M.Phys. C6 Theory Option/M.Math - Phys. Advanced Quantum Theory Lecture Notes For 2020-2021

M.Phys. C6 Theory Option/M.Math.Phys.
Advanced Quantum Theory

Lecture Notes for 2020-2021
Lectures given by Sid Parameswaran in Michaelmas 2020
Lecture Notes written by F.H.L. Essler
The Rudolf Peierls Centre for Theoretical Physics
Oxford University, Oxford OX1 3PU, UK
November 19, 2020
These notes cover Parts I-IV of the lectures (Functional Methods; Transfer Matrices; Many-Particle
Quantum Systems; and Landau Theory), which are common to both M.Phys. C6 Theory Option and to
the M.Math.Phys. Advanced Quantum Theory.
Please report errors and typos to sid.parameswaran@physics.ox.ac.uk.
c
2015 F.H.L. Essler.
Contents
I Functional Methods in Quantum Mechanics 4
1 Some Mathematical Background 4

1.1 Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Functional differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Multidimensional Gaussian Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Homework Questions 1&2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Path Integrals in Quantum Mechanics 9

2.1 The Propagator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Propagator as a “Functional Integral” . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Quantum Mechanics à la Feynman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Classical Limit and Stationary Phase Approximation . . . . . . . . . . . . . . . . . . 12
2.4 The Propagator for Free Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Homework Questions 3-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3 Path Integrals in Quantum Statistical Mechanics 16

3.1 Harmonic Oscillator at T > 0: a first encounter with Generating Functionals 17
3.1.1 Imaginary Time Green’s Function of the Harmonic Oscillator . . . . . . . 21
3.2 Homework Question 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Correlation Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3.1 Wick’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4 Probability distribution of position . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Perturbation Theory and Feynman Diagrams . . . . . . . . . . . . . . . . . . . . . . . 28
3.5.1 Partition Function of the anharmonic oscillator . . . . . . . . . . . . . . . 28
3.6 Homework Question 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1
II Path Integrals and Transfer Matrices 35
4 Relation of D dimensional quantum systems to D + 1 dimensional classical ones 35

4.1 Some Facts from Statistical Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Quantum Mechanical Particle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5 The Ising Model 38

5.1 Statistical mechanics of the Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2 The One-Dimensional Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2.1 Transfer matrix approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2.2 Averages of observables in the transfer matrix formalism . . . . . . . . . . 41
5.2.3 The related zero-dimensional quantum model . . . . . . . . . . . . . . . . . . 42
5.3 The Two-Dimensional Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3.1 Transfer Matrix Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3.2 Spontaneous Symmetry Breaking . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3.3 The Related One-Dimensional Quantum Problem . . . . . . . . . . . . . . . . 46
5.4 Homework Questions 8-10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
III Many-Particle Quantum Mechanics 50
6 Second Quantization 50
6.1 Systems of Independent Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.1.1 Occupation Number Representation . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2 Fock Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.2.1 Creation and Annihilation Operators . . . . . . . . . . . . . . . . . . . . . . . 52
6.2.2 Basis of the Fock Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.3 Homework Questions 11-13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.3.1 Change of Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.4 Second Quantized Form of Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.4.1 Occupation number operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.4.2 Single-particle operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.4.3 Two-particle operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.5 Homework Question 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7 Application I: The Ideal Fermi Gas 61

7.1 Quantization in a large, finite volume . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.1.1 Ground State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.1.2 Excitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.1.3 Density Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
8 Application II: Weakly Interacting Bosons 68

8.1 Ideal Bose Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.2 Bogoliubov Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.3 Bogoliubov Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8.4 Ground State and Low-lying Excitations . . . . . . . . . . . . . . . . . . . . . . . . . 70
8.5 Ground state correlation functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
8.6 Depletion of the Condensate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2
9 Application III: Spinwaves in a Ferromagnet 72
9.1 Heisenberg model and spin-rotational SU(2) symmetry . . . . . . . . . . . . . . . . . 73
9.2 Exact ground states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
9.3 Spontaneous Symmetry Breaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
9.4 Holstein-Primakoff Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
9.4.1 Heisenberg Antiferromagnet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
10 Path Integral for interacting Bose systems 78

10.1 Coherent States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
10.2 Partition Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
IV Landau Theory of Phase Transitions 81
11 Phase Transitions 81
12 The Ising Model 81

12.1 Peierls Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
12.2 Mean Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
12.3 Solution of the self-consistency equation for h=0 . . . . . . . . . . . . . . . . . . 85
12.4 Vicinity of the Phase Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
13 Critical Behaviour and Universality 86

13.1 Universality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
14 Landau Theory 88
14.1 Thermodynamic Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
14.2 Beyond the Landau Free Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
14.3 Saddle Point Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
14.4 Mean Field Exponents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
15 Other Examples of Phase Transitions 96

15.1 Isotropic-Nematic Transition in Liquid Crystals . . . . . . . . . . . . . . . . . . . . . 96
15.2 Superfluid Transition in Weakly Interacting Bosons . . . . . . . . . . . . . . . . . 98
16 Ising Model and Φ4 Theory 98
17 Transverse-Field Ising Model 100

17.1 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
17.2 Phase Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
17.2.1 Disordered Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
17.2.2 Ordered Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
17.3 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
17.4 Mean Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
17.5 Jordan-Wigner ‘Fermionization’ of the TFIM . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
17.6 Solution of the Ising Model via Bogoliubov Transformation . . . . . . . . . . . . . . . . . . . 110
17.7 Continuum Limit and Critical Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
17.7.1 Path Integral and Critical Theory for the TFIM . . . . . . . . . . . . . . . . . . . . . 114
17.7.2 Correlation Functions, Structure Factors, and Long-Range Order . . . . . . . . . . . . 118
17.7.3 Ising Symmetry in CoNb2 O6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3
Some general remarks:
These notes aim to be self-contained. Homework questions are marked in red, and are placed at appro-
priate positions in the text, i.e. to work them out you will require only the preceeding material. Passages
marked in blue give details on derivations we don’t have time to go through in the lectures, or present ma-
terial that goes beyond the core of the course. In some cases this material will be very useful for particular
homework problems. All of the material covered in the course can be found in some form or other in a
variety of books. These is no book that covers everything. Some useful references are
Sections of the text written in blue are not formally examinable.
• Path Integrals
R.P. Feynman and A.R. Hibbs, Quantum Mechanics and Path Integrals, Dover Books.
J. Zinn-Justin, Path Integrals in Quantum Mechanics, Oxford.
• Many-particle QM
R.P. Feynman, Statistical Mechanics: A Set of Lectures, Westview Press.
A. Altland and B.D. Simons, Condensed Matter Field Theory, Cambridge.
• Landau Theory of Phase Transitions
M. Kardar, Statistical Physics of Fields, Cambridge.
Part I
Functional Methods in Quantum Mechanics
1 Some Mathematical Background
Functional Methods form a central part of modern theoretical physics. In the following we introduce the
notion of functionals and how to manipulate them.
1.1 Functionals
What is a functional? You all know that a real function can be viewed as a map from e.g. an interval [a, b]
to the real numbers
f : [a, b] → R , x → f (x). (1)
A functional is similar to a function in that it maps all elements in a certain domain to real numbers,
however, the nature of its domain is very different. Instead of acting on all points of an interval or some
other subset of the real numbers, the domain of functionals consists of (suitably chosen) classes of functions.
In other words, given some class {f } of functions, a functional F is a map
F : {f } → R , f → F [f ]. (2)
We now consider two specific examples of functionals.
1. The distance between two points. A very simple functional F consists of the map which assigns to all
paths between two fixed points the length of the path. To write this functional explicitly, let us consider
a simple two-dimensional situation in the (x, y) plane and choose two points (x1 , y1 ) and (x2 , y2 ). We
consider the set of paths that do not turn back, i.e. paths along which x increases monotonically as
we go from (x1 , y1 ) to (x2 , y2 ). These can be described by the set of functions {f } on the interval
[x1 , x2 ] satisfying f (x1 ) = y1 and f (x2 ) = y2 . The length of a path is then given by the well-known
expression Z x2
dx0 1 + f 0 (x0 )2 .
p
F [f (x)] = (3)
x1
4
2. Action Functionals. These are very important in Physics. Let us recall their definition in the context
of classical mechanics. Start with n generalised coordinates q(t) = (q1 (t), . . . , qn (t)) and a Lagrangian
L = L(q, q̇). Then, the action functional S[q] is defined by
Z t2
S[q] = dt L(q(t), q̇(t)) . (4)
t1
It depends on classical paths q(t) between times t1 and t2 satisfying the boundary conditions q(t1 ) = q1
and q(t2 ) = q2 .
1.2 Functional differentiation

In both the examples given above a very natural question to ask is what function extremizes the functional.
In the first example this corresponds to wanting to know the path that minimizes the distance between two
points. In the second example the extremum of the action functional gives the solutions to the classical
equations of motion. This is known as Hamilton’s principle. In order to figure out what function extremizes
the functional it is very useful to generalize the notion of a derivative. For our purposes we define the
functional derivative by
δF [f (x)] F [f (x) + δ(x − y)] − F [f (x)]

= lim .
δf (y) →0
(5)
Here, as usual, we should think of the δ-function as being defined as the limit of a test function, e.g.
1 2 2
δ(x) = lim √ e−x /a , (6)
a→0 πa
and take the limit a → 0 only in the end (after commuting the limit with all other operations such as
the lim→0 in (5)). Importantly, the derivative defined in this way is a linear operation which satisfies the
product and chain rules of ordinary differentiation and commutes with ordinary integrals and derivatives.
Let us see how functional differentiation works for our two examples.
1. The distance between two points. In analogy with finding stationary points of functions we want to
extremize (3) by setting its functional derivative equal to zero
δF [f (x)]
0= . (7)
δf (y)
We first do the calculation by using the definition (5).
q q
δF [f (x)]
Z x2 1 + [f (x ) + δ (x − y)] − 1 + [f 0 (x0 )]2
0 0 0 0 2
= lim dx0 . (8)
δf (y) →0 x
1

√
The Taylor expansion of the square root is 1 + 2 = 1 + + . . ., which gives
f 0 (x0 )δ 0 (x0 − y)
q q
1 + [f 0 (x0 ) + δ 0 (x0 − y)]2 = 1 + [f 0 (x0 )]2 + q + O(2 ) , (9)
2
1 + [f 0 (x0 )]
where δ 0 (x) is the derivative of the delta-function and O(2 ) denote terms proportional to 2 . Substi-
tuting this back into (8) we have 1
δF [f (x)] x2
δ 0 (x0 − y)f 0 (x0 ) d f 0 (y)
Z
= dx0 q =− q . (11)
δf (y) dy
x1 1 + [f 0 (x0 )]2 1 + [f 0 (y)]2
1
In the last step we have used Z b
dx0 δ 0 (x0 − y)g(x0 ) = −g 0 (y) , (10)
a
which can be proved by “integration by parts”.
5
The solution to (7) is thus
f 0 (y) = const, (12)
which describes a straight line. In practice we don’t really go back to the definition of the functional
derivative any more than we use the definition of an ordinary derivative to work it out, but proceed
as follows.
• We first interchange the functional derivative and the integration

Z x2
δF [f (x)] δ
q
= dx 0
1 + [f 0 (x0 )]2 . (13)
δf (y) x1 δf (y)
• Next we use the chain rule

p
δ 1 + f 0 (x0 )2 1 δ(1 + f 0 (x0 )2 ) f 0 (x0 ) δf 0 (x0 )
= p =p . (14)
δf (y) 2 1 + f 0 (x0 )2 δf (y) 1 + f 0 (x0 )2 δf (y)
• Finally we interchange the functional and the ordinary derivative

δf 0 (x0 ) d δf (x0 ) d
= 0 = 0 δ(x0 − y) . (15)
δf (y) dx δf (y) dx
The last identity follows from our definition (5).
Now we can put everything together and arrive at the same answer (11).
Exercise 1
2. Next we want to try out these ideas on our second example and extremize the classical action (4)
in order to obtain the classical equations of motion. We first interchange functional derivative and
integration and then use the chain rule to obtain
Z t2
δS[q] δ
= dt̃ L(q(t̃), q̇(t̃)) (16)
δqi (t) δqi (t) t1
Z t2
∂L δqj (t̃) ∂L δ q̇j (t̃)
= dt̃ (q, q̇) + (q, q̇) (17)
t1 ∂qj δqi (t) ∂ q̇j δqi (t)
(18)
δ q̇j (t̃) d δqj (t̃)
We now use that δqi (t) =and integrate by parts with respect to t̃
dt̃ δqi (t)
Z t2
δS[q] ∂L d ∂L δqj (t̃)
= dt̃ (q, q̇) − (q, q̇) (19)
δqi (t) t1 ∂qj dt̃ ∂ q̇ j δqi (t)
Z t2
∂L d ∂L ∂L d ∂L
= dt̃ (q, q̇) − (q, q̇) δij δ(t̃ − t) = (q, q̇) − (q, q̇) . (20)
t1 ∂qj dt̃ ∂ q̇j ∂qi dt ∂ q̇i
In the second last step we have used
δqj (t̃)
= δij δ(t̃ − t) , (21)
δqi (t)
which follows straightforwardly from our general definition (5). Thus we conclude that the extrema
of the classical action are given by paths that fulfil the equations of motion
∂L d ∂L
(q, q̇) − (q, q̇) = 0.
∂qi dt ∂ q̇i
(22)
Nice.
6
1.3 Multidimensional Gaussian Integrals
As a reminder, we start with a simple one-dimensional Gaussian integral over a single variable y. It is given
by
Z ∞ r
1 2 2π
I(z) ≡ dy exp(− zy ) = ,
−∞ 2 z
(23)
where z is a complex number with Re(z) > 0. The standard proof of this relation involves writing I(z)2
as
p a two-dimensional integral over y1 and y2 and then introducing two-dimensional polar coordinates r =
y12 + y22 and ϕ. Explicitly,
Z ∞ Z ∞ Z ∞ Z ∞
2 1 2 1 2 1
I(z) = dy1 exp(− zy1 ) dy2 exp(− zy2 ) = dy1 dy2 exp(− z(y12 + y22 )) (24)
−∞ 2 −∞ 2 −∞ −∞ 2
Z 2π Z ∞
1 2π
= dϕ dr r exp(− zr2 ) = . (25)
0 0 2 z
Next we consider n-dimensional Gaussian integrals

1 T
Z
n
W0 (A) ≡ d y exp − y Ay , (26)
2
over variables y = (y1 , . . . , yn ), where A is a symmetric, positive definite matrix (all its eigenvalues are
positive). This integral can be reduced to a product of one-dimensional Gaussian integrals by diagonalising
the matrix A. Consider an orthogonal rotation O such that A = ODOT with a diagonal matrix D =
diag(a1 , . . . , an ). The eigenvalues ai are strictly positive since we have assumed that A is positive definite.
Introducing new coordinates ỹ = OT y we can write
n
X
T T
y Ay = ỹ Dỹ = ai ỹi2 , (27)
i=1
where the property OT O = 1 of orthogonal matrices has been used. Note further that the Jacobian of
the coordinate change y → ỹ is one, since |det(O)| = 1. Hence, using Eqs. (23) and (27) we find for the
integral (26)
n Z
Y 1
W0 (A) = dỹi exp(− ai ỹi2 ) = (2π)n/2 (a1 a2 . . . an )−1/2 = (2π)n/2 (detA)−1/2 . (28)
2
i=1
To summarise, we have found for the multidimensional Gaussian integral (26) that
W0 (A) = (2π)n/2 (detA)−1/2 ,

(29)
a result which will be of some importance in the following. We note that if we multiply the matrix A by a
complex number z with Re(z) > 0 and then follow through exactly the same steps, we find
n/2
2π
W0 (zA) = (detA)−1/2 . (30)
z
One obvious generalisation of the integral (26) involves adding a term linear in y in the exponent, that is

1 T
Z
n T
W0 (A, J) ≡ d y exp − y Ay + J y . (31)
2
7
Here J = (J1 , . . . , Jn ) is an n-dimensional vector. Changing variables y → ỹ, where
y = A−1 J + ỹ (32)
this integral can be written as

Z
1 T −1 n 1 T
W0 (A, J) = exp J A J d ỹ exp − ỹ Aỹ . (33)
2 2
The remaining integral is Gaussian without a linear term, so can be easily carried out using the above
results. Hence, one finds

n/2 −1/2 1 T −1
W0 (A, J) = (2π) (detA) exp J A J .
2
(34)
1.4 Homework Questions 1&2

Question 1. Consider paths X = X(τ ), where τ is a parameter, and the functional
Z τ1 r
dX dX
l[X] = dτ n(X) · ,
τ0 dτ dτ
where n = n(X) is a function. (The minima of this functional can be interpreted as light rays propagating in a
medium with refractive index n.)
a) Derive the differential equation which has to be satisfied by minimal paths X.
b) Consider a two-dimensional situation with paths X(τ ) = (X(τ ), Y (τ )) in the x, y plane and a function
n = n0 + (n1 − n0 ) θ(x). (The Heaviside function θ(x) is defined to be 0 for x < 0 and 1 for x ≥ 0. Recall
that θ0 (x) = δ(x).) Solve the differential equation in a) for this situation, using the coordinate x as parameter τ
along the path.
c) Show that the solution in b) leads to the standard law for refraction at the boundary between two media with
refractive indices n0 and n1 .
Question 2. a) Evaluate the Gaussian integral

Z ∞ 1 2
dx e− 2 zx (35)
−∞
for a complex constant z. What is the requirement on z for the integral to exist?
b) The gamma function Γ is defined by
Z ∞
Γ(s + 1) = dx xs e−x .
0
c) Show that Γ(1) = 1 and Γ(s + 1) = sΓ(s). (Hence Γ(n + 1) = n!)

d) Take s to be real and positive. Evaluate Γ(s + 1) in the steepest descent approximation: write the integrand
in the form ef (x) and argue that for large s 1 the dominant contribution to the integral arises from the minima
of f (x). Expand the function to quadratic order around the minimum, argue that you may extend the integration
boundaries to ±∞, and then carry out the resulting integral. Your result is known as Stirling’s approximation: it
tells you what n! is when n becomes large.
e)∗ The following extension is for complex analysis afficionados, so simply omit it if you haven’t taken the short
option. Take s to be complex with positive real part. Deform the contour in a suitable way, so that you can
again apply a steepest descent approximation. Ponder the name of the method. What is Stirling’s approximation
for complex s?
8
2 Path Integrals in Quantum Mechanics
So far you have encountered two ways of doing QM:
1. Following Schrödinger, we can solve the Schrödinger equation for the wave function → Fun with
PDEs...
2. Following Heisenberg, we can work with operators, commutation relations, eigenstates → Fun with
Linear Algebra...
Historically it took some time for people to realize that these are in fact equivalent. To quote the great
men: I knew of Heisenberg’s theory, of course, but I felt discouraged, not to say repelled, by the methods
of transcendental algebra, which appeared difficult to me, and by the lack of visualizability. (Schrödinger in
1926)
The more I think about the physical portion of Schrödinger’s theory, the more repulsive I find it. What
Schrödinger writes about the visualizability of his theory is probably not quite right, in other words it’s crap.
(Heisenberg, writing to Pauli in 1926)
There is a third approach to QM, due to Feynman. He developed it when he was a graduate student,
inspired by a mysterious remark in a paper by Dirac. Those were the days! Feynman’s approach is partic-
ularly useful for QFTs and many-particle QM problems, as it makes certain calculations much easier. We
will now introduce it by working backwards. The central object in Feynman’s method is something called
a propagator. We’ll now work out what this is using the Heisenberg/Schrödinger formulation of QM you
know and love. After we have done that, we formulate QM à la Feynman.
2.1 The Propagator

Our starting point is the time-dependent Schrödinger equation
∂
i~ |ψ(t)i = H|ψ(t)i. (36)
∂t
We recall that the wave function is given by
ψ(~x, t) = h~x|ψ(t)i. (37)
Eqn (36) can be integrated to give

i
|ψ(t)i = e− ~ Ht |ψ(0)i (38)
The time-evolution operator in QM is thus (assuming that H is time-independent)
i
U (t; t0 ) = e− ~ H(t−t0 ) . (39)
A central object in Feynman’s approach is the propagator
h~x0 |U (t; t0 )|~xi,

(40)
where |~xi are the simultaneous eigenstates of the position operators x̂, ŷ and ẑ. The propagator is the
probability amplitude for finding our QM particle at position ~x0 at time t, if it started at position ~x at time
t0 . To keep notations simple, we now consider a particle moving in one dimension with time-independent
Hamiltonian
p̂2
H = T̂ + V̂ = + V (x̂). (41)
2m
We want to calculate the propagator
hxN |U (t; 0)|x0 i. (42)
9
It is useful to introduce small time steps
tn = n , n = 0, . . . , N, (43)
where = t/N . Then we have by construction

i N
U (t; 0) = e− ~ H . (44)
The propagator is
i i
hxN |U (t; 0)|x0 i = hxN |e− ~ H · · · e− ~ H |x0 i
Z Z
i i i
= dxN −1 · · · dx1 hxN |e− ~ H |xN −1 ihxN −1 |e− ~ H |xN −2 i . . . hx1 |e− ~ H |x0 i, (45)
where we have inserted N − 1 resolutions of the identity in terms of position eigenstates

Z
1 = dx |xihx| . (46)
This expression now has a very nice and intuitive interpretation, see Fig. 1: The propagator, i.e. the
Figure 1: Propagator as sum over paths.
probabilty amplitude for finding the particle at position xN and time t given that it was at position x0 at
time 0 is given by the sum over all “paths” going from x0 to xN (as x1 ,. . . , xN −1 are integrated over).
In the next step we determine the “infinitesimal propagator”
i
hxn+1 |e− ~ H |xn i. (47)
Importantly we have [T̂ , V̂ ] 6= 0 and concomitantly
eα(T̂ +V̂ ) 6= eαT̂ eαV̂ . (48)
However, using that is infinitesimal, we have

i i
e− ~ (T̂ +V̂ ) = 1 − (T̂ + V̂ ) + O(2 ) ,
~
i i i
e− ~ T̂ e− ~ V̂ = 1 − (T̂ + V̂ ) + O(2 ). (49)
~
10
So up to terms of order 2 we have
i i i i i
hxn+1 |e− ~ H |xn i ' hxn+1 |e− ~ T̂ e− ~ V̂ |xn i = hxn+1 |e− ~ T̂ |xn ie− ~ V (xn ) , (50)
where we have used that V̂ |xi = V (x)|xi. As T̂ = p̂2 /2m it is useful to insert a complete set of momentum
eigenstates 2 to calculate
dp dp − ip2 −i p (xn −xn+1 )
Z Z
i ip̂2
hxn+1 |e− ~ T̂ |xn i = hxn+1 |e− 2m~ |pihp|xn i = e 2m~ ~
2π~ 2π~
r
m im 2
= e 2~ (xn −xn+1 ) . (51)
2πi~
In the second step we have used that p̂|pi = p|pi and that
ipx
hx|pi = e ~ . (52)
The integral over p is performed by changing variables to p0 = p + m (xn − xn+1 ) (and giving a very
small imaginary part in order to make the integral convergent). Substituting (51) and (50) back into our
expression (45) for the propagator gives
N −1
!
h m iN Z i X m xn+1 − xn 2

2
hxN |U (t; 0)|x0 i = lim dx1 . . . dxN −1 exp − V (xn ) . (53)
N →∞ 2πi~ ~ 2
n=0
Note that in this expression there are no operators left.
2.1.1 Propagator as a “Functional Integral”

The way to think about (53) is as a sum over trajectories:
• x0 , . . . , xN constitute a discretization of a path x(t0 ), where we set xn ≡ x(tn ).
• We then have
xn+1 − xn x(tn+1 ) − x(tn )
= ' ẋ(tn ), (54)
tn+1 − tn
and
N −1 2 t t
xn+1 − xn

m hm
X Z i Z
0 0
− V (xn ) ' dt 2
ẋ (t ) − V (x) ≡ dt0 L[ẋ, x], (55)
2 0 2 0
n=0
where L is the Lagrangian of the system. In classical mechanics the time-integral of the Lagrangian
is known as the action Z t
S= dt0 L. (56)
0
• The integral over x1 , . . . xN −1 becomes a functional integral, also known as a path integral, over all
paths x(t0 ) that start at x0 at time t0 = 0 and end at xN at time t0 = t.
• The prefactor in (53) gives rise to an overall (infinite) normalization and we will denote it by N .
These considerations lead us to express the propagator as the following formal expression
i 0
Dx(t0 ) e ~ S[x(t )] .
R
hxN |U (t; 0)|x0 i = N
(57)
What is in fact meant by (57) is the limit of the discretized expression (53). The ultimate utility of (57) is
that it provides a compact notation, that on the one hand will allow us to manipulate functional integrals,
and on the other hand provides a nice, intuitive interpretation.
2 dp
R
We use a normalization hp|ki = 2π~δ(p − k), so that 1 = 2π~
|pihp|.
11
2.2 Quantum Mechanics à la Feynman
Feynman’s formulation of Quantum Mechanics is based on the single postulate that the probability amplitude
for propagation from a position x0 to a position xN is obtained by summing over all possible paths connecting
x0 and xN , where each path is weighted by a phase factor exp ~i S , where S is the classical action of the

path. This provides a new way of thinking about QM!
2.3 Classical Limit and Stationary Phase Approximation

An important feature of (57) is that it gives us a nice way of thinking about the classical limit “~ → 0” (more
precisely in the limit when the dimensions, masses, times etc are so large that the action is huge compared
to ~). To see what happens in this limit let us first consider the simpler case of an ordinary integral
Z ∞
g(a) = dt h1 (t)eiah2 (t) , (58)
−∞
when we take the real parameter a to infinity. In this case the integrand will oscillate wildly as a function of
t because the phase of exp iah2 (t) will vary rapidly. The dominant contribution will arise from the points
where the phase changes slowly, which are the stationary points
h02 (t) = 0. (59)
The integral can then be approximated by expanding around the stationary points. Assuming that there is
a single stationary point at t0
Z ∞ ah00
2 (t0 ) 2
dt h1 (t0 ) + (t − t0 )h01 (t0 ) + . . . eiah2 (t0 )+i 2 (t−t0 ) ,

g(a 1) ≈ (60)
−∞
Changing integration variables to t0 = t − t0 (and giving a a small imaginary part to make the integral
converge at infinity) as obtain a Gaussian integral that we can take using (23)
s
2πi
g(a 1) ≈ h1 (t0 )eiah2 (t0 ) . (61)
ah002 (t0 )
Subleading contributions can be evaluated by taking higher order contributions in the Taylor expansions
into account. If we have several stationary points we sum over their contributions. The method we have
just discussed is known as stationary phase approximation.
The generalization to path integrals is now clear: in the limit ~ → 0 the path integral is dominated by
the vicinity of the stationary points of the action S
δS
= 0. (62)
δx(t0 )
The condition (62) precisely defines the classical trajectories x(t0 )!
2.4 The Propagator for Free Particles

We now wish to calculate the functional integral (57) for a free particle, i.e.
V (x) = 0. (63)
Going back to the explicit expression (53) we have

N −1
!

2
hxN |U (t; 0)|x0 i = lim dx1 . . . dxN −1 exp . (64)
N →∞ 2πi~ ~ 2
n=0
12
It is useful to change integration variables to
yj = xj − xN , j = 1, . . . , N − 1, (65)
which leads to an expression

h m iN Z
1 T

im 2
dy exp − y Ay + J · y e 2~ (x0 −xN ) .
2 T
hxN |U (t; 0)|x0 i = lim (66)
N →∞ 2πi~ 2
Here
im
JT =

(xN − x0 ), 0, . . . , 0 , (67)
~
and A is a (N − 1) × (N − 1) matrix with elements
−im
Ajk = [2δj,k − δj,k+1 − δj,k−1 ] . (68)
~
For a given N (66) is a multidimensional Gaussian integral and can be carried out using (34)
h m iN N −1 1

1 T −1

im 2
(2π) 2 [det(A)]− 2 exp J A J e 2~ (x0 −xN ) .
2
hxN |U (t; 0)|x0 i = lim (69)
N →∞ 2πi~ 2
The matrix A is related to the one dimensional lattice Laplacian, see below. Given the eigenvalues and
eigenvectors worked out below we can calculate the determinant and inverse of A (homework problem).
Substituting the results into (69) gives
im
m (x0 −xN )2
p
hxN |U (t; 0)|x0 i = 2πi~t e
2~t .
(70)
For a free particle we can evaluate the propagator directly in a much simpler way.
Z ∞ Z ∞
dp p̂2 t dp −i p2 t −i p(x0 −xN )
hxN |U (t; 0)|x0 i = hxN |e−i 2m~ |pihp|x0 i = e 2m~ ~
−∞ 2π~ −∞ 2π~
r
m im (x0 −xN )2
= e 2~t . (71)
2πi~t
Exercise 2
The matrix A is related to the one dimensional Lattice Laplacian. Consider functions of a variable z0 ≤
z ≤ zN with “hard-wall boundary conditions”
f (z0 ) = f (zN ) = 0. (72)
The Laplace operator D acts on these functions as
d2 f (z)
Df ≡ . (73)
dz 2
Discretizing the variable z by introducing N − 1 points
zn = z0 + na0 , n = 1, . . . , N − 1 (74)
where a0 = (zN − z0 )/N is a “lattice spacing”, maps the function f (z) to a N − 1 dimensional vector
f (z) → f = (f (z1 ), . . . , f (zN −1 )). (75)
13
Recalling that
d2 f f (z + a0 ) + f (z − a0 ) − 2f (z)
(z) = lim , (76)
dz 2 a0 →0 a20
we conclude that the Lapacian is discretized as follows
Df → a−2
0 ∆f , (77)
where
∆jk = δj,k+1 + δj,k−1 − 2δj,k . (78)
im
Our matrix A is equal to ~ ∆. The eigenvalue equation
∆an = λn an , n = 1, . . . , N − 1 (79)
gives rise to a recurrence relation for the components an,j of an
an,j+1 + an,j−1 − (2 + λn )an,j = 0. (80)
The boundary conditions an,N = an,0 = 0 suggest the ansatz
πnj
an,j = Cn sin . (81)
N
Substituting this in to (80) gives
πn
λn = 2 cos −2 , n = 1, . . . , N − 1. (82)
N
The normalized eigenvectors of ∆ are
πn πn
   
sin N sin N
2πn 2πn
1  sin N
 r2  sin N

an = qP .. = .. (83)
   
N
  
N −1
sin2 πnj  .   . 
j=1 N π(N −1)n π(N −1)n
sin N . sin N .
2.5 Homework Questions 3-5

Question 3. Consider a free quantum mechanical particle moving in one dimension. The Hamiltonian is
~2 d2
H=− . (84)
2m dx2
We have shown in the lecture that the propagator can be represented in the form
N −1
!

− ~i tH 2
hxN |e |x0 i = lim dx1 . . . dxN −1 exp . (85)
N →∞ 2πi~ ~ 2
n=0
a) Change variables from xj to yj = xj − xN to bring it to the form

h m iN Z
1 T

− ~i tH im 2
dy exp − y Ay + J · y e 2~ (x0 −xN ) .
2 T
hxN |e |x0 i = lim (86)
N →∞ 2πi~ 2
Give expressions for J and A.

b) Carry out the integrals over yj to get an expression for the propagator in terms of A and J.
14
c) Work out the eigenvalues λn and eigenvectors an of the matrix A. You may find helpful hints in the lecture
notes.
d) What is det(A)? A useful identity you may use is
N
Y −1 √
2 sin(πj/2N ) = N. (87)
j=1
Now work out JT A−1 J by working in the eigenbasis of A−1 (Hint: write this as JT A−1 J = JT OT OA−1 OT OJ,
where OT O = 1 and OA−1 OT is a diagonal matrix you have already calculated above.). A useful identity you
may use is
N −1
X N −1
cos2 (πj/2N ) = . (88)
2
j=1
e) Use the result you have obtained to write an explicit expression for the propagator.
Question 4. Denote the propagator by

i 0
K(t, x; t0 x0 ) = hx|e− ~ H(t−t ) |x0 i. (89)
Show that the wave function ψ(t, x) = hx|Ψ(t)i, where |Ψ(t)i is a solution to the time-dependent Schrödinger
equation
∂
i~ |Ψ(t)i = H|Ψ(t)i, (90)
∂t
fulfils the integral equation Z ∞
ψ(t, x) = dx0 K(t, x; t0 x0 ) ψ(t0 , x0 ). (91)
−∞
Question 5. Diffraction through a slit. A free particle starting at x = 0 when t = 0 is determined to

pass between x0 − b and x0 + b at time T . We wish to calculate the probabilty of finding the particle at position
x at time t = T + τ .
a) Argue on the basis of Qu 5. that the (un-normalized) wave function can be written in the form
Z b
ψ(T + τ, x) = dy K(T + τ, x; T, x0 + y) K(T, x0 + y; 0, 0) , (92)
−b
where
i 0
K(t, x; t0 x0 ) = hx|e− ~ H(t−t ) |x0 i. (93)
b) Using that the propagation for 0 ≤ t < T and T ≤ t < T + τ is that of a free particle, obtain an explicit
integral representation for the wave function.
c) Show that the wave function can be expressed in terms of the Fresnel integrals
Z x Z x
2
C(x) = dy cos(πy /2) , S(x) = dy sin(πy 2 /2) . (94)
0 0
Hint: make a substitution z = αy + β with suitably chosen α and β.

Derive an expression for the ratio P (T + τ, x)/P (T + τ, x0 ), where P (T + τ, x)dx is the probability of finding
the particle in the interval [x, x + dx] at time T + τ .
d)∗ If you can get hold of Mathematica (the default assumption is that you will not), plot the result as a function
of the dimensionless parameter x/[b(1 + τ /T )] for x0 = 0 and different values of the ratio
mb2 (1 + τ /T )
γ= . (95)
~τ
Discuss your findings.
15
3 Path Integrals in Quantum Statistical Mechanics
Path integrals can also be used to describe quantum systems at finite temperatures. To see how this works
we now consider a quantum mechanical particle coupled to a heat bath at a temperature T . An important
quantity in Statistical Mechanics is the partition function
h i
Z(β) = Tr e−βH , (96)
where H is the Hamiltonian of the system, Tr denotes the trace over the Hilbert space of quantum mechanical
states, and
1
β= . (97)
kB T
Ensemble averages of the quantum mechanical observable O are given by
1 h i
hOiβ = Tr e−βH O . (98)
Z(β)
Taking the trace over a basis of eigenstates of H with H|ni = En |ni gives
1 X −βEn
hOiβ = e hn|O|ni ,
Z(β) n
X
Z(β) = e−βEn . (99)
n
Assuming that the ground state of H is non-degenerate we have
lim hOiβ = h0|O|0i , (100)

T →0
where |0i is the ground state of the system. Let us consider a QM particle with Hamiltonian
p̂2
H= + V (x̂), (101)
2m
coupled to a heat bath at temperature T . The partition function can be written in a basis of position
eigenstates Z Z Z
−βH
Z(β) = dxhx|e |xi = dx dx0 hx|x0 i hx0 |e−βH |xi. (102)
Here
hx0 |e−βH |xi (103)
is very similar to the propagator
i(t−t0 )
hx0 |e− ~
H
|xi. (104)
Formally (103) can be viewed as the propagator in imaginary time τ = it, where we consider propagation
from τ = 0 to τ = β~. Using this interpretation we can follow through precisely the same steps as before
and obtain
N −1
!
h m iN Z X m xn+1 − xn 2

−βH 2
hxN |e |x0 i = lim dx1 . . . dxN −1 exp − + V (xn ) , (105)
N →∞ 2π~ ~ 2
n=0
where now
~β
= . (106)
N
16
We again can interpret this in terms of a sum over paths x(τ ) with
x(τn ) = xn , τn = n. (107)
Going over to a continuum description we arrive at an imaginary-time functional integral
1
hxN |e−βH |x0 i = N Dx(τ ) e− ~ SE [x(τ )] ,
R
(108)
where SE is called Euclidean action

" 2 #
~β
m dx
Z
SE [x(τ )] = dτ + V (x) , (109)
0 2 dτ
and the path integral is over all paths that start at x0 and end at xN . Substituting (108) into the expression
for the partition function we find that
1
Dx(τ ) e− ~ SE [x(τ )] ,
R
Z(β) = N
(110)
where we integrate over all periodic paths
x(~β) = x(0). (111)

R
The restriction to periodic paths arises because Z(β) is a trace. Please note that the notation Dx(τ )
means very different things in (108) and (110), as a result of involving very different classes of paths that
are “integrated over”. In the first case the path integral is over all paths from x0 to xN , in the latter over all
periodic paths starting an an arbitrary position. In practice it is always clear from the context what paths
are involved, and hence this ambiguous notation should not cause any confusion.
3.1 Harmonic Oscillator at T > 0: a first encounter with Generating Functionals

We now consider the simplest case of the potential V (x), the harmonic oscillator
p̂2 κ
H= + x̂2 . (112)
2m 2
The physical quantities we want to work out are the averages of powers of the position operator
dx hx|e−βH x̂n |xi dx xn hx|e−βH |xi

R R
n
hx̂ iβ = R = R . (113)
dxhx|e−βH |xi dxhx|e−βH |xi
If we know all these moments, we can work out the probability distribution for a position measurement
giving a particular result x. At zero temperature this is just given by the absolute value squared of the
ground state wave function. The coupling to the heat bath will generate “excitations” of the harmonic
oscillator and thus affect this probability distribution. We have
h i
dx 2 κ 2
Z
2 ( dτ )
− 1 ~β dτ m
R
+2x
hx|e−βH |xi = N Dx(τ ) e ~ 0 , (114)
where the path integral is over all paths with x(0) = x(β~). Integrating by parts we can write the action as
" #
1 1 ~β m dx 2 κ 2 1 ~β m
Z Z ~β
− SE [x(τ )] = − dτ + x =− dτ x(τ )D̂x(τ ) − x(τ )ẋ(τ ) , (115)

~ ~ 0 2 dτ 2 2 0 2~ 0
17
where
m d2 κ
D̂ = − 2
+ . (116)
~ dτ ~
The contributions from the integration boundaries in (115) don’t play a role in the logic underlying the
following steps leading up to (124) and work out in precisely the same way as the “bulk” contributions. In
order to show that we’re not dropping anything important we’ll keep track of them anyway. We now define
the generating functional Z
1
R ~β
W [J] ≡ N Dx(τ ) e− ~ SE [x(τ )]+ 0 dτ J(τ )x(τ )
. (117)
Here the functions J(τ ) are called sources. The point of the definition (117) is that we can obtain hx̂n iβ by
taking functional derivatives

n 1 δ δ
hx̂ iβ = ... W [J]. (118)
W [0] δJ(0) δJ(0)

J=0
We now could go ahead and calculate the generating functional by going back to the definition of the the
path integral in terms of a multidimensional Gaussian integral. In practice we manipulate the path integral
itself as follows. Apart from the contribution from the integration boundaries the structure of (115) is
completely analogous to the one we encountered for Gaussian integrals, cf (31). The change of variables
(32) suggests that we should shift our “integration variables” by a term involving the inverse of the integral
operator D̂. The latter corresponds to the Green’s function defined by
D̂τ G(τ − τ 0 ) = δ(τ − τ 0 ) , G(0) = G(β~). (119)
We then change variables in the path integral in order to “complete the square”
Z
y(τ ) = x(τ ) − dτ 0 G(τ − τ 0 )J(τ 0 ). (120)
Under this change of variables we have

Z Z Z Z
dτ x(τ )D̂τ x(τ ) − 2 dτ x(τ )J(τ ) = dτ y(τ )D̂τ y(τ ) − dτ dτ 0 J(τ )G(τ − τ 0 )J(τ 0 )
+boundary terms. (121)
Exercise 3
Verify that
Z Z Z
dτ y(τ )D̂τ y(τ ) = dτ x(τ )D̂τ x(τ ) + dτ dτ 0 dτ 00 G(τ − τ 0 )J(τ 0 )D̂τ G(τ − τ 00 )J(τ 00 )
Z h i
− dτ dτ 0 x(τ )D̂τ G(τ − τ 0 )J(τ 0 ) + G(τ − τ 0 )J(τ 0 )D̂τ x(τ )
Z Z Z
0 0 0
= dτ x(τ )D̂τ x(τ ) + dτ dτ G(τ − τ )J(τ )J(τ ) − 2 dτ x(τ )J(τ )
m ~β m ~β
+ x(τ )ẋ(τ ) − y(τ )ẏ(τ ) , (122)

~ 0 ~ 0
In the last step you need to use (119) and integrate by parts twice to simplify the last term in the second
line.
Putting everything together we arrive at
1 1 1 ~β
Z ~β Z
− SE [x(τ )] + dτ J(τ )x(τ ) = − SE [y(τ )] + dτ dτ 0 J(τ )G(τ − τ 0 )J(τ 0 ). (123)
~ 0 ~ 2 0
18
On the other hand, the Jacobian of the change of variables (120) is 1 as we are shifting all paths by the
same constant (you can show this directly by going back to the definition of the path integral in terms of
multiple Gaussian integrals). Hence we have Dy(τ ) = Dx(τ ) and our generating functional becomes
1
dτ dτ 0 J(τ )G(τ −τ 0 )J(τ 0 )
R
W [J] = W [0] e 2 .
(124)
Now we are ready to calculate (118). The average position is zero

1 δ 1
Z
0 0 0 0 0

hx̂iβ = W [J] = dτ dτ δ(τ )G(τ − τ )J(τ ) + J(τ )G(τ − τ )δ(τ ) W [J] = 0.

W [0] δJ(0) 2W [0]

J=0 J=0
(125)
Here we have used that
δJ(τ )
= δ(τ − τ 0 ). (126)
δJ(τ 0 )
The expression (125) vanishes, because we have a “left over” J and obtain zero when setting all sources to
zero in the end of the calculation. By the same mechanism we have
hx̂2n+1 iβ = 0. (127)
Next we turn to

1 δ δ
hx̂2 iβ = W [J]
W [0] δJ(0) δJ(0)

J=0
1 δ 1
Z
dτ dτ 0 δ(τ )G(τ − τ 0 )J(τ 0 ) + J(τ )G(τ − τ 0 )δ(τ 0 ) W [J] = G(0). (128)

=
W [0] δJ(0) 2

J=0
So the mean square deviation of the oscillator’s position is equal to the Green’s function evaluated at zero.
1: Aside: Propagators vs. Green’s Function vs. Correlation functions
A frequent point of confusion is between propagators, correlation functions, and Green’s functions.
Often, QFT textbooks use these terms interchangeably, and in particular refer to Green’s functions of
classical equations of motion (such as the ones defined above) simply as ‘propagators’. This is true in
some sense” given the value of a field at some time, the Green’s function tells you how to determine
the field at some other time, by ‘propagating’ it forward in time. (In our QM examples, x is a field,
since QM can be thought of as a 0+1D QFT; one part of the confusion in translating between QM
and QFT is that the position x in QM is a field, whereas in a QFT x labels a field.) Viewed in the
language of path integrals in imaginary time, this is the same as a correlation function. Given certain
analytic properties of how to treat poles, the imaginary-time and real-time expressions can be linked
by analytic continuation (a trick known as ‘Wick rotation’).
However, what we have called the propagator
i 0
K(x, t; x0 , t0 ) = hx|Û (t, t0 )|x0 i = hx|e− ~ H(t−t ) |x0 i (129)
is a little bit different: it is really the representation of the unitary time evolution operator (as
encoded by the Sc hrödinger equation) in the position basis. Crucially, it is not a Green’s function of
the Schrödinger equation. To see this, consider the example of the free particle propagator (70)
m
r im
(x−x0 )2
K(x, t; x0 , t0 ) = 0
e 2~|t−t0 | . (130)
2πi~|t − t |
19
One can readily check that
~2 ∂ 2

∂ 0 0 ∂
i~ − Hx K(x, t; x , t ) = i~ + K(x, t; x0 , t0 ) = 0 (131)
∂t ∂t 2m ∂x2
where we use Hx to remind ourselves that we are working in the position basis. In other words,
K is a homogeneous solution to the differential equation. A Green’s function GS (x, t; x0 , t0 ) for the
Schrödinger equation should instead satisfy

∂
i~ − Hx GS (x, t; x0 t0 ) = δ(t − t0 )δ(x − x0 ) (132)
∂t
– i.e., it is a solution with a δ-function source term. Consider now the object i~ 1
Θ(t − t0 )K(x, t; x0 ; t0 ).
Evidently, we have

∂
i~ − Hx Θ(t − t0 )K(x, t; x0 ; t0 ) = δ(t − t0 )K(x, t; x0 ; t0 ) (133)
∂t
where we have used that K is a homogenous solution of i~∂t − Hx and the fact that the derivative of
a step function is a δ-function. Examining the limiting behaviour of K for |t − t0 | → 0 (the only part
of it that is relevant to the RHS of the expresssion above, due to the δ(t − t0 ) prefactor) we see that
it reduces to a spatial δ-function: in other words, GS (x, t; x0 , t0 ) ≡ Θ(t − t0 )K(x, t; x0 ; t0 ) is a Green’s
function for the Schrödinger operator i~∂t − Hx .
However, notice that this is not especially useful to us: a Green’s function solves a linear differential
equation with driving terms, but this is not what we want to do with regard to the Schrödinger
equation. (In contrast, we frequently do want to do this with respect to classical equations of motion.)
Rather than a problem with a driving term, we want to solve an initial value problem: given ψ(x0 , 0),
what is ψ(x, t)? The answer is
Z ∞
ψ(x, t) = dx0 K(x, t; x0 , 0)ψ(x0 , 0). (134)
∞
To see this, we can use the property that (i~∂t − Hx )K(x; t, x0 , t0 ) = 0, which means ψ(x, t) satisfies
the time-dependent Schrödinger equation, and the limiting property limt→t0 K(x, t; x0 , t0 ) = δ(x − x0 )
to confirm that limt→0 ψ(x, t) = ψ(x, 0) so that ψ(x, t) satisfies the correct initial value condition.
We could have also defined what we have called the propagator directly using the language of differ-
ential equations: K is simply the solution to the Schrödinger equation (i~∂t − Hx )K(x; t, x0 , t0 ) = 0
(a homogenous PDE) with the time-Dirichlet boundary conditionK(x, t; x0 t0 ) = δ(x − x0 ). In the
PDE literature, such a solution is usually termed the kernel. (134) makes it clear what the kernel
propagates: a quantum state (written as a wavefunction) forward in time. In other words, what we
have called the propagator is the kernel of the Schrödinger PDE, not its Green’s function.
Often in field theory, one works with various types of Green’s functions (things that solve an inho-
mogenous PDE with δ-function sources) but they are naturally written in terms of step functions and
kernels. The kernels themselves are usually related to correlators. Therefore, the distinction is often
ignored, but this leads to a perennial source of confusion. The key is to understand what equation
you are solving, and what you are using the quantity in question for – the rest is just names, and
to paraphrase the father of a famous physicist, there is a difference between knowing the name of
something and knowing something!
20
3.1.1 Imaginary Time Green’s Function of the Harmonic Oscillator
To determine G(τ ) we need to solve the differential equation (119). As G(0) = G(β~) we are dealing with
a periodic function and therefore may employ a Fourier expansion
∞
1 X
G(τ ) = √ gn eiωn τ , (135)
β~ n=−∞
where the Matsubara frequencies ωn are

2πn
ωn = . (136)
β~
Substituting this into the differential equation gives
∞ 2

1 X
iωn τ mωn κ
D̂G(τ ) = √ gn e + = δ(τ ). (137)
β~ n=−∞ ~ ~
dτ e−iωk τ on both sides fixes the Fourier coefficients and we obtain

R ~β
Taking the integral 0
∞
1 X ω2
G(τ ) = eiωn τ , (138)
βκ n=−∞ ω 2 + ωn2
p
where ω = κ/m. Using contour integration techniques this can be rewritten as
" #
~ω eω|τ | e−ω|τ |
G(τ ) = + . (139)
2κ e~βω − 1 1 − e−~βω
Setting τ = 0 gives
~ω ~ω 1 1
G(0) = = + . (140)
2κ tanh β~ω/2 κ eβ~ω − 1 2
We can relate this result to things we already know: using equipartition
hHiβ = hT iβ + hV iβ = 2hV iβ = κhx̂2 iβ = κG(0), (141)
we find that the average energy of the oscillator at temperature T is

1 1
hHiβ = ~ω β~ω + . (142)
e −1 2
Recalling that the Hamiltonian of the harmonic oscillator can be expressed as

1
H = ~ω n̂ + , (143)
2
where n̂ = a† a is the number operator, we recover the Bose-Einstein distribution
1
hn̂iβ = . (144)
eβ~ω −1
2: Aside: Contour integration for imaginary time Green’s function.
When working with finite-temperature Green’s functions, it is convenient to work in (Matsubara)

frequency space. This means that to go back to imaginary time, one frequently needs to compute
21
Matsubara summations of the form
∞
1 X
S= g(iωn ) (145)
β n−∞
where ωn = 2πn β and g is some (for now unspecified) function. The key to this is to use contour
integration. Recall that the basic result behind the utility of contour integration is that a contour
integral is completely determined by the values of its integrand at the poles, i.e.
1
I X
f (z)dz = Res[f (z)]|z=zn , (146)
C 2πi
poles zn of f ∈C
where we use the standard convention that contour integrals are positive when the contour is traversed
in an anticlockwise sense. The strategy is as follows: first, we find a function h(z) that has the following
properties:
1. the only singularities of h(z) are simple poles along the imaginary axis at the Matsubara fre-
quencies, i.e. at zn = i 2πn
β~
2. all residues of h(z) are identical: Res[h(zn )] = ζ;
3. the function f (z) = g(z)h(z) has the property that |g(z)h(z)| → 0 faster than 1/|z| as |z| → ∞.
We will also assume that any singularities of g(z) lie away from the imaginary axis. (Note that g more
or less must have singularities:a function that is bounded and analytic (as most physical correlators
will be) in the entire complex plane is constant, i.e. every ‘interesting’ function will have singularities!)
To write our sum S as a contour integral, we integrate the product f (z) over a contour Ca that runs
parallel to the imaginary axis from −i∞ to i∞ and ‘encircles’ it (Fig 2a; we can also view this as
hugging the imaginary axis and making little detours around the poles – the two are the same because
of the properties of analytic functions.) We have
I X X
I= f (z) = 2πi Res[g(z)h(z)]z=iωn = 2πi g(iωn )ζ = 2πiζβS (147)
Ca n n
where we have successively used (i) the residue theorem, (ii) our assumption that g has no poles on
the imaginary axis and the choice of poles of h; and (iii) the fact that all the residues of h are identical
and equal to ζ.
Now, we evaluate I in a different way. Recall that we can deform the contour in the complex plane,
as long as we don’t cross any singularities; the value of the contour integral is unchanged. Using this
freedom, we deform Ca into Cb . We push Cb out as far as it will go, in particular pulling it far from
any singularities of h; by ‘pinching’ around singularities of g (whether poles or branch cuts), we can
deform Cb all the way until it consists just of the ‘pinches’ and an integral around a large circle of
radius R → ∞ (Fig.2b). This gives us another expression for I
X I
I = IR + f (z)dz. (148)
Cs encircling singularities of g Cs
We now use the property (3) related to our choice of h above: this allows us to take IR = 0. Thus,
comparing (147) and (148), we have
1 X I
S= g(z)h(z)dz (149)
2πiβζ Cs
Cs encircling singularities of g
Crucially, the RHS only depends on the analytic structure of g, and we can evaluate the necessary
integrals using standard complex analysis. Usually, there will only be a finite number of siingularities,
22
meaning that it is much easier to do this than perform a complicated infinite sum. In the special case
where all the singularities of g are simple poles, we can use the residue theorem on the RHS, to find
1 X
S=− h(zk )Res[g(zk )], (150)
βζ
poles zk of g
where the − comes from the fact that the poles of g are encircled in the clockwise sense. Branch cuts
can be handled with slightly more effort.
Evidently, the above approach relies on our ability to choose a simple form for h. There are two
especially convenient choices:
1 1
h+ (z) = or h− (z) = . (151)
eβ~z − 1 1 − e−β~z
Notice that h± (z) are respectively exponentially vanishing for Re(z) → ±∞, and have residue ζ = β1
at each of their poles at z = iωn . Our choice of which of h± to use is determined by the choice of g in
(145). For the example of the Green’s function, f (z) is only convergent as demanded by property (3)
above for one choice of h = h± , depending on the sign of τ , but there is always such a choice. You
are encouraged to try this out in order to reproduce the Green’s function above.
Figure 2: Contours used to evaluate Matsubara sums.
3.2 Homework Question 6

Question 6. In this question the objective is to evaluate the Feynman path integral in one of the relatively few
cases, besides those treated in lectures, for which exact results can be obtained. The system we consider consists
of a particle of mass m moving on a circle of circumference L. The quantum Hamiltonian is
~2 d2
H=−
2m dx2
and wavefunctions obey ψ(x + L) = ψ(x). We want to determine the imaginary time propagator
hx1 | exp(−βH)|x2 i .
23
a) What are the eigenstates and eigenvalues of H? As we are dealing with a free particle, we can determine
the propagator as in the lectures in a simple way by inserting resolutions of the identity in terms of the eigenstates
of H. Show that this leads to the following result
∞
β(2πn)2 ~2 [x1 − x2 ]

X 1
hx1 | exp(−βH)|x2 i = exp − + 2πin . (152)
n=−∞
L 2mL2 L
b) Next, approach this using a path integral in which paths x(τ ) for 0 ≤ τ ≤ β~ satisfy the boundary
conditions x(0) = x1 and x(β~) = x2 . The special feature of a particle moving on a circle is that such paths
may wind any integer number l times around the circle. To build in this feature, write
τ
x(τ ) = x1 + [(x2 − x1 ) + lL] + s(τ ),
β~
where the contribution s(τ ) obeys the simpler boundary conditions s(0) = s(β~) = 0 and does not wrap around
the circle. Show that the Euclidean action for the system on such a path is
m ds 2
Z β~
m 2
S[x(τ )] = Sl + S[s(τ )] where Sl = [(x2 − x1 ) + lL] and S[s(τ )] = dτ .
2β~ 0 2 dτ
c) using the results of b) show that

∞
X m 2
hx1 | exp(−βH)|x2 i = Z0 exp − [(x 1 − x 2 ) + lL] , (153)
2β~2
l=−∞
where Z0 is the diagonal matrix element hx|e−βH |xi for a free particle (i.e. without periodic boundary conditions)
moving in one dimension.
d) Argue on the basis of the result you obtained in Qu 3. for the propagator of a free particle that
1/2
m
Z0 = . (154)
2πβ~2
e) Show that the expressions in Eq. (152) and Eq. (153) are indeed equal. To do so, you should use the Poisson
summation formula
X∞ X∞
δ(y − l) = e−2πiny
l=−∞ n=−∞
(think about how to justify this). Introduce the left hand side of this expression into Eq. (153) by using the
relation, valid for any smooth function f (y),
∞
X Z ∞ ∞
X
f (l) = dy δ(y − l)f (y) ,
l=−∞ −∞ l=−∞
substitute the right hand side of the summation formula, carry out the (Gaussian) integral on y, and hence
establish the required equality.
3.3 Correlation Functions

It is clear from the above that we can calculate more general quantities from the generating functional W [J],
namely
n n
1 Y δ N
Z Y 1
W [J] = Dx(τ ) x(τj ) e− ~ SE [x(τ )] (155)
W [0] δJ(τj ) W [0]

j=1 J=0 j=1
24
τ
x(hβ )
hβ
τn ..
x(τn )
.
τ3 x( τ3 )
τ2
τ1
x( 0 )
Figure 3: Path integral corresponding to (155).
What is their significance? Graphically, the path integral in (155) is represented in Fig. 3. It consists of
several parts. The first part corresponds to propagation from x(0) to x(τ1 ) and the associated propagator
is
hx(τ1 )|e−Hτ1 /~ |x(0)i. (156)
The second part corresponds to propagation from x(τ1 ) to x(τ2 ), and we have a multiplicative factor of x(τ1 )
as well. This is equivalent to a factor
hx(τ2 )|e−H(τ2 −τ1 )/~ x̂|x(τ1 )i. (157)
Repeating this analysis for the other pieces of the path we obtain
 
n
Y
 hx(τj+1 )|e−H(τj+1 −τj )/~ x̂|x(τj )i hx(τ1 )|e−Hτ1 /~ |x(0)i , (158)
j=1
where τn+1 = ~β. Finally, in order to represent the full path integral (155) we need R to integrate over
the intermediate positions x(τj ) and impose periodicity of the path. Using that 1 = dx|xihx| and that
W [0] = Z(β) we arrive at
1
Z
dx(0)hx(0)|e−H(β−τn )/~ x̂e−H(τn −τn−1 )/~ x̂ . . . x̂e−H(τ2 −τ1 )/~ x̂e−Hτ1 /~ |x(0)i
Z(β)
1 h i
= Tr e−βH x̄(τn )x̄(τn−1 ) . . . x̄(τ1 ) , (159)
Z(β)
where we have defined operators
x̄(τj ) = eHτj /~ x̂e−Hτj /~ . (160)
There is one slight subtlety: in the above we have used implicitly that τ1 < τ2 < . . . < τn . On the other
hand, our starting point (155) is by construction symmetric in the τj . The way to fix this is to introduce a
time-ordering operation Tτ , which automatically arranges operators in the “right” order. For example
Tτ x̄(τ1 )x̄(τ2 ) = θ(τ1 − τ2 )x̄(τ1 )x̄(τ2 ) + θ(τ2 − τ1 )x̄(τ2 )x̄(τ1 ), (161)
where θ(x) is the Heaviside theta function. Then we have
n

1 Y δ 1 h i
W [J] = Tr e−βH Tτ x̄(τ1 )x̄(τ2 ) . . . x̄(τn ) .
W [0] δJ(τj ) Z(β)

j=1 J=0
(162)
25
Finally, if we analytically continue from imaginary time to real time τj → itj , the operators x̄(τ ) turn into
Heisenberg-picture operators
it it
x̂(t) ≡ e ~ H x̂e− ~ H . (163)
The quantities that we get from (162) after analytic continuation are called n-point correlation functions
1 h i
hT x̂(t1 )x̂(t2 ) . . . x̂(tn )iβ ≡ Tr e−βH T x̂(t1 )x̂(t2 ) . . . x̂(tn ) .
Z(β)
(164)
Here T is a time-ordering operator that arranges the x̂(tj )’s in chronologically increasing order from right
to left. Such correlation functions are the central objects in both quantum field theory and many-particle
quantum physics.
3.3.1 Wick’s Theorem

Recalling that
1
dτ dτ 0 J(τ )G(τ −τ 0 )J(τ 0 )
R
W [J] = W [0] e 2 , (165)
then taking the functional derivatives, and finally setting all sources to zero we find that
n

1 Y δ X
W [J] = G(τP1 − τP2 ) . . . G(τPn−1 − τPn ) . (166)
W [0] δJ(τj )

j=1 J=0 P (1,...,n)
Here the sum is over all possible pairings of {1, 2, . . . , n} and G(τ ) is the Green’s function (138). In particular
we have
hTτ x̄(τ1 )x̄(τ2 )iβ = G(τ1 − τ2 ). (167)
The fact that for “Gaussian theories” 3 like the harmonic oscillator n-point correlation functions can be
expressed as simple products over 2-point functions is known as Wick’s theorem.
3.4 Probability distribution of position

Using Wick’s theorem it is now straightforward to calculate all moments hx̂2n iβ for the harmonic oscillator.
It is instructive to calculate the corresponding probability distribution directly. Let us first work out the
relevant expectation value to consider. Let |ψi be an arbitrary state and consider
Z Z
hψ|δ(x̂ − x0 )|ψi = dxhψ|δ(x̂ − x0 )|xihx|ψi = dx δ(x − x0 )hψ|xihx|ψi = |ψ(x0 )|2 . (168)
So the expectation value of the delta-function indeed gives the correct result for the probability distribution
of a position measurement, namely the absolute value squared of the wave function. We then have
Z ∞
dk ik(x̂−x0 )
hδ(x̂ − x0 )iβ = he iβ
−∞ 2π
Z ∞
dk −ikx0 N
Z
1 ~β
R R ~β
= e Dx(τ ) e− 2 0 dτ x(τ )D̂x(τ )+ 0 dτ x(τ )ikδ(τ ) . (169)
−∞ 2π W [0]
This is a special case of our generating functional, where the source is given by J(τ ) = ikδ(τ ). We therefore
can use (124) to obtain
Z ∞ Z ∞
dk −ikx0 W [ikδ(τ )] dk −ikx0 1 R0~β dτ R0~β dτ 0 ikδ(τ ) G(τ −τ 0 ) ikδ(τ 0 )

hδ(x̂ − x0 )iβ = e = e e 2
−∞ 2π W [0] −∞ 2π
1 2
= p e−x0 /2G(0) . (170)
2πG(0)
3
These are theories in which the Lagrangian is quadratic in the generalized co-ordinates.
26
To go from the first to the second line we have taken the integrals over τ and τ 0 (which are straightforward
because of the two delta functions) and finally carried out the k-integral using the one dimensional version
of (34). We see that our probability distribution is a simple Gaussian with a variance that depends on
temperature through G(0). Note that at zero temperature (170) reduces, as it must, to |ψ0 (x0 )|2 , where
ψ0 (x) is the ground state wave function of the harmonic oscillator.
Exercise 4
If you have never worked out the harmonic oscillator at finite temperature before using “Heisenberg
QM”, here is how to do it. Defining creation and annihilation operators by
r r
mω ip̂ † mω ip̂
a= x̂ + , a = x̂ − , (171)
2~ mω 2~ mω
brings the Hamiltonian to the form

† 1
H = ~ω a a + . (172)
2
The creation/annihilation operators act on energy eigenstates |ni as
√ √
a|ni = n|n − 1i , a† |ni = n + 1|n + 1i. (173)
The energy eigenvalues are
1
En = ~ω n + . (174)
2
The partition function is
∞
X e−β~ω/2
Z(β) = e−βEn = (175)
1 − e−β~ω
n=0
Similarly you can calculate
∞
1
−βH †
X 1
Tr e aa = ne−βEn = β~ω . (176)
Z(β) e −1
n=1
Time dependent operators are defined following (160)

ā(τ ) = eHτ /~ ae−Hτ /~ , ā† (τ ) = eHτ /~ a† e−Hτ /~ . (177)
Note that
ā† (τ ) 6= (ā(τ ))† . (178)
Taking derivatives this gives the following equations of motion
d 1
ā(τ ) = eHτ /~ [H, a]e−Hτ /~ = −ωā(τ ). (179)
dτ ~
Integrating this as well as the analogous equation for ā† (τ ) gives
ā(τ ) = e−ωτ a , ā† (τ ) = eωτ a† . (180)
This allows us to compute imaginary-time dependent thermal averages
1 e−ωτ
Tr e−βH ā(τ )ā† (0) = ,
Z(β) 1 − e−β~ω
1 eωτ
Tr e−βH ā† (τ )ā(0) = β~ω . (181)
Z(β) e −1
Finally, recalling the relation between x̂ and a and a† we find
~ω eωτ e−ωτ

1
−βH
Tr e x̄(τ )x̄(0) = + . (182)
Z(β) 2κ eβ~ω − 1 1 − e−β~ω
27
3.5 Perturbation Theory and Feynman Diagrams
Let us now consider the anharmonic oscillator
p̂2 κ λ
H= + x̂2 + x̂4 . (183)
2m 2 4!
As you know from QM2, this Hamiltonian is no longer exactly solvable. What we want to do instead is
p̂2
perturbation theory for small λ > 0. As the Hamiltonian is of the form H = 2m + V (x̂) our previous
construction of the path integral applies. Our generating functional becomes
Z
1 λ 4
R ~β R ~β
Wλ [J] = N Dx(τ ) e− ~ SE [x(τ )]+ 0 dτ J(τ )x(τ )− 4!~ 0 dτ x (τ ) . (184)
The partition function is

Zλ (β) = Wλ [0] . (185)
The idea is to expand (184) perturbatively in powers of λ

λ
Z Z ~β
1
R ~β
Wλ [J] = N Dx(τ ) 1 − dτ x (τ ) + . . . e− ~ SE [x(τ )]+ 0 dτ J(τ )x(τ )
0 4 0
4!~ 0
" 4 #
λ δ
Z Z ~β
− ~1 SE [x(τ )]+ 0~β dτ J(τ )x(τ )
R
0
= N Dx(τ ) 1 − dτ + . . . e
4!~ 0 δJ(τ 0 )
Z h i 4
− λ ~β dτ 0 δJ(τ
δ
R
1
R ~β
= N Dx(τ ) e 4!~ 0 0)
e− ~ SE [x(τ )]+ 0 dτ J(τ )x(τ )
h i4
λ δ
dτ 0
R ~β
− 4!~ δJ(τ 0 )
= e 0
W0 [J]. (186)
We already know W0 [J]

1
dτ dτ 0 J(τ )G(τ −τ 0 )J(τ 0 )
R
W0 [J] = W0 [0] e 2 , (187)
which will enable us to work out a perturbative expansion very efficiently.
3.5.1 Partition Function of the anharmonic oscillator

By virtue of (185) the perturbation expansion for Zλ (β) is
4
R ~β 0 h δ i4
λ δ
λ Z ~β
− 4!~ dτ 0 0
Zλ (β) = e 0 δJ(τ ) W0 [J] = Z0 (β) − dτ W0 [J]

4!~ 0 δJ(τ 0 )

J=0 J=0
4
1 λ 2 ~β 0 00
Z 4
δ δ
+ dτ dτ W0 [J] + . . .

2 4!~ 0
δJ(τ ) 00
δJ(τ )

0
J=0
= Z0 (β) 1 + λγ1 (β) + λ2 γ2 (β) + . . . .

(188)
1. First order perturbation theory:

Carrying out the functional derivatives gives
λ ~β
λβ
Z
λγ1 (β) = − dτ 0 [G(τ − τ )]2 = − [G(0)]2 . (189)
8~ 0 8
This contribution can be represented graphically by a Feynman diagram. In order to do so we introduce the
following elements:
1. The two-point function G(τ − τ 0 ) is represented by a line running from τ to τ 0 .
28
Figure 4: Graphical representation of the Green’s function.
Figure 5: Graphical representation of the interaction vertex.
Figure 6: Feynman diagram for the 1st order perturbative contribution to the partition function.
λ
R ~β
2. The interaction vertex − 4!~ 0 dτ is represented by by the 4-point vertex shown in Fig. 5.
Combining these two elements, we can express the integral λγ1 (β) by the diagram shown in Fig. 6.
Here the factor of 3 is a combinatorial factor associated with the diagram. We will discuss this more
below. To see why it is quite natural to introduce Feynman diagrams in the way we have done let us go
back to taking functional derivatives. Let us introduce a short-hand notation

1 1 1
W [J] = W [0]e 2 J1 G12 J2 = W [0] 1 + J1 G12 J2 + 3 J1 G12 J2 J3 G34 J4 + . . . . (190)
2 2
The last term we have written is the one that gives rise to our first order diagram, so we have a factor 213
to begin with. The first order contribution is worked out in Fig. 7. There are 4 ways of connecting the first
Figure 7: Combinatorics for the first order diagram
line to the vertex. In the next step, we can connect the second line in the vertex either to the remaining
29
external line (giving a factor of 2 as we could connect to either end), or to the line in the vertex that was
connected in the first step. In the last step we complete the diagram.
Let us now turn to the combinatorial factor associated with the Feynman diagram. This is simply equal
to the ways of connecting the lines in our vertex in such a way as to reproduce the Feynman diagram we
want. As shown in Fig. 8 there are three ways.
Figure 8: Combinatorial factor for the first order vacuum diagram.
2. Second order perturbation theory.

To second order we obtain a number of different contributions upon taking the functional derivatives.
The full second order contribution is
1 λ 2
Z ~β Z ~β
2
λ γ2 (β) = 72 dτ dτ 0 G(τ − τ )G2 (τ − τ 0 )G(τ 0 − τ 0 )
2 4!~ 0 0
1 λ 2
Z ~β Z ~β
+ 24 dτ dτ 0 G4 (τ − τ 0 )
2 4!~ 0 0
2
1 λ 2
Z ~β
+ 9 dτ G2 (τ − τ ) . (191)
2 4!~ 0
The corresponding Feynman diagrams are shown in Fig.9. They come in two types: the first two are
connected, while the third is disconnected. The factor of 1/2 that has been separated off is associated with
second order perturbation theory.
Figure 9: Feynman diagram for the 2nd order perturbative contribution to the partition function.
The point about the Feynman diagrams is that rather than carrying out functional derivatives and then
representing various contributions in diagrammatic form, in practice we do the calculation by writing down
the diagrams first and then working out the corresponding integrals! How do we know what diagrams to
30
draw? As we are dealing with the partition function, we can never produce a diagram with a line sticking
out: all (imaginary) times must be integrated over. Such diagrams are sometimes called vacuum diagrams.
An important part of the process is to work out the correct combinatorial factors. For the second order
vacuum diagrams this is shown in Figs 10, 11 and 12.
Figure 10: To produce the first diagram in Fig. 9 we need to make one loop at each vertex (there are six
ways to do this), and then to connect the two resulting pieces (there are two ways of doing this).
Figure 11: To produce the second diagram in Fig. 9 we need to connect all lines of the first vertex with the
second vertex. There are 4 ways of connecting the first line, 3 ways of connecting the second, and two ways
of connecting the third.
Figure 12: To produce last second diagram in Fig. 9 we need to connect all lines at each individual vertex.
As discussed before there are 3 ways of doing this.
3: Aside: Symmetry Factors.
You might have wondered why we have the 4!1 normalization in front of the λx4 term. This is designed
so that if we had to, e.g. consider a process with 4 external legs (a 4-point correlation function) then
the leading connected contribution which appears at O(λ) comes with a coefficient of 1. This follows
because there are 4! different ways in which we can connect an external leg to the vertex. However,
when we have loops, then two different contractions could give the same diagrammatic result. There
are two ways to account for this. The most straightforward is the route taken in these notes, which is
to simply compute the number of ways a particular diagram can appear and keep all the factors of
1/4! in front, and cancel off against the combinatorics as needed. The more elegant approach, adopted
by expert practitioners, is to work with symmetry factors: it turns out that the ratio between the
combinatorics (the number of ways of getting a particular diagram) and the factorials is related to the
symmetries of the Feynman diagram. So one can ‘read off’ this ratio without going through tedious
combinatorics.
31
Put simply: one can either compute a large combinatorial factor and divide with various powers of
the 4! terms, or else one can just compute the symmetry factor S and divide the diagram by it.
For example, the symmetry factor for the ‘figure of 8’ diagram in the partition function is given by
S −1 = 18 = 4!3 .
Familiarity with symmetry factor technology is often crucial to doing high order diagrammatic per-
turbation theory, but not terribly important for our present purposes.

Question 7. Anharmonic Oscillator. Consider the anharmonic oscillator
p̂2 κ λ1 λ2
H(λ1 , λ2 ) = + x̂2 + x̂3 + x̂4 . (192)
2m 2 3! 4!
where κ, λ1,2 > 0 and λ21 − 3κλ2 < 0. Define a generating functional by
Z nh i o
− ~1 SE [x(τ )]+ 0~β dτ J(τ )x(τ ) +U
R
x(τ )
Wλ1 ,λ2 [J] = N Dx(τ ) e , (193)
where
m d2

1 ~β
λ1 3 λ2 4 κ
Z

U x(τ ) = − dτ x (τ ) + x (τ ) , D̂ = − + . (194)
~ 0 3! 4! ~ dτ 2 ~
a) Show that the partition function is equal to
Zλ1 ,λ2 (β) = Wλ1 ,λ2 [0]. (195)
b) Show that the generating functional can be expressed in the form

δ
Wλ1 ,λ2 [J] = exp U W0,0 [J]. (196)
δJ(τ )
c) Determine the first order perturbative corrections in λ1 and λ2 to the partition function. Draw the corresponding
Feynman diagrams.
d) Determine the perturbative correction to the partition function proportional to λ21 . Draw the corresponding
Feynman diagrams. Are there corrections of order λ1 λ2 ?
e)∗ Determine the first order corrections to the two-point function
hTτ x̄(τ1 )x̄(τ2 )iβ . (197)
Draw the corresponding Feynman diagrams. What diagrams to you get in second order in perturbation theory?
4: Aside: Response Functions
To see why correlation functions are important for applications to experiments let us briefly summarize
the main facts on linear response theory. To ease notations let us introduce the thermal density matrix
1 −βH 1 X −βEn
ρ= e = e |nihn|, (198)
Z(β) Z(β) n
where H|ni = En |ni. Then
hOiβ = Tr (ρO) . (199)
32
The time evolution of ρ follows from that of the energy eigenstates
|n(t)i = U (t, t0 )|n(t0 )i ⇒ ρ(t) = U (t, t0 )ρ(t0 )U † (t, t0 ) (200)
Let us now consider a time-dependent Hamiltonian
H = H0 + λ(t)O1 , (201)
where we take λ(t) to be very small at all times. In practice the perturbation λ(t)O1 will correspond
to some external experimental probe. Time-dependent perturbation theory tells us that the time-
evolution operator in this case can be written to leading order in λ(t) as
i t 0
Z
0 †
U (t, t0 ) = U0 (t, t0 ) 1 − dt λ(t )U0 (t, t0 )O1 U0 (t, t0 ) + . . . , (202)
~ t0
where
i
U0 (t, t0 ) = e− ~ H0 (t−t0 ) . (203)
Let us denote Heisenberg-picture operators in the unperturbed theory (λ(t) = 0) by
O(0) (t) = U0† (t, t0 )O(t0 )U0 (t, t0 ). (204)
Let us now look at the thermal average of an observable O
Tr (ρ(t)O) = Tr (ρO(t)) , (205)
where
t
i
Z
† (0)
O(t) = U (t, t0 )OU (t, t0 ) ≈ O (0)
(t) − dt0 λ(t0 )[O(0) (t), O1 (t0 )] (206)
~ t0
Let us now imagine that we switch on the perturbation at t0 = −∞ and calculate the change in the
thermal average of an observable O induced by the perturbation
1 ∞ 0
Z
Tr (ρ(t)O) − Tr (ρ(−∞)O) ≈ dt λ(t0 ) χOO1 (t, t0 ) ,
~ −∞
(0)
χOO1 (t, t0 ) = −iΘ(t − t0 )h[O(0) (t), O1 (t0 )]iβ . (207)
Here χOO1 (t, t0 ) is called a generalized susceptibility. Next comes a bit of magic. First, observe
that the susceptibility in fact depends only on the time difference, i.e. χOO1 (t, t0 ) = χOO1 (t − t0 ).
Considering the Fourier transform then gives
Z ∞ Z ∞
iωt (0)
χeOO1 (ω) = dt e χOO1 (t) = −i dt eiωt h[O(0) (t), O1 (0)]iβ (208)
−∞ 0
Inserting a resolution of the identity in terms of the eigenstates of H0 gives the following Lehmann
representation
Z ∞ X e−βEn h i
iωt (0) (0) (0) (0)
eOO1 (ω) = −i
χ dt e hn|O (t)|mihm|O1 (0)|ni − hn|O1 (0)|mihm|O (t)|ni
0 n,m
Z(β)
X e−βEn − e−βEm Z ∞

(0) (0)
= hn|O (0)|mihm|O1 |ni(−i) dt ei(ω−Em +En )t
n,m
Z(β) 0
1 X e−βEn − e−βEm

(0)
= hn|O(0) (0)|mihm|O1 |ni. (209)
Z(β) n,m ω + i0 − Em + En
33
Now consider the imaginary-time correlator
−hTτ O(τ )O1 (0)iβ (210)
This is a periodic function of τ and its Fourier transform is

Z ~β
χ̄OO1 (ω` ) = − dτ eiω` τ hTτ O(τ )O1 (0)iβ
0
1 X e−βEn − e−βEm

(0)
= hn|O(0) (0)|mihm|O1 |ni. (211)
Z(β) n,m iω` − Em + En
Comparing (211) and (209) we see that the Fourier transform of the dynamical susceptibility is equal
to the Fourier transform of the imaginary-time correlator, if we analytically continue from Matsubara
frequencies to real frequencies
iω` −→ ω + i0. (212)
5: Aside: Response Functions for the Harmonic Oscillator
Let us now consider the explicit example on a harmonic oscillator that is perturbed by an infinitesimal
electric field E(t)
p̂2 κ
H(t) = + x̂2 − E(t)x̂ (213)
2m 2
Then according to our general result (207) the change in the average position is
1 ∞ 0
Z
Tr (ρ(t)x̂) − Tr (ρ(−∞)x̂) ≈ dt E(t0 ) χx̂x̂ (t − t0 ) ,
~ −∞
1 ∞ 0
Z ∞
dω −iω(t−t0 )
Z
= dt E(t0 ) e χ
ex̂x̂ (ω). (214)
~ −∞ −∞ 2π
According to (167), (138) we have

∞
1 X ω2
hTτ x̄(τ )x̄(0)iβ = G(τ ) = eiωn τ , (215)
βκ n=−∞ ω + ωn2
2
which results in
~ ω2
χ̄x̂x̂ (ω` ) = − . (216)
κ ω 2 + ω`2
Analytically continuing this to real frequencies we obtain

~ ω ω
χ
ex̂x̂ (z) = − , (217)
2κ z + i0 − ω z + i0 + ω
and Fourier transforming back gives

~ω
χx̂x̂ (t) = −θ(t) sin(ωt) . (218)
κ
This susceptibility equals the retarded correlation function of the position operator and, interestingly,
is independent of temperature.
34
Part II
Path Integrals and Transfer Matrices
4 Relation of D dimensional quantum systems to D + 1 dimensional
classical ones
Let’s start by defining what we mean by the spatial “dimension” D of a system. Let us do this by considering
a (quantum) field theory. There the basic objects are fields, that depend on time and are defined at all
points of a D-dimensional space. This value of D defines what we mean by the spatial dimension. For
example, in electromagnetism we have D = 3. In this terminology a single quantum mechanical particle or
spin are zero-dimensional systems. On the other hand, a linear chain of spins is a one-dimensional system,
while a bcc lattice of spins has D = 3. Interestingly, there is a representation of D dimensional quantum
systems in terms of D + 1 dimensional classical ones. We will now establish this for the particular case of
the simple quantum mechanical harmonic oscillator.
4.1 Some Facts from Statistical Physics

Consider a classical many-particle system coupled to a heat bath at temperature T . The partition function
is defined as X 1
Z= e−βE(C) , β= . (219)
kB T
configurations C
Here the sum is over all possible configurations C, and E(C) is the corresponding energy. Thermal averages
of observables are given by
1 X
hOiβ = O(C)e−βE(C) , (220)
Z
configurations C
where O(C) is the value of the observable O in configuration C. The average energy is
1 X ∂
E= E(C)e−βE(C) = − ln(Z). (221)
Z ∂β
configurations C
The free energy is

F = −kB T ln(Z). (222)
The entropy is
E−F ∂ ∂
S= = kB ln(Z) − kB β ln(Z) = kB [T ln(Z)] . (223)
T ∂β ∂T
4.2 Quantum Mechanical Particle

Let us revisit the path-integral representation (108) for the partition function of our QM particle at tem-
perature β
N −1
!
h m iN Z X m xn+1 − xn 2
Z
2
Z(β) = lim dx dx1 . . . dxN −1 exp − + V (xn ) , (224)
N →∞ 2π~ ~ 2
n=0
where we have set x0 = xN = x. For a given value of N , this can be interpreted as the partition function
of N classical degrees of freedom xj , that can be thought of as deviations of classical particles from their
equilibrium positions, cf. Fig. 13. In this interpretation V (xj ) is simply a potential energy associated with
moving the j th particle a distance xj away from its equilibrium position, while m 2 2
2 (xn+1 − xn ) / describes
35
Figure 13: Periodic array of classical particles.
an interaction energy that favours equal displacements, i.e. xn = xn+1 . Importantly, the temperature Tcl of
this one-dimensional classical model equals
~
kB Tcl = = N kB T. (225)

So for large values of N (and fixed T ) this temperature is very large. A convenient way for working out
partition functions in classical statistical mechanics is by using transfer matrices. In the case at hand, this
is defined as an integral operator T̂ with kernel
r
0 m − β Ecl (x,x0 )
T (x, x ) = e N ,
2π~
m x − x0 2 V (x) + V (x0 )

0
Ecl (x, x ) = + . (226)
2 2
The integral operator T̂ acts on functions f (x) as

Z
(T̂ ∗ f )(x) = dx0 T (x, x0 )f (x0 ). (227)
In terms of this transfer matrix the partition function can be written as

Z
Z(β) = lim dxdx1 . . . dxN −1 T (x, x1 )T (x1 , x2 ) . . . T (xN −1 , x)
N →∞
Z
= lim dx T N (x, x) = lim Tr(T̂ N ). (228)
N →∞ N →∞
By construction T̂ is a real, symmetric operator and can therefore be diagonalized. Hence the partition
function can be expressed in terms of the eigenvalues of T̂ using
X
Tr(T̂ N ) = λN
n. (229)
n
In order to get a clearer idea how to use transfer matrices in statistical mechanics problems we now turn
to a simpler example, the celebrated Ising model. This is in fact the key paradigm in the theory of phase
transitions.
36
6: Aside: Recap of the quantum-classical correspondence
It is important to realise that we have used the path integral formulation to make links between
the transition amplitude and two different partition functions. It’s useful at this stage to review
the logical development to see what these are, working with the quantum-mechanical particle as an
example, so that everything is collected in one place.
The fundamental object is the Feynman propagator for a real time interval TR , written for specificity
in the basis of position eigenstates:
F (x0 , x1 , TR ) = hx1 |e−iHTR /~ |x0 i (230)
We can write this as a path integral following the treatment so far:

( Z !)
i TR m dx 2
Z
F (x0 , x1 , TR ) = x(0)=x Dx(t) exp dt − V (x(t)) (231)
0 ~ 0 2 dt
x(TR )=x1
Let us imagine propagating the particle for an imaginary time TI . The relevant propagator is obtained
by replacing TR → −iTI in F :
Z(x0 , x1 , TI ) = hx1 |e−HTI /~ |x0 i

= F (x, x0 , −iTI )
( Z 2 !)
i −iTI

m dx
Z
= Dx(t) exp dt − V (x(t)) . (232)
x(0)=x0 ~ 0 2 dt
x(−iTI )=x1
Now, let’s change variables and define τ = it. Then, we see that
( !)
1 TI m dx 2
Z Z
Z(x0 , x1 , TI ) = x(0)=x Dx(τ ) exp − dτ + V (x(τ )) . (233)
0 ~ 0 2 dτ
x(TI )=x1
Note that the conditions on x in the path integral are now written for x(τ ) rather than x(t). The
expression in the exponential is the Euclidean action defined previously. We can now use (233) in
two ways.
Statement I: We can link F (via Z) to a classical

statistical mechanical problem. Define a classical
1 TI m dx 2
R
energy Ecl [x(t)] = δ 0 dτ 2 dτ + V (x(τ ) . This is the energy of an elastic string of length TI
and tension m, vibrating in one dimension are captured by coordinate x(τ ), and which has its ends
fixed at x(0) = x0 and x(TI ) = x1 . Here δ is an arbitrary constant with units of time, since the
integral manifestly has units of action=energy×time. Thus, we have (writing TI = T )
Z
δ
F (x0 , x1 , −iT ) = Z(x0 , x1 , T ) = x(0)=x e− ~ Ecl [x(τ )] (234)
0
x(T )=x1
Notice that the choice of δ is arbitrary – the physical quantity is the dimensionless ratio ~δ Ecl [x(τ )].
However given some choice of ‘energy scale’, one can interpret the RHS as the partition function of a
classical statistical mechanical problem at temperature Tcl = ~/δkB ). [In the main part of the notes,
we took a slightly different route in which we also discretized the integrals. This introduced a similar
small time scale (see (225)), with the arbitrariness being the number of slices N , but we were able
to take a sensible N → ∞ limit. Note also that it is very tempting to just take δ = TI so that there
is proportionality between βcl and the total imaginary time TI . However, this choice would challenge
the interpretation of Ecl as a classical energy, since it would then no longer be extensive in the ‘length’
37
parameter TI . So, it is better to let δ remain finite as we take TI → ∞ so that the classical statistical
mechanical analogy really makes sense.]
We have thus demonstrated that the transition amplitude for a quantum particle over an imaginary
time −iTI is equal to the classical partition function for a string of length TI at ends fixed at x0 and
x1 , at a classical temperature ∝ 1/~. Note that the classical problem has D + 1 dimensions, since
the time direction of the quantum problem has become a classical ‘spatial’ coordinate. This is OK,
because classical statistical mechanics does not have any notion of time — we average over ensembles
since we have assumed (by the ergodic hypothesis) that this is the same as a time average — and
once you average over a quantity, the theory does not depend on it anymore.
Statement II: We can also link F to the partition function of a quantum statistical mechanical
problem (this was the thing we did first in the lecture notes). Since the partition function of a quantum
statistical mechanical prooblem with Hamiltonian H and inverse temperature β is Zqu (β) = Tre−βH ,
we see that
Z Z
−βH
Zqu (β) = dxhx|e |xi = dxF (x, x, −iβ~). (235)
We have already seen how this leads to a path integral over periodic paths. So, we have argued that
the quantum mechanical partition function at temperature β is given by the probability amplitude for
the particle to return to its original position in an imaginary time β~.
The above discussion should make it clear that there are really two separate statements being made
about the transition amplitude. To summarize: F (x, x0 , −iT ) captures two different statistical me-
chanical problems depending on how we choose to interpret the imaginary time parameter T . We can
view T as a ‘length’ of classical system with fixed end points at some effective classical temperature
Tcl ∝ β −1 , and thereby identify a classical statistical mechanical problem with one extra dimension,
or we can set T = β~ take the trace of F to obtain the quantum partition function.
What we have done in effect in the preceding discussion in these notes is to combine these two pictures.
From (234) and (235) we have
Z Z
Zqu (β) = dxF (x, x, −iβ~) = dxZ(x, x, β~) = Zclβ~ (Tcl ∝ ~−1 ) (236)
where on the RHS because of the trace we now have the classical partition function for a string that
obeys periodic boundary conditions and finite extent β~ in the ‘extra’ spatial direction corresponding
to imaginary time. Since none of our results relied on starting with a single QM particle, we can
extend them to arbitrary D. We have therefore formally related a quantum statistical mechanical
partition function for a D dimensional quantum system to the classical statistical mechanical partition
function for a D + 1 dimensional problem.
The quantum classical correspondence is subtle, and at this point it is best to view as a beautiful
formal step. To quote Alexandre Polyakov (arguably the greatest living quantum field theorist): Our
derivation of these analogies was technical. I feel that there are deeper reasons for them, connected
with the properties of space-time. [...] At the moment our aims are more modest – we are going to
exploit these analogies in concrete problems. — A.M. Polyakov, in Gauge Fields and Strings.
5 The Ising Model

Ferromagnetism is an interesting phenomenon in solids. Some metals (like Fe or Ni) are observed to acquire a
finite magnetization below a certain temperature. Ferromagnetism is a fundamentally quantum mechanical
effect, and arises when electron spins spontaneously align along a certain direction. The Ising model is a
38
very crude attempt to model this phenomenon. It is defined as follows. We have a lattice in D dimensions
with N sites. On each site j of this lattice sits a “spin” variable σj , which can take the two values ±1.
These are referred to as “spin-up” and “spin-down” respectively. A given set {σ1 , σ2 , . . . , σN } specifies a
configuration. The corresponding energy is taken to be of the form
X N
X
E({σj }) = −J σi σj − h σj , (237)
hiji j=1
where hiji denote nearest-neighbour bonds on our lattice and J > 0. The first term favours alignment
on neighbouring spins, while h is like an applied magnetic field. Clearly, when h = 0 the lowest energy
states are obtained by choosing all spins to be either up or down. The question of interest is whether the
Ising model displays a finite temperature phase transition between a ferromagnetically ordered phase at low
temperatures, and a paramagnetic phase at high temperatures.
5.1 Statistical mechanics of the Ising model

The partition function of the model is
X X X
Z= ··· e−βE({σj }) . (238)
σ1 =±1 σ2 =±1 σN =±1
The magnetization per site is given by

N
1 X 1 ∂
m(h) = h σj iβ = ln(Z). (239)
N N β ∂h
j=1
The magnetic susceptibility is defined as
∂m(h) 1 ∂2
χ(h) = = ln(Z). (240)
∂h N β ∂h2
Substituting the expression (238) for the partition function and then carrying out the derivatives it can be
expressed in the form
N
β X
χ(h) = hσl σm iβ − hσl iβ hσm iβ . (241)
N
l,m=1
5.2 The One-Dimensional Ising Model

The simplest case is when our lattice is one-dimensional, and we impose periodic boundary conditions. The
energy then reads
N X N
X h
E= −Jσj σj+1 − (σj + σj+1 ) ≡ E(σj , σj+1 ), (242)
2
j=1 j=1
where we have defined

σN +1 = σ1 . (243)
The partition function can be expressed in the form
N
X Y
Z= e−βE(σj ,σj+1 ) . (244)
σ1 ,...,σN j=1
It can be evaluated exactly by means of the transfer matrix method.
39
5.2.1 Transfer matrix approach
The general idea is to rewrite Z as a product of matrices. The transfer matrix T is taken to be a 2 × 2
matrix with elements
0
Tσσ0 = e−βE(σ,σ ) . (245)
Its explicit form is
eβ(J+h) e−βJ

T++ T+−
T = = . (246)
T−+ T−− e−βJ eβ(J−h)
The partition function can be expressed in terms of the transfer matrix as follows
X
Z= Tσ1 σ2 Tσ2 σ3 . . . TσN −1 σN TσN σ1 (247)
σ1 ,...,σN
As desired, this has the structure of a matrix multiplication
Z = Tr T N .

(248)
The trace arises because we have imposed periodic boundary conditions. As T is a real symmetric matrix,
it can be diagonalized, i.e.
† λ+ 0
U TU = , (249)
0 λ−
where U is a unitary matrix and
q
λ± = eβJ cosh(βh) ± e2βJ sinh2 (βh) + e−2βJ . (250)
Using the cyclicity of the trace and U U † = I, we have
λN

† N

† N

† N
0
Z = Tr U U T = Tr U T U = Tr [U T U ] = Tr + = λN N
+ + λ− . (251)
0 λN
−
But as λ− < λ+ we have

N !
λ−
−N ln(λ+ /λ− )

Z = λN
+ 1+ = λN
+ 1+e . (252)
λ+
So for large N , which is the case we are interested in, we have with exponential accuracy
Z ' λN
+.
(253)
Given the partition function, we can now easily calculate the magnetization per site
1 ∂
m(h) = ln(Z). (254)
N β ∂h
In Fig. 14 we plot m(h) as a function of inverse temperature β = 1/kB T for two values of magnetic field h.
We see that for non-zero h, the magnetization per site takes its maximum value m = 1 at low temperatures.
At high temperatures it goes to zero. This is as expected, as at low T the spins align along the direction
of the applied field. However, as we decrease the field, the temperature below which m(h) approaches unity
decreases. In the limit h → 0, the magnetization per site vanishes at all finite temperatures. Hence there is
no phase transition to a ferromagnetically ordered state in the one dimensional Ising model.
40
mHh=0.1L mHh=10-5 L
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
Β Β
2 4 6 8 10 2 4 6 8 10
Figure 14: Magnetization per site as a function of inverse temperature for two values of applied magnetic
field. We see that when we reduce the magnetic field, the temperature region in which the magnetization is
essentially zero grows.
5.2.2 Averages of observables in the transfer matrix formalism

The average magnetization at site j is
1 X
hσj iβ = σj e−βE({σj }) . (255)
Z σ1 ,...,σN
We can express this in terms of the transfer matrix as

1 X
hσj iβ = Tσ1 σ2 Tσ2 σ3 . . . Tσj−1 σj σj Tσj σj+1 . . . TσN σ1 . (256)
Z σ1 ,...,σN
Using that
(T σ z )σj−1 σj = Tσj−1 σj σj , (257)

z 1 0
where σ = is the Pauli matrix, we obtain
0 −1
1 j−1 z N −j+1 1
= Tr T N σ z .

hσj iβ = Tr T σ T (258)
Z Z
Diagonalizing T by means of a unitary transformation as before, this becomes
N
1 h † N † z
i 1 λ+ 0 † z
hσj iβ = Tr U T U U σ U = Tr U σ U . (259)
Z Z 0 λN −
The matrix U is given in terms of the normalized eigenvectors of T
T |±i = λ± |±i (260)
as
U = (|+i, |−i). (261)
For h = 0 we have
1 1 1
U |h=0 =√ . (262)
2 1 −1
This gives
hσj iβ = 0. (263)

h=0
41
For general h the expression is more complicated
 α
√ + 2 √ α− 2

q
1+α 1+α−  2βJ
U = 1 + , α± = e sinh(βh) ± 1 + e4βJ sinh2 (βh) . (264)
√ 2 √ 1 2
1+α+ 1+α−
The magnetization per site in the thermodynamic limit is then

2 2
α+ −1 N + α− −1 λN
α2+ +1
λ + α2− +1 −
2
α+ − 1

lim hσj iβ = lim = 2 +1 . (265)
N →∞ N →∞ λN+ + λ−
N α+
This now allows us to prove, that in the one dimensional Ising model there is no phase transition at any
finite temperature:
lim lim hσj iβ = 0 , β < ∞.
h→0 N →∞
(266)
Note the order of the limits here: we first take the infinite volume limit at finite h, and only afterwards
take h to zero. This procedure allows for spontaneous symmetry breaking to occur, but the outcome of our
calculation is that the spin reversal symmetry remains unbroken at any finite temperature.
Similarly, we find
r N −r
1 j−1 z r z N +1−j−r 1 † z λ+ 0 † z λ+ 0
hσj σj+r iβ = Tr T σ T σ T = Tr U σ U U σ U −r . (267)
Z Z 0 λr− 0 λN
−
We can evaluate this for zero field h = 0

−r r −r r r
λN
+ λ− + λN
− λ+ λ−
hσj σj+r iβ = ≈ = e−r/ξ . (268)

h=0 N N
λ+ + λ− λ+
So in zero field the two-point function decays exponentially with correlation length
1
ξ= . (269)
ln coth(βJ)
5.2.3 The related zero-dimensional quantum model

The 1D classical Ising model is related to a 0D quantum mechanical system as follows. Given the discussion
leading to (226), we are looking to write the transfer matrix of the Ising model in the form
r
c −HQ /~
T = e , (270)

where c is a constant with dimension of time. Our transfer matrix can be written as
T = eβJ cosh(βh)I + eβJ sinh(βh)σ z + e−βJ σ x (271)
We see that if we tune h to zero such that for large β

K
e−2βJ = , βhe2βJ = λ = fixed, (272)
~
then r
~ K z x 2
T = I+ [λσ + σ ] + O( ) . (273)
K ~
We conclude that the one dimensional classical Ising model is related to the quantum mechanics of a single
spin-1/2 with Hamiltonian
HQ = −Kσ x − Kλσ z . (274)
42
5.3 The Two-Dimensional Ising Model
We now turn to the 2D Ising model on a square lattice with periodic boundary conditions. The spin variables
have now two indices corresponding to rows and columns of the square lattice respectively
σj,k = ±1 , j, k = 1, . . . , N. (275)
The boundary conditions are σk,N +1 = σk,1 and σN +1,j = σ1,j , which correspond to the lattice “living” on
σk,j
k
Figure 15: Ising model on the square lattice.
the surface of a torus. The energy in zero field is

X
E({σk,j }) = −J σk,j σk,j+1 + σk,j σk+1,j . (276)
j,k
5.3.1 Transfer Matrix Method

The partition function is given by X
Z= e−βE({σk,j }) . (277)
{σj,k }
The idea of the transfer matrix method is again to write this in terms of matrix multiplications. The
difference to the one dimensional case is that the transfer matrix will now be much larger. We start by
expressing the partition function in the form
X PN
Z= e−β k=1 E(k;k+1) , (278)
{σj,k }
where
N
X 1
E(k; k + 1) = −J σk,j σk+1,j + [σk,j σk,j+1 + σk+1,j σk+1,j+1 ] . (279)
2
j=1
This energy depends only on the configurations of spins on rows k and k + 1, i.e. on spins σk,1 , . . . , σk,N
and σk+1,1 , . . . , σk+1,N . Each configuration of spins on a given row specifies a sequence s1 , s2 , . . . , sN with
sj = ±1. Let us associate a vector
|si (280)
with each such sequence. By construction there 2N such vectors. We then define a scalar product on the
space spanned by these vectors by
N
Y
ht|si = δtj ,sj . (281)
j=1
43
With this definition, the vectors {|si} form an orthonormal basis of a 2N dimensional linear vector space.
In particular we have X
I= |sihs|. (282)
s
Finally, we define a 2N × 2N transfer matrix T by
hσk |T |σk+1 i = e−βE(k;k+1) . (283)
The point of this construction is that the partition function can now be written in the form
XX X
Z= ··· hσ1 |T |σ2 ihσ2 |T |σ3 i . . . hσN −1 |T |σN ihσN |T |σ1 i (284)
σ1 σ2 σN
We now may use (282) to carry out the sums over spins, which gives
Z = Tr T N

, (285)
where the trace is over our basis {|si|sj = ±1} of our 2N dimensional vector space. Like in the 1D case,
thermodynamic properties involve only the largest eigenvalues of T . Indeed, we have
2 N
X
Z= λN
j , (286)
j=1
where λj are the eigenvalues of T . The free energy is then

   
2N N N
X λj X λj
F = −kB T ln(Z) = −kB T ln λN max
 = −kB T N ln(λmax ) − kB T ln   , (287)
λmax λmax
j=1 j
where λmax is the largest eigenvalue of T , which we assume to be unique. As |λj /λmax | < 1, the second
contribution in (287) is bounded by −kB T N ln(2), and we see that in the thermodynamic limit the free
energy per site is
F kB T
f = lim = lim − ln(λmax ).
N →∞ N 2 N →∞ N
(288)
Thermodynamic quantities are obtained by taking derivatives of f and hence only involve the largest eigen-
value of T . The main complication we have to deal with is that T is still a very large matrix. This poses
the question, why we should bother to use a transfer matrix description anyway? Calculating Z from its
2 2
basic definition (277) involves a sum with 2N terms, i.e. at least 2N operations on a computer. Finding
the largest eigenvalue of a M × M matrix involves O(M 2 ) operations, which in our case amounts to O(22N ).
For large values of N this amounts to an enormous simplification.
5.3.2 Spontaneous Symmetry Breaking

Surprisingly, the transfer matrix of the 2D Ising model can be diagonalized exactly. Unfortunately we don’t
have the time do go through the somewhat complicated procedure here (but see the Aside below), but the
upshot is that the 2D Ising model can be solved exactly. Perhaps the most important result is that in the
thermodynamic limit the square lattice Ising model has a finite temperature phase transition between a
paramagnetic and a ferromagnetic phase. The magnetization per site behaves as shown in Fig.16. At low
temperatures T < Tc there is a non-zero magnetization per site, even though we did not apply a magnetic
field. This is surprising, because our energy (276) is unchanged if we flip all spins
σj,k → −σj,k . (289)
44
m(T)
1 T/Tc
Figure 16: Phase Transition in the square lattice Ising model.
The operation (289) is a discrete (two-fold) symmetry of the Ising model. Because we have translational
invariance, the magnetization per site is
m = hσj,k iβ . (290)
Hence a non-zero value of m signifies the spontaneous breaking of the discrete symmetry (289). In order to
describe this effect mathematically, we have to invoke a bit of trickery. Let us consider zero temperature.
Then there are exactly two degenerate lowest energy states: the one with all spins σj,k = +1 and the one
with all spins σj,k = −1. We now apply a very small magnetic field to the system, i.e. add a term
X
δE = − σj,k (291)
j,k
to the energy. This splits the two states, which now have energies
E± = −JNB ∓ N , (292)
where NB is the number of bonds. The next step is key: we now define the thermodynamic limit of the free
energy per site as
−kB T ln(Z)
f (T ) ≡ lim lim . (293)
→0 N →∞ N
The point is that the contributions Z± = e−βE± of the two states to Z are such that
Z−
= e−2N/kB T . (294)
Z+
This goes to zero when we take N to infinity! So in the above sequence of limits, only the state with all
spins up contributes to the partition function, and this provides a way of describing spontaneous symmetry
breaking! The key to this procedure is that
lim lim Z 6= lim lim Z.

→0 N →∞ N →∞ →0
(295)
The procedure we have outlined above, i.e. introducing a symmetry breaking field, then taking the infinite
volume limit, and finally removing the field, is very general and applies to all instances where spontaneous
symmetry breaking occurs.
45
5.3.3 The Related One-Dimensional Quantum Problem
As we have seen, the transfer matrix approach allows up to map between classical systems in D+1 dimensions
and quantum systems in D dimensions, with the ‘extra’ dimension in the classical language mapping to the
time variable. This mapping is implemented via the transfer matrix. For the 1D classical Ising model the
related 0+1D quantum problem was of a single spin. What about the 2D classical Ising model? Recall that
T is a 2N × 2N matrix, whose matrix elements are given by
hs|T |σi = e−L(s,σ) (296)
Notice that we have changed our notation very slightly, to emphasize that the energies are specified by the
2N -vectors s, σ, each element of which can be seen as a spin, viz. s(j), σj ∈ {−1, 1}. We have also replaced
the k label previously used, by letting s, σ stand in for σk , σk+1 . As we have noted above, the set of vectors
{|σi} with σj = ±1 is a basis for a system of 2N spins-1/2 labeled by index j. If we introduce Pauli
matrices X̂j , Ŷj , Ẑj that act on the spin-1/2 labeled by j, then our convention is that Ẑj |σi = σj |σi, and
0(j)
X̂j |σi = |σ 0(j) i, where σ 0(j) is the vector obtained from σ by flipping the jth spin: σi = σi (1 − δij ) − σi δij .
(We use the X̂, Ŷ , Ẑ notation to avoid having too many confusing σs, but you can replace them with the
more familiar σ̂x , σ̂y , σ̂z if it makes you uncomfortable.)
Our aim is to write T = e−∆τ Ĥ , where Ĥ is a quantum Hamiltonian, and ∆τ is a small time interval.
(296) specifies the matrix elements of Ĥ (via e−∆τ Ĥ ). We can already see that Ĥ acts on the 2N -dimensional
Hilbert space of N spins-1/2. Comparing (296) and (283), we have
N
X βJx X
L(s, σ) = −βJ s j σj − sj sj+1 + σj σj+1
2
j=1 j
βJ βJx X
≡ const. + (sj − σj )2 − sj sj+1 + σj σj+1 (297)
2 2
j
Notice that we have also changed the problem very slightly, to allow for different couplings J, Jx in the
horizontal and vertical directions (where x is the direction that connects the N spins; the other direction
plays the role of time) ; the reasons for this will become clear below.
In order to determine Ĥ, observe that its matrix elements are given by
e−L(s,σ) = Tsσ = hs|T |σi ' hs|1 − ∆τ Ĥ|σi, (298)
where we will always work to leading order in ∆τ . For this to make sense will adjust parameters J, Jx so
that ∆τ is small. Since the space of states is very large (after all T is a big matrix), we should think about
how to organize our calculation of matrix elements. It turns out a particularly convenient approachcount
the number of spin flips needed to go from σ to s.
First of all, let’s consider the case when s = σ. This means that the two configurations are 0 spin flips
apart. Clearly,
βJx X X
L(0 flips) = − sj sj+1 + σj σj+1 = −βJx σj σj+1 . (299)
2
j j
Next, consider the case where the two configurations are identical except for one spin, which is flipped (let’s
say it’s the ith spin, so that in the notation above, s = σ 0(i) . It follows that
βJ βJx X βJx X
L(1 flip) = (1 − (−1))2 − sj sj+1 + σj σj+1 = 2βJ − sj sj+1 + σj σj+1 . (300)
2 2 2
j j
It is not too hard to see that this generalizes, so that

βJx X
L(n flips) = 2nβJ − sj sj+1 + σj σj+1 . (301)
2
j
46
Returning to (298) but now organising things in terms of spin flips, we have

−L(0 flips)
P
esσ = δsσ eβJx j σj σj+1 ' hs|1 − ∆τ Ĥ|σi , (302)

0 flips
βJx

−L(1 flip)
P
esσ = e−2βJ e j σj σj+1 +sj sj+1
' −∆τ hs|Ĥ|σi , (303)
2

1 flip
βJx

−L(n flips)
P
esσ = e−2nβJ e j σj σj+1 +sj sj+1
' −∆τ hs|Ĥ|σi . (304)
2

n flips
where in the second and third line we have recognized that if s and σ differ in any way the identity cannot

make a contribution. Expanding exponentials and taking matrix elements, while recognizing that Ĥ

0 flips
must be diagonal, we see that
X X
δsσ (1 + βJx σj σj+1 ) ' δsσ (1 − ∆τ hσ|Ĥ|σi) ↔ βJx ∼ ∆τ and hσ|Ĥ|σi = − σj σj+1 . (305)
j j
A similar expansion of the 1 flip term though suggests that e−βJ ∼ ∆τ , and we also have that ∆τ should
be small. The above are consistent if we make ∆τ small by taking βJ, βJx → ∞ while holding their ratio
g = (βJx )−1 e−2βJ fixed. Then, we see that all matrix elements of H between states that differ by n spin
flips scale like (∆τ )n and so we can ignore all those with n > 1 since we are working to leading order in
∆τ . Notice that the scaling limit makes the Ising model very anisotropic – this is the price one pays for a
convenient mapping to the classical problem but the power of universality means that this anisotropy will
not matter.
The upshot of the above argument is that all we need to do is guess a Ĥ which is good in the zero and
one-flip sectors, and then in the appropriate scaling limit introduced above we will have found the right
quantum Hamiltonian. Now, here is a clue: in the one-flip sector, from our discussion above it follows that
we must have |si = X̂i |σi for some i; allowing i to range over 1, . . . , N for a fixed σ we clearly will range
over all states connected to σ by a single spinflip. Consider the following quantum Hamiltonian:
N
ĤTFIM = −J¯
X
g X̂j + Ẑj Ẑj+1 . (306)
j=1
where J¯ is an overall constant (not simply related to J) with dimensions of energy (that we can pull out
of ∆τ by making it dimensionless – this is the freedom implied by the ‘∼’ in our expressions for ∆τ ).
Clearly, the diagonal elements of ĤTFIM satisfy the condition (302) (or equivalently, (305)). What about
the off-diagonal ones? Let us take |si = |σ 0(i) i, so that we are in the 1-flip sector. We then have (using the
fact that we are working to leading order in ∆τ to expand and/or re-exponentiate as desired)
 

¯ ¯
X X
−∆τ hs|Ĥ|σi = hs| J∆τ Ẑj Ẑj+1 + Jg∆τ X̂j + |σi
 
1 flip
j j
1 flip

¯
X
= Jg∆τ hs| X̂j |σi
j
1 flip
−2βJ
'e (307)
¯
where in the first step we used the fact that Ẑj Ẑj+1 is diagonal and in the last step we defined J∆τ =
¯
βJx , Jg∆τ = e −2βJ 0(i)
(consistent with our requirements above), and used the fact that if |si = |σ i then
hs|X̂j |σi is nonzero only for j = i, and every matrix element of H in the single-flip sector is of this
form. Now, notice that the final expression we obtain differs from what we want (cf. (303)) by a factor
βJx P
of e 2 j σj σj+1 +sj sj+1 but since βJx ∼ ∆τ and the overall term goes as e−2βJ ∼ ∆τ , the discrepancy is
O((∆τ )2 ) so we can ignore it.
47
ĤTFIM is known as the“quantum Ising chain” or “1D transverse-field Ising model” (whence TFIM). It
is one of the best studied problems in all of many-body quantum physics. It is central to conformal field
theory as well. It also has a long history at Oxford, dating to work of Elliot, Pfeuty, and Wood in the 1970s.
In another aside, we will use an idea of Pfeuty’s (that uses the so-called Jordan-Wigner mapping) to map
(306) to a model of free fermions. Here ‘free’ means that the result is quadratic in the fermionic creation
and annihilation operators, which makes the model exactly solvable.

Question 8. A lattice model for non-ideal gas is defined as follows. The sites i of a lattice may be empty or
occupied by at most one atom, and the variable ni takes the values ni = 0 and ni = 1 in the two cases. There
is an attractive interaction energy J between atoms that occupy neighbouring sites, and a chemical potential µ.
The model Hamiltonian is X X
H = −J ni nj − µ ni , (308)
hiji i
P
where hiji is a sum over neighbouring pairs of sites.
(a) Describe briefly how the transfer matrix method may be used to calculate the statistical-mechanical properties
of one-dimensional lattice models with short range interactions. Illustrate your answer by explaining how the
partition function for a one-dimensional version of the lattice gas, Eq. (1), defined on a lattice of N sites with
periodic boundary conditions, may be evaluated using the matrix
eβµ/2

1
T= .
eβµ/2 eβ(J+µ)
(b) Derive an expression for hni i in the limit N → ∞, in terms of elements of the eigenvectors of this matrix.
(c) Show that
1
hni i = ,
1 + e−2θ
where
sinh(θ) = exp(βJ/2) sinh(β[J + µ]/2) .
Sketch hni i as a function of µ for βJ 1, and comment on the physical significance of your result.
Question 9. The one-dimensional 3-state Potts model is defined as follows. At lattice sites i = 0, 1, . . . , L
“spin” variables σi take integer values σi = 1, 2, 3. The Hamiltonian is then given by
L−1
X
H = −J δσi ,σi+1 , (309)
i=0
where δa,b is the Kronecker delta, J > 0.

(a) What are the ground states and first excited states for this model?
(b) Write down the transfer matrix for (309). Derive an expression for the free energy per site f in the limit
of large L in terms of the transfer matrix eigenvalues. Show that vectors of the form (1, z, z 2 ) with z 3 = 1 are
eigenvectors, and hence find the corresponding eigenvalues. Show that at temperature T (with β = 1/kB T ) and
in the limit L → ∞
f = −kB T ln 3 + eβJ − 1 . (310)
(c) The boundary variable σ0 is fixed in the state σ0 = 1. Derive an expression (for large L), that the variable
at site ` 1 is in the same state, in terms of the transfer matrix eigenvalues and eigenvectors. Show that your
result has the form
1 2
hδσ` ,1 i = + e−`/ξ . (311)
3 3
48
How does ξ behave in the low and high temperature limits?
Question 10. Consider a one dimensional Ising model on an open chain with N sites, where N is odd. On all
even sites a magnetic field 2h is applied, see Fig. 17. The energy is
N −1 (N −1)/2
X X
E = −J σj σj+1 + 2h σ2j . (312)
j=1 j=1
(a) Show that the partition function can be written in the form
2h 2h 2h 2h
1 J J J J J J J J N
Figure 17: Open Ising chain with magnetic field applied to all even sites.
Z = hu|T (N −1)/2 |vi , (313)
where T is an appropriately constructed transfer matrix, and |ui and |vi two dimensional vectors. Give explicit
expressions for T , |ui and |vi.
(b) Calculate Z for the case h = 0.
49
Part III
Many-Particle Quantum Mechanics
In the basic QM course you encountered only quantum systems with very small numbers of particles. In
the harmonic oscillator problem we are dealing with a single QM particle, when solving the hydrogen atom
we had one electron and one nucleus. Perhaps the most important field of application of quantum physics
is to systems of many particles. Examples are the electronic degrees of freedom in solids, superconductors,
trapped ultra-cold atomic gases, magnets and so on. The methods you have encountered in the basic QM
course are not suitable for studying such problems. In this part of the course we introduce a framework,
that will allow us to study the QM of many-particle systems. This new way of looking at things will also
reveal very interesting connections to Quantum Field Theory.
6 Second Quantization
The formalism we develop in the following is known as second quantization.
6.1 Systems of Independent Particles

You already know from second year QM how to solve problems involving independent particles
N
X
H= Hj (314)
j=1
where Hj is the Hamiltonian on the j’th particle, e.g.
p̂2j ~2 2
Hj = + V (r̂j ) = − ∇ + V (r̂j ). (315)
2m 2m j
The key to solving such problems is that [Hj , Hl ] = 0. We’ll now briefly review the necessary steps,
switching back and forth quite freely between using states and operators acting on them, and the position
representation of the problem (i.e. looking at wave functions).
• Step 1. Solve the single-particle problem
Hj |φl i = El |φl i . (316)
The corresponding wave functions are
φl (rj ) = hrj |φl i. (317)
The eigenstates form an orthonormal set

Z
hφl |φm i = δl,m = dD rj φ∗l (rj )φm (rj ). (318)
• Step 2. Form N -particle eigenfunctions as products

   
N
X N
X
 Hj  φl1 (r1 )φl2 (r2 ) . . . φlN (rN ) =  Elj  φl1 (r1 )φl2 (r2 ) . . . φlN (rN ) . (319)
j=1 j=1
This follows from the fact that in the position representation Hj is a differential operator that acts
only on the j’th position rj . The corresponding eigenstates are tensor products
|l1 i ⊗ |l2 i ⊗ · · · ⊗ |lN i. (320)
50
• Step 3. Impose the appropriate exchange symmetry for indistinguishable particles, e.g.
1
ψ± (r1 , r2 ) = √ [φl (r1 )φm (r2 ) ± φl (r2 )φm (r1 )] , l 6= m. (321)
2
Generally we require
ψ(. . . , ri , . . . , rj , . . . ) = ±ψ(. . . , rj , . . . , ri , . . . ) , (322)
where the + sign corresponds to bosons and the − sign to fermions. This is achieved by taking
X
ψl1 ...lN (r1 , . . . , rN ) = N (±1)|P | φlP1 (r1 ) . . . φlPN (rN ),
P ∈SN
(323)
where the sum is over all permutations of (1, 2, . . . , N ) and |P | is the number of pair exchanges required
to reduce (P1 , . . . , PN ) to (1, . . . , N ). The normalization constant N is
1
N =√ , (324)
N !n1 !n2 ! . . .
where nj is the number of times j occurs in the set {l1 , . . . , lN }. For fermions the wave functions can
be written as Slater determinants
 
φl1 (r1 ) . . . φl1 (rN )
1
ψl1 ...lN (r1 , . . . , rN ) = √ det  ... ..
. (325)
 
.
N!
φlN (r1 ) . . . φlN (rN )
The states corresponding to (323) are

X
|l1 , . . . , lN i = N (±1)|P | |lP1 i ⊗ · · · ⊗ |lPN i .
P ∈SN
(326)
6.1.1 Occupation Number Representation

By construction the states have the symmetry
|lQ1 . . . lQN i = ±|l1 . . . lN i , (327)
where Q is an arbitrary permutation of (1, . . . , N ). As the overall sign of state is irrelevant, we can therefore
choose them without loss of generality as
| |1 .{z
. . 1} |2 .{z . . 3} 4 . . . i ≡ |n1 n2 n3 . . . i.
. . 2} 3| .{z (328)
n1 n2 n3
In (328) we have as many nj ’s as there are single-particle eigenstates, i.e. dim H 4 . For fermions we have
nj = 0, 1 only as a consequence of the Pauli principle. The representation (328) is called occupation number
representation. The nP j ’s tell us how many particles are in the single-particle state |ji. By construction the
states {|n1 n2 n3 . . . i| j nj = N } form an orthonormal basis of our N -particle problem
Y
hm1 m2 m3 . . . |n1 n2 n3 . . . i = δnj ,mj , (329)
j
where we have defined hm1 m2 m3 . . . |=|m1 m2 m3 . . . i† .

4
Note that this is different from the particle number N .
51
6.2 Fock Space
We now want to allow the particle number to vary. The main reason for doing this is that we will encounter
physical problems where particle number is in fact not conserved. Another motivation is that experimental
probes like photoemission change particle number, and we want to be able to describe these. The resulting
space of states is called Fock Space.
1. The state with no particles is called the vacuum state and is denoted by |0i.
P
2. N -particle states are |n1 n2 n3 . . . i with j nj = N .
6.2.1 Creation and Annihilation Operators

Given a basis of our space of states we can define operators by specifying their action on all basis states.
• particle creation operators with quantum number l
(
0 if nl = 1 for fermions
c†l |n1 n2 . . . i = √ Pl−1
nj
nl + 1(±1) j=1 |n1 n2 . . . nl + 1 . . . i else.
(330)
Here the + (−) sign applies to bosons (fermions).
• particle annihilation operators with quantum number l
√ Pl−1
nj
cl |n1 n2 . . . i = nl (±1) j=1 |n1 n2 . . . nl − 1 . . . i .
(331)
We note that (331) follows from (330) by
hm1 m2 . . . |c†l |n1 n2 . . . i∗ = hn1 n2 . . . |cl |m1 m2 . . . i . (332)
The creation and annihilation operators fulfil canonical (anti)commutation relations
[cl , cm ] = 0 = [c†l , c†m ] , [cl , c†m ] = δl,m bosons,

(333)
{cl , cm } = cl cm + cm cl = 0 = {c†l , c†m } , {cl , c†m } = δl,m fermions.

(334)
7: Proof of the anticommutation relation

Let us see how to prove these in the fermionic case. For l < m we have
√ Pm−1
c†l cm | . . . nl . . . nm . . . i = c†l nm (−1) j=1 nj | . . . nl . . . nm − 1 . . . i
√ √ Pm−1
= nl + 1 nm (−1) j=l nj | . . . nl + 1 . . . nm − 1 . . . i. (335)
Similarly we have
√ √ Pm−1
cm c†l | . . . nl . . . nm . . . i = nl + 1 nm (−1)1+ j=l nj | . . . nl + 1 . . . nm − 1 . . . i. (336)
This means that for any basis state |n1 n2 . . . i we have
{c†l , cm }|n1 n2 . . . i = 0 , if l > m. (337)
52
This implies that
{c†l , cm } = 0 , if l > m. (338)
The case l < m works in the same way. This leaves us with the case l = m. Here we have
√ Pl−1
c†l cl | . . . nl . . . nm . . . i = c†l nl (−1) j=1 nj | . . . nl − 1 . . . i = nl | . . . nl . . . i. (339)
( √ Pl−1
cl nl + 1(−1) j=1 nj | . . . nl + 1 . . . i if nl = 0 ,
cl c†l | . . . nl ...i =
0 if nl = 1 ,
(
| . . . nl . . . i if nl = 0 ,
= (340)
0 if nl = 1 ,
Combining these we find that
{c†l , cl }| . . . nl . . . i = | . . . nl . . . i , (341)
and as the states | . . . nl . . . i form a basis this implies
{c†l , cl } = 1. (342)
Note that here 1 really means the identity operator 1.
6.2.2 Basis of the Fock Space

We are now in a position to write down our Fock space basis in a very convenient way.
• Fock vacuum (state without any particles)
|0i. (343)
• Single-particle states
1 0 . . . i = c†l |0i .
|0 . . . 0 |{z} (344)
l
• N -particle states
Y 1 † nj
|n1 n2 . . . i = c |0i . (345)
nj ! j
p
j

Question 11. Consider a fermion ‘system’ with just one single-particle orbital, so that the only states of the
system are |0i (unoccupied) and |1i (occupied). Show that we can represent the operators a and a† by the
matrices
0 C∗

† 0 0
a = , a= .
C 0 0 0
You can do this by checking the values of aa, a† a† and a† a + aa† . What values may the constant C take?
Question 12. A quantum-mechanical Hamiltonian for a system of an even number N of point unit masses
interacting by nearest-neighbour forces in one dimension is given by
N
1X 2
pr + (qr+1 − qr )2 ,

H=
2
r=1
53
where the Hermitian operators qr , pr satisfy the commutation relations [qr , qs ] = [pr , ps ] = 0, [qr , ps ] = iδrs , and
where qr+N = qr . New operators Qk , Pk are defined by
1 X 1 X
qr = √ Qk eikr and pr = √ Pk e−ikr ,
N k N k
where k = 2πn/N with n = −N/2 + 1, . . . , 0, . . . , N/2.

Show that:
N
1 X −iks PN
(a) Qk = √ qs e Pk = √1 iks
and N s=1 ps e
N s=1
(b) [Qk , Pk0 ] = iδkk0

1 P 2
(c) H = 2 k Pk P−k + ω Qk Q−k , where ω 2 = 2(1 − cos k).
Similarly to the treatment of the simple harmonic oscillator in QM I we then define annihilation operators ak by
1
ak = (ωk Qk + iP−k ).
(2ωk )1/2
Show that the Hermitian conjugate operators are
1
a†k = (ωk Q−k − iPk ),
(2ωk )1/2
and determine the canonical commutation relations for ak and a†p . Construct the Fock space of states and de-
termine the eigenstates and eigenvalues of H.
Question 13. Bosonic creation operators are defined through their action on basis states in the occupation
number representation as √
c†l |n1 n2 . . . i = nl + 1|n1 n2 . . . nl + 1 . . . i , (346)
a) Deduce from this how bosonic annihilation operators act.
b) Show that the creation and annihilation operators fulfil canonical commutation relations
[cl , cm ] = 0 = [c†l , c†m ] , [cl , c†m ] = δl,m . (347)
6.3.1 Change of Basis

The Fock space is built from a given basis of single-particle states
single-particle states |li N-particle states |n1 n2 . . . i Fock Space

−→ −→ . (348)
You know from second year QM that it is often convenient to switch from one basis to another, e.g. from
energy to momentum eigenstates. This is achieved by a unitary transformation
{|li} −→ {|αi} , (349)
where X
|αi = hl|αi |li. (350)
| {z }
l Ulα
By construction X X
†
Ulα Uαm = hl|αihα|mi = hl|mi = δlm . (351)
α α
54
We now want to “lift” this unitary transformation to the level of the Fock space. We know that
|li = c†l |0i ,

|αi = d†α |0i . (352)
On the other hand we have

Ulα c†l |0i.
X X
|αi = Ulα |li = (353)
l l
This suggests that we take

Ulα c†l ,
X
d†α =
l
(354)
and this indeed reproduces the correct transformation for N -particle states. Taking the hermitian conjugate
we obtain the transformation law for annihilation operators
†
X
dα = Uαl cl .
l
(355)
We emphasize that these transformation properties are compatible with the (anti)commutation relations (as
they must be). For fermions
{dα , d†β } =
X † X †
Uαl Umβ {cl , c†m } = Uαl Ulβ = (U † U )αβ = δα,β . (356)
| {z }
l,m l
δl,m
6.4 Second Quantized Form of Operators

In the next step we want to know how observables such as H, P , X etc act on the Fock space.
6.4.1 Occupation number operators

These are the simplest hermitian operators we can build from cl and c†m . They are defined as
n̂l ≡ c†l cl . (357)
From the definition of cl and c†l it follows immediately that
n̂l |n1 n2 . . . i = nl |n1 n2 . . . i. (358)
6.4.2 Single-particle operators

Single-particle operators are of the form X
Ô = ôj , (359)
j
where the operator ôj acts only on the j’th particle. Examples are kinetic and potential energy operators
X p̂2j X
T̂ = , V̂ = V (x̂j ). (360)
2m
j j
We want to represent Ô on the Fock space built from single-particle eigenstates |αi. We do this in two steps:
55
• Step 1: We first represent Ô on the Fock space built from the eigenstates of ô
ô|li = λl |li = λl c†l |0i. (361)
Then, when acting on an N -particle state (326), we have

 
XN
Ô|l1 , l2 , . . . , lN i =  λj  |l1 , l2 , . . . , lN i. (362)
j=1
This is readily translated into the occupation number representation

" #
X
Ô|n1 n2 . . . i = nk λk |n1 n2 . . .i. (363)
k
As |n1 n2 . . . i constitute a basis, this together with (358) imply that we can represent Ô in the form
λk c†k ck .
X X
Ô = λk n̂k = (364)
k k
• Step 2: Now that we have a representation of Ô on the Fock space built from the single-particle states
|li, we can use a basis transformation to the basis {|αi} to obtain a representation on a general Fock
space. Using that hk|Ô|k 0 i = δk,k0 λk we can rewrite (364) in the form
hk 0 |Ô|kic†k0 ck .
X
Ô = (365)
k,k0
Then we apply our general rules for a change of single-particle basis of the Fock space
c†k =
X †
Uαk d†α . (366)
α
This gives
†
XX X
hk 0 |Uαk Ukβ |ki d†α dβ .

Ô = 0 Ô (367)
α,β k0 k
| {z } | {z }
hα| |βi
where we have used that

†
X
|ki = Uαk |αi . (368)
α
This gives us the final result

X
Ô = hα|Ô|βi d†α dβ .
α,β
(369)
We now work out a number of explicit examples of Fock space representations for single-particle operators.
1. Momentum Operators P in the infinite volume:

(i) Let us first consider P in the single-particle basis of momentum eigenstates
P̂|ki = k|ki , hp|ki = (2π~)3 δ (3) (p − k). (370)
56
8: Remark
These are shorthand notations for
P̂a |kx , ky , kz i = ka |kx , ky , kz i , a = x, y, z. (371)
and
hpx , py , pz |kx , ky , kz i = (2π~)3 δ(kx − px )δ(ky − py )δ(kz − pz ) . (372)
Using our general result for representing single-particle operators in a Fock space built from their
eigenstates (364) we have
d3 p
Z
P̂ = pc† (p)c(p) , [c† (k), c(p)} = (2π~)3 δ (3) (p − k). (373)
(2π~)3
Here we have introduced a notation
(
c† (k)c(p) − c(p)c† (k) for bosons
[c† (k), c(p)} = (374)
c† (k)c(p) + c(p)c† (k) for fermions.
(ii) Next we want to represent P̂ in the single-particle basis of position eigenstates
X̂|xi = x|xi , hx|x0 i = δ (3) (x − x0 ). (375)
Our general formula (369) gives

Z
P̂ = d3 xd3 x0 hx0 |P̂|xic† (x0 )c(x) . (376)
We can simplify this by noting that
hx0 |P̂|xi = −i~∇x0 δ (3) (x − x0 ), (377)
which allows us to eliminate three of the integrals

Z h i Z
P̂ = d3 xd3 x0 −i~∇x0 δ (3) (x − x0 ) c† (x0 )c(x) = d3 xc† (x) (−i~∇x ) c(x). (378)
2. Single-particle Hamiltonian:
N
X p̂2j
H= + V (x̂j ). (379)
2m
j=1
(i) Let us first consider H in the single-particle basis of energy eigenstates H|li = El |li, |li = c†l |0i.
Our result (364) tells us that
El c†l cl .
X
H= (380)
l
(ii) Next we consider the position representation, i.e. we take position eigenstates |xi = c† (x)|0i as a
basis of single-particle states. Then by (369)
Z
H = d3 xd3 x0 hx0 |H|xi c† (x0 )c(x). (381)
Substituting (379) into (381) and using
hx0 |V (x̂)|xi = V (x)δ (3) (x − x0 ) , hx0 |p̂2 |xi = −~2 ∇2 δ (3) (x − x0 ) , (382)
57
we arrive at the position representation
2 2
~ ∇
Z
3 †
H= d x c (x) − + V (x) c(x).
2m
(383)
(iii) Finally we consider the momentum representation, i.e. we take momentum eigenstates |pi =
c† (p)|0i as a basis of single-particle states. Then by (369)
Z 3 3 0
d pd p
H= hp0 |H|pi c† (p0 )c(p). (384)
(2π~)6
Matrix elements of the kinetic energy operator are simple
hp0 |p̂2 |pi = p2 hp0 |pi = p2 (2π~)3 δ (3) (p − p0 ). (385)
Matrix elements of the potential can be calcuated as follows

Z Z
i i 0 0
hp0 |V̂ |pi = d3 xd3 x0 hp0 |x0 ihx0 |V̂ |xihx|pi = d3 xd3 x0 hx0 |V̂ |xi e ~ p·x− ~ p ·x
| {z }
V (x)δ (3) (x−x0 )
Z
i 0
= d3 x V (x)e ~ (p−p )·x = Ṽ (p − p0 ), (386)
where Ṽ (p) is essentially the three-dimensional Fourier transform of the (ordinary) function V (x).
Hence
Z 3 3 0
d3 p p2 † d pd p
Z
H= c (p)c(p) + Ṽ (p − p0 )c† (p0 )c(p).
(2π~)3 2m (2π~)6
(387)
6.4.3 Two-particle operators

These are operators that act on two particles at a time. A good example is the interaction potential V (r̂1 , r̂2 )
between two particles at positions r1 and r2 . For N particles we want to consider
N
X
V̂ = V (r̂i , r̂j ). (388)
i<j
On the Fock space built from single-particle position eigenstates this is represented as
1
Z
V̂ = d3 rd3 r0 c† (r)c† (r0 )V (r, r0 )c(r0 )c(r).
2
(389)
Note that when writing down the first quantized expression (388), we assumed that the operators acts
specifically on states with N particles. On the other hand, (389) acts on the Fock space, i.e. on states
where the particle number can take any value. The action of (389) on N -particle states (where N is fixed
but arbitrary) is equal to the action of (388).
58
9: Derivation of (389)
Let us concentrate on the fermionic case. The bosonic case can be dealt with analogously. We start
with our original representation of N -particle states (326)
X
|r1 , . . . , rN i = N (−1)|P | |r1 i ⊗ . . . |rN i . (390)
P ∈SN
Then X 1X
V̂ |r1 , . . . , rN i = V (ri , rj )|r1 , . . . , rN i = V (ri , rj )|r1 , . . . , rN i . (391)
2
i<j i6=j
On the other hand we know that

N
Y
|r1 , . . . , rN i = c† (rj )|0i. (392)
j=1
Now consider
N
Y N
Y
c(r)|r1 , . . . , rN i = c(r) c† (rj )|0i = [c(r), c† (rj )}|0i , (393)
j=1 j=1
where is the last step we have used that c(r)|0i = 0, and [A, B} is an anticommutator if
both A and B involve an odd number of fermions and a commutator otherwise.
In our case we have a commutator for even N and an anticommutator for odd N .
By repeatedly adding and subtracting terms we find that
N
Y N
Y N
Y
[c(r), c† (rj )} = {c(r), c† (r1 )} c† (rj ) − c† (r1 ){c(r), c† (r2 )} c† (rj )
j=1 j=2 j=3
N
Y −1
+... + c† (rj ){c(r), c† (rN )}. (394)
j=1
Using that {c(r), c† (rj )} = δ (3) (r − rj ) we then find

N N N missing
X Y X
n−1 (3) † n−1 (3)
c(r)|r1 , . . . , rN i = (r − rn ) (r − rn )|r1 . . . rn . . . rN i.
z}|{
(−1) δ c (rj )|0i = (−1) δ
n=1 j6=n n=1
(395)
Hence
N N missing
X X
† 0 0 n−1 (3) (3) 0
c (r )c(r ) c(r)|r1 , . . . , rN i = (r − rn ) (r − rm ) |r1 . . . rn . . . rN i,
z}|{
(−1) δ δ (396)
| {z }
n=1 m6=n
number op.
and finally
N
X N
X
c† (r)c† (r0 )c(r0 )c(r)|r1 , . . . , rN i = δ (3) (r − rn ) δ (3) (r0 − rm ) |r1 . . . rn . . . rN i. (397)
n=1 m6=n
This implies that

1 1 X
Z
d3 rd3 r0 V (r, r0 ) c† (r)c† (r0 )c(r0 )c(r)|r1 , . . . , rN i = V (rn , rm )|r1 , . . . , rN i. (398)
2 2
n6=m
As {|r1 , . . . , rN i} form a basis, this establishes (389).
59
Using our formula for basis transformations (354)
hl|ri c†l ,
X
c† (r) = (399)
l
we can transform (389) into a general basis. We have

1 X
Z
V̂ = d3 rd3 r0 V (r, r0 )hl|rihl0 |r0 ihr0 |m0 ihr|mic†l c†l0 cm0 cm . (400)
2 0 0
ll mm
We can rewrite this by using that the action of V̂ on two-particle states is obtained by taking N = 2 in
(388), which tells us that V̂ |ri ⊗ |r0 i = V (r, r0 )|ri ⊗ |r0 i. This implies
V (r, r0 )hl|rihl0 |r0 ihr0 |m0 ihr|mi = V (r, r0 ) hl| ⊗ hl0 | |ri ⊗ |r0 i| hr| ⊗ hr0 | |mi ⊗ |m0 i|

= hl| ⊗ hl0 | V̂ |ri ⊗ |r0 i| hr| ⊗ hr0 | |mi ⊗ |m0 i|

(401)
Now we use that Z

d3 rd3 r0 |ri ⊗ |r0 i hr| ⊗ hr0 | = 1

(402)
to obtain
1
hl| ⊗ hl0 |V̂ | mi ⊗ |m0 i c†l c†l0 cm0 cm .
X
V̂ = (403)
2
l,l0 ,m,m0
Finally we can express everything in terms of states with the correct exchange symmetry
1
|mm0 i = √ |mi ⊗ |m0 i ± |m0 i ⊗ |mi (m 6= m0 ).

(404)
2
in the form
hll0 |V̂ |mm0 ic†l c†l0 cm0 cm .
X
V̂ =
(ll0 ),(mm0 )
(405)
Here the sums are over a basis of 2-particle states. In order to see that (403) is equal to (405) observe that
X 1 X 1 X
[|mi ⊗ |m0 i]cm0 cm = [|mi ⊗ |m0 i ± |m0 i ⊗ |mi]cm0 cm = √ |mm0 i cm0 cm (406)
2 2
m,m0 m,m 0 m,m 0
Here the first equality follows from relabelling summation indices m ↔ m0 and using the (anti)commutation
relations between cm and cm0 to bring them back in the right order. The second equality follows from the
definition of 2-particle states |mm0 i. Finally we note that because |mm0 i = ±|m0 mi (the minus sign is for
fermions) we have
1 X √ X
√ |mm0 i cm0 cm = 2 |mm0 i cm0 cm , (407)
2 m,m0 0
(mm )
where the sum is now over a basis of 2-particle states with the appropriate exchange symmetry. The
representation (405) generalizes to arbitrary two-particle operators O.

Question 14. Consider the N -particle interaction potential
N
X
V̂ = V (r̂i , r̂j ),
i<j
60
where V (r̂i , r̂j ) = V (r̂j , r̂i ). Show that in second quantization it is expressed as
1
Z
V̂ = d3 rd3 r0 V (r, r0 ) c† (r)c† (r0 )c(r0 )c(r).
2
To do so consider the action of V̂ on a basis of N -particle position eigenstates

N
1 X 1 Y
|r1 . . . rN i = √ (±1)|P | |r1 i ⊗ |r2 i ⊗ |rN i = √ c† (rj )|0i ,
N !n1 !n2 ! . . . P n1 !n2 ! . . . j=1
where nj is the occupation number of the j th single-particle state. Argue that in an arbitrary basis of single-particle
eigenstates |li V̂ can be expressed in the form
hll0 |V̂ |mm0 ic†l c†l0 cm0 cm .

X
V̂ =
ll0 mm0
7 Application I: The Ideal Fermi Gas

Consider an ideal gas of spin-1/2 fermions. The creation operators in the momentum representation (in the
infinite volume) are
c†σ (p) , σ =↑, ↓ . (408)
They fulfil canonical anticommutation relations
{cσ (p), cτ (k)} = 0 = {c†σ (p), c†τ (k)} , {cσ (p), c†τ (k)} = δσ,τ (2π~)3 δ (3) (k − p). (409)
The Hamiltonian, in the grand canonical ensemble, is
d3 p
Z 2 X
p
H − µN̂ = − µ c†σ (p)cσ (p). (410)
(2π~)3 2m
| {z } σ=↑,↓
(p)
Here µ > 0 is the chemical potential. As c†σ (p)cσ (p) = n̂σ (p) is the number operator for spin-σ fermions
with momentum p, we can easily deduce the action of the Hamiltonian on states in the Fock space:
h i
H − µN̂ |0i = 0 ,
h i
H − µN̂ c†σ (p)|0i = (p) c†σ (p)|0i ,
n
" n # n
h iY X Y
H − µN̂ c†σj (pj )|0i = (pk ) c†σj (pj )|0i . (411)
j=1 k=1 j=1
7.1 Quantization in a large, finite volume

In order to construct the ground state and low-lying excitations, it is convenient to work with a discrete set
of momenta. This is achieved by considering the gas in a large, periodic box of linear size L. Momentum
eigenstates are obtained by solving the eigenvalue equation e.g. in the position representation
p̂ψk (r) = −i~∇ψk (r) = kψk (r). (412)
The solutions are plane waves

i
ψk (r) = e ~ k·r . (413)
61
Imposing periodic boundary conditions (ea is the unit vector in a direction)
ψk (r + Lea ) = ψk (r) for a = x, y, z, (414)
gives quantization conditions for the momenta k

i 2π~na
e ~ Lka = 1 ⇒ ka = , a = x, y, z. (415)
L
To summarize, in a large, periodic box the momenta are quantized as
 
n
2π~  x 
k= ny (416)
L
nz
Importantly, we can now normalize the eigenstates to 1, i.e.

1 i
ψk (r) = 3 e ~ k·r . (417)
L 2
Hence Z
0
hk|k i = d3 rψk∗ (r)ψk0 (r) = δk,k0 . (418)
As a consequence of the different normalization of single-particle states, the anticommutation relations of

creation/annihilation operators are changed and now read
{cσ (p), cτ (k)} = 0 = {c†σ (p), c†τ (k)} , {c†σ (p), cτ (k)} = δσ,τ δk,p . (419)
The Hamiltonian is
X X
H − µN̂ = (p) c†σ (p)cσ (p).
p σ=↑,↓
(420)
We define a Fermi momentum by
p2F
= µ. (421)
2m
7.1.1 Ground State

Then the lowest energy state is obtained by filling all negative energy single-particle states, i.e.
Y
|GSi = c†σ (p)|0i.
|p|<pF ,σ
(422)
The ground state energy is X X

EGS = (p). (423)
σ |p|<pF
This is extensive (proportional to the volume) as expected. You can see the advantage of working in a finite
volume: the product in (422) involves only a finite number of factors and the ground state energy is finite.
The ground state momentum is X X
PGS = p = 0. (424)
σ |p|<pF
The ground state momentum is zero, because is a state with momentum p contributes to the sum, then so
does the state with momentum −p.
62
ε (p)
Figure 18: Ground state in the 1 dimensional case. Blue circles correspond to “filled” single-particle states.
7.1.2 Excitations
• Particle excitations
c†σ (k)|GSi with |k| > pF . (425)
Their energies and momenta are
E = EGS + (k) > EGS , P = k. (426)
• Hole excitations
cσ (k)|GSi with |k| < pF . (427)
E = EGS − (k) > EGS , P = −k. (428)
• Particle-hole excitations
c†σ (k)cτ (p)|GSi with |k| > pF > |p|. (429)
E = EGS + (k) − (p) > EGS , P = k − p. (430)
ε (p) ε (p) ε (p)
p p p
Figure 19: Some simple excited states: (a) particle (b) hole (c) particle-hole.
63
7.1.3 Density Correlations
Consider the single-particle operator
o = |rihr| (431)
It represents the particle density at position |ri as can be seen by acting on position eigenstates. In second
quantization it is
XZ X
ρ(r) = d3 r0 d3 r00 hr0 |o|r00 i c†σ (r0 )cσ (r00 ) = c†σ (r)cσ (r). (432)
σ σ
1. One-point function.
We now want to determine the expectation value of this operator in the ground state
X
hGS|ρ(r)|GSi = hGS|c†σ (r)cσ (r)|GSi. (433)
σ
A crucial observation is that the ground state has a simple description in terms of the Fock space built
from momentum eigenstates. Hence what we want to do is to work out the momentum representation
of ρ(r). We know from our general formula (355) that
X
cσ (r) = hr|pi cσ (p). (434)
| {z }
p i
L−3/2 e ~ p·r
Substituting this as well as the analogous expression for the creation operator into (433), we obtain
X 1 X i 0
hGS|ρ(r)|GSi = 3
e ~ (p−p )·r hGS|c†σ (p0 )cσ (p)|GSi. (435)
σ
L 0 p,p
For the expectation value hGS|c†σ (p0 )cσ (p)|GSi to be non-zero, we must have that c†σ (p0 )cσ (p)|GSi
reproduces |GSi itself. The only way this is possible is if |p| < pF (so that the c pokes a hole in the
Fermi sea) and p0 = p (so that the c† precisely fills the hole made by the c). By this reasoning we
have
hGS|c†σ (p0 )cτ (p)|GSi = δσ,τ δp,p0 θ(pF − |p0 |). (436)
Similarly we can show that
hGS|cσ (p0 )c†τ (p)|GSi = δσ,τ δp,p0 θ(|p| − pF ),

hGS|cσ (p0 )cτ (p)|GSi = 0 = hGS|c†σ (p0 )c†τ (p)|GSi. (437)
Substituting (436) back into (435) we find

X 1 X i 1 X N
(p−p0 )·r
hGS|ρ(r)|GSi = 3
e ~ δp,p 0 θ(pF − |p|) = 2 3
θ(pF − |p|) = 3 . (438)
σ
L 0
|{z} L p L
p,p spin
So our expectation value gives precisely the particle density. This is expected because our system is
translationally invariant and therefore hGS|ρ(r)|GSi cannot depend on r.
2. Two-point function.
Next we want to determine the two-point function
X 1 XX i 0 i 0 0
hGS|ρ(r)ρ(r0 )|GSi = 6
e ~ (p−p )·r e ~ (k−k )·r hGS|c†σ (p0 )cσ (p)c†τ (k0 )cτ (k)|GSi. (439)
σ,τ
L 0 0 p,p k,k
The expectation value hGS|c†σ (p0 )cσ (p)c†τ (k0 )cτ (k)|GSi can be calculated by thinking about how the
creation and annihilation operators act on the ground state, and then concentrating on the processes
that reproduce the ground state itself in the end (see Fig. 20).
64
Figure 20:
The result is
hGS|c†σ (p0 )cσ (p)c†τ (k0 )cτ (k)|GSi = δk,k0 δp,p0 θ(pF − |p|)θ(pF − |k|)
+δσ,τ δp,k0 δk,p0 θ(|k0 | − pF )θ(pF − |k|). (440)
Observe that by virtue of (436) and (437) this can be rewritten in the form
hGS|c†σ (p0 )cσ (p)|GSihGS|c†τ (k0 )cτ (k)|GSi + hGS|c†σ (p0 )cτ (k)|GSihGS|cσ (p)c†τ (k0 )|GSi . (441)
The fact that the 4-point function (440) can be written as a sum over products of two-point functions is a
reflection of Wick’s theorem for noninteracting spin-1/2 fermions. This is not part of the syllabus and we
won’t dwell on it, but apart from extra minus signs, this says that 2n-point functions are given by the sum
over all possible “pairings”, giving rise to a product of two-point functions. In our particular case this gives
hc†σ (p0 )cσ (p)c†τ (k0 )cτ (k)i = hc†σ (p0 )cσ (p)ihc†τ (k0 )cτ (k)i − hc†σ (p0 )c†τ (k0 )ihcσ (p)cτ (k)i
+ hc†σ (p0 )cτ (k)ihcσ (p)c†τ (k0 )i, (442)
and using that the two point function of two creation or two annihilation operators is zero we obtain (441).
Substituting (440) back in to (439) gives
X 1 X
hGS|ρ(r)ρ(r0 )|GSi = θ(pF − |k|)θ(pF − |p|)
0
L6
σ,σ k,p
X 1 X i 0 0
+ 6
θ(|k| − pF )θ(pF − |k0 |)e ~ (k−k )·(r−r )
σ
L 0 k,k
1 X i k·(r−r0 ) 1 X − i k0 ·(r−r0 )
= hGS|ρ(r)|GSihGS|ρ(r0 )|GSi + 2 e~ e ~ . (443)
L3 L3 0
|k|>pF |k |<pF
65
10: Evaluating k-sums
Evaluting the k sums for large L: The idea is to turn sums into integrals
Z ∞ Z π Z 2π
1 X i k·R d3 k θ(pF − ~p) ip|R| cos ϑ
Z
i
k·R 2
3
e~ −→ 3
θ(pF − |k|)e ~ = dpp dϑ sin ϑ dϕ e
L (2π~) 0 0 0 (2π)3
|k|<pF
Z pF /~
dp 2p sin(p|R|) sin(pF |R|) − pF |R| cos(pF |R|)
= 2
= ≡ h(|R|). (444)
0 (2π) |R| 2π 2 |R|3
Here we have introduced spherical polar coordinates such that the z-axis of our co-ordinate system is
along the R direction, and
kx = ~p sin ϑ cos ϕ ,
ky = ~p sin ϑ sin ϕ ,
kz = ~p cos ϑ. (445)
The other sum works similarly

1 X i k·R 1 X i k·R 1 X i k·R
e ~ = e ~ − e~ . (446)
L3 L3 L3
|k|>pF k |k|<pF
The second part is evaluated above, while the first part is

1 X i k·R
e~ = δ (3) (R). (447)
L3
k
The equality can be proved by multiplying both sides by a test-function f (R)

and then integrating over R:
1 X i k·R 1 X 1 X˜
Z Z
i
d3 R 3 e~ f (R) = 3 d3 Re ~ k·R f (R) = 3 fk = f (0). (448)
L L L
k k k
Here we have used standard definitions for Fourier series, cf Riley/Hobson/Bence 12.7.
Using these simplifications for large L we arrive at our final answer

2
hGS|ρ(r)ρ(r0 )|GSi = hGS|ρ(r)|GSi2 + hGS|ρ(r)|GSiδ (3) (r − r0 ) − 2 h(|r − r0 |) .

(449)
The first two terms are the same as for a classical ideal gas, while the third contribution is due to the
fermionic statistics (Pauli exclusion: “fermions don’t like to be close to one another”).

Question 15. Consider a system of fermions moving freely on a one-dimensional ring of length L, i.e. periodic
boundary conditions are applied between x = 0 and x = L. The fermions are all in the same spin state, so that
spin quantum numbers may be omitted. Fermion creation and annihilation operators at the point x are denoted
by ψ † (x) and ψ(x).
a) Write down the complete set of anticommutation relation satisfied by ψ † (x1 ) and ψ(x2 ).
b) Write down the wave-functions of single-particle momentum eigenstates (make sure to take the boundary
conditions into account!). What are the allowed values of momentum? Using this result, derive an expression for
the momentum space creation and annihilation operators Ψ†p and Ψp in terms of ψ † (x) and ψ(x) (hint: use the
general result for basis transformation obtained in the lecture notes).
66
c) Starting with your expression for the anticommutator {ψ † (x1 ), ψ(x2 )}, evaluate {Ψ†p , Ψq }.
d) Derive an expression for ψ(x) in terms of Ψk .
e) The density operator ρ(x) is defined by ρ(x) = ψ † (x) ψ(x). The number operator is
Z L
N= dx ρ(x) .
0
Express ρ(x) in terms of Ψ†p and Ψq , and show from this that
X †
N= Ψk Ψk .
k
Let |0i be the vacuum state (containing no particles) and define |φi by
|φi = A (uk + vk Ψ†k )|0i,

Y
where uk and vk are complex numbers depending on the label k, and A is a normalisation constant.
Evaluate (i) |A|2 , (ii) hφ|N |φi, and (iii) hφ|N 2 |φi. Under what conditions is |φi an eigenstate of particle
number?
Question 16. Consider a system of fermions in which the functions ϕ` (x), ` = 1, 2 . . . N , form a complete
orthonormal basis for single particle wavefunctions.
a) Explain how Slater determinants may be used to construct a complete orthonormal basis for n-particle states
with n = 2, 3 . . . N . Calculate the normalisation constant for such a Slater determinant at a general value of n.
How many independent n-particle states are there for each n?
b) Let C`† and C` be fermion creation and destruction operators which satisfy the usual anticommutation relations.
The quantities ak are defined by
XN
ak = Uk` C` ,
`=1
where Uk` are elements of an N × N matrix, U . Write down an expression for a†k . Find the condition which
must be satisfied by the matrix U in order that the operators a†k and ak also satisfy fermion anticommutation
relations.
c) Non-interacting spinless fermions move in one dimension in an infinite square-well potential, with position
coordinate 0 ≤ x ≤ L. The normalised single particle energy eigenstates are
1/2
2 `πx
ϕ` (x) = sin ,
L L
and the corresponding fermion creation operator is C`† .

Write down expressions for C † (x), the fermion creation operator at the point x, and for ρ(x), the particle
density operator, in terms of C`† , C` and ϕ` (x).
d) What is the ground state expectation value hρ(x)i in a system of n fermions?
In the limit n → ∞, L → ∞, taken at fixed average density ρ0 = n/L, show that

sin 2πρ0 x
hρ(x)i = ρ0 1 − .
2πρ0 x
Sketch this function and comment briefly on its behaviour for x → 0 and x → ∞.
67
8 Application II: Weakly Interacting Bosons
As you know from Statistical Mechanics, the ideal Bose gas displays the very interesting phenomenon of
Bose condensation. This has been observed in systems of trapped Rb atoms and led to the award of the
Nobel prize in 2001 to Ketterle, Cornell and Wiemann. The atoms in these experiments are bosonic, but the
atom-atom interactions are not zero. We now want to understand the effects of interactions in the framework
of a microscopic theory. The kinetic energy operator is expressed in terms of creation/annihilation operators
single-particle momentum eigenstates as
X p2
T̂ = c† (p)c(p). (450)
p
2m
Here we have assumed that our system in enclosed in a large, periodic box of linear dimension L. The
boson-boson interaction is most easily expressed in position space
1
Z
V̂ = d3 rd3 r0 c† (r)c† (r0 )V (r, r0 )c(r0 )c(r) (451)
2
A good model for the potential V (r, r0 ) is to take it of the form
V (r, r0 ) = U δ (3) (r − r0 ), (452)
i.e. bosons interact only if they occupy the same point in space. Changing to the momentum space
description
1 X i p·r
c(r) = 3/2 e ~ c(p), (453)
L p
we have
U X
V̂ = c† (p1 )c† (p2 )c(p3 )c(p1 + p2 − p3 ). (454)
2L3 p1 ,p2 ,p3
8.1 Ideal Bose Gas

For U = 0 we are dealing with an ideal Bose gas and we know that the ground state is a condensate: all
particles occupy the lowest-energy single-particle state, i.e. the zero-momentum state
1 † N
|GSi0 = √ c (p = 0) |0i. (455)
N!
So p = 0 is special, and in particular we have
†
0 hGS|c (p = 0)c(p = 0)|GSi0 = N. (456)
8.2 Bogoliubov Approximation

For small U > 0 we expect the Bose-Einstein condensate to persist, i.e. we expect
hGS|c† (p = 0)c(p = 0)|GSi = N0 1. (457)
However,
[c† (0)c(0), V̂ ] 6= 0, (458)
so that the number of p = 0 bosons is not conserved, and the ground state |GSi will be a superposition of
states with different numbers of p = 0 bosons. However, for the ground state and low-lying excited states
we will have
hΨ|c† (0)c(0)|Ψi ' N0 , (459)
68
where N0 , crucially, is a very large number. The Bogoliubov approximation states that, when acting on the
ground state or low-lying excited states, we in fact have
c† (0) '
p p
N0 , c(0) ' N0 ,
(460)
i.e. creation and annihilation operators are approximately diagonal. This is a much stronger statement than
(459), and at first sight looks rather strange. It amounts to making an ansatz for low-energy states |ψi that
fulfils
hψ 0 |c(0)|ψi = N0 hψ 0 |ψi + . . .
p
(461)
√
where the dots denote terms that are small compared to N0 . We’ll return to what this implies for the
structure of |ψi a little later. Using (460) we may expand H in inverse powers of N0
X p2
H = c† (p)c(p)
p
2m
U U N0 X †
+ 3
N02 + 2c (k)c(k) + 2c† (−k)c(−k) + c† (k)c† (−k) + c(−k)c(k)
2L 2L3
k6=0
+... (462)
Now use that X

N0 = c† (0)c(0) = N − c† (p)c(p), (463)
p6=0
where N is the (conserved) total number of bosons, and define

N
ρ= = density of particles. (464)
L3
Then our Hamiltonian becomes
Uρ X p2
Uρ h † i
H= N+ + U ρ c† (p)c(p) + c (p)c† (−p) + c(−p)c(p) + . . .
2 2m 2
p6=0 | {z }
(p)
(465)
The Bogoliubov approximation has reduced the complicated four-boson interaction to two-boson terms. The
price we pay is that we have to deal with the “pairing”-terms quadratic in creation/annihilation operators.
8.3 Bogoliubov Transformation

Consider the creation/annihilation operators defined by

b(p) cosh(θp ) sinh(θp ) c(p)
= (466)
b† (−p) sinh(θp ) cosh(θp ) c† (−p).
It is easily checked that for any choice of Bogoliubov angle θp
[b(p), b(q)] = 0 = [b† (p), b† (q)] , [b(p), b† (q)] = δp,q . (467)
Exercise 5
Let f (p), f † (p) be fermionic annihilation and creation operators
{f (p), f † (q)} = δp,q , {f (p), f (q)} = 0. (468)
69
Show that the operators defined by

g(p) cos(θp ) i sin(θp ) f (p)
= , θ−p = −θp , (469)
g † (−p) i sin(θp ) cos(θp ) f † (−p)
fulfil canonical anticommutation relations for any value of the angle θp . Eqn (469) is a Bogoliubov
transformation for fermions.
In terms of the Bogoliubov bosons the Hamiltonian becomes

2 h
1X p i
H = const + + U ρ cosh(2θp ) − U ρ sinh(2θp ) b† (p)b(p) + b† (−p)b(−p)
2 2m
p6=0
2 h
p i
− + U ρ sinh(2θp ) − U ρ cosh(2θp ) b† (p)b† (−p) + b(−p)b(p) + . . . (470)
2m
Now we choose
Uρ
tanh(2θp ) = p2
, (471)
2m + Uρ
as this removes the b† b† + bb terms, and leaves us with a diagonal Hamiltonian
X
H = const + E(p)b† (p)b(p) + . . .
p6=0
(472)
where s 2
p2
E(p) = + Uρ − (U ρ)2 . (473)
2m
We note that
p2
E(p) −→ for |p| → ∞, (474)
2m
which tells us that at high momenta (and hence high energies) we recover the quadratic dispersion. In this
limit θp → 0, so that the Bogoliubov bosons reduce to the “physical” bosons we started with. On the other
hand r
Uρ
E(p) −→ |p| for |p| → 0. (475)
m
So here we have a linear dispersion.
8.4 Ground State and Low-lying Excitations

We note that the Hamiltonian (472) involves only creation/annihilation operators with p 6= 0. Formally, we
can define zero-momentum Bogoliubov bosons as simply being equal to the original ones
b(0) = c(0) . (476)
Let us now define the Bogoliubov vacuum state |0̃i by
b(p)|0̃i = 0 . (477)
Clearly, for p 6= 0 we have E(p) > 0, and hence no Bogoliubov quasiparticles will be present in the ground
state. On the other hand, a basic assumption we made was that
p
hGS|b(0)|GSi ' N0 . (478)
70
In order to get an idea what this implies for the structure of the ground state, let us express it in the general
form
∞
X n
|GSi = αn b† (0) |0̃i . (479)
n=0
Eqn (478) then implies that √

N0
αn+1 ' αn . (480)
n+1
Replacing this approximate relation by an equality leads to a coherent state
√
N0 b† (0)
|GSi = e−N0 /2 e |0̃i. (481)
Low-lying excited states can now be obtained by creating Bogoliubov quasipartices, e.g.
b† (q)|GSi, (482)
is a particle-excitation with energy E(q) > 0.
8.5 Ground state correlation functions

We are now in a position to work out correlation functions in the ground state such as
hGS|c† (p)c(q)|GSi , p, q 6= 0. (483)
Inverting the Bogoliubov transformation (466) we have
c† (p) = cosh(θp )b† (p) − sinh(θp )b(−p) ,

c(q) = cosh(θq )b(q) − sinh(θq )b† (−q) . (484)
Using that
hGS|b† (p) = 0 = b(q)|GSi, (485)
we find that
hGS|c† (p)c(q)|GSi = sinh(θq ) sinh(θq )hGS|b(−p)b† (−q)|GSi

= sinh2 (θp )δq,p (p, q 6= 0). (486)
This tells us that, in contrast to the ideal Bose gas, in the ground state of the interacting Bose gas we have
a finite density of bosons with non-zero momentum
hGS|c† (p)c(p)|GSi = sinh2 (θp ) . (487)
Another feature of the ground state is that the two-point function of two annihilation/creation operators is
non-zero
hGS|c(p)c(q)|GSi = hGS|c† (q)c† (p)|GSi = − cosh(θp ) sinh(θq )δp,−q . (488)
These imply that boson number is not a good quantum number in the ground state. More formally, we
say that the ground state spontaneously breaks the U(1) symmetry of the Hamiltonian H = T̂ + V̂ . Let us
explain that statement. The Hamiltonian is invariant under the symmetry operation
Û c(p)Û † = eiφ c(p) , φ∈R,

† † −iφ †
Û c (p)Û = e c (p) , (489)
i.e.
Û H Û † = H. (490)
71
The reason for this is that all terms in H involve the same number of creation as annihilation operators,
and the total particle number is therefore conserved. This is referred to as a global U(1) symmetry (as
the transformations (489) form a group called U(1)). Let us now investigate how ground state expectation
values transform. We have
hGS|c(p)c(q)|GSi = hGS|Û † Û c(p)Û † Û c(q)Û † Û |GSi = e2iφ hGS|Û † c(p)c(q)Û |GSi . (491)
If the ground state were invariant under the symmetry, we would have Û |GSi = |GSi. Eqn (491) would then
imply that hGS|c(p)c(q)|GSi = 0. Reversing the argument, we see that a non-zero value of the expectation
value (488) implies that the ground state cannot be invariant under the U(1) symmetry, and in fact “breaks
it spontaneously”.
8.6 Depletion of the Condensate

We started out by asserting that for small interactions U > 0 we retain a Bose-Einstein condensate, i.e. the
consensate fraction N0 /N remains large. We can now check that this assumption is self-consistent. We have
X
N0 = N − c† (p)c(p). (492)
p6=0
Thus in the ground state

N0 1 X 1 X
=1− hGS|c† (p)c(p)|GSi = 1 − sinh2 (θp ), (493)
N N N
p6=0 p6=0
where we have used (486). This equals

 
 
N0 1 X 1 1 X 1 
= 1− q − 1 = 1 − r
i2 − 1 .
 (494)
N 2N 2
1 − tanh (2θp ) 2N  h
Uρ
p6=0 p6=0 1− (p)
We again turn this into an integral and evaluate it in spherical polar coordinates, which gives
 
N0 2π ∞ dp 1
Z
p2 
 
≈1− 3
r − 1
. (495)
N ρ 0 (2π~)  h
Uρ
i 2
1 − (p)
√
By means of the substitution p = 2mU ρz we can see that the integral is proportional to U 3/2 and thus
indeed small for small U .
9 Application III: Spinwaves in a Ferromagnet

Consider the following model of a magnetic insulator: at each site r of a D-dimensional with N sites lattice
we have a magnetic moment. In QM such magnetic moments are described by three spin-operators
Srα , α = x, y, z , (496)
which fulfil the angular momentum commutation relations
[Srα , Srβ0 ] = δr,r0 iαβγ Srγ . (497)
We will assume that the spin are large in the sense that
X 2
S2r = Srα = s(s + 1) 1. (498)
α
72
Let us begin by constructing a basis of quantum mechanical states. At each site we have 2s + 1 eigenstates
of Srz
Srz |mir = m|mir , m = s, s − 1, . . . , −s. (499)
They can be constructed from |sir using spin lowering operators Sr− = Srx − iSry
1
|s − nir = S − )n |sir , n = 0, 1, . . . , 2s, (500)
Nn r
where Nn are normalization constants. A basis of states is then given by
Y
|sr ir , −s ≤ sr ≤ s spin on site r. (501)
r
9.1 Heisenberg model and spin-rotational SU(2) symmetry

An appropriate Hamiltonian for a ferromagnetic insulator was derived by Heisenberg
X
H = −J Sr · Sr0 .
hr,r0 i
(502)
Here hr, r0 i denote nearest-neighbour pairs of spins and we will assume that J > 0. The model (502) is
known as the ferromagnetic Heisenberg model. You can check that the Hamiltonian (502) commutes with
the three total spin operators X
[H, S α ] = 0 , S α = Srα . (503)
r
These imply that the Hamiltonian is invariant under general rotations (in spin space)
eiα·S He−iα·S = H. (504)
The transformations (504) form a group known as SU(2), and the Heisenberg Hamiltonian (502) is invariant
under them.
9.2 Exact ground states

One ground state of H is given by
Y
|GSi = |sir .
r
(505)
Its energy is X
H|GSi = −J s2 |GSi = −Js2 NB |GSi, (506)
hr,r0 i
where NB is the total number of bonds in our lattice. The total spin lowering operator S − = −
P
r Sr
commutes with H by virtue of (503) and hence
1 n
|GS, ni = S − |GSi , 0 ≤ n ≤ 2sN (507)
Nn
are ground states as well (as they have the same energy). Here Nn is a normalization.
73
11: Proof that |GSi is a ground state
We note that the spin-spin interaction can be written in the form
2Sr · Sr0 = (Sr + Sr0 )2 − S2r − S2r0 = J2 − 2s(s + 1). (508)
Here J2 is the total angular momentum squared. Its eigenvalues follow from the theory
of adding angular momenta to be
J2 |j, mi = j(j + 1)|j, mi , j = 2s, 2s − 1, . . . , 1, 0. (509)
This tells us that the maximal eigenvalue of J2 is 2s(2s + 1), and by expanding |ψi in a
basis of eigenstates of J2 we can easily show that
X
hψ|J2 |ψi = hψ|j, mihj, m|J2 |j 0 , m0 ihj 0 , m0 |ψi
j,m,j 0 ,m0
X X
= |hψ|j, mi|2 j(j + 1) ≤ 2s(2s + 1) |hψ|j, mi|2 = 2s(2s + 1). (510)
j,m j,m
This tells us that

hψ|Sr · Sr0 |ψi ≤ s2 . (511)
This provides us with a bound on the eigenvalues of the Hamiltonian, as
X
hψ|H|ψ|i ≥ −J s2 = −Js2 N z. (512)
hr,r0 i
The state we have constructed saturates this bound, so must be a ground state.
Let us now see how the SU(2) symmetry is reflected in expectation values of operators O. At finite
temperature we have
1 h i
hOiβ = Tr e−βH O , (513)
Z(β)
where Z(β) = Tr[e−βH ] is the partition function and β = 1/kB T . In the T → 0 limit we have
2sN
1 X
hOi∞ = hGS, n|O|GS, ni, (514)
2sN + 1
n=0
i.e. we average over all ground states. The thermal average, as well as its T = 0 limit, are invariant under
rotations in spin space. Indeed, under a rotation in spin space we have
1 h i
heiα·S Oe−iα·S iβ = Tr e−βH eiα·S Oe−iα·S (515)
Z(β)
P
where S = r Sr are the global spin operators. Using the cyclicity of the trace and the fact that H
commutes with the global spin operators, we see that this equals hOiβ . If we choose as our operator O any
of the global spin operators, and consider a rotation by π around one of the orthogonal axes, we see that
the magnetization always vanishes
hS α iβ = 0 , α = x, y, z. (516)
Physically this is what one would expect for a system that is spin rotationally invariant, i.e. looks the same
in any direction in spin space.
74
9.3 Spontaneous Symmetry Breaking
In a real system, the 2sN + 1-fold ground state degeneracy is usually broken through imperfections. In
practice the details of these imperfections are not important, the only thing that matters is that the symmetry
gets broken. To keep things simple, one retains the spin-rotationally symmetric Hamiltonian, and says that
the ground state breaks the symmetry “spontaneously”.
A convenient P mathematical description of this effect is as follows. Imagine adding an infinitesimal
magnetic field − r Srz to the Hamiltonian. This will break the symmetry and hence the degeneracy of the
ground states, which now will have energies
EGS,n = −Js2 NB − (sN − n). (517)
Now consider the sequence of limits

(
n
1 0 if limN →∞ N =0,
lim lim [EGS,n − EGS,0 ] = (518)
→0 N →∞ N ∞ else.
This means that if we define the thermodynamic limit in the above way, then the only surviving ground states
(defined as states whose energy density is minimum in the thermodynamic limit) will have magnetization
per site s, i.e. contain only a non-extensive number of spin flips. In all of these remaining ground states
the spin rotational symmetry has been broken. As we have taken → 0 our Hamiltonian is again SU(2)
symmetric, but the remaining ground states “spontaneously” break this symmetry.
9.4 Holstein-Primakoff Transformation

We succeeded in finding the ground states of H because of their simple structure. For more general spin
Hamiltonians, or even the Hamiltonian (502) with negative value of J, this will no longer work and we need
a more general, but approximate way of dealing with such problems. This is provided by (linear) spinwave
theory.
As shown by Holstein and Primakoff, spin operators can be represented in terms of bosonic creation and
annihilation operators as follows:
s
√ a†r ar
Srz = s − a†r ar , Sr+ = Srx + iSry = 2s 1 − ar . (519)
2s
You can check that the bosonic commutation relations
[ar , a†r0 ] = δr,r0 (520)
imply that
[Srα , Srβ0 ] = δr,r0 iαβγ Srγ . (521)
However, there is a caveat: the spaces of QM states are different! At site r we have
n
Sr |sir , n = 0, . . . , 2s (522)
for spins, but for bosons there are infinitely many states
n
a†r |0ir , n = 0, . . . , ∞. (523)
To make things match, we must impose a constraint, that there are at most 2s bosons per site. Now we take
advantage of the fact that we have assumed s to be large: in the ground state there are no bosons present,
because
hGS|s − a†r ar |GSi = hGS|Srz |GSi = s = (524)
75
Low-lying excited states will only have a few bosons, so for large enough s we don’t have to worry about
the constraint. Using the Holstein-Primakoff transformation, we can rewrite H in a 1/s expansion
h i
s2 − s a†r ar + a†r0 ar0 − a†r ar0 − a†r0 ar + . . .
X
H = −J
hr,r0 i
(525)
Here the dots indicate terms proportional to s0 , s−1 , etc. Once again using that s is large, we drop these
terms (for the time being). We then can diagonalize H by going to momentum space
1 X ik·r
ar = √ e a(k) , [a(k), a† (p)] = δk,p , (526)
N k
which gives
X
H = −Js2 N z + (q)a† (q)a(q) + . . .
q
(527)
For a simple cubic lattice the energy is
(q) = 2Js [3 − cos qx − cos qx − cos qz ] . (528)
For small wave numbers this is quadratic
(q) ≈ Js|q|2 for |q| → 0. (529)
In the context of spontaneous symmetry breaking these gapless excitations are known as Goldstone modes.
Let us now revisit the logic underlying our 1/s expansion. For things to be consistent, we require that
the terms of order s in (527) provide only a small correction to the leading contribution proportional to s2 .
This will be the case as long as we are interested only is states |Ψi such that
hΨ|a† (q)a(q)|Ψi s. (530)
This condition is certainly fulfilled for the ground state and low-lying excited states.
9.4.1 Heisenberg Antiferromagnet

Another example to which spinwave theory can be fruitfully applied is the model
X
H=J Sr · Sr0 , (531)
hr,r0 i
where hr, r0 i denote nearest-neighbour pairs of spins on a simple cubic lattice and J > 0. Compared to (502)
all we have done is to switch the overall sign of H, but this has important consequences. In particular, it is
no longer possible to obtain an exact ground state for the model. Instead, we start by considering our spins
to be classical. This is justified if we are interested only in states with large spin quantum numbers. We
will assume this to be the case and check the self-consistency of our assumption later. In the classical limit
we can think of the spins as three-dimensional vectors. The lowest energy configuration is then one, where
all neighbouring spins point in opposite directions, i.e. along the three cystal axes the spin configuration
looks like ↑↓↑↓↑↓ .... This is known as a Néel state. It is convenient to subdivide our lattice into two
sublattices: on sublattice A all spins point in the same direction, while on sublattice B all spins point in
the opposite direction. Like the ferromagnet, the model (531) has a global spin-rotational symmetry, that
will be spontaneously broken in the ground state. By choosing our spin quantization axis appropriately, the
classical ground state can then be written in the form
Y Y
|sir | − sir0 (532)
r∈A r0 ∈B
76
The idea is now to map this state to a ferromagnetic one, by inverting the spin quantization axis in the B
sublattice. After that we can employ the Holstein-Primakoff transformation to carry out a 1/s expansion.
As a result of the rotation of spin quatization axis on the B sublattice, the part of the Hamiltonian of order
s now contains terms involving two annihilation or two creation operators. Diagonalizing the Hamiltonian
then requires a Bogoliubov transformation.

Question 17. (optional) Consider the ideal Fermi gas at finite density N/V in a periodic 3-dimensional box of
length L.
(a) Give an expression of the ground state in terms of creation operators for momentum eigenstates.
(b) Calculate the single-particle Green’s function
Z Z
iω(t−t0 ) 0
Gστ (ω, q) = dt e d3 r e−iq·(r−r ) Gστ (t, r; t0 , r0 ) ,
Gστ (t, r; t0 , r0 ) = −ihGS|T cσ (r, t) c†τ (r0 , t0 )|GSi, (533)
where T denotes time-ordering (i.e. T O(t1 )O(t2 ) = θ(t1 − t2 )O(t1 )O(t2 ) − θ(t2 − t1 )O(t2 )O(t1 ) for fermionic
operators), and
i i
cσ (r, t) ≡ e ~ Ht cσ (r)e− ~ Ht . (534)
First express the creation/annihilation operators c†σ (r, t), cσ (r, t) in terms of creation/annihilation operators in
momentum space c†σ (p, t), cσ (p, t). Then show that for annihilation operators in momentum space we have
i i i
cσ (p, t) ≡ e ~ Ht cσ (p)e− ~ Ht = cσ (p)e− ~ t(p) , (535)
where (p) = p2 /2m − µ. Use this to show that

1 X i
cσ (r, t) = e− ~ t(p)+ip·r cσ (p). (536)
L3/2 p
Now insert (536) into (533) and evaluate the ground state expectation value to obtain an integral representation
for Gστ (t, r; t0 , r0 ). Why does the Green’s function only depend on t − t0 and r − r0 ? Finally, calculate Gστ (ω, q).
Note: the imaginary part of the single-particle Green’s function is (approximately) measured by angle resolved
photoemission (ARPES) experiments.
Question 18. A magnetic system consists of two types of Heisenberg spin SA and SB located respectively
on the two inter-penetrating sublattices of an NaCl crystal structure (i.e. a simple cubic structure with alternate
A and B in any Cartesian direction). Its Hamiltonian is
X
H=J SA B
i · Sj
i,j
where the i, j are nearest neighbours, respectively on the A and B sublattices. J is positive. Show that the
classical ground state has all the A spins ferromagnetically aligned in one direction and all the B spins ferromag-
netically aligned in the opposite direction. Assume the quantum mechanical ground state is well approximated
by the classical one. To a first approximation the spin operators are given in terms of boson operators a, b by
A sublattice B sublattice
Siz = SA − a†i ai Sjz = −S B + b†j bj
Si+ ≡ Six + iSiy ' (2S A )1/2 ai Sj+ ≡ Sjx + iSjy ' (2S B )1/2 b†j
Si− ≡ Six − iSiy ' (2S A )1/2 a†i Sj− ≡ Sjx − iSjy ' (2S B )1/2 bj
Discuss the validity of this approximation. Use these relations to express the Hamiltonian in terms of the boson
operators to quadratic order.
77
Transforming to crystal momentum space using (with N the number of sites on one sublattice)
X X
ai = N −1/2 e−ik·ri ak , bj = N −1/2 eik·rj bk
k k
show that your result can be expressed in the form

Xh i
H = E0 + Ak a†k ak + Bk b†k bk + Ck (a†k b†k + bk ak )
k
and determine the coefficients. Hence calculate the spectrum of excitations at low momenta. Consider both the
cases with S A = S B and S A 6= S B and comment on your results.
10 Path Integral for interacting Bose systems

A key ingredient in our construction of the path integral for a single particle was the resolution of the
identity Z
1 = dx |xihx| , (537)
which allowed us to break up e−βH into a product over infinitesimal propagators.
10.1 Coherent States

In order to deal with many-boson systems, we require a convenient analog on the Fock space. This is
provided by coherent states
!
†
X
|φi = exp φ` a` |0i , φ` ∈ C ,
`
(538)
where a` denotes the bosonic annihilation operator for the single-particle state labeled by ` and |0i is the
Fock vacuum. If NSP is the number of single-particle states, then φ is a NSP -dimensional complex vector.
The states (538) are simultaneous eigenstates of all annihilation operators
aj |φi = φj |φi.
(539)
In order to prove (539) consider

! !  
h i ih i
φ` a†` |0i = aj , exp φ` a†` φ` a†`  aj , exp φj a†j |0i.
X X X
aj |φi = aj exp |0i = exp  (540)
` ` `6=j
The commutator is easily calculated by expanding the exponential in its power series
h i
aj , exp φj a†j = φj exp φj a†j , (541)
and substituting this back into (540) establishes (539). Coherent states are not mutually orthogonal. In
fact, they fulfil
ψ`∗ φ`
P
hψ|φi = e ` .
(542)
This result for the scalar product can be obtained by applying the Baker-Campbell-Hausdorff (BCH) formula,
which states that for two operators such that [A, [A, B]] = 0 = [B, [A, B]] we have
1
eA eB = eA+B e 2 [A,B] = eB eA e[A,B] . (543)
78
Setting A = ` ψ`∗ a` , B = j φj a†j , using the BCH formula, and then noting that A|0i = 0 = h0|B, we
P P
obtain (542). While coherent states do not form an orthogonal set, they nevertheless provide a resolution
of the identity on the Fock space
Z Y 2
d φj P
|φ` |2
1= e− ` |φihφ|.
π
j
| {z }
d(φ,φ∗ )
(544)
Here d2 φ` denotes the integration over the complex variable φ` , e.g. in polar co-ordinates we have
Z Z ∞ Z 2π
d2 φj = drj rj dϕj , φj = rj eiϕj . (545)
0 0
To prove (544) we note that

∞ ∞
φ1 a†1 φ2 a†2 φ3 a†3 (φ1 )n1 X (φ2 )n2
. . . (a†1 )n1 (a†2 )n2 (a†3 )n3 . . . |0i
X
|φi = e e e . . . |0i =
n1 ! n2 !
n1 =0 n2 =0
∞ ∞ ∞
X (φ1 )n1 X (φ2 )n2 X (φ3 )n3
= √ √ √ . . . |n1 n2 n3 . . .i , (546)
n =0
n1 ! n =0 n2 ! n =0 n3 !
1 2 3
where |n1 n2 . . . i is a state in the occupation number representation. Hence

X X φn1 1 (φ∗1 )m1 φn2 2 (φ∗2 )m2 . . .
|φihφ| = √ |n1 n2 n3 . . . ihm1 m2 m3 . . . | . (547)
n1 ,n2 ,... m1 ,m2 ,...
n1 !m1 !n2 !m2 ! . . .
Inspection of (544) and (547) shows that the integral over φj and φ∗j is
Z ∞ Z 2π
2
drj dϕj rnj +mj +1 e−rj eiϕ(nj −mj ) = nj ! δnj ,mj . (548)
0 0
Carrying out all integrals we obtain

Z
2
P X
d(φ, φ∗ )e− ` |φ` | |φihφ| = |n1 n2 n3 . . . ihn1 n2 n3 . . . | . (549)
n1 ,n2 ,...
The right hand side is a resolution of the identity in the occupation number representation.
10.2 Partition Function

Let us now consider a general many-boson Hamiltonian of the form
hij a†i aj + Vijkl a†i a†j ak al .

X X
Ĥ =
i,j i,j,k,l
(550)
We first want to derive a path integral representation for the partition function
h i X
Z(β) = Tr e−β Ĥ = hn|e−β Ĥ |ni. (551)
n
Inserting a resolution of the identity in terms of coherent states this can rewritten as
Z Z
∗ − ` |ψ` |2 2
P X P X
−β Ĥ
d(ψ, ψ )e hn|ψi hψ|e |ni = d(ψ, ψ ∗ )e− ` |ψ` | hψ|e−β Ĥ |nihn|ψi
n n
Z
|2
P
= d(ψ, ψ ∗ )e− ` |ψ`
hψ|e−β Ĥ |ψi. (552)
79
In the next step we break up e−βH
β
e−βH = e−H e−H . . . e−H , = , (553)
N
and then insert resolutions of the identity between each of the factors. This leaves us with matrix elements
of the form
hψ (n+1) |e−Ĥ |ψ (n) i = hψ (n+1) |1 − Ĥ|ψ (n) i + O(2 )

∗
h i
= hψ (n+1) |ψ (n) i 1 − H ψ (n+1) , ψ (n) + O(2 )
(n+1) ∗ ,ψ (n)

= hψ (n+1) |ψ (n) i e−H ψ + O(2 ) , (554)
where X X
H ψ∗, ψ0 = hij ψi∗ ψj0 + Vijkl ψi∗ ψj∗ ψk0 ψl0 . (555)
i,j i,j,k,l
In going from the first to the second line in (554) we have used that
aj |ψ 0 i = ψj0 |ψ 0 i , hψ|a†i = ψi∗ hψ| . (556)
After these steps we end up with a representation of the form
X ψ (n) ∗ − ψ (n+1) ∗
Z YN Z
" N −1 #
(m) (m) ∗ (n) (n+1) ∗ (n)

Z(β) = lim d(ψ , ψ ) exp − ·ψ +H ψ ,ψ . (557)
N →∞
m=1 n=0
In complete analogy to what we did in the single-particle case, we can now interpret the sequence ψ (1) ,
ψ (2) ,. . . ψ (N −1) as a discretization of a path on the space of NSP dimensional complex vectors
ψ(τn ) = ψ (n) , τn = n. (558)
In the limit N → ∞ this goes over into a vector-valued function of imaginary time ψ(τ ), and the partition
function acquires the following formal expression
Z
∗
Z(β) = D ψ ∗ (τ ), ψ(τ ) e−S[ψ (τ ),ψ(τ )] ,

(559)
where the action S is given by

 
Z β X X
S[ψ ∗ (τ ), ψ(τ )] = dτ  ψi∗ (τ ) [δij ∂τ + hij ] ψj (τ ) + Vijkl ψi∗ (τ )ψj∗ (τ )ψk (τ )ψl (τ ) . (560)
0 i,j i,j,k,l
80
Part IV
Landau Theory of Phase Transitions
11 Phase Transitions
Physically a phase transition is a point in parameter space, where the physical properties of a many-particle
system undergo a sudden change. An example is the paramagnet to ferromagnet transition in Fe or Ni
shown in Fig. 21.
T
m(T,h)
T<Tc
Tc
h
h
Figure 21: Phase diagram for the paramagnetic to ferromagnetic transition. The magnetization per site
m(T, h) jumps when crossing zero for T < Tc .
Mathematically a phase transition is a point in parameter space, where the free energy F = −kB T ln(Z)
becomes a nonanalytic function of one of its parameters (i.e. F or some of its derivatives becomes singular
or discontinuous) in the thermodynamic limit.
For a finite system this can never happen, because
X
Z= e−E(C)/kB T (561)
configurations C
is a finite sum over finite, positive terms. Hence all derivatives are finite and well defined as well.
Phase transitions are usually divided into two categories:
1. First Order Phase Transitions.
Here the free energy is continuous, but a first derivative is discontinuous. At the transition there is
phase coexistence. The magnetization per site is a first order derivative of the free energy with respect
to the magnetic field h. Therefore the phase transition at h = 0 and T < Tc in Fig. 21 is first order.
2. Second Order Phase Transitions.
These are characterized by a divergence in one of the higher order derivatives (“susceptibilities”) of
the free energy. The phase transition as a function of T for h = 0 in Fig. 21 is second order.
12 The Ising Model

The two-dimensional square-lattice Ising model is one of the simplest statistical models to show a phase
transition and as a result has played a hugely important role in the development of the theory of critical
81
phenomena. The model can in fact be solved exactly by the transfer matrix method introduced above.
Interestingly, it is possible to establish the existence of a finite-temperature phase transition in the model
without solving it. This was pioneered by Sir Rudolf Peierls, who was the Wykeham Professor of Theoretical
Physics at Oxford for many years and during his career made numerous ground breaking contributions in
particular to condensed matter physics.
12.1 Peierls Argument

The Peierls argument is a nice way of establishing that the 2D square lattice Ising model has magnetic
long-range order at sufficiently low temperatures without actually solving the model. Given that at very high
temperature there is no magnetic order, this shows that there must be at least one phase transition at a
finite temperature.
Consider the Ising model on the square lattice with boundary conditions such that all spins on the
boundary are up, i.e. take the value +1. You can think of these boundary conditions as a symmetry
breaking field. The bulk magnetic field is taken to be zero. Configurations look like the one shown in
Fig. 22, and can be characterized by domain walls. These are lines separating + and − spins such that
1. The + (−) spins lie always to the left (right) of the wall.
2. Where ambiguities remain, the wall is taken to bend to the right.
3. The length of the wall is defined as the number of lattice spacings it traverses.
Figure 22: A configuration of spins and the corresponding domain walls.
A wall of length b encloses at most b2 /16 spins. The total number of domain walls of length b, m(b), is
bounded by
m(b) ≤ 4Nt 3b−1 ,
(562)
where Nt is the total number of sites. This can be seens as follows:
• the first link can go into less than 4Nt positions (starting at an arbitrary site and going in any of the
four possible directions).
• subsequent links have at most 3 possible directions each.

Let us denote the i’th domain wall of length b by (b, i). Next consider a particular configuration σ = {σj,k }
of spins on the lattice, and define
(
1 if (b,i) occurs in σ
Xσ (b, i) = (563)
0 else
82
Then the total number of − spins in σ is bounded by
X b2 m(b)
X
N− ≤ Xσ (b, i), (564)
16
b i=1
because each spin is enclosed by at least one domain wall due to our choice of boundary conditions. Taking
thermal averages, we have
X b2 m(b)
X
hN− iβ ≤ hXσ (b, i)iβ . (565)
16
b i=1
Here the thermal average of Xσ (b, i) can be written as

1 X0 −βE(σ)
hXσ (b, i)iβ = e , (566)
Z σ
where the sum is only over configurations, in which (b, i) occurs. Now consider the configuration σ 0 obtained
from σ by reversing the spins inside the domain wall (b, i). Clearly the energies of the two configurations
Figure 23: Configurations σ and σ 0 related by reversing all spins inside the domain wall (b, i). Shown are
all the bonds whose energies have been changed from −J to J.
are related by
E(σ) = E(σ 0 ) + 2bJ. (567)
This can be used to obtain a bound on Z
X0 0
X0
Z≥ e−βE(σ ) ≥ e−βE(σ) e2bJβ , (568)
σ0 σ
83
where the first sum is only over configurations in which (b, i) occurs, and where we then have flipped all
spins inside the domain wall. This gives us a bound on
1 X0 −βE(σ)
hXσ (b, i)iβ = e ≤ e−2bJβ . (569)
Z σ
Now we put everything together
X b2 m(b)
X X b2 Nt X h ib
hN− iβ ≤ e−2βJb ≤ 4Nt 3b−1 e−2βJb = b2 3e−2βJ . (570)
16 16 12
b i=1 b b=4,6,8,...
The sum over b can now be easily carried out, and the results at small T (large β) is
hN− iβ ≤ 108Nt e−8βJ .

(571)
So, at low temperatures we have

hN− iβ 1
.
Nt 2
(572)
This proves the existence of a spontaneous magnetization at low temperatures.
12.2 Mean Field Theory

Consider the Ising model on a D-dimensional lattice with coordination number (number of nearest neighbour
sites) z X X
E = −J σi σj − h σj . (573)
hiji j
Here hiji denotes nearest neighbour bonds, and each bond is counted once. The magnetization per site is
N
1 X
m= hσj iβ . (574)
N
j=1
We now rewrite the energy using

σi = m + (σi − m). (575)
In particular we have
σi σj = m2 + m(σj − m) + m(σi − m) + (σi − m)(σj − m). (576)
The idea of the mean-field approximation is to assume that the deviations σj − m of the spins from their
average values are small, and to neglect the terms quadratic in these fluctuations. This gives
X X
EMF = −J −m2 + m(σi + σj ) − h σj .
hiji j
(577)
Physically, what we have done is to replace the interaction of a given spin with its neighbours by an average
magnetic field. We can simplify (577) further by noting that
X Nz
−J −m2 = Jm2 ,
2
hiji
X X
σi + σj = z σj . (578)
hiji j
84
The mean-field energy then becomes
N
JN z 2 X
EMF = m − (Jzm + h) σj .
2
j=1
(579)
The partition function in the mean-field approximation is
N
X N Jzβm2 X XY
ZMF = e−βEMF = e− 2 ··· eβ(Jzm+h)σj
{σj } σ1 σN j=1
" # " #
2
− N Jzβm
X X
β(Jzm+h)σ1 β(Jzm+h)σN
= e 2 e ... e
σ1 σN
2
− N Jzβm
= e 2 [2 cosh(Jzmβ + hβ)]N . (580)
The magnetization per site is
N
1 X
m= hσj iβ = hσN iβ , (581)
N
j=1
where we have used translational invariance in the last step. In mean field theory we have
N
1 N Jzβm2 X X Y
m = hσj iβ = e− 2 ··· σN eβ(Jzm+h)σj = tanh(Jzmβ + βh).
ZMF σ1 σN j=1
(582)
This is a self-consistency equation for m.
12.3 Solution of the self-consistency equation for h = 0

For zero field the self-consistency equation reads
m = tanh(Jzmβ). (583)
This can be solved graphically by looking for intersections of the functions g1 (m) = tanh(Jzmβ) and
g2 (m) = m. There are either one or three solutions
(
0 if Jzβ < 1
m= . (584)
±m0 , 0 if Jzβ > 1
We still have to check whether these solutions correspond to mimima of the free energy per site. The latter
is
1 Jzm2 1
fMF = − ln(ZMF ) = − ln[2 cosh(Jzmβ)]. (585)
βN 2 β
We have
∂ 2 fMF

Jzβ
= Jz 1 − . (586)
∂m2 cosh(Jzmβ)
This is negative for m = 0 and Jzβ > 1, and hence this solution corresponds to a maximum of the free
energy and hence must be discarded. This leaves us with
(
0 if T > Tc
m= ,
±m0 if T < Tc
(587)
where the transition temperature is
Jz
Tc = . (588)
kB
85
12.4 Vicinity of the Phase Transition
Let us define a dimensionless variable, that measures the distance in temperature to the phase transition
T − Tc
t= . (589)
Tc
For |t| 1 we obtain the following results
1. Magnetization per site
(
0 if T > Tc ,
m ' √

h=0 −3t if T < Tc .
(590)
2. Magnetic susceptibility in zero field

(
1 −1
∂m kB Tc t if T > Tc ,
χ= ' 1 −1
∂h h=0

2kB Tc (−t) if T < Tc ,
(591)
3. Free energy per site and heat capacity per volume

(
0 if T > Tc
fMF ' −kB T ln 2 +

h=0 − 3kB4 Tc t2 if T < Tc .
(592)
(
C ∂ 2 fMF 0 if T > Tc
= −T ' 3kB
V ∂T 2 h=0

2 if T < Tc .
(593)
13 Critical Behaviour and Universality

Close to a critical point thermodynamic functions display power-law behaviours characterized by critical
exponents. We will now discuss various such exponents, using as a specific example the paramagnet to
ferromagnet transition.
1. Order Parameter
This is a quantity that is different in the various phases and can be used to characterize the phase
transition. For the paramagnet to ferromagnet transition in zero magnetic field the appropriate order
parameter is the magnetization per site
1
m(T ) = lim lim M (h, T ).
h→+0 V →∞ V
(594)
Here M (h, T ) is the magnetization. Where T ≈ Tc , one has

(
0 if T > Tc T − Tc
m(T ) ∼ t= .
|t|β if T < Tc Tc
(595)
β is a critical exponent.
86
m(T)
paramagnet
ferromagnet 1 T/Tc
Figure 24: Order parameter for the paramagnet to ferromagnet transition in zero field.
2. Susceptibilities
At the critical point the system is very sensitive to external perturbations. The singularity in the
response of the order parameter to a field “conjugate” to it is characterized by critical exponents γ± .
For our magnet
∂ 1
χ± (T ) = lim M (h, T ) ∼ |t|−γ± .
∂h h=0 V →∞ V

(596)
Figure 25: Critical behaviour of the magnetic susceptibility.
3. Heat Capacity
A third critical exponent is associated with the heat capacity
(
∂2F A+ |t|−α+ if T > Tc
C(T ) = −T ∼
∂T 2 A− |t|−α− if T < Tc .
(597)
Depending on the signs of α± this may or may not be singular, see Fig. 26.
13.1 Universality
The critical exponents are insensitive to microscopic details of the system under consideration and are
characteristic of the critical point. A consequence of this is that completely different systems can exhibit
87
Figure 26: Critical behaviour of the heat capacity.
the same critical behaviour!
14 Landau Theory
Landau Theory is a general approach to phase transitions that
• is phenomenological in nature and deals only with macroscopic quantities;
• applies only to the neighbourhood of a critical point, where the order parameter is small.
Landau theory is constructed as follows.
1. Identify the order parameter(s) M (r) characterizing the phase transition. Depending on which phase
transition we are dealing with M can be a real scalar, a complex scalar, a real or complex vector or
something more complicated. For the paramagnet to ferromagnet transition the order parameter is
the magnetization per site, i.e. a real number.
2. Form a “coarse-grained” order parameter density Φ(r). Think of this as the microscopic order param-
eter averaged over atomic distances. This is a continuum field. For the example of the paramagnet to
ferromagnet transition we are dealing with a real scalar field.
3. Consider the free energy density to be a functional of the order parameter field Φ(r). The free energy
is then
Z
βF = dD rf [Φ(r)].
(598)
4. By construction of the order parameter(s), the latter is (are) small close to our critical point. This
allows us to expand f [Φ(r)] is a power series around Φ = 0. From now on we will focus on the simplest
case of a real scalar order parameter φ(r). Then the series expansion is
1
f [φ] ' const − hφ + α2 φ2 + |∇φ|2 + α3 φ3 + α4 φ4 + . . .
2 (599)
where the coefficient of the gradient term is fixed by convention to be 1/2. This makes φ in general
dimensionful
dim[φ(r)] = (length)1−D/2 . (600)
The only linear term that is not a total derivative is −hφ, where h is an external field (a “source”) cou-
pling to the order parameter. Total derivative terms can be dropped, because the only give boundary
contributions to F . The coefficients αn are a priori all functions of temperature.
88
5. In translationally invariant systems the free energy is minimized by r-independent order parameters
(i.e. ∇φ(r) = 0). The reason is that 12 |∇φ|2 ≥ 0, and hence this contribution to F is minimized by
constant solutions. For constant field h the potential V (φ(r)) = −hφ + α2 φ2 + α4 φ4 is also minimized
by constant solutions. In order to understand the nature of the phase transition, we therefore can
simply look at the minima of the potential V (φ).
6. Finally, we use symmetries and the fact that we are interested in the vicinity of a critial point to
constrain the αj .
• If we truncate our expansion at order φ4 , then thermodynamic stability requires
α4 > 0.
(601)
If α4 < 0 the free energy density would be unbounded from below and could become infinitely
negative, which is forbidden.
• If we know that the system is invariant under certain symmetry operations, e.g.
φ → −φ, (602)
then the free energy must respect this symmetry. A ferromagnet has the symmetry (602) in
absence of a magnetic field because of time-reversal invariance. Hence we must have α3 = 0 in
this case.
• In the case h = α3 = 0, for a translationally invariant system, we can obtain the temperature
dependence of α2 as follows. As discussed above, the nature of the phase transition can be inferred
from the minima of the potential V (φ) = α2 φ2 + α4 φ4 . This is done in Fig. 27. We see that the
phase transition corresponds to α2 changing sign at T = Tc . So in the vicinity of the transition
we have (by Taylor expanding α2 in T − Tc )
T − Tc
α2 (t) = At + O(t2 ) , t= , A > 0.
Tc
(603)
The parameter α4 is also temperature dependent, but this dependence is subleading
α4 (t) = α4 (0) + O(t). (604)
7. If we have α3 < 0 the transition is generically first order. To see this we again use that in a trans-
lationally invariant system the minima of the free energy density will be r-independent, so that we
merely need to scrutinize the potential V (φ) to understand the nature of the phase transition. In
Fig. 28 we plot V (φ) when α2 is decreased at fixed α3 , α4 . We see that initially the minimum occurs
at φ = 0 (no order), and at some critical value of α2 then jumps from zero to a finite value φ0 . This
is characteristic of a first order transition.
14.1 Thermodynamic Equilibrium

The state of thermodynamic equilibrium is obtained by minimizing the free energy (we assume that α3 = 0)

1
Z
D 2 2 4
βF = d r −hφ + α2 φ + |∇φ| + α4 φ . (605)
2
In our case we are searching for the order parameter configuration φ(r) that gives the smallest contribution
to βF . This is found by functional extremization
δF
= 0. (606)
δφ(r)
89
Figure 27: Minima of the potential V (φ). For α2 > 0 the minima occurs at φ = 0, so there is no ferromagnetic
order. For α2 < 0 there are two minima at φ = ±φ0 , corresponding to the emergence of a spontaneous
magnetization.
The resulting nonlinear differential equation is
−∇2 φ(r) + 2α2 φ(r) + 4α4 φ3 (r) − h = 0.

(607)
It is easy to see (see the discussion above) that the solutions to (607) that minimize the free energy are
in fact r-independent (if we ignore boundary conditions). Hence for zero field h = 0 the order parameter
configuration that minimizes the free energy is
(
0 if α2 > 0 ↔ T > Tc
φ(r) = q
α2 .
±φ0 = ± − 2α 4
if α2 < 0 ↔ T < Tc
(608)
We observe that we are dealing with a second order phase transition (because φ0 vanishes at the critical
point) from a paramagnetic to a ferromagnetic phase.
14.2 Beyond the Landau Free Energy

So far we have focussed entirely on the state of thermodynamic equilibrium. We now want to extend Landau
theory to take into account fluctuations as well. Let us go back to the microscopic model underlying our
Landau free energy. The partition function for this microscopic theory is
X
Zmicro = e−βE(C) . (609)
configurations C
90
Figure 28: Minima of the potential V (φ) for α3 < 0. Decreasing the value of α2 leads to a discontinuous
jump in the order parameter at some critical value. The transition is therefore first order.
In order to define our order-parameter field, we used a coarse-graining procedure. Hence after coarse-graining
Z should become X
Z −→ e−βH . (610)
order parameter
configurations
Because under the coarse-graining procedure many microscopic configurations C map onto the same order
parameter configuration φ(r), the “Landau-Ginzburg Hamiltonian” βH incorporates certain entropic effects.
By employing the same logic as before, we can argue that
Z
βH = dD r f [φ(r)] , (611)
where f [φ(r)] is the same functional we constructed when considering the Landau free energy. As the order
parameter is really a continuous field, what we mean by the sum in (610) is really the functional integral
Z
Z = Dφ(r) e−βH . (612)
The latter is defined as follows:
• We first discretize our D-dimensional spatial variable
r −→ a0 n = a0 (n1 , . . . , nD ), (613)
where a0 is a lattice spacing and the total number of points on our discrete grid is N D .
91
• We then discretize the order parameter field and its derivatives
φ(r) −→ φn ,
φn+ej − φn
∂rj φ(r) −→ , (614)
a0
where ej are unit vectors in the j-direction.
• The Landau-Ginzburg Hamiltonian is discretized as
X
βH −→ f [φn ]aD
0 . (615)
n
• The functional integral is then defined as follows

Z Z Y
D
P
−βH
Z = Dφ(r) e ≡ lim dφn e− m f [φm ]a0 .
N →∞
n
(616)
Crucially, the functional integral defined in this way can be manipulated according to the same rules we
derived for path-integrals in Quantum Mechanics. In this new way of looking at things we now can analyze
properties that are not directly related to the free energy. For example, we may ask about properties of
correlation functions like
1
Z
hφ(r)φ(0)iβ ≡ Dφ φ(r)φ(0) e−βH . (617)
Z
14.3 Saddle Point Approximation

The Landau-Ginzburg field theory
Z
dD rf [φ(r)]
R
Z= Dφ(r) e− , (618)
is still difficult to analyze. For the example we have discussed, it reduces to the Euclidean space version of
the λφ4 theory you have encountered in the field theory part of the course. In order to proceed we therefore
resort to further approximations. The saddle-point approximation takes into account the thermodynamically
most likely configuration φ(r), i.e. the configuration that minimizes
Z
βH = dD rf [φ(r)]. (619)
We see that the saddle-point approximation precisely recovers the results of the Landau free energy approach!
However, using our new formulation we now go beyond this approximation and take into account fluctuations.
We will see below how to do this.
14.4 Mean Field Exponents

Using the saddle point solution we can determine the corresponding approximation for the critical exponents.
• Order parameter.
Using that α2 = At for t = (T − Tc )/Tc < 0, we have
r
A 1
φ0 = ± |t| 2 .
2α4
(620)
This gives the critical exponent
1
β= .
2 (621)
92
• Magnetic susceptibility.
Differentiating (607) with respect to h gives for r-independent solutions
∂φ ∂φ
2α2 + 12α4 φ2 = 1. (622)
∂h ∂h
The zero-field susceptibility is thus
∂φ 1
χ= = . (623)
∂h h=0 2α2 + 12α4 φ2

Using that α2 = At and φ20 = A|t|/2α4 this becomes

(
1
2At if t > 0,
χ= 1
4A|t| if t < 0.
(624)
This gives the critical exponents

γ± = 1.
(625)
• Heat capacity.
The heat capacity is defined by
∂2F
C(T, h = 0) = −T . (626)
∂T 2
The saddle point contribution to the free energy is (letting β ≈ βc to obtain F ):
(
F 0 if t > 0,
∼ A2 kB Tc t2 (627)
V − 4α4 if t < 0,
giving
(
0 if t > 0,
C(T, h = 0) ∼ A2 k B V
2α4 if t < 0,
(628)
Note that we have set T ∼ Tc in the prefactor. This has a finite jump at Tc , which corresponds to the
critical exponent
α = 0. (629)
• Correlation length exponent.

The exponents described above can all be obtained from the saddle point solution, or equivalently the
Landau free energy. This is not the case for the correlation length exponent ν, which is related to
fluctuations around the saddle point.
Away from the critical point the (connected) order-parameter two-point function decays exponentially
with distance
hφ(r)φ(0)iβ − hφ(r)iβ hφ(0)iβ = e−|r|/ξ , |r| → ∞. (630)
The correlation length ξ diverges when the critical point is approached
ξ ∼ |t|−ν .
(631)
93
The relation (631) defines the exponent ν. We now determine ν in what is known as the Gaussian
approximation. In the disordered phase this amounts to simply dropping the φ4 term in the free energy
density, i.e. by setting
Z
1 D 0 1
Z
0 2 2 0
hφ(r)φ(0)iβ ≈ Dφ φ(r)φ(0) exp − d r |∇φ(r )| + α2 φ (r ) . (632)
Z 2
It is not a priori clear that the Gaussian approximation will give a good account of th two point
function. It turns out to be good if the spatial dimensionality D is sufficiently high.
The two-point function (632) can be calculated using a generating functional
Z
D 0 1
Z
0 2 2 0 0 0
Z[h] = Dφ exp − d r |∇φ(r )| + α2 φ (r ) − h(r )φ(r ) . (633)
2
Indeed, we have
δ δ
hφ(r)φ(0)iβ = ln(Z[h]). (634)
δh(r) δh(0) h=0

We calculate the generating functional by going to Fourier space
dD p −ip·r dD p −ip·r
Z Z
φ(r) = e φ̃(p), h(r) = e h̃(p). (635)
(2π)D (2π)D
This gives
dD p p2
Z
βH = + α2 φ̃(p)φ̃(−p) − h(p)φ̃(−p) . (636)
(2π)D 2
Next we “complete the square” by changing variables to
h̃(p)
ϕ̃(p) = φ̃(p) − . (637)
p2 + 2α2
As the Jacobian of the change of variables is 1, this gives
!
dD p p2 dD p h̃(p)h̃(−p)
Z
1
Z Z
Z[h] = Dϕ̃ exp − + α 2 ϕ̃(p) ϕ̃(−p) exp . (638)
(2π)D 2 2 (2π)D p2 + 2α2
The first factor is merely a constant, which we will denote by N , while the second factor is rewritten
as
Z
1 D D 0 0 0
Z[h] = N exp d rd r h(r)G(r − r )h(r ) ,
2
dD p e−ip·r
Z
G(r) = . (639)
(2π)D p2 + 2α2
Taking functional derivatives we have
√
e−|r| 2α2
hφ(r)φ(0)iβ = G(r) ∼ , |r| → ∞. (640)
|r|D−2
This gives the correlation length
1 1
ξ=√ ∼ 1, (641)
2α2 t2
and thus the critical exponent
1
ν= .
2 (642)
94
Given the explicit calculation we have just done, we are now in the position to introduce a shortcut
for obtaining the two point function in the Gaussian approximation in similar situations. In absence
of a source h(r) the Landau-Ginzburg Hamiltonian (and the Landau free energy βF for that matter)
is written as
1 dD p h 2
Z i

βH = p + 2α 2 φ̃(p) φ̃(−p) + quartic . (643)
2 (2π)D
From this expression we can simply read off the result for the two point function in the Gaussian
approximation
(2π)D δ D (p + q)
hφ̃(p)φ̃(q)iβ = . (644)
p2 + 2α2
Here the denominator is given by whatever factor multiplies the quadratic term in βH. The delta-
function expresses momentum conservation.
In the ordered phase t < 0 we expand f [φ(r)] around one of the minima at ±φ0 . The choice of
minimum implements spontaneous symmetry breaking. We have
V (φ) = α2 φ2 + α4 φ4 ' α2 φ20 + α4 φ40 + (α2 + 6α4 φ20 )(φ − φ0 )2 + . . . (645)
We may drop the constant and retain only the contribution quadratic in δφ = φ − φ0 (this is the
Gaussian approximation in the ordered phase), which gives
1
f [δφ(r)] ' |∇δφ|2 + α̃2 δφ2 . (646)
2
Here α̃2 = −2α2 > 0. We may now copy the calculation in the disordered phase and obtain for the
connected correlation function
√
e−|r| −4α2
hδφ(r)δφ(0)iβ = hφ(r)φ(0)iβ − hφ(r)iβ hφ(0)iβ ∼ , |r| → ∞. (647)
|r|D−2
We see that the correlation length scales as ξ ∝ |t|−1/2 , giving again ν = 1/2.

Question 19. Consider a Landau expansion of the free energy of the form
a 2 b 4 c 6
F = m + m + m
2 4 6
with c > 0. Examine the phase diagram in the a−b plane, and show that there is a line of critical transitions
a = 0, b > 0 which joins a line of first order transitions b = −4(ca/3)1/2 at a point a = b = 0 known as a
tricritical point.
Supposing that a varies linearly with temperature and that b is independent of temperature, compare
the value of the exponent β at the tricritical point with its value on the critical line.
From Yeomans, Statistical Mechanics of Phase Transitions
Question 20.
(a) Discuss how an order parameter may be used to characterise symmetry breaking at a phase transition.
95
(b) Argue that the uniaxial ferromagnet-paramagnet transition can be described by a Landau free energy
of the form
1
Z
3 2 2 3 4
F = d r |∇φ(r)| − hφ(r) + α2 φ (r) + α3 φ (r) + α4 φ (r) . (648)
2
What can you say about α4 ?
(c) What is the nature of the transition for h = 0 if α3 6= 0? Explain your answer.
(d) Now assume that α3 = h = 0. Argue that close to the critical point
T − Tc
α2 = At , t= and A > 0. (649)
Tc
(e) Derive the equation characterizing the saddle point solution for α3 = h = 0. What are the configurations
φ with the lowest free energy for h = 0, at T > Tc and at T < Tc ? Why are these r independent?
(f) Now consider more general solutions to the saddle point equation in the low-temperature phase. With
suitable boundary conditions the saddle point solutions for the order parameter are functions of x only, i.e.
φ = φ(x). Show that in this case
1 dφ(x) 2

E= − α2 φ2 − α4 φ4 (650)
2 dx
is independent of x. Construct a solution φ(x) such that
lim φ(x) = φ1 , lim φ(x) = φ2 , (651)

x→∞ x→−∞
where φ1,2 are the solutions found in (d). Hint: determine E for such solutions first.
Question 21. A system with a real, two-component order parameter (φ1 (r), φ2 (r)) has a free energy

1 1 1 2 1 2
Z
d 2 2 2 2
2
F = d r |∇φ1 (r)| + |∇φ2 (r)| − φ (r) + φ2 (r) + φ (r) + φ2 (r) .
2 2 2 1 4 1
Find the order-parameter values Φ1 , Φ2 that minimise this free energy. Consider small fluctuations around
such state, with (φ1 (r), φ2 (r)) = (Φ1 + ϕ1 (r), Φ2 + ϕ2 (r)) and expand F to second order in ϕ.
Assuming that the statistical weight of thermal fluctuations is proportional to exp(−F ), calculate ap-
proximately the correlation function
hϕ1 (r)ϕ1 (0) + ϕ2 (r)ϕ2 (0)i
by evaluating a Gaussian functional integral. How does your result depend on the dimensionality d of the
system?
15 Other Examples of Phase Transitions

15.1 Isotropic-Nematic Transition in Liquid Crystals
Liquid crystals are fluids of rod-like molecules. At high temperatures their centres of mass are randomly
distributed and the rods are randomly oriented. At low temperatures, in the nematic phase the rods
spontaneously align along a common axis, see Fig. 29. What is the order parameter characterizing this
transition?
Let us associate a unit vector ~n(r) with a molecule at position r. The first guess one may have is to take
h~n(r)iβ as the order parameter. This will not work, because the two vectors ~n(r) and −~n(r) describe the
same orientation of the molecule. Hence the order parameter must be invariant under
~n(r) −→ −~n(r). (652)
96
high temperatures low temperatures
→
n
Figure 29: At low temperatures the rod-like molecules spontaneously align along an axis ~n.
So how about something quadratic like

hni (r)nj (r)iβ . (653)
The problem with this expression is that it is different from zero even for randomly oriented molecules
(which is what one has at very high temperatures). Indeed, using a parametrization of the unit vector for
a single molecule in terms of polar coordinates we have
 
sin θ cos ϕ
~n =  sin θ sin ϕ  . (654)
cos θ
Then averaging over all possible orientations gives

Z π Z 2π
1 1
hhni nj ii = dθ dϕ sin θni nj = δi,j 6= 0. (655)
4π 0 0 3
This consideration suggests to try

1
Qij = hni nj iβ − δi,j
3 (656)
as our order parameter. At very high temperatures, when molecules have random orientations, this is zero.
On the other hand, if the molecules are aligned in the z-direction, i.e. ~n = ~ez , we have
 1 
−3 0 0
Q =  0 − 13 0  . (657)
0 0 23
So this seems to work. Given the order parameter, how do we then determine the free energy? In the high
temperature phase the free energy must be invariant under rotations of the molecules, i.e. under
Q(r) −→ RQ(r)RT . (658)
97
This suggests the following expansion for the free energy density
1 2
f [Q(r)] = |∇Q|2 + α2 Tr[Q2 ] + α3 Tr[Q3 ] + α4 Tr[Q2 ] + . . .
2 (659)
Here |∇Q|2 = 3i,j,k=1 (∂k Qij )2 . The presence of a cubic term suggests that the transition is first order,
P
which is indeed correct.
15.2 Superfluid Transition in Weakly Interacting Bosons

Let us recall the second quantized Hamiltonian for weakly repulsive bosons
∇2

U †
Z
4 † †
H = d r c (r) − c(r) + c (r)c (r)c(r)c(r) . (660)
2m 2
In the superfluid phase we have macroscopic occupation of the zero momentum single-particle state
p
hc(p = 0)iβ = N0 . (661)
This implies that

ψ(r) = hc(r)iβ 6= 0 , T < Tc . (662)
At high temperatures the expectation value vanishes, so that we may take ψ(r) as a complex valued order
parameter. The Landau free energy describing the transition can then be constructed by noting that it
should be invariant under U (1) transformations
ψ(r) −→ eiθ ψ(r). (663)
This gives

1
Z
D 2 2 4
F [ψ] = d r |∇ψ| + α2 |ψ(r)| + α4 |ψ| .
2m
(664)
16 Ising Model and Φ4 Theory

The 3-dimensional Ising model has a paramagnet to ferromagnet phase transition that can be described by
Landau theory. An obvious question is then whether it is possible to derive the appropriate Landau theory
starting from the lattice model. Let us therefore consider a general Ising model with Hamiltonian
X X
H=− eij σj − h̃
σi C σi . (665)
i,j i
Let’s work on a hypercubic lattice in D dimensions, and take
C e i − rj |)
eij = C(|r (666)
to be translationally invariant. The patition function is

X P P
Z= e j,` σj Cj` σ` +h ` σ` , (667)
{σj }
where we have defined h = β h̃ and Cij = β C

eij . Recalling the Gaussian integration identity
Z Y
1 P 0 −1 0
dψ`0 e− 4 j,k ψj Cjk ψk ,
p
1 = det(4πC ) −1 (668)
| {z }
N | ` {z }
Dψ 0
98
and then shifting the integration variables
X
ψj0 = ψj − 2 Cjk σk , (669)
k
we arrive at the following identity
Z
1 P −1 P P
1=N Dψ e− 4 j,k ψj Cjk ψk + j σj ψj − j,k σj Cjk σk
. (670)
Substituting this into our expression for the partition function gives
Z X − 1 P ψ C −1 ψ +P σ (ψ +h)
Z = N Dψ e 4 j,k j jk k j j j .
{σj }
(671)
What we have done is to remove the interaction between the spins by introducing auxiliary variables ψj
at every site. This is an example of a Hubbard-Stratonovich transformation. Now we can perform the
summations over spin variables
X P Y
e k σk (ψk +h) = [2 cosh(ψk + h)] . (672)
{σj } k
Thus Z
1 P −1 P
Z=N Dψ e− 4 j,k ψj Cjk ψk + j ln(2 cosh(h+ψj ))
(673)
Finally we change variables to
1 X −1
φj = Cjk (ψk + h), (674)
2
k
which leads to the following expression for the partition function
Z P P P P
Z=N 0
Dφ e− j,k φj Cjk φk +h j φj + j ln(cosh(2 k Cjk φk )) . (675)
This is still at the level of the lattice model. We now want to focus on large distances/small energies, in order
to obtain a field theory description. We furthermore will now assume that the matrix C, which depends
on temperature, is such that large fluctuations of φj are supressed. We then can take φj to be small and
expand
X X2 X4
ln[cosh(2 Cjk φk )] ' − + .... (676)
2 12
| k {z }
X
Finally, we go to Fourier space using
1 X −ik·r 1 X −ik·(rj −rk )
φj = e φ(k) , Cjk = e C(k), (677)
N N
k k
and focus on the small-|k| regime. This gives
Z c P
−1 φ(k)[c1 +c2 k2 +... ]φ(−k)+c3 hφ(0)+ 43
P
Z=N 00
Dφ e N k N k1 ,k2 ,k3 φ(k1 )φ(k2 )φ(k3 )φ(−k1 −k2 −k3 )+... , (678)
where the coefficients cl depend on the matrix C

ejk specifying the interactions of Ising spins. Going back to
4
real space we recover Euclidean Φ theory
Z
Dφ e− d x[c1 (∇φ(x)) +c2 φ (x)+c3 hφ(x)+c4 φ (x)] .
R D 2 2 4
00
Z=N (679)
Voilà. Note however, that the relation of the field φ(x) to the original Ising spin is rather indirect. In order
to get a better feeling of “what happens to the spin” under all the above transformations, we could repeat
the above steps for the generating functional rather than the partiton function. This will show us that Φ is
indeed a good order parameter for the Ising transition.
99
17 Transverse-Field Ising Model
We now use the methods of many-particle quantum mechanics to study aspects of the transverse-field
Ising model (TFIM). As a reminder this model was obtained by interpreting the transfer matrix of the 2D
classical Ising model as the transition amplitude for a quantum 1-dimensional Hamiltonian (see (306)). The
TFIM is given by the following Hamiltonian for spin-1/2 particles arranged on a 1D lattice: The degrees of
freedom of this model are quantum spins- 12 on a lattice, with dynamics described by the following quantum
Hamiltonian, " L #
X X
HTFIM = −J σiz σi+1
z
+g σix . (680)
i=1 i
where σiµare Pauli matrices associated to sites i (we have gone back to using σ instead of X, Y, Z; also,
to unclutter notation, we have dropped the bar on J introduced in (680) and also the hats on operators.)
Pauli matrices on sites i 6= j commute. We may or may not choose to impose P periodic boundary conditions
(in which case we take ZL+1 ≡ Z1 , etc. Also, we we will often just write i with the understanding that
the limits are as abov. In this section we will discuss various properties of this model.
Model (680) describes a set of spins- 21 on a line, in the presence of an external transverse magnetic field
(we will explain the meaning of ‘transverse’ in a moment.) The ‘spins’ can arise in a variety of different
contexts, and it’s best to think of these as ‘generic’ two-level systems — for instance, in a condensed-matter
context these could arise due to some conspiracy of interactions of electrons with ions and each other, that
forces the electrons to remain tied to individual lattice sites, and interact only through their spin degrees
of freedom. Another example is provided by spinless fermions or hard-core5 bosons but now free to hop
around; the presence or absence of a particle on a site is our ‘Ising’ variable. For our purposes, the origin of
the spins is not terribly important; we will just work with the simple model that we will introduce in the
next section. In this section6 , we will study the transverse field Ising model in one spatial dimension, often
also called the quantum Ising chain.
Note that J simply sets the overall energy scale and hence plays a rather minor role; the physics is
controlled by the dimensionless parameter g. The reason we refer to the second term as a transverse field
is that it ‘points’ in a direction different from that favored by the first term, and as such competes with
the first term. The essentially quantum mechanical nature of the model stems from the fact that the
two terms don’t commute: as such, the uncertainty principle guarantees for us that if we are in a limit
where one term is ‘pinned’ to a definite value, we have infinite quantum-mechanical uncertainty in the
other term. However, one might ask why P the firstPterm takes such a complex form; after all, a simpler
Hamiltonian, of the form Hboring = −J [ i σiz + g i σix ] also has such competition. The difference lies in
the evolution of the eigenstates as a function of g: it is easy to see that in Hboring , we may simply rotate
our spin coordinate
p system about the y-axis by an angle θ = tan−1 g, so that in the new coordinate system,
Hboring → −J 1 + g 2 i σiz , and there is no competition — thus, as we change g, we do not expect to see
P
anything but smooth behavior of various quantities computed from Hboring . In contrast, we cannot ‘rotate
away’ the competition in HTFIM in this way, and (as we will see) the evolution of the physics as a function
of g must become singular somewhere between g 1 and g 1 (and in fact, occurs at precisely g = 1.)
The key distinction between Hboring and HTFIM is one of symmetry, that we now explain.
17.1 Symmetries
HTFIM possesses an Ising symmetry (also called a Z2 symmetry) of taking up spins to down spins, imple-
mented via the mapping
U : (σx , σy , σz ) → (σx , −σy , −σz ) (681)
Clearly, doing this twice is equivalent to the identity — hence the symmetry group is Z2 , and its elements
are (1, U ) where U 2 = 1. Note that the transformation (681) is equivalent to a π rotation of all the spins
5
The term ‘hard-core’ here simply means no more than a single boson can sit at a site.
6
The treatment here closely parallels John McGreevy’s excellent notes, but will differ on some points.
100
about the x-axis, generated by the operator
P 1 x Y i π σx Y π π Y Y
U = eiπ j 2 σj = e 2 j = cos + iσjx sin = iN σjx ∼ σjx (682)
2 2
j j j j
where we have used elementary results from the quantum theory of spin, as well as the fact that spin
operators on different sites commute, and finally ignored an overall constant prefactor of iN . Clearly,
U † HTFIM U = HTFIM , hence U is a symmetry. We now analyze the manner in which this symmetry is
realized by the ground state of HTFIM in the limits g 1 and g 1.
17.2 Phase Diagram

17.2.1 Disordered Phase
Consider HTFIM in the limit g 1. As a first step in this limit, we may ignore the first ‘ferromagnetic’
term, and just satisfy the second ‘transverse field’ term. This gives a unique ground state, where all the
spins point along the x direction,Q that we represent as |0i = | →→→ . . . →i. Clearly, this ground state
preserves the symmetry: U |0i = σi | →→→ . . . →i = (1)N | →→→ . . . →i = |0i. The first excited states
x
in this limit are obtained by reversing a single spin at point i:

|1i = | ←→→ . . . →i, |2i = | →←→ . . . →i, . . . |ii = |1i = | →→ . . . ←i . . . →i, etc. (683)
There are clearly N such ‘spin flip’ states in an N -site system, all of which are degenerate in the absence
of the ferromagnetic term, and have energy ∆ = 2gJ above the ground state (whose energy we label E0 .)
Adding the ferromagnetic term as a perturbation gives the spin flips a dispersion: let us for the moment
only work in the subspace of a single spin flip. It is easy to see that the term σiz σi+1
z just moves a spin flip
from site i to site i + 1 (or vice-versa). Thus, to leading order in perturbation theory in the ferromagnetic
term in this sector, we may write the (approximate) eigenvalue equation in terms of ‘hopping’ of spin flips:
H|ii = −J [|i + 1i + |i − 1i] + (E0 + 2gJ)|ii. (684)
It is easiest to consider the system to live on a circle with N sites, so that when we may identify 0 ≡ N ,
and thus interpret the hopping terms in this manner. Then, we may readily solve by means of a Fourier
transform: let us write
1 X ikrj
|ki = √ e |ji (685)
N j
where the locations of the spins are at sites rj = ja. Periodic boundary conditions require k = 2π Nma with m
an integer, for |ki to be single-valued; also, only m = 1, 2, . . . , N are distinct (which is good, since we can’t
have more than N states arising from any linear combination of spin flips!). In the thermodynamic limit of
N → ∞, k becomes a continuous variable in the first Brillouin zone, i.e. the segment k ∈ [−π/a, π/a].
Inserting this into our effective hopping problem for spin flips, we find
(H − E0 )|ki = (2gJ − 2J cos k)|ki (686)
In other words, we find the gap to the first excited state at k follows the pattern (k) = 2J[g − cos k]. A
spin flip thus behaves like a particle excitation: for a given momentum, it has a fixed energy (k). Hence,
the lowest-lying excitations above the symmetric g 1 ground state are gapped particles, with an energy
gap ∆ = 2J(g − 1) and a dispersion (k) ∼ ∆ + Jk 2 for small momenta. σix behaves like a particle ‘number
operator’ as it measures the number of spin flips, while σiz inserts a single spin flip at site i and is associated
with creating a particle. Note that creating two particles at a single site is equivalent to doing nothing —
suggesting that the particle is its own antiparticle, a point we will return to below.
In the above discussion, we assumed g 1. Suppose we simply extend the calculation well beyond its
regime of validity, to g ∼ 1; we anticipate a transition at g = 1, where the gap to excitations closes, and
the spin-flips ‘condense’. Serendipitously, it turns out that this is the exact value at which the ground state
changes its fundamental nature — a fact which is generally never true (one should not be able to predict
critical parameters from such a theory in most cases) but lies on the fact that the Ising model in d = 1 is
self-dual. Before discussing duality, we will first try to understand the ordered phase
101
17.2.2 Ordered Phase
We now turn to the opposite limit of HTFIM , namely the ‘weak coupling’ limit when g 1. Let us begin
as before, by first setting g = 0 and then perturbatively considering what hapens at finite but small g.
At g = 0. we retain only the ferromagnetic term, which is clearly minimized by either of the two states
|+i = | ↑↑ . . . ↑i or |−i = | ↓↓ . . . ↓i. Clearly, these ground states break symmetry: U |+i = |−i, since the
action of σ x is to flip a spin pointing up along the z-axis to one pointing down. Note however that linear
combinations of the two states of the form |+i ± |−i actually do preserve the symmetry; for any non-zero
g, there is a non-zero matrix elements to mix these states, since the transverse field terms can flip spins.
Let us now estimate this matrix element in perturbation theory. The perturbing term here is the transverse
field; this can flip the spin on a single site. Thus, in order to obtain a non-vanishing matrix element between
the state |+i, with all spins up, and the state |−i with all spins down, we need to go to N th order in
perturbation theory, where N is the size of the system. The resulting effective Hamiltonian takes the form
(here s, s0 = +, −),
E0 g N

0
hs|Heff |s i = (687)
g N E0
from which we see that the eigenstates are |ψ± i ∝ |+i±|−i, with a small splitting δ = |E+ −E− | ∼ O(g N ) =
−N log g1
O(e ) (remember, g 1) — in other words, the splitting between the states vanishes exponentially
as N → ∞. We will provide a more physical description of this in a moment, but let us assume that
this splitting does indeed vanish; then, we can see how spontaneous symmetry breaking appears. While
at any finite system size the ground states are indeed of the ‘Schrödinger cat’ form |+i ± |−i and hence
actually respect the symmetry, if one initially prepares the system in (say) all up, it takes an extremely
long time (exponential in the size of the system) to evolve out of this state to demonstrate that it is not an
eigenstate. So for all practical purposes, we may consider |+i and |−i to be eigenstates, and can improve
the approximation by going to larger and larger systems while keeping other parameters fixed. (Note that
we could also impose a symmetry-breaking longitudinal field, take the thermodynamic limit, and then take
the field to zero, just as we did when diagnosing SSB in a classical system.) We will see later that this
degeneracy is mirrored by a certain topological robustness in a different description of the TFIM in terms
of fermions.
To understand the excited-state spectrum in the g 1 limit, let us consider the state |+i = | ↑↑ . . . ↑i.
What is the lowest-energy excitation here? For simplicity, w econsider an open chain, and we might naively
answer that this is again a spin flip, where at a single site we send ↑→↓. This costs energy ∆ = 4J. However,
we can do better. Consider a domain wall in the spin configuration, where between site i, i + 1 the spin
goes from ↑ to ↓: |. . . ↑↑ . . . ↑i |↓i+1 ↓ . . .i. This has a lower energy than a spin flip: ∆ = 2J. With periodic
boundary conditions, we have to make a pair of domain walls (can you see why?), but these are in fact
independent excitations — it turns out that single spin flip in the Ising model is not a ‘particle’ in that it
does not in general have a sharp energy-momentum relation. Indeed, the domain walls are the true ‘particle’
excitations, as we will show in a moment; so a spin flip of a given momentum will decay by partitioning its
energy and momentum into a pair of domains walls, i.e. it merges with the ‘two-particle continuum’. The
domain walls are however non-local objects: one needs to change the state of a macroscopic number of spins
to create a single domain wall far from any others.
Let us analyze the spectrum of the single domain wall sector. To avoid the complications of disentangling
the behavior of a single domain wall from the two-domain-wall sector that is the minimal accessible one
with periodic boundary conditions, we work directly in the N → ∞ limit. It is most convenient to think
of the domain walls as living on links of the lattice, and label a state with a domain wall P between sites i,
1 x
i + 1 as |ī = i + 2 i. As before, in the absence of the perturbation (in this case, the g i σi term) there
are N degenerate, localized domain wall 7 . These states acquire a dispersion for g 6= 0: within the
Pstates
x
single-domain-wall sector, all that the g i σi term is ‘hop’ the domain wall one spacing to the left or right.
7
We are glossing over the fact of boundary conditions in making this kind of counting argument; it is in fact possible to
consider single domain walls on a ring by also allowing boundary conditions that are antiperiodic, rather than periodic.
102
Hence, we arrive at a very similar problem in the one-domain-wall sector as we did in the single-spin-flip
sector:
(H − E0 )|īi = −gJ [|ī + 1i + |ī − 1i] + 2J|īi (688)
Fourier transforming again (though now observe that the spatial degrees of freedom lie on the mid-points of
the links, rather than on lattice sites) we obtain the dispersion (k) = 2J[1 − g cos k] for the domain walls.
Once again, we see that the (k) becomes negative at k = 0 when g = 1, and we may expect a transition,
now driven by the condensation of domain walls.
This argument also allows us to see why the broken-symmetry state is stable: one needs to create a
pair of domain walls and make them move around the system ‘unzipping’ up spins to down spins, and then
annihilate when they span the system. However, these are gapped excitations: it first costs energy 4J to
nucleate the domain wall pair, and then each step that the domain wall moves costs an energy of order gJ.
−N log g1
Computing the action for this process, we find S ∼ e as we computed earlier, which vanishes in the
thermodynamic limit N → ∞.
Note that in the g → 0 limit, the system has long range order in the z-basis, since all the spins point
in the same direction. We may also think of the ordered phase as a condensate of spin flips out of the
paramagnetic (disordered) phase that forms at large g: consider the operator σiz , that creates a single spin
flip; we have for its expectation value at g = 0
hσjz i = h±|σjz |±i = ±1 for any j (689)
In the opposite limit of g → ∞, we see that hσjz i = 0, and the value does not deviate from this until we
start condensing spin flips, which occurs in the vicinity of g = 1. We can thus think of hσjz i as an order
parameter for the Ising transition, as it is nonzero only in the broken-symmetry phase.
17.3 Duality
We have thus far studied the TFIM in two limits: the disordered or ‘paramagnetic’ phase for g 1, and
the Ising-ordered broken symmetry ferromagnet for g 1. We found that in both limits there is a gap to
excitations, with low-lying particle-like excitations: spin flips and domain walls, in the two limits. We found
that the dispersion of these excitations have a simple form:
spin flip (k) = 2J[g − cos k]

domain wall (k) = 2J[1 − g cos k] (690)
There is a suggestive symmetry between these expressions, that seems to require

1
g→ , J → Jg. (691)
g
It turns out there is a rigorous duality between strong (g 1) and weak (g 1) coupling in the quantum
Ising model, that maps spin flips to domain walls and vice versa.
First, assume open boundary conditions, so that the system just stops at j = 1 and j = N , so we can
avoid the even-domain-wall constraint. Any state in the σ z basis can be specified by the direction (sz =↑, ↓)
of the first spin along the z axis, and the locations of the domain walls. Consider the operator8 , diagonal in
this basis, that measures whether there is a domain wall at j̄ = j + 21 , i.e. between sites j, j + 1:
+1 if szj̄− 1 = szj̄+ 1
(
τj̄x ≡ z
σj̄− 1σ
z
j̄+ 12
= 2 2 = (−1)is there a domain wall at j̄? . (692)
2 −1 if szj̄− 1 = −szj̄+ 1
2 2
8
When discussing lattice dualities, we must be careful to pay attention to site variables, since dual variables live on links
and ‘direct’ variables live on sites.
103
Note that τj̄2 = 1, and τj̄† = τj̄ . Now, consider the operator that inserts a domain wall at j̄: we implement
this by flipping all spins to the right of j̄, i.e.
Y
τj̄z ≡ σj̄+
x
1σ
x
j̄+ 3 ··· = σjx . (693)
2 2
j>j̄
This operator is also Hermitian, and squares to the identity. Now, we observe the following neat identity:
τj̄z τj̄x0 = (−1)δj̄ j̄ 0 τj̄x0 τj̄z (694)
— just like the original Pauli matrices, σjz , σjx0 ! (To see this, note that if j̄ 0 < j̄, τj̄z and τj̄x0 don’t share any
sites and hence commute. If j̄ 0 = j̄, then τj̄z and τj̄x0 share exactly one site, namely j̄ + 21 ; as τj̄z contains σj̄+
x
1
2
and τj̄x contains σj̄+
z
1 , commuting these past each other ‘inherits’ a factor of (-1) from the anticommutation
2
of two Pauli matrices. Finally, if j̄ 0 > j̄, then τj̄z and τj̄x0 have two sites in common, and the overall sign after
commuting the relevant Pauli matrices σiµ is (−1)2 and thus cancels out.) This anticommutation relation
tells us that domain walls can be represented in terms of two-state systems living on the links, suggesting
that their character is quite similar to the original Ising spins.
Now, the inverse map from σ x , σ z to τ x , τ z is simply given by
σjx = τj−
z
1τ
z
j+ 1
(695)
2 2
reflecting the fact that a single spin flip is equivalent to a pair of adjacent domain walls.
So, in the bulk, we have
X X
σjz σj+1
z
+ gσjx = −J τj̄x + gτj̄z τj̄+1
z

HTFIM = −J (696)
j j̄
Note that the second expression is a new TFIM, with couplings mapped via (691). Thus, we have ‘dualized’:
the new degrees of freedom are the domain walls in the original Ising spins. It turns out that there is a
similar mapping in the D = 2 classical Ising model due to Kramers and Wannier; due to the quantum to
classical mapping, we will see that this is the same duality as discussed here. One important thing to note
is the following: suppose the model has a single phase transition. Since the transition occurs as we move
from strong to weak coupling, and (691) interchanges the role of these, if there is a single phase transition,
then it must occur at the so-called ‘self-dual point’ g = 1.
Note that the paramagnetic (disordered) phase of the original variables is a condensate of domain walls
in the following sense: the operator that creates a domain wall has an expectation value
Y
hτj̄z i = hgsg=∞ |τj̄z |gsg=∞ i = hgsg=∞ | σj̄z |gsg=∞ i = 1 for any j̄. (697)
j>j̄
For 1 < g < ∞, this expectation value is between 0 and 1 — mirroring the behavior of σjz but in reverse.
While there is a condensate, there is no order: a nonzero expectation value of σ x does not break any
symmetry of HTFIM . (Note that there is another state with the opposite sign, one with all spins pointing
‘left’, but for large g it’s a very high-energy state.) The domain wall creation operator is thus often referred
to as a disorder parameter, or a disorder operator.
An important point to note (and why you should always treat the word ‘duality’ with some suspicion!)
is that the two sides of the phase diagram are not the same despite our mapping. On one side, there are
two ground states related by the broken symmetry, on the other there is a unique symmetric ground state
– indeed, it is this fact that requires the existence of an intervening phase transition!
It seems that since the duality transformation exchanges ordered and disordered phases, it also exchanges
the ground state degeneracies. The disordered phase of the dual theory corresponds to the broken-symmetry
phase of the Ising variables, and as such should have a twofold ground state degeneracy; similarly, the ordered
104
phase of the dual theory naively has two distinct ground states, but as it maps to the Ising paramagnet,
there is again an apparent mismatch. To see that the dual theory actually does represent the phase diagram
and the ground state degeneracies correctly, we need to be more precise than we have been about what
happens at the boundaries. Consider N sites in a row, where the j = 1 and j = N sites have only one
neighbor. We have
N
X −1
σjz σj+1
z
+ gσjx + gσN
x

HTFIM = −J (698)
j=
Let’s now perform the duality map: we have (not using our ‘barred’ link notation for clarity)
σjz σj+1
z x
= τj+ 1 , j = 1, 2, . . . , N − 1
2
σjx = τj−
z
1τ
z
j+ 1
, j = 1, 2, . . . , N (699)
2 2
so that in terms of domain-wall variables, we have

 
N −1
X
z z 1 x
HTFIM = −Jg  τj− 1τ
j+ 12
+ τj+ 1 + τNz − 1 τNz + 1  (700)
2 g 2 2 2
j=1
We have two special cases:
• τNz + 1 = 1: this flips all spins with j > N ; as there are no such spins, it must be the identity operator.
2
• τx
1 = σ0z σ1z : this measures whether or not there is a domain wall between j = 1 and j = 0; but there
2
is no spin at j = 0. One way to resolve this is to put a fiducial ‘ghost spin’ at j = 0 that always points
up. Then, τ 1x = σ1z : it measures the value of the ‘reference spin’ at the origin, with respect to which
2
we define the domain wall.
PN −1 x , and the ground state has τ x
At g = 0, we have Hg=0 = −J j=1 τj+ 1
j+ 1
for j = 1, 2 . . . , N − 1. Missing
2 2
in this list is τ 1x , so there are two degenerate ground states, eigenstates of τ 1x with eigenvalues ±1 — these
2 2
are just the ‘no domain wall’ states |±i: since we have τ 1x , all the spins to the right agree with the first one.
2
PN −1 z z z
At g = ∞, we have Hg=∞ = −Jg j=1 τj− 1 τj+ 1 + τ
N− 1
(where we have set τNz + 1 = 1). The first
2 2 2 2
z
term requires that τj− z for j = 1, 2, . . . , N − 1 but the second term forces τNz − 1 = 1, thus fixing the
1 = τ
2
j+ 12 2
sign. So, we obtain the unique paramagnetic state.
In other words, in the dual language, the freedom to choose τ 1x maps into the twofold degeneracy of the
2
ordered phase of the original ising spins, and the fixed value of τNz + 1 = 1 ensures that the paramagnetic
2
phase of the original model has a unique ground state.
12: Aside: Duality with Periodic Boundary Conditions
How do we carry out the duality transformation with periodic boundary conditions? The specific
problem is to understand how to define the creation operator for a single domain wall with periodic
boundary conditions on the spins. It turns out this is impossible, since the fact we can only make an
even number means that the sign of the operator is ill-defined — a single domain wall operator cannot
be measured (we will see this also in a different sense, in that a single domain wall is a Majorana
fermion that does not have a sense in which we can think of it as being occupied or empty.)
105
In spite of the sign ambiguity, we can ‘fix’ this by fiat as follows: and define the domain wall operators
on a ring as follows:
τj̄x = σj̄−
z
1σ
z
j̄+ 12
2
Y
τj̄z = σjx (701)
N ≥j>j̄
In the second line, observe that we have made a choice: even though σN z z
+1 = σ1 (as per periodic
boundary conditions), we’ve decided to ‘mark’ the link j̄ = N + 21 to stop the strong of σ x s. This is
arbitrary; we could have marked any other link, the important thing is that we have some prescription
for terminating the string. As in the case of open boundary conditions, this means that τ 1z and τNz + 1
2 2
are special: we have
τNz + 1 = 1 (702)
2
since the product is over the empty set, while

N
Y
τ 1z = σjz = U, (703)
2
j=1
the generator of the global Ising symmetry. This is significant, since [HTFIM , U ] = 0, so that eigen-
states of τ 1z are energy eigenstates. Also, we see that it determines the boundary conditions on the
2
τjz as follows: we observe that
τ 1z σjx = U = 1
( Q
if j
1= τNz + 1 = 2 (704)
−τ 1z σjx = U = −1
Q
2 if j
2
Thus, the distinction between the two symmetry sectors of the Ising model on a ring gets mapped into
the distinction between periodic and antiperiodic boundary conditions for he domain-wall operator
in the dual description; the operator that says whether the domain wall insertion operator is periodic
or antiperiodic is now a dynamical variable.
17.4 Mean Field Theory

So far, we have studied the phase structure of the d = 1 TFIM in detail in two distinct limits, g 1 and
g 1. Let us now try and guess the phase diagram over all g, using a variational approximation9 known as
‘mean field theory’, in which we completely ignore the entanglement10 between different sites, and assume
that the ground state is a product state,
|MFTi = |ψ1 i ⊗ |ψ2 i ⊗ · · · ⊗ |ψN i. (705)
Let us further assume a translationally invariant solution, so that each site is identical. Then, the state at
each site is parametrized by a unit vector ~n, i.e. a point on the Bloch sphere, and our uniform variational
9
Note that the energy expectation value calculated within any variational state is an upper bound for the ground state
energy: since the true ground state minimizes the energy, it had better improve on our naive guesswork wavefunction!
10
You might wonder if it’s possible to come up with variational wavefunctions that systematically build in ‘just enough’
entanglement to capture correlations beyond mean-field theory. Indeed, it turns out that the density-matrix renormalization
group can be viewed as a variational optimization over a class of states that do just that: so-called matrix product states, that
provide a remarkably efficient representation of one dimensional systems. Generalizations of this to higher dimensions using
tensor networks is a line of research that is being vigorously pursued.
106
state is given by the product state

θ θ −iϕ
|~ni = ⊗j |n~j i = ⊗j cos | →ij + sin e | ←ij (706)
2 2
Here, we have parametrized the angles with reference to the x-axis in spin space: θ measures the deviation
of the spin from the x-axis, and ϕ represents its azimuthal angle in the yz plane. To compute the varia-
tional energy of our mean-field solution, we must compute expectation values, which (owing to the lack of
entanglement) follow straightforwardly from those for a single site11 :
hn~j |σjx |n~j i = cos θ, hn~j |σjz |n~j i = sin θ cos ϕ (707)
So, we find the variational energy
E(θ, ϕ) = −N J sin2 θ cos2 ϕ + g cos θ .

(708)
Clearly, this is extremized when ϕ = 0, π, independent of θ; the value of θ that optimizes the energy is then
given by 0 = ∂θ E(θ, ϕ = 0, π) = N J sin θ(2 cos θ − g).
We see that when θ = 0, the two solutions for ϕ are the same since the azimuthal angle loses its meaning.
This makes sense: looking back, θ = 0 corresponds to the paramagnet, and we expect a unique solution, so
having two different ϕ solutions makes no sense. The θ = 0 state clearly corresponds to the solution where
the spins are all polarized along the x-axis. Now, this is indeed the unique solution for g > 2, since there is no
real solution for cos θ = g/2 in this regime; we may verify that it is a minimum of the energy by computing
its second derivative. For g < 2, however, we find that in fact the θ = 0 solution is a maximum (you should
verify this) and instead two new minima appear at θ = cos−1 g/2 6= 0, and we have hσiz i = sin θ 6= 0. Thus,
in the mean field approximation, the critical coupling, gc = 2. We see a hallmark of mean-field theory:
the ordered state is overestimated as we ignore fluctuations that tend to suppress order. (Recall from the
duality argument that order is lost at g = 1, assuming a single transition, as we have done in the present
instance.)
Close to the critical point, we anticipate that θ is small, and so we can Taylor expand: we up to quartic
order, we have (setting ϕ = 0, π):
" 2 #
θ3 θ2 θ4 g − 2 2 θ4

E(θ) ' −N J θ− + ... + g 1 − + + ... ' −2N J + θ + (709)
6 2 24 2 4
Note that in writing the above expression, we have set g = 2 everywhere except in the quadratic term, since
the corrections are small except in this case.
It is easy to see that the minima of (709) occur when
√
hσiz i = sin θ ' θ = ± gc − g (710)
and the energy near the critical point and in the ordered phase behaves as
3 2
EMFT (g) = 4 (gc − g) , g ≤ gc (711)
0, , g ≥ gc
Note that ∂g E is continuous at the transition, indicating that the transition is second-order. This is perhaps
a good place to pause and comment on the nature of phase transitions, as we will be spending a lot of time
with second-order transitions.
What is going on in a phase transition? Note that clearly E(g) has a singular point: we certainly cannot
perform a Taylor series expansion near g = gc , since the behavior is strictly zero on one side, and rising on
the other, and a moment’s thought will convince you that this is never going to be duplicated in any Taylor
11
The state we have put each site into is known as a spin coherent state; it has the neat property that the expectation of the
spin components along any axis are the same as the projection of a classical spin along the same axis.
107
E(✓, 0)
3NJ
g>2
g=2
g<2
-p p
✓
-3NJ
Figure 30: Mean-Field Energy Function. The energy functional goes smoothly from having a single
minimum (g > 2) to having a double minimum (g < 2), indicating broken symmetry. At the critical point
(gc = 2 in mean-field theory) a careful analysis reveals that a Taylor expansion of the free energy has a
vanishing quadratic term.
series. Such a non-analyticity in the energy (for classical transition, in the free energy) is the hallmark of
a dramatic change in the organization of the ground state (for a classical transition, in the nature of the
Boltzmann weight across the statistical ensemble.) How can such non-analyticities arise? There are two
common situations:
1. A first-order quantum phase transition is simply a level crossing: two eigenstates with totally
different character cross in energy as a parameter (call it g, in sympathy with our present example) is
varied, and the ground state simply follows the lower branch. If the states truly are totally different
in character (e.g., have different quantum numbers), then the states go right through each other as
there are no matrix elements between them to provide ‘level repulsion’. It is clear that in this case, the
first derivative of the energy is discontinuous. Near a first order phase transition, the two states don’t
need to ‘know anything about each other, nor do the excited states above the crossing levels have any
indication about the dramatic change in the ground state: the transition comes as a surprise to the
ground states, which suddenly find themselves no longer the optimal solution of the Hamiltonian, and
the excited states have no idea that anything happened!
2. A second-order quantum phase transition (often called a continuous phase transition, though
this term could apply just as well to higher-order transitions) is one in which the derivative of the
energy is continuous, as in the mean-field Ising problem discussed here. Loosely speaking, these are
transitions at which the gap above the ground state also vanishes as the system is tuned to its critical
point, i.e. as g → gc . Put differently, rather than having a level crossing as in the first-order case,
the non-analyticity in the energy as a function of the tuning parameter is because of the spectrum
‘collapsing’ so that we have an accumulation point of energy eigenvalues near gc . In this case, both
the excitations and the ground states on either side are ‘aware’ of the phase transition, and we might
expect to see significant experimental signatures of the transition. It also turns out that the vanishing
of the gap implies that correlation length for local observables diverges near the transition.
Returning to the Ising model, let’s pause and see where we stand. We have performed a mean field
analysis, and found a phase transition as a function of g, at gc = 2. We observe that the transition is
continuous, and furthermore, we find that the magnetization scales with exponent 21 near the transition
(recall hσiz i ∝ (gc − g)1/2 ). How much can we trust these results?
It turns out that the seriousness of the neglect of fluctuations in mean field theory (in other words,
assuming a product state that is uniform) is highly dependent on dimension. For sufficiently high dimensions
(d ≥ du , the upper critical dimension), mean field theory becomes exact, and we can trust its predictions
for exponents, etc. For sufficiently low dimensions (d ≤ dl , the lower critical dimension), the neglect of
108
fluctuations is extremely problematic: the fluctuations are so strong that they destroy the order in the
system. For the quantum Ising model12 , it turns out that du = 3 and dc = 0. As we will show, for d = 1
(our present case) mean field theory correctly predicts a transition, but gets gc wrong, and, more seriously13
gets the wrong exponent for the rise in the magnetization, which an exact solution reveals is really 18 .
17.5 Jordan-Wigner ‘Fermionization’ of the TFIM

We now introduce a very useful tool in the study of a one-dimensional systems: the so-called Jordan-Wigner
transformation, that allows us to map back and forth between spin systems and lattice models of fermions.
σjy ±iσjz
First, observe that we can define for each j raising and lowering operators σj± = 2 which satisfy
the commutation and anticommutation relations
1
[ σix , σj± ] = ±σj± δij and {σix , σj± } = 0 (712)
2
and also satisfy
1 + σjx 1 − σjx
σj+ σj− = and σj− σj+ = =⇒ [σ + , σ − ] = σjx and {σ + , σ − } = 1 (713)
2 2
Observe that σ ± are related to spin flips in the x-basis. Now, let us define the following operators
σlx and c†i = σi+
Y Y
ci = σi− σlx (714)
l>i l>i
These can be viewed as an operator that creates a spin flip in the paramagnet, attached to a ‘string’ of
operators attached that together create a domain wall in the ferromagnet (the string is non-local as it
should be, since domain walls are nonlocal excitations). So the new object created/annihilated by c†i , ci
is something like a bound state of a spin flip and a domain wall, so in a sense combines properties of the
excitations that are ‘natural’ in either phase. So it is perhaps not so surprising that they are linked to
thinking about the critical fluctuations
We will now show that these new operators are fermionic, i.e., they obey the canonical anticommutation
relations
{ci , cj } = 0 = {c†i , c†j } and {ci , c†j } = δij (715)
For the first case, let us assume without loss of generality that i < j. Then,
!    !
Y Y Y Y
− x  − x − x − x
{ci , cj } = σi σl σj σl0 + σj σl 0  σi σl
l>i l0 >j l0 >j l>i
Y Y
= −σi− σj− σlx + σj− σi− σlx =0 (716)
l>j l>j
where the different signs for the two terms follows from the fact that because i < j we had to commute a
σj− past a σjx only in the first term (which we did using the anticommutation relation (712)), and we have
used the fact that (σjx )2 = 1 to simplify terms where possible. A similar argument gives {c†i , c†j } = 0. For
the last commutator, let us first consider i < j. Using the same logic, we find
!    !
{ci , c†j } =
Y Y Y Y
σi− σlx σj+ σlx0  + σj+ σlx0  σi− σlx
l>i l0 >j l0 >j l>i
Y Y
= −σi− σj+ σlx + σj+ σi− σlx = 0, (717)
l>j l>j
12
Actually, we can understand this more simply from the quantum-classical mapping by studying Du , Dc for the classical
Ising model in dimension D = d + 1; we find Dc = 1, Du = 4, in accord with our discussion.
13
Thinking in terms of universality, we care more about exponents than precise values of the critical parameters; nevertheless,
overall scales are often worth worrying about if you want to access the properties within some limited parameter space, e.g. in
a solid state setting we’d be somewhat dismayed if we needed temperatures in the nanoKelvin!
109
where we have used the fact that spin operators commute when i 6= j. However the i = j case is special,
since in that case
! !   !
{ci , c†i } = σi−
Y Y Y Y
σlx σj+ σlx0 + σj+ σlx0  σi− σlx = σi− σj+ + σi+ σi− = 1, (718)
l>i l0 >i l0 >j l>i
since the string does not include σix and so we don’t need to worry about the associated −1.
The Jordan-Wigner transformation is invertible — you can try to see how to reverse this starting from
the fermion language (hint: you will need a fermionic version of the ‘string’). The lesson is that spins and
fermions can be mapped into each other by a nonlocal transformation in 1D.
Let us now use this to rewrite the TFIM as a model of fermions. Using (713) we see that the transverse
field term is
σjx = 1 − 2σj+ σj− = 1 − 2c†j cj . (719)
Now, consider
−(c†j+1 + cj+1 )(c†j − cj ) = −(σj+1 y

Y Y
− +
+ σj+1 ) σlx × (σj− − σj+ ) σlx = −(σj+1 )σjx (−iσjz ) = σj+1
z
σjz
l>j+1 l>j
Putting these together, we find:

X †
HTFIM = −J (cj+1 − cj+1 )(c†j + cj ) + g(1 − 2c†j cj )
j
X
= −N gJ + J 2gc†j cj − c†j+1 cj − c†j cj+1 − c†j+1 c†j − cj cj+1 . (720)
j
To simplify further, we introduce Fourier modes of the fields, ck = √1N j cj e−ikxj , and from the results
P
above, we may rewrite (you should check that you can do this!) the TFIM Hamiltonian as
Xh i
HTFIM = J 2 (g − cos ka) c†k ck − i sin ka c†−k c†k + c−k ck − g . (721)
k
This Hamiltonian does not conserve fermion number: the second (‘pairing’14 ) term violates it in units
of two. It turns out any physical fermion Hamiltonian must conserve fermion number modulo two: the
corresponding variable, fermion parity is always a good quantum number. Note that we may ignore the
constant term in the above expression for most purposes unless we wish to compute the ground state energy.
There are subtleties to do with boundary conditions that we have been a little bit sloppy about above.
Technically speaking, our construction is a bit ambiguous with periodic boundary conditions (since we
made a special choice when defining the ‘string’). This is closely related to the issues around implementing
dualities, and we will not dwell on it, though we have implicitly assumed that things go through correctly
with periodic bcs, as we need those to be able to go to Fourier space. Also, a more elegant solution introduces
so-called Majorana fermions. We will not discuss these either but they are a very active area of research,
as there are interesting properties associated to these fermions on open chains (when the fermionic chain is
realised directly in a condensed matter setting.)
17.6 Solution of the Ising Model via Bogoliubov Transformation

The fermionized version of HTFIM written above is quadratic in the ck s but is not quite diagonal, as the
pairing terms mix k and −k. It turns out this can be solved quite easily. We perform a fermionic Bogoliubov
transformation by introducing new operators
γk ≡ uk ck − ivk c†−k (722)

14
The reason for the terminology is that such terms typically arise from superconducting attraction between the electrons.
110
and demand that the transformation is canonical, i.e. preserves the fermionic anticommutators {γk , γk†0 } =
δk,k0 . This restricts uk = cos φk /2, vk = sin φk /2. (Note that the difference between preserving commutators
and anticommutators are why fermionic Bogoliubov transformations involve ordinary rather than hhyprbolic
trigonometric functions — and can be viewed as a linear transformation that mixes creation and annihilation
operators by a unitary rather than antiunitary matrix.) The remaining ambiguity is in the angle φk which we
fix by asking that the the Hamiltonian in terms of γk is diagonal, and contains no pairing terms. The resulting

condition is tan φk = 1 (k) where 1 (k) = 2J(g − cos ka), 2 (k) = −2J sin ka and HTFIM = k k γk† γk − 12
2 (k) P
p
with k = 21 + 22 .
At the end of the procedure, we see that the exact single-particle (here, ‘particle’ = ‘quanta created by
γk† ’) dispersion is p
k = 2J 1 + g 2 − 2g cos ka (723)
Note that for any g ≥ 0, the argument of the square root is positive, so we get a real energy; furthermore
the minimum of the energy is at k = 0, which gives us the exact gap at all g:
k ≥ 0 = 2J|1 − g| ≡ ∆(g) (724)
— amazingly, exactly what we got from first order perturbation theory15 . Note that the gap has a nonana-
lyticity at g = gc = 1; this gets transmitted to other quantities, such as the energy16 .
An important observation we can make is that at criticality, the system is scale invariant: at g = 1, the
fermions at k = 0 become gapless:
g=gc
k = c|k| (725)
where the speed of propagation is c = 2Ja. There is an emergent Lorentz symmetry (at least in the
spectrum) at the critical point; furthermore, the fact that the long-wavelength (k ∼ 0) modes have the
lowest energy allows a continuum description that forgets the details of the lattice. Near the critical point
(|g − gc | gc ) we gave v
g − gc 2
u
g∼gc u 2+
k = cu k (726)
u
t a
| {z }
m2 →0
We identify the second term with a ‘mass’, or alternatively, with a divergent length scale, ξ = m−1 =
a|g − gc |−1 . We may now introduce some (standard) terminology for critical exponents:
• the correlation length exponent, or ν in ξ ∼ |g − gc |−ν controls the rate of divergence of the
correlation length near the critical point. For the d = 1 TFIM, ν = 1.
• the dynamical critical exponent is the power-law in the dispersion at criticality, namely (k, gc ) ∝
k z . This controls ‘length-energy scaling’, as the transformation t → λx, t → λz t preserves the disper-
sion, and if we wish to perform RG near the critical point, this tells us the ‘correct’ RG transformations
to use. We expect the continuum critical theory will also be invariant under this scale transformation.
For the 1D Ising model, z = 1 as in a relativistic (Lorentz-invariant) theory, where space and time are
equivalent and there are operations — boosts — that rotate between them.
17.7 Continuum Limit and Critical Theory

So far, we have seen how to exactly rewrite the TFIM in terms of fermionic degrees of freedom, and obtained
a nice understanding of how the ferromagnetic and paramagnetic limits work. We already had nice variables
to describe those, though — in a sense, the fermions are overkill in those limits, when we can think in terms
15
Don’t expect such yeomanlike service from first order perturbation theory outside of this very special example.
16
It is clear that for any g we can compute the ground state energy of the fermions; by the quantum-to-classical mapping, we
can relate this to the free energy of the classical Ising model as a function of temperature, the famous Onsager solution!
111
of domain walls or spins. Where the fermionic description comes into its own, though, is in addressing the
critical properties.
In this section, we will show how to take a quantum lattice model (with continuous time but discrete
space) into a continuum field theory, in which both space and time are continuous variables. This is known
as the continuum scaling limit of the model, and is the most convenient way to access universal properties
of the transition using analytical tools. We have actually already has some experience with the subtleties of
continuum scaling limits hen we discussed the quantum-classical correspondence (e.g. for the 2D classical
Ising model), but for now it’s perhaps best to jump right into an example.
A continuum field theory emerges near the critical point where the correlation length is large. For ξ a,
define the continuum fermion field
1
Ψ(xj ) ≡ √ cj (727)
a
√
where the factor of a is designed to convert Kronecker deltas to Dirac deltas: we want
{cj , c†j 0 } = δjj 0 ↔ {Ψ(x), Ψ† (x0 )} = δ(x − x0 ) (728)
Now, we have
a X√ e−ikx
r
1 X X a Z
ck = √ cj e−ikxj = aΨ(xj )e −ikxj
= √ Ψ(xj )e −ikxj
= dx √ Ψ(x) (729)
N j L L a→0 L
j j|x =ja j
We can now use this in HTFIM , and work in the ka 1 limit (as these are the lightest modes and hence
dominate the continuum limit). We find
Z
(g − cos ka)c†k ck → (g − gc ) dx Ψ† (x)Ψ(x)
X
J (730)
k
where we have dropped higher order terms in an expansion17 in terms of a∂x . The second term gives us,
again to leading order
c
X Z
−iJ sin kac−k ck → dx Ψ† (x)∂x Ψ† (x). (731)
2
k
Putting it all together, we have at leading order in our gradient expansion the continuum field theory
ξa c
Z h i Z
HTFIM −→ dx Ψ (x)∂x Ψ (x) − Ψ(x)∂x Ψ(x) + ∆ dx Ψ† Ψ
† †
(732)
2
with c = 2Ja and ∆ = 2J|g − 1|. In order to proceed, we need to build some technology to construct
fermionic path integrals.
13: Aside: Grassmann Variables and Fermionic Coherent States
We start somewhat obliquely, by asking the question: what does it mean to have a coherent state for
a fermionic operator? By a coherent state we mean a state with the property that
c|ψi = ψ|ψi (733)
where c is a (complex) fermion operator and ψ is a c-number, i.e. is not an operator. However, ψ
cannot be your run-of-the-mill number, since we must have c2 = 0, it must follow that ψ 2 = 0. Such
a number is called a Grassmann number. They have many peculiar properties, but it turns out that
they are invaluable (and quite simple) to work with in theories that involve fermions. In particular,
they are absolutely essential in order to build a path-integral description.
17
A useful trick to remember is that k ↔ −i∂x .
112
The first rule to remember is that the Grassmann numbers anticommute with each other and with
fermionic operators
ψ1 ψ2 = −ψ2 ψ1 , ψc = −cψ (734)
while commuting with bosonic operators and ordinary numbers. With Grassmanns, it’s easy to define
a fermionic coherent state as
|ψi = |0i − ψ|1i (735)
where |0i and |1i = c† |0i are the empty and filled states. You may verify quite easily that this works.
Similarly, if we take hψ̄| = h0| + ψ̄h1|, it is easy to show that this is a left eigenvector of the creation
operator: hψ̄|c† = hψ̄|ψ̄. One should note that the fermionic coherent states are somewhat subtle: we
have enlarged the Hilbert space to allow Grassmann coefficients rather than ordinary numbers. Also,
ψ̄ is not the complex conjugate of ψ and hψ̄| is not the adjoint of |ψi: indeed, their overlap is
hψ̄|ψi = 1 + ψ̄ψ = eψ̄ψ . (736)
We will also need to do calculus with Grassmann variables; while this might seem scary, were we
allowed to use Grassmans instead of ordinary variables, you could summarize your entire A-level
calculus textbook in less than a page! The previous calculation of an overlap has already given a nice
example of how simple Taylor’s theorem is for Grassman variables:
f (ψ) = f0 + f1 ψ (737)
since higher powers all vanish! Integration is as easy as taking derivatives, since
Z Z
ψdψ = 1, 1dψ = 0, (738)
though with more than one Grassmann we have to worry about the order:
Z Z
1 = ψ̄ψdψdψ̄ = − ψ̄ψdψ̄dψ (739)
(You can remember this by noting that the positive sign is associated with the order of Grassmanns
in the integrand being exactly the reverse of the order of integration.) Note that the definition of the
integral is counterintuitive. More properly, we should view the integral as analogous to an indefinite
integral for ordinary numbers,
R with limitsR ±∞. For this analogy to be sensible, we require ‘shift
invariance’: namely, that f (ψ + η)dψ = f (ψ)dψ for a constant R (Grassmann) η. However, taking
f (ψ + η) = ψ + η and assuming η is a constant, this requires dψ = 0. Then, given that the most
general Grassmann
R function is f (ψ) = f0 + f1 ψ, the only way to have any nonzero integrals at all is
R that ψdψ 6= 0. This leaves freedom of normalization; we choose the simplest consistent
to assert
choice, ψdψ = 1.
Really the only integral we need to care about is the Gaussian integral,
Z
e−aψ̄ψ dψ̄dψ = a (740)
and its multivariate generalization

Z
eψ̄·A·ψ dψ̄dψ = det A (741)
where  
A11 A22 · · ·  
ψ1
 .. 
ψ̄ · A · ψ = ψ̄1 , . . . ψ̄M 
 A21 . ···   .  (742)
.. ..

.. ψM
. . .
113
and we solve by performing a change of variable that brings A to diagonal form. Note that the results
of Gaussian integration over Grassmann variables is exactly the inverse of that for ordinary numbers.
The crucial result we will need is the resolution of the identity in fermionic coherent states:
Z
1 = dψ̄dψ e−ψ̄ψ |ψihψ̄| (743)
and, finally, for any operator A that is bosonic (made up of an even number of fermionic operators
and Grassmanns) we have Z
trA = dψ̄dψ e−ψ̄ψ h−ψ̄|A|ψi (744)
The sign is important — we’ll come back to it in a bit.
17.7.1 Path Integral and Critical Theory for the TFIM

The reason for all this rigmarole is the following: we wish to understand the quantum statistical mechanics
of the many-body problem of the Ising chain in its fermionic language. The derivation of the fermionic path
integral closely parallels that for bosons, except that it uses Grassmannn variables.
We would like to compute the partition function (β = 1/T )
L
Z Y
Z(β) = tr e−βH = dψ̄0 (xj )dψ0 (xj ) e−ψ̄0 ψ0 h−{ψ̄0 (x)}|e−βH |{ψ0 (x)}i (745)
j=1
where we have introduced a composite resolution of the identity for all the sites in the system: |ψ0 (x)i =
⊗Lj=1 |ψ0 (xj )i, i.e. we have introduced a Grassman variable parametrizing a coherent state for each fermionic
site xj . How are we to compute this partition function? The key tool is to use the path integral approach
adapted to many-fermion problems, which we now develop.
As a warm-up, let us compute a simpler example, the partition function for a single fermionic mode,
with Hamiltonian H = c† c(ω0 − µ). A particularly useful trick is to Trotterize: we write
eβH = (1 − ∆τ H)(1 − ∆τ H) . . . (1 − ∆τ H) (746)

| {z }
M times
β
where ∆τ = M . You will recognise that since lim∆τ →0 (1 − ∆τ H)β/∆τ = e−βH , this is a reasonable
approximation as long as M → ∞.
The reason for Trotterizing is that we can now insert resolutions of the identity of the form (743) between
the different (1 − ∆τH ): we see that there are M − 1 such points of insertion, so that
−1
Z MY
Z(β) = dψ̄l dψl e−ψ̄l ψl hψ̄l+1 |(1 − ∆τ H)|ψl i. (747)
l=0
where we have defined

ψ̄M = −ψ¯0 , ψM = −ψ0 (748)
so that we can replace h−ψ̄0 | = hψ̄N |. We can now use the coherent state property to turn matrix elements
into Grassman-valued functions:
∆τ →0
hψ̄l+1 |(1 − ∆τ H(c† , c))|ψl i = hψ̄l+1 |(1 − ∆τ H(ψ̄l+1 , ψl ))|ψl i = eψ̄l+1 ψl e−∆τ H(ψ̄l+1 ,ψl ) (749)
114
Note that it was crucial that H was normal ordered, i.e. had all its c† s to the right of its cs.
−1
Z MY
Z(β) = dψ̄l dψl e−ψ̄l ψl eψ̄l+1 ψl e−∆τ H(ψ̄l+1 ,ψl )
l=0
  
−1
Z MY
  ψ̄l+1 − ψ̄l 
= dψ̄l dψl exp 
∆τ 
 ∆τ ψl − H(ψ̄l+1 , ψl )

l=0 | {z }
=∂τ ψ̄
Z Z β Z
' [Dψ̄Dψ] exp dτ ψ̄(τ )(−∂τ − ω0 + µ)ψ(τ ) = [Dψ̄Dψ]e−S[ψ̄,ψ] (750)
0
We should note that:
• In the penultimate step, we defined continuum (Grassmann) fields
ψ(τl = ∆τ l) ≡ ψl , ψ̄(τl = ∆τ l) ≡ ψ̄l (751)
• We also wrote H(ψ̄l+1 , ψl ) = H(ψ̄l , ψl ) + O(∆τ ). While this is usually a negligible distinction, it does
sometimes matter (we will see an example in the homework).
• The antiperiodic boundary conditions (cf. (748)) that require ψ(τ + β) = −ψ(τ ) mean that in the
Fourier representation18
1X 1X
ψ(τ ) = ψ(ωn )e−iωn τ , ψ̄(τ ) = ψ̄(ωn )eiωn τ , (752)
β n β n
the Matsubara frequencies ωn = (2n + 1)π/β, n ∈ Z are half-integer multiples of π/β.
• The measure [Dψ̄Dψ] is implicitly defined. We should view this the same way as we viewed the
continuum limit of the scalar field theory: with mathematical suspicion but physical confidence. We’ll
encounter several examples in the future.
• The derivative of a Grassmann function is also implicitly defined; however, there is no sense in which
one can think of Grassmann numbers as ‘big’ or ‘small’, so the numerator is heuristic at best.
• In the final step, we integrated by parts i.e. relabeled terms in the sum, as follows
X X X X X X
(ψ̄l+1 − ψ̄l )ψl = ψ̄l+1 ψl − ψ̄l ψl = ψ̄l0 ψl0 −1 − ψ̄l1 ψl = − ψ̄l (ψl+1 − ψl ) (753)
l l l l0 =l−1 l l
without moving Grassmanns past each other, and hence the sign makes sense.
The most important thing to take away is that the quantum partition function can be computed using a
path integral formulation where we use the Euclidean action19
Z

S[ψ̄, ψ] = dτ ψ̄∂τ ψ + H(ψ̄, ψ) . (754)
It’s straightforward to extend this to higher dimensional fermionic integrals: we would just insert a reso-
lution of identity at each spatial point at each time slice (as in the definition of |{ψ0 }i, above), and then
18
Though ψ̄ 6= ψ ∗ , the relative sign in the Fourier transformation is still convenient.
19
The reason this is ‘Euclidean’ is because τ plays the role of time, but it’s imaginary time — and as a result, we measure
spacetime distance using the Euclidean metric, ds2 = −dτ 2 − dx2 rather than the Minkowski metric ds2 = dt2 − dx2 .
115
it’s straightforward to show that we
Q must now integrate over all space-time paths (since we may write,
generalizing the previous example, j [Dψ̄(xj )Dψ(xj )] = [Dψ̄Dψ]) so that Z = [Dψ̄Dψ]e−S[ψ̄,ψ] , with
R
Z Z
d
dτ dd x L

S[ψ̄, ψ] = dτ d x Ψ̄(x)∂τ Ψ(x) + H(Ψ̄(x), Ψ(x)) ≡ (755)
where we have introduced the Hamiltonian density H, which when integrated over spatial coordinates, yields
the Hamiltonian H, and implicitly defined the (Euclidean) Lagrangean density L
In both cases, it’s important to note that the kinetic term is first-order in time; this is sometimes called
a ‘Berry phase term’. We note that this term is actually
M −1
X 1X
ψ̄l+1 (ψl+1 − ψl ) = ψ̄(ωn )[1 − eiωn ∆τ ]ψ(ωn ) (756)
β n
l=0
but in writing it as a derivative, we have only kept the leading nonzero term, as we substituted 1 − eiωn ∆τ →
−iωn ∆τ . This is cleary innocuous if ωn ∆τ 1 for all ωn that matter. It turns out this is indeed the case
for reasonable Hamiltonians and reasonable quantities (and like good physicists, we are cagey on the precise
definition of ‘reasonable’, by claiming ‘we’ll know it when we see it.’)
Armed with these results, we are ready to return to our study of the Ising chain. Recall that we had, in
the continuum limit, that
ξa c
Z h i Z
HTFIM −→ dx Ψ (x)∂x Ψ (x) − Ψ(x)∂x Ψ(x) + ∆ dx Ψ† Ψ
† †
(757)
2
R R
with c = 2Ja and ∆ = 2J|g − 1|. The Euclidean action is S[Ψ̄, Ψ] = dτ dx L, where
c
L = Ψ̄∂τ Ψ + H = Ψ̄∂τ Ψ + Ψ̄∂x Ψ̄ − Ψ∂x Ψ + ∆Ψ̄Ψ (758)
2
obtained by simply replacing field operators Ψ† (x), Ψ(x) with their Grassmann eigenvalues Ψ̄(x), Ψ(x). It’s
actually much nicer to change variables again, by defining new variables ψ± : we write
Ψ = ψ+ + ψ− + i(ψ+ − ψ− ), Ψ̄ = ψ+ + ψ− − i(ψ+ − ψ− ) (759)
in terms of which we have X

L= ψ± (i∂τ ± ∂x ) ψ± + ∆ψ+ ψ− . (760)
±
Note that the ψ± are counterpropagating modes; if we think of periodic boundary conditions in space, each
has a definite ‘handedness’. They are sometimes referred to as chiral modes. The mass term serves to couple
the counterpropagating modes — colloquially, ‘backscattering’ them into each other.
We observe that the partition function now looks like a classical statistical mechanics problem (except
that the classical fields are Grassman variables!) If you’ve taken (or plan to take) a course on RG 20 , you
know (will soon know) what to do now: we simply ask what operators are relevant, irrelevant or marginal
under the RG. Recall that at its heart, the RG is scaling momenta to smaller and smaller scales, and distances
to longer and longer scales. So we should study how the action behaves under scale transformations. In
order to keep S invariant, we had better (remembering that z = 1) scale space and time the same way, i.e.,
x → λx, t → λt. How should we scale the field? The action has two pieces, and they would give us different
answers. Which is correct? Here’s a useful way to think about this: we could always set the mass term to
zero — then we’d be at the critical point. This is the scale-invariant theory we’re looking for, and it is only
scale invariant if we specify the correct transformation for the Ψs. This leads us to the scaling rule
1
x → λx, t → λt, Ψ → λ− 2 Ψ (761)
20
In the MMathPhys, it’s taught in Hilary Term.
116
So, just by scaling, we see that21
1
hΨ† (x)Ψ(0)i ∼
. (762)
x
Indeed, as the theory is Gaussian — has a quadratic action — one could even calculate the precise coefficient.
More generally the scaling dimension δ of an operator O(x) in O?λ−δ O in a scale-invariant quantum
field theory determines its vacuum autocorrelation functions to be
1
hO† (x)O(0)i ∼ . (763)
x2δ
As we mentioned before, the mass perturbation has a different scaling than the first (derivative) piece. Given
the scaling in (761), we see that Z Z
dτ dx Ψ̄Ψ → λ dτ dx Ψ̄Ψ, (764)
or in other words, that as we makes length bigger and bigger, the mass term grows — this is what we call
a relevant perturbation, and it will actually dominate the behavior at long distances. It turns out that we
knew this: if we tip the Ising model on either side of its critical point, it is described by a theory that at
the longest length scales eventually looks ‘boring’, as all the action lies at very high energies (of the order
the gap). It is only at the critical point that we have gapless excitations, and therefore interesting behavior
at all length scales. The scaling dimension of the relevant perturbation determines the correlation length
exponent: again, this should make sense, since the correlation length must diverge as the critical point is
approached, and the proximity to the critical point is controlled by the relevant perturbation. To see this
more formally recall that by definition, the correlation length is the length scale ξ = aλ at which the relevant
coupling ∆ = λν ∆0 has evolved to order one (∆0 is the ‘bare’ value at lattice scale a, and ν its scaling
dimension.) From this, we find
1
1 = (ξ/a)ν ∆0 =⇒ ξ ∼ ν (765)
∆0
and since ν is the scaling dimension of the relevant operator ∆, we have confirmed that it is the correlation
length exponent.
Are there any other relevant operators? If there were, the Ising transition would actually be multicritical :
we’d have to tune more than one parameter to zero in order to stay at criticality. We demand that Ising
symmetry is preserved; in the fermion language, the Ising symmetry is captured by Ψ(x) → −Ψ(x). So
we should only consider operators containing an even number of fermions. All other operators that can be
made of an even number of fermions are irrelevant: they shrink as we go to larger distances. A candidate
marginal operator is Ψ̄ΨΨ̄Ψ, that vanishes due to Fermi statistics. Operators with higher derivatives have
no chance to survive: Z Z
−1
2
dτ dx Ψ̄∂x Ψ ∼ λ , dτ dx Ψ̄∂x ΨΨ̄∂x Ψ ∼ λ−2 (766)
So, we only need to tune one parameter to access the transition, so we have verified that the transition in
the TFIM is really critical. Note that this is more powerful than just cooking up a model where we only
tune one parameter; such models are always approximate and you might worry in principle that other terms
that you ignored may push you off the critical point. However, we have analyzed the continuum scaling limit
and concluded that the transition is a universality class with exactly one relevant perturbation, and once
we find it (g in our case) that’s it: we have exhausted the list. Note also that having no classically marginal
operators (‘classical’ here being shorthand for ‘just do naive dimensional analysis without worrying about
higher order terms in the RG’) we need not worry about higher terms coming in to change things — small
quantum corrections cannot beat the order-1 irrelevancy we have computed. Were there exactly marginal
operators at the classical level, this higher level analysis would be absolutely necessary.
21
In a z = 1 theory, it’s enough to specify the scaling of spacelike-separated correlators as we can boost to get the full x, t
dependence.
117
There are operators of dimension22 less than 2, namely the operator σ z itself. This operator creates a
branch cut in the fermion fields, since they are made from domain wall operators. It turns out that the
scaling dimension of this operator is 1/8, and this is important as it tells us the scaling of the magnetization
as g → gc from below (since σ z is the quantity that measures the magnetization.) However, it is not
an allowed perturbation to the Ising critical theory as it violates the Ising symmetry, and breaking this
symmetry destroys the transition!
Notice that the critical theory is completely different from the naive guess we could have made from
mean field analysis: we might have said that we should just ‘add fluctuations to mean field theory’ and thus
written a theory in terms of the (bosonic) mean field variable θ. At g = gc we’d have said that the theory
should be Gaussian with a fixed-point action
Z
S0 [θ] = dτ dx ((∂τ θ)2 + (∂x θ)2 ) (767)
that preserves the θ → −θ symmetry. However, in this case, many relevant operators preserve the symmetry:
θ2 , θ4 , ... and indeed all even functions are potentially relevant. This would be a grave error in d = 1, though
for d > 1 we have the correct variables (though not the right fixed point), and is perfectly OK for d ≥ 3:
we have identified both the variables and the fixed point correctly!
14: Experiments
We close our discussion of the Ising model by giving a brief discussion of experimental studies of the
various quantum-critical properties that we have studied. In order to do this, we need to introduce
some technology.
17.7.2 Correlation Functions, Structure Factors, and Long-Range Order

A useful set of physical observables to study are the correlations functions of the spins:
C(r, r0 ) ≡ h0|σrz σrz0 |0i. (768)
Notice that this is an equal-time correlator. For a translation-invariant ground state, this is only a
function of the separation of the spins, C(r, r0 ) = C(r − r0 ) (we are using the convention that r = ja
to go from lattice units to actual distances. The translation invariance allows us to take a Fourier
transform of the correlation function, called the static structure factor :
X
S(q) ≡ e−iqr C(r) (769)
r
where the sum is over the discrete positions r ∈ a{1, 2, . . . , N }; assuming periodic boundary con-
ditions, we have q ∈ N2πa {1, 2, . . . , N }. How do the correlation functions behave across the phase
diagram?
• At g = ∞, deep into the paramagnetic phase, we have

h0|σjz σjz0 |0ig→∞ = δjj 0

(770)
which is the zero correlation length limit of a stable phase. Different sites are completely
uncorrelated, and at finite but large g 1, the correlations decay exponentially,
|xj −xj 0 |a
h0|σjz σjz0 |0ig1 e−|xj −xj 0 |/ξ .

' (771)
In other words, fluctuations of spin flips communicate between sites, but only over a short range
as the excitations are massive.
22
The dimension usually counts the operator without the two dimensions in the integrand, so a dimension of 2 marks the
dividing line between ‘blow up’ (dimension less than 2) and ‘shrink’ (dimension greater than 2), i.e. marginal operators have
dimension 2. Note the quirk of terminology that we count the energy dimension rather than the time dimension!
118
• In the ferromagnetic phase, for g 1, we have
|xj −xj 0 |a
h0|σjz σjz0 |0ig1 N02 (g)

' (772)
where N02 (g) = 1 for g = 0, and vanishes as g → gc . More simply, if we train the system (in the
sense if having an infinitesimal field that tells the order parameter which way to point, we have
h0|σjz |0i = ±N0 (g). Here N0 is the order parameter, the magnetization. Note that the ‘training
procedure’ entails a sensitive order of limits: in any finite system N0 must vanish, so we must
have
lim lim hσiz i = 0, but lim lim hσiz i = N0 . (773)
N →∞ h→0 h→0 N →∞
It is in this sense that we say that the ferromagnet has long-range order.
One thing should bother you about the ferromagnet: it looks like it has a divergent correlation length,
and we earlier claimed that this is a property only of critical theories. This is indeed true, but it turns
out the correct correlation function is the so called connected correlation function, that measures the
extent to which two spins are correlated relative to the fact that their two-point function would have
an expectation value just by virtue of each having a nonzero mean: for the Ising case, this is the
quantity
Gzz z z z z
jl = hσj σl i − hσj ihσl i (774)
which you will recognize as the second cumulant of the distribution (also known as the variance). You
may argue quite easily that the connected correlator indeed has a vanishing length scale as g → 0,
and a divergent one as g → gc , showing the expected behavior. In other words, the correct thing
to measure are the fluctuations about the mean — it is these that diverge at criticality. The most
important thing to take away, though, is that the long-range order gives a sharp distinction between
phases: no smooth function of g can interpolate between the exponential decay and long-range ordered
behavior, so something singular must occur at g = gc .
How does the structure factor reflect this physics? It turns out that when there is long-range order
(and for simplicity, assuming N → ∞) so we can use continuous Fourier transforms, the structure
factor has a delta function singularity at the origin, whose weight is propertional to the square of the
magnetization:
S(q) = 2πδ(q)N02 + regular (775)
where ‘regular’ means terms smooth as q → 0. This reflects the broken symmetry.
2
(using σqz = σ z† ) then we have a sum rule
P
If we writea S(q) = h0|σqz σ−q
z |0i = z
n hn|σq |0i

−q
π
dq
Z
S(q) = C(r − r0 = 0) = h0|σrz σrz |0i = 1. (777)
−π 2π
which basically remembers that we have a normalization condition in somewhere. A consequence of a

sum rule is that spectral weight can only rearrange — the total weight is conserved, and so the delta
function at the origin has to come from somewhere.
More information can be obtained from the dynamical structure factor, that measures correlations
both in space and in time. We begin with the correlation function
C(r, t) ≡ h0|σrz (t)σ0z (0)|0i (778)
which reduces to the equal-time correlator when t = 0, and is a temporal autocorrelator when r = 0.
It is useful to think of σrz (t) in the Heisenberg picture: σrz (t) = eiHt σrz e−iHt . The dynamic structure
119
factor is thenb
X Z ∞
−iqr
S(q, ω) = e dt eiωt C(r, t)
r −∞
X Z∞
= e−iqr dt eiωt h0|σrz (t)σ0z (0)|0i
r −∞
X Z∞
= e−iqr dt eiωt h0|eiHt σrz e−iHt σ0z |0i
r −∞
X Z∞ X
= e−iqr dt eiωt h0|eiHt σrz e−iHt |nihn|σ0z |0i
r −∞ n
X Z ∞
= e−iqr dt ei(ω+E0 −En )t h0|σrz |nihn|σ0z |0i
r −∞
hn|σqz |0i2 δ(ω + E0 − En )

X
= 2π (779)
n
where in the last step we recognized the result from the calculation of S(q) previously. S(q, ω) is the
number of states that differ by momentum q and energy ω, weighted by (the square of) their overlap
with the state created by acting on the ground state with σjz — this piece is often called the ‘matrix
element’, for obvious reasons. It turns out (and I won’t go into detail here but you should remember
a version of it from courses on condensed matter and scattering theory) that an application of Fermi’s
golden rule tells us that this is the transition rate in an (inelastic) scattering experiment; the obvious
tool is scattering of neutrons as they will have a nonzero matrix elements for exciting the spins, as a
neutron couples to spin via its magnetic dipole moment.
17.7.3 Ising Symmetry in CoNb2 O6
Information about the excitations may be obtained by studying the inleastic neutron data: inelastic
scattering of a neutron can create a spin flip, as we have argued above spin flips behave very differently
in the two phases (the corresponding evolution of S(q, ω) is shown in Fig. 32.)
From our knowledge of the spectrum, we can infer the behavior of S(q, ω) in different limits of the
phase diagram. At large g, where the low-energy spectrum is mostly built from spin flips As these
are created by σ z directly, we should see a contribution to S(q, ω) from spin flips — but only if there
is the correct energy momentum relation ω = q where q is the spin flip dispersion. So we should
see a sharp peak in S(q, ω) along this curve for ω in the right range; for large ω, we merge into the
‘multiparticle’ continuum where we have many spin flips with the energy distributed haphazardly
among them. It turns out this continuum begins near 3∆, as the matrix element vanishes for an
even number of spin flips. At large g, if we create a spin flip by scattering neutrons, it is usntable
— the elementary excitations here are domain walls, and so a spin flip decays into a pair of domain
walls which then slosh around. (Recall that the domain wall operator is non-local and is made up
of a string of σ x s, so can’t be accessed by σjz ). So all we see is a continuous mess at small g – a
multiparticle continuum of domain walls with no sharp peaks.
This behavior has been observed in a system well-described by the TFIM, the material CoNb2 O6
(cobalt niobate). This material is a quasi-one-dimensional compound: it can be approximated as a
collection of decoupled chains of spins with nearest- neighbor ferromagnetic Ising interactions. The
transverse field is produced by, well, applying an external transverse field, allowing g to be varied.
All the data and figures in this section have been reproduced from R. Coldea et. al., Science 327,
177 (2010).
120
First, we see that the system displays a phase transition as a function of transverse field. This
is accomplished by tracking how the magnetic Bragg peak intensity (measured via elastic neutron
scattering) evolves with a transverse field (Fig. 31).
Next, the system is analyzed via inelastic neutron scattering, which measures S(q, ω). (Fig. 32). As
we have argued, this should show distinct behavior in the paramagnetic and ferromagnetic regimes.
Thus, we experimentally confirm that spin excitations change character from pairs of domain walls
in the ordered phase to spin-flips in the paramagnetic phase.
So far, this is in accord with our picture of the Ising model. If we turn off the field, it turns out
that there is a weak effect owing to the fact that model is not ‘truly’ 1D: the coupling between the
zigzag Co2+ chains generates a weak, longitudinal field (i.e., a field parallel to the direction of spin
alignment.) One can further examine the model by examining the behavior in zero external field as
a function of temperature. Note that a longitudinal field actually acts as a linear confining potential
for domain walls. Intuitively, a field pointing along the +z direction prefers domains of +z. A pair
of domain walls in the + phase are separated by a − region, with costs an energy growing linearly
with its length (Fig. 33, left top.).
This leads to the formation of domain wall bound states; this is a simple quantum-mechanical model
that can be solved quite easily, leading to the theoretical predictions in Fig. 33 for temperatures
above (C) and below (D) the experimentally determined transition temperature into the ordered
phase Tc ∼ 2.95 K (recall that the Ising model has a thermally tuned transition only for d ≥ 2; by
including the weak interchain couplings, this is allowed as the model is effectively three-dimensional!).
The experimental data Fig. 33(A,B) are in good agreement with the theory, including the theoretical
simulation of the neutron result Fig. 33(F). Reading off the ‘spectrum’ of the bound state energies
(called ‘masses’ in accord with the field-theoretic terminology) we see excellent match to experiment
Fig. 33(F, and left bottom).
Finally, one can study the fine structure of the inelastic scattering data near the critical point. Owing
to the small longitudinal field, the critical point is no longer gapless (Fig. 34, left top), and instead
has a massive spectrum. It turns out that there is an emergent E8 symmetry that allows one to
solve for the ‘particle’ spectrum (the set of discrete excitations). It turns out there are 8 ‘particles’
(i.e., distinct collective excitation modes, see Fig. 34, left bottom). While in principle these could
all give sharp peaks in S(q, ω), it turns out that the mass of the third particle is just a bit bigger
than twice the mass of the second, m3 & 2m1 ; so particles of mass M ≥ m3 are unstable to decay
into the 2m1 continuum. As a result, we can only see m1 , m2 as sharp neutron peaks. The theory
predicts, somewhat
√
amazingly, that at the critical point the ration m2 /m1 is precisely the golden
mean, ϕ = 1+2 5 . As shown in Fig. 34(A-D), the experiment is in remarkable agreement with theory.
In fact it turns out that to match fine details of the structure, one needs to do a bit more work, and
really the model has some slightly more subtle hidden symmetry that makes the story very slightly
different from the simplified Ising picture we have given above [M. Fava et al, Proc. Nat. Acad. Sci.
USA 117, 25219 (2020) ].
a
The first equality can be a bit tricky but here’s the easiest way to see it: recall that σqz = √1N r e−iqr σrz . Then,
P
we have
1 X −iq(r1 −r2 ) 1 X −iq(r1 −r2 ) X −iqr
h0|σqz σ−q
z
|0i = e h0|σrz1 σrz2 |0i = e C(r1 − r2 ) = e C(r) = S(q) (776)
N r ,r N r ,r r
1 2 1 2
where we used translational

R invariance and relabeled the summation.
b
Observe that S(q) = dω 2π
S(q, ω), not S(q, 0).
121
(a) (b) (c)
Figure 31: Ising Symmetry in CoNb2 O6 . (a) Lattice structure of CoNb2 O6 .; a single zig-zag chain of
Co2+ is well-described by HTFIM (b) Elastic neutron scattering data allows us to track the magnetization
via Bragg peaks, showing classic TFIM behavior (c) Inelastic neutron scattering allows us to excite single
spin flips and study their dynamics.
Figure 32: Inelastic Structure Factor S(q, ω) across the Ising transition as measured by neutron
scattering. Note the sharp structure seen in a 6 T field, reflecting the fact that the spin flip is a sharp
excitation in the paramagnetic phase; in the ferromagnet, S(qω) is incoherent, as spin flips decay into pairs
of domain walls.
122
Figure 33: Domain-wall bound states in the absence of external transverse field. Sketch of the
bound state formation (‘kink confinement’) and theoretical spectrum of bound states is shown alongside
experimental data.
p
m2 1+ 5
=
m1 2
Figure 34: Emergent E8 Symmetry. Particle spectrum computed from theory and corresponding ex-
perimental signatures of the emergent symmetry in the Ising model with both longitudinal and transverse
fields.
123

M.Phys. C6 Theory Option/M.Math - Phys. Advanced Quantum Theory Lecture Notes For 2020-2021

Uploaded by

M.Phys. C6 Theory Option/M.Math - Phys. Advanced Quantum Theory Lecture Notes For 2020-2021

Uploaded by

M.Phys. C6 Theory Option/M.Math.Phys.

Advanced Quantum Theory

November 19, 2020

I Functional Methods in Quantum Mechanics 4

1 Some Mathematical Background 4

2 Path Integrals in Quantum Mechanics 9

3 Path Integrals in Quantum Statistical Mechanics 16

4 Relation of D dimensional quantum systems to D + 1 dimensional classical ones 35

5 The Ising Model 38

III Many-Particle Quantum Mechanics 50

7 Application I: The Ideal Fermi Gas 61

8 Application II: Weakly Interacting Bosons 68

10 Path Integral for interacting Bose systems 78

IV Landau Theory of Phase Transitions 81

12 The Ising Model 81

13 Critical Behaviour and Universality 86

15 Other Examples of Phase Transitions 96

16 Ising Model and Φ4 Theory 98

17 Transverse-Field Ising Model 100

1.2 Functional differentiation

δF [f (x)] F [f (x) + δ(x − y)] − F [f (x)]

• We first interchange the functional derivative and the integration

• Next we use the chain rule

• Finally we interchange the functional and the ordinary derivative

Next we consider n-dimensional Gaussian integrals

W0 (A) = (2π)n/2 (detA)−1/2 ,

this integral can be written as

1.4 Homework Questions 1&2

Question 2. a) Evaluate the Gaussian integral

c) Show that Γ(1) = 1 and Γ(s + 1) = sΓ(s). (Hence Γ(n + 1) = n!)

2.1 The Propagator

ψ(~x, t) = h~x|ψ(t)i. (37)

Eqn (36) can be integrated to give

A central object in Feynman’s approach is the propagator

h~x0 |U (t; t0 )|~xi,

where  = t/N . Then we have by construction

where we have inserted N − 1 resolutions of the identity in terms of position eigenstates

Figure 1: Propagator as sum over paths.

Importantly we have [T̂ , V̂ ] 6= 0 and concomitantly

eα(T̂ +V̂ ) 6= eαT̂ eαV̂ . (48)

However, using that  is infinitesimal, we have

Note that in this expression there are no operators left.

2.1.1 Propagator as a “Functional Integral”

path. This provides a new way of thinking about QM!

2.3 Classical Limit and Stationary Phase Approximation

h02 (t) = 0. (59)

The condition (62) precisely defines the classical trajectories x(t0 )!

2.4 The Propagator for Free Particles

Going back to the explicit expression (53) we have

which leads to an expression

f (z0 ) = f (zN ) = 0. (72)

The Laplace operator D acts on these functions as

f (z) → f = (f (z1 ), . . . , f (zN −1 )). (75)

gives rise to a recurrence relation for the components an,j of an

an,j+1 + an,j−1 − (2 + λn )an,j = 0. (80)

The boundary conditions an,N = an,0 = 0 suggest the ansatz

2.5 Homework Questions 3-5

a) Change variables from xj to yj = xj − xN to bring it to the form

Give expressions for J and A.

Question 4. Denote the propagator by

Question 5. Diffraction through a slit. A free particle starting at x = 0 when t = 0 is determined to

Hint: make a substitution z = αy + β with suitably chosen α and β.

Assuming that the ground state of H is non-degenerate we have

lim hOiβ = h0|O|0i , (100)

x(τn ) = xn , τn = n. (107)

Going over to a continuum description we arrive at an imaginary-time functional integral

where SE is called Euclidean action

where we integrate over all periodic paths

x(~β) = x(0). (111)

3.1 Harmonic Oscillator at T > 0: a first encounter with Generating Functionals

dx hx|e−βH x̂n |xi dx xn hx|e−βH |xi

D̂τ G(τ − τ 0 ) = δ(τ − τ 0 ) , G(0) = G(β~). (119)

Under this change of variables we have

Putting everything together we arrive at

Now we are ready to calculate (118). The average position is zero

where the Matsubara frequencies ωn are

δF [f (x)] F [f (x) + δ(x − y)] − F [f (x)]

where = t/N . Then we have by construction

However, using that is infinitesimal, we have

x(τn ) = xn , τn = n. (107)

Statement I: We can link F (via Z) to a classical