Lectures Stat Field Theory Vnew

Laboratoire de physique de l’École normale supérieure de Paris.
Centre national de la recherche scientifique, Institut de physique.

Université PSL, Sorbonne université, Université de Paris.
Statistical Field Theory and Applications :

An Introduction for (and by) Amateurs
by
Denis BERNARD
Laboratoire de Physique de l’École Normale Supérieure,

CNRS, ENS & Université PSL,
Sorbonne Université, Université de Paris, 75005 Paris, France.
Web: https://www.phys.ens.fr/∼dbernard/
June 2020*
Preliminary version : 2020.....

Chapter 3 was written in part by Jesper Jacobsen.
Comments are welcome.
D. Bernard Statistical Field Theory for (and by) Amateurs
Contents
1 What are we aiming at describing ? 7
2 Brownian motions and random paths 13

2.1 Random walks and random paths . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Scaling limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Brownian motions and path integrals . . . . . . . . . . . . . . . . . . . . . . 18
2.4 The 2D Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 Brownian motions and stochastic differential equations . . . . . . . . . . . . 22
2.6 Path integral representation for SDEs . . . . . . . . . . . . . . . . . . . . . 26
2.7 The small noise limit and the Arrhenius law . . . . . . . . . . . . . . . . . . 31
2.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3 Statistical lattice models 39

3.1 Examples of statistical lattice models . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Transfer matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 The 2D Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 Partition functions and path integrals . . . . . . . . . . . . . . . . . . . . . 51
3.5 The classical/quantum correspondence . . . . . . . . . . . . . . . . . . . . . 55
3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4 From statistical models to field theories 60

4.1 Why do we care about second order phase transitions ? . . . . . . . . . . . 60
4.2 Field theory representation of statistical models . . . . . . . . . . . . . . . . 62
4.3 Critical systems: phenomenology . . . . . . . . . . . . . . . . . . . . . . . . 66
4.4 Landau-Ginzburg theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.5 Upper and lower critical dimensions . . . . . . . . . . . . . . . . . . . . . . 73
4.6 Deviation from mean field theory . . . . . . . . . . . . . . . . . . . . . . . . 75
4.7 Symmetry breaking and Goldstone modes . . . . . . . . . . . . . . . . . . . 76
4.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5 The renormalisation group and universality 80

5.1 Block spins and RG transformations . . . . . . . . . . . . . . . . . . . . . . 80
5.2 RG fixed points and universality . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3 Scaling functions and critical exponents . . . . . . . . . . . . . . . . . . . . 85
5.4 Corrections to scaling and finite size effects . . . . . . . . . . . . . . . . . . 88
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6 Free field theory 91

6.1 Euclidean (free) field theories . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.2 Gaussian field theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.3 Green functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.4 Products and composite operators . . . . . . . . . . . . . . . . . . . . . . . 99
6.5 Quantization of free field theories . . . . . . . . . . . . . . . . . . . . . . . . 101
2
6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7 Interacting field theory: basics 109

7.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.2 Symmetries and Ward identities . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.3 Generating functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.4 Perturbation theory and Feynman rules . . . . . . . . . . . . . . . . . . . . 119
7.5 Diagrammatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.6 One-loop effective action and potential . . . . . . . . . . . . . . . . . . . . . 126
7.7 The O(N ) vector models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8 Conformal field theory: basics 135

8.1 The group of conformal transformations . . . . . . . . . . . . . . . . . . . . 135
8.2 Conformal invariance in field theory . . . . . . . . . . . . . . . . . . . . . . 138
8.3 Fields and operator product expansions . . . . . . . . . . . . . . . . . . . . 142
8.4 Massless gaussian free field in 2D . . . . . . . . . . . . . . . . . . . . . . . . 143
8.5 2D conformal field theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.6 Operator formalism in 2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
9 Scaling limits and

the field theory renormalisation group 156
9.1 Field RG transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9.2 Momentum RG transformations in field theory . . . . . . . . . . . . . . . . 161
9.3 The perturbative renormalization group . . . . . . . . . . . . . . . . . . . . 163
9.4 The Wilson-Fisher fixed point . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.5 Scaling limits and renormalized theories . . . . . . . . . . . . . . . . . . . . 168
9.6 Covariance of the renormalization . . . . . . . . . . . . . . . . . . . . . . . . 171
9.7 Perturbatively renormalized φ4 theory . . . . . . . . . . . . . . . . . . . . . 174
9.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
10 Miscellaneous applications 179

10.1 The XY model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
10.2 Self avoiding walks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.3 The coulomb gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
10.4 The stress tensor and diffeomorphisms . . . . . . . . . . . . . . . . . . . . . 189
10.5 Irreversibility of RG flows and the c-theorem . . . . . . . . . . . . . . . . . 191
10.6 Exact RG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
10.7 Etc... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
11 Fermionic techniques 194

11.1 Fermions and Grassmann variables . . . . . . . . . . . . . . . . . . . . . . . 194
11.2 The 2D Ising model again . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
11.3 Dirac fermions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
11.4 Chiral anomaly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
3
11.5 Fermionisation-Bosonisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

11.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
12 Stochastic field theories and dynamics 194

12.1 Et un... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
12.2 Et deux... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
12.3 Et trois... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
12.4 Zero... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
13 Exercise corrections 194
4
An exemple of lecture series:

The lecture series may be far from covering the materials written here. For instance:
• Brownian motions and random paths

1.0 What we are aiming at describing?
2.1 Random walks and random paths
2.2 Scaling limits
• Statistical lattice models
3.1 Examples of statistical lattice models
3.2 Transfer matrices
3.3 2D Ising model
• From statistical model to field theory
4.1 Why do we care about second order phase transitions ?
4.2 Field theory representation of statistical models
(4.5 Landau-Ginzburg theory)
4.7 Deviation from mean field theory
• Renormalization group and universality
5.1 Block spins and RG transformations
5.2 RG fixed points and universality
5.3 Scaling functions and critical exponents
• Free field theory
6.3 Gaussian field theories
6.4 Green functions
6.5 Products and composite operators
• Interacting field theory
7.3 Generating functions
7.4 Perturbation theory and Feynman rules
7.5 Diagrammatics
7.0 The O(N) vector model with N → ∞ in D = 3
• Conformal field theory
8.4 Massless gaussian free field in 2D
8.1 The conformal group
8.2 Conformal invariance in field theory
8.3 Operator product expansion
• Scaling limits and the field theory renormalization group
(9.2 Back to basics & field transformations)
9.3 The perturbative renormalization group
9.4 The Wilson-Fisher fixed point
(9.5 Scaling limits and renormalized theories)
5
6
1 What are we aiming at describing ?
Statistical Field Theory is an important topic in theoretical physics, with a wide range
of interdisciplinary applications from statistical physics or condensed matter to high energy
physics or random geometry, which has undergone many progresses in recent years. It is
of course intimately linked to Quantum Field Theory.
Statistical Field Theory aims at dealing with the behavior of systems (classical or quan-
tum) with a large —actually infinitely large— number of interacting degrees of freedom.
These systems have very interesting and peculiar behaviours: they have different phases
with different characters, they manifest phase transitions, their behaviors are dominated
by collective modes and/or refined geometrical patterns, etc. Their understanding and
analysis make contact with very elegant mathematical structures (say probability theory,
representation theory, geometry) and with remarkable concepts, notably the renormaliza-
tion group which is nowadays a cornerstone of Physics and its ramification.
Statistical Field Theory aims at an understanding of those behaviors on the basis of a
few physical principles. This is particularly true for its application to critical phase tran-
sitions through their universality property. These are characterized by sharp transitions
in the physical properties of statistical systems controlled by external parameters. They
are induced by collective phenomena which involve large fluctuations over long distances
without scale separation. Statistical Field Theory provides tools to deal with many nested
degrees of freedom, with large fluctuations, over a cascade of scales.
A sample of Samples of 2D self-avoiding walks,

a Brownian curve in 2D. alias polymers.
Understanding random patterns is at the core of the comprehension of many physical

phenomena or mathematical structures, and the Brownian motion is a historical example
of such structures. Although the relevant geometries can be as simple as gentle curved or
surfaces, many relevant patterns are however not well-described by an integer dimension
but have a fractal character, at least over some length scale. For instance, the singular
behavior of second order phase transitions are compatible with fractal surfaces separating
phases at the critical point. In general, the deeper our understanding of these geometries,
7
the more complete our predictions can be. See Figure.

Polymers provide simple examples of such fractal geometry going beyond the simple
model of Brownian motions. These can be modeled by self avoiding random walks which
may be viewed as the paths drawn by a random walker on a square lattice constrained
not to visit twice any site of the lattice. This constraint mimics the self-repulsion of
the polymers. They undergo a phase transition when the temperature is varied: at low
temperature the polymers is compactly curled up while at large temperature it extends
macroscopically and resembles a fluctuating smooth curve. At the critical point, the poly-
mers possess macroscopic fractal shapes which nowadays can be described using statistical
field theory tools. See Figure.
Percolation is another geometrical example manifesting phase transition and generating
random fractal structures. Imagine that we randomly color the cells of a honeycomb lattice
black or white. The rule is that each cell has a probability p of being white and 1 − p
of being black. If p is small, most of the cells will be black, with a few small islands of
white cells. As p is increased, these white clusters grow larger until a critical value pc at
which there is a non-zero probability that one of these clusters spans the whole domain,
no matter how large it is. Percolation is important as a model for random inhomogeneous
systems, for instance if the black clusters represent untapped oil pockets it is much easier
to extract the oil if they percolate. See Figure.
A small sample of definition (left).

A critical percolating clusters (right).
Critical phase transitions (second order phase transitions) of statistical systems are
the main actors in statistical field theory. There is a large variety of physical systems
exhibiting second order phase transitions. Standard examples are the para-to-ferro mag-
netic transition in magnetic materials, the superfluidity transition in quantum fluids, the
superconductivity transition of certain metallic materials at low temperature, etc. Sec-
ond order phase transitions have a universality property in the sense that the types of
phase transitions fall into a relatively small number of categories, known as universality
classes, which all behave similarly. Close to the transition there are singularities in the
thermodynamical functions and, in parallel, large anomalous fluctuations and power law
behavior of the correlation functions. These singular behaviors of thermodynamical func-
tions, characterized by so-called scaling exponents, are understood within statistical field
theory. The size of the fluctuations is characterized by a length, the correlation length
ξ, which is the typical extension of the domain over which the degrees of freedom are
correlated. At the critical point, fluctuations are all over the scale and as a consequence
8
this length diverges. On one hand, this has dramatic consequences because all degrees of
freedom are then coupled, at any scale, making the analysis difficult (to say the least),
but one the other hand, it renders a continuous description within statistical field theory
possible and it is at the origin of the universality property.
Magnetization versus temperature Heat capacity versus temperature

in the 2D Ising model. in the 2D Ising model.
The archetypical statistical model of phase transition is the Ising model for magnetic
transition. The Ising degrees of freedom are simple spin variables si , whose values are
either + or −, defined on the sites i of a lattice. The statistical weight of a given spin
configuration is proportional to the Boltzmann −E[{s}]/T , with T the temperature
P weight e
and with interaction energy E[{s}] = −J i,j si sj where the sum is restricted to the
neighbour spins on the lattice. For J > 0, the configuration with aligned spins are the most
probable at low temperature. There is a phase transition at a certain critical temperature
Tc . At T > Tc , spins have a tendency not to be aligned and the mean magnetization
vanishes, while at T < Tc spins get ordered, they form large clusters, and there is a non
zero spontaneous magnetization. At T = Tc , fluctuations are large, clusters of identical
spins form scale invariant fractal patterns. See Figure.
Ising configurations
Statistical Field Theory is also applicable to statistical systems out-of-equilibrium in-

volving a large number of coupled degrees of freedom fluctuating over large range of scales.
Here, fluid turbulence or turbulent transports are the main —yet unsolved— classical ex-
amples. Turbulence in fluids is one of those natural phenomena for which we think we
9
understand the underlying physics but are nevertheless unable to make analytic precise
predictions. Turbulence occurs when a piece of fluid with low viscosity is stirred at large
length scales. There is then a continual cascade of energy from large to small scales where
the energy is dissipated. This cascade of energy may be described statistically in terms
of random velocity fluctuations, with a power-law spectrum of their Fourier components.
Those fluctuations are all over the scales over which the cascade takes place and, as a
consequence, are potentially describable with statistical field theory methods. In two di-
mensions the fluid vorticity clusters form random mutli-scale patterns bearing similarities
with clusters of critical statistical transitions. See Figure.
Vorticity clusters A sample of the 2D massless

in 2D turbulence discrete Gaussian free field
As its name suggests, statistical field theory has grown from the merging of field theory,
classical or quantum, and statistical physics. As in classical field theory, statistical field
theory deals with the continuum of degrees of freedom encompassed in field variables.
Those degrees of freedom, which for instance encode for the large distance behaviour
of relevant fluctuations close to a critical phase transition, are characterized by their
geometrical properties and symmetries. Concretely the fields take values in manifolds
specified by those properties. Because it deals with the fluctuations of these degrees of
freedom, statistical field theory gives statistical weights to any of the field configurations.
At equilibrium these weights are specified according to the Boltzmann rules for statistical
physics. Naively, statistical field theory may be viewed as an attempt to define probability
measures on the spaces of field configurations —which are indeed very infinite dimensional
spaces. The construction of these measures is of course guided by statistical physics but
also by a few other general principles, say local property inherited from the local nature
of the underlying degrees of freedom and their interaction, symmetry reflecting global or
internal geometrical properties depending on the systems, etc.
The renormalization group, which provides both a conceptual framework and efficient
tools to deal with these infinitely many coupled degrees of freedom, is a key principle
underlying the construction and the use of statistical field theory. Its impact on Physics
—in general not only in statistical or quantum field theory— cannot be underestimated.
See Figure.
By its very nature, Statistical Field Theory borrows tools from statistics or probabil-
ity theory, say Gaussian variables or fields. It also uses techniques from representation
theory, in particular because of its relation with Quantum Field Theory. Geometry, group
10
theory and dynamics have been historical cross points between mathematics and theoret-
ical physics, as recently illustrated by the impact of geometry or symmetry algebra on
gauge theory, on the standard model of particle physics, or on critical Conformal Field
Theory. The recent evolution advocates for a deeper role of probability theory in this
mathematics-and-physics interaction, especially in statistical field theory.
11
A non-exhaustive list of references:
• L. Kadanoff, Statistical Physics: Statics, Dynamics and Renormalization, World

Scientific Publishing Co., (2000).
• C. Itzykson and J.M. Drouffe, Statistical Field Theory, vol. 1 & 2, Cambridge
University Press, (1989); or in French, Théorie statistique des champs, vol. 1 & 2,
Editions du CNRS, (1989).
• J. Zinn-Justin, Quantum Field Theory and Critical Phenomena, Oxford University

Press, (1989).
• G. Parisi, Statistical Field Theory, Addison-Wesley Publishing Co., (1988).
• J. Cardy, Scaling and Renormalization in Statistical Physics, Cambridge University

Press, (1996).
• Ph. Di Francesco, P. Mathieu, D. Sénéchal, Conformal Field Theory, Springer-Verlag

NY, (1996).
• M. Kardar, Statistical Physics of Fields, Cambridge University Press, (2007).
• G. Mussardo, Statistical Field Theory: An Introduction to Exactly Solved Models in

Statistical Physics, Oxford University Press, (2010).
12
2 Brownian motions and random paths
The aim of this chapter is to introduce simple but fundamental random objects and to
get acquainted with their path integral representations. Both —the objects and the path
integrals— are baby examples of structures involved in statistical field theory. Illustrating
such structures in the case of stochastic processes yields tools which find applications in
field theory but also in other scientific domains. We are going to start with random walks
and random paths defined over discrete lattices and then take the limit to continuous
spaces. This will allow us to introduce the important notion of scaling limits that we will
be encountered many times in the following.
2.1 Random walks and random paths

We start by discussing simple examples of random paths and random curves. The simplest
is of course that of random walks which we will later use to define the Brownian motion,
one of the basic objects random geometry deals with. We will introduce a few other
examples whose statistics are coded into sums over paths.
• Random walks: basics.
For simplicity let us (first) consider random walks on Z (generalizations are easy). Let
Λ ≡ aZ be a one-dimensional lattice (alias the line) with edges of size a. We consider
symmetric walks, with equal probability to move to the left and to the right. Let q0 be the
initial walker’s position and Qn be its position after n step. The rules are Qn → Qn+1 =
Qn + an+1 with n+1 = ± with probability 21 . Hence,
n−1
X
Qn = q0 + a k+1 ,
k=0
with k independent and identically distributed random variables (i.i.d.) with P[k = ±] =
1
2 . The mean and the covariance of the position Qn are easy to compute (set q0 = 0 for
simplicity): X X
E[Qn ] = 0, E[Q2n ] = a2 E[ k l ] = a2 E[ 2k ] = a2 n.
k,l k
This last equation tells us that the typical displacement scales as the square of the number
√
of steps: ` ' n. We can compute other expectations, for instance the two point functions
which is given by:
n X
m min(n,m)
X X
2 2
E[Qn Qm ] = a E[ k l ] = a E[ 2k ] = a2 min(n, m).
k=1 l=1 k=1
Note that Qn is not a Gaussian variable but a sum of binomial variables.
13
We can of course consider more general random walks, say asymmetric walks with
P[k = +] = p and P[k = −] = q with p + q = 1. Then (for q0 = 0) the mean
is E[Qn ] = a(p − q)n and the variance E[Q2n ] − E[Qn ]2 = a2 4pq n, so that the walk
displacement is subject to a drift. We can also imagine random walks in inhomogeneous
(potential) landscape (i.e. with the probability to move to the left or to the right depending
on the instantaneous position of the walker).
• Interlude: sum over paths.
Expectation values of functions of random walks are tautologically represented as (dis-
crete) path sums or integrals. A walk ω, say of length N , can be coded with its se-
ries of steps ω = (1 , · · · , N ) or by its successive positions ω = (q0 , q1 , · · · , qN ) with
qk − qk−1 = k . Observables may for instance be the positions at different times whose
correlation functions are the multipoint functions E[Qk1 · · · Qkp ]. Let f be such a (general)
observable, i.e. a function defined over the space of walks: f : ω → f (ω). Its expecta-
tion reads (as usual in probability theory, or in statistical physics dealing with Boltzmann
sums) X
E[f ] = pω f (ω),
ω: path
where pω is the probability of occurence of the walk ω. For random walks, pω = p#right q #left .
This is a sum over configurations (with Boltzmann weights proportional to pω ) or a sum of
discrete paths (with a weight assign to each path). Path integrals for Brownian motion, for
stochastic processes, etc., or for field theories will be ‘formal’ extension this construction
to continuous paths (which are maps from some interval to some manifolds) or to higher
dimensional analog maps from manifolds.
Of course we are free to change the measure on path pω (if it yields interesting enough
models). For instance we can choose to weight the paths according to their lengths, say
pω ∝ µ|ω| with µ > 0 and |ω| the number of the steps of the walk ω. This ensemble is
called that of free random paths, and we will discuss it more in details in a little while. We
can either consider the canonical or grand-canonical ensembles, depending whether we fix
the length of the curves or not. For µ = 1/2 (in D = 1) this ensemble of walks is that of
standard symmetric random walks with p = q = 1/2. This alternative formulation offers
possible extensions. For instance, we can also look at curves with free ends or connecting
two preselected end points. We can then look at the continuous (scaling) limit of those
models, etc.
We can also restrict the set of paths that are sampled. For instance, we may demand
that the paths are not self-touching, and hence also not self-intersecting. Such paths
are called self avoiding walks and we make them random by equipping them with the
probability measure pω ∝ µ|ω| . Self avoiding walks are sometimes used as a model for
polymer physics. We leave all of these as exercises and, in this Section, we shall concentrate
ourselves on the scaling limit of the random walks and on the Brownian motion.
2.2 Scaling limits

We give here two presentations of the Brownian motion: via a continuous limit of random
walks or via its characteristic properties. We also discuss a scaling limit of the free random
14
paths defined above which will turn out to be related to the Gaussian free field of statistical
field theory. The notion of scaling limits is very important: it encompasses the passage
from a discrete model, defined on a lattice, to a continuous model. Making sure that this
limit exits requires ensuring that the random objects under study should macroscopic sizes
even when the mesh of the underlying lattice shrinks.
• Brownian motion: basics.
We define the Brownian motion from a scaling limit of random walks. We want to
take the limit of random walks with the lattice spacing going to zero, i.e. a → 0. Since
E[Q2n ] = a2 n, we see that we have to simultaneously take the limit n → ∞ keep the
product a2 n fixed in order to get a non trivial result. Let t = a2 n be called the time.
Then the Brownian motion is defined by:
[t/a2 ]
X
Wt := lim Qn = lim a k ,
a→0, n→∞ a→0
a2 n=t k=1
with [t/a2 ] the integer part of t/a2 . This is a scaling limit (with scaling dimensions assign
to physical quantities). From the previous computation of the mean displacement of the
variance of random walks we have
E[Wt ] = 0, E[Wt Ws ] = min(t, s).
In particular E[(Wt − Ws )2 ] = |t − s|. The central limit theorem also implies that Wt is a
Gaussian variable. Indeed, a direct application of this theorem (see the exercise Section)
gives:
z2
E eizWt = e−t 2 .

The Brownian motion possesses a few remarkable properties, some of which are directly
inherited from those of random walks:
— Markov property, in the sense that the future of a Brownian trajectory after a time s
is independent of the past but only dependent of the present position Ws ;
— Identically distributed increments, in the sense that the increments Wt+s − Ws and
Wt − W0 , for any t and s, have identical distributions. They are also independent from
the past. These two properties are inherited from similar properties which hold true for
random walks;
— Translation invariance, in the sense that the increments of Wt and Wt + a have the
same law;
— Dilatation invariance, in the sense that Wt and λ−1 Wλ2 t have the same law (notice the
change in time).
It is worth knowing that these properties, together with continuity of the curves t →
Wt are characteristic of the Brownian motion. A Brownian path is almost surely not
differentiable.
The generalization to higher dimensions in RD is simple: one takes D copies of inde-
pendent Brownian motions in correspondance with each of the D-orthogonal direction:
~ t = (W 1 , · · · , W D ),
W t t
15
with Wti , i = 1, · · · , D, independent identical distributed (i.i.d.) Brownian motions. The

D-dimension Brownian motion possesses all the properties of the 1D Brownian motion
(translation and dilatation invariance, Markov properties, i.i.d. of the increments) plus
rotation invariance W ~t →R·W ~ t with R ∈ SO(D) in dimension D ≥ 2.
• Scaling limit of asymmetric random walks

We now look how to take the continuum limit of asymmetric random walks with
different probability to step to the left or to the right. That is: asymmetric walks with
P[k = +] = p and P[k = −] = q. Since E[Qn ] = a(p − q)n, the continuous limit under
the Brownian scaling a2 n = t exists only if the probability p and q depend appropriately
on the mesh a. We should have (p − q) = aν with ν fixed, or equivalently p = 21 (1 + νa)
and q = 12 (1 − νa). Then,
E[Qn ] = a(p − q)n ' ν t, E[Q2n ] − E[Qn ]2 = a2 4pq n ' t.
As in the symmetric case, we can thus define the limit of the asymmetric walks by
1 1
Xt = lim Qn , with p = (1 + νa), q = (1 − νa).
a→0, n→∞ 2 2
a2 n=t
By the central limit theorem again, Xt is a Gaussian process with mean νt and covariance
t:
z2
E eizXt = e−t 2 +iνt .

There is a constant drift, with velocity ν, in the sense that it satisfies
dXt := Xt+dt − Xt = ν dt + dWt ,
with Wt a Brownian motion. This type of noisy evolution equation is an example of

stochastic differential equations.
Notice that we could have chosen to make the probabilities to jump to the left or to
the right inhomogeneous, so that the drift ν would also be inhomogeneous, position space
dependent, say ν = u(x). Then we would have got a more general stochastic differential
equation of the form dXt = u(Xt ) dt + dWt . We will come back to stochastic differential
equation below.
The noticeable point about this limit is that we have to let the parameters p and q
of the discretized microscopic model to approach the ‘critical’ value 12 appropriately when
a → 0 for the continuous limit to exist. This is a simple instance of what is called the
scaling limit. This is a typical procedure for extracting non-trivial continuous theories
from discrete versions thereof.
• Scaling limit of free random paths
We now describe the scaling limit of free paths and its connexion with what you will
later recognise as Gaussian free field. It also provides a very simple statistical model of
random geometrical structures, namely curves.
Recall that free random paths were previously defined as the statistical ensemble of
curves weighted according to their lengths. Let us be a bit more precise. Let Λ be a D
16
dimensional square lattice with mesh size a: Λ = (aZ)D . We let ej , jP= 1, · · · , D, be

a basis of orthonormal vectors in RD , so that points x ∈ Λ are x = a j nj ej with nj
integers. We shall deal with paths Γ starting at the origin 0 and ending at fixed point x,
i.e. paths Γ = (x0 , · · · , xi , · · · , x|Γ| ) with x0 = 0 and x|Γ| = x. The successive points in Γ
have to be lattice neighbors so that the segment [xi , xi+1 ] from an edge of the lattice.
The statistical ensemble of free random paths is defined as the set of random paths
drawn on Λ whose statistics is specified by assigning a Boltzmann weight wΓ := µ|Γ| to
each path Γ, with µ a real number (µ > 0) and |Γ| be the number of bonds of the path Γ,
so that a|Γ| is its length. The parameter µ is often called the fugacity. The probability of
a given path Γ from 0 to x is thus |Γ|
P µ /Z(x) |Γ|
with Z(x) the partition function conditioned
on paths from 0 to x: Z(x) = Γ: 0→x µ .
As usual in statistical physics, the partition function is a generating function for the
number of configuration of a given ‘energy’ —here of a given length. Thus, by construction,
X X X
Z(x) = µ|Γ| = µN WNfree (x) = δx;0 + µN WNfree (x),
Γ: 0→x N ≥0 N >0
with WNfree (x) the number of paths from 0 to x with N bonds. We note that WNfree =0 (x) =
δx;0 because a path of zero length does not escape from its original position.
We are now aiming at understanding the properties of this statistical ensemble of
curves and their limit in continuous space obtained by sending the mesh of the lattice to
zero.
Because paths arriving at a given point have to visit one of its neighbor points before
reaching its destination (if it has more than one bond) the partition function Z(x) has to
satisfy the following difference equation:
D
X
Z(x) = δx;0 + µ Z(x + aej ) + Z(x − aej ) .
j=1
This is a linear equation, involving the discrete Laplacian on Λ. As a consequence, it can

be solved by Fourier transform giving
dD k eik·x
Z
Z(x) = D
P .
BZ (2π/a) 1 − 2µ j cos(ak · ej )
with BZ the Brillouin zone of the square lattice: BZ:= [− πa , πa ]D . This explicit expression
makes clear that Z(x) converges if the fugacity is smaller than a critical value: i.e. it
converges for |µ| < µc . This value actually depends on which lattice the model is defined
but µc = 1/2D is for Λ the square lattice. The partition function Z(x) has a simple
pole singularity near the critical fugacity, i.e. Z(x) ' const.
µc −µ . This singular behaviour
codes for the asymptotic number of curves from 0 to x with a large number N of bonds:
WNfree (x) ' const. µ−N
c at large N .
So good so far for the discrete model. Let us now look at its continuous limit a → 0.
The first naive attempt would be to take this limit keeping all other parameters fixed,
say µ, x fixed. Then, the above equation for Z(x) implies that (1 − 2Dµ)Z(x) ' δx;0 as
17
a → 0, and hence that Z(x) is trivial in this limit (unless µ = µc ) in the sense that all
geometrical objects it codes for have collapse around the origin.
That is: we learn that, if we keep all parameters fixed, the naive continuous limit do
not describe extended structures in space. We have to adjust the parameters (here the
fugacity µ while taking the continuous limit). This is called the ‘scaling limit’. Letting
µ → µc put more and more weights on long paths and gives a chance to get a meaningful
continuous limit of extended geometrical macroscopic objects.
Let us thus look at the continuous limit a → 0 but adjusting µ → µc as a function of
a. Expanding the discrete equation for Z in power of a yields (recall that µ−1
c = 2D)
(µ−1 − µ−1 −1 2
c )Z(x) = µ δx;0 + a (∆x Z)(x) + · · · ,
with ∆x the Laplacian in RD . Thus, we see that to get a non trivial limit we have to let
µ−1 − µc−1 = a2 m2 as a → 0 with the parameter m fixed (m has the dimension of the
inverse of a length). Let us then define the scaling limit of Z(x) by
1
G(x) := lim a2−D Z(x),
2D a→0
with the limit understood as the scaling limit with µ−1 − µ−1 2 2
c = a m . By construction,
it is non-trivial and it satisfies
− ∆x + m2 G(x) = δ(x),

where we use the fact that a−D δx;0 → δ(x), the Dirac measure centred at the origin, as
a → 0. Notice that we had to incorporate an extra factor a−D in G(x) to get a meaningful
continuous limit, so that G(x) has the dimension of the (2 − D)-th power of a length —we
shall say that its scaling dimension is D − 2—, whereas Z(x) was initially dimensionless.
The scaling function G(x) is thus the Green function of the massive Laplacian in RD . Its
explicit expression
dD k eik·x
Z
G(x) = ,
(2π)D m2 + k 2
can either be found by taking the limit of the explicit formula for Z(x) or by solving the
differential equation it satisfies. As we shall see in a few Chapters, G(x) is the two point
function for a Gaussian free field.
Let us insist on the notion of scaling limit: to get a meaningful continuous limit we
had to let the parameter µ of the model to approach its critical value µc in a specific way
as a function of the lattice cutoff.
2.3 Brownian motions and path integrals

Here, we define and compute the Brownian transition kernel. This will then be used to
give a path integral representation of the Brownian motion. This type of path integrals is
a baby exemple of those we will encounter later when dealing with random fields (it is also
a baby exemple of (Euclidean version of) Feynman path integrals in Quantum Mechanics).
18
• Brownian motion: its transition kernel.

We now characterize and determine the transition kernel Pt (x0 , x) dx which is the
probability for a Brownian motion, started at position x0 , to be at point x, up dx, at time
t. We will get the standard result that
dx (x − x)2
0
Pt [x0 ; x] dx = √ exp − . (1)
2πt 2t
Note that this is the solution of the heat equation ∂t − 21 ∂x20 Pt [x0 ; x] = 0 with initial

condition Pt=0 [x0 ; x] = δ(x0 − x). The Markov property implies that it satisfies the
convolution property Z
Pt+t0 [x0 ; x] = dy Pt0 [x0 ; y]Pt [y; x].
The Brownian process is indeed Gaussian with i.i.d. increments.

More precisely, let us compute the probability Pt [x0 ; Ω] for a Brownian motion started
at x0 to be in the interval Ω at time t, i.e. the probability that Wt ∈ Ω. Let us start with
the discretizing model and look for the probability Pn [x0 ; Ω] for a random walker to be in
Ω after n step staring at x0 . Since the walker would have done a first step either to the
left or to the right with probability 1/2, we have:
1
Pn+1 [x0 ; Ω] = Pn [x0 + a; Ω] + Pn [x0 − a; Ω] ,
2
In the continuous limit a → 0, n → ∞, with t = a2 n, we get:
Pn+1 [x0 ; Ω] − Pn [x0 ; Ω] ' a2 ∂t Pt [x0 ; Ω]

Pn [x0 + a; Ω] + Pn [x0 − a; Ω] − 2Pn [x0 ; Ω] ' a2 ∂x20 Pt [x0 ; Ω]

The probability Pt [x; Ω] is therefore solution of the heat equation:

1
∂t Pt [x0 ; Ω] = ∂x20 Pt [x0 ; Ω].
2
If Ω is small and centered around a point x we may write Pt [x; Ω] as a probability density
Pt [x0 ; x] dx, and for a domain ω of arbitrary size we have:
Z
Pt [x0 ; Ω] = dx Pt [x0 ; x]
Ω
By linearity, the probability density is also solution of the heat equation with initial
condition Pt=0 [x0 ; x] = δ(x0 − x).
• The path integral representation of the Brownian motion.
Brownian motion are random curves. We aim at representing their probability measure
as an integral sum over paths. Consider the Brownian motion on the time interval [0, t]
starting at initial point xi = x0 . We are going to show that its transition kernel from xi
to final point xf admits the following path integral representation
Z
1 t 2
R
Pt (xi , xf ) = x(0)=x [Dx] e− 2 0 ds ẋs , (2)
i
x(t)=xf
19
R
where the symbols [Dx]... “means” the integral sum over continuous paths.
This may sounds a bizarre definition because the integral over continuous paths does
not seem to be well-defined (at least we didn’t defined it) and furthermore the Rderivative
ẋs does not exist for the Brownian motion so that the meaning of the integrals dx ẋ2 in
unclear (it depends on which curves/paths the measure [Dx] is supported). Path integrals
are almost, if not always, not well defined mathematically except in few cases. The
Brownian motion is such exceptional case: the Brownian path integral is “tautologically”
defined by: Z
1
Rt
ds ẋ2s
[Dx] e− 2 0 (· · · ) := E[ (· · · ) ],
with (here) E[· ·R· ] denoting the Brownian expectations. Alternatively, we may formally
1 t 2
write: [Dx] e− 2 0 ds ẋs = dP, with P the Brownian measure which is a mathematically well
defined notion.
The construction of the path integral goes by discretizing the time interval [0, t] and
taking a limit when the mesh of this discretization goes to zero (hence the number of
intermediate points becomes infinite). Consider a Brownian curve on interval [0, t] and look
at the transition kernel Pt (xi , xf ). We use a trick consisting in introducing successive cuts
and intermediate points. Let us first introduce a single intermediate point t0 , 0 < t0 < t.
Thanks to the convolution property of the transition kernel we may write (See Figure)
Z
Pt (xi , xf ) = dx0 Pt0 (xi , x0 )Pt−t0 (x0 , xf ).
By iterating and dividing the interval [0, t] in N sub-intervals [kδ, (k + 1)δ], for k =
0, · · · , N − 1, of equal length δ := t/N , we write
Z NY−1
Pt (xi , xf ) = [ dxk ] Pδ (xi , x1 ) · · · Pδ (xN −1 , xf ).
k=1
(x−y)2
Using Pδ (x, y) = √1
2πδ
e− 2δ , this can equivalently be written as (with x0 = xi and
xN = xf )
−1
NY
Z
dx (N )
Pt (xi , xf ) = √ k e−S ,
k=1
2πδ
1 PN −1
where the function S (N ) , call the discretized action, is S (N ) = 2δ 2
k=0 (xk+1 − xk ) .
Up to this stage the formula is exact. We have simply sampled the Brownian curve
at a large, but finite, number of points. The path integral formulation comes from taking
the limit N → ∞ and thus δ = t/N → 0. By definition the path integral representation
is set to be
−1 dxk −S (N )
Z Z NY
−S[x]
[Dx] e := lim √ e .
N →∞ 2πδ
k=1
P −1 xk+1 −xk 2
Now the discretized action can alternatively be written as S (N ) = δ N k=0 δ .
xk+1 −xk 2 2
The terms ( δ ) are the discretized version of the velocity ẋ , and the formal limit
20
of S (N ) is:
N −1
xk+1 − xk 2 1 t
X Z
S (N ) = δ −→ S[x] := ds ẋ2s .
δ 2 0
k=0
Q −1 dxk
Denoting by [Dx] the formal limit of [ N k=0
√
2πδ
] when the number of sampled points
increase, we get (by definition) the path representation of the transition kernel
1 t
Z Z
−S[x]
Pt (xi , xf ) = [Dx] e , S[x] = ds ẋ2s ,
2 0
with the boundary condition x(0) = xi and x(t) = xf .
2.4 The 2D Brownian motion

The 2D Brownian motion is the simple two component generalization of the 1D Brownian
motion. Besides the basic properties and symmetries (translation, rotation, dilatation
invariances) it possesses an extra remarkable property: it is conformally invariant.
The conformal invariance of 2D Brownian trajectories has first been understood by
P. Lévy. This invariance concerns the set of traces formed by these trajectories, i.e.
their graphs, it does not concern the time parametrized trajectories. These traces are
not invariant realisation by realisation, independently, but their statistical ensemble is
conformal invariant. This invariance manifests itself in the following way. Let us imagine
sampling a Brownian motion in the unit disk started from the origin and stopped the first
instance it touches the boundary of the disk. By conformal transformation, these curves
can be transported in another planar domain having a topology identical to that of the
disk, for example a rectangle. One then obtains a statistical ensemble of curves inscribed
in this new planar domain. Conformal invariance asserts that this set of curves possesses
a statistical distribution identical to that of the Brownian motion started from a point in
the new domain (the image of the origin by the conformal map) and stopped as soon as it
touches the boundary of this new domain. That is: we cannot distinguish the Brownian
curves transported by conformal transformations to the Brownian curves themselves.
• 2D conformal transformations.
Conformal transformations in 2D are geometric transformations that preserve angles.
Let (x, y) ∈ R2 be the coordinates of a point in the plane and z = x + iy its complex
coordinate. Conformal transformations are locally holomorphic transformations:
z → w = f (z)
Let z1 = z0 + δz1 and z2 = z0 + δz2 two neighbour points of z0 . The two small vectors
δz1 and δz2 joining z0 to z1 or z2 are transformed into two vectors δw1 and δw2 joining
w0 = f (z0 ) to w1 = f (z1 ) or w2 = f (z2 ). To first order, we have:
δw1 = f 0 (z0 ) δz1 , δw2 = f 0 (z0 ) δz2 .
The angle between the vectors δw1 and δw2 is therefore identical to that between the
vectors δz1 and δz2 . Each of these vectors has been rotated, by an angle equal to the
argument of f 0 (z0 ), and has been dilated proportionally to the modulus of f 0 (z0 ).
21
• Conformal invariance of the 2D Brownian motion.

Lévy’s argument consists in promoting the global invariance by rotation and dilatation
into a local invariance (hence, the conformal invariance) using the locality properties of
the Brownian motion. This principle has a domain of applicability that extends beyond
the study of Brownian curves. Let us consider a Brownian curve stopped at time T .
Let us divide the time intervalle [0, T ] in a large number N of intervalles [ti , ti+1 ] with
0 = t0 < t1 < · · · < tN = T , and decompose the trajectory as the sum of all its increments
between these successive times:
Xt − X0 = (Xt1 − Xt0 ) + (Xt2 − Xt1 ) + · · · + (Xt − Xtj−1 ),
for t ∈ [tj−1 , tj ]. All the increments Xtj+1 − Xtj are statistically independent and dis-
tributed identically to Xtj+1 −tj . This decomposition means that Brownian curves can be
reconstructed by concatenating its increments.
Let us now transform each of these increments by a rotation Rj and by a dilatation
with scale factor λj , which may vary from one increment to the next. We get:
(Xtj+1 − Xtj ) → λj Rj · (Xtj+1 − Xtj ) ≡in law Xλ2 (tj+1 −tj )

j
where in the last equivalence we used the global invariance of the Brownian motion. Thus,
after concatenation of the rotated and dilated increments, we obtain curves with the same
statistic as the Brownian curves but with a different temporal parameterization (since the
time increments tj+1 − tj have been transformed into λ2j (tj+1 − tj )). Arbitrary conformal
transformations can be applied to the Brownian curves by naively taking a limit where
the above discretization is increasingly fine.
This argument, which can be made rigorous, indicates that the image of a Brownian
curve by a conformal transformation is yet another Brownian curve up to time reparame-
terization. The conformal invariance of Brownian curves is at the origin of many of their
peculiar properties. For instance, it has recently been proved that the exterior perimeter
of the graph a Brownian trajectory is a fractal curve of dimension 4/3.
To conclude, let us note that this property is based on the following principle: “Global
invariance under dilatations and rotations plus locality implies invariance under local
dilations and rotations, i.e. under conformal transformations”. This is a principle that
find applications in many physical systems.
2.5 Brownian motions and stochastic differential equations

Stochastic differential equations provide a framework to model and analyse the effects of
noise or random external forces on dynamical systems. In the physics literature they are
often represented as Langevin-type equations of the form
Ẋt = a(Xt ) + ξt ,
with a(x) some function (or vector field) driving the dynamical variable X in absence of
noise and ξt representing the noise with a parameter coding for the amplitude of that
noise. Depending on the physical problem various assumption can be made about the
noise statistics, e.g, it can be correlated or uncorrelated in time.
22
• Stochastic differential equations: discrete and continuous.

The simplest case —but nevertheless quite rich— amounts to assume that ξt is a
white-in-time Gaussian noise with zero mean and covariance E[ξt ξs ] = δ(t − s). The δ-
correlation codes for the absence of memory so that ξt is a highly fluctuating quantity.
Since the process ξt is ill-defined mathematically, we shall write the above equation is a
slightly different form:
dXt = a(Xt )dt + dBt , (3)
with dXt = Xt+dt − Xt and Bt a (normalized) Brownian motion with zero mean and
covariance E[Bt Bs ] = min(t, s). In particular, dBt = Bt+dt − Bt and E[dBt2 ] = dt. The
parameter has been introduced by convenience to scale the amplitude of the noise. Since
ξt can be viewed as the time derivative of the Brownian motion —which, mathematically
speaking, almost surely does not exists— these are two (equivalent) ways of writing the
same equation. Of course one can also consider more general SDEs, say with more vari-
ables. A proper definition
Rt of solution of this equation is the integrated version of the SDE
in the form Xt = 0 a(Xs )ds + Bt .
The process Xt defined by the SDEs dXt = a(Xt )dt + dBt may be viewed as the
scaling limit of asymmetric random walks but space dependent with probabilities to move
to the left or right, so that they induce a space dependent drift. Concretely one consider
the scaling limit of asymmetric random walks, defined on the lattice hZ with mesh size h,
and probability p(x) (resp. q(x) to move to the right (resp. to the left) scaling with the
lattice mesh as p(x) = 21 (1 + a(x)h) and p(x) = 12 (1 − a(x)h) as h → 0.
We may also look at more general SDEs of the form dXt = a(Xt )dt+b(Xt )dBt with b(x)
a non-constant function. Dealing with them however requires some care as the irregularity
of the Brownian motion as function of t demands to make precise what is meant by the
product b(Xt )dBt . This will be briefly discussed below.
A discretized version of this equation is (with δ := δt the elementary time step):
Xn+1 − Xn = δ a(Xn ) + ξn+1 . (4)
The ξk are Gaussian i.i.d. with zero mean, E[ξk ]√= 0 and covariance E[ξk ξl ] = δk;l δ. It is
important to note that ξk are typically of order δ. The integrated version is
n−1
X n
X
Xn − X0 = δ a(Xk ) + ξk .
k=0 k=1
Recall that Bn = nk=1 ξk is a discretized version of the Brownian motion, and we may
P
view ξn as discrete time derivative, ξn+1 = Bn+1 − Bn . The form of this discretization, in
which we have chosen to sample a(X) at the starting point of the discretized intervals, is
called the Itô convention.
• Transition kernels and the Fokker-Planck equation.
We are interested in the transition kernel dx Pt (x0 , x) or dx Pt (x0 → x) which is the
probability to be at position x up to dx at time t starting from point x0 . It is such that
Z
Ex0 [F (Xt )] = dx Pt (x0 , x) F (x),
23
for any test function F . Another (formal) way to define it uses Dirac delta function:
Pt (x0 , x) = Ex0 [δ(Xt − x)]. By the Markov property the kernel satisfies the composition
property: Z
Pt+s (x0 , x) = dy Pt (x0 , y)Ps (y, x).
It defines a semi-group. The transition kernel transports probability distributions (by

duality). If dx0 Q0 (x0 ) is the initial distribution, then its transport at time t is
Z
Qt (x) = dx0 Q0 (x0 ) Pt (x0 , x).
R
Similar formulas apply in the discrete version: Ex0 [F (Xn )] = dx Pn (x0 , x) F (x), etc.
We are know going to prove that the transition kernel is solution of a partial differential
equation (similar to the Schrödinger equation).
2
∂t Pt (x0 , x) = H · Pt (x0 , x), H := ∂x ∂x − a(x) . (5)
2
This is called the Fokker-Planck equation. Notice its interpretation as a conservation
2
law for probabilities: ∂t Pt (x0 , x) = ∂x Jt (x), with probability current Jt (x) := ( 2 ∂x −
a(x))Pt (x0 , x). We can write Pt = etH . By construction H is a non positive operator.
This operator actually fully determine the process: its data is equivalent to that of the
process. At large time Pt converges toward the stationary measure if it exists. The
approach to stationary is governed by the spectrum of H. Note the analogy with quantum
mechanics.
The invariant measure Pinv (x) should satisfies H · Pinv = 0. A solution with vanishing
2
probability current should satisfies ( 2 ∂x − a(x))Pinv (x) = 0. Let us set a(x) = −U 0 (x)
(this is always possible in 1D, with one variable x, but not in higher dimension). The
function U is called the potential. Then, if normalizable, the invariant measure (with zero
probability flux) is:
2
Pinv (x) = Z −1 e−2U (x)/ ,
R
with the normalization constant Z such that dx Pinv (x) = 1 (if normalizable). This is
Boltzmann
√ distribution for a potential U (x) and temperature 2 /2, so we can set =
2kB T .
Let us now derive the differential equation satisfied by the kernel. We compute the
conditioned expectation E[F (Xn+1 )|Fn ], conditioned on the knowledge of the process up
to step n. We have to compute E[F (Xn +δAn + ξn+1 |Fn ] with An = a(Xn ). Conditioning
on Fn means that Xn , An are fixed (non random) number in this expectation, so that the
expectation is only about the gaussian variable
√ ξn+1 . We compute by expanding in power
of δ (recall that ξk are typically of order δ):
E[F (Xn + δAn + ξn+1 )|Fn ]

2 00
= F (Xn ) + δ An F 0 (Xn ) + F 0 (Xn )E[ξn+1 ] + 2
F (Xn )E[ξn+1 ] + O(δ 3/2 ),
2
2 00
= F (Xn ) + δ An F 0 (Xn ) +

F (Xn ) + · · ·
2
24
The r.h.s. depends only on data up to step n (i.e. it is Fn -measurable) as it should be.
As a consequence
2 00
E[F (Xn+1 ] = E[F (Xn ) + δ a(Xn ) F 0 (Xn ) +

F (Xn ) ] + · · ·
2
Up to an integration by part, this is equivalent to (recall that δ is the time step δt)
2 2
∂x − ∂x a(x))Pn (x0 , x) + O(δ 2 ).

Pn+1 (x0 , x) − Pn (x0 , x) = δ
2
Hence we get the Fokker-Planck equation in the continuum limit.
We could also have derived a differential equation for Pt (x0 , x) but acting the initial
2
point x0 . This equation is the dual of the one acting x, that it ∂t Pt (x0 , x) = 2 ∂x0 +
a(x0 ) ∂x0 Pt (x0 , x). As for the Brownian motion, it can be derived by looking at what
happened after one step but taking into account that the probabilities to move to the
right or to the left of space dependent with the appropriate scaling as the mesh of the
lattice goes to zero.
• Itô versus Stratonovich.
We can of course consider more general SDEs —and those naturally occur in physical
problems. Instead of Sees of the form dXt = a(Xt )dt + dBt we can more generally look
at SDEs
dXt = a(Xt )dt + b(Xt )dBt ,
with a(x) and b(x) smooth function and with Bt a Brownian motion, dBt2 = dt. A proper
definition of solution of this equation is in the integrated version:
Z t Z t
Xt = a(Xs )ds + b(Xs )dBs .
0 0
We have to be careful how we define the integrals, in particular the stochastic integral.
Let us defined them by discretization. Different conventions differ from where we sample
the function in the Riemann sum. Two important conventions are the so-called Itô or
Stratonovich convention, with ξk+1 = (Bk+1 − Bk ):
n−1
X X
Ito : Xn = δ a(Xk ) + b(Xk )ξk+1 ,
k=0 k=0
n−1
X 1 X1
Stratonovich : Xn = δ (a(Xk+1 + a(Xk )) + (b(Xk+1 ) + b(Xk ))ξk+1 ,
2 2
k=0 k=0
in the large N limit with δ = 1/N and nδ = t. With the Itô convention we sample
the function at the beginning of the sub-intervals, with the Stratonovich convention with
sample it with the mean of its value at both ends of the sub-intervals. Of course there are
intermediate alternative choices in which we weight differently the sums. The important
property of Itô convention is that ξk+1 = Bk+1 − Bk is independent of b(Xk ) and all b(Xl ),
l ≤ k.
25
Itô and Stratonovich convention for stochastic integrals are related by

Z Z Z T
1
b(Bs ) dBs |Ito = b(Bs ) dBs |Strato − b0 (Bs ) ds.
2 0
This equation is simply checked in the case b(B) = B. The simplest way to check it is
RT RT
to verify it for Bt dBt . One has 0 Bt dBt |Ito = 12 BT2 − 12 T and 0 Bt dBt |Strato = 12 BT2 .
R
Here is a more generalR but naive proof. Let I be the difference between the Itô minus
Stratonovich integrals b(Bs )dBs . We have:
1X
I = − (b(Bk ) − b(Bk−1 )) (Bk − Bk−1 ),
2
k
Z
1X 0 1X 0 1
' − 2
b (B̂k−1 ) (Bk − Bk−1 ) ' − b (B̂k−1 ) (δt)k = − b0 (Bs ) ds.
2 2 2
k k
The Itô calculus is very efficient “computationally” because (i) it is simply based on
Taylor expansions in dBt and dt with the rule dBt2 = dt, and (ii) in product of the form
b(Bt )dBt the random quantities b(Bt ) and dBt are independent. In particular, for any
regular enough function F we have
1
dF (Bt ) =Ito F 0 (Bt )dBt + F 00 (Bt )dt.
2
In the Stratonovich convention the formula will be dF (Bt ) =Strato F 0 (Bt )dBt as with
Leibniz rules but the quantities F 0 (Bt ) and dBt are there not independent.
2.6 Path integral representation for SDEs

The path integral representation (also called the MSR representation) of a SDE of one
variable of the form dXt = a(Xt )dt + dBt with dBt2 = dt is specified by the action (in
the Itô convention):
t
(ẋs − a(xs ))2
Z
S[x] = ds (6)
0 22
Here ẋs := ∂s xs . The action codes for the Boltzmann weight e−S in the path integral
measure. This means that expectation of function F (Xt ) testing the process can be written
as path integral. For instance the probability Pt (x0 , x) to go from x0 to x can be written
as
Z
Pt (x0 , x) = x(0)=x [Dx] e−S[x] (7)
0
x(t)=x
We shall discuss some property of this representation and use it in the following section
to derive exact —and non perturbative— results in the small noise limit.
We are going to present two derivations: the first using a discretization of the SDE,
the second directly using the continuous formulation.
26
• An approach by discretization.
With the Itô convention, the discrete form of the SDE reads
Xn+1 − Xn = δ a(Xn ) + ξn+1 ,
with δ the small time interval, δ = dt, and ξk ’s Gaussian i.i.d. with covariance E[ξk2 ] = δ,
with initial data X0 = x0 . Imagine we aim at computing the expectation E[F (X1 , · · · , XN )]
testing the process at different times. The expectation is with respect to the Gaussian
noise ξk so that it corresponds to the measure (up to a proportionality coefficient):
N −1
Y 2
dξk e−ξk+1 /2δ
k=0
Now, given the ξk ’s we have to recursively solve the difference equations xk+1 − xk =
δ a(xk ) + ξk+1 for k = 0 up to N − 1. These are recursive equations (in the sense that
given the n first xk ’s we can solve for xn+1 ). At this point there are two possible routes:
either we integrate directly on the ξk ’s, or we implement these constraints by introducing
series of δ-functions in the integrals. We choose the second way because it has a wider
application (in more general context). So let us enforce these constraints by inserting the
following series of integrated δ-functions:
N
Y −1

[Dx] δ xk+1 − xk − δ a(xk ) − ξk+1 .
k=0
Because the constraints are recursively linear in xk+1 , their implementations do not in-
volve any Jacobian. This would not have been true if we would have chosen another
discretization form (i.e. it is only true for the Itô prescription). We can then represent
these delta-functions via their Fourrier transforms: Fourier integral:
Z
dφk+1 −iφk+1 (xk+1 −xk −δ a(xk )− ξk+1 )
δ(xk+1 − xk − δ a(xk ) − ξk+1 ) = e .
2π
Gathering, we get the measure [ k dξk dφk dxk ] e−S with action
Q
N −1 N −1
1 X 2 X
S= ξk+1 + i φk+1 (xk+1 − xk − δak − ξk+1 ).
2δ
k=0 k=0
Doing successively the Gaussian integrals over the ξ’s first and the φ’s yields:
N −1 N −1
δ 2 X 2 X
S ≡ φk+1 + i φk+1 (xk+1 − xk − δak )
2
k=0 k=0
N −1
X (xk+1 − xk − δak )2
≡ ,
2δ 2
k=0
27
By convention notation ≡ means “equivalent up to Gaussian integration”. The integra-

tions are over ξ1 , · · · , ξN , and φ1 , · · · , φN but x1 , · · · , xN −1 with x0 and xN = xf fixed.
The last form can be expanded into
N −1 N −1 N −1
1 X (xk+1 − xk )2 X δ X 2
S= − a k (x k+1 − xk ) + ak
2 2δ 2
k=0 k=0 k=0
1
first term is the Gaussian Brownian measureR 2 ds ẋ2s , the second is the Itô integral
R
The
as dxs and the third is the Riemann integral 12 ds a2s . Taking the large N limit we get
R
the action:
Z Z Z
1 1 2 1
ds a2s

S[x] = 2 ds ẋs − as dxs +
2 2
Z T 2
(ẋs − a(xs ))
= ds
0 22
Note that the formula for the action depends on the fact that we use the Itô convention.
It will also be different if we were dealing with a SDE of the form dXt = a(Xt )dt+b(Xt )dBt .
The construction is slightly different if we use the Stratonovich convention.
• A continuous description.
Here we do the same computation in the ‘physicist’s way’ directly in the continuous
formulation. This will give us the opportunity to become acquainted with formal manip-
ulations of path-integrals. We start again from the SDE
dXt = a(Xt )dt + dWt ,
with Wt a normalized Brownian motion with dWt2 = dt. Let us code this equation into
F(xt ) := Rẋt − a(xt ) − ẇt . We aim at integrating over ẇt , with the Brownian measure
1 2
[Dw] e− 2 ds ẇs , with the contraint that xt is the solution (supposed to be unique) of
the equation F(xt ) = 0. We have Q to enforce this constraint by inserting in the path-
integral a multi-variable δ-function t δ(xt − x∗t ) with x∗t the solution of the differential
equation. Recall the properties of Dirac delta function of single variable is such that
δ(f (x))|f 0 (x)|dx = δ(x − x∗ )dx for any function f (x) with single zero at x∗ . The extension
of this formula to the infinite dimensional case corresponding to the curve xt reads
Y Y δF
δ xs − x∗s = [Dx] ×

[Dx] × δ(F(xs )) × Det ,

s s
δx
where Det δF

δx is the infinite dimensional (functional) determinant of the functional oper-
δF(xs )
ator with entries δx 0 . This operator is [∂t − a0 (xt )] viewed as acting on functions of
s
the variable t. Thus the path-integral measure we start with reads
1 2
R Y
[Dw][Dx] e− 2 ds ẇs × δ ẋs − a(xs ) − ẇs Det[∂t − a0 (xt )].

28
As before, we can represent the Dirac delta function using Fourier transform. The only
delicate point is that now this is continuous product of δ-function so that the conjugated
variable, which we denote φs , as to be a function of the time variable, so that
Y Z R
δ ẋs − a(xs ) − ẇs = [Dφ] ei ds φs (ẋs −a(xs )−ẇs ) .

s
The field φs plays the role of Lagrange multipliers imposing the constraint. Gathering all
the bits, we can write the path-integral measure as
1
dsẇs2 i
R R
[Dw][Dx][Dφ] e− 2 Det[∂t − a0 (xt )] .
ds φs (ẋs −a(xs )−ẇs )

e
We now have to deal with the determinant (the Jacobian). We know from the discrete
formula we previously discuss that this Jacobian is 1, at least with the Itô convention (this
will be proved in more detail in the exercise section)1 :
Det[∂t − a0 (xt )]
= 1, (Ito convention).
Det[∂t ]
We give an alternative naive argument (if not proof) working directly in the continuum.
This determinant is the Jacobian for the change of variable from the function xt to the
function zt , solution of żt = ẋt − a(xt ) − wt . With the Itô convention, which
R t is a strict
forward in time convention, the function zt as to be thought as zt = xt − 0 − ds (a(xs ) +
ws ). Hence, with this convention, the Jacobian is one because it is the determinant of
δzt
the ‘matrix’ [ δxs
] which is (in this convention) strictly triangular with 1 on the diagonal:
Hence, with this convention, we got the measure [Dw][Dx][Dφ] e−S[x] with action:
Z Z
1
S[x] = ds ẇs2 + i ds φs (ẋs − a(xs ) − ẇs ).
2
The term ẇs2 represents the Brownian Gaussian measure, and φs plays the role of Lagrange
multiplier enforcing the SDE. We can now do the Gaussian integrals successively the
Gaussian integrals over ws and then φs to obtain:
2
Z Z
S[x] ≡ ds φ̇2s + i ds φs (ẋs − a(xs ))
2
Z
1
≡ ds (ẋs − a(xs ))2
22
By convention notation ≡ means “equivalent up to Gaussian integration”. Note that this
Jacobian, and hence this convention, do not matter in the small nose limit → 0.
1
Actually, we have
R 0
Det[∂t − a0 (xt )] = Det[∂s ] Det δ(t − s) − θ(t − s)a0 (xs ) = Det[∂s ] e−θ(0) ds a (xs ) .

As usual, there is an ambiguity in defining θ(0), which is linked to ordering factors. The different conven-
tions for stochastic calculus, say Itô versus Stratonovich, correspond to different convention for θ(0). Itô
convention corresponds to θ(0) = 0, Stratonovich convention to θ(0) = 1/2.
29
• Path integral perturbation and the Girsanov theorem.

The path integral representation has a simple interpretation in terms of Radon-Nikodym
derivatives (i.e. changes of probability measures), martingales and the Girsanov theorem.
This provides a formal way to make it rigorous mathematically.
Recall the action (with = 1 for simplicity)
Z t
1
S[x] = ds (ẋs − a(xs ))2 .
2 0
1
R 2
RT 1
RT 2
The weight e−S factorizes into the product of two terms: e− 2 ds ẋs and e 0 ds ẋs a(xs )− 2 0 ds a(xs ) .
Even though there is clearly a question of convention (‘Itô versus Stratonovich’) let us write
RT RT
formally 0 ds ẋs a(xs ) = 0 a(xs ) dxs . Then we have:
1
RT RT RT
ds ẋ2s a(xs ) dxs − 12 ds a(xs )2
[Dx] e−S[x] = [Dx] e− 2 0 ×e 0 0 .
The first term is the Brownian measure and the second codes for the change of measure,
weighting differently the trajectories of the process. Let MT be this extra weight:
RT RT
a(xs ) dxs − 12 ds a(xs )2
MT = e 0 0 .
We can then write the measures as

1
RT
ds ẋ2s
[Dx] e−S[x] = [Dx] e− 2 0 × MT .
Hence we may interpret the path integral representation in terms of the Brownian measure
as : Z
[Dx] e−S[x] · · · = Eo MT · · · ,

with Eo [· · · ] the Brownian measure for X and Mt the exponential martingale defined
above. The process t → Mt (defined as an Itô integral) is known to be a martingale under
the Brownian measure, solution of the stochastic differential equation dMt = a(Xt )Mt dXt
for dXt2 = dt. Alternatively we can write this relation as dP = dPo × MT , with Po the
Brownian measure and P that for the process associated to the SDE dXt = a(Xt )dt + dWt .
This representation of the measure of the process Xt is known as the Girsanov theorem.
• Itô versus Stratonovich.
Here is a comment on the echo of the Itô and Stratonovich conventions on the path
integral representation of SDEs. Although they yield naively different representations,
these conventions are of course compatible. Let us again consider the (simple) SDE dXt =
a(Xt )dt + dWt with dWt2 = dt. The action of path integral representation depends on the
convention we are using because, as mentioned above and proved in an exercice below, the
determinant Jacobian is convention dependent. One has:
ds a0 (xs )
R
Det[∂t − a0 (xt )] = Det[∂t ] × e−α ,
30
with α = 12 with the Stratonovich convention and α = 0 with the Itô convention. As a
consequence the path integral actions are respectively:
1 T
Z
SIto = ds (ẋ2s − a(xs ))2 ,
2 0
1 T 1 T
Z Z
SStrato = 2 2
ds (ẋs − a(xs )) + ds a0 (xs ).
2 0 2 0
The two representations are of course compatibleR because when expanding the square
2 2
R
R ds(ẋs − a(xs )) we have to view the cross term dsẋs a(xs ) as the stochastic integral
dxs a(xs ) which is convention dependent by definition. Compatibility is recovered be-
RT
cause Itô and Stratonovich integrals are related as follows (see above): 0 a(xs ) dxs |Ito =
RT 1 T 0
R
0 a(xs ) dxs |Strato − 2 0 a (xs ) ds. Each convention has his own avantages: Ito calculus
encodes directly for the independence of the increment of the Brownian motion but chain
rules and changes of variables are simpler in the Stratonovich convention.
2.7 The small noise limit and the Arrhenius law

The path integral representation may sound a bit formal –actually, it is quite formal.
Except in a few cases, Gaussian integrals for instance, it cannot be computed. It is
nevertheless adapted to formal manipulations which allow to extract general properties
that the measure should satisfy —or it expected to satisfy— if it can be defined. It is
also computable when there is a small parameter which can be used to approximate the
integral via a saddle point. In the context of quantum mechanics this approximation is
called the semi-classical approximation, in the context of field theory it is called instantons
for historical reasons. This saddle point approximation allows us to derive results which
are exact in the small noise limit. For instance, the presence of noise may allow the
dynamical systems to follows paths, say to through an energy barrier, which would be
forbidden in absence of noise. The small noise saddle point approximation provides a way
to derive exact results for the probability of the occurence of these forbidden moves even
though these probabilities are non perturbative in (they are asymptotically zero as → 0
and do not admit a power series expansion in ).
• Small noise and saddle points.
We want to study the small noise Rlimit (i.e. the limit → 0). In this case the
action S[x] = 22 ds (ẋs − a(xs ))2 + α ds a0 (xs ) is independent of the convention (i.e.
1
R
independent on α). So we take

Z T
1 1
S[x] = 2 ds (ẋs − a(xs ))2 =: A.
2 0 2
The path integral, with measure [Dx] e−S is dominated in the limit → 0 by the saddle
points (actually the minimum) of S. We may write S = A/2 to make the dependence
explicit. Let us call “classical” the trajectory minimizing the action A. These trajectories
31
are solutions of the Euler-Lagrange equations of motion derived from this action. These
classical solutions are also/sometimes called “instantons”. Naively, we may write
Z
−2
[Dx]e−S[x] [...] →0 e− Aclassical [...]classical . (8)
A more precise formulation of this equation consists in taking the logarithm of both sides
(and this is what the symbol means). There could be a sum over all classical solutions
in case there are many (but the one with the minimal action contributes the most).
A more precise formulation proved rigorously (Freidlin-Wentzell theory) is the follow-
ing. Let x̂s be a given smooth path from xi to xf on the time interval [0, T ]. Then the
probability that the stochastic path xs is (arbitrary) close the given path is such that
1 T
h i Z
2
2
lim lim log P sup |xs − x̂s | < δ = − ds x̂˙ s − a(x̂s )
δ→0 →0 0≤s≤T 2 0
Note that this action (and hence minus the logarithm of the probability) is always positive.
It is zero (and hence maximal) on the noiseless trajectories x̂˙ s = a(x̂s ). Given two points xi
and xf , such noiseless trajectories do not always exist. This probability is maximal on the
classical trajectory (minimizing the classical action). The formula codes more information
because it also measures how fast the probability decreases for trajectories which deviate
from the classical ones.
• Saddle points and classical instantons.
Instantons are solutions of the Euler-Lagrange (classical) equation (these are not the
noiseless equations of motion, they are more general):
1
ẍ = a(x)a0 (x) = −∂x Veff (x), with Veff = − a2 (x).
2
This corresponds to the motion of a classical particle in a potential Veff . The ‘energy’
is conserved e = 21 ẋ2 − 12 a2 (x) = 21 ẋ2 + Veff . The momentum is p = ẋ − a(x) and the
2
hamiltonian is h(p, x) = 21 ẋ2 − 21 a2 (x) = p2 + pa(x). The action A(T, x), evaluated along
the classical trajectory and view as a function of the time T and final position x satisfies
the hamilton-Jacobi equation
1
∂T A = h(p = ∂x A, x) = (∂x A)2 + a(x)∂x A.
2
We can also write the equation of motion keeping track of the auxiliary field (the
RT 2 RT
Lagrange multiplier). Then S = i 0 ds φs (ẋs − a(xs )) + 2 0 ds φ2s . To factorize we
RT RT
choose to redefine φs = −i−2 ξs ). Then 2 S = 0 ds ξs (ẋs − a(xs )) − 12 0 ds ξs2 . The
(classical) equations of motion are ξs = ẋs − a(xs ) and ξ˙s = −a0 (xs )ξs . They are clearly
equivalent to the equations above.
Notice that noiseless trajectories, solutions of ẋ = a(x) are always solutions (because
they satisfy ẍ = a(x)a0 (x)). They are zero ‘energy’ solution e = 0. These are not the
only zero ‘energy’ trajectories. The other set are those with reversed/opposite velocity,
32
i.e. ẋ = −a(x). The former are sometime called the ‘downhill trajectories’, the later the
‘uphill’ trajectoires (the names comes from looking that these trajectories in the landscape
specified by the potential u(x) such that a(x) = −U 0 (x)).
downhill : ẋt = a(xt ), pt = 0, e = 0, A = 0

uphill : ẋt = −a(xt ), pt 6= 0, e = 0, A = 2 U (xf ) − U (xi ) .
We have given the value of the action on these trajectories. The uphill trajectories exist
only if U (xf ) ≥ U (xi ), because A ≥ 0 (hence they are effectively uphill; they are induced
by the noise). The downhill trajectories exist only if U (xf ) ≤ U (xi ).
• Application to the Arrhenius law.
Let a(x) = −U 0 (x) and take the potential U to be a double well. Figure. The
2
stationary measure is P (x)dx = Z −1 e−2U (x)/ dx; it is very localized around the minima
when → 0 (which means that a population of particle in such landscape, if stationary
w.r.t. to this process, will mostly be concentrated around the minima). We may identify
2 with the temperature (recall 2 = 2kB T ). We aim at evaluating the probability that
thermal noise/fluctuations induce transitions of particle from one minima to the next. So
we want to evaluate the probability that there is transition from xi (close to one minima)
to xf (close to another minima; take xi < xf to fix the setup), that we may define by2
P[xi → xf ] := SupT PT [xi → xf ].

−2
In the small noise limit PT [xi → xf ] '→0 e− A(T ;xi →xf ) . Maximizing over the time T
to reach the final position amounts to extremize the action A over T , hence to impose
∂T A = 0. By the Hamilton-Jacobi equation, this corresponds to zero ‘energy’ solution.
Hence the transition probability is
−2 A(x
P[xi → xf ] →0 e− i →xf ;e=0)
.
Now which zero energy trajectory we have to chose depends on the profil of the potential
U (x) (recall that we set a(x) = −U 0 (x)). When the potential goes down we take the
noiseless (classical) downhill trajectories with A = 0, when the potential goes up we have
to take the “anti-noiseless” uphill trajectoires with A = 2(∆U ). Hence,
2
P[xi → xf ] →0 e−2(∆U )tot / = e−(∆U )tot /kB T ,
with (∆U )tot the sum of the potential differences along uphill trajectories. Recall that 2 /2
is identified with the temperature. This is the Arrhenius law/formula. We can actually
do a bit better and compute the pre-factor (but we leave this as an exercise...).
2
It is clear that if xi and xf are separated by an energy barrier the maximum of PT [xi → xf ] is attained
asymptotically when T goes to infinity, and then PT [xi → xf ], as a function of xf , is asymptotically
2
proportional to the invariant measure e−2U (xf )/ .
33
2.8 Exercises
• Exercise 2.1: Random variables and generating functions.
Let X be a real random variable. Let its characteristic function (also called generating
function) be defined by
Φ(z) := E[ eizX ].
It is always well defined for real z (why?), but its properties under analytical continuation
depend on the distribution. Let W (z) be defined as W (z) := log Φ(z) or Φ(z) = eW (z) .
(i) Expand Φ and W in power of z and identify the first few Taylor coefficients.
n
(ii) Suppose that X is a Poisson variable taking integer values with P[X = n] = λn! e−λ .
What are its means, its covariance and its generating function?
(iii) Suppose that X is a Gaussian variable with probability distribution density
dx −x2 /2σ
P[ X ∈ [x, x + dx] ] = √ e .
2πσ
Verify that P is correctly normalized and compute its characteristic function.
• Exercise 2.2: Random Gaussian vectors.
Let X ~ be a N -dimensional Gaussian random vector with real coordinates X i , i =
1, · · · , N . By definition it probability distribution is
det[G] 21 1
P(X) dN X = dN X

N
exp − hX|G|Xi ,
(2π) 2
X i Gij X j , where the real symmetric form Gij is supposed to be non

P
with hX, G, Xi = ij
degenerate. Denote by G b its inverse: Gij Gb jk = δ i .
k
(i) Verify that this distribution is normalized, that is:
dN X − 1 hX|G|Xi
Z
− 1
e 2 = det[G] 2
.
(2π)N/2
(ii) Verify that E[X i ] = 0 and E[X i X j ] = G

b ij .
ihU |Xi ] with hU |Xi = i Ui X i is
P
(iii) Show the generating function E[e
1
E[eihU |Xi ] = e− 2 hU |G|U i .
b
Notice that U and X belong to dual spaces.

• Exercise 2.3: The law of large number and the central limit theorem.
The aim of this exercise is to prove (a simplified version of) the central limit theorem,
which we used above. Let k , k = 1, · · · , n, be independent identically distributed (i.i.d.)
variables. To simplify matter, let us suppose that = ± with probability 12 . This theorem
says that the sum Ŝn = √1n k k , is a Gaussian variable in the limit n → ∞ (the more
P
precise statement is that it converges in law).
34
(i) Prove that

z n z2
E[eiz Ŝn ] = cos( √ ) −−−→ e− 2 ,

n n→∞
and conclude.
z2
Hint: Recall the Taylor expansion cos( √zn ) = 1− 2n +o( n1 ) and use limn→∞ [1− ny +o( n1 )]n =
e−y (which can be proved by taking the logarithm).
• Exercise 2.4: Free random paths.
The aim of this exercise is to complete the study of free paths presented in the main
text. See the main text for a precise definition of the statistical ensemble of free paths.
Recall the definition of Z(x) as the partition function of free paths from 0 to x:
D
X
j=1
(i) Compute the Fourier transform of Z(x) and prove the explicit expression of Z(x) as
an integral over the Brouillon zone given in the text.
(ii) Let ∆dis. be the discrete Laplacian and write ∆dis. = Θ − 2D I with Θ the lattice
adjacencyPmatrix and I the identity matrix. We view Θ as acting on functions via (Θ ·
D
f )(x) = j=1 f (x + aej ) + f (x − aej ) . Show that:
1
Z(x) = hx| |0i,
I − µΘ
with |xi the δ-function peaked at x, i.e. hy|xi = δy;x . Deduce from this formula that Z(x)
converges for |µ| P
< µc with µc = 1/2D.
(iii) Let Z(x) = N µN WNfree (x) for |µ| < µc . Give an expression of WNfree (x) as matrix
elements of powers of the matrix Θ and give a geometrical interpretation of this formula.
(iv) Prove the formula for the Green function G(x) given in the main text.
• Exercise 2.5: Computation of a path integral Jacobian determinants.
The aim of this exercise is to compute the determinant Det[∂t − A(t)] of the linear
map acting functions f (t) as follows f (t) → (Jf )(t) = f 0 (t) − A(t)f (t) with A(t) a given
function. Instead of computing directly this determinant we factorize the derivation op-
erator and we write Det[∂t − A(t)] := Det[∂t ] × Det[1 − K]. The operator K is defined by
integration as follows:
Z t
K : f (t) → (Kf )(t) = dsA(s)f (s),
0
for any function f defined on the finite interval [0, T ]. The aim of this exercise is thus to
compute the determinant Det[1 − K] and to prove that
RT
Det[1 − K] = e−α 0 ds A(s)
,
with α a parameter depending on the regularization procedure (α = 0 for Itô and α = 21
for Stratonovich conventions). This illustrates possible strategy to define and compute
functional –infinite dimensional– determinants.
35
To define the determinant Det[1 − K] we need to discretize it by representing the

integral of any function by a Riemann sum. Let us divide the interval [0, T ] in N sub-
interval [nδ, (n+1)δ] with n = 0, · · · N −1 and δ = T /N . We will then take the limit N → 0.
To simplify notation we denote fn := f (nδ). There are many possible discretizations
but we shall only consider two of them (which correspond to the Itô and Stratonovich
conventions):
Z t n−1
X
Ito : f (t)dt := lim δ fk ,
0 N →∞
k=0
Z t n−1
X 1
Stratonovich : f (t)dt := lim δ (fk + fk+1 ).
0 N →∞ 2
k=0
(i) Write the regularized action of the operator K on function f by writing the expression
of (Kf )n .
(ii) Show that the operator 1 − K is lower triangular and determine the diagonal entries
(which are convention dependent).
(iii) Deduce, by taking the large N limit, the formula for the determinant:
Ito : Det[1 − K] = 1,
1
RT
Stratonovich : Det[1 − K] = e− 2 0 ds A(s)
.
MoreR general discretizationPare defined by sampling differently the Riemann sum as

t
follows: 0 f (t)dt = limN →∞ δ n−1k=0 ((1 − α)fk + αfk+1 ). Following the same strategy as
RT
above, it is then clear that Det[1 − K] = e−α 0 ds A(s)
.
• Exercise 2.6: Levy’s construction of the Brownian motion.
The path integral representation is actually closely related to an older (!) construction
of the Brownian motion due to P. Levy. The aim of this exercise is to present the main
point of Levy’s approach which constructs the Brownian paths by recursive dichotomy.
We aim at constructing the Brownian curves on the time interval [0, T ] starting point
x0 . The construction is recursive.
2
(a) First, pick the end point xT with the Gaussian probability density √dx T
2πT
e−(xT −x0 ) /2T
and draw (provisionally) a straight line from x0 to xT .
(b) Second, construct the intermediate middle point xT /2 at time T /2 by picking it ran-
domly from the Gaussian distribution centered around the middle of the segment joining
x0 to xT , and with the appropriate covariance to be determined. Then, draw (provision-
ally) two straight lines from x0 to xT /2 and from xT /2 to xT .
(ic) Next, iterate by picking the intermediate points at times T /4 and 3T /4, respectively,
from the Gaussian distribution centered around the middle point of the two segments
drawn between x0 and xT /2 and between xT /2 and xT , respectively, and with the ap-
propriate covariance. Then draw (provisionally) all four segments joining the successive
points x0 , xT /4 , xT /2 , x3T /4 and xT /2 .
(d) Iterate ad infinitum...
36
Show that this construction yields curves sampled with the Brownian measure.
Hint: This construction works thanks to the relation
x +x
(xi − x)2 (x − xf )2 (xi − xf )2 (x − i 2 f )2
+ = +
2(t/2) 2(t/2) 2t 2(t/4)
• Exercise 2.7: The over-damped limit of the noisy Newtonian particle.

Consider Newton’s equation for a particle of mass m subject to a friction and random
forcing (white noise in time). That is, consider the SDEs:
Pt
dXt = dt, dPt = −γ dXt + dBt ,
m
with Xt the position and Pt the momentum. We are interested in the limit m → 0 (or
equivalently γ large). Let us set m = 2 to match the Brownian scaling. Then show that:
(i) the process γXt converges to a Brownian motion Bt ;
(ii) Yt := Ẋt converges to a finite random variable with Gaussian distribution.
That is: Introducing the mass, or , is a way to regularize the Brownian curves in the
sense that Xt admits a time derivative contrary to the Brownian motion. But quantities
such as Yt , which are naively expected to vanish in the limit → 0, actually do not
disappear because the smallness of is compensated by the irregularities in Ẋt as → 0.
For instance E[ 21 mẊt2 ] is finite in the limit m → 0. Such phenomena –the existence of
naively zero but nevertheless finite quantities due to the emergence of irregular structures
in absence of regularizing– is commun phenomena in statistical field theory, (sometimes)
called ‘anomaly’.
• Exercise 2.8: SDEs with ‘multiplicative’ noise.
Generalize the results described above for a more general SDE of the form
dXt = a(Xt )dt + b(Xt )dBt
with a(x) and b(x) smooth non constant functions. To deal with the small noise limit one
may introduce a small parameter by rescaling b(x) via b(x) → b(x).
(i) Prove that the Fokker-Planck operator for SDEs reads
1
∂x b2 (x) − a(x)

H = ∂x
2
(ii) Verify that the invariant measure (if normalizable) is
Z x
a(y)
Pinv (x) dx = b−2 (x) e−2s(x) dx, s(x) := − dy .
b2 (y)
What is the invariant measure if the later is not normalizable?
What is then the physical interpretation of this new measure?
(iii) Show that the action of the path integral representation of these SDEs is
1 T (ẋs − a(xs ))2

Z
S= ds .
2 0 b2 (xs )
37
in the small noise limit 1. Verify (by going back to the discret formulation) that this
way of writing the action is still valid away from the small noise limit provided that one
carefully defined the integrals.
• Exercise 2.9: Feynman-Kac formula
RT
Write the path integral representation of E e− 0 ds V (xs ) , with s → xs a standard

Brownian motion, wohse measure is denoted E[·], and V (x) a smooth bounded potential.
• Exercise 2.10: Multivariable SDEs
Generalize all these results for multivariable SDEs of the form dX i = ai (X) dt +
bij (X) dBtj
where B j are Brownian motions with covariance E[Bti Bsj ] = δ ij min(t, s).
38
3 Statistical lattice models
In this Chapter we present basic examples of lattice models of statistical mechanics

and the 2D Ising model with more details. This allows us to introduce the important
concept of transfer matrix which is based on an algebraic rewriting of the Boltzmann
sums. The correspondences between D dimensional statistical lattice models and D − 1
quantum systems is spelled out.
3.1 Examples of statistical lattice models

By a statistical lattice model is meant a model of statistical mechanics defined on a graph
G = (V, E) with vertices V and edges E. Most of the time, G is a regular lattice embedded
in D dimensions, such as a square or a (hyper)cubic lattice. In some situations it is of
interest to study models on more general classes of graphs, for instance the set of all
planar triangulations. The statistical sum then carries over both the statistical degrees of
freedom and the graphs themselves, weighted by the number of vertices, and it provides a
discretisation of 2D quantum gravity.
The statistical degrees of freedom are typically discrete variables, called spins and
denoted Si . In most situations the spins are defined on the vertices V , so that i ∈ V , and
the interaction takes place along the edges E, via the definition of some (dimensionless)
energy functional, called Hamiltonian and denoted H = E/(kB T ), that depends only on
pairs of neighbouring spins. The properties of the system can be studied via the correlation
functions of spins, by which is meant the mean (expectation) values of functions of spins
at selected vertices, such as hSi i, hSi Sj i, etc.
Although the interaction in such models is short-ranged, they can have a very rich
behaviour at long distances. When D ≥ 2 there typically exists a critical temperature
Tc so that the system is ordered for T < Tc and disordered at T > Tc . In the low-
temperature phase, the majority of the spins takes a particular value, meaning that we
have spontaneous symmetry breaking with a non-zero value of a suitably defined order
parameter. The transition at Tc can be first-order (with latent heat), or second (or maybe
higher) order. In the latter case, the system is said to be critical, and it will display
statistical fluctuations at all length scales, meaning more precisely that its correlation
functions transform covariantly under a change of the length scale. Concretely, the two-
point function of spins will decay with distance as a power law (sometimes involving also
a logarithmic factor), hSi Sj i ∝ |i − j|−α , defining some critical exponent α. For T 6= Tc , it
would instead decay as an exponential, hSi Sj i ∝ e−|i−j|/ξ , defining a characteristic length
ξ, called correlation length, beyond which the effective interaction becomes negligible.
• Ising model
The simplest —and most well-studied— model of this type is the Ising model with
39
Si = ±1 and X X
H = −K Si Sj − B Si . (9)
(ij)∈E i∈V
Here K = J/(kB T ) is the (dimensionless) coupling constant, and B the corresponding

magnetic field. The Ising model can be solved exactly —by which we mean that the
partition function and correlation functions can be computed analytically— in D = 1,
and also for B = 0 on regular lattices in D = 2, for any value of the temperature T . The
exact solutions in D = 2 with B 6= 0, and in D > 2 with B = 0 seem beyond reach, despite
considerable effort; but in both cases we have nevertheless a good understanding of the
long distance behaviour from field theory. The Ising model has a Z2 symmetry, Si → −Si ,
which is explicitly broken for B 6= 0, and spontaneously broken for B = 0 and T < Tc .
Below we present the exact solution of the D = 2 zero-field Ising model.
The properties of statistical models depend subtly on which correlation functions one
is interested in. In the case of the 2D Ising model, it is also of interest to study the
behaviour of domain walls, which are the boundaries between regions with Si = 1 and
Si = −1. To be precise, they are closed curves on the dual lattice G∗ . One may then ask
questions like:
— What is the fractal dimension of a domain wall, or of a region of spins surrounded by
one domain wall?
— What is the decay of the probability that two (or more) domain walls extend from the
neighbourhood of a point i to the neighbourhood of a distant point j?
This type of questions have highly non-trivial answers even in the case T = ∞. So for
T = ∞, the 2D Ising model behaves trivially in terms of the spins (with ξ = 0), but
non-trivially in terms of the non-locally defined domain walls. This gives rise to an active
research field known as random geometry.
• Potts model
The Ising model is a special case of the Q-state Potts model. Let Q ∈ N and define,
for each i ∈ V , a spin Si = 1, 2, . . . , Q. The Hamiltonian is
X
H = −K δ(Si , Sj ) , (10)
(i,j)∈E
where the Kronecker delta function is defined as

1 if Si = Sj
δ(Si , Sj ) = (11)
0 otherwise
Note that is we had defined the Q = 2 model with Si = ±1, then 2δ(Si , Sj ) = Si Sj + 1,
so this is equivalent to the Ising model with KPotts = 2KIsing .
The Potts model has a second order phase transition also for Q = 3 and Q = 4. But
more interestingly, it is possible to reformulate it geometrically for real values of Q. To
see this, notice first that by (11) we have the identity
eKδ(Si ,Sj ) = 1 + vδ(Si , Sj ) , (12)
40
where we have defined v = eK − 1. Now, it is obvious that for any edge-dependent factors
he one has Y X Y
(1 + he ) = he . (13)
e∈E E 0 ⊆E e∈E 0
where the subset E 0 is defined asQthe set of edges for which we have taken the term he in
the development of the product e∈E . In particular, taking he = vδ(σi , σj ) we obtain for
the partition function
0 0 0
X X X Y X
Z= e−H = v |E | δ(Si , Sj ) = v |E | Qk(E ) , (14)
S E 0 ⊆E S (ij)∈E 0 E 0 ⊆E
where k(E 0 ) is the number of connected components in the graph G0 = (V, E 0 ), i.e. the
graph obtained from G by removing the edges in E \ E 0 . Those connected components
are called clusters, and (14) is the Fortuin-Kasteleyn cluster representation of the Potts
model partition function. The sum over spins S in the original definition of Z has now
been replaced by a sum over edge subsets, and Q appears as a parameter in (14) and no
longer as a summation limit. Therefore it makes sense to take any real Q > 0.
• Bond percolation
For Q = 1 the Potts model is seemingly trivial, with partition function Z = (1 + v)|E| .
Instead of setting Q = 1 brutally, one can however consider taking the limit Q → 1. This
leads to the important special case of bond percolation.
Let p ∈ [0, 1] and set v = p/(1 − p). We then consider the rescaled partition function
0 0 0
X
Z(Q)
e ≡ (1 − p)|E| Z = p|E | (1 − p)|E|−|E | Qk(E ) . (15)
E 0 ⊆E
We have of course Z(1)

e = 1. But formally, what is written here is that each edge is
0
present in E (i.e., percolating) with probability p and absent (i.e., non percolating) with
probability 1 − p. Appropriate correlation functions and derivatives of Z(Q)
e in the limit
Q → 1 furnish valuable information about the geometry of the percolation clusters. For
instance
dZ(Q)
e
lim Q = hk(E 0 )i
Q→1 dQ
gives the average number of clusters.

Bond percolation is an extreme example of random geometry. The local occupation
numbers of edges are completely uncorrelated, and all corresponding correlation functions
are trivial. However, non-trivial questions about the fractal geometry of the clusters can
be asked in analogy with the above discussion of Ising domain walls.
3.2 Transfer matrices

The transfer matrix T is a linear operator that builds the partition function Z of a D-
dimensional statistical lattice model with short-range interactions, by relating the states
41
of two adjacent (D − 1) dimensional slices. With suitable modification T can be used to

study correlation functions as well.
The basic strategy is as follows. Divide the lattice into two parts, called past and
future, by intersecting it with a (D − 1) dimensional (hyper)plane, called a time slice. The
vertices contained in the time slice belong to the present. Let S ~t denote the states of all
~
spins corresponding to vertices in the present, and let Zt (St ) be the partition function
of the past part of the system, conditioned by the state S ~t of the present vertices. We
interpret t as a time parameter, and we consider building up the whole lattice by letting
the time evolve in discrete steps.
Consider now another time slice at time t + 1. We suppose that it differs from the
one at time t by having had the present shifted a bit into the future, in such a way that
vertices belonging to the present at time t and at time t + 1 are either identical or nearest
neighbours. Knowing S ~t and S~t+1 it is then possible to infer the part w(t, t + 1) of the
Boltzmann weights that describe the interaction between the spins in S ~t and those in S
~t+1 .
We can then write X
~t+1 ) =
Zt+1 (S ~t ) ,
w(t, t + 1)Zt (S (16)
~t
S
where w(t, t + 1) are the matrix elements of the linear operator Tt+1 :
~t+1 |Tt+1 |S
w(t, t + 1) = hS ~t i .
Eq. (16) is just the component form of Zt+1 = Tt+1 Zt , where Zt and Zt+1 are vectors and
T is a matrix, whence the name transfer matrix. Iterating this relation we obtain finally
the complete partition function
X X
Z= ZT = ~T |TT · · · T3 T2 T1 |S
hS ~0 i , (17)
~T
S ~ 0 ,S
S ~T
where S~0 and S ~T describe the possible initial and final states. When writing this, care
must be taken so that the interaction along each edge of the lattice is represented by one
and only one transfer matrix, since otherwise some double-counting would occur.
For a regular lattice, we can choose always to shift the time slices in the same way, in
the direction orthogonal to the time slice, so that both the state space S~t and the transfer
matrix Tt are independent of t. If we further assume periodic boundary conditions in
the direction perpendicular to the time slice, the initial and final spins can be identified,
~T = S
viz. S ~0 , and (17) reduces to
Z = Tr T T .

(18)
In most situations T is diagonalisable, and since Boltzmann weights are strictly positive
the Perron-Frobenius theorem guarantees that the largest (dominant) eigenvalue Λ0 is
non-degenerate. We have then
Z ' (Λ0 )T . (19)
for T large. The behaviour of the partition function in the thermodynamic limit (large
system size) is thus determined by the largest transfer matrix eigenvalue.
42
• Example: 1D Ising model

The Hamiltonian is
T T
X BX
H = −K St St+1 − (St + St+1 ) , (20)
2
t=1 t=1
where each time slice consists of a single spin St . Note that we have rewritten the interac-
tion with the magnetic field B in a convenient way that makes St and St+1 play symmetric
roles. Since St = ±1 takes two values, T is just a 2 × 2 matrix:
e−K
K+B
e
T = .
e−K eK−B
• Example: Dimer-monomer mixture in 1D

Consider a 1D lattice with T sites and free boundary conditions. We wish to cover
the lattice with two different kinds of objects —dimers that are small rods covering two
adjacent sites, and monomers that are points covering just one site— in such a way that
each lattice site is covered by precisely one object. The Boltzmann weights for dimers and
monomers are wd and wm respectively.
To deal with this situation, we let St describe the states of the edges rather than the
vertices. More precisely, St = 0 or 1, depending on whether the edge just to the left of
the site t is empty or occupied by a dimer. Then

wm 1
T = ,
wd 0
and the initial and final state vectors are

1
|S0 i = , hST | = 1 0 ,
0
In (17) there is no sum over S0 and ST , since the boundary states have been fixed in this
way, by requiring that no dimer can stick out of the system. So Z = hST |T T |S0 i. We
leave as an exercise to show that
bT /2c
X T −n
Z= (wd )n (wm )T −2n
T − 2n
n=0
by using the transfer matrix formalism. Can you provide also a direct combinatorial
argument?
• Further remarks
In D = 2, each time slice describes the set of spins along a (say) horizontal line
intersecting the lattice. Therefore, for a lattice of size Nh × Nv , the transfer matrix of the
Ising model has dimension 2Nh . To describe an infinite lattice, we must take the Nh → ∞
43
limit, and dealing with this is the crux of solving such a model exactly. We shall see it
done below.
The transfer matrix formalism is quite malleable and can deal with a variety of situ-
ations, provided that the concepts of time slices and state spaces are carefully rethought.
Here are a few possibilities:
• Interactions of longer (but finite range): Make the time slices thicker (i.e. consist of
several adjacent layers).
• Continuous degrees of freedom: In each matrix product, replace the summation over
the discrete degrees of freedom by an integral.
• Inhomogeneities: Replace T by a time-dependent Tt .
• Random graphs: Sum over both the statistical degrees of freedom and the graphs
themselves. Use a time slice whose size varies as it is swept over the graph.
3.3 The 2D Ising model

We now consider the zero-field Ising model on a 2D square lattice. The Hamiltonian is
Ny
Nx X
X
H=− [Kx Sx,y Sx+1,y + Ky Sx,y Sx,y+1 ] , (21)
x=1 y=1
where we have chosen different coupling constants, Kx and Ky , in the horizontal and
vertical directions. We assume periodic boundary conditions in both directions.
• Existence of a phase transition
We first present an argument due to Peierls that the 2D Ising model has a phase
transition at some non-zero temperature T . We assume here Kx = Ky ≡ K = J/T , where
J is the physical (temperature independent) coupling constant.
The ground state in which all spins are aligned has energy E0 = 2JNx Ny . An excitation
consisting of a droplet of overturned spins within the ordered ground state has energy
E1 = E0 + 2J`, where ` is the perimeter of the droplet. The number of possible shapes of
the droplet equals the number of closed walks on the square lattice of length `, with self-
intersections at the vertices being allowed. Viewing such walks as an exploration process,
one has at least 1 and at most 3 possibilities for each step (the walk cannot backtrace).
So the number of walks is of order c` , with 1 < c < 3 (the precise value of c is known but
does not matter for this argument).
The contribution to Z of such droplet configurations is c` e−E1 /T , so the free energy is
F = E − T S = E0 + (2J − T log c)` .

2J
Now, if T < Tc ≡ log c we have F > 0 for any `, so droplets will be exponentially
suppressed. Conversely, if T > Tc we have F < 0 for large enough `, so droplets will
proliferate. Therefore we expect a phase transition at Tc > 0.
44
By contrast, in 1D a droplet corresponds to just two units of domain wall, so the

excitation energy is constant. Therefore we will have Tc = 0 in that case, as is easily
confirmed by solving the model exactly.
• High-temperature expansion
The Boltzmann weight of a configuration of Ising spins S reads
Y
W [S] = exp(Kx Sx,y Sx+1,y ) exp(Ky Sx,y Sx,y+1 ) .
x,y
Since the product of two spins Si Sj = ±1, we have the identity
exp(KSi Sj ) = cosh K + sinh K Si Sj = cosh K × (1 + w Si Sj ) , (22)
where w = tanh K. If we drop the overall multiplicative factor, we get:

Y
W [S] ∝ (1 + wx Sx,y Sx+1,y )(1 + wy Sx,y Sx,y+1 ) ,
x,y
with
wx = tanh Kx , wy = tanh Ky .
This product can be expanded graphically, associating the term 1 with an empty edge,
and the term wSi Sj with an occupied edge. Thus
 
wx|Ex (A)| wy|Ey (A)|
XX Y
Z = (cosh Kx )|Ex | (cosh Ky )|Ey | Si Sj  ,
{S} A⊆E (ij)∈A
where |Ex (A)| and |Ey (A)| are the number of horizontal and vertical edges contained in
the edge subset A. To get a non-zero contribution, every factor Si should occur as an even
power, so a non-zero contribution is obtained only if A is a set of closed polygons:
|E (A)|
X
Z = (cosh K1 )|Ex | (cosh K2 )|Ey | 2|V | wx|Ex (A)| wy y . (23)
A polygons
This exact rewriting of Z is called a high-temperature expansion, since w = tanh K 1

when K 1.
• Low-temperature expansion
We can also expand Z around a fully ordered state. The excitations are then droplets
of spins of the opposite sign, which are bordered by domain walls that live on the dual
graph G∗ , again a square lattice. These domain walls are again polygon configurations:
∗ ∗
X
Z = 2eKx |Ex |+Ky |Ey | (wx∗ )|Ey (A )| (wy∗ )|Ex (A )| ,
A∗ ⊆E ∗
where
wx∗ = e−2Kx , wy∗ = e−2Ky .
45
Notice how wx∗ is now conjugate to |Ey (A∗ )|, since a horizontal edge in E is dual to a
vertical edge in E ∗ , and vice versa. This is called a low-temperature expansion.
• Duality
It follows that if the high-temperature rewriting of Z has a singularity at some param-
eters (wx , wy ), then the low-temperature rewriting of Z must have the same singularity
at the dual parameters (wy∗ , wx∗ ). This is called a duality transformation. Since the square
lattice is self-dual, the two rewritings are in terms of the same polygon expansion, so we
have related singularities of Z at two sets of parameter values.
It might of course be that Z really has a pair of distinct singularities (critical points),
but if we suppose —as seems more likely— that there is a unique critical point, then it
must occur where
wx = wy∗ , wy = wx∗ .
It is easy to show the involution property (wx∗ )∗ = wx , so there is actually only one relation:
tanh Kx = e−2Ky . (24)
This fixes the selfdual manifold along which the Ising model is critical.
The critical properties (critical exponents) are identical all along the self-dual variety
(24). Indeed, in the continuum limit we can just scale the x and y directions with opposite
scale factors, so that the system becomes isotropic. This is known as anisotropic scaling.
The advantage is that we may then solve the Ising model by going to the completely
anisotropic limit, where Kx 1 and Ky 1, while maintaining the relation (24). In this
limit the dynamics of the system simplifies considerably, since all interactions are close to
the identity in the transfer matrix formalism (where we transfer along the y-direction).
We now give the details of this solution.
• Transfer matrix and Hamiltonian
A spin configuration along a horizontal row of the lattice is denoted S ~1 , with S

~1 =
{S1,y , S2,y , . . . , SNx ,y }. The transfer matrix T transfers this into another configuration S ~2
in which y has been replaced by y + 1: S ~2 = {S1,y+1 , S2,y+1 , . . . , SNx ,y+1 }. Explicitly this
reads
Nx
Y
~2 |T |S
hS ~1 i = exp (Kx Sx,y Sx+1,y + Ky Sx,y Sx,y+1 ) .
x=1
We can separate the horizontal and vertical interactions by writing T = Tv Th , with

Nh
Y
~ 1 |Th |S
hS
e ~1 i = ~ 1, S
exp (Kx Sx,y Sx+1,y ) δ(S
e ~1 ) ,
x=1
Nh
Y
~2 |Tv |S
hS ~ 1i =
e
exp Ky Sex,y Sx,y+1 .
x=1
46
The individual factors in each product commute, but Th does not commute with Tv .
However, in the completely anisotropic limit (Kx 1 and Ky 1) both matrices are
close to the identity, up to an unimportant overall factor in Tv . Indeed we can write
Th = e−Hh , Tv = eKx Nx e−Hv ,
where the matrices Hh , Hv 1. In particular we have
e−Hh e−Hv ≈ (1 − Hh )(1 − Hv ) ≈ 1 − Hh − Hv ≈ e−Hh −Hv .
To leading order, Th reads

Nx
!
X
~ 1 |Th |S
hS
e ~ 1, S
~1 i = δ(S
e ~1 ) 1 + Kx Sx,y Sx+1,y
x=1
and in terms of the Pauli matrices

1 0 1 2 0 −i 3 1 0
τ = , τ = , τ =
1 0 i 0 0 −1
we have
Nx
X
Th = 1 + Kx τx3 τx+1
3
.
x=1
Here the subscript of the Pauli matrices indicate on which tensorand they act. Re-
exponentiating we get the horizontal part of the Hamiltonian:
Nx
X
Hh = −Kx τx3 τx+1
3
.
x=1
Let us now examine one factor in Tv , acting at position x:
e−2Ky

ex,y Sx,y+1 Sx,y+1 1
h i
Ky S Ky
e = e
S
ex,y e−2Ky 1
= eKy 1 + e−2Ky τ 1 ≈ eKy exp e−2Ky τ 1 .

Combining this, we arrive at

Nx
X
Hv = −e−2Ky τx1 .
x=1
Summarising our achievements this far, if we write the transfer matrix as
T = eKy Nx e−H ,
and introduce the parameters
γ = e−2Ky , β = Kx ,
47
then the corresponding 1D quantum Hamiltonian is

Nx
X Nx
X
H = −γ τx1 −β τx3 τx+1
3
, (25)
x=1 x=1
This is known as the 1D transverse field Ising spin chain. By the argument of anisotropic
rescaling, its critical behaviour is identical to that of the original 2D Ising model.
• Ordered phase
When γ β, there are two degenerate ground states of H with energy E0 = −βNx :

1 1 0 0
| ⇑i = ⊗ ··· ⊗ , | ⇓i = ⊗ ··· ⊗ .
0 0 1 1
They correspond of course to the two ordered states in the 2D formulation of the Ising
model.
An elementary excitation is obtained by reversing one of the Nx spins, and the corre-
sponding energy is E1 = −βNx + 2β. Recalling (25) the corresponding eigenvalues of T
are then
λ0 = eKy Nx e−E0 , λ1 = eKy Nx e−E1 ,
so the correlation length is finite:

λ0 1 1
ξ = log = = .
λ1 E1 − E0 2β
• Disordered phase
When γ β, we look instead for the dominant eigenvector of τ 1 . The ground state
of H is then

1 1
⊗ ··· ⊗
1 1
with energy E0 = −γNx . This corresponds to the completely disordered state in the 2D
formulation: a equally weighted superposition of all possible states.
An elementary excitation is obtained by replacing one of the factors in the tensor
product by the other eigenvector of τ 1 , which is (1, −1) with eigenvalue −1. The energy
is then E1 = −γNx + 2γ, and we have again a finite correlation length:
1 1
ξ= = .
E1 − E0 2γ
Below we shall show that the exact expression for ξ, valid for any values of γ and β is
1
ξ= .
2|γ − β|
48
Notice that this diverges for γ = β, so this must correspond to Tc . This matches indeed
the selfduality criterion (24) in the completely anisotropic limit Kx 1:
Kx = e−2Ky ,
in agreement with (25).

• Jordan-Wigner transformation
We rotate the Pauli matrices by means of a unitary transformation to obtain the
equivalent Hamiltonian
Nx
X Nx
X
H = −γ τx3 − β τx1 τx+1
1
.
x=1 x=1
Introduce the raising and lowering operators τx± = 12 (τx1 ± iτx2 ). They satisfy the anticom-
muting relations
(τx+ )2 = (τx− )2 = 0 , τx+ τx− + τx− τx+ = 1 ,
which makes the problem start looking fermionic. However, these τ ± commute on two
different sites, so a trick is needed to obtain the required anticommutativity also in this
case.
This is accomplished by the Jordan-Wigner transformation:
X x x
X
ax = exp iπ Sy− Sy+ Sx+ , a†x = exp − iπ Sy− Sy+ Sy− ,
y=1 y=1
It is straightforward to show that we now have the required fermionic relation
a†x ay + ay a†x = δx,y .
The Hamiltonian is then

Nx Nx
a†x − ax a†x+1 + ax+1 .
X X
H=γ a†x ax − ax a†x − β (26)
x=1 x=1
It is quadratic in the fermion operators.

• Diagonalisation of H
By means of a discrete Fourier transformation
1 X ikx 1 X −ikx †
ax = √ e ak , a†x = √ e ak
Nx k Nx k
this becomes
X X †
H=γ a†k ak − ak a†k − β ak − a−k a†−k + ak eik ,
k k
49
2π
where the sum is over the Brillouin zone, k ∈ [−π, π], in steps of Nx . Note that this
couples only the wave numbers k and −k, so we can write
X
H= Hk ,
k≥0
where Hk reads in matrix notation

γ − β cos k −iβ sin k

ak

†
Hk = 2 ak a−k .
iβ sin k −γ + β cos k a†−k
a†−k
This latter form can be diagonalised by successively applying the relabelling a†−k = ie
and a−k = −ie a−k , and the Bogoliubov transformation (orthogonal rotation)

ak cos φk sin φk ck
= .
a†−k
e − sin φk cos φk c†−k
Omitting the details, the result is

hk c†k ck − c−k c†−k ,
X
H= (27)
k>0
where 1/2
hk = (γ − β)2 + 4γβ sin2 (k/2)

. (28)
The ground state of (27) is obtained by making only the minus term in the parenthesis
contribute, viz., by leaving all the fermionic modes empty. An elementary excitation
corresponds to populating just one fermionic mode q. The excitation energy is
E1 (q) − E0 = 2hq .
The lowest-lying excitation corresponds to q → 0, so we get the correlation length

1 1
ξ= = (29)
E1 (q → 0) − E0 2|γ − β|
as already announced. √
If we define the “mass” m = |γ −β|/ γβ, then the excitation energy has the relativistic
form
(E1 (q) − E0 )2 ∝ m2 + q 2 , (30)
and the mass vanishes at Tc . In other word, at the critical point we have a massless field
theory. According to (27) it is a theory of free fermions, but of a particular type. Namely,
the modes with positive (resp. negative) momentum k have an energy that increases when
k becomes more positive (resp. negative), so they move to the right (resp. left) in Fourier
space. Therefore they are often called right-movers (resp. left-movers). Such particles are
called Majorana fermions.
By contrast, the standard Dirac fermions move left or right for any k, independently
of the sign. (There is one more type of free fermions known as symplectic fermions.)
50
3.4 Partition functions and path integrals

The aim of the two following sections is to point out analogies and relations between models
of statistical physics and of quantum mechanics. These relations are already present, more
or less explicitly, in the transfer matrix formulation of lattice statistical models —which
is grounded on time evolution of time slice ‘spin configuration states’.
We first start by describing a path integral representation of the partition function
of a quantum system at finite temperature. We assume the reader is familiar enough
—or at least know or heard about it— with Feynman path integral representation of
Quantum Mechanics. This will be done by implementing an analytic continuation in the
time variable known as the “Wick rotation”. We then extend this construction to field
theory.
• Feynman path integrals
In quantum mechanics the evolution operator U (t) is solution of the Schrödinger equa-
tion (we shall later put ~ = 1)
i~∂t U (t) = H(t)U (t),
with H(t) the system hamiltonian, so that the system state |ψ(t)i at time t is U (t)|ψ0 i,
with |ψ0 i its initial state. For time independent hamiltonian H, it is simply U (t) = e−itH .
It may be explicitly written if one knows the spectral decomposition of the hamiltonian.
Feynman path integral is an alternative representation of this evolution operator as some
over all possible system histories. Let q be the dynamical variables of the system and
|qi are the eigen-position states. Then, the path integral representation of the matrix
elements of the evolution operator reads:
Z
i
0
hq |U (t)|qi = q(0)=q [Dq] e ~ S[q] (31)
q(t)=q 0
with S(q) the classical action evaluated along the trajectories s → qs . The integral sum
is formally over all paths starting at q and ending at q 0 . This is similar to the path
integral representation of the Brownian motion discussed in a previous chapter except for
the important complex extra factor ‘i’ inherent to quantum mechanics. See any Quantum
Mechanics text book for more details.
• The Euclidean evolution operator and Wick’s rotation.
Wick rotation is based on the fact that the evolution operator U (t) is analytic in
the lower half complex plane =m t < 0. Indeed, if ψα form an ortho-normalized eigen-
basis of H (which is supposed to be time independent) with eigenvalue Eα , we have
i
U (t) = α |ψα i e− ~ tEα hψα |. Its matrix elements between position eigenstates |qi and hq 0 |
P
are X i
hq 0 |U (t)|qi = ψα (q 0 ) e− ~ tEα ψα (q),
α
with ψα (q) = hq|ψα i. All these matrix elements are indeed analytic for =m t < 0 if
the hamiltonian is bounded from below (because the series is absolutely convergent for
51
=m t < 0). This analyticity property is valid provided its hamiltonian is bounded from
below.
Thanks to this property, we can analytically continue the time evolution operator from
real time to complex time t → −iτ with τ real and called the “Euclidean time”. This
analytic continuation is called the Wick rotation. Let UE be this analytic continuation.
It satisfies the differential equation
−∂τ UE = H UE ,
where the hamiltonian H can be time dependent or not. For H time independent, UE =
e−τ H (because −it = −τ ). For a particle in a potential V (q), this is a kind of heat kernel,
solution of a generalized heat equation of the form
∆q
∂ τ UE = − − + V (q) UE .
2m
The Euclidean evolution operator also admits a path integral representation
Z
hq |UE (τ )|qi = q(s=0) =q [Dq] e−SE [q] ,
0
(32)
q(s=t) =q 0
with Euclidean action SE defined as (here the integration variable s is again the Euclidean
time) Z τ
1
ds q̇s2 + V (q) .

SE [q] =
0 2
Notice the difference of signs between the real time action (defined by a Lagrangian)
and the Euclidean action (defined after analytic continuation). These differences can
be justified/derived in different ways: either using the previous construction based on a
discretization of the path integral —as for the path integral of quantum mechanics or for
the Brownian motion—, R or directly by noticing that under the Wick rotation q̇ 2 → −q̇ 2
2 2
R
and ids → ds so that i ds[q̇ −V (q)] → − ds[q̇ +V (q)], or equivalently iS[q] → −SE [q].
For a free particle, in absence of external potential, UE is simply the heat kernel and
the Euclidean path integral with action SE [q] = 12 ds q̇s2 is that of the Brownian motion.
R
Hence the Euclidean path integral for a particle in a potential V can be written as Brownian
expectation (and hence it is better defined its the real time analogue). Namely,
Z
R
[Dq] e−SE [q] = E e− ds V (Xs ) ,

where E[· · · ] the expectation over the Brownian motion Xs . This representation of the
Euclidean evolution operator as a Brownian expectation is called the Feynman-Kac rep-
resentation. Note that it is slightly different form the representation of SDEs with drift
which we previously discussed.
• Path integral representation of partition functions.
The euclidean evolution operator is UE = e−τ H . It may be compared with the density
matrix of a system with hamiltonian H at temperature kB T = 1/β which reads:
ρGibbs = Z −1 e−βH .
52
Here Z = Tr(e−βH ) is the partition function, a key quantity of statistical physics at

equilibrium (recall that F = −kB T log Z is the free energy). Computing the partition
function amounts the trace of the Euclidean evolution operator with τ = β = 1/kB T , that
is Z
−βH
Z = Tr(e ) = dq hq|UE (τ = β)|qi.
This of course can be represented by a path integral with periodic boundary condition,
with period β:
Z
Z= [Dq] e−SE (q) . (33)
q(β) =q(0)
Periodicity is a way to implement the trace sum. This is an important formula, because
it generalizes to any quantum systems and in particular to quantum field theory.
We can similarly write path integral representations of thermal expectations of opera-
tors. If A is an observable of the quantum system with hamiltonian H, its expectation at
temperature 1/β is defined to be
Tr(A e−βH )
hAiβ = .
Tr(e−βH )
Its path integral representation is thus given by the ratio

−SE (q)
R
q(β) =q(0) [Dq] A(τ0 ) e
hAiβ = R
−S (q)
.
q(β) =q(0) [Dq] e E
For instance, if the observable A is a function A(q) of the operator q, its insertion amounts
to insert A(q(τ0 ) ) in the path integral. Again we have periodic boundary conditions to close
the trace. The time τ0 at which we insert the operator A does not matter (because of the
periodicity we can always split e−βH is two pieces of move them around the trace).
Note that physical quantities, such as expectations or correlations, are ratio of path
integrals. Taking this ratio improves some of the problems of the (any way ill-defined)
path integral measure because the undefined normalization cancels between the numerator
and the denominator.
• Application to field theory.
It is a small step to go (formally) from high dimensional hamiltonian systems to field
theory. To make it concrete one may discretized space on a lattice with vertices at points
xk . The dynamical variables are then countable variables φ(xk ) so that one has the
correspondance
{q} ↔ {φ(xk )}.
One may also take (formally) the continuous limit and view the value of the field φ(x)
at space point x as the dynamical variables. They don’t form a countable set but we
nevertheless apply the previous formal manipulations.
53
A (classical) field theory comes equipped with an action S[φ], depending on the field
configurations, which determine the field dynamics. By hypothesis, it can be written in
terms a Lagrangian density L(φ, φ̇) as
Z
S[φ] = dtdx L(φ, φ̇).
The simplest example is that of free massive field for which the action is:
Z Z
1 1
S[φ] = dtdx L = dtdx (φ̇)2 − (∂x φ)2 − V (φ)

2 2
where V (φ) = 21 m2 φ2 with m the mass of the field.
This classical field theory may (formally) be quantized using the previous path integral
formalism. Upon quantization, a basis of the Hilbert space of the quantum field theory
is (formally) made of eigen-field configuration states |{φ(x)}i. These are the analogue of
the eigen-position states |qi. To specify the quantum dynamical we may give ourselves
two configurations at two different times (at two different time slices), {φ0 (x)} at time t0
and {φ1 (x)} at time t1 , and postulate that the quantum amplitude is (formally) defined
by the path integral
Z
h{φ1 (x)}|U (t1 , t0 )|{φ0 (x)}i = φ(x,t )=φ (x) [Dφ] e(i/~) S[φ] ,
0 0
φ(x,t1 )=φ1 (x)
with the boundary condition specified by the two field configurations. The integral is over
‘path’ in the space of field configurations.
We can of course Wick’s rotate this path integral, as we did for quantum mechanics.
This yields a Euclidean path integral over field configurations in one dimension higher (we
set ~ = 1): Z
[Dφ] e−SE [φ] ,
with Euclidean action SE [φ] = dx 21 (∇φ)2 + V (φ) where ∇ is the gradient derivative
R
along all directions (space and Euclidean time). Here we use the notation x = (τ, x),
with τ the Euclidean time, which effectively span a space of one dimension higher than
that of x. In particular the partition function Z(β) = Tr(e−βH ) of a quantum field at
temperature 1/β is given by the Euclidean path integral with periodic condition along the
Euclidean time: Z
Z(β) = [Dφ] e−SE (φ) .
φ(β,x)=φ(0,x)
To make these formal definitions workable is what quantum and statistical field theory is
about.
It is worth having in mind a geometrical picture related to these path integral rep-
resentations. If the quantum field theory is defined on the infinite line the path integral
representation of its partition function takes place over a cylinder of circumference β and
infinite height (the long axis of the cylinder is the infinite line on which the quantum
system is defined, the periodic circle is associated with the finite temperature and the pe-
riodicity comes form taking the traces). If the quantum field theory is defined on a finite
54
interval, then the path integral takes place on a cylinder of finite length. If the quantum
system is periodic with period R (i.e. it is defined on a circle of radius R), the path
integral takes place on a torus. Generalizations to higher dimensional quantum systems
is clear. If the field theory is defined over a manifold M, then its partition function path
integral takes values on M × S 1 with S 1 the circle of period β. It may be useful to have
these geometrical pictures in mind. See Figure.
3.5 The classical/quantum correspondence

The previous constructions provide instances of a correspondence between the physics
of quantum systems and that of classical systems but in one dimension higher —a cor-
respondence which was already transparent in the transfer matrix formulation of lattice
statistical models. The aim this section is to spell out this correspondence and to establish
the dictionary between these two setups
• D-classical/(D − 1)-quantum correspondence.
By interpreting the Euclidean path integral as the partition function of a classical
system whose Boltzmann weights e−SE we get a correspondence between the physics of
quantum system in dimension d at finite temperature and that of classical system in
D = d + 1 dimension.
Consider a lattice statistical model defined on hyper-cubic lattice in dimension D
with spin variables Si0 ,i1 ,··· ,iD−1 on each point of the lattice with (i0 , i1 , · · · , iD−1 ) the
coordinates the vertices. We can choose one of the direction, say the first one corresponding
to the index i0 , and declare it to be the Euclidean time direction. Accordingly we can
slice the D-dimensional hyper-cube as (D − 1) hyper-surfaces equally spaced along the
Euclidean time direction. The set of the spin variables on these (D − 1) hyper-surfaces
may be viewed as the dynamical variables of a quantum systems. Collecting all values of
those variables on all slices along the chosen Euclidean time direction gives a Euclidean
time trajectories. The energy of the configuration of all spins variables on all hyper-
surfaces slices (i.e. on the original hyper-cube) is then identified with the Euclidean action
of these trajectories. Hence, if this identification applies, the partition sum of the classical
statistical model is identified with the (discretized) path integral of these trajectories.
This mapping is particularly explicit in the transfer matrix formulation of lattice sta-
tistical models. The transfer matrix T is identified with the Euclidean evolution over a
lattice distance δ, that is we have the identification T = e−δH with H the hamiltonian. For
a statistical model in dimension D, it acts on Hilbert space defined over an hypersurface
of dimension d = D − 1 which is identified with the quantum Hilbert space.
• Thermal expectations, energy gap and correlation length.
Through the previous correspondance we associate a D-dimensional classical statistical
systems with partition function Z with a (D − 1) quantum systems with hamiltonian H.
The quantum system is thought to be defined on the hyper-surface orthogonal to one of
the direction classical system which, by convention is called the Euclidean time direction.
This correspondance is such that if the classical system is made periodic, then
Z(L) = Tr(e−LH ),
55
where the trace is over the quantum system Hilbert space.

Accordingly we have the correspondance:
D − dim. ”classical” system (D − 1) − dim. ”quantum” system

lattice spacing ↔ discretization
configuration energy βEc ↔ Euclidean action SE
periodic system L ↔ finite temperature β
inverse correlation length ξ −1 ↔ energy gap ∆E = E1 − E0
The first three statement are clear and were already discussed at length. The last
statement about the correlation length needs an explanation. It requires computing one
and two point correlation functions.
Let us first look at the expectation of some observable A. We first (and provisionally)
makes the Euclidean time direction periodic of period L. We will later take the large L
limit. Following the above strategy we may represent the expectation values as
Tr e−LH A

hAiL = ,
Tr e−LH )
where Z(L) = Tr e−LH ) is the partition function with H the hamiltonian coding for the
evolution along the Euclidean time. By decomposing on the eigen-state of H, we write
1 X −LEn
hAiL = e Ann ,
Z(L) n
where En are the eigen-values of HPand Ann the matrix elements of A in the eigen-basis.
The partition function is Z(L) = n e−LEn . In the large L limit, the minimum energy
E0 dominates (assuming that there is a gap). This projects the sum on the ground state
|0i. The vacuum energy compensate (as it should be) in the ratio, and we get
hAiL 'L→∞ A00 .
A formula which may be summarized as: taking the infinite size limit projects on the
ground state.
Let us now pick B and A two operators and consider their two point functions at
different sites at position τ1 and τ2 along the Euclidean time direction. Let ` = τ2 − τ1 > 0
be their distance. We may then represent this correlation function as follows:
Tr e−(L−`)H B e−`H A

hB(`)A(0)iL = ,
Tr e−LH )
with ` = τ2 − τ1 > 0 the distance between the two insertion points. We want to know how
this correlation function behavior at large distance but small compare to the system size,
i.e. 1 ` L. Decomposing the trace on the hamiltonian eigenstates gives
1 X −(L−`)En
hB(`)A(0)iL = e Bnm e−`Em Amn .
Z n,m
56
Imposing 1 ` L amounts to take the limits L → ∞ then ` → ∞ is this order. Again

the eigen-state with minimum energy dominates in the large L limit, and
X
hB(`)A(0)iL=∞ = B0m e−`(Em −E0 ) Am0 .
m
The dominating contribution in this last sum again comes from the vacuum energy with
the terms B00 A00 = hBihAi. Let us defined the connected correlation function by
hB(`)A(0)iconn.
L := hB(`)A(0)iL − hB(`)iL hA(0)iL .
In the large L limit we have hB(`)A(0)iconn. −`(Em −E0 ) A

P
L=∞ = m6=0 B0m e m0 . Hence, the
dominating contribution at large distance ` 1 comes the first excited state and we have
−`/ξ
hB(`)A(0)iconn.
L=∞ ' const.e
with
ξ −1 = E1 − E0 .
By definition, ξ is then the correlation length. Notice that the decay is exponentially fast
only if the gap is not vanishing (in the infinite size limit).
3.6 Exercises
• Exercise 3.1: Fermionic representation of the 2D Ising model

The aim of this exercise is to complete the study of the 2D Ising model presented in
the main text. Recall the definition of the 2D Ising model given in the text.
(i) Prove –or argue for– the expression for the Ising transfer matrix.
(ii) Recall the Jordan-Wigner transformations given in the main text which construct
fermionic operators in terms of Pauli matrices via
Px − + Px − +
ax = eiπ y=1 τy τy τx+ , a†x = e−iπ y=1 τy τy τx− .
Show that we may alternatively write

x−1
Y x−1
Y
τyz τx+ , a†x = τyz τx− .

ax =
y=1 y=1
Verify that they satisfy the canonical fermionic relation a†x ay + ay a†x = δx,y .
(iii) Complete the proof of the diagonalisation of the Ising hamiltonian and its spectrum.
Proof that, after an appropriate Bogoliubov transformation on the fermion operators, the
Ising hamiltonian (25) can be written in the form (27) given in the main text, which we
recall here, X †
H= hk ck ck − c−k c†−k ,
k>0
1/2
with single particle spectrum hk = (γ − β)2 + 4γβ sin2 (k/2)

.
57
• Exercise 3.2: Spin operators, disorder operators and parafermions.

The aim of this exercise –and the following two– is to study some simple consequences
of group symmetry in lattice statistical models.
Let us consider a lattice statistical model on a two dimensional square lattice Λ :=
a Z2 with spin variables s on each vertex of the lattice. These variables take discrete or
2
continuous values, depending on the models. We consider neighbour spin interactions with
a local hamiltonian H(s, s0 ) so that the Boltzmann weight of any given configuration [c] is
Y
W ([c]) := w[i,j] , w[i,j] = e−H(si ,sj ) ,
[i,j]=edge
where, by convention, [i, j] denotes the edge connecting the vertices i and j. Let Z :=
P
[c] W ([c]) be the partition function.
Let us suppose that a group G is acting the spin variables. We denote by R the
corresponding representation. Furthermore we assume that the interaction is invariant
under this group action so that, by hypothesis,
H(R(g) · s, R(g) · s0 ) = H(s, s0 ), ∀g ∈ G.
(i) Transfer matrix: Define and construct the transfer matrix for these models.
(ii) Spin operators: Spin observables, which we denote σ(i), are defined as the local inser-
tions of the spin variables at the lattice site i. That is: σ(i) is the function which to any
configuration associate the variable si .
Write the expectations of the spin observables hσ(i1 ) · · · σ(iN )i as a sum over configura-
tions weighted by their Boltzmann weights.
Write the same correlation functions in terms of the transfer matrix.
(iii) Disorder operators: Disorder observables are defined on the dual lattice and are
indexed by group elements. Let Γ be a closed anti-clockwise oriented contour on the
square lattice Λ̃ dual to Λ –the vertices of Λ̃ are the center of the faces of Λ. Let ` denote
an oriented edge of Γ. It crosses an edge of Λ and we denote by `− and `+ the vertices of
this edge with `− inside the loop Γ. The disorder observable µg (Γ) for g ∈ G is defined as
X
µΓ (g) := exp (H(s`− , s`+ ) − H(s`− , R(g)s`+ ) ,
`∈Γ
Inserting µΓ (g) in the Boltzmann sum amounts to introduce a defect by replacing the
hamiltonian H(s`− , s`+ ) by its rotated version H(s`− , R(g)s`+ ) on all edges crossed by Γ.
Write the expectations of disorder observables in terms of the transfer matrix.
• Exercise 3.3: Symmetries, conservation laws and lattice Ward identities
The aim of this exercise is to understand some of the consequences of the presence of
symmetries. The relations we shall obtain are the lattice analogue of the so-called Ward
identities valid in field theory.
We consider the same two dimensional lattice model as in previous exercise. We recall
that we assume the Bolztmann weight to be invariant under a symmetry group G in the
sense that
H(R(g) · s, R(g) · s0 ) = H(s, s0 ), ∀g ∈ G.
58
(i) Let ik be points on the lattice Λ and Γ a contour as in previous exercise.. Show that
the group invariance implies that
Y Y Y
hµΓ (g) σ(ik )i = Rik (g) · h σ(ik )i,
k ik inside Γ k
where Rik (g) denote the group representation R acting on the spins at site ik .
(ii) Show that µg (Γ) is invariant under any smooth continuous deformation of Γ as long
as the deformation does not cross points of spin insertions (it is homotopically invariant).
We now look at the consequences of these relations for infinitesimal transformations.
Suppose that G is a Lie group and Lie(G) its Lie algebra. Let us give a name to small
variations of H by defining ∂X H. For g = 1 + X + · · · with X ∈ Lie(G), we set
H(s, R(g)s0 ) − H(s, s0 ) =: ∂X H(s, s0 ) + · · · .
For ` = [`− , `+ ] an oriented edge of Γ as in previous exercise and X ∈ Lie(G), we let
∗J`X := ∂X H(s`− , s`+ ),
They are specific observables, called currents, whose correlation functions are defined as
usual via insertion into the Bolztmann sums.
(i) Show that the following equality holds:
X Y X Y
h ∗J`X · σ(i)i = h Rik (X) · σ(i)i,
`∈Γ i ik inside Γ i
if some spin observables are inserted inside Γ.

(ii) Deduce that, if there is no observables inserted inside Γ, then the following equality
holds inside any expectation values:
X
∗J`X = 0,
`∈Γ
That is: The Rsecond of these two equations is a conservation law (i.e. it is the analogue of
the fact that ∗J = 0 if ∗J is a closed form, or equivalently, if J is a conserved current),
the first tells about the consequences of this conservation law when insertion of observables
are taken into account. It is analogous to the Gauss law in electrodynamics. They are
called Ward identities in field theory.
• Exercise: Lattice gauge theory.
[...To be completed...]
• Exercise: Self-avoiding walks and the O(n)-models
59
4 From statistical models to field theories
The Chapter aims at making the first steps into the field theory landscape by illustrat-
ing how field theories provide representations of statistical models. These are particularly
efficient when the underlying statistical models exhibit second order phase transitions,
because the latter are accompanied with large scale fluctuations such that coarse-grained
field theory descriptions are relevant. We describe basics facts about second order phase
transitions and critical systems. These include scaling exponents, coding for the singular
behavior of thermodynamic quantities at the transition, and the important notion of uni-
versality. We explain the Landau theory which is based on mean field approximations as
well as its domain of validity. First steps on how to go beyond mean field theory, away
from its domain of validity, are discussed.
4.1 Why do we care about second order phase transitions ?

Classical thermodynamics classifies phase transitions according to the behavior of the
thermodynamical functions at the transition. (See standard textbook for basic facts about
phases transitions). For first order phase transition the first derivative of the free energy
is discontinuous at the phase transition point. For second order phase transition, the first
order derivative is continuous but there is a discontinuity in a higher order derivative of
the free energy. Second order phase transition are characterized by large scale fluctuations
and correlations.
• Why do we care about second order phase transitions ?
(i) Second order phase transitions (with singularities in the free energy) can only occur
in the thermodynamic limit (i.e. when there are an infinite number of degrees of freedom
in the system, as otherwise the partition function is a finite polynomial sum). Phase
transitions are key examples of collective phenomena. Analysing them lead to understand
how many (actually infinitely many) constituents act together to give new physics.
(ii) Phase transitions have universal properties. Phases of matter are so diverse so that
many of the properties of these phases are “non-generic”. However, second order phase
transitions turn out to fall into a relatively small number of classes known as “univer-
sality classes” which all behave similarly. Understanding the phase transitions also helps
understanding the key differences between these phases.
(iii) The mathematical structure developed to study phase transitions (statistical field
theory and the renormalization group) turns out to be one of the most fundamental tool
to understand non-trivial physics (say involving infinitely large numbers of degrees of
freedom, at all scales), with applications to condensed matter and high energy physics.
Standard examples of second order phase transitions include :
— Para/Ferromagnet transition: Above the critical temperature Tc , the magnetization is
zero, whereas below the critical temperature the magnetization becomes non zero. For a
second order (and hence continuous) phase transition the magnetization vanishes at the
60
critical temperature (in absence of external magnetic field);

— Bose condensation: Below a critical temperature, the lowest energy orbital becomes
macroscopically occupied (i.e. it becomes on the order of the macroscopic number of
bosons in the system);
— Superconducting or superfluid transition: Below a critical temperature, certain systems
display dissipation-less flow of current or of fluid mass.
— And many more examples,...
All these transitions share a great deal of similarities and they often can be mapped to
the others.
• The Ising model and its mean field theory

The Ising model is one of the simplest model of statistical physics exhibiting a second
order phase transition. Although non exact the mean field theory provides a simple scheme
illustrating the emergence of singular behaviors at a transition.
Let us consider the Ising model on a hyper-cubic lattice in dimension D, with a the
mesh of the lattice a, i.e. the lattice is aD ZD . On each lattice site we put a spin variable
si , i ∈ Λa , which possibly takes two values si = ±1. A configuration [s] is the data of
all spin values at all points of the lattice. The energy of a configuration codes for the
interaction of neighbor spins. By definition, it is given by the formula:
X X
E[s] = −J0 si sj − h si
i∼j i
where the sum is over pairs of neighbour sites. The constant J0 > 0 measures the strength
of the interaction and h is the external magnetic field. The statistical distribution is
specified by the Boltzmann rules: the probabilities for a configuration [s] is proportional
to e−E[s]/T where T is the temperature. Since J0 > 0, aligned configurations are favored
at low temperature. The Ising model possesses a second order phase transition (at zero
external field) at some critical temperature Tc .
The mean field theory consists in making the approximation that at each site the local
spin is effectively sensible only to the“mean field” created by the neighbour spins. There
are different ways of deriving the mean field theory (either using a local field approximation
as done below, or using a variational ansatz as done in the exercise Section). If z is the
number of neighbour (z = 2D in the case of the hyper-cubic lattice), the mean field is
h + zJ0 m̄ with m̄ the local magnetization (m̄ is the meanPvalue hsi of the spin P si ). This
mean field approximation consists in replacing the sum <ij> si sj by z m̄ i si so that
the energy of a spin configuration is then reduced to a sum of independent terms:
X
E[s] → Ē[s] = − (zJ0 m̄ + h) si .
i
The local mean energy is e(si ) := (zJ0 m̄ + h) si . Self-consistency condition demands that
the mean spin (computed using the Boltzmann rules specified by Ē[s]) should be m̄, i.e.
e−βe(+) − e−βe(−)
hsi = m̄, alias m̄ = = tanh(β(zJ0 m̄ + h)),
e−βe(+) + e−βe(−)
61
with β = 1/kB T the inverse temperature (in the following we take kB = 1). This condition
means that we equate the mean computed using a spatial averaging procedure with the
thermodynamic mean. As consequence, and according to the central limit theorem, we
expect this approximation to be better as the spatial dimension increases.
At zero external field (see Figure), the mean field equation is trivial with m̄ = 0 if
T > Tc but non-trivial solution with m̄ 6= 0 for T < Tc (the other solution m̄ = 0 is
unstable). The critical temperature is Tc = zJ0 . That is, we have:
(
if βzJ0 < 1, m̄ = 0, (para phase);
if βzJ0 > 1, m̄ = ±m̄0 6= 0, (ferro phase).
At h = 0, and close to the critical point, one obtains the magnetization by simply
expanding the self consistency condition in power of the magnetization. This yields
1 Tc − T
(zβJ0 m̄)3 = m̄.
3 T
Hence for T > Tc , we recover m̄ = 0, while for T < Tc we get m̄0 ' (Tc − T )1/2 . The
phase transition is thus continuous. It is a second order phase transition.
For h 6= 0 but at the critical temperature, the mean field equation reads m̄ = tanh(m̄+
βc h). Hence m̄0 ' h1/3 since tanh x ' x − 31 x3 + O(x5 ).
Similarly, writing the mean field equation at small external field and at T 6= Tc but
close to the critical point yields (∂ m̄/∂h)h=0 ' |T − Tc |−1 for the magnetic susceptibility
(for instance by evaluating the derivative of the mean field equation w.r.t to external field
at h = 0).
4.2 Field theory representation of statistical models

The aim of this section to show on exemples how models of statistical physics can written,
or reformulated, in terms of fields theory. The latter may be defined on a lattice or in
the continuum depending whether the original statistical model is defined in the discrete
or in the continuum, or depending whether one is adopting a large scale coarse-grained
description or not. We shall make explicit the dictionary between objects, concepts, or
quantities used in field theory or in usual models of statistical physics.
• Heuristics
Field theory representation of macroscopic statistical models may be viewed as aris-
ing from a coarse-grained approximation, at mesoscopic scales, but taking into account
fluctuations of the mesoscopic degrees of freedom.
Let us consider a statistical spin system and let us imagine dividing our system into
mesoscopic cells, large enough compared to the microscopic scale but still very small
compared to lengths of interest (i.e. we want to be able to take averages within the cells
and to deal with them as if they were macroscopic enough sub-systems). We can then
divide our partition function into sums over a large number of individual cells
X X X
Z= e−βE[s] = e−βE[s] ,
P
[s] [m] [s: i∈cell si =m]
62
with [s] the microscopic spin configurations, E[s] their energy and β the inverse tempera-
ture. The last sum is restricted such that the sum over all the spins in the cell number k
is equal to the magnetization mk . Since all cells are macroscopic sub-systems, we assign
them a free energy given the values of the magnetizations in the different boxes, so that
the partition function reads
Z Y
Z → [ dmk ] e−F [m1 ,··· ,mk ,··· ] .
k
We are defining a free energy which is a function of a magnetization, the only difference
is that here the free energy is a function of the magnetization in all of the different cells.
We have absorbed β into the definition of F . The sum over cells has been converted into
integrals (since the cells are large enough that the averaged magnetization is essentially
continuous) and then the integration over all of the small cells is essentially what we mean
by a functional integral (over the magnetization configurations)
Z
Z ' [Dm] e−F [m] .
Under the assumption that there are no long range interactions and that the mesoscopic
cells are larger than the interaction length scale, we may be convinced Rthat the free energy
functional F [m] as an integral over a local energy functional F [m] = dD x f [m(x)].
• Field theory representation of the Ising model
Here we present an exact reformulation of the Ising model in terms of dual variables.
As a consequence the Ising model is going to appear as a discrete scalar field theory. This
reformulation gives a precise illustration of the previous heuristics and goes beyond the
Landau theory that we shall discuss in the following Section.
We start with the Ising model on the square lattice with partition sum (we have
absorbed the inverse temperature β into J and h)
X P P
Z= e ij Jij si sj + i hi si .
{sk }
We represent the spin-spin interaction as follows (this is called the Hubbard-Stratonovich

transformation, it is simply based on Gaussian integral techniques):
Z Y
P 1 P −1 P
e ij Jij si sj
= const. [ dφk ] e− 4 ij φi Jij φj + i φi si .
k
Here, the φk ’s are scalar variables at each site of the lattice. Once this identity is inserted
into the partition sum we can explicitly do the sum over the spin configurations (at fixed
φ) using
X P Y P
e i (φi +hi )si = [2 cosh(φi + hi )] = const. e i log[cosh(φi +hi )] .
[s] i
63
Hence, up to an irrelevant multiplicative constant (proportional to [det[J]|−1 ), the Ising

partition sum reads
Z Y
1 P −1 P
Z = [ dφk ] e− 4 ij φi Jij φj + i log[cosh(φi +hi )] ,
k
P
Equivalently, we can change variables φi → j Jij φj (at the prize of a Jacobian propor-
1P
tional to [det[J]|) so
P that the first quadratic
P term is transformed into 4 ij φi Jij φj and the
second term into i log[cosh(hi + j Jij φj ). The integral is convergent by construction
(and the expansion we are going to do are legitimate if we do not spoil this convergence
property). This Ising partition function can thus be written in the form
Z Y
Z = [ dφk ] e−S[φ;h] ,
k
with Boltzmann weight e−S[φ;h] where S[φ; h] the so-called action is given by
1X X X
S[φ; h] = − φi Jij φj + log[cosh(hi + Jij φj )].
4
ij i j
We have thus alternatively written the partition function in terms of scalar variables
φk . It specified a probability measure on configurations of the discrete field φ. We thus
have reformulated the Ising model as a discrete field theory, that is,
2 4
Let us look at what happens at hi = 0. Expanding log[cosh(x)] = x2 − x12 + · · · , the
action for φ is then
1X 1X 1
S[φ; h = 0] = φi Jij φj − φj Jij Jik φk + · · · = hφ, (J − 2J 2 )φi + · · · ,
4 2 4
ij ijk
where,
P to simplify notation, we view Jij as a matrix acting the vector φi and we let hφ, ψi =
j j j . The dots correspond to higher order terms which would induce interactions
φ ψ
between the various components of the field φ.
Remember that theP coupling constants Jij couple only neighbour sites on the lattice
(recall that the terms ij Jij si sj in the energy), i.e. for instance for a square lattice
ˆ J0
Jij = 2 (δi,j−1 + δi,j+1 ) with the convention that adding one to the site index amounts
to move by one step on the lattice. We have Jij = J20 (∆ij + zIij ) with ∆ the Laplacian
on the lattice (say square lattice), I the identity matrix and the number z related to the
number of neighbors. The action for the field φ can then be written as
S[φ; h = 0] = const. hφ, (−∆ + M 2 )φi + higher derivatives × higher order terms,
with M 2 ∝ (1 − J0 z) ∝ (T − Tc ). At this approximation order, the action is quadratic in

the field variable φ.
To be exact the action should include many more terms (actually, an infinite number of
terms) coming form the expansion log[cosh(h + J · φ)], or equivalently log[cosh(h + J20 (∆ +
z) · φ)]. However, as we have discussed in the following Sections, scaling arguments tell us
that only the φ4 term matters, if we are aiming at describing the long distance physics.
64
• Field theory / Statistical model dictionary

Let us spell out the dictionary between objects and tools of (standard) statistical
physics and field theory. The rules of statistical physics assign a probability weight, the
Boltzmann weight, to each configuration of a physical system. Say for a spin system, the
weight of a spin configuration [s] is Z −1 e−E[s] , with E[s] the energy of the spin configura-
tion [s]
P (with the temperature absorb in its definition) and where Z is the partition sum,
Z = [s] e −E[s] .
Discrete field theory defined on a lattice Λ, say covering a domain D ⊂ RD , may be
thought as usual statistical systems but with configurations specified by the values of the
field at each point of the lattice, say φi for all i ∈ Λ. For simplicity, let us assume that the
field takes value on the real line so that φi ∈ R. To each field configuration is assigned a
Boltzmann weight e−S[φ] where S[φ], which is often called the action, is a function of the
field configuration φ, usually with some locality property. Since the field takes values in
a continuous set, summing over all field configurations involves integrations over all the
possible field values.
We can thus draw the following correspondence:
configurations : [s] −→ R [φ], −S[φ]

e−E[s]
P
partition sum : Z= [s] −→ Z = Dφ e ,
Q
with Dφ the integration measure on the field configuration space, i.e. Dφ = k dφk .
Similarly as spin systems can be coupled to an external magnetic field h, the field
configurations can also be coupled to an external source, say j, with the correspondence
P P
external source : E[s; h] = E[s] + i hi si −→ S[φ; j] = S[φ] + i ji φi .
The corresponding partition sums, Z[h] = [s] e−E[s;h] or Z[j] = Dφ e−S[φ;j] , are respec-
P R
tively the generating functions for the moments of the spin or field variables. Correlation
functions of the spin (resp. the field) variables are obtained by taking successive derivatives
of log Z[h] (resp. log Z[j]).
Field theory aims at giving a meaning to the continuous version of these discrete
field theory. They are naively “defined” by taking the formal limit in which the mesh
of the lattice Λ vanishes so that the based points over which the field is defined vary
continuously. As a consequence, the field becomes a fonction φ(x) of the based point
x ∈ D, the integration measure becomes a sum over all possible field configurations and
is formally represented as an integration over the field values at all points in D, and the
action S[φ] becomes a (local) functional of the field configuration:
configurations : [φ] −→ {φ(x),

Q x ∈ D},
integration sum : [Dφ] −→ R Q x∈D dφ(x),
Z = Dφ e−S[φ] −S[φ(x)] ,
R
partition function : −→ Z = [ x∈D R dφ(x)] e
action : S[φ] −→ S[φ] = D dxR L[φ(x)],
external source : S[φ; j] −→ S[φ; j] = S[φ] + D dx j(x)φ(x),
with L[φ(x)] some local functional of φ, often called the energy density or the Lagrangian.
Q
Of course this definition/correspondence is formal (in particular the measure x∈D φ(x)
65
over the infinite dimension space of field configuration has no mathematical meaning) and
one of the aim of field theory is to give a sense of this correspondence.
4.3 Critical systems: phenomenology

Critical systems refer to second order phase transitions. They are characterized by sin-
gularities in the thermodynamic response functions which are echoes of large scale fluctu-
ations. We introduce below the standard notation to parametrize the singular behaviors
of thermodynamical function at a critical point. These singular behaviors are coded into
series of critical exponents which we list just below.
• Second order phase transitions.
Second order phase transitions are continuous phase transitions (no discontinuity in
the response function). The fact that the transition is continuous does not mean that the
behavior is not singular. Although continuous at the transition, the response functions
are not a smooth functions of the external parameters (some of their derivatives develop
singularities).
We shall use the (most standard) example of para-to-ferro transition to illustrate the
phenomena. Typical behavior of response functions at a second order phase transitions
are the followings:
At high temperature T > Tc , the system is in a disordered phase in which the local
spins vary randomly and rapidly from one site to the others. For instance, in the infi-
nite temperature Ising model (T J) the spins at different sites are independent. The
response of the system to a small magnetic field is linear: the mean magnetization m̄ is
linear in the external field and the magnetic susceptibility χ(T ) := ∂ m̄(h)/∂h|h=0 is finite.
Figure.
At low temperature T < Tc , the system is ordered. The spins have tendency to be
aligned, with a large probability, and they are correlated at large distances. For instance,
in the zero temperature Ising model, the most probable configurations are those in which
all spins point in the same direction. There is a non zero spontaneous magnetization and
hence the system is in a ferromagnetic phase. More precisely,
lim m̄(h) 6= 0.
h→0+
in such case, on says that the symmetry is broken (the symmetry is the Z2 action reversing
all spins). This is a key concept in the theory of second order phase transition but also in
particule physics through the so-called Higgs mechanism. Figure
The spontaneous magnetization m̄ (at h = 0) is a function of the temperature: m̄ 6= 0
at T < Tc and m̄ = 0 at T > Tc and the magnetization vanishes continuously at T = Tc .
The magnetization is not a smooth function of the temperature but develops a power law
behavior m̄(T ) ' |T − Tc |β for T < Tc close to the critical temperature. Figure.
At the critical point T = Tc , the system possesses peculiar and anormal properties:
the magnetization at small external field is singular, m̄(h)|T =Tc ∝ h1/δ , the susceptibility
diverges close to the transition, χ(T ) ∝ |T − Tc |−γ , etc. All these exposants β, δ, γ, etc,
are non trivial.
66
These behaviors are linked to the presence of large fluctuations at all scales so that
the physics is dominated by collective phenomena. Critical systems (at T = Tc ) are
characterized by infinite correlation length. In particular there is no intrinsic characteristic
length scale so that these systems are invariant under dilatation.
• Critical exponents.
— For the magnetization as function of the temperature (magnetization exponent):
(
C |Tc − T |β , for T < Tc ,
m̄(T ) '
0, for T > Tc .
Both the critical temperature and the amplitude C are not universal (meaning that they
depend on details of the model systems, say shape of the lattice), but the exponent is
(meaning that it does not depend on such details).
— The magnetic susceptibility (at zero magnetic field) is defined as χ(T ) := ∂ m̄(h)/∂h|h=0 .
It is singular as the temperature approches Tc (it diverges):
(
χ+ |Tc − T |−γ , for T < Tc ,
χ(T ) = ∂ m̄(h)/∂h|h=0 '
χ− |T − Tc |−γ , for T > Tc .
The amplitudes χ± are not universal but their ratio is. The exponent γ is universal (it
could have been different on both side of the transition but this turns out not to be the
case).
— The magnetisation at the critical temperature is also singular as a function of the
magnetic field:
m̄(h)|T =Tc ' σ0 h1/δ .
The writing 1/δ (and not the reverse) is a convention.
— The heat capacity (which is the second order derivative of the free energy w.r.t. the
temperature) diverges at the critical temperature (and this is often taken as a characteristic
of the transition):
(
A+ |Tc − T |−α , for T > Tc ,
C(T ) '
A− |T − Tc |−α , for T < Tc .
The amplitudes A± are not universal but their ratio is. The exponent α is universal.
— The correlation length diverges at the transition. The correlation length codes how
much two distant spins (for magnetic systems) are correlated. It could be defined via the
two point connected correlation spin function, G(i, j) = hSi Sj ic = hSi Sj i − hSi ihSj i, for
spins at positions xi and xj . At T > Tc , hSi i = 0, whereas hSj i 6= 0 at T < Tc . At
distances large enough we expect an exponential decreases of the correlation function
G(i, j) ' const. e−|xi −xj |/ξ ,
with ξ the (so-called) correlation length. It depends on the temperature (and on the
magnetic field which we set here to zero). For T Tc spins are uncorrelated (by the very
67
nature of the disordered phase) but they become more and more correlated as T approches
Tc so that ξ increases as T → Tc+ until it gets infinite as one approaches the ferromagnetic
phase (the direction of all spins in the system are correlated in the ferromagnetic phase).
This divergence also occurs when approaching Tc from below. We have
(
ξ+ |Tc − T |−ν , for T > Tc ,
ξ(T ) '
ξ− |T − Tc |−ν , for T < Tc .
Surprisingly the exponents are identical on both sides of the transition. The amplitudes
ξ± are not universal but their ratio is.
It should be noticed that the correlation length is infinite at the critical temperature
ξ(Tc ) = ∞. This tells us that all length scales are becoming correlated as we approach the
critical temperature. At the critical point there is no characteristic length and the system
behavior is thus expected to be invariant under scale (global dilatation) transformation.
By the usual (folklore) argument, locality promotes this invariance to global conformal
invariance.
At criticality the correlation function decreases as a power law (not exponentially)
1
G(i, j)|T =Tc ' .
|xi − xj |D−2+η
This defines a new exponent (the scaling dimension of the spin operator).
All these exponents are in general not independent. They satisfy (so-called) scaling
relations (not to be proven at this point of the lectures):
α + 2β + γ = 2 , α + βδ + β = 2,
α + Dν = 2 , γ = ν(2 − η).
These relations come from assuming (or proving) that all relevant quantities, such as the
(singular part) of the free energy or the correlation functions, are essentially homogeneous
functions of the external parameters and of the correlation length. But a proof of them
requires analyzing the RG transformations, see below.
To conclude, we present a table of the Ising critical exponents in various dimensions:
ISING ; mean field ; exact 2D ; 3D approx.

−−−−−−−−−−−−−−−−−−− ; − − −− ; − − −− ; − − −−
magnetization : m(T ) ' (Tc − T )β ; β = 1/2 ; β = 1/8 ; β = 0.3264...
critical magnetization : m(h) ' h1/δ ; δ=3 ; δ = 15 ; δ = 4.7898...
magnetic susceptibility : χ(T ) ' |Tc − T |−γ ; γ=1 ; γ = 7/4 ; γ = 1.2371...
correlation length : ξ(T ) ' |Tc − T |−ν ; ν = 1/2 ; ν=1 ; ν = 0.6299...
We observe deviations from the mean field values. They are bigger in lower dimensions.
They are echoes of fluctuations not taken into account in the mean field theory. As we
will discuss in a following Section, below a certain dimension Di , called the lower critical
dimension, those fluctuations are so strong that they destroy any order, above another
dimension Ds , called the upper critical dimension, fluctuations are weak and mean field is
68
exact. In the intermediate regime, Di < D < Ds the fluctuations modify drastically the
nature of the transition.
• Universality
We have already mentioned this important property. Universality means that the
critical behavior near the transition does not depend on the details of the systems or of
the model systems, but only on a few properties such as symmetries, the representation
of this symmetry group associated to the order parameters, etc. For instance the Ising
model is a representative model of the universality class for which the order parameter is
a scalar —here the magnetization— and the symmetry is the Z2 reflexion symmetry of
this scalar. The universality property is an echo the renormalization group analysis.
4.4 Landau-Ginzburg theory

The Landau theory is an effective theory to deal with second order phase transitions, very
much close the Ising mean field theory, but with much wider perspectives and domains of
applications. It is based on the following methodological idea:
(i) Identify a local order parameter (often called m or φ), which describes (macroscopi-
cally) the relevant degrees of freedom close to the transition. This parameter can be a
scalar, a vector, a tensor, etc. but symmetry arguments usually helps identifying it.
(ii) Construct a local free energy functional F [m] by expanding in power of m (or power of
its gradient ∇m) assuming that all fluctuations are small enough. Again symmetry argu-
ments help restricting the possible terms in this expansion. By hypothesis, the expansion
coefficients depend smoothly on the external parameters (T , h, etc.).
(iii) Minimize the free energy to evaluate the thermodynamical functions and the critical
exponents.
The Landau-Ginzburg theory is an extension of Landau theory taking spatial inhomo-
geneities into account.
• The Ising (mean field) universality class.
The order parameter is φ = m, the local magnetization. There is a Z2 -symmetry. For
homogeneous configurations, and homogeneous applied external field h, the free energy is
proportional to the volume F [m; T, h] = Vol. f [m; T, h] where f is the free energy per unit
volume. The coupling to the external field is
f [m; T, h] = f0 [T, m] − hm.
Z2 -symmetry demands that f0 to be even: f0 [T, m] = f0 [T, −m]. Hence we expand in

power of m:
f0 [T, m] = g0 (T ) + g2 (T ) m2 + g4 (T ) m4 + · · · .
We truncate the expansion at order four. For the free energy to be bounded from below
we assume g4 > 0 (otherwise we would have to expand to the next order hoping the next
expansion coefficient is positive). The Taylor coefficient gk are smooth functions of the
temperature. The behavior of f0 as function of m (here the variational parameter) is
69
different whether g2 is positive or negative. Hence the transition temperature Tc occurs

when g2 vanishes. We write (a > 0):
a g
g2 (T ) = (T − Tc ) + · · · , g4 = + ··· ,
2 4!
where the dots refer to higher order terms in T − Tc .
The thermodynamical quantities are determined by minimizing the free energy with
respect to m (this comes from usual arguments of thermodynamics). Writing ∂f /∂m = 0
yields
g
a(T − Tc )m̄ + m̄3 = h.
3!
At h = 0 we get
(
0, for T > Tc ,
m̄|h=0 ' 1
6a 2 1
± g (Tc − T ) 2 , for T < Tc .
This corresponds to an exponent β = 1/2.

At non zero magnetic field h 6= 0, evaluating the derivative (at h = 0) of the above
minimizing equation yields
(
[a(T − Tc )]−1 , for T > Tc ,
χ(T ) = ∂ m̄/∂h|h=0 '
[2a(Tc − T )]−1 , for T < Tc .
This corresponds to an exponent γ = 1.

g 3
At the critical temperature, the minimizing equation is 3! m̄ = h. Hence,
m̄(h)|T =Tc ' h1/3 .
This corresponds to an exponent δ = 3.

∂2f
Similarly, the singular part of the heat capacity is C ' −T ∂T 2 . For T > Tc , m = 0 and
2
f0 = g0 (T ). For T < Tc , the magnetization is non zero and f0 = g0 (T )+ 3a 2
2g (Tc −T ) +· · · .
Hence, there is no divergence of the heat capacity at the transition but a jump with an
extra heat capacity
δC ' 3a2 Tc /g.
This corresponds to an exponent α = 0. We of course recover all exponents of the Ising
mean field theory.
The key point of this section is that we don’t actually need any microscopic details
in order to calculate exponents. All we needed to know was the symmetry of the order
parameter. All we needed was a generic Taylor series expansion, and we didn’t even
need to know much about the value of any of the expansion coefficients. One could have
ferromagnets on any shaped lattice, or any sort of microscopic physics, and still the same
expansion pertains. Furthermore, we can translate this calculation to many other systems,
such as ferro-electrics (where h is replaced by the externally applied electric field), if one
can argue that the necessary symmetry holds.
70
• Landau-Ginzburg theory
The Landau-Ginzburg theory is same thing as Landau theory but taking into account
more carefully spatial the dependence of the magnetization. It generalizes the previous
construction by taking into account inhomogeneity (in the external field or in the local
magnetization) and hence gives access to the correlation length. It only describes the long
wave length of these inhomogeneities (not inhomogeneities at small distance). To lowest
order in the derivative, the free energy reads
Z h κ(T ) i
F [m; h] = dD x (∇m)2 (x) + g2 (T ) m(x)2 + g4 (T ) m(x)4 − h(x)m(x) + · · · ,
2
up to the irrelevant constant (in m) term g0 (T ). Higher order derivatives are neglected
assuming that fluctuations are small and that the long wave length modes are the only
relevant ones.
One may justify the use of this functional free energy from the heuristics of the pre-
vious section. Recall that we argued that a coarse graining approximation lead to a
representation of the partition sum as a functional integral
Z
Z ' [Dm] e−F [m;h] ,
where the free energy

R Dfunctional F [m] is assumed to be an integral over a local energy
functional F [m] = d x f [m(x)] with m(x) the local magnetization at mesoscopic scales.
The mean field approximation, which reproduces the Landau-Ginzburg theory, corre-
sponds to estimate this integral by a saddle approximation, (even though there is no small
parameter) so that
log Z = −Fmin , with Fmin = min[m(x)] F [m; h].
In this approximation, fluctuations around the mean magnetization are neglected, and the
latter is determined by a minimization problem, m(x) = Arg[min[m] F [m; h]], as usual in
thermodynamics.
• Correlation functions
To compute the correlation exponent requires to slightly generalize the construction of
the previous section by taking into account inhomogeneity (in the external field or in the
local magnetization). To lowest order in the derivative, the Landau-Ginzburg free energy
reads
Z h κ(T ) i
F [m; h] = dD x (∇m)2 (x) + g2 (T ) m(x)2 + g4 (T ) m(x)4 − h(x)m(x) + · · · ,
2
up to the irrelevant constant term g0 (T ). The minimization is now with respect to m(x).
The minimizing condition is the differential equation
−κ(T ) ∆m(x) + 2g2 (T ) m(x) + 4g4 (T ) m(x)3 = h(x).
Again we Taylor expand the coefficient around the critical temperature, so that
a g
κ(T ) = κ + · · · , g2 (T ) = (δT ) + · · · , g4 = + · · · ,
2 4!
71
with δT := T − Tc , the gap to the critical point, and a > 0. Up to a redefinition of m, we

can set κ = 1. The equation for the local magnetization then reads
g
−∆m(x) + a(δT ) m(x) + m(x)3 = h(x).
3!
We look at the response function
∂m(x)
G(x, y) := ,
∂h(y) h=0
which is linked to the correlation functions (coding for the fluctuations) through the
fluctuation-dissipation theorem. It satisfies
g
− ∆x + a(δT ) + m̄20 G(x, y) = δ(x − y),

2!
with m0 the magnetization at zero external
q field. This is the equation for the Euclidean
Green function with mass Meff = a(δT ) + 3!g m̄20 . This function decreases at large
−1
distances G(r) ' e−Meff r with a correlation length ξ = Meff . See next Chapter. Recall
that m0 = 0 for δ(T ) > 0 and 3!g m̄20 = −a(δT ) for δT < 0. Hence,
(
−1 [a(T − Tc )]−1/2 , for T > Tc ,
ξ(T ) = Meff '
[2a(Tc − T )]−1/2 , for T < Tc .
This corresponds to the exponent ν = 1/2.

At the critical point T = Tc , the mass Meff vanishes and the correlation length ξ
diverges. The Green function satisfies −∆x G(x − y) = δ(x − y), so that
G(x)|T =Tc ∝ |x|2−D .
This corresponds to the exponent η = 0.

Remark that speaking about correlation functions requires talking about the probabil-
ity measure used to define and to compute them. Here we took a short cut arguing that
we could use the fluctuation-dissipation theorem and the Landau-Ginzburg free energy to
compute the response functions. Hence we implicitly assume that this free energy can be
used as a generating functional for correlation functions (called an effective action). This
requires arguing why and when the mean free approximation is exact, or not. This will
be discussed below.
We can a posteriori justify the truncation of the free energy expansion within the mean
field approximation. The Landau-Ginzburg free energy describes the transition close the
transition, i.e. for δT = |T − Tc | Tc . Let us evaluate the typical order of magnitude of
each term. We have seen that m ∼ (δT )1/2 . By the minimization equation this means that
h ∼ (δT )3/2 (more precisely that the previous description does not hold if the magnetic
field is much bigger that this typical order of magnitude). We also evaluate the correlation
length ξ ∼ (δT )−1/2 . This gives the typical order of the space variation the magnetization
so that the spatial derivatives scale as ∇ ∼ ξ −1 ∼ (δT )1/2 . Hence, all the term is the LG
free energy scale the same way, namely
(∇m)2 ∼ (δT ) m2 ∼ m4 ∼ hm ∼ ξ −4 ∼ (δT )2 .
72
All the higher order terms, compatible with the Z2 -symmetry, such as (∇m)4 , (∇2 m)2 or
m2 (∇m)2 , m6 etc. scale with higher power of (δT ). They can be neglected close to the
transition point. This scaling argument will be made more precise below and within the
renormalization group formulation.
4.5 Upper and lower critical dimensions

This section is concerned with the domain of validity of the mean field Landau-Ginzburg
description. The Landau model, as any mean field model, assumes that the fluctuations are
not too important and do no play a relevant role. The upper and lower critical dimensions
are the dimensional thresholds at which those fluctuations become relevant.
The fluctuations typically increase as the dimension is lowered. There are two notions,
the upper and the lower critical dimension.
— If the dimension is small enough, i.e. if D < Di with Di the lower critical dimension,
the fluctuations are strong enough to destroyed any order at any temperature. That is:
for D < Di , there could not be any ordered phase: the fluctuations prevent them to exist.
— If the dimension is big enough, i.e. D > Ds , the fluctuations are small enough so
that the mean field approximation is correct. That is, for D > Ds , the fluctuations are
irrelevant and the Landau description is exact.
In the intermediate regime,
Di < D < D s ,
the fluctuations are relevant They do not destroy the possible existence of a phase tran-
sition but modify drastically its characteritics. They have to be taken into account in
the description of this phase transition (which is then quite non-trivial, see the following
Chapters).
• The upper critical dimension
Let us start with the estimation of the upper critical dimension Ds . It depends on the
universality class, so we shall do it for the Ising universality class (the φ4 Landau theory).
For the Ising class Ds = 4. We consider the Ising model at zero external field h = 0.
Recall the expression for the free energy
Z 1 a g
F [m] = dD x (∇m)2 (x) + (δT ) m(x)2 + m(x)4 ,
2 2 4!
with a > 0. The Landau theory will be correct if the typical fluctuations of the magneti-
zation ∆m(x) are small compared to the typically magnetization m0 , i.e.
(∆m)2 m20 .
This is called the Ginzburg criteria.

Since spins are uncorrelated at distances bigger than the correlation length, the natural
scale/size ` of such fluctuations is the correlation length: ` ∼ ξ. In the Landau mean
field theory ξ 2 ∼ (δT )−1 and m20 ∼ (δT ). Let us estimate the free energy variation,
∆F = F [m0 + ∆m] − F [m0 ], associated to fluctuations of typical amplitude ∆m. The first
73
two terms in F scale the same ways (using that `2 ∼ ξ 2 ∼ (δT )−1 ) but the second scale
differently with the size ` ∼ ξ:
Z
1
dD x (∇m)2 (x) ∼ ξ D−2 (∆m)2
2
Z
dD x a(δT ) m2 (x) ∼ ξ D (δT )(∆m)2 ∼ ξ D−2 (∆m)2 ,
Z
g
dD x m4 (x) ∼ ξ D g(∆m)4 .
4!
According to the Boltzmann rules, the relative statistical weight of such fluctuations is of
order e−∆F . Thus the typical sizes of the maximum fluctuations are so such that ∆F is
of order O(1) —as, if it is much bigger it occurs with a much smaller probability.
For D > 4, the typical size is determined by the first terms, so that (∆m)2 ∼ ξ 2−D .
Indeed, then ξ D (∆m)4 ∼ ξ 4−D → 0 for large ξ. Thus for D > 4, we have
(∆m)2 ∼ ξ 2−D m20 ∼ ξ −2 .
Hence, the Landau mean field approximation is justified for D > 4.

For D < 4, the typical size is determined by the second term, so that (∆m)2 ∼ ξ −D/2 .
Indeed, then ξ D−2 (∆m)2 ∼ ξ (D−4)/2 → 0. Thus for D < 4, we have
(∆m)2 ∼ ξ −D/2 m20 ∼ ξ −2 .
Hence, the Landau mean field approximation breaks down for D < 4. And the upper
critical dimension is Ds = 4.
We can refine this argument a little. Consider still D < 4. The parameters of the
model define a length scale `c by dimensional analysis via
`D−4
c ' g.
Let us repeat the previous analysis but putting back the dependence on g via `c . This yields
(D−4)/2
that the typical fluctuations are of size (∆m)2 ∼ ξ −D/2 `c for D < 4. Equivalently,
(∆m)2 ξ (4−D)/2
2 ∼ .
m0 `c
Hence, (∆m)2 m20 iff ξ `c . Since the correlation length has of course also to be much
bigger than the microscopic length, we thus learn that for D < 4 the Landau theory is
correct is
ξmicro ξ(T ) `c .
Since the correlation length increases close to the critical point, ξ(T ) ∼ (δT )−1 , this
is equivalent to saying that the Landau theory is correct if the distance to the critical
temperature is big enough. In other words, the breakdown of the mean field approximation
occurs very close the critical point (it only occurs as one approches the critical point).
Before closing this discussion recall that the upper critical dimension depends on the
universality class, i.e. it depends on the symmetry, on the Landau expansion, etc.
74
• The lower critical dimension

Let us now discuss the lower critical dimension. A standard argument by Peierls tells
that there is no ordered phase at positive temperature T > 0 in dimension 1. Hence the
lower critical dimension is at minimum bigger than one: Di > 1.
This is the lower critical dimension for universality class with discrete symmetries.
For continuous symmetry, ordered phase associated to symmetry breaking comes with
massless Goldstone modes (see below). A theorem from Mermin and Wagner tells that
this cannot occurs in dimension less than 2 (at least for a compact symmetry group). The
argument is essentially based the IR behavior of massless Green function,
dD k eikx
Z
GM =0 (x) = .
(2π)D k 2
It scales like ∼ k D−2 for k small, and hence it is infrared divergent in D < 2. A way to
regularize in D = 2 is to put the system in a finite volume of typical size L (see the more
in-depth discussion in the chapter on free conformal field theory below), then at distances
large enough
G(x) ' log(|x|/L), if D = 2,
The fluctuations are therefore large at large distance and the spins cannot be ordered
(because if they were their two point function G(x) = hS0 Sx i will asymptotically factor-
izes/clusters as hSi2 ). Thus the lower critical dimension is Di = 2.
4.6 Deviation from mean field theory

Here we describe how going away from the mean field approximation leads to deal with
large fluctuations encoded in statistical field theory.
• Naive scaling dimensions
Let us derive again the previous results using scaling arguments –or let us make contact
between this result and scaling arguments. This will be a tiny step towards reasonings
used when formulating the renormalization group. We aim at understanding how the
theory changes when changing the size of the fluctuations. Let us write again the free
energy action as
Z 1 τ g
F [m] = dD x (∇m)2 (x) + m(x)2 + m(x)4 .
2 2 4!
Here, we set τ :' (δT ) that we view as an external parameter similar to the coupling
constant g. We want to compare this action for a profile m(x) and a scaled profile mλ (x) in
which we have scaled (scaled up if λ 1) the distance (hence the size of the magnetization
profile). Thus
m(x) → mλ (x) = λ−∆ m(x/λ),
where we used the freedom to scale the height of the profile. ∆ is free parameter called
the (bare) scaling dimension. We fix it by demanding that the first term in the free energy
is preserved, hence
2∆ = D − 2.
75
We now look how to transform the parameter g and τ as g → gλ and τ → τλ in such way
to preserve the free energy, i.e. such that
F [mλ ; g, τ ] = F [m; gλ , τλ ].
This is may alternatively be interpreted as follows: the free energy (or the action) for
configurations of size λ` with parameter τ and g is identical as those of size ` but with
parameter gλ and τλ . This is a key equation, because it will generalize within the renor-
malisation group in the following Chapter.
It is then a simple scaling exercise to check that
τλ = λ2 τ, gλ = λ4−D g.
Hence, the non-linear effects become more and more important at large scale (i.e. for large
scale configuration) if D < 4 = Ds . That is: the non-linear quartic terms are relevant at
large distances, at large scales, if D < 4.
Let us then determine the scaling dimensions of other operators, 2k
2p 2k
R D i.e. 2km (x) or
∇ m (x). They for instance correspond to terms of the form gk d x m (x) in the
action. Under the scaling transformation m → mλ , these couplings change (in order to
preserve the key relation F [mλ ; gk ] = F [m; gk (λ)] as
gk (λ) = λD−k(D−2) gk .
All these couplings decrease as λ → ∞, i.e. at large distances, for D > 3 (the coupling to
m6 is marginal in D = 3). Adding derivatives as in ∇2p m2k renders this decrease faster.
Thus, in D > 3 only the coupling to m2 and m4 are relevant. We can thus discard all
other operators as long as we are only looking at the large distance physics.
Remark that this analysis shows that the dimension D = 2 is peculiar, in the sense
that m has zero scaling dimension in D = 2. All polynomial terms scale the same way in
D = 2 and this allow for a wide landscape of multi-critical points.
• Renormalization group
The above analysis relies on estimating how the fields
R and the coupling constants scale
under dilation using the first quadratic term, namely 12 dD x(∇m)2 (x), in the free energy.
It does not take into account neither the other terms in the free energy nor the effect of the
fluctuations (which may arise from the fact that the partition function involves summing
over all magnetisation configurations). The renormalization group that we shall discuss
in the following Chapter gives a general framework to go beyond this naive (mean field)
scaling analysis.
4.7 Symmetry breaking and Goldstone modes

Here we discuss why and how continuous symmetry breaking is associated with the emer-
gence of massless modes, called Goldstone bosons.
• U (1) symmetry breaking
Do first the standard example of U (1) broken symmetry with Mexican hat potential....
76
[....To be completed...]
• Goldstone modes and coset spaces.
We present here the semi-classical (alias mean field like) argument but for a general
symmetry group G broken to H. We consider situations in which the order parameter is
a multi-component field that we shall denote Φ again - we denote Φa the field component
- with action Z h1 i
S[Φ] = dD x |∇Φ|2 + V (Φ) .
2
The potential V depends on all components of the field (so it is a multi-variable function).
We assume that the field Φ takes values in a representation of the group G so that the
group acts linearly on it: Φ → g · Φ for g ∈ G. The action, both the kinetic and the
potential terms, are supposed to be invariant under this action. In particular
V (g · Φ) = V (Φ).
In absence of external potential, mean field solutions are minima of the potential V .
Let us pick one of this minima that we shall denote ϕ0 : ∇V (ϕ0 ) = 0 or in components
∂a V (ϕ0 ) = 0. Let us then suppose that this so-called vacuum solution is non trivial ϕ0 6= 0
so that the symmetry G is broken, g · ϕ0 6= 0 for at least some g ∈ G, and let H ⊂ G be
the subgroup that preserves ϕ0 :
h · ϕ0 = ϕ0 , h ∈ H ⊂ G.
H is supposed to be the maximal subgroup preserving ϕ0 . We aim at looking at the

fluctuations around this solution. We set Φ(x) = ϕ0 + φ(x) and expand the action (to
quadratic order):
Z h1 X 1X a i
S[ϕ0 + φ] = dD x |∇φa |2 + φ Hab φb + · · · , Hab := ∂a ∂b V (ϕ0 ).
2 a 2
a,b
The Hessian matrix is thus the mass matrix of the fluctuating modes φ. The Goldstone
theorem asserts that these fluctuating modes contain massless modes, that is: the mass
matrix Hab possesses a number of zero eigen-values (and the eigen-vectors are the massless
modes).
This properties comes from the G-invariance of the potential (with G a continuous
group). This invariance reads V (g · Φ) = V (Φ). Let us apply this relation to the vacuum
solution, hence V (g · ϕ0 ) = V (ϕ0 ) for g ∈ G. Take g infinitesimal g = 1 + εX + · · · , with
X ∈ Lie(G) and ε 1. Expanding the relation to second order in ε, using the fact that
ϕ0 is minimum, yields
X
Hab (Xϕ0 )a (Xϕ0 )b = 0, ∀X ∈ Lie(G).
a,b
However, Xϕ0 = 0 for X ∈ Lie(H) ⊂ Lie(G), because ϕ0 is preserved by the subgroup H.

Let us decompose Lie(G), as a vector space as,
Lie(G) = Lie(H) ⊕ TG/H ,
77
where TG/H is the complement vector space of Lie(H) in Lie(G). (It is isomorphic to
the tangent space to the quotient G/H at the identity point and form a representation of
H). For any X ∈ TG/H , Xϕ0 6= 0 because H is supposed to be the maximal subgroup
preserving ϕ0 .
Hence, any element of TG/H is associated to a massless mode. In other words, the
massless zero modes of the G-invariant action broken to H ⊂ G are in correspondance
with the quotient G/H. If no other relevant physics play a role at intermediate scale, this
implies that the large distance phenomena are governed by this G/H-zero modes.
4.8 Exercises
• Exercise 4.1: Mean field from a variational ansatz
The aim of this exercise is to derive the Ising mean field approximation form a vari-
ational ansatz. We consider the P Ising in homogeneous
P external field hi so that the con-
figuration energy is E[s] = − i,j Jij si sj − i hi si , with Jij proportional to the lattice
adjacency matrix. The Ising spins take values si = ±. Let Z[h] be its partition function.
(Note that we introduce the external magnetic field with a minus sign).
As an ansatz we consider the model of independent
P ospins in an effective inhomogeneous
external field hoi with ansatz energy E o [s] = −
i hi si , so that the ansatz Boltzmann
−1 β i hoi si
P
weights are Z0 e Q with Z0 the ansatz partition function.
(i) Show that Z0 = i [2 cosh(βhoi )].
(ii) Using a convexity argument, show that E0 [e−X ] ≥ e−E0 [X] for any probability measure
E0 and measurable variable X.
(iii) Choose to be E0 the ansatz measure and X = β(E[s] − E o [s]) to prove that
o [s]]
Z[h] ≥ Z0 e−βE0 [E[s]−E ,
or equivalently, F [h] ≤ F0 − E0 [E o [s] − E[s]], with F [h] and F0 the Ising and ansatz free
energy respectively.
The best variational ansatz is that which minimizes F0 − E0 [E o [s] − E[s]].
(iv) Compute F0 , E0 [E o [s]] and E0 [E[s]] and show that the quantity to minimize is
X X X
F0 [ho ] + hoi m̄i − Jij m̄i m̄j − hi m̄i ,
i ij i
o
where m̄i = − ∂F∂h
0 [h ]
o = tanh(βhoi ) is the local mean magnetization evaluated with the
i
ansatz measure. Show that this minimization problem reduces to the Ising mean field
equations.
• Exercise 4.2: Thermodynamic functions and thermodynamic potentials
The aim of this exercise is to recall a few basic fact about generating functions, ther-
modynamic functions and their Legendre transforms.
Let us consider a (generic) spin model and let E[{s}] be the energy of a spin configu-
ration {s} with local spin si . We measure the energy
P in unit of the temperature so that
the Boltzmann weights are e−β E[{s}] . Let Z[0] = {s} e−β E[{s}] be the partition function.
78
(i) Explain why the partition function Z[h] in presence an external inhomogeneous external
field h is the generating function for spin correlations.
What is the expression for Z[h]?
Show that P the generating function for this spin correlation functions can written as (with
(s, h) = i si hi )
Z[h]
E[e(s,h) ] = ,
Z[0]

(ii) Let F [h] be the free energy and let W [h] = − F [h] − F [0] . Verify that
log E[e(s,h) ] = W [h].
(iii) Let Γ(m) be the thermodynamic potential defined as the Legendre transform of W [h].
Recall that
∂W
Γ(m) = (m, h∗ ) − W [h∗ ], with [h∗ ] = m.
∂h
Verify that this transformation is inverted by writing
∂Γ
W [h] = (m∗ , h) − Γ[m∗ ], with [m∗ ] = h.
∂m
• Exercise 4.3: An alternative representation of the Ising partition function.

The aim of this exercise is to explicitly do the computation leading to the representation
of the Ising partition function in terms of a bosonic field. It uses a trick —representing
the interaction terms via a Gaussian integral over auxiliary variables— which find echoes
in many other problems.
(i) Prove the following representation of the Ising partition function given in the text
(without looking at its derivation given there):
Z Y
Z = [ dφk ] e−S[φ;h] ,
k
with the action

1X X X
S[φ; h] = − φi Jij φj + log[cosh(hi + Jij φj )].
4
ij i j
(ii) Deduce what is the representation of the Ising spin variables si in terms of the bosonic
variables φi .
• Exercise 4.4: Mean field vector models
See the exercise booklet....
79
5 The renormalisation group and universality
The aim of this important Chapter is to present the basics ideas of the renormaliza-
tion group (RG), the techniques associated to its implementation, and the output of its
analysis. Following K. Wilson, RG transformations will be introduced by doing real space
renormalization via spin blockings. This amounts to find iteratively the effective hamil-
tonians coding for the interactions of coarse grained blocks of spins of increasing sizes.
It may alternatively be viewed as arising from a partitioning of the configuration space
though a conditioning procedure which amounts to fix the effective spins labelling the
block spin configurations. The way the effective hamiltonians evolve under RG transfor-
mations is encoded into so-called beta functions which generate the RG flows on the space
of coupling constants. Critical theories are in correspondence with fixed points of the
renormalization group. The renormalisation group yields an understanding of the origin
of scaling exponents and scaling functions and the universality of critical phenomena.
5.1 Block spins and RG transformations

• Block-spin transformations in the 1D Ising model.
P
We start with 1D Ising hamiltonian H[s] = −J i si si+1 with si = ± (we absorb
the inverse temperature β in the energy scale J so that J ∝ 1/T ). The partition sum is
Z = [s] e−H[s] . There are many ways to solve this (trivial) problem, say by using the
P
transfer matrix formalism. But here we will do it using a complicated method to illustrate
the idea of the RG.
Imagine grouping the spins by blocks of size 3, i.e. (· · · ][s1 s2 s3 ][s4 s5 s6 ][· · · ). Each
blocks may be in 23 = 8 configurations. We can group these eight configurations in two
disjoint sets to which we assign an effective spin s0 . We can for instance choose the majority
rule so that s0 = + if the three internal spins of the block are [+ + −] or a permutation
thereof, and s0 = − if the internal spins are [− − +] up to permutation. It will actually be
simpler to choose to assign to each block the spins of the middle site, so that the effective
spin for the block [s1 s2 s3 ] is s0 = s2 , or alternatively s0 = ± for the configurations [s1 ± s3 ].
We then imagine computing the partition function in two steps: first summing over the
internal spins of each blocks conditioned on their effective spins and second on the block
effective spins.
Consider two adjacent blocks, say (· · · ][s1 s2 s3 ][s4 s5 s6 ][· · · ), and denote by s01 := s2 and
0
s2 := s5 the two block spins. Doing the partial sum induces effective interaction between
the block spins. The Boltzmann weights are of the form:
0 0 0 0
· · · eJs1 s1 × eJs1 s3 eJs3 s4 eJs4 s2 × eJs2 s6 · · · .
We sum over s3 and s4 at s01 and s02 fixed (the other spins s1 , s6 , · · · do not play a role).
0
Using eJss = cosh J(1 + xss0 ) with x = tanh J, we may write this as the product of three
80
terms (cosh J)3 (1 + xs01 s3 )(1 + xs3 s4 )(1 + xs4 s02 ). The sum is done by expanding this
product. It yields
22 (cosh J)3 (1 + x3 s01 s02 ).
Up to a multiplicative constant (independent of the spins) this expression, coding for the
interaction between the effective blocks spins, is of the same form as that for the original
spins but with a new interaction constant J 0
x0 = x3 , i.e. tanh J 0 = (tanh J)3 .
The (new) hamiltonian for the block spin is thus identical to the original 1D Ising hamil-
tonian up to an irrelevant constant,
X
H 0 ([s0 ]) = N e(J) − J 0 s0i s0i+1 ,
i
with log e(J) = 22/3 (cosh J)/(cosh J 0 )1/3 . The original partition function can thus be
rewritten as
0 0
X X
Z(J) = e−H[s] = e−H [s ] = e−N e(J) Z(J 0 ).
[s] [s0 ]
Via this blocking procedure, we have effectively reduced the number of degrees of freedom
from 2N to 2N/3 .
By iteration the effective coupling transforms as xn → xn+1 = x3n at each step. There
is only two fixed points: x = 1 which corresponds to zero temperature and x = 0 which
corresponds to infinite temperature. Since x < 1, unless T = 0, the effective couplings
xn converge to zero, and the effective temperature increases towards infinite temperature.
Hence, the long distance degrees of freedom are effectively described by an infinite tem-
perature: they are in the disordered paramagnetic phase (a statement that we already
knew: no phase transition in 1D).
Since the system is in the disordered phase, its correlation length is finite. The ‘phys-
ical’ correlation length has of course the dimension of a length, but we can measure it
in units of the lattice spacing a. This dimensionless correlation length only depends on
J, or equivalently on x. Since the block spin transformations preserve the long distance
physics, the dimension-full correlation length is preserved by these transformations. Since
the lattice size has been dilated by a factor 3, i.e. a → λa (λ = 3), the dimensionless
correlation length transforms as
1
ξ(x0 ) = ξ(x),
3
with x0 = x3 . This implies that
const. const.
ξ(x) = = .
log x log(tanh J)
Of course the correlation length is always finite, for all value of the coupling constant x,
(because the system is in the disordered phase) but it diverges exponentially close to zero
temperature: ξ ' econst./T near T → 0.
81
• Framework for a general theory: RG transformations.

Consider a lattice model with spin variables si on each lattice site (of mesh size a)
with hamiltonian H (again the inverse temperature is included in H). As we will soon
see, we need to think about the hamiltonian as having all possible interactions included
(compatible with the symmetry). That is: we have to think as H being the “more general
hamiltonian” H([s]|{g}) with coupling constants {g} (a possibly infinite number) of the
form (if it preserves the symmetry)
X (2) X (4)
H([s]|{g}) = gij si sj + gijkl si sj sk sl + · · · .
ij ijkl
Let us do a block spin transformation. Each block is supposed to be of dimensionless

size λ (this is its size measured in lattice units), i.e. its ‘physical’ size is λa. At each
block we affect an effective spin sλ , say by the majority rule or via the middle spin. Each
effective spin corresponds to a set of configuration of the original spins (in each block).
That is, each block spin configuration [sλ ] indexes a partition of the set of original spin
configurations. So we can decompose the sum of the original spin configurations as
X X X
(· · · ) = (· · · ), (34)
[s] [sλ ] [s]↓[sλ ]
where [s] ↓ [sλ ] means that the configuration [s] belongs to the set of configurations indexed
by [sλ ], i.e. the configuration [s] yields the block spin configuration [sλ ]. What we have
done here is simply to partition (or to condition) the configuration space, the partition
being indexed by the [sλ ].
In particular for the partition function we have
X X X X
Z[{g}] = e−H([s]|{g}) = e−H([s]|{g}) = Z 0 ([sλ ]|{g}),
[s] [sλ ] [s]↓[sλ ] [sλ ]
where Z 0 ([sλ ]|{g}), the conditioned partition function, is obtained by conditioning

X
Z 0 ([sλ ]|{g}) := e−H([s]|{g}) .
[s]↓[sλ ]
We may define an effective hamiltonian H 0 for the effective block spins via H 0 ([sλ ]|{g}) :=
− log Z 0 ([sλ ]|{g}).
The RG hypothesis (which is here a bit tautological as we consider the most general
hamiltonians) is that this hamiltonian H 0 is of the same nature as the orginal one (up to
an additive constant) but with new coupling constants {g λ }:
H 0 ([sλ ]|{g}) = N eλ ({g}) + H([sλ ]|{g λ }), (35)
or equivalently
λ ({g}) λ ]|{g λ })
Z 0 ([sλ ]|{g}) = e−N e e−H([s ,
82
In other words, the conditioned partition function Z 0 is the new Boltzmann weight for the
effective block spins (Boltzmann weights and partition functions are always defined up to
a multiplicative factor, so the function e does not matter much).
The RG transformation is the map from {g} to {g λ }:
{g} → {g λ } = Rλ ({g}). (36)
Of course, we can iterate these transformations. We can first group spins into blocks of
λ and then group the new effective spins into blocks of size λ0 . This will of course be
equivalent to group the spins in blocks of size λ λ0 (dilatation factors multiply). Hence
0 0
Rλ ·λ = Rλ · Rλ ,
and the RG transformations form a group (a pseudo-group has they cannot be inverted:
λ > 1 always).
Since they form a (pseudo)-group, the RG transformations are generated by the in-
finitesimal transformations. Here, we are implicitly making the assumption that we can
view the scaling factor λ as a continuous variable (initially, on the lattice this scale factor
was discrete but may assume that when describing the long distance physics we may take
it to be continuous). So let gα be coordinates for the set of coupling constants and define
the vector fields
βα ({g}) := λ∂λ gαλ λ=1 .

(37)
These are called “beta-functions” (they are actually vector fields). They generate the RG
map {g} → {g λ } in the sense that {g λ } is solution of the flow equation
λ∂λ gαλ = βα ({g λ }), (38)
with initial condition {g λ=1 } = {g}. The solutions g λ are called the “running coupling
constants”.
As in the case of the 1D Ising model, the physical correlation length remains unchanged
under RG transformations (because this is just a reorganization of the statistical sum).
However the lattice mesh is rescaled from a to λa under the transformation. Hence, the
dimensionless correlation length satisfies
ξ({g}) = λ ξ({g λ }).
In practice we can never exactly compute Rλ as a map acting on an infinite set of

variables. So we will have to do approximations (constructive field theory and/or exact
renormalization group aims at controlling exactly these RG transformations) by truncating
the set of coupling constants (keeping only the relevant ones, or those which are expected
to be the most relevant ones).
Of course we cannot iterate the RG group ad-finitum, we have to stop once the size
of the block is comparable to the correlation length, that is a λ a aξ. This is self
consistent because spins inside a block of size much smaller than aξ are correlated (they are
a distance much smaller that the correlation length) and hence behave almost collectively.
This is the physical rational behind the renormalization group idea.
83
The understanding of many physical phenomena, ranging from fundamental interac-

tions to classical or quantum extended systems, requires extracting the relevant large-scale
degrees of freedom, which can manifest themselves in different guises, say collectives modes,
shapes, structures or variables. Extracting these relevant variables is of course one of the
main aim of the renormalization group, but making sure that the RG strategy works often
requires having recognized the appropriate setup (in the opposite case, one cannot make
the RG transformations contracting towards a relevant effective model, and as an illustra-
tion, the absence of such appropriate setup is for instance why the RG has not been yet
successfully applied to turbulence).
5.2 RG fixed points and universality

We now look at what are the consequences of the existence of fixed points of the RG
transformations. To simplify matter we assume that the RG flow is defined on a set of N
coupling constants (”rigorously” this number is, or could be, infinite). We also assume that
the RG transformations have enough analytical properties (smoothness, regularities,...) to
justify the formal developments we are going to make.
• RG fixed points.
A fixed point (in the coupling constant manifold) is a point {g∗ } such that
Rλ ({g∗ }) = {g∗ }.
Fixed points are zero of the beta-functions (because the beta-functions are the vector fields
generating the RG transformations):
βα ({g∗ }) = 0. (39)
At a fixed point the correlation length is either infinite or zero (because it satisfies
ξ({g∗ }) = λξ({g∗ }) and by iteration ξ({g∗ }) = λn ξ({g∗ }) for any λ > 1). Fixed points
with zero correlation length are called trivial fixed point (but, non-trivial model with zero
correlation length but topological property, have recently been defined). Critical fixed
points are those with infinite correlation length.
• RG eigenvalues at a fixed point.
We aim at describing the behaviour of the RG flow closed to a fixed point (this is
actually a standard exercice in dynamical systems). Let us linearize the flow in the vicinity
of the point. Let {g α } be some local coordinates in the space of all coupling constants.
We can then linearize the RG flow (to first order) and write near a fixed point (with
coordinates {g∗α })
X
Rλ ({g∗α + ε δg α + · · · }) = {ε λ
Bασ δg σ + · · · },
σ
with g α = g∗α + ε δg α + · · · . The matrix Bασ

λ is simply the matrix of derivatives of Rλ at
the fixed point. The nature of the flow locally around the fixed point is determined by
the eigenvalues of Bασ λ , because we can then change variables locally to diagonalize the
84
flow. Indeed, let ψ i be the left eigenvectors, with component ψαi . These eigenvalues are
necessarily of the form λyi for some yi because of the group law of the RGP
transformation:
i λ y i i i i α
P
α ψα Bασ = λ ψσ . Let us define locally new coordinates u by u = α ψα δg . The
i
new coordinates transform diagonally (at least in a small neighbourhood of the fixed point)
Rλ ({ui }) = {λyi ui }.
The number yi are called the RG eigenvalues and the ui scaling variables (or scaling fields).
There are three cases:
— yi > 0: the ui are said to be “relevant”. These directions are pushed away from the
fixed point from by the RG flow;
— yi = 0: the ui are said to be “marginal”. The nature of the flows depends on the higher
order terms: it can be marginally relevant or marginally irrelevant.
— yi < 0: the ui are said to be “irrelevant”. These directions are attracted to the fixed
point by the RG flow;
Suppose that they are n relevant directions (n finite !). These n directions flow away
from the fixed points, the N − n remaining ones are attracted by the fixed point. Suppose
that n is relatively small (or at least finite!... and this is always the case we shall consider
—it corresponds to renormalizable field theories). Then, there is, locally around the fixed
point, a hypersurface of (high) dimension N − n whose points flow towards the fixed
point under RG transformations. This hypersurface is called the “critical surface”. For
all points on this surface the long distance physics is critical and described by the fixed
point because all those points flow towards the fixed point. In the remaining directions
the flow moves the point away from the fixed points. Thus n (relevant) parameters have
to be adjusted for the system to be critical (to be on the critical surface). For instance, for
the Ising model, two parameters have to be adjusted, the temperature T and the magnetic
field h, so there are two relevant directions whose coordinates can be named uT and uh .
The very large dimensionality of the critical hypersurface explains (or expresses) the
universality of the critical phenomena: all points on this hypersurface possess the same
critical behavior, i.e. all hamiltonians on the hypersurface, irrespective of the interactions
they encode, have the same critical behavior. See Figure 1.
5.3 Scaling functions and critical exponents

We now explained why RG transformations and the existence of fixed points imply that
the existence of scaling functions and critical exponents. These exponents are going to be
related to the RG eigenvalues as we are going to explain.
• Universal scaling functions.
Let us go back to the basic definition of the RG transformations coding the way the
partition function decomposes. Let f ({g}) be the free energy per site, so that the partition
function for N site is Z[{g}] = e−N f ({g}) . We now prove that RG transformations imply
that
f ({g}) = eλ ({g}) + λ−D f ({g λ }),
85
Figure 1: Schematic representation of coupling constant RG flow.
with eλ ({g}) a function, which is expected to be smooth, representing a local contribution

to the free energy. The factor λ−D comes from the fact that a RG transformation by a
scale λ reduces the number of sites from N to N λ−D .
Recall that X
Z[{g}] = Z 0 ([sλ ]|{g}),
[sλ ]
λ ({g}) λ ]|{g λ })
with Z 0 ([sλ ]|{g}) = e−N e e−H([s , so that
λ λ λ
X
Z[{g}] = e−N e ({g}) e−H([s ]|{g }) .
[sλ ]
Doing the sum of the block spin [sλ ] but remembering that the number of blocks is N λ−D
proves the result.
The function eλ ({g}) is expected to be non singular because it comes from summing
over the spins inside the blocks (so it comes from finite sum of Boltzmann weights and
hence cannot developed singularities). As a consequence the singular part in the free
energy transforms homogeneously under RG transformations:
fsing ({g}) = λ−D fsing ({g λ }).
After a few iterations of the RG transformations, we may expect to be close to a fixed

point (assuming the RG transformations have good contracting properties, at least close
to a fixed point). There we can use the scaling variables. The relation then becomes
fsing ({ui }) = λ−D fsing ({λyi ui }). (40)
Similarly for the (dimensionless) correlation length
ξ({ui }) = λ ξ({λyi ui }). (41)
When iterating the RG transformations (λ increases) only the relevant and marginal
variables with positive eigen-values remain (See the Section about corrections to scaling
to look for the effect of the irrelevant variables).
86
There are many ways to express these above relations. The simplest case is when there
is only one relevant scaling variable. Let us still denote it as uT . The singular part of
the free energy satisfies the scaling relation (40) which in present case reads fsing (uT ) =
λ−D fsing (λyT uT ). Hence, we have
fsing (uT ) = fˆ± |uT |D/yT ,
with fˆ± a constant which however may depend on the sign of uT . Similarly, the correlation
length ξ is a scaling function of variable uT :
ξ(uT ) ∝ |uT |−1/yT .
This is the only scale of the problem (if we are dealing with an infinite volume system)
and all physical quantities can be expressed in terms this scale. That is: we can measure
everything (say the singular part of the free energy) in terms of the correlation length
instead of the scaling variable uT .
Let us now consider cases with more relevant variables. Recall that the set of {ui }
refers to a collection of scaling variables uT , uh , etc., associated to external parameters,
the temperature, the magnetic field, etc.... Let us again single out one of the variable, say
(0)
uT . Then let us pick λ = |uT |−1/yT so that λyT uT = ±1 (or any other reference point uT
still in the perturbative domain), and we get
uj uj
ξ(uT , {uj }j6=T ) = |uT |−1/yT ξ(±1, { }j6 = T ) =: |uT | −1/yT ˆ
ξ± ({ }j6=T ),
|uT |yj /yt |uT |yj /yT
and
uj D/yT ˆ uj
fsing (uT , {uj }j6=T ) = |uT |D/yT fsing (±1, { }j6 = T ) =: |uT | f± ({ }j6=T ).
|uT |yj /yT |uT |yj /yT
Recall that we only kept the relevant variables.
In cases with only two relevant scaling variables, say uT and uh as in the Ising model,
we have
uh
ξ(uT , uh ) = |uT |−1/yT ξˆ± ( ),
|uT |yh /yT
and
uh
fsing (uT , uh ) = |uT |D/yT fˆ± ( ).
|uT |yh /yT
The functions ξˆ± , fˆ± are called “universal scaling functions”.
• Critical exponents.
The existence of critical exponents follows from these relations.
— At zero magnetic field uh = 0 and the correlation length scales like ξ ' |uT |−1/yT , so
that
ν = 1/yT .
— The specific heat at zero magnetic field is ∂ 2 f /∂u2T |uh =0 ' |uT |D/yT −2 , so that
α = 2 − D/yT .
87
— The spontaneous magnetization is ∂f /∂uh |uh =0 ' |uT |(D−yh )/yT so that
β = (D − yh )/yT .
— The magnetic susceptibility is ∂ 2 f /∂u2h |uh =0 ' |uT |(D−2yh )/yT so that
γ = (2yh − D)/yT .
— The magnetization at critical temperature is ∂f /∂uh |uT =0 . To compute requires know-

ing the behaviour of fˆ± at infinity. Demanding that fsing (uT , uh ) has finite limit as
uT → 0 requires that fˆ± (x) ' xD/yh (so that the various powers of |uT | cancels). Hence
D/y −1
∂f /∂uh |uT =0 ' uh h , so that
δ = yh /(D − yh ).
Since the renormalization eigenvalues yi are given by the dimensions of the operators at
the fixed point (see below), all critical exponents are determined from these dimensions.
5.4 Corrections to scaling and finite size effects

Corrections to scaling comes from taking irrelevant variables into account. Finite size
effects are treated by looking at the RG transformation fixing the size of the system L
fixed. Let us concentrate on the finite size effects.
Suppose that we consider a critical system but in a finite size box of linear system
L, say a square box L × L × L. The free energy is going to depend on the coupling
constants {g} and on the size L. But, since L is dimension-full, the dependence of the
free energy on L comes via the ratio L/a with a the lattice microscopic mesh size – or
alternatively the free energy depends on the coupling constants {g} and on the number of
lattice sites N = (L/a)D . Let us now imagine implementing a RG transformation –say a
block spin transformation– scaling a to λa at L fixed. As explained above, this reduces the
number of lattice sites –and hence of degrees of freedom– from N to λ−D N , or equivalently
L/a → λ−1 L/a. Hence, under a RG transformation, the singular part of the free energy
transforms as
fsing ({g}, L/a) = λ−D fsing ({g λ }, λ−1 L/a),
with g λ the RG dressed coupling constants. Here we are implicitly making the hypothe-
sis that the coupling constants transform in the same way in infinite and finite volume.
Similarly, the correlation length transform as
ξ({g}, L/a) = λ ξ({g λ }, λ−1 L/a).
As in infinite volume, we may expect that we approach a fixed point after a few iterations
of the RG transformations. There we can use the scaling variables and write the above
relation as
fsing ({ui }, L/a) = λ−D fsing ({λyi ui }, λ−1 L/a),

ξ({ui }, L/a) = λ ξ({λyi ui }, λ−1 L/a).
88
As in infinite volume, the scaling form can found by singling out on variable, say uT , and
choosing the RG scale λ = |uT |−1/yT .
Let us for instance assume that there is only one relevant scaling variable, again say
uT . Then the scaling relation for the correlation length, in finite volume, reads
ξ(uT , L/a) = |uT |−1/yT ξˆ± (|uT |1/yT L/a) = |uT |−1/yT ξˆ± (L/ξ∞ ),
where we introduce the (dimension-full) correlation length in infinite volume ξ∞ = a |uT |−1/yT .
In other words, the correction to infinite volume limit is obtained by comparing the only
two relevant macroscopic length: the size of the system and the system correlation length.
Similarly, the behaviour the finite size correction to the critical behaviour of the specific
heat can be obtained form the scaling relation fsing (uT , L/a) = λ−D fsing (λyT uT , λ−1 L/a),
in case there is only one relevant variable. For the specific heat χ ∼ ∂ 2 f /∂ 2 uT this gives
χ ∼ |uT |−α ϕ± (L/ξ∞ ) = |uT |−α ϕ± (|uT |ν L/a),
with again ξ∞ = a |uT |−ν the infinite volume correlation length (with ν = 1/yT the
correlation length exponent). The infinite volume is recovered in the limit L → ∞ so that
ϕ(∞) is supposed to be finite. At finite volume, there is no sharp phase transition (because
there is then a finite number of degrees of freedom and the partition is a polynomial in
the Boltzmann weights), and the heat capacity is non singular so that ϕ(x) ∼ xαyT for
x → 0. As a consequence, the heat capacity is not singular in finite volume, but it becomes
a rounded fonction of uT with a maximum value scaling as Lα/ν and a width scaling as
L−1/ν .
5.5 Exercises
• Exercise 5.1: Real-space renormalisation: Ising on the triangular lattice.
[... Cf. the exercise booklet...]
• Exercise 5.2: Correction to scaling.
The aim of this exercise is to understand how the irrelevant variables induce sub-
leading corrections to scaling behaviours. To simplify matter, let us suppose that the
critical system possesses only one relevant scaling variable, say uT with RG eigen-value
yT > 0, and one irrelevant variable, say uirr with RG eigen-value yirr < 0. (Of course
generic physical systems have an infinite number of irrelevant variables but considering
only one will be enough to understand their roles).
(i) By iterating RG transformations as in the main text, show that the singular part of
the free energy can be written as
fsing = |uT |D/yT ϕ± (u0irr |uT ||yirr |/yT ),
where ϕ± are functions possibly different for uT > 0 or uT < 0, and u0irr is the initial value
(before RG transformations) of the irrelevant coupling.
(ii) Argue (without formal proof) that the functions ϕ± may raisonnably be expected to
be smooth.
Under this assumption, prove that
fsing = |uT |D/yt A0 + A1 u0irr |uT ||yirr |/yT + · · · ,

89
where A0 and A1 are non-universal constants.

• Exercise 5.3: Change of variables and covariance of RG equations.
Let us consider a theory with a finite number of relevant coupling constants that we
generically denote {g i }. Let us write the corresponding beta functions as (no summation
in the first term)
1X i j k
β i (g) = yi g i − Cjk g g + · · · .
2
jk
(i) Prove that, if all yi are non-vanishing, then there exist a change of variables from {g i }
to {ui }, with ui = g i + O(g 2 ), which diagonalizes the beta functions, up to two loops, i.e.
such that β i (u) = yi ui + O(u3 ).
(ii) Prove that, if all yi are zero, then the second and third Taylor coefficient are invariant
under a change of variables from {g i } to {ui }, with ui = g i + O(g 2 ).
That is: For marginal perturbation, the second and third loop beta function coeffcients
are independent on the renormalization scheme (alias on the choice of coordinate in the
coupling constant space)
(iii) Let expand the beta functions to all orders in the coupling constants:
X X
β i (g) = yi g i − Cji1 ,··· ,jn g j1 · · · g jn .
n>0 j1 ,··· ,jn
Prove that, if there is no integers pi , pj such that pi yi − pj yj ∈ Z, for i 6= j (in such

cases, one says they that there is non resonances), then there exists a change of variables
from {g i } to {ui }, with ui a formal power series in the g i ’s, with ui = g i + O(g 2 ), which
diagonalizes the beta functions as a formal power series in the ui ’s.
That is: There exist scaling variables, at least as formal power series.
90
6 Free field theory
The aim of this Chapter is to present the basic concepts of free statistical field theories.
These are Gaussian theories. Basics tools, including the Wick’s theorem, its graphical
representation, are introduced. The connection between free statistical field theories and
free quantum theories via analytic continuation is outlined.
6.1 Euclidean (free) field theories

We aim at describing statistical sums whose configurations are the possible values of fields
of the form Z
Z := [Dφ] e−S[φ] ,
where φ is a field x → φ(x) defined over a base space of dimension D and taking value
in some target space (say RN for simplicity). If we view this integral sum as arising
from a discrete lattice statistical model, the continuous description applies if the typical
correlation length ξ is much bigger than the lattice space a, i.e. ξ R a.
For a scalar field theory, the typical form of the action is S[φ] = dD x 21 (∇φ)2 +V (φ) ,

for some potential V .

The above statistical sum formally defines a measure on the random field configura-
tions. The least we can ask for giving a meaning to this measure is to specify what are
the correlation functions of the field at different points, that is
Z
1
hφ(x1 ) · · · φ(xn )i := [Dφ] e−S[φ] φ(x1 ) · · · φ(xn ).
Z
When dealing with field theory, we shall adopt the “standard” convention to denote ex-
pectation values by h· · · i instead of E[· · · ]. If we were more precise mathematically, we
would have to define more than just these n-points functions. This is actually beyond the
present understanding of generic field theory (except in some special cases). Nevertheless,
n-point functions of local fields is all what is needed as long as we don’t look at non-local
properties (say property related to extended objects or structures of field configurations).
We may imagine that the fields code for the shape of some kind of landscape or
membrane, say imbedded in some higher dimensional manifold and parametrized whose
coordinate(s) are the field components φ. We are then describing fluctuating shapes, a
notion which may be view as part of what random geometry can be. We may also want to
describe magnetic material, the coordinate will then be parametrizing the long wave length
behavior of the local magnetization. There are of course many other possibilities. The
field can be a scalar, a vector, etc., with multi-component, and/or take values in manifold
with/without internal structures depending on the physical problem. The form of the
potential of course also depend on the physical setup, say on relevant symmetries. For
instance, if we aim at describing the long distance physics of Goldstone bosons associated
91
to the symmetry breaking of a continuous group G down to a subgroup H, the Goldstone

field take values in the coset space G/H. These are then non free field theory.
• Free field theory in continuous space.
In this chapter we restrict ourselves to Gaussian scalar free field theory (we shall
deal with interacting theory in the following Chapters). They correspond to quadratic
potentials. One sets V (φ) = 21 m2 φ2 , so that the action becomes
Z
1 1
S[φ] = dD x (∇φ)2 + m2 φ2 .

(42)
2 2
The classical equation of motion is linear. It is the well-known Klein-Gordon equation:
−∆φ + m2 φ = 0, so that m is identified with (or more precisely called) the mass of the
field. As discussed in previous chapters, higher terms in the potential may be relevant
(not all polynomial are relevant but only the first few depending on the dimension) which
mean that they potentially modify the long distance behavior of the field configurations.
These terms induce self-interactions, we shall analyse in the following chapter.
By integration by part (with appropriate boundary condition), the action can be writ-
ten as Z
1
dD x φ(x) − ∆x + m2 φ(x),

S[φ] =
2
with ∆ = ∇2 the D-dimensional Laplacian. It is quadratic in the field φ. If we view the
field φ as an infinite dimensional vector with component φ(x), the action is a bilinear form
with matrix the positive operator (−∆ + m2 ).
Of primordial importance is the Green function G(x, x0 ), the inverse of this operator,
defined by the differential equation
(−∆x + m2 )G(x, x0 ) = δ(x, x0 ),
with δ(x, x0 ) the Dirac distribution (viewed as the identity operator acting on the space
of functions).
Since, the total action is quadratic, the field is going to be Gaussian (hence simple to
describe, see next Section) with zero mean and covariance the Green function. That is:
hφ(x)i = 0, hφ(x)φ(x0 )i = G(x, x0 ).
More will be described in a little while after a digression on Gaussian theories.

• Lattice free field theory.
Field theories emerging from statistical lattice models may also be defined on lattices.
Let us describe their free field version. For simplicity we consider theories on a square
lattice aZD , with a the lattice mesh (lattice spacing). Let eα , with α = 1, · · · , D, be unit
vector basis in RD . Let x = an be points of the lattice with n integer (i.e. point on ZD )
and φx be the field at that point. Notice that here x is the dimensional position, not the
dimensionless integer n, as otherwise we would use the notation φn . The lattice action
simply reads
aD X
S[φ] = φx [(−∆dis + m2 )φ]x ,
2 x
92
with ∆dis the discrete analogue of the Laplacian defined through finite difference by:
X
[(−∆dis + m2 )φ]x = a−2 − φx+aeα + (2 + (am)2 )φx − φx−aeα .

α
Clearly the discrete Laplacian is an infinite dimensional matrix [−∆dis ]x,x0 , indexed by
points on the lattice, which is symmetric with +2 on the diagonal and −1 on the first line
off the diagonal. The lattice action can alternatively be written as
aD X
S[φ] = φx [(−∆dis + m2 )]x,x0 φx0 ,
2 0
x,x
from which it is clear that the theory is Gaussian.

This action is diagonalized via Fourier transformed.Let
dD k
X Z
−ix·k
φ̂k = e φx , φx = D
eix·k φ̂k
x BZ (2π/a)
where the integration is over the Brillouin zone. For a D-dimensional square lattice of
mesh size a, i.e. aZD , the Brillouin zone is the hyper-cube BZ ≡ [− πa , πa ]D . The Laplacian
is of course diagonal in the Fourier basis. The action becomes
dD k
Z
1
S[φ] = φ̂−k (−∆dis + m2 )k φ̂k .
2 BZ (2π/a)D
In Fourier space, free field theory is thus a collection of i.i.d. Gaussian variables, indexed
by the momentum k, with mean and covariance determined by the discrete Laplacian.
Notice however, that there is no zero mode if m2 6= 0 but one in the massless case, and
that if the model was defined on the periodic torus the integration on the momentum
would be replaced by a discrete sum. See the exercise Section for further details.
6.2 Gaussian field theories

Here we summarize basic facts about Gaussian models. They all share the same structure,
only the explicit expressions for their two-point functions —their covariance— differ from
model to model.
• Discrete Gaussian models.
Let Λ be a lattice, or a domain on a lattice with a finite number of points (we may
have then to take the infinite volume limit to define the thermodynamic or scaling limit).
To any point j ∈ Λ we associate a (scalar) variable φj , which is the value of the field at
point j. We assume that this field is Gaussian (with zero mean), so that the measure is
1 Y
dφk e−S[φ] ,

Z
k∈Λ
with
1X 1
S[φ] = φj Mjk φk =: (φ, M φ), (43)
2 2
j,k
93
with M a real positive definite symmetric matrix, Mjk = Mkj . The factor Z in the above
formula, which is identified as the partition sum, ensures the proper normalization of the
measure –its total sum is one. The set of φk ’s form a (big) vector φ whose component are
those φk ’s. This is a random vector, a Gaussian vector.
We need to collect basic information on multi-dimensional Gaussian variables.
— The first is about the partition sum. We have (by diagonalization and from the 1D
result) Z
1 1 Y dφk
Z = [Dφ] e− 2 (φ,M φ) = [det M ]− 2 , [Dφ] = √ .
k∈Λ
2π
— The second is about the two point function. We have (by change of variable and contour
deformation)
E[φj φk ] ≡notation hφj φk i = [M −1 ]jk =: Gjk ,
with G = M −1 the inverse of the matrix M . The one point function vanishes: hφk i = 0.
— The third is about generating function (coding for higher moments and cumulants).
Let u be a dual vector (i.e. a vector in thePdual space) with component uj , and define its
contraction with the vector φ by (u, φ) = j uj φj . Then,
1 −1 u)
E[e−(u,φ) ] ≡notation he−(u,φ) i = e 2 (u,M . (44)
The proof is done by change of variables and contour deformations; convergence of the
integrals has to be discussed but there is no problem if M is positive definite. In order to
avoid any convergence problem (in case one does not have a good explicit control on the
measure) one usually defines the generating function with a purely imaginary vector, that
is through the formula hei(u,φ) i, and look for its analytic continuation. See the Exercises
section in Chapter 2.
Remark that this last expectation, as any correlation, can be viewed as the ratio of
the two partition functions, one with the external field, the other without
Z[u]
he−(u,φ) i = ,
Z[0]
with Z[u] the partition with u-dependent modified action
Z
Z = [Dφ] e−S[φ;u] , S[φ; u] = S[φ] + (u, φ).
The dual vector u plays the role of a source for the field φ.
• Gaussian generating functions.
The function Z[u] is a generating function for multi-point correlation functions in the
sense that its Taylor expansion near u = 0 yields the multi-point expectations:
∂ ∂ Z[u]
hφk i = − k he−(u,φ) iu=0 = − k

,
∂u ∂u Z[0] u=0
∂2 −(u,φ)
∂ 2 Z[u]
hφj φk i = he i u=0
= ,
∂uj ∂uk ∂uk ∂uj Z[0] u=0
∂3 ∂3 Z[u]
hφj φk φl i = − j k l he−(u,φ) i|u=0 = − j k l , etc.
∂u ∂u ∂u ∂u ∂u ∂u Z[0] u=0
94
It is actually useful to deal with the generating function of the simpler correlation func-
tions, called connected correlation functions. Let W [u] defined by
Z[u]
eW [u] := he−(u,φ) i = ,
Z[0]
or alternatively
W [u] = log(Z[u]/Z[0]).
By construction, it is the generating function of connected correlation functions:
hφk ic := ∂uk W [u]u=0 = hφk i,

hφj φk ic := ∂uj ∂uk W [u]

= hφj φk i − hφj ihφk i, etc.
u=0
It is sometimes alternatively called the ‘effective action’. The origin of the name ‘con-
nected’ comes from the diagrammatic representation of the correlation functions in terms
of Feynman diagrams that we shall discuss in the following Chapter. They correspond
to connected diagrams, those which cannot be split in parts without breaking one of the
lines of the diagrams.
• Wick’s theorem and graph representations.
For Gaussian theory, W [u] is quadratic (cf above)
1 1
W [u] = (u, M −1 u) = (u, Gu).
2 2
Thus, the two point function is the only non vanishing connected correlation function.
All the odd correlation functions vanish and for the even correlation functions –
not necessarily connected– can be written in terms of the two-point functions. Since
Z[u]/Z[0] = eW [u] is the generating function of the correlation function, one has,
hφj1 φj2 i = G(j1 , j2 ),

hφj1 φj2 φj3 φj4 i = G(j1 , j2 )G(j3 , j4 ) + G(j1 , j3 )G(j2 , j4 ) + G(j1 , j4 )G(j2 , j3 ), etc.
To simplify the notation we write G(j, k) instead of Gjk , with G = M −1 .

The way to compute the multi-point correlation functions in terms of the two-point
function for Gaussian model is called the “Wick’s theorem”. It allows to compute explicitly,
and recursively, all correlation functions. The general formula is
X
hφj1 φj2 · · · φj2n i = G(jσ1 , jσ2 ) · · · G(jσ2n−1 , jσ2n ) (45)
pairings
How many pairings do we have? To enumerate them, we pick (i) the number 1 and
associate to it another number, this yields (2n − 1) possibilities, (ii) then we pick the
next smallest number and we associated to it another number left, this yields (2n − 3)
possibilities, and so on. Altogether there are (2n − 1)(2n − 3) · · · 3 · · · 1 pairings, or else
(2n)!
2n n! pairings. For 2n = 2 this gives 1, for 2n = 4 this gives 3, for 2n = 6 this gives 15, etc.
√ number
This
n
grows faster than exponentially: the number of pairing
√ of 2n fields/points
n
is
2(2n/e) asymptotically (thanks to the Stirling formula n! ∼ 2πn(n/e) ).
95
One uses simple graphical representation expressing the relation between the multi-
point and 2-point functions. The multi-point functions are sum of monomials made of
product of the 2-point function G(j, k), say G(jσ1 , jσ2 ) · · · G(jσ2n−1 , jσ2n ). Each graph
represent of those monomial. The vertices of the graph label the field components and
there is one line connecting those vertices for each insertion of the 2-point function. The
construction of the graph is better explained by a few illustrative drawing than by words:
[...Give examples f or
the 2 and 4 point and 6 f unctions...]
Feynman diagrams are made of those graphs.

As it is clear from the construction a Gaussian field is specified by its two-point func-
tion. Thus, every statement made above apply to the Gaussian free field with the discrete
index j replaced by the continuous position x of the fields —or its momentum k if we work
in Fourier space— and with the Green function G(x, x0 ) as two-point function.
• Back the Gaussian free field theories.
All previous remarks apply directly to Gaussian field theories.
For the lattice theories, the mapping is {φj } → {φx } with x ∈ Λ in position space, or
{φj } → {φk } with k ∈ BZ in Fourier space, with the covariance matrix Gjk mapped to
the Green function
1
G(k, p) = (2π/a)D δ(k + p)
(−∆dis. + m2 )k
in Fourier space.
For continuous theories defined over RD the mapping is {φj } → {φ(x)} with x ∈ RD ,
with the covariance matrix Gjk mapped to the Green function G(x, x0 ) which is the kernel
of the massive Laplacian, so that
hφ(x)φ(x0 )i = [(−∆ + m2 )]−1 (x, x0 ) =: G(x, x0 ).
All higher order correlation functions are computed using Wick’s theorem.
Of course, continuous field theories over RD can also be formulated in Fourier space.
See below for explicit expression for the Green function in Fourier space.
Gaussian field theory can also be formulated on curve spaces (one then have to deal
with the Green function of the Laplacian defined using the curved metric) or on finite
domain (one then have to specify the field boundary conditions). See the exercise Section
for examples.
• Another definition of Gaussian free fields.
One can use the previous remarks to present a variant definition (possibly more rig-
orous...) of Gaussian free fields via its generating function against test function. Under
suitable conditions, the distribution of a random variable is specified by its characteristic
function —one is the Fourier transform of the other. In particular, given a Hilbert space V
with norm ||f ||2 for f ∈ V , on can defined a random Gaussian variable φ in V by defining
its generating function as
1 2
E[ei(f,φ) ] = e− 2 ||f || .
96
We can apply this to the Gaussian free field: we choose the space of function with
compact support for V equipped with the norm
Z
||f || = dD xdD x0 f (x) G(x, x0 ) f (x0 ),
2
with G(x, x0 ) the Green-function. The integral is convergent (except in the massless D = 2
case in infinite volume in which an extra finite volume regularization is needed).
The random variables are values of the form f → (f, φ), so that the field φ is identified
with a distribution. The duality is implemented as usual by testing the distribution against
f, Z
(f, φ) = dD x f (x)φ(x).
As generic distribution, the value of the distribution at a point does not exist (although
we use the notation φ(x), as usual in physics literature). This is reflected that the 2-point
function of the field at coinciding points is infinite
hφ2 (x)i = G(x, x) → ∞.
But the pairings (f, φ) makes mathematical sense.

More mathematically rigorous definitions of bosonic free field start from this type of
constructions.
6.3 Green functions

Here we collect information on Green functions and propagators.
• Green functions in the continuum.
The Green function is the kernel of the inverse of the operator (−∆ + m2 ). Thus it is
the kernel G(x, x0 ) solution of the differential equation
(−∆x + m2 )G(x, x0 ) = δ(x, x0 ).
In Fourier space, this reads
dD k ik·(x−x0 )
Z
0 0
G(x, x ) = G(x − x ) = e Ĝ(k),
(2π)D
with (recall that −∆ acts by multiplication by k 2 in Fourier space)

1
Ĝ(k) = . (46)
k2 + m2
Of course G is rotation invariant. Let r2 = x2 . In radial coordinates we have (here k is

the norm of the momentum)
π
dk k D−1
Z Z
SD−1
G(r) = dθ (sin θ)D−2 eikr cos θ .
(2π)D−1 2π k 2 + m2 0
97
with SD = 2π D/2 /Γ(D/2) the volume of the unit sphere in RD . We have

Z
1 1 kr
dθ (sin θ)2a eikr cos θ = Γ(a + )Γ( ) ( )−a Ja (kr)
2 2 2
Z
Ja (kr)
dk k a+1 2 = ma Ka (mr)
k + m2
Here Ka (·) is the modified Bessel function. This gives the formula
m D−2
G(x) = (2π)−D/2 ( ) K D−2 (mr). (47)
r 2
Explicit computations can be done in dimension D = 2 and D = 3.

Let us start with D = 3. We then have
Z π
k 2 dk
Z
1
G(r) = dθ sin θ eikr cos θ
(2π)2 k 2 + m2 0
eikr − e−ikr
Z +∞
k 2 dk −i k eikr
Z
1
= = dk .
(2π)2 k 2 + m2 ikr 4π 2 r −∞ k 2 + m2
We can further evaluate the integral using contour integral techniques. There is simple
poles in k-space at points ±im. To ensure the convergence of the integrals, we deform
and close the contour on the upper-half plane. By the residue theorem, we thus pick the
residue at k = im which yields
1 1 −mr
GD=3 (r) = e .
4π r
In D = 2, the formula does not simplify so much and we have
Z Z
dk k 1
GD=2 (r) = dθ eikr cos θ = K0 (mr),
(2π)2 k 2 + m2 2π
which cannot be written in terms of elementary function.

• Short and large distance behaviors.
1
At short distances we have Gr (r) ∝ rD−2
with a logarithmic divergence in dimension
D = 2. More precisely
1 2
GD=2 (r) = log( ) + ··· ,
2π mr
1
GD=3 (r) = + ··· ,
4πr
1
GD=4 (r) = + ··· .
4π 2 r2
At large distances we have an exponential decrease if the mass is non vanishing
e−mr
GD (r) ∝ , r m−1 ,
rD−2
98
This means that the correlation length (of the field with covariance G) can be identified
as the inverse of the mass: ξ = 1/m. This is important to remember.
R dD k eik·x
At m = 0 (and hence at infinite correlation length), G(x) = (2π) D k2
and
k D−3 dkdΩ ikr cos θ

Z
1
Gm=0
D (r) = D
e ∝ D−2 .
(2π) r
Note that the massless Green function are pure power laws. There isRa logarithmic
R dθ diver-
gence in dimension D = 2 (UV and IR divergence because GD=2 (r) = |k|≤Λ dkk (2π)2
eikr cos θ ).
We thus have to introduce the cut-off Λ ' 1/a and we have (or can choose)
1
Gm=0
D=2 (r) = log(Λr).
2π
It is easy to verify that (−∆x )G(x) = δ(x) in d = 2. The arbitrariness in choosing Λ
comes form the fact that the operator (−∆) possesses zero mode (the constant function).
• Lattice Green function.

On the lattice, the action is
dD k
Z
1
S[φ] = φ̂k [−∆dis. + m2 ]k φ̂k .
2 BZ (2π/a)D
and the Green function reads
dD k eik·x
Z
G(x) = D dis. + m2 ]
,
BZ (2π/a) [−∆ k
where [−∆dis. + m2 ]k = 2a−2 α η − cos(ak.eα ) , with η = (2 + a2 m2 )/2, is the discrete

P
Laplacian in Fourier space. We see that there is a natural momentum cut-off |k| ≤ Λ ' 1/a
with a the lattice mesh size. It is also important to realize that m has the dimension of
the inverse-length.
How the continuous limit is taken? First the limit is taken at x fixed. Since x = an
is a point on the lattice, this means we take the limit fixing the “physical” distance (not
the lattice distance defined as the minimal number of step to arrive at point x). The mass
is also fixed in this limit, which means that the correlation length, measured in terms
of the physical distances (not the lattice distance) is kept fixed. For instance, in 1D the
continuous limit consists in the limit a → 0, n → ∞ with x = an fixed. Similarly in
[−∆dis. + m2 ]k , we keep m fixed, that is we keep ξ = 1/m the ‘physical’ correlation length
fixed (not the lattice correlation length which would be the product am).
6.4 Products and composite operators

• Normal ordering and composite operators.
Product of operators at coincident points are singular. Let us for instance look at
the product of the Gaussian field φ at two neighbour points. Let us image inserting φ(x)
99
and φ(y), with x near y, in a correlation function containing many other insertions of
φ at points at finite distance away from x and y, say at points ξk . We can compute
such correlation function using the Wick’s theorem. There are then two kinds of Wick’s
contactions: (a) those for which x and y are contracted together producing the Green
function G(x, y) and the other points are contracted together, and (b) those for which x
is contracted to one of the points ξk and y with another one. Since G(x, y) is singular as
x → y, the Wick’s contraction of the first type are singular as x → y whereas those of the
second type are regular is that limit. Using G(x, y) ' const./|x − y|D−2 as x → y we can
write
φ(x)φ(y) = G(x, y) + regular 'x→y const./|x − y|D−2 + · · · ,
where the ‘regular’ terms are regular at x = y.
Alternatively, we can give a name of this remaining (regular) part and write
φ(x)φ(y) = G(x, y)+ : φ(x)φ(y) :,
where : φ(x)φ(y) : is called to the ‘normal ordered product’ of φ(x) and φ(y). This equation
may be taken as a definition of the normal ordered operator. Alternatively, : φ(x)φ(y) : is
defined inside any correlation functions as φ(x)φ(y) but without self Wick’s contractions.
This normal order product is now regular (because there is no self contraction) so that we
can take the limit of coincident points and define the product φ2 as
: φ2 (x) : = : φ(x)φ(x) : = lim φ(x)φ(y) − G(x, y) .

y→x
There are actually some freedom on the way to define this product operator as its con-
struction only requires subtracting the singular part of the Green which is defined up to
regular terms.
Of course this can be extended to higher point functions and higher products. It is a
good exercise to compute the correlation functions h: φ2 (x) : : φ2 (y) :i. For instance,
: φ2 (x) : φ(y) = 2G(x, y)+ : φ2 (x)φ(y) :,

: φ2 (x) : : φ2 (y) : = 2G(x, y)2 + 4G(x, y) : φ(x)φ(y) : + : φ2 (x)φ2 (y) :,
where : φ2 (x)φ(y) : means the insertion of the product φ2 (x)φ(y) without self Wick’s
contraction, and similarly for : φ2 (x)φ2 (y) :. In particular we can now define the fourth
power of φ as
: φ4 (x) : = lim : φ2 (x) : : φ2 (y) : −2G(x, y)2 + 4G(x, y) : φ(x)φ(y) : .

y→x
And ad-finitum, recursively.

From this analysis, we understand that field theory operators, or observables, are
sensible to —or are testing— the neighbourhood of the points they are attached to.
• Geometrical interpretation.
We now describe a geometrical interpretation of these operators in terms of random free
path. Recall the definition of random free path given in Chapter 2. There, we computed
100
partition function Zpath (0, x) (we call it Z(x) in that Chapter) as the Boltzmann sum over
free paths from the origin to the point x and we found that
I
Zpath (0, x) ∝ hx| |0i,
I − µΘ
with Θ the lattice adjacency matrix and µ the fugacity. In the scaling limit (a → 0, µ →
µc ), it satisfies (−∆ + m2 )Zpath (0, x) = δ(x). We thus have the identification
Zpath (0, x) = hφ(x)φ(0)i,
with φ a massive Gaussian free field. Alternatively, by the expanding the inverse I/(I − µ Θ)
in power of µ, we may view the Gaussian correlation function as a sum over paths (because
the N -th power of the adjacency matrix ΘN codes for the sum of paths of length N ). We
may then interpret the field φ(x) as conditioning on a curve starting or arriving at the
point x.
Similarly, given four points x1 , x2 , x3 , x4 , the partition function Zpath ([x1 , x2 ]; [x3 , x4 ])
for pairs of free (possibly intersecting) paths joining x1 to x2 and x3 to x4 respectively is
Zpath ([x1 , x2 ]; [x3 , x4 ]) = G(x1 , x2 )G(x3 , x4 ).
Thus, the partition sum Zpath (x1 , x2 , x3 , x4 ) over free paths joining these points pairs,
independently of the set of connected pairs, is the four point functions
Zpath (x1 , x2 , x3 , x4 ) = hφ(x1 )φ(x2 )φ(x3 )φ(x4 )i.
As a consequence we can identify the normal operator : φ2 (x) : as conditioning on two

curves starting (or arriving) at point x.
Similar, the higher order normal ordered powers : φk : condition on k curves starting
the operator insertion point:
: φk (x) : = create/absord k curves at x.
The fact that : φ2 :, the operator creating two curves at a point, has scaling dimension
D − 2 can be used to argue that the fractal dimension of free paths is Dfrac = 2. Notice
also that we may identify the diagrams representing the Wick’s contraction with the free
path configuration we are summing over.
6.5 Quantization of free field theories

We go back to Minkowski (real-time) quantum field theory and make connection with the
Euclidean and Minkowski versions of the field theory.
• Path integral quantization.
Remember that the weight of the QM path integral is e(i/~)S with S the
R action. We use
the notation x = (x, t) and D = d + 1. The Euclidean action was S[φ] = dD x 1 (∇φ)2 +
2
V (φ) with V (φ) = 21 m2 φ2 for a free field. The analytic continuation to go from the

101
Euclidean time tE to the Minkowski time t is t → itE . The Minkoswki version of the
action is
Z Z Z
1 2 1 1
d d
S[φ] = dtd x L = dtd x (φ̇) − (∂x φ) − V (φ) = dtdd x (∇φ)2 − V (φ) .
2

2 2 2
where here the gradient square (∇φ)2 is defined using the Minskowky metric. The Feyn-
man path integral reads Z
[Dφ] e(i/~) S[φ] .
The dynamical variables are the field configurations (at all space points x)
q → φ(x), p → π(x).

The canonical commutation relations are φ(x) , π(y) = i~δ(x − y). They are formally
represented on functional of field configurations with the field operator acting by multi-
plication and the conjugated momentum by derivative. We formally have a basis a state
made of field configuration |{φ(x)}i which are eigen-states of the field operator
φ(x0 )|{φ(x)}i = φ(x0 )|{φ(x)}i.
If we give ourselves two configurations at two different times (at two different time slices),
{φ0 (x)} at time t0 and {φ1 (x)} at time t1 , the quantum amplitude is (formally again)
defined by the path integral with specified boundary condition:
Z
h{φ1 (x)}|U (t1 , t0 )|{φ0 (x)}i = φ(x,t )=φ (x) [Dφ] e(i/~) S[φ] .
0 0
φ(x,t1 )=φ1 (x)
Of course the normalization is not very well specified, but this ill-defined normalisation
cancels when computing normalized correlation functions.
Below we shall argue that taking the limit t0 → −∞ and t1 → +∞ this amplitude
projects onto the vacuum state, that is the path integral from −∞ to +∞ (without
specified boundary condition) is the vacuum to vacuum amplitude. The simplest (naive)
argument for the limit to exist requires introducing an infinitesimal small imaginary part
(or doing an analytic continuation) to the time parameters which then produce the Eu-
clidean transfer matrix which projets on the vacuum at large time. See Chapter 3 and
4.
Unless otherwise specified, we now set ~ = 1.
• Canonical quantization.
The canonical momentum is π(x, t) = φ̇(x, t) with canonical commutation relation

π(x, t) = φ̇(x, t), φ(x, t) , π(y, t) = iδ(x − y).
The action can be written in the hamiltonian form

Z
S[φ] = dtdd x π(x)φ̇(x) − H ,

102
with Z Z
d 1 1
dtdd x π(x)2 + (∂x φ(x))2 + V (φ(x) .

H= d xH =
2 2
For free field theory of non-zero mass m2 , this reads
Z
1
dtdd x π 2 + (∂x φ)2 + m2 φ2 .

H=
2
Let us go to Fourier space
dk ik·x
Z Z
φ̂(k) = dd x e−ik·x φ(x), φ(x) = e φ̂(k),
(2π)d
and similarly for the momentum. The canonical commutation relation now reads [π̂(k), φ(k0 )] =
−i2πδ(k − k0 ). The hamiltonian (density) reads
dd k
Z
1 2 2 2
ωk2 = k2 + m2 .

H= π̂(k) + ωk φ̂(k) ,
2 (2π)d
This hamiltonian codes for a collection of independent harmonic oscillators, indexed by

the momentum k. It is diagonalized by introducing annihiliation-creation operators Ak
and A∗k with commutation relation
Ak , A∗p = (2π)d δ(k − p),

defined by r r
ωk 1
π̂(k) = (Ak + A∗k ), φ̂(k) = −i (Ak − A∗k ).
2 2ωk
To be finite the hamiltonian needs renormalization by normal ordering. See any book on
basics quantum field theory for more detailed.
The representation of these commutations relations is done as usual on the Fock space
F ' ⊗k Fk over the vacuum |0i annihilated by all Ak , that is
Ak |0i = 0, h0|A∗k = 0,
for all k. States in the Fock space are obtained by repeated action of the creation operators
on the vacuum: A∗k1 · · · A∗kN |0i.
Alternatively, we can solve directly the wave equation and ‘quantize’ them. The field
φ satisfies the wave equation
− ∂t2 + ∂x2 + m2 φ(x, t) = 0.

Hence, by Fourier transform
dd k
Z
1
ei(k·x−ωk t) Ak + e−i(k·x−ωk t) A∗k

φ(x, t) = d
√
(2π) 2ωk
103
The momentum π(x, t) = φ̇(x, t) is
dd k
Z r
ωk i(k·x−ωk t)
Ak − e−i(k·x−ωk t) A∗k

π(x, t) = −i e
(2π)d
2
∗ d

We impose the canonical commutation relation Ak, Ap = (2π) δ(k − p). They imply
the canonical commutation relation φ(x, t) , π(y, t) = iδ(x − y), as it should be.
We could alternatively set Ak = √a2ωk
(if ωk does not vanish). The formula for field φ
k
looks then a bit simpler because there is no square root:
dd k
Z
ei(k·x−ωk t) ak + e−i(k·x−ωk t) a∗k .

φ(x, t) = d
(2π) 2ωk
The operators ak then satisfy
ak , a∗p = (2π)d 2ωk δ(k − p).

The representation of these canonical commutations relations is as above.

• Wick’s rotation and time ordering.
Let us compute the vacuum two point functions. We have (using only the commutation
relations of the creation-annihilation operators and the fact the defining property of the
vacuum)
dd k 1 ik·(x−y) −iωk (t−s)

Z

h0|φ(x, t)φ(y, s)|0i t>s
= e e ,
(2π)d 2ωk
for t > s. We can (again) check that from this formula that the canonical commutation
relations are satisfied, that is h0|[φ̇(x, t) , φ(y, t)|0i = −iδ(x − y).
We now compare this formula to the analytic continuation of the Euclidean Green
function. Recall that (with τ denoting the Euclidean time)
dd kdk0 ik·(x−x0 )+ik0 (τ −τ 0 )

Z
0
GE (x − x ) = e ĜE (k),
(2π)D
with
1
ĜE (k) = .
k2 + k02 + m2
As above, we do the integration over k0 by contour integral. There is a pole at k0 =
±iωk . We pick one or the other depending whether τ > τ 0 or the reverse (because the
contour deformation is chosen such that the integral converges). Since the Euclidean
Green function is symmetric under the exchange of x and x0 , we are free to choose either
case. Suppose τ > τ 0 . For the integral to converge we then have to close the contour on
the upper half plane and pick the pole at +iωk . Hence, for τ > τ 0 ,
dd k 1 ik·(x−y) −ωk (τ −τ 0 )
Z
0
GE (x − x ) = e e .
(2π)d 2ωk
104
If τ 0 > τ , we have to close the contour in the lower half plane and pick the other pole,
so that we get the same result but with τ 0 and τ exchanged. That the general formula is
with |τ − τ 0 |, that is
dd k 1 ik·(x−y) −ωk |τ −τ 0 |
Z
GE (x − x0 ) = e e .
(2π)d 2ωk
We now implement the analytic continuation. Remember that to go from real time
quantum mechanic to Euclidean time we had to set t = −iτ with τ > 0 (for convergence).
That is, we analytically continue the real time to its lower complex plane. The reverse it
τ = it (but we have to keep track of the imaginray part to keep track of the ordering). So
let τ = i(t − i) and τ 0 = i(s − i0 ). For > 0 we have <(τ − τ 0 ) > 0, so the Euclidean
0
integral produces e−ωk (τ −τ ) . And hence the analytic continuation
dd k 1 ik·(x−y) −iωk (t−s)

Z
0 0
lim GE (x, τ = it + ; y, τ = is + ) = e e
,0 →0+ (2π)d 2ωk
>0

= h0|φ(x, t)φ(y, s)|0it>s .
Thus we learn that real time correlation functions are obtained by analytic continuation
of the Euclidean expectation values, and that the ordering of the operator is induced by
the precise way the analytic continuation is done.
The fact that we have to add a tiny negative imaginary part to order them has a
natural interpretation. Remember that when going to the unitary evolution e−itH to the
Euclidean one e−τ H we had to set t = −iτ . So adding a negative imaginary part to t
amount to slightly evolve the operator forward in Euclidean time. We order the operator
according to their Euclidean time (e.g. as suggested by the transfer matrix formalism).
This is summarized by the fact the analytic continuation of the Euclidean propagator
gives the Feynman propagator corresponding to time ordered correlation functions. Under
the wick rotation, x0 = τ = it and k0 = −iω, we have

i
GE → GF (x, y) = = hT · φ(x)φ(y)i,
∆ − m2 + i0+ x,y
with ∆ = −∂t2 + ∂x2 the Minkowski Laplacian and T the time order operator. But one has
to be careful how the analytical continuation is done (cf the above discussion).
6.6 Exercises
• Exercise 6.1: Translation invariance and the stress-tensor
The aim of this exercise is to see some aspect of the relation between translation invari-
ance and the stress-tensor. R Let us consider classical scalar field theory with Lagrangian
L[φ, ∂φ] and action S[φ] = dD x L[φ, ∂φ]. Recall that maps extremalizing this action are
said to be solution of the classical equations of motion, which reads
∂L ∂L
∂µ = .
∂(∂µ φ)(x) ∂φ(x)
105
These equations are the Euler-Lagrange equations.

(i) Consider an infinitesimal field transformation φ(x) → φ(x) + (δφ)(x). Suppose that,
under such transformation the Lagrangian variation is δL[φ, ∂φ] = ∂µ Gµ so that the
action is invariant. Show that the following Noether current
∂L
J µ = (δφ) − Gµ ,
∂(∂µ φ)
is conserved on solutions of the equations of motion.
(ii) Let us look at translations x → x + ε a. How does a scalar field φ transforms under
such translation? Argue that if the Lagrangian density is a scalar, then δL = ε aµ ∂µ L.
Deduce that the action is then translation invariant and that associated conserved Noether
current is Jaµ = Tνµ aν with
∂L
Tνµ = (∂ν φ) − δνµ L.
∂(∂µ φ)
This tensor is called the stress-tensor. It is conserved: ∂µ Tνµ = 0.
R(iii)DFind the expression of the stress-tensor Tνµ for a scalar field theory with action S[φ] =
1
d x 2 (∇φ)2 + V (φ) .
• Exercise 6.2: Lattice scalar field and lattice Green function
Recall that lattice scalar free theory is defined by the action
aD X
S[φ] = φx [(−∆dis + m2 )]x,x0 φx0 ,
2 0
x,x
where φx are the value of the field at point x on the lattice and ∆dis discrete Laplacian on
that lattice. We here consider only D-dimensional square lattice of mesh size a, i.e. aZD .
Let us also recall that the Fourier transforms in aZD are defined by
dD k
X Z
φ̂k = e−ix·k φx , φx = D
eix·k φ̂k
n BZ (2π/a)
where the integration is over the Brillouin zone, which is the hyper-cube BZ ≡ [− πa , πa ]D .
i) Verify that the Laplacian acts diagonally in the Fourier basis, with
X
(−∆dis + m2 )k = 2a−2

η − cos(a kα ) ,
α
with η = (2 + a2 m2 )/2 and kα the component of the momentum k in the direction α.

ii) Verify that in the Fourier basis the free field action reads
dD k
Z
1
S[φ] = φ̂−k (−∆dis + m2 )k φ̂k .
2 BZ (2π/a)D
(iii) Deduce that in Fourier space, a scalar free field is thus equivalent to a collection of
i.i.d. Gaussian variables, indexed by the momentum k, with mean and covariance
1
hφk i = 0, hφk φp i = (2π/a)D δ(k + p).
(−∆dis + m2 )k
106
• Exercise 6.3: The Green function in 2D

[... To be completed...]
• Exercise 6.4: Fractal dimension of free paths.
The fractal dimension Dfrac of a set embedded in a metric space may be defined through
the minimal number N of boxes of radius need to cover it by Dfrac = lim→0 log N / log(1/).
(i) Prove that the fractal dimension of free paths is Dfrac = 2 using the fact that the
composite operator φ2 , with φ a (massless) Gaussian free field, is the operator conditioning
on two paths emerging from its insertion point.
• Exercise 6.6: Two ways to compute the free energy
The aim of this exercise is to compute the free energy, or the partition function, of a
massless free boson in space dimension d = 1 at temperature T = 1/β. Let D = d + 1.
Recall that the partition function is defined as Z = Tr(e−βH ) where the trace is over the
quantum Hilbert space with H the hamiltonian. Let us suppose that the quantum theory
is define dover an interval I of length L. We shall be interested in the large L limit.
(i) Argue (see Chapter 3) that the partition function is given by the Euclidean path integral
on the cylinder I × S1 with a radius β:
Z
Z= [Dφ] e−S[φ] .
φ(x,β)=φ(x,0)
We shall compute the partition function by quantizing the theory along two different
channels (see Figure):
(a) either taking the direction S1 as time, this Euclidean time is then period with period
β;
(b) or taking the direction I as time, this time then runs from 0 to L with L → ∞. Global
rotation invariance implies that this to way of computing gives identical result. Let us
check. On the way this will give us a nice relation about the Riemann ζ-function.
(ii) Explain why the first computation gives Z = e−βLF (β) , where F the free energy.
(iii) Explain why the second computation gives Z = e−LE0 (β,A) with E0 (β) = β E0 (β)
where E0 is the vacuum energy and E0 is the vacuum energy density (this is the Casimir
effect).
(iv) Show that the free energy density of a massless boson in one dimension is:
Z Z ∞
1 dk 1 dx
F= log(1 − e−β|k| ) = 2 log(1 − e−x ).
β 2π β 0 π
(v) Compute the integral to write this free energy density as
1
F =− ζ(2).
πβ 2
We have introduce the so-called zeta-regularisation. Let ζ(s) := n>0 n1s . This function
P
was introduced by Euler. This series is convergent for <s > 2. It is defined by analytic
107
continuation for other value of s via an integral representation.

(iv) Show that the vacuum energy density is E0 (β) = β1 n 12 2nπ
P
β . (v) This is divergent.

Argue that a regularization based on analytic continuation gives
2π
E0 (β) = ζ(−1).
β2
π2
(vi) Conclusion: A remarkable fact is that ζ(2) = 6 and that the analytic continuation
1
of ζ gives ζ(−1) = − 12 . Thus
π
F(β) = E0 (β) = − .
6β 2
1
Actually, we could reverse the logic: physics tells us that ζ(−1) has to be equal to − 12
because E0 has to be equal to F.
• Exercise 6.7: Radial quantization (at least in 2D).
[... To be completed...]
108
7 Interacting field theory: basics
In this Chapter we go deeper in analysing the structures and the basic tools of statistical
field theories. We first write consequences of symmetries on correlation functions which are
called Ward identities. We then go on by introducing generating functions of multi-point
correlation functions and make the analogy with thermodynamic functions. This leads to
the important concept of effective action. We describe how perturbation theory can be
formulated and how it can be encoded in synthetic way in Feynman diagrams. We make
the connection between the irreducible components of the Feynman diagram expansion
and the effective action. Finally, we study in different ways –via loop expansion or via
saddle point approximation– the O(N ) vector model at large N .
7.1 Preliminaries
For an interacting scalar theory the action is of the form S[φ] = S0 [φ] + SI [φ] with S0 [φ]
the free action. Standard examples are
Z Z
1
S0 [φ] = dD x[(∇φ)2 + m2 φ2 ], SI [φ] = dD x V (φ),
2
with V a potential coding for self-interaction. The Boltzmann weights of field configura-
tions then read
[Dφ] e−S[φ] = [Dφ] e−S0 [φ] e−SI [φ] .
The first part is the Boltzmann measure for a free Gaussian theory so that we can (for-
mally) factorize the measure on field configurations as
dD x V (φ)
R
dP[φ] = dPfree [φ] × e− .
This formula is however only formal for many reasons (the field φ is not a random func-
tion but a random distribution, the interacting and free measures may be singular, etc).
Making sense of such formula is part of what statistical field theory is aiming at and the
renormalization group is a key tool to reach this goal. It requires introducing a UV cut-off
that we may think as coding for an underlying lattice. We shall implicitly assume below
that such UV cut-off is present.
Keeping this prerequisite in mind, we can look at what constraints these measures, or
the correlation functions they code for, should satisfy. This is the purpose of the Ward
identities and the introduction of generating
R D functions. Perturbation theory consists in
developing the interaction terms e − d x V (φ) in power of V and analyzing each terms of
this expansion. These two techniques —symmetries and perturbative expansion— are two
basic, fundamental, ways of computing —or of making sense— of correlation functions
in statistical field theory. A third possible ways may be applied when there is a small
parameter such that the statistical path integral can be evaluated via a saddle point
109
approximation —in a way analogous to the WKB approximation in Quantum Mechanics

or to the small noise problems we discussed in the first Chapter. This last method will be
applied to the O(N ) vector model at large N at the end of this Chapter.
Field configurations are random, by construction, and a way to test their statistics is
to look at expectations of functions, often called observables, of those configurations. If
they depend locally on the field configuration at a point x, those observables are called
either local observables, operators or fields, indifferently. Standard (formal) observables
are products of polynomials of field values at given points, say monomials φp (x) with p
integer, but we may aim at considering any local functions of the field, say
F1 [φ(x1 )] · · · FN [φ(xN )].
Of course these naive writings are purely formal since, on the one hand, statistical fields are
singular objects (in the previous Chapter dealing with Gaussian free fields we gave them a
meaning in terms of random distributions), and on the other hand, product of statistical
field at coinciding points is ill-defined (as discussed in previous Chapter in the case of
free Gaussian field) and we will have to give a meaning to such products. Again different
strategies can be developed to attain this goal, say either using symmetries to specify
the properties of fields and observables, or using perturbation expansions to recursively
defined those observables, or else.
As mentioned in the previous Chapter we may look at more general theories with fields
of different types (scalar, vectors, etc) taking values in different manifolds (flat or curve
spaces, group manifolds or quotient spaces, etc.) with more or less simple structures. For
instance, we may consider theories not over the flat space RD but over a metric space
M, equipped pwith a metric g. For scalar field theories the kinetic term is then going to
be 21 dD x |g| g µν ∇µ φ∇ν φ, with |g| the metric determinant. More generally, we may
R
consider scalar fields taking values in another target space E, equipped with a metric G,
with local coordinates φA . The maps x → φA (x) are then maps from the base space M
—sometimes called the world-sheet— to the target space E, and the statistical field theory
codes from statistical properties ofpsuch maps. The kinetic part of the action could for
instance be chosen to be 12 dD x |g| g µν ∇µ φA ∇ν φB GAB , which is not a free action if
R
the metric G is φ-dependent. Such models are generically called sigma-models.

Thus in general, statistical field theories and their expectations,
hO1 (x1 ) · · · iM,g;E,G,··· ,
depends on collection of data, of geometrical nature (say, spaces, metrics, etc), or not (say,
external fields, sources, etc).
Below, we shall restrict ourselves to scalar field (mostly with one component) on flat
space.
7.2 Symmetries and Ward identities

In classical field theory, conservation laws are associated to symmetries –and reciprocally.
If a classical action is invariant under some symmetries, there exists an associated con-
served Noether current. The aim of this section is to understand how this translates in
statistical field theory.
110
• Symmetry and invariance

Symmetries refer to invariances of statistical expectations or of correlations functions.
There are different kinds of transformations: those which leave the space points fixed and
act only the field components (for example, group symmetries, etc), and those which act
on the space point (for example, translation, rotation, etc). If all the geometrical data
are preserved, the transformations act on observables/operators and transform them as
O(x) → Ô(x) say. In this case invariance of the correlation functions means that
hO1 (x1 ) · · · iM,g;··· = hÔ1 (x1 ) · · · iM,g;··· .
The aim of this Section is to show some of consequences of these symmetries and their
relations with the (classical) symmetries of the action defining (maybe formally) the statis-
tical field measure. These invariance relations give constraints on the correlation functions.
Sometimes, the logic can be reversed and the correlation functions in a symmetric enough
field theory could be defined from their symmetry properties (this is in particular the case
in a large class of conformal field theory in dimension two).
• Noether’s current and invariances.
Classically, symmetry and conservation law are tied, and if the classical action possesses
some symmetries then there exists a conserved Noether current J µ , ∂µ J µ = 0. The
Noether current is in general defined by looking at the the variation of the action. If the
field is transformed according to φ(x) → φ(x) + (δφ)(x), the variation of the action for
a non constant is going to be proportional to the derivative of , so that we may write
the variation of the action for a non constant transformation parameter as
Z Z
δS[φ] = − d x (∂µ )(x) J (x) = dD x (x) (∂µ J µ )(x),
D µ
by integration by part. This defines the Noether current, and it is a simple exercise in
classical field theory to check that it is conserved on solution of the classical equation of
motion.
We now look at what are the consequences of this transformation law in the path
integral formulation of statistical field theory. Those will be consequences of a change of
variables in the path integral.
Let us start with the path integral [Dφ] e−S[φ] , and let us image implementing the
R
change of variables φ(x) → φ̂(x) with φ̂(x) = φ(x) + (x) (δφ)(x) with (x) small but
position dependent. The point to understand is : how does the path integral measure
varies under such transformation? They are two contributions to this variation, one from
the action, the other from the change of variables.
Let us first look at the variation of the action S[φ] = dD x L[φ, ∂φ]. As we assumed
R
above, the variation of the action is

Z
δS[φ] = dD x (x) (∂µ J µ )(x). (48)
Here we are computing the variation of the action for any field configuration and not only
for solution of the classical equations of motion (because we are going to integrate over
111
them), so (∂µ J µ ) maybe non zero (it is a fluctuating quantity). Under the change of field
variables, φ → φ̂, the path integral measure change according
Z
−S[φ] −S[φ̂]
[Dφ] h i
[Dφ]e = [Dφ̂] e × Det 1 − dD x (x) (∂µ Jˆµ )(x) ,

[Dφ̂]

where Det [Dφ] is the Jacobian of the transformation.

[Dφ̂]
We will suppose that the Jacobian of the transformation φ → φ̂ is trivial:
[Dφ]
Hypothesis : Det = 1.

[Dφ̂]
This is an important but delicate point. To be defined and evaluated, this determinant
needs to be regularized (say by discretizing the theory, by defining it on a lattice, etc). This
regularization amounts to discretize or regularize the path integral measure. It may happen
that it is impossible to regularize this measure in such way to preserve its symmetry. In
such case the Jacobian will not be trivial. This is call a (quantum) anomaly, see below.
Under this hypothesis, we get that
Z
[Dφ]e−S[φ] = [Dφ̂] e−S[φ̂] × 1 − dD x (x) (∂µ Jˆµ )(x) + · · · .

for (smooth enough) function (x). Integrating over φ (or φ̂) gives:
h(∂µ J µ )(x)i = 0. (49)
That is the current is conserved in mean (away from field insertions, see below). This is
an analogue of the Noether theorem. (See the Exercise section of Chapter 4 to see the
lattice analogue of this construction).
• Field insertions and Ward identities.
Things becomes more interesting when generalizing this construction with insertions
of operators, say O(y), or products of such operators. This operator may be thought as a
local function of the field φ. We consider the expectation value:
Z
1
hO(y)i = [Dφ] e−S[φ] O(y),
Z
with Z the partition function, Z
Z= [Dφ] e−S[φ] .
Let us do the change of variable φ → φ̂. Suppose that under such transformation the
operator O(y) transforms as O(y) → O(y) + (y)(δO)(y). We then get (after renaming φ̂
into φ):
Z hZ i Z
−S[φ]
[Dφ]e d x (x) (∂µ J )(x) O(y) = [Dφ]e−S[φ] (y)(δO)(y),
D µ
112
for any (x). Hence

h(∂µ J µ )(x) O(y)i = δ(x − y) h(δO)(y)i.
Generalization to products of operators is clear.
We can also use the Stoke’s theorem to write differently the same equation3 . Let us
pick a ball B and integrate the current normal
R to the ball (i.e. evaluate the flux of this
current through B). Let us call this flux B n · J. Then:
Z N
Y X
h n·J · Ok (yk )i = hO1 (y1 ) · · · (δOk )(yk ) · · · ON (yN )i. (50)
B k=1 yk inside B
These relations are called “Ward identities”. They express the fact that (∂µ J µ )(x) = 0
away from field insertions but that there is a contact term coding for the field transfor-
mation when the current is near to the operator:
(∂µ J µ )(x) O(y) ' δ(x − y)(δO)(y), locally (51)
where (δO)(y) is the infinitesimal transformation of the field. Here, locally means inside
any correlation functions with other fields inserted at points away from x and y.
Ward identities are analogous to Gauss law in electrodynamic –but within statistical
field theory– because, if we view the variation (δO) of a field O as coding for its charge, we
infer from the Ward identities that the flux of the current through a ball B is equal to the
sum of all charges inside this ball. They give constraints and information on correlation
functions.
• Charge currents.
Let us give a simple example: u(1) symmetry for a complex field φ with action
Z
1
S[φ] = dD x (∂µ φ)(∂ µ φ∗ ) + V (φφ∗ ) .

2
The Lagrangian is clearly invariant under the u(1) symmetry φ → eiα φ with α real. The
infinitesimal transformation is φ → φ + iαφ and φ∗ → φ∗ − iαφ∗ . It is clear that the
measure [Dφ] is invariant under this transformation, so that the Jacobian is one. The
Noether current is (we have absorbed a factor i)
J µ = i(φ ∂ µ φ∗ − φ∗ ∂ µ φ∗ )
This current is conserved ∂µ J µ = 0 away from operator insertions (and inside correlation
functions).
Let us look at what happens if there are field insertions. We start with the expectation
values hφ(y1 ) · · · φ(yp ) · φ∗ (z1 ) · · · φ∗ (zq )i. When doing the change of variables, we have to
implement the field transformation. Hence we get
h(∂µ J µ )(x) φ(y1 ) · · · φ(yp ) · φ∗ (z1 ) · · · φ∗ (zq )i

X X
δ(x − zk ) hφ(y1 ) · · · φ(yp ) · φ∗ (z1 ) · · · φ∗ (zq )i.

=i δ(x − yj ) −
j k
3
This way of writing the Ward identities will be useful in conformal field theories, see Chapter 8.
113
That is, locally within any correlation functions (with other field insertions away from the
point x and y) we have
(∂µ J µ )(x) φ(y) ' iδ(x − y)φ(y),

(∂µ J µ )(x) φ∗ (y) ' −iδ(x − y)φ∗ (y), locally.
The operator φ has charge +1, its complex conjugated φ∗ charge −1. Note that if we
integrate the previous Ward identity over a ball of infinitely large radius, the l.h.s. vanishes,
assuming that J µ decrease fast enough at infinite (which is a statement about the absence
of symmetry breaking), while the r.h.s. is proportional to (p − q). Hence the above
expectation is non-zero only if p = q, i.e. only if the total charge is balanced (a result
that we could have obtained directly from the invariance of the measure). This is charge
conservation.
• The stress-tensor.
Most of the field theories we are studying here are translation and rotation invariant.
As a consequence there are Ward identities associated to those symmetries. They are
related –by construction or by definition– to properties of the stress-tensor which codes
for the response of the action —or more generally for the statistical measure— under a
small space diffeomorphism. If we change variable x → x + ξ(x) with 1 and ξ µ (x)
some vector field then, by definition, the action varies as
Z
δS[Φ] = dD x(∂µ ξ ν (x)) Tνµ (x),
with the stress-tensor. See the exercise Section for an illustration.

• Anomalies.
We here make a few comments on cases for which the Jacobian associated to the change
of variable φ → φ̂ is non-trivial. In such case, the current is not anymore conserved. This
is call a (quantum) anomaly 4 .
Recall that the transformations we look at are local in the field variables. They are
of the form φ(x) → φ̂(x) = φ(x) + (x)(δφ)(x) + · · · . Thus it is natural to expect that
the default for the Jacobian to be trivial is going to be local in the small parameter (x).
That is: we expect to first order in (x) that
Z
Dφ Dδφ
+ · · · = 1 + dD x (x) F (x) + · · · ,

Det = Det I −
Dφ̂ Dφ̂
for some field functional F (x). Such ansatz is justified by using a formal generalization
to infinite determinant of the expansion Det[1 + M ] = 1 + Tr(M ) + · · · valid for finite
dimensional matrices. To compute F (x) is of course much harder, it requires regularizing
the infinite dimensional Jacobian, and it is very much model dependent.
4
A baby example illustrating this fact has been described in the Chapter 2: there, we have seen that the
irregularity of the Brownian path requires regularizing the stochastic integrals and this induces possibly
non trivial Jacobians, depending on the regularization scheme.
114
Going back to the previous derivation of the Ward identities but taking this non-
trivial Jacobian into account, it is easy to deduce (following exactly the same steps) that
the following relation,
∂µ J µ (x) = F (x),
is valid inside any correlation functions (with other fields inserted away from point x). That
is: due to the non-triviality of the Jacobian, the current conservation law is transformed
into the new equation of motion above. This is call an anomaly: something which was
expected to be zero by classical consideration turns not to vanish in the statistical theory
because of fluctuations.
7.3 Generating functions

We consider Euclidean field theory with action of the form S[φ] = dD x 12 (∂φ)2 + V (φ)
R
and its generalizations (with more field components, non-trivial background, etc...). In this
Section we introduce standard generating function for field correlation functions. They
are analogous to generating functions of random variables and/or to thermodynamical
functions of statistical physics.
• Generating functions.
QFT or SFT deals with random objects and their statistics. The random variables
are the fields and the field configurations. As for any random variables one may introduce
their generating functions (which almost characterize their distributions). Here it amounts
to introduce a source so that the action now becomes (by definition)
Z
S[φ; j] := S[φ] − dD x j(x)φ(x).
One defines the partition function in presence of the source

Z
Z[j] := [Dφ] e−S[φ;j] . (52)
A convenient normalization is Z[j = 0] = 1 which fixes the normalization of the mea-

sure (but this may hides some properties, say the dependence on the background met-
ric,...). Alternatively one may not normalize the generating function but consider the
ration Z[j]/Z[0]. It may be worth comparing this definition with that of the generating
functions in the case of the free field.
This is a generating function for correlation functions. The latter can be obtained by
differentiation:
1 δ N Z[j]
hφ(x1 ) · · · φ(xN )i = .

Z[0] δj(x1 ) · · · δj(xN ) j=0
As for random variables, one may (preferably) introduce the generating function of
the cumulants which in diagrammatic perturbation theory are going to correspond to
connected correlation functions. It is usually denoted as W [j] and defined by
W [j] = log Z[j], Z[j] = eW [j] , (53)
115
normalized by W [j = 0] = 0 (that is: we use the normalization Z[j = 0] = 1 as otherwise

we would have to consider the ratio Z[j]/Z[0]). It is the generating function of (so-called)
connected correlation functions:
δ N W [j]
hφ(x1 ) · · · φ(xN )ic = .

δj(x1 ) · · · δj(xN ) j=0
We could alternatively write W [j] as a series expansions:

X 1 Z
W [j] = dD x1 · · · dD xN hφ(x1 ) · · · φ(xN )ic j(x1 ) · · · j(xN ).
N!
N ≥0
Of course there is an analogy with thermodynamical potentials: Z[j] is the partition

function, and W [j] is minus the free energy (up to a coefficient proportional to the tem-
perature). Hence (as usual in thermodynamics) we expect W [0] to be proportional to the
space volume, and W [j] − W [0] to be finite if the source are localized in space.
Of course one can write all these formula in Fourrier space. See below.
Before closing this subsection, let us recall the formula for the generating function in
the case of Gaussian free field
Z
1
dD xdD y j(x) G(x, y) j(y) ,

Z[j] = exp
2
with G(x, y) the Green function. Hence W [j] is quadratic for a free Gaussian theory (and
this is characteristic of Gaussian theory).
• The effective action.

One may introduce other functions via Legendre transform as done with thermody-
namic functions. In field theory, a particularly important one is the so-called effective
action (or vertex function) defined as the Legendre transform of W [j]:
Z
δW [j]
Γ[ϕ] := dD x j(x)ϕ(x) − W [j], ϕ(x) = . (54)
δj(x)
The field ϕ(x) is sometimes called the background field. One may of course expand the
effective action in power of ϕ:
X 1 Z
Γ[ϕ] = dD x1 · · · dD xN ϕ(x1 ) · · · ϕ(xN ) Γ(N ) (x1 , · · · , xN ),
N!
N
or equivalently
(N ) δ N Γ[ϕ]
Γ (x1 , · · · , xN ) = .

δϕ(x1 ) · · · δϕ(xN ) ϕ=0

The functions Γ(N ) are called the N -point vertex functions. In perturbation theory they
will be related to one-particle irreducible diagrams (see below).
The effective action contains all the information on the theory, in a synthetic way if
not the most possible compact way.
116
• The effective potential and vacuum expectation values.

As usual with Legendre transform, the inverse formula for reconstructing the function
W [j] from its Legendre transform Γ[ϕ] is:
Z
δΓ[ϕ]
W [j] = dD x j(x)ϕ(x) − Γ[ϕ], j(x) = .
δϕ(x)
δΓ[ϕ]
The relation j(x) = δϕ(x) , or reciprocally ϕ(x) = δW [j]
δj(x) allow to express the source j as a
function of the background field ϕ, or reciprocally the background field as a function of
the source j.
Let ϕvac (x) := hφ(x)i|j=0 be the field expectation value (in absence of external source),
also called the vacuum expectation value. Because hφ(x)i|j=0 = δW [j]
δj(x) |j=0 by definition of
the generating function W [j], we have :
ϕvac (x) = ϕ(x)|j=0 = hφ(x)i|j=0 .

δΓ[ϕ]
From the relation j(x) = δϕ(x) taken at j = 0 we learn that the field expectation at zero
external source is a minimum of the effective action:
δΓ[ϕ]
= 0. (55)
δϕ(x) ϕvac

Hence the name of effective action: the field expectation value is the minimum of the
effective action.
One also defines the effective potential Veff (v) as the effective action evaluated on
constant field configuration (in particular in case of translation invariance). For a constant
field configuration the effective action is expected to be proportional to the space volume
(by analogy with thermodynamic function which are extensive and thus proportional to
the volume of the system). The definition of the effective potential is
Γ[ϕ]|ϕ(x)=v = Vol. Veff (v), (56)
with Vol. the volume of the system. For homogeneous system, the vacuum expectation
value is thus the minimum of the effective potential. This is important –and very useful–
in analyzing physical phenomena, in particular in case of symmetry breaking (because
then the existence of non trivial minimum of the effective potential will be the sign of
symmetry breaking).
In case of non-homogeneous background, the vacuum expectation hφ(x)i is not uniform,
it may be x-dependent. Eq.(55) then yields a series of equations satisfied by ϕvac (x), which
may be thought of as the equations of motion for ϕvac (x). In general, these equations
may be non-local in space because the effective action can be non-local. However, if we
are only looking at slowly varying vacuum configurations, so that ϕvac (x) is a smooth,
slowly varying, function we may (sometimes) do a gradient expansion of the expansion
and approximate it by truncating this expansion to lowest order.
• Generating functions in Fourier space.
117
Of course we can write all these generating function in Fourier space. For instance
for the effective action, we decompose the background R field on its Fourier components via
dD k ik·x
R D −ik·x
ϕ̂(k) = d x e ϕ(x), and reciprocally ϕ(x) = (2π)D e ϕ̂(k). Then the effective
action becomes
X 1 Z dD k1 dD kN
Γ[ϕ] = · · · ϕ̂(−k1 ) · · · ϕ̂(−kN ) Γ̃(N ) (k1 , · · · , kN ),
N! (2π)D (2π)D
N
with N -point functions
N N
dD kj Y ikj ·xj (N )
Z Y
(N )
Γ (x1 , · · · , xN ) = e Γ̃ (k1 , · · · , kN ),
(2π)D
j=1 j=1
If translation invariance holds (which is often the case) then

Γ̃(N ) (k1 , · · · , kN ) = (2π)D δ(k1 + · · · + kN ) Γ̂(N ) (k1 , · · · , kN ).
by momentum conservation (imposing the conservation law k1 + · · · + kN = 0 is a conse-
quence of translation invariance).
Similar formula hold for the other generating functions, say that of connected correla-
tion functions.
• Two-point correlation and vertex functions.
There are of course relations between the vertex functions and the connected corre-
lation functions. We shall see below that all correlation functions can (perturbatively)
reconstructed from the vertex functions, so that the latter contain all the information on
the theory. Let us start from the two point functions. We have the connected 2-point
(2)
function Gc (x, y) and the 2-point vertex function Γ(2) (x, y). By translation invariance,
we can write them in Fourier space as (we have extracted the delta-function ensuring
momentum conservation).
dD k ik·(x−y) (2) dD k ik·(x−y) (2)
Z Z
(2) (2)
Gc (x, y) = e Ĝ c (k), Γ c (x, y) = e Γ̂ (k).
(2π)D (2π)D
(2) (2)
We wrote Gc (k, p) = (2π)D δ(k + p) Ĝc (k), and similarly for Γ, in order to extract the
delta-function ensuring momentum conservation.
We are now going to prove that these two functions are kernels of inverse operators.
In Fourier space this translates into
Γ̂(2) (k) Ĝ(2)
c (k) = 1. (57)
Thus knowing Γ̂(2) (k) is enough to reconstruct the connected two-point function.
δW [j]
Indeed recall that ϕ(x1 ) = δj(x 1)
. Differentiating this relation with respect to ϕ(x2 )
and then taking the limit j = 0 yields:
δ 2 W [j]
Z
δ δW [j] δj(y)
δ(x1 − x2 ) = = dD y
δϕ(x2 ) δj(x1 ) j=0 δϕ(x2 ) δj(x1 )δj(y) j=0

Z
δj(y)
= dD y hφ(x1 )φ(y)ic
δϕ(x2 ) j=0

Z Z
= d y hφ(x1 )φ(y)i Γ (y, x2 ) = dD y G(2)
D c (2) (2)
c (x1 , y) Γ (y, x2 )
118
In the first line we used the chain rule to compute the derivative. In the second and
third line we used the fact that W [j] and Γ[ϕ] are the generating function of the con-
nected and vertex functions, respectively. In Fourier space, this gives the above relation
(2)
Γ̂(2) (k) Ĝc (k) = 1.
7.4 Perturbation theory and Feynman rules

There is not so many ways to evaluate a path integral: if it is not gaussian, the only
available generic methods are either via saddle point approximations, via symmetry ar-
guments, or via perturbation theory. One has to remember that, although generic and
sometimes useful, perturbation theory give, at best, series which are only asymptotic series
(not convergent series) and they miss all non perturbative effects (say exponentially small
but important effect).
We shall exemplify the perturbation 4
R D 4 theory in the case of the φ theory for which the
interaction potential is V (φ) ∝ d x φ (x).
• Perturbative expansion.
To compute perturbatively we split the action in its quadratic part plus the rest:
S[φ] = S0 [φ] + SI [φ].
In the case of φ4 theory, the interacting part is

Z
g
SI [φ] = dD y φ4 (y).
4!
Let h· · · i0 denote the expectation with respect to the measure defined by the unperturbed
action S0 [φ]. We may be interested in computing the partition function and the N -point
functions. The partition function reads
Z
g R D 4
Z = [Dφ] e−S0 [φ]−SI [φ] = Z0 he− 4! d y φ (y) i0 ,
with Z0 = 1 by convention. The partition function Z is also called the vacuum expectation.
The correlation functions are
g
dD y φ4 (y)
R
hφ(x1 ) · · · φ(xN ) e− 4! i0
hφ(x1 ) · · · φ(xN )i = g
− 4!
R
dD y φ4 (y)
he i0
Notice that one has to divide by the partition function in the formula for the N -point
functions in order to preserve the normalization (the expectation value of 1 is 1).
It is convenient to introduce the “un-normalized” expectation values:
g
dy φ4 (y)
R
G (N ) (x1 , · · · , xN ) := hφ(x1 ) · · · φ(xN ) e− 4! i0 ,
so that
hφ(x1 ) · · · φ(xN )i = G (N ) (x1 , · · · , xN )/Z.
119
Figure 2: First order expansion of the partition function and the connected two-point
function.
Expanding in power of g we get the perturbative series:

(N )
X
G (N ) (x1 , · · · , xN ) = (−g)k Gk (x1 , · · · , xN ),
k≥0
Z
(N ) 1 1 k
Gk (x1 , · · · , xN ) = ( ) dD y1 · · · dD yk hφ(x1 ) · · · φ(xN ) φ4 (y1 ) · · · φ4 (yk )i0
k! 4!
Of course we can write similar formula for the perturbative series in Fourier space.
They are evaluated using Wick’s theorem with two-point function G0 (x − y). It is
important to remember that the theory comes equipped with its UV cut-off (a the lattice
spacing or Λ = 1/a the momentum cut-off). The unperturbed two-point function is
dD k eik·(x−y)
Z
G0 (x − y) = .
|k|<Λ (2π)D k 2 + m2
It is convenient to denote its Fourier transform by ∆(k):

1
∆(k) = . (58)
k2 + m2
It is called the “propagator”.
Notice that doing this expansion amounts to exchange (path) integration and series
expansion, a process which is (often) ill-defined. In general we are going to get, at best,
series which are only asymptotic series (not convergent series) missing non perturbative
effects.
• Feynman rules.
The perturbation expansion is going to be coded into diagrams representing the dif-
ferent terms or contractions involved in Wick’s theorem.
Let us look at the first few terms. The first order in the vacuum expectation is
Z
g
Z =1− dD y hφ4 (y)i0 + · · · .
4!
120
To evaluate this integrated expectation value, it is better and simpler to imagine splitting
the four coincident points in four neighbor points at position y1 , · · · y4 . Then we need to
evaluate hφ(y1 )φ(y2 )φ(y3 )φ(y4 )i0 using Wick’s theorem. We get three terms,
hφ(y1 )φ(y2 )i0 hφ(y3 )φ(y4 )i0 + hφ(y1 )φ(y3 )i0 hφ(y2 )φ(y4 )i0 + hφ(y1 )φ(y4 )i0 hφ(y2 )φ(y3 )i0 .
It is better to draw the diagrams with the four points slightly split (because then the
enumeration of those diagrams is simpler). We now remember that the points are actually
coinciding, so that we get three times the same diagrams corresponding to hφ(x)φ(x)i20 .
Hence (by translation invariance)
Z
g g
Z =1− dD y G0 (0)2 + · · · = 1 − Vol. G0 (0)2 + · · · .
8 8
Recall that
dD p
Z
1
G0 (0) = .
|k|<Λ (2π) p + m2
D 2
SD−1 D−2
Notice that G0 (0) ' (2π)D
Λ + ··· at large Λ, with SD the volume of the unit sphere
in dimension D: SD = 2π D/2 /Γ(D/2). Hence it diverges for D > 2.
Let us now look at the two point function. At first order we have
Z
(2) g
G (x1 , x2 ) = G0 (x1 − x2 ) − dD y hφ(x1 )φ(x2 )φ4 (y)i0 + · · · .
4!
Again, each term is evaluated using Wick’s theorem which is done by pairing the points.
To correctly count the number of pairings it is again simpler to slightly split the point y
into y1 , · · · , y4 . They are two type diagrams: connected or disconnected (the disconnected
one are cancelled when dividing by the vacuum expectation value). The number of ways
to get the disconnected diagrams is 3 (the same as above for Z), the number of ways to
obtained the connected diagram is (4 · 3) (use the splitting method to avoid mistake when
counting the number of ways to obtain a given diagram). Hence
Z
(2) g 2 g
G (x1 , x2 ) = G0 (x1 −x2 )− G0 (x1 −x2 )G0 (0) − dD y G0 (x1 −y)G0 (y−x2 )G0 (0)+· · ·
8 2
The factor 81 comes from 4!1 × 3 and 12 comes from 4!1 × (4 · 3). We are only interested in
the connected diagrams as the other ones can be deduced from them. See Figure 7.4 for
the graphical representation of this formula.
We now understand what are the Feynman rules for connected diagrams.
We write them for the φ4 theory, first in position space (but the generalization is clear):
(i) draw all topologically distinct connected diagrams with N external lines and each
internal vertex attached to 4 lines;
(ii) to each line associate a factor G0 (x − x0 ) (where x and x0 can be either an internal or
external vertex);
(iii) to each internal vertex associate a factor (−g);RQ D
(iv) integrate over internal vertices with measure k d yk ;
(v) multiply by the symmetry factor 1/(integer) (this symmetry factor is more easily
121
Figure 3: Second order expansion of the 2 and 4 point connected functions.
computed by splitting virtually the internal points: it combines the number of equivalent
pairings with the factor k!(1/4!)k coming form the expansion).
The elements, line and vertices, of the Feynman graphs are:
x y
= G0 (x − y) ; X = (−g).
The convention is to sum over all internal labels yk with the measure k dD yk .
Q
A way to draw all diagrams may consist in drawing all external points as small circles
with on leg emerging from each of them, all internal vertices with dots at the center of
small four legs crossing (or a different number of legs if the interaction is different from φ4 ),
and connect these seeds of legs in all possible ways. Given a diagram obtained that way,
the symmetry factor is k!(1/4!)k times the number of ways of obtaining it by connection
the emerging legs. It may a good idea to practice a bit.
• Feynman rules in momentum space.
We can write the perturbative expansion in momentum space. Let us look only at the
connected diagrams. Consider for instance the connected two-point functions.
dD k ik·(x1 −x2 ) (2)
Z
(2)
Gc (x1 , x2 ) = e Gc (k),
(2π)D
Recall its perturbative expansion given above. If Fourier space, this reads
g
(k 2 + m2 ) Gc(2) (k) = 1 − ∆(k) G0 (0) + · · ·
2
dD p
Z
g 1 1
= 1− ( 2 2
) D 2 2
+ ···
2 k +m |k|<Λ (2π) p + m
with ∆(k) = 1/(k 2 +m2 ). The important point is that integration on the internal positions
imposes momentum conservation at the vertex.
122
We then get the Feynman rules in the momentum space for the connected diagrams
(again we write them for the φ4 theory):
(i) draw all topologically distinct connected diagrams with N external lines and each in-
ternal vertex attached to 4 lines;
(ii) assign momenta flowing along each line so that the external lines have momenta kj
and momentum is conserved at each internal vertex;
(iii) to each internal vertex associate a factor (−g); RQ D
(iv) integrate over remaining loop momenta with measure D
k d pk /(2π) ;
(v) multiply by the symmetry factor 1/(integer).
See Figure 7.4 for exemples.
Another way to read the Feynman diagram in Fourier space is to write directly the
action is
R Fourier space. Recall that we define the Fourier components of the field via
dD k ik·x
φ(x) = (2π) D e φ̂(k). We can write the free and the interaction parts of the action as
dD k
Z
S0 [Φ] = φ̂(−k)(k 2 + m2 )φ̂(k),
(2π)D
Z D
g d k1 dD k4
SI [Φ] = · · · (2π)D δ(k1 + k2 + k3 + k4 ) φ̂(k1 )φ̂(k2 )φ̂(k3 )φ̂(k4 ).
4! (2π)D (2π)D
From which we read the graphical representation:

k k1 k3
→ ←p = (2π)D δ(k + p) ∆(k) ; k2 X k4 = (−g) (2π)D δ(k1 + k2 + k3 + k4 ).
The convention is to sum over al the internal labels pk with the measure k dD pk /(2π)D .
Q
A point which could be tricky sometimes (it is actually often tricky...) is the symmetry
factor. One can give a formal definition, by look at the order of the symmetry group of
the diagram (based on the fact that indexing the point in the splitting method amounts
to index the lines arriving at a vertex). But this is probably not useful at this level as one
can always go back to the Wick’s theorem expansion to get it (via the splitting method).
7.5 Diagrammatics
• Connected and one-particle irreducible (1-PI) graphs.
It is clear that the generating function of connected graph is the logarithm of the
generating function of all graphs (connected and disconnected). This simply follows from
usual combinatorics encoded into the exponential function. Let us recall the order two
expansion (at order g 2 included) of the 2 and 4 point connected functions. See Figure 7.4.
We see that diagrams of higher order includes sub-diagrams of lower order. This
leads to the following definition of irreducible graphs, also called one-particle irreducible
diagrams and denote 1-PI. By definition, the irreducible graphs are the connected graphs
which do not become disconnected if one of their lines is cut. See Figure 7.5.
All these diagrams have external legs, which are simple multiplicative factors in the
momentum representation. These factors are the propagators ∆(k)−1 with k the momen-
tum of the corresponding leg. Thus we define the truncated irreducible diagrams as the
123
Figure 4: Second order expansion of the 2 and 4 point vertex functions.
irreducible diagrams but with the external legs removed (i.e. truncated). We represent
these truncated diagrams by erasing the external propagators but indicating the incoming
momenta by arrows.
Finally, we define the irreducible N -point functions. Their differ for N = 2 and N ≥ 2:
— For N 6= 2: the irreducible N -point functions is “minus” the sum of the N -point
irreducible diagrams;
— For N = 2: we add to this sum the zeroth order contribution so that the irreducible
2-point functions is the propagator minus the sum of all 2-point irreducible diagrams.
These generating functions are also called vertex functions.
• Two-point functions
It can be proved that the effective action Γ[ϕ] is the generating function of the ir-
reducible N -point functions. See below and the exercise Section. This justifies that we
denote them with the same letter.
The last statement can be checked on the two point function. According to the above
definition, let us decompose the two-point vertex function as
(2)
Γ̂(2) (k) = Γ̂0 (k) − Σ(k), (59)
(2)
with Γ̂0 (k) = ∆−1 (k) the propagator, so that Σ(k) is the sum of irreducible graphs
containing at least one loop which remains once the external legs have been removed.
We now explain the relation between Σ(k) and the connected two-point function
(2)
Ĝc (k). Let us look at their diagram expansions. Each time we cut a link which dis-
connect a diagram of the two point functions, we produce two diagrams with two external
line each, that is we produce two diagrams which contribute to the diagram expansion
of these functions. Hence, cutting all the lines up to the point we reach the irreducible
components produces product of diagrams of Σ(k). As a consequence, the connected two-
point function can be decomposed as a chain of Σ joined by intermediate propagators. See
Figure 7.5.
124
Figure 5: Series expansion of the 2 point connected functions.
In equation, this translates into
Ĝ(2)
c (k) = ∆(k) + ∆(k)Σ(k)∆(k) + ∆(k)Σ(k)∆(k)Σ(k)∆(k) + · · ·
1
= −1
= Γ̂(2) (k)−1 .
∆ (k) − Σ(k)
(2)
Thanks to the relation Γ̂(2) (k) Ĝc (k) = 1, this shows that the diagrammatic definition of
the 2-point irreducible function coincide with the vertex function defined via the effective
action. We will see below that this statement hold for all N -point functions.
• Effective action and 1-PI diagrams
First let us observe that any connected diagram can be decomposed uniquely as a tree
whose vertices are its 1-PI components (some of the irreducible components may reduced to
single vertex). The simplest proof is graphical. As a consequence, all connected correlation
functions can written as tree expansions with propagator the two-point connection function
Γ(2) and as vertices the irreducible 1-PI vertex functions Γ(N ) . See Figure 7.5.
One can prove that the effective action is the generating function for 1-PI diagrams.
See the exercise Section. All correlation functions, connected or not, can but recursively
reconstructed from the 1-PI diagrams. These diagrams contain all information on the
theory, and all information on the theory is contained in the effective action or equivalently
in the irreducible functions.
See the exercise Section for a proof of the equivalence between the generating function
of 1-PI diagrams and the effective action.
We stop here for the (slightly boring) diagrammatic gymnastic.
• The Euler characteristic of a graph and two useful formula.

The give here two tools which are often useful in analysing perturbative graphs: one
relates the topology of the graph to its order of the perturbative expansion, the other is
useful representation of elements of Feynman loop integrants.
— Let G be a graph, and let V its number of vertices, E its number of edges and F
its number of faces. Then the Euler characteristic is defined as χ(G) = V − E + F . One
has χ(G) = 2 − g where g is the so-called genus g ≤ 0). This number codes the topology
of the surface on which the graph can be drawn: g is the (minimal) number of handle
of that surface, g = 0 corresponds to the sphere, g = 1 to the torus, etc. One has to
pay attention that in this formula one has to count the external vertices —so that V the
number of internal vertices (equal to the order in the perturbative expansion) plus the
number of external points— as well as the external face of the graph —so that F is the
125
Figure 6: The 3 and 4 point connected functions in terms of the vertex functions.
number of loops plus one. For instance, for a tree the number of faces is 1. It is a good
exercise to check this on examples.
— Two useful formulas to compute Feynman integral:
Z ∞
1 2 2
= du e−u(p +m ) ,
p2 + m2 0
n
δ(x1 + · · · xn − 1)
Z
1 1 Y
= dxj .
a1 · · · an (n − 1)! [xj >0] [x1 a1 + · · · xn an ]n
j=1
This leads to the so-called parametric representation of the Feynman graphs.
7.6 One-loop effective action and potential

Here, we compute the effective action at one-loop (i.e. first order in perturbation theory)
2
for the φ4 theory with action S[φ] = dD x 12 (∂φ)2 + m2 φ2 + 4!g φ4 . This amounts to a
R
Gaussian integral around the solution of the classical equations of motion. We are going
to prove that, to leading order in the loop counting parameter ~ (see below), the effective
action is:
~
Tr log S 00 [ϕ] + O(~2 ),

Γ[ϕ] = S[ϕ] + (60)
2
where S 00 [ϕ] is the Hessian of the action (i.e. the matrix operator of double derivatives)
evaluated at the field configuration ϕ. Thus, the correction to the effective action at one-
loop is given by the logarithm of the Hessian of the classical action (in a way very similar
to the WKB approximation).
• The one-loop effective action
126
The strategy, which follows closely the definition, is to first compute the generating
function W [j] as function of the external source and then do the Legendre transformation.
The partition function with source is
Z
1
Z[j] = [Dφ] e− ~ S[φ;j] , S[φ; j] := S[φ] − (j, φ).
We have introduce the notation (j, φ) := dD x j(x)φ(x), and the coefficient ~ which
R
serves as loop counting parameter. It justifies the saddle point approximation (one-loop)
as ~ → 0. The classical equation of motion (for φc ) is,
δS[φ]
j(x) = .
δφ(x) φc
In the case of φ4 this reads

g
−∇2 φc + m2 φc + φc = j.
3!
We expand around this classical solution (saddle point): φ = φc + ~φ̂. We only keep the
term quadratic in φ̂. Then (here the dots refer to higher order in ~)
~2
Z
S[φc + φ̂; j] = S[φc ; j] + dD xdD y φ̂(x) S 00 [φc ](x, y) φ̂(y) + · · ·
2
The quadratic term S 00 is the Hessian of S. It defines a kernel acting on functions. At this
order (one-loop) the integration over φ̂ is Gaussian. ,
−1/2 − 1 S[φc ;j]
Z[j] = Z[0] × Det[S 00 [φc ]]

e ~ .
Since W [j] = ~ log(Z[j]/Z[0]) (here we put back the factor ~ in the definition), we thus
have
~
W [j] = −S[φc ] + (j, φc ) − Tr log S 00 [φc ] + O(~2 ).

2
To compute the effective action, we have to implement the Legendre transform. The
background field ϕ is determined by
δW [j] δ ~ δ
Tr S 00 [φc ] + O(~2 ).

ϕ(x) := = (j, φc ) − S[φc ] −
δj(x) δj(x) 2 δj(x)
Recall that here φc is viewed as function of j. Thanks to the equations of motion, j(x) =
δS[φ]
δφ(x) φc , it is easy to verify that
δ
(j, φc ) − S[φc ] = φc .
δj(x)
The second term is of order ~. Hence, at this one-loop order, the background field is the
classical field:
δW [j]
ϕ(x) := = φc (x) + O(~).
δj(x)
127
Then we compute the effective action by Legendre transform: Γ[ϕ] := hj, ϕi − W [j]. We
use the fact that S[ϕ] = S[φc ]+(ϕ−φc , S 0 [φc ])+O(~) and j = S 0 [φc ], to get the announced
result: Γ[ϕ] = Γ1−loop [ϕ] + O(~2 ) with
~
Γ1−loop [ϕ] = S[ϕ] + Tr log S 00 [ϕ] .

2
One can compute diagrammatically this action for φ4 . The Hessian S 00 [ϕ] is an operator
acting on function via
g
f (x) → S 00 [ϕ] · f (x) := −∇2 f (x) + m2 f (x) + φ2 (x) f (x).

2
00
We aim at computing Tr log S [ϕ] . We factorize the free part so that
1 1 g
Γ1−loop [ϕ] = Tr log − ∇2 + m2 + Tr log 1 + (−∇2 + m2 )−1 ϕ2

2 2 2
We can then expand (perturbatively) the logarithm in power series of g. Then (up to the
trivial pre-factor, trivial in absence of background)
∞
X g k (−)k+1 (k)
Γ1−loop [ϕ] = Γ1−loop [ϕ],
2k+1 k
k=1
(k) k
with Γ1−loop [ϕ] = Tr log (−∇2 + m2 )ϕ2 , or else

Z
(k)
Γ1−loop [ϕ] = dD x1 · · · dD xk ϕ2 (x1 )G0 (x1 −x2 )ϕ2 (x2 ) · · · G0 (xk−1 −xk )ϕ2 (xk )G0 (xk −x1 ).
with G0 the Green function (the inverse) of (−∇2 + m2 ). These are closed polygonal
diagrams with k vertices. They also are all the irreducible one-loop diagrams with k
vertex and two truncated external lines at each vertex. These diagrams can be computed
for a constant background field which is enough to determine the effective potential.
Notice that for a general potential V (φ) we would just have to replace m2 + g2 ϕ2 by
V 00 (ϕ) in the previous formula.
• One-loop effective action for the Φ4 theory
The effective action can be computed for a constant background field ϕ. This determine
the effective potential. This Section can be viewed as part of the main text or as an
exercise (in that case the reader is advised not to read the answer before trying to compute
independently the one-loop effective potential).
Recall that the effective action is Γ1−loop [ϕ] = S[ϕ] + ~2 Tr log S 00 [ϕ] with S 00 [ϕ] =

−∇2 + V 00 (ϕ). We will specify the potential to the φ4 theory in a little while. For constant
back-ground field, the effective action is proportional to the volume of the space and the
proportionality coefficient is by definition the effective potential:
eff
Γ1−loop [ϕ] = Vol. V1−loop (ϕ).
128
For constant back-ground field the trace of the differential operator can be computed by
going to Fourier space. Thus the one-loop correction to the effective potential is
dD k
Z
~
eff
log k 2 + V 00 (ϕ) .

V1−loop = V (ϕ) + D
2 (2π)
This is clearly UV-divergent. We have to introduce a cut-off Λ and write:
dD k
Z h
eff ~ 2 00
2i
V1−loop = V (ϕ) + log k + V (ϕ) /Λ . (61)
2 |k|<Λ (2π)D
We have introduced the cut-off scale Λ2 in the logarithm to make it dimensionless.

Let now fix the dimension D = 4. Recall that SD /(2π)D = 2/(4π)2 in D = 4. We can
then explicitly do the integral over the momentum to obtain 5
eff ~ h 2 00 1 00 2
Λ2 + V 00 (ϕ) −2
i
V1−loop (ϕ) = V (ϕ) + const.Λ + Λ V (ϕ) + [V (ϕ)] log + O(Λ ) .
32π 2 2 Λ2
Here const.Λ is a Λ-dependent diverging constant.
m2
Let us now go back to the φ4 theory to write explicitly V (ϕ) = 20 ϕ2 + g4!0 ϕ4 and
V 00 (ϕ) = m20 + g20 ϕ2 . (We introduce the notation m0 and g0 for consistency with the
following discussion on renormalization). Let us introduce an arbitrary scale µ2 in or-
00 00 2
der to extract the logarithmic divergence by writing log V Λ(ϕ)
2 = log V µ(ϕ)
2 − log Λ µ2
.
Gathering all the terms, the one-loop effective potential can be written as
eff 1 1 ~ 00
2 V 00 (ϕ)
V1−loop (ϕ) = AΛ ϕ2 + BΛ ϕ4 + V (ϕ) log[ ] + O(~2 ), (62)
2! 4! (8π)2 µ2
m20 2 g0 4
with V (ϕ) = 2 ϕ + 4! ϕ and
~g0 Λ2 m20 Λ2
m2 := AΛ = m20 + − log( ) + O((~g0 )2 ),
2 (4π)2 (4π)2 µ2
3 Λ2
g := BΛ = g0 − ~g02 log( ) + O(g0 (~g0 )2 )
2(4π)2 µ2
All the diverging terms are at most of degree 4 in ϕ. They can thus be absorbed in a
renormalization of the mass m and the coupling constant g, see following Chapter. It is an
interesting exercise to analyse this potential, its minima, as a function of the parameters
m0 , g0 and Λ.
7.7 The O(N ) vector models

We here present basic properties of the O(N ) vector model at large N . It provides an
instance of models which can be solved via a saddle point approximation —a method
which is a valuable tool in statistical field theory. This approximative method, which can
5
After a change of variable to x = k2 , the integral to reduce to
R
xdx log(x + a). This is evaluated by
R 2 dx 2
integration by part and using xx+a = x2 + a2 log(x + a) − ax.
129
be applied when there exist an appropriate small parameter —in a way similar to the
small noise problem in stochastic diffusion equations— becomes exact when N → ∞.
The action of the O(N ) vector model is
Z h1
~ 2 + 1 m20 Φ
~ 2 + g (Φ
i
~
S[Φ] = dD x (∇Φ) ~ 2 )2 ,
2 2 4!
where Φ~ = (Φ1 , · · · , ΦN ) is a N -component scalar field.

There are various ways to solve this model. We can for instance use diagrammatic
expansion techniques and derive exact equations for the propagator which are valid in the
large N limit. See the Exercise Section for detail computations concerning this diagram-
matic approach. Another way consists in using a saddle point approximation and this is
the method we shall pursue.
• The Hubbard-Stratonovich transform.
This is a technique, based on Gaussian integrals, which allows to disentangle the in-
teraction at the prize of introducing an extra field. It is similar to the trick we used when
representing the lattice Ising partition function in terms of a scalar field. It amounts to
represent the contribution of the interaction terms dD x (Φ ~ 2 )2 by a Gaussian integral.
R
So, let σ(x) be a scalar field. Then, by a standard Gaussian integral we have:
Z √ g ~2
g R D ~ 2 )2 − dD x N σ 2 +i 12
R
− 4! d x (Φ Φ σ
e = [Dσ] e 2 .
As a consequence, instead of dealing with a theory field with a single field Φ ~ with a
quadratic interaction, we can deal with a theory involving two fields with action
Z hN r
~ D 2 1 ~ 2 1 2 ~2 g ~2 i
S[Φ; σ] = d x σ + (∇Φ) + m0 Φ + i Φ σ .
2 2 2 12
~ although there is an
The advantage we gain is that now the action is quadratic in Φ,
~ and σ.
interaction term between Φ
• Saddle point approximation.
Since the action S[Φ;~ σ] is quadratic in Φ,
~ we can easily integrate over this field. The
q
~ is 1 dD x Φ~ − ∆x + m2 + i g σ(x) Φ, ~ with ∆x = ∇2x the
R
part of the action involving Φ 2 0 3
Laplacian. By integrating over Φ ~ we thus get the following contribution to the Boltzmann
weights: r
h g i−N/2
−N
R D
d x σ(x)2 2
[Dσ] e × Det − ∆x + m0 + i σ(x) ,
2

3
or equivalently, we get the following effective action for σ:
Z r
N D 2
h
2 g i
Seff [σ] = d x σ(x) + log Det − ∆x + m0 + i σ(x) .
2 3
The noticeable point is that this action is proportional to N so that the σ-path integral
can be evaluated via a saddle point approximation in the large N limit.
130
It is clear that the saddle point is for a uniform configuration σ(x) = σ0c . For such
configuration the effective reads (by using the formula log Det M = Tr(log M ) and repre-
senting the action of the Laplacian in Fourier space):
dD k
Z r
N
2
2 2 g
Seff [σ] = Vol. σ0 + log k + m 0 + i σ0 .
2 (2π)D 3
The saddle is thus for σ0c with
dD k
r Z
g 1
σ0c +i = 0,
3 (2π)D k 2 + m2R
q
with m2R = m20 + i g3 σ0c . This equation is analysed in the Exercise Section, and we
recommend to do the corresponding exercise.
Let us just comment, that inserting the saddle point value σ0c in the action S[Φ, ~ σ]
~
yields an effective action for the field Φ,
Z
1
~
Seff [Φ] = dD x (∇Φ) ~ 2 + 1 m2 Φ
~2 ,

R
2 2
q
with again m2R = m20 + i g3 σ0c , so that mR is the renormalized mass. The critical theory
is for mR = 0 (because the physical correlation length is ξ = 1/mR ). The above saddle
point equation fixes the value of the bare critical mass mc0 as a function of the cut-off
and the interaction strength g. By looking at the behaviour of the correlation length as
m0 approach its critical value mc0 one get information on the correlation length critical
exponent ν, which here takes the value ν ∞ = 1.
A refined analysis of this model, including the diagrammatic expansion, the analysis of
the critical theory, as well as a study of fluctuations around the saddle point approximation
and 1/N corrections, is given in a dedicated exercise (that we recommend!). See the
Exercise Section.
7.8 Exercises
• Exercise 7.1: The effective potential and magnetization distribution functions
The aim of this exercise to probability distribution function of the total magnetization
is governed by the effective potential — and this gives a simple interpretation of the
effective potential.
Let Mφ := dD x φ(x) be the total magnetization. It is suppose to be typically exten-
R
sive so let mφ be the spatial mean magnetization, mφ = Vol.−1 Mφ .

(i) Find the expression of the generating function of the total magnetization, E[ezMφ ], in
terms of the generating function W [·] of connected correlation functions.
Recall that if the source J(x) is uniform, i.e. J(x) = j independent of x, then W [J] is
extensive in the volume: W [J(x) = j] = Vol. w(j).
(ii) Let P (m)dm be the probability density for the random variable mφ . Show that at
large volume, we have
P (m) ' e−Vol. Veff (m) ,
131
with Veff (m) the effective potential, defined as the Legendre transformed of w(j).
This has important consequence, in particular the most probable mean magnetization is
at the minimum of the effective potential, and phase transition occurs when this minimum
changes value.
• Exercise 7.2: Two-point correlation and vertex functions.
Prove that the two-point connected correlation function and the two-point vertex func-
tion are inverse one from the other, that is:
Γ̂(2) (k) Ĝ(2)

c (k) = 1,
as mentioned in the text.

• Exercise 7.2bis: Ward identities for the stress-tensor.
The aim of this exercise is to derive the Ward identities associated to translation
symmetry. This will allows us to make contact with the stress tensor.
We consider a scalar field φ in D-dimensional Euclidean flat space with action
Z
1
S[φ] = dD x (∇φ)2 + V (φ) .

2
Translations act on the field as φ(x) → φ(x − a) for any vector a. The infinitesimal
transformation is φ(x) → φ(x) − aµ (∂µ φ)(x).
(i) let us consider an infinitesimal transformation φ(x) → φ(x) − µ (x)(∂µ φ)(x) with the
space dependent vector fields (x).
Prove that the variation of the action is (assuming that the boundary terms do not con-
tribute) Z Z
δS[φ] = − dD x (∂ µ σ )(x) Tµσ (x) = dD x σ (x) (∂ µ Tµσ )(x),
with Tµσ (x) the so-called stress-tensor (gµσ is the Euclidean flat metric):
1
(∇φ)2 + V (φ) .

Tµσ (x) = ∂µ φ ∂σ φ − gµσ
2
(ii) Prove that the stress tensor is conserved, that is: ∂µ Tµν (x) = 0 inside any correlation
functions away from operator insertions.
(iii) Prove the following Ward identities (here we use the notation ∂jν = ∂/∂yjν ):
X
h(∂µ Tµν )(x) φ(y1 ) · · · φ(yp )i = δ(x − yj ) ∂jν hφ(y1 ) · · · φ(yp )i,
j
in presence of scalar field insertion of the form φ(y1 ) · · · φ(yp ).

(iv) Do the same construction but for rotation symmetry.
• Exercise 7.3: Generating functions Z[J], W[J] and ?[?] for a ?3-theory in D = 0.
[... See the exercise booklet...]
• Exercise 7.4: Effective action and one-particle irreducible diagrams.
132
The aim of this exercise is to prove the equality between the effective action and the
generating function of 1PI diagrams. To simplify matter, we consider a ‘field’ made of N
(N 1) components φj , j = 1, · · · , N . We view φj as random variables.
Let us define a ‘partition function’ Z [J] by
Z
−1
Y dφj
Z [J] = [Dφ] e− Γ[φ]−(J,φ)
, with [Dφ] = [ √ ]
j
2π
with J a source (J, φ) = Jj φj , and Γ[φ] an action which we define via its (formal) series
expansion (summation over repeated indices is implicit):
1 (2) X 1 (n)
Γ[φ] = Γjk φj φk − Γ φj1 · · · φjn .
2 n! j1 ···jn
n≥3
We shall compute this partition function in two different ways: via a saddle point
approximation or via a perturbation expansion.
(i) Justify that this integral can be evaluating the integral via a saddle-point when → 0.
Prove that
1
log Z = W [J] 1 + O() ,

where W [J] is the Legendre transform of the action Γ: W [J] = (J, φ∗ ) − Γ(φ∗ ) with φ∗
∂Γ
determined via ∂φ j (φ∗ ) = Jj .
Hint : Do the computation formally which amounts to assume that the integral converges
and that there is only one saddle point.
Let us now compute Z [J] in perturbation theory. Let us decompose the action
as the sum of its Gaussian part plus the rest that we view as the interaction part:
(2)
Γ[φ] = 21 Γjk φj φk − Γ̂[φ].
(ii) Write Z
1 (2) j k −1 Γ̂[φ] −1 (J,φ)
Z [J] = [Dφ] e− 2 Γjk φ φ
e e .
We view J/ as source, and we aim at computing the connected correlation function using
Feynman diagrams perturbative expansion.
(n)
Show that the propagator is Gjk with G = (Γ(2) )−1 and the vertices are −1 Γj1 ···jn with
n ≥ 3.
(iii) Compute the two-, three- and four-point connected correlations G(n) , n = 1, 2, 3, at
the level tree, defined by
1 ···jn
∂n
Gj(n) = log Z [J] .

∂Jj1 · · · ∂Jjn tree
Show that they are of order −1 . Draw their diagrammatic representations (in terms of
propagators and vertices) and compare those with the representations of the connected
correlation functions in terms of 1PI diagrams.
133
(iv) Prove that, when → 0, the leading contribution comes only from the planar tree
diagrams and that all these diagrams scale like 1/. That is:
1
log Z [J] = planar tree diagrams + O() .

Hint : Recall that, for a connected graph drawn on a surface of genus g (i.e. with g
handles, g > 0), one has V − E + L + 1 = 2 − g with V its number of vertices, E its number
of edges and L its numbers of loops (this is called the Euler characteristics). Then, argue
that each Feynman graph contributing to the N point connected functions is weighted by
(symbolically) ( G)E (−−1 Γ(n) )Vint (−1 J)N with Vint + N total number of vertices.
(v) By inverting the Legendre transform, deduce the claim that the effective action is the
generating function of 1-PI diagrams.
• Exercise 7.5: Computation of the one-loop effective potential.
Prove the formula for the one-loop effective potential of the φ4 -theory given in the
text. Namely
eff 1 1 ~ 00 V 00 (ϕ)
V1−loop (ϕ) = AΛ ϕ2 + BΛ ϕ4 + V (ϕ) log[ ],
2! 4! (8π)2 µ2
with
~g0 Λ2 m20 Λ2
AΛ = m20 + − log( ) + O((~g0 )2 ),
2 (4π)2 (4π)2 µ2
3 Λ2
BΛ = g0 − ~g02 log( ) + O(g0 (~g0 )2 )
2(4π)2 µ2
with µ2 an arbitrary scale that we introduced by dimensional analysis.

[...Analyse this potential and conclude...]
• Exercise 7.6: Computation of one-loop Feynman diagrams.
• Exercise 7.7: The O(N ) vector model in D = 3.
134
8 Conformal field theory: basics
This Chapter deals with field theory invariant under conformal transformations, called
conformal field theory. These are important because they are fixed point of the renor-
malization group. They describe statistical field theory with infinite correlation length.
They have numerous applications. They will also provide simple examples of (interacting)
statistical field theory which will illustrate basics, but fundamental and generic, properties
of statistical field theories.
8.1 The group of conformal transformations

Conformal transformations act locally as dilatations and rotations, but with dilatation
coefficients and rotation angles which may vary from point to point. We first look at the
group of conformal transformations and its structure, which is different in dimension two
and higher.
• Conformal transformations in dimension D
A conformal transformation on a metric manifold, —that is on a manifold equipped
with a metric allowing to measure locally distances and angles— is a transformation which
locally modifies the metric by a dilatation factor possibly depending on the position the
manifold.
To simplify matter, let us consider the Euclidean space in dimension D, equipped with
the flat Euclidean metric ds2 = dx21 + · · · + dx2D . The metric tensor gµν , is defined by
ds2 = gµν dxµ dxν where the summation over the repeated indices is implicit, gµν = δµν
with δµν Kronecker symbol.
Under a diffeomorphism x → y = f (x), the metric tensor transforms as gµν → ĝµν :=
∂xσ ∂xρ
∂y ∂y ν gσρ . A conformal transformation is thus a transformation x → y = f (x), locally
µ
defined on Rd , such that the two metrics ĝ and g are locally proportional with a propor-
tionality coefficient which may vary from point to point, namely
∂xσ ∂xρ
ĝµν (x) = ( )( ) gσρ (x) = e2ϕ(x) gµν (x),
∂y µ ∂y ν
where ϕ is the local dilatation factor, called the conformal factor. Clearly, such transfor-
mation preserves angles, locally.
Our aim is now to identify the set of conformal transformations in RD , equipped with
the Euclidean metric. Clearly, conformal transformations form a group, because they can
be composed and the composition of two conformal transformations is again a conformal
transformation. We look at infinitesimal transformations xµ → xµ + ξ µ (x)+· · · generated
by a vector field ξ µ (x). To first order in , the condition for conformal invariance reads
∂µ ξν + ∂ν ξµ = 2(δϕ) δµν , (63)
135
with e2(δϕ) the infinitesimal dilatation coefficient and ξµ := ξ ν δνµ . The dilatation factor
is not independent from the generating vector field ξ µ . Indeed, contracting the indices
yields D(δϕ) = (∂µ ξ µ ).
Identifying all possible infinitesimal conformal transformations amounts to find the
general solution of this condition. This can be done via simple manipulations (e.g tak-
ing derivatives, contracting indices, etc). See the exercise Section for a more detailed
description. The outcome is that all infinitesimal conformal transformations, xµ →
xµ + ξ µ (x) + · · · , in dimension D > 2, are generated by vector fields ξ µ of the form:
ξν (x) = aν + kxν + θνσ xσ + [2(b · x)xν − (x · x)bν ], (64)
where aν , k, bµ and θνσ = −θσν are constant and parametrize the infinitesimal transfor-
mations. The infinitesimal dilatation factor is (δϕ) = k + 2(b · x) . It is uniform if b = 0
but not otherwise.
Each term in the above equation for ξ µ possesses a simple interpretation:
— the first corresponds to translation: xν → xν + aν ,
— the second to dilatation: xν → xν + kxν ,
— the third to rotation: xν → xν + θνσ xσ ,
— the last to so-called special conformal transformation: xν → xν + [2(b · x)xν − (x · x)bν ].
The dimension of the group of conformal transformation is thus D + 1 + D + D(D−1)2 =
1
2 (D + 1)(D + 2). One can show –the exercise Section– that this group is isomorphic to
so(D + 1, 1).
• Conformal transformations in D = 2
In dimension D = 2, the special conformal transformations are the so-called homo-
graphic transformations, also called the Mobius transformations. In complex coordinates
z = x + iy, they read
az + b
z → f (z) = , a, b, c, d ∈ C.
cz + d
These are the only holomorphic bijections of the complex plane with a point at infinity
added (it is important to add the point at infinity; with this point added the complex
plane is isomorphic to the 2D sphere). By composition, they form a group isomorphic to
the group of linear transformation of C2 with unit determinant whose elements are:
a b
∈ SL(2, C)/{±1} = PSL(2, C).
c d
Looking back to the previous analysis of the conformal transformations, we see that in
dimension D = 2 the dilatation factor as to be a harmonic function: ∆(∂ · ξ) = 0. Hence
locally (∂ · ξ) = ∂z v(z) + ∂z̄ v̄(z̄) with v(z) holomorphic. This corresponds to infinitesimal
transformations
z → z + v(z)
with v(z) locally holomorphic. These transformations are in general defined only locally,
not globally on the sphere, because there is no holomorphic vector on the sphere except
vn (z) = z n+1 with n = 0, ±1.
136
As a consequence, conformal transformations in 2D are locally holomorphic transfor-

mations:
z → w = f (z)
Let z1 = z0 + δz1 and z2 = z0 + δz2 two neighbour points of z0 . The two small vectors
δz1 and δz2 joining z0 to z1 or z2 are transformed into two vectors δw1 and δw2 joining
w0 = f (z0 ) to w1 = f (z1 ) or w2 = f (z2 ). To first order, we have:
δw1 = f 0 (z0 ) δz1 , δw2 = f 0 (z0 ) δz2 .
The angle between the vectors δw1 and δw2 is therefore identical to that between the
vectors δz1 and δz2 . Each of these vectors has been rotated, by an angle equal to the
argument of f 0 (z0 ), and has been dilated proportionally to the modulus of f 0 (z0 ). Confor-
mal transformations act locally as rotations and dilatations but with dilatation factors or
rotation angles which may vary from point to point.
Let us give a few examples of (simple) conformal transformations:
— The transformation
z−i
z→w=
z+i
is holomorphic, without singularity, on the upper half-plane H = {z ∈ C, Imz > 0}, and
the image of the upper half-plane is the unit disk centred at the origin, D = {w ∈ C, |w| <
1}.
More generally, Riemann’s theorem asserts that any planar domain with the topology
the unit disk, i.e. any connexe, simply connected, open set of the complex plane, different
from C, is in conformal bijection with the unit disk, and hence also in bijection with the
upper half-plane.
— The transformation
z → w = β log z, w → z = ew/β
maps the complex z-plane (with the point at the origin and at infinity removed) to the
the w-cylinder with radius β. The circle of radius eτ centred at the origin in the complex
plane if mapped to the circle loop of the cylinder at azimutal altitude τ /β.
• A classical application of conformal transformations.
Conformal transformations have many applications in numerous problems of classical
physics governed by the two-dimensional Poisson-Laplace equation,
∆z φ(z, z̄) = 0 avec ∆z = ∂x2 + ∂y2 = 4 ∂z ∂z̄ .
Any solution of this equation is called a harmonic function. Since the laplacian can be
factorized as ∆z = 4 ∂z ∂z̄ , any harmonic functions can be decomposed into the sum of a
holomorphic function and an anti-holomorphic function,
φ(z, z̄) = ϕ(z) + ϕ̄(z̄).
The conformal invariance of the Poisson-Laplace equation follows form this decomposition.
Indeed, if φ(z, z̄) is a solution, so it φ(f (z), f (z)). Alternatively, let w = f (z). The
137
laplacian in the new coordinate w can be written in a simple way in terms of the laplacian
in the coordinate z as: ∆w = 4∂w ∂w̄ = |f 0 (z)|2 ∆z .
This equation obviously applies to electrostatic problems where φ is the electrostatic
potential in the vacuum. These are generally formulated in the form of Dirichlet problems
for which one looks for the value of the electrostatic potential in a planar domain D –
which is the exterior of perfectly conducting material body – knowing the value φ|D of the
potential at the boundary of this domain.
It is also applied in fluid mechanics to the description of two-dimensional movements
of incompressible and irresistible fluids. Indeed, if the fluid vorticity ω = ∇ ∧ u vanishes,
the velocity field u is (locally) the gradient of some function φ, i.e. u = −∇φ, and the
incompressibility condition ∇ · u = 0 imposes ∆φ = 0. The boundary conditions require
that the gradient of φ is tangent to the surface of the domain so that the flow is tangent
to this surface. They are called Neumann boundary conditions, depend on the shape of
the domain in which the fluid moves.
8.2 Conformal invariance in field theory

The aim of this section is to understand the basic simple echoes of conformal invariance
in statistical field theory. One of the take-home message is that conformal invariance fixes
the structures and the values of the one, two and three point field correlation functions.
Before starting let us first recall that to any symmetry correspond conserved currents.
For conformal symmetries, there are three conserved currents respectively associated to
translation, dilatation and special conformal transformation. They are of the form Jξµ =
Tµν ξ ν with ξ ν any of the vector fields generating conformal transformations, i.e. ξν (x) =
aν + kxν + θνσ xσ + [2(b · x)xν − (x · x)bν ] and T µν the stress-tensor. (Recall that the
stress-tensor is the generator of diffeomorphisms). The conservation laws are:
∂ µ (Tµν ξ ν ) = 0.
This is equivalent to the conservation for Tµν and for its traceless-ness:
∂ µ Tµν = 0, Tµµ = 0.
Thus a conformal field theory is characterized by a trace-less, conserved, stress-tensor.

These relations are valid away from field insertions (cf. the discussion about symmetries
and Ward identities).
• Global conformal invariance and the 2-point functions

We first describe simple consequences of conformal invariance on the two and three
point functions, using a simple step-by-step construction.
Let us first look at 2 point functions. Pick two field φ1 and φ2 located at x1 and x2
and consider their correlation functions (in flat space, with the Euclidean metric)
G(2) (x1 , x2 ) := hΦ1 (x1 )Φ2 (x2 )i.
By definition a conformal field theory is such that the correlation functions of local fields
are invariant under the global conformal transformation.
138
To simplify matter, let assume that the fields are scalar fields (no spin). We shall
generalize later, especially in dimension 2. We shall prove that the two point functions of
‘scaling fields’ have to be of the following form:
const.
hΦ1 (x1 )Φ2 (x2 )i = δh ;h .
|x1 − x2 |2h1 1 2
The numbers h1 and h2 are the so-called scaling dimensions of the fields Φ1 and Φ2 respec-
tively. What scaling fields are and how scaling dimensions are defined is explained in a
few lines. In particular, the 2-point functions in a conformal field theory are non vanishing
only if the two fields have identical scaling dimensions. We can (usually) normalize the
fields such that the above constant is 1.
Global translation is easy to implement: it says that hΦ1 (x1 )Φ2 (x2 )i = hΦ1 (x1 +
a)Φ2 (x2 + a)i for any shift a. To write the condition of rotation invariance, we have to
specify how the fields transform under rotation (say, whether they are scalar, vectors, etc).
To do that, we have to associate a representation of the rotation group SO(D) to each
fields (or more precisely to each multiplet of fields), so that the fields under rotation as
Φ(x) → (ρ(R) · Φ)(R−1 x), with R → ρ(R) the representation of rotation group associated
to the field Φ. Rotation invariance is then the statement that
−1 −1
hΦ1 (x1 ) Φ2 (x2 )i = hΦR R
1 (R x1 ) Φ2 (R x2 )i, ΦR
j = ρj (R) · Φj ,
for any rotation R ∈ SO(D). A scalar field transform trivially under rotation so that
Φ = ΦR , by definition. Then, translation and rotation invariance impose that the two
point functions of scalar fields depends only in the distance between the two points,
G(2) (x1 , x2 ) = F2 (r), r2 = |x1 − x2 |2 .
Now to use symmetry under dilatation, we have to specify how fields transform under
dilatation. Again this depends on the nature of the field –that is: how it transform under
the conformal group. For scalar field, this is specified by a number, called the field scaling
dimension. Namely, by definition a field of scaling dimension h transform under dilatation
x → y = λ x as
Φ(x) → Φ̂(y) = λ−h Φ(x = y/λ).
Dilatation invariance is then the statement that
hΦ1 (y1 )Φ2 (y2 )i = λ−h1 −h2 hΦ1 (y1 /λ)Φ2 (y2 /λ)i.
By taking derivative with respect to λ and integrating the resulting differential equation,
it is easy to verify that this implies that
G(2) (x1 , x2 ) = const. r−(h1 +h2 ) .
Notice that the above transform law tells that the scalar field of scaling dimension h
transform as Φ(x) → Φ̂(y) = Φ(y) − (δΦ)(y) with (δΦ)(y) = [h + y ν ∂ν ]Φ(y) for an
infinitesimal dilatation x → y = x + x.
Next we should impose invariance under global special conformal transformation.
These are infinitesimal diffeomorphism transformation xµ → y µ = xµ + ξ µ (x), for which
139
distances (defined via the metric) are scaled by a factor δϕ = D(∂.ξ). So, we impose that
the field transforms as above but with a position dependent scaling factor, that
−h
Φ(x) → Φ̂(y) = eϕ(x(y)) Φ(x(y)), xµ (y) = y µ − ξ µ (y) + · · · .

with ϕ = δϕ. For infinitesimal transformation this corresponds to:
Φ(x) → Φ̂(y) = Φ(y) − [h(δϕ)(y) + ξ µ (y)∂µ ]Φ(y).
Of course this reproduces the above transformation for translation, rotation or dilatation.
For special conformal transformation, ξν (x) = 2(b · x)xν − (x · x)bν and δφ = (b · x) this
gives the field variation
δΦ(x) = 2h(b · x) + [2(b · x)xν − (x · x)bν ]∂ν Φ(x).

Such fields are called ‘scalar conformal fields’ (or pseudo-primary fields, or scaling fields)
and h is their scaling dimensions. Not all fields are scaling fields, for instance if Φ is a
scaling fields its derivatives are not.
The condition for invariance of the two point function then under special conformal
transformation then reads
hΦ1 (y1 ) Φ2 (y2 )i = e−h1 ϕ(x(y1 )) e−h2 ϕ(x(y2 )) hΦ(x(y1 )) Φ(x(y2 ))i,
This yields an extra condition on G(2) . Since we already know that G(2) (x1 , x2 ) =
const. r−(h1 +h2 ) , it gives a constraint on the scaling dimensions h1 and h2 which is that
G(2) vanishes unless h1 = h2 . Thus we get
G(2) (x1 , x2 ) = const. r−(h1 +h2 ) δh1 ;h2 .
See the exercise Section for a detailed proof of all the technical steps. Notice that global
dilatation invariance only is not enough we need the special conformal transformation to
prove that two scaling dimensions have to be identical.
• Global conformal invariance and 3-point functions
Similar computations can be done for 3-point functions:
G(3) (x1 , x2 , x3 ) = hΦ1 (x1 )Φ2 (x2 )Φ3 (x3 )i.
We look for the correlation of scalar conformal fields Φj , j = 1, 2, 3, of respective scaling di-
mensions hj . These are fields which, by definition, transform as follows under infinitesimal
conformal transformations x → y = x + ξ(x):
Φ(x) → Φ̂(y) = Φ(y) − ε [h (δϕ)(y) + ξ µ (y)∂µ ]Φ(y),
with h the scaling dimension fo the field Φ and (δϕ)(y) = D−1 (∂µ ξ µ )(y) the infinitesimal
conformal factor. Invariance of the 3-point functions then demands that
hj (δϕ)(xj ) + ξµ (xj )∂jµ G3 (x1 , x2 , x3 ) = 0.

X
j=1,2,3
140
It is a simple matter of computation to verify that this determines the three point functions
as (See the exercise Section)
c123
G3 (x1 , x2 , x3 ) = ,
|x1 − x2 |h1 +h2 −h3 |x2 − x3 |h2 +h3 −h1 |x1 − x3 |h1 +h3 −h2
with c123 a constant (which depends on the structure of the theory). No condition is
imposed on the scaling dimension h1 , h2 and h3 . Whether the three point function is
non vanishing and what is the value of the above constant depend on the specific theory
one is considering. Note that once the fields Φj have been normalized by their two point
functions the above constant is specified (there is no freedom to redefine it).
Notice that these results for the 2- and 3-point functions are valid in any dimensions.
• Invariance
In previous Chapters, we discussed symmetries and associated Ward identities, but we
restricted ourselves to cases in which the symmetry transformations were preserving the
geometrical data involved in the specification of the statistical field theory. We here go
back to this point but for symmetry transformation involving the geometrical data.
We aim at comparing statistical field theory over two spaces M0 and M, respectively
equipped with a metric g 0 and g. As data, we have series of field correlation functions of
the form
hΦ1 (y1 ) · · · iM,g , hΦ1 (x1 ) · · · iM0 ,g0 .
with y1 , · · · points in M and x1 , · · · points in M0 . These series of objects cannot be
compared because they are not defined over the same set. So let us give ourselves a
mapping from M0 to M, say x → y = f (x), or reciprocally the inverse mapping y → x
from M to M0 . As discussed above, via this mapping we get a metric ĝ 0 on M by
transporting that from M0 :
∂xσ ∂xρ
dŝ2 = ĝµν
0
dy µ dy ν = gσρ
0
( )( )dy µ dy ν .
∂y µ ∂y ν
We can then use it to push forward all correlation functions hΦ1 (x1 ) · · · iM0 ,g0 into cor-
relation functions over M, by viewing the points x in M0 as function of the point y in
M. Along such transformation the field Φ are transformed into Φ → Φ̂ —we will make
explicit this transformation below in the meantime you may think about how vectors,
forms, etc, transform on diffeomorphism. Thus from the correlation functions in M0 we
get correlation functions
hΦ̂1 (x(y1 )) · · · iM0 ,g0 , yj ∈ M,
which are correlation functions over M. In these correlation functions, distances are
measured with the metric ĝ 0 defined above, by construction. Hence, by pushing forward
the correlation functions from M0 to M, we got two series of correlations functions over
M differing by the metrics used to measure distances (one is the original one g, the other
is the induced one ĝ 0 ).
Conformal invariance is the statement that these correlation functions are identical if
the two metrics g and ĝ 0 are conformally equivalent, that is ĝ 0 (y) = eϕ(y) g(y):
hΦ1 (y1 ) · · · iM,g = hΦ̂1 (x(y1 )) · · · iM0 ,g0 , if ĝ 0 = e2ϕ g.
141
We have to specify how the fields transform. As discussed above, for scalar conformal
h
fields of scaling dimension h, Φ̂1 (x(y)) = eϕ(x(y)) Φ(x(y)) so that conformal invariance

reads h
hΦ1 (y1 ) · · · iM,g = h eϕ(x(y1 )) Φ1 (x(y1 )) · · · iM0 ,g0 , if ĝ 0 = e2ϕ g.

This can alternatively, but formally, be written as the following transformation rule:
|dy|hg φ̂h (y) = |dx|hg0 φh (x),
where |dy|g (resp. |dx|g0 ) is the distance measure with the metric g (resp. g 0 ).
This means that, if conformal invariance holds, we can transport correlation functions
from M0 equipped with the metric g 0 to M equipped with the metric g, provided that the
metric ĝ 0 and g are conformally equivalent.
The important special case we consider above is when M and M0 are both the D-
dimensional Euclidean space and g and g 0 are both the Euclidean flat metric.
8.3 Fields and operator product expansions

We now introduce an important concept in field theory (which turns out to be especially
powerful in two dimensional conformal field theory). Let us look at two operators at
nearby positions x and y. Say x = R + r and y = R − r with r R. When viewed from a
point far away from x and y, this product operator looks as defined at medium point r or
equivalent at one of the two points, say y. Thus it is ‘natural’ to expect that this product
can be expanded in all the other local operators. That it is, we expect —or assume— that
we can write the expansion:
X
l
Φj (x) Φk (y) = Cjk (x, y) Φl (y),
l
where the sum is over all set of local operators. Such expansion is called the operator
product expansion (OPE).
The existence of the operator product expansion is one of the basics assumption of
statistical field theory (which may either view as a tautological statement or as axiom,
impossible to prove but part of the definition of what a statistical field theory could be...).
The group of conformal transformations acts on the space of local fields of a conformal
field theory (by definition and/or constitutive properties of a conformal field theory). We
can hence group the fields into conformal multiplets, forming representation and of the
group of conformal transformations, and we can organize the space of fields as sum of
those representations (assuming complete reducibility).
By a choice of basis, we order the basis field Φl according to their scaling dimension
(under global dilatation). Scaling invariance then fixes the way the function Cjk l (x, y)
l (x, y) ∝ |x − y|hl −hj −hk (if all the fields involved are spinless). Hence
scales: Cjk
X cljk
Φj (x) Φk (y) = Φl (y) + · · · ,
|x − y|hj +hk −hl
l
The numbers cljk are called the OPE coefficients. Of course they coincide with the coeffi-
cient of the three point functions. Here the dots refers to contribution to spin full fields
142
(say derivative of spinless fields) that we do not write explicitly because their writing is a
bit cumbersome (but without any conceptual difficulties).
This OPE can be checked on the three point functions. Let G(3) (x1 , x2 , x3 ) be the
expectations of three scalar conformal fields, Φ1 , Φ2 and Φ3 of respective scaling dimension
h1 , h2 and h3 , at three different positions. Its explicit expression was just derived above.
Let us then fuse x1 → x2 in this formula and observe the result. We get
1 c123
G(3) (x1 , x2 , x3 ) 'x1 →x2 h +h −h
+ ···
|x1 − x2 | 1 2 3 |x2 − x3 |2h3
1
'x1 →x2 c123 hΦ3 (x2 )Φ3 (x3 )i + · · ·
|x1 − x2 |h1 +h2 −h3
Hence, the fusion of Φ1 (x1 ) and Φ2 (x2 ) effectively generate the field Φ3 if the OPE co-
efficient c123 does not vanishes. This coefficient is then identified with the three point
normalization coefficient. Recall that conformal fields with different scaling dimension
have vanishing two point function so that only the contribution from Φ3 in the OPE
survives when computing the limit of three point function when x1 → x2 .
The OPE can be used to write series expansion for correlations. As explained above,
the two- and three- point functions of a CFT are fixed by conformal symmetry. Let us
thus look at a four-point function
hΦ1 (x1 )Φ2 (x2 )Φ3 (x3 )Φ4 (x4 )i.
We can image evaluating this correlation function by series expansion by fusing using the
OPE the fields Φ1 and Φ2 in one hand and Φ3 and Φ4 on the other hand. We then are
reduced to compute two-point functions of the fused fields. The result is
X
k
C12 (x1 , x2 ) δk;l |x2 − x4 |−hk −hl C34
l
(x3 , x4 ).
k,l
We can alternatively choose to fuse the fields Φ2 and Φ3 in one hand and Φ1 and Φ4 on
the other hand. The result is then
X
k
C23 (x2 , x3 ) δk;l |x3 − x4 |−hk −hl C14
l
(x1 , x4 ).
k,l
This is two expressions have to be equal. This yields a relation fulfilled by the OPE
coefficients. In many respect, it is similar to the condition for associativity of a product
algebra. This relation is called crossing symmetry.
We will encounter examples of OPE in the following sections.
8.4 Massless gaussian free field in 2D

Here we describe the statistical field theory of a massless free bosonic field in two dimen-
sion. This is the simplest example of 2D conformal field theory. We will explicit the
transformation properties of the basic fields of the theory, the structure of both the Ward
identities and the operator product expansion.
143
We are going to use complex coordinate z = x + iy, z̄ = x − iy and ∂z = 12 (∂x + i∂y ),

∂z̄ = 21 (∂x − i∂y ). The Euclidean metric is ds2 = dx2 + dy 2 = dzdz̄ so that its components
in complex coordinates are g z z̄ = g z̄z = 1/2.
• Action and Green function
We aim at describing the statistical field theory of a free scalar bosonic field φ defined
by the action S
Z Z
g g
S[φ] = d2 x (∂µ φ)(∂ µ φ) = dzdz̄ (∂z φ)(∂z̄ φ),
8π 8π
Here we made a slight abuse of notation: d2 x is the Lebesgue measure on R2 (d2 x is the
infinitesimal flat area dA) which we denote in complex coordinate as dzdz̄. The (classical)
equation of motion is ∆z φ = 0, with ∆z the 2D Laplacian, which is conformally invariant
in 2D. It is in particular simple to verify that the action is invariant under dilatation, if
we assume that φ is a scalar field, invariant under dilatation (recall that a scalar field a
scaling dimension (D − 2)/2 in dimension D so that its dimension vanish in D = 2).
The parameter g is dimensionless. It depends on the chosen normalization of the field
φ. (To simplify the notation we shall later choose g = 1.)
The action S is actually invariant under conformal transformation z → w = w(z), if
we assume that the field transforms as
φ(z, z̄) → φ̂(w, w̄) = φ(z(w), z̄(w̄)).
Indeed, we have
Z Z
g g
S[φ̂] = dwdw̄ (∂w φ̂)(∂w̄ φ̂) = dzdz̄ |∂z w(z)|2 (∂w φ̂)(∂w̄ φ̂)
8π 8π
Z
g
= dzdz̄ (∂z φ)(∂z̄ φ) = S[φ].
8π
where we used, on one hand, that the Jacobian to change variable from w to z is |∂w/∂z|2
(recall that d2 x is the infinitesimal area), and one the other hand, the chain rule for
derivatives to write ∂z φ = (∂w/∂z)∂w φ̂ and similarly for the complex conjugate.
Recall that the Green function of the 2D Laplacian is such that:
1
G(z, w) = − log |z − w|, up to a constant.
2π
It satisfies (−∆z )G(z, w) = δ (2) (z − w). The Green function is only determined up to
a constant because the Laplacian possesses a constant zero mode. The Laplacian is not
invertible and we cannot define/compute the gaussian integral. We thus have to deal with
this constant zero mode. A way to it is to define the theory on a finite domain and then
let the size of this domain increases infinitely. The simplest is to consider the theory on a
disc DR of radius R and impose Dirichlet condition at the boundary, that is φ = 0 at the
boundary. By integration by part, we can alternatively write the action as (with ∆x the
Laplacian) Z
g
S[φ] = − d2 x φ(x) (∆x φ)(x),
8π DR
144
because there is no contribution from the boundary terms thanks to the Dirichlet boundary
condition.
With the Dirichlet boundary condition, the Laplacian is invertible (there is then no
zero mode because an harmonic fonction which vanish at the boundary is zero). The
Green function is then
1 R|z − w|
GR (z, w) = − log .
2π |z w̄ − R2 |
Note that it has the same short distance expansion as the naive Green function, GR (z, w) '
1
− 2π log |z −w|+reg(z, w), and it satisfies (−∆z )GR (z, w) = δ (2) (z −w) with the boundary
condition GR (z, R) = 0. In the large R limit we have
1
GR (z, w) 'R→∞ − log(|z − w|/R).
2π
Since the theory is Gaussian, the expectation values of product of fields (or vertex
operators) are computable using its two point function. Taking the normalisation pre-
factor g/2π into account, the field two-point function reads
4π
hφ(z, z̄)φ(w, w̄)iDR = GR (z, w),
g
and higher order multi-point correlation functions can be computed using Wick’s theorem.
As discussed in previous Chapter, their generating function can (of course) be computed
exactly
− 2π d2 xd2 y J(x)GR (x,y)J(y)
R 2 R
he− d x J(x)φ(x) iDR = e g .
The factor R is going to be irrelevant as long as we consider correlation functions invariant
under translation of the field φ. That is the large volume limit R → ∞ (IR limit) exists
for all correlation functions invariant under translation of the field (and this invariance
is going to be related to Rsome u(1) symmetry). In the above generating function this
amounts to demand that d2 x J(x) = 0. We shall denote these correlation functions as
h· · · iC .
To simply notation, we shall choose the normalisation g = 1 now, and in the following.
The two point function is then normalized such that hφ(z, z̄)φ(w, w̄)iDR ' − log |z −
2
w| + reg(z, w) and
hφ(z, z̄)φ(w, w̄)iDR ' − log(|z − w|2 /R2 ) + · · ·
at large radius.
• u(1) current.
The theory possesses two conserved currents Jµ := ∂µ φ and Jˆµ := µν ∂ν φ (one is
topological, the other is the Noether current associated to a u(1) symmetry): ∂µ Jµ = 0
and ∂µ Jˆµ = 0. In complex coordinates, this translates into two currents, J and J,
¯ defined
by:
J = i∂z φ with ∂z̄ J = 0,

J¯ = −i∂z̄ φ with ∂z J = 0.
145
Thus, J = J(z) is holomorphic: it only depends on the z-coordinate, while J¯ = J(z̄) ¯

is anti-holomorphic, it only depends on the z̄ coordinate. These conservation laws are
equivalent to the equation of motion ∂z ∂z̄ φ = 0.
As usual in the statistical field theory, these conservation laws valid only when inserted
in correlation functions provided all other fields are located away from the current location.
For instance
∂z̄ hJ(z) Vα1 (v1 , v̄1 ) · · · iDR = 0,
with Vα1 (v1 , v̄1 ) some operator (constructed using the field φ) with v1 6= z. This equation
can be checked explicitly by computing using Wick’s theorem these correlation functions
(one the operator Vα1 (v1 , v̄1 ) has been specified) and using the fact the Green function
is a harmonic function. This equation is similar to those we discussed when looking at
symmetries and associated Ward identities.
We can compute explicitly all the current correlation functions. Since the currents are
invariant under a translation of φ, these correlation functions have a large volume limit
R → ∞. The current two point function reads
1
hJ(z1 )J(z2 )iC = −∂z1 ∂z2 G(z1 , z2 ) = ,
(z1 − z2 )2
as a result of the Wick’s contraction between the two currents. This is (of course) com-
patible with the fact that the chiral/holomorphic current has dimension one (recall that
J = i∂φ where φ has scaling dimension zero and ∂ scaling dimension one). All other
multipoint correlation functions can be computed using the Wick’s theorem.
We can as well use the Wick’s theorem to find the operator product expansion (OPE)
between two currents. Indeed, inside any correlation functions
1
J(z1 )J(z2 ) = + : J(z1 )J(z2 ) :,
(z1 − z2 )2
where, again, the double dots : · · · : means normal ordering (forbidding self-contraction
of the field inside the double dots). This is proved by imaging how the Wick’s theorem
would apply inside a correlation function (via Wick’s contractions). Since : J(z1 )J(z2 ) :=
− : ∂z1 φ(z1 )∂z2 φ(z2 ) :, we can expand this formula in Taylor series and write the current-
current OPE
1 1
J(z1 )J(z2 ) = + J 2 (z2 ) + (z1 − z2 )∂z2 J 2 (z2 ) + · · · ,
(z1 − z2 )2 2
with J 2 (z) = − : (∂z φ)2 (z) : by definition of the operator we named J 2 . This computation
illustrate how operator product expansions arise in statistical field theory (and it proves its
existence in this peculiar example). We may notice that it involves composite operators,
such as (∂φ)2 , made of regularized version of product of operators.
Because φ is a scalar, the current J = i∂z φ transforms like a holomorphic one form:
under a conformal transformation z → w = w(z), it transforms as
ˆ
J(z) → J(w) = [z 0 (w)] J(z(w)).
146
because φ̂(w) = φ(z(w)) and hence ∂w φ̂(w) = (∂z/∂w) ∂φ(z(w)), as a direct consequence
of the chain rule for derivatives. This may also be written in a more geometrical way as
ˆ
dw J(w) = dz J(z), a formula which encodes the fact the current is a holomorphic one
form.
• The stress-tensor and its OPE
The classical stress tensor Tµν , deduced form the action, is 2π Tµν = ∂µ φ∂ν φ −
1 2
2 gµν (∂φ) . By construction it is conserved ∂µ Tµν = 0 (because of the equation of motion)
and traceless g µν Tµν = 0 (because g µν gµν = 2 in dimension two). The traceless-ness of
the the stress-tensor ensures the conservation law of the current Dµ = Tµν xν associated
to dilatation: ∂µ Dµ = 0. Actually the conservation law and the traceless-ness of the
stress-tensor ensures the conservation law of all the Noether currents Dµξ := Tµν ξ ν asso-
ciated to conformal transformations xµ → xµ + ξµ , because the vector fields ξµ satisfy
∂µ ξν + ∂ν ξµ = (δφ)δµν .
In complex coordinates z and z̄, we have (absorbing the factors of 2π):
1 1
Tzz = − (∂z φ)2 , Tz̄ z̄ = − (∂z̄ φ)2 , Tz z̄ = 0.
2 2
The conservation law ∂µ Tµν = 0 then becomes
∂z̄ Tzz = 0, ∂z Tz̄ z̄ = 0,
which are of course consequences of the equation of motion ∂z ∂z̄ φ = 0.

Since product of fields at coincident points are ill-defined, these have to be regularized,
or renormalized, to make sense in statistical field theory. This can be done by normal
ordering. That is, in the statistical field theory we define the stress tensor for a massless
bosonic field by
1 1
Tzz = − : (∂z φ)2 :, Tz̄ z̄ = − : (∂z̄ φ)2 : .
2 2
Since normal ordering amounts to cancel all the self contraction in the Wick theorem, this
is equivalent to regularized the product (∂z φ)(∂z φ) by slightly splitting the two field apart
at nearby position z and w → z and subtracting the Wick contraction between these two
field. Hence
1 h
0 1 i
T (z) := Tzz (z) = lim − ∂ z φ(z)∂ z φ(z ) − ,
2 z 0 →z (z − z 0 )2
1 h
0 1 i
T̄ (z̄) := Tz̄ z̄ (z̄) = lim − ∂ z̄ φ(z̄)∂z̄ φ(z̄ ) − .
2 z 0 →z (z̄ − z̄ 0 )2
It is worth comparing this definition with the structure of the OPE of the currents. In
particular we see that the stress tensor is the first regular term in the OPE of two currents.
Correlation functions of products of the stress tensor components can be computed (say
using Wick’s theorem). For instance their two point function is
c/2
hT (z1 )T (z2 )i = , with c = 1.
(z1 − z2 )4
147
This is of course compatible with the fact that T is a chiral (holomorphic) field of scaling
dimension 2 (because it is the product of two derivatives of the scalar field φ).
We leave as an exercice to compute the OPE between two insertion of stress tensor at
nearby points (c = 1):
c/2 2 1
T (z1 )T (z2 ) = 4
+ 2
+ ∂z T (z2 ] + · · · .
(z1 − z2 ) (z1 − z2 ) (z1 − z2 )
This is an important formula for the general theory of conformal field theory. See below
(or in a more advance course).
The stress tensor has scaling dimension h = 2. Because φ is a scalar and T = − 12 (∂φ)2 ,
it naively resembles a holomorphic 2-form. However, the regularization we used to define
it modifies the way it transforms: under a conformal transformation z → w = w(z), the
holomorphic component of the stress-tensor transforms as (c = 1):
c
T (z) → T̂ (w) = [z 0 (w)]2 T (z(w)) + S(z; w),
12
where S(z; w) is the so-called Schwarzian derivative
z 000 (w) 3 z 00 (w) 2
S(z; w) = − .
z 0 (w) 2 z 0 (w)
This is quite general fact in statistical field theory: dealing with singularities, which are
unavoidable due to the roughness of the field configuration, requires regularization, or
renormalization, which then modifies the way objects transform (because the rules of
transformation do not preserve the rules of regularization).
Let us do the computation. Recall that the definition of T which is the w-coordinate
reads
1 h
ˆ J(wˆ 0) − 1 i
T̂ (w) = lim J(w) .
2 w0 →w (w − w0 )2
The important point is that we have used the same subtraction as before but in the w-
coordinate, i.e. we substract −1/(w − w0 )2 . Now, recall that φ is scalar so that φ̂(w) =
ˆ
φ(z(w)) and hence J(w) = z 0 (w) J(z(w)). Thus
ˆ J(w
J(w) ˆ 0 ) = [z(w)z 0 (w0 )] J(z(w))J(z(w0 )).
Recall now the OPE of the currents found above only involves J 2 at leading orders and
thus can be written as J(z1 )J(z2 ) = (z1 − z2 )−2 + 2T (z2 ) + · · · . Hence
1ˆ 0 0 0
ˆ 0 ) = [z 0 (w)]2 T (z(w)) + [z (w)z (w )] + reg · · ·
J(w)J(w
2 2(z(w) − z(w0 ))2
Equivalently,
1 h [z 0 (w)z 0 (w0 )] 1 i
T̂ (w) = [z 0 (w)]2 T (z(w)) + lim − .
2 w0 →w (z(w) − z(w0 ))2 (w − w0 )2
The extra term has a finite limit equals to the Schwarzian derivative (as it is easy to verify
00 000
by Taylor expansion z(w + δw) − z(w) = (δw)z 0 (w) + 21 (δw)2 z (w) + 16 (δw)3 z (w) + · · · ).
148
This is an example of an anomaly, or extra contribution, arising as echoes of renor-

malization, which is due to the fact the regularization breaks the symmetry. Here, the
definition of the stress-tensor needs the regularization consisting in subtracting 1/(z −z 0 )2 .
But the quantities subtracted is not invariant under the transformation symmetry. That is:
the subtraction in the image space 1/(w − w0 )2 is not equal to the image of the subtraction
1/(z − z 0 )2 . There is this extra contribution because the conformal transformation has dis-
torted the quantities we subtracted. This reflects —or originates from— the fact that the
fields are not ultra-local on the lattice but codes configuration on a small neighbourhood
around their position.
• Vertex operators and their correlation functions
Vertex operators are operators of the form Vα (z, z̄) = eiαφ(z,z̄) . Since they are defined
from products of field localized at identical positions they need to be regularized (this is
an UV regularization, not to be confused with the previous, armless, IR regularization).
By Gaussian integration (or by definition of the generating function of Gaussian vari-
ables) we have:
Y 1X
h eiαj φ(zj ,z̄j ) i = exp −

αj αk G(zj , zk ) .
2
j j,k
P
To be able to take the IR limt R → ∞ we have to impose that j αj = 0 (which is charge
neutral condition).
We have a divergence from the diagonal terms in the sum: 21 j αj2 G(zj , zj ). This is
P
infinite. So we have to regularized it by defining the theory on a lattice (and then taking
the limit in which mesh of the lattice shrinks to zero). On the (say a square) lattice the
Green function at coincident point is finite and equal to log(1/a2 ). This yields a diverging
2
factor j aαj which can be absorb in the definition of the operator/observable.
Q
Thus we define the (re-normalized) vertex operator as the UV limit
2
Vα (z, z̄) = lim a−α eiαφ(z,z̄) |reg. .
a→0
where the label ‘reg.’ means that an UV regularisation, with UV cut-off a, has been
implicit used to make sense of this operator and to compensate the divergence of the
Green function at short distances.
Naively, since φ is a scalar field, eiαφ should also be a scalar field. But this naive guess
forget about the renormalization procedure which will modify the transformation laws of
the vertex operators under conformal transformation. One has to take into account the
2
renormalization procedure which amounts to multiply by a−α . Recall the definition of the
renormalized operator above. This definition is formulated in the z-plane. In the image
w-plane the definition reads,
2
V̂α (w, w̄) = lim â−α eiαφ(z(w),z̄(w)) |reg. ,
â→0
where â in the UV lattice cut-off in the image w-plane. However, this short distance cut-
off is not the same the initial cut-off a in the z-plane because such transformations dilate
distances. For an operator located at point z, the ratio of the distances is a/â = |z 0 (w)|.
As a consequence,
2
V̂α (w, w̄) = |z 0 (w)|α Vα (z(w), z̄(w̄)),
149
2
The ‘anomalous’ factor |z 0 (w)|α are again an echo of the renormalization procedure does
not transform covariantly under the conformal transformation. From there we read the
scaling dimension of the vertex operator Vα is α2 .
This can be alternatively (more rigorously) by using a regularization directly defined in
the continuum (not calling the lattice formulation of the model) which amounts to smear
the field over a small neighbourhood. See the exercice Section. P
The correlation functions of the vertex operators are (recall that j αj = 0):
Y 1X
h Vαj (zj , z̄j )i = exp − αj αk G(zj , zk )
2
j j6=k
Y
= |zj − zk |2αj αk
j<k
In particular we have the formula for the two point functions:

2
hVα (z, z̄)V−α (w, w̄)i = |z − w|−2α .
Again we read that the scaling dimension of Vα is hα = α2 . Remark that this (so called
‘anomalous’) scaling dimension (which would naively be zero because φ is dimensionless)
is an echo of the renormalization procedure used in the definition of the operator.
From these formulas we can read the operator product expansion (OPE) of vertex
operators. Indeed, let us single out two vertex operators Vα1 (z1 , z̄1 ) and Vα2 (z2 , z̄2 )
in the above N -point correlation functions. Since normal ordering amounts to sup-
press all self contractions in the Wick’s theorem, we learn that insertion of the prod-
uct Vα1 (z1 , z̄1 ) Vα2 (z2 , z̄2 ) is equivalent to the insertion of the normal order product |z1 −
z2 |2α1 α2 : eiα1 φ(z1 ,z̄1 )+iα2 φ(z2 ,z̄2 ) :. Hence, by Taylor expansion we have the operator prod-
uct expansion:
Vα1 (z, z̄) Vα2 (w, w̄) = |z − w|2α1 α2 : eiα1 φ(z,z̄)+iα2 φ(w,w̄) :
= |z − w|2α1 α2 Vα1 +α2 (w, w̄) + · · ·

We localized one of the field at the origin (without loss of generalities by translation
invariance). This OPE holds true in any correlation function.
In particular for α1 = −α2 = α:
2
Vα (z, z̄) V−α (0, 0) = |z|−2α : eiα(φ(z,z̄)−φ(0,0)) :
2
= |z|−2α 1 + iα(z∂z φ(0, 0) + z̄∂z̄ φ(0, 0)) + · · ·

¯ =0
We leave as an exercice to compute the higher order term (but remember that ∂ ∂φ
in any correlation function).
For later use, let us compute the OPE between the vertex operators and the u(1)-
current or the stress-tensor. For instance, we have
α
J(z) Vα (w) ' Vα (w) + · · · ,
z−w
150
a formula which can be derived from the Wick’s theorem or from the z-derivative at ε = 0
of the OPE between the two vertex operators V (z, z̄) and Vα (0). Similarly, one has
α2 /2 1
T (z) Vα (w) ' 2
Vα (w) + ∂Vα (w) + · · · ,
(z − w) z−w
1
which again follow from Wick’s theorem and T (z) = 2 : J 2 (z) :.
8.5 2D conformal field theory

Here we extract the basics principles and structures of the 2D massless boson field theory
which remain valid for generic/arbitrary conformal field theory. These structures emerge
from the Ward identities associated to the conformal transformations.
In 2D the conservation laws, ∂ µ Tµν = 0 and Tµµ = 0, arising form conformal invariance
read in complex in complexe coordinates
Tz z̄ = 0, ∂z̄ Tzz = 0, ∂z Tz̄ z̄ = 0,
These relations are valid away from field insertions.

• Ward identities and primary field OPEs
Recall the Ward identities (or its generalized form with product of operators) for
translation, rotation and dilatation
symmetries.µ Recall also that (scalar) conformal field
transformations read δξ Φ(y) = h (∂ · ξ)(y) + ξ (y)∂µ Φ(y) or equivalently as Φh (z) →
|z 0 (w)|2h Φh (w). Write them in complex conformal coordinates. This leads to the integral
form of the (chiral) ward identities:
I
dz(z)hT (z)Φ(y)i = hδ Φ(y)i.
This is equivalent to the following OPE between primary fields and the stress-tensor:
h 1
T (z)Φh (w) = 2
Φh (w) + ∂w Φh (w) + · · ·
(z − w) z−w
Give a few consequences of these relations, e.g. for multipoint correlation functions with
a single stress-tensor.
• The stress-tensor, its OPE and the Virasoro algebra
Explain the stress-tensor OPE:
c/2 2 1
T (z1 )T (z2 ) = 4
+ 2
+ ∂z2 T (z2 ] + · · · .
(z1 − z2 ) (z1 − z2 ) (z1 − z2 )
Here c is the so-called central charge (and c 6= 1 in general, c = 1 only for free massless
boson).
Explain the relation between this OPE and the transformation rules of the stress-tensor
under conformal transformations:
c
T (z) → T̂ (w) = [z 0 (w)]2 T (z(w)) + S(z; w),
12
151
or its infinitesimal version.

Elements of consequences: determine all correlation functions of the stress tensor (on
the sphere/plane) using its analytic and pole structure. For instance:
c/2
hT (z1 )T (z2 )i = ,
(z1 − z2 )4
hT (z1 )T (z2 )T (z3 )i = ....
Idem for multipoint correlation functions T and Φ insertions.

• The Casimir effect (again)
[See the exercise Section...]
8.6 Operator formalism in 2D

• Bosonic free field
• Virasoro representations and field spaces
• Null vectors and differential equations
[.... It is probably better to reserve this last section to a more advanced course.....]
8.7 Exercises
• Exercise 8.1: Conformal mappings in 2D.
(i) Verify that the map z → w = z−i z+i is holomorphic map from the upper half plane
H = {z ∈ C, Imz > 0} to the unit disc D = {w ∈ C, |w| < 1} centred at the origin 0.
(ii) Similarly verify that the map w → z = ew/β is holomorphic map from the cylinder
with radius β to the complex z-plane with the origin and the point at infinity removed.
• Exercise 8.2: The group of conformal transformations.
The aim of this exercise is to fill the missing steps in determining all infinitesimal
conformal transformations in the flat Euclidean space RD . Recall that an infinitesimal
transformation xµ → xµ + ε ξ µ (x) + · · · is conformal if ∂µ ξν + ∂ν ξµ = 2(δϕ) δµν with
D(δϕ) = (∂µ ξ µ ).
(i) Take derivatives of the previous equation to deduce that D ∆ξν = (2 − D)∂ν (∂ · ξ),
with ∆ the Euclidean Laplacian.
(ii) Take further derivatives, either w.r.t ∂ν or w.r.t. ∂µ , to get two new equations: (D −
1)∆(∂ · ξ) = 0, and 2(2 − D) ∂µ ∂ν (∂ · ξ) = D ∆(∂µ ξν + ∂ν ξµ ).
(iii) Deduce that (2−D)∂µ ∂ν (∂ ·ξ) = 0, and hence that, in dimension D > 2, the conformal
factor δϕ(x) is linear in x.
Let us write δϕ(x) = k + bν xν with k and bν integration constantes. We thus have
∂µ ξν + ∂ν ξµ = 2(k + bσ xσ ) δµν .
152
A way to determine ξ consists in getting information on the difference ∂µ ξν − ∂ν ξµ .

(iv) By taking derivates of the previous equation w.r.t ∂σ and permuting the indices,
deduce that ∂ν (∂σ ξµ − ∂µ ξσ ) = 2(bσ δµν − bµ δνσ ), and hence, by integration, that
∂σ ξµ − ∂µ ξσ = 2(bσ xµ − bµ xσ ) + 2θµσ ,
where θσµ = −θµσ are new integration constants.

(v) Integrate the last equations to prove that
1
ξν (x) = aν + kxν + θνσ xσ + [(b · x)xν − (x · x)bν ],
2
where aν are new, but last, integration constants.
(vi) Find the explicit formula for all finite –not infinitesimal– conformal transformations
in dimension D.
(vii) Verify that the Lie algebra of the group of conformal transformation in dimension D
is isomorphic to so(D + 1, 1).
• Exercise 8.3: The two- and three-point conformal correlation functions.
The aim of this exercise is to fill the missing steps in determining the two and three
point function of conformal fields in conformal field theory. Let G(2) (x1 , x2 ) = hΦ1 (x1 )Φ2 (x2 )i
be the two point function of to scalar conformal fields of scaling dimension h1 and h2 re-
spectively.
(i) Prove that translation and rotation invariance implies that G(2) is a function of the
distance r = |x1 − x2 | only.
(ii) Prove that dilatation invariance of the 2-point function demands that
[h1 + x1 · ∂1 + h2 + x2 · ∂2 ]G2 (x1 , x2 ) = 0.
Deduce that G2 (x1 , x2 ) = const. r−(h1 +h2 ) .

(iii) Prove that invariance under special conformal transformations (also called inversions)
implies that
X 1
hj (b · xj ) + [(b · xj )xνj − (xj · xj )bν ]∂xνj G2 (x1 , x2 ) = 0.

2
j=1,2
Deduce that G(2) (x1 , x2 ) vanishes unless h1 = h2 .

Let us now look at the three point functions of scalar conformal fields. Let G(3) (x1 , x2 , x3 ) =
hΦ1 (x1 )Φ2 (x2 )Φ3 (x3 )i, be their correlation functions.
(iv) Prove that invariance under infinitesimal conformal transformations demands that
X
hj D−1 (∂.ξ)(xj ) + ξ µ (xj )∂xµj G(3) (x1 , x2 , x3 ) = 0,

j=1,2,3
for any conformal vector ξ µ (x). See previous exercise.

(v) Integrate this set of differential equations to determine the explicit expression of
G(3) (x1 , x2 , x3 ) up to constant.
153
• Exercise 8.4: Diff S1 and its central extension.

The aim of this exercise is to study the Lie algebra Diff S1 of vector fields in the
circle and its central extension the Virasoro algebra. Let z = eiθ coordinate on the
unit circle. A diffeomorphism is on application θ → f (θ) from S1 onto S1 . Using the
coordinate z, we can write it as z → f (z) so that it is, at least locally, identified with a
holomorphic map (again locally holomorphic). They act on functions φ(z) by composition:
φ(z) → (f · φ)(z) = φ(f −1 (z)). For an infinitesimal transformation, f (z) = z + v(z) + · · ·
avec 1, the transformed function is
(f · φ)(z) = φ(z) + δv φ(z) + · · · , with δv φ(z) = −v(z) ∂z φ(z).
(i) Take v(z) = z n+1 , with n integer. Verify that δv φ(z) = `n φ(z) with `n ≡ −z n+1 ∂z .
Show t
`n , `m = (n − m) `n+m .
This Lie algebra is called the Witt algebra.
(ii) Let us consider the (central) extension of the Witt algebra, generated by the `n and
the central element c, with the following commutation relations
c
`n , `m = (n − m)`n+m + (n3 − n)δn+m;0 ,

[c, `n ] = 0.
12
Verify that this set of relation satisfy the Jacobi identity. This algebra is called the Virasoro
algebra.
(iii) Prove that this is the unique central extension of the Witt algebra.
• Exercise 8.5: The stress-tensor OPE in 2D CFT
Let φ be a massless Gaussian free field in 2D with two point function hφ(z, z̄)φ(w, w̄)i =
− log(|z − w|2 /R2 ). Recall that the (chiral component of the) stress-tensor of a massless
2D Gaussian field is T (z) = − 21 : (∂z φ)2 (z) :. Prove, using Wick’s theorem, that it satisfies
the OPE
c/2 2 1
T (z1 )T (z2 ) = + + ∂z T (z2 ] + reg.
(z1 − z2 )4 (z1 − z2 )2 (z1 − z2 )
• Exercise 8.6: Transformation of the stress-tensor in 2D CFT.

Recall that the transformation rules for the stress-tensor in 2D CFT, under a conformal
transformation z → w = w(z). :
c
T (z) → T̂ (w) = [z 0 (w)]2 T (z(w)) + S(z; w),
12
with z 0 (w) the derivative of z w.r.t. w and S(z; w) the Schwarzian derivative: S(z; w) =
000 (w) 00 (w)
[ zz 0 (w) ] − 32 [ zz 0 (w) ]2 .
(i) Let us consider two conformal transformations z → w = w(z) and w → ξ = ξ(w) and
their composition z → ξ = ξ(z). Prove that consistency of the stress-tensor transformation
rules demands that:
S(z; ξ) = S(w; ξ) + [ξ 0 (w)]2 S(z, w).
154
Verify this relation from the definition of S(z; w).

(ii) Use this formula to compute the stress-tensor expectation for a CFT defined over a
infinite cylinder of radius R. Show that
π
hT (z)icylinder = −c .
12 R2
• Exercise 8.7: Regularization of vertex operators

In the text, we use the connection with lattice model to argue for the anomalous
transformation of vertex operators in gaussian conformal field theory. The aim of this
exercise is to derive (more rigorously) this transformation within field theory (without
making connection with lattice models).
Let φ(z, z̄) a Gaussian free field normalized by hφ(z, z̄)φ(w, w̄)i = − log(|z − w|2 /R2 )
with R the IR cut-off tending to infinity. In order to regularized the field we introduce a
smeared version φ of φ defined by integrating it around a small circle, of radius , centred
at z: Z 2π
dθ
φ (z, z̄) = φ(z (θ), z̄ (θ)),
0 2π
with z (θ) be point on this circle, 0 < θ < 2π. The small radius play the role of U V
cutoff.
(i) Prove that (notice that we consider the smeared at the same central position z but
with two different cutoff and 0 )
hφ (z, z̄)φ0 (z, z̄)i = min(log(R/)2 , log(R/0 )2 ).
In particular hφ (z, z̄)2 i = log(R/)2 .

2
(ii) Verify that heiαφ (z,z̄) i = (/R)α , for α real. Let us define the vertex operator by
2
Vα (z, z̄) = lim −α eiαφ (z,z̄) .
→0
Argue that this limit exists within any expectation values.

(iii) Let us now consider a conformal transformation z → w = w(z) or inversely w → z =
z(w). Show that a small circle of radius ˆ, centred at point w, in the w-plane is deformed
into a small close curve in the z-plane which approximate a circle of radius = |z 0 (w)| ˆ,
centred at z(w).
Deduce that under such conformal transformation the vertex operator transforms as fol-
lows:
2
V̂α (w, w̄) = |z 0 (w)|α Vα (z, z̄).
That is: the anomalous scaling transformation of the vertex operator arises from the fact
that the regularization scheme/geometry is not preserved by the conformal transforma-
tions.
• Exercise 8.8: Free field in a finite domain
...Dirichlet Green function in a finite domain and free field in this domain...
155
9 Scaling limits and

the field theory renormalisation group
The aim of this Chapter is, one hand, to apply concepts and techniques of the renor-
malization group to field theory, and on the other hand, to use information from the
renormalization group to define field theories as scaling limits of lattice statistical mod-
els. We shall first go back to the block spin renormalization group to understand how
observables and fields transform under the renormalization group. This will leads to an
understanding the origin of the anomalous scaling dimension of fields. We shall then dis-
cuss how RG transformations can be analyzed perturbatively, but concretely, around fixed
points. This will amount to analyse field theory as perturbations of critical field theories
and will lead to general formulas for the beta functions in the neighbourhood of RG fixed
points. At leading orders closed to a fixed point, properties of RG transformations are
encoded in simple data of the corresponding conformal field theory (scaling dimensions of
fields, OPE coefficients). This in particular applies to perturbative (one-loop) formulas
for the beta functions. These formula have tremendous applications to phenomenological
analysis of critical phenomena. Finally, we explain how analysing RG transformations
yields to a scheme to define renormalized field theories via scaling limits of lattice regu-
larized theories.
9.1 Field RG transformations

Before entering the analysis of field theory RG, let us recall how RG transformations of
effective hamiltonians were introduced in Chapter 5 using spin blocking. Let us also recall
a comment we made in Chapter 5 which should be kept in mind. RG transformations is a
very effective tool to extract the relevant variables adapted at describing a given physical
phenomena. But making sure the RG strategy works often requires having recognized
the appropriate setup (in the opposite case, one cannot make the RG transformations
contracting towards a relevant effective model).
• Back to RG transformations.
Let us recall that in Chapter 5 we introduced RG transformations in statistical spin
models by blocking together spins within cells of increasing sizes and looking at the trans-
formation rules of the effective hamiltonians of the effective spins. To each block of spins of
physical size λa, with a the original lattice cut-off, we affect an effective spin sλ , such that
block spin configurations [sλ ] label a partition of the original spin configuration space. In
other words, blocking the spins amounts to partition the configuration space, the partition
being indexed by the [sλ ]. As a consequence, sums over the original spin configurations
can be done in two steps: first summing over the spins within a given block spin configu-
ration [sλ ] (i.e. within a given element of the partition) and then over all possible block
spin configurations (i.e. over all constituents of the partition).
156
In particular, if we denote by [s] the original spin configurations and H([s]|{g}) the
original spin hamiltonian, with coupling constants {g}, the partition sum Z[{g}] :=
P −H([s]|{g})
[s] e can be written as
X
Z[{g}] = Z 0 ([sλ ]|{g}),
[sλ ]
with [sλ ] labelling the block spin configurations and where Z 0 ([sλ ]|{g}) is the condi-
tioned partitionPfunction obtained by conditioning on the value of the block spins [sλ ]:
Z 0 ([sλ ]|{g}) := [s]↓[sλ ] e−H([s]|{g}) . This conditioned partition function Z 0 can be viewed
as the Boltzmann weights for the effective block spins. The effective hamiltonian H 0
for the effective block spins is then defined by H 0 ([sλ ]|{g}) := − log Z 0 ([sλ ]|{g}). As-
suming that the set of coupling constants is complete, this hamiltonian can written
as the original hamiltonian but with a new set of coupling constants {g λ }, that is:
H 0 ([sλ ]|{g}) = H([sλ ]|{g λ }), or equivalently
λ ]|{g λ })
Z 0 ([sλ ]|{g}) = e−H([s ,
The important point is that the RG transformation leaves the partition function invariant
in the sense that summing over the block spin configurations [sλ ] with Boltzmann weights
specified by their effective hamiltonian H([sλ ]|{g λ }) reproduces the original partition sum:
λ λ
Z[{g}] = [sλ ] e−H([s ]|{g }) .
P
The transformation map from {g} to {g λ },
{g} → {g λ } = Rλ ({g}).
0 0
is the RG transformation on coupling constants. They form a group Rλ λ = Rλ · Rλ
(actually a pseudo-group since they cannot be inverted). As a consequence, the flow of the
coupling constants obtained by progressively increasing λ is generated by the infinitesimal
transformations. This leads to define the corresponding vector fields (on the space of
coupling constants), called “beta-functions”, by
βα ({g}) := λ∂λ gαλ λ=1 .

They generate the RG map {g} → {g λ } in the sense that {g λ } is solution of the flow
equation λ∂λ gαλ = βα ({g λ }), with initial condition {g λ=1 } = {g}. The solutions g λ are
called the “running coupling constants”.
• Block-spin transformation on observables.
We now look at how operators transform under the RG. This will give us access to
their anomalous scaling dimensions. Recall that an observable O is some kind of function
(a measurable function), [s] → O([s]) on the configuration space. In a way similar to what
has been done with the partition function, RG transformations for observables O → Oλ
can be defined using conditioned sums as follows
1 X
Oλ ([sλ ]) = e−H([s]|{g}) O([s]). (65)
Z 0 ([sλ ]|{g}) λ
[s]↓[s ]
157
λ λ λ λ
We may
λ
λ understand O ([s ]) as the following conditioned expectation values: O ([s ]) =
E[O [s ]] with E he measure on the original spin configuration [s] specified by the
Boltzmann weights e−H([s]|{g}) . Of course the transformed operators depend both on
the scale transformation factor λ and on the coupling constants. Recall that we set
0 λ −N eλ ({g}) −H([sλ ]|{g λ }) P 0 λ
Z ([s ]|{g}) = e e and that Z({g}) = [sλ ] Z ([s ]|{g}). Hence by
construction, the transformed operators are such that
1 X λ λ λ 1 X
e−N e ({g}) e−H([s ]|{g }) Oλ ([sλ ]) = e−H([s]|{g}) O([s]).
Z({g}) λ Z({g})
[s ] [s]
The l.h.s. is the expectation with respect to the transformed action with coupling con-
stants {g λ }, the r.h.s. is the original expectation with the original action with coupling
constant {g}. The relation thus tells us that expectations of the transformed operators Oλ
with respect to the transformed hamiltonian is the same as the expectation of the initial
operator with the initial hamiltonian:
hOλ i{gλ } = hOi{g} . (66)
This is a simple but important equation.

• Local field transformations and mixing matrix.
Suppose now that the observable is a product of local observables Φk , localized at
different lattice points (say nk ): Y
O= Φk (nk ),
k
where each Φ is a local operator sensitive to the local configuration of the spin variables
(hence the definition of local operators involves open neighbourhoods of their insertion
points). It is reasonable to expect that the transformed operators are still going to be
quasi-local (because if the rescaled lattice λa is much smaller than the correlation length
ξ, the spins inside a block behave almost collectively). It can thus be decomposed on the
set of local operators so that we have (or more precisely we expect)
Φλ (n) = Γλ ({g}) · Φ(n/λ),
where Γλ is a matrix – coding for the decomposition of the transformed operator on the
basis of the original operators. It is called the mixing matrix. We wrote
P λthe above equation
λ σ
in a matrix form: if we prefer to keep indices it reads Φα (n) = σ [Γ ]α Φσ (n/λ) where
α, σ indexes all possible local operators. The rescaling of the position of the operators
comes about from the same argument for the rescaling of the correlation length: After a
block spin transformation the dimensionless distances between the operators (counted by
the number of lattice site to cross to go from operator to the other) has been divided by
λ. (Here n is the integer dimensionless distance, counted in unit of lattice spacing).
The matrices Γλ (g) are called the mixing matrices.
They clearly depend both on the rescaling factor λ and on the coupling constant.
They inherit a composition law —more precisely a cocycle structure— from that of the
158
RG transformations. Applying successively twice a block spin transformations on the

operators implies
0 0 0
Γλ (g) · Γλ (g λ ) = Γλλ (g).
Of course Γλ=1 (g) = I.
The same relation applies (locally) if we consider an operator made of products of
operators localized at different position, say ⊗k Φk (nk ). Under RG transformations, each
of this operators transforms as above with an associated Γλk matrix. For operators and
their expectations the fundamental RG relation is thus the equality:
Y Y
h Γλk (g) · Φk (nk /λ)i{gλ } = h Φk (nk )i{g} . (67)
k k
Again: this tells that the expectations of the transformed operators, at the transformed
positions and with the transformed action, is the original expectations. This is a fun-
damental relation (the lattice analogue of the so-called Callan-Symanzik equation, see
below).
0
If we sit at a fixed point g∗ then it transforms multiplicatively: Γλ (g∗ ) · Γλ (g∗ ) =
0
Γλλ (g∗ ). The scaling operators Φi of the fixed (conformal) theory are those diagonalising
this matrix so that (at a fixed point)
Φλi (n) = Γλ (g∗ ) · Φi (n/λ) = λ−∆i Φi (n/λ),
with ∆i the scaling dimension of Φi . At a fixed point, Γλ (g∗ ) acts multiplicatively and
Γλ (g∗ ) = λ−∆i on a scaling field of dimension ∆i .
• Scaling operators and operator scaling dimensions.
We now relate the RG eigenvalues to the scaling dimensions of the operators at the
conformal point. The continuum theory corresponding to a fixed point has no typical scale
and it is scale invariant. It is thus described by a conformal field theory (with an action
that we denote S∗ ). That is: conformal field theories are fixed points of the renormalization
group. Exploring the neighbor of the critical point amounts to perturb the critical action
S∗ with by some operators Oα .
X X
S = S∗ + (δg)α Oα (n),
α n
with n labeling the integer points on the lattice. However, we can change basis of operators.
Recall that a conformal field theory comes equipped with its set of operators Φi (x) of
scaling dimension ∆i . They transform homogeneously under dilatation Φi → λ−∆i Φi .
Using this basis, we can alternatively write
X X
S = S∗ + ui Φi (n).
i n
the behaviour near a critical point is thus governed by scaling operators at the critical
point. In particular the critical exponents directly are related to the scaling dimension of
the operators at the critical point. This statement is made more precise by looking at the
RG transform close to a fixed point.
159
A comment about the normalization of fields on the lattice and in the continuum is
needed. We initially start with dimensionless fields defined on the integer lattice point,
say Φlatt
i (n) where n refers to the integer point of lattice of mesh size a (and we temporally
added the label ‘latt’ to specify that these fields are defined within the lattice model). At
the fixed point, their two point functions are power laws with unit normalization (because
they are dimensionless), i.e.
1
hΦlatt latt
i (n)Φi (m)i∗ = .
|n − m|2∆i
In the continuous limit (a → 0), the scaling fields are defined as (recall how we define the
scaling limit of free random paths or how we define the renormalized vertex operators in
conformal field theory)
Φcont.
i (x) = lim a−∆i Φlatt (n) = lim a−∆i Φlatt (x/a).
a→0,n→∞ a→0
x=an fixed
They are now dimension-full (with the dimension of [length]−∆i ) with finite two-point
correlation functions (defined in the fixed point conformal field theory) normalized to
1
hΦcont.
i (x)Φcont.
i (y)i∗ = .
|x − y|2∆i
Away from the critical point the action can thus be written as
X X X Z
i latt i ∆i −D
S = S∗ + u Φi (n) = S∗ + u a dD x Φcont.
i (x).
i n i
The two above expressions are equivalent (in the limit P alatt→ 0) because
P ∆ithe discrete
cont.
sum over
R lattice points approximates ∆ the integral: n Φi (n) = n a Φi (n) =
a∆i −D dD x Φcont.
i (x). The prefactor a i −D makes the action dimensionless (because u
i
is dimensionless while Φi as dimension ∆i ).
Under dilatation (RG transformation) the scaling variable transforms as ui → λyi ui ,
hence consistency of the coupling tells us that
yi = D − ∆ i . (68)
Critical exponents are thus encoded in the scaling dimensions of the operators at the
conformal points. Operators are gathered in three classes:
— Relevant operators have dimension ∆i < D;
— Marginal operators have dimensions ∆i = D;
— Irrelevant operators have dimension ∆i > D.
• Relation with the beta-functions.
The mixing matrices are not independent of the beta functions, because local operators
are dual to coupling constants. Let us slightly deformed the initial hamiltonian into
X X
H({g}) → H({g + δg}) = H({g}) + δg α Φα (n).
α n
160
To first order in δg the expectation values are

X X
h[· · · ]i{g+δg} = h[· · · ]i{g} − δg α h[· · · ]Φα (n)i{g} .
α n
Let us now do a block spin transformation with scale factor λ. By taking into account
factors λD coming from the dilatation of the volume, consistency gives
X
Rλ (g + δg)α = Rλ (g)α + δg σ λD [Γλ (g)]ασ + · · · ,
σ
or alternatively,
∂ λ α
R (g) = λD [Γλ (g)]ασ .
∂g σ
This means that the mixing matrix Γλ (g) is, up to a factor λD the derivative of the
renormalization group, a non so-surprising result in view of the duality between local fields
and coupling constants. Let us define the so-called “matrix of anomalous dimensions” γ(g)
(whose interpretation will be further developed below) by
γ(g) = −λ∂λ Γλ (g)||λ=1 .
Recall the definition of the beta-function, β(g) = λ∂λ g λ |λ=1 . Then, we have the relation
∂α β σ (g) = Dδασ − γασ (g). (69)
At a fixed point, it is reduces to the relation yi = D − ∆i between the RG eigenvalues and

the dimensions of the scaling fields.
9.2 Momentum RG transformations in field theory

We can alternatively formulate the RG transformations in momentum space. This idea is
then to integrate recursively over shells of high momenta and to rescale space appropriately.
To make it more precise let us assume that we look at a scalar field φ with action
S[φ] = S0 [φ] + S1 [φ].
where S0 is a free quadratic action and S1 is coding for the interaction. To simplify matter,
we choose S0 to be the action of a massless scalar field in dimension D > 2 with a (smooth)
momentum cut-off Λ = 1/a.
The field theory defined by S0 is therefore Gaussian with two-point function
dD k eik·(x−y)
Z
G(x − y) = hφ(x)φ(y)i|S0 = ϕΛ (|k|),
(2π)D k2
where ϕΛ (|k|) is the smooth cut-off function: ϕΛ (k) ' 1 for |k| ≤ Λ and ϕΛ (k) ' 0 for
|k| ≥ Λ, that is: ϕΛ varies only abruptly from 1 to 0 when |k| crosses the cut-off value Λ.
Let us denote by dµG [φ] this Gaussian measure: (in)-formally dµG [φ] = [Dφ] e−S0 [φ] . The
−S
partition function of the interacting theory is Z = dµG [φ] e 1 [φ] .
R
161
To implement the RG idea, we want to pick a rescaling factor λ > 1 and to integrate
over all modes with momenta in the shell between λ−1 Λ and Λ. The part of the Gaussian
covariance with support in this shell is Γλ , with
dD k eik·(x−y)
Z

Γλ (x) = D 2
ϕΛ (|k|) − ϕΛ (λ|k|) ,
(2π) k
because the difference ϕΛ (|k|) − ϕΛ (λ|k|) is significantly non zero only for |k| ∈ [λ−1 Λ, Λ].
Note then that G = Gλ + Γλ with Gλ (x) = λ−(D−2) G(x/λ), by construction. This
decomposition of the covariance tells us that we can decompose the field φ as a sum
φ = φ̂λ + ϕλ where φ̂λ and ϕλ have covariances Gλ and Γλ respectively (because the sum
of two independent Gaussian variables is a Gaussian variable with covariance equals to
the sum of the two covariances).
Since Gλ is defined from G be rescaling, we can construct the field φ̂λ by taking an
independent copy of φ, which we denote φ̂, and set φ̂λ (x) = λ−(D−2)/2 φ̂(x/λ). We thus
get the decomposition of the Gaussian measure as
dµG [φ] = dµG [φ̂] dµΓλ [ϕλ ], φ = φ̂λ + ϕλ ,
with dµG [φ̂] the initial Gaussian measure but for the independent variable φ̂ and dµΓλ [ϕλ ]
the Gaussian measure with covariance Γλ .
With all this preparation we can now integrate over the modes with momenta in the
shell [λ−1 Λ, Λ] by integrating over the field ϕλ . This indeed gives a precise formulation of
integrating over the high momenta and rescaling space appropriately. For the Boltzmann
weights, or equivalently for the action, this yields the definition of the flow S1 → Sλ by
Z
−Sλ [φ̂]
e := dµΓλ [ϕλ ] e−S1 [φ̂λ +ϕλ ] .
As above with block spin transformations, this map is such that it preserves the partition
function: Z Z
−S1 [φ]
Z = dµG [φ] e = dµG [φ] e−Sλ [φ] ,
by construction. This defines the RG flow of the (interacting) part of the action:
Rλ : S1 → Sλ .
0 0
As above it forms a semi-group under composition: Rλ ◦ Rλ = Rλλ . The RG flow is here
defined on the space of actions and not on the space of coupling constants, but these two
formulations are equivalent because coupling constants are simply all possible parameters
specifying an action.
We can also describe how operators or observables evolve under momentum RG trans-
formations. Let F1 [φ] be some functional of the field which we identify with an observable.
We can again define its RG transform by decomposing the field as φ = φ̂λ + ϕλ and by
integrating over the modes ϕλ . This yields the definition of the transformed observable
Fλ as Z
Fλ [φ̂] = e +Sλ [φ̂]
× dµΓλ [ϕλ ] e−S1 [φ̂λ +ϕλ ] F1 [φ̂λ + ϕλ ].
162
Note the presence of the prefactor e+Sλ [φ̂] which ensures that the identify function is
mapped to the identity. This transformation is such that expectations are preserved,
namely
hF1 [φ]i|S0 +S1 = hFλ [φ̂]i|S0 +Sλ .
These rules are analogous to those for RG transformation of operators in lattice models,
see eq.(66).
9.3 The perturbative renormalization group

The aim of this section is to compute the β-function perturbatively near a fixed point and
to analyse the consequences of this formula.
• One-loop beta-functions.
A fixed point is described by a conformal field theory. Let S∗ be its action. Away from
the critical point, the field theory action is a perturbation of S∗ by some operators Φi :
X Z
S = S∗ + gi a∆i −D dD x Φi (x).
i
The operators Φi are supposed to be relevant operators (as otherwise they don’t change
the long distance physics) of dimensions ∆i . The factors a∆i −D have been introduced to
make the action explicitly dimensionless. The operators Φi are assumed to be normalized
by their two point function
1
hΦi (x)Φj (y)i∗ = δi;j .
|x − y|2∆i
Here and below h· · · i∗ refers to the expectation values at the fixed point. A conformal
field theory is characterized by its operator product expansion (OPE):
X k
Cij
Φi (x)Φj (y) = Φk (y) + · · · .
|x − y|∆i +∆j −∆k
k
k are the structure constants of the OPE, they are determined by the 3-
The coefficients Cij
point functions of the scaling operators at the conformal point (see Chapter 8). Existence
of OPE is expected in any conformal field theory but not really proved in full generalities
(unless we tautologically put the existence of OPEs as part of the axioms defining a
conformal field theory).
We are going to prove that to lowest orders, the beta function is given by:
SD X k i j
β k (g) := a∂a g k = (D − ∆k )g k − Cij g g + · · · . (70)
2
ij
We see that the one-loop beta-function is fully determined by the OPE structure of the
conformal fixed point. We even can absorb the factor SD in a redefinition of the coupling
constant. The rules is that all relevant operators compatible with the symmetries which
can be generated under OPE should be included in the beta function.
163
Formula (70) is an important formula.

To prove (70) let us now imagine computing the partition functions by perturbing the
conformal field theory. (We will deal with correlation functions a bit later. That is we
first deal with the ‘measure’ and then with the observables.). It is defined by the action
S,
Z Z
∆ −D
P R D
Z × hO(z) · · · i := [Dϕ] e Z × hO(z) · · · i = [Dϕ] e−S∗ − i gi a i
−S d x Φi (x)
O(z) · · · .
We expand in perturbation theory up to second order (also called ”one-loop” order).

X Z
Z × hO(z) · · · i = hO(z) · · · i∗ − g i a∆i −D dD x hΦi (x) O(z) · · · i∗
i
Z
1 X i j ∆i −D ∆j −D
+ gg a a dD x1 dD x2 hΦi (x1 )Φj (x2 ) O(z) · · · i∗ + · · ·
2!
ij
We insert operators O(z) · · · just to remember that we are dealing with correlation func-
tions but do not really take care of them yet (we shall do it just in a while when computing
the matrix of anomalous dimensions). The integrals are UV divergent because of the singu-
larity in the OPE (due to large fluctuations in the microscopic model). They are actually
also IR divergent, but those IR divergencies can be treated by defining the systems is
in finite box. We will not care about the IR divergences because understanding the RG
behavior closed to the fixed point only requires dealing with the UV divergences. To cure
the UV divergences we imagine that the system is defined on a lattice so that the field
insertions are always at a distance a away each other. That is: the integral have to be
understand with a short distance cut-off (and implicitly also an IR cut-off), namely:
Z
dD x1 dD x2 hΦi (x1 )Φj (x2 ) · · · i∗
|x1 −x2 |>a
The renormalization problem consists now in answering the following question: How
should the coupling constant be a-dependent such that the partition function is unchanged
when dilating the cut-off a → λa with λ = 1 + ε close to one? This is equivalent to
answering the question: How should the coupling constants be a-dependent in order to
compensate for the UV divergences? Thus, let us replace gi → gi (a) and demand that the
derivative of the partition function w.r.t. a vanishes when a → 0. We get, keeping only
the important terms (the others terms either do not contribute or contribute to the higher
orders of the beta functions):
h X Z X Z
i ∆i −D D i ∆i −D
0 = − (a∂a g ) a d x Φi (x) − g (∆i − D) a dD x Φi (x)
i i
Z
1 X i j ∆i −D ∆j −D i
− gg a a dD x1 dD x2 δ(|x1 − x2 | − a) Φi (x1 )Φj (x2 ) + · · ·
2
ij
The last term comes from slightly moving the short distance cut-off (say, the Dirac δ-
function δ(|x1 − x2 | − a) comes form the derivative of the Heaviside function Θ(|x1 −
164
x2 | − a) defining the cut-off integrals6 : this term reflects the block spin transformation.
As a → 0 we can evaluate the last integrals using the operator product expansion (OPE)
k
Cij
Φi (x + a)Φj (x) ' ∆i +∆j −∆k Φk (x). The Dirac δ-function δ(|x1 − x2 | − a) reduces the
|a|
integral to the angular variables over the D-dimensional sphereR with volume aD SD , with
SD = 2π D/2 Γ(D/2). Hence to lowest order, the term linear in dD x Φk (x), are
h X Z X Z
0 = − (a∂a g i ) a∆i −D dD x Φi (x) − g i (∆i − D) a∆i −D dD x Φi (x)
i i
Z
1 X k i j ∆k −D D
i
− Cij g g a SD d x Φk (x) + · · ·
2
ijk
Demanding this to vanish yields the announced claim (70). Computing the beta function
to higher order is more complicated (in particular the short distance cut off we used is not
very much adapted and not easy to deal with at higher orders).
• One-loop anomalous dimensions.
We now perturbatively compute the matrix of anomalous dimension. The computation
is very similar. As we know there is operator mixing. We have to consider families of
operators Oα , of dimension ∆α , whose correlation function are perturbatively defined by
the expansion
X Z
i ∆i −D
hOα (y) · · · i = hOα (y) · · · i∗ − g a dD x hΦi (x) Oα (y) · · · i∗
α
Z
1 X i j ∆i −D ∆j −D
+ gg a a dD x1 dD x2 hΦi (x1 )Φj (x2 ) O(y)α · · · i∗ + · · ·
2!
ij
We know look how to redefined them and the coupling constants,
g i → g i (a), Oα → a−∆α Γσα (a) Oσ ,
such as to compensate the a-divergences. We adopt the same computational strategy

as before. The cut-off integral dD x Φi (x) Oα (z) yields Φi (y + a) Oα (y) which can be
R
evaluated using the OPE structure

σ
Ciα
Φi (y + a) Oα (y) ' Oσ (y).
|a|∆i +∆α −∆σ
Looking at the terms linear in Oσ in the a-derivative of the expectation of hOα (y) · · · i
give to lowest order
X
(a∂a Γ · Γ−1 )σα + ∆α δασ + SD g i Ciα
σ
= 0.
i
6
R R
Be careful with the signs: ∂a Θ(|x − y| − a) = δ(|x − y| − a), or equivalently |x−y|>a+da
− |x−y|>a
=
R
− a+da>|x−y|>a
165
Figure 7: Graph of a beta function with a nearby zero; Schematic representation of a RG

flow between two fixed points, from UV to IR.
Hence, the matrix of anomalous dimensions at one-loop order is

X
γασ = ∆α δασ + SD g i Ciα
σ
+ ··· . (71)
i
Again it is completely determined by the OPE structure of the fixed point. Note that
we have γij = Dδij − ∂i β j as we should. Again higher orders in the coupling constant are
more difficult to compute (and are not encoded in simple universal structures as the OPEs
coefficients).
9.4 The Wilson-Fisher fixed point

Here we illustrate, on a famous exemple, the use of the perturbative expression of the beta-
function to perturbatively learn about new fixed points which are near another controllable
fixed point.
• Almost marginal perturbations.
Consider perturbing a fixed point by an almost marginal operator Φ of dimension
∆ = D − with 1. The beta function is then of the form β(g) = g + O(g 2 ). The
one-loop terms depend on the OPE structure at the fixed point. Suppose for simplicity
that the OPE is such that Φ × Φ = Φ and that no other operator is generated by the RG.
The beta function is then
C
β(g) = g − SD g 2 + · · · .
2
The flow is as described in the Figure 7. The remarkable fact is that this beta function
vanishes for g of order so that we can trust the approximate one-loop beta function. The
new fixed point is at point g∗ ,
1
g∗ = + O(2 ),
SD C
whose sign depends on the sign of the OPE coefficient C. If g has the appropriate sign,
the large distance physics of the system is governed by this new fixed point. The new
anomalous dimension of the operator Φ at the new IR fixed point is obtained by linearizing
166
the beta function at the new fixed point (or by using the relation γ = D−∂β): β(g∗ +δg) =
−δg + · · · . The operator Φ is thus irrelevant at the new fixed point (as it should be since
we are approaching the new fixed point by the operator) with new dimension δ∗ = D + .
• The Wilson-Fisher fixed point.
We look at the φ4 theory with action
Z h1 1 1 i
S = dD x (∇φ)2 + g2 φ2 + g4 φ4 ,
2! 2 4!
which
R D we view as a perturbation of the massless gaussian fixed point with action S∗ =
1 2
2 d x(∇φ) . In the Gaussian theory, perturbative expectation can be computed using
Wick’s theorem. The massless Green function in dimension D is
1
hφ(x)φ(y)i∗ = .
|x − y|D−2
Thus φ has dimension (D − 2)/2 at the Gaussian fixed point. To avoid self Wick’s contract
we define the operator φ2 and φ4 by normal ordering. This amounts to subtract their
vacuum expectation (self contractions), so that
: φ2 := φ2 − hφ2 i∗ , : φ4 := φ4 − 3hφ2 i∗ φ2 .
At the massless Gaussian fixed point, the dimension of the operator φn :=: φn : is n(D −
2)/2. For 3 < D < 4, the operators φ2 =: φ2 : and φ4 =: φ4 : are the only even relevant
operators. So the action is
Z h1 1 i
S = S∗ + dD x g2 φ2 + g4 φ4 .
2! 4!
We should actually also included (∇φ)2 which is marginal with dimension D but it has no
effect at one-loop. It is a simple exercice to compute the OPE between φ2 and φ4 to get:
φ2 × φ2 = 2 I + 4 φ2 + φ4 ,
φ2 × φ4 = 12 φ2 + 8 φ4 ,
φ4 × φ4 = 24 I + 96 φ2 + 72 φ4 .
We did not write the terms corresponding to operators φn with n > 4 or to higher deriva-
tives (because they are irrelevant). The coefficients here comes from Wick’s theorem and
thus from computing the number of different pairings.
Let D = 4 − so that φ2 has dimension 2 − and φ4 has dimension 4 − 2 (it is almost
marginal). The beta functions are (we absorb the factor S4 /2 in the coupling constant):
β2 = a∂a g2 = 2 g2 − 4 g22 − 24 g2 g4 − 96 g42 + · · ·

β4 = a∂a g4 = g4 − 72 g42 − 16 g2 g4 − g22 + · · ·
For 1 there is a nearby fixed with g4∗ = O() and g2∗ = O(2 ):

g4∗ = + O(2 ), g2 = O(2 ).
72
167
The dimension of the operator φ2 at the new fixed point can be read from the linearization
of the beta function at this fixed point: β2 = (2 − 24 g4∗ )g2 + · · · = 2(1 − 6 )g2 + · · · . The
operator φ2 is coupled to the temperature so that g2 ∼ δT . We thus have yT = 2(1 − 6 )
and hence the critical exponent
1 1
ν= = + + O(2 ).
yT 2 12
This is the critical exponent for the Ising model in dimension 4 − which, once (naively)
extended to dimension 3 by setting = 1, is in reasonably good agreement with the
exponent obtained by numerical simulations. Other exponents are found by computing
the dimension of the operators (say φ which is coupled to the magnetic field) at the new
fixed points. Those are obtained from the OPE coefficient as explained above. We leave
this computation as an exercise.
9.5 Scaling limits and renormalized theories

Constructing the renormalized theory amounts to define the field theory in the continuum
by taking the limit of vanishing lattice spacing. As we will see this requires approaching
simultaneously the critical surface is such way as to preserve the dimension-full correlation
length. The (finite) continuous theory is called the renormalized theory.
• Real space correlation length and scaling limit.
Suppose for a little while that there is only one relevant scaling variable uT . The
(dimensionless) correlation length diverges as ξ ' |uT |−1/yT when close to the fixed point
(i.e. |uT | → 0). The physical dimension-full correlation length thus behaves as a|uT |−1/yT
with a the lattice mesh size. If we aim at defining a continuous theory with a physical
finite correlation length we thus have to take the limit a → 0, |uT | → 0, with a|uT |−1/yT
fixed. Equivalently we may take the continuous limit a → 0 fixing a mass scale (inverse
of length scale):
m := a−1 |uT |1/yT .
That is: to define the continuous field theory with fixed finite mass scale we have to
simultaneously approach the critical hyper-surface, uT → 0, as we take the small lattice
size limit. This is called a double scaling limit: a → 0, uT → 0 with a−1 |uT |1/yT fixed.
ˆ
Doing this we fix a dimension-full correlation length ξ,
ξˆ = m−1 = a |uT |−1/yT ,
or equivalently, we scale uT appropriately as a → 0,
uT (a) = |am|yT , as a → 0.
At this point may be a good idea to go back Chapter 2 and to recall how the continuous
(scaling) limit of free random path was defined in that Chapter.
If ξ is the dimensionless correlation length, the two point functions behave as G(n) '
e −|n|/ξ with ξ ' |u|−1/yT . Here |n| is the dimensionless distance, counted in unit of the
168
lattice mesh, so that the ‘physical distance’ is |x| = a|n|. The double scaling limit ensures
that the existence of the continuous two-point functions
1/yT
G(x) = lim G(n) 'x large lim e−|n||uT ' e−m|x| .
a→0,n→∞ a→0,n→∞
|x|=a|n| fixed |x|=a|n| fixed
If there are more than one relevant coupling constant, say two uT and uh , we have to
take simultaneously the scaling limit in all these variables in order to approach the critical
surface as a → 0, that is
a → 0, with mT = a−1 |uT |1/yT and mh = a−1 |uh |1/yh fixed.
The continuous theory will then depend on the two independent mass scales mT and mh ,
or alternatively ony one mass scale, say mT , and on the scale invariant ratio of the scaling
|T |yh
variables, say |u T| h
y
|uh | T ' y
|h| T .
We can alternatively write these scaling relations in terms of RG transformations.
Recall that the scaling variables scale homogeneously under RG transformation, say uT →
uλT = λyT uT . If we pick uT (a) = (amT )yT as initial condition as required for the double
scaling limit, then the running coupling constant at scale λ = `R /a is fixed, that is:
uT (a)λ=`R /a = (`R mT )yT .
We set uR = (`R mT )yT so that uT (a)λ=`R /a = uR . The new length `R (which makes
λ = `R /a dimensionless) can be chosen at will. It is called the “renormalization scale”.
In terms of generic coupling constants the scaling limit thus requires taking the con-
tinuous limit a → 0 with coupling constants g(a) with fixed values at scale λ = `R /a,
i.e.
a → 0, with g(a)λ=`R /a = gR fixed.
Imposing this condition imposes to the coupling constants g(a) to approach the critical
hypersurface as a → 0.
• Scaling limits of fields.
To define the continuous limit of expectation of products of operators we have to take
their anomalous dimension into account. Consider first scaling operators Φi of scaling
dimension ∆i . At the fixed point, the lattice two point functions are
1
hΦi (n)Φi (m)ilatt
∗ = .
|n − m|2∆i
The continuous theory is defined by taking the limit a → 0, n → ∞ with x = an fixed.
This demands to define the scaling operator in the continuous theory by
x
Φcont.
i (x) = lim a−∆i Φlatt
i (n = ).
a→0 a
This shows that to take the continuous limit we have to dress the field in a-dependent
way according to their anomalous dimensions. Hence, away from the fixed point we have
to combine this dressing with the running of the coupling constant g(a) explained above.
169
The best we can hope for is that there exist matrices (acting on the set of operators) Γ(a)
such that the following limits:
Y
lim h Γ−1 latt
k (a) · Ok (nk )ig(a) exist,
a→0,nk →∞
xk =ank fixed k
for any operators Ok . Alternatively, we look for matrices Γ(a) (acting on the set of
operators) such the field operators in the continuous theory could be defined from the
lattice ones by the following limiting procedure
x
Ocont. (x) = lim Γ−1 (a) · Olatt (n = ).
a→0, a
g(a)λ=`R /a =gR fixed
This relation is simply the (formal) generalization of the relation Φcont. (x) = a−∆i Φlatt
i (n =
x/a) in case of scaling fields.
We shall show that we can (formally) construct this matrix using input from the mixing
matrices of the RG transformations. Namely, choosing,
Γ(a) = Γλ=`R /a (g(a)), (72)
ensures that the correlation functions h k Γ−1 latt

Q
k (a) · Ok (nk )ig(a) are (formally) independent
of the lattice cut-off a, asymptotically as a → 0 at xk = ank fixed.
If these limits exist, we could define the ‘renormalized’ continuous theory as (see below)
Y Y xk latt
h Ok (xk )iR gR = lim h Γ−1 k (a) · Ok (nk = a )ig(a) . (73)
a→0
k λ=` /a
g(a) R =gR fixed k
Let us argue that the RG transformation (formally) ensures that such limits exist (if
the mixing matrices of the RG transformations can be sufficiently controlled, see previous
section). Recall from previous sections that RG transformations on fields yield that:
Y Y
h Γλk (g) · Ok (nk /λ)ilatt
{g λ } = h Ok (nk )ilatt
{g}
k k
xk
for a small enough and any λ, where Γλ are the RG mixing matrices. Picking nk = a
and λ = `R /a yields
Y λ=1/a xk Y xk latt
h Γk (g) · Ok ( )ilatt = h Ok (nk = )i .
a {g}
{g λ=` /a
R }
`R
k k
Choosing now the coupling constant to be g(a) such that g(a)λ=`R /a = gR fixed and
multiplying both sides of the equation by Γ−1
k (a) = [Γ
λ=`R /a (g(a))]−1 we get that
Y λ=` /a xk latt
h [Γk R (g(a))]−1 Ok (nk = )i ,
a {g(a)}
k
is independent of a, at least for a small (a → 0) as claimed. The limit a → 0 is thus

expected to exist and define the renormalized correlation functions (73).
170
• Renormalized correlation functions.

The renormalized continuous theory is thus (tentatively) defined by
Y Y λ=` /a xk latt
h Ok (xk )iRgR := lim h [Γk R (g(a))]−1 · Ok (nk = )i . (74)
a→0 a g(a)
k λ=` /a
g(a) R =gR k
It is defined from data encoding the behavior under RG transformations. Of course proving
rigorously that the limit exists requires more work (this is clearly much beyond the scope
of this simple introduction). By construction, the renormalized expectations are functions
of the renormalized parameters gR but not of the lattice coupling constants g(a).
Renormalization procedures to construct continuous field theory encode all the infor-
mation about critical systems, and reciprocally RG transformations are the data needed
construct continuous field theory. The relation between the renormalized and lattice cor-
relation functions can be inverted, so that
Y xk latt Y
h Ok (nk = )ig(a) ∼ h Γk (a) · Ok (xk )iR
gR , with g(a)λ=`R /a = gR ,
a
k k
for a → 0. Hence, the renormalized correlation functions code for the large distance
behavior of the lattice correlation functions, provided the coupling constants approach the
critical hypersurface (in the appropriate double scaling limit) as the distance are scaled
away.
Let us assume that there is only one relevant variable uT . Then the scaling relation
is uT (a) = (am)yT with ξˆ = m−1 the dimensionfull correlation length, which is the only
renormalized parameter of the theory. If Φi are scaling fields, then the RG mixing matrix
is Γλ = λ−∆i so that Γi (a) = Γλ=1/ma = (ma)∆i and
Y Y
h Φi (xi )ilatt
uT (a) ' h (ma)∆i Φi (xi )icont
ξ̂=1/m
.
i i
This is again another way to express that Φcont

i = a−∆i Φlatt
i . It expresses that the corre-
lation length is the only relevant macroscopic length for a physical system near criticality
(if we are dealing the system in infinite volume).
9.6 Covariance of the renormalization

Here, we discuss a covariance property of the renormalized field theory with respect to
changing the (arbitrary) renormalization scale. As a consequence, the renormalized corre-
lation functions satisfy a differential equation, encoding for their behaviors under dilation,
which has very important consequences.
• RG transformations and the Callan-Symanzik equation.
The renormalization scale `R at which we fixed the running constant is arbitrary: a
change in the renormalization scale `R → λ0 `R can be compensated by a change in the
renormalized coupling constant gR → gRλ since
0 0 0
g(a)λ=λ `R /a = [g(a)λ=`R /a ]λ = gR
λ
.
171
For instance, if the coupling constant is a scaling variable uT then uT (a) is such that
uT (a)λ=`R /a = (`R mt )yT so that changing the renormalization scale `R → λ0 `R amounts
to change the renormalized coupling uR = (`R mT )yT into u0R = λ0 yT uR .
The physical correlation length ξˆ = 1/mT is left invariant by this transformation. By
dimensional analysis, the two point correlation function GR R
∆ (x, y) = hφ∆ (x)φ∆ (y)i of a
field of scaling dimension ∆ can be written as
1 ˆ = 1
GR
∆ (x, y) = 2∆
F (|x − y|/ξ) F (|x − y|mT ),
|x − y| |x − y|2∆
since ξˆ = 1/mT is the only scale of the problem. Since uR = (`R mT )yT this can alter-
1 1/yT
natively be written as GR ∆ (x, y) = |x−y|2∆ F (uR |x − y|/`R ). By construction this two
point function is under the simultaneous transformation `R → λ0 `R and u0R = λ0 yT uR and
satisfies
x · ∂x + y · ∂y + 2∆ − yT uR ∂uR GR

∆ (x, y) = 0.
This set of invariances and equations reflects the redundancy in parametrizing the renor-
malized theory via the renormalized coupling constant uR and renormalisation length `R .
They are called the Callan-Symanzik equation.
More generally, this implies that the renormalized expectations are also covariance
under RG. Let us first state this covariant relation and then prove it. The covariance says
that
Y 0 xk
h Γλk (gR ) · O( 0 )iR 0 λ0 − independent. (75)
λ gRλ
k
This equation reflects the fact that the renormalization length scale `R was not present
in the initial formulation of the problem. It expresses the covariance of the renormalized
expectations w.r.t. to new arbitrary length scale. Taking the derivative w.r.t to λ0 gives
the so-called Callan-Symanzik equations
N
hX N
X X i
(k)
[ x k · ∂ xk + γ (gR ) − β (gR )∂gi hO1 (x1 ) · · · ON (xN )iR
i
{gR } = 0, (76)
R
k=1 k=1 i
with, as before,
λ
β(gR ) = λ∂λ gR |λ=1 , γ(g) = −λ∂λ Γλ (gR )|λ=1 .
A (maybe not so usual) way to derive this equation is by using the composition law
for the RG mixing matrices. Let start as above with a RG invariance of the lattice
expectations but for scale dilatation λλ0 instead of λ0 . We have (as above)
0
Y Y
0 latt
h Γλλ k (g) · Ok (nk /λλ )i{g λλ0 } = h Ok (nk )ilatt
{g}
k k
0 0
Recall the composition law for the mixing matrices Γλ (g) · Γλ (g λ ) = Γλλ (g). In particular
for λ = `R /a and g(a)λ=`R /a = gR , this becomes,
0 0
Γλ=`R /a (g(a)) · Γλ (gR ) = Γλ /a (g(a)). (77)
172
Hence RG invariance now reads

Y 0
Y xk latt
h Γλ=1/a (g(a)) · Γλ (gR ) · Ok (xk /λ0 `R )ilatt0 = h Ok (nk = )i ,
{gRλ }
a {g(a)}
k k
for a-small enough. Multiplying by [Γλ=1/a (g(a))]−1 as above, and taking (formally) the
limit a → 0, we get
Y 0 xk Y
h Γλk (gR ) · Ok ( 0 )iR λ0 = h O(xk )iR
{gR } .
λ {gR }
k k
as claimed.
• Solutions of the RG equations.
Let us assume that there is only one coupling constant g with beta-function β(g).
Given the beta-function, the RG flow equation is λ∂λ g(λ) = β(g(λ)) whose solution (with
initial condition g(λ = 1) = g) is:
g(λ) dg 0
Z
λ = exp .
g β(g 0 )
Reciprocally this defines a mass scale m(g) solution of β(g)∂g m(g) = m(g):
dg 0
Z g
m(g) = exp[ .
β(g 0 )
This is the only scale of the problem. It is an interesting exercise to compute it for a rele-
vant perturbation with β(g) = ε g+· · · or for a marginally relevant/irrelevant perturbation
β(g) = cg 2 + · · · . See the exercise Section.
By construction the integrated form of the Callan-Symanzik equation is that given
above. Let us detailed it in the case of a two-point function of scalar fields and for a scalar
matrix of anomalous dimension with only one coupling constant (i.e. there is no mixing
and γ(g) is a pure number). Let G be the two-point function:
G(|x − y|; g) = hΦ(x)Φ(y)iR
g.
In this case the Callan-Symanzik equation reads (with r = |x − y|)

[r∂r + 2γ(g) − β(g)∂g ]G(r; g) = 0.
Its solutions are of the form
G(r; g) = Z 2 (g) Ĝ(m(g)r),
Rg γ(g 0 )
dg 0
with Z(g) solution of β(g)∂g Z(g) = γ(g)Z(g), explicitly Z(g) = e β(g0 ) . It is called
the wave function renormalization factor. And m(g) clearly defines a mass scale. It is by
construction RG covariant in the sense that
G(r/λ; g(λ)) = Ẑ 2 (λ) G(r; g), for all λ,
with Ẑ(λ) = Z(g(λ)) solution of λ∂λ log Ẑ(λ) = γ(g(λ)). This equation allows to decipher
the IR and/or UV behaviours.
Simple, but important, examples of beta functions and solutions of the Callan-Symanzik
equations are detailed in the Exercise Section.
173
9.7 Perturbatively renormalized φ4 theory

We now look at the perturbative RG analysis of the φ4 theory. For 3 < D < 4 the two
operators φ2 , φ4 are relevant and (∇φ)2 is marginal. We have to included them in the
action which then reads:
Z
z0 m0 2 g0 4
S = S∗ + dD x (∇φ)2 + φ + φ ,
2 2 4!
with S∗ the Gaussian massless action S∗ = 21 dD x (∇φ)2 . Alternatively, we may write

R
Z
1 1 1
S = dD x Z0 (∇φ)2 + m0 φ2 + g0 φ4

2 2 4!
We have introduced a cut-off that we are going to represent as a momentum cut-off Λ =
1/a. To renormalized the theory we have to find how to make all coupling constants Λ-
dependent such that the limit Λ → ∞ exists. That is we have to find the functions Z0 (Λ),
m0 (Λ) and g0 (Λ) in such way to cancel all possible divergencies in the limit Λ → ∞.
Let G(N ) (x1 , · · · , xN ) be the N -point functions hφ(x1 ) · · · φ(xN )i. The functions Z0 (Λ),
m0 (Λ) and g0 (Λ) are such the N -point correlation functions
N/2
Z0 (Λ) G(N ) ({xj }; g0 (Λ), m0 (Λ), Λ), finite
have a finite limit when Λ → ∞ (here Z0 , g0 and m0 are implicit functions the cut-off Λ).
To make these finite we have to identify physical parameters and re-express Z0 , g0 and
m0 in terms of these physical parameters.
• Perturbative one-loop renormalization in D = 4.
Let us do it explicitly, to first order (i.e. one-loop) in dimension D = 4. All correlation
functions are computed perturbatively using Wick’s theorem. We regularized them by
introducing a momentum cut-off Λ. We just quote the results when needed. At one-loop
the two-point functions is (recall that Γ(2) (p) = G(2) (p)−1 ): [... Draw the diagrams...]
g0
Γ(2) (p) = p2 + m0 + I2 (m0 ; Λ) + O(g02 ),
2
where
dD p Λ2 m2 Λ2
Z
1
I2Λ (m) = = − log( ) + O(1).
|p|<Λ (2π)D p2 + m2 (4π)2 (4π)2 m2
This divergence may be cancelled by setting
g0 Λ2 m2 Λ2
m2 = m20 + − log( ) + O(g02 ),
2 (4π)2 (4π)2 µ2
where µ is an arbitrary mass scale. Since I2 does not depend on the external momen-
tum, there is no p-dependent divergencies to cancel and hence there is not need for a
renormalization of Z0 , i.e.
Z0 = 1 + O(g02 ).
174
To compute the four point function is a bit more complicated. According to the pertur-
bative Feynman graphs, its structure is: [... Draw the diagrams...]
g02 h Λ i
Γ(4) (pj ) = g0 − I4 (p1 + p2 ; m0 ) + I4Λ (p1 + p3 ; m0 ) + I4Λ (p1 + p4 ; m0 ) + O(g03 ).
2
The function I4Λ (p; m) are given by a Feynman diagram and can be computed to be (See
the Exercise Section)
dD k Λ2
Z
1 1
I4Λ (p; m) = D 2 2 2 2
= 2
log( 2 ) + O(1).
|k|<Λ (2π) (k + m )((p + k) + m ) (4π) m
This divergence may be cancelled by setting
3 1 Λ2
g = g0 − g02 log( ) + O(1) + O(g03 ).
2 (4π)2 µ2
where again we have introduced the arbitrary mass scale. Notice the possible extra O(1)
term which is related to the way we choose to defined the physical coupling constant. The
statement of renormalizability is now that if m0 , g0 are expressed in terms of m and g,
then all correlation functions are finite in the limit Λ → ∞ (at m and g fixed).
The procedure becomes simpler (especially at higher orders) if we adopt a slightly
different strategy. We start as above from the action with Z0 , m0 and g0 but we write
m20 = m2R + δm2 , g0 = gR + δg, Z0 = 1 + δz,
and we treat the term δm2 , δg and δz as perturbation, so that the Feynman propagator
contains the mass m. The corrections δm2 , δg and δz are perturbatively computed to
cancel the divergences. The 2 and 4 point functions are then (at one-loop order)
g
Γ(2) (p) = p2 + mR + I2Λ (mR ) + δm2 + p2 δz + O(gR 2
),
2
g2
Γ(4) (pj ) = gR + − R I4Λ (p1 + p2 ; mR ) + I4Λ (p1 + p3 ; mR ) + I4Λ (p1 + p4 ; mR ) + δg

2
3
+ O(gR ).
As above, the 2-point function is made finite (renormalized) by setting

2
δz = 0 + O(gR ),
gR Λ2 m2R Λ2
δm2 = − + log( 2
) + O(gR ).
2 (4π)2 (4π)2 µ2
(2)
The renormalized 2-point functions ΓR is defined by the limit Λ → ∞ at m2R and gR
(2)
fixed. It is ΓR (p) = p2 + m2phys + O(gR
2 ) with the physical mass is (the p-dependance is
more complicated at higher order):
m2R m2R
m2phys = m2R + gR log( 2
) + O(gR ).
2(4π)2 µ2
175
The 4-point function is renormalized by setting as above (we do not include the extra O(1)
term because it is included in choosing m2R not egal to m2phys )
2 3 Λ2 3
δg = gR log( ) + O(gR ).
2(4π)2 µ2
The renormalized 4-point function is similarly defined by the limit Λ → ∞ at m2R and
(4) g2 P
gR fixed. It is (of course) finite and ΓR (pj ) = gR − 2R 4k=2 I4R (p1 + pk ; mR ) + O(gR
3 ).
It may be checked (by power counting) that this is enough to make all N -point function
Γ(N ) finite at one-loop. The renormalized N -point of course dependent on gR and mR but
also on the arbitrary scale µ.
• Renormalization of the φ4 theory.
The renormalizability of the φ4 theory is that this construction extends to arbitrary
higher order. The logic is the same as above. We look for three functions Z0 (Λ), m0 (Λ)
N/2
and g0 (Λ) such the N -point correlation functions Z0 G({xj }; g0 , m0 , Λ) have a finite
limit when Λ → ∞ (here Z0 , g0 and m0 are implicit functions the cut-off Λ). Since there
is three functions, fixing them requires imposing three conditions: two conditions involve
the two-point functions for Z0 and m0 and the third involve the 4-point function for g0 at
some reference momenta. For instance, we can choose
Γ(2) (p)|p2 =−mR = 0;

∂Γ(2) (p)
|p2 =µ2 = 1;
∂p2
Γ(4) (pi )|pi =pref = gR , |pref ref 2 2
1 + p2 | = µ .
i
These conditions are parametrized by the renormalized coupling constant mR and gR .

The need for references momenta introduce the reference scale µ, called the renormaliza-
tion scale. There are some freedom in choosing these renormalization conditions. Spec-
ifying them specify a (so-called) ‘renormalization scheme’. Changing them amounts to
reparametrization of the renormalized coupling constants (and the theory is covariant un-
der those reparametrization). The renormalized correlation functions are then defined
as
(N ) N/2
GR ({xj }; mR , gR , µ) = lim Z0 G(N ) ({xj }; g0 , m0 , Λ).
Λ→∞
This is analogue of the equation above defining renormalized correlation functions from
the lattice data, with Λ playing the role of the (inverse of the) lattice cut-off a and the
renormalization scale µ that of the scale `R . To prove that this procedure works to all
orders is of course much more complicated than the one-loop computation we did!
The renormalization scale is arbitrary (as was `R in the previous section) and there is
an equation coding for the covariance under changes of µ. It follows by writing that the
‘bare’ correlation functions G(N ) ({xj }; g0 , m0 , Λ) are independent of µ. Indeed the wave
function renormalization Z0 dependence of the parameter and we may choose to express
it as function of mR , gR and µ. Then we have
h i
−N/2 (N )
µ∂µ Z0 (mR , gR , µ) GR ({xj }; mR , gR , µ) = 0.
176
This is the Callan-Symanzik equation we discussed in previous section.

• Back to the effective potential
The renormalization procedure can also be implemented —or read— on the effective
potential. Recall that in a previous Chapter we computed the effective potential for the
φ4 theory. We got, to one-loop order, that (see the formula eq.(62) in which we set ~ = 1)
1 1 1 2 h V 00 (ϕ) i
eff 00
V1−loop (ϕ) = AΛ ϕ2 + BΛ ϕ4 + V (ϕ) log ,
2! 4! (8π)2 µ2
m20 2 g0 4
with V (ϕ) = 2 ϕ + 4! ϕ , the bare potential, and
g0 Λ2 m20 Λ2
AΛ = m20 + − log( ) + O(g02 ),
2 (4π)2 (4π)2 µ2
3 Λ2
BΛ = g0 − g02 log( ) + O(g03 )
2(4π)2 µ2
All the diverging terms are in the terms of degree 2 or 4 in ϕ. They can thus be absorbed
in a renormalization of the mass and the coupling constant. So, let m20 = m2R + δm2 and
g0 = gR + δg. As above, the one-loop effective potential is made finite by choosing
gR Λ2 m2R Λ2
δm2 = − − log( 2
) + O(gR ),
2 (4π)2 (4π)2 µ2
2 3 Λ2 3
δg = +gR log( ) + O(gR )
2(4π)2 µ2
These are —of course— the exact counter-terms that we had to choose for renormalizing
the 2 and 4 point vertex functions at one loop. The renormalized one-loop effective
potential then reads
eff;R 1 00 VR00 (ϕ)

V1−loop (ϕ) = VR (ϕ) + V R (ϕ) log[ ],
(8π)2 µ2
1 2 2 1 4
with VR (ϕ) = 2! mR ϕ + 4! gR ϕ . By looking at the ϕ2 term we can read what the physical
mass is:
2 m2R
m2phys = m2R 1 + g R log + O(g 2
R .
)
(8π)2 µ2
It of course coincides with that found by looking at the 2-point vertex function.
In other words, the same procedure works for renormalizing the effective action —to
be more precise we should had looked at the complete effective action, not only at the
effective potential. This had to be expected because the effective action is the generating
function of the vertex functions.
9.8 Exercises
• Exercise 9.1: Explicit RG flows.
177
The aim of this exercise is to study simple, but important, examples beta functions
and solutions of the Callan-Symanzik equation.
(i) Consider a field theory with only one relevant coupling constant g and suppose that
its beta function is β(g) = yg.
Show that the RG flow, solution of λ∂λ g(λ) = β(g(λ)) is g(λ) = g1 λy .
Show that the RG mass scale, solution of β(g)∂g m(g) = m(g) is m(g) = m∗ g 1/y .
Consider the two point function G(r; g) of a scaling field Φ of scaling dimension ∆, i.e.
G(r, g) = hΦ(r)Φ(0)ig . Prove (using the Callan-Symanzik equation) that
G(r; g) = r−2∆ F (m(g)r),
with m(g) the RG mass scale defined above.
(ii) Consider a field theory with only one marginal coupling constant g and suppose that its
beta function is β(g) = cg 2 (c > 0 corresponds to marginally relevant, c < 0 to marginally
irrelevant).
Prove that the RG flow, solution of λ∂λ g(λ) = β(g(λ)) is g(λ) = gµ /(1 − cgµ log(λ/µ)).
Notice that gλ → 0+ , if c < 0, while gλ flows up if c > 0, as λ → ∞ (with gµ > 0 initially).
Prove that the RG mass scale, solution of β(g)∂g m(g) = m(g) is m(g) = m∗ e−1/cg .
Notice that this mass scale is non perturbative in the coupling constant.
Consider the two point function G(r; g) of a scaling field Φ whose matrix of anoma-
lous dimension is γ(g) = ∆ + γ0 g. Prove (using the Callan-Symanzik equation) that
G(r/λ; g(λ)) = Z(λ)2 G(r, g) with
Z(λ) = const. λ∆ [g(λ)]γ0 /c .
Deduce from this that, in the case marginally irrelevant perturbation (i.e. c < 0) and
asymptotically for r large,
G(r; ga ) ' const. r−2∆ [log(r/a)]γ0 /c .
This codes for logarithmic corrections to scaling.
• Exercise 9.2: Anomalous dimensions and beta functions.
(i) Prove the relation γασ (g) = Dδασ −∂α β σ (g) between the matrix of anomalous dimensions
and the beta functions.
(ii) Give two proofs of the formula γασ (g) = ∆α δασ +SD i g i Ciα
σ for the matrix of anomalous
P
i
dimensions to first order in perturbation theory (Here g are the perturbative coupling
constant and SD the volume of the D-dimensional unit sphere): one proof comes from using
the previous result, the second proof comes from analysing the perturbative expansion of
the correlation functions.
• Exercise 9.3: Renormalisation of φ3 in D = 6.
• Exercise 9.4: Current-current perturbations and applications.
• Exercise 9.5: Disordered random bound 2D Ising model.
178
10 Miscellaneous applications
10.1 The XY model

• The physics of the XY model and the Kosterlitz-Thouless transition
(...) Bla Bla (...)
The rest of this section is a problem whose aim is to study the XY model using
techniques from lattice statistical physics and from statistical field theory. The problem
is divided in three parts:
IA/IB- The XY model on a lattice;
II- The role of vortices in the XY field theory;
IIIA/IIIB- The XY field theory and the sine-Gordon model.
The lattice models we shall consider are defined on the two dimensional square lattice Λ
with mesh size a, Λ = (aZ)2 , that we shall view as embedded in the Euclidean plane.
statistical field theories we shall consider are defined in the 2D Euclidean space R2 .
• I- The XY model on a lattice
~i , on each site i of the
The XY model is a statistical spin model with spin variables S
~ 2
lattice Λ, which are two component unit vectors, Si = 1. The energy of a configuration
~ is defined as E[S]
~ = −P S ~ ~
[S] [ij] i · Sj where the sum runs over neighboor points on Λ.
Parametrising the unit spin vectors S ~i by an angle Θi defined modulo 2π, we write the
configuration energy as X
~ =−
E[S] cos(Θi − Θj ).
[ij]
P
dΘi
R Q
The partition function is Z = [ i 2π ] exp β [i,j] cos(Θi − Θj ) with β = 1/kB T the
inverse temperature.
• IA- High temperature expansion
The aim of this section is to study the high temperature (β 1) behavior of the XY
model. It is based on rewriting the Boltzmann sums in terms of dual flow Pvariables. inΘ
IA-1 Explain why we can expand e β cos Θ in series as e β cos Θ = I(β) 1+ n6=0 tn (β)e ,
where I(β) and tn (β) are some real β-dependent coefficients. We set t0 (β) = 1.
IA-2 By inserting this series in the defining expression of the partition function and by
introducing integer variables u[ij] on each edge [ij] of the lattice Λ, show that the partition
function can be written as Z = I(β)Ne · Ẑ with Ne the number of edges and
X Y
Ẑ = tu[ij] (β),
[u], [∂u=0] [ij]
where the partition sum is over all configurations [u] of integer edge variablesP
u[ij] such that,
for any vertex i ∈ Λ, the sum of these variables arriving at i vanishes, i.e. j u[ij] = 0.
179
Remark: The variables u are attached to the edge of the lattice and may be thought
of as ‘flow variables’. The condition that their sum vanishes at any given vertex is a
divergencePfree condition. The divergence at a vertex i of a configuration [u] is defined as
(∂u)i := j u[ij] .
IA-3 Let i1 and i2 be two points of Λ and hS ~i i be the two-point spin correlation
~i · S
1 2
function.
Explain why hS ~i i = Rehe−i(Θi1 −Θi2 ) i.
~i · S
1 2
Show that,
1 X Y
he−i(Θi1 −Θi2 ) i = · tu[ij] (β),
Ẑ [u] [ij]
[∂u=δ·;i1 −δ·;i2 ]
where the sum is over all integer flow configurations such that their divergence is equal to
+1 at point i1 , to −1 at point i2 , and vanishes at any other vertex.
n
IA-4 Show that tn (β) = t−n (β) ' 2βn n! as β → 0.
Argue, using this asymptotic expression for the tn (β)’s, that the leading contribution to
the spin correlation functions at high temperature comes from flow configurations with
u = 0 or u = ±1 on each edge of the lattice.
IA-5 Deduce that, at high temperature, the correlation function hS ~i · S
~i i decreases
1 2
exponentially with the distance between the two points i1 and i2 .
Show that the correlation length behaves as ξ ' a/ log(2/β) at high temperature.
• IB- Low temperature expansion
The aim of this section is to study the low temperature (β 1) behavior of the XY
model. It consists in expanding the interaction energy cos(Θi − Θj ) to lowest order in the
angle variables so that we write the configuration energy as (up to an irrelevant additive
constant)
~ = const. + 1
X
E[S] (Θi − Θj )2 + · · · .
2
[i,j]
This approximation neglects the 2π-periodicity of the angle variables.

IB-1 Argue that the higher order terms in this expansion, say the terms proportional
to [i,j] (Θi − Θj )4 , are expected to be irrelevant and can be neglected.
P
IB-2 Write the expression of the partition function Z of the model within this approx-
imation.
Explain why, in this approximation, the theory may be viewed as a Gaussian theory.
IB-3 Let Gβ (x) be the two-point function of this Gaussian theory. Show that Gβ (x) =
−1
β G(x) with
Z +π/a
d2 p eip·x
G(x) = 2
,
−π/a (2π/a) 4 − 2(cos ap1 + cos ap2 )
with p1 , p2 the two components of the momentum p and a the lattice mesh.
IB-4 Let i1 and i2 be two points on Λ and x1 and x2 be their respective Euclidean
positions. Let Cα (x1 , x2 ) = heiα(Θi1 −Θi2 ) i with α integer. Show that
2

− αβ G(0)−G(x1 −x2 )
Cα (x1 , x2 ) = e .
180
IB-5 Explain why G(x) is actually IR divergent7 and what is the origin of this diver-
gence, but that G(0) − G(x) is finite for all x. Show that
1
G(0) − G(x) = log(|x|/a) + const. + O(1/|x|).
2π
IB-6 Deduce that the correlation functions Cα decrease algebraically at large distance
according to
2
Cα (x1 , x2 ) ' const. (a/|x1 − x2 |)α /2πβ .
Compare with the high temperature expansion.
• II- The role of vortices in the XY field theory
The previous computations show that the model is disordered at high temperature
but critical at low temperature with temperature dependent exponents. The aim of this
section is to explain the role of topological configurations, called vortices, in this transition.
We shall now study the model in continuous space, the Euclidean plane R2 , but with
an explicit short distance cut-off a. We shall consider the XY system in a disc of radius
L.
In the continuous formulation, the spin configurations are then maps Θ from R2 to
[0, 2π] modulo 2π. The above Gaussian energy is mapped into the action
Z
κ
S0 [Θ] = d2 x(∇Θ)2 ,
2
with a coefficient κ proportional to β.
II-1 Argue that the coefficient κ cannot be absorbed into a rescaling of the field variable
Θ?
II-2 A vortex, centred at the origin, is a configuration such that Θ± v (z) = ±Arg(z),
with z the complex coordinate on R2 , or in polar coordinates8 , Θ± v (r, φ) = ±φ.
± 2 ±
Show that HΘv is an extremum of S0 in the sense that ∇ Θv = 0 away from the origin.
Show that C0 dΘ± v = ±2π for C0 a small contour around the origin.
II-3 Let a0 be a small short distance cut-off and let D(a0 ) be the complex plane with
small discs of radius a0 around the vortex positions cut out. Prove that, evaluated on Θ± v,
the action S0 integrated over D(a0 ) (with an IR cut-off L) is
Z
(1) κ
d2 x (∇Θ± 2

Svortex = v ) = πκ log L/a0 .
2 D(a0 )
Give an interpretation of the divergence as a0 → 0.

II-4 What is the entropy of single vortex configurations? Show that the contribution
of single vortex configurations to the free energy is
(1) L 2 −πκ log[L/a0 ]
e−Fvortex ' const. e
a0
7
So that, when defining G(x), we implicitly assumed the existence of an IR cut-off, say |p| > 2π/L with
L the linear size of the box on which the model is considered.
8
We recall the expression of the gradient in polar coordinates: ∇Θ = (∂r Θ, r1 ∂φ Θ). The Laplacian is
∇ F = r1 ∂r (r∂r )F + r12 ∂φ2 F .
2
181
Conclude that vortex configurations are irrelevant for πκ > 2 but relevant for πκ < 2.
• III- The XY field theory and the sine-Gordon model
The aim of this section is to analyse this phase transition using renormalization group
arguments via a mapping to the so-called sine-Gordon field theory.
(M )
We shall consider a gas of vortices. The field configuration Θv for a collection of
M vortices of charges qa centred at positions xa is given by the sum of single vortex
configuration:
M
X
Θ(M
v
)
= qa Arg(z − za ).
a=1
We shall admit that the action of such configuration is
(M ) κ X |xa − xb | κ X 2 L X
Svortex = −2π( ) qa qb log + 2π( )( qb ) log + βc ,
2 a0 2 a0 a
a6=b b
where c is a ‘core’ energy (which is not taken into account by the previous continuous
description).
• IIIA- Mapping to the sine-Gordon theory
This mapping comes about when consideringPa gas of pairs of vortices of opposite
−
charges ±, so that the vortex system is neutral ( a qa = 0). We denote x+ j (resp. xj )
the positions of the vortices of charge + (resp. −).
The vortex gas is defined by considering all possible vortex pair configurations (with
arbitrary number of pairs) and fluctuations around those configurations. We set Θ =
(2n)
Θv + θsw and associate to each such configuration a statistical weights e−S with action
given by
(2n) −
S = Svortex [x+
j , xj ] + S0 [θsw ],
with S0 [θsw ] the Gaussian action κ2 d2 x(∇θsw )2 . We still assume a short-distance cut-off
R
a.
(2n) −
IIIA-1 Write the expression of the action Svortex [x+ j , xj ] for a collection of n pairs of
vortices at positions x± j , j = 1, · · · , n.
IIIA-2 Argue that the partition function of the gas of vortex pairs is given by the
product Z = Zsw × Zvortex with Zsw the partition function for the Gaussian free field θsw
and
n n + + 2πκ (|x− − x− |/a)2πκ
Q
i<j (|xi − xj |/a)
X µ2n Z Y Y i j
2 + 2 −
Zvortex = × ( d xj d xj ) + − ,
n! · n! 2πκ
Q
n≥0 j=1 j=1 i,j (|xi − x j |/a)
πκ −β
with µ = aa0 e c.
IIIA-3 The aim of the following questions R is2 to express Zvortex as a path integral over
1 2
an auxiliary bosonic field ϕ. Let S̃κ [ϕ] = 2κ d x(∇ϕ) be a Gaussian action. Show that,
computed with this Gaussian action,
1
hei2πϕ(x) e−i2πϕ(y) iS̃κ = .
|x − y|2πκ
182
κ

Hint: The Green function associated to the action S̃κ [ϕ] is G(x, y) = − 2π log |x − y|/a .
IIIA-4 What is the scaling dimension (computed with the Gaussian action S̃κ [ϕ]) of
the operators (∇ϕ)2 and cos(2πϕ)?
Deduce that the perturbation cos(2πϕ) is relevant for πκ < 2 and irrelevant for πκ > 2.
Is the the perturbation (∇ϕ)2 relevant or irrelevant?
IIIA-5 Show that Zvortex can be written as the partition function of Gaussian bosonic
field with action SsG [ϕ], Z
Zvortex = [Dϕ] e−SsG [ϕ] ,
where the action SsG is defined as

Z h1 i
2 2
SsG [ϕ] = d x (∇ϕ) − 2µ cos(2πϕ) .
2κ
This is called the sine-Gordon action.
Hint: Compute perturbatively the above partition function as a series in µ while paying
attention to combinatorial factors.
• The renormalization group analysis
IIIB-1 We now study the renormalization group flow of the action SsG for κ close to
the critical value κc = 2/π. We let κ−1 = κ−1c − δκ and write
Z h1 i
SsG [ϕ] = S̃κc [ϕ] − d2 x (δκ)(∇ϕ)2 + 2µ cos(2πϕ)
2
Show that, to lowest order, the renormalization group equations for the coupling con-
stants δκ and µ are of the following form:
˙ = `∂` (δκ) = b µ2 + · · ·
(δκ)
µ̇ = `∂` µ = a (δκ) µ + · · ·
with a and b some positive numerical constants.

Hint: It may be useful to first evaluate the OPE of the fields (∇ϕ)2 and cos(2πϕ).
√
IIIB-2 We redefine the coupling constants and set X = a (δκ) and Y = ab µ such
that the RG equations now reads Ẋ = Y 2 and Ẏ = XY .
Show that Y 2 − X 2 is an invariant of this RG flow.
Draw the RG flow lines in the upper half plane Y > 0 near the origin.
IIIB-3 We look at the flow with initial condition XI < 0 and YI .
Show that if YI2 − XI2 < 0 and XI < 0, then the flow converges toward a point on the line
Y = 0.
Deduce that for such initial condition the long distance theory is critical. Compare with
section I-B.
IIIB-4 Show that if YI2 − XI2 > 0 and XI < 0, the flow drives X and Y to large values.
Let Y02 = YI2 − XI2 with Y0 > 0. Show that the solution of the RG equations are
` 1h X(`) XI i
log = arctan − arctan .
a Y0 Y0 Y0
183
IIIB-5 The initial condition XI and YI are smooth functions of the temperature T of
the XY model. The critical temperature Tc is such that XI + YI = 0. We take the initial
condition to be near the critical line XI + YI = 0 with XI < 0. We let XI = −YI (1 + τ )
in which τ 1 is interpreted at the distance from the critical temperature: τ ∝ (T − Tc ).
For τ > 0, we define the correlation length as the length ξ at which X(`) is of order 1.
Why is this a good definition?
Show that √
ξ/a ' const. econst./ τ .
IIIB-6 Comment and discuss.
10.2 Self avoiding walks

(...) Bla Bla (...)
The rest of this section is a problem whose aim is to study the self-avoiding walks using
techniques from lattice statistical physics and from statistical field theory. This problem
deals with statistical properties of random curves (say polymers) and their connections
with statistical field theory. It contains two parts (which are almost independent):
I- Non-interacting paths and their scaling limit;
II- Self-avoiding walks and the O(n → 0) model.
All lattice models we shall consider are defined on the D dimensional square lattice Λ
with mesh size a: Λ = (aZ)D . We let ej , j = 1, · · · , D be a basis of orthornormal vectors
in RD , so that points x ∈ Λ are x = a D
P
j=1 nj ej with nj integers.
A path Γ in Λ is an ordered collection of neighbour points in Λ, i.e. Γ = (x0 , · · · , xi , · · · , x|Γ| )
with xi , xi+1 neighbour on the lattice. We let |Γ| be the number of bonds of the path Γ,
so that a|Γ| is its length. A loop in Λ is a closed path with x0 = x|Γ| .
A self avoiding walk Γsaw is a path such that all xi are distinct9 .
The statistical field theories we shall consider are defined in RD .
• I- Non-interacting paths and their scaling limit
Non-interacting paths are random paths whose statistics is specified by assigning a

Boltzmann weight wΓ := µ|Γ| to each path Γ, with µ a real number (µ > 0). Let x ∈ Λ.
We shall deal with paths Γ starting at the origin 0 and ending at point x, i.e. paths
Γ = (x0 , · · · , xi , · · · , x|Γ| ) with x0 = 0 and x|Γ| = x. Let Z(x) be the partition function
conditioned on paths from 0 to x:
X
Z(x) = µ|Γ| .
Γ: 0→x
The probability of a given path Γ from 0 to x is µ|Γ| /Z(x).
• IA: The discrete lattice model

9
Hence, none of the points of a self-avoiding walk Γsaw is visited twice by Γsaw .
184
IA-1: Show that

X X
Z(x) = µN WNfree (x) = δx;0 + µN WNfree (x),
N ≥0 N >0
with WNfree (x) the number of paths from 0 to x with N bonds.

IA-2: Show (without long computation) that Z(x) satisfies the following difference
equation:
D
X
j=1
IA-3: Compute the Fourier transform of Z(x) and give an explicit expression of Z(x)
as an integral over the Brouillon zone.
IA-4: Let ∆dis. be the discrete Laplacian and write ∆dis. = Θ − 2D I with Θ the
lattice adjacency
P matrix and I the identity matrix. We view Θ as acting on functions via
(Θ · f )(x) = D j=1 f (x + aej ) + f (x − aej ) . Show that:
1
Z(x) = h0| |xi,
I − µΘ
with |xi the δ-function peaked at x, i.e. hy|xi = δy;x .

IA-5: Prove that there exists a critical µc such that Z(x) converges for |µ| < µc but
diverges as µ → µc . What is the value of µc ?
• IB: The scaling limit
We now study the continuous limit a → 0 of the above discrete model.

IB-1: Show that, for µ and x fixed, lima→0 a−D Z(x) = const. δ(x), with an explicit
formula for the ‘const.’ to be determined. Give a geometrical interpretation of this result.
The continuous limit is thus meaningful only if we scale µ appropriately as a → 0. Let

G(x) be defined by
1
G(x) = lim a2−D Z(x),
2D a→0
where the limit is understood with µ an appropriate function of a to be determined below.
IB-2: Show that

− ∆x + m2 G(x) = δ(x),

with ∆x the D-dimensional Laplacian, if µ approaches µc such that µ−1 − µ−1 2 2

c = a m as
a → 0 with m fixed.
IB-3: Verify that the scaling limit – i.e. the limit a → 0, µ → µc with a−2 (µ−1 −µ−1
c )=
m2 fixed – of the explicit expression of Z(x) found above is a solution of this equation.
185
IB-4: Show that Z(x) can be written in the scaling form Z(x) ' (ma)D−2 F (mx), for
some function F , when µ → µc at mx fixed.
• II- Self-avoiding walks and the O(n → 0) models
This part develops the connection between self-avoiding walks and a specific O(n → 0)
lattice model on one hand, and between large distance properties of self-avoiding walks
and an O(n → 0) field theory10 on the other hand.
The O(n) lattice model is defined as follows: a spin variable S ~x , of dimension n, is
attached to each lattice site x ∈ Λ with components a
Sx , a = 1, · · · , n. The statistical
Q
Boltzmann weight of a spin configuration is hx,x0 i (1 + µ S ~x · S
~x0 ) where the product is
over neighbour sites (i.e. product over lattice edges) and µ a real number. The sum
over the spin configurations is represented by an integration with measure d[S] on spin
configurations so that the partition function Zn is
Z Y
Zn = d[S] ~x · S
(1 + µ S ~x0 ),
hx,x0 i
We choose the measure to be factorized over lattice points, d[S] = x dm(S

Q ~x ), so that
spins at different positions are uncorrelated under the measure d[S] and any integral of
a
products of spins (Sxa11 · · · Sxqq ) factorizes into products of integrals over spins localized at
the same position. The measure and spin variables are normalized such that
Z Z
d[S] (Sxa Syb ) = δ ab δx;y and d[S] = 1.
We shall
R further assume that all higher moments of spins at a given point x vanish11 so
a
that d[S] (Sxa1 · · · Sxp ) = 0 as soon as p 6= 2.
The O(n) field theory we shall consider is a Φ4 -theory whose action is given below.
• IIA: The O(n) lattice model and self-avoiding walks

Q
IIA-1: By expanding the product hx,x0 i (1 + µ S ~x · S~x0 ) into a sum of clusters of
occupied edges (declaring that an edge is occupied if it is weighted by µ), show that Zn
can be written as a sum over non-crossing loop configurations12 of the form
X
Zn = n`(C) µ|C| ,
C: non crossing loops
where |C| is the total number of bonds forming C and `(C) the number of connected
components (i.e. the number of loops).
Hint: It may help to draw a picture of the lattice (say in D = 2) representing each
term of this expansion with the edges occupied marked (say with a bold line) and to look
at which of those graphs survive after integration over the spin variables.
10
This connection was first revealed by P.G. de Gennes.
11
This is actually an approximation, but it can be shown that higher moments vanish as n → 0, so that
we can (hopefully legitimately) neglect them in the limit n → 0 we are interested in.
12
A loop is said to be non-crossing if all of its edges and vertices are distinct.
186
This formula shows that we can define Zn for any number n and any fugacity µ by
analytic continuation13 . We adopt this definition from now on.
IIA-2: Show that Zn→0 = 1.
We now consider the spin correlation function hS0a Sxb i between spins at the origin 0
and a point x in Λ. It is defined by:
Z
a b 1 Y
~y · S
~y0 ) (S a S b ).
hS0 Sx i = d[S] (1 + µ S 0 x
Zn 0 hy,y i
Global O(n)-invariance implies that hS0a Sxb i = δ ab Gn (x).

IIA-3: Write Gn (x) as a weighted sum of non-crossing loops C in Λ and a self-avoiding
walk Γsaw from 0 to x:
X
Gn (x) = n`(C) µ|C|+|Γsaw | .
C+(Γsaw :0→x)
IIA-4: Show that the analytical continuation of this two point

Pfunction can be written
as a sum over self-avoiding walks Γsaw from 0 to x: Gn→0 (x) = Γsaw : 0→x µ|Γsaw | .
Show that, as a consequence,
X
Gn→0 (x) = µN WNsaw (x)
N ≥0
with WNsaw (x) the number of self-avoiding walks with N bonds from 0 to x.
• IIB: An O(n) statistical field theory
We now consider the O(n)-invariant Φ4 field theory in RD with action

Z
1
S = dD x (∇Φ) ~ 2 + 1 µ(Φ ~ + 1 g(Φ
~ · Φ) ~ · Φ)
~ 2 .

2 2! 4!
~ has n components Φa , a = 1, · · · , n.
The field Φ
IIB-1: Explain why symmetry and universality arguments allow us to restrict to a
~ ~ + (Φ
(Φ · Φ) ~ · Φ)
~ 2 interaction.
IIB-2: What are the Feynman rules (in momentum space) for this theory?
IIB-3: Compute the beta functions βµ and βg at one loop (i.e. at order g 2 , µ2 and gµ)
for the mass µ and coupling constant g renormalization and show that :
βµ = 2µ − 8(n + 2)gµ − 4µ2 − 32(n + 2)g 2 + · · ·

βg = (4 − D)g − 8(n + 8)g 2 − µ2 − 16gµ + · · ·
13
Positivity of the Boltzmann weights is then no guaranteed.
187
Hint: It may be useful to first evaluate the OPE of the perturbing fields φ2 =: (Φ ~ · Φ)
~ :
~ · Φ)
and φ4 =: (Φ ~ :.
2
Assume that the renormalization of (∇Φ) ~ 2 can neglected at this order.

Remark: Only the first coefficients (proportional to µ and to gµ in βµ and to g and
2
g in βg are going to be relevant. If time is lacking, you may restrict in proving that
βµ = 2µ − 8(n + 2)gµ + · · · and βg = (4 − D)g − 8(n + 8)g 2 + · · · , and argue that the other
terms are indeed irrelevant.
IIB-4: Show that there is a pertubative IR fixed point in dimension D = 4 − at
µc = 0 + O(2 ) and gc = g∗ + O(2 ) for some g∗ .
What is the value, to first order in , of the correlation length scaling exponent ν
defined by ξ ' |µ − µc |−ν with ξ the correlation length?
• IIC: Scaling theory of self-avoiding walks
Based on universality arguments we may claim that the large distance behaviour of
self-avoiding walks is encoded into the O(n → 0) field theory. We make this hypothesis in
the following.
IIC-1: Let WN (x) be the number of walks with N bonds, free or self-avoiding, from
0 to x: W free saw
PN (x) =NWN (x) for free paths, WN (x) = WN (x) for self-avoiding walks. Let
W (x) = N ≥0 µ WN (x). Show that
I
dµ −N −1
WN (x) = µ W (x),
2iπ
where the integration is over a small contour around the origin.
Note: We shall admit the following technical result: If a series f (µ) = N >0 µN cN is
P
analytic for |µ| < µc but singular at µc with f (µ) ' const. (µc − µ)−γ−1 as µ → µc , then
cN ' const. N γ µ−N
c for N large.
IIC-2: Explain why the hypothesis of a scaling theory implies that, for µ → µc , the
function Z(x) behaves as
W (x) ' ξ −2∆ F (|x|/ξ),
with ξ the correlation length and F a smooth function. What is the meaning of ∆?
How does the correlation length ξ depend on µ close to the critical point µc ?
IIC-3: Let hR2 iN be the mean square distance defined by
2
P
2 x |x| WN (x)
hR iN = P .
x WN (x)
Show that if the scaling hypothesis applies then
hR2 iN ' const. N 2ν
for N large, with ν the correlation length exponent.

Hint: It may be useful to consider both generating functions x W (x) and x |x|2 W (x).
P P
188
IIC-4: What is the value of ν for free paths?

What is the value of ν for self-avoiding walks in dimension D = 4 − to first order in
, assuming that the O(n → 0) field theory describes the scaling theory of self-avoiding
walks?
Compare this value with the exact value ν = 3/4 in D = 2 and the approximate
numerical value ν ' 0.588 · · · in D = 3.
IIC-5: The fractal dimension Dh of a set embedded in a metric space may be de-
fined through the minimal number Nr of boxes of radius r needed to cover it by Dh =
limr→0 log Nr / log(1/r).
Use the previous scaling theory to find an estimate of the fractal dimension of self-
avoiding walks in dimension D = 4 − .
10.3 The coulomb gas

(...) Bla Bla (...)
10.4 The stress tensor and diffeomorphisms

(...) Bla Bla (...)
The aim of this problem is to study some properties of the stress-tensor in free field
theories an dits relation with diffeomorphisms, and to use them to derive the transforma-
tion law of the Green function of the Laplacian under domain deformations, see eq.(78)
below.
The statistical field theories we shall consider are defined in a subdomain of the Eu-
clidean space RD . The metric on RD is denoted dx2 = gµν dxµ dxν , with xµ , µ = 1, · · · , D,
the Euclidean coordinates of a point x ∈ RD .
We consider a massive free scalar field theory over a finite domain D ⊂ RD whose
action is Z
1 1
dD x (∇φ(x))2 + m2 φ(x)2 ,

SD [φ] :=
D 2 2
with Dirichlet boundary conditions : namely, φ(x) = 0 for x ∈ ∂D, or alternatively
φ|∂D = 0. We shall assume that the domain D has the topology of a sphere, that is: D is
a smooth connected and simply connected domain.
We let n(x), with components nµ (x) normalized to n2 = 1, be the outwards normal of
D at x ∈ ∂D. We denote by t(x), with components tµ (x), any vector tangent to ∂D at a
point x on the boundary of D.
The aim of this problem is to compare this field theory in two different domains D and
D0 .
We shall consider infinitesimal diffeomorphisms, x ∈ D → x0 ∈ D0 with x0 = x+ ξ(x)+
· · · , where ξ(x) is a vector field, with components ξ µ (x), and 1 a small parameter.
We recall that the transformation rules for a scalar field are φ(x) → φ0 (x0 ) = φ(x(x0 )).
• IA- Diffeomorphisms fixing the domain
189
We first consider the case where D = D0 , so that ξ is tangent to ∂D at the boundary

and its normal component vanishes at the boundary: nµ ξ µ (x) = 0 for x ∈ ∂D.
IA-a Express the transformed field φ0 in terms of φ to first order in 1 (order
included).
IA-b Show that the variation of the action δ̂SD [φ] = SD [φ0 ] − SD [φ] is
Z
δ̂SD [φ] = − dD x Tµν (x)∇µ ξ ν (x) + O(2 ),
D
with Tµν the stress-tensor defined by

1
Tµν = ∇µ φ∇ν φ − gµν (∇φ)2 + m2 φ2 .

2
IA-c Let Tnt (x) := nµ tν Tµν (x) and Tnn (x) := nµ nν Tµν (x) with n the normal of D and
t any vector tangent to the boundary of D at a point x on the boundary. Show that
1
Tnt (x) = 0, Tnn (x) = (∇n φ(x))2 , for x ∈ ∂D,
2
with ∇n φ = nµ ∇µ φ.
IA-d Let GD (x, y) = hφ(x)φ(y)iD , for x, y ∈ D, be the scalar field two point function
in the domain D.
Give its path integral representation and explain why GD (x, y) is the Green function of
the massive Laplacian, that is
− ∇2x + m2 GD (x, y) = δ(x − y),

with δ(x) the Dirac measure at x.

What are the boundary conditions for GD (x, y)?
IA-e Show that the following identity is fulfilled:
Z h i
dD z ξ ν (z) h(∇µ Tµν (z)) φ(x)φ(y)iD − h(∇µ Tµν (z))iD hφ(x)φ(y)iD
= − ξµ (x)∇µx + ξµ (y)∇µy GD (x, y),

Give an interpretation of this formula.

Hint: Use Wick’s theorem to evaluate h(∇µ Tµν (z)) φ(x)φ(y)iD and properties of the Green
function to prove the above formula.
• IB- Diffeomorphisms deforming the domain
We now consider the case where D0 is a slight deformation of the domain D, so that
the normal component of the vector field ξ does not vanish at the boundary. We suppose
ξ µ to be normal to ∂D at the boundary (i.e. tµ ξ µ = 0 at the boundary for any tangent
vector t) and we set ξ µ (x) = ζ(x) nµ for x ∈ ∂D.
IB-a Let GD0 (x0 , y 0 ) = hφ0 (x0 )φ0 (y 0 )iD0 , for x0 , y 0 ∈ D0 , be the scalar field two point
function in D0 . Give its path integral representation and show that, to first order in ,
GD0 (x0 , y 0 ) − GD0 (x, y) = ξµ (x)∇µx + ξµ (y)∇µy GD (x, y) + O(2 ).

190
Please pay attention to the prime and un-prime symbols.

IB-b Let δSD [φ] = SD0 [φ0 ] − SD [φ] be the variation of the action to first order in
(again, please pay attention to the prime and un-prime symbols). Show that
Z
δSD [φ] = − dD x Tµν (x)∇µ ξ ν (x) + O(2 ).
D
Hint: Please, pay attention that there are two contributions to δSD [φ]: one arising from
the difference between φ0 and φ, the other from the change of variables from x0 to x.
IB-c Let ZD and ZD0 be the free field partition function in D and D0 .
Write a path integral representation of ZD .
Show that the variation of the partition function can be written as the sum of two integrals,
one on the bulk of D, the other over the boundary of D, namely:
Z Z
D ν µ
ZD0 = ZD 1 − d x ξ (x)h(∇ Tµν (x))iD + ds(x) ζ(x)hTnn (x)iD + · · · ,
D ∂D
with ds(x) the surface element on ∂D and Tnn (x) = nµ nν Tµν (x) with n the normal of D.
IB-d Using the previous results from IIA & IIB, show that the variation of the Green
function, δGD (x, y) := GD0 (x, y) − GD (x, y) can be written as
Z
δGD (x, y) = ds(z) ζ(z) hφ(x)φ(y)Tnn (z)iD − hφ(x)φ(y)iD hTnn (z)iD + O(2 ) ,
∂D
Please pay attention to the prime and un-prime symbols (the difference δGD (x, y) is not
the same as that considered in IIB-a).
IB-e Deduce that
Z
δGD (x, y) = ds(z) ζ(z) ∇zn GD (x, z) ∇zn GD (y, z) + O(2 ), (78)
∂D
This formula, which expresses the variation of the Green function under domain defor-
mations, in terms of the boundary integral of product of normal derivatives of the Green
function is called the Hadamard formula.
IB-f Let DR be a disc of radius R centred at the origin in the plane R2 . Recall that
the Green function for a massless free field with Dirichlet boundary condition on DR is
1 |z − w|R
GR (z, w) = − log ,
2π |z w̄ − R2 |
with z and w complex coordinates on the plane.

Verify the formula (78) when varying the radius of the disc.
Hint: The computation is going to be simpler by reducing the integral (78) to a complex
contour integral.
10.5 Irreversibility of RG flows and the c-theorem

(...) Bla Bla (...)
191
The aim of this problem is to study a few properties of RG flows in (generic) 2D

statistical field theories. RG fixed points correspond to conformal field theories (CFT).
Perturbing a CFT by relevant operators yields a theory with finite correlation lengths
which flows in the IR to another CFT.
In 1986, A. Zamolodchikov proves that, in the (appropriate) space of 2D renormalisable
field theories, there exists a function of the coupling constants which is monotonically non-
increasing along the RG flows and stationary at the fixed points. This encodes for the
irreversibility of RG flows.
The field theories we shall consider are those satisfying the ‘positive reflexion property’
which, in particular, claims that two point functions of scalar operators O(x) are positive,
that is: hO(x)O(y)i ≥ 0.
We shall denote by {g i } the set of coupling constants perturbing a CFT and by β i (g)
the beta functions associated to this perturbation. We recall that the RG flow equations
are λ∂λ g i (λ) = β i (g(λ)). We shall assume that such field theory generically admits a
stress-tensor.
We consider field theory in R2 . For (x1 , x2 ) the Euclidean coordinates, we let z =
x1 + ix2 be the complex coordinate and z̄ = x1 − ix2 its complex conjugate.
• IA- The stress-tensor in two dimension
Let Tµν = Tνµ be the stress-tensor of a generic statistical field theory. We recall that,
by covariance symmetry, this tensor is such that ∇µ Tµν = 0 inside correlation functions,
away from field insertions.
IA-a Taking the free field theory as an example, explain why Tµν transforms as a 2-
tensor under diffeomorphism, that is: under a change of coordinate x → y, it transforms
as Tµν (x) → Tασ (y) with
Tασ (y)dy α dy σ = Tµν (x)dxµ dxν .
IA-b Let T = Tzz , T̄ = Tz̄ z̄ and Θ = Tµµ = 4 Tz z̄ . Express T , T̄ and Θ in terms of the
stress-tensor components Tµν in an Euclidean frame.
IA-c How does T , T̄ and Θ transform under rotation z → eiθ z and z̄ → e−iθ z̄?
IA-d Show that the conservation law ∇µ Tµν = 0 translates into
1
∂z̄ T + ∂z Θ = 0.
4
IA-e Taking the free field theory as an example, explain why Tµν has scaling dimension
2.
What would be its scaling dimension in dimension D > 2?
We shall assume that, in any 2D generic statistical field theory, Tµν has scaling dimen-
sion 2 (without any anomalous dimension).
IA-f We consider the stress-tensor two point functions. Explain why translational plus
rotational symmetry dedicate that these correlation functions have the following form,
F (z z̄) G(z z̄) H(z z̄)

hT (z, z̄)T (0, 0)i = , hΘ(z, z̄)T (0, 0)i = 4 , hΘ(z, z̄)Θ(0, 0) = 16 ,
z4 z 3 z̄ z 2 z̄ 2
192
with F , G and H scalar functions, depending only on the distance r.
• IB- Irreversibility of 2D RG flows
We use the same notations as in IIA.

IB-a Explain why F , G and H are all dimensionless functions of both the distance r
and the coupling constants {g}.
IB-b Show that the stress-tensor conservation law implies
Ḟ + Ġ − 3G = 0, Ġ − G + Ḣ − 2H = 0,
where Ḟ = z z̄ F 0 (z z̄).
IB-c Let C = 2F − 4G − 6H. Show that Ċ ≤ 0.
IB-d Deduce from this inequality that, for fixed r0 ,
X ∂
β i (g) C(r0 , {g}) ≤ 0.
∂g i
i
Conclude that λ → C(r0 , {g λ }) is decreasing along the RG flows.

Comment and discuss this property.
IB-e What are the values of C at the IR and UV fixed points?
Give an interpretation of these values.
IB-f Show the sum rule
Z
3
CU V − CIR = d2 x |x|2 hΘ(x)Θ(0)i.
4π
IB-g Check this last relation for a free massive field theory.
Explain why CU V = 1/4π 2 .
What is the value of CIR ?
Give an interpretation of this result.
Hint: We recall that the Green function of the massive Laplacian in R2 is G(x) =
1
2π K0 (mr) with r = |x| and K0 the 0-th R Bessel function. At short distance G(x) =
1
− 2π log(mr) + · · · for mr 1. We give dR R3 K0 (R)2 = 1/3.
10.6 Exact RG
(...) Bla Bla (...)
10.7 Etc...
193
11 Fermionic techniques
11.1 Fermions and Grassmann variables
11.2 The 2D Ising model again
11.3 Dirac fermions
11.4 Chiral anomaly
11.5 Fermionisation-Bosonisation
11.6 Exercises
12 Stochastic field theories and dynamics

12.1 Et un...
12.2 Et deux...
12.3 Et trois...
12.4 Zero...
13 Exercise corrections
See the exercise booklet...
THE END
194

Lectures Stat Field Theory Vnew

Uploaded by

Copyright:

Available Formats

Lectures Stat Field Theory Vnew

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lectures Stat Field Theory Vnew

Uploaded by

Copyright:

Available Formats

Laboratoire de physique de l’École normale supérieure de Paris.

Centre national de la recherche scientifique, Institut de physique.

Statistical Field Theory and Applications :

Laboratoire de Physique de l’École Normale Supérieure,

Preliminary version : 2020.....

1 What are we aiming at describing ? 7

2 Brownian motions and random paths 13

3 Statistical lattice models 39

4 From statistical models to field theories 60

5 The renormalisation group and universality 80

6 Free field theory 91

6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

7 Interacting field theory: basics 109

8 Conformal field theory: basics 135

9 Scaling limits and

10 Miscellaneous applications 179

11 Fermionic techniques 194

11.5 Fermionisation-Bosonisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

12 Stochastic field theories and dynamics 194

13 Exercise corrections 194

An exemple of lecture series:

• Brownian motions and random paths

1 What are we aiming at describing ?

A sample of Samples of 2D self-avoiding walks,

Understanding random patterns is at the core of the comprehension of many physical

the more complete our predictions can be. See Figure.

A small sample of definition (left).

Magnetization versus temperature Heat capacity versus temperature

Statistical Field Theory is also applicable to statistical systems out-of-equilibrium in-

Vorticity clusters A sample of the 2D massless

A non-exhaustive list of references:

• L. Kadanoff, Statistical Physics: Statics, Dynamics and Renormalization, World

• J. Zinn-Justin, Quantum Field Theory and Critical Phenomena, Oxford University

• G. Parisi, Statistical Field Theory, Addison-Wesley Publishing Co., (1988).

• J. Cardy, Scaling and Renormalization in Statistical Physics, Cambridge University

• Ph. Di Francesco, P. Mathieu, D. Sénéchal, Conformal Field Theory, Springer-Verlag

• M. Kardar, Statistical Physics of Fields, Cambridge University Press, (2007).

• G. Mussardo, Statistical Field Theory: An Introduction to Exactly Solved Models in

2 Brownian motions and random paths

2.1 Random walks and random paths

Note that Qn is not a Gaussian variable but a sum of binomial variables.

2.2 Scaling limits

E[Wt ] = 0, E[Wt Ws ] = min(t, s).

with Wti , i = 1, · · · , D, independent identical distributed (i.i.d.) Brownian motions. The

• Scaling limit of asymmetric random walks

E[Qn ] = a(p − q)n ' ν t, E[Q2n ] − E[Qn ]2 = a2 4pq n ' t.

There is a constant drift, with velocity ν, in the sense that it satisfies

dXt := Xt+dt − Xt = ν dt + dWt ,

with Wt a Brownian motion. This type of noisy evolution equation is an example of

dimensional square lattice with mesh size a: Λ = (aZ)D . We let ej , jP= 1, · · · , D, be

This is a linear equation, involving the discrete Laplacian on Λ. As a consequence, it can

2.3 Brownian motions and path integrals

• Brownian motion: its transition kernel.

The Brownian process is indeed Gaussian with i.i.d. increments.

Pn [x0 + a; Ω] + Pn [x0 − a; Ω] − 2Pn [x0 ; Ω] ' a2 ∂x20 Pt [x0 ; Ω]

The probability Pt [x; Ω] is therefore solution of the heat equation:

2.4 The 2D Brownian motion

• Conformal invariance of the 2D Brownian motion.

Xt − X0 = (Xt1 − Xt0 ) + (Xt2 − Xt1 ) + · · · + (Xt − Xtj−1 ),

(Xtj+1 − Xtj ) → λj Rj · (Xtj+1 − Xtj ) ≡in law Xλ2 (tj+1 −tj )

E[F (Xn + δAn + ξn+1 )|Fn ]

Xn+1 − Xn = δ a(Xn ) + ξn+1 ,

dXt = a(Xt )dt + dWt ,

This exact rewriting of Z is called a high-temperature expansion, since w = tanh K 1

where the matrices Hh , Hv 1. In particular we have