ‘freeze out’ exponentially, as in the case of Einstein phonons. Indeed, the Einstein
model is appropriate in describing the d (r−1) optical phonon branches, though it fails miserably for the
acoustic branches.
In the vicinity of the zone center k = 0 (also called Γ in crystallographic notation) the d acoustic modes
obey a linear dispersion, with ωa (k) = ca (k̂) k. This results in an acoustic phonon density of states in
d = 3 dimensions of
V0 ω 2 X dk̂ 1
Z
g̃(ω) = Θ(ωD − ω)
2π 2 a 4π c3a (k)
(5.120)
3V0 2
= 2 3 ω Θ(ωD − ω) ,
2π c̄
where c̄ is an average acoustic phonon velocity (i.e. speed of sound) defined by

dk̂ 1
Z
3 X
= (5.121)
c̄3 a
4π c3a (k)

and ωD is a cutoff known as the Debye frequency. The cutoff is necessary because the phonon branch
does not extend forever, but only to the boundaries of the Brillouin zone. Thus, ωD should roughly be
equal to the energy of a zone boundary phonon. Alternatively, we can define ωD by the normalization
condition
Z∞
dω g̃(ω) = 3 =⇒ ωD = (6π 2 /V0 )1/3 c̄ . (5.122)
0

This allows us to write g̃(ω) = 9ω 2 /ωD3 Θ(ωD − ω).

The specific heat due to the acoustic phonons is then

ZωD
~ω 2
   
9N kB 2 2 ~ω
CV (T ) = dω ω csch
ωD3 2kB T 2kB T
0 (5.123)
2T 3
 


= 9N kB φ ΘD /2T ,
ΘD

where ΘD = ~ωD /kB is the Debye temperature and


1 3
Zx 3x
 x→0
φ(x) = dt t4 csch 2 t = (5.124)
 π4

0
30 x→∞.

Therefore,   3
12π 4 T

 5
 N k B Θ T  ΘD
D
CV (T ) = (5.125)


3N kB T  ΘD .

242 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Element Ag Al Au C Cd Cr Cu Fe Mn
ΘD (K) 227 433 162 2250 210 606 347 477 409
Tmelt (K) 962 660 1064 3500 321 1857 1083 1535 1245
Element Ni Pb Pt Si Sn Ta Ti W Zn
ΘD (K) 477 105 237 645 199 246 420 383 329
Tmelt (K) 1453 327 1772 1410 232 2996 1660 3410 420

Table 5.1: Debye temperatures (at T = 0) and melting points for some common elements (carbon is
assumed to be diamond and not graphite). (Source: the internet!)

Thus, the heat capacity due to acoustic phonons obeys the Dulong-Petit rule in that CV (T → ∞) = 3N kB ,
corresponding to the three acoustic degrees of freedom per unit cell. The remaining contribution of
3(r − 1)N kB to the high temperature heat capacity comes from the optical modes not considered in the
Debye model. The low temperature T 3 behavior of the heat capacity of crystalline solids is a generic
feature, and its detailed description is a triumph of the Debye model.

5.6.4 Melting and the Lindemann criterion

Atomic fluctuations in a crystal

For the one-dimensional chain, eqn. 5.105 gives

 1/2
~
ak − a†−k .


ũk = i (5.126)
2mωk

Therefore the RMS fluctuations at each site are given by

1 X
hu2n i = hũk ũ−k i
N
k
(5.127)
1 X ~
n(k) + 21 ,


=
N mωk
k

 −1
where n(k, T ) = exp(~ωk /kB T ) − 1 is the Bose occupancy function.
Let us now generalize this expression to the case of a d-dimensional solid. The appropriate expression for
the RMS position fluctuations of the ith basis atom in each unit cell is

dr
1 XX ~
hu2i (R)i = na (k) + 21 .


(5.128)
N Mia (k) ωa (k)
k a=1

Here we sum over all wavevectors k in the first Brilliouin zone, and over all normal modes a. There are
dr normal modes per unit cell i.e. d branches of the phonon dispersion ωa (k). (For the one-dimensional
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 243

chain with d = 1 and r = 1 there was only one such branch to consider). Note also the quantity Mia (k),
(a)
which has units of mass and is defined in terms of the polarization vectors eiα (k) as
d
1 X  (a) 2
e (k) .
= iµ (5.129)
Mia (k)
µ=1

The dimensions of the polarization vector are [mass]−1/2 , since the generalized orthonormality condition
on the normal modes is
(a) ∗ (b)
X
Mi eiµ (k) eiµ (k) = δ ab , (5.130)
i,µ

where Mi is the mass of the atom of species i within the unit cell (i ∈ {1, . . . , r}). For our purposes we
can replace Mia (k) by an appropriately averaged quantity which we call Mi ; this ‘effective mass’ is then
independent of the mode index a as well as the wavevector k. We may then write
Z∞ ( )
~ 1 1
h u2i i ≈ dω g(ω) · + , (5.131)
Mi ω e~ω/kB T − 1 2
0

where we have dropped the site label R since translational invariance guarantees that the fluctuations
are the same from one unit cell to the next. Note that the fluctuations h u2i i can be divided into
a temperature-dependent part h u2i ith and a temperature-independent quantum contribution h u2i iqu ,
where
Z∞
2 ~ g(ω) 1
h ui ith = dω · ~ω/k T
Mi ω e B −1
0
(5.132)
Z∞
~ g(ω)
h u2i iqu = dω .
2Mi ω
0

Let’s evaluate these contributions within the Debye model, where we replace g(ω) by
d2 ω d−1
ḡ(ω) = Θ(ωD − ω) . (5.133)
ωDd
We then find
d−1
d2 ~

kB T
h u2i ith = Fd (~ωD /kB T )
Mi ωD ~ωD
(5.134)
d2 ~
h u2i iqu = · ,
d − 1 2Mi ωD

where  d−2
x
Zx  d−2 x→0
sd−2

Fd (x) = ds s = . (5.135)
e −1 
ζ(d − 1) x → ∞

0

We can now extract from these expressions several important conclusions:

244 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

1) The T = 0 contribution to the the fluctuations, h u2i iqu , diverges in d = 1 dimensions. Therefore
there are no one-dimensional quantum solids.

2) The thermal contribution to the fluctuations, h u2i ith , diverges for any T > 0 whenever d ≤ 2. This
is because the integrand of Fd (x) goes as sd−3 as s → 0. Therefore, there are no two-dimensional
classical solids.

3) Both the above conclusions are valid in the thermodynamic limit. Finite size imposes a cutoff on
the frequency integrals, because there is a smallest wavevector kmin ∼ 2π/L, where L is the (finite)
linear dimension of the system. This leads to a low frequency cutoff ωmin = 2πc̄/L, where c̄ is the
appropriately averaged acoustic phonon velocity from eqn. 5.121, which mitigates any divergences.

Lindemann melting criterion

An old phenomenological theory of melting due to Lindemann says that a crystalline solid melts when
the RMS fluctuations in the atomic positions exceeds a certain fraction η of the lattice constant a. We
therefore define the ratios
h u2i ith ~2 T d−1
 
2 2
xi,th ≡ = d · · · F (ΘD /T )
a2 Mi a2 kB ΘDd
(5.136)
h u2i iqu d2 ~2
 
2 1
xi,qu ≡ = · · ,
a2 2(d − 1) Mi a2 k B ΘD
q q .
with xi = xi,th + xi,qu = h u2i i a.
2 2

Let’s now work through an example of a three-dimensional solid. We’ll assume a single element basis
(r = 1). We have that
9~2 /4kB
2 = 109 K . (5.137)
1 amu Å
According to table 5.1, the melting temperature always exceeds the Debye temperature, and often by a
great amount. We therefore assume T  ΘD , which puts us in the small x limit of Fd (x). We then find
s
Θ? Θ? 4T 4T Θ?

2 2
xqu = , xth = · , x= 1+ . (5.138)
ΘD ΘD ΘD ΘD ΘD

where
109 K
Θ∗ =
2 . (5.139)
M [amu] · a[Å]
The total position fluctuation is of course the sum x2 = x2i,th + x2i,qu . Consider for example the case of
copper, with M = 56 amu and a = 2.87 Å. The Debye temperature is ΘD = 347 K. From this we find
xqu = 0.026, which says that at T = 0 the RMS fluctuations of the atomic positions are not quite three
percent of the lattice spacing (i.e. the distance between neighboring copper atoms). At room temperature,
T = 293 K, one finds xth = 0.048, which is about twice as large as the quantum contribution. How big
are the atomic position fluctuations at the melting point? According to our table, Tmelt = 1083 K for
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 245

copper, and from our formulae we obtain xmelt = 0.096. The Lindemann criterion says that solids melt
when x(T ) ≈ 0.1.
We were very lucky to hit the magic number xmelt = 0.1 with copper. Let’s try another example. Lead
has M = 208 amu and a = 4.95 Å. The Debye temperature is ΘD = 105 K (‘soft phonons’), and the
melting point is Tmelt = 327 K. From these data we obtain x(T = 0) = 0.014, x(293 K) = 0.050 and
x(T = 327 K) = 0.053. Same ballpark.
We can turn the analysis around and predict a melting temperature based on the Lindemann criterion
x(Tmelt ) = η, where η ≈ 0.1. We obtain

η 2 ΘD
 
Θ
TL = ?
−1 · D . (5.140)
Θ 4

We call TL the Lindemann temperature. Most treatments of the Lindemann criterion ignore the quantum
correction, which gives the −1 contribution inside the above parentheses. But if we are more careful and
include it, we see that it may be possible to have TL < 0. This occurs for any crystal where ΘD < Θ? /η 2 .
Consider for example the case of 4 He, which at atmospheric pressure condenses into a liquid at Tc = 4.2 K
and remains in the liquid state down to absolute zero. At p = 1 atm, it never solidifies! Why? The number
density of liquid 4 He at p = 1 atm and T = 0 K is 2.2 × 1022 cm−3 . Let’s say the Helium atoms want
to form a crystalline lattice. We don’t know a priori what the lattice structure will be, so let’s for the
sake of simplicity assume a simple cubic lattice. From the number density we obtain a lattice spacing of
a = 3.57 Å. OK now what do we take for the Debye temperature? Theoretically this should depend on the
microscopic force constants which enter the small oscillations problem (i.e. the spring constants between
pairs of helium atoms in equilibrium). We’ll use the expression we derived for the Debye frequency,
ωD = (6π 2 /V0 )1/3 c̄, where V0 is the unit cell volume. We’ll take c̄ = 238 m/s, which is the speed of
sound in liquid helium at T = 0. This gives ΘD = 19.8 K. We find Θ? = 2.13 K, and if we take η = 0.1
this gives Θ? /η 2 = 213 K, which significantly exceeds ΘD . Thus, the solid should melt because the RMS
fluctuations in the atomic positions at absolute zero are huge: xqu = (Θ? /ΘD )1/2 = 0.33. By applying
pressure, one can get 4 He to crystallize above pc = 25 atm (at absolute zero). Under pressure, the unit
cell volume V0 decreases and the phonon velocity c̄ increases, so the Debye temperature itself increases.
It is important to recognize that the Lindemann criterion does not provide us with a theory of melting
per se. Rather it provides us with a heuristic which allows us to predict roughly when a solid should
melt.

5.6.5 Goldstone bosons

The vanishing of the acoustic phonon dispersion at k = 0 is a consequence of Goldstone’s theorem

which says that associated with every broken generator of a continuous symmetry there is an associated
bosonic gapless excitation (i.e. one whose frequency ω vanishes in the long wavelength limit). In the
case of phonons, the ‘broken generators’ are the symmetries under spatial translation in the x, y, and z
directions. The crystal selects a particular location for its center-of-mass, which breaks this symmetry.
There are, accordingly, three gapless acoustic phonons.
Magnetic materials support another branch of elementary excitations known as spin waves, or magnons.
246 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

In isotropic magnets, there is a global symmetry associated with rotations in internal spin space, described
by the group SU(2). If the system spontaneously magnetizes, meaning there is long-ranged ferromagnetic
order (↑↑↑ · · · ), or long-ranged antiferromagnetic order (↑↓↑↓ · · · ), then global spin rotation symmetry is
broken. Typically a particular direction is chosen for the magnetic moment (or staggered moment, in the
case of an antiferromagnet). Symmetry under rotations about this axis is then preserved, but rotations
which do not preserve the selected axis are ‘broken’. In the most straightforward case, that of the
antiferromagnet, there are two such rotations for SU(2), and concomitantly two gapless magnon branches,
with linearly vanishing dispersions ωa (k). The situation is more subtle in the case of ferromagnets, because
the total magnetization is conserved by the dynamics (unlike the total staggered magnetization in the
case of antiferromagnets). Another wrinkle arises if there are long-ranged interactions present.
For our purposes, we can safely ignore the deep physical reasons underlying the gaplessness of Goldstone
bosons and simply posit a gapless dispersion relation of the form ω(k) = A |k|σ . The density of states
for this excitation branch is then
d
−1
g(ω) = C ω σ Θ(ωc − ω) , (5.141)

where C is a constant and ωc is the cutoff, which is the bandwidth for this excitation branch.4 Normalizing
the density of states for this branch results in the identification ωc = (d/σC)σ/d .
The heat capacity is then found to be

Zωc
~ω 2
d
   
−1 2 ~ω
CV = N kB C dω ω σ
csch
kB T 2kB T
0 (5.142)
 d/σ
d 2T

= N kB φ Θ/2T ,
σ Θ

where Θ = ~ωc /kB and


σ d/σ
Zx d d x
 x→0
+1
φ(x) = dt t σ csch 2 t = (5.143)

 −d/σ
Γ 2 + σd ζ 2 + σd



0 2 x→∞,

which is a generalization of our earlier results. Once again, we recover Dulong-Petit for kB T  ~ωc , with
CV (T  ~ωc /kB ) = N kB .
In an isotropic ferromagnet, i.e.a ferromagnetic material where there is full SU(2) symmetry in internal
‘spin’ space, the magnons have a k 2 dispersion. Thus, a bulk three-dimensional isotropic ferromagnet will
exhibit a heat capacity due to spin waves which behaves as T 3/2 at low temperatures. For sufficiently
low temperatures this will overwhelm the phonon contribution, which behaves as T 3 .

−d d
−σ
4
If ω(k) = Akσ , then C = 21−d π σ −1 A
2

g Γ(d/2) .
5.7. THE IDEAL BOSE GAS 247

5.7 The Ideal Bose Gas

5.7.1 General formulation for noninteracting systems

Recall that the grand partition function for noninteracting bosons is given by
∞
!
Y X Y −1
β(µ−εα )nα
Ξ= e = 1 − eβ(µ−εα ) , (5.144)
α nα =0 α

In order for the sum to converge to the RHS above, we must have µ < εα for all single-particle states |αi.
The density of particles is then
  ∞
∂Ω g(ε)
Z
1 1 X 1
n(T, µ) = − = β(ε −µ)
= dε β(ε−µ) , (5.145)
V ∂µ T,V V α e α −1 e −1
ε0

where g(ε) = V1 α δ(ε − εα ) is the density of single particle states per unit volume. We assume that
P
g(ε) = 0 for ε < ε0 ; typically ε0 = 0, as is the case for any dispersion of the form ε(k) = A|k|r , for
example. However, in the presence of a magnetic field, we could have ε(k, σ) = A|k|r − gµ0 Hσ, in which
case ε0 = −gµ0 |H|.
Clearly n(T, µ) is an increasing function of both T and µ. At fixed T , the maximum possible value for
n(T, µ), called the critical density nc (T ), is achieved for µ = ε0 , i.e.
Z∞
g(ε)
nc (T ) = dε . (5.146)
eβ(ε−ε0 ) −1
ε0

The above integral converges provided g(ε0 ) = 0, assuming g(ε) is continuous5 . If g(ε0 ) > 0, the integral
diverges, and nc (T ) = ∞. In this latter case, one can always invert the equation for n(T, µ) to obtain
the chemical potential µ(T, n). In the former case, where the nc (T ) is finite, we have a problem – what
happens if n > nc (T ) ?
In the former case, where nc (T ) is finite, we can equivalently restate the problem in terms of a critical
temperature Tc (n), defined by the equation nc (Tc ) = n. For T < Tc , we apparently can no longer invert to
obtain µ(T, n), so clearly something has gone wrong. The remedy is to recognize that the single particle
energy levels are discrete, and separate out the contribution from the lowest energy state ε0 . I.e. we write
n0 n0
z }| { z }| {
Z∞
1 g0 g(ε)
n(T, µ) = β(ε −µ)
+ dε , (5.147)
V e 0 −1 eβ(ε−µ) −1
ε0

where g0 is the degeneracy of the single particle state with energy ε0 . We assume that n0 is finite, which
means that N0 = V n0 is extensive. We say that the particles have condensed into the state with energy
5
OK, that isn’t quite true. For example, if g(ε) ∼ 1/ ln ε, then the integral has a very weak ln ln(1/η) divergence, where
η is the lower cutoff. But for any power law density of states g(ε) ∝ εr with r > 0, the integral converges.
248 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

ε0 . The quantity n0 is the condensate density. The remaining particles, with density n0 , are said to
comprise the overcondensate. With the total density n fixed, we have n = n0 + n0 . Note that n0 finite
means that µ is infinitesimally close to ε0 :
 
g0 g k T
µ = ε0 − kB T ln 1 + ≈ ε0 − 0 B . (5.148)
V n0 V n0

Note also that if ε0 − µ is finite, then n0 ∝ V −1 is infinitesimal.

Thus, for T < Tc (n), we have µ = ε0 with n0 > 0, and
Z∞
g(ε)
n(T, n0 ) = n0 + dε . (5.149)
e(ε−ε0 )/kB T −1
ε0

For T > Tc (n), we have n0 = 0 and

Z∞
g(ε)
n(T, µ) = dε . (5.150)
e(ε−µ)/kB T −1
ε0

The equation for Tc (n) is

Z∞
g(ε)
n = dε . (5.151)
e(ε−ε0 )/kB Tc −1
ε0

For another take on ideal Bose gas condensation see the appendix in §5.10.

5.7.2 Ballistic dispersion

We already derived, in §5.3.3, expressions for n(T, z) and p(T, z) for the ideal Bose gas (IBG) with ballistic
dispersion ε(p) = p2 /2m, We found

n(T, z) = g λ−d
T Li d (z)
2
(5.152)
p(T, z) = g kB T λ−d
T Li d +1 (z),
2

where g is the internal (e.g. spin) degeneracy of each single particle energy level. Here z = eµ/kB T is the
fugacity and
∞
X zm
Lis (z) = (5.153)
ms
m=1

is the polylogarithm function. For bosons with a spectrum bounded below by ε0 = 0, the fugacity takes
values on the interval z ∈ [0, 1]6 .
6
It is easy to see that the chemical potential for noninteracting bosons can never exceed the minimum value ε0 of the
single particle dispersion.
5.7. THE IDEAL BOSE GAS 249

1 3 5
Figure 5.6: The polylogarithm function Lis (z) versus z for s = 2, s = 2, and s = 2. Note that
Lis (1) = ζ(s) diverges for s ≤ 1.

Clearly n(T, z) = g λ−d

T Li d (z) is an increasing function of z for fixed T . In fig. 5.6 we plot the function
2
Lis (z) versus z for three different values of s. We note that the maximum value Lis (z = 1) is finite if
s > 1. Thus, for d > 2, there is a maximum density nmax (T ) = g Li d (z) λ−d
T which is an increasing
2
function of temperature T . Put another way, if we fix the density n, then there is a critical temperature
Tc below which there is no solution to the equation n = n(T, z). The critical temperature Tc (n) is then
determined by the relation

mkB Tc d/2
2/d
2π~2
  
d n
n = gζ 2 =⇒ kB Tc = . (5.154)
2π~2 m g ζ d2

What happens for T < Tc ?

As shown above in §5.7, we must separate out the contribution from the lowest energy single particle
mode, which for ballistic dispersion lies at ε0 = 0. Thus writing
1 1 1 X 1
n= −1
+ , (5.155)
V z − 1 V α z e α BT − 1
−1 ε /k
(εα >0)

where we have taken g = 1. Now V −1 is of course very small, since V is thermodynamically large, but if
µ → 0 then z −1 − 1 is also very small and their ratio can be finite, as we have seen. Indeed, if the density
of k = 0 bosons n0 is finite, then their total number N0 satisfies
1 1
N0 = V n 0 = =⇒ z= . (5.156)
z −1−1 1 + N0−1
The chemical potential is then
k T
µ = kB T ln z = −kB T ln 1 + N0−1 ≈ − B → 0− .


(5.157)
N0
In other words, the chemical potential is infinitesimally negative, because N0 is assumed to be thermo-
dynamically large.
250 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

According to eqn. 5.11, the contribution to the pressure from the k = 0 states is

kB T k T
p0 = − ln(1 − z) = B ln(1 + N0 ) → 0+ . (5.158)
V V
So the k = 0 bosons, which we identify as the condensate, contribute nothing to the pressure.
Having separated out the k = 0 mode, we can now replace the remaining sum over α by the usual integral
over k. We then have

n = n0 + g ζ d2 λ−d


T < Tc : T
(5.159)
d −d


p = g ζ 2 +1 kB T λT

and

T > Tc : n = g Li d (z) λ−d

T
2
(5.160)
p = g Li d +1 (z) kB T λ−d
T .
2

The condensate fraction n0 /n is unity at T = 0, when all particles are in the condensate with k = 0, and
decreases with increasing T until T = Tc , at which point it vanishes identically. Explicitly, we have
d/2
g ζ d2


n0 (T ) T
=1− =1− . (5.161)
n n λdT Tc (n)

Let us compute the internal energy E for the ideal Bose gas. We have

∂ ∂Ω ∂Ω
(βΩ) = Ω + β =Ω−T = Ω + TS (5.162)
∂β ∂β ∂T

and therefore
∂
E = Ω + T S + µN = µN + (βΩ)
∂β
 ∂ 
(5.163)
= V µn − (βp)
∂β
= 2 d gV kB T λ−d
1
T Li d +1 (z) .
2

This expression is valid at all temperatures, both above and below Tc . Note that the condensate particles
do not contribute to E, because the k = 0 condensate particles carry no energy.

We now investigate the heat capacity CV,N = ∂E

∂T V,N . Since we have been working in the GCE, it is
very important to note that N is held constant when computing CV,N . We’ll also restrict our attention
to the case d = 3 since the ideal Bose gas does not condense at finite T for d ≤ 2 and d > 3 is unphysical.
While we’re at it, we’ll also set g = 1.
5.7. THE IDEAL BOSE GAS 251

Figure 5.7: Molar heat capacity of the ideal Bose gas (units of R). Note the cusp at T = Tc .

The number of particles is

V λ−3
 3


N0 + ζ
 2 T (T < Tc )
N= (5.164)
V λ−3

T Li3/2 (z) (T > Tc ) ,


and the energy is

3 V
E= kB T Li (z) . (5.165)
2 λ3T 5/2
For T < Tc , we have z = 1 and
 
∂E 15 5

V
CV,N = = 4 ζ 2 kB . (5.166)
∂T V,N λ3T

The molar heat capacity is therefore

CV,N
−1
15 5
R · n λ3T


cV,N (T, n) = NA · = 4 ζ 2 . (5.167)
N
For T > Tc , we have

15 V dT V dz
+ 32 kB T Li3/2 (z) 3 ·

dE V = 4 kB T Li5/2 (z) 3 · , (5.168)
λT T λT z

where we have invoked eqn. 5.52. Taking the differential of N , we have

 3 V dT V dz
dN V = 2 Li3/2 (z) 3 · + Li1/2 (z) 3 · . (5.169)
λT T λT z

We set dN = 0, which fixes dz in terms of dT , resulting in

"5 3
#
3 2 Li5/2 (z) 2 Li3/2 (z)
cV,N (T, z) = 2 R · − . (5.170)
Li3/2 (z) Li1/2 (z)
252 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

To obtain cV,N (T, n), we must invert the relation

n(T, z) = λ−3
T Li3/2 (z) (5.171)

in order to obtain z(T, n), and then insert this into eqn. 5.170. The results are shown in fig. 5.7. There
are several noteworthy features of this plot. First of all, by dimensional analysis the function cV,N (T, n)
is R times a function of the dimensionless ratio T /Tc (n) ∝ T n−2/3 . Second, the high temperature limit
is 23 R, which is the classical value. Finally, there is a cusp at T = Tc (n).
For another example, see §5.11.

5.7.3 Isotherms for the ideal Bose gas

Let a be some length scale and define

2π~2 2π~2
va = a3 , pa = , Ta = (5.172)
ma5 ma2 kB
Then we have
 3/2
va T
= Li3/2 (z) + va n0 (5.173)
v Ta
 5/2
p T
= Li5/2 (z) , (5.174)
pa Ta

where v = V /N is the volume per particle7 and n0 is the condensate number density; n0 vanishes for
T ≥ Tc , where z = 1. One identifies a critical volume vc (T ) by setting z = 1 and n0 = 0, leading to
vc (T ) = va (T /Ta )3/2 . For v < vc (T ), we set z = 1 in eqn. 5.173 to find a relation between v, T , and n0 .
For v > vc (T ), we set n0 = 0 in eqn. 5.173 to relate v, T , and z. Note that the pressure is independent
of volume for T < Tc . The isotherms in the (p, v) plane are then flat for v < vc . This resembles
the coexistence region familiar from our study of the thermodynamics of the liquid-gas transition. The
situation is depicted in Fig. 5.8. In the (T, p) plane, we identify pc (T ) = pa (T /Ta )5/2 as the critical
temperature at which condensation starts to occur.
Recall the Gibbs-Duhem equation,
dµ = −s dT + v dp . (5.175)
Along a coexistence curve, we have the Clausius-Clapeyron relation,
 
dp s − s1 `
= 2 = , (5.176)
dT coex v2 − v1 T ∆v
where ` = T (s2 − s1 ) is the latent heat per mole, and ∆v = v2 − v1 . For ideal gas Bose condensation,
the coexistence curve resembles the red curve in the right hand panel of fig. 5.8. There is no meaning to
the shaded region where p > pc (T ). Nevertheless, it is tempting to associate the curve p = pc (T ) with
the coexistence of the k = 0 condensate and the remaining uncondensed (k 6= 0) bosons8 .
7
Note that in the thermodynamics chapter we used v to denote the molar volume, NA V /N .
8
The k 6= 0 particles are sometimes called the overcondensate.
5.7. THE IDEAL BOSE GAS 253

Figure 5.8: Phase diagrams for the ideal Bose gas. Left panel: (p, v) plane. The solid blue curves are
isotherms, and the green hatched region denotes v < vc (T ), where the system is partially condensed. Right
panel: (p, T ) plane. The solid red curve is the coexistence curve pc (T ), along which Bose condensation
occurs. No distinct thermodynamic phase exists in the yellow hatched region above p = pc (T ).

The entropy in the coexistence region is given by

5
ζ 52
 
 
1 ∂Ω n0
5 5
kB v λ−3 2


s=− = 2 ζ 2 T =
kB 1 − . (5.177)
N ∂T V ζ 32 n

All the entropy is thus carried by the uncondensed bosons, and the condensate carries zero entropy.
The Clausius-Clapeyron relation can then be interpreted as describing a phase equilibrium between the
condensate, for which s0 = v0 = 0, and the uncondensed bosons, for which s0 = s(T ) and v 0 = vc (T ).
So this identification forces us to conclude that the specific volume of the condensate is zero. This is
certainly false in an interacting Bose gas!
While one can identify, by analogy, a ‘latent heat’ ` = T ∆s = T s in the Clapeyron equation, it is
important to understand that there is no distinct thermodynamic phase associated with the region p >
pc (T ). Ideal Bose gas condensation is a second order transition, and not a first order transition.

5.7.4 The λ-transition in Liquid 4 He

Helium has two stable isotopes. 4 He is a boson, consisting of two protons, two neutrons, and two electrons
(hence an even number of fermions). 3 He is a fermion, with one less neutron than 4 He. Each 4 He atom
can be regarded as a tiny hard sphere of mass m = 6.65 × 10−24 g and diameter a = 2.65 Å. A sketch
of the phase diagram is shown in fig. 5.9. At atmospheric pressure, Helium liquefies at Tl = 4.2 K. The
gas-liquid transition is first order, as usual. However, as one continues to cool, a second transition sets in
at T = Tλ = 2.17 K (at p = 1 atm). The λ-transition, so named for the λ-shaped anomaly in the specific
heat in the vicinity of the transition, as shown in fig. 5.10, is continuous (i.e. second order).
254 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.9: Phase diagram of 4 He. All phase boundaries are first order transition lines, with the exception
of the normal liquid-superfluid transition, which is second order. (Source: University of Helsinki)

If we pretend that 4 He is a noninteracting Bose gas, then from the density of the liquid n = 2.2×1022 cm−3 ,
2 3 2/3
we obtain a Bose-Einstein condensation temperature Tc = 2π~


m n/ζ( 2 ) = 3.16 K, which is in the right
ballpark. The specific heat Cp (T ) is found to be singular at T = Tλ , with
 −α
Cp (T ) = A T − Tλ (p) . (5.178)

α is an example of a critical exponent. We shall study the physics of critical phenomena later on in this
course. For now, note that a cusp singularity of the type found in fig. 5.7 corresponds to α = −1. The
behavior of Cp (T ) in 4 He is very nearly logarithmic in |T − Tλ |. In fact, both theory (renormalization
group on the O(2) model) and experiment concur that α is almost zero but in fact slightly negative, with
α = −0.0127 ± 0.0003 in the best experiments (Lipa et al., 2003). The λ transition is most definitely not
an ideal Bose gas condensation. Theoretically, in the parlance of critical phenomena, IBG condensation
and the λ-transition in 4 He lie in different universality classes 9 . Unlike the IBG, the condensed phase in
4 He is a distinct thermodynamic phase, known as a superfluid .

Note that Cp (T < Tc ) for the IBG is not even defined, since for T < Tc we have p = p(T ) and therefore
dp = 0 requires dT = 0.

5.7.5 Fountain effect in superfluid 4 He

At temperatures T < Tλ , liquid 4 He has a superfluid component which is a type of Bose condensate. In
fact, there is an important difference between condensate fraction Nk=0 /N and superfluid density, which
is denoted by the symbol ρs . In 4 He, for example, at T = 0 the condensate fraction is only about 8%,
while the superfluid fraction ρs /ρ = 1. The distinction between N0 and ρs is very interesting but lies
beyond the scope of this course.
9
IBG condensation is in the universality class of the spherical model. The λ-transition is in the universality class of the
XY model.
5.7. THE IDEAL BOSE GAS 255

Figure 5.10: Specific heat of liquid 4 He in the vicinity of the λ-transition. Data from M. J. Buckingham
and W. M. Fairbank, in Progress in Low Temperature Physics, C. J. Gortner, ed. (North-Holland, 1961).
Inset at upper right: more recent data of J. A. Lipa et al., Phys. Rev. B 68, 174518 (2003) performed
in zero gravity earth orbit, to within ∆T = 2 nK of the transition.

One aspect of the superfluid state is its complete absence of viscosity. For this reason, superfluids can flow
through tiny cracks called microleaks that will not pass normal fluid. Consider then a porous plug which
permits the passage of superfluid but not of normal fluid. The key feature of the superfluid component
is that it has zero energy density. Therefore even though there is a transfer of particles across the plug,
there is no energy exchange, and therefore a temperature gradient across the plug can be maintained10 .
The elementary excitations in the superfluid state are sound waves called phonons. They are compres-
sional waves, just like longitudinal phonons in a solid, but here in a liquid. Their dispersion is acoustic,
given by ω(k) = ck where c = 238 m/s.11 The have no internal degrees of freedom, hence g = 1. Like
phonons in a solid, the phonons in liquid helium are not conserved. Hence their chemical potential van-
ishes and these excitations are described by photon statistics. We can now compute the height difference
∆h in a U-tube experiment.

10
Recall that two bodies in thermal equilibrium will have identical temperatures if they are free to exchange energy.
11
The phonon velocity c is slightly temperature dependent.
256 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.11: The fountain effect. In each case, a temperature gradient is maintained across a porous
plug through which only superfluid can flow. This results in a pressure gradient which can result in a
fountain or an elevated column in a U-tube.

Clearly ∆h = ∆p/ρg. so we must find p(T ) for the helium. In the grand canonical ensemble, we have

d3k
Z
ln 1 − e−~ck/kB T


p = −Ω/V = −kB T 3
(2π)
Z∞ (5.179)
(kB T )4 4π 2 −u π 2 (kB T )4
=− du u ln(1 − e ) = .
(~c)3 8π 3 90 (~c)3
0

Let’s assume T = 1 K. We’ll need the density of liquid helium, ρ = 148 kg/m3 .
3
2π 2

dh kB T kB
=
dT 45 ~c ρg
3 (5.180)
2π 2 (1.38 × 10−23 J/K)(1 K) (1.38 × 10−23 J/K)

= × ' 32 cm/K ,
45 (1.055 × 10−34 J · s)(238 m/s) (148 kg/m3 )(9.8 m/s2 )

a very noticeable effect!

5.7.6 Bose condensation in optical traps

The 2001 Nobel Prize in Physics was awarded to Weiman, Cornell, and Ketterle for the experimental
observation of Bose condensation in dilute atomic gases. The experimental techniques required to trap
and cool such systems are a true tour de force, and we shall not enter into a discussion of the details
here12 .
The optical trapping of neutral bosonic atoms, such as 87 Rb, results in a confining potential V (r) which
12
Many reliable descriptions may be found on the web. Check Wikipedia, for example.
5.7. THE IDEAL BOSE GAS 257

is quadratic in the atomic positions. Thus, the single particle Hamiltonian for a given atom is written
~2 2 1
∇ + 2 m ω12 x2 + ω22 y 2 + ω32 z 2 ,


Ĥ = − (5.181)
2m
where ω1,2,3 are the angular frequencies of the trap. This is an anisotropic three-dimensional harmonic
oscillator, the solution of which is separable into a product of one-dimensional harmonic oscillator wave-
functions. The eigenspectrum is then given by a sum of one-dimensional spectra, viz.

En1 ,n2 ,n3 = n1 + 21 ) ~ω1 + n2 + 12 ) ~ω2 + n3 + 12 ) ~ω3 . (5.182)

According to eqn. 5.13, the number of particles in the system is

∞
X ∞
X ∞ h
X i−1
N= y −1 en1 ~ω1 /kB T en2 ~ω2 /kB T en3 ~ω3 /kB T − 1
n1 =0 n2 =0 n3 =0
∞     (5.183)
X
k 1 1 1
= y ,
k=1
1 − e−k~ω1 /kB T 1 − e−k~ω2 /kB T 1 − e−k~ω3 /kB T

where we’ve defined

y ≡ eµ/kB T e−~ω1 /2kB T e−~ω2 /2kB T e−~ω3 /2kB T . (5.184)
Note that y ∈ [0, 1].
Let’s assume that the trap is approximately anisotropic, which entails that the frequency ratios ω1 /ω2
etc. are all numbers on the order of one. Let us further assume that kB T  ~ω1,2,3 . Then

kB T ∗

 k~ωj k<∼ k (T )
1 
≈ (5.185)
1 − e−k~ωj /kB T 
 ∗
1 k > k (T )

where k ∗ (T ) = kB T /~ω̄  1, with

1/3
ω̄ = ω1 ω2 ω3 . (5.186)
We then have
∗ k∗
3 X
y k +1 yk

kB T
N (T, y) ≈ + , (5.187)
1−y ~ω̄ k3
k=1
where the first term on the RHS is due to k > k ∗ and the second term from k ≤ k ∗ in the previous sum.
Since k ∗  1 and since the sum of inverse cubes is convergent, we may safely extend
−1the limit on the
above sum to infinity. To help make more sense of the first term, write N0 = y −1 − 1 for the number
of particles in the (n1 , n2 , n3 ) = (0, 0, 0) state. Then
N0
y= . (5.188)
N0 + 1
This is true always. The issue vis-a-vis Bose-Einstein condensation is whether N0  1. At any rate, we
now see that we can write
kB T 3
 

∗
−1 −k
N ≈ N0 1 + N0 + Li3 (y) . (5.189)
~ω̄
258 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

As for the first term, we have


0
 N0  k ∗

∗
−1 −k
N0 1 + N0 = (5.190)

N0 N0  k∗


Thus, as in the case of IBG condensation of ballistic particles, we identify the critical temperature by the
condition y = N0 /(N0 + 1) ≈ 1, and we have

 1/3  
~ω̄ N ν̄
Tc = = 4.5 N 1/3 [ nK ] , (5.191)
kB ζ(3) 100 Hz

where ν̄ = ω̄/2π. We see that kB Tc  ~ω̄ if the number of particles in the trap is large: N  1. In this
regime, we have

kB T 3
 
T < Tc : N = N0 + ζ(3)
~ω̄
 3 (5.192)
kB T
T > Tc : N= Li3 (y) .]
~ω̄

It is interesting to note that BEC can also occur in two-dimensional traps, which is to say traps which are
very anisotropic, with oblate equipotential surfaces V (r) = V0 . This happens when ~ω3  kB T  ω1,2 .
We then have
6N 1/2
 
(d=2) ~ω̄
Tc = · (5.193)
kB π2

1/2
with ω̄ = ω1 ω2 . The particle number then obeys a set of equations like those in eqns. 5.192, mutatis
mutandis 13 .
For extremely prolate traps, with ω3  ω1,2 , the situation is different because Li1 (y) diverges for y = 1.
We then have
k T

N = N0 + B ln 1 + N0 . (5.194)
~ω3

Here we have simply replaced y by the equivalent expression N0 /(N0 +1). If our criterion for condensation
is that N0 = αN , where α is some fractional value, then we have

~ω3 N
Tc (α) = (1 − α) · . (5.195)
kB ln N
π2
13

3
2
Explicitly, one replaces ζ(3) with ζ(2) = 6
, Li3 (y) with Li2 (y), and kB T /~ω̄ with kB T /~ω̄ .
5.8. THE IDEAL FERMI GAS 259

5.8 The Ideal Fermi Gas

5.8.1 Grand potential and particle number

The grand potential of the ideal Fermi gas is, per eqn. 5.11,
X  
Ω(T, V, µ) = −V kB T ln 1 + eµ/kB T e−εα /kB T
α
Z∞   (5.196)
= −V kB T dε g(ε) ln 1 + e(µ−ε)/kB T .
−∞

The average number of particles in a state with energy ε is

1
n(ε) = , (5.197)
e(ε−µ)/kB T +1

hence the total number of particles is

Z∞
1
N = V dε g(ε) . (5.198)
e(ε−µ)/kB T +1
−∞

5.8.2 The Fermi distribution

We define the function

1
f () ≡ , (5.199)
e/kB T +1
known as the Fermi distribution. In the T → ∞ limit, f () → 21 for all finite values of ε. As T → 0, f ()
approaches a step function Θ(−). The average number of particles in a state of energy ε in a system at
temperature T and chemical potential µ is n(ε) = f (ε − µ). In fig. 5.12 we plot f (ε − µ) versus ε for
three representative temperatures.

5.8.3 T = 0 and the Fermi surface

At T = 0, we therefore have n(ε) = Θ(µ − ε), which says that all single particle energy states up to
ε = µ are filled, and all energy states above ε = µ are empty. We call µ(T = 0) the Fermi energy:
εF = µ(T = 0). If the single particle dispersion ε(k) depends only on the wavevector k, then the locus
of points in k-space for which ε(k) = εF is called the Fermi surface. For isotropic systems, ε(k) = ε(k)
is a function only of the magnitude k = |k|, and the Fermi surface is a sphere in d = 3 or a circle in
d = 2. The radius of this circle is the Fermi wavevector , kF . When there is internal (e.g. spin) degree
of freedom, there is a Fermi surface and Fermi wavevector (for isotropic systems) for each polarization
state of the internal degree of freedom.
260 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

 −1
Figure 5.12: The Fermi distribution, f () = exp(/kB T ) + 1 . Here we have set kB = 1 and taken
1
µ = 2, with T = 20 (blue), T = 34 (green), and T = 2 (red). In the T → 0 limit, f () approaches a step
function Θ(−).

Let’s compute the Fermi wavevector kF and Fermi energy εF for the IFG with a ballistic dispersion
ε(k) = ~2 k2 /2m. The number density is

g k /π (d = 1)
 F




g Ωd kFd
Z d 
dk 
n=g Θ(k F − k) = · = g kF2 /4π (d = 2) (5.200)
(2π)d (2π)d d 





 3
g kF /6π 2 (d = 3) ,

where Ωd = 2π d/2 /Γ(d/2) is the area of the unit sphere in d space dimensions. Note that the form
of n(kF ) is independent of the dispersion relation, so long as it remains isotropic. Inverting the above
expressions, we obtain kF (n):


 πn/g (d = 1)



d n 1/d 
  
kF = 2π = (4πn/g)1/2 (d = 2) (5.201)
g Ωd 





 2
(6π n/g)1/3 (d = 3) .

The Fermi energy in each case, for ballistic dispersion, is therefore

 2 2 2
π ~ n


 2g2 m
(d = 1)



~2 kF2 2π 2 ~2 d n 2/d  2π~2 n
  
εF = = = gm (d = 2) (5.202)
2m m g Ωd 




 ~2 6π2 n
2/3

(d = 3) .

2m g
5.8. THE IDEAL FERMI GAS 261

Another useful result for the ballistic dispersion, which follows from the above, is that the density of
states at the Fermi level is given by

g Ωd mkFd−2 d n
g(εF ) = · = · . (5.203)
(2π)d ~2 2 εF

−1 −1
For the electron gas, we have g = 2. In a metal, one typically has kF ∼ 0.5 Å to 2 Å , and εF ∼
1 eV − 10 eV. Due to the effects of the crystalline lattice, electrons in a solid behave as if they had an
effective mass m∗ which is typically on the order of the electron mass but very often about an order of
magnitude smaller, particularly in semiconductors.
Nonisotropic dispersions ε(k) are more interesting in that they give rise to non-spherical Fermi surfaces.
The simplest example is that of a two-dimensional ‘tight-binding’ model of electrons hopping on a square
lattice, as may be appropriate in certain layered materials. The dispersion relation is then

ε(kx , ky ) = −2t cos(kx a) − 2t cos(ky a) , (5.204)

where kx and ky are confined to the interval − πa , πa . The quantity t has dimensions of energy and is
 

known as the hopping integral . The Fermi surface is the set of points (kx , ky ) which satisfies ε(kx , ky ) =
εF . When εF achieves its minimum value of εmin F = −4t, the Fermi surface collapses to a point at
(kx , ky ) = (0, 0). For energies just above this minimum value, we can expand the dispersion in a power
series, writing
ε(kx , ky ) = −4t + ta2 kx2 + ky2 − 12 1
ta4 kx4 + ky4 + . . . .



(5.205)
If we only work to quadratic order in kx and ky , the dispersion is isotropic, and the Fermi surface is a
circle, with kF2 = (εF + 4t)/ta2 . As the energy increases further, the continuous O(2) rotational invariance
is broken down to the discrete group of rotations of the square, C4v . The Fermi surfaces distort and
eventually, at εF = 0, the Fermi surface is itself
a square. As εF increases further, the square turns back
into a circle, but centered about the point πa , πa . Note that everything is periodic in kx and ky modulo
2π
a . The Fermi surfaces for this model are depicted in the upper right panel of fig. 5.13.

Fermi surfaces in three dimensions can be very interesting indeed, and of great importance in under-
standing the electronic properties of solids. Two examples are shown in the bottom panels of fig. 5.13.
The electronic configuration of cesium (Cs) is [Xe] 6s1 . The 6s electrons ‘hop’ from site to site on a body
centered cubic (BCC) lattice, a generalization of the simple two-dimensional square lattice hopping model
discussed above. The elementary unit cell in k space, known as the first Brillouin zone, turns out to be
a dodecahedron. In yttrium, the electronic structure is [Kr] 5s2 4d1 , and there are two electronic energy
bands at the Fermi level, meaning two Fermi surfaces. Yttrium forms a hexagonal close packed (HCP)
crystal structure, and its first Brillouin zone is shaped like a hexagonal pillbox.

Spin-split Fermi surfaces

Consider an electron gas in an external magnetic field H. The single particle Hamiltonian is then

p2
Ĥ = + µB H σ , (5.206)
2m
262 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.13: Fermi surfaces for two and three-dimensional structures. Upper left: free particles in two
dimensions. Upper right: ‘tight binding’ electrons on a square lattice. Lower left: Fermi surface for
cesium, which is predominantly composed of electrons in the 6s orbital shell. Lower right: the Fermi
surface of yttrium has two parts. One part (yellow) is predominantly due to 5s electrons, while the other
(pink) is due to 4d electrons. (Source: www.phys.ufl.edu/fermisurface/)

where µB is the Bohr magneton,

e~
µB = = 5.788 × 10−9 eV/G
2mc (5.207)
µB /kB = 6.717 × 10−5 K/G ,
where m is the electron mass. What happens at T = 0 to a noninteracting electron gas in a magnetic
field?
Electrons of each spin polarization form their own Fermi surfaces. That is, there is an up spin Fermi
surface, with Fermi wavevector kF↑ , and a down spin Fermi surface, with Fermi wavevector kF↓ . The
individual Fermi energies, on the other hand, must be equal, hence
~2 kF↑
2 ~2 kF↓
2
+ µB H = − µB H , (5.208)
2m 2m
which says
2 2 2eH
kF↓ − kF↑ = . (5.209)
~c
The total density is
3
kF↑ 3
kF↓ 3 3
n= + =⇒ kF↑ + kF↓ = 6π 2 n . (5.210)
6π 2 6π 2
5.8. THE IDEAL FERMI GAS 263

Clearly the down spin Fermi surface grows and the up spin Fermi surface shrinks with increasing H.
Eventually, the minority spin Fermi surface vanishes altogether. This happens for the up spins when
kF↑ = 0. Solving for the critical field, we obtain

~c
1/3
Hc = · 6π 2 n . (5.211)
2e
In real magnetic solids, like cobalt and nickel, the spin-split Fermi surfaces are not spheres, just like the
case of the (spin degenerate) Fermi surfaces for Cs and Y shown in fig. 5.13.

5.8.4 The Sommerfeld expansion

In dealing with the ideal Fermi gas, we will repeatedly encounter integrals of the form
Z∞
I(T, µ) ≡ dε f (ε − µ) φ(ε) . (5.212)
−∞

The Sommerfeld expansion provides a systematic way of expanding these expressions in powers of T and
is an important analytical tool in analyzing the low temperature properties of the ideal Fermi gas (IFG).
We start by defining
Zε
Φ(ε) ≡ dε0 φ(ε0 ) (5.213)
−∞

so that φ(ε) = Φ0 (ε). We then have

Z∞
dΦ
I = dε f (ε − µ)
dε
−∞
(5.214)
Z∞
= − dε f 0 (ε) Φ(µ + ε) ,
−∞

where we assume Φ(−∞) = 0. Next, we invoke Taylor’s theorem, to write

∞
X εn dnΦ
Φ(µ + ε) =
n ! dµn
n=0 (5.215)
 
d
= exp ε Φ(µ) .
dµ

This last expression involving the exponential of a differential operator may appear overly formal but it
proves extremely useful. Since
1 eε/kB T
f 0 (ε) = −
, (5.216)
kB T eε/kB T + 1 2
264 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.14: Deformation of the complex integration contour in eqn. 5.219.

we can write
Z∞
evD
I = dv Φ(µ) , (5.217)
(ev + 1)(e−v + 1)
−∞

with v = ε/kB T , where

d
D = kB T (5.218)
dµ
is a dimensionless differential operator. The integral can now be done using the methods of complex
integration:14
Z∞ ∞
" #
evD X evD
dv = 2πi Res
(ev + 1)(e−v + 1) (ev + 1)(e−v + 1)
−∞ n=1 v=(2n+1)iπ
∞
X (5.219)
= −2πi D e(2n+1)iπD
n=0
2πiD eiπD
=− = πD csc πD .
1 − e2πiD

Thus,
I(T, µ) = πD csc(πD) Φ(µ) , (5.220)
which is to be understood as the differential operator πD csc(πD) = πD/ sin(πD) acting on the function
Φ(µ). Appealing once more to Taylor’s theorem, we have

π2 d2 7π 4 d4
πD csc(πD) = 1 + (kB T )2 2 + (kB T )4 4 + . . . . (5.221)
6 dµ 360 dµ
14
Note that writing v = (2n + 1) iπ +  we have e±v = −1 ∓  − 21 2 + . . . , so (ev + 1)(e−v + 1) = −2 + . . . We then expand
evD = e(2n+1)iπD 1 + D + . . .) to find the residue: Res = −D e(2n+1)iπD .
5.8. THE IDEAL FERMI GAS 265

Thus,
Z∞
I(T, µ) = dε f (ε − µ) φ(ε)
−∞
(5.222)
Zµ
π2 7π 4
= dε φ(ε) + (kB T )2 φ0 (µ) + (k T )4 φ000 (µ) + . . . .
6 360 B
−∞

If φ(ε) is a polynomial function of its argument, then each derivative effectively reduces the order of the
polynomial by one degree, and the dimensionless parameter of the expansion is (T /µ)2 . This procedure
is known as the Sommerfeld expansion.

Chemical potential shift

As our first application of the Sommerfeld expansion formalism, let us compute µ(n, T ) for the ideal
Fermi gas. The number density n(T, µ) is
Z∞
n = dε g(ε) f (ε − µ)
−∞
(5.223)
Zµ
π2
= dε g(ε) + (kB T )2 g 0 (µ) + . . . .
6
−∞

Let us write µ = εF + δµ, where εF = µ(T = 0, n) is the Fermi energy, which is the chemical potential at
T = 0. We then have
εF
Z+δµ
π2
n= dε g(ε) + (kB T )2 g 0 (εF + δµ) + . . .
6
−∞
(5.224)
ZεF
π2
= dε g(ε) + g(εF ) δµ + (kB T )2 g 0 (εF ) + . . . ,
6
−∞

from which we derive

π2 g 0 (εF )
δµ = − (kB T )2 + O(T 4 ) . (5.225)
6 g(εF )
Note that g 0 /g = (ln g)0 . For a ballistic dispersion, assuming g = 2,
Z 3
~2 k 2
  
dk m k(ε) 
g(ε) = 2 δ ε− = 2 2  (5.226)
(2π)3 2m π ~ k(ε)= 1 √2mε
~

Thus, g(ε) ∝ ε1/2 and (ln g)0 = 1

2 ε−1 , so
π 2 (kB T )2
µ(n, T ) = εF − + ... , (5.227)
12 εF
~2 2 2/3 .
where εF (n) = 2m (3π n)
266 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Specific heat

The energy of the electron gas is

Z∞
E
= dε g(ε) ε f (ε − µ)
V
−∞
Zµ
π2 d  
= dε g(ε) ε + (kB T )2 µ g(µ) + . . .
6 dµ
−∞ (5.228)
ZεF
π2 π2
= dε g(ε) ε + g(εF ) εF δµ + (kB T )2 εF g 0 (εF ) + (kB T )2 g(εF ) + . . .
6 6
−∞
π2
= ε0 + (kB T )2 g(εF ) + . . . ,
6
RεF
where ε0 = dε g(ε) ε . is the ground state energy density (i.e. ground state energy per unit volume).
−∞
Thus, to order T 2 ,
π2
 
∂E
CV,N = = V kB2 T g(εF ) ≡ V γ T , (5.229)
∂T V,N 3
π2
kB2 g εF (n) . Note that the molar heat capacity is


where γ(n) = 3

π2 π 2 kB T
 
NA kB T g(εF )
cV = · CV = R· = R, (5.230)
N 3 n 2 εF
where in the last expression on the RHS we have assumed a ballistic dispersion, for which
g(εF ) g mkF 6π 2 3
= 2 2
· 3
= . (5.231)
n 2π ~ g kF 2 εF

The molar heat capacity in eqn. 5.230 is to be compared with the classical ideal gas value of 23 R. Relative
to the classical ideal gas, the IFG value is reduced by a fraction of (π 2 /3) × (kB T /εF ), which in most
metals is very small and even at room temperature is only on the order of 10−2 . Most of the heat capacity
of metals at room temperature is due to the energy stored in lattice vibrations.
A niftier way to derive the heat capacity15 : Starting with eqn. 5.225 for µ(T ) − εF ≡ δµ(T ) , note that
2
g(εF ) = dn/dεF , so we may write δµ = − π6 (kB T )2 (dg/dn) + O(T 4 ) . Next, use the Maxwell relation
(∂S/∂N )T,V = −(∂µ/∂T )N,V to arrive at

π 2 2 ∂g(εF )
 
∂s
= k T + O(T 3 ) , (5.232)
∂n T 3 B ∂n
where s = S/V is the entropy per unit volume. Now use S(T = 0) = 0 and integrate with respect to the
density n to arrive at S(T, V, N ) = V γT , where γ(n) is defined above.
15
I thank my colleague Tarun Grover for this observation.
5.8. THE IDEAL FERMI GAS 267

5.8.5 Magnetic susceptibility

Pauli paramagnetism

Magnetism has two origins: (i) orbital currents of charged particles, and (ii) intrinsic magnetic moment.
The intrinsic magnetic moment m of a particle is related to its quantum mechanical spin via

q~
m = gµ0 S/~ , µ0 = = magneton , (5.233)
2mc

where g is the particle’s g-factor, µ0 its magnetic moment, and S is the vector of quantum mechanical
spin operators satisfying S α , S β = i~αβγ S γ , i.e. SU(2) commutation relations. The Hamiltonian for


a single particle is then

1  q 2
Ĥ = p− A − H ·m
2m∗ c (5.234)
1  e 2 g
= p+ A + µB H σ ,
2m∗ c 2

where in the last line we’ve restricted our attention to the electron, for which q = −e. The g-factor
for an electron is g = 2 at tree level, and when radiative corrections are accounted for using quantum
electrodynamics (QED) one finds g = 2.0023193043617(15). For our purposes we can take g = 2, although
we can always absorb the small difference into the definition of µB , writing µB → µ̃B = ge~/4mc. We’ve
chosen the ẑ-axis in spin space to point in the direction of the magnetic field, and we wrote the eigenvalues
of S z as 12 ~σ, where σ = ±1. The quantity m∗ is the effective mass of the electron, which we mentioned
earlier. An important distinction is that it is m∗ which enters into the kinetic energy term p2 /2m∗ , but
it is the electron mass m itself (m = 511 keV) which enters into the definition of the Bohr magneton. We
shall discuss the consequences of this further below.
In the absence of orbital magnetic coupling, the single particle dispersion is

~2 k2
εσ (k) = + µ̃B H σ . (5.235)
2m∗

At T = 0, we have the results of §5.8.3. At finite T , we once again use the Sommerfeld expansion. We
then have
Z∞ Z∞
n = dε g↑ (ε) f (ε − µ) + dε g↓ (ε) f (ε − µ)
−∞ −∞
Z∞ n o
1
= 2 dε g(ε − µ̃B H) + g(ε + µ̃B H) f (ε − µ) (5.236)
−∞
Z∞ n o
= dε g(ε) + (µ̃B H)2 g 00 (ε) + . . . f (ε − µ) .
−∞
268 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.15: Fermi distributions in the presence of an external Zeeman-coupled magnetic field.

We now invoke the Sommerfeld expension to find the temperature dependence:

Zµ
π2
n = dε g(ε) + (k T )2 g 0 (µ) + (µ̃B H)2 g 0 (µ) + . . .
6 B
−∞
(5.237)
ZεF
π2
= dε g(ε) + g(εF ) δµ + (k T )2 g 0 (εF ) + (µ̃B H)2 g 0 (εF ) + . . . .
6 B
−∞

Note that the density of states for spin species σ is

1
gσ (ε) = 2 g(ε − µ̃B Hσ) , (5.238)

where g(ε) is the total density of states per unit volume, for both spin species, in the absence of a magnetic
field. We conclude that the chemical potential shift in an external field is
 2  0
π 2 2 g (εF )
δµ(T, n, H) = − (kB T ) + (µ̃B H) + ... . (5.239)
6 g(εF )

We next compute the difference n↑ − n↓ in the densities of up and down spin electrons:

Z∞ n o
n↑ − n↓ = dε g↑ (ε) − g↓ (ε) f (ε − µ)
−∞
Z∞ n o (5.240)
1
= 2 dε g(ε − µ̃B H) − g(ε + µ̃B H) f (ε − µ)
−∞
= −µ̃B H · πD csc(πD) g(µ) + O(H 3 ) .
5.8. THE IDEAL FERMI GAS 269

We needn’t go beyond the trivial lowest order term in the Sommerfeld expansion, because H is already
assumed to be small. Thus, the magnetization density is

M = −µ̃B (n↑ − n↓ ) = µ̃2B g(εF ) H . (5.241)

in which the magnetic susceptibility is

 
∂M
χ= = µ̃2B g(εF ) . (5.242)
∂H T,N

This is called the Pauli paramagnetic susceptibility.

Landau diamagnetism

When orbital effects are included, the single particle energy levels are given by
~2 kz2
ε(n, kz , σ) = (n + 12 )~ωc + + µ̃B H σ . (5.243)
2m∗
Here n is a Landau level index, and ωc = eH/m∗ c is the cyclotron frequency. Note that
µ̃B H ge~H m∗ c g m∗
= · = · . (5.244)
~ωc 4mc ~eH 4 m
Accordingly, we define the ratio r ≡ (g/2) × (m∗ /m). We can then write

1 ~2 kz2
+ 12 rσ ~ωc +


ε(n, kz , σ) = n + 2 . (5.245)
2m∗
The grand potential is then given by
Z∞ ∞
HA dkz X X h 1 1 2 2 ∗
i
Ω=− · Lz · kB T ln 1 + eµ/kB T e−(n+ 2 + 2 rσ)~ωc /kB T e−~ kz /2m kB T . (5.246)
φ0 2π
−∞ n=0 σ=±1

A few words are in order here regarding the prefactor. In the presence of a uniform magnetic field, the
energy levels of a two-dimensional ballistic charged particle collapse into Landau levels. The number of
states per Landau level scales with the area of the system, and is equal to the number of flux quanta
through the system: Nφ = HA/φ0 , where φ0 = hc/e is the Dirac flux quantum. Note that

HA V
· Lz · kB T = ~ωc · 3 , (5.247)
φ0 λT
hence we can write
∞ X
X
1
+ 12 rσ)~ωc − µ ,


Ω(T, V, µ, H) = ~ωc Q (n + 2 (5.248)
n=0 σ=±1

where
Z∞
V dkz h 2 2 ∗
i
Q(ε) = − 2 ln 1 + e−ε/kB T e−~ kz /2m kB T . (5.249)
λT 2π
−∞
270 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

We now invoke the Euler-MacLaurin formula,

∞
X Z∞
F (n) = dx F (x) + 21 F (0) − 1
12 F 0 (0) + . . . , (5.250)
n=0 0

resulting in

X
( Z∞
dε Q(ε − µ) + 21 ~ωc Q 1


Ω= 2 (1 + rσ)~ωc − µ
σ=±1 1
2
(1+rσ)~ωc (5.251)
)
1
(~ωc )2 Q0 1


− 12 2 (1 + rσ)~ωc − µ + . . .

We next expand in powers of the magnetic field H to obtain

Z∞
1 2 1
(~ωc )2 Q0 (−µ) + . . . .


Ω(T, V, µ, H) = 2 dε Q(ε − µ) + 4r − 12 (5.252)
0

Thus, the magnetic susceptibility is

1 ∂ 2Ω 2

2
 2 0
∗ 2

χ=− = r − 1
3 · µ̃ B · m/m · − Q (−µ)
V ∂H 2 V
 2 (5.253)
m2

g
= − · µ̃2B · n2 κT ,
4 3m∗ 2

where κT is the isothermal compressibility16 . In most metals we have m∗ ≈ m and the term in brackets is
positive (recall g ≈ 2). In semiconductors, however, we can have m∗  m; for example in GaAs we have
m∗ = 0.067 m . Thus, semiconductors can have a diamagnetic response. If we take g = 2 and m∗ = m, we
see that the orbital currents give rise to a diamagnetic contribution to the magnetic susceptibility which
is exactly − 31 times as large as the contribution arising from Zeeman coupling. The net result is then
paramagnetic (χ > 0) and 23 as large as the Pauli susceptibility. The orbital currents can be understood
within the context of Lenz’s law.
Exercise : Show that − V2 Q0 (−µ) = n2 κT .

5.8.6 Moment formation in interacting itinerant electron systems

The Hubbard model

A noninteracting electron gas exhibits paramagnetism or diamagnetism, depending on the sign of χ,

but never develops a spontaneous magnetic moment: M (H = 0) = 0. What gives rise to magnetism
∂ 2Ω
16
We’ve used − V2 Q0 (µ) = − V1 ∂µ2
= n2 κT .
5.8. THE IDEAL FERMI GAS 271

in solids? Overwhelmingly, the answer is that Coulomb repulsion between electrons is responsible for
magnetism, in those instances in which magnetism arises. At first thought this might seem odd, since
the Coulomb interaction is spin-independent. How then can it lead to a spontaneous magnetic moment?
To understand how Coulomb repulsion leads to magnetism, it is useful to consider a model interacting
system, described by the Hamiltonian

X 
c†iσ cjσ + c†jσ ciσ + U
X X †
Ĥ = −t ni↑ ni↓ + µB H · ciα σαβ ciβ . (5.254)
ij,σ i i,α,β

This is none other than the famous Hubbard model , which has served as a kind of Rosetta stone for
interacting electron systems. The first term describes hopping of electrons along the links of some regular
lattice (the symbol ij denotes a link between sites i and j). The second term describes the local (on-
site) repulsion of electrons. This is a single orbital model, so the repulsion exists when one tries to put
two electrons in the orbital, with opposite spin polarization. Typically the Hubbard U parameter is
on the order of electron volts. The last term is the Zeeman interaction of the electron spins with an
external magnetic field. Orbital effects can be modeled by associating a phase exp(iAij ) to the hopping
matrix element t between sites i and j, where the directed sum of Aij around a plaquette yields the
total magnetic flux through the plaquette in units of φ0 = hc/e. We will ignore orbital effects here.
Note that the interaction term is short-ranged, whereas the Coulomb interaction falls off as 1/|Ri − Rj |.
The Hubbard model is thus unrealistic, although screening effects in metals do effectively render the
interaction to be short-ranged.
Within the Hubbard model, the interaction term is local and written as U n↑ n↓ on any given site. This
term favors a local moment. This is because the chemical potential will fix the mean value of the total
occupancy n↑ + n↓ , in which case it always pays to maximize the difference |n↑ − n↓ |.

Stoner mean field theory

There are no general methods available to solve for even the ground state of an interacting many-body
Hamiltonian. We’ll solve this problem using a mean field theory due to Stoner. The idea is to write the
occupancy niσ as a sum of average and fluctuating terms:

niσ = hniσ i + δniσ . (5.255)

Here, hniσ i is the thermodynamic average; the above equation may then be taken as a definition of the
fluctuating piece, δniσ . We assume that the average is site-independent. This is a significant assumption,
for while we understand why each site should favor developing a moment, it is not clear that all these local
moments should want to line up parallel to each other. Indeed, on a bipartite lattice, it is possible that the
individual local moments on neighboring sites will be antiparallel, corresponding to an antiferromagnetic
order of the pins. Our mean field theory will be one for ferromagnetic states.
272 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

We now write the interaction term as

(flucts )2
z }| {
ni↑ ni↓ = hn↑ i hn↓ i + hn↑ i δni↓ + hn↓ i δni↑ + δni↑ δni↓
(5.256)
= −hn↑ i hn↓ i + hn↑ i ni↓ + hn↓ i ni↑ + O (δn)2

= 14 (m2 − n2 ) + 12 n (ni↑ + ni↓ ) + 12 m (ni↑ − ni↓ ) + O (δn)2

,
where n and m are the average occupancy per spin and average spin polarization, each per unit cell:
n = hn↓ i + hn↑ i
(5.257)
m = hn↓ i − hn↑ i ,

i.e. hnσ i = 21 (n − σm). The mean field grand canonical Hamiltonian K = Ĥ − µN , may then be written
as
X  † 
tij ciσ cjσ + c†jσ ciσ − µ − 12 U n

X †
KMF = − 12 ciσ ciσ
i,j,σ iσ

X † (5.258)
+ µB H + 21 U m σ ciσ ciσ + 41 Nsites U (m2 − n2 ) ,
iσ

where we’ve quantized spins along the direction of H, defined as ẑ. You should take note of two things
here. First, the chemical potential is shifted downward (or the electron energies shifted upward ) by an
amount 12 U n, corresponding to the average energy of repulsion with the background. Second, the effective
magnetic field has been shifted by an amount 21 U m/µB , so the effective field is
Um
Heff = H + . (5.259)
2µB
The bare single particle dispersions are given by εσ (k) = −t̂(k) + σµB H, where
X
t̂(k) = t(R) e−ik·R , (5.260)
R

and tij = t(Ri −Rj ). For nearest neighbor hopping on a d-dimensional cubic lattice, t̂(k) = −t dµ=1 cos(kµ a),
P
where a is the lattice constant. Including the mean field effects, the effective single particle dispersions
become
εeσ (k) = −t̂(k) − 21 U n + µB H + 12 U m σ .


(5.261)

We now solve the mean field theory, by obtaining the free energy per site, ϕ(n, T, H). First, note that
ϕ = ω + µn, where ω = Ω/Nsites is the Landau, or grand canonical, free energy per site. This follows
from the general relation Ω = F − µN ; note that the total electron number is N = nNsites , since n is the
electron number per unit cell (including both spin species). If g(ε) is the density of states per unit cell
(rather than per unit volume), then we have17
Z∞     
1 2 2 1 (µ̄−ε−∆)/kB T (µ̄−ε+∆)/kB T
ϕ= 4 U (m + n ) + µ̄n − 2 kB T dε g(ε) ln 1 + e + ln 1 + e (5.262)
−∞
17
Note that we have written µn = µ̄n + 12 U n2 , which explains the sign of the coefficient of n2 .
5.8. THE IDEAL FERMI GAS 273

where µ̄ ≡ µ − 12 U n and ∆ ≡ µB H + 21 U m. From this free energy we derive two self-consistent equations
for µ and m. The first comes from demanding that ϕ be a function of n and not of µ, i.e. ∂ϕ/∂µ = 0,
which leads to
Z∞ n o
1
n = 2 dε g(ε) f (ε − ∆ − µ̄) + f (ε + ∆ − µ̄) , (5.263)
−∞
 −1
where f (y) = exp(y/kB T ) + 1 is the Fermi function. The second equation comes from minimizing f
with respect to average moment m:
Z∞ n o
1
m= dε g(ε) f (ε − ∆ − µ̄) − f (ε + ∆ − µ̄) .
2 (5.264)
−∞

Here, we will solve the first equation, eq. 5.263, and use the results to generate a Landau expansion of
the free energy ϕ in powers of m2 . We assume that ∆ is small, in which case we may write
Z∞ n o
n = dε g(ε) f (ε − µ̄) + 12 ∆2 f 00 (ε − µ̄) + 1
24 ∆4 f 0000 (ε − µ̄) + . . . . (5.265)
−∞

We write µ̄(∆) = µ̄0 + δ µ̄ and expand in δ µ̄. Since n is fixed in our (canonical) ensemble, we have

Z∞


n = dε g(ε) f ε − µ̄0 , (5.266)
−∞

which defines µ̄0 (n, T ).18 The remaining terms in the δ µ̄ expansion of eqn. 5.265 must sum to zero. This
yields

D(µ̄0 ) δ µ̄ + 21 ∆2 D0 (µ̄0 ) + 12 (δ µ̄)2 D0 (µ̄0 ) + 21 D00 (µ̄0 ) ∆2 δ µ̄ + 1

24 D000 (µ̄0 ) ∆4 + O(∆6 ) = 0 , (5.267)

where
Z∞
D(µ) = − dε g(ε) f 0 (ε − µ) (5.268)
−∞

is the thermally averaged bare density of states at energy µ. Note that the k th derivative is
Z∞
D (k)
(µ) = − dε g (k) (ε) f 0 (ε − µ) . (5.269)
−∞

Solving for δ µ̄, we obtain

δ µ̄ = − 21 a1 ∆2 − 1
3a31 − 6a1 a2 + a3 ∆4 + O(∆6 ) ,


24 (5.270)
18
The Gibbs-Duhem relation guarantees that such an equation of state exists, relating any three intensive thermodynamic
quantities.
274 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

where
D(k) (µ̄0 )
ak ≡ . (5.271)
D(µ̄0 )

After integrating by parts and inserting this result for δ µ̄ into our expression for the free energy f , we
obtain the expansion
 0 2 !
D (µ̄ 0 )
ϕ(n, T, m) = ϕ0 (n, T ) + 41 U m2 − 12 D(µ̄0 ) ∆2 + 18 − 13 D00 (µ̄0 ) ∆4 + . . . , (5.272)
D(µ̄0 )

where prime denotes differentiation with respect to argument, at m = 0, and

Z∞
1 2


ϕ0 (n, T ) = 4Un + nµ̄0 − dε N (ε) f ε − µ̄0 , (5.273)
−∞

where g(ε) = N 0 (ε), so N (ε) is the integrated bare density of states per unit cell in the absence of any
magnetic field (including both spin species).
We assume that H and m are small, in which case

U χ0
ϕ = ϕ0 + 12 am2 + 41 bm4 − 12 χ0 H 2 − Hm + . . . , (5.274)
2µB

where χ0 = µ2B D(µ̄0 ) is the Pauli susceptibility, and

!
(D0 )2 1 00
a= 1
2U 1− 1
2 U D) , b= 1
32 −3D U4 , (5.275)
D

where the argument of each D(k) above is µ̄0 (n, T ). The magnetization density (per unit cell, rather than
per unit volume) is given by
∂ϕ U χ0
M =− = χ0 H + m. (5.276)
∂H 2µB
Minimizing with respect to m yields

U χ0
am + bm3 − H=0, (5.277)
2µB

which gives, for small m,

χ0 H
m= . (5.278)
µB 1 − 12 U D
We therefore obtain M = χ H with
χ0
χ= , (5.279)
U
1− Uc

where
2
Uc = . (5.280)
D(µ̄0 )
5.8. THE IDEAL FERMI GAS 275

Figure 5.16: A graduate student experiences the Stoner enhancement.

The denominator of χ increases the susceptibility above the bare Pauli value χ0 , and is referred to as – I
kid you not – the Stoner enhancement (see Fig. 5.16).
It is worth emphasizing that the magnetization per unit cell is given by

1 δ Ĥ
M =− = µB m . (5.281)
Nsites δH

This is an operator identity and is valid for any value of m, and not only small m.
When H = 0 we can still get a magnetic moment, provided U > Uc . This is a consequence of the simple
Landau theory we have derived. Solving for m when H = 0 gives m = 0 when U < Uc and

 1/2 p
U
m(U ) = ± U − Uc , (5.282)
2b Uc

when U > Uc , and assuming b > 0. Thus we have the usual mean field order parameter exponent of
β = 21 .

Antiferromagnetic solution

In addition to ferromagnetism, there may be other ordered states which solve the mean field theory. One
such example is antiferromagnetism. On a bipartite lattice, the antiferromagnetic mean field theory is
obtained from

hniσ i = 21 n + 12 σ eiQ·Ri m , (5.283)

276 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

where Q = (π/a, π/a, . . . , π/a) is the antiferromagnetic ordering wavevector. The grand canonical Hamil-
tonian is then
X  † 
tij ciσ cjσ + c†jσ ciσ − µ − 12 U n

X †
KMF = − 12 ciσ ciσ
i,j,σ iσ

σ c†iσ ciσ
X
iQ·Ri 2
+ 1
2Um e + 1
4 Nsites U (m − n2 ) (5.284)
iσ
!
X  ε(k) − µ + 1 U n 1
σ Um

ck,σ
= 1
2 c†k,σ †
ck+Q,σ 1
2 2
1
kσ 2σ Um ε(k + Q) − µ + 2 U n ck+Q,σ
+ 14 Nsites U (m2 − n2 ) , (5.285)
where ε(k) = −t̂(k), as before. On a bipartite lattice, with nearest neighbor hopping only, we have
ε(k+Q) = −ε(k). The above matrix is diagonalized by a unitary transformation, yielding the eigenvalues
p
λ± = ± ε2 (k) + ∆2 − µ̄ (5.286)
with ∆ = 21 U m and µ̄ = µ − 12 U n as before. The free energy per unit cell is then
ϕ = 41 U (m2 + n2 ) + µ̄n (5.287)
∞
√ √
Z     
− 21 kB T dε g(ε) ln 1 + e(µ̄− ε +∆ )/kB T + ln 1 + e(µ̄+ ε +∆ )/kB T
2 2 2 2
.
−∞

The mean field equations are then

Z∞ n  p  p o
n = 12 dε g(ε) f − ε2 + ∆2 − µ̄ + f ε2 + ∆2 − µ̄ (5.288)
−∞
Z∞
1 1 g(ε) n  p  p o
= 2 dε √ f − ε2 + ∆2 − µ̄ − f ε2 + ∆2 − µ̄ . (5.289)
U ε2 + ∆ 2
−∞

As in the case of the ferromagnet, a paramagnetic solution with m = 0 always exists, in which case the
second of the above equations is no longer valid.

Mean field phase diagram of the Hubbard model

Let us compare the mean field theories for the ferromagnetic and antiferromagnetic states at T = 0 and
H = 0. Due to particle-hole symmetry, we may assume 0 ≤ n ≤ 1 without loss of generality. (The
solutions repeat themselves under n → 2 − n.) For the paramagnet, we have
Zµ̄
n = dε g(ε) (5.290)
−∞
Zµ̄
1 2
ϕ= 4Un + dε g(ε) ε , (5.291)
−∞
5.8. THE IDEAL FERMI GAS 277

with µ̄ = µ − 21 U n is the ‘renormalized’ Fermi energy and g(ε) is the density of states per unit cell in the
absence of any explicit (H) or implicit (m) symmetry breaking, including both spin polarizations.
For the ferromagnet,
µ̄−∆
Z µ̄+∆
Z
1 1
n= 2 dε g(ε) + 2 dε g(ε) (5.292)
−∞ −∞
µ̄+∆
Z
4∆
= dε g(ε) (5.293)
U
µ̄−∆
µ̄−∆ µ̄+∆
∆2
Z Z
1 2
ϕ= 4Un − + dε g(ε) ε + dε g(ε) ε . (5.294)
U
−∞ −∞

Here, ∆ = 21 U m is nonzero in the ordered phase.

Finally, the antiferromagnetic mean field equations are
Z∞ Z∞
nµ̄<0 = dε g(ε) ; nµ̄>0 = 2 − dε g(ε) (5.295)
ε0 ε0
Z∞
2 g(ε)
= dε √ (5.296)
U ε2 + ∆ 2
ε0

2 Z∞
∆ p
ϕ = 14 U n2 + − dε g(ε) ε2 + ∆2 , (5.297)
U
ε0
p
where ε0 = µ̄2 − ∆2 and ∆ = 21 U m as before. Note that |µ̄| ≥ ∆ for these solutions. Exactly at
half-filling, we have n = 1 and µ̄ = 0. We then set ε0 = 0.
The paramagnet to ferromagnet transition  may be first or second order, depending on the details of
g(ε). If second order, it occurs at Uc = 1 g(µ̄P ), where µ̄P (n) is the paramagnetic solution for µ̄. The
F

paramagnet to antiferromagnet transition is always second order in this mean field

∞theory, since the RHS
dε g(ε) ε−1 . Note that
R
of eqn. (5.296) is a monotonic function of ∆. This transition occurs at UcA = 2
µ̄P
UcA → 0 logarithmically for n → 1, since µ̄P = 0 at half-filling.
For large U , the ferromagnetic solution always has the lowest energy, and therefore if UcA < UcF , there
will be a first-order antiferromagnet to ferromagnet transition at some value U ∗ > UcF . In fig. 5.17, I
plot the phase diagram√ obtained by solving the mean field equations assuming a semicircular density of
states g(ε) = π2 W −2 W 2 − ε2 . Also shown is the phase diagram for the d = 2 square lattice Hubbard
model obtained by J. Hirsch (1985).
How well does Stoner theory describe the physics of the Hubbard model? Quantum Monte Carlo calcula-
tions by J. Hirsch (1985) found that the actual phase diagram of the d = 2 square lattice Hubbard Model
278 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.17: Mean field phase diagram of the Hubbard model, including paramagnetic (P), ferromagnetic
(F), and antiferromagnetic (A) phases. Left panel: results using a semicircular density of states function
of half-bandwidth W . Right panel: results using a two-dimensional square lattice density of states with
nearest neighbor hopping t, from J. E. Hirsch, Phys. Rev. B 31, 4403 (1985). The phase boundary
between F and A phases is first order.

exhibits no ferromagnetism for any n up to U = 10. Furthermore, he found the antiferromagnetic phase
to be entirely confined to the vertical line n = 1. For n 6= 1 and 0 ≤ U ≤ 10, the system is a param-
agnet19 . These results were state-of-the art at the time, but both computing power as well as numerical
algorithms for interacting quantum systems have advanced considerably since 1985. Yet as of 2018, we
still don’t have a clear understanding of the d = 2 Hubbard model’s T = 0 phase diagram! There is
an emerging body of numerical evidence20 that in the underdoped (n < 1) regime, there are portions of
the phase diagram which exhibit a stripe ordering, in which antiferromagnetic order is interrupted by a
parallel array of line defects containing excess holes (i.e. the absence of an electron)21 . This problem has
turned out to be unexpectedly rich, complex, and numerically difficult to resolve due to the presence of
competing ordered states, such as d-wave superconductivity and spiral magnetic phases, which lie nearby
in energy with respect to the putative stripe ground state.
In order to achieve a ferromagnetic solution, it appears necessary to introduce geometric frustration,
either by including a next-nearest-neighbor hopping amplitude t0 or by defining the model on non-bipartite
lattices. Numerical work by M. Ulmke (1997) showed the existence of a ferromagnetic phase at T = 0 on
19
A theorem due to Nagaoka establishes that the ground state is ferromagnetic for the case of a single hole in the U = ∞
system on bipartite lattices.
20
See J. P. F. LeBlanc et al., Phys. Rev. X 5, 041041 (2015) and B. Zheng et al., Science 358, 1155 (2017).
21
The best case for stripe order has been made at T = 0, U/t = 8, and hold doping x = 81 (i.e. n = 87 ).
5.8. THE IDEAL FERMI GAS 279

the FCC lattice Hubbard model for U = 6 and n ∈ [0.15, 0.87] (approximately).

5.8.7 White dwarf stars

There is a nice discussion of this material in R. K. Pathria, Statistical Mechanics. As a model, consider
a mass M ∼ 1033 g of helium at nuclear densities of ρ ∼ 107 g/cm3 and temperature T ∼ 107 K. This
temperature is much larger than the ionization energy of 4 He, hence we may safely assume that all helium
atoms are ionized. If there are N electrons, then the number of α particles (i.e. 4 He nuclei) must be 12 N .
The mass of the α particle is mα ≈ 4mp . The total stellar mass M is almost completely due to α particle
cores.
The electron density is then

N 2 · M/4mp ρ
n= = = ≈ 1030 cm−3 , (5.298)
V V 2mp

since M = N · me + 12 N · 4mp . From the number density n we find for the electrons

kF = (3π 2 n)1/3 = 2.14 × 1010 cm−1

pF = ~kF = 2.26 × 10−17 g cm/s (5.299)
−28 10 −17
mc = (9.1 × 10 g)(3 × 10 m/s) = 2.7 × 10 g cm/s .

Since pF ∼ mc, we conclude that the electrons are relativistic. The Fermi temperature will then be
TF ∼ mc2 ∼ 106 eV ∼ 1012 K. Thus, T  Tf which says that the electron gas is degenerate and may
be considered to be at T ∼ 0. So we need to understand the ground state properties of the relativistic
electron gas.
The kinetic energy is given by p
ε(p) = p2 c2 + m2 c4 − mc2 . (5.300)
The velocity is
∂ε pc2
v= =p . (5.301)
∂p p2 c2 + m2 c4
The pressure in the ground state is

p0 = 31 nhp · vi
ZpF
1 p2 c2
= 2 3 dp p2 · p
3π ~ p2 c2 + m2 c4
0
ZθF (5.302)
m4 c5 4
= 2 3 dθ sinh θ
3π ~
0
m4 c5

= 2 3
sinh(4θF ) − 8 sinh(2θF ) + 12 θF ,
96π ~
280 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

Figure 5.18: Mass-radius relationship for white dwarf stars. (Source: Wikipedia).

where we use the substitution

 
1 c+v
p = mc sinh θ , v = c tanh θ =⇒ θ= 2 ln . (5.303)
c−v

Note that pF = ~kF = ~(3π 2 n)1/3 , and that

M 9π M
n= =⇒ 3π 2 n = . (5.304)
2mp V 8 R 3 mp

Now in equilibrium the pressure p is balanced by gravitational pressure. We have

dE0 = −p0 dV = −p0 (R) · 4πR2 dR . (5.305)

This must be balanced by gravity:

GM 2
dEg = γ · dR , (5.306)
R2
where γ depends on the radial mass distribution. Equilibrium then implies

γ GM 2
p0 (R) = . (5.307)
4π R4

To find the relation R = R(M ), we must solve

γ gM 2 m4 c5

4
= 2 3
sinh(4θF ) − 8 sinh(2θF ) + 12 θF . (5.308)
4π R 96π ~
5.9. APPENDIX I : SECOND QUANTIZATION 281

Note that 
96 5
 15 θF
 θF → 0
sinh(4θF ) − 8 sinh(2θF ) + 12θF = (5.309)

1
2 e4θF θF → ∞ .
Thus, we may write
  5/3
~2 9π M

2
 15π m 8
 R3 mp
θF → 0
γ gM 2 
p0 (R) = = (5.310)
4π R4 
  4/3
9π M
~c
θF → ∞ .


12π 2 8 R 3 mp

In the limit θF → 0, we solve for R(M ) and find

~2
R= 3
40γ (9π)2/3 5/3
∝ M −1/3 . (5.311)
G mp m M 1/3

In the opposite limit θF → ∞, the R factors divide out and we obtain

 1/2  3/2
9 3π ~c 1
M = M0 = . (5.312)
64 γ3 G m2p

To find the R dependence, we must go beyond the lowest order expansion of eqn. 5.309, in which case
we find
 1/3   "  #1/2
M 1/3 M 2/3
 
9π ~
R= 1− . (5.313)
8 mc mp M0

The value M0 is the limiting size for a white dwarf. It is called the Chandrasekhar limit.

5.9 Appendix I : Second Quantization

5.9.1 Basis states and creation/annihilation operators

Second quantization is a convenient scheme to label basis states of a many particle quantum system. We
are ultimately interested in solutions of the many-body Schrödinger equation,

ĤΨ(x1 , . . . , xN ) = E Ψ(x1 , . . . , xN ) (5.314)

where the Hamiltonian is

N N
~2 X 2 X
Ĥ = − ∇i + V (xj − xk ) . (5.315)
2m
i=1 j<k

To the coordinate labels {x1 , . . . xN } we may also append labels for internal degrees of freedom, such as
 
spin polarization, denoted {ζ1 , . . . , ζN }. Since Ĥ, σ = 0 for all permutations σ ∈ SN , the many-body
282 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

wavefunctions may be chosen to transform according to irreducible representations of the symmetric

group SN . Thus, for any σ ∈ SN ,
 

1
Ψ xσ(1) , . . . , xσ(N ) = Ψ(x1 , . . . , xN ) , (5.316)
sgn(σ)

where the upper choice is for Bose-Einstein statistics and the lower sign for Fermi-Dirac statistics. Here
xj may include not only the spatial coordinates of particle j, but its internal quantum number(s) as well,
such as ζj .

A convenient basis for the many body states is obtained from the single-particle eigenstates |αi of
some single-particle Hamiltonian

 0 Ĥ0 , with h x | α i = ϕα (x) and Ĥ0 |αi = εα |αi. The basis may be taken
as orthonormal, i.e. α α = δαα0 . Now define


1 X  1

Ψα1 ,...,α (x1 , . . . , xN ) = p ϕα (x1 ) · · · ϕα (xN ) . (5.317)
sgn(σ)
Q
N
N! α nα ! σ∈S
σ(1) σ(N )
N

Here nα is the number of times the index α appears among the set {α1 , . . . , αN }. For BE statistics,
nα ∈ {0, 1, 2, . . .} , whereas for FD statistics, nα ∈ {0, 1} . Note that the above states are normalized22 :
Z Z X  N Z
 Y
2 1 1
d d
ddxj ϕ∗α

d x1 · · · d xN Ψα1 ···α (x1 , . . . , xN ) = Q (xj ) ϕα (xj )
N N! α nα ! σ,µ∈S sgn(σµ) σ(j) µ(j)
N
j=1
N
1 X Y
=Q δαj ,α =1 . (5.318)
α nα !
σ(j)
σ∈SN j=1

Note that
X 
ϕα (x1 ) · · · ϕα (xN ) ≡ per ϕαi (xj )
σ(1) σ(N )
σ∈SN
X  (5.319)
sgn(σ) ϕα (x1 ) · · · ϕα (xN ) ≡ det ϕαi (xj ) ,
σ(1) σ(N )
σ∈SN

which stand for permanent and determinant, respectively. We may now write

 
Ψα1 ··· α (x1 , . . . , xN ) = x1 , · · · , xN  α1 · · · αN , (5.320)
N

where
  1 X  1      
 α1 · · · αN = p Q α
σ(1) ⊗ ασ(2) ⊗ · · · ⊗ ασ(N ) . (5.321)
 
N ! α nα ! σ∈S sgn(σ)
N

Note that | ασ(1) · · · ασ(N ) i = (±1)σ | α1 · · · αN i , where by (±1)σ we mean 1 in the case of BE statistics
and sgn (σ) in the case of FD statistics.
22
ddx implicitly includes a sum
R P
In the normalization integrals, each ζ over any internal indices that may be present.
5.9. APPENDIX I : SECOND QUANTIZATION 283

We may express | α1 · · · αN i as a product of creation operators acting on a vacuum | 0 i in Fock space.

For bosons,
  Y (b†α )nα    
 α1 · · · αN = p  0 ≡  {nα } , (5.322)
α nα !
with
b†α , b†β = 0 bα , b†β = δαβ
     
bα , bβ = 0 , , , (5.323)
where [ • , • ] is the commutator. For fermions,

 α1 · · · αN = c† c† · · · c†  0 ≡  {nα }
     
α α α , (5.324)
1 2 N

with
c†α , c†β = 0 cα , c†β = δαβ
  
cα , cβ = 0 , , , (5.325)
where {• , •} is the anticommutator.

5.9.2 Second quantized operators

2 PN 2
Now consider the action of permutation-symmetric first quantized operators such as T̂ = − 2m
~
i=1 ∇i
PN
and V̂ = i<j v̂(xi − xj ). For a one-body operator such as T̂ , we have
Z Z Y −1/2  Y −1/2
α1 · · · αN  T̂  α10 · · · αN
0
ddx1 · · · ddxN n0α !

  
= nα ! × (5.326)
α α
X N
X
(±1)σ ϕ∗α (x1 ) · · · ϕ∗α (xN ) T̂i ϕα0 (x1 ) · · · ϕα0
σ(1) σ(N ) σ(1) σ(N )
σ∈SN k=1

X Y N Y
−1/2 X Z
= (±1)σ nα ! n0α ! δα 0 ddx1 ϕ∗αi (x1 ) T̂1 ϕα0 (x1 ) .
j ,ασ(j) σ(i)
σ∈SN α i=1 j
(j6=i)

One may verify that any permutation-symmetric one-body operator such as T̂ is faithfully represented
by the second quantized expression,
X
  
T̂ = α  T̂  β ψα† ψβ , (5.327)
α,β

where ψα† is b†α or c†α as the application determines, and

Z
α  T̂  β = ddx1 ϕ∗α (x1 ) T̂1 ϕβ (x1 )

  
. (5.328)

Similarly, two-body operators such as V̂ are represented as

αβ  V̂  γδ ψα† ψβ† ψδ ψγ
X

V̂ = 12
  
, (5.329)
α,β,γ,δ
284 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

where Z Z
d
ddx2 ϕ∗α (x1 ) ϕ∗β (x2 ) v(x1 − x2 ) ϕδ (x2 ) ϕγ (x1 ) .

  
αβ  V̂  γδ = d x1 (5.330)

The general form for an n-body operator is then

1 X

α1 · · · αn  R̂  β1 · · · βn ψα† n · · · ψα† n ψβ · · · ψβ1

  
R̂ = . (5.331)
n! α ··· α n
1 n
β1 ··· βn

PN
Finally, if the Hamiltonian is noninteracting, consisting solely of one-body operators Ĥ = i=1 ĥi , then
X
Ĥ = εα ψα† ψα , (5.332)
α

where {εα } is the spectrum of each single particle Hamiltonian ĥi .

5.10 Appendix II : Ideal Bose Gas Condensation

We begin with the grand canonical Hamiltonian K = H − µN for the ideal Bose gas,
√ X
(εk − µ) b†k bk − N νk b†k + ν̄k bk
X

K= . (5.333)
k k

Here b†k is the creation operator for a boson in a state of wavevector k, hence bk , b†k0 = δkk0 . The
 

dispersion relation is given by the function εk , which is the energy of a particle with wavevector k. We
must have εk − µ ≥ 0 for all k, lest the spectrum of K be unbounded from below. The fields {νk , ν̄k }
break a global O(2) symmetry.
Students who have not taken a course in solid state physics can skip the following paragraph, and be
aware that N = V /v0 is the total volume of the system in units of a fundamental ”unit cell” volume.
The thermodynamic limit is then N → ∞. Note that N is not the boson particle number, which we’ll
call Nb .
Solid state physics boilerplate : We presume a setting in which the real space Hamiltonian is defined
by some boson hopping model on a Bravais lattice. The wavevectors k are then restricted to the first
Brillouin zone,
Ω̂, and assuming periodic boundary conditions are quantized according to the condition
exp iNl k · al = 1 for all l ∈ {1, . . . , d}, where al is the lth fundamental direct lattice vector and Nl is
Q of the system in the al direction; d isP
the size the dimension of space. The total number of unit cells is
N ≡ l Nl . Thus, quantization entails k = l (2πnl /Nl ) bl , where bl is the lth elementary reciprocal
lattice vector (al · bl0 = 2πδll0 ) and nl ranges over Nl distinct integers such that the allowed k points form
a discrete approximation to Ω̂ .
To solve, we first shift the boson creation and annihilation operators, writing
X |ν |2
(εk − µ) βk† βk − N
X
k
K= , (5.334)
εk − µ
k k
5.10. APPENDIX II : IDEAL BOSE GAS CONDENSATION 285

where √ √
N νk N ν̄k
βk = bk − , βk† = b†k − . (5.335)
εk − µ εk − µ
Note that βk , βk† 0 = δkk0 so the above transformation is canonical. The Landau free energy Ω =
 

−kB T ln Ξ , where Ξ = Tr e−K/kB T , is given by

Z∞ X |ν |2
Ω = N kB T dε g(ε) ln 1 − e(µ−ε)/kb T − N k


, (5.336)
k ε k−µ
−∞

where g(ε) is the density of energy states per unit cell,

Z d
1 X
dk

g(ε) = δ ε − εk −−−−→ d
δ ε − εk . (5.337)
N N →∞ (2π)
k
Ω̂

Note that
1
 1 ∂Ω νk
ψk ≡ √ bk = − = . (5.338)
N N ∂ ν̄k ε k−µ

In the condensed phase, ψk is nonzero.

The Landau free energy (grand potential) is a function Ω(T, N, µ, ν, ν̄). We now make a Legendre
transformation, X

Y (T, N, µ, ψ, ψ̄) = Ω(T, N, µ, ν, ν̄) + N νk ψ̄k + ν̄k ψk . (5.339)
k
Note that
∂Y ∂Ω
= + N ψk = 0 , (5.340)
∂ ν̄k ∂ ν̄k
by the definition of ψk . Similarly, ∂Y /∂νk = 0. We now have
Z∞ X
Y (T, N, µ, ψ, ψ̄) = N kB T dε g(ε) ln 1 − e(µ−ε)/kb T + N (εk − µ) |ψk |2


. (5.341)
−∞ k

Therefore, the boson particle number per unit cell is given by the dimensionless density,
Z∞
N 1 ∂Y X g(ε)
n= b =− = |ψk |2 + dε , (5.342)
N N ∂µ e(ε−µ)/kB T −1
k −∞

and the condensate amplitude at wavevector k is

1 ∂Y
νk = = (εk − µ) ψk . (5.343)
N ∂ ψ̄k

Recall that νk acts as an external field. Let the dispersion εk be minimized at k = K . Without loss of
generality, we may assume this minimum value is εK = 0 . We see that if νk = 0 then one of two must
be true:
286 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

(i) ψk = 0 for all k

(ii) µ = εK , in which case ψK can be nonzero.

Thus, for ν = ν̄ = 0 and µ > 0, we have the usual equation of state,

Z∞
g(ε)
n(T, µ) = dε , (5.344)
−∞
e(ε−µ)/kB T −1

which relates the intensive variables n, T , and µ. When µ = 0, the equation of state becomes
n> (T )
n0 z }| {
z }| { Z∞
X g(ε)
n(T, µ = 0) = |ψK |2 + dε , (5.345)
e ε/k BT − 1
K −∞

where now the sum is over only those K for which εK = 0 . Typically this set has only one member,
K = 0, but it is quite possible, due to symmetry reasons, that there are more such K values. This last
equation of state is one which relates the intensive variables n, T , and n0 , where
X
n0 = |ψK |2 (5.346)
K

is the dimensionless condensate density. If the integral n> (T ) in eqn. 5.345 is finite, then for n > n0 (T )
we must have n0 > 0. Note that, for any T , n> (T ) diverges logarithmically whenever g(0) is finite. This
means that eqn. 5.344 can always be inverted to yield a finite µ(n, T ), no matter how large the value of
n, in which case there is no condensation and n0 = 0. If g(ε) ∝ εα with α > 0, the integral converges and
n> (T ) is finite and monotonically increasing for all T . Thus, for fixed dimensionless number n, there will
be a critical temperature Tc for which n = n> (Tc ). For T < Tc , eqn. 5.344 has no solution for any µ and
we must appeal to eqn. 5.345. The condensate density, given by n0 (n, T ) = n − n> (T ) , is then finite for
T < Tc , and vanishes for T ≥ Tc .
In the condensed phase, the phase of the order parameter ψ inherits its phase from the external field
ν, which is taken to zero, in the same way the magnetization in the symmetry-broken phase of an Ising
ferromagnet inherits its direction from an applied field h which is taken to zero. The important feature
is that in both cases the applied field is taken to zero after the approach to the thermodynamic limit.

5.11 Appendix III : Example Bose Condensation Problem

PROBLEM: A three-dimensional gas of noninteracting bosonic particles obeys the dispersion relation ε(k) =
 1/2
A k  .

(a) Obtain an expression for the density n(T, z) where z = exp(µ/kB T ) is the fugacity. Simplify your
expression as best you can, adimensionalizing any integral or infinite sum which may appear. You
5.11. APPENDIX III : EXAMPLE BOSE CONDENSATION PROBLEM 287

may find it convenient to define

Z∞ ∞
1 tν−1 X zk
Liν (z) ≡ dt = . (5.347)
Γ(ν) z −1 et − 1 kν
0 k=1

Note Liν (1) = ζ(ν), the Riemann zeta function.

(b) Find the critical temperature for Bose condensation, Tc (n). Your expression should only include
the density n, the constant A, physical constants, and numerical factors (which may be expressed
in terms of integrals or infinite sums).
1
(c) What is the condensate density n0 when T = 2 Tc ?

(d) Do you expect the second virial coefficient to be positive or negative? Explain your reasoning. (You
don’t have to do any calculation.)

SOLUTION: We work in the grand canonical ensemble, using Bose-Einstein statistics.

(a) The density for Bose-Einstein particles are given by

Z 3
dk 1
n(T, z) = 3 −1
(2π) z exp(Ak 1/2 /kB T ) − 1
 Z∞
1 kB T 6 s5

= 2 ds −1 s (5.348)
π A z e −1
0
120 kB T 6
 
= 2 Li6 (z) ,
π A

where we have changed integration variables from k to s = Ak 1/2 /kB T , and we have defined the
functions Liν (z) as above, in eqn. 5.347. Note Liν (1) = ζ(ν), the Riemann zeta function.

(b) Bose condensation sets in for z = 1, i.e. µ = 0. Thus, the critical temperature Tc and the density
n are related by
120 ζ(6) kB Tc 6
 
n= , (5.349)
π2 A
or
π 2 n 1/6
 
A
Tc (n) = . (5.350)
kB 120 ζ(6)

(c) For T < Tc , we have

120 ζ(6) kB T 6
 
n = n0 +
π2 A
 6 (5.351)
T
= n0 + n,
Tc
288 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS

1
where n0 is the condensate density. Thus, at T = 2 Tc ,

n0 T = 12 Tc = 63


64 n. (5.352)

(d) The virial expansion of the equation of state is

 
p = nkB T 1 + B2 (T ) n + B3 (T ) n2 + . . . . (5.353)

We expect B2 (T ) < 0 for noninteracting bosons, reflecting the tendency of the bosons to condense.
(Correspondingly, for noninteracting fermions we expect B2 (T ) > 0.)
For the curious, we compute B2 (T ) by eliminating the fugacity z from the equations for n(T, z) and
p(T, z). First, we find p(T, z):
Z 3
dk 
1/2

p(T, z) = −kB T ln 1 − z exp(−Ak /k B T )
(2π)3
 Z∞
kB T kB T 6

ds s5 ln 1 − z e−s


=− 2 (5.354)
π A
0
 6
120 kB T kB T
= Li7 (z).
π2 A

Expanding in powers of the fugacity, we have

6 n
z2 z3

120 kB T o
n= 2 z+ 6 + + . . .
π A2 36
6 n (5.355)
z2 z3

p 120 kB T o
= 2 z+ 7 + + . . . .
kB T π A 2 37

Solving for z(n) using the first equation, we obtain, to order n2 ,

2
π 2 A6 n π 2 A6 n
  
1
z= − 6 + O(n3 ) . (5.356)
120 (kB T )6 2 120 (kB T )6

Plugging this into the equation for p(T, z), we obtain the first nontrivial term in the virial expansion,
with
π2 A 6
 
B2 (T ) = − , (5.357)
15360 kB T
which is negative, as expected. Note that the ideal gas law is recovered for T → ∞, for fixed n.