February 2, 2011

Euclidean Field Theory

Kasper Peeters & Marija Zamaklar

Lecture notes for the M.Sc. in Elementary Particle Theory at Durham University.
Copyright
c 2009-2011 Kasper Peeters & Marija Zamaklar

Department of Mathematical Sciences

University of Durham
South Road
Durham DH1 3LE
United Kingdom

http://maths.dur.ac.uk/users/kasper.peeters/eft.html
[email protected]
[email protected]
1 Introducing Euclidean Field Theory 5
1.1 Executive summary and recommended literature . . . . . . . . . . . . 5
1.2 Path integrals and partition sums . . . . . . . . . . . . . . . . . . . . . 7
1.3 Finite temperature quantum field theory . . . . . . . . . . . . . . . . . 8
1.4 Phase transitions and critical phenomena . . . . . . . . . . . . . . . . 10

2 Discrete models 15
2.1 Ising models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.1 One-dimensional Ising model . . . . . . . . . . . . . . . . . . . 15
2.1.2 Two-dimensional Ising model . . . . . . . . . . . . . . . . . . . 18
2.1.3 Kramers-Wannier duality and dual lattices . . . . . . . . . . . 19
2.2 Kosterlitz-Thouless model . . . . . . . . . . . . . . . . . . . . . . . . . 21

3 Effective descriptions 25
3.1 Characterising phase transitions . . . . . . . . . . . . . . . . . . . . . 25
3.2 Mean field theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Landau-Ginzburg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4 Universality and renormalisation 31

4.1 Kadanoff block scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Renormalisation group flow . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.1 Scaling near the critical point . . . . . . . . . . . . . . . . . . . 33
4.2.2 Critical surface and universality . . . . . . . . . . . . . . . . . 34
4.3 Momentum space flow and quantum field theory . . . . . . . . . . . 35

5 Lattice gauge theory 37

5.1 Lattice action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Wilson loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3 Strong coupling expansion and confinement . . . . . . . . . . . . . . 39
5.4 Monopole condensation . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3
4
 Introducing Euclidean Field Theory
1
1.1. Executive summary and recommended literature
This course is all about the close relation between two subjects which at first sight
have very little to do with each other: quantum field theory and the theory of criti-
cal statistical phenomena. On the one hand, the methods and insight from quantum
field theory have helped tremendously to understand the concept of universality
in statistical mechanics. This is the fact that the effective theory relevant for long
distance behaviour of statistical systems depends only on a small number of param-
eters. On the other hand, the theory of critical phenomena has shed new light on the
role of renormalisation in particle physics. Meant here is the way in which subtle
tuning of parameters of the model is required to remove, or make sense of, spurious
infinities in loop computations. It is the goal of these lectures to explain how these
two fields are related.
Given that this course is about the common ground of two quite different disci-
plines, it is no surprise that there is a vast literature to help you with many technical
and conceptual issues. The list below is highly incomplete, but you are strongly en-
couraged to consult some of these books to get a wider view on the topics discussed
here.

M. Le Bellac, Quantum and statistical field theory, Oxford University Press,

1991.
Good for a number of explicit computations, though not terribly transparent
in its connection between the two fields, at least on first reading.

J. L. Cardy, Scaling and renormalization in statistical physics, Cambridge

University Press, 1996.
Standard book on the topic, gives a good overview, but is somewhat thin on
explicit computations.

J. M. Yeomans, Statistical mechanics of phase transitions, Oxford University

Press, 1992.
Very good, compact and explicit book on the statistical aspects of phase tran-
sitions.

J. B. Kogut and M. A. Stephanov, The phases of quantum chromodynamics:

From confinement to extreme environments, 2004.
Mainly focussed on applications to QCD, but contains some good introductory
chapters as well.

5
1.1 Executive summary and recommended literature

A. M. Polyakov, Gauge fields and strings, Harwood, 1987. Contemporary

concepts in physics, volume 3.
An original take on many of the topics discussed here, though sometimes a bit
too compact for a first introduction.
M. H. Thoma, New developments and applications of thermal field theory,
hep-ph/0010164.
Very clear text on quantum field theory at finite temperature, in particular
the connection between the imaginary time and real time formalisms. Also
contains a number of explicit applications to QCD.
G. P. Lepage, What is Renormalization?, hep-ph/0506330.
Lectures on the meaning of renormalisation from the point of view of a particle
physics theorist.
M. Creutz, Quarks, gluons and lattices, Cambridge University Press, 1983.
Another very accessible intro to lattice field theory.

I. Montvay and G. Munster,

Quantum fields on a lattice, Cambridge Uni-
versity Press, 1994.
More advanced book on lattice field theory. Good examples on the strong-
coupling expansion.

J. Smit, Introduction to quantum fields on a lattice: A robust mate, Cam-

bridge Lect. Notes Phys., 2002.
Nice and compact book about lattice field theory. Used heavily in the prepa-
ration of these lectures.

J. Zinn-Justin, Quantum field theory and critical phenomena, Int. Ser. Monogr.
Phys. 113 (2002) 11054.
A thousand page reference book, not for the faint of heart, and not always or-
ganised in the most pedagogical way. Nevertheless, a very good source if you
want to look up details.
J. Zinn-Justin, Quantum field theory at finite temperature: An introduction,
hep-ph/0005272.
A much smaller set of lectures exposing the relation between finite tempera-
ture quantum field theory and statistical mechanics. The first few sections are
worth reading as an introduction to the topic.

6
 1.2 Path integrals and partition sums

1.2. Path integrals and partition sums

The root of the connection between quantum field theory and statistical systems is
the functional integral (or path integral). This functional integral is closely related
to an object in classical statistical mechanics: the partition sum. In order to illustrate
this, let us briefly recall the quantum mechanics of a non-relativistic particle in the
path integral language.
Let us start with the transition amplitude for such a particle to go from q to q0 in
time . It is given by
Z q(t=)=q0 hi Z m  i
A(q0 , ; q, 0) = D q(t) exp q i (t)q i (t) V (q(t)) dt . (1.1)
q(t=0)=q h 0 2

As you have seen in the course on quantum field theory, this path integral expres-
sion can be related to the usual expression for a transition amplitude in the operator
formalism. For a suitable normalisation of the path integral measure, we have
h i i p 2
A(q0 , ; q, 0) = hq0 | exp H |qi , with H = + V (q) . (1.2)
h 2m
Here |qi is a time-independent state in the Heisenberg picture, and H the Hamilto-
nian corresponding to the particle action used in (1.1). For the details of this relation,
see your favourite quantum mechanics book.
Consider now what happens to the path integral expression if we make a substi-
tution t i. This is called a Wick rotation (it rotates the integration in the complex
time plane by 90 , as in the figure). Provided the Lagrangian does not depend ex-
plicitly on t (which, in our case, it does not), this changes the path integral weight
into a nice real expression,
Z q( =)=q0 h 1 Z m  i
A(q0 , ; q, 0) = D q( ) exp q i ( )q i ( ) + V (q( )) d ,
q(t=0)=q h 0 2
(1.3)
where the dot now indicates a derivative with respect to . Similarly, if we substitute
i in the operator expression, we get
The Wick rotation in the complex time
1 i h plane.
0 0
A(q , ; q, 0) = hq | exp H |qi . (1.4)
h
This expression should make some bells ring. If we would set |qi = |q0 i, and sum
over all position eigenstates, then (1.4) looks exactly like the expression for the par-
tition function Z [ ] of a statistical system at inverse temperature := 1/(kT ) given
by = /h, The quantum mechanical
transition amplitude for a system
h i
Z [ ] = hn| exp H |ni , (1.5) with periodic boundary conditions
n is, after summing over all
where we sum over a complete basis of states. This gives us our first relation be- boundary conditions, equal to a
quantum statistical partition
tween a quantum mechanical transition amplitude with periodic boundary condi- function, with /h.
tions and a quantum statistical system at non-zero temperature.
Let us now take a somewhat different look at the exponent in (1.3). We can write
it as
1 m i
Z 
Vstring = q (, )q i (, ) + V (q( )) d , (1.6)
h 0 2
i.e. as the potential energy of a field qi (, ) in two dimensions, where we now view
as the time (the total energy would have an additional kinetic term involving
derivatives). The path integral then simply sums over time-independent configu-
rations of this field. We thus see that we can also interpret (1.3) as computing the

7
 1.3 Finite temperature quantum field theory

partition function of a classical string of length , at an inverse temperature = 1/h.

This gives us our second relation, between a quantum mechanical transition function
The transition amplitude for a for a particle and a classical statistical system for a string.
quantum particle for a time i is The second of these relations is easily generalised to quantum field theory. In a
equal to the classical partition
function of a string of length ,
Lorentzian field theory, the generating functional in the presence of a source takes
with 1/h. the form
 
i
Z Z Z
d +1 e d +1 2
Z [ J ] = D exp S[] + d x J ( x )( x ) d x ( x ) . (1.7)
h h
(end of lecture 1) where S[] is the action corresponding to the field . We have included a damping
factor dependent on e, which ensures that field configurations which do not fall off
at space-like and time-like infinity are suppressed in the path integral. This ensures
that all correlation functions which we compute from Z [ J ] are correlation functions
in the vacuum state, e.g.
  1 
h0| T ( x ) (y) |0i = Z [ J ] . (1.8)

Z [ J ] J ( x ) J (y) J =0

We can now do a Wick rotation, and then interpret the exponent in (1.7) as the po-
tential of a field in one extra dimension. The path integral then computes for us
the classical partition sum of a model at inverse temperature = 1/h.
The first relation is a bit more tricky to generalise to field theory, because of the
explicit dependence of the statistical model on the time interval. In field theory we
usually take the time interval to . Moreover, we never consider arbitrary field
configurations at the endpoint of this time interval, but rather force the field to sit
in the vacuum state. So summing over boundary conditions does not produce an
sensible object in quantum field theory. However, what we can do is turn the first
relation around, and use it to describe a quantum field theory at finite temperature
in terms of a Euclidean path integral. This is the topic of the next section.

I Summary: The generating functional of a Euclidean quantum field theory in d

space dimensions can alternatively be interpreted as a partition sum in a classical
statistical model in d dimensions. The corresponding inverse temperature is then
related to Plancks constant by 1/h.

ISee also: Discussions like these can be found in many places, e.g. chapter 2 of the
book by Smit [10] or chapter V.2 of the book by Zee [13].

1.3. Finite temperature quantum field theory

In standard quantum field theory courses one almost always works with systems at
vanishing temperature. That is to say, the state which is used to define all correlation
functions is taken to be the vacuum state |0i of the free theory, with no particles. The
propagator, for instance, is defined as
 
GF ( x y) = h0| T ( x ) (y) |0i . (1.9)

However, we can certainly also construct a quantum theory of fields in which we use
a thermal state instead of the no-particle vacuum. Instead of computing a vacuum
expectation value, we then have to compute
 
hn| T (x) (y) |ni eEn
GFT >0 ( x y) = n . (1.10)
eEn
n

8
 1.3 Finite temperature quantum field theory

Here |ni is an eigenstate of the Hamiltonian with eigenvalue En . Such states are
generically complicated multi-particle states. As particles can be created with any
three-momentum, and in any number, they take the form
n
(~k ) i

M
a i
{n1 , . . . , n M ; ~k1 , . . . ,~k M } = p
 
|0i . (1.11)
i =1 ni !

This state has i ni particles, and the generic label n that appears in (1.10) is seen to
describe both the set of momenta that appear, as well as the number of particles of
each momentum.
We will leave this expression for the time being, and instead ask the question
whether we can write (1.10) back in a path integral language. It is not immediately
clear how to do this, because the Gaussian factor in the path integral (1.7) enforces
the appearances of the |0i state in the correlator. The way to continue is to observe
that (1.10) is an expression similar to (1.4). We can thus take it from there, and
go backwards, i.e. rewrite the thermal sum over energy eigenstates in terms of a
path integral, as we did for the non-relativistic particle. We view the right-hand
side of (1.10) as a propagator for a field theory over a compact Euclidean time with
period , with periodic boundary conditions in this direction.
In fact, this additional Euclidean direction is simply the imaginary part of the
time coordinate which is already present in x and y . That is to say, the Green
function (1.10) is periodic under a shift of the time difference by ih :
1 h H  0 i
GFT >0 ( x0 , ~x; 0, ~y) =
Tr e T ( x , ~x ) (0, ~y)
Z
1 h
i
= Tr e H ( x0 , ~x ) (0, ~y) (assuming x0 > 0)
Z
1 h
i
= Tr (0, ~y)e H ( x0 , ~x ) (1.12)
Z
1 h
i
= Tr e H e H (0, ~y)e H ( x0 , ~x )
Z
1 h
 i
= Tr e H T ( x0 , ~x ) (ih , ~y) = GFT >0 ( x0 , ~x; ih , ~y) .
Z
The last step, in which we re-introduce time-ordering, requires that we can compare
x0 with ih . That is possible if x0 = i, i.e. if the time coordinate was imaginary
to start with. The derivation thus goes through only for correlation functions which
are independent of the real time. For time-dependent situations, we need what is
called the real-time formalism for thermal quantum field theory, which goes beyond
the Euclidean Field Theory concepts which we focus on in these lectures.

I Summary: The static properties of a quantum field theory at finite temperature can
be obtained by doing a Wick rotation and imposing periodic boundary conditions
in imaginary time.1 The periodic boundary conditions of the path integral corre-
spond to the trace of the thermal sum, and the period corresponds to the inverse
temperature .

I See also: The lectures by Thoma [6].

1 You may have heard about Hawking radiation of black holes, and the fact that it has a certain

temperature. The simplest way to see that black holes have a temperature is by looking at the analytic
continuation of a black hole metric in the complex plane. It turns out that the imaginary part of time
needs to be compact, and hence that any field theory on top of the black hole background is in a thermal
state.

9
 1.4 Phase transitions and critical phenomena

1.4. Phase transitions and critical phenomena

Although we have quickly stepped over the relation between the path integral and
the quantum mechanical transition amplitude, a crucial ingredient in the derivation
of that relation is a proper definition of the path integral measure. The infinite-
dimensional integral is only well-defined if we regularise the model somehow, for
instance by putting it on a lattice (which, in the quantum-mechanical model, means
discretising time). As a result, all the systems that we are dealing with in these
lectures are in a sense discrete systems, with a countable number N of degrees of
freedom. The main question we are then interested in is how such systems behave
when we change their parameters, when we change the lattice spacing, and when
we take the thermodynamic limit N .
From statistical systems, we know that many interesting things can happen. To
illustrate some of these, let us focus on an example which will return several more
times during these lectures: the Ising model. This model has its roots in the search
for a simple model which describes a ferromagnet. The main property of a ferro-
magnet is that it retains its magnetisation in the absence of an external magnetic
(end of lecture 2) field. However, for sufficiently large temperature, this magnetisation disappears.
The key ingredient is the interaction between the electron spins at different sites
of the atomic lattice. A simple model of this type is the Heisenberg model, with Hamil-
tonian operator
H Heisenberg = J ~i ~j , J > 0 . (1.13)
hi,ji

Here the ~i are the Pauli matrices acting on the spin degrees of freedom at site i. The
notation hi, ji indicates a pair of spins at site i and j. Unfortunately, this model can
only be solved on a one-dimensional lattice (Bethe 1931). In that case, for J > 0 the
ground state is one in which all spins are aligned (while for J < 0 the ground state is
the so-called Neel state, with alternating spins). Even in the classical limit, in which
we replace the spin operator ~i at each site by a classical three-vector ~Si , the model
is not solvable except in one dimension.
The Ising model incorporates one last simplification, namely the reduction of
the three-vector at each site to a simple scalar Si , which can take the values 1. Its
classical Hamiltonian, in the presence of an external magnetic field, reads

HIsing = J Si S j h Si . (1.14)
hi,ji i

This model was introduced by Lenz in the 1920s, and solved by Ising in the one-
dimensional case. In the absence of a magnetic field h, the model was solved by
Onsager in two dimensions. No other exact solutions are known.
The parameters of this model are thus the coupling J, the external field h and
the number of sites N. All distances between spins are determined by the lattice
spacing a (which in this simple model does not actually appear in the Hamiltonian).
The partition sum, finally, adds the temperature T to the set of parameters. What we
want to learn is how we can do measurements on this system, how we can express
those measurements in terms of J, h, N, a and T, and how these behave when we
vary the parameters and take the N limit. This same story holds for all the
other discrete models that appear in these lectures.
What are the observables which we can measure in a statistical model of the
type (1.14)? In general, these are correlation functions, just like in a quantum field
theory. The simplest correlation function is the expectation value of a single spin,
which is called the local magnetisation. The expectation value of the sum of all spins

10
 1.4 Phase transitions and critical phenomena

is the total magnetisation,2

" #
Si HIsing [{Si }]


e
{ Si } i
M := . (1.16)
eHIsing [{Si }]
{ Si }

Naively, one would perhaps be tempted to say that M = 0 when the background
magnetic field h vanishes: for every configuration {Si } there exists another config-
uration in which all the spins are flipped, and these two configurations contribute
with opposite sign to the partition sum.
While this is mathematically true, we have to ask the question what happens The 2d Ising model at high temper-
physically to a configuration of aligned spins, i.e. with net magnetisation. When the ature, in the absence of a magnetic
field. The spins are randomly fluctuat-
temperature is very high, the thermal fluctuations will indeed wash out such mag- ing, there is no magnetisation, and the
netisation, as in the figure at the top. However, for sufficiently low temperatures, model is thus paramagnetic (dark and
light denote spin up and down, respec-
it becomes very hard for the thermal fluctuations to sample the entire configura- tively).
tion space, and we do see large connected regions with a net magnetisation. In this
regime, the spin up/down symmetry of the Hamiltonian is spontaneously broken by
the state of the system.
In the continuum limit N , this transition becomes sharp. Mathematically,
the absence of a phase transition for a finite system is expressed by the fact that the
two limits N and h 0 do not commute,
lim lim M(h) = 0 ,
N h 0
(1.17)
lim lim M(h) 6= 0 .
h 0 N

From a practical point of view these expressions are, however, of limited value.
We will more generically see that discrete models exhibit remnants of phase tran- The 2d Ising model at low temperature.
sitions, which are in some sense smoothed out versions of the continuum transi- Large connected regions have formed,
in which the spin is predominantly in
tion. The more degrees of freedom we include (the larger N) the closer we get to the one direction and the magnetisation is
proper phase transition of the continuum model. non-zero. The model is ferromagnetic.

Moving on to higher order correlation functions, the next one is the correlator of
spins at two different sites,
h i
Si Sj eHIsing [{Si }]
{ Si }
h Si S j i : = . (1.18)
eHIsing [{Si }]
{ Si }

This correlator will be a non-trivial function of the distance between the two sites.
If there are n sites between i and j, so that the distance is r = n a, the correlator
typically behaves at large distances as
 2  
M 1 r
hSi S j i hSi ihS j i = hSi S j i D+ 2 exp . (1.19)
N r
The length is called the correlation length, D is the dimension of the system, and is
called the anomalous dimension (to which we will return in much greater detail later).

2 The total magnetisation can be obtained from the partition sum in the presence of a magnetic field

by simply taking a derivative,

1 ln Z
M= . (1.15)
h
For more complicated correlators, a similar trick requires that we know the partition function coupled to
a non-homogeneous field.

11
1.4 Phase transitions and critical phenomena

As is clear from the figures, the correlation length becomes large when we approach
the phase transition, and we will see that it in fact diverges in a very particular way
as we approach the critical temperature.
We will often use a thermodynamical language to describe the large-distance
behaviour of statistical systems. Without trying to include a full course on ther-
modynamics, let us briefly recall some essential concepts. The basic quantity we
need is the entropy. At the microscopic level, it is defined in terms of the number of
microstates given the macroscopic parameters (like magnetisation),

S := k B ln . (1.20)

It can also be expressed in terms of the probability of a single state p({Si }) to occur,
using Shannons formula,
h i
S = k B p({Si }) ln p({Si }) . (1.21)
{ Si }

This allows us to relate the entropy to the energy and eventually to the partition
function, as we have
1 h i h i
p({Si }) = exp H({Si }) , Z = exp H({Si }) . (1.22)
Z {S } i

All of thermodynamics follows from this microscopic starting point, together with
the laws of thermodynamics.
The second law of thermodynamics states that a system out of equilibrium will
eventually evolve to an equilibrium in which the entropy is at least as large as before.
If we have two connected systems at energy E A and EB respectively, equilibrium will
be reached when the total entropy reaches an extremum, i.e. when

dStot 
= 0. (1.23)
dE A eq.

This leads to the fundamental definition of temperature in terms of entropy,

  1
dS
T := . (1.24)
dE

From here, all the other usual thermodynamical potentials follow by making use
of the first law of thermodynamics, which essentially expresses that energy is con-
served. The change in internal energy of a system is related to the heat absorbtion
and the performed work as
dU = Q W . (1.25)
In terms of state functions, this gives

dU = TdS PdV + hdM . (1.26)

This expression holds true for a closed system. If the external parameters S and
V are fixed, the system in thermodynamical equilibrium will be determined by a
minimum of the internal energy U.
However, we often deal with open systems, which are coupled to a heat bath
which can provide an arbitrary amount of energy to keep the temperature T con-
stant. In this case, the energy flow means that S is no longer constant. The second
law now imposes that a minimum is found for

F = U TS hM . (1.27)

12
 1.4 Phase transitions and critical phenomena

This quantity F is the Helmholtz free energy, and we can relate it to the partition
function by making use of (1.21),

1 h i h i
S = k B Z
exp H H ln Z ) = k B (U hM ) + ln Z , (1.28)
{ Si }

so that we find (end of lecture 3)

F = k B T ln Z . (1.29)
There are also thermodynamic potentials which are relevant when the pressure is
kept fixed (and thus the volume is allowed to change), but these are not very rel-
evant for systems like the Ising model. For completeness, they are all listed in the
table below.

Internal energy U (S, V, M) dU = TdS PdV + hdM

Helmholtz free energy F ( T, V, h) = U TS hM dF = SdT PdV Mdh
Gibbs free energy G ( T, P) = F + PV
Enthalpy H (S, P) = U + PV
In the last column of this table, you can see how the potentials change when the
parameters are modified.

I See also: A more extensive explanation of entropy and free energy can be found
in [14].

13
1.4 Phase transitions and critical phenomena

14
 Discrete models
2
As announced, our main interest will be in systems which have a finite number of
degrees of freedom, either because they are defined on a lattice or because they are
obtained by cutting off high-momentum modes of a continuum theory. In chap-
ter 4 we will be interested in analysing how such systems behave when we send the
lattice spacing to zero, or remove the cutoff. For the time being, however, we are
interested in analysing the discrete models by themselves.

2.1. Ising models

2.1.1 One-dimensional Ising model
We have already seen the Ising model Hamiltonian,

HIsing = J Si S j h Si . (2.1)
hi,ji i

The partition function Z [ , J, h] for the one-dimensional model is then given by

N 1 h 1
i
Z= exp JSi Si+1 + h(Si + Si+1 ) .
2
(2.2)
{Si =1} i =0

It can be computed by a method known as the transfer matrix method. For this
purpose, we make the spin chain periodic, i.e. the spin to the right of S N 1 is S0
again. The partition can then be expressed as the trace of a product of 2 2 matrices,

Z= TS0 S1 TS1 S2 TS N 1 S0 ,
{Si =1}
h 1
i
where TSi S j = exp JSi S j + h(Si + S j ) . (2.3)
2
The matrix T is the transfer matrix. This procedure is in spirit very similar to the
way in which we derive the path integral from the quantum mechanical transition
amplitude: we cut up the amplitude in a number of steps, glue those steps together,
and sum over the states at the intermediate points. Here that sum over intermediate
states is given by the sum over spins.
In more explicit form, we can write the matrix elements of T as

e ( J h) e J
!
T= . (2.4)
e J e ( J +h)

15
 2.1 Ising models

Since the trace of a matrix is equal to the sum of its eigenvalues, we can express the
partition sum as h i
Z = Tr T N = +
N N
+ , (2.5)

where are the eigenvalues of T, with + > . It is not hard to compute these
eigenvalues explicitly; from (2.4) we get
 q 
= e J cosh( h) sinh2 ( h) + e4J . (2.6)

When we take the continuum limit N , the largest of these two eigenvalues will
tend to dominate the physics.
We can immediately compute a number of observables, following the general
logic outlined in section 1.4. The simplest is the free energy F. In the N limit,
this one becomes
 q 
2 4J
F = k B TN ln + = JN k B TN ln cosh( h) + sinh ( h) + e . (2.7)

The free energy in the continuum It is an extensive quantity (scaling with the size of the system), but the strict linear
limit is determined by the largest scaling with N holds true only in the large-N limit. When h = 0 this reduces to
eigenvalue of the transfer matrix.
F = k B TN ln [2 cosh( J )] . (2.8)

This is an analytic function in T except at T = 0. Hence there is no phase transition

in the one-dimensional Ising model.
The next quantity is the magnetisation. We can compute this in two different
ways. Given that we already have the free energy, we can compute it simply by
using  
F sinh( h)
M= = Nq . (2.9)
h T
sinh2 ( h) + e4J
We can alternatively compute it by doing a direct computation of the expectation
value of the sum of spins (this method will prove useful later when we compute
the correlation length). Since the system is translationally invariant, we can simply
compute the expectation value of a single spin and multiply that with N, to get

N
M = N h Sk i =
Z TS0 S1 TS1 S2 TSk1 Sk Sk TSk Sk+1 TSN1 S0
{ Si }
h i h i
Tr T k ST N k Tr S T N 
1 0

=N =N h i , where S := . (2.10)
0 1
h i
Tr T N Tr T N

Translation invariance is reflected in the fact that this result is independent of k, and
related to the cyclic invariance of the trace. In order to compute the trace on the
right-hand side of (2.10), we note that T is a symmetric matrix and hence can be
diagonalised with an orthogonal matrix P,
 
1 0
T = PP , where = . (2.11)
0 +

If we find P, we can then compute the magnetisation using

h i
Tr P1 SP N
M=N h i . (2.12)
Tr N

16
 2.1 Ising models

By writing out the right-hand side of (2.11) explicitly and comparing with the com- M
1.0
ponents of T on the left-hand side, one finds that
0.5
 
cos sin sinh( h)
P= , with cos(2) = q . (2.13) h
sin cos -3 -2 -1 1 2 3

sinh2 ( h) + e4J
-0.5

We then get -1.0

! The magnetisation of the one-

1
cos(2) sin(2) dimensional Ising model as a function
P SP = , of the external magnetic field. The
sin(2) cos(2) flatter curves correspond to higher
temperatures.
N N ) cos(2 )
(+ N
M=N N + N
N cos(2) , (2.14)
+

which together with (2.13) reproduces the result obtained in (2.9). Note that for
h = 0 the magnetisation is always zero, indicating that there is indeed no phase
transition in the one-dimensional Ising model. A plot of the magnetisation as a
function of the external field, for various values of the temperature, is given in the
figure. (end of lecture 4)
The correlation length can now be computed using a very similar technique.

1
h S0 S r i =
Z S0 TS0 S1 TS1 S2 TSr1 Sr Sr TSr Sr+1 TSN1 S0
{ Si }
h i h i
Tr ST r ST N r Tr S ( Pr P1 )S( P N r P1 )
= h i = h i , (2.15)
Tr T N Tr T N

Here is again the matrix of eigenvalues as given in (2.11). In the large-N limit
the N matrix can be approximated by keeping only the largest eigenvalue + , but
because r is finite, the r matrix depends on both eigenvalues. A bit of matrix
multiplication yields

r
 
2
hS0 Sr i = cos (2) + sin2 (2) . (2.16)
+

3.5

Using our previous result for the magnetisation, the variance becomes 3.0

2.5

r
  2.0

hS0 Sr i hS0 ihSr i = sin2 (2) . (2.17) 1.5

+ 1.0

0.5

The only r-dependence sits in the ratio of eigenvalues. From the definition of the 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
h

correlation length (1.19) we then find

The correlation length of the one-
dimensional Ising model as a function
1

 , = 1. (2.18) of the external magnetic field. Again,
the flatter curves correspond to higher
ln /+ 
temperatures.

At zero external magnetic field the correlation length reduces to = 1/ ln tanh( J ).

The behaviour as a function of the magnetic field is plotted for various temperatures
in the figure.
We state here without proof that for many other models, the correlation length
is determined by the same type of expression (2.18), with + and the two largest The correlation length is
eigenvalues of the transfer matrix. Note that if = 0 the correlation length van- determined in terms of the two
largest eigenvalues of the transfer
ishes. This happens when J = 0, i.e. when there is no interaction between the spins. matrix.

17
 2.1 Ising models

A phase transition happens when , which requires that the two largest
eigenvalues become degenerate. For the one-dimensional Ising model this only hap-
pens when h = 0 and (T 0), but this does not connect two phases.

I See also: The one-dimensional Ising model is discussed in many places, see
e.g. [15].

2.1.2 Two-dimensional Ising model

The two-dimensional Ising model can also be solved by the transfer matrix tech-
nique, as was shown by Onsager in 1944. However, this requires a much more sub-
stantial analysis. Instead, we will in these lectures use a different method to extract
The unique configuration which
changes 4 bounds (multiplicity N).
properties of the two-dimensional model. This goes under the name of Kramers-
Wannier duality, and is a relation between the low-temperature perturbative expan-
sion and the high-temperature perturbative expansion. The advantage of doing this
is that such expansions and duality relations have analogues in all of the other sys-
tems which we will discuss, and hence these techniques are of more general use (in
contrast to the transfer matrix method which, although it can sometimes give exact
results, is in general too complicated to be of any use). We will in this section only
consider the case of a vanishing external magnetic field, i.e. h = 0, and for the time
being just look at the low- and high-temperature expansions in isolation.
For the low-temperature expansion, we start with the ground state at T = 0, which
has all spins in one direction (here taken to be up). We have to sum over all config-
urations in the partition sum. For low temperature, the more spins we flip the more
bonds will change their energy, and hence we can look at things order by order in the
number of flipped spins. If we flip one spin, 4 bonds are changed. Each such bond
yields an energy increase of 2J. If we flip two spins, we can do it in two different
ways, depending on how close the spins are. For the first possibility (top figure) we
can place this at any of the N lattice sites, but there are two orientations, so there are
2N possibilities, each changing 6 bonds. For the second possibility (bottom figure)
we have N ( N 5)/2 possible ways of placing it, and we change 8 bonds. There are
three other configurations which change 8 bonds, obtained by flipping more spins.
For the partition sum, we obtain the low-temperature expansion,
 
1
Z = e2NJ 1 + Ne8J + 2Ne12J + N ( N + 9)e16J + . (2.19)
2
Here e2NJ is energy of the all-up state (explained by the fact that there are 2 links
for every site, e.g. the one pointing up and to the right). Since T is small, J is large
and hence e J is a small number.
Let us now turn to the high-temperature expansion. When T is large, J is small.
A crucial role in finding an expansion in small J is played by the following sim-
ple identities, valid for a variable S which takes values 1 (and easily verified by
inserting S = 1 explicitly),

0 if n odd
e JS = cosh( J ) + S sinh( J ) , Sn = (2.20)
S=1 2 if n even .

Using these identities we can rewrite the partition sum in the following form,
 
Z = cosh( J ) + Si S j sinh( J )
{Si =1} hi,ji
h i2N  
Configurations which change 6 bonds
= cosh( J ) 1 + Si S j tanh( J ) . (2.21)
(top figure, multiplicity 2N) or 8 bonds {Si =1} hi,ji
(other figures, multiplicity 12 N ( N 5),
2N, 4N and N respectively).
18
 2.1 Ising models

At this stage we have not done any approximation yet, we have just rewritten the
partition sum in a somewhat different form. However, note that tanh( J ) is small
at high temperature. We can thus use (2.21) to attempt an expansion in powers of
tanh( J ). (end of lecture 5)
If we expand (2.21) in powers of tanh( J ), the only terms which contribute are
those in which all spins which occur do so in even powers (by virtue of the identities
in (2.20)). It is convenient to think about such terms in a graphical way, since that
will help us to count the number of terms at each order. To do this, first observe that
in the expression (2.21), a given product of two spins Sn Sm appears only once. If we
denote this product by a line from site n to site m, we need to look for those graphs in
which the contours are closed, i.e. do not end in a single spin factor. Multiple closed
contours are allowed though. Every line segment contributes a power of tanh( J ),
and the total contour length thus sets the order in the perturbative expansion. For
the partition sum we find
 2N 
Z= cosh J 2 N 1 + N (tanh J )4 + 2N (tanh J )6

1
+ N ( N + 9)(tanh J )8 + . (2.22)
2

The main idea behind the high-temperature expansion we have just seen is that
their are only nearest-neighbour couplings in the Hamiltonian. These couplings are
the lines in our graphs, and what the high-temperature expansion does is to group
lattice variables into linked clusters. For this reason, it is also known as the cluster
expansion or hopping expansion. We will see that it has its use also in much more
complicated discrete models, like lattice gauge theory.

2.1.3 Kramers-Wannier duality and dual lattices

Both the low-temperature and high-temperature expansions of the two-dimensional
Ising model partition sum are untransparent and complicated. However, we see that
these two perturbative series get transformed into each other if we simultaneously
replace
e2J tanh( J ) ,

Zlow Zhigh (2.23)

N .
e 2NJ 2 (cosh J )2N
This is a remarkable property. It says that if we understand the low-temperature,
magnetised phase of the model, we will also understand the high-temperature, dis-
ordered phase. As we will see shortly, this symmetry is not just a property of the
first few terms in the perturbative expansions. In fact, it is possible to start with
Various closed contours with non-
a partition sum Z [ ], and rewrite it such that it looks like the partition sum of the overlapping edge segments. In the
Ising model at a different temperature, Z [(1/2) ln tanh( J )]. More precisely, third figure, there are N ways to put
the first square and N 5 to put the
h i N second (since the 5 squares denoted by
)
Z [ ] = sinh(2 J Z [ ] , (2.24) red dotted lines are disallowed). The
multiplicities are N, 2N, 12 N ( N 5),
2N, 4N and N respectively.
where and are related by the relation we found from the perturbative analysis,

= 1 ln tanh( J ) .
J (2.25)
2
The two-dimensional Ising model is called self-dual, in the sense that we can rewrite The two-dimensional Ising model
its partition function as that of another two-dimensional Ising model, but at a dif- is self-dual.
ferent temperature.

19
 2.1 Ising models

If we assume that the system has only one phase transition, then this had better
(end of lecture 6) happen when the inverse temperatures and are the same. This requires

2J
sinh(2 c J ) = 1 Tc = . (2.26)
k B ln(1 + 2)

Even though we have not yet shown that there is in fact a phase transition, we
already know that it has to happen at this particular critical temperature. No exact
solution of the model was required.
In this particular case, the result (2.26) for the critical temperature can be veri-
fied by making use of Onsagers solution of the two-dimensional Ising model using
transfer matrix techniques. Needless to say, it agrees with the explicit computation.
Let us now look closer at how the exact duality (2.24) can be proven. What we
want to show is how the Ising model partition sum can be rewritten in terms of
another partition sum, but at a different temperature. The starting point is the full
expression for the partition sum at vanishing magnetic field. Setting J = 1 (as only
the combination J appears), this is
h i
Z = exp Si S j . (2.27)
{ Si } hi,ji

The first step is to rewrite this expression in a way similar to what we did in the first
step of the high-temperature treatment, now written in a somewhat more compact
form,

Z= Ckij ( )(Si S j )kij ,

{Si } hi,ji k ij =0,1

where C0 ( ) = cosh , C1 ( ) = sinh . (2.28)

The new variables k ij are associated to every nearest-neighbour pair, i.e. to every
link on the lattice (they are link variables). As we already noted earlier, the only con-
tributions to this sum are those which have an even power of each spin variable Si .
This means that the only non-zero contributions are those in which the distribution
of values of the link variables is such that the sum of all links connected to each and
every spin Si is an even number. If it is, the sum over Si will yield twice the same
contribution one, so that we can simplify Z as

Z= Ckl ( ) 2 (sum of values of links incident at i modulo 2) . (2.29)

{k l } l i

In other words, we have rewritten the partition sum as a sum over the link variables,
but these link variables have to satisfy a particular constraint.
The trick to solve this constraint is to realise that there is yet another set of vari-
ables which we can introduce, such that the link constraint is always satisfied. These
are the dual variables Sei , which sit on the nodes of the dual lattice, as indicated in the
The original Ising lattice (open dots) to-
gether with the dual lattice (filled dots). figure. The link variables are related according to
Also indicated are the four links inci-
dent on S1 , one of which is k12 . The 1
dual variables Se1 and Se2 are used to de- k12 = (1 Se1 Se2 ) , (2.30)
fine k12 . 2
and so on. For the sum of the four link variables incident at S1 , this gives

1 1
2 (Se1 Se2 + Se2 Se3 + Se3 Se4 + Se4 Se1 ) = 2 (Se1 + Se3 )(Se2 + Se4 ) . (2.31)
2 2
This expression is manifestly even for any choice of dual variables.

20
 2.2 Kosterlitz-Thouless model

So we can now write the partition function as

1 N
Z=
2
2 C(1Sei Sej )/2 ( ) . (2.32)
{Sei } hi,ji

The extra factor of 1/2 arises because there are two configurations of dual variables
which give rise to the same value of the link variable, and thus all configurations of
link variables are counted twice. The only thing that is left is to express the Ck ( ) in
terms of the dual variables. This gives


Ck ( ) = cosh 1 + k(tanh 1)
 
= cosh exp k ln tanh (2.33)

= (cosh sinh )1/2 exp 12 Sei Sej ln tanh .

 

If we insert this into the partition sum we arrive at

1
(2 cosh sinh ) N exp 21 ln tanh Sei Sej
 
Z=
2 hi,ji
{Sei }

1
(sinh 2 ) N exp Sei Sej
 
= . (2.34)
2 hi,ji
{Sei }

This is again an Ising model, now in terms of the dual variables, and with a cou-
pling which is related to the original one by the relation we found earlier in the
perturbative analysis, namely = (1/2) ln tanh .

I See also: A good text on duality in statistical systems and field theory is Savits
report [16]. The Ising model is discussed in e.g. [3].

2.2. Kosterlitz-Thouless model

The next model we will discuss is the Kosterlitz-Thouless model. It is again a lattice
model, but instead of the spin variables Si of the Ising model, which can only take
on the values 1, it has a two-dimensional unit-vector ~S at each site. That is to say,
there is one continuous angle variable at each site, with

~Si = cos i , sin i .

(2.35)

It is thus the same as the classical Heisenberg model in two dimensions. The Hamil-
tonian is again given by a nearest-neighbour one,

~Si ~Sj = J cos

HKT = J i j (2.36)
hi,ji hi,ji

The crucial ingredients which set this model apart from the Ising model is that we
are dealing with a continuous variable, and that this continous variable is periodic,
i.e. takes values in a compact space. These properties will come back again when
we discuss lattice gauge theories for QCD, so the Kosterlitz-Thouless model is a nice
warm-up example.
One might perhaps have expected that, just like for the Ising model, there should
be a high-temperature phase in which the spins are disordered and there is no mag-
netisation, and a low-temperature phase in which the spins align and there is a net

21
 2.2 Kosterlitz-Thouless model

magnetisation (the expectation value of the sum of spins does not vanish). However,
that intuition turns out to be wrong. What we will see is that the Kosterlitz-Thouless
model never has a net magnetisation, but it does exhibit two different phases, char-
(end of lecture 7) acterised by two different behaviours of the correlation function h~Si ~S j i.
At high temperature, we can make an expansion which is very similar to the one
in the Ising model. Write the correlation function as

1 1
Z  
Z
hei0 ein i = dm ei0 in exp J e i j + e j i .
i i i i
(2.37)
m 2 hi,ji

Here we have expanded the cosine in terms of exponentials. For large tempera-
tures, we can expand in powers of and obtain integrals over exponentials. Using
identities which take the role of those in (2.20),
Z 2 Z 2
dm = 2 , dm eim = 0 , (2.38)
0 0
we then see (again by similar logic as in the Ising model at high temperature) that
the only terms which will contribute to the correlator are those in which we have
exponentials corresponding to a contour which starts at site 0 and ends at site n.
Any other contour will have a left-over eik for some site, and integrate to zero.
Assuming that the sites 0 and n are on an axis, the lowest order contribution comes
from a contour that goes straight between these sites, and is of order ( J )n . The
correlator thus behaves as
h
i
hei0 ein i exp |n| ln 1/( J ) . (2.39)


The Kosterlitz-Thouless model at The correlation length is thus 1/ ln 1/( J ) at high temperature.
high temperature has a finite In the low-temperature phase we expect the phases n to vary slowly as a func-
correlation length, which goes to
zero as T .
tion of the position on the lattice. We can then expand the cosine in the Hamiltonian,
and obtain for the correlator

1 1
Z

Z
hei0 ein i = dm ei0 in exp J J (i j )2 . (2.40)
m 2 hi,ji

In preparation for the jump to continuum field theories, the argument of the Boltz-
mann weight is often interpreted as a discretised derivative, i.e.
1 1
J (i j )2 = J ( i )2 , (2.41)
2 hi,ji 2 i,

where labels the direction away from site i, and (in a somewhat sloppy notation),
i : = i + i . (2.42)
The integral in (2.40) is purely Gaussian, so can be performed exactly. Using the
continuum language (2.42), we find1
  1
 1  2J
hei0 ein i   . (2.44)
n
The Kosterlitz-Thouless model at Since there is no exponential suppression, the correlation length is infinite. How-
low temperature has an infinite
correlation length, but no 1 This is similar to the manipulation with path integrals, which states that
spontaneous symmetry breaking.
Z 1   Z 
1
Z Z
D exp dx ( ( x ))2 + J ( x )( x ) exp dx dy J ( x )( x y) J (y) , (2.43)
2 2
with ( x y) the inverse of the kinetic operator. In the case here, J ( x ) = ( x ) (n x ), and the
propagator in two dimensions behaves as (n) 2 1
ln |n|.

22
 2.2 Kosterlitz-Thouless model

ever, the correlation function does fall off to zero for large distance. Because we
have hei0 ein i hei0 ihein i for large distances, this immediately also implies that
there is no magnetisation, i.e. no spontaneous symmetry breaking. This is in agree-
ment with the Coleman-Mermin-Wagner theorem which says that a continuous global
symmetry in two dimensions cannot be spontaneously broken in a system with local
interactions.
The question now remains: what is the crucial qualitative difference between
the low-temperature and the high-temperature regime? The key ingredient is the
periodicity of the i variables. If we consider the continuum formulation of the
model (using (2.41)), the solutions to the equations of motion are solutions to the
two-dimensional Laplacian, with an additional periodicity condition,
 
2x + 2y = 0 , 0 < < 2 . (2.45)

The solutions to this equation can be labelled by the number of times the phase
goes around when we circle the origin. These vortices are configurations which
A vortex of the Kosterlitz-Thouless
have a lower free energy for sufficiently high temperatures. They condense above model, arrows denoting the unit-
this Tc , and lead to a finite correlation length. vectors ~Si .
It is not so easy to show that these vortices indeed dominate the partition sum
above Tc . However, it can be made plausible by estimating the free energy of a
vortex, and showing that condensation of a vortex lowers the free energy above a
certain temperature. The free energy is

F = E TS , (2.46)

where S is the entropy. The energy can be determined from

1
Z Z
E= d2 x ( )2 = ddr rg = J ln( L/a) , (2.47)
2
where we approximated = , and have written the answer in terms of the total
size of the lattice L and the lattice spacing a. Now the entropy is simply given by k B
times the logarithm of the number of ways in which we can put down the vortex,
which is the number of lattice sites ( L/a)2 . So we find


F = J 2k B T ln( L/a) . (2.48)

Above Tc = J/(2k B ) the system prefers to condense a vortex. This is one of the
classic examples of a system in which entropy wins over energy. Working out
the precise details of the condensation mechanism goes beyond the scope of these
lectures. (end of lecture 8)

ISee also: Section 2.5 of [4] for a deeper analysis of vortex condensation in the
Kosterlitz-Thouless model.

23
2.2 Kosterlitz-Thouless model

24
 Effective descriptions
3
In the previous chapter we have seen a number of explicit examples of systems with
a phase transition. We have, however, not discussed much of the details of phase
transitions themselves, or the techniques to compute properties of these systems
close to the phase transition. Exact computations are often hard (as witnessed by
e.g. the two-dimensional Onsager solution of the Ising model). In the present chap-
ter we will introduce two ideas for approximate methods to capture the essentials
of systems with phase transitions. Before we do that, we will however need a little
bit more terminology.

3.1. Characterising phase transitions

A phase transition occurs when the external parameters of a system are such that
one or more obervables are singular (i.e. discontinuous or divergent). These points
in the phase diagram are called critical points. The critical exponents show how var-
ious quantities vanish or diverge near the critical points. An important property
of phase transitions is that various systems which naively look very different may
still, when they approach the critical point, have identical behavior. More precisely,
many different systems turn out to have the same critical exponents. Examples of
such systems are the Ising model on a square and triangular lattice. While the value
of the critical temperature depends on the details of the interaction, these two mod-
els do turn out to have the same critical exponents.
Phase transitions are commonly classified into roughly three different groups.
For first order transitions,

1. The singular behaviour is manifest in the form of discontinuities. The phase diagram of the two-
dimensional Ising model. For T < Tc
2. The correlation length stays finite. the transition between h < 0 and
h > 0, along the blue curve, is
first-order (the magnetisation jumps
3. The latent heat can be non-zero, i.e. going from one phase to the other can discontinuously). The line of first-
release or absorb energy. order transitions in the h T plane
(red dashed line) ends in the critical
4. At the transition two or more phases may coexist in equilibrium (for instance, point T = Tc .
liquid-ice in water, or domains of magnetisation in ferromagnets).
5. The symmetries of the system in the phases on either side of the transition are
often unrelated.

A first order transition occurs for example in the Ising model below the critical tem-
perature (see the figure), when we change the external magnetic field from negative
to positive. Typically, the line of first-order transitions ends in a critical point. If we

25
 3.1 Characterising phase transitions

follow a path at zero magnetic field, passing through this point leads to a second-
order transition.
Second-order transitions are transitions in which

1. The singular behaviour is visible in the form of divergences.

2. The correlation length diverges.
3. The latent heat is zero.
4. There are no mixed phases at the transition point.
5. The symmetries of the system in the phases on either side of the transition
are related (usually the symmetries for T < Tc form a subgroup of those at
T > Tc ).
Using again the Ising model as an example, and focussing on h = 0, we see that the
spontaneous magnetisation for T < Tc breaks the symmetry of the phase at T > Tc .
Finally, there exist a class of infinite order transitions, in which no symmetries get
spontaneously broken. We have actually seen such a transition when we discussed
the Kosterlitz-Thouless model.1
For the purposes of this course, the second-order transitions are most relevant,
as the correlation length diverges and the system loses knowledge about the mi-
croscopic details. In the language of discretised quantum field theory, adjusting
the parameters such that the correlation length diverges is also known as taking the
continuum limit. The way in which the correlation length diverges is commonly pa-
Critical exponents determine the rameterised by the critical exponent , defined as
behaviour of observables near a
second-order phase transition.  
 T Tc  . (3.1)

In fact, the two-point correlator yields more information than just the correlation
length; recalling the definition (1.19),
 
1 r
hSi S j i hSi ihS j i D+ 2 exp . (3.2)
r

we also have the exponent , which sets the power-law behaviour of the two-point
correlator. The exponent is called the anomalous dimension. Typically, many other
correlation functions will diverge too. The first of those is the magnetisation, which
is the order parameter of the phase transition (it can be used to determine whether
we are in one phase or another). We thus have a number of other standard critical
Phase transitions are labelled by exponents,
the value of one or more order  
parameters.
specific heat : C  T Tc  (at h = 0) ,
 
order parameter : M  T Tc  (at h = 0) ,
  (3.3)
susceptibility :  T Tc  (at h = 0) ,

equation of state : M h1/ (at T = Tc ) .

The specific heat is defined as C = (2 F/T 2 )
V and the susceptibility as =
(M/h)T . For systems other than the Ising model, the role of M and h can of
course be played by some other parameters. We will see later that not all of these
The critical exponents , , , , critical exponents are independent; they are related by so-called scaling relations.
and the anomalous dimension
label a universality class.
1 The naming convention of phase transitions is historical and is vaguely connected to the order in

derivatives of the thermodynamic potentials at which a discontinuity is present. This historical detail is
best forgotten.

26
 3.2 Mean field theory

Experimentally, it is observed that many quite different systems nevertheless

have the same critical exponents. This goes under the name of universality. From
the point of view of the microscopic theory, the existence of such universality classes
can be quite suprising. Mean field theory and the Landau-Ginzburg approximation
both explain to some extent why universality exists, and how we can make use of it
to compute properties of systems near a phase transition.

3.2. Mean field theory

Mean field theory and Landau-Ginzburg models are phenomenological theories
which were developed to describe phase transitions. Both give the same predictions
for the critical exponents. Both ignore local fluctuations of spins (or other variables)
about their mean value. Hence, neither of them give correct critical exponents in
low dimensions, where these fluctuations can be very wild. As the dimension is
increased, the predictions of these theories is improved, and they essentially predict
the correct critical exponents at an ordinary critical point if D > 4. (end of lecture 9)
In the mean-field approach, we are still dealing with the microscopic degrees
of freedom, but instead of coupling them to their neighbours, we couple them to
the average field produced by all other degrees of freedom. Let us illustrate this
approach at the level of the Ising model. Recall the Hamiltonian (1.14),

HIsing = J Si S j h Si . (3.4)
hi,ji i

The idea is now to write Si = m + Si , i.e. write the spin as a difference with respect
to the average spin. We use this to write

Si S j = m2 + m(Si + S j ) + Si S j . (3.5)

The idea will now be to neglect Si S j in all computations (as well as higher order
terms in the variation).2 With this assumption, the Hamiltonian becomes
 
HMFT = J m 2 + m ( Si + S j ) h Si
hi,ji i

= JNDm2 (h + 2JDm) Si , (3.6)

i

where we have made use of the fact that in D dimensions, there are D links for each
site. The partition sum now becomes
" #
h i
ZMFT = exp JNDm 2
exp (h + 2JDm) Si
{ Si } i

h i
 N
= exp JNDm2 2 cosh (h + 2JDm) , (3.7)

where we have used once more the identities (2.20). The free energy becomes
h 
i
FMFT = N JDm2 k B T ln 2 cosh (h + 2JDm) . (3.8)

2 As the average spin m can be between plus and minus one, the difference S can be of order one
i
depending on the site. However, the idea is that deviations for which this value is large are rare, and for
most spins Si will in fact be small.

27
 3.3 Landau-Ginzburg

The magnetisation follows, as in section (2.1.1), either by computing the expec-

tation value of the spin directly,
" #
Sk exp (h + 2JDm) Si
{ Si } i  
m = h Sk i = " # = tanh (h + 2JDm) , (3.9)
exp (h + 2JDm) Si
<S>
1.5

1.0
{ Si } i
0.5

-1.5 -1.0 -0.5 0.5 1.0 1.5

m or by taking the h-derivative of the partition sum. The value for the magnetisation
-0.5 can be obtained graphically from this equation, see the figure. Plotting the magneti-
-1.0 sation as a function of T then gives us the mean-field approximation to the phase
-1.5 diagram of the Ising model.
The critical exponents can be obtained directly. Firstly, the critical point (that
Solving for the magnetisation of the
Ising model in the mean-field approx-
temperature above which (3.9) only has the trivial solution m = 0) is at T = 2DJ/k B .
imation. At h = 0, there are three solu- When h = 0 we can expand m near the critical point, to obtain
tions, m = 0 and m = m0 .
m
1
m m (1 t ) m3 (1 t )3 ,
1.0
(3.10)
0.8
3
0.6
1
0.4 where t = ( T Tc )/Tc (so that 2DJ (1 t). This leads to m (t)1/2 (of course
0.2
valid only for t < 0), so that the critical exponent = 1/2. Similar logic leads to the
0.0
1 2 3 4 5 6 7
T
other critical exponents, = 1 and = 3.
The mean-field approximation gets several things wrong, in particular the fact
Magnetisation of the two-dimensional
Ising model in the mean-field approxi-
that it predicts a second-order phase transition for any dimension (whereas we have
mation, for h = 0 (positive branch) and seen that no such transition exists for the one-dimensional Ising model).
h > 0, as a function of temperature.

3.3. Landau-Ginzburg
The Landau-Ginzburg approximation which assumes that the only relevant degrees
of freedom near a phase transition are those which are associated to an order pa-
rameter. Concretely, for the Ising model, this means that it attempts to describe
the model purely in terms of the average magnetisation m( x ) of the spins. In its
crudest form, it assumes that the average magnetisation is the same for all spins, in
(end of lecture 10) which case it obtains results which are like those of the mean-field approximation.
In contrast to that approximation, however, the Landau-Ginzburg approach does
not explicitly deal with the microscopic degrees of freedom. Rather, the goal is to
write down the most general expression for the free energy as a function of the order
parameter, such that it is compatible with the symmetries of the underlying discrete
F
model. The idea is that the order parameter is in some sense the slowest degree of
25
freedom of the system.
20
For the Ising model, the underlying system at h = 0 has a symmetry Si Si ,
15
so the Landau-Ginzburg energy has to be symmetric under a sign flip of m( x ). We
10
write
5
F = F0 + F2 m2 + F4 m4 + . . . . (3.11)
m
-2 -1 1 2

The value F2 = 0 corresponds to the critical temperature. For F2 > 0 only the
Behaviour of the Landau-Ginzburg
free energy for positive, zero and nega-
m = 0 minimum exists, but for F2 < 0 we get spontaneous symmetry breaking
tive value of F2 . For negative values, and two equivalent minima at m 6= 0 develop. Note that this happens despite the
two minima appear, which break the
symmetry.
fact that the free energy is completely smooth. In this broken phase, we can write
F2 = t F2 with t the reduced temperature t = ( T Tc )/Tc . The other coefficients stay
finite when t 0. For t < 0, the minimum of the free energy is given by
F
= 0 = 2 F2 tm + 4F4 m3 , (3.12)
m

28
 3.3 Landau-Ginzburg

F
from which we have m (t)1/2 . This recovers the mean-field exponent. Of 25

course, this had to happen, because the mean-field expression for the free energy, 20

15
expanded for small magnetisation, reads at h = 0 10

5
2 4
FMFT = F0 + DJN (1 2DJ)m + O(m ) . (3.13) 1.0 0.5 0.5 1.0
m

Here we also see again that the critical temperature corresponds to the zero of F2 . 10

We can easily extend this model to a non-vanishing external magnetic field. In

this case, the symmetry Si Si of the underlying microscopic model is broken, Free energy in the F4 > 0 case, for large
positive h
and so we should allow for terms in the free energy which are odd powers of m. The F
25
simplest one is of course the hm interaction, so that
20

15
2 4
F = F0 + hm + F2 m + F4 m + . . . (3.14) 10

There are good reasons to exclude e.g. the m3 term. Assuming again F4 > 0, we 1.0 0.5 0.5 1.0
m

find that for any h 6= 0 there is only one global minimum. By changing slowly and 5

10
continously from h > 0 to h < 0, the system undergoes a process by which it is
temporarily overheated, i.e. sitting in a local minimum with a higher energy than
the one in the global minimum. Near the critical point (where F2 = 0) the equation Free energy in the F4 > 0 case for h = 0
which determines the minimum of the free energy yields m h1/3 . So we see that 25
F

the critical exponent = 3. This is the same as in mean-field theory. 20

If F4 < 0, we need to include more terms in the free energy to prevent it from 15

10
being unbounded from below. Let us briefly discuss the system determined by the 5

Landau-Ginzburg free energy 1.0 0.5 0.5 1.0

m

F = F0 + F2 m2 | F4 |m4 + F6 m6 , (3.15) 10

where F6 > 0 in order to keep the free energy from running off to minus infinity. Free energy for small negative h; the
system is overheated and still sits in
The minima are determined by a local rather than global minimum.
F
25
F
= 0 = m(2F2 4| F4 |m2 + 6F6 m4 ) , (3.16) 20

m 15

10
which yields p 5

| F4 | | F4 |2 3F6 F2 m

m=0 or m2 = . (3.17) 1.0 0.5

5
0.5 1.0

3F6 10

When F2 > | F4 |2 /(3F6 )

only the m = 0 minimum exists. However, as long as | F4 |
is not zero, we can always make two minima appear by moving the temperature Free energy for large negative h; the
system has now rolled to the global
sufficiently close to Tc (i.e. making F2 positive but sufficiently small). By tuning minimum.
F2 , we can make these two extra minima appear at F = F0 , hence creating three
degenerate ground states.
There are thus two temperature scales that play a role. For T > T0 (where T0 is
the temperature at which the three minima become degenerate), there is only one
global minimum at m = 0. For Tc < T < T0 there are three minima, one of which
sits at m = 0, but the other two have lower free energy. Finally, for T < Tc the
minimum at m = 0 disappears (turns into a maximum).
In the F4 0 limit, the two local minima that are typically present just below
the critical temperature disappear. The special point at F2 = F4 = 0 is the tricritical
point. At this point, the equation of motion yields m h1/5 so that = 5. This
is quite different from the mean-field behaviour at the ordinary critical point. This
behaviour is found in e.g. a mixture of 3 He and 4 He.
Finally, the assumption that m is a constant is only the first step as far as the
Phase diagram of the model with F6 =
Landau-Ginzburg approach is concerned. It makes perfect sense to assume that in 1, showing the tricritical point. At F4 =
fact m = m( x ) and include terms in the free energy which contain gradients of the 0, the system sits on the vertical axis,
with the unbroken phase at T < Tc and
average magnetisation. the broken phase at T > Tc , as before.
(end of lecture 11)
29
3.3 Landau-Ginzburg

30
 Universality and renormalisation
4
We have seen that if we restrict ourselves to large distance physics, the idea of
Landau-Ginzburg is that we can restrict ourselves to the dynamics of the slow, long-
wavelength modes, and this then essentially forces us to conclude that many sys-
tems which are quite different at microscopic scale will behave in the same way
when we get close to a phase transition. This phenomenon is called universality.
The only things which really seem to matter are the symmetries of the underlying
model, and possibly the dimension of the system.
While the Landau-Ginzburg approach captures this idea nicely in the form of
a mathematical model, it would be much nicer if we could derive this universality
property from the microscopic models. The techniques used to do this go under
the name of the renormalisation group. They allow us to start from a microscopic
model and, in successive steps, integrate out microscopic variables to end up with
a model which describes the physics at a somewhat larger distance scale. Follow-
ing this many times, we can then see which microscopic interactions determine the
behavour at large distances, and which ones are irrelevant for the purpose of de-
scribing a phase transition.

4.1. Kadanoff block scaling

Before we discuss more general aspects, let us first illustrate the renormalisation
group idea at the level of the one-dimensional Ising model. We start with a lattice
with spacing a. The idea is now that we are only interested in what happens at larger
distances, e.g. 2a. We can then try to construct a new effective model with lattice
spacing 2a, in such a way that we reproduce all physics at these larger distances
from the simpler model with half the number of degrees of freedom.
The Ising partition sum is
N h i
Z= exp JS i S i + 1 + hS i + F0 = TS1 S2 TS2 S3 TSN S1 = Tr T N , (4.1)
{ Si } i = 1 { Si }

with the transfer matrix

A 1
!
AB
T=C A = e J , B = e h , C = e F0 . (4.2)
A 1 AB1

We have added a constant F0 to the energy, as we will shortly see that a renormalisa-
tion group analysis requires us to keep track of it. The idea is now to do a so-called
block-spin transformation, in which we group the dynamics of two spins together,

31
4.1 Kadanoff block scaling

i.e. integrate over every second spin. We can do that by grouping the transfer ma-
trices in groups of two, and define

T ( A, C ) = T 2 ( A, B, C ) .
B, (4.3)

This gives rise to

C B A = C2 ( A2 B2 + A2 ) ,

C A 1 = C2 ( B + B1 ) , (4.4)

C A B 1 = C2 ( A2 B2 + A2 ) .
These are discrete renormalisation group equations: they relate the parameters (cou-
pling constants) of the original model to those of the new model. The number of de-
grees of freedom has been reduced by a factor of two, but the long-range behaviour
of these two models is the same.
Isolating the parameters of the coarse grained model, and introducing the fol-
lowing shorthands,

x = e4J , y = e2h , z = C 4 , (4.5)

the system of equations becomes

x (1 + y )2
x = ,
( x + y)(1 + xy)
y( x + y)
y = , (4.6)
1 + xy

z2 xy2 (1 + y)2
z = .
( x + y)(1 + xy)

Starting from some initial point ( x, y, z), we can now follow the images of this point
under successive renormalisation group transformations. Restricting ourselves to
the x y plane, we get the following plot,
y

1.0

0.8

0.6

0.4

0.2

0.0 x
0.0 0.2 0.4 0.6 0.8 1.0

Under n blocking transformations, the correlation length (expressed in terms of

the number of lattice sites) should scale as

= 2n . (4.7)

However, at a fixed point of the renormalisation group flow, the parameters do not
change, and hence the system will have the same correlation length before and after
the transformations,
fixed point : = . (4.8)

32
 4.2 Renormalisation group flow

This can only be true if = 0 (a trivial fixed point) or = (a non-trivial fixed

point). (end of lecture 12)
We have here seen the simplest way in which we can implement the renormal-
isation group. The process of summing over every other spin is called decimation.
It has the property that the remaining variables, i.e. the variables in which the new
system is expressed, are again Ising spins. The one-dimensional case is even more
special in the sense that the new system is not just expressed in terms of Ising spins,
but does not have anything more than the nearest-neighbour interactions that we
started from.
In general, things are not so simple. In two dimensions, for example, you might
be tempted to sum over all spins on odd lattice sites. While this leaves us with a
system of Ising spins (the even lattice spins), it turns out to introduce new couplings,
e.g. next-to-nearest neighbour ones.

I See also: This standard example appears e.g. in [3].

4.2. Renormalisation group flow

4.2.1 Scaling near the critical point
Let us now look at the renormalisation group in a somewhat more formal setting.
The renormalisation group equations in general map a set of parameters (couplings)
to another one; we will denote the set of parameters by K and its individual mem-
bers by K i ,
ren.group
K i = { J, h, . . .}i K i = { J, . . . }i = Ri ( K ) .
h, (4.9)

A fixed point Ki of the equations is one where K i = K i = Ki . In general, the

renormalisation group transformation is highly non-linear. Close to a fixed point,
however, we can expand the transformation in powers of the small deviation V i
from the fixed point, or more formally,

Ri (K ) j
K i = K i + V i = Ki + V , or V i = Ri j ( K ) V j , (4.10)
j K j j

which implicitly defines the matrix components Ri j . The problem becomes linear
(because R is evaluated at the fixed point).
The eigenvalues j and eigenvectors ~e j of R play an important role.1 We will
write the eigenvalues as
j by j , (4.11)
where b is the scale factor associated to the renormalisation group transformation
(in the Ising case we discussed, b = 2). We can now expand a point K i close to the
fixed point in terms of the eigenvectors,

K i = Ki + g j e j i . (4.12)
j

The expansion coefficients g j are called the scaling fields. Applying one step of the
renormalisation group transformation gives

g j = by j g j . (4.13)

1 In general there is of course no guarantee that R is symmetric, so that the eigenvalues do not have

to be real and the eigenvectors do not have to span a complete basis. We will assume, however, that they
do.

33
 4.2 Renormalisation group flow

This shows that there are particular combinations of the parameters of the theory
which scale with a simple factor under the renormalisation group. The exponents
y j are directly related to the critical exponents we have seen in (3.3); we will see this
illustrated explicitly on the Ising model shortly.2
It is important to emphasise once more that the partition function does not change
under the renormalisation group flow. We simply change from a partition function
expressed in terms of N spins to one of N spins, with modified couplings,

Tr{Si } e N f ({K }) = Tr{Si } e N f ({K }) . (4.15)

= b D N, we have that the free energy per spin has to satisfy

Because N

f ({K }) = b D f (K ) . (4.16)

If we express the couplings using the scaling fields g j instead of the components K i ,
The free energy near a critical point and then use (4.13), we get
is a generalised homogeneous
function, with exponents related to
f ( g1 , g2 , . . . ) = b D f ( b y 1 g1 , b y 2 g2 , . . . ) . (4.17)
the eigenvectors and -values of the
renormalisation group matrix at
that point. Functions satisfying this property are called generalised homogeneous. We see that the
free energy near a critical point is constrained by the eigenvectors and -values of the
renormalisation group transformation matrix.

4.2.2 Critical surface and universality

We have seen that a good way to parameterise the system near a critical point is in
terms of the expansion coefficients g j , defined in (4.12). These linear combinations
of the parameters in K scale in a nice way under scale transformations, as we have
seen in (4.13). We can distinguish three different cases:

yi > 0
The operators multiplying these coefficients are called relevant. As we do more
and transformations, the coefficients grow larger and larger, driving the sys-
tem away from the critical point.

yi < 0
These operators are called irrelevant. If you give them a small non-zero co-
efficient, a renormalisation group transformation will drive these coefficients
back to zero, i.e. back to the critical point.
y=0
These are marginal. In order to understand what happens, you have to go
beyond the linear approximation.
In the vicinity of a critical point, the irrelevant directions span a subspace of
points which are attracted to the critical point. By continuity, this subspace will exist
also in some larger neighbourhood of K . We call this subspace the critical surface.
On the critical surface, the correlation length is infinite. This follows because the

2 For the one-dimensional Ising model, we find from (4.6) that the matrix R at the critical point

( x, y) = (0, 1) is  
4 0
R (0, 1) = . (4.14)
0 2
The eigenvectors are simply ~e1 = (1, 0) and ~e2 = (0, 1), with eigenvalues 1 = 4 and 2 = 2 respectively,
so that y1 = 2 and y2 = 1. The scaling fields are x and y. In terms of the nomenclature to be introduced
below, these couplings (or equivalently, J and h) are both relevant.

34
 4.3 Momentum space flow and quantum field theory

correlation length can only decrease under renormalisation group transformations,

but it has to be infinite at the critical point.
The directions orthogonal to the critical surface are spanned by the relevant pa-
rameters. If you want to achieve infinite correlation length, these parameters have
to be tuned perfectly, because any small deviation from the critical surface will get
amplified by successive renormalisation group transformations. (end of lecture 13)

I See also: [2] chapter 3, [1].

4.3. Momentum space flow and quantum field theory

We have so far discussed the renormalisation group transformation in terms of co-
ordinate space, explicitly constructing blocks of microscopic degrees of freedom
and clumping them together, integrating out the fluctuations within the block. In
quantum field theory, we are more accustomed to a discussion in momentum space,
as it is there that an interpretation of asymptotic states (free particles with some
given momentum) becomes most transparent. So let us transcribe the renormalisa-
tion group idea to that setting.
In coordinate space, the renormalisation group has been implemented by sum-
ming over fluctuations with wavelength between a and ba, so that we obtain a new
lattice system with lattice spacing ba. All distances (measured in units of the lattice
spacing) get scaled down, i.e. = b1 . If we do more complicated transformations
than the simple decimation we have seen in one dimension, it may also be neces-
sary to rescale the remaining variables by a factor, in order to get back to the original
form of the Hamiltonian (plus possible new interaction terms). In momentum space
the story will be very similar. Summing over the small wavelength fluctuations is
now replaced by summing or integrating over high-momentum modes. After that,
all distances are scaled with b1 , and thus all momenta are scaled with b. Sometimes
a rescaling of the variables is required.
Let us make this more explicit, by considering a scalar field on a lattice. The
Hamiltonian is given by (we will start straight in the continuum)
 
1 1 1
Z
H = dD x ()2 + m20 2 + 0 4 + . . . , (4.18)
2 2 4!

and of course it is possible to write down an arbitrary number of higher-order terms.

Here we immediately see that we will have to rescale the field after the blocking
transformation: if x = b1 x then = b and in order to keep the same normalisa-
tion of the gradient term, we need to scale

( x ) = b( D2)/2 ( x ) =: bd ( x ) . (4.19)

The exponent d = ( D 2)/2 is called the engineering dimension of the field, or per-
haps more useful, the mass dimension (if the mass dimension is one, the field scales
in a way opposite to length). In a similar way we can work out the dimensions of
the coupling constants that appear in the Hamiltonian. If we now go to momentum
space using a Fourier transformation, we get

dD x i~k~x
Z
(~k) := e (~x ) . (4.20)
L D/2
The normalisation L D/2 is there also in the inverse transform. From this we find
that the field in momentum space transforms as

(k ) = bd D/2 (k) . (4.21)

35
 4.3 Momentum space flow and quantum field theory

The idea of the momentum space renormalisation group transformation is now

to split the Hamiltonian in modes with momentum below /b and those between
/b and . We will then integrate over the latter. This integration is trivial when
0 = 0, because we then just have a Gaussian integral. If the coupling is non-zero,
we generate additional terms. It is not easy to do the integral in that case, but one
way to compute them is to resort to perturbation theory in 0 . At the very lowest
order, there is only one diagram that contributes, namely

(4.22)

This generates an effective interaction term for the long wavelength modes,
 
D 2d 2 1 2 1 2  2 0  2
Z
D 22d 0 4
H = d x b ( ) + b m0 + G + b . (4.23)
2 2 2 4!

This gives us for the renormalisation group transformation

 
20 = b2 m20 + 0 G ,
m
2 (4.24)
0 = b4 D 0 .

There is a simple fixed point of these equations, at m20 = 0 = 0. To make things

more explicit and prepare for linearisation, we at this stage in perturbation theory
only need the integral
Z D 1
k D 2
dk = (1 b 2 D ) . (4.25)
/b k2 D2

Linearising around the fixed point gives us a renormalisation group matrix

b2 B ( b 2 b 4 D ) K D D 2
 
R= , B= . (4.26)
0 b 4 D 2( D 2)

The eigenvectors and -values are

~e1 = (1, 0) , 1 = b2 , y1 = 2 , ~e2 = ( B, 1) , 2 = b4 D , y2 = 4 D . (4.27)

For D > 4 this has one critical exponent y1 = 2 and one exponent y2 = 4 D < 0.
That is, there is one relevant direction, corresponding to the m0 parameter, and one
(end of lecture 14) irrelevant direction, corresponding to 0 .

36
 Lattice gauge theory
5
To tie the various element of this course together, let us finally have a look at a lattice
formulation of the theory of strong interactions, described by a SU(3) gauge theory.
We will see various things which we have encountered in earlier chapters some
back in a somewhat more complicated setting, in particular the strong coupling ex-
pansion and the condensation of vortex-like configurations in certain phases of the
model.

5.1. Lattice action

To formulate gauge theory on the lattice, we need to find a discretised version of the
action
1
Z
S= d4 x Tr( F F ) , (5.1)
4
where the gauge field strength is given by
a
F = Aa Aa + gYM f abc Ab Ac . (5.2)

Here f abc are the structure constants of the group. This action is invariant under the
non-abelian gauge transformations
1 1
A g 1 A g + g g ,
gYM (5.3)
1
F g F g .
So we better make sure that this invariance exists in the lattice version of the theory
as well.
Let us first introduce some lattice language. We will denote a point on the lattice
by
x = m a , m = 0, 1, . . . , N 1 . (5.4)
For derivatives of a field at a point x, we write
1
x = (x+a x ) , (5.5)
a
where is a unit-vector in the direction.
The idea will now be to incorporate gauge invariance by associating the vector
potential A not to the nodes of the lattice, but to the links connecting them. More
precisely, we will attach a link variable U to each link, given by
)
U ( x ) = eigYM aA ( x+ a/2 . (5.6)

37
5.2 Wilson loops

Here a is, as always, the lattice spacing, and the relation to the continuum A is
made by evaluating the latter at the midpoint between the nodes x and x + a. If
we now perform a gauge rotation g( x ) at each site, the link variables transform as

U ( x ) g( x )U ( x ) g1 ( x + a ) . (5.7)

This is precisely what we need to know to construct a gauge invariant action: if we

string together a set of link variables such that the contour is closed, we obtain a
gauge invariant object.
The simplest such object is the plaquette action, which is obtained by going around
a small two-dimensional square of the lattice and multiplying all link variables that
we encounter,
1  
Splaq = 2 < Tr Uij Ujk Ukl Uli . (5.8)
2gYM
We can then construct an action by summing over this expression over all possible
plaquettes in the lattice. This indeed reproduces, in the small-a limit, the continuum
action. To see that, use the definition of the link variable and expand for small a.
Using CBH then gives
2
UUUU eia gYM F . (5.9)
Upon inserting this in the trace and expanding to second order in a we get the re-
quired continuum result.

5.2. Wilson loops

Having access to a lattice action, how now do we go about and construct quantities
which correspond to interesting observables?1 One thing we would like to com-
pute is the interaction potential between two static quarks. We can compute that by
creating a quark-antiquark state, evolving it in time and annihilating it again,

hqq|e HT |qqi = eV ( R)T , (5.10)

where R is the distance between the quarks. In a path integral setting, what we need
to compute is the path integral in the presence of a source, which peaks at the two
positions of the quarks, i.e.
Z h Z i
D Aa exp S + igYM Aa Ja d4 x , (5.11)

with suitably chosen Ja ( x ). Incorporating the quark-antiquark creation process, we

are computing the expectation value of a link variable evaluated along the closed
path,
1 H
V ( R) = lim lnhTr PeigYM C A dx i (5.12)
T T

The object on the right-hand side is called a Wilson loop. The symbol P denotes path
ordering, i.e. the matrices A dx are multiplied together in the order in which they
appear on the contour C. In the lattice version, this becomes a bit more well-defined:
we now have
1
V ( R) = lim h U (n)i , (5.13)
T T
C

1 We cannot simply compute the expectation value of a link variable, because it is not gauge invari-

ant, and so it has to vanish. The simplest gauge-invariant object is obtained by stringing together link
variables in a closed loop.

38
 5.4 Monopole condensation

that is, the expectation value of the products of link variables, evaluated in the order
in which they appear on the contour C.
If this expression scales linearly with distance, we have a confining potential,
V ( R) = R + . . . . (5.14)
We will see that this behaviour is indeed present when we look at the strong-coupling
limit of the model, for which we can use techniques seen earlier (the high-temperature
expansion).

5.3. Strong coupling expansion and confinement

We would like to see that at strong coupling, the potential is a linear function of R.
This can be done using methods very similar to the strong coupling expansion we
have seen in the Ising model. What we need to compute is
Z .Z
h U (n)i = dU (n) U (n)eS dU (n)eS . (5.15)
C n, C n,

This involves integrating over the group for every link variable. We need some
identities for these group integrals,
1
Z Z Z

dU = 1 , dU Uij = 0 , dU Uij Ukl = . (5.16)
2 il jk
These are similar, in some sense, to the relations we had for the exponential integrals
in the Kosterlitz-Thouless model (2.38). Since we have one link variable for every
link in the contour C, we will need to bring down, from the action, a set of plaquettes
which fills the contour, as in the figure. Since the plaquette action has an overall
2 (see (5.8)), this gives
factor 1/gYM
! Nc
1
h U (n)i = 2
= exp(Area ln g2 ) . (5.17)
C g YM

We thus see that we indeed obtain an area law, and the string tension at strong
coupling is given by
2
= ln gYM +... . (5.18)
The conclusion is that for large bare coupling gYM , the colour force is confining.

I See also: [4] section 3.2.

5.4. Monopole condensation

Unfortunately, the fixed point in QCD is the one where the bare coupling goes to
zero. That is to say, the continuum limit requires us to go away from the strong-
coupling regime just analysed. It then becomes less clear what is responsible for the
confining force (which, in contrast to the bare coupling, is a physically measurable
quantity). However, if we just do perturbation theory at small coupling, no sign of
the confining force arises.
We thus need some new effects, very similar to what we have seen in the Kosterlitz-
Thouless model, to understand what makes the regime of small bare coupling pro-
duce a confining force. Electron condensation leads to magnetic flux tubes con-
necting monopoles. Monopole condensation leads to electric flux tubes connecting
electrons. (end of lecture 15/16)

I See also: See e.g. [10] section 5.8 for more literature pointers.

39
5.4 Monopole condensation

40
Bibliography

[1] M. Le Bellac, Quantum and statistical field theory, Oxford University Press,
1991.
[2] J. L. Cardy, Scaling and renormalization in statistical physics, Cambridge
University Press, 1996.
[3] J. M. Yeomans, Statistical mechanics of phase transitions, Oxford University
Press, 1992.

[4] J. B. Kogut and M. A. Stephanov, The phases of quantum chromodynamics:

From confinement to extreme environments, 2004.
[5] A. M. Polyakov, Gauge fields and strings, Harwood, 1987. Contemporary
concepts in physics, volume 3.

[6] M. H. Thoma, New developments and applications of thermal field theory,

hep-ph/0010164.
[7] G. P. Lepage, What is Renormalization?, hep-ph/0506330.
[8] M. Creutz, Quarks, gluons and lattices, Cambridge University Press, 1983.

[9] I. Montvay and G. Munster, Quantum fields on a lattice, Cambridge
University Press, 1994.
[10] J. Smit, Introduction to quantum fields on a lattice: A robust mate,
Cambridge Lect. Notes Phys., 2002.

[11] J. Zinn-Justin, Quantum field theory and critical phenomena, Int. Ser.
Monogr. Phys. 113 (2002) 11054.
[12] J. Zinn-Justin, Quantum field theory at finite temperature: An introduction,
hep-ph/0005272.
[13] A. Zee, Quantum field theory in a nutshell, 2003.

[14] K. Peeters, A guide to biophysics and biomathematics,

http://maths.dur.ac.uk/~dma6kp/biophysics.html.
[15] K. Huang, Statistical mechanics, Wiley, 1987.
[16] R. Savit, Duality in field theory and statistical systems, Rev. Mod. Phys. 52
(1980) 453.

41