SQM PDF

Lecture Notes in Statistical Mechanics and Mesoscopics
Doron Cohen
Department of Physics, Ben-Gurion University, Beer-Sheva 84105, Israel
(arXiv:1107.0568)
These are the lecture notes for quantum and statistical mechanics courses that are given by
DC at Ben-Gurion University. They are complementary to Lecture Notes in Quantum Mechanics
[arXiv:quant-ph/0605180]. Some additional topics are covered, including: introduction to master
equations; non-equilibrium processes; fluctuation theorems; linear response theory; adiabatic trans-
port; the Kubo formalism; and the scattering approach to mesoscopics.
Thermal Equilibrium
1 The statistical picture of Mechanics 2
2 Spectral functions 9
3 The canonical formalism 15
4 Thermodynamics 26
5 Chemical equilibrium and the Grand Canonical state 33
6 Quantum ideal gases 40
Systems with interactions

7 Interactions and phase transitions 46
8 The Ising model 55
9 Phase transitions - heuristic approach 62
10 Phase transitions - field theory 71
Fluctuations and Response

11 Fluctuations 82
12 Linear response theory 88
13 The fluctuation dissipation relation 96
System interacting with a bath

14 The modeling Hamiltonian 103
15 Stochastic picture of the dynamics 108
16 Quantum master equations 114
Additional topics
17 The kinetic picture 121
18 Scattering approach to mesoscopic transport 128
19 The theory of electrical conductance 133
20 Irreversibility and Nonequilibrium processes 138
Detailed table of contents is available in the last two pages.

2
Thermal Equilibrium
[1] The statistical picture of Mechanics

Before we start discussing the canonical formalism of statistical mechanics, we would like to dedicate the first lecture
for some preliminaries regarding: Random variables and probability functions; The statistical picture of classical
dynamics in phase space; The notion of chaos; Stationary states in general; and the canonical state in particular.
This lecture is quite terse, and possibly will be expanded in the future.
====== [1.1] Random variables

Here is a list of topics that should be covered by a course in probability theory:
Random variable/observation x̂ (1.1)

Distribution function ρ(x) (1.2)
for discrete spectrum ρ(x) ≡ Prob (x̂ = x) (1.3)
for continuous spectrum ρ(x)dx ≡ Prob (x < x̂ < x + dx) (1.4)
Changing variables ŷ = f (x̂) , ρ̃ (y) dy = ρ(x)dx (1.5)
X
Expectation value of the random variable hx̂i ≡ ρ(x)x (1.6)
x
X
Expectation value of some other observable hÂi ≡ ρ(x)A(x) (1.7)
x
2
Variance Var(x̂) = h(x̂ − hx̂i) i = hx̂2 i − hx̂i2 (1.8)
Moment generating function Z(λ) = heλx̂ i (1.9)

Comulant generating function is defined through Z(λ) ≡ exp[g(λ)] (1.10)
" 2 #
1 x−µ
Gaussian distribution, definition ρ(x) ∝ exp − (1.11)
2 σ
1
Gaussian distribution, comulant g(λ) = µλ + σ 2 λ2 (1.12)
2
Legendre transform.– We can write the probability function as ρ(x) = exp(−F (x)), and redefine the comulant
generating function as G(λ) = −g(λ). We have by definition
Z ∞
−G(λ)
e = e−F (x) + λx
dx (1.13)
−∞
If we are allowed to use a saddle point approximation, it follows that G(λ) is related to F (x) by a Legendre transform:
n o
G(λ) ≈ min F (x) − λx = F (x̄) − λx̄ (1.14)
x
where the most probable value x̄ is determined by solving λ = F 0 (x). We shall see that this is formally the same
mathematics as going from the Helmholtz to Gibbs free energy. Below we explain that the inverse of this relation is
the large deviation theory.
3
====== [1.2] Several random variables

In classical probability theory we can define a joint distribution function for random variables, and then characterize
this distribution by correlation functions.
Joint distribution function of two variables ρ (x, y) (1.15)

Correlation between two variables Cxy = hx̂ŷi − hx̂ihŷi (1.16)
In the quantum framework, known as “measurement theory”, it is not possible in general to define joint distribution
function. Instead one defines a probability matrix. See the lecture regarding the first and the second quantum
postulates in quant-ph/0605180
If we have a sequence of random variable {x̂j } it is called a stochastic process, and the common notation for the
correlation function is Cij . For time-continuous process the notations is C(t0 , t00 ) where t0 and t00 are the two “sampling”
times of the “signal”.
Adding random variables.–

Adding two independent random Adding N independent and identi-
variables: cally distributed random variables:
Ŝ = x̂ + ŷ N
X
Ŝ = x̂j
j=1
hŜi = hx̂i + hŷi
hŜi = N µ
Var(Ŝ) = Var(x̂) + Var(ŷ)
Var(Ŝ) = N σ 2
gs (λ) = gx (λ) + gy (λ)
gs (λ) = N g(λ)
The are two useful results for large N . One is the central limit theorem and the other is the large deviation theory.
Central limit theorem.– We define the scaled variable

P
j x̂j − N µ
ŷ ≡ √ (1.17)
N σ
The statement is that in the large N limit it has a normal distribution with zero average and unit dispersion. This
follows by taking the limit of

λ λµ
gy (λ) = N g √ −√ (1.18)
Nσ Nσ
P
Large deviation theory.– Define the scaled variable x̂ = (1/N ) x̂j . Accordingly

λ
gx (λ) = N g (1.19)
N
The statement regarding its distribution is
Prob(x̂ > x) ≈ e−N f (x) , f (x) = max{λx − g(λ)} (1.20)

λ
In order to prove this result note that Θ(x) < eλx for any positive λ. Consequently
D hX iE D P E
Prob(x̂ > x) = Θ x̂j − N x < eλ[( x̂j )−N x] = eN (g(λ)−λx) (1.21)
And the approximation is obtained by optimizing the value of λ to get the lowest bound. Note that the optimization
parameter λ is formally like λ/N , where λ is conjugate to the random variable x̂.
4
====== [1.3] The statistical description of a classical particle

The statistical state of a classical particle with one degree of freedom is described by a probability function:
dxdp
ρ(x, p) ≡ PROB (x < x̂ < x + dx, p < p̂ < p + dp) (1.22)
2π~
where the normalization is

x dxdp
ρ (x, p) = 1 [in the next lectures ~ = 1] (1.23)
2π~
The generalization of this definition to the case of d freedoms is straightforward with Planck cell volume (2π~)d . The
expectation values of observables are defined in the usual way:
x dxdp
hAi = ρ (x, p) A(x, p) (1.24)
2π~
We note that in the quantum context one can define a quasi distribution that corresponds to ρ(x, p), known as
the Wigner function. Furthermore with any observable Â we can associate a phase apace function A(x, p) such
that the expectation value can be calculated using classical look-alike formulas. This is known as the Wigner-Weyl
formalism. This formalism can be regraded as generalization of WKB: Roughly speaking one may say that each
Planck cell in phase space can be regarded as representing a quantum state. The volume of Planck cell is (2π~)d
where d is the number of freedoms. Above we have assumed d = 1. Note that the normalization convention allows a
sloppy interpretation of ρ (x, p) as the probability to occupy a Planck cell in phase space. We also remark that the
quantum requirement trace(ρ2 ) ≤ 1 implies that a wavepacket in space space cannot occupy a volume that is less than
a Planck cell. The probability function of x is
Z
dp
ρ(x) = ρ (x, p) (1.25)
2π
The ”spreading” of a wavepacket is characterize by
2
σx2 ≡ Var(x̂) = h(x̂ − hx̂i) i = hx̂2 i − hx̂i2 (1.26)
2
σp2 ≡ Var(p̂) = h(p̂ − hp̂i) i = hp̂ i − hp̂i 2 2
(1.27)
In the quantum context σx σp > (~/2). The ”energy” of the system is defined as follows:
x dxdp
E = hH (x̂, p̂)i = ρ (x, p) H(x, p) (1.28)
2π~
Later we shall define some other ”spectral” functions that are related to H. Those can be written as an expectation
value of functions of H.
====== [1.4] Dynamics in phase space

The difference between “classical mechanics” and “classical statistical mechanics” parallels the distinction between
“Heisenberg picture” and “Schrodinger picture” in quantum mechanics. The former describes the evolution of the
system using a set of dynamical variables that obey some equations of motion, while the latter describe the evolution
of the associated probability function. In order to make the above distinction clear we consider the simplest example:
a free particle. The Hamiltonian is
p2
H = + V (x), for free particle V (x) = 0 (1.29)
2m
5
Say that at t = 0 the particle is at (x0 , p0 ). The equations of motion are
∂H p
ẋ = = (1.30)
∂p m
∂H
ṗ = − = 0 (1.31)
∂x
The solution is:
t
x(t) = x0 + p0 (1.32)
m
p(t) = p0 (1.33)
In the Heisenberg picture we regard x̂0 and p̂0 as random variables that have some probability function ρ (x, p). Then
we define new random variables
t
x̂t = x̂0 + pˆ0 (1.34)
m
p̂t = pˆ0 (1.35)
It follows from the composition law of random variables that there is spreading in space as a function of time:
s
σp (0) σp (0)
σx (t) = σx2 (0) + t2 ∼ t (1.36)
m m
It should be clear that “spreading” is a classical effect that originates if we assume that there is some dispersion
in the momentum. In quantum mechanics this effect is unavoidable because preparations with zero dispersion are
non-physical.
In the optional Schrodinger picture we define ρt (x, p) as the probability distribution of x̂t and p̂t . So instead of talking
about the time evolution of x̂ and p̂ we talk about the time evolution of ρ (x, p). In statistical mechanics we prefer
the latter point of view. Evolution takes place in phase space. Liouville theorem applies. Let us see how we use the
“Schrodinger picture” in the above example. Assume that the free particle has been prepared in a “classical pure
state” at the point (x0 , p0 ) in phase space. Accordingly
ρt=0 (x, p) = 2πδ (p − p0 ) δ (x − x0 ) (1.37)
After time t the state is

p0
ρt (x, p) = 2πδ (p − p0 ) δ x − x0 + t (1.38)
m
If the preparation is not a “classical pure state”, but say a Gaussian wave-packet that has some finite momentum
spread σp , then one observes spreading as explained previously. More generally we can discuss the spreading of
a wavepacket in the case of a non-linear oscillator. In such case V (x) has either sub-quadratic or super-quadratic
variation, and consequently the oscillation frequency ω(E) depends on the energy: decreases or increases with energy
respectively. If the initial distribution has some finite spread σE in energy, there will be angular spreading that
leads to a quasi-ergodic distribution within the energy shell. It is not really ergodic because if we started with a
mono-energetic distribution (σE = 0) it would not fill uniformly the energy surface: here the energy surface is merely
a one-dimensional “ellipse”. For graphical illustrations see figures in the next section.
6
====== [1.5] The route to ergodicity

Let us outline some major observations with regard to the dynamics of classical Hamiltonian systems.
Simple 1D system:– The student is expected to be familiar with the dynamics of harmonic oscillator; potential
well; pendulum. In the case of non-linear oscillations we have the spreading effect. In the case of a pendulum we
have a multi-component phase space with separatrix. The dynamics is not chaotic. One can define the oscillation
frequency ω(E) as a function of energy. In the quantum case ω(E) corresponds to the level spacing at the vicinity of
the energy E.
Chaotic system:– The student is expected to be familiar with the dynamics in simple billiards. The visualization
can be achieved using a Poincare section. In the case of a Sinai billiard (motivated by the discussion of Lorentz gas)
the dynamics is fully chaotic, leading to ergodization. More generally we might have mixed phase space that contains
”chaotic sea” as well as ”islands”.
Ergodization:– The evolution of a chaotic system leads to an ergodization on the energy shell. This can be
mathematically described using the Boltzamnn approach: course graining of phase space by dividing it into cells;
definition of Boltzamnn entropy. Eventually the system will become stationary-like, as if it were prepared in a state
that has maximum entropy.
Driven system:– There is a complicated route to chaos in the case of driven integrable (1D) systems. In contrast
to that in the case of driven globally chaotic systems the picture is qualitatively simple: if we prepare the system
initially within an energy shell, it will ”evolve” with this energy shell, along with diffusion transverse to the energy
shell. This diffusion leads in general to increase of the average energy (heating).
Convex Billiard Concave Billiard
Phase space illustration. – The dynamics of

a particle in a convex (Sinai) Billiard is completely
chaotic. In contrast to that, in the case of a concave
billiard, we have a mixed phase space that contains
both quasi-integrable regions and chaotic sea. The
Spreading illustration. – We consider the evolu- phase space is 3-dimensional (x, y, ϕ) where ϕ is the
tion of an initial Gaussian distribution (left panels) direction of the velocity. It is illustrated in the left
in the case of a non-linear oscillator. After a short lower panel. The dotted line indicates the normal di-
time (middle panels) the spreading is like that of a rection on the boundary. The reflections are specular
free particle. After a long time (right panels) one with regard to this direction. The right lower panel is
observes an ergodic-like distribution within the en- the two-dimensional (s, θ) Poincare section of phase
ergy shell. However, this is not really ergodic: if we space: each trajectory is represented by a sequence
started with a mono-energetic distribution, it would of points that indicate successive collisions with the
remain localized on the energy shell, as in the case of boundary, where s is the boundary coordinate, and θ
an harmonic oscillator. is the collision angle (relative to the normal).
7
====== [1.6] Stationary states

The evolution of a statistical state is determined by the Lioville equation of classical mechanics, which becomes the
von-Neumann Lioville equation in quantum mechanics.
∂ρ
= [H, ρ]PB (1.39)
∂t
We consider non-driven bounded systems, and focus our attention on stationary states that do not change in time.
This means ∂ρ/∂t = 0. In the classical language ρ can be regarded as a mixture of ”energy shells”, while in the
quantum language it means that ρ 7→ diag{pr } is a mixture of energy eigenstates labelled by r. In particular the
classical microcanonical state corresponds to an energy eigenstate, and is formally written as
1
ρ(x, p) = δ(H(x, p) − E) (1.40)
g(E)
The canonical state is
1
pr = e−βEr (1.41)
Z(β)
and in a classical context it is written as
1
ρ(x, p) = e−βH(x,p) (1.42)
Z(β)
The density of states and the partition function are defined as
X
g(E) = trace(δ(E − H)) = δ(E − Er ) (1.43)
r
X Z
Z(β) = trace(e−βH ) = e−βEr = g(E)dE e−βE (1.44)
r
We note that the probability distribution of the energy can be written as ρ(E) = g(E) f (E), where the occupation
probability function is f (E) ∝ δ(E − E) and f (E) ∝ e−βE in the microcanonical and canonical cases respectively. If
we have a many body system of non-interacting participles we can re-interpret f (E) as the occupation function, and
accordingly ρ(E) becomes the energy distribution of the particles (with normalization N ).
====== [1.7] The microcanonical and canonical states

Let us assume the following total Hamiltonian for a universe that consists of system and environment:
Htotal = H (Q) + Henv (Qα ) + Hint (Q, Qα ) (1.45)
For sake of presentation we do not write the conjugate momenta, so Q stands for (Q, P ) or it may represent spin
freedoms. If one neglect the interaction the eigenenergies are written as ErR = Er + ER , where r labels system states
and R labels environmental states.
It is argued that the weak interaction with the environment leads after relaxation to a canonical state which is
determined by the parameter
d
β = log(genv (E)) (1.46)
dE
8
where genv (E) is the density of states, which is assumed to be huge and rapidly growing with energy. The argument
is based on the assumption that the universe (system+environment) is (say) a closed system with some total energy
Etotal . After ergodization the system get into a stationary-like state that resembles a microcanonical states:

pr,R ∝ δ Etotal − (Er + ER ) (1.47)
with finite width (broadened) delta function. The probability pr to find the system in a state Er is proportional to
genv (Etotal −Er ) ≈ genv (Etotal )e−βEr . Accordingly
1 −βEr
pr = e (1.48)
Z
where the so-called partition function provides the normalization
X
Z(β) = e−βEr (1.49)
r
The partition function may depend on parameters that appear in the system Hamiltonian. Therefore we use in general
the notation Z(β, X).
====== [1.8] Mathematical digression

Sometimes is is more appropriate to expand the log of a function. Specifically this would be the case if the function
is definite positive and span many decades. Let us see what is the error which is involved in such an expansion:
f (x) = xN (1.50)
1
f (x + δx) = xN + N xN −1 δx + N (N − 1)xN −2 δx2 (1.51)
2
δx x/N (1.52)
Optionally we expand the log of the function:
S(x) ≡ ln f (x) = N ln(x) (1.53)

N 1N 2
S (x + δx) = N ln(x) + δx − δx (1.54)
x 2 x2
δx x (1.55)
Thus we have the recipe:
∂ ln f (x)
f (x + δx) ≈ f (x)eβδx where β ≡ (1.56)
∂x
In complete analogy we have:
g (E0 + ) ≈ g (E0 ) eβ (1.57)
where β is the log derivative of the density of states.

9
[2] Spectral functions

Various types of spectral functions are defined in mathematical physics. In the quantum context they characterize
the spectrum {En } of energies of as given Hamiltonian H. In the continuum or classical limit it is essential to define a
measure. Below we focus on the most popular spectral functions in statistical mechanics: the density of states g(E),
and the partition function Z(β). We shall see later that the state equations of a system in equilibrium can be derived
from, say, the partition function. Hence the spectral function serves as a generating function.
In the section below we define g(E) and Z(β), and show how they are calculated using standard examples: Two
level system; Harmonic oscillator; Particle in a box; Particle with general dispersion relation; The effect of A(x), V (x)
potential; Several particles; Identical classical particles, the Gibbs factor; Particles with interactions; In particular
two quantum particles; Molecules of type AA and AB (exercise).
====== [2.1] The definition of counting and partition functions

We consider a time independent bounded system which is described by a Hamiltonian H whose eigenvalues are Er .
We can characterize its energy spectrum by the functions
X X
N (E) ≡ Θ (E − Er ) = 1 (2.1)
r Er <E
X
Z(β) ≡ e−βEr (2.2)
r
If we have a large system we can smooth N (E), and then we can define the density of states as
dN (E) X
g(E) ≡ = δ (E − Er ) (2.3)
dE r
Note that
Z
Z(β) = g(E)dE e−βE (2.4)
For a classical particle in 1D we can use the above definitions with the prescription
X x dxdp
7−→ (2.5)
r
2π~
Each ”Planck cell” in phase space represents a state. Accordingly
x dxdp x dxdp
N (E) = Θ (E − H (x, p)) = (2.6)
2π 2π
H(x,p)<E
x dxdp Z
Z(β) = e−βH(x,p) = g(E)dE e−βE (2.7)
2π
In what follows the Gaussian integral is useful:
Z 12
− 12 ax2 2π
e dx = (2.8)
a
10
====== [2.2] Two level system or spin

The Hamiltonian of spin 1/2 in magnetic field is
1
H = hσz (2.9)
2
The eigenstates are |+i and |−i with eigenvalues E± = ±h/2. Accordingly

1
Z(β) = e−β (− 2 ) + e−β ( 2 ) = 2 cosh
h h
βh (2.10)
2
Optionally we can write the energies of any two level system as Er = n with n = 0, 1 then
Z(β) = 1 + e−β

(2.11)
====== [2.3] Two spins system in interaction

If we have N is interacting spins the sum over states can be factorized and we simply get
N
ZN (β) = Z1 (β) (2.12)
For two spins in the absence of magnetic field we get Z2 = 22 = 4. Let us see what happens if there is an interaction:
1 a 1 b
H = εσ a · σ b = 2S 2 − 3 ε,

S= σ + σ (2.13)
2 2
The energy levels are Esinglet = −3ε and Etriplet = ε. With and added magnetic field the partition function is
Z(β) = e3βε + eβh + e−βh + 1 e−βε

(2.14)
which factorized for ε = 0, but not in general.
====== [2.4] Harmonic oscillator

The Hamiltonian of Harmonic oscillator is
p2 1
H= + mω 2 x2 (2.15)
2m 2
1

The eigenstates are |ni with eigenvalues En = 2 + n ω. Accordingly
∞
1
e−β ( 2 +n)ω
X 1
Z(β) = = 1
(2.16)
n=0
2 sinh 2 ωβ
Note that if we write the energies as Er = n with n = 0, 1, 2, 3, 4, ... then
1
Z(β) = (2.17)
1 − e−β
11
Now let us see how the classical calculation is done.
1
1 1 2E 2 1 E
N (E) = ellipse area = π (2mE) 2 = (2.18)
2π 2π mω 2 ω
Z Z 1 1
−β 21 mx2 dp −β p2 2π 2
m 2 T
Z(β) = dx e e 2m = = (2.19)
2π βmω 2 2πβ ω
One can verify the validity of WKB quantization.
====== [2.5] Particle in a 1D box

The simplest is to assume periodic boundary conditions
p2
H = x ∈ [0, L] (ring) (2.20)
2m
The eigenstates are the momentum states |pi with
2π
p = n where n = 0, ±1, ±2... (2.21)
L
Hence the eigenvalues are
2
1 2π
En = n (2.22)
2m L
The number of states up to energy E is
L 1 1 L
N (E) = 2 (2mE) 2 ≡ kE L ≡ 2 (2.23)
2π π λE
The density of states is
L
g(E) = (2.24)
πvE
The 1D case here is pathological because in general the density of states grows rapidly with energy. Nevertheless in
the limit of ”infinite volume” we may treat the spectrum as a continuum:
∞ Z ∞ 12
2
n2 m L
dn e−β 2m ( L )
X 1 2π
Z(β) = e−βEn ≈ = L ≡ (2.25)
n=−∞ −∞ 2πβ λT
Let us see how the calculation is carried out classically. We can still consider a ring, or optionally we can write the
Hamiltonian with a box potential VL (x). Then we get
1 1 L
N (E) = rectangle area = × L × 2 (2mE) 2 = 2 (2.26)
2π λE
Z Z 1
dp −β p2 m 2 L
Z(β) = dx e 2m = L = (2.27)
2π 2πβ λT
12
One can verify the validity of WKB quantization (but without the 1/2 shift).
====== [2.6] A particle in 3D box, or higher dimensions

Consider a particle in a d=3 box of volume V = Ld .
3
X p2i
H = + [implicit boundary conditions with volume Ld ] (2.28)
i=1
2m
The eigenstates are
2π
p~ = (n1 , n2 , n3 ) (2.29)
L
2
1 2π
n21 + n22 + n23

En1 n2 n3 = (2.30)
2m L
The summation over the states factorizes:

!3
X
−βEn1 n2 n3
X
−βEn V
Z(β) = e = e = (2.31)
n1 n2 n3 n
λ3T
The above calculation gives Z = (L/λT )d in d dimensions. For the counting function we get:
 
1 Ωd  2  L d
d
N (E) = (kE L) = π (2.32)
(2π)d d 4π/3 λE
and accordingly
Ωd d−1 L
g(E) = (kE L) ∝ E (d/2)−1 (2.33)
(2π)d vE
d−1
The factor (kE L) can be interpreted as the number of open modes. For d = 2 the DOS is independent of energy
and reflects the mass of the particle.
As far as the classical calculation is concerned, N particle systems is formally like one particle system with d 7→ N d.
In the quantum treatment the Fermonic or Bosonic nature of identical particles should be taken into account: see
later how the calculation is done e.g. for two particles).
====== [2.7] Classical particle in magnetic field

For a particle in an arbitrary scalar potential V (r) in 3D we get
Z 3 Z
dr dp −βH 1
Z(β) = e = dr e−βV (r) (2.34)
(2π)3 λT
Let us include also a vector potential:
1 2
H= (p − A(r)) + V (r) (2.35)
2m
13
dr dp0
Z Z
dr dp
h 2
i
Z = e−β [ 2m (p−A(r))
1 2
+V (r)]
= e
−β 1
2m (p0 ) +V (r)
(2.36)
3 2
(2π) (2π)
The result does not depend on A(r). The energy spectrum is not affected from the existence of A(r). The energies are
E = (1/2)mv 2 + V (r) irrespective of A(r). This is no longer the case upon quantization. Note the implicit assumption
of having background relaxation processes that make the dynamics irrelevant.
====== [2.8] Gas of classical particles in a box

Let us consider N particles:
N 2
X p~ α
H = + V (rα ) + U (r1 , ..., rN ) (2.37)
α=1
2m
In the absence of interaction the partition function is
N Z N
1
ZN (β) = Z1 (β) = d3 r e−βV (r) (2.38)
λ3T
From now on we assume gas of identical particles and therefore include the Gibbs factor:
1
ZN (β) 7→ ZN (β) (2.39)
N!
For N interacting particles we get

N Z
1 1
ZN (β) = dr1 ...drN e−βU (r1 ,...,rN ) (2.40)
N! λ3T
====== [2.9] Two quantum identical particles

Let us see what is the partition function for a system that consists of two identical particles, say in a box. The total
energy is written as Eab = Ea + Eb . The partition function is
X
1 X −β(Ea +Eb ) 1
Z(β) = e + e−β(2Ea ) (2.41)
2 0
a6=b a
 
1 X −β(Ea +Eb ) X −2βEa  1h i
= e ± e = Z1 (β)2 ± Z1 (2β) (2.42)
2 a
2
a,b
For a particle in a d dimensional box

d
L
Z1 = (2.43)
λT
1 2
Z2 = Z1 ± 2−d/2 Z1 (2.44)
2
Note that for d = 3 we get
λ3T

1 2 1
Z2 (β) = Z × 1 ± 3/2 (2.45)
2 1 2 volume
14
The Fermi case is similar to hard sphere:

1 2 sphere volume
Z2 (β) = Z1 × 1 − (2.46)
2 box volume
====== [2.10] Two quantum particles in a box with interaction

The calculation of the partition function Z2 for two identical quantum particle in a box, is both interesting and later
on useful for the purpose of calculating the second virial coefficient of an N particle gas. The Hamiltonian is:
P2 p2
H= + + V (r) (2.47)
4m m
In order to be able to do the calculation using separation of variables we cheat with the boundary conditions as
follows: The center of mass motion is confined to a box of volume V = (4π/3)R3 , and the relative motion is confined
by |r| < R independently. Accordingly the partition function is factorizes as follows:
 0 
X "X Z ∞ #
V V 2
Z2 = 23/2 3  e−βEn`m  = 23/2 3 e−βEb + g(k)dk e−(β/m)k (2.48)
λT λT 0
n`m b
where (n, `, m) are the good quantum numbers for the relative motion. Ignoring the possibility of spin, the sum is
over even or odd values of `, for Bosons or Fermions respectively. In the second equality we separate the bond states
from the scattering (continuum) states. In order to determine the DOS of the latter we recall that the radial wave
functions are phase shifted spherical Bessel functions. Accordingly the box quantization condition for the allowed kn
values is
π
kR − ` + δ` = nπ (2.49)
2
From here one deduce a formula for the effect of the phase shifts on the DOS:
1
1X ∂δ`
g(k) − g(0) (k) = (2` + 1) (2.50)
π % ∂k
Using this result we get after integration by parts the following expression for the interaction effect on the partition
function:
 0

2 XZ ∞
X
(0) V λ 2
Z2 − Z2 = 23/2 3  e−βEb + T2 kdk δ` (k) e−(β/m)k  (2.51)
λT π 0
b `
V(x)
111111111
000000000
000000000
111111111
111111111
000000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
1111111
0000000
1111111
0000000
111111111
000000000
111111
000000
1111
0000
111
000
R
111
000
11
00
15
[3] The canonical formalism
====== [3.1] The energy equation of state

Consider some system, for example particles that are confined in a box. The Hamiltonian is
H = H(r, p; X) (3.1)
where X is some control parameter, for example the length of the box. Assuming that we are dealing with a stationary
state, the energy of the system is
X
E ≡ hHi = trace(Hρ) = pr Er (3.2)
r
If the system is prepared in a canonical states, then it is a mixture of energy eigenstates with probabilities
1 −βEr
pr = e (3.3)
Z
where the partition function is
X
Z(β, X) = e−βEr (3.4)
r
One observes that the energy of a system that is prepared in a canonical state can be derived from the partition
function as follows:
1 ∂ ∂
E = hHi = − Z = − ln Z (3.5)
Z ∂β ∂β
Also one can find expressions for the higher moments, for example
1 ∂ ∂
hH2 i = Z (3.6)
Z ∂β ∂β
In particular one deduces the relation
∂2 ∂E
Var(E) = hH2 i − hHi2 = ln Z = T2 = T 2C (3.7)
∂β 2 ∂T
where in the latter equality we have defined the temperature as T = 1/β and the heat capacity as C = dE/dT . The
notion of temperature will be discussed further below.
====== [3.2] The Equipartition theorem

In the classical context the Hamiltonian might be a sum of quadratic terms
X
H = cj qj2 (3.8)
j
16
where qj are either coordinates of conjugate momenta. The partition function factorizes, where each quadratic term
contributes a ∝ T 1/2 term. It follows that each quadratic term contributes T /2 to the energy, and hence 1/2 to the
heat capacity.
This observation can be applied to the analysis of ”balls connected by springs”. We can always go to normal
coordinates. The center of mass degree of freedom contributes T /2 to the energy, while each vibrational mode
contributes T .
A formal extension of this so-called ”Equipartition Theorem” is as follows:

∂H
qi = T δij (3.9)
∂qj
The proof is as follows: The measure of integration over phase space can be written as dqi dqj dq 0 , where q 0 represents
all the other coordinates. Applying integration by parts we have
Z Z Z
∂H(q) −βH(q) 1 ∂ h −βH(q) i 1
dqi dqj dq 0 qi e =− dqi dqj dq 0 qi e = δij dqi dqj dq 0 e−βH(q) (3.10)
∂qj β ∂qj β
and form here follows the Equipartition Theorem. This generalized version is useful in discussing particles that have
interaction u(xi − xj ) ∝ |xi − xj |α , which constitutes a generalization of the harmonic (α = 2) case.
====== [3.3] Heat capacity

From the Equipartition Theorem one deduce that the heat capacity of an ”ideal” system equals to the effective number
of freedoms: Each independent quadratic term in the Hamiltonian contributes 1/2 to the heat capacity. This simple
prescription should be refined for two reasons: (i) Degrees of freedom can ”freeze” in the quantum treatment; (ii) In
general a many body system is not ideal due to interactions. We first discuss the quantum issue referring to spins
and oscillators.
Spin and oscillator.– For spin (+) or oscillator (-) with level spacing ω we have
ln(Z(β)) = ± ln(1 ± e−βω ) (3.11)

∂ ln Z ω
E = − = (3.12)
∂β eβω ± 1
dE 1 ω 2
C(T ) = = 2 , ”csnh” is cosh or sinh (3.13)
dT 2csnh ω T

2T
In both case the low temperature behavior of C is identical, namely, for T ω it is dominated by the Boltzmann
factor e−βω . At high temperature C of the spin drop down because energy reaches saturation, while C of the oscillator
approaches unity reflecting the classical prediction E ≈ T . Since E = ωn it is more illuminating to re-write the above
results as follows:
1
hni = ≡ f (ω) (3.14)
eβω ± 1
Var(n) = [1 ∓ f (ω)]f (ω) (3.15)
where f (ω) is known as the occupation function. In the case of an oscillator the result for the number variance can be
regarded as a sum of a shot-noise particle-like term Var(n) = hni, and a classical term Var(n) = hni2 . In the case of a
spin the fluctuations go to zero in both the ”empty” and ”full” occupation limits. It is customary in quantum-optics
to characterize the fluctuations by g (2) = (hn2 i − hni)/hni2 and to say that the bosonic (oscillator) result g (2) = 2
corresponds to bunching, while the fermionic (spin) result g (2) = 0 corresponds to anti-bunching. The value g (2) = 1
reflects Poisson statistics and would apply in the case of coherent state preparation.
17
Debye model.– Let us refer further to a system that can be described as consisting of many harmonic freedoms, e.g.
modes of vibrations. The spectral density of the modes might be ∝ ω α−1 . For example in Debay model α = d = 3,
with some cutoff frequency ωc . Then we get for the heat capacity
Z ωc ω 2
1 ω
c
C(T ) = const 2 ω α−1 dω = const T α F (3.16)
ω T T

0 2 sinh 2T
where
ν
ex
Z
F (ν) ≡ xα+1 dx (3.17)
0 (ex − 1)2
The quantum result can be described as emerging from ”freezing” of freedoms due to the quantization of energy. This
phenomena has lead to the birth of quantum mechanics in the context of blackbody radiation (Planck law).
Glasses.– The standard model for glasses regard them as a large collection of ”two level” entities with splitting ω
that has roughly uniform distribution. Hence the calculation of the heat capacity is formally as in the Debye model
model with sinh replaces by cosh, and α = 1, leading to a linear dependence C(T ) ∝ T .
Quantum gases.– In the classical treatment, disregarding prefactors of order unity, a gas of N particles have total
energy E ∼ N T , hence the heat capacity is C ∼ N . If we have a gas of Fermions in low temperatures, then the number
of excited particles is Neff ∝ T , hence the energy is E ∝ T 2 , and the heat capacity is C(T ) ∝ T . In contrast to that
Bosons in 3D condense into the ground states. Hence the occupation of an excited state of energy r is formally the
same as the occupation of an oscillator with the same frequency. Consequently one observes C(T ) ∝ T α as in Debye
model.
Phase transitions.– We shall discuss phase transitions in later lectures. As the temperature is lowered towards
a critical temperature Tc the system becomes ”correlated”, which means that the effective number of freedoms is
reduced. We assume T > Tc and note that similar picture applies if one approaches Tc from below. We can easily
explain why the heat capacity diverges as Tc is approached. For an ideal gas, or better to think about a collection
of non-interacting oscillators, the partition function is Z = g N , where N is the number of freedoms, and g ∝ T is
the number of accessible states for a single freedom at temperature T . For a correlated system Z = g Neff , where
Neff = N/ξ d is the effective number of independent regions, and ξ is called the correlation length. The prototype Ising
model consist of spins (g = 2) rather than oscillators and ξ ∝ |T − Tc |−ν where ν ≈ 1/2. Either way we can write the
expression for the heat capacity as follows:
d2 ln Z
C(T ) = β 2 ≡ Cg (T ) + Cξ (T ) (3.18)
dβ 2
where the non-singular Cg (T ) originates from the temperature dependence of g, and equals N for non-interacting
oscillators, reflecting the effective number of freedoms. The singular term Cξ (T ) originates from the temperature
dependence of ξ. For an Ising system its divergence near the critical temperature is described by |T − Tc |νd−2 . Note
the significance of the space dimension d.
1 N Debye
(α=d=3)
exp OSC Tα
exp spin
w wc
ε
N Fermi
Bose
N
Tα
Phase transition
fF Tcondensation Tcritical
18
====== [3.4] Generalized forces

Assume that X is a parameter that appears in the Hamiltonian. We define the generalized force F which is associated
with the parameter X as
∂H
F = − (3.19)
∂X
This definition will be motivated later on when we discuss the notion of work. We shall explain that for an isolated
system that undergoes a quasi-static adiabatic process the change in energy is dE = − hFi dX, meaning that the work
that has been dome by the system is dW = hFi dX. Here are some examples for generalized forces:
parameter generalized force

piston displacement in cylinder with gas Newtonian force
volume of a box with gas Newtonian pressure
length of a polymer Newtonian force (tension)
homogeneous electric field total polarization
homogeneous magnetic field total magnetization
magnetic flux through a ring electric current
Flux and Current.– We would like to better clarify why magnetic flux and electrical current are conjugate variables.
Note that for an homogeneous magnetic field the flux through a ring is Φ = AB, and the magnetization is ad-hock
defined as M̃ = AI, where A is the area of the ring. Using the notation X = Φ, the direct identification of the
conjugate operator F as the current, is rationalized in a simple-minded manner as follows: If we make a change dX
of the flux during a time dt, then the electro-motive force (EMF) is −dX/dt, leading to a current I in the ring. The
energy increase of the ring is the EMF times the charge, namely dE = (−dX/dt) × (Idt) = −IdX.
Magnetic field.– Usually we shall denote the applied magnetic field by the letter h, possibly absorbing into it
definition the coupling constant. For example we write the interaction of a spin with a vertical magnetic field as
−hσz . But there are circumstance in which the sample affect the the magnetic field in a way that cannot be ignored.
For example: if we place a typeI superconductor inside a solenoid, it expels sideways the magnetic field, such that
the total magnetic field is B = 0 inside the sample. We therefore have to be careful in how we write the Hamiltonian.
Schematically we write
X 1
Htotal = (pj − ej A)2 + U (r1 , r2 , ...) (3.20)
j∈system
2m j
Z
h i 1
+ similar expression for the solenoid + B(x)2 d3 x (3.21)
8π
Our focus is on the system, so we keep only the interaction of the system with the solenoid:
Z Z
1
Htotal = Hsystem (rj , pj ; A) + B(x)2 d3 x − A · Jsolenoid d3 x (3.22)
8π
The current density of the solenoid defines the applied magnetic field through the relation
∇ × h = 4π Jsolenoid (3.23)
Substitution of this definition into the last term, and doing integration by parts, the Hamiltonian that described the
19
interaction of the system with the applied magnetic field takes the following form:
Z Z
1 2 3 1
Htotal = Hsystem (rj , pj ; A) + B(x) d x − h(x)B(x)d3 x (3.24)
8π 4π
Schematically the interaction is described by a term that looks like −hB, where h is a control parameter, and B is
the conjugate dynamical variable. This sounds less strange if we think of h as an external current. Note that the
role of the ”current” and the ”flux” have been switched. Here the current is the source, and the magnetic ”flux” is a
dynamics variable. The expression for the work for a bulk sample will take the form
V V
dW = B(h) dh ≡ h dh + M̃ (h) dh (3.25)
4π 4π
If the state equation M̃ (h) is known, we can deduce from the above relation how the free energy of the sample
changes as the magnetic field is turned on. In the next lecture regarding thermodynamics Htotal as defined above will
be identified as a ”grand Hamiltonian” with which a ”Gibbs” free energy can be associated.
====== [3.5] Susceptibility and fluctuations

Given X and assuming that the system is prepared in a canonical state characterized by some β, we can derive the
average value y of the generalized force F from the partition function as follows:

X dEr 1 ∂ ln Z
y(X) ≡ hF iX = pr − = (3.26)
r
dX β ∂X
The so-called thermodynamics “constants” describe the dependence of y(X) on the the parameter X,
∂y
χ(X) = generalized susceptibility ≡ (3.27)
∂X
Specifically, let us consider the dependence of of the length of a polymer, or the volume of a gas, on the applied
tension or pressure. In these examples the total Hamiltonian can be written as
H(λ) = H − λV (3.28)
where the parameter λ is the applied field, and V is the conjugate dynamical variable (length or volume in the above
mentioned examples). Consequently we get in the presence of the applied field
trace [V exp (−βH(λ))]

= hV i + βλ hV 2 i − hV i2 + higher orders

hV iλ = (3.29)
trace [exp (−βH(λ))]
where both numerator and denominators have been expanded, without much caring about commutation relations.
From the above we deduce the following classical relation between the compressiblility and the fluctuations:

∂hV iλ 1
κ ≡ = Var(V ) (3.30)
∂λ λ=0 T
In a later lecture we shall introduce generalizations of this relation that are known as the ”Onsager regression theorem”
and as the ”Fluctuation dissipation relation”.
The relation κ = (1/T )Var(V ) parallels the relation C = (1/T 2 )Var(E) between the heat capacity and the fluctuations
in energy. It automatically implies that these constants have to be positive. Another way of looking on it is to say
that C > 0 and κ > 0 are stability conditions. Negative value means that that the system will undergo a ”phase
separation” process. See discussion of the ”Maxwell construction” is the ”Interactions and phase transitions” lecture.
20
====== [3.6] Empirical temperature

In practice we would like to be able to probe the β of the environment. For this purpose we use a thermometer. The
simplest thermometer would be an ideal gas in a box, for which the partition function is
3N
m 2
Z (β, V) = VN (3.31)
2πβ
1 ∂ ln Z N
P = = β −1 (3.32)
β ∂V V
The empirical temperature is defined as follows:
PV 1
θ = = (3.33)
N β
We can of course define different thermometers. The idea is simply to identify a measurable quantity that reflects the
parameter β.
====== [3.7] The Virial theorem

Somewhat related to the equipartition theorem, is the Virial theorem. It is used to relate the expectation value of
the “kinetic” and “potential” terms in Hamiltonian of the type H = K(p) + U (r).
Consider any observable G. It is clear that if the system is prepared in a stationary (not necessarily canonical) state,
then the expectation value hGi is constant in time. By the rate of change equation of motion it follows that
D E
[H, G] = 0 (3.34)
In particular let us consider the generator of dilations
1X
G = (rj · pj + pj · rj ) [the symetrization is required in the quantum case] (3.35)
2 j
For the Hamiltonian H = K(p) + U (r) we get

∂K ∂U
p· − r· = 0 (3.36)
∂p ∂r
with implicit summation over j. If the classical equipartition theorem applies, each term equals T multiplied by the
number of freedoms. For quadratic K(p) and U (r) the first term equals 2hKi, and the second term equals −2hU i.
More generally, for two-body interaction of the type
X X
U (r) = u(ri − rj ) = |ri − rj |α (3.37)
hiji hiji
the second term in the Virial theorem equals −αhU i. This is a meaningful statement for α > 0, otherwise there should
be a “box” that confines the particles. Writing the full Hamiltonian as H = K(p) + U (r) + VL (r) we deduce that

∂K ∂U ∂VL
p· − r· − r· = 0 (3.38)
∂p ∂r ∂r
21
In the next section we shall see how this relation helps us to derive an expression for the “pressure” on the walls of
the box.
====== [3.8] Pressure on walls

Possibly the simplest point of view about pressure is to regard it as arising from collisions of particles with the walls.
This is the so called the kinetic picture point of view. However, within the framework of the canonical formalism
the pressure is defined as the generalized force that is associated with the volume, such that dW ¯ = P dV. It is quite
puzzling that in the formal classical calculation the kinetic part factors out and the mass of the particles does not
appear in the result:
3N
ln(Z(β, V)) = − ln β + N ln V + const (3.39)
2
∂ ln Z 3
E = − = NT (3.40)
∂β 2
1 ∂ ln Z NT
P = = (3.41)
β ∂V V
With interactions we have to calculated a complicated configuration (dr1 dr2 ...drN ) integral. This calculation will be
discussed in later sections. In the absence of interactions we see that the pressure is related to the kinetic energy,
namely P = (2/3)E/V. Below we generalize this relation using the Virial theorem: we shall see that quite generally,
both classically and quantum mechanically, the pressure is related to the kinetic and potential energy of the gas.
The volume deformation of a box is represented by a deformation field D(r). To be specific let us write the Hamiltonian
of N gas particles in a box as follows:
H = K(p) + U (r) + VL (r − λD(r)) (3.42)
Here K(p) is the kinetic term, and U (r) are the interactions, and VL (r) is box defining potential, and λ is the
deformation parameter. We want λ to equal the extra volume due to the deformation, such that V = V0 + λ. We
therefore normalize the displacement field such that
{ 1
D · ds = 1, standard choice: D(r) = r (3.43)
3V0
Accordingly the definition and the expression for the pressure are

∂H ∂H 1 ∂VL 1 ∂K ∂U
P = − = − = r· = p· − r· (3.44)
∂V ∂λ λ=0 3V ∂r 3V ∂p ∂r
where in the last equality we have used the Virial theorem. Note that this extends the discussion of the Virial theorem
in previous section. The case of inter-atomic interactions with α > 0 (bounded system with no walls) can be regarded
formally as a special case of the above relation with P = 0. If α < 0 there is non-zero pressure. We can use the
equipartition theorem to obtain in the classical case

1 1 ∂U
P = NT − r· (3.45)
V 3 ∂r
where the first term is the same as in the law of ideal gases, while the second is due to the interactions, and can be
expressed using moments of the inter-particle separation.
22
====== [3.9] Tension of a polymer

The calculation of a tension of a polymer is very similar to the calculation of pressure. The parameter in the
Hamiltonian is the length X of the polymer, which is analogous to the length or the volume of the box that contains
the gas particles. In both cases the formal result depends only on the configuration integral, while the kinetic term
factors out. Thus in both cases the result does not depend on the mass of the gas particles or on the mass of the
monomers from which the polymer is composed. The partition function in the case of a polymer is
X
Z(β, X) = [kinetic term] × δ(X − (r1 + r2 + ... + rN )) e−βU (configuration) (3.46)
conf.
Typically, in analogy with the case of hard spheres, the potential energy merely restricts the space of allowed con-
figurations. Without the extra X restriction the summation would give a value Z(β). One observes that the ratio
Z(β, X)/Z(β) would be the probability of observing length X if the polymer were unconstrained at its endpoints.
According to the central limit theorem, for a long polymer
" 2 #
Z(β, X) 1 X
P(X) = ∝ exp − (3.47)
Z(β) 2 L0
Above we assume that the polymer can stretch either sides, hence its average ”algebraic” length is hXi = 0, while the
RMS average is denoted L0 . The force that is exerted on the endpoint obeys Hooke’s law:
1 ∂ ln Z
hFiX = = −(T /L20 ) X (3.48)
β ∂X
Optionally, if we insist on calculating directly the partition sum, we can write

Z Z Z
dk dk i k X
Z(β, X) = dr1 dr2 ...drN ei(X−(r1 +r2 +...+rN ))k e−βU (r1 ,r2 ,...,rN ) ≡ e Z̃(β, k) (3.49)
2π 2π
Above the we have omitted the irrelevant kinetic term. The integral over all possible configurations factorizes if the
potential energy of the polymer can be written as a sum u(r1 ) + u(r2 ) + ... + u(rN ) of independent terms. Notably
this is the case for a polymer whose all possible configurations have the same energy.
Assume that a tension f is applied on the polymer: this can be regarded as an ”electric” field that is applied on
the endpoint of the polymer, such that the Hamiltonian is changed by a constant H̃(r̂; X) = H + f X. To keep sign
consistency the field is applied in the negative direction, namely, the equilibrium point Xeq is determined by the
condition hFi˜
X = 0, leading to hFiX = f .
P
Next we consider the possibility that X̂ = xn becomes an unconstrained dynamical variable. The new system
is described by a grand-Hamiltonian HG (r̂; f ) = H + f X̂ that has one extra degree of freedom. If fluctuations are
neglected we expect hXif = Xeq to be consistent with hFiX = f . If we blur the distinction between the tension hFi
in the sense of expectation value, and the tension f in the sense of an external parameter (applied force), then, under
the same assumption, the relation dEG = Xdf is consistent with dE = −f dX. In other words: if the conjugate of X
is f , then in the grand-Hamiltonian framework the conjugate of f is −X. The grand partition function is
Z Z
ZG (β, f ) = dr1 dr2 ...drN e−β[U (r1 ,r2 ,...,rN )+(r1 +r2 +...+rN )f ] = Z(β, X) e−βf X dX (3.50)
The factorization of this partition function implies that the total length hXi of the polymer, for a given applied field f ,
is the sum of lengths of the monomers for the same field (the field determines the tension of the polymer). We realize
that Z(β, X) and ZG (β, f ) are related by a Laplace transform. From strict mathematical point the former is like the
probability function, and the latter is like the associated moment generating function. What we were doing is in fact
a generalization of the ”convolution theorem”, as used in the derivation of the central limit theorem.
23
Finally, in the large N limit the relation between ZG (β, f ) and Z(β, X) can be formulated as a Legendre transfor-
mation. We shall encounter the Legendre transformation in the next section, in a formally identical context, as the
relation between the Gibbs free energy G(T, P ) and the Helmholtz free energy F (T, V). Later we use the same ”trick”
in the analysis of quantum gases, when we go from the canonical to the so called “grand-canonical” framework.
====== [3.10] Polarization

The polarization is the generalized force that is associated with electric field. Let us assume that we have a bounded
system of particles with an added uniform electric field:
X p2 X
α
H = + interactions + potential − qα Exα (3.51)
α
2mα α
∂H X
P̂ = − = qα x̂α (3.52)
∂E α
The polarization P̃ is the expectation value of P̂ . One simple example is the calculation of the polarization of an
”atom”, where we have (say) a negative particle that is bounded by a ”spring” to a positive charge. Another simple
example concerns a diatomic molecule that has a permanent dipole moment µ. Here the Hamiltonian is
p2θ p2φ
H(θ, φ, pθ , pφ ) = + − µE cos(θ) (3.53)
2I 2I sin2 (θ)
where I is the moment of inertia. For the polarization we get

" −1 #
1 ∂ ln Z µE µE
P̃ = = µ coth − (3.54)
β ∂E T T
Note that expansion for weak field implies the electric susceptibility χ = (1/3)µ2 /T .
====== [3.11] Magnetization

The magnetization is the generalized force that is associated with magnetic field. It is either due to having spin degree
of freedom (Pauli) or due to the orbital motion. Here we clarify the definition using the three simplest examples.
Pauli magnetism.– Consider a collection of N spins. We denote the magnetic filed by h. The Hamiltonian is
N
X
H = − gα hSzα (3.55)
α=1
N
∂H X
M̂ = − = gα Szα (3.56)
∂h α=1
The magnetization M̃ is the expectation value of M̂ . For a single spin 1/2 entity we get the following result:

1 ∂ ln Z g gh
M̃ = = tanh (3.57)
β ∂h 2 2T
Note that expansion for weak field implies the magnetic susceptibility χ = (1/4)g 2 /T . Note also that in the classical
limit (”large spin”) the problem becomes formally identical to that of calculating polarization of electric dipoles.
24
Orbital magnetism (classical).– In the following we shall identify what is the magnetization M̂ for charged spinless
particles, using the formal definition −∂H/∂h. In the 1D case (ring) it is identified as arising from a circulating current.
In the 2D case it is more convenient to bypass the question what is M̂ and to go directly to the M̃ calculation via the
partition function. In the classical case one obtains M ˜= 0. But in the quantum calculation one obtains finite result.
The classical result is puzzling because we would like to interpret M̂ as arising from circulating currents as in the
1D case. Indeed such interoperation is possible. The point to realize that within the bulk we indeed have circulating
electrons that give rise to a diamagnetic response. But this is compensated by ”Hall currents” that flow along the
boundary. The exact cancellation of these two contributions is spoiled upon quantization, instead we get the de Haas
van Alphen (dHvA) oscillations. Details below.
Orbital magnetism (1D).– Consider a spinless particle in a ring of length L, and area A. The magnetic flux is
Φ = hA. The Hamiltonian, the velocity-operator, the current-operator and the magnetization-operator are
2
1 Φ
H = p−e + V (x) (3.58)
2m L

1 Φ
v̂ = i[H, x] = p−e (3.59)
m L
∂H e
Iˆ = − = v̂ (3.60)
∂Φ L
∂H
M̂ = − = AIˆ (3.61)
∂h
The magnetization M̃ is the expectation value of M̂ , or optionally we can refer to the circulating current I, which is
ˆ
the expectation value of I.
Orbital magnetism (2D).– The more interesting case is the magnetization of electrons in a 2D box (3rd dimension
does not play a role) due to the formation of Landau levels. We recall again that classically the energy spectrum of
the system is not affected by magnetic field. But quantum mechanically Landau levels are formed (see ”Lecture notes
in Quantum mechanics”). Let us consider a box of area A that contains N spinless electrons. In the bulk, the energy
of a Landau state that belongs to the ν level is εν = (ν + (1/2))ωB where ωB = eB/m is the cyclotron frequency. The
degeneracy of each Landau level is gB = eBA/2π. The calculation of the single particle partition function is the same
as that of harmonic oscillator (multiplied by the degeneracy). Assuming N electrons that can be treated as an ideal
Boltzmann gas we get
N e 2 B
M̃ = − + O(B 3 ) (3.62)
12 m T
This result does not hold for a low temperature electron gas, because the Fermi statistics of the occupation becomes
important. Assuming zero temperature we define Bn with n = 1, 2, 3, ... as the threshold value for which n Landau
levels are fully filled. This values are determined by the equation ngB = N . Considering first strong field B > B1 , the
(N )
energy of the system is E0 = N ωB /2 and hence
(N )
∂E0 e
M̃ = − = −N , for B > B1 (3.63)
∂B 2m
This result has a simple interpretation using ”Bohr picture” of an orbiting electron: each electron performs a minimum
energy cyclotron motion with unit angular momentum L, and associated magnetic moment −(e/2m)L. If the magnetic
field is Bn+1 < B < Bn , one has to sum the energy of the electrons in n filled Landau levels, where the upper one is
only partially filled. One obtain a quadratic expression from which it follows that the magnetization grows linearly
from −N (e/2m) to +N (e/2m). Hence there is saw-tooth dependence of M̃ on the field, which is known as the de
Haas van Alphen (dHvA) oscillations.
Semiclassical interpretation.– There is a simple way to understand the dHvA result. For this purpose assume
that A looks like a circle. Each ”Landau state” occupies a thin strip that has a finite width. Within each strip there
is a diamagnetic cyclotron motion whose net effect is like having an inner anticlockwise current (I > 0), and an
outer clockwise current (I < 0). In the bulk the net current of a strip is zero, but nevertheless it has a diamagnetic
25
contributions to the magnetization, because I encloses a larger area compared with I . As we come close to the
boundary, near the potential wall, the net current of the strip becomes positive, and its value is determined by the
potential gradient. This is known as Hall effect. In the case of hard wall there is a nice semi-classical illustration of the
trajectories that bounce along the boundary. Upon quantization the ”strips” support so-called ”edge states”. When
B crosses Bn we get a jump in the magnetization that corresponds to the occupation of an additional edge states: The
total Hall conductance of n Landau levels is GH = (e/2π)n, residing in a region that experiences a potential difference
ωB . Hence the drop in the magnetization is (GH ωB ) × A = N (e/m). It is now easy to understand why in the classical
limit we do not have magnetization: the Hall current along the edges compensates the diamagnetic currents of the
bulk. It is only upon quantization that the balance is violates, and instead we have the dHvA oscillations as a function
of B.
26
[4] Thermodynamics
====== [4.1] Absolute temperature and entropy

Let us formally vary the parameters X and β. The implied change in the energy is
" ! ! # " ! #
X X X dpr X dpr X dEr
dE = dpr Er + pr dEr = Er dβ + Er dX + pr dX (4.1)
r r r
dβ r
dX r
dX
The second term in the formal dE expression is identified as the work dW

¯ that would be done on the system during
a reversible quasi-static process:
!
X X dEr
pr dEr = pr dX = −y(X) dX (4.2)
r r
dX
In the next section we shall identify the first term in the formal dE expression as the heat dQ
¯ that would be absorbed
during a reversible quasi-static process. This expression is not an “exact differential”, but it has an “integration
factor” that we call “absolute temperature”. Namely,
1
T = integration factor = (4.3)
β
such that
! !
X X dpr X dpr
dpr Er = Er dβ + Er dX = T dS (4.4)
r r
dβ r
dX
The implied definition of the thermodynamic entropy is

X
S = − pr ln pr (4.5)
Note that the thermodynamic entropy is an extensive quantity in the thermodynamic limit. It should not be confused
with other types of ”entropy”. In particular we shall discuss the ”Boltzmann entropy” in a later section with regard
to the 2nd law of thermodynamics.
We see that the formal expression for dE can be written as follows:
dE = T dS − ydX (4.6)
It is important to emphasize that the above formal expression is a valid mathematical identity that holds irrespective
of whether it reflects an actual physical process. However, it is only for a reversible quasi-static process that ydX is
identified as the work, and T dS as the heat. For a non-reversible process these identifications are false.
====== [4.2] The Thermodynamic potentials

From the basic relation dE = T dS − ydX one concludes that if E is formally expressed as a function of S and X, then
we can derive from it the state equations T (S, X) and y(S, X). Accordingly ee say that E(S, X) is a thermodynamic
potential. At this stage it is convenient to define also the Helmholtz thermodynamic potential:
1
F (T, X) ≡ − ln Z(β; X) (4.7)
β
27
Within the framework of the canonical formalism the energy is obtained taking to the derivative of Z with respect
to β. This translates to the relation E = F + T S. The relation between F (T, X) and E(S, X) is formally a Legendre
transform. Consequently dF = −SdT − ydX and the associated state equations are
∂F ∂F
S=− , y=− , (4.8)
∂T ∂X
Within the framework of the thermodynamic formalism state equations that describe physical systems are derived
from thermodynamic potentials. The latter should be expressed using their canonical variables. The common ther-
modynamic potentials are:
E (S, X) dE = T dS − ydX (4.9)

F (T, X) = E − T S, dF = −SdT − ydX (4.10)
G (T, y) ≡ F + yX, dG = −SdT + Xdy (4.11)
1 y
S (E, X) , dS = dE + dX (4.12)
T T
The derivatives of the state equations are know as the ”thermodynamic constants” though they are not really con-
stant...
∂S ∂y
C≡T χ≡ (4.13)
∂T ∂X
In the context of gases
1 ∂V
Compressibility ≡ − [common notation - ”beta” or ”kappa”] (4.14)
V ∂P
1 ∂V
ExpansionCoeff ≡ [common notation - ”alpha”] (4.15)
V ∂T
====== [4.3] The Grand Hamiltonian approach

It is customary in thermodynamics to define ”thermodynamic potentials” that are obtained from the Helmholtz free
energy by means of Legendre transform. This can be regarded as a formal mathematical trick for switching the role of
conjugate variables, but it also can be motivated physically. It is the same procedure that we had discussed regarding
the calculation of the tension of a polymer. Here we repeat it with regard to a gas in a box with piston.
Let us regard the position of the piston (the parameter X) as a dynamical variable (let us call it x). We can apply
force, say ”electric” field E on the piston. Accordingly the ”grand Hamiltonian” of the system is
HG = H(· · · , x) + Ex + [optional kinetic term] (4.16)
The optional kinetic term is required if the piston has finite mass, but its inclusion will not affect the analysis because
it factors out of the calculation. Given x = X the force that the system exerts on the piston is y(X) = h−∂H/∂xiX .
Once x becomes a dynamical variable, and E is introduced, the equilibrium point of the piston is determined by the
equation E = y(x), hence the sign convention for the second term in HG .
The partition function of HG is related to that of H by Laplace transform:
X X
ZG (β, E) = e−βEx,r = Z(β, x) e−(βE)x (4.17)
x,r x
28
This can be written as

−G(T,E)/T
X F (T, x) + Ex
e = exp − (4.18)
x
T
In the thermodynamic limit fluctuations can be neglected, and a saddle point approximation implies
n o
G(T, E) ≈ min F (T, x) + Ex = F (T, x̄) + E x̄ (4.19)
x
where the most probable value x̄ is determined by solving the state equation E = −F 0 (x). Accordingly we realize that
G(T, E) is the Legendre transform of F (T, X). The roles of the conjugate variable X and E have been switched. If X
and E are the volume V and the pressure P , then G(T, P ) is known as the Gibbs function.
====== [4.4] The chemical potential

Consider a gas that consists of N identical particles. This can be either classical or quantum gas (contrary to prevailing
misconception, quantum mechanics is irrelevant to this issue - this will be explained in the ”chemical equilibrium”
lecture). Within the framework of the canonical formalism we define the chemical potential as follows:
∂F
µ (T, V, N ) ≡ (4.20)
∂N
Accordingly we have
dF = −SdT − P dV + µdN (4.21)

dG = −SdT + VdP + µdN (4.22)
The above definition of the chemical potential can be motivated by adopting a ”grand Hamiltonian” perspective. Let
us define a ”grand system” that consists of the system and of a reservoir of particles. This reservoir consists of a
huge collection of sites that hold a huge number of particles with binding energy µ. If we transfer N particle from
the reservoir to the system the energy of the ”grand system” becomes
HG = H − µN (4.23)
The so called grand partition function ZG (β, µ) of the Grand system will be discussed in future lecture.
====== [4.5] The extensive property

At this stage it is appropriate to remark on a relation between the Gibbs function and the chemical potential that
holds is the so-called thermodynamic limit. In this limit the system acquires an extensive property that can be
formulated mathematically. Relating to the Gibbs function G(T, P ; N ), one observes that if N is multiplied by some
factor, then the volume V and the entropy S for the same (T, P ) are expected to be multiplied by the same factor,
and hence also G should be multiplied by the same factor. We therefore write
G(T, P, N ) = N G(T, P, 1) (4.24)
From µ = −dG/dN we deduce that the chemical potential is merely the Gibbs energy per particle. Consequently
from the expression for dG it follows that
S V
dµ = − dT + dP (4.25)
N N
29
====== [4.6] Work

In the definition of work the system and the environment are regarded as one driven closed unit. If we change X in
time then from the “rate of change formula” we have the following exact expression:

dE ∂H
= = −hFit Ẋ (4.26)
dt ∂t
it follows that
Z
W ≡ work done on the system = Efinal − Einitial = − hF it dX (4.27)
This is an exact expression. Note that hFit is calculated for the time dependent (evolving) state of the system. In
a quasi-static adiabatic process one replaces hFit by hFiX(t) , where the notation hFiX implies that the system is
assumed to be in a canonical state at any moment. More generally, within the framework of linear response theory
hFit ≈ hF iX − η Ẋ = y(X) − η Ẋ (4.28)
The first terms is the conservative force, which is a function of X alone. The subscript implies that the expectation
value is taken with respect to the instantaneous adiabatic state. The second term is the leading correction to the
adiabatic approximation. It is the “friction” force which is proportional to the rate of the driving. The net conservative
work is zero for a closed cycle while the “friction” leads to irreversible dissipation of energy with a rate
Ẇirreversible = η Ẋ 2 (4.29)
More generally it is customary to write
W = −W + Wirreversible (4.30)
where the first term is the conservative work, or so to say “the work which is done by the system”
Z Z XB
W = hFiX dX = y(X) dX (4.31)
XA
The two main examples that illustrate the above discussion are:
Example 1: box with piston Example 2: ring with flux
X = position of a wall element (or scatterer) X = magnetic flux through the ring
Ẋ = wall (or scatterer) velocity −Ẋ = electro motive force
hFi = Newtonian force hFi = electrical current
−η Ẋ = friction law −η Ẋ = Ohm law
η Ẋ 2 = rate of heating η Ẋ 2 = Joule law
30
In the first example instead of having a displaceable wall (”piston”) we can have a moveable object inside the box
(”scatterer”). In the latter case there is friction while the conservative force is zero (because the volume of the box is
not changing).
====== [4.7] Heat

In order to understand which type of statements can be extracted form the canonical formalism we have to discuss
carefully the physics of work and heat. We distinguish between the system and the environment and write the
Hamiltonian in the form
Htotal = H(r, p; X(t)) + Hint + Henv (4.32)
It is implicit that the interaction term is extremely small so it can be ignored in the calculation of the total energy.
The environment is characterized by its temperature. More generally we assume that the environment consists of
several “baths” that each has different temperature, and that the couplings to the baths can be switched on and off.
Below we consider a process in which both the initial and the final states are equilibrated with a single bath. This
requires that at the end of the driving process there is an extra waiting period that allows this equilibration. It is
implied that both the initial and the final states of the system are canonical. Now we define

W = work ≡ hHtotal iB − hHtotal iA (4.33)

Q = heat ≡ − hHenv iB − hHenv iA (4.34)
Efinal − Einitial ≡ hHiB − hHiA = W + Q (4.35)
It is important to emphasize that the definition of work is the same as in the previous section, because we regard
Htotal as describing an isolated driven system. However, E is redefined as the energy of the system only, and therefore
we have the additional term Q in the last equation.
====== [4.8] Quasi static process

In general we have the formal identity:
X X
dE = dpr Er + pr dEr (4.36)
r
We would like to argue that for an ideal (reversible) quasi-static process we can identify the first term as the heat
¯ and the second term is the work −¯
dQ dW . One possible scenario is having no driving. Still we have control over the
temperature of the environment. Assuming a volume preserving quasi-static process we have
dX = 0 (4.37)
X
dE = dpr Er = T dS (4.38)
dQ
¯ = T dS (4.39)
dW
¯ = 0 (4.40)
Q = [E(B) − E(A)] (4.41)
W = 0 (4.42)
31
A second possible scenario is having an isolated system going through an adiabatic process:
dpr = 0 (4.43)
X
dE = pr dEr = −ydX (4.44)
r
dQ
¯ = 0 (4.45)
dW
¯ = ydX (4.46)
Q = 0 (4.47)
W = −[E(B) − E(A)] (4.48)
Any general quasi-static process can be constructed from small steps as above, leading to
Z B
Q = T dS (4.49)
A
Z B
W = y(X)dX (4.50)
A
In particular for isothermal process we get
Q = T × [S(B) − S(A)] (4.51)

W = −[F (B) − F (A)] (4.52)
If a process is both isothermal (constant T ) and isobaric (constant P ) we can still get work being done by changing
some other parameter X. For example X might be a ”reaction coordinate”. Then we get
Q = T × [S(B) − S(A)] (4.53)

W = −[G(B) − G(A)] (4.54)
====== [4.9] Cycles

It is possible to design cycles in (X, T ) space, such that the net effect is to convert heat into work (engine) or in reverse
(heat pump). Consider for example a gas in a cylinder with a piston. If there is no restriction on the availability of
baths the simplest engine could work as follows: Allow the gas to expand at high temperature; Lower the temperature;
Compress the gas back by moving the piston back to its initial position; Raise back the temperature. The net effect
here is to convert heat into work. This is known as the Stirling cycle. A traditional version of a Stirling engine is
displayed in the following figure [left panel taken from Wikipedia]:
32
In order to see the relation between the engine and the cycle it is proposed to analyze the operation as follows. Denote
by XH and XC the volumes of the hot and cold cylinders. As the wheel is rotated it defines a cycle in (XC , XH )
space. The XH > XC segment of the cycle represents expansion of gas during the stage when most of it is held in
high temperature. The XH < XC segment represents the compression of the gas during the stage when most of it is
held in low temperature.
The disadvantage of the Stirling cycle is that in order to realize it in a reversible manner we need infinitely many
intermediate baths in the heating and cooling stages. The way to do it in practice is to use a “heat exchange” device.
This device can be regarded as layered structure that is attached in one end to the hot bath and in the other end to
the cold bath. As a result each layer is held in a different temperature. We assume that the layers are quesi-isolated
from each other. The trick is to couple the pipes that lead the gas between the hot and the cold cylinders to this
layered structure, such that they can exchange heat with the layers without net effect on the temperature of the layer.
If we want to use a reversible cycle that uses two baths only, we can consider the Carnot cycle. See block diagram
above [taken form Wikipedia]. Note that if we operate this cycle in reverse we get a heat pump instead of an engine.
Let us analyze what happens during a Carnot cycle. Assuming that the levels become more dense as X is increased,
it follows that the result of an adiabatic process would be a lower temperature (adiabatic cooling). To make this point
clear consider just two levels E1 and E2 with occupation probabilities p1 and p2 respectively. The implied temperature
is T = (E2 − E1 )/[− ln(p2 /p1 )]. In an adiabatic process the probabilities do not change, hence as the level get closer
the implied temperature become lower. If the process is isothermal rather than adiabatic there will be heat absorption
(isothermal absorption) and re-distribution of the probabilities such that p2 /p1 = exp[−(E2 − E1 )/T0 ] . These “rules
of thumb” allow to gain intuition with regard to the operation of engines and heat-pumps.
Besides the piston example, the other simplest example for a thermodynamic cycle concerns spin S 1 in magnetic
field. In order to be consistent with the piston example we define X = −|h|, so larger X is like larger volume, i.e. higher
density of states. We consider a cycle that consists of 4 stages: adiabatic cooling to lower temperature; isothermal
absorption stage (QC > 0); adiabtic heating to higher temperature; and isothermal emission stage (QH < 0). The net
effect is is to transfer heat from the cold bath to the hot bath, which requires to invest work.
At each stage the work W is positive or negative depending on whether the occupied levels go down or up respectively.
The inclination is to say that during the adiabatic cooling stage the work is positive. This is true in the piston example,
but not in general, as can be verified with the spin example. It should be clear that doing work on the system does
not imply that its temperature becomes higher: the simplest demonstration would be to take an isolated container
with gas to the top of Eifel Tower: it requires work but the temperature is not affected. What is essential for the
operation of the cycle is the variation in the density of the levels, irrespective of whether they go up or down during
the cycle.
33
[5] Chemical equilibrium and the Grand Canonical state
====== [5.1] The Gibbs prescription

In this lecture we are going to discuss chemical equilibrium. We shall see that the condition of chemical equilibrium
involves the chemical potentials of the participating gases. For the purpose of calculating µ it is essential to find how
the partition function depends on the number of particles. Classically the calculation of ZN for a gas of identical
particles is done using the Gibbs prescription:
1
ZN [Gibbs] = ZN [distinguishable particles] (5.1)
N!
We shall motivate this prescription in the following discussion of chemical equilibrium. For an ideal gas we get
1 N V X −βεbound V
ZN = Z , where Z1 = e ≡ g0 3 (5.2)
N! 1 3
λT λT
The summation is over the non-translational freedoms of the particle. Hence we get

∂F N N 3
µ = = T ln = ε0 + T ln λ (5.3)
∂N Z1 V T
where in the last equality we have assumed that the particle has a single well defined binding energy. The inverse
relation is
V −(ε0 −µ)/T
N = Z1 eβµ = e (5.4)
λ3T
The notion of identical particles:– The notion of identical particles does not require extra explanations if they
are indistinguishable as in the quantum mechanical theory. Still we can ask what would happen if our world were
classical. The answer is that in a classical reality one still has to maintain the Gibbs prescription if one wants to
formulate a convenient theory for Chemical equilibrium. Namely, the condition for ”chemical equilibrium” that we
derive below has a simple form if we adopt the Gibbs prescription. Without the Gibbs prescription one would be
forced to formulate an equivalent but non-friendly version for this condition.
====== [5.2] Chemical equilibrium

Consider the following prototype problem of chemical equilibrium:
A[a]
A[b] (5.5)
where ”a” and ”b” are two phases, say to be in one of two regions in space, or to be either in the bulk or on the
boundary of some bounded region. Given N identical particles we characterize the macroscopic occupation of the two
phases by a reaction coordinate n, such that N − n particles are in phase [a] and n particles are in phase [b]. The
partition function is
X N!

Zab
N = ZaN −n Zbn (5.6)
n
(N − n)! n!
The combinatorial ”mixing” factor in the curly brackets counts the number of possibilities to divide N particles into
two groups. It should be excluded if the particles are indistinguishable, as in the quantum theory. In the classical
34
theory, where the particles are distinguishable it should be included, but it can be absorbed into the definition of the
partition function. This is what we call the “Gibbs prescription”. Using the Gibbs prescription the above sum can
be re-written as follows:
X
ab a b
ZN = ZN −n Zn (5.7)
n
The probability to have (N −n, n) occupation is proportional to the nth term in the partition sum:
ZaN −n Zbn a b

N! ZN −n Zn
= C exp −β F a (N − n) + F b (n)

p(n) = × = (5.8)
(N − n)!n! Zab
N ZNab
One should appreciate the usefulness of the Gibbs prescription. It is thanks to this prescription that the Free Energy
is additive. If we did not use the Gibbs prescription we would be compelled to add in F a term that reflects ”mixing
entropy”. The most probable value n̄ is determined by looking for the largest term. This leads to the Chemical
equilibrium condition:
F a (N − n) + F b (n) = minimum (5.9)

; −µa (N − n) + µb (n) = 0 (5.10)
Let us consider the case of ideal gases. Using the expression for µ we get
n Zb Z1b
= 1a ; n̄ = N (5.11)
N −n Z1 Z1a + Z1b
This example is simple enough to allow a determination of the average value hni too. The probability distribution of
the reaction coordinate is
N −n n
N! (Z1a ) Z1b
p(n) = (5.12)
(N − n)!n! (Z1a + Z1b )N
leading to
X
hni = p(n) n = n̄ (5.13)
n
We see that the expectation value of n coincides with its typical (most probable) value. In the more general case of
chemical equilibrium, as discussed below, this is an approximation that becomes valid for N 1 in accordance with
the central limit theorem.
====== [5.3] The law of mass action

This procedure is easily generalized. Consider for example
2C
5A + 3B (5.14)
Given that initially there are NA particles of type A, NB particles of type B, and NC particles of type C we define a
macroscopic reaction coordinate n such that NC −2n is the number of particles of type C, and NA +5n is the number
of particles of type A, and NB +3n is the number of particles of type B. Accordingly
X
Z abc = c
ZN Za Zb
C −2n NA +5n NB +3n
(5.15)
n
35
and
c
(NC −2n)+F a (NA +5n)+F b (NB +3n))
p(n) = const e−β (F (5.16)
leading to the equation
−2µc (NC −2n) + 5µa (NA +5n) + 3µb (NB +3n) = 0 (5.17)
which with Boltzmann/Gibbs approximation becomes
5 3
(NA +5n)5 (NB +3n)3 (Z1a ) Z1b
= (5.18)
(NC −2n)2 (Z1c )2
or, assuming that [a],[b],[c] are all volume phases,
NA +5n 5 NB +3n 3

V V
= κ(T ) (5.19)
NC −2n 2

V
where the reaction rate constant κ(T ) ∝ e−ε/T depends on the reaction energy ε = 5εa + 3εb − 2εc . In this sign
convention ε < 0 means exotermic reaction.
====== [5.4] Equilibrium in pair creation reaction

Consider the reaction
γ+γ
e+ + e− (5.20)
This can be analyzed like a chemical reaction C

A + B, which is of the same type as considered in the previous
version. The important point to notice is that Z c is independent of n, and therefore the chemical potential of the
electromagnetic field is formally µc = 0. The electromagnetic field is like a ”bath”, and we can regard it as part of
the environment, hence we could have written vacuum
e+ + e− . In any case we get at equilibrium
+ −
µe (n1 ) + µe (n2 ) = 0 (5.21)
where in the Boltzmann/Gibbs approximation
nλ3T

2
µ(n) ≈ mc + T ln (5.22)
V
leading to
2
V 2
n1 n2 = e−2mc /T
(5.23)
λ3T
This problem is formally the same as that of a semiconductor where e+ and e− are the holes and the electrons,
and 2mc2 corresponds to the energy gap between the valance and the conduction bands. Accordingly, an optional
derivation of the latter equilibrium condition can be based on a grand-canonical perspective (see next lecture) with
regard to the occupation of the electrons.
36
====== [5.5] Equilibrium in liquid-gas system

The equilibrium between a liquid phase and a gaseous phase is just another example for a chemical equilibrium. We
can write the equation that determines the coexistence curve in (T, P ) diagram as [µa (T, P ) − µb (T, P )] = 0. By
implicit differentiation of this equation with respect to T we get the Clausius-Clapeyron relation

dP ∂T [µa − µb ] ∆S 1 [Latent heat]
= − = = (5.24)
dT coexistence ∂P [µa − µb ] ∆V T [Volume change]
Outside of the coexistence curve either µa or µb are smaller, and accordingly all the particles occupy one phase only.
====== [5.6] Site system

The chemical potential can be calculate easily for a system of N identical particles that occupy a set of M sites (or
modes) that have the same binding energy ε. Since we assume that the biding energy is the same for all sites, it follows
that estimating Z1 is essentially a combinatorial problem. We assume n 1 so we can approximate the derivative of
ln(n!) as ln(n). We also write the result for the most probable n which is obtained given µ. Note that the result for
n̄ is meaningful only for large M .
Fermionic site:– Each site can have at most one particle
M!
Zn = e−βεn (5.25)
n!(M − n)!

n
µ = ε + T ln (5.26)
M −n
n̄ = M (eβ(ε−µ) + 1)−1 (5.27)
Bosonic site:– Each site can have any number of particles. The combinatorial problem is solved by asking how
many ways to divide n particles in a row with M − 1 partitions. If the particles were distinct the result would be
(n + (M − 1))!. Taking into account that the particles are indistinguishable we get
(n + M − 1)! −βεn
Zn = e (5.28)
n!(M − 1)!

n
µ = ε + T ln (5.29)
(M − 1) + n
n̄ = (M − 1)(eβ(ε−µ) − 1)−1 (5.30)
Electromagnetic mode:– Each mode of the electromagnetic field can be regarded as a Bosonic site that can occupy
photons with binding energy ω. Since n is not constrained it follows formally that
µ = 0 (5.31)
n̄ = (eβω − 1)−1 (5.32)
37
Boltzmann approximation:– Assuming dilute occupation (1 n M ) we get a common approximation for

both Fermi and Bose case:
M n −βεn
Zn = e (5.33)
n! n
µ = ε + T ln (5.34)
M
n̄ = M e−β(ε−µ) (5.35)
General system of sites:– If we want to consider the partition function of N particles in M sites that have different
binding energies we have to calculate
X
ZN (β) = e−β(ε1 n1 +...+εM nM ) (5.36)
n1 +...+nM =N
Because of the constraint the sum cannot be factorized. We therefore adopt the ”Grand Hamiltonian” strategy and
calculate the Grand partition function Z(β, µ) that corresponds to HG = H − µN . In principle we can get ZN (β)
from Z(β, µ) via an inverse transform, but in practice it is more convenient to stay with the Grand Hamiltonian
framework.
====== [5.7] The grand canonical formalism

We can regard the grand canonical formalism as a special case of the canonical formalism, where the Grand Hamilto-
nian HG = H − µN describes a Grand system that consists of the gas particles and a hypothetical reservoir. Optionally
we can motivate the introduction of a the grand canonical formalism following the same justification strategy as in
the case of the canonical formalism. First we have to specify the many body eigenstates R of the system:
N̂ |Ri = NR |Ri (5.37)

Ĥ|Ri = ER |Ri (5.38)
Then we assume that the system can exchange particles as well as energy with the environment. The probability of
a many body eigenstate R is
e−βER ZN̄
env
−N env
pR = , with ZN̄ −N ∝ eβµN (5.39)
Z sys+env
We deduce that
1 −β(ER −µNR )
pR = e (5.40)
Z
where the normalization constant is

X
Z(β, µ) ≡ e−β(ER −µNR ) (5.41)
R
The Grand Canonical Z(β, µ) is defined in complete analogy with the canonical case as sum over the many body
states ”R”. For some purposes it is convent to write is as a function Z(z; β) of the fugacity:
1 ∂ ∂
z ≡ eβµ , =z (5.42)
β ∂µ ∂z
38
The Grand Canonical Z(β, µ) can serve as a generating function as follows:
X 1 ∂ ln Z
N ≡ hN̂ i = pR NR = (5.43)
β ∂µ
R
∂ ln Z
E − µN = − (5.44)
∂β

∂H 1 ∂ ln Z
P ≡ − = (5.45)
∂V β ∂V
Equivalently
1
FG (T, V, µ) ≡ − ln Z (5.46)
β
∂FG
N =− (5.47)
∂µ
∂FG
P =− (5.48)
∂V
∂FG
S=− (5.49)
∂T
E = FG + T S + µN (5.50)
In the thermodynamic limit FG is extensive, also in the case of non ideal gas. Consequently
FG (T, V, µ) = −VP (T, µ) (5.51)

S N
dP = dT + dµ (5.52)
V V
In other words rather then using the notation FG , we can regard P (T, µ) as the generating function. Note that this
is the ”Grand canonical” version of the ”canonical” Gibbs function relation
S V
dµ = − dT + dP (5.53)
N N
For constant T , a variation in the chemical potential is related to a variation dP = ndµ in the pressure, where n = N/V
is the density. In the canonical setup N is fixed, while in the grand-canonical setup V is fixed. The compressibility of
the gas can be expressed as follows:

1 dV 1 dN 1 dn 1 dn
κT = − = = = (5.54)
V dP N N dP V n dP n2 dµ
====== [5.8] Fermi occupation

A site or mode can occupy n = 0, 1 particles. The binding energy is . the site is in thermochemical equilibrium with
a gas in temperature β and chemical potential µ.
Nn = n (5.55)
En = n (5.56)
1 −β(−µ)n
pn = e (5.57)
Z
39
and accordingly,

Z(β, µ) = 1 + e−β(−µ) (5.58)
X 1
N (β, µ) = hn̂i = pn n = ≡ f ( − µ) (5.59)
n
eβ(−µ) + 1
E(β, µ) = hn̂i = f ( − µ) (5.60)
We have defined the Fermi occupation function 0 ≤ f ( − µ) ≤ 1
====== [5.9] Bose occupation

A site or mode can occupy n = 0, 1, 2, 3... particles. The binding energy is . the site is in thermochemical equilibrium
with a gas in temperature β and chemical potential µ.
Nn = n (5.61)
En = n (5.62)
1 −β(−µ)n
pn = e (5.63)
Z
and accordingly,
−1
Z(β, µ) = 1 − e−β(−µ) (5.64)
X 1
N (β, µ) = hn̂i = pn n = ≡ f ( − µ) (5.65)
n
eβ(−µ) −1
E(β, µ) = hn̂i = f ( − µ) (5.66)
We have defined the Bose occupation function 0 ≤ f ( − µ) ≤ ∞. If < µ then hni → ∞. If = µ then the site may
have any occupation. If < µ then hni is finite.
====== [5.10] Bosonic mode occupation

The occupation of a mode of vibration, say the number photons in an electromagnetic mode, or the number of
phonons in a vibration mode, are described by the canonical ensemble, by can be optionally regarded as described by
the grand-canonical ensemble with µ = 0. With slight change in notations we have:
Nn = n (5.67)
En = nω (5.68)
1 −βωn
pn = e (5.69)
Z
and accordingly,
−1
Z(β) = 1 − e−βω (5.70)
X 1
N (β) = hn̂i = pn n = βω
≡ f (ω) (5.71)
n
e −1
E(β) = hn̂ωi = ωf (ω) (5.72)
40
[6] Quantum ideal gases
====== [6.1] Equations of state

In what follows, unless written otherwise = 0 is the ground state and
X Z ∞
→ g()d (6.1)
r 0
The stationary states of the multi particle system are occupation states
|ni = |n1 , n2 , n3 , ..., nr , ...i (6.2)
where nr = 0, 1 for Fermi occupation and nr = 0, 1, 2, 3, 4, ... for Bose occupation. For these states we have
X
Nn = nr (6.3)
r
X
En = n r r (6.4)
r
P
−β r (r −µ)nr
pn ∝ e (6.5)
which can be factorized. This means that each site or mode can be treated as an independent system. We use E
and N without index for the expectation values in an equilibrium state. For the Fermionic and Bosonic case we have
respectively (±)
X Z ∞
ln Z = ± ln(1 ± e−β(−µ) ) = β N ()d f ( − µ) (6.6)
r 0
X X Z ∞
N = hn̂r i = f (r − µ) = g()d f ( − µ) (6.7)
r r 0
X X Z ∞
E = r hn̂r i = f (r − µ)r = g()d f ( − µ) (6.8)
r r 0
∞
1 ln Z
Z
1
P = = N ()d f ( − µ) (6.9)
β V V 0
It is good to remember that P (T, µ) can serve as a generating function for all other state equations. This would be
true also if the gas were not ideal. In particular N/V = dP/dµ relates the density to the chemical potential, which
implies a relation between the the pressure P and the density N/V.
====== [6.2] Explicit expressions for the state equations

We assume one particle states |ri that have the density
1
g() = Vc α−1 , N (E) = g() (6.10)
α
For a particle in d dimensional box α = d/ν where ν is the exponent of the dispersion relation ∝ |p|ν , and c is a
constant which is related to the mass m. For example, in the case of spin 1/2 particle in 3D space we have
(2m)3/2 1
g() = 2 × V 2 (6.11)
(2π)2
41
The following integral is useful (upper sign for Bose, lower sign for Fermi):
∞
xα−1 dx
Z
Fα (u) ≡ ≡ ±Γ(α)Liα (±z), z ≡ eu (6.12)
0 ex−u ∓ 1
where the upper/lower sign refers to the Bose and the Fermi case respectively. Details of the Polylogarithm function
Liα (z) can be found in Wikipedia. In the physics community it is commonly denoted as gα (z). Note that
∞
X 1 ` d 1
Liα (z) ≡ z = z + ..., Liα (1) ≡ ζ(α), Liα (z) = Liα−1 (z) (6.13)
`α dz z
`=1
As u becomes larger the function Fα (u) grows faster in the case of a Bose occupation, and it either diverges or attains
√
a finite value as u → 0. The finite value Fα (0) = Γ(α)ζ(α) is attained for α > 1. In particular we have Γ(3/2) = π/2
and ζ(3/2) ≈ 2.612. For α = 1 one obtains Li1 (z) = − ln(1 − z), which has logarithmic divergence as z → 1. For α < 1
it is easily shown that Fα (u) ∼ [1/(1−α)](−u)−(1−α) as u approach zero from below. In the Fermi case the integral
is always finite. Using the step-like behavior of the Fermi occupation function we obtains for z 1 the so-called
Sommerfeld expansion:
" 2 #
1 α π2 1
Fα (u) = u 1 + α(α−1) + ... (6.14)
α 6 u
Fα α>1
Bose
Fermi
1 z
We can express the state equations using this integral, where z is identified as the fugacity. We get
N µ
= cT α Fα (6.15)
V T
E α+1
µ
= cT Fα+1 (6.16)
V T
while P is related trivially to the energy:

1 ln Z 1 E
P = = (6.17)
β V α V
The grand-canonical free-energy is FG = −V P , from which one can derive the entropy S = −(dFG /dT )µ . Optionally
the canonical free energy for N particles can be calculated via Legendre transform F = FG + µN . The specific results
in the case of a spinless non-relativistic Bose particles are [see also Huang p.231-232;242]:
N 1
= 3 Li3/2 (z) (6.18)
V λT

E 3 T 2 E T
= Li5/2 (z), P = = Li5/2 (z) (6.19)
V 2 λ3T 3 V λ3T
42
====== [6.3] Ideal gases in the Boltzmann approximation

We take = 0 as the ground state energy of the one-particle states. The Boltzmann approximation is
f ( − µ) ≈ e−β(−µ) (6.20)
It holds whenever the occupation is f () 1. If it is valid for the ground state = 0, then it is valid globally for all the
higher levels. Accordingly the validity condition is z 1, meaning eβµ 1. Under such condition one can make the
approximation Li(z) ≈ z. In the case of standard 3D gas the Boltzmann approximation condition can be rewritten as
1
N λ3T V ; T ; T Tc (Bosons), ; T TF (Fermions) (6.21)
m`2
where ` = (V /N )1/3 is the typical distance between particles. Is later sections we shall defined the condensation
temperature (Tc ) and the Fermi energy (TF ). Within the framework of the Boltzmann approximation we can re-
derive the classical equation of an ideal gas:
N 1 1
= 3 z = 3 eµ/T (6.22)
V λT λT
E 3 T 3N
= 3 z = T (6.23)
V 2 λT 2V
T N
P = 3z = T (6.24)
λT V
Note that within this approximation E and P do not depend on the mass of the particles.
E E
fermi
boltzman
0
bose
g(E)
µ f(E−µ )
1/2 1
====== [6.4] Bose Einstein condensation

Let us write again the general expression for the occupation of the orbitals:
Z
X 1 α−1 1
N (β, µ) = hn0 i + hnr i = + cV d (6.25)
r>0
e−βµ − 1 eβ(−µ) − 1
In the limit µ → 0− this expression always diverges, so we can invert it and find µ as a function of N . But the physics
is more illuminating if the ground-orbital occupation (n0 ) is dropped from the above expression. Then we realize
that for α > 1, notably for α = 3/2, the total occupation remains finite, namely N (µ → 0− ) = cVΓ(α)ζ(α)T α . It is
43
implied that the excited states can accommodate only a finite fraction N/V of particles in the thermodynamic limit
(V → ∞). Any additional amount of particles forces µ = 0, and has to condense into the ground state orbital. The
conclusion if different for α < 1. For clarity we change notation to V = Ld and α = d/2. The integral is dominated
by the implicit lower cutoff k ∼ 1/L. Hence we get N (µ → 0− ) ∝ Ld (L2 )1−(d/2) T . It is implies that the excited
states can accommodate any fraction N/Ld of particles in the thermodynamic limit (L → ∞). So in the latter case
condensation is not forced. The figure below illustrates the reasoning of extracting µ versus T for a given N in both
cases .
N
N
Τ1 <Τ2<Τ3 V
T1 <T2 <T3 V
µ µ
for α <1 for 1<α
µ µ
(α < 1) (1< α )
T Τc T
Considering α > 1, having “µ = 0” implies that
N = n0 + cVΓ(α)ζ(α)T α (6.26)
E = cVΓ (α+1) ζ (α+1) T α+1 (6.27)

1 E
P = (6.28)
α V
In particular the standard results for condensation in 3D are

3
3 m 2 3
N = n0 + Vζ T2 (6.29)
2 2π
3
5 m 2 5
P = ζ T2 (6.30)
2 2π
The pressure P is independent of the total number of particles, because the condensate does not have any contribution.
Hence the compressibility κ ∝ (∂P /∂V)−1 = ∞. If we change the volume the extra/missing particles just come from
the ground state, which is like a reservoir of µ = 0 particles.
Given T , if we push N particles into a box, the condition to have condensation is N > N (β, µ → 0− ). The condensation
temperature, below which µ = 0, is
1/α
1 N 1
Tc = ∼ (6.31)
cΓ (α) ζ (α) V m`2
where ` is the average distance between the particles. Given N , if one tries to eliminate µ, and writes it as a function
of T , then one observes that below the condensation temperature µ is forced to become zero. Under such circumstances
44
all the particles that cannot be occupied in the excited states have to condense in the ground state:
α
−
T
hn0 i = N − N β, µ → 0 = 1− N (6.32)
Tc
The common phrasing is that a macroscopic fraction of the particles occupies the ground state. This fraction is
determined by (T /Tc )α or equivalently by [V/λ3T ]/N . Note that T Tc is an optional way to write the Boltzmann
condition.
====== [6.5] Fermi gas at low temperatures

At zero temperatures the Fermi function is a step function. At finite temperatures the step is smeared over a range T .
In order to find explicit expressions for the state functions we have to perform an integral that involves the product
of f () with a smooth function g(). The latter is the density of states g() if we are interested in N , or g() if we
are interested in E. The Sommerfeld expansion is a procedure to get an approximation, say, to second-order in T .
For this purpose we first define the zero temperature result
Z µ
G(µ) ≡ g()d (6.33)
−∞
And then proceed with the finite temperature calculation using integration by parts:
Z ∞ Z ∞ Z ∞
d g() f ( − µ) = d G() [−f 0 ( − µ)] ≡ d G() δT ( − µ) (6.34)
−∞ −∞ −∞
Z ∞
1
= d G(µ) + G0 (µ)( − µ) + G00 (µ)( − µ)2 + ... δT ( − µ) (6.35)
−∞ 2
π 2 2 00
= G(µ) + T G (µ) + O(T 4 ) (6.36)
6
We can apply this formula to the N = N (µ) calculation. First we do the zero temperature integral, and from it
eliminate µ as a function of N . This zero temperature result is known as the Fermi energy µ = F . Then we
substitute µ = F + δµ in the above second order expression, expand G(µ) ≈ G(F ) + g(F )δµ , and find
π 2 g0 (F ) 2
µ(T ) ≈ F − T + O(T 4 ) (6.37)
6 g (F )
E E
µ
f(E−µ )
0 g(E)
45
Ν
Liquid
Bose
Fermi
Gas
µ
The specific result for N fermions in system with α−1 density of orbitals, and in particular for spinless non-relativistic
fermions in 3D box is:
" 2 # " 2 #
1 α π2 T 1 3 3 π 2
T
N = cVµ 1 + α (α − 1) + ... = V 2 (2m) 2 µ 2 1 + + ... (6.38)
α 6 µ 6π 8 µ
leading after elimination to
α1 23
αN 1 2N
F = = 6π (6.39)
c V 2m V
" 2 # " 2 #
2
π2

π T T
µ = 1 − (α−1) + ... F = 1− + ... F (6.40)
6 F 12 F
For the energy we get

" 2 # " 2 #
3 1 3 5 5π 2 T 5π 2 T 3
E = V 2
(2m) µ 2 1 +
2
+ ... = 1+ + ... N F (6.41)
5 6π 8 µ 12 F 5
The pressure is given by the equation
53
2 E 1 2 1 N
P = = 6π 2 3 + O(T 2 ) (6.42)
3 V 5 m V
The grand-canonical free-energy is FG = −V P from which one can derive the entropy S = −(dFG /dT )µ ∝ T . The
canonical free energy for N particles can be calculated via Legendre transform F = FG + µN leading to
" 2 #
5π 2

T 3
F (T, V, N ) = 1− + ... N F (6.43)
12 F 5
From here one can recover the expression for the entropy S = −(dF/dT )N , and additionally calculate the heat capacity
CV = T (dS/dT )V,N ∝ T for a closed system of fermions.
46
Systems with interactions

[7] Interactions and phase transitions
Energy scales:– With regard to the prototype models of systems with interactions there are generically two energy
scales. One is the interaction strength ε, and the other is the temperature T . For T ε a perturbative treatment is
appropriate. See below the cluster expansion. For T ε it is advised to re-model the system with Hamiltonian that
describes its collective excitations. The interesting regime is T ∼ ε where the phase transition takes place.
Models of interest:– It is natural to start with the discussion with the phenomenology of the gas-liquid phase
transition, as implied by the Van-der-Waals equation of state. Later one realizes that the essential physics is captured
by the ”lattice gas” version, which is formally equivalent to the ”Ising model” that describes a ferromagnetic phase
transition. Its generalization is known as the ”Potts model”. The system consists of N sites. At each site there is
a ”spin” that can be in one of s possible states. The Ising model is a special case with s = 2, and the interaction
is σi σj , where σ = ±1. The ”Ising model” has a discrete up/down symmetry. Its Field theory version is known as
the Landau model. The Heisenberg model is a different lattice model that has Si · Sj interaction. This interaction
has a continuous symmetry with respect to rotations. Its 2D version is known as the XY model. There are also
corresponding Field theory models that are known as non-linear sigma models.
First order phase transition.– There are systems where there are (say) two families of states, such that each
family has different DOS with different minimum. In such case a control parameter (call it h) might induce a crossover
from the dominance of one family to the dominance of the second family. This crossover is reflected in the partition
function and hence in the heat capacity and in the state equations. In the thermodynamic limit the crossover might
be abrupt. In such case it is a ”first order phase transition”. If a change in a parameter leads to a bifurcation in the
calculation of the partition function, it is called a ”second order phase transition”. The prototype example for phase
transition is ferromagnetism where the magnetization might be ”up” or ”down”.
Order parameter.– In order to analyze a second order phase transition it is useful to identify the ”order parameter”,
which is a field ϕ(x) that describes the coarse grained state of the system. In the prototype example of ferromagnetism
it is the magnetization density in the sample. Defining an entropy functional S[ϕ] that reflects the number of
microscopic states that have the same field configuration, we can express the partition function as
X 1h i
Z = e−A[ϕ] , A[ϕ] = E[ϕ] − T S[ϕ]
ϕ
T
Symmetry breaking.– Second order phase transition is due spontaneous symmetry breaking leading to long range
order. At T = 0 the definition of symmetry breaking is very simple. It means that E[ϕ] attains (say) two distinct
minimum that are described by different field configurations (different ”order”). However, at finite temperature the
canonical state is not the minimum of the energy functional E[ρ] but of the free energy functional F [ρ] = E[ρ]−T S[ρ].
Accordingly entropic contribution may wash away the broken symmetry.
There is an implicit assumption with regard to the possibility to observe ”symmetry breaking”. It is assumed that
ϕ(x) has slow dynamics. If a magnet is prepared (say) with ”up” magnetization then it takes a huge time until it
flips to the quasi degenerate ”down” magnetization.
Long range order.– In the prototype examples at high temperatures there is no ”order” meaning that the correlation
function g(r) = hϕ(r)ϕ(0)i decays exponentially. As the critical temperature is approached from above the correlation
length diverges. Below the critical temperature there is ”long range order” and the correlation function should be
re-defined with respect to the new order. There is a possibility to witness ”infinite order phase transition” where
below the critical temperature there is no long range order, but instead the correlation function become powerlaw.
See discussion of the XY model.
Formal analysis.– Disregarding a few models that possess exact solutions, the analysis of the partition function
can be done by adopting the following stages: (1) Primitive mean field theory evaluates the partition function by
calculating A[ϕ] for the field configuration that minimizes it. This corresponds mathematically to an inferior saddle
point approximation. (2) Proper mean field theory is based on proper saddle point approximation, which means
that the calculation takes into account the Gaussian fluctuations around the minimum. (3) Renormalization Group
(RG) treatment is required in the critical regime, whenever the Gaussian approximation in not valid. It explains the
scaling anomalies that are witnessed in the vicinity of the critical temperature.
47
====== [7.1] Gas of weakly interacting particles

Consider a classical gas of interacting particles:
N
X p~i 2 X
H = + u (~xi − ~xj ) (7.1)
i=1
2m
hiji
The partition function without the Gibbs factor is

 
N Z
1 X
ZN (β, V) = dx1 ...dxN exp −β u (xij ) (7.2)
λ3T
hiji
N Z
1 Y
≡ dx1 ...dxN (1 + f (xij )) , f (r) ≡ e−βu(r) − 1 (7.3)
λ3T
hiji
Note that the configuration space integral has the dimensions of VN . It equals VN if there are no interaction. If there
are interactions we can regard the f (r) as a perturbation. Then we can expand the product and perform integration
term by term. The result can be organized as an expansion:
N " 2 #N
V N N
ZN (β, V) = 1 + coef2 + coef3 + ... (7.4)
λ3T V V
Note that we have raised an N using (1 + N S) ≈ (1 + S)N , such that S is an expansion in powers of the density
(N/V ). From here we can derive the so called Virial expansion for the pressure:
∞
" 2 # `
NT N N X N
P = 1 + a2 + a3 + ... = T a` (T ) (7.5)
V V V V
`=1
The a` are known as the virial coefficients. From the above it is implied that
Z Z h
1 1 i
a2 = − f (r)d r = 3
1 − e−βu(r) d3 r [classical] (7.6)
2 2
More generally it is implied from the discussion in the next sections that in order to get a2 we just have to calculate
the two-body partition function Z2 . Namely:
(λ3T )2 1
Z2 − Z21

a2 = − [general, no Gibbs prescription here!] (7.7)
V 2!
The calculation of Z2 for two interacting quantum particles, given the scattering phase-shifts, has been outlined
in a past lecture. In the classical case it is standard to assume that the gas particles are like hard spheres, each
having radius R, with some extra attractive part that has depth ∼ 0 , similar to Lenard-Jones potential. Using high
temperature expansion in β we get in leading order
1h 0 i 4π ā
a2 ≈ 1− (2R)3 ≡ b̄ − [Van-der-Waals] (7.8)
2 T 3 T
The coefficients ā and b̄ appear in the phenomenological Van-der-Waals equation of state that we shall discuss in a
later stage. They are related to the attraction between the particles, and to their hard-core radius. Note that b̄ is the
48
excluded volume per particle multiplied by 2d−1 , where d = 3. Contrary to a common misconception it is only in 1D
that b̄ equals the excluded volume.
====== [7.2] The grand canonical perspective

It is simplest to deduce the Virial expansion from the grand canonical formalism. From now on the dependence on the
temperature is implicit, and we emphasize the dependence on the fugacity z. The grand canonical partition function
using the Gibbs prescription is
∞
X 1
Z(z) = ZN z N , [Here ZN is defined without Gibbs factor] (7.9)
N!
N =0
For an ideal classical gas all the ZN are determined by the one-particle partition function, namely ZN = ZN 1 . Ac-
cordingly ln(Z) includes a single term, namely ln(Z) = Z1 z. It makes sense to assume that interactions and quantum
effects will add higher order terms. Hence we postulate an expansion
∞
X 1
ln Z(z) = Bn z n (7.10)
n=1
n!
The relation between the Bn and the Zn is formally the same as the relation between commulants and moments in
probability theory:
Z1 = B1 (7.11)
Z2 = B21 + B2 (7.12)
Z3 = B31 + 3B1 B2 + B3 (7.13)
Or backwards:
B1 = Z1 (7.14)
B2 = Z2 − Z21 (7.15)
B3 = Z3 − 3Z2 Z1 + 2Z31 (7.16)
We can use these relations both directions: First we can evaluate a few ZN , typically Z1 and Z2 , in order to get the
leading order Bn coefficients, say B1 and B2 . Once the leading order Bn coefficients are known, we can generate from
them a generalized Gibbs approximation for all(!) the ZN .
====== [7.3] The cluster expansion

Our objective is to calculate the Bn coefficients in the expansion of ln(Z). For convenience we define their scaled
versions bn through the following substitution:
n
1 1
Bn ≡ V bn (T ) (7.17)
n! λ3T
We turn to outline a general diagrammatic procedure to evaluate the bn for a classical gas of interacting particles. A
graph (network, diagram) is a set of vertices (nodes) that are connected by edges (connectors, bonds). In the present
context each diagram represents an integral. The sum over all the connected diagrams that have n nodes gives the
expansion coefficient Bn of the ”comulant” generating function ln(Z), while the sum over all diagrams (including
reducible diagrams) gives the expansion coefficient ZN of the moments generating function Z. Formally we write
N Z N X
1 Y 1
ZN = dx1 ...dxN (1 + f (xij )) = [diagrams with N nodes] (7.18)
λ3T λ3T
hiji
49
In this expression a diagram represents an integral of the type

Z
C[30 , 1, 2, 3] × [f (x12 )f (x23 )] [f (x56 )] [f (x78 )f (x89 )f (x97 )] dx1 ...dx9 (7.19)
where C is a combinatorial factor that arise because we identify diagrams that differ only in the labelling of
the vertices. One should realize that if a diagram is reducible, say N = n1 + n2 + n3 , then C[n1 , n2 , n3 ] =
[N !/(n1 !n2 !n3 !)]C[n1 ]C[n2 ]C[n3 ]. In the above example C[30 ] = 3 is the number of ways to have a triangle with
2 bonds, while C[3] = 1. Using this observation it is not difficult to prove that
n X
1
Bn = [connected diagrams with n nodes] (7.20)
λ3T
The implied expression for the bn is the same diagrammatic sum, but the prefactor is replaced by 1/(n!V). The
expressions for the leading coefficients are:
Z
1
b1 = dx = 1 (7.21)
V
Z Z
1 1
b2 = f (x12 ) dx1 dx2 = f (r) dr (7.22)
2!V 2!
Z
1
b3 = [3f (x12 )f (x23 ) + f (x12 )f (x23 )f (x31 )] dx1 dx2 dx3 (7.23)
3!V
====== [7.4] The Virial coefficients

Having found the bn the grand canonical partition function is
∞ n
X z
ln Z(z) = V bn (T ) (7.24)
n=1
λ3T
where b1 = 1, and bn has the dimension of lengthn−1 . Note that for an ideal Bose or Fermi gas one obtains
bn = (±1)n+1 n−5/2 (λ3T )n−1 . We would like to find a procedure to determine these coefficients if there are weak
interactions between the particles. Once they are known we get the state equations from
∂
N = z ln Z (7.25)
∂z
T
P = ln Z (7.26)
V
leading to
∞ n
N X z
= nbn (T ) (7.27)
V n=1
λ3T
∞ n
P X z
= bn (T ) (7.28)
T n=1
λ3T
It is customary to eliminate z/λ3T from the first equation and to substitute into the second equation, thus getting the
virial expansion with the coefficients
a1 = b1 = 1 (7.29)
a2 = −b2 (7.30)
a3 = 4b22 − 2b3 (7.31)
50
====== [7.5] The Van-der-Waals equation of state

Consider a classical gas that is composed of N particles in volume V. The particles have hard core of radius R, and
the two-body interaction is assumed to be attractive, with depth ∼ 0 . We have formally obtained from the virial
expansion the following equation of state:

NT ā N
P ≈ 1 + b̄ − (7.32)
V T V
where b̄ ∼ R3 and ā ∼ 0 R3 . The effect of hard-core repulsion is under-estimated in this leading order perturbative
expansion. The add-hock correction is to re-write the equation of state as follows:
2
NT N
P = − ā (7.33)
V − N b̄ V
Roughly this equation can be derived by assuming that the partition function is like that of an ideal gas, where each
particle experiences volume Veff = (V − N b̄), and mean potential hU i = −N ā/V. Optionally the a term could have
been deduced from the virial theorem, using the estimate hr · (∂U /∂r)i ∼ N 2 (0 R3 /V). If we plot P versus V we find
that it becomes non-monotonic if the temperature is lower than a critical value. For a detailed analysis see [Huang,
section 2.3]. The critical value of the temperature is
8ā
Tc = ∼ 0 (7.34)
27b̄
The P dependence for T < Tc is illustrated in the figure below [taken from Wikipedia]. From the relation
P = −dF/dV one can deduce the free energy F (V). One can argue that there is a V range of instability where the
free energy can be lowered via phase separation. This is known as Maxwell construction (details below). A similar
reasoning can be applied to the ferromagntic phase transition where the role of V is played by the magnetization.
Maxwell construction.– Assume that we divide the volume such that V = Ṽ1 + Ṽ2 , and the particles are partitioned
such that N = Ñ1 + Ñ2 . Using the extensive property of F (V; N ) we deduce that the free energy of the mixed phase
is

Ñ1 N Ñ2 N
Fmix = F Ṽ1 + F Ṽ2 ≡ xF (V1 ) + (1 − x)F (V2 ) (7.35)
N Ñ1 N Ñ2
where x ∈ [0, 1] is the fraction of particles in phase1, while V1 and V2 are the volumes that would be occupied if all the
particles were in phase1 or in phase2 respectively. Observing that the mixture occupies a volume V = xV1 + (1−x)V2
we deuce that the mixture is represented by a point that is located on a chord that connects point1 and point2 of
the F (V ) plot above (right panel). It follows that any concave segment of F (V ) is unstable: the free energy can be
lowered via phase separation. Observing that P (V ) is the derivative of F (V ), the concave segment can be optionally
51
determined by the ”equal area” law. In reality we expect, as the volume is increased, to go along the constant pressure
equilibrium line, until all the particles evaporate from the ”liquid” phase to the ”gas” phase.
A possibly simpler perspective on Maxwell construction is to regard the applied P as the free variable, and see how
V depends on it. The grand Hamiltonian is HG = H + P V , and the grand partition function is related by Laplace
transform:
Z
1
ZG (P ) = dV e−A(V ;P ) A(V ; P ) ≡ (F (V ) + P V ) (7.36)
T
Note that a plot of A(V ) versus V is related trivially to the plot of F (V ). The integral is dominated by the minimum
of A(V ), which provide the most probable value of V . This leads to the standard Legendre prescription for the
determination of the Gibbs function. But here the situation is somewhat subtle. As P is increased we get at some
point two minima that represent stable and meta-stable solutions. As P is further increased, at some stage the two
minima will swap, implying a jump at V . This swap corresponds to the Maxwell construction. We note that the
volume (V ) as a function of the pressure (P ) is analogous to the magnetization (M ) as a function of the field (h),
which we discuss in more detail later on.
====== [7.6] From gas with interaction to Ising problem

Consider classical gas with interactions U (r~1 , ..., r~N ). The N particle partition function is
3N Z
1 1
ZN = d3N r e−βU (r1 ,...,rN ) (7.37)
N! λT
We see that the kinetic part factors out, hence the whole physics of the interactions is in the configuration integral.
Therefore, without loss of generality we can consider ”static gas”. To further simplify the treatment we consider a
”lattice gas” version:
X
H = U (r~1 , ..., r~N ) = u (x, x0 ) n(x)n (x0 ) (7.38)
hx,x0 i
We can represent graphically the interaction between two sites x and x0 by “bonds”. The notation hx, x0 i means
summation over all the bonds without double counting. In the simplest case there are interactions only between
near-neighbor sites. The grand partition function is
  
X X X
Z = exp −β  u (x, x0 ) n(x)n (x0 ) − µ n(x) (7.39)
n(·) hx,x0 i x
where n(x) = 0, 1. We define
1 + σ(x)
n(x) = , σ(x) = ±1 (7.40)
2
Then we get
  
X X X
Z= exp −β − ε (x, x0 ) σ(x)σ (x0 ) − h σ(x) + const (7.41)
σ(x) hx,x0 i x
where h = [µ − ū]/2. Here ū is the interaction energy per site for full occupation. Note that h = 0 implies that a fully
occupied lattice has the same energy as an empty lattice. We also have changed notation u(x, x0 ) = −4ε(x.x0 ), with
ε > 0 corresponding to attractive interaction.
52
We see that the calculation of Z for static lattice gas is formally the same as calculation of Z for an Ising model. The
following analogies should be kept in mind
occupation N ←→ magnetization M = 2N − N (7.42)

chemical potential µ ←→ magnetic field 2h (7.43)
fugacity z = eβµ ←→ define z = e2βh (7.44)
grand canonical Z(β, µ) ←→ canonical Z(β, h) (7.45)
From now on we refer to Ising model, but for the formulation of some theorems in the next section it is more convenient
to use the lattice gas language for heuristic reasons. Note also that N is more convenient than M because it does not
skip in steps of 2.
We shall clarify that in the thermodynamic limit (large N ) the magnetization density M/N , when plotted as a
function of h, might exhibits a jump at h = 0. This is called 1st order phase transition. Similarly, in the lattice gas
model, the density N/N , when plotted as a function of µ, might exhibits a jump at µ = ū. This can be regarded
a gas-to-liquid phase transition. In the canonical perspective we fix N and solve for µ. It is best to illustrate the
procedure graphically. Having a plot of N vs µ, we get a ”gas” or a ”liquid” state provided the horizontal N line does
not intersect the discontinuity. Otherwise there is a phase separation, where the gas and the liquid phases coexist
with the same chemical potential µ = ū.
M T<Tc
N
M vs h N vs µ µ
In the phase diagram (T, h) of the Ising model it is customary to indicate the discontinuity of the first order transition
by a solid line that stretches from (0, 0) to (Tc , 0). Similarly, in the phase diagram (T, µ) of the lattice gas the
discontinuity is indicated by a solid line that stretches from (0, ū) to (Tc , ū). However in practice it is customary to
use a (T, P ) phase diagram. Here we bring the phase diagrams for conventional gas-liquid-solid transition, for water,
for Helium-II and for Helium-III [taken from the web]:
53
====== [7.7] Yang and Lee theorems

Consider the lattice gas or the equivalent Ising model. We can use n(x) or σ(x) in order to specify whether a cell is filled
(spin ”up”) or empty (spin ”down”). The probability of a given configuration is determined by the grand-canonical
energy Eσ − µNσ , namely
pσ ∝ e−β(Eσ −µNσ ) (7.46)

P
Here Eσ is the Ising energy due to the interactions, and Nσ = x n(x) is the number of ”up” spins, which we call
below ”total magnetization”. In fact the total magnetization is 2Nσ − N , hence µ in the lattice-gas model is like 2h in
the Ising model. Either way we use the fugacity as the free variable and write the probability of a given configuration
as follows:
1 −βEσ Nσ
pσ = e z , z ≡ eβµ ≡ e2βh (7.47)
Z
Note that z = 1 means zero field. The partition function is
N
X
Z (z; β) = ZN (β)z N (7.48)
N =0
where ZN sums over all the configurations that have N spins up, and Z sums over all the possible values of the ”total
magnetization”. The Helmholtz function is
1
F (z; β) = − ln Z(z; β) (7.49)
β
The expectation value of the ”total magnetization” is
∂
hN i = −βz F (z; β) (7.50)
∂z
As we increase z we expect the magnetization hN i to grow, and hN i/N to reach a well defined value in the limit
N → ∞. Moreover, below some critical temperature we expect to find a phase transition. In the latter case we
expect hN i to have a jump at zero field (z = 1). The Yang and Lee theorems formulate these expectations in a
mathematically strict way. Given N it is clear that we can write the polynomial Z as a product over its roots:
N
Y
Z (z) = const × (z − zr ) (7.51)
r=1
Consequently
N
1X
F (z) = − ln(z − zr ) + const (7.52)
β r=1
N
X 1
hN i = z (7.53)
r=1
z − zr
There is a strict analogy here with the calculation of an electrostatic field in a 2D geometry. In the absence of
interactions (infinite temperature) we get that all the roots are at z = −1. Namely,
N
X
Z (z; β) = N N
CN z = (1 + z)N [non-interacting sites] (7.54)
N =0
54
So we do not have phase transition since the physical axis is 0 < z < 1, where this function is analytic. The questions
are what happens to the distribution of the roots as we increase the interaction (lower the temperature), and what
is the limiting distribution in the thermodynamics limit (N → ∞). There are three statements that give answers to
these questions due to Yang and Lee. The first statement is regarding the existence of the thermodynamics limit:
F (z)
lim = exists (7.55)
N →∞ N
The second statement relates specifically to the standard Ising model, saying that all the roots are lying on the circle
|zr | = 1. In general other distributions are possible. The third statement is that below the critical temperature the
density of roots at z = 1 becomes non-zero, and hence by Gauss law hN i/N has a jump at zero field. This jump is
discontinuous in the thermodynamic limit.
h<0 h>0
55
[8] The Ising model
====== [8.1] Model definition

The energy of a given Ising model configuration state is
X X
E[σ (·)] = − ε (x, x0 ) σ(x)σ (x0 ) − h(x)σ(x) (8.1)
hx,x0 i x
The canonical state is
1 −βE[σ(·)]
p[σ (·)] = e (8.2)
Z
where the partition function is

  
X X X
Z[h (·) , β] = exp β  ε (x, x0 ) σ(x)σ (x0 ) + h(x)σ(x) (8.3)
σ(·) hx,x0 i x
We expand the Helmholtz function as
1 X
G(x, x0 )h(x)h (x0 ) + O h4

F [h (·) , T ] = F0 (T ) − (8.4)
2T 0
x,x
In the absence of external field F (T ) = F0 (T ) and we have the usual relations
E (T ) = F (T ) + T S (T ) (8.5)
∂F (T )
S (T ) = − (8.6)
∂T
∂S dE
C (T ) = T = (8.7)
∂T dT
Next we assume weak field, leading to a linear response relation between hσ(x)i and h(x). Namely,
∂F 1X
hσ(x)i = − = G(x, x0 )h (x0 ) (8.8)
∂h(x)T 0
x
∂F
hσ(x)σ (x0 )i0 = G(x, x0 )

= −T (8.9)
∂h(x)∂h (x0 ) 0
For an homogeneous field we get hσ(x)i = χh, where
1X
χ = G(r), [fluctuation-response relation] (8.10)
T r
For the total magnetization we get
X
hM̃ i = hσ(x)i = N χh (8.11)
x
56
Optionally we could get the same result by replacing h(x) 7→ h, and using M̃ = −∂F /∂h. For the fluctuations of the
total magnetization we get
X X
hM̃ 2 i = hσ(x)σ(x0 )i = N G(r) (8.12)
x,x0 r
Form here we deduce that
1 hM̃ 2 i
χ = [fluctuation-response relation] (8.13)
T N
This is merely another version of the same ”fluctuation-response relation”.
====== [8.2] The spatial correlation function

It is possible to measure G(r) via a scattering experiment. Given a configuration σ(x) the intensity of the scattering
in the Born approximation is
Z 2
I (q) ∝ σ(x)e−i~q·~x d~x

(8.14)
If we average over configurations we get

Z
0
I (q) ∝ dxdx0 hσ(x)σ (x0 )i e−iq·(~x−~x ) ∝ G̃(q) (8.15)
Here G̃(q) is the FT of the correlation function G(r) = hσ(x)σ (x0 )i, where r = |x − x0 |.
We would like to discuss what happens to G(r) as the temperature is lowered. Specifically we would like to illuminate
what is the fingerprints of approaching a critical temperature of a phase transition, below which the system is
“ordered”. We note that all the discussion below can be repeated if we apply an infinitesimal field h = +0 and
approach the critical temperature from below. In the latter scenario the correlation function should be redefined by
subtracting the constant hσi2 .
We shall see in the next section that Landau’s approach in the Gaussian approximation leads to the Ornstein-Zernike
expression for the FT of the correlation function:

(1/ξ)
G̃(q) = (8.16)
q 2 + (1/ξ)2
This leads to
G(r) ∼ exp(−r/ξ) if ξ < ∞ (8.17)

G(r) ∼ 1/rd−2 for d>2 if ξ = ∞ (8.18)
Using the scaled variable r = r/ξ the exact FT can be expressed in terms of the modified Bessel function of the second
kind:
(d/2)−1
1 1
G(r) = K(d/2)−1 (r) (8.19)
(2π)d/2 r
In 1D it is G(r) = [1/2]e−r , and in 3D it is G(r) = [1/(4πr)]e−r .

57
The information about order-disorder transition is in G(r). If ξ < ∞ there is no long range order, and we get χ < ∞.
As ξ → ∞ the susceptibility diverges, which implies a phase transition. Note that for finite ξ the total magnetization
M̃ can be regarded as a sum of random variables, its variance scales like N , and consequently χ comes out finite, as
implied by the ”fluctuation-response relation”. At the critical temperature the fluctuations are strongly correlated
over arbitrarily large distances, and χ diverges.
====== [8.3] Critical behavior and the scaling hypothesis

Below we display the phases diagram in (T, h) space, and qualitative plots of the state equations. For the 2D Ising
model with near neighbor interactions Tc ≈ 2.27.
h
T
Tc
E(T) M~ χ h M(h)
T
χ ∼ t −γ h
T > Tc
M(h)
C(T) M~ h 1/δ
C ~ t −α h
T = Tc
M(T) h=+0 M(h)

M~ t β h
T < Tc
T
The state equations in the critical region are characterized by the exponents α, β.γ, δ (see below). Two other exponents
ν and η are defined via the critical behavior of the correlation function, which is assumed to be a variation on the
Ornstein-Zernike expression. Namely, one conjectures that the divergence of the correlation length as T → Tc is
described by
ξ ∼ |T − Tc |−ν (8.20)
and that the correlation function is
1
G(r) ∼ exp(−r/ξ) (8.21)
rd−2+η
Here we combined the T = Tc and the T > Tc into one expression. This expression describes the long range behavior.
Note that the ”microscopic” short range behavior is not too interesting because it is bounded by G(0) = 1. The
divergence of χ is due to the slow power-law tails. Below Tc the behavior is similar to T > Tc provided the correlation
function is properly defined. Going on with the same phenomenology the conjecture is that away from Tc the
correlation distance ξ is the only relevant length scale in the thermodynamic limit. This means that each ”correlated
block” of the system has the same contribution to the Free energy irrespective of the temperature, hence
V
F0 (T ) ∼ ∝ |T − Tc |νd (8.22)
ξd
where d is the dimensionality. It is customary to define the scaled temperature as t = |T − Tc |/Tc , and to summarize
58
the above conjectures as a scaling hypothesis that involves the critical exponents:
G(sr, s−1/ν t) = s−(d−2+η) G(r, t) (8.23)

−1/ν −d
F0 (s t) = s F0 (t) (8.24)
From here it follows that
χ(s−1/ν t) = s2−η χ(t) (8.25)

−1/ν (2/ν)−d
C(s t) = s C(t) (8.26)
From the combined scaling relation

F s−1/ν t, s−(d+2−η)/2 h = s−d F (t, h) (8.27)
we can deduce similar relations for the magnetization. These scaling relations allow to deduce the critical exponents
α, β.γ, δ from d, ν, η, leading to
C ∼ |t|−α , α = 2 − νd (8.28)
M ∼ |t|β , β = (d − 2 + η)ν/2 (8.29)
χ ∼ t−γ , γ = (2 − η)ν (8.30)
M ∼ |h|1/δ , δ = (d + 2 + η)/(d − 2 + η) (8.31)
The so called “classical” mean-field exponents that we derive later are
ν = 1/2, η = 0, α = 0, β = 1/2, γ = 1, δ=3 (8.32)
In order to get a non-trivial result for α we have to take into account Gaussian fluctuations around the mean field
leading to α = [2 − (d/2)], in consistency with the scaling relations. However, one observes that the classical mean-
field exponents satisfy the other scaling relations with d = 4, and not with d = 3. This implies that we have to go
beyond mean field theory in order to establish the experimentally observed scaling behavior.
====== [8.4] Digression regarding scaling

A function of one variable has a scaling property if
F (sx) = sDF F (x) (8.33)
where DF is the scaling exponent. It follows that F (x) = const xDF . For example F (x) = x2 has the scaling exponent
DF = 2. If we have say two variables then the more general definition is
F (sDx x, sDy y) = sDF F (x, y) (8.34)
Note that the scaling exponents can be multiplied by the same number, and still we have the same scaling relation.
It follows that there is a scaling function such that

x
F (x, y) = y DF /Dy f (8.35)
y Dx /Dy
For example F (x, y) = x2 + y 3 has the scaling exponents Dx = 1/2, Dy = 1/3, DF = 1. More generally any “physical”
function has an “engineering” scaling property that follows trivially from dimensional analysis.
59
====== [8.5] Solution of the 1D Ising Model

Assuming only near neighbor interactions
X X
E [σ] = −ε σi σj − hi σi (8.36)
hiji i
The partition function is
X
Z[h, β] = e−βE[σ] (8.37)
σ(·)
For ε = 0 we get
N
Y
Z[h, β] = 2 cosh (βhi ) (8.38)
i=1
and hence
N N
X hi 1 X 2
F [h, T ] = −T ln 2 cosh ≈ −N T ln (2) − h (8.39)
i=1
T 2T i=1 i
The correlation function is
∂F
G(r) = −T = δij = δr,0 (8.40)
∂hi ∂hj
and hence the susceptibility is
1X 1
χ = G(r) = (8.41)
T r T
The magnetization is

∂F h
≈ N χh + O h3

M̃ = − = N tanh (8.42)
∂h T
We turn now to the case ε 6= 0. Without anPexternal field the calculation is very easy. We can define shiji = σi σj .
Then the interaction can be written as −ε b sb . Instead of summing over spins, we can sum over the bonds sb .
Assuming a chain of N spins the sum factorizes and we get Z = 2[2 cosh(βε)]N −1 . Next we would like to assume
that there is non zero homogeneous field h. The calculation becomes somewhat more complicated, and requires the
so called “transfer matrix” method. Let us define the matrix
!
eε̃+h̃ e−ε̃

0 00 1 0 00
Tσ0 σ00 ≡ exp ε̃σ σ + h̃ (σ + σ ) = , ε̃ ≡ βε, h̃ ≡ βh (8.43)
2 e−ε̃ eε̃−h̃
The eigenvalues of this matrix are
r
−ε̃
λ± ε̃
= e cosh h̃ ± e 1 + e4ε̃ sinh2 h̃ (8.44)
60
The partition function of N site Ising model on a ring can be calculated as

X
= trace T N = λN N

Z (β, h) = Tσ0 ,σ1 Tσ1 ,σ2 ...TσN −1 ,σ0 + + λ− (8.45)
σ(·)
and hence for very large N we get
F (T, h) = −N T ln (λ+ ) (8.46)
Expanding we get
1 exp 2 Tε 2
ε
F (T, h) ≈ −N T ln 2 cosh − N h (8.47)
T 2 T
Hence
1 ε
χ= exp 2 (8.48)
T T
Now we would like to calculate the correlation function at zero field.
1 X
G(r) ≡ hσ0 σr i = σ0 Tσr0 σr σr TσNr σ−r (8.49)
Zσσ 0
0 r
We have
! ! !
√1 √1 λ+ 0 √1 √1
Tσ0 σ00 = 2 2 2 2 (8.50)
√1 − √12 0 λ− √1 − √12
2 2
with
λ+ = 2 cosh (ε̃) (8.51)

λ− = 2 sinh (ε̃) (8.52)
Using standard Pauli matrix notations and denoting the digonalized matrix T as Λ we get
1 1 λr λN −r + λr− λN −r
trace σz T r σz T N −r = trace σx Λr σx ΛN −r = + − N +

G(r) = (8.53)
Z Z λ+ + λN −
For very large N we get

r
λ−
G(r) = = e−r/ξ (8.54)
λ+
where
h ε i−1 1 2ε/T
ξ = ln coth ≈ e (8.55)
T 2
P
The calculation of G(r) involves a geometric summation, and it can be verified that it agree with the the result
for χ. The same result as the exact one is obtained from the approximated exponential expression if the summation
is replaced by an integral.
61
====== [8.6] Solution of the 2D Ising model

The full details of the Onsager solution for this problem is in Huang. Also here the transfer matrix approach is used.
Recall that the zero field solution of the 1D model is
1
ln Z = ln (2) + ln (cosh (ε̃)) (8.56)
N
The 2D solution is
dθdθ0
Z Z
1 1 h
2
i
ln Z = ln (2) + 2 ln (cosh (2ε̃)) + sinh (2ε̃) (cos θ + cos θ0 ) (8.57)
N 2 (2π)
dθdθ0
Z Z
1 h κ 0
i
= ln (2) + ln (cosh (ε̃)) + 2 ln 1 + (cos θ + cos θ ) (8.58)
2 (2π) 2
The integral is determined by the dimensionless parameter
2 sinh (2ε̃)
κ ≡ 2 ≤ 1 (8.59)
(cosh (2ε̃))
The value κ = 1, for which ln Z exhibits discontinuity in its derivative, is attained for sinh(2ε) ˜ = 1, from which it
follows that the critical value of the interaction is ε̃ = 0.44, leading to Tc = 2.27ε. This is almost half compared with
the heuristic “mean field” value Tc ≈ 4ε that will be derived in the next lecture.
62
[9] Phase transitions - heuristic approach
====== [9.1] The ferromagnetic phase transition

The standard Ising Hamiltonian is
X X
H = −ε σi σj − h σi (9.1)
hiji i
Let us assume that in equilibrium we can regard the spins as quasi-independent, each experiencing an effective field h̄,
such that the effective Hamiltonian for the spin at site i is H(i) = −h̄σi . This means that the equilibrium state is
" #
X
pσ1 ...σN ∝ exp β h̄ σi (9.2)
i
We have to find what is h̄. The formal way is to use a variational scheme. We shall introduce this procedure later.
In this section we guess the result using a self-consistent picture. By inspection of the Hamiltonian if the mean
magnetization of each spin is hσi, then it is reasonable to postulate that
X
h̄ = h + ε hσj i = h + cεhσi (9.3)
neighbors
where c is the coordination number. Form H(i) we get the self-consistent requirement

1
hσi = tanh (h + cεhσi) (9.4)
T
This equation should be solved for hσi, and then we get h̄ as well.
<σ>
zero h
nonzero
<σ>
By inspection of the plot we observe that for h = 0 the condition for getting a non trivial solution is cε/T > 1.
Therefore Tc = cε. If we want to explore the behavior in the critical region it is convenient to re-write the equation
in the following way:
h = T tanh−1 hσi − Tc hσi (9.5)
and to approximate it as
1
h = (T − Tc ) hσi + Tc hσi3 (9.6)
3
63
For T > Tc we get the Curie-Weiss law, which implies a critical exponent γ = 1. Namely,
1
hσi = h (9.7)
T − Tc
For T = Tc the dependence of h is characterized by the critical exponent δ = 3. Namely,

13
3
hσi = h (9.8)
Tc
For zero field (h = +0), below Tc , the temperature dependence is characterized by β = 1/2. Namely,
21
Tc − T
hσi = 3 (9.9)
T
In the mean field approximation the spins are independent of each other, and therefore hσi σj i = hσi ihσj i. It follow
that the energy is
1
E = hHi = − cN ε hσi2 (9.10)
2
For the heat capacity we get

dE ∂hσi
C(T ) = = −cN hσi (9.11)
dT h=0 ∂T h=0
For T > Tc we get C(T ) = 0, and from below we approach a constant value. The implied critical exponent is α = 0.
To get the non-trivial mean-field result α = [2 − (d/2)] we have to take into account Gaussian fluctuations.
C (Τ)
====== [9.2] The anti-ferromagnetic phase transition

Let us consider a less trivial example for the use of the heuristic approach. An anti-ferromagnet is described by the
Ising Hamiltonian with ε 7→ −ε. Specifically we consider a 2D square lattice that consists of to sub-lattices (for other
lattices we might have frustrations). We mark the magnetization of the two sub lattices by Ma and Mb . We define
1
M = (Ma + Mb ) (9.12)
2
1
Ms = (Ma − Mb ) (9.13)
2
Without the magnetic field, the problem is the same as the ferromagnetic one with Ms as the order parameter. With
magnetic field h the heuristic mean field equations become

1 1
Ma = tanh (h − Tc Mb ) Mb = tanh (h − Tc Ma ) (9.14)
T T
64
Following the same algebraic simplification procedure as in the ferromagnetic case, we get after addition and subtrac-
tion of the two resulting equations,
1
(T − Tc ) Ms + Tc 3M 2 Ms + Ms3 = 0

(9.15)
3
1
(T + Tc ) M + Tc 3Ms2 M + M 3 = h

(9.16)
3
From here it follows that (see graphical illustration):

2 Tc − T
Ms = 0 or 3M + Ms2 = 3 (9.17)
T
1 h
(2 + Ms2 )M + M 3 = (9.18)
3 Tc
As expected from the second equation we get M = 0 in the absence of an external field, and from the first equation
we get the order parameter Ms (T ), which satisfies the same equation as in the ferromagnetic problem. If we switch
on the magnetic field Tc is shifted to a lower temperature.
Μs h critical line
T<Tc (2nd order)
Μ T
If the magnetic field h is strong enough, it destroys the anti-ferromagnetic order and causes Ms = 0. This is implied
by the identification of the ground state:

1
E (↑↓↑↓) = N · − c , [for weak field] (9.19)
2

1
E (↑↑↑↑) = N· c − h , [for strong field] (9.20)
2
In the region where T ∼ Tc and h ∼ 0 we get for the magnetization

 
1
M =   h ≡ χh (9.21)
2
Tc + T 1 + 16 Ms (T )
We can get a better general expression for all of the temperature range by differentiation of the heuristic equations
1
χ= 2 Tc
(9.22)
Tc + T cosh T Ms (T )
In the region T ∼ Tc substitution of Ms (T ) gives
(
1
Tc +T Tc < T
χ= 1
(9.23)
4Tc −2T T < Tc
65
====== [9.3] Beyond the Ising model

We now make a slight generalization of the Ising model. We consider coupled non-inertial oscillators, meaning that
the kinetic term in the Hamiltonian is neglected:
X X
H = [U (sj ) − hsj ] − ε si sj (9.24)
j hiji
For the Ising model U (s) = 0 for s = ±1 and U (s) = ∞ otherwise. But more generally we assume, say,
U (s) = (α/2)s2 + (u/4)s4 . In the absence of interaction
M ≡ hsi = TANH(χh) (9.25)
where TANH is a function that has by definition slope unity at the origin (TANH0 (0) = 1), and χ is a constant that
has the meaning of zero-field susceptibility. For zero non-linearity (u = 0) we get TANH(x) = x with slope χ = 1/α.
Otherwise the TANH becomes a concave function with temperature dependent χ(T ). For the Ising model χ(T ) = 1/T .
The heuristic mean field equation is
M = TANH [χ (h + cM )] (9.26)
In the absence of an external field it possesses a non-trivial solution provided cχ(T ) > 1, leading to a finite Tc . But
if we consider dynamical degrees of freedom (see discussion of coupled rotors below), the susceptibility might be finite
also at zero temperature due to quantum fluctuations. Thus, if cχ(0) < 1, phase transition does not take place.
Then, by tuning the model parameters at T = 0, we can witness a quantum phase transition once we cross to a regime
where cχ(0) > 1.
Note: Considering again zero non-linearity (u = 0), the heuristic approach implies that the system becomes unstable
for α < c. This condition becomes more illuminating if we wrote the interaction between to oscillators as (/2)[si −sj ]2 .
The price for that is to write U (s) = (a/2)s2 instead of U (s) = (α/2)s2 with a = α − c. The condition for instability
becomes simply a < 0.
====== [9.4] The mean-field Hamiltonian

The heuristic approach for phase-transition can be regarded as the formal outcome of a mean-field approximation for
the system Hamiltonian. We set in the Hamiltonian sj = M + δsj , expand the interaction term, and get
X X
H ≈ H (j) − ε δsi δsj (9.27)
j hiji
were the first sum is the mean field Hamiltonian with
1
H (j) = U (sj ) − (h + cεM )sj + cεM 2 (9.28)
2
Assuming that the fluctuation are uncorrelated we deduce

1
E = hHi ≈ N hU (s)i − cεM 2 − hM (9.29)
2
where hU (s)i should be calculated from the mean field Hamiltonian. Note that it is zero for the Ising model, while here
we consider a more general class of systems. If the non-linear term and the fluctuation are neglected hU (s)i = (α/2)M 2 ,
hence the symmetry breaking is implied if a = (α − c) < 0. Taking non-linearity into account, symmetry breaking is
implied if the temperatures is low enough.
66
====== [9.5] Coupled rotors

We now consider coupled rotors. The rotors are dynamical entities, they have finite mass. We define γ as the inverse
moment of inertia. Note that if we started with [ϕ, p] = i~, then with n = p/~ we get that γ ∝ ~2 . Accordingly infinite
mass is like taking the classical limit. The Hamiltonian is
X hγ i X
H = n2j − h cos(ϕj ) − ε cos(ϕj − ϕi ) (9.30)
j
2
hiji
If we ignore the kinetic term it is formally like coupled non-inertial oscillators with sj = cos(ϕj ). In a classical context
if we take the kinetic term into account it has no effect because it factorizes out of the partition function. The mean
field Hamiltonian is
γ 2 1
H (j) = n − (h + cεM ) cos(ϕj ) + cεM 2 (9.31)
2 j 2
In the quantum treatment the energy shift of the ground-sate is not −(1/2)cεM 2 because of quantum fluctuations:
the price of small ϕ is large uncertainty in the conjugate momentum n. The implications is that quantum fluctuations
are able to diminish M at zero-temperature.
Quantum phase transition.– Let us find the condition for diminished “order” at zero-temperature. The simplest
perspective is the heuristic approach. At zero temperature standard quantum-mechanical calculation using second
order perturbation theory shows that the zero temperature susceptibility of a rotor is χ = 2/γ. It follows that
symmetry-breaking is avoided if
2cε < γ [Mott phase] (9.32)
We see that zero-temperature “order” is diminished either by having T or γ that are larger than ∼ ε, reflecting
strong quantum or thermal fluctuations respectively. An equivalent way to deduce the above condition is to consider
the ground state √energy of E0 (h) of H0 = (γ/2)n2 − h cos(ϕ). For large h using harmonic-oscillator approximation
E0 = −h + (1/2) γh, but for small h using 2nd order perturbation E0 = −h2 /γ. Using the latter result we get at
the vicinity of M = 0 that the mean-field energy per rotor is
D E a (cε)2
E = H (j) = M 2 , with a = −2 + cε (9.33)
2 γ
Symmetry-breaking is avoided if a > 0.
====== [9.6] The variational approach

A different way to derive the heuristic mean-field equations is to use the variational approach. The canonical state
minimizes the free energy functional. Accordingly we look for a solution to the variation problem
F [ρ] ≡ hHi − T S [ρ] = minimum (9.34)
with implicit constraint on the normalization. In the mean-field approach the canonical state is assumed to be well
approximated by ρ = {pσ }, where
" #
1 X
pσ1 ...σN = N exp −β h̄ σk (9.35)
2 cosh β h̄ k
Here the variational parameter h̄ is the effective mean field. We would like to determine the optimal valuePof h̄
for which F [ρ] is minimal. For the calculation we use the identity F [ρ] = F0 [ρ] + hH − H0 i, where H0 = −h̄ i σi ,
67
leading to

1
F [ρ] = N f (h̄) − c m(h̄)2 − h − h̄ m(h̄)

(9.36)
2
and m(h̄) = −f 0 (h̄) is the mean-field magnetization. The variational equation for

where f (h̄) = −T ln 2 cosh h̄/T
h̄ is as expected

h̄
h̄ = h + c tanh (9.37)
T
Hence, we get the variational free energy

"
2 #
h̄ 1 h̄
F (T, h) = N −T ln 2 cosh + c tanh (9.38)
T 2 T
This is not a pleasant expression because the dependence on h is implicit in h̄. We can differentiate this equation to
find M̃ , which involves ∂ h̄/∂h. The calculation is lengthy, but we can skip it because the result is obvious

∂F (T, h) h̄
M̃ = − = N tanh (9.39)
∂h T
To make calculations of the state equations more convenient we notice that F (T, h) depends in a very simple way
on M̃ , hence it is useful to make the Legendre transformation
A(T, M̃ ) ≡ F (T, h) + hM̃ (9.40)
such that dA = −SdT + hdM̃ . Note that the mean field equation for h̄ implies that
M̃
h = T tanh−1 (M ) − cM where M ≡ = hσi (9.41)
N
Using the identity tanh−1 (x) = (1/2) ln((1 + x)/(1 − x)) one obtains

1 2
1 1+M 1 2
A (T, M ) = N −T ln 2 + T ln 1 − M + T M ln − cM (9.42)
2 2 1−M 2

1+M 1+M 1−M 1−M 1
= NT ln + ln − N cM 2 (9.43)
2 2 2 2 2
From this expression it is convenient to derive explicit results for the state equations. In particular S = −∂A/∂T and
one can recover the result for the heat capacity.
68
====== [9.7] The Bragg Williams formulation

Consider an Ising model with N sites, at any dimension, and with any coordination number. Given a spin configuration
define
N = total number of spins (9.44)

m = total magnetization (9.45)
M = m/N (9.46)
1 1
N+ = number of up spins = (N + m) = N (1 + M ) (9.47)
2 2
1 1
N− = number of down spins = (N − m) = N (1 − M ) (9.48)
2 2
N+− = number of bonds connecting spins with opposite direction (9.49)
The total number of bonds is (1/2)cN , where c is the coordination number. It follows that
X
σi = m (9.50)
X 1
σi σj = cN − 2N+− (9.51)
2
hiji
If we look on two connected spins, there is a probability (N+ /N ) to have the first up, and a probability (N− /N ) to
have the the second down. Or we can have the first down and the second up. This motivates the Bragg Williams
approximation:

N+ N− Nc Nc
N+− ≈ 2 = (1 − M 2 ) (9.52)
N N 2 4
Assuming that it holds for typical configurations we approximate the energy functional as

1
E[σ] ≈ −N × cεM 2 + hM (9.53)
2
We note that this expression with c = N if formally exact for a fully connected cluster of spins. The number of
configuration with total magnetization m is

N! 1 2 1
gm = ≈ const exp −N M + M 4 + ... (9.54)
(N+ )! (N− )! 2 12
In order to derive the latter approximation note that
∂ 1 1 1
− ln gm = (ln(1 + M ) − ln(1 − M )) ≈ M + M 3 + M 5 + ... (9.55)
∂m 2 3 5
With this approximation we get

XX X X
Z = e−βE[σ] ≈ gm e−βE(m) = e−A(M ) (9.56)
m σ∈m m M

1 1
A(M ) = N × (1 − βcε) M 2 + M 4 − βhM (9.57)
2 12
In the next section we are going to clarify the following points: (1) The sum can be evaluated via Gaussian integration.
(2) This Gaussian approximation can be justified if N is large. (3) Phase transition is implied. (4) But nevertheless
69
the result is false for d = 1. In the next lecture we shall further explain that the Bragg-Williams formulation fails in
providing the correct description of the symmetry-breaking if the critical temperature is approached.
====== [9.8] The Gaussian approximation

The expression that we have obtained for Z using the Bragg Williams formulation is a typical approximation that
can obtain for various models. We rewrite it as follows:
Z ha u 4 i
Z = dϕ e−A(ϕ) A(ϕ) = N × ϕ2 + ϕ − hϕ (9.58)
2 4
This sum can be evaluated via Gaussian integration. The dominant contribution comes from the ϕ for which A(ϕ) is
minimal. One can easily verify that A0 (ϕ) = 0 coincides with the heuristic mean field equation that has been discussed
in a previous lecture. Non trivial solutions appear for a < 0 which implies Tc = cε. Note that a ≈ (T − Tc )/Tc .
Above the critical temperature there is a single minimum at ϕ̄ = (1/a)h and one obtains
1/2
2π N 2
Z ≈ exp h (9.59)
Na 2a
In the absence of an external field, as a becomes negative, the trivial minimum ϕ̄ = 0 bifurcates into two minima,
namely ϕ̄ = ±(|a|/u)1/2 . For these values
1/2
a2

|a|
A(ϕ̄) = − ∓ h (9.60)
4u u
Approximating Z as the sum of two Gaussian integrals, one realizes, after expanding A(ϕ̄ + ϕ̃), that the coefficient
of the ϕ̃2 term is the same as above Tc , with a replaced by 2|a| (positive). We get that the partition function is like
that of a spin:
1/2 " #
1/2
a2

π N 2 |a|
Z ≈ exp N + h 2 cosh N h (9.61)
N |a| 4u 4|a| u
From here one deduces that for T < Tc the susceptibility becomes χ = [1/(2|a|)] + N |a|/u instead of χ = 1/a.
At this point one can ask whether it was allowed to ignore the quartic term in A(ϕ). This should be checked self
consistently. For a > 0 the dispersion of ϕ in the Gaussian approximation is (N a)−1/2 . The quartic term can be
neglected if u|ϕ|4 a|ϕ|2 leading to the condition a (u/N )1/2 . This condition is always satisfied if N is large
enough. The same condition also guarantees that for a < 0 the dispersion is much smaller compared with the non-zero
mean field ϕ̄. Generalization of this condition in the field-theory treatment will be discussed later and lead to the
Ginzburg criterion.
It is now appropriate to point out that the above treatment implies a phase-transition in the thermodynamic limit.
We first note that the existence of the thermodynamic limit for A(ϕ)/N could have been anticipated from general
considerations. From Z we can get the free energy F (h)/N that will have a thermodynamic limit too. The question
is whether the subsequent limits h → +0 and h → −0 lead to the same magnetization, or optionally whether the
susceptibility χ beomes infinite below Tc . Indeed this is what we found.
70
====== [9.9] The importance of fluctuations

The above analysis is misleading. The Bragg Williams approximation underestimates the effect of fluctuations. We
already know from the exact solution of the Ising model that in 1D there is no phase transition at finite temperature.
We would like to explain in detail why the fluctuations in 1D smear away the phase transition. Later we shall see
that also for d = 2, 3 the fluctuations are important: they do not smear away the phase transition, but they modify
the state equations in the critical region, which explains the failure of mean field theory there, and the observed
anomalous values of the scaling exponents.
T=0
T>Tc T<Tc
The problem with the Bragg Williams approximation is implied by the figure above. The action A(ϕ) is plotted. It
is determined by the energy term E[ϕ], and by the entropy term S[ϕ]. Recall that

1 −A(ϕ) 1
p(ϕ) = e ∝ exp − E[ϕ] + S[ϕ] (9.62)
Z T
In the T = 0 panel the energy of the states is indicted by bars. By taking the ”typical” value of the energy (thick
solid line) we ignore a dominant fraction of less typical states that have a very low energy. These states corresponds
to configurations where the spins a bunched in ”zones”. The simplest arrangement in 1D has two zones and its energy
is Eb = E0 + 2, where E0 = −N is the ground state energy. Consequently the effective barrier between the ”all up”
and ”all down” states is very low (thick dashed line), and the symmetry breaking is avoided at any finite temperature.
The formal argument is outlined below. In contrast to that, in 2D the effective barrier is Eb = E0 + N 1/2 and
therefore symmetry breaking is realized in the thermodynamic limit.
Domain walls.– It is possible to argue that in 1D there is no phase transition at finite temperature. The argument
goes as follows: consider a mixture of states with one domain wall. For such state E[ρ] = E0 +2, where is the cost of
the domain wall. But the entropic contribution is S[ρ] = ln[N ] where N is the length of the chain. It follows that for
any finite T the ground state, or any (exclusive) mixture of ground-like states, do not minimize F [ρ] at the thermody-
namic limit. We can lower E[ρ] by adding states that have with equal probability any magnetization. Consequently
we get huge fluctuations whose relative amplitude does not diminish, in contrast with the (N )−1/2 prediction of the
mean-field Gaussian estimate. Therefore spontaneous magnetization at finite temperature is impossible.
The above argument fails in 2D because the energy cost of a domain E[ρ] = E0 + N 1/2 domeiniate over the entropic
contribution. In fact it is possible to refine the domain wall approach and deduce that for the d = 2 Ising model there
exists spontaneous magnetization at finite temperature [see Huang p.349], in consistency with the exact solution.
However, the possibility to witness phase transition at d = 2 is limited to systems with discrete symmetries. The
Mermin-Wagner theorem states that a continuous symmetry cannot be spontaneously broken at finite temperature
in d ≤ 2 systems with short-range interactions. See discussion of the Heisenberg model.
71
[10] Phase transitions - field theory
====== [10.1] The Landau model

We would like to take into account the spatial fluctuations of the magnetization ϕ(x) in the calculation of the partition
function. We therefore use a refined version of the Bragg-Williams approach. Namely, first we sum over all microscopic
configuration that corresponds to a magnetization ϕ(x), and then we are left with a so-called functional integral:
X Z
Z[h, parameters; Λ, L] = e−A[ϕ(·)] = Dϕ e−A[ϕ(·)] (10.1)
ϕ(·)
where the sum over configurations becomes an integral with the measure
Z
Y Y 1
Dϕ = dϕx = dϕ̃k , ϕ̃k ≡ d ϕ(x) e−ikx dx (10.2)
L
x∈Ld |k|<Λ
In this definition ϕ̃k are the Fourier components of ϕ(x), and we are sloppy about an uninteresting overall prefactor.
It is implicit that one assumes a finite volume Ld , and a finite momentum cutoff Λ, otherwise the functional-integral
that gives Z is ill defined. To have a momentum cutoff is like to assume that space is discretized with lattice spacing
2π/Λ. Accordingly the number of freedoms of the model is N = (LΛ/2π)d . Technical remark: for presentation
purpose it is more convenient to work with the complex ”exp” Fourier basis, but from mathematical point of view
the meaning of the Dϕ integration is more obvious if we work with real ”sin” and ”cos” Fourier basis, corresponding
to the real and imaginary parts of ϕ̃k . Either way the ϕ field is represented by N independent real amplitudes.
In the Landau model the assumed action is

Z c
2 a u X 1
A[ϕ(·)] = dx (∇ϕ) + ϕ2 + ϕ4 − hϕ = Ld (ck 2 + a)|ϕ̃(k)|2 + ... (10.3)
2 2 4 2
k
The summation over the k components of the field is conventionally written as an integral with the measure
[L/(2π)]d dk. The convention c = 1 with regard to the prefactor of the first term fixes the dimensions of ϕ, and
hence of all the other model parameters. We write these dimensions as Ld , accordingly
d−2 d+2
dϕ = − , dh = − , da = −2, du = −(4 − d), (10.4)
2 2
The model has a thermodynamic limit, hence L is not significant, and we can calculate the Helmholtz free energy F
per unit volume. In contrast to that Λ is significant. In particular we note that the model contains two significant
dimensionless parameters that are related to the underlying microscopic Hamiltonian:
ã = a/Λ2 Note: later we see that ξ = a−1/2 is the correlation length (10.5)
ũ = u/Λ4−d Note: later we see that a/u2/(4−d) is the Ginzburg parameter (10.6)
(0) (0)
Relating to the Bragg-Williams approximation we identify a ∝ (T − Tc ), where Tc is the mean field critical
temperature. In fact we shall see that the field theory analysis implies that for 2 ≤ d < 4 the actual critical temperature
is pushed down due to the fluctuations (ac < 0). In the d = 1 case there is no phase transition.
Coarse graining.– The Landau model can be regarded as the outcome of coarse graining on scale Λ. Therefore its
parameters a(Λ) and u(Λ) are “running coupling constants”. The cutoff Λ is in the range [Λ0 , Λ∞ ], where Λ∞ reflects
a limiting microscopic scale, while Λ0 reflects the maximal spatial range over which coarse-graining is meaningful.
Clearly it is the correlation distance, hence Λ0 ∼ 1/ξ. We shall see that as Λ is decreased, as the result of successive
course-graining operations, we get ũ(Λ) → ũc , where ũc = 1/9 for 3D. So in some sense there is only one relevant
parameter (a) in this model, and results for different values of a are related by scaling.
72
====== [10.2] Related models

The Landau model stands by itself as a prototype model. It might have different physical interpretations. Possibly
the simplest is to regard it as the continuum model of ”coupled oscillators”. If a = u = 0 it is formally like the Debye
model. Having a > 0 means that the oscillators have a positive spring constant that stabilizes them at ϕ = 0. For
a < 0 each oscillator is pushed away from ϕ = 0 and its new equilibrium position is determined by the nonliterary u
of the spring. Thanks to the non-linearity of the springs the ϕ cannot diverge to infinity.
Above we have regarded the Landau model as a coarse grained version of the Ising model, using the Bragg-Williams
approximation for each coarse-grained cell. There is an optional possibility to motivate the Landau model as an
approximation for the Ising model using a somewhat more direct procedure. For this purpose one replaces the
discrete summation over σ = ±1 by an integration over ϕ with a weight function:
Z Y
2 2
dϕx e− 4 u(ϕx −1)
X 1
→ (10.7)
σ(·) x
One should realize that the ferromagnetic interaction −σ(x)σ(x0 ) corresponds to differences (ϕ(x) − ϕ(x0 ))2 , and
hence translates to the gradient term in the Landau model.
The field theory that corresponds to the Ising model contains a real field. It reflects the discrete mirror (up/down)
symmetry of the system. More generally we can consider a system that has a continuous rotational symmetry. In
such case the Action is A[S] with vector field S =
√ (S1 , S2 , S3 ). Of particular interest is to have a gauge-invariant
Action A[Ψ] with a complex field Ψ = (ψ1 , ψ2 ) ≡ n exp(iϕ). Note that in two dimensions gauge-invariance can be
regarded as a rotational-invariance.
====== [10.3] The Gaussian approximation

Let us start with the simplest possibility of having u = 0. Regarded as an approximation it is meaningful only if
a > 0. If we had c = 0 the result would be the same as that of the Bragg-Williams model. If c is non-zero (equal to
unity by convention) the summation still factorizes, but in k space. Assuming for simplicity h = 0, and not caring
about a global prefactor we get the following:
YZ
Ld 2
Y 1 1/2
k + a ϕ̃2k

Z = dϕ̃k exp − = (10.8)
2 k2 + a
k k
The free energy in the mean-field approximation was F = T A(0) = 0. Now we have taken the Gaussian fluctuations
into account. Consequently we get a non-trivial result for the free energy:
T X
ln k 2 + a

F (T ) = (10.9)
2
k
In particular we can derive from this expression the Gaussian prediction for the heat capacity. Contrary to the mean
field approximation, it is no longer zero. The singular contribution at the vicinity of Tc originates from the second
derivative with respect to a. Accordingly
Λ
d2 F k d−1 dk
Z
C(T ) = −T ∼ ∼ |T − Tc |−(4−d)/2 , for d < 4 (10.10)
dT 2 0 (k 2 + a)2
Hence the modified mean field exponent α = 0 is replaced by α = [2 − (d/2)].
Below Tc we can perform a Gaussian approximation around the mean-field ϕ̄ that will be discussed in the next section.
The calculation is essentially the same, with an offset A[ϕ̄] that is added to the action, hence F (T ) 7→ F (T ) + T A[ϕ̄].
The non-singular contribution of this additional ”mean field” term implies a discontinuity of C(T ) at T = Tc as
discussed in past lecture, and has no effect on the Gaussian value of α.
73
====== [10.4] Digression - Gaussian integrals

The partition function can be calculated exactly whenever the action is a quadratic form. The so-called Gaussian
integral reduces to the product of N one dimensional integrals if we transform it to a basis in which the quadratic
form is diagonal. For a system that has translation symmetry it is momentum space.
Z YZ
1
P P 1 2
Dϕ e− 2 i,j Aij ϕi ϕj + i hi ϕi
= dϕ̃k e− 2 ak ϕ̃k +h̃k ϕi (10.11)
k
 
Y 2π 1/2 1

1

h2k p 1X
= e 2 ak
= det(2πG) exp  Gij hi hj  (10.12)
ak 2 i,j
k
here G = A−1 , and note that det(G) = 1/ det(A). Note also that going back to the original basis, in the case of
position-to-momentum transformation implies that G(r) is the Fourier transform of 1/a(k).
From the above result, it follows that G(r) is the correlation function hϕ(r)ϕ(0)i for h = 0. Otherwise hϕi is non-zero,
and it equals to field ϕ̄ that minimizes that action. It satisfies the equation Aϕ = h, whose solution is ϕ̄ = Gh. Hence
G can be regarded as the ”Green function”.
====== [10.5] The mean field equation

We define the mean field ϕ̄ via the equation A (ϕ) = minimum. This gives the equation
−∇2 + a ϕ + uϕ3 = h(x)

(10.13)
The mean field for an homogeneous h(x) = h is obtained from aϕ + uϕ3 = h. In particular for h = ±0 we get
( (
0, for a > 0 0, for a > 0
ϕ̄0 = −a
12 A[ϕ̄0 ] = a2
(10.14)
± u , for a < 0 − 4u , for a < 0
For a > 0 we neglect

the non linear term in the action, define ξ = a−1/2 , and write the mean field equation as
2 2
−∇ + (1/ξ) ϕ(x) = h(x). For a < 0 we make the substitution ϕ 7→ ϕ̄0 + ϕ, expand the action around the new
minimum, and then neglect the non-linear term. The mean-field equations takes the same form as for a > 0, with
ξ = (−2a)1/2 . Accordingly the first-order solution in h(x) is
Z
ϕ̄(x) = ϕ̄0 + G(x − x0 )h (x0 ) dx0 + O(h2 ) (10.15)
0
(
eiq(x−x )
1
a− 2 ,
Z
0 dq for a > 0
where G(x − x ) = d 2 ξ = − 21 (10.16)
(2π) q 2 + (1/ξ) (−2a) , for a < 0
We recall that G(r) is both the Green function and the correlation function in the Gaussian approximation. Hence
the Gaussian critical exponents are ν = 1/2 and η = 0. This ν value is consistent with what we have obtained for the
heat capacity exponent α. In 3D we get α = 1/2.
====== [10.6] Symmetry breaking

Let us first assume that a > 0 and ask whether we can trust the Gaussian approximation. Higher non-Gaussian
terms in the expansion of A[ϕ] around ϕ̄ were neglected in the above treatment. The condition for this neglect is
found in the same way as in the Bragg-Williams analysis. Namely, the neglect of the non-Gaussian term is justified
if uϕ4 aϕ2 . This leads to the condition Var(ϕ) (a/u). As in
p the Bragg-Williams analysis the same condition
is deduced if we approach a = 0 from below, from the condition Var(ϕ) |ϕ̄|. So we would like to estimate the
74
fluctuations and see if this condition is satisfied. Within the framework of the Gaussian approximation the variance
of each Fourier component of the field is

1 1
Var(ϕk ) = (10.17)
Ld k + (1/ξ)2
2
The field amplitude ϕ(x) at a given point in space is the sum of N independent Fourier components, and accordingly
X
Var(ϕ) = Var(ϕk ) = G(0) (10.18)
k
If we kept only the k = 0 contribution, as in the Bragg-Williams analysis, we would get Var(ϕ) ∼ ξ 2 /Ld , which would
imply that the Gaussian approximation is always OK in the thermodynamic limit. If on the other hand we keep all
the terms in the above some we get for d > 2 a huge result that depends on Λ. This ”bare” estimate of the variance
has no significance because it reflects the renormalization of a by the large k fluctuations as discussed by [Amit 1974].
Consequently, as suggested by Ginzburg the effective cutoff for the purpose of estimating the Gaussian fluctuations
is Λ0 ∼ 1/ξ, hence
1/ξ
k d−1 dk
Z
1
Var(ϕ) ∼ ∼ ∼ G(ξ) (10.19)
0 k 2 + (1/ξ)2 ξ d−2
This value is determined by (L/ξ)d effective modes that each contribute to the variance ξ 2 /Ld , hence it is Λ indepen-
dent unlike the bare value G(0). Substitution into the condition Var(ϕ) (a/u) leads to the Ginzburg Criterion
|T − Tc(0) | C u2/(4−d) (10.20)
where C is a constant. This condition defines the border of the critical region. Within the critical region the Gaussian
approximation breaks down because of non-Gaussian fluctuations.
T<Tc Tc <T
T~Tc
We now turn to the question whether there is a phase transition within the critical region. The other possibility is
that the fluctuations smear away the the phase transition leading to a smooth r rather than abrupt crossover. Namely,
(0)
for T Tc there is a symmetry breaking, such that the mean field ϕ̄ jumps from positive finite value to negative
finite value as h is varied across h = 0. Obviously, in order to observe a phase transition we have to require that this
abrupt jump is not smeared by the (non-Gaussian) fluctuations within the critical region. We shall discuss below two
cases in which fluctuations completely destroy the possibility to observe a phase transition.
Goldstone excitations. – First we refer to the case where the order parameter has a continuous rather than a
discrete symmetry. To be specific let us assume a complex order parameter Ψ = (ψ1 , ψ2 ). In such case the potential
V (Ψ) = (1/2)a|Ψ|2 + (1/4)|Ψ|4 looks like a Mexican hat for r < 0. It means that instead of having two minima we
have a continuous set of minima. This implies that the low frequency excitations of the system are phonon-like, called
magnons in the ferromagnetic context. In general the excitations that appear if a continuous symmetry is broken
are known as Goldstone excitations. They have some dispersion ω = c|k|. Coming back to the partition sum, we
see that we have to include the Goldstone excitations in the Var(Ψ) calculation, leading formally to an integral over
1/k 2 instead of 1/(k 2 + (1/ξ)2 ). The integral is ”infrared divergent” unless d > 2. We conclude that the fluctuations
destroy the possibility to observe a phase transition at d = 2. This is know as the Mermin-Wagner theorem.
Non-Gaussian fluctuations. – Going back to the real field case, there are no Goldstone excitations, still we might
have non-Gaussian excitations that smear the phase-transition. For d > 1 the implication of such fluctuation is not too
75
dramatic: the critical point is shifted down (ac < 0) but remains finite (see discussion of the RG analysis). For d = 1
we already have discussed the absence of a finite temperature phase-transition using a ”domain walls” perspective.
Let us re-phase the explanation using a field-theory perspective. It is clear that if we have a phase transition, then
formally Var(ϕ) ∼ ϕ̄2 on the critical line. The question is whether it becomes Var(ϕ) ϕ̄2 for any finite h away
from the critical line. This depends on the height of the ”barrier” between the two minima, and therefore cannot
be deduced from the Ginzburg criterion: the latter is based on a local Gaussian approximation that does not know
anything about the height of the ”barrier”. We shall see in the next section that in the d = 1 case the crossover has
a finite width: there is no abrupt change in hϕi as h = 0 is crossed.
Quantum phase transition. – The Landau model is “classical” in the sense that its Hamiltonian commutes with
the order parameter. This is not the case e.g. for coupled rotors (see previous lecture). In a field theory treatment the
partition function Z = trace[e−βH ] can be written as a Feynman path integral over the field ϕ(x, τ ), where τ ∈ [0, β] is
the so-called imaginary time. The integral is over all field configurations in ([0, β] × [0, L]d ). Accordingly the analysis
of the ground state in the thermodynamic limit (β, L → ∞) maps formally to a classical field theory with dcl = d + 1
dimensions. This implies that is is feasible to observe a zero temperature “quantum phase transitions”, as a control
parameter is varied, even for d = 1 and notably at d = 2.
====== [10.7] The one dimensional model

The one-dimensional field model can be solved exactly. This is merely a variation on the ”transfer matrix” method.
The Dϕ integral is sliced and written as a trace over the product of N matrices. Each matrix can be written as
exp(−dxH) where H is the ”Hamiltonian”. One realize that this is nothing else but the Feynman path integral in
”imaginary time”. Let us define
1 ∂2 1 ∂2 h1 1 i
H = − 2
+ V (ϕ) = − 2
+ aϕ2 + ϕ4 (10.21)
2 ∂ϕ 2 ∂ϕ 2 4
using the notation ϕ̇ = dϕ/dx and x = τ with periodic boundary conditions over [0, L], the calculation of the partition
function goes as follow:
Z RL 1 2
X
Z = Dϕ e− 0 2 ϕ̇ +V (ϕ)dτ
= trace(e−LH ) = e−Lµn (10.22)
n
where µn are the eigenvalues of H. In the thermodynamic limit F (T, h) = LT µ0 where µ0 is the ground state energy
of H. Similarly hϕi is just the ground state expectation value. For the correlation function we get
h i X
G(r) ∝ trace e−(L−r)H ϕe−rH ϕ ∝ |hn|ϕ|0i|2 e−r(µn −µ0 ) (10.23)
n
where in the last equality we already dropped the terms that vanish in the thermodynamic limit. We see that the
long tails are characterized the correlation length ξ = 1/(µ1 − µ0 ). This correlation length does not diverge, reflecting
that the variation of hϕi is smooth. The crossover at h = 0 has a width that equals the tunnel splitting (µ1 − µ0 ).
76
====== [10.8] Coarse graining and scaling

The free energy F (a, u, h, c; Λ, L) of an homogeneous system that is described by the Landau model, and the associated
correlation function G(r, a, u, c; Λ, L) depend on the following parameters:
L = linear size of the model (10.24)

Λ = largest momentum scale (10.25)
(c=1, a, u) = microscopic related parameters (10.26)
r = distance between two test points (10.27)
Schematically we write the free energy as F (g; Λ, L), where g stands for any of the action parameters. The c = 1
convention fixes the units of the field ϕ, as well as the engineering dimension dg of any of the action parameters. The
microscopic-related parameters have been determined, as in the Bragg-Williams approximation, by summation over
all the microscopic configurations that correspond to the same coarse-grained ϕ(x). Accordingly these parameters
depend on the value of Λ. To emphasize this aspect one may use the notation g(Λ).
Assume that we have used a coarse-graining cutoff Λ to construct the action. But later we might prefer to work with
an action that corresponds to a somewhat lower cutoff Λ0 . Obviously the result of the calculation should be the same.
Accordingly we write
F (g; Λ, L) = F (g 0 ; Λ0 , L) = F (sdg g 0 ; Λ, sL) = sd F (sdg g 0 ; Λ, L) (10.28)
In the second equality we scaled the units by factor s = Λ0 /Λ in order to restore the original Λ cutoff, and in the last
equality we have used the thermodynamic limit in order to restore the original L cutoff. Using a compact notation
we have deduced the scaling relation
F (g) = sd F (gs ), gs ≡ sdg g(sΛ) (10.29)
With regard to the correlation function we note that G, unlike F , does not depend on L, but the units of the field
have been modified, hence sd should be replaced by a different scaling factor that we discuss in the next section.
It should be clear that the units of length are arbitrary, hence F should be a well defined function of the dimensionless
model parameters. It follows that we can write the scaling relation for the microscopic related parameters without
giving explicit reference to Λ. Namely,
g̃s = R(s) g̃, g̃s ≡ Λdg gs (10.30)
where R(s) is a non-linear transformation that depends on s. This transformation relates values of F along
a trajectory in parameter space, and by definition has the semi-group property R(s2 )R(s1 ) = R(s2 s1 ). Using
the parametrization s = e−τ we can write the transformation as g̃τ = R(τ )g̃0 and the semi-group property as
R(τ2 )R(τ1 ) = R(τ2 + τ1 ). Clearly we can generate R(τ ) from infinitesimal steps, so we define a β function via the
expansion R(τ )g̃ = g̃ + τ β(g̃) + O(τ 2 ), and write the so called renormalization group (RG) equation as
dg̃
= β(g̃), [opposite sign convention if dτ 7→ d ln Λ] (10.31)
dτ
Increasing the course graining parameter τ , we get a flow in g space. A fixed point of this flow represents a critical
point of the model, where the system look-alike on any scale. If we start the RG trajectory at a point close to the
fixed point it will flow away, meaning that on coarse-grained scale the system looks like having larger |T − Tc |. For
the Landau model in 3D we shall see below that the RG equation for a becomes
dã 1 1 3
= (ã − ãc ), ãc = − , ν = (10.32)
dτ ν 5 5
77
Defining t = (a − ac ) we get the solution ts = s−1/ν t, leading to the scaling relation F (t) = sd F (s−1/ν t). Hence for
the heat capacity exponent we get α = 2 − νd = 1/5, and not the mean-filed value α = 0, neither the Gaussain value
α = 1/2.
====== [10.9] Renormalization Group (RG) analysis

We outline the RG procedure that is used in order to find the β(g) function, where g stands for the parameters (a, u, c)
of the Landau model with the convention c = 1. For extra technical details see Section 18.7 of Huang.
Step1 of RG.– Perturbation theory allows to integrate the high Fourier components within a shell Λ0 < k < Λ.
where Λ0 = Λ−δΛ. Namely any field configuration can be written as a sum of smooth and erratic components:
X X
ϕ(x) = ϕk eikx + ϕk eikx ≡ ϕ̄(x) + ϕ̃(x) (10.33)
|k|<Λ0 Λ0 <|k|<Λ
The action can be expanded with respect to ϕ̃ up to quadratic order. This is allowed because δΛ is chosen as arbitrarily
small. Using abstract notation with regard to field indexes we write the expansion as
A[ϕ(·)] = A[ϕ̄(·)] + hinduced [ϕ̄] ϕ̃ + ainduced [ϕ̄] ϕ̃2 (10.34)
Now it is possible to use Gaussian integration over ϕ̃ to get an effective expression for A[ϕ(·)] that involves new values
for the model parameters. Accordingly
F (a, u, c; Λ, L) = F (a0 , u0 , c0 ; Λ0 , L) (10.35)
Doing the algebra the result is

h i
a0 = a + δΛ 3Ωd Λd−3 u − Λd−5 au (10.36)
h i
u0 = u − δΛ 9Ωd Λd−5 u2 (10.37)
0
c = c (10.38)
Note that the u of Huang should be identified with our u/4, and r of Huang is identified with our a. Though not the
case here, one should be aware that in general the elimination of the high Fourier components might spoils the c = 1
convention.
Step2 of RG.– In ”step2” of the RG procedure the original value of Λ is restored vis engineering scaling, and then
the thermodynamic limit is assumed to restore L as well. Accordingly
F (a, u, c; Λ, L) = sd F (sda a0 , sdu u0 , c0 ; Λ, L), s ≡ e−τ , τ = δΛ/Λ 1 (10.39)
Note that according to the common convention dc = 0.
Step3 of RG.– In ”step3” of the RG procedure the field ϕ is re-scaled such that the convention c = 1 is restored.
Using the notation c0 = s−η , and dropping reference to the restored parameters, we get
F (a, u) = sd F (as , us ) (10.40)

G(r, a, u) = s−2dϕ +η G(sr, as , us ) (10.41)
Here we suppressed c, because it has been restored to unity. In the Landau model c0 = c and therefore η = 0. For the
two other parameters we get
h i
as = s−2 a0 = a + τ 2a + 3Ωd Λd−2 u − Λd−4 au (10.42)
h i
us = s−(4−d) u0 = u + τ (4 − d)u − 9Ωd Λd−4 u2 (10.43)
78
where s = 1 − τ applies for an infinitesimal step. Without the high frequency contribution the above RG relation is
just a complicated way to write how the parameters are affected by engineering scaling.
RG equation.– We can illustrate the RG flow in the (a, u) space. Increasing τ means lower resolution description
of the system, with effective parameters (as , us ). It is convenient to use dimensionless parameters g̃ = Λdg g, such that
the transformation R(s) becomes free of Λ. Considering an infinitesimal τ one finds that the RG-equations of the
Landau model are
dã
= 2ã + 3ũ − 3ãũ (10.44)
dτ
dũ
= (4 − d)ũ − 9ũ2 (10.45)
dτ
where ã = a/Λ2 and ũ = Ωd u/Λ4−d .
RG flow.– The RG equation defines flow in (a, u) space. This flow is illustrated in the figure below. In the Landau
model we have two fixed points. The Gaussian fixed point is for ũ0 = 0 at ã0 = 0. The nontrivial fixed point is
(4−d)
ũc = (10.46)
9
−1
(4−d) (4−d)
ãc = − 1− (10.47)
6 6
For d < 4 the Gaussian fixed point is unstable and the flow is dominated by the nontrivial fixed point. One observes
that the critical temperature (ac ) is shifted below the mean field Gaussian value.
====== [10.10] Implications of the RG results

The results of the RG analysis are used slightly differently in Statistical Mechanics (”Ising”) and in high energy
physics (”HEP”). We first would like to explain the subtle difference, and then to focus on the the ”Ising” context.
√
HEP.– In the HEP context a = m is the so-called ”bare mass”, and u is the ”bare interaction”. They are associated
with the momentum exchange q ∼ Λ in, say, electron-electron scattering calculation. The low energy scale Λ0 is the
Compton wavelength that corresponds to the electron mass. For larger Λ the scattering involves ”loop” corrections
reflecting the virtual appearance of electron-positron pairs. Consequently the mass and the interaction become
79
“running coupling constants”. With different choices of Λ one associates different a(Λ) and u(Λ). The physical mass
and the physical interaction for q-scattering are defined through the measured dispersion-relation and through the
cross-section respectively. Both are functions of (q; a, u), and are calculated from the bare parameters. With different
choices of Λ we can associate different points in (ã, ũ) space. It follows that the relevant physics is along a specific
HEP line of the RG flow (see figure). Accordingly what we get from the RG analysis is how a(Λ) and u(Λ) depend
on Λ. It turns out that the d = 4 Landau model has the essentially the same beta function as in QED. Namely, the
beta function for the QED coupling g = e2 /(4π) is β(g) = Cg 2 with C = 2/(3π) instead of C = 1/9. One obtains
g0 1
g(Λ) = ≡ (10.48)
1 − Cg0 ln(Λ/Λ0 ) C ln(Λ∞ /Λ)
where Λ0 corresponds to the Compton length of the electron and g0 ≈ 1/137 is the asymptotic QED coupling. At
large distances (q < Λ0 ) the interaction strength is renormalized to a universal value that is independent of the bare
parameters. For short distances the interaction becomes stronger, and diverges at the Landau pole Λ∞ . Note that
this divergence is possibly not physical because the derivation of the RG equation assumes that g is small.
Ising.– In the Ising context the reasoning is different. After Λ coarse-graining we get an action with a and u, which
we represent as a point in (ã, ũ) space. If we vary the temperature it is like going along a horizontal line. If we use a
lower cutoff Λ0 it would takes us to a lower horizontal line (see figure). Assuming that we have done enough coarse-
graining, the value of u would become u ≈ uc . We therefore say that u is ”irrelevant”. Without loss of generality we
can set u = uc in our calculations. We end up with a single RG equation for the variable t = (a − ac ), namely
dt dg
= (2 − 3uc ) t generic notation: = λg (10.49)
dτ dτ
The solution of this equation is tτ = eλτ t0 hence we deduce the scaling relation F (t) = sd F (s−λ t). We therefore
conclude that F (t) ∝ |t|d/λ , from which the dependence of the heat capacity on the temperature can be derived.
The critical exponents.– We want to understand how the RG flow explains the scaling hypothesis. In the vicinity
of the fixed point we can linearize the RG equation. After linear transformation and shift we get an equation of the
type dg/dτ = λg, whose solution is gτ = eλτ g. Parameters with negative eigenvalues λ < 0 vanish from the model due
to the coarse graining and therefore are called ”irrelevant”. We keep only the relevant parameters and deduce that
F (g) = sd F (sDg g), Dg ≡ −λ (10.50)
The anomalous dimension Dg of Huang is defined with opposite sign (he is using inverse length units). In general ”g”
stands for a collection of (relevant) parameters, each having its own ”dimension”. In the Landau model the relevant
parameter is related to the temperature, namely t = (a − ac ) ∝ (T − Tc ). From the RG equation we deduce

(4−d) 1
Dt = − [2 − 3uc ] = − 2 − ≡ − (10.51)
3 ν
Hence the scaling relations take the familiar form
F (t) = sd F (s−1/ν t) (10.52)

d−2+η −1/ν
G(r, t) = s G(sr, s t) (10.53)
This means that the coarse grained system looks less and less critical as τ is increased. Having determined ν and
η = 0 using the RG procedure, we can deduce all the other critical exponents from the scaling relations.
80
====== [10.11] The Heisenberg model

The Heisenberg model relates to a three component vector field S. It is described by the Hamiltonian
X
H = −J Si · Sj (10.54)
hi,ji
The field theory version of the Heisenberg model is the nonlinear sigma model (NLSM), where S(r) is a vector field
that obeys the constraint |S(r)| = 1. The NLSM energy functional is commonly written as
Z
1
E[S] = |∇S|2 dr (10.55)
2g
The Mermin-Wagner theorem:– In the Heisenberg and Landau-Ginzburg models the order parameter has a
continuous symmetry with respect to spin rotations. This is different from the Ising and Landau models where
the symmetry is discrete (up/down). The Mermin-Wagner theorem states that continuous symmetries cannot be
spontaneously broken at finite temperature in d ≤ 2 systems with short-range interactions. This is due to long-range
fluctuations, corresponding to massless ”Goldstone modes”. These fluctuations cost little energy, while contribute
large entropy. Hence at finite temperatures the T = 0 broken symmetry is washed away.
Let us outline the argument that explains the dimensionality issue. Assume that there is a long range order. Use the
Gaussian approximation to calculate the correlation function around the selected minimum. The Goldstone modes
are massless (which is like r = 0 in the Landau model). Hence g(r) = FT[1/k 2 ]. The area under g(r) diverges due to
the low wavelength fluctuations if d ≤ 2, indicating that the minimum is washed away.
====== [10.12] The XY model

The XY model relates to a two component vector field S. It can be regarded as describing coupled rotors. The
orientation of the ith spin in the XY plane is ϕi . Accordingly we can write
X
H = −J cos(ϕi − ϕj ) (10.56)
hi,ji
The associated continuous field version is defined by the energy functional

Z Z
K K
E[ϕ] = (∇ϕ)2 dr = |Ψ0 (r)|2 dr (10.57)
2π 2π
where drp integrates over space, and Ψ(r) = eiϕ(r) . Note that the Landau-Ginzburg model has an order-parameter
iϕ(r)
Ψ(r) = n(r)e , and can be regarded as a variation of the XY model, where n(r) is not constrained to unity. The
low lying excitations of the XY model in 2D are vortexes. For example, a single vortex at the origin is represented by
ϕ(r) = q arg(x, y) = q tan−1 (y/x) (10.58)

|∇ϕ| = (q/r) [tangential] (10.59)
E[ϕ] = Kq 2 ln(L/a) (10.60)
where L is the radius of the system, and a is the lattice spacing, and q is an integer, say q = ±1. One realizes
that F~ = ∇ϕ is like a 90deg rotated version of a field dues to a charge q in a 2D electrostatic problem, while E[ϕ]
corresponds to the electroststic energy. Accordingly if we have several vortexes we get
X
E[ϕ] = const − K qi qj ln |ri − rj | (10.61)
hi,ji
81
which is formally the expression for the energy of a Coulomb gas in 2D.
Note about complex representation.– Consider a field F~ = (Fx , Fy ) in 2D that has no sources (∇ · F = 0) and
no circulations (∇ × F = 0). These two conditions are the Cauchy-Riemann equations that allow to represent the field
by a complex differentiable function f (z) = Fx − iFy , where z = x + iy. Note that in sloppy notations F~ = f (z)∗ .
Note that such functions are called holomorphic and hence they are analytic, i.e. have a convergent Taylor expansion
at any point. Note also that a holomorphic function can be regarded as a conformal map w = f (z) that induces a
local transformation df = Cdz, where the local derivative C can be regarded as a rotation (due to its phase) combined
with dilation (due to its modulo). From the Cauchy-Riemann no-circulation condition it follows that the field can be
derived from a scalar potential (F~ = −∇V ). Optionally from the Cauchy-Riemann no source condition it follows that
the field can be derived from a vector potential (F~ = ∇ × (0, 0, A)). We can summarize these two options by defining
a complex potential Ψ(z) = V (z) + iA(z) such that f (z) = −Ψ0 (z). The lines of constant A(z) are called ”stream
lines”, while the lines of constant V (z) are called ”propagation fronts”. Differences of A have the meaning of ”flux”,
while differences of V have the meaning of ”work”. For a point charge in a 2D electrostatic problem the Coulomb
field is f (z) = q/z and Ψ(z) = −q ln(z) corresponding to V (x, y) = −q ln(r). The vortex in the XY model is just a
rotated version of a Coulomb field with f (z) = −iq/z and Ψ(z) = −iq ln(z).
The Kosterlitz-Thouless transition.– Considering the XY model in 2D space, let us see what happens if we have
a collection of vortexes. The entropy which is associated with the number of possibility to position the vortex in the
lattice is S = ln[(L/a)2 ]. Accordingly the introduction of a vortex into the system implies
F [ρ] = E[ρ] − T S[ρ] = (K − 2T ) ln[(L/a)] (10.62)
Hence for T > (K/2) the creation of a vortex is favoured, so we expect to have a gas of quasi-independent vortexes.
For T < (K/2) these vortexes have to pair, which is like not having vortexes. The above argumentation implies
an ”infinite order phase transition”, known as the Kosterlitz-Thouless transition. In consistency with the Mermin-
Wagner theorem it does not lead to ”long range order”. Rather it describes a transition from ”conventional disordered
phase” at high temperature, to ”quasi-long-range ordered phase” at low-temperature. The correlation function goes
from exponential above the critical temperature, to powerlaw at the critical temperature and below. At the critical
temperature g(r) = 1/rη , with η = 1/4. Below the critical temperature a Gaussian estimate gives η = T /(2K).
82
Fluctuations and Response
[11] Fluctuations
In order to know the expectation value of an operator we need only spectral information which is present in g(E, X)
or in Z(β, X). Note that these functions contains only spectral information about the system (no information on
the dynamics). Still it is enough for the calculation of the conservative force. For example, in case of a canonical
preparation

∂H X ∂En 1 ∂ ln(Z)
hFi0 = − = pn − = (11.1)
∂X n
∂X β ∂X
In contrast to that, the fluctuations of F(t) − hF i0 require knowledge of the dynamics, and cannot be calculated from
the partition function. For simplicity of presentation we assume below that the fluctuating quantity of interest is
re-defined such that hFi0 = 0.
====== [11.1] The classical power spectrum

Consider a stationary stochastic classical variable F (t), and define its correlation function as
C(t2 − t1 ) = hF (t2 )F (t1 )i (11.2)
The power spectrum C̃(ω) is the Fourier transform of C(τ ), where τ = t2 − t1 is the time difference. In practice a
realization of F within time interval 0 < t0 < t can be Fourier analyzed as follows:
Z t
0
Fω = F (t0 )eiωt dt0 (11.3)
0
and we get the ”Wiener-Khinchin theorem”
h|Fω |2 i = C̃(ω) × t (11.4)
where we assume that t is much larger compared with the correlation time.
A related perspective concerns the intensity of the power spectrum:

Z ∞
ν = C(τ )dτ = C̃(ω=0) (11.5)
−∞
If we regard F (t) as the velocity (”steps”) in a random-walk process, then the zeroth Fourier component is
Z t
r(t) = F (t0 ) dt0 = total displacement (11.6)
0
On the average hri = 0 but the variance is linear in time:
Z tZ t
Var(r) = hF (t0 )F (t00 )i dt00 dt0 = ν t ≡ 2D t (11.7)
0 0
Accordingly the intensity ν is trivially related to the coefficient D of a diffusion process that is generated by the noisy
signal F (t).
83
====== [11.2] The quantum power spectrum

We consider a system whose dynamics is generated by Hamiltonian H. We assume that it is prepared in a stationary
state, possibly but not necessarily a thermal states. Our interest is in the fluctuations of an observable F. We use the
common interaction picture notation F(t) = eiHt Fe−iHt . The non-symmetrized correlation function of F is defined
as
S(t) = hF(t)F(0)i (11.8)
This function is complex, but its Fourier transform is real and can be calculated as follows:
∞
Em − En
Z X X
iωt 2
S̃(ω) = S(t)e dt = pn |Fmn | 2πδ ω − (11.9)
−∞ n m
~
In the case of a microcanonical preparation at some energy E, this is the same object that appears in the Fermi-
Golden-rule (FGR) for rate of decay due to transitions to other levels, namely
ΓFGR = S̃E (Ω) × f02 , for H − f (t)F, with f (t) = f0 e−iΩt (11.10)
See further discussion of the FGR in the Kubo formula section. In the above formula ω > 0 corresponds to absorption
of energy (upward transitions), while ω < 0 corresponds to emission (downward transitions).
====== [11.3] The detailed balance relation

It is a straightforward algebra to show that for a canonical preparations at temperature T , where pn ∝ exp(−En /T ),
there is a detailed balance relation:

~ω
S̃T (−ω) = exp − S̃T (ω) (11.11)
T
This implies that if we couple to the system another test system (e.g. a two level “thermometer”) it would be driven
by the fluctuations into a canonical state with the same temperature.
The disadvantage of S̃T (ω) is that it has meaning only in the quantum mechanical context. We want a formulation
that treat the quantum and the classical on equal footing. We therefore define spectral functions that have well
defined classical limit:
1D E 1
C̃(ω) ≡ FT F(t)F(0) + F(0)F(t) = S̃(ω) + S̃(−ω) (11.12)
2 2
iD E i
K̃(ω) ≡ FT [F(t), F(0)] = S̃(ω) − S̃(−ω) (11.13)
~ ~
and deduce that at thermal equilibrium they are related as follows:

2 ~ω
K̃(ω) = i tanh C̃(ω) (11.14)
~ 2T
We shall see later that K̃(ω) determines the absorption coefficient of the system, hence the above relation is going to
be the basis for a “fluctuation-dissipation relation”.
It is interesting to look on the classical limit of the detailed balance relation. The classical canonical version can be
regarded as the low frequency limit of the quantum relation:
1
K̃T (ω) = iω × C̃T (ω) [classical canonical version] (11.15)
T
84
It looks very nice in time domain:
1
KT (t) = − ĊT (t) (11.16)
T
We shall use the latter version in order to derive what we call later the “DC version” of the a generalized fluctuation-
dissipation relation.
====== [11.4] The classical version of ”detailed balance”

The purpose of the present section is to show how a general relation between K(t) and C(t) can be derived within
the framework of classical mechanics, assuming that the system is prepared in some arbitrary stationary state
ρ(x, p) = f (H(x, p)). The classical canonical version that we have deduced in the previous version by taking the
~ω → 0 limit can be regarded as a special case with f (H) ∝ exp(−βH). We first define K(t) and C(t) in a more
general way:
C(t) = trace {f (H) At B} (11.17)

K(t) = trace {f (H) [At , B]} (11.18)
The trace means dxdp phase-space integration, with obvious generalization to more than one freedom. We define
z = (x, p) and ∂ = (∂x , ∂p ). The Poisson brackets of two functions are:
X
[A, B]PB = (∂x A)(∂p B) − (∂p A)(∂x B) ≡ (∂i A)Ji,j (∂j B) (11.19)
i,j
where Ji,j = matrix{0, 1; −1, 0}. One easily proves that trace{[A, B]C} = trace{A[B, C]}. Note also the chain rule
[f (H), A] = f 0 (H)[H, A]. It is now possible to derive the following identity:
d
trace {f (H)[At , B]} = trace {[f (H), At ]B} = trace {f 0 (H)[H, At ]B} = − trace {f 0 (H)At B} (11.20)
dt
where the time evolved function At obeys the Hamilton equation of motion (d/dt)At = −[H, A]t , and we use the
canonical invariance [H, A]t = [Ht , At ]. For a microcanonical distribution f (H) = g(E)−1 δ(H − E) we get
1 d h i
KE (t) = − g(E) ĊE (t) [classical microcanonical version] (11.21)
g(E) dE
For a canonical distribution f (H) = Z(β)−1 exp(−βH) the weighted energy derivative is replaced by 1/T as antici-
pated from the quantum version. Optionally the canonical version can be obtained from the microcanonical version
by averaging over the energy with the canonical weight ∝ g(E) exp(−E/T ), and integration by parts. A physically
appealing deduction of the connection between K and C will be implied by the derivation of the classical fluctuation-
dissipation relation.
====== [11.5] Fluctuations of a many body system

A single particle dynamics at energy can be characterized by the single particle microcanonical fluctuation spectrum
C (ω), and by the associated response function
1 dh i
K̃ (ω) = iω × g() C̃ (ω) (11.22)
g() d
Let us assume that we have a many body occupation that is described by an occupation function f (n ). If we consider
a one-body operator Â = mn Amn a†m an , the expectation function an additive property that relates it to the single
P
85
particle expectation values:
X
hÂi = f (En ) hAin (11.23)
n
If C is the two-body operator such property does not in general exist: the total ”interaction” is greater than the
sum of the interactions within subsets. But if the two-body operator C = [A, B] is defined as the commutator of two
one-body operators, it is easy to show that the additive property is re-gained. According we deduce that for an N
body system
Z
[N ]
K̃ (ω) = d g()f () K̃ (ω) (11.24)
Expression K̃ (ω) using C̃ (ω) we get
(
g(F ) C̃F (ω) degenerate Fermi occupation (T F )
K̃ [N ] (ω) = iω × (11.25)
(N/T ) C̃T (ω) dilute Boltzman occupation (T F )
where the subscripts F and T implies that C (ω) is evaluated at the Fermi energy, or averaged according to the
thermal occupation, respectively. Note that in practice g(F ) ∼ N/F , so we have a smooth crossover at T ∼ F . The
fluctuations of the many-body current are deduced from the detailed balance relation:
h i
~ ~ω
C̃ [N ] (ω) = coth Im K̃ [N ] (ω) (11.26)
2 2T
Considering a system of Fermions we get

(
~ω ~ω
[N ] 2 coth 2T g(F ) C̃F (ω) degenerate Fermi occupation
C̃ (ω) = (11.27)
N C̃T (ω) dilute Boltzmann occupation
Contrary to the classical reasoning the zero temperature Fermi sea is not noisy νN = C̃ [N ] (0) = 0. The intensity νN
of the current-fluctuations is not simply the sum over the one-particle fluctuations. The classical result is νN = N νT
where νT is the thermally averaged one-particle fluctuation-intensity. In contrast, the quantum result is νN = NT νF ,
where νF is evaluated at the Fermi energy, and NT = T g(F ) ≡ T /∆ is the effective number of particles that contribute
to the noise. Strangely enough (see next lecture) the classical and the quantum calculations give the same result within
the framework of the Drude model: the temperature dependence merely shifts from the “ν” to the “N ”.
Many body calculation.– It is interesting to see how the relation between many-body fluctuations and single
particle fluctuations is deduced for a low temperature system of Fermions using a direct calculation. We consider the
fluctuations of a general observable A. If we treat the many body system as a whole then we have to employ second
quantization to write Â = mn Amn a†m an . Excluding the irrelevant diagonal n=m terms we get for a non-interacting
P
system in a thermal state
D E
S̃ [N ] (ω) = FT Â(t)Â(0) (11.28)
X
|Amn |2 a†n am a†m an 2πδ(ω−(m −n ))

=
nm
X
= (1−f (m ))f (n )|Amn |2 2πδ(ω−(m −n ))
nm
Z
d
= (1−f (+ω))f () C̃ (ω)
∆
86
If we had bosons the last expression would be (1+f )f instead of (1−f )f . In the dilute Boltzmann limit (f 1)
we recover additivity, as expected from classical considerations with regard to uncorrelated motions of independent
particles. But for non-dilute occupations the results depends of the “many body statistics”. For bosons there are
enhanced fluctuations due to the ”bunching” of particles in the same orbital. For fermions it is the opposite effect.
In the latter case let us assume that the single particle power spectrum has a well-defined mean level spacing ∆ at
the energy range of interest, namely at the thermal vicinity of the Fermi energy. At the limit of zero temperature the
result of the integral is clearly zero for ω < 0 reflecting that a zero-temperature the system can only absorb energy.
For positive frequencies the zero temperature result (ω/∆)C̃F (ω) is proportional to the number levels (ω/∆) from
which transitions to empty orbitals can take place. For finite temperature we get:
ω/∆
S̃ [N ] (ω) = C̃F (ω) (11.29)
1 − e−ω/T
ω ω
C̃ [N ] (ω) = coth C̃F (ω) (11.30)
2∆ 2T
ω
K̃ [N ] (ω) = i C̃F (ω) (11.31)
∆
Needless to say that these results, that have been deduced here from a direct many-body calculation, are consistent
with the former deduction that has been based on detailed-balance considerations.
====== [11.6] Fluctuations of several observables

Give several observables F j , and assuming that the system is prepared in a stationary state, the fluctuations can be
characterized by the correlation function
S kj (t) = hF k (t)F j (0)i (11.32)
The associated spectral function S̃ kj (ω) is defined by a Fourier transform. For simplicity we use the notations F 1 = A
and F 2 = B, and write the spectral decomposition

AB
X X Em − En
S̃ (ω) = pn Anm Bmn 2πδ ω − (11.33)
n
~
m(6=n)
It is convenient to write S AB (t) as the sum of two real spectral functions that have a good classical limit:
~
S AB (t) = C AB (t) − i K AB (t) (11.34)
2
1 D E
C AB (t) ≡ A(t)B(0) + B(0)A(t) (11.35)
2
iD E
K AB (t) ≡ [A(t), B(0)] (11.36)
~
We use the notations S̃ AB (ω), and C̃ AB (ω), and K̃ AB (ω) for their Fourier transforms. With regard to the spectral
decomposition of C̃ AB (ω) and K̃ AB (ω) we note that it is convenient to write pn = f (En ). We can simplify these
expressions by interchanging the dummy indexes n,m in the second term. Thus we get

AB 1X Em −En
C̃ (ω) = (f (En ) + f (Em )) Anm Bmn 2πδ ω − (11.37)
2 n,m ~
X f (En ) − f (Em )
AB Em −En
K̃ (ω) = iω Anm Bmn 2πδ ω − (11.38)
n,m
Em − En ~
Note that for a canonical state f (En )−f (Em ) = tanh((En −Em )/(2T )) × (f (En )+f (Em )). Note also that in the
expression for K̃ AB (ω), the ω cancels the Em −En denominator. The reason for pulling ω out of the sum is to
emphasize the low frequency dependence.
87
====== [11.7] Reciprocity relations and detailed balance

∗
There are some reciprocity relations that should be noticed. First we note that by definition S AB (t) = S BA (−t) .
In practice it is more illuminating to write the FTed version of this reciprocity relation, which can be directly deduced
by inspection of the spectral decomposition. Namely,
BA ∗
S̃ AB (ω) = S̃ (ω) (11.39)
∗ ∗
It follows that C̃ AB (ω) = C̃ BA (ω) , while K̃ AB (ω) = − K̃ BA (ω) . There is a second reciprocity relation that
follows from time reversal invariace. Also here it is simpler to look on the spectral decomposition and to remember
that [An,m ]∗ = An∗ ,m∗ , where n∗ and m∗ are the eigenstates of the time reversed Hamiltonian. In practical terms it
∗
means that one has to reverse the sign of the magnetic field h. Consequently S̃ BA (ω; h) = [±]S̃ BA (ω; −h), where
the plus (minus) applies if the signs of A and B transform (not) in the same way under time reversal. Combining
with the trivial reciprocity relation we get the Onsager reciprocity relation
S̃ AB (ω; h) = [±] S̃ BA (ω; −h) (11.40)
The Kubo formula that we discuss in the next section implies that the same reciprocity relations hold for the response
kernel αkj , to the susceptibility χkj and to the DC conductance Gkj . These are called Onsager reciprocity relations
Finally, we can also generalize what we called the “detailed balance relation”. In the quantum context this is a relation
between “positive” and “negative” frequencies. Assuming that the system is prepared in a canonical state we have

~ω
S̃TAB (−ω) = exp − S̃TBA (ω) (11.41)
T
From here it follows that

2 ~ω
K̃TAB (ω) = i tanh C̃TAB (ω) [quantum canonical version] (11.42)
~ 2T
In the classical limit this relation takes the form
1 AB
KTAB (t) = − Ċ (t) [classical canonical version] (11.43)
T T
where the dot indicates time derivative.
The Kubo formula that we discuss in the next section is expressed using K̃ AB (t). But it is more convenient to
use C̃ AB (t). The canonical relation between the two is the basis for the Fluctuation-Dissipation relation.
88
[12] Linear response theory
====== [12.1] The notion of linear response

Let us assume that X(t) is an input signal, while F (t) is the output signal of some black box. Linear response means
that the two are related by
Z ∞
F (t) = α(t − t0 ) X(t0 ) dt0 (12.1)
−∞
The response kernel α(t − t0 ) can be interpreted as the output signal that follows a δ(t − t0 ) input signal. We assume
a causal relationship, meaning that α(τ ) = 0 for τ < 0. The linear relation above can be written in terms of Fourier
components as:
Fω = χ(ω) Xω (12.2)
where χ(ω) is called the generalized susceptibility. Because of causality χ(ω) is analytic in the upper complex plane.
Consequently its real and imaginary parts are inter-related by the Hilbert transform:
∞
Im[χ(ω 0 )] dω 0
Z
Re[χ(ω)] = (12.3)
−∞ ω0 − ω π
(the reverse Hilbert transform goes with an opposite sign). The imaginary part of χ(ω) is the sine transforms of α(τ ),
and therefore it is proportional to ω for small frequencies. Consequently it is useful to define
Z ∞ Z ∞
χ0 (ω) ≡ Re[χ(ω)] = α(τ ) cos(ωτ )dτ ∼ α(τ )dτ (12.4)
Z0 ∞ Z0 ∞
Im[χ(ω)] sin(ωτ )
η(ω) ≡ = α(τ ) dτ ∼ α(τ ) τ dτ (12.5)
ω 0 ω 0
The asymptotic expressions apply for small frequencies: in this ”DC driving” limit one can regard χ0 and η as
constants. In any case with the above notations one can write
Fω = [χ0 (ω) + iωη(ω)] Xω = χ0 Xω − η Ẋω (12.6)
====== [12.2] Rate of energy absorption

We first consider the case of one parameter driving, where F (t) = hFit − hF i0 is the deviation from equilibrium value
of a generalized force that is conjugate to X. In the postulated LRT relation Fω = χ0 Xω − η Ẋω , the driving X(t)
is assumed to be a small deviation from some reference value X0 = 0. What we call “DC response” means that we
regard χ0 and η as constants in the frequency range of interest. Then we deduce that
hFit = hFi0 + χ0 (X − X0 ) − η Ẋ = hFiX − η Ẋ (12.7)
The in-phase response gives the conservative effect, and in the second equality it has been absorbed into the first
term. The out-of-phase response gives the dissipation. The dissipative term is responsible to the irreversible work as
discussed in the ”Work” section. The rate of dissipation is Ẇ = η Ẋ 2 .
The above considerations regarding dissipation can be generalized to a source that has a wide power spectrum
(”AC driving”). The irreversible work equals the time integral over F (t)Ẋ, and hence to the integral over Fω Ẋω .
89
Assuming linear response we get an integral over η(ω)Ẋω2 . Note that in the context of electrical engineering X(t)
might represent magnetic flux, hence Ẋω are the Fourier components of the voltage. For a stationary driving source
Ẋω2 is proportional to the measurement time and is characterized by a power spectrum S̃Ẋ (ω). Consequently the rate
of energy absorption is
Z ∞
dω
Ẇ = η(ω) S̃Ẋ (ω) ≡ η̄AC RMS[Ẋ]2 (12.8)
−∞ 2π
Possibly it is more transparent to consider a pure AC source that has a definite frequency Ω. In such a case we write
h i
X(t) = Re Ae−iΩt (12.9)
h i
F (t) = Re χ(Ω) Ae−iΩt = χ0 (Ω)X − η(Ω)Ẋ (12.10)
Ẇ = h−ẊFit = η(Ω) × (1/2)[AΩ]2 , [averaged over cycle] (12.11)
√
Note again that only the out-of-phase response gives dissipation, and that AΩ/ 2 is the RMS value of sinusoidal
driving.
====== [12.3] LRT with several variables

Commonly the Hamiltonian H(r, p; X1 , X2 , X3 ) depends on several control parameters. Then we can define gener-
alized forces in the usual way:
∂H
Fk = − (12.12)
∂Xk
It is implicit above that the Xj represent some small deviation form some initial value X(0) = X0 = 0. Consequently
the postulated linear response relation is written as
XZ ∞
hF k it = αkj (t − t0 ) Xj (t0 )dt0 (12.13)
j −∞
The low frequency limit of the linear relation between the generalized forces and the rate of the driving can be regarded
as a generalized Ohm law that includes an additional ”geometric” term. Disregarding the conservative contribution
and changing notation for the dissipation coefficient the one parameter version hFi = −GẊ is generalized as follows:
X X X
hF k i = − Gkj Ẋj = − η kj Ẋj − B kj Ẋj (12.14)
j j j
where η kj and B kj are the symmetric and anti-symmetric parts of Gkj . In an abstract notation this formula can be
written as follows:
hFi = −η Ẋ − B ∧ Ẋ (12.15)
Note that second term is analogous to a magnetic Lorentz force. Later we shall see that it can be derived form the
theory of adiabatic processes, where it can be expressed as a ”rotor” of the Berry connection A. The derivation of
the first term requires the larger perspective of the Kubo formula which we discuss in the next section. It is the first
term that is responsible for dissipation. Namely, the rate of dissipation is given by
X X
Ẇ = − hF k iẊk = ηkj Ẋk Ẋj (12.16)
k k,j
90
====== [12.4] The Kubo formula

The Kubo formula is an expression for the response kernel that relates the expectation value hAit of some observable A
to driving field f (t), where the driving term in the Hamiltonian −f (t)B involves the operator B.
h i
AB i
α (t) = Θ(t) A(t), B ≡ Θ(t) K AB (t) [Kubo formula] (12.17)
~
The formula has a good classical limit, and has various derivations. See ”Lecture notes in quantum mechanics”. One
option is to deduce hA(t)i = trace(Aρ(t)) from the time evolution of the probability matrix. Another is to regard
the Kubo formula as the interaction picture version of the “rate of change formula”. Namely, the rate of change of
the expectation value of A is determined by the expectation value of the commutator [H, A], hence in the interaction
picture it is related to [B, A].
Yet there is a very simple way to derive the Kubo formula in ”one line”. Assume that the system is prepared in a
stationary state of H0 , and that we provide a pulse f (t) = λδ(t) at t = 0. Accordingly the evolution operator after
time t is is U (t) = exp(−iH0 t) exp(iλB). By definition α(t) is the first-order approximation for the response to this
pulse. Hence we get
hA(t)i0 = hU (t)† AU (t)i0 = he−iλB A(t)e+iλB i0 = hAi0 + iλh[A(t), B]i0 (12.18)
From here the Kubo formula follows.
====== [12.5] Memory and Sensitivity

The so called out-of-time-order correlator (OTOC) of operators A and B is defined as follows:
D E
K ABAB (t) = [A(t), B]† [A(t), B] (12.19)
In order to understand its significance let us write the this correlator and the Kubo correlator for a classical particle,
with the substitutions A = x and B = p.
D E
K AB (t) = − [x(t), p]PB (12.20)
D 2 E
K ABAB (t) = [x(t), p]PB (12.21)

An infinitesimal perturbation at t = 0 shifts the initial position of the particle a distance λ = δx, and consequently
x(t) = x(0) − λ[x(t), p]PB . By definition the Kubo response kernel reflects the memory for the perturbation of the
initial conditions, namely K AB (t) = (1/λ) hδx(t)i. In contrast to that, the the OTOC reflects the sensitivity for
the perturbation of the initial conditions. For chaotic
system we have |δx(t)| ∼ δx(0)eγt , where γ is known as the
Lyapunov exponent. Accordingly K ABAB (t) = (1/λ2 ) |δx(t)|2 reflect the exponential sensitivity of the system to
any small perturbation of the initial conditions. The quantum mechanical version of the OTOC in general suppresses
this sensitivity.
====== [12.6] The Onsager regression formula

A related deduction of the Kubo formula is based on the analysis of a “quench process”. We assume that a system
is described by the Hamiltonian Hλ = H0 − λB. For example B might be the volume of the gas, and then the
conjugate field λ is the applied pressure. At t = 0 the field λ is instantly turned off, such that the dynamics for t > 0
is described by the unperturbed Hamiltonian H0 . In the lecture regrading generalized forces it has been shown that
the compressibility equals Var(B)/T . This means that the expectation value hBit should decay from hBiλ to hBi0 ,
where
1
hBiλ = hBi0 + λ Var(B) (12.22)
T
91
Here we repeat essentially the same calculation, but in the context of a time dependent scenario, considering an
arbitrary observable A. Not caring about commutation relations (”classical limit”) we get
trace [A(t) exp (−βHλ )] h i

hA(t)i = = hA(t)i0 + βλ hA(t)Bi0 − hA(t)i0 hBi0 + higher orders (12.23)
trace [exp (−βHλ )]
Note that hA(t)i0 = hAi0 . Thus, in the limit of very weak field,
1
hA(t)i = hAi0 + λ C AB (t) [f (t) is a step function] (12.24)
T
We see that the re-equilibration mimics the decay of the pertinent correlation function. This can be regarded as a
formal way to justify the “Onsager regression hypothesis” (see later). From here we can derive the classical version
of the Kubo formula. We simply have to notice that the response for a “delta pulse” is simply the derivative of the
response for a “step function”. Accordingly
1
hA(t)i = hAi0 − λ Ċ AB (t) [f (t) is a delta pulse] (12.25)
T
Using the relation K(τ ) = −(1/T )Ċ(τ ), that has been derived in the lecture regarding fluctuations, we deduce the
Kubo formula.
====== [12.7] The Onsager regression hypothesis

The Onsager regression hypothesis states that “the average regression of fluctuations should obey the same laws
as the corresponding irreversible process”. The regression scenario is defined as follows. We allow the system to
equilibrate in the presence of an applied field. Then we turn off the field, and watch the time dependence of some
observable. We already demonstrated that the Onsager regression hypothesis can be deduced in the classical limit
via first order-perturbation theory treatment of the response. The quantum generalization of the Onsager regression
formula is known as the fluctuation-dissipation relation, to be discussed in the next lecture.
The ”regression formula” describes the relaxation of the system back to equilibrium. Consider the case of having
one fluctuating variable A. If we assume that its relaxation obeys an exponential decay law Ȧ = −γA, then we can
deduce that the rate of relaxation is

Ȧ Ċ(0)
γ = − = − (12.26)
A C(0)
t=0
In the next section we describe the departure from equilibrium with a scaled variable XA = A/C(0), and write the
relaxation as Ȧ = −γAA XA with γAA = −Ċ(0). Then, for several variables, it would be possible to get a generalized
formula γij = −Ċij (0) with a reciprocal relation between γij and γji .
====== [12.8] Onsager reciprocity

Within the Hamiltonian framework the Onsager reciprocity is the statement that response coefficients obey relations
of the type GAB (h) = GBA (−h), where h is the magnetic field. This follows from the observation that they are
related to cross-correlation functions that obey reciprocity as discussed in previous lecture. Below we provide a more
general perspective of Onsager reciprocity, which is not based on Hamiltonian formulation, and hence can be applied
to a wider range of problems in thermodynamics. We no longer assume the canonical Boltzmann ensemble. Instead
we assume that the probability of a ”configuration” is given by
p(ϕ1 , ϕ2 , ...) ∝ e−A(ϕ1 ,ϕ2 ,...) (12.27)

92
For example the ϕj might represent a set of chemical reaction coordinates. Or it can stand for the amount of energy
that is transferred from one body to another body. For an isolated system that is described by a microcanonical
ensemble the function A(ϕ) is the Boltzmann entropy of a given configuration. For a system in contact with a heat
bath that is described by a canonical ensemble A(ϕ) = βF (ϕ), where F is the Helmholtz function.
We use the convention that ϕ = 0 is the most probable value if there are no constraints nor additional fields. We
assume that the deviations from equilibrium are small, such that A(ϕ) can approximate by a quadratic expression:
1X
A(ϕ) = Aij ϕi ϕj (12.28)
2 ij
As for the temporal aspect we assume that the fluctuations are characterized by some correlation function:
Cij (t) = hϕi (t)ϕj (0)i (12.29)
The Aij determines the correlations Cij (0) = hϕi ϕj i = (A−1 )i,j , but give no
P information on the temporal aspect. At
this point it is convenient to define conjugate variables Xk = −∂k A = − i Akj ϕj , which are like restoring forces,
and to realize that hXk ϕj i = −δk,j . Next we assume that the relaxation of the system is described by a linear relation
that reflects the tendency of the system to restore equilibrium
X
ϕ̇i = γik Xk [which implies ϕ̇ = γAϕ where both A and γ are ”nice” matrices] (12.30)
k
From here it follows that hϕ̇i ϕj i = −γij . Thus, as expected, we can derive response coefficients from correlation
functions
γij = −Ċij (τ = 0) (12.31)
Note that the absence of the 1/T prefactor is because we defined the conjugate variables not from the Hamiltonian
but from the ”Action”. The Onsager reciprocity relation follows automatically from the symmetry of the correlation
function. In the absence of magnetic field, assuming time-reversible dynamics, we have Cij (τ ) = Cij (−τ ), and hence
γij = γji . A non-trivial example for the Onsager reciprocity relation is discussed with regard to the thermo-electric
effect in the kinetic theory lecture.
====== [12.9] The Kubo formula for AC/DC driving

The DC value of the dissipation coefficient is obtained by integration:
Z ∞
η AB [DC limit] = K AB (τ ) τ dτ (12.32)
0
More generally, an expression for the generalized susceptibility follows from the convolution theorem:
∞
iK̃ AB (ω 0 ) dω 0
h i Z
AB AB
χ (ω) ≡ FT α (τ ) = (12.33)
−∞ ω − ω 0 + i0 2π
Of particular interest is the case where A = B is a generalized force that is conjugate to the variation of some
parameter X such that f (t) = (X(t) − X0 ). This is the case of interest in the study of friction (where f˙ is the
displacement velocity) and in the study of electrical conductance (where f˙ is the electromotive field). From the
definition it follows that K(−τ ) = −K(τ ), hence K(ω) is pure imaginary, and consequently the friction coefficient is
Im[χ(ω)] 1
η(ω) ≡ = K̃(ω) [Kubo formula for the dissipation coefficient] (12.34)
ω i2ω
93
====== [12.10] The Kubo formula - FGR version

So far we have used versions of the Kubo formula that are ”good” both classically and quantum mechanically. In the
quantum case one can write a version of this formula that involves the non-symmetrized correlation function:
1 h i
η(ω) = S̃(ω) − S̃(−ω) [Quantum FGR version of the Kubo formula] (12.35)
2~ω
This expression can be deduced directly from the FGR picture as follows. Assume that Hdriving = −f (t)B with
f (t) = f0 sin(Ωt). From the FGR it follows that the rate of energy absorption due to upward transitions is
(f0 /2)2 S̃(Ω)Ω. Similarly the rate of energy emission is (f0 /2)2 S̃(−Ω)Ω. The net rate of heating is the difference. By
definition it is written as Ẇ = η(Ω)[f˙2 ], where [f˙2 ] = (1/2)[f0 Ω]2 . Hence one deduce the above expression for η.
Below we discuss the non-trivial generalization of the Kubo linear response formalism for the case of Hamiltonian
that depends on several parameters. We start with the dissipation-less quantum adiabatic limit, and continue with
the full linear response analysis.
====== [12.11] Adiabatic response

For an extended presentation see ”Lecture notes in quantum mechanics”. Given an Hamiltonian H(r, p; X1 , X2 , X3 )
that depends on several control parameters, we find the zero order adiabatic basis |n(X)i with eigenenergies En (X).
Then, for a given level, we define in parameter space the ”Christoffel symbols” that are known in this context as
”Berry connection”, and the associated ”curvature field” as follows:

∂
Ajnm = i~ n(X)
m(X) (12.36)
∂Xj
Bnij = ∂i Ajn − ∂j Ain (12.37)
We use the notation Ajn = Ajnn , and note the following identities:
j
−i~Fmn
Ajnm = [n 6= m] (12.38)
Em −En
i j

X 2~Im Fnm Fmn
Bnij = (12.39)
(Em − En )2
m(6=n)
If we have 3 control variables it is convenient to use notations suggesting that we can formally regard An as a vector
potential whose rotor Bn is formally like a magnetic field:
X −7 → (X1 , X2 , X3 ) (12.40)
A − 7 → (A1nn , A2nn , A3nn ) (12.41)
B 7−→ (B 23 , B 31 , B 12 ) (12.42)
With the above definitions the Schrodinger equation can be written as follows:
d i
|ψi = − H(X(t)) |ψi (12.43)
dt ~
We expand the state in the zero order adiabatic basis
X
|ψ(t)i = an (t) |n(X(t))i (12.44)
n
94
and get the equation
dan i iX
= − (En −Ẋ · An )an − Wnm am (12.45)
dt ~ ~ m
where
X
Wnm ≡ − Ẋj Ajnm for n6=m, else zero (12.46)
j
It follows that the first order adiabatic state that is associated with the nth level is
X Wmn
|ψ(t)i = |n(X(t))i + |m(X(t))i (12.47)
En − Em
m6=n
Consequently the first order adiabatic response of a system that has been prepared in the nth adiabatic state is
X
hF k i = − Bnkj Ẋj = −B ∧ Ẋ (12.48)
j
We shall explain in the next section that this corresponds to the geometric part of the response in the Kubo formula.
The Kubo formula contains an additional non-adiabatic (dissipative) term that reflects FGR transitions between
levels.
====== [12.12] Low frequency response

Here we go beyond adiabatic response and discuss both the adiabatic and dissipative terms that are implied by the
Kubo formula. Recall that the Kubo expression for the response kernel is αkj (τ ) = Θ(τ ) K kj (τ ), whose Fourier
transform is the generalized susceptibility:
∞
iK̃ kj (ω 0 ) dω 0
Z
χkj (ω) = (12.49)
−∞ ω − ω 0 + i0 2π
Taking into account that Re[χkj (ω)] is symmetric with respect to ω we have
∞
Im[χkj (ω)]
Z
kj d kj
G = lim = lim χ (ω) = K kj (τ )τ dτ (12.50)
ω→0 ω ω→0 dω 0
The last expression (in time domain) is mentioned for completeness. In practice it is more convenient to proceed in
frequency domain. After some straightforward algebra we get
∞
Im[K̃ kj (ω)] dω Re[K̃ kj (ω)]
Z
kj 1
G = lim − ≡ η kj + B kj (12.51)
2 ω→0 ω −∞ 2π ω2
We notice that η kj is a symmetric matrix while B kj is anti-symmetric. Hence in abstract notation the linear-response
relation can be written as a generalized Ohm law:
X
hF k i = − Gkj Ẋj = −η Ẋ − B ∧ Ẋ (12.52)
j
95
This is a generalization of the adiabatic response formula. The additional term takes into account the FGR non-
adiabatic transitions between levels. To see clearly the connection we substitute the spectral decomposition of K̃ kj (ω)
and get the following expressions:
∞
iK̃ kj (ω 0 ) dω 0 X k j j k

−Fnm Fmn Fnm Fmn
Z X
kj
χ (ω) = = f (En ) + (12.53)
−∞ ω − ω 0 + i0 2π n m
~ω−(Em −En )+i0 ~ω+(Em −En )+i0
and
1 Im[K̃ kj (ω)] X f (En ) − f (Em )

η kj = lim = −π~ k
Fnm j
Fmn δ(Em − En ) (12.54)
2 ω→0 ω n,m
E n − E m
Z ∞
dω Re[K̃ kj (ω)] X k
−i~Fnm j
Fmn
B kj = − = (f (E n ) − f (E m )) (12.55)
−∞ 2π ω2 n,m
(Em − En )2
The Re[] and Im[] are not required because the doubles summation cares for that. The expression for B kj can be
written in a way that better emphasize the relation to the analysis of the adiabatic response:
k j

X X 2~Im Fnm Fmn X
kj
B = f (En ) = f (En )Bnkj (12.56)
n
(Em −En )2 n
m(6=n)
96
[13] The fluctuation dissipation relation
====== [13.1] General formulation

The essence of the fluctuation dissipation relation (FDR) is to relate the response of a system to its fluctuations in
equilibrium. In order to derive this relation we have to supply information on the preparation of the system, which
is typically assumed to be canonical. In the classical context there is a useful microcanonical version from which
the canonical version can be derived. The formal derivation of the FDR is based on the generalized detailed balance
relation that allows to express K̃ kj (ω) using C̃ kj (ω).
We first consider what we call “the AC version” of the FDR. For simplicity we consider the one-variable version: the
driving term in the Hamiltonian is −X(t)F, and our interest is in the observable F that is conjugated to the driving
field X. Recall that K̃(ω) is imaginary, and Im[χ(ω)] = [1/(2i)]K̃(ω). Assuming canonical preparation the detailed
balance relation implies that

1 ~ω
Im[χ(ω)] = tanh C̃ F F (ω) [FDR, the AC version] (13.1)
~ 2T
What we call “the DC version” of the FDR is obtained by taking the small ω limit of the AC version, which is formally
equivalent to the classical limit (small ~). One deduces that the low frequency dissipation coefficient is relate to the
equilibrium intensity of the fluctuations:
Z ∞
νT
η = , νT ≡ hF(τ )F(0)iT dτ [FDR, the DC version] (13.2)
2T −∞
Note that in the above writing we assume that the equilibrium value of the fluctuating force is hFi = 0, else F should
be re-defined so as to have a zero average.
For completeness we also point out the multi-variable version of the FDR in the DC limit. Here we change notion
and use GAB for the generalized conductance. In the DC case we know from Kubo that G is an integral over τ K(τ ).
In the classical treatment K(τ ) is the derivative of C(τ ), hence after integration by parts
Z ∞
AB 1
G = C AB (τ ) dτ [generalized FDR, classical DC version] (13.3)
T 0
Note that the (1/2) prefactor is absent, and that the integration is over positive τ , and that the cross-correlation
function C AB (τ ) does not have to be symmetric in time. The asymmetry is responsible for the geometric part of the
conductance matrix.
====== [13.2] The diffusion-dissipation picture

We can illuminate the physics of FD for DC driving using a simple diffusion-dissipation picture. We show below that
the DC energy absorption rate is related to the induced diffusion in energy space. To simplify the presentation we
use a classical language. We can deduce that the driving induce diffusion in energy space from the relation
Z t
E(t) − E(0) = −Ẋ F(t0 )dt0 (13.4)
0
leading to
Z tZ t
2
h(E(t) − E(0)) i = Ẋ 2
hF(t0 )F(t00 )idt0 dt00 (13.5)
0 0
97
where the averaging assumes a microcanonical preparation. Thus we get
δE 2 (t) = 2DE t (13.6)
where the leading order estimate for the diffusion is
Z ∞
1 2 1
DE = Ẋ hF(τ )F(0)iE dτ = νE Ẋ 2 (13.7)
2 −∞ 2
On long times we assume that the probability distribution in the democratic variable n = N (E) satisfy a standard
diffusion equation. Transforming to the non-democratic variable E we get

∂ρ ∂ ∂ 1
= g(E)DE ρ (13.8)
∂t ∂E ∂E g(E)
where g(E) reflects the ratio between the proper phase-space

R measure dn and the distorted measure dE. For more
details see [arXiv]. The energy of the system is hHi = Eρ(E)dE. Taking the time derivative and integrating by
parts, it follows that the rate of energy absorption is
Z ∞
d ∂ ρ(E)
Ẇ = hHi = − dE g(E) DE (13.9)
dt 0 ∂E g(E)
For a microcanonical preparation ρ(E) = δ(E − E). Substitution and integration by parts leads to
d 1 d
Ẇ = hHi = [g(E) DE ] (13.10)

dt g(E) dE E=E
By definition Ẇ = η Ẋ 2 and DE = (1/2)ν Ẋ 2 . Consequently the diffusion-dissipation relation reduces immediately to

the microcanonical version of the fluctuation-dissipation relation:
1 1 d
η = [g(E)νE ] (13.11)
2 g(E) dE
The canonical version η = νT /(2T ) can be derived from the integral expression for Ẇ, upon the substitution
ρ(E) = (1/Z)g(E)e−βE . Optionally it can be obtained from the microcanonical version by canonically averaging
over E and performing integration by parts.
====== [13.3] The wall formula

The first prototype application of the FD relation is to the calculation of the friction in the Brownian motion problem.
Consider a gas of particles in a box. The system is driven by moving in it a ”spoon”, or a ”piston” or a ”Brownian
body”.
• The parameter in H(X) represents the position of the spoon.

• The generalized force is the Newtonian force hFi on the spoon.
• The DC linear response relation hFi = −η Ẋ describes friction.
• The dissipation rate is Ẇ = η Ẋ 2 .
Our purpose below is to find an expression for the friction coefficient η using the FD relation. For this purpose we
have to calculate the intensity νT of the fluctuations of F at equilibrium, and to use the relation η = νT /(2T ).
98
Due to random collisions of the gas particles, the Brownian body experiences a ”random force” that can be written
as the sum of short impulses:
X
F (t) = 2mvj δ(t − tj ) (13.12)
j
Here tj is the time of the jth collision with velocity vj at the x direction. Note that |vj | ∼ vT , where vT = (T /m)1/2
is the thermal velocity. The rate of collision for N particles is

1 A vT
= N× × (13.13)
τ0 L2 L
where L3 is the volume of the box that holds the gas particles, and A is the effective area of the moving wall.
Accordingly the intensity of fluctuations is
1 N
νT = C̃(ω = 0) = (mvT )2 = m2 vT3 3 A (13.14)
τ0 L
and for the friction we get
1
η = C̃(ω = 0) = ρvT × A (13.15)
2T
where ρ = (N/L3 )m is the mass density of the gas particles.
We note that if the dynamics of the Brownian body is described by a Langevin equation, then ν/η = 2T implies that a
canonical equilibrium is reached. For more details see the Langevin section. This was in fact the historical deduction
of the FD relation by Einstein in the context of Brownian motion study.
If the Brownian particle is moving in an incompressible fluid the above result does not apply. Instead the friction is
given by Stokes Law (see “Additional topics / The Kinetic picture / Viscosity”), and we can use the FD relation “in
reverse” in order to deduce the intensity of fluctuations.
====== [13.4] The Drude formula

The second prototype application of the FD relation is to the calculation of electrical conductance. Here we show
how to derive the Drude formula for a gas of classical particles in an EMF driven ring. For an extended discussion of
electrical conductance see the ”additional topics” section of the lecture notes.
We consider a ring driven by an electro-motive force (EMF). The interaction term originates from the kinetic term
[p − (e/L)Φ(t)]2 /(2m), where Φ is the flux and I = (e/L)v is the conjugate current. Optionally we can say that the
interaction is −A(t)v, where A(t) is the vector potential and the velocity v is the conjugate variable. Summarizing:
• The parameter in H(Φ) represents the magnetic flux.

• The generalized force is the current hIi in the ring.
• The rate in which the flux is varied determines the EMF = −Φ̇ by Faraday law.
• The DC linear response relation hIi = G × EMF is Ohm law.
• The dissipation rate Ẇ = GΦ̇2 describes Joule heating.
Our purpose below is to find an expression for the conductance G using the FD relation. For this purpose, following
Drude, we postulate what is the velocity-velocity correlation function; calculate the intensity νT of the fluctuations
of I at equilibrium, and use the relation G = νT /(2T ).
Following Drude we assume an exponential velocity-velocity correlation function with a time constant τ0 , and RMS
velocity v0 , such that the mean free path is ` = v0 τ0 . Hence we deduce that C̃vv (ω) is a Lorentzian. The displacement
99
of the particle (x(t) − x(0)) is the integral over the velocity v(t0 ), hence the variance is h(x(t) − x(0))2 i = 2Dt where
1 1 2
D = C̃vv (0) = v τ0 (for a 3D sample) (13.16)
2 3 0
For a single particle the current operator is I = (e/L)v, hence the intensity of the fluctuations of the current is
ν = [(e/L)2 ] 2D. For N classical particles at thermal equilibrium we get νT = N [(e/L)2 ] 2DT , where DT is calculated
with the thermal velocity that is defined via (1/2)mvT2 = (3/2)T . For N Fermions at low temperatures we get
νT = NT (e/L)2 2DF , where NT = T /∆ is the effective number of participating electron at the Fermi energy. This
result has been derived in the lecture about fluctuations. Here ∆ is the mean level spacing at the Fermi energy. Note
that N = (2/3)(F /∆). The diffusion coefficient DF is calculated with the Fermi velocity which is determined via
(1/2)mvF2 = F . Either way we get
N e2
e 2
νT = Neff 2Deff = 2 τ0 T (13.17)
L L2 m
and for the conductance we get
1 N e2 A
G[N ] = νT = τ0 ≡ σ (13.18)
2T L2 m L
where A is the cross section of the ring. As a byproduct of this derivation we see clearly why the conductivity σ is
related to the diffusion coefficient D.
Optionally the Drude expression can be written in a way that allows to make an association with the Landauer
formula of mesoscopic physics. Considering zero temperature Fermi occupation:
e2

N ` `
G[N ] = e2 ≡ M (13.19)
mvF L L 2π L
where M corresponds to the effective number of open modes. There is a very simple toy model for which the ”expo-
nential” velocity-velocity correlation can be deduced, and hence `/L can be evaluated analytically. Consider a ring
with a single stochastic scatterer that has a transmission g. The current-current correlation function C(t) P = hI(t)Ii
at given energy E can be calculated as detailed in [arXiv]. The procedure is to use the identity hBAi = a pa hBia a,
where where A and B are any two operators,
P and hBia is the expectation value of B given that A = a. Applying
this rule in our case we get hI(t)Ii = r pr hI(t)ir Ir , where r = (x, v) labels all the possible states of the particle in
the ring, and I = evδ(x) is the current through the measurement point x = 0. The P current hI(t)ix,v , given that the
particle has been launched at x with velocity v, can be written as a sum of pulses j qj δ(t − tj ). If the measurement
point x = 0 is situated right across the barrier, such that the barrier is at x = ±L/2, one obtains
∞
evE 2 vE
X
|n| L
C(t) = hI(t)i0,vE = e (2g − 1) δ t− n (13.20)
L L n=−∞
vE
which exhibits exponential decay of correlations as in the Drude model. Assuming low temperature Fermi occupa-
tion, with NT = T /∆ thermal particles, that occupy levels whose spacing is ∆ = πvF /L, we use the FD relation
G = νT /(2T ) and get
∞
e2
Z
1 g
G[N ] = C(t) = (13.21)
∆ 0 2π 1−g
For small g one can neglect the 1−g denominator, and this formula becomes identical with the Landauer formula.
For larger g the two formulas differ. The reason for this difference concerns the geometry: Here we consider the
conductance of a barrier that is integrated into a closed ring, while Landauer concerns the conductance of a barrier
that is connected to open reservoirs. In the latter case the particle cannot circulate multiple times via the barrier.
100
====== [13.5] Conductor in electric field

A straightforward generalization of the driven ring problem applies for an extended piece of metal that in placed in a
time dependent electric field. The electric field is described by a vector potential such that E = −Ȧ. The interaction
term in the Hamiltonian is an extended version of the simplified −A(t)v that we have assumed in previous discussion:
Z
Hint = − J(x) · A(x) d3 x (13.22)
In linear response theory the current is proportional to the rate in which the parameters are being changed in time.
Regarding the values of A at different points in space as independent parameters the postulated linear response
relation takes the form
Z
hJ(x)i = σ(x, x0 ) E(x0 ) d3 x (13.23)
where σij (x, x0 ) is called the conductivity matrix. The FD relation states that the conductivity is proportional to
the temporal FT of hJi (x, t)Jj (x0 , t0 )i with respect to the time difference τ = t − t0 . The proportionality constant is
1/(2T ) in the DC limit.
====== [13.6] Forced oscillator

Consider a particle that is (say) bounded to a spring. Let us assume that the motion of x(t) ≡ hx̂it obeys the equation
mẍ + η ẋ + mΩ2 x = E, where the external driving is due to an interaction term −E(t)x̂. Accordingly,
• The parameter in H(E) represents an electric field.

• The generalized force is the polarization hxi of the particle.
• The AC linear response relation is hxi = χ(ω) E.
The FD relation implies that x has fluctuations at equilibrium, that are related to the susceptibility:

~ω h i 1
C̃xx (ω) = ~ coth Im χ(ω) , χ(ω) = (13.24)
2T −mω 2 − iηω + mΩ2
Note that the fluctuations of the velocity are C̃vv (ω) = ω 2 C̃xx (ω). Integrating over ω we get Cxx (0) and Cvv (0), from
which can deduce the average energy of the oscillator. The results are consistent with the canonical expectation in
the limit of zero damping.
====== [13.7] Forced particle

The limit Ω → 0 of the forced harmonic oscillator corresponds formally to a Brownian particle. In the classical limit
we get for the power spectrum of the velocity:
2(η/m)
C̃vv (ω) = (T /m) × (13.25)
ω 2 + (η/m)2
The area of this Lorentzian is Cvv (0) = T /m, as expected from the canonical formalism. The corresponding velocity-
velocity correlation is Drude type (exponential), with damping constant γ = η/m. The integral over the velocity-
velocity correlation function determines the diffusion coefficient, namely
1 T
D = C̃vv (0) = = µT (13.26)
2 η
This is known as the Einstein relation between D and the mobility µ = 1/η. There is an optional shortcut in the
application of the FD relation, that leads directly to the above Einstein relation. Let us write the electric field as
E = −Ȧ. The interaction term is −A(t)v. Accordingly,
101
• The parameter in H(A) represents the vector potential.

• The generalized force is the velocity hvi of the particle.
• The DC linear response relation hvi = µ E describes drift motion.
• The dissipation rate Ẇ = µE 2 describes Joule heating (per particle).
The FD relation in this notations implies that v has fluctuations at equilibrium, that are related to the mobility
µ. The ”intensity” of the velocity fluctuations is 2D. Hence the classical FD relation implies that the ratio of the
diffusion (D) to the mobility (µ) equals the temperature (T ).
====== [13.8] Duality between friction and mobility

In the “Forced particle” problem we have considered interaction of the type −x̂F (t), and defined the mobility µ,
which is the response of hx̂it to the control field F (t). In the “wall formula” problem we have considered interaction
of the type −x(t)F̂ , and defined the friction coefficient η, which is the response of hF it to the control parameter x(t).
The two points of view on the system are dual, and with the standard definitions we have the identification µ = 1/η.
====== [13.9] The fluctuations of an Ohmic system

Nyquist Noise.– The FD relation in the electrical context is known as Nyquist theorem. It can be used ”in reverse”
in order to deduce the Nyquist noise ν = 2GT , provided G is known from experiment. It should be clear that in
non-equilibrium conditions we might have extra fluctuations, which in this example are known as shot noise.
Ohmic response.– Sometimes it is convenient to characterize the system by its response, and from this to deduce
the power spectrum of the fluctuations. So we regard K̃(ω) as the input. Inspired by jargon of electrical engineering,
so-called Ohmic response is characterized by a dissipation coefficient η that is independent of ω up to some implicit
high frequency cutoff ωc . It follows that the DC intensity of the fluctuations is ν = 2ηT , and the associated spectral
functions are:
K̃ohmic (ω) = i2ηω (13.27)

h i
~ ~ω ~ω
C̃ohmic (ω) = coth Im K̃(ω) = η~ω coth (13.28)
2 2T 2T
~ ~ω
S̃ohmic (ω) = C̃ohmic (ω) − i K̃ohmic (ω) = 2η (13.29)
2 1 − e−~ω/T
====== [13.10] The fluctuations of the potential in metals

The dielectric constant of a metal is defined via the linear relation between the total electrostatic potential Utotal and
an external test charge density ρext
4πe2

1
Utotal = ρext (13.30)
ε(q, ω) q2
For simplicity we relate here and below to one component q of the fields. The total electrostatic potential is the sum
of the external potential Uext = (4πe2 /q 2 )ρext , and the induced potential Uelct = (4πe2 /q 2 )ρelct , where ρelct is the
total density of the electrons. The dielectric constant can be deduced from the equations of motion ∂ρelct /∂t = −∇J
with J = −(σ/e2 )∇Utotal − D∇ρelct that leads to the relation
(σ/e2 )q 2
ρelct = Utotal (13.31)
iω − Dq 2
102
and hence to Utotal = (1/ε)Uext , where
4πσ
ε(q, ω) = 1 − . (13.32)
iω − Dq 2
Note that

−1 4πσω ω
Im = ≈ (13.33)
ε(q, ω) (Dq 2 + 4πσ)2 + ω 2 4πσ
The interaction between the electrons and an external electrostatic field is described by Hext = Uext ρelct which can be
also written as Hext = ρext Uelct . The fluctuation dissipation relation expresses S̃ [N ] (q, ω) using the response function
α(q, ω) that relates Uelct to −ρext which is
4πe2

1
α(q, ω) = 2 1 − (13.34)
q ε(q, ω)
Using the fluctuation dissipation relation
h i 2

[N ]
S̃ (q, ω) = Im α(q, ω) (13.35)
1 − e−ω/T
we deduce
e2 1

2ω
S̃ [N ] (q, ω) ≈ (13.36)
σ q2 1 − e−ω/T
The Ohmic behavior is cut-off by |ω| . 1/τc and |q| . 1/` where ` = vF τc is the elastic mean free path, and vF is the
Fermi velocity. Recalling the Einstein relation σ = e2 νD, where ν = ∆−1 /Ld is the density of states per unit volume,
we can write this result more conveniently as follows:

[N ] 1 2ω
S̃ (q, ω) ≈ (13.37)
νDq 2 1 − e−ω/T
Note that the electron charge e cancels out from this final result for the Nyquist noise spectrum. This well-known
fact is due to the effects of screening: A larger value of the charge would be canceled by a correspondingly stronger
suppression of density fluctuations.
103
System interacting with a bath

[14] The modeling Hamiltonian
====== [14.1] The Born-Oppenheimer Hamiltonian

We first discuss system that is coupled to some other degrees of freedom that can be eliminated using an adiabatic
scheme. This leads to the Born-Oppenheimer picture. It is strongly related to Linear response theory, and the
presentation below is arranged accordingly.
Linear response theory is the leading formalism to deal with driven systems. Such systems are described by a
Hamiltonian
H = H(Q, P ; X(t)) (14.1)
where (Q, P ) is a set of canonical coordinates (in case that the Hamiltonian is the outcome of ”quantization”), and
X(t) is a set of time dependent classical parameters (”fields”). For example, X can be the position of a piston. In
such case Ẋ is its velocity. More interesting is the case where X is the magnetic flux through a ring. In such a case
Ẋ is the electro motive force. The Kubo formula allows the calculation of the response coefficients. In the mentioned
examples these are the “friction coefficient” and the “conductance of the ring” respectively.
In the limit of a very slow time variation (small Ẋ), linear response theory coincides with the “adiabatic picture”. In
this limit the response of the system can be described as a non-dissipative “geometric magnetism” effect (this term was
coined by Berry and Robbins). If we increase Ẋ beyond a certain threshold, then we get Fermi-golden-rule transitions
between levels, leading to absorption of energy (“dissipation”). Then linear response theory can be regarded as a
generalization of “Ohm law”.
The Born-Oppenheimer picture allows to deal with Hamiltonians of the type
Htotal = H0 (x, p) + H(Q, P ; x) (14.2)
Here we replaced the parameter X(t) by a dynamical variable x. The standard textbook example is the study of
diatomic molecules. In such case x is the distance between the nuclei. It is evident that the theory of driven systems
is a special limit of this problem, which is obtained if we treat x as a classical variable. For presentation purpose let
us consider the Hamiltonian
1 X 2
Htotal = p + H(Q, P ; x) (14.3)
2M j j
We define the basis |x, n(x)i = |xi ⊗ |n(x)i, and expand the state as
X
|Ψi = Ψn (x) |x, n(x)i (14.4)
n,x
Using
hx, n(x)|H|x0 , m(x0 )i = δ(x−x0 ) × δnm En (x) (14.5)

hx, n(x)|pj |x0 , m(x0 )i = (−i∂j δ(x−x0 )) × hn(x)|m(x0 )i = −i∂j δ(x−x0 )δnm − δ(x−x0 )Ajnm (x) (14.6)
we deduce that pj 7→ −i∂j − Ajnm (x), and the Hamiltonian can be written as
1 X
Htotal = (pj − Aj (x))2 + E(x) (14.7)
2M j
104
The adiabatic approximation is obtained if one neglects the n 6= m terms that couple the motion on different energy
surfaces. These couplings are responsible to the dissipation effect.
====== [14.2] The bath model Hamiltonian

The Hamiltonian of a system that interact with and environment is conveniently arranged as
Htotal = H0 (x, p) + H(Q, P ; x) (14.8)
For an interaction with a general (possibly chaotic) environment we write
Htotal = H0 (x, p) + xB + E (14.9)
where E = {En } is the bath Hamiltonian that can be written is some diagonal representation, while B = {Bnm }
represents that interaction term with x. However, it is more conventional to consider an interaction with an harmonic
bath:
H0 (x, p) + U(x, Qα ) + Hbath (Qα , Pα ) (14.10)
where (say)
1 2
H0 (x, p) = p + V (x) (14.11)
2M
X Pα 2

1

Hbath (Qα , Pα ) = + mα ωα2 Qα 2 (14.12)
α
2mα 2
and the interaction is (say)
X
UZCL = −x cα Qα (14.13)
α
X
UDLD = − cα Qα u(x−xα ) (14.14)
α
where the DLD possibility is appropriate for the motion of a particle in an extended environment.
x dissipation
fluctuations Q
driving source driven system
"slow" DoF "fast" DoF
"system" "environment"
105
V
V
(Q,P)
====== [14.3] Harmonic bath

It is common to model the environment as a huge collection of harmonic oscillators, and to say that the system if
subject to the fluctuations of a field variable F which is a linear combination of the bath coordinates:
1/2
X X 1
F = cα Qα = cα (aα + a†α ) (14.15)
α α
2mα ωα
For preparation of the bath in state n = {nα } we get

XX
S̃(ω) = c2α |hnα ±1|Qα |nα i|2 2πδ(ω ∓ ωα ) (14.16)
α ±
Using
1/2
√

1
hnα +1|Qα |nα i = 1 + nα (14.17)
2mα ωα
1/2
1 √
hnα −1|Qα |nα i = nα (14.18)
2mα ωα
we get
X 1 h i
S̃(ω) = 2πc2α (1+nα )δ(ω − ωα ) + nα δ(ω + ωα ) (14.19)
α
2mα ωα
For a canonical preparation of the bath we have hnα i = f (ωα ) ≡ 1/(eω/T − 1). It follows that
(
(1 + f (ω)) 1
S̃(ω) = 2J(|ω|) × = 2J(ω) (14.20)
f (−ω) 1 − e−βω
106
where we used f (−ω) = −(1 + f (ω)), and defined the spectral density of the bath as
π X c2α
J(ω) = δ(ω − ωα ) (14.21)
2 α m α ωα
with anti-symmetric continuation. For an Ohmic bath J(ω) = ηω, with some cutoff frequency ωc .
====== [14.4] Spin bath

We define F
X X
F= cα Qα = cα (aα + a†α ) (14.22)
α α
Thus Qα is the first Pauli matrix. Its non-trivial matrix elements are
√
hnα −1|Qα |nα i = n (14.23)
√ α
hnα +1|Qα |nα i = 1 − nα (14.24)
In complete analogy we get

X h i
S̃(ω) = 2πc2α (1−nα )δ(ω − ωα ) + nα δ(ω + ωα ) (14.25)
α
For canonical preparation hnα i = f (ωα ) where (from here on ~ = 1)
1
f (ω) = (14.26)
eβω + 1
1
f (−ω) = = 1 − f (ω) (14.27)
1 + e−βω
Thus we get
(
(1 − f (ω)) 1
S̃(ω) = 2J(|ω|) × = 2J(ω) (14.28)
f (−ω) 1 + e−βω
and
C̃(ω) = J(ω) (14.29)
where we define
X
J(ω) = π c2α δ(ω − ωα ) (14.30)
α
with symmetric continuation. For Ohmic bath J(ω) = ν, with some cutoff frequency ωc .
====== [14.5] Spatially extended environment

In this section we describe fluctuations of an extended environment in space and time using the form factor S̃(q, ω).
We define
h i
S̃(q, ω) = FT hU(x2 , t2 )U(x1 , t1 )i (14.31)
107
where the expectation value assumes that the bath is in a stationary state of its unperturbed Hamiltonian. The force-
force correlation function is obtained via differentiation. In particular the local power spectrum of the fluctuating
force is
Z
dq 2
S̃(ω) = q S(q, ω) (14.32)
2π
and the intensity of the fluctuations at a given point in space is

Z
dq 2
ν ≡ S̃(ω=0) = q S(q, ω=0) (14.33)
2π
For the one dimensional DLD bath we get
X
U = − cα Qα u(x−xα ) (14.34)
α
Taking into account that the oscillators are independent of each other we get
X
hU(x2 , t2 )U(x1 , t1 )i = c2α hQα (t2 )Qα (t1 )iu(x2 −xα )u(x1 −xα ) (14.35)
α
Z " #
X
= dx c2α hQα (t2 )Qα (t1 )iδ(x − xα ) u(x2 −x)u(x1 −x) (14.36)
α
Z
= u(x2 −x)u(x1 −x) dx S(t2 − t1 ) (14.37)
= w(x2 − x1 ) S(t2 − t1 ) (14.38)
Where we have assumed homogeneous distribution of the oscillators, and S(τ ) is defined implicitly by the above
equality. With the convention w00 (0) = −1 it is identified as the local force-force correlation function. Consequently
we get for the form factor
h i
S(q, ω) = FT hU(x2 , t2 )U(x1 , t1 )i = w̃(q) S(ω) (14.39)
As an example we may consider the following correlation function:

1 r 2
w(r) = `2 exp − (14.40)
2 `
If the spatial correlation distance is very large we get ZCL model:
1
w(r) = const − r2 (14.41)
2
leading to
2π
S(q, ω) = δ(q) S̃(ω) (14.42)
q2
This means that the force is homogeneous in space, and fluctuates only in time, which is effectively the case if a
particle or an atom interacts with long wavelength modes.
108
[15] Stochastic picture of the dynamics

There are various ”levels” in which the dynamics of a non-isolated system can be treated. We start with the random
walk problem that can describe the motion of a Brownian particle in the absence of friction. Then we discuss the
Langevin equation where friction is included. The dynamics in the above problem is described by a diffusion equation
and Fokker-Planck equation respectively. More generally we can talk about Master equations and in particular their
simplest stochastic version which is known as rate equations.
====== [15.1] Random walk and diffusion

Consider a particle that can hope from site to site in a stochastic manner. Each step can be represented by a random
number ft = ±a, where a is the lattice constant and t is the integer time index. The total displacement is
t
X
x(t) − x(0) = f (t0 ) (15.1)
t0 =0
Assuming a stationary stochastic process in which the correlation function is
hf (t1 )f (t2 )i = C(t1 − t2 ) (15.2)
we get that the variance is
0
t X
X t t
X +t
X t
X
Var[x] = hf (t1 )f (t2 )i = C(τ ) ≡ 2D(t0 ) −→ 2Dt (15.3)
t1 =0 t2 =0 t0 =0 τ =−t0 t0 =0
where the asymptotic value of the diffusion coefficient is
∞
1 X
D = C(τ ) (15.4)
2 τ =−∞
Most significant is to realize that there is a continuum limit of the random walk problem where the dynamics is
described by the following ”Langevin” equation of motion
Z
ẋ = f (t) ; x(t) − x(0) = f (t0 ) dt0 (15.5)
and accordingly
Z ∞
1 1
D = C(τ )dτ = C̃(ω=0) (15.6)
2 −∞ 2
There are various generalizations of the random walk problem, where the dwell time or the size of the steps are
random variables, leading in general to sub-diffusive or super diffusive behavior respectively. The latter case is known
as Levi-flight.
In the random walk problem the stochastic dynamics can be described by an equation for the time evolution of the
probabilities pn to find the particle in site n. This has the form of a ”rate equation”. In the continuum limit it
becomes the ”diffusion equation” for the probability density ρ(x) which we describe below. More generally this type
of ”master equation” is known as the ”Fokker Planck equation” which we discuss later on.
109
====== [15.2] The diffusion equation

It is natural to ask what is the ”master equation” that describes the time evolution of the probability density ρt (x) in
the case of a diffusion process. We assume that the stochastic equation of motion is ẋ = f (t) with stochastic f (t) that
has a zero average. A trivial generalization is to include a drift term such that the equation is ẋ = u + f (t), where
u is the so-called drift velocity. In order to derive the diffusion equation, note that for any particular realization of
f (t) the probability ρt+dt (xt+dt )dxt+dt must equal ρt (xt )dxt . Since the phase space element preserves its volume one
obtains the Liouville equation (d/dt)ρt (xt ) = 0, from which one deduces the continuity equation
∂ ∂ h i
ρt (x) = − (u + f (t))ρt (15.7)
∂t ∂x
From this equation it follows that ρt0 +dt can be expressed as an integral that involves ρt0 within t0 < t0 < t0 + dt.
The equation can be solved iteratively. In order to simplify notations we set without loss of generality t0 = 0 and
t = t0 + dt. Consequently we get an expansion that involves nested terms with higher order ∂/∂x derivatives of ρ0 .
For sake of clarity we drop the drift term and write
t t t0
∂ 2 ρ0
Z Z Z
0 0 ∂ρ0 0 0
ρt = ρ0 − dt f (t ) + dt f (t ) dt00 f (t00 ) + higher order terms (15.8)
0 ∂x 0 0 ∂x2
Averaging over realizations of f (), and neglecting the higher order terms, one obtains a diffusion equation, to which
we add back the drift term:
∂ ∂ρ ∂2ρ
ρ = −u +D 2 (15.9)
∂t ∂x ∂x
The neglect of higher order terms, say O(dt3 ) terms, is justified in the limit where the correlation time goes to
zero. This is sometimes known as the Markovian approximation. It is possible to regard the diffusion equation as a
combination of a continuity equation
∂ ∂
ρt (x) = − It (x) (15.10)
∂t ∂x
with expression to the current that includes a drift term and a diffusion term. The diffusion term is known as Fick’s
law. Fick’s law can be explained heuristically as reflecting a non-zero net net flow of particles across a section, due to
a difference of concentrations between its two sides. Ignoring the drift, if we have a sample of length L with a steady
state current then
D h i
I = − × ρ(L) − ρ(0) (15.11)
L
This means that there is a strict analogy here to Ohm law, implying that D is formally like the conductivity of the
chain, and accordingly can be obtained from a resistor network calculation. This observation is useful in analyzing
diffusion is non-homogeneous networks.
The drift velocity is typically related to a the gradient of an external potential, u = −µV 0 (x), with a coefficient which
is called mobility. Accordingly we write
∂ ∂V ∂ρ
I(x) = uρt (x) − D ρt (x) = −µρ −D (15.12)
∂x ∂x ∂x
If this expression is applied to a system in canonical equilibrium with ρ(x) ∝ exp(−βV (x)), it follows from the
requirement I(x) = 0 that µ = (1/T )D. This is called Einstein relation. It is useful in semiconductor physics.
For electrons
R in metal it is common to define the conductivity σ = µρ, and postulate that at equilibrium
ρ(x) = dEg(E − V (x))f (E − EF ). It follows that the Einstein relation for metals is σ = g(EF )D. Note that
g(EF ) is defined here as the density of one-particle states per unit volume, and it is proportional to ρ/EF .
110
====== [15.3] The Langevin equation

Consider a test particle subject to a homogeneous but fluctuating field of force F, leading to stochastic dynamics that
is described by the Langeving equation mẍ = F. It is convenient to isolate the average (= “friction”) term from F,
and accordingly to redefine F as a stochastic variable (= “noise”) that has zero average. Consequently the Langevin
equation is written as
mẍ = −η ẋ + F(t)
where F is a stochastic variable that satisfies hF(t)i = 0, and
hF(t2 )F(t1 )i = C(t2 − t1 )
It is assumed that C(τ ) has a short correlation time. We are interested in the dynamics over larger time scales (we
have no interest to resolve the dynamics over very short times). We also note that if F were a constant force, then
the particle would drift with velocity (1/η)F. The coefficient µ = 1/η is called mobility. The equation for the velocity
v = ẋ can be written as
d (η/m)t 1 (η/m)t
e v(t) = e F(t) (15.13)
dt m
leading to the solution
Z t
1 0
v(t) = dt0 e−(η/m)(t−t ) F(t0 ) (15.14)
m −∞
We see that τη = m/η is the damping time. After time τη the initial velocity is forgotten, hence the lower limit
of the integration can be extended to −∞. Evidently the average velocity is zero. We turn now to calculate the
velocity-velocity correlation. ”Squaring” and averaging over realizations we get
Z t1 Z t2
1 0 00
hv(t2 )v(t1 )i = 2 dt0 dt00 e−(η/m)(t1 +t2 −t −t ) C(t0 − t00 ) (15.15)
m −∞ −∞
We treat C(t0 − t00 ) like a delta function. Then it is not difficult to find that
Z ∞
1 −(η/m)|t2 −t1 | 1 ν
hv(t2 )v(t1 )i = e C(τ )dτ = e−|t2 −t1 |/τη (15.16)
2ηm −∞ m 2η
There is an optional shorter derivation of the latter result: In Fourier-space the Langevin equation is solved easily
vω = [−imω + η]−1 Fω , leading to C̃vv (ω) = [(mω)2 + η 2 ]−1 C̃(ω). With C̃(ω) = ν we get after FT the same result.
The correlation function hv(t2 )v(t1 )i for t1 = t2 = t should be consistent with h 21 mv 2 i = 12 T . From this one deduces
an FD relation ν/(2η) = T with regard to the response characteristics of the bath. The displacement x(t) − x(0) of
the particle is the integral over its velocity v(t0 ). On the average it is zero, but the second moment is
Z tZ t
ν
h(x(t) − x(0))2 i = dt0 dt00 hv(t00 )v(t0 )i = × t ≡ 2Dt (15.17)
0 0 η2
Hence we have diffusion in space. From the above we deduce the Einstein relation
D ν
= = Temperature (15.18)
µ 2η
The two results for D/µ, and for ν/η, can be regarded as special consequences of the general FD relation, as demon-
strated in a previous lecture.
111
====== [15.4] The Fokker-Planck equation

As in the case of a ”random walk” one can ask what is the ”master equation” that described the evolution of the
probability density ρ(x, p). This leads to the Fokker-Planck equation. The derivation is the same as in the case of a
diffusion process. Here the diffusion is in momentum with a coefficient ν/2. Including the v(p) = p/m drift in the
position, we get the continuity equation
∂ ∂ h i ∂ h ν ∂ρ i
ρ = − vρ − − V 0 (x)ρ − ηvρ − (15.19)
∂t ∂x ∂p 2 ∂p
There are quantum generalizations of the Fokker-Planck equation which we discuss in a separate section.
====== [15.5] Rate equations

A rate equation is merely a discrete version of the diffusion or Fokker-Planck equation. It can be regarded as describing
a generalized ”random walk” problem, where the transition rates Wnm are not necessarily equal in the n 7→ m and
m 7→ n directions. In the context of the ”system-bath” paradigm it is common to model the system as a set of
levels {En } with transition rates
β
wnm
ε
Wnm = wnm β
+ wnm , ε
wnm ε
= wmn , β
= e−(En −Em )/TB (15.20)
wmn
The first term describes the transitions that are induced by a TA =∞ driving source, as in the standard random walk
problem. The second term describes the bath induced transitions, with Boltzmann ratio for n ⇔ m transitions, as
required by detailed balance considerations.
One can regard the average value (Wnm + Wmn )/2 as the ”noise” which is introduced into the system by the bath,
while the difference Wnm − Wmn is the friction. However this point of view is strictly correct only for constant density
of states. If the level density grows with energy there will be a heating effect even if TB =∞.
The dynamics of the population probabilities p = {pn } is described by a rate equation
dpn X X
= [Wnm pm − Wmn pn ] ≡ −Γn pn + Wnm pm (15.21)
dt m m(6=n)
P equation can be written schematically as ṗ = Wp. There is an implicit

This Pconservation of probability requirement
pn = 1. The diagonal elements of W are −Γn , with the decay rates Γn = m Wmn . The steady state is determined
from the matrix equation Wp = 0. In the presence of driving the detailed balance is disturbed leading in general to
a non-canonical steady state.
Diffusion.– It is useful to note that a rate equation that describes motion in 1D, with transition rates
wnm = w(n − m) that depends on the hopping distance r = n − m, leads to diffusion with a coefficient
∞
1 X 2
D = r w(r) (15.22)
2 r=−∞
For near neighbor hoping between equally spaced sites D = wa2 , where a is the lattice constant. If we discretize a
diffusion equation by slicing the x axis into cells of width dx, the effective hopping rate w should be chosen such that
the diffusion coefficient is D = wdx2 .
Two level system.– There is an elegant way to write the general rate equation for a two level system. Defining
Sz = p+ − p− , and recalling that p+ + p− = 1, one realizes that the two equations for ṗ+ and ṗ− can be written as a
single equation (whose full quantum version will be discuss in the next lecture):
1
Ṡz = (w+− − w−+ ) − (w+− + w−+ )Sz ≡ − (Sz − Seq ) (15.23)
T1
112
====== [15.6] Rate equations - formalism

In the remaining subsections of this lecture we discuss the formal aspects of treating rate equations. The state of
the system is described by a column vector p whose entries are the occupation probabilities pn . The dynamics is
determined by the rate equation
dp
= W p, W = diagonal{−γn } + offdiagonal{wnm } (15.24)
dt
Probability is conserved hence W has the left eigenvector q 0 = {1, 1, ...} with eigenvalue λ0 = 0. The associated right
eigenvector p0 is the NESS. The other eigenvalues −λr of W might be complex, but the real part of λr has to be
positive. This follows from the observation that for t → ∞ only the NESS survives, while all the higher has to
diminish. The proof is based on the Perron-Frobenius theorem with regard to U (t) = exp(W t). If we have detailed
balance (see below) the λr have to be real and positive.
In general we can write the transition rates as follows:

Em;n
wnm = exp −Bnm + (15.25)
2
where B is a symmetric matrix while E is the anti-symmetric part. The latter can be decomposed into conservative
and solenoid components in a unique way:

wnm X
(α)
Em;n = ln = (Vm − Vn ) + α Am;n (15.26)
wmn α
Note that the “solenoid gauge” implies that A is uniform along the Cα loop. One can use a non-solenoid gauge, e.g.
to have it zero on one bond only. Anyway we fix its gauge and normalized its circulation as follows:
X
A(α)
x = 1 (15.27)
x∈Cα
We can define the height of the barrier relative to the potential reference level:
Vn + Vm
Bnm = Bnm + (15.28)
2
Then the expression for the transition rates takes the following form:
" #
1X
wnm = exp − (Bnm − Vm ) + α A(α)
m;n (15.29)
2 α
The detailed balance condition of having no circulations implies that there exist a diagonal matrix V such that
eV W e−V = W† (15.30)
Then we can perform a “gauge” transformation to a symmetric matrix
W̃ = eV /2 W e−V /2 , W̃ = W̃ † (15.31)
It follows that all the λr have non-negative real values. If we spoil the detailed-balance the matrix W̃ will becomes
parameterized by the affinities α, and the λr might become complex. Instead of the detailed balance condition we
h i†
†
get eV W α e−V = [W −α ] or equivalently W̃ α = W̃ −α .
113
====== [15.7] Rate equations - counting statistics

We add a pointer q with conjugate variable ϕ such that [q, ϕ] = i. Now the Hilbert space is spanned by |n, qi. The
dynamics of the joint probability distribution pn (q) is generated by
X X
Ŵ α = |n, q + Am;n i wnm hn, q| = wnm Jˆ(m;n) ⊗ e−iAm;n ϕ̂ (15.32)
n,m n,m
We change basis to |n, ϕi and use the Laplace transform convention iϕ 7→ ϕ, such that
X X
p̃n (ϕ) eiϕq pn (ϕ) eϕq pn (ϕ) = e−ϕq

pn (q) ≡ ≡ ⇐⇒ (15.33)
ϕ ϕ
The equation for pn (ϕ) involves the matrix
W α (ϕ) = wmn e−ϕAn;m ≡ W α−2ϕ (15.34)
We no longer have detailed balance but
†
eV W α (ϕ)e−V = [W α (α − ϕ)] (15.35)
The comulants are determined by the lowest eigenvalue λ0 (ϕ), leading to the NFT
g(ϕ) = g(α − ϕ) ; P (−q)/P (q) = exp(−αq) (15.36)
====== [15.8] Rate equations - ergodicity

It is convenient to define a weighted distribution qn = pn /p0n , such that qn =uniform once the NESS is reached. The
rate equation takes the form ṗ = Gq, where
X X
Gnm = Wnm p0m Gnm = Gnm = 0 (15.37)
n m
Accordingly
1X
q † Gq = − Gnm (qn − qm )2 (15.38)
2 nm
For an eigen-mode Gq r = λr pr and it is implied that λr r 2 0

P
n [(pn ) /pn ] > 0, leading to the conclusion that λr > 0.
In order to characterize the approach to steady state we pick a convex function f (x), for example f (x) = x ln(x) and
define an ergodicity measure
X X
H(t) = p0n f (qn (t)) = e.g. = pn ln(pn /p0n ) (15.39)
n n
Then we get
d X
H(t) = − Gnm [(f (qn ) − f (qm )) − (qn − qm )f 0 (qn )] < 0 (15.40)
dt n
114
[16] Quantum master equations
====== [16.1] General perspective

The description of the reduced dynamics of a system that is coupled to a bath using a Master equation is commonly
based on the following working hypothesis: (i) The bath is fully characterized by a single spectral function. (ii) There
is a way to justify the neglect of memory effects. The latter is known as the Markovian approximation. In particular it
follows that the initial preparation, whether it is factorized or not, is not an issue. If the master equation is regarded
as exact description of the reduced dynamics it should be of the Lindblad form. Otherwise is should be regarded
merely as an approximation.
There are two common approximation schemes: (A) In the Microscopic regime of atomic physics (e.g. two level atom)
it is assumed that the bath induced rates are much smaller than the level spacing, and a ”secular approximation” is
employed. (B) In the Mesoscopic regime of condense matter physics (e.g. Brownian motion) it is assumed that the
bath is Ohmic, and accordingly its effect can be treated as a generalization of ”white noise”.
====== [16.2] The general Lindblad form

A master equation for the time evolution of the system probability matrix is of Lindblad form if it can be written as
dρ X 1 X
= −i[H, ρ] + νr Lr ρL†r − [Γρ + ρΓ] , Γ= νr L†r Lr (16.1)
dt r
2 r
where Lr are called Lindblad generators, and νr are positive coefficients. Lindblad equation is the most general form
of a Markovian master equation for the probability matrix. The heart of the argument is as follows: The most general
linear relation between matrices is
X
ρ̃αβ = K(αβ|α0 β 0 ) ρα0 β 0 (16.2)
α0 β 0
This linear transformation can be regarded as “quantum operation” if it preserves the hermiticity and the positivity of
ρ. See [arXiv (lecture 53)] for details. Changing notation to K(αβ|α0 β 0 ) = Kαα0 ,ββ 0 one observes that Kαα0 ,ββ 0 should
be hermitian, with non-negative eigenvalues λr . Accordingly we can find a spectral decomposition with P a transforma-
tion matrix T (αα0 |r). Changing notation to Kα,α
r 0
0 = T (αα |r) we get the Kraus representation ρ̃ =
r r †
r λr [K ]ρ[K ] .
r † r
P
Conservation of probability implies r λr [K ] [K ] = 1. Looking on the incremental change of ρ during a small time
interval dt, one obtains the Lindblad form of the Master equation.
Detailed derivation of Lindblad form from the Kraus representation can be found in [arXiv]. For completeness we
present here a brief outline. We note that for zero time evolution we have as single non-zero eigenvalue λ(0) = N
which is associated with the normalized identity matrix T (αα0 |0) = N −1/2 δα,α0 ≡ L0 , where N is the dimension of
Hilbert space. For a small time step we substitute λ(r6=0) ≡ νr dt and T (αα0 |r) ≡ Lr . Note that the νr are positive,
and that the L(r6=0) are traceless due to the orthogonality with L0 . Accordingly
dρ X
= νr Lr ρL†r + ... (16.3)
dt
r6=0
The remaining terms in the Lindblad form are the H-term and the Γ-term, that are related to the r = 0 term, and can
be straightforwardly deduced. However we prefer to point out an indirect approach. From the sameP argument as for
quantum operations it is clear that any linear expression that preserves hermiticity can be written as r νr [Lr ]ρ[Lr ]† .
We transform to some general orthonormal basis where one of the operators is F (0) = N −1/2 1, while by orthonormality
all the other F r are traceless. We get
0
dρ X 1 X
= ν̃r,s [F r ]ρ[F s ]† = −i[H, ρ] − {Γ, ρ} + ν̃r,s [F r ]ρ[F s ]† (16.4)
dt r,s
2 r,s
115
P0
where the last summation excludes the ”0” terms. The definition of the Hamiltonian H ∝ i r ([F r ] − [F r ]† ) is
implied. Also the Γ-term is implied, and optionally
P0 can be deduced from the requirement of obtaining a trace-
preserving map. It is now argued that the sum r,s should coincide, upon diagonalization, with the summation over
the Lr terms. Hence the ν̃r,s must have positive eigenvalues νr .
====== [16.3] Quantum Fokker-Planck equation

Consider the the classical Langevin equation. Using canonical phase-space coordinates it reads
ṗ = −V 0 (x) − ηv + f (t), where v = ẋ = p/m is the velocity, and f (t) is white noise that has intensity ν. The
corresponding master equation for ρ(x, p) is the Fokker-Planck equation:
dρ ∂ h i ∂ h ν ∂ρ i
=− pρ − − V 0 (x)ρ − ηvρ − (16.5)
dt ∂x ∂p 2 ∂p
This equation can be written with Poisson Brackets, which are replaced in the quantum context by commutators:
dρ ν η
= −i[H, ρ] − [x, [x, ρ]] − i [x, {v, ρ}] (16.6)
dt 2 2
We shall discuss later the general procedure to derived this master equation from an Hamiltonian, where the interaction
with the bath is via the system operator W = x. The same procedure can be uses for any W , leading to
dρ ν η νη
= −i[H, ρ] − [W, [W, ρ]] − i [W, {V, ρ}] − [V, [V, ρ]] (16.7)
dt 2 2 2
where v has been replaced by V = i[H, W ], and where νη = 0. This Ohmic master equation does not have the
Lindblad form (see below), and hence in general complete positivity is not guaranteed. For example: if we consider
the relaxation of a wavepacket in damped harmonic oscillator, then at low temperatures we end up with a sub-minimal
wavepacket that violates the uncertainty relation.
The Ohmic master equation involves the bilinear form r,s ν̃r,s [F r ]ρ[F s ]† with F (1) = W , and F (2) = V , and
P
!
ν −i η2
ν̃r,s = (16.8)
i η2 νη
In order for this equation to be Lindblad, the matrix ν̃r,s should be positive. The minimal modification would be to
set a non-zero νη = η 2 /(4ν). With this substitution, after diagonalization, one ends up with a single Lindbald term
with the generator
η
L = W +i V (16.9)
2ν
Note that the pre-factors of the three terms in the modified Ohmic version are ν/2 and ν/(2T ) and ν/(32T 2 ) respec-
tively. These terms can be regarded as arsing from an expansion in powers of (Ω/T ), where Ω is the frequency of the
motion. Accordingly in the high temperature regime the deviation of the standard Fokker-Planck equation from the
Lindblad form is negligible.
We can also go in reverse and provide an “Ohmic interpretation” for each term in the Linblad form. Namely, consider
1 1
Lindblad = LρL† − L† Lρ − ρL† L (16.10)
2 2
116
Writing
L = A + iB (16.11)
C = i[A, B] (16.12)
D = (1/2){A, B} (16.13)
and using the identity
1 i
[A, {B, ρ}] = [D, ρ] − {C, ρ} + AρB − BρA (16.14)
2 2
we get the following optional expressions for the Lindblad term:
1
Lindblad = −i[D, ρ] − {A2 + B 2 , ρ} + AρA + BρB − i[A, {B, ρ}] (16.15)
2
1 1
= −i[D, ρ] − [A, [A, ρ]] − [B, [B, ρ]] − i[A, {B, ρ}] (16.16)
2 2
The first term represents “Lamb shift”, the second and the third are ”noise” induced diffusion terms, and the last is
the “friction” term.
====== [16.4] System-bath interaction

In the following presentation we assume that the full Hamiltonian is
Htotal = H − W F + Hbath (16.17)
where W and F are system and bath operators respectively. Neglecting the interaction, the bath is characterized by
the spectral function
h i
C̃(ω) = FT hF (t)F (0)i (16.18)
and the convention hF (t)i = 0. Whether the bath is composed of harmonic oscillators or not is regarded by the
working hypothesis as not important. There is a well known discussion of this point in Feynman-Vernon paper.
The spectral function C̃(ω) is characterized by temperature T and by a cutoff frequency ωc . The latter is assumed
below to be large compared with any other temporal frequency. What we call ”noise” means C̃(−ω) = C̃(ω). What
we call ”finite temperature” means
C̃(−ω)/C̃(ω) = exp(−ω/T ) (16.19)
What we call ”white noise” or ”infinite temperature Ohmic bath” corresponds to C̃(ω) = ν, leading to
C(t) = hF (t)F (0)i = νδ(t) (16.20)
What we call “high temperature Ohmic bath” takes into account that C̃(ω) possesses an antisymmetric component
that at small frequencies can be approximated as ν × [ω/(2T )]. Consequently
C(t) = hF (t)F (0)i = νδ(t) + iηδ 0 (t) (16.21)
where η = ν/(2T ) is the so called friction coefficient.

117
For a general bath, non-necessarily Ohmic, it is useful to define a bath spectral function via a Fourier-Laplace transform
Z ∞
1
G(ω) ≡ C(t)e−iωt dt ≡ C̃(ω) − i∆(ω) (16.22)
0 2
It is also useful to look on W in the interaction picture:
X X
W (t) = eiHt W e−iHt = |niWnm ei(En −Em )t hm| ≡ e−iΩt WΩ (16.23)
n,m Ω
P
We can say that the unperturbed system Hamiltonian
P H induces spectral decomposition W = Ω WΩ of the in-
teraction. For non-degenerated spectrum W0 = n |ni hn| is the diagonal part of the W -matrix in the energy rep-
resentation, and each WΩ† = W−Ω with Ω 6= 0 corresponds to a pair of coupled levels. Additionally it is useful to
define
Z ∞ X
W̃ ≡ C(t)W (−t)dt = G(Ω)WΩ (16.24)
0 Ω
Coming back to the Ohmic case, it is useful to define a “velocity” operator V = i[H, W ]. Accordingly, in the Ohmic
case, we get
ν η
W̃ ≈ W +i V (16.25)
2 ν
The notations above are useful for the purpose of writing down the Master equation for the time evolution of the
reduced probability matrix.
====== [16.5] Derivation of the master equation

We first demonstrate the derivation of the Master equation in the case of white noise. The Hamiltonian is
H(t) = H + f (t)W , were f (t) represents white noise: that means that upon ensemble average hf (t)i = 0, while
hf (t)f (t0 )i = νδ(t − t0 ). Given ρ(t0 ) ≡ ρ, the Liouville von-Neumann equation can be solved iteratively to deter-
mine ρ(t0 +dt), where dt is a small time interval. Without loss of generality we set t0 =0 and t = t0 +dt and get:
Z t Z tZ t0
0 0
ρ(t) = ρ − i dt [H(t ), ρ] − dt0 dt00 [H(t0 ), [H(t00 ), ρ]] + ... (16.26)
0 0 0
Averaging over realizations of f (t) all the odd orders in this expansion vanish, while the leading dt contribution comes
only from the zero order term that involves H and from the second order term that involves W . Consequently we get
the following Master equation:
dρ 1 1
= −i[H, ρ] − ν[W, [W, ρ]] = −i[H, ρ] − {Γ, ρ} + νW ρW (16.27)
dt 2 2
where Γ = νW W . Note that the first two terms in the second expression generate so called non-Hermitian dynamics
with the effective Hamiltonian Heff = H − (i/2)Γ, while the last term represents “continuous measurement”.
The generalization of the ”white noise” derivation for a system that is coupled to a high temperature Ohmic bath
is straightforward. It is based on the assumption that at any moment the system-bath state is ”factorized”, which
can be justifies if ωc−1 is sufficiently small. We define the interactions-representation of the probability matrix via
ρ(t) ≡ U (t)ρ̃(t)U (−t) where U (t) = e−iHt . The iterative procedures provides for ρ̃(t) the same expansion as in the
118
previous subsection with H(t) replaced with F (t)W (t), where W (t) = U (−t)W U (t). Consequently we get the so-called
Redfield equation
dρ
= −i[H, ρ] + W̃ ρW + W ρW̃ † − W W̃ ρ − ρW̃ † W (16.28)
dt
Note that an optional style of writing this expression is with −[W, W̃ ρ − ρW̃ † ], which reduces to −[W, [W̃ , ρ]] for an
hermitian W̃ . If instead one substitutes the non-hermitian Ohmic expression for W̃ , one obtains the Ohmic master
equation, that contains both noise and friction terms. As noted above, one can add to the Ohmic master equation a
term [V, [V, ρ]] that represents an extra white noise coupled via V .
====== [16.6] The Pauli master equation

We come back one step, and consider again general bath, not necessarily Ohmic. Instead of assuming small correlation
time, we shall assume weak interaction. Specifically, in atomic physics applications the induced rate of transitions w
becomes much smaller compared with the Rabi-Bloch frequency Ω of the coherent oscillations. Accordingly it is
appropriate to write that master equation in the interaction picture:
dρ̃
= W̃ (t)ρ̃W (t) + W (t)ρ̃W̃ (t)† − W (t)W̃ (t)ρ̃ − ρ̃W̃ (t)† W (t) (16.29)
dt
Substitution of the H-induced spectral decomposition of the W -s one observes terms that oscillate with frequencies
Ω + Ω0 . We keep only the terms that oscillate with ∼ 0 frequency, and hence do not average to zero. For example, in
W̃ ρW we keep only the G(Ω)WΩ ρW−Ω terms. Consequently we obtain the so called secular approximation
dρ Xh i
= −i[H, ρ] + C̃(Ω) WΩ ρWΩ† − G(Ω) WΩ† WΩ ρ − G(−Ω) ρWΩ† WΩ (16.30)
dt
Ω
The imaginary part of G(Ω), aka Lamb shift, can be absorbed into the Hamiltonian H, so we end up with a simple
sum over Lindblad terms that are weighted by the spectral intensities C̃(Ω) at the Ω 6= 0 frequencies of the inter-level
transitions, and additionally an Ω = 0 term that is weighted by the spectral intensity ν = C̃(0).
We now write the above equation in the H basis, and realize that the dynamics of the diagonal elements decouples
from that of the off-diagonal elements. Namely, the first term in the secular approximation induces FGR transitions
with rates
wnm = C̃(−(En −Em )) |Wnm |2 (16.31)

P0
The corresponding decay constants are Γn = m wnm . The ratio wnm /wmn is not unity unless we consider white
noise source (infinite temperature). For finite temperatures the FGR rates favor downwards transitions. Consequently
we get the so-called Pauli rate equation for the probabilities pn
 
−Γ1 w12 ...
dp
= W p, W =  w21 −Γ2 ... (16.32)
 
dt
... ... ...
For the off-diagonal terms we get
dρnm h i
= − i(En − Em ) − γnm ρnm , [for n 6= m] (16.33)
dt
with dephasing rates
ν 1
γnm = |Wnn − Wmm |2 + (Γn + Γm ) ≡ γϕ + γrlx (16.34)
2 2
119
where the first term originates from the Ω = 0 generator, while the second term originates from the Ω 6= 0 transitions.
We note that the above results can by derived from heuristic consideration, without going through the heavy machinery
of the master equation formalism. Taking the white noise master equation as a starting point, it is enough to realize
that the elements of ρnm can be classified according to their unperturbed frequencies (En − Em ). Elements that are
oscillating with different frequencies, have a negligible cross interaction. In particular the dynamics of the pn , that
have ∼ 0 frequencies, decouple from the dynamics of the off-diagonal elements, leading to FGR picture of transitions.
For the off diagonal terms the reasoning is similar, and there is an additional dephasing γϕ due to the noisy detuning.
The finite temperature case is merely a variation on the same reasoning.
====== [16.7] The Bloch equation

The time dependence of an expectation values is given by the adjoint equation:

d d
hQi = trace Q ρ = trace [QLρ] = trace (L† Q)ρ = L† Q

(16.35)
dt dt
If the master equation is written in the Lindblad form, the expression for L† Q is the same as Lρ with H 7→ −H.
Let us consider a two level system as an example. The probability matrix is conveniently expressed using the Bloch
~ = (Sx , Sy , Sz ), were Sj = hσj i, namely,
vector S
1
ρ(t) = 1 + Sx σx + Sy σy + Sz σz (16.36)
2
Note that Sz = p+ − p− is the population probability difference, while Sx and Sy are the so called ”coherences”.
Using the adjoint equation one can easily show that the equation of motion for the Bloch vector takes the form
dS ~ × S − γ (S − Seq )
= −Ω (16.37)
dt
where S is regarded as a column vector, and γ = diag(γx , γy , γz ) is a diagonal matrix. The first term is generated by
the unperturbed Hamiltonian: We assume H = −(Ω/2)σz , hence Ω ~ = (0, 0, Ω).
We consider the effect of having a coupling term −W F (t), where F (t) represents a bath or a noise source. Even
without going through the master equation formalism it is clear that consistency with the canonical formalism implies
that the equilibrium states is

Ω
Seq = 0, 0, tanh (16.38)
2T
We now refer separately to different versions of the Bloch equation. The different versions are distinguished by the
assumptions regarding W , the intensity ν of the F (t) fluctuations, and their spectral characteristics.
Pure dephasing.– The simplest possibility is to have a so-called pure dephasing effect due to a W = σz inter-
action with a white noise source that has an intensity νϕ . In the master equation it introduces a diffusion term
(νϕ /2)[W, [W, ρ]]. The implication is to have in the Bloch equation
γ[Dephasing] = diag (2νϕ , 2νϕ , 0) (16.39)
The interaction with the noise commutes with H therefore there is no equilibration in the Sz direction. For this reason
if we replace the noise source by a finite temperature bath, it will have a similar effect.
Ohmic version.– Next in complexity is to consider a high temperature Ohmic bath coupled via W = σx . Using the
notations of the previous sections we have here a ”position” coordinate W = σx and a conjugate ”velocity” coordinate
V = Ωσy . Consequently, after some straightforward algebra we deduce that
γ[Ohmic] = diag (0, 2ν, 2ν) (16.40)

120
Due to the lack of commutation we have an additional ”friction” term 2ηΩ in the master equation for dSz /dt, which
implies Seq = (0, 0, ηΩ/ν). This is consistent with the canonical expectation, upon the substitution ν/η = 2T , provided
the condition (Ω/T ) 1 is satisfied. This is the regime where the high temperature Ohmic approximation is valid.
The dephasing in the Ohmic version of the Bloch equation is non-isotropic in the transverse XY plane. Note that the
Sy transverse component satisfy the equation
S̈y + 2ν Ṡy + Ω2 Sy = 0 (16.41)
√
which leads to damped frequency Ωeff = Ω2 − ν 2 . In the secular and NMR versions that we discuss in the next
paragrpahs the dephasing is isotropic in the XY plane and therefore Ω is not affected.
Secular version.– We now consider what comes out, for the same coupling, within the framework of the secular
approximation. Note that this approximation, unlike the high temperature Ohmic version, assumes large Ω. Using
the Pauli equation prescription we realize that the FGR average transition rate is ν. Hence we get
γ[Secular] = diag (ν, ν, 2ν) (16.42)
One observes that due to the perturbative nature of this approximation the transverse relaxation looks isotropic.
Disregarding this artifact, one should keep in mind that the secular approximation allows to consider the case of non-
Ohmic bath. From the general derivation it should be realized that ν in the above equation is determined exclusively
by the C̃(±Ω) components of the fluctuations.
NMR version.– The so called nuclear-magnetic-resonance version of the Bloch equation consider a general W .
Formally it is like to add to the secular version of the previous paragraph an additional pure dephasing effect.
Accordingly we write the Bloch equation as
dSz 1
= − (Sz − Seq ) (16.43)
dt T1
dSx,y 1
= −[Ω × S]x,y − Sx,y (16.44)
dt T2
where the equilibrium value is
w+− − w−+
Seq = (16.45)
w+− + w−+
as in the secular version. The rates for the diagonal relaxation and for the off-diagonal transverse depahsing are:
1
= w+− + w−+ ≡ 2γrlx (16.46)
T1
1
= γrlx + γϕ ≡ γ (16.47)
T2
The pure dephasing rate γϕ originates from the diagonal elements of Wnm and hence is formally proportional to
the intensity C̃(0), while the FGR transition rates originate from the off-diagonal elements of Wnm , and hence are
proportional to C̃(±Ω), were Ω = |E+ − E− | is the level spacing.
121
Additional topics
[17] The kinetic picture
====== [17.1] The Boltzmann distribution function

The number of one particle states within a phase space volume is dN = d3 rd3 p/(2π~)3 . The occupation of this phase
space volume is:
drdp
dN ≡ f (r, p) (17.1)
(2π~)3
where f (r, p) is called Boltzmann distribution function. In equilibrium we have

f (r, p) = fβ (p − µ) (17.2)

eq
where fβ ( − µ) is either the Bose or the Fermi occupation function, or possibly their Boltzmann approximation. If
we use (r, v) with measure d3 rd3 v instead of (r, p) we have
m 3
f (r, v) = f (r, p) (17.3)
2π
By integrating over r and over all directions we get the velocity distribution
m 3
1 2
F (v) = L3 × 4πv 2 fβ mv − µ (17.4)
2π 2
If we use Boltzmann approximation for the occupation function and express µ using N and T we get
m 3/2 1 2
F (v) = N 4πv 2 e− 2 mv /T (17.5)
2πT
We note that
x drdp Z Z
N = f (r, p) = dg()f β ( − µ) = F (v)dv (17.6)
(2π~)3
====== [17.2] The Boltzmann equation

The Liouville equation for ρ(r, p) is merely a continuity equation in phase space. It can be written as (d/dt)ρ = 0
where d/dt unlike ∂/∂t is the total derivative reflecting the change in the occupation of a phase space cell. The
Boltzmann equation for f (r, p) is formally identical to the Liouville equation in the absence of collisions, and with
collisions becomes (d/dt)ρ = g(r, p), where g(r, p) is the net rate in which particles are generated at (r, p) due to
collisions. Accordingly the Boltzmann equation is
h∂ ∂ ∂ i
+ v(p) · + F(r) · f (r, p) = g[f ] (r, p) (17.7)
∂t ∂x ∂p
where v(p) = p/m is the dispersion relation, and F(r) = −V 0 (r) is the force due to some external potential. The
notation emphasizes that g(r, p) is a functional of the f distribution. It can be written as a difference of ingoing
122
and outgoing fluxes due to collisions. A distribution that gives zero in the left hand side of the Boltzmann equation
is called ”ergodic”. A distribution that gives g = 0 at any point is ”locally equilibrated”. If a locally equilibrated
distribution is ergodic it constitutes an equilibrium solution of the Boltzmann equation. If there is no such solution,
one can look for a non-equilibrium steady-state (NESS) solution. For example, assume that g is different in two
regions of space reflecting the presence baths with different temperatures; In such case there exists a non-ergodic
NESS solution that features a non-zero heat transport through the system.
The standard Boltzmann expression for g(r, p) is based on 2body collision mechanism and ”molecular chaos” assump-
tion. See chapter 4 of Huang or chapter 14 of Reif. Assuming that collisions from (p, p0 ) to (p1 , p2 ) has the same
rate as that of the inverse process, it takes the following form:
Z Z
dp0 dp1 dp2
g[f ] (r, p) = w(p, p0 |p1 , p2 ) [f (r, p2 )f (r, p1 ) − f (r, p0 )f (r, p)] (17.8)
2π 2π 2π
The gas reaches a steady state in accordance with the Boltzmann H theorem. The formal solution for the steady state
implies the Maxwell-Boltzmann distribution for the velocities. A much simpler expression for g(r, p) appears while
discussing electronic transport. See chapters 16 and 13 of Ashcroft & Mermin. Here the scattering mechanics is
1body collisions of the electrons with the lattice, leading to
dp0
Z
g[f ] (r, p) = [wp,p0 f (r, p0 ) − wp0 ,p f (r, p)] (17.9)
2π
Note that the first term, that corresponds to electrons that are scattered out of the phase-space cell can be written
as −(1/τ )f where (1/τ ) ∼ N w is the decay rate to the other N cells to which it is connected. If we assume that f
is close to an equilibrium solution fβ , it follows that we can approximate f ≈ fβ in the ingoing flux term. Hence we
get the so called relaxation time approximation:
fβ ((p) − µ) − f (r, p)
g[f ] (r, p) ≈ (17.10)
τ
NESS is reached if β(r) or µ(r) or the potential V (r) are non-uniform in space. Using the relaxation time approxi-
mation the solution that can be written schematically as follows:
t
t − t0
Z
1
f (t) = exp − fβ (t0 ) dt0 (17.11)
−∞ τ τ
where fβ (t0 ) stands for fβ that is evaluated at the point (r(t0 ), p(t0 )), which is connected by an unscattered classical
trajectory that ends at (r, p) at time t. Assuming τ to be small one can easily obtain a first order solution. For
example, in the presence of a constant field of force the NESS becomes

∂fβ
f (r, p) ≈ fβ − τ Fv(p) − (17.12)
∂
where fβ ((p)) is the the equilibrium occupation function that is calculated in the absence of the field. The above
NESS carries current whose density can be written as J = σF, where
d3 p
Z Z
2 ∂fβ g() 2 0 g(F ) 2
σ = τ v − = τ v −fβ ( − F ) = v τ (17.13)
(2π) 3 ∂ V V F
The last equalities assume a zero temperature Fermi occupation. If g() corresponds to the standard dispersion
relation, one obtains the Drude formula σ = (N/V)(τ /m).
123
====== [17.3] The calculation of incident flux

Given N gas particles that all have velocity v we can calculate the number of particles that hit a wall element per unit
time (=flux), and also we can calculate the momentum transfer per unit time (=force). Using spherical coordinates,
such that normal incidence is θ=0, one obtains (per unit area):
x
dΩ N
Z 1
1

N

1 N
J = v cos(θ) = cos(θ)d cos(θ) v = v (17.14)
4π V 2 0 V 4 V
|θ|<π/2
x dΩ N Z 1
N 1

N

2
P = v cos(θ) 2mv cos(θ) = cos (θ)d cos(θ) mv 2 = mv 2 (17.15)
4π V 0 V 3 V
|θ|<π/2
If we have the distribution F (v) of the velocities, or optionally if we are provided with the one-particle energy
distribution, the total flux is given by an integral:
Z ∞ Z ∞
1 F (v)dv 1 g()f ()d
Jincident = v = v (17.16)
0 4 V 0 4 V
Similar expression holds for the pressure P , where one can make the identification = (1/2)mv2 , and recover the
familiar Grand canonical result.
====== [17.4] Blackbody radiation

The modes of the electromagnetic field are labeled by the wavenumber k and the polarization α. For historical reasons
we use k instead of p for the momentum and ω instead of for the energy. The dispersion relation is linear ω = c|k|.
The density of modes is
V 2
g(ω)dω = 2 × 3 4πω dω (17.17)
(2πc)
Recall that the canonical state of oscillators can be formally regarded as the grand canonical equilibrium of µ = 0
Bose particles, with the occupation function
1
hnkα i = ≡ f (ωkα ) (17.18)
eβωkα −1
For the total energy E we have
∞ ∞
ω3
Z Z
1
E = ωdω g(ω) f (ω) = V dω 2 3 (17.19)
0 0 π c eβω − 1
For the total number of photons N we have a similar integral but without the ω. The calculation of the incident flux
of photons is the same as in the case of non-relativistic particles with v 7→ c. Accordingly we get J = (1/4)(N/V)c.
For the flux of energy we just have to replace N by E, namely
∞
ω3
Z
1 E 1
Jincident [energy] = c = dω (17.20)
4 V 0 4π 2 c2 βω
e −1
The calculation of the pressure (the rate of momentum transfer) is again the same as in the case of non-relativistic
particles, with mv 7→ (ω/c). Accordingly we get for the radiation pressure P = (1/3)(E/V). Note that P 7→ (1/2)P
for an absorbing surface.
124
Considering the thermal equilibrium between blackbody radiation from the environment, and an object that has an
absorption coefficient a(ω), detailed balance consideration implies that
Jemitted (ω) dω = a(ω)Jincident (ω) dω (17.21)
It follows that we can regard a(ω) as the emissivity of the object. From here we get the Planck formula
ω3 ν3

1 a(ω) 3
Jemitted (ω) = a(ω) 2 2 = T (17.22)
4π c eβω − 1 4π 2 c2 eν − 1
where ν = ω/T is the scaled frequency. Note that the peak of a blackbody radiation is at ν ≈ 3 which is known as
Wein’s law. Upon integration the total blackbody radiation is
∞
π4
Z
1
Jemitted = Jemitted (ω)dω = T4 (17.23)
0 4π 2 c2 14
which is know as Stephan-Boltzmann Law. Note that the units of the flux are energy per time per unit area.
Je (w)
====== [17.5] Viscosity

We have considered above the rate in which momentum is transferred to a wall due to ballistic collisions, leading
to pressure. There is a somewhat related effect that is called ”viscosity”. It is simplest to explain the concept with
regard to a gas whose particles have a short mean free path `, such that in equilibrium each gas particle has a diffusion
coefficient D = vT `, where vT is the mean thermal velocity.
Assume out of equilibrium steady state in which the average velocity of the gas particles v(x, y, z) is in the x direction
but its magnitude varies in the y direction. Due to the transverse diffusion there will be momentum transfer across
the y = 0 plane, which implies that the ”upper” flow exerts a force on the ”lower” region of the fluid (which is possibly
the boundary layer of some ”wall”). We shall explain below that if the area of the boundary region is A, then the
force on it is given by the ”stress-shear equation”
dvx
Fx = µA (17.24)
dy
where µ = ρD is the viscosity, and ρ is the mass density of the gas. The argument goes as follows: Divide the y axis
into layers of width dy. Define w, the effective transition rate of particles between layers, such that D = wdy 2 . Define
the flow of momentum between layers as J(y) = [(ρdy)vx (y)] × w. Hence J(dy/2) − J(−dy/2) is the rate in which
momentum is transferred across y = 0, leading to the desired result.
125
====== [17.6] The Navier-Stokes equation

The variables that describe the flow of gas or liquid are the density ρ(x) and the velocity u(x). More generally we
can add also temperature θ(x), and heat flow q(x). The conservation laws are for the mass, for the momentum, and
for the energy. The conservation of the momentum corresponds to the second law of Newton, and formally can be
written as [Huang section 5.3]:
! " #
duj ∂ X ∂ 1 X ∂
≡ + ui uj = fj − Pij (17.25)
dt ∂t i
∂xi ρ i
∂xi
Following the presentation as in Huang one obtains the Navier-Stokes equation [Wiki] that describes the rate of
change of the velocity due to momentum transfer in a vicious fluid:
du µ
ρ = f − ∇ P − ∇ · u + µ∇2 u (17.26)
dt 3
This equation is valid in the hydrodynamics regime, where the mean free path is small compared with to the geometrical
length scales. It includes a scalar-pressure term, a viscosity term, and an optional external force term (say gravitation).
For incompressible flow ∇ · u = 0, and Euler equation is obtained if the viscosity is completely neglected.
Sound waves.– As we see the viscosity µ plays a major role in the Navier-Stokes equation. Usually the equation is
supplemented by a continuity equations for the mass, and for the energy, as well as by state equation that connects
the pressure to the density. For compressible fluid with state equation P [ρ] the continuity and the Euler equations
for the time derivatives of ρ(x) and u(x) lead to sound waves with velocity
∂P κ
c2 = = (17.27)
∂ρ ρ
Once the viscosity is taken into account we get damping of the sound waves.
Stokes law.– A well known result that comes from the Navier-Stokes equation is Stokes law for the friction force
that is exerted on a spherical object of radius R
F [stick] = −6πµR vsphere , F [slip] = −4πµR vsphere (17.28)
Roughly the shear is 1/R while the area is R2 hence the friction is proportional to R. and not to the area of the
sphere. The traditional version (with 6π) assumes no-slip boundary conditions. The way to derive it is to find the
velocity field for the flow, and then to use the “stress-shear equation” of the previous section. For details see Huang
p.119 and [PRA 1970]. The optional derivation via a microscopic theory is quite complicated, see [JCP 1980].
Reynolds number.– The dimensionless parameter in the Navier-Stokes equation is the Reynolds number
1 µ
= (17.29)
Re ρu Length
If we have a wall that moves with some velocity v in a transverse distance L from another parallel wall the ”simple”
solution apparently becomes unstable for large ”Re”. One way to understand the instability is to consider the transient
√
solution for moving cube of side L. There is a boundary region vt ∼ L in the movement direction and hence Dt
in the transverse direction. Once the width of the boundary region becomes like L the instability develops. Stokes
law is valid for small ”Re”, meaning small translation velocity v. For large ”Re” we cannot ignore the boundary
region. Once the boundary becomes of order of the radius R, it indicates that the flow becomes turbulent, and the
v dependence acquire a fractional exponent. If the translation velocity becomes very large the the friction would
become proportional to v 2 reflecting transfer of momentum by a moving wall. This should be contrasted by Stoke
law where the ”pushed” particles flow to the sides of the sphere and join to the back of the sphere.
126
====== [17.7] Heat current in an open geometry

If we have two boxes, and energy can flow from one to the other, then the heat current IQ is simply defined as the
rate IE of energy transfer. But if we have a flow IN of particles, this simple-minded reasoning fails. We would like to
argue below the the correct expression for the heat current is
IQ = IE − µIN (17.30)
This expression assumes quasi-reversible flow at well-defined temperature and energy.
First of all let us recognize that if dN particles are transferred from one box to a second box, then the transferred
energy dE is ill-defined. Assume for example that the particles are transferred from an energy level of the first
box to an energy level with the same energy at the second box. We get dE = dN , which depends on the arbitrary
energy reference of the Hamiltonian, and hence has no physical significance. But the quantity dQ = dE − µdN is well
defined. The question is how to rationalize that dQ is indeed the appropriate definition of heat in this context.
Referring to a box with N particles and energy E, recall that a quasi-reversible process of taking an energy −dE
requires the supply of energy dQ = dE, such as dS = 0. In complete analogy, taking −dN particles with energy −dE
requires compensation dQ = dE − µdN , such as dS = 0.
====== [17.8] Thermo-electricity

Reversible flow through a conductor, with no entropy production due to Joule heating, can be regarded as a sequence
of quasi-reversible transfer operations. In each step heat is taken or given to the phonons that dwell in another
segment of the conductor. The net results is the transfer of energy from one end of the wire to the other end. This
reasoning leads to the Mott analysis of the Peltier effect. The expression for the electric current of charge e carriers
has the form (Here J = I/A is the current density):
Z
JN = e2 E c(ε) (−f 0 (ε − µ)) dε ≡ σE (17.31)
while the expression for the quasi-reversible heat current is

Z
JQ = eE (ε − µ) c(ε) (−f 0 (ε − µ)) dε ≡ σST E (17.32)
Using the Sommerfeld expansion we deduce that
π 2 2 c0 (µ)
IQ = T IN ≡ ST IN (17.33)
3e c(µ)
where Π = ST is known as the Peltier coefficient, and S is known as the Seebeck coefficient. From the Onsager
reciprocity it follows that the same coefficient appears in the linear relation between JN and ∇T , as explained below.
A thermal current can be induced also by a temperature gradient. Namely,
JQ = −κ∇T (17.34)
A relation between the thermal conductivity κ and the electrical conductivity σ can be obtained using a straightforward
extension of the above derivation (see Ashcroft p.253):
π2
κ = Tσ (17.35)
3e2
127
This is known as the Wiedemann-Franz law. We now turn to discuss more general circumstances of having both
potential and temperature gradients.
Considering again two boxes, the natural thermodynamic coordinates are ϕN = (N2 − N1 )/2 and ϕE = (E2 − E1 )/2.
Note that JN = ϕ̇N and JE = ϕ̇E . In the continuum limit J is re-defined as the current density. The entropy function
is S(XN , XE ), and the conjugate variables are XN = ∇(µ/T ) and XE = −∇(1/T ). Here ∇ is the difference, while in
the continuum limit it becomes the gradient. The linear relation between the Js and the Xs involves coefficients γij .
The Onsager relation is invariant under the change of reference energy, but the coefficients depend on the choice of
reference. It is customary to set µ = 0 as the reference. Accordingly the linear relations take the following form:

E 1
JN = γσ + γ⊥ ∇ ≡ σ (E − S∇T ) (17.36)
T T

E 1
JQ = γ⊥ + γκ ∇ ≡ σST E − κ∇T (17.37)
T T
From the JN equation it follows that in an open circuit a temperature gradient would be balanced by an induced
electric field E = S∇T . This is called Seebeck effect.
128
[18] Scattering approach to mesoscopic transport

The most popular approach to transport in mesoscopic devices takes the scattering formalism rather than the Kubo
formalism as a starting point, leading to the Landauer and the BPT formulas. We first cite these formulas and then
summarize their common derivation. This should be compared with the Kubo-based derivation of the previous section.
====== [18.1] The Buttiker-Pretre-Thomas-Landauer formula

We assume without loss of generality that there are three parameters (x1 , x2 , x3 ) over which we have external control,
where x3 = Φ is the AB flux. The expression for the current IA that goes out of lead A, assuming DC linear response,
can be written as
X
IA = − G3j ẋj (18.1)
j
where −ẋ3 = −Φ̇ is the EMF, and therefore G33 is the conductance in the usual sense. The Büttiker-Prétre-Thomas-
Landauer formula for the generalized conductance matrix is

e ∂S †
G3j = trace PA S (18.2)
2πi ∂xj
In particular for the Ohmic conductance we get the Landauer formula:
e2
G33 = trace(tt† ) (18.3)
2π~
In order to explain the notations in the above formulas we consider a two lead system. The S matrix in block form
is written as follows:
!
rB tAB e−iφ
S= (18.4)
tBA eiφ rA
where r and t are the so called reflection and transmission (sub) matrices respectively. We use the notation φ = eΦ/~.
In the same representation, we define the left lead and the right lead projectors:
! !
0 0 1 0
PA = , PB = (18.5)
0 1 0 0
The following identity is important in order to obtain the Landauer formula from the BPT formula:
dS e e e
= i (PA SPB − iPB SPA ) = i (PA S − SPA ) = −i (PB S − SPB ) (18.6)
dΦ ~ ~ ~
Another important identity is
XX
trace(PA SPB S † ) = trace(tt† ) = |tab |2 (18.7)
a∈A b∈B
The trace() operataion is over the channel indexes.

129
====== [18.2] Floque theory for periodically driven systems

The solution of the Schrodinger equation
dψ
i = Hψ (18.8)
dt
with time independent H is
X
|ψ(t)i = e−iEt |ψ (E) i (18.9)
E
where the stationary states are found from
H|ψ (E) i = E|ψ (E) i (18.10)
Consider now the more complicated case where H depends periodically on time. Given that the basic frequency is ω
we can write
X
H(t) = H(n) e−inωt (18.11)
The solution of the Schrodinger equation is
∞
X X
|ψ(t)i = e−i(E+nω)t |ψ (E,n) i (18.12)
E n=−∞
where the Flouqe states are found from
X 0 0
H(n−n ) |ψ (E,n ) i = (E + nω)|ψ (E,n) i (18.13)
n0
and E is defined modulo ω.
====== [18.3] The Floque scattering matrix

In scattering theory we can define a Flouqe energy shell E. The solution outside of the scattering region is written as
∞ X
X
−i(E+nω)t 1 −ikan r 1 +ikan r
|ψ(t)i = e Aan √ e − Ban √ e ⊗ |ai (18.14)
n=nfloor a
van van
where van and kan are determinedPby the available energy E + nω. The current in a given channel is time dependent,
but its DC component is simply n (|Ban |2 − |Aan |2 ). Therefore the continuity of the probability flow implies that
we can define an on-shell scattering matrix
X
Bbnb = Sbnb ,ana Aana (18.15)
ana
We can write this matrix using the following notation
nb −na
Sbnb ,ana ≡ Sb,a (E + na ω) (18.16)
130
Unitarity implies that
X X
|Sbnb ,ana |2 = n
|Sb,a (E)|2 = 1 (18.17)
bnb bn
X X
|Sbnb ,ana |2 = n
|Sb,a (E + nω)|2 = 1 (18.18)
ana an
If the driving is very slow we can use the adiabatic relation between the incoming and outgoing amplitudes
X
Bb (t) = Sba (X(t)) Aa (t) (18.19)
a
where Sba (X) is the conventional on-shell scattering matrix of the time independent problem. This relation implies
that
Z ω/2π
n ω
Sb,a (E) = Sba (X(t)) einωt dt (18.20)
2π 0
For sake of later use we note the following identity
Z 2π/ω
X
n 2 i Sba (X(t))
n|Sb,a | = dt Sba (X(t)) (18.21)
n
2π 0 ∂t
====== [18.4] Current within a channel

Consider a one dimensional channel labeled as a. Let us take a segment of length L. For simplicity assume periodic
boundary condition (ring geometry). If one state n is occupied the current is
e
Ia,n = va,n (18.22)
L
If several states are occupied we should integrate over the energy

Z Z
X LdE e e
Ia = fa (En )Ia,n = fa (E) va = fa (E)dE (18.23)
n
2πva L 2π
For fully occupied states withing some energy range we get
e e2
Ia = (E2 − E1 ) = (V2 − V1 ) (18.24)
2π 2π
If we have a multi channel lead, then we have to multiply by the number of open channels.
====== [18.5] The Landauer formula

Consider a multi channel system which. All the channels are connected to a scattering region which is described by
an S matrix. We use the notation
gba = |Sba |2 (18.25)

131
Assuming that we occupy a set of scattering states, such that fa (E) is the occupation of those scattering states that
incident in the ath channel, we get that the outgoing current at channel b is
Z " ! #
e X
Ib = dE gba fa (E) − fb (E) (18.26)
2π a
P
Inserting 1 = a gba in the second term we get
Z " #
e X
Ib = dE gba (fa (E) − fb (E)) (18.27)
2π a
Assuming low temperature occupation with
fa (E) = f (E − eVa ) ≈ f (E) − f 0 (E)eVa (18.28)
we get
e2 X
Ib = − gba (Vb − Va ) (18.29)
2π a
which is the multi channel version of the Landauer formula. If we have two leads A and B we can write
" #
e2 X X
IB = − gba (VB − VA ) (18.30)
2π
b∈B a∈A
Form here it follows that the conductance is
e2 X X e2
G= gba = trace(PB SPA S † ) (18.31)
2π 2π
b∈B a∈A
where PA and PB are the projectors that define the two leads.
====== [18.6] The BPT formula

Assuming that the scattering region is periodically driven we can use the Floque scattering formalism. The derivation
of the expression for the DC component Ib of the current in channel b is very similar to the Landauer case, leading to
Z " ! #
e X
n 2
Ib = dE |Sba (E − nω)| fa (E + nω) − fb (E) (18.32)
2π a,n
Z " #
e X
n
= dE |Sba (E − nω)|2 (fa (E − nω) − fb (E)) (18.33)
2π a,n
Z " #
e X
n
= dE |Sba (E)|2 (fa (E) − fb (E + nω)) (18.34)
2π a,n
Z " # " #
e X e X
≈ dE n
nω|Sba (E)|2 (−fa0 (E)) = n
nω|Sba (E)| 2
(18.35)
2π a,n
2π a,n
132
In the last two steps we have assumed very small ω and zero temperature Fermi occupation. Next we use an identity
that has been mentioned previously in order to get an expression that involves the time independent scattering matrix:
Z 2π/ω
e X ω Sba (X(t))
Ib = i dt Sba (X(t)) (18.36)
2π a 2π 0 ∂t
which implies that the pumped charge per cycle is
I I
e X X Sba (X)
Q=i dX Sba (X) ≡ − G(X)dX (18.37)
2π a
∂X
b∈B
with

e X X Sba (X) e ∂S †
G(X) = −i Sba (X) = −i trace PB S (18.38)
2π a
∂X 2π ∂X
b∈B
Note: since S(X) is unitary it follows that the following generator is Hermitian
∂S †
H(X) = i S (18.39)
∂X
The trace of a product of two hermitian operators is always a real quantity.
====== [18.7] BPT and the Friedel sum rule

If only one lead is involved the BPT formula becomes

1 ∂S †
dN = −i trace S dX (18.40)
2π ∂X
where dN is the number of particles that are absorbed (rather than emitted) by the scattering region due to change
dX in some control parameter. A similar formula known as the Friedel sum rule states that

1 ∂S †
dN = −i trace S dE (18.41)
2π ∂E
where N (E) counts the number of states inside the scattering region up to energy E. Both formulas have a very
simple derivation, since they merely involve counting of states. For the purpose of this derivation we close the lead
at r = 0 with Dirichlet boundary conditions. The eigen-energies are found via the equation
det(S(E, X) − 1) = 0 (18.42)
Let us denote the eigenphases of S as θr . We have the identity
X
i δθr = δ[ln det S] = trace[δ ln S] = trace[δSS † ] (18.43)
r
Observing that a new eigenvalue is found each time that one of the eigenphases goes via θ = 0 we get the desired
result.
133
[19] The theory of electrical conductance
====== [19.1] The Hall conductance

The calculation of the Hall conductance is possibly the simplest non-trivial example for adiabatic non-dissipative
response. The standard geometry is a 2D ”hall bar” of dimension Lx × Ly . In ”Lecture notes in quantum mechanics”
we have considered what happens if the electrons are confined in the transverse direction by a potential V (y). Adopting
the Landauer approach it is assumed that the edges are connected to leads that maintain a chemical potential
difference. Consequently there is a net current in the x direction. From the ”Landau level” picture it is clear that the
Hall conductance Gxy is quantized in units e2 /(2π~). The problem with this approach is that the more complicated
case of disorder V (x, y) is difficult for handling. We therefore turn to describe a formal Kubo approach. From now
on we use units such that e = ~ = 1.
We still consider a Hall bar Lx × Ly , but now we impose periodic boundary condition such that ψ(Lx , y) = eiφx ψ(0, y)
and ψ(x, Ly ) = eiφy ψ(x, 0). Accordingly the Hamiltonian depends on the parameters (φx , φy , ΦB ), where ΦB is the
uniform magnetic flux through the Hall bar in the z direction. The currents Ix = (e/Lx )vx and Iy = (e/Ly )vy
are conjugate to φx and φy . We consider the linear response relation Iy = −Gyx φ˙x . This relation can be written as
dQy = −Gyx dφx . The Hall conductance quantization means that a 2π variation of φx leads to one particle transported
in the y direction. The physical picture is very clear in the standard V (y) geometry: the net effect is to displace all
the filled Landau level ”one step” in the y direction.
We now proceed with a formal analysis to show that the Hall conductance is quantized for general V (x, y) potential.
We can define a ”vector potential” An on the (φx , φy ) manifold. If we performed an adiabatic cycle the Berry phase
would be a line integral over An . By Stokes theorem this can be converted into a dφx dφy integral over Bn . However
there are two complementary domains over which the surface integral can be done. Consistency requires that the
result for the Berry phase would come out the same modulo 2π. It follows that
Z 2π Z 2π
1
Bn dφx dφy = integer [Chern number] (19.1)
2π 0 0
This means
P that the φ averaged Bn is quantized in units of 1/(2π). If we fill several levels the Hall conductance is
the sum n Bn over the occupied levels, namely
X X 2Im[I y I x ]
nm mn
Gyx = (19.2)
m
(E m − En )2
n∈band
If we have a quasi-continuum it is allowed to average this expression over (φx , φy ). Then we deduce that the Hall
conductance of a filled band is quantized. The result is of physical relevance if non-adiabatic transitions over the gap
can be neglected.
====== [19.2] The Drude formula

The traditional derivation of the Drude formula is based on the Boltzmann picture. Optionally one can adopt a
Langevine-like picture. The effect of the scattering of an electron in a metal is to randomize its velocity. This
randomization leads to a statistical ”damping” of the average velocity with rate 1/t` . On the other hand the electric
field accelerates the particle with rate eE/m. In steady state the drift velocity is vdrif t = (eE/m)t` , and the current
density is J = (N/V)evdrif t leading to the Drude conductivity σ = (N/V)(e2 /m)t` . Consequently the conductance of
a ring that has a length L and a cross-section A is
e2 e2

A N 2 N ` `
G = σ = t` = e ≡ M (19.3)
L L2 m mvF L L 2π~ L
where ` = vF t` is the mean free path at the Fermi energy, and M is the effective number of open modes. Below we
would like to derive this result formally from the FD relation.
134
The canonical version of the FD relation takes the form G = [1/(2T )]νT , where ν is the intensity of the current
fluctuations and G is the conductance. This is known as Nyquist version of the FD relation. One way to go is to
calculate νT for a many body electronic system, see how this is done in a previous lecture. But if the electrons are non-
interacting it is possible to do a shortcut, relating the conductance of the non-interacting many body electronic system
to its single particle fluctuations. This can be regarded as a generalizations of the canonical Nyquist formula. The
generalization is obtained by re-interpretation of f (E) as the Fermi occupation function (with total occupation N ),
rather than probability distribution. Assuming a Boltzmann occupation one obtains G[N ] = [N/(2T )]νT . A similar
generalization holds for a microcanonical occupation, from which one can deduce results for other occupations. In
particular for low temperature Fermi occupation of non-interacting particles one obtains:
e 2 Z e 2
1 1
G[N ] = g(EF ) νEF = g(EF ) hvk (t)vk (0)idt = g(EF )D0 (19.4)
2 2 L L
The crossover from the high temperature ”Boltzmann” to the low temperature ”Fermi” behavior happens at
T ∼ EF . Assuming exponential velocity-velocity correlation function with time constant τ0 , such that the mean
free path is ` = vF τ0 , we get D0 = vF `. disregarding numerical prefactors the density of states can be written as
g(EF ) = (L/vF )M, where M is the number of open modes. From here we get the Drude formula
e 2 e2 `
G[N ] = g(EF )D0 = M (19.5)
L 2π~ L
Relating to electrons that are moving in a lattice with static disorder, the mean free path can be deduced from the
Fermi Golden Rule (FGR) picture as follows: Assuming isotropic scattering, the velocity-velocity correlation function
is proportional to the survival probability P (t) = e−t/t` . Ignoring a factor that has to do with the dimensionality
d = 2, 3 of the sample the relation is
2 2 −|t|/t`
hv(t)v(0)i ≈ vE P (t) = vE e (19.6)
where the FGR rate of the scattering is
1 πa 2
= 2π%E |Umn |2 = W (19.7)
t` vE
In the last equality we have used |Unm |2 ≈ [a/(ML)]W 2 , where a is the lattice spacing, and W is the strength of the
disorder. Disregarding prefactors of order unity we deduce the so-called Born approximation for the mean free path:
1 vE 2
` = vE t` ≈ (19.8)
a W
====== [19.3] Formal calculation of the conductance

The DC conductance G of a ring with N non-interacting electrons is related by Kubo/FD expression to the density
of one-particle states g(EF ) at the Fermi energy, and to the C̃vv (ω ∼ 0) fluctuations of velocity. The latter can be
deduced semi-classically from the velocity-velocity correlation function, or from the matrix elements of the velocity
operator using the quantum-mechanical spectral decomposition. Optionally one can use path integral or Green
function diagrammatic methods for the calculation.
Let us summarize some optional ways in which the Kubo/FD expression for the Ohmic conductance can be written.
If we use the spectral decomposition with pn = g(EF )−1 δ(En − EF ), we get
1 e 2 X e 2
G = g(EF ) C̃vv (0) = π |vnm |2 δ(En − EF ) δ(Em − En ) (19.9)
2 L nm
L
135
It is implicit that the delta functions are ”broadened” due to the assumed non-adiabaticity of the driving, else we
shall get vanishing dissipation. The calculation of G is the adiabatic regime requires a more careful treatment, and
possibly goes beyond LRT. As long as the broadening is large compared to the level spacing, but small compared with
other energy scales, the result is not too sensitive to the broadening parameter, and small corrections are commonly
ignored, unless the ring is very small. Schematically the above expression can be written as
e 2
G = π~ g(EF )2 |vnm |2 = π~ g(EF )2 |Inm |2 (19.10)
L
where the bar indicates that an average should be taken over the near diagonal matrix elements of the velocity operator
near the Fermi energy. A somewhat more fancy way to write the same is
h i 1
G = π trace I δ(EF − H) I δ(EF − H) = trace [I Im[G(EF )] I Im[G(EF )]] (19.11)
π
where G = 1/(E − H + i0) is the one-particle retarded Green function. This opens the way to formal calculations that
are based on path integral or diagrammatic methods.
For a chaotic ring, the dispersion |Inm |2 of the off-diagonal matrix elements is equal, up to a symmetry factor, to the
dispersion of the diagonal matrix elements. Note that Inn = −∂En /∂Φ. It is common to use the notation φ = (e/~)Φ.
Hence one obtains the Thouless relation:
2
e2

[N ] 1 ∂En
G = factor × × 2 (19.12)
~ ∆ ∂φ
where the numerical factor depends on symmetry considerations, and ∆ is the mean level spacing at the Fermi energy.
There is a more refined relation by Kohn. The Thouless relation is a useful staring point for the formulation of the
scaling theory for localization.
====== [19.4] Conductivity and Conductance

Consider a ring geometry, and assume that the current is driven by the flux Φ. In order to have a better defined
model we should specify what is the vector potential A(r) along the ring. We can regard the valuesH of A at different
points in space as independent parameters (think of tight binding model). Their sum (meaning A(r)·dr) should be
Φ. So we have to know how Φ is ”distributed” along the ring. This is not just a matter of ”gauge choice” because the
electric field E(r) = −Ȧ(r) is a gauge invariant quantity. So we have to know how the voltage is distributed along
the ring. However, as we explain below, in linear response theory this information is not really required. Any voltage
distribution that results in the same electro-motive force, will create the same current.
In linear response theory the current is proportional to the rate in which the parameters are being changed in time.
Regarding the values of A at different points in space as independent parameters, linear response theory postulates
a linear relation between hJ(r)i and E(r 0 ) that involves the conductivity matrix σ(r, r0 ) as a kernel. The current
density has to satisfy the continuity equation ∇ · hJ(r)i = 0. From here it follows that if we replace A by A + ∇Λ(r),
then after integration by parts we shall get the same current. This proves that within linear response theory the
current should depend only on the electromotive force −Φ̇, and not on the way in which the voltage is distributed.
Note that A 7→ A + ∇Λ(r) is not merely a gauge change: A gauge transformation of time dependent field requires a
compensating replacement of the scalar potential (which is not the case here).
In the following it is convenient to think of a device which is composed of ”quantum dot” with two long leads, and
to assume that the two leads are connected together as to form a ring. We shall use the notation r = (r, s), where r
is the coordinate along the ring, and s is a transverse coordinate. In particular we shall distinguish a left ”section”
r = rB and a right section r = rA of the two leads, and we shall assume that the dot region is described by a scattering
matrix Sab .
We further assume that all the voltage drop is concentrated across the section r = rB , and we measure the current
IA through the section r = rA . With these assumptions we have two pairs of conjugate variables, which are (ΦA , IA )
136
and (ΦB , IB ). Note that the explicit expression for the current operator is simply
1
IA = e (v δ(r − rA ) + δ(r − rA )v) (19.13)
2
where v is the r component of the velocity operator. We are interested in calculating the conductance, as define
through the linear response relation hIA i = −GAB Φ̇B . The Kubo expression takes the form
~
GAB = trace [IA Im[G] IB Im[G]] (19.14)
π
This is yet another version of the Kubo formula. Its advantage is that the calculation of the trace involves integration
that is effectively restricted to two planes, whereas the standard version (previous section) requires a double integration
over the whole ”bulk”.
====== [19.5] From the Kubo formula to the Landauer formula

Before we go on we recall that it is implicit that for finite system Im[G] should be ”smeared”. In the dot-leads setup
which is described above, this smearing can be achieved by assuming very long leads, and then simply ”cutting” them
apart. The outcome of this procedure is that G± is the Green function of an open system with outgoing wave (ingoing
wave) boundary conditions. As customary we use a radial coordinate in order to specify locations along the lead,
namely r = ra (r), with 0 < r < ∞. We also define the channel basis as
hr, s|a, ri = χa (s) δ(r − ra (r)) (19.15)
The wavefunction in the lead regions can be expanded as follows:

X
Ca,+ eika r + Ca,− e−ika r

|Ψi = |a, ri (19.16)
a,r
We define projectors P + and P − that project out of the lead wavefunction the outgoing and the ingoing parts
respectfully. It follows that P + G+ = G+ , and that P − G+ = 0, and that G− P − = 0 etc. We define the operator
X
ΓA = |a, rA i~va ha, rA | (19.17)
a∈A
X
= δ(r − rA ) ⊗ |ai~va ha| (19.18)
a∈A
where va = (~ka /mass) is the velocity in channel a. The matrix elements of the second term in Eq.(19.18) are
X
ΓA (s, s0 ) = χa (s) ~va χ∗a (s0 ) (19.19)
a∈A
The operator ΓB is similarly defined for the other lead. Note that these operators commute with the projectors P ± .
It is not difficult to realize that the current operators can be written as
IA = (e/~)[−P + ΓA P + + P − ΓA P − ] (19.20)
IB = (e/~)[+P + ΓB P + − P − ΓB P − ] (19.21)
Upon substitution only two (equal) terms survive leading to the following version of Kubo formula:
e2
GBA = trace ΓB G+ ΓA G−

(19.22)
2π~
137
There is a well known expression (Fisher-Lee) that relates the Green function between plane A and plane B to the S
matrix. Namely:
X 1 1
G+ (sB , sA ) = −i χb (sB ) √ Sba √ χ∗ (sA ) (19.23)
a,b
~vb ~va a
Upon substitution we get
e2 X X
GBA = |Sba |2 (19.24)
2π~
a∈A b∈B
This is the Landauer formula. Note that the sum gives the total transmission of all the open channels.
====== [19.6] From the Kubo formula to the BPT formula

It should be emphasized that the original derivations of the Landauer and the BPT formulas are based on a scattering
formalism which strictly applies only in case of an open system (= system with leads which are connected to reservoirs).
In contrast to that Kubo formula is derived for a closed system. However, it can be shown that by taking an appropriate
limit it is possible to get the BPT formula from the Kubo formula. Namely,
~
η kj = trace F k Im[G+ ] F j Im[G+ ]

(19.25)
π
†
~ ∂S ∂S
= trace (19.26)
4π ∂xi ∂xj
i~
B 3j = − trace F 3 (G+ +G− ) F j Im[G+ ]

(19.27)
2π
∂S † ∂S †

e
= trace PA S − S + intrf (19.28)
4πi ∂xj ∂xj
So the sum is

e ∂S †
G3j = trace PA S (19.29)
2πi ∂xj
For more details see Phys. Rev. B 68, 201303(R) (2003).

138
[20] Irreversibility and Nonequilibrium processes
====== [20.1] The origin of irreversibility

Assume an isolated system with Hamiltonian H(X), where X is a set of control parameters that determine the “fields”.
For simplicity assume that at t = 0 the system is in a stationary state. A driving process means that X = X(t) is
changed in time. In particular a cycle means that X(tfinal ) = X(t=0). A driving process is called reversible is we
can undo it. In the latter case the combined process (including the ”undo”) is a closed cycle, such that at the end
of the cycle the system is back in its initial state. Generally speaking a driving cycle becomes reversible only in the
adiabatic limit. Otherwise it is irreversible.
Micro-reversibility.– One should not confuse reversibility with micro-reversibility. The latter term implies that the
mechanical evolution has time reversal symmetry (TRS). This TRS implies that if we could reverse that state of the
system at some moment (and also the magnetic field if exists) then ideally the system would come back to its initial
state. This is called Lodschmit Echo. In general it is impossible to reverse the state of the system, and therefore in
general micro-reversibility does not imply reversibility!
Sudden process.– The irreversibility of typical systems is related to chaos. The simplest example is free expansion.
In this example X is the location of a piston. At t = 0 the system is prepared in an ergodic state, say a microcanonical
state on the energy surface H(XA ) = E. The piston is moved outwards abruptly form XA to XB . After some time
of ergodization the system will become ergodic on H(XB ) = E. There is no way to reverse this process.
Slow driving.– The more interesting scenario is a slow cycle. Using the assumption of chaos it can be argued that
at the end of the cycle the state will occupy√a shell around H(XA ) = E. If the system is driven periodically (many
cycles), the thickness of this shell grows like DE t with DE ∝ Ẋ 2 . This diffusion in energy space implies (with some
further argumentation) monotonic increase of the average energy. Thus irreversibility implies dissipation of energy:
The system is heated up on the expense of the work which is being done by the driving source.
Non equilibrium steady state.– Another reason for irreversibility is having a ”frustrated” system that is connected
to several baths, each in different temperature, as in the prototype problem of heat conduction. Typically, after a
transient, a steady state is reached. But this steady state is not a canonical thermal equilibrium state. With such
configuration one can associate a rate of ”entropy production”.
====== [20.2] The notion of Entropy

The term ”entropy” is used in a diverse way. In order to avoid confusion we distinguish between the Shanon entropy,
the Von-Neumann entropy,Pthe Boltzmann entropy, and the Thermodynamic entropy. All are calculated by the same
look-alike formula S = − r pr log(pr ), but the context and the meaning of the pr is in general not the same.
Information entropy:– If {pr } are the probabilities to get an output r of a measurement, then S provides a
measure for the uncertainty which is involved in our knowledge of the statistical state. This point of view that regards
S as an information measure has been promoted by Shanon. In the quantum mechanical context we define “classical
state” as implying 100% certainty for any measurement. Such states do not exist in Nature. Rather the states of
minimum uncertainty in N dimensional Hilbert space are the pure states, and they have finite information entropy.
See quant-ph/0401021. They should be contrasted with the worst mixed state whose entropy is S = log(N ).
Von-Neumann entropy:– Von-Neumann has used a Shanon look-alike formula in order to characterize the purity
of a quantum state. In the Von-Neumann definition, the pr are the weights of the pure states in the mixture, namely
S = −trace[ρ ln ρ], where ρ is the probability matrix. It is important to realize that the Von-Neumann entropy has
nothing to do with the theory of irreversibility. If we consider for example the free expansion of a wavepacket in a big
chaotic box, then we have S = 0 at any moment. Still it is nice that the Von-Newman entropy of a canonical state
coincides with the thermodynamic definition.
Boltzmann entropy:– Boltzmann has defined S is a way that allows to discuss irreversibility. The idea was to
divide the phase space of a system into small cells, and to characterize the statistical state in this representation
using S. Then Boltzmann has proposed that S has the tendency to increase with time if the motion is chaotic (a
variation of this idea is the “H theorem” that refers to the reduced one-particle description of a gas particle). The
139
same reasoning can be carried out in the quantum mechanical context where the division of Hilbert space into “cells”
is defined by a complete set of projectors. Obviously in the latter context recurrences imply that the dynamics of a
quantized closed chaotic system looks irreversible only for a limited time.
Thermodynamic entropy:– Using ideal gas thermometer we have identified the empirical temperature P θ = 1/β,
a notion that is postulated by the 0th law of thermodynamics. Later we have shown that dQ ¯ = r dpr Er has an
integration factor T = 1/β as postulated in thermodynamics (see below). Thus we can write dQ ¯ = T dS, where the
definition of S is implied. It turns out that S(T ) = −trace[ρeq ln ρeq ], where ρeq is a canonical state of temperature T .
====== [20.3] Digression - traditional thermodynamics

Let us discuss how ”entropy” is defined in ”traditional thermodynamics” without relaying on Statistical Mechanics.
The first step is to characterize any thermal state by empirical temperature θ. This is well defined by the ”zeroth
law” of thermodynamics. The second step is to represent all the thermal states as points in a X = (X, θ) space. Now
we can define ”adiabatic surface” as the set of states that can be reached via a reversible adiabatic process that does
not involve exchange of energy with the environment. We can label each surface by an number S(X), that we call
”entropy” (with quotations marks). The convention would be that S[A] < S[B] if we can get from A to B via an
irreversible process.
If we have a reversible process that starts at point X, and ends at point X + dX the change in ”entropy” is
dS = ∇S · dX. At the same time we can write for the heat d̄Q = F · dX. By definition both dS = 0 and d̄Q = 0
define the same adiabatic surfaces. It follows that there is an ”integration factor” such that F = T (X, θ)∇S, and
hence one can write dQ
¯ = T dS. We now postulate that there is a possibility to define S such that T is a function of
θ alone. This leads to the definitions of the “absolute temperature” and of the “thermodynamic entropy”.
Let us rephrase the thermodynamic postulate in a more illuminating way. Consider a reversible isothermal process at
temperature θH that connects two adiabatic surfaces. Consider a second reversible isothermal process at temperature
θC that connects the same surfaces. To say that d̄Q has an integration factor that depends on θ alone means that the
ratio dQ
¯ H /¯
dQC depends only on the temperatures θH and θC . Hence we can define “absolute temperature” using the
definition of Carnot, and the definition of S is implied.
In Carnot’s picture the ratio dQ

¯ H /¯dQC has to do with the efficiency of the heat transfer process. According to Carnot
the maximal ratio dQ
¯ H /¯dQC depends only on the temperatures θH and θC . In ”traditional thermodynamics” Carnot’s
statement is regarded as the consequence of either Clausius or Kelvin’s statements that we derive later. If Carnot’s
statement were false, one would be able to combine two reversible processes that do not have the same ”efficiency”
in order to produce a device that can pump heat from cold to hot bath without investing work.
R
During a reversible
R quasi-static process the change of the entropy of a system a is dQ ¯ a /Ta , while that of a second
system is dQ¯ b /Tb . If we have Ta = Tb = θ, it follows that the total entropy change has an additive property, hence
entropy is an extensive quantity.
====== [20.4] The space of all possible states

Canonical states:– The following visualization is useful. Consider a systems that has energy levels n (X). Any
canonical state ρ of the system can be represented as a point in a (X, T ) plane, and has some entropy S(T ; X). Note
that T = 0 states have zero entropy and energy E = 0 (X). We can use S as an optional coordinate instead of T ,
and define Eeq (X, S) as the energy of the canonical state that has entropy S.
Excited states.– We now add a 3rd vertical axis for the energy. In this extended (X, S, E) representation the
canonical states form a surface Eeq (X, S). We refer to this surface as the floor. Non-canonical states with the same
entropy as the canonical state have a higher energy and accordingly are represented by points above the floor.
E ∗ [energy of an excited state that has entropy S] > Eeq (X, S) (20.1)
These excited states are represented in the extended (X, S, E) space as points that reside ”above” the canonical state
Eeq (X, S). Accordingly, all states along a vertical line have the same entropy, but only the lowest state ”on the floor”
is canonical. The trivial example is of course the excited pure states, that by definition have zero entropy, while their
energy E ∗ = n (X) is larger than Eeq = 0 (X).
140
General processes:– We visualize a thermodynamic process as a trajectory in the (X, S, E) space, or optinaly we
can project is on the (X, T ) plane. A reversible quasi-static process that connects points A and B on the floor is
represented by a solid line in the (X, T ) plane. An actual non-reversible process, that resides ”above” the floor, is
represented by a dashed line in the (X, T ) plane. In a closed system Boltzmann told us that the entropy during a
process always increases. Loosely speaking this means that the probability at the end of the process is scattered on
more ”energy levels”.
Thermodynamic processes:– If the system can be attached to baths we can consider a more restricted set of
processes that we call ”Thermodynamic processes”. Such processes start and end at the ”floor”. In other words,
we assume that before and after the process the system is found in equilibrium with a heat bath. The process is
irreversible if during the intermediate stages it is represented by a dashed line that resides ”above” the floor.
====== [20.5] The Statistical-Mechanics version of the second law

P
The Boltzmann entropy is defined as S = − r pr ln pr , where pr is the probability to be in the rth cell in phase
space. The Boltzmann version of the second law states that for any process from state ”A” to state ”B”
S universe [B] − S universe [A] > 0 (20.2)
The Boltzmann entropy is a theoretical construct and hence the statistical version of the second law has no practical
value. We have to ”translate” both the definition of entropy and the ”second law” into a thermodynamic language.
For this purpose it is essential to assume that both A and B are equilibrium states (while during the process the system
may be out of equilibrium). Then we can identify the Boltzmann entropies S(A) and S(B) with the thermodynamic
entropies of states A and B. We shall see in the next section how it helps to formulate a thermodynamic version of
the second law in terms of ”Heat” and ”Work”.
====== [20.6] The thermodynamic version of the second law

In order to translate this microscopic formulation of Boltzmann into the practical language of thermodynamics one
assumes: (1) In the initial and final states the system is in equilibrium with bodies that have well defined temperatures
TA and TB respectively; (2) During the process the system absorbs heat from equilibrated bodies that have well
defined temperatures, while the system itself might be out of equilibrium; (3) The change in the entropy of an
equilibrated body that has a well defined temperature T is −¯ dQ/T , where d̄Q is the heat transfer with the usual sign
convention. With these assumptions we get the thermodynamic version of the second law:
i Z B
h
sys sys dQ
¯
S [B] − S [A] − > 0 (20.3)
A Tbaths
In particular for a closed cycle we get the Clausius inequality

I
dQ
¯
Entropy production ≡ − > 0 (20.4)
Tbaths
Clausius statement:– The simplest application of the Clausius inequality concerns the direction of heat flow.
Consider a cycle (AB)# in which the system is in contact with TA , and later in contact with TB (work is not
involved). The result of such cycle is the transfer of an amount q of energy from TA to TB . Assuming TB < TA it
follows from the Clausius inequality that q must be positive, which loosely speaking means that heat can flow only
from the high to the low temperature. (work-free heat pumps do not exist).
Kelvin statement:– Another immediate implication of the Clausius inequality is that there exist no process whose
sole result is to transfer heat into work. If such process existed one would have at the end of each cycle a single bath
with Q > 0, and hence the total entropy of the universe would decrease. Also the inverse statement is true: if it were
possible to device a work-free pump that violates Clausius statement, then it would be possible to violate Kelvin’s
statement. The proof is based on the possibility to combine such pump device with a Carnot engine.
141
Maximum work principle:– Consider an isothermal process. We use the standard assumptions: the temperature
of the bath is T0 , the initial state is equilibrium, and also at the end of the process we wait until an equilibrium is
reached. Using the first law of thermodynamics (energy conservation) we can substitute Q = (E(B) − E(A)) − W,
where W is the work that has been done on the system. Using F (A) = E(A) − T0 S(A) and F (B) = E(B) − T0 S(A)
we deduce from the second law
W > [F (B) − F (A)] = minimal work required to carry out the process (20.5)
The work that can be extracted from an engine is W = −W. Obviously in order to extract positive work W we need
F (A) > F (B). The maximum work that can be extracted is equal to the free energy difference [F (A) − F (B)]. In
particular it follows that if the universe included only one bath, with one temperature, it would not be possible to
extract work from a closed cycle.
Irreversible work:– Assuming that the state of the system is canonical-like at any instant of time, with a well
define temperature Tsys at any moment along the A 7→ B process. We have established that the change of energy can
be written as dE = −ydX + Tsys dS sys . The second term originates from transitions between levels. These transitions
are induced by the coupling to the environment and/or by the non-adiabaticity of the driving. On the the other hand
by definition dE = dW
¯ + d̄Q. We have identified dW = ydX as the reversible work that could be done by the system.
¯ − (−dW ). Accordingly Tsys dS sys is identified as the sum of heat
The irreversible work is the difference d̄Wirvrs = dW
dQ
¯ and irreversible work d̄Wirvrs . Namely,
Tsys dS sys = d̄Q + dW

¯ irvrs (20.6)
In an actual experiment the irreversible work Wirvrs can be determined by subtracting −ydX from the total work
that has been done on the system, or it can be deduced from the above relation by integrating over Tsys dS sys − dQ.
¯
Obviously the result would be the same, which reflects the first law of thermodynamics (conservation of energy).
Optionally the above relation can help us to express dS sys using dQ

¯ and dW
¯ irvrs . Then it is possible to rewrite the
Clausius statement as follows:
Z B Z B
dW
¯ irvrs 1 1
+ − dQ
¯ > 0 (20.7)
A Tsys A Tsys Tbaths
We see that the origin of reversibility is (i) irreversible work, e.g. frictional effects; (ii) temperature difference between
the system and the bath during heat conduction.
142
====== [20.7] The Carnot Cycle paradigm

A strict adiabatic process is a quasi-static process during which the system is totally isolated from the environment.
For such process we have the adiabatic theorem. Namely, assuming that the motion is chaotic the system that has
been prepared with definite energy E will remain on the the same adiabatic surface (classical version) or in the
same energy level (quantum version) if a parameter X is being changed very slowly. If the system is prepared with
probability pn in some energy shell (classical) or energy level (quantum) then this probability will not change during
the process, and hence also the entropy will remain constant. In the classical version n is the phase space volume of
the evolving energy surface, while in the quantum mechanical formulation it is the index that labels the energy levels.
In the classical limit n ∈ [0, ∞], and the associated energy is denoted as E = n (X).
We can represent all the possible states of a system as points in (X, S, E) space as described in a previous section. The
thermo-adiabatic lines connects canonical points that have the same entropy. Such lines are going along the ”floor”
of the (X, S, E) space. A thermo-adiabatic process is defined as a quasi-static process along a thermo-adiabatic line.
We can think of such process as composed of many infinitesimal steps, where each step consists of a strict adiabatic
process followed by a contact interaction with a bath that has the appropriate temperature.
To see that the quasi-static limit exists, note the following: If a system is prepared in a canonical state Eeq (X0 , S),
Then its energy after a strict adiabatic process is E ∗ > Eeq (X, S) for any X away from X0 . For a small variation
dX the energy difference can be expanded as dE ∗ ∝ dX 2 . If after such a variation the system is connected to a bath
that has the appropriate temperature, such that S(T ; X) = S, it would relax to a canonical state with the same
entropy, but with the lower energy Eeq (X, S). This relaxation involves an entropy production dS env = dE ∗ /T due to
the release of energy to the bath. Integrating dS env over the whole process we see that in the quasi-static limit the
entropy production goes to zero.
A strict Carnot cycle involves only two heat baths. The cycle (ABB ∗ CDD∗ ) is illustrated in the Figure. The initial
preparation is canonical at A(T1 , XA ). The process from A(T1 , XA ) to B ∗ (XB ) is strictly adiabatic. At the end of
this stage the obtained state is not canonical. The process from B ∗ (XB ) to B(XB , T2 ) is the equilibration due to
contact with a bath that has the temperature T2 . It is an irreversible relaxation process in which the system goes to
a lower energy with the same entropy. At the end of this process the obtained state is canonical. The process form
B(XB , T2 ) to C(XC , T2 ) is quasi-static in contact with the same heat bath. The process from C(XC , T2 ) to D∗ (XD )
is strictly adiabatic. The process from D∗ (XD ) to D(XD , T1 ) and later back to A(T1 , XA ) is in contact with the heat
bath T1 .
thermo−adiabatic line
T1 A
T1 D
D*
X
B*
T2 T2 B C
XA XB XD XC
143
====== [20.8] Fluctuations away from equilibrium

It is customary to say that at equilibrium the expectation value of an observable reflects the typical value of this
observable,
√ while the fluctuations are relatively small. If the central limit theorem applies the RMS/mean should scale
as 1/ N . However, it turns out that the full statistics might reveal interesting information about the underlying
dynamics. In the following we shall discuss processes where the distribution function of work or entropy production
does not satisfy the symmetry relation P (−s) = P (s). Rather it satisfies a detailed-balance look-alike relation:
P (−s) = e−βs P (s), [beta-symmetric distribution] (20.8)
It follows that P (s) can be written as a product of a symmetric function and an exponential factor eβs/2 . Another
consequence of the β-symmetry is
he−βs i = 1, [convex average] (20.9)
The latter equality can be re-phrased as follows: In analogy with the definition of harmonic average and geometric
average that are defined as the inverse of h(1/s)i and as the exp of hlog(s)i respectively, here we can define a convex
average that is defined as the log of the above expression. The convex average is zero for a β-symmetric distribution,
while the standard algebraic average is positive
hsi > 0, [convex inequality] (20.10)
While for a symmetric distribution the average value hsi has to be zero, this is no longer true for a β-symmetric
distribution. Rather the average should be related to the variance. To be specific let us assume that s has Gaussian
distribution. It can be easily verified that such distribution has β-symmetry with β = 2µ/σ 2 , where µ = hsi is the
average value and σ 2 = Var(s) is the variance. This relation between the first and second moment can be regarded
as a fluctuation dissipation relation:
1
hsi = β Var(s), [”fluctuation dissipation” relation] (20.11)
2
We can formalize this relation for non-Gaussian distribution in terms of comulant generating function g(λ) which is
defined through
he−λs i ≡ eg(λ) (20.12)
Note that due to normalization g(0) = 0, while g 0 (0) = −µ and g 00 (0) = σ 2 . In particular for a Gaussian
g(λ) = −µλ + (1/2)σ 2 λ2 . For a symmetric distribution g(−λ) = g(λ). But for β-symmetry we must have
g(β − λ) = g(λ), [characterization of beta-symmetric distribution] (20.13)
Again we see that for a Gaussian β-symmetry implies a relation between the mean and the variance.
In the following we shall consider two versions of the non-equilibrium fluctuation theorem. In one version we consider
the statistics P (W) of the work W that is done by an agent during a cycle that involves a thermally isolates system. In
the second version we consider the statistics P (S) of the entropy production S during a cycle that involves exchange
of energy with several heat baths.
144
====== [20.9] The distribution function of the work

The Crooks relation and Jarzynski equality concern the probability distribution of the work that is done during a
non-equilibrium process. For presentation purpose let us consider a gas in cylinder with a movable piston. Initially
the piston is in position A, and the gas in equilibrium with temperature T0 . The canonical probabilities are

1 −(1/T0 )Er(A) F (A)
p(A)
r = e , where Z(A) = exp − (20.14)
Z(A) T0
Now we displace the piston to position B doing work W. After that we can optionally allow the system to relax to
the bath temperature T0 , but this no longer affects W. The distribution of work is defines as
X
PA;B (W) = p(A)
r δ W − (Er(B) − Er(A) ) (20.15)
r
It is implicit here that we assume a conservative deterministic classical system with a well-defined invariant measure
that allows division of phase space into ”cells”. The phase-space states |r(B) i are associated with |r(A) i through the
dynamics in a one-to-one manner. In other words, the index r in the above definition labels a trajectory that starts
(A)
at r. If the dynamics is non-adiabatic the order of the cells in energy space is likely to be scrambled: if the Er are
(B)
indexed in order of of increasing energy; it is likely that Er will become disordered.
If the dynamics is not deterministic the above definition can be modified in an obvious way. To be specific let
us consider the quantum case, where the probability to make a transition form an eigenstate |n(A) i of the initial
Hamiltonian, to an eigenstate |m(A) i of the final Hamiltonian, is given by
2
PA;B (m|n) = hm(B) |UA;B |n(A) i (20.16)

Then we define the spectral kernel:

X
PA;B (ω) = p(A)
n PA;B (m|n) δ ω − (E (B)
m − E (A)
n ) (20.17)
n,m
Since we consider here a closed system, we can identify the work as the energy difference W = ω. For further discussion
of how work can be defined in the quantum context see arXiv:1202.4529
====== [20.10] The Crooks relation

We have defined the probability distribution PA;B (W) for a process that starts at equilibrium with the piston at
position A. The probability distribution PB;A (W) is defined in the same way for a reversed process: initially the
piston is in position B, and the gas in equilibrium with temperature T0 , then the piston is displaced to position A.
The Crooks relation states that

PB;A (−W) W − (F (B) − F (A))
= exp − (20.18)
PA;B (W) T0
The derivation of this relation using the ”quantum” language is trivial and follows exactly the same steps as in the
derivation of the detailed balance relation for any spectral function S̃(ω). The only difference is that here we have an
(A)
extra factor exp[F (B) − F (A)], on top of the Boltzmann factor, that arises because the pn in the forward process
(B)
involves a normalization factor 1/Z(A), while the pm is the reversed process involves a normalization factor 1/Z(B).
====== [20.11] The Jarzynski equality

Multiplying both sides of the Crooks relation by PB;A (W), integrating over W, and taking into account the normal-
145
ization of P (−W), one obtains the Jarzynski equality

W F (B) − F (A)
exp − = exp − (20.19)
T0 T0
It follows from the Jarzynski equality that hWi > [F (B) − F (A)], which is equivalent to the maximum work principle.
It reduces to W = (F (B) − F (A)) in the the case of a quasi-static adiabatic process.
An optional one line derivation of the Jarzynski equality in the context of deterministic classical dynamics is as follows:

W 1 X −(1/T0 )Er (A) Er (B) − Er (A) F (B) − F (A)
exp − = e exp − = exp − (20.20)
T0 Z(A) r T0 T0
The Crooks relation could have been derived in a similar way, but we had preferred to get it using the ”quantum”
language, and to regard the Jarzynski equality as its implication.
====== [20.12] The fluctuation dissipation relation

Let us see what is the implication on the Crooks relation with regard to a simple closed cycle for which F (B) = F (A).
In such case P (W) is a β-symmetric distribution. It follows that there is a ”fluctuation dissipation relation”
1
hWi = Var(W) (20.21)
2T
Considering a multi-cycle process Var(W) = 2DE t and hWi = Ẇt, leading to the dissipation-diffusion relation that
we have derived in past lecture Ẇ = (1/T )DE , from which follows the dissipation-fluctuation relation η = ν/(2T ).
====== [20.13] The non-equilibrium fluctuation theorem

The non-equilibrium fluctuation theorem (Bochkov, Kuzovlev, Evans, Cohen, Morris, Searles, Gallavotti) regards the
probability distribution of the entropy production during a general non-equilibrium process. The clearest formulation
of this idea assumes that the dynamics is described by a rate equation. The transition rates between state n and
state m satisfies

w(m|n) Em − En
= exp − (20.22)
w(n|m) Tnm
Where Tnm is the temperature that controls the nm transition. We can regard the rate equation as describing
a random walk process. Consider a trajectory x(t). If the particle makes a transition from m to n the entropy
production is (Em − En )/Tnm . Hence we get for example

w(1|2)w(2|3)w(3|4) E1 − E2 E2 − E3 E3 − E4
= exp − − − ≡ e−S[1;2;3;4] (20.23)
w(4|3)w(3|2)w(2|1) T1,2 T2,3 T3,4
In general we write
P [x(−t)] h i
= exp − S[x] (20.24)
P [x(t)]
From this ”microscopic” relation we deuce that the probability distribution of the energy production S is a β symmetric
function. A simple example for the practicality of this relation concerns the fluctuations of the current I that emerge
due to the motion of a particle in a ring. Given a trajectory q ≡ It is the winding number and S ≡ qS is the entropy
production. The non-equilibrium fluctuation theorem implies that the ratio P (−q)/P (q) should equal exp(−qS ).
Note that in the case of an electric current S = eV /T , where V is the electro-motive force.
146
====== [20.14] Analysis of heat conduction

A prototype application of the non-equilibrium fluctuation theorem concerns the analysis of heat flow form hot bath TH
to cold bath TC . The temperature difference is = TH − TC . We assume that the conductor that connects the two
baths can be modeled using a master equation. The transition between states of the conductor are induced by the
bath and are like a random walk. With any trajectory we can associate quantities QH and QC that represent that
heat flow from the baths into the conductor. From the fluctuation theorem it follows that

P (−QH , −QC ) QC QH
= exp + (20.25)
P (QH , QC ) TC TH
Next we define the absorbed energy Q̄ = QH + QC and the heat flow Q = (QH − QC )/2. We realize that in the long
time limit Q ∼ t while the fluctuations of Q̄ are bounded. Accordingly we get

P (−Q) 1 1
= exp − − Q (20.26)
P (Q) TC TH
If we use a Gaussian approximation, we get a ”fluctuation-dissipation” relation

1 1 1
hQi = − Var(Q) (20.27)
2 TC TH
The relation can be linearized with respect to = TH − TC . The thermal conductance is defined through hQi = K × t,
and the intensity of fluctuations through Var(Q) = ν × t. Thus we deduce that
1
hQ̇i = K × (TH − TC ), with K = ν (20.28)
2T 2
147
Detailed Table of Contents
======= Thermal Equilibrium. (page 2)
The statistical picture of Mechanics: Random variables; Several random variables; The statistical
description of a classical particle; Dynamics in phase space; The route to ergodicity; Stationary states; The
microcanonical and canonical states; Mathematical digression;
Spectral functions: The definition of counting and partition functions; Two level system or spin; Two spins
system in interaction; Harmonic oscillator; Particle in a 1D box; A particle in 3D box, or higher dimensions;
Classical particle in magnetic field; Gas of classical particles in a box; Two quantum identical particles; Two
quantum particles in a box with interaction;
The canonical formalism: The energy equation of state; The Equipartition theorem; Heat capacity;
Generalized forces; Susceptibility and fluctuations; Empirical temperature; The Virial theorem; Pressure on walls;
Tension of a polymer; Polarization; Magnetization;
Thermodynamics: Absolute temperature and entropy; The Thermodynamic potentials; The Grand Hamil-
tonian approach; The chemical potential; The extensive property; Work; Heat; Quasi static process; Cycles;
Chemical equilibrium and the Grand Canonical state: The Gibbs prescription; Chemical equilib-
rium; The law of mass action; Equilibrium in pair creation reaction; Equilibrium in liquid-gas system; Site system;
The grand canonical formalism; Fermi occupation; Bose occupation; Bosonic mode occupation;
Quantum ideal gases: Equations of state; Explicit expressions for the state equations; Ideal gases in the
Boltzmann approximation; Bose Einstein condensation; Fermi gas at low temperatures;
======= Systems with interactions. (page 46)
Interactions and phase transitions: Gas of weakly interacting particles; The grand canonical perspective;
The cluster expansion; The Virial coefficients; The Van-der-Waals equation of state; From gas with interaction to
Ising problem; Yang and Lee theorems;
The Ising model: Model definition; The spatial correlation function; Critical behavior and the scaling
hypothesis; Digression regarding scaling; Solution of the 1D Ising Model; Solution of the 2D Ising model;
Phase transitions - heuristic approach: The ferromagnetic phase transition; The anti-ferromagnetic
phase transition; Beyond the Ising model; The mean-field Hamiltonian; Coupled rotors; The variational approach;
The Bragg Williams formulation; The Gaussian approximation; The importance of fluctuations;
Phase transitions - field theory: The Landau model; Related models; The Gaussian approximation;
Digression - Gaussian integrals; The mean field equation; Symmetry breaking ; The one dimensional model; Coarse
graining and scaling; Renormalization Group (RG) analysis; Implications of the RG results; The Heisenberg model;
The XY model;
148
======= Fluctuations and Response. (page 82)
Fluctuations: The classical power spectrum; The quantum power spectrum; The detailed balance relation;
The classical version of ”detailed balance”; Fluctuations of a many body system; Fluctuations of several observables;
Reciprocity relations and detailed balance;
Linear response theory: The notion of linear response; Rate of energy absorption; LRT with several
variables; The Kubo formula; Memory and Sensitivity; The Onsager regression formula; The Onsager regression
hypothesis; Onsager reciprocity; The Kubo formula for AC/DC driving; The Kubo formula - FGR version; Adiabatic
response; Low frequency response;
The fluctuation dissipation relation: General formulation; The diffusion-dissipation picture; The wall
formula; The Drude formula; Conductor in electric field; Forced oscillator; Forced particle; Duality between
friction and mobility; The fluctuations of an Ohmic system; The fluctuations of the potential in metals;
======= System interacting with a bath. (page 103)
The modeling Hamiltonian: The Born-Oppenheimer Hamiltonian; The bath model Hamiltonian; Har-
monic bath; Spin bath; Spatially extended environment;
Stochastic picture of the dynamics: Random walk and diffusion; The diffusion equation; The Langevin
equation; The Fokker-Planck equation; Rate equations; Rate equations - formalism; Rate equations - counting
statistics; Rate equations - ergodicity;
Quantum master equations: General perspective; The general Lindblad form; Quantum Fokker-Planck
equation; System-bath interaction; Derivation of the master equation; The Pauli master equation; The Bloch
equation;
======= Additional topics. (page 121)
The kinetic picture: The Boltzmann distribution function; The Boltzmann equation; The calculation of
incident flux; Blackbody radiation; Viscosity; The Navier-Stokes equation; Heat current in an open geometry;
Thermo-electricity;
Scattering approach to mesoscopic transport: The Buttiker-Pretre-Thomas-Landauer formula;

Floque theory for periodically driven systems; The Floque scattering matrix; Current within a channel; The
Landauer formula; The BPT formula; BPT and the Friedel sum rule;
The theory of electrical conductance: The Hall conductance; The Drude formula; Formal calculation
of the conductance; Conductivity and Conductance; From the Kubo formula to the Landauer formula; From the
Kubo formula to the BPT formula;
Irreversibility and Nonequilibrium processes: The origin of irreversibility; The notion of Entropy;
Digression - traditional thermodynamics; The space of all possible states; The Statistical-Mechanics version of the
second law; The thermodynamic version of the second law; The Carnot Cycle paradigm; Fluctuations away from
equilibrium; The distribution function of the work; The Crooks relation; The Jarzynski equality; The fluctuation
dissipation relation; The non-equilibrium fluctuation theorem; Analysis of heat conduction;

SQM PDF

Uploaded by

Copyright:

Available Formats

SQM PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SQM PDF

Uploaded by

Copyright:

Available Formats

Lecture Notes in Statistical Mechanics and Mesoscopics

3 The canonical formalism 15

5 Chemical equilibrium and the Grand Canonical state 33

6 Quantum ideal gases 40

Systems with interactions

8 The Ising model 55

9 Phase transitions - heuristic approach 62

10 Phase transitions - field theory 71

Fluctuations and Response

12 Linear response theory 88

13 The fluctuation dissipation relation 96

System interacting with a bath

15 Stochastic picture of the dynamics 108

16 Quantum master equations 114

18 Scattering approach to mesoscopic transport 128

19 The theory of electrical conductance 133

20 Irreversibility and Nonequilibrium processes 138

Detailed table of contents is available in the last two pages.

[1] The statistical picture of Mechanics

====== [1.1] Random variables

Random variable/observation x̂ (1.1)

Moment generating function Z(λ) = heλx̂ i (1.9)

====== [1.2] Several random variables

Joint distribution function of two variables ρ (x, y) (1.15)

Adding random variables.–

Central limit theorem.– We define the scaled variable

The statement regarding its distribution is

Prob(x̂ > x) ≈ e−N f (x) , f (x) = max{λx − g(λ)} (1.20)

====== [1.3] The statistical description of a classical particle

where the normalization is

The ”spreading” of a wavepacket is characterize by

====== [1.4] Dynamics in phase space

Say that at t = 0 the particle is at (x0 , p0 ). The equations of motion are

The solution is:

ρt=0 (x, p) = 2πδ (p − p0 ) δ (x − x0 ) (1.37)

After time t the state is

====== [1.5] The route to ergodicity

Convex Billiard Concave Billiard

Phase space illustration. – The dynamics of

====== [1.6] Stationary states

The canonical state is

and in a classical context it is written as

The density of states and the partition function are defined as

====== [1.7] The microcanonical and canonical states

Htotal = H (Q) + Henv (Qα ) + Hint (Q, Qα ) (1.45)

where the so-called partition function provides the normalization

====== [1.8] Mathematical digression

Optionally we expand the log of the function:

S(x) ≡ ln f (x) = N ln(x) (1.53)

Thus we have the recipe:

In complete analogy we have:

g (E0 + ) ≈ g (E0 ) eβ (1.57)

where β is the log derivative of the density of states.

[2] Spectral functions

====== [2.1] The definition of counting and partition functions

Each ”Planck cell” in phase space represents a state. Accordingly

In what follows the Gaussian integral is useful:

====== [2.2] Two level system or spin

====== [2.3] Two spins system in interaction

g (E0 + ) ≈ g (E0 ) eβ (1.57)

Note that if we write the energies as Er = n with n = 0, 1, 2, 3, 4, ... then