Main

AQFT Lecture Notes February 15, 2023
Contents
1 Topic-I: Path Integrals 4
1.1 Idea of path integrals from quantum mechanics . . . . . . . . . . . . . . . . . . . 4
1.2 Path integral in quantum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Path integral in QFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 The Classical limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Correlators in the path integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Generating functional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.7 Schwinger-Dyson Equation in Canonical Quantization . . . . . . . . . . . . . . . 12
1.8 Schwinger-Dyson Equation in Path Integral Formalism . . . . . . . . . . . . . . . 14
1.9 Property of path integrals: time-ordering . . . . . . . . . . . . . . . . . . . . . . 15
1.10 Example: solving the free theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.11 Perturbation theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.12 Connected generating functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.13 1PI Green’s functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.13.1 Quantum effective action and its relation to W [J] . . . . . . . . . . . . . 21
1.13.2 The role of Γ[φ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.14 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2 Topic-II: Symmetries 28
2.1 Why symmetries? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2 Noether’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 The Ward-Takahashi Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 The energy-momentum tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 Spontaneous Symmetry Breaking . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5.1 Spontaneous breaking of discrete symmetries: a Z2 example . . . . . . . . 37
2.5.2 Spontaneous breaking of global symmetries and Goldstone’s theorem . . . 38
2.5.3 The linear sigma model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6 The Higgs mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.6.1 The Abelian Higgs model . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.7 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3 Topic-III: Wilsonian Renormalization Group (in Real Space) 45

3.1 The Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 RG on Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2.1 Step one: reducing the degrees of freedom . . . . . . . . . . . . . . . . . . 47
3.2.2 Step two: new effective Hamiltonian . . . . . . . . . . . . . . . . . . . . . 48
3.3 RG fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4 Critical surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
1
4 Topic-IV: Wilsonian Renormalization Group (in Momentum Space) 56
4.1 Integrating out degrees of freedom . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2 Running couplings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3 Anomalous dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4 Renormalization group trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.5 Counterterms and continuous limits . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.6 Polchinski’s equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.7 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5 Topic-V: Conformal Field Theory 68

5.1 Why should we care about conformal field theories? . . . . . . . . . . . . . . . . 68
5.2 Conformal transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3 Algebra of local fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.4 Crossing symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.5 Quasi-primary fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.6 Two-dimensional conformal transformations . . . . . . . . . . . . . . . . . . . . . 84
5.7 Primary fields and Ward identities . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.8 The free boson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.9 Energy-momentum tensor and the Virasoro algebra . . . . . . . . . . . . . . . . . 98
5.10 Ground state energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.11 Radial quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.12 State-operator correspondence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.13 Representation theory of conformal algebra . . . . . . . . . . . . . . . . . . . . . 109
5.14 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
2
Topic-I: Path Integrals
Instructor: Miranda C. N. Cheng

Notes taken by Naveen Balaji Umasankar
It’s a draft. Do not distribute. Corrections and comments are welcome!
3
1 Topic-I: Path Integrals
The path integral formulation of quantum field theory leads not only new computational tech-
niques but also conceptual insights into what quantum field theories are.
1.1 Idea of path integrals from quantum mechanics

The path integral is built on an action. In this picture, the evolution of a quantum mechanical
system is represented in terms of a sum over classical trajectories in the phase space. The
intuition for a path integral comes from a simple thought experiment we can do in quantum
mechanics, namely the double-slit experiment where the wave-function of the particle is the sum
of two interfering wave-functions.
Figure 1: (From left to right) The standard double-slit used in scattering experiments in QM,
replacing the vacuum with infinitely many slits and hence, infinitely many screens, and the
diagrammatic representation of a path integral with interference between infinitely many paths.
1.2 Path integral in quantum mechanics

Let us consider a single-particle Hamiltonian in 1-D with the following form
p̂2
Ĥ = + V (x̂), (1)
2m
where Ĥ, p̂ and x̂ are operators acting on the Hilbert space. The time evolution operator is
given by

i
U (tf , ti ) = exp − Ĥ(tf − ti ) , (2)
~
so a state at tf is determined by the state |ii at the initial time ti and the operator to be
U (tf , ti )|ii. Note also that U (tf , ti ) only depends on the duration tf − ti in the case of a time-
independent Hamiltonian. From now on we will often work with natural units where ~ = 1. Now,
suppose our initial state |ii = |xi i and the final state |f i = |xn i are both position eigenstates,
then the projection of the former onto the latter is expressed as follows,
hf |ii = hxn |e−i(tn −t0 )Ĥ |xi i. (3)
We can now express this amplitudes in terms of a path integral, without any mention of screens
4
and slits, by adopting a mathematical treatment, which is the method of Dirac who initiated
the idea of path integral before Feynman developed it further. First split the time interval into
n segments, each with duration δt = (tf − ti )/n:
hf |ii = hxn |e−i(tn −tn−1 )Ĥ . . . e−i(t1 −t0 )Ĥ |xi i = hxn |e−iδtĤ . . . e−iδtĤ |xi i. (4)
Our states are normalized by hx1 |x2 i = δ (x1 − x2 ), where δ is the Dirac delta function. Using
the fact that the kets |xi form a complete set of states, we insert the completeness relation,
´
dx |xihx| = 1, in between all factors of e−iδtĤ to express the transition amplitude as a coherent
sum over all states xi through which the system might pass at some intermediate time t
ˆ n−1
 
Y
hf |ii =  dxj  hxn |e−iδtĤ |xn−1 ihxn−1 |e−iδtĤ |xn−2 i . . . hx1 |e−iδtĤ |xi i. (5)
j=1
Let us now focus on the factor hxi+1 |e−iδtĤ |xi i. Recall the existence of |pi, a momentum eigenket
of the operator p̂, i.e p̂|pi = p|pi, and that the plane wave is expressed by the braket hx|pi = eixp
´
and inserting a complete set of momentum states dp |pihp| = 1, we obtain the following for
the special case of a free particle with Ĥ = p̂2 /2m
i 2
´ i 2
hxi+1 |e− 2m δtp̂ |xi i = dphxi+1 |e− 2m δtp̂ |pihp|xi i
´ i 2
= dp e− 2m δtp hxi+1 |pihp|xi i (6)
´ i 2
= dp e− 2m δtp eip(xi+1 −xi ) .
The integral over p is the standard Gaussian integral, which leads to a result proportional to

xj+1 −xj 2
iδt m
2 δt
e .
We will loosely denote the constant of proportionality by N . Such factors will not play a physical
role as they will be divided out when we extract physical observables from the path integral.
For a more general Hamiltonian with a potential piece, we get1

p̂2
1 A+B A B −1 [A,B] −iδt 2m +V (x̂)
We use the Baker-Campbell-Hausdorff equation e = e e e 2 to write e =
p̂2 2
e−iδt 2m e−iδtV (x̂) eO(δt )
and ignore the O(δt2 ) terms.
5
´
hxi+1 |e−iδtĤ |xi i = dphxi+1 |pihp|e−iδtĤ |xi i
´

p̂2
−iδt 2m +V (x̂)
= dp hxi+1 |pihp|e |xi i
´

p2
−iδt 2m +V (xi ) (7)
= dp eipxi+1 e e−ipxi (1 + O(δt))

xj+1 −xj 2

iδt m
2 δt
−V (xi )
=N e
= N eiδt L(xi ,ẋi ) ,
where N is a normalization constant we ignore and
m 2
L(xi , ẋi ) = ẋ − V (x) (8)
2
is the Lagrangian. Note that the Gaussian integration has the effect of a Legendre transform
from the H(x, p) to L(x, ẋ). The Lagrangian has emerged naturally from the Hamiltonian! (At
least we have shown this for Hamiltonian of the form (1). ) Using 7 to each term in the product
in the integrand in 5 and taking the limit that the time interval goes to zero (n → ∞), we obtain
´ n−1 i
P
δt L(xj ,ẋj )
hf |ii = lim N n
Q
dxj e j
n→∞ j=1
(9)
´x xn f
=N Dx(t) eiS[x] ,
x0 =xi
where Dx means sum over all paths x(t) with the correct boundary conditions, and
ˆ
S[x(t)] = dt L[x(t), ẋ(t)].
This infinite-dimensional integral is known as a functional integral, and the representation of

the matrix element is called a functional or path-integral representation of quantum mechanics.
We have just derived the path integral for quantum mechanics! The appearance of t in the path
integral measure Dx(t) is to denote the fact that we are integrating over paths in spacetime. In
the path integral representation, x(t) is a classical function and defines a trajectory in the phase
space of the system. We count on the integral over all the paths to converge. The heuristic
idea behind this is the oscillatory phase factors from different paths tend to cancel out. We
will return to this idea when we discuss the stationary phase (or semi-classical in the present
context) approximation.
6
1.3 Path integral in QFT
In what follows we will work in 3 + 1 dimensions. We will use ~x, ~y etc to denote a vector in the
3 spatial dimensions, and x, y etc to denote a vector in the 3 + 1 dimensional spacetime. We
will use the convention that the signature is (3,1); in other words, x2 := t2 − |~x|2 .
The path integral formalism in QFT is akin to what we just saw in the case of QM. Note: the
QFT equivalent of x̂ are the (Schrödinger picture) fields φ̂(~x). Put differently, φ̂ maps a point
~x in space to an operator. These fields and their conjugate momenta, denoted by π̂(~x), at an
intermediate time can be expressed as follows
´ d3 p

φ̂(~x) = (2π)3
√1 ap ei~p·~x + a†p e−i~p·~x ,
2ωp
(10)
´ d3 p
q
ap ei~p·~x − a†p e−i~p·~x ,
ωp
π̂(~x) = −i (2π)3 2
p|2 + m2 and ap , a†p satisfy the usual commutation relation. As a result we have
p
where ωp = |~
h i
φ̂(~x), π̂(~y ) = iδ (3) (~x − ~y ). (11)
The QM and QFT equivalents are shown in table 1. For the sake of simplicity, we will focus on
one quantum field Φ(~x).
We first begin the derivation by writing out the Hamiltonian density as follows
QM QFT

(x̂, p̂) φ̂(~x), π̂(~x)
h i
[x̂, p̂] = i φ̂(~x), π̂(~y ) = iδ (3) (~x − ~y )
x̂|xi = x|xi φ̂(~x)|Φi = Φ(~x)|Φi
p̂|pi = p|pi π̂(~x)|Πi =´ Π(~x)|Πi
3
hp|xi = e−ipx hΠ|Φi = e−i´ d x Π(~x)Φ(~x)
´ 0 ´ 3 0
hx|x i = dp e−ip(x−x )
0 hΦ|Φ0 i = DΠ e−i d x Π(~x)(Φ(~x)−Φ (~x))
Table 1: QM→QFT map
1
Ĥ = π̂ 2 + V (φ̂), (12)
2
where the potential term V (φ̂) can include interactions. The idea behind the derivation can be
broken down into three main steps- 1) compute the VEV by insertion of a complete set of states,
2) insert yet another complete set of states, the eigenstates of the conjugate operator, so as to
replace the intermediate brakets with the exponential representation, and 3) finally bring the
integral representation to the standard form to perform a Gaussian integral. With this in mind,
we first find the VEV by inserting a complete set of intermediate states, exactly analogous to
the approach taken in QM, as follows. Focusing on the normalized vacuum state
7
ˆ
−iĤ(tf −ti )
hΩ, ti |Ω, tf i := hΩ|e |Ω, tf i = DΦ1 (x) . . . DΦn (x)hΩ|e−iδt Ĥ(tn )
|Φn ihΦn | . . . |Φ1 ihΦ1 |e−iδt Ĥ(t0 )
|Ωi.
(13)
´
The next step is to insert the complete set of conjugate states, DΠj |Πj ihΠj |. In order to
elucidate this better we first consider the simple case of a free particle with Hamiltonian density
´
Ĥ = π̂ 2 /2 and hence, Ĥ = d3 x 21 π̂(~x)2 , where an individual factor of hΦj+1 |e−iδt Ĥ(t) |Φj i takes
the following form
´
d3 x 12 π̂ 2
hΦj+1 |e−iδt Ĥ(t) |Φ
ji = hΦj+1 |e−iδt |Φj i
´ ´
(14)
−iδt d3 x 12 π(~
x)2
= DΠ e hΦj+1 |ΠihΠ|Φj i,
where we have used the fact that the action of π̂ on the eigenstate calls for replacing it with the
eigenvalue π. Now, from table 1, we replace the braket hΠ|Φj i to obtain
ˆ ´ ´
−iδt Ĥ(t) d3 x 21 π(~
x)2 i d3 yΠ(~
hΦj+1 |e |Φj i = DΠ e−iδt e y )(Φj+1 (~
y )−Φj (~
y ))
. (15)
This integral over Π is a Gaussian integral and it is easy to show that performing the integral
results in the following expression
1 ˆ 2 !
2πi 2 1 Φj+1 − Φj
hΦj+1 |e−iδt Ĥ(t)
|Φj i = exp iδt d3 x , (16)
δt 2 δt
where we can replace (Φj+1 − Φj )/δt with Φ̇ under the limit of δt → 0 so that we finally obtain
ˆ
hΦj+1 |e−iδt Ĥ(t)
|Φj i = N exp iδt d3 x L[∂t Φj ](~x) , (17)
where L = 12 (∂t Φj )2 is the Lagrangian density of a free particle and N is the constant prefactor.
This is the result for one matrix element, as for n-number of them we are to sum over the index
P ´
j and this sum together with δt conspires under the limit δt → 0 to yield δt → dt. Since
´ j
S[Φ] = d4 x L[Φ], we can now replace this in the definition of the VEV to obtain the desired
answer.
Similar to the QM case, a more general theory with non-vanishing potential term can be treated
in an analogous way, with the Lagrangian given by L[Φj , ∂t Φj ] = 21 (∂t Φj )2 − V [Φj ] 2 , and we
have
ˆ
hΩ, ti |Ω, tf i = N DΦ(~x, t)eiS[Φ] , (18)
2
Here we put the spatial derivatives terms also in the potential; elsewhere we combine them with the time
derivative term as ∂x2 Φ. This is imposed by Lorentz invariance.
8
n
Q
where the measure is given by DΦ(~x, t) = DΦj . For your convenience, we include Table 2
j=1
which provides a list some useful Gaussian integral results to help aid calculations.
Type Gaussian form Result

´ 1 2 1
Standard Gaussian dx e− 2 x (2π) 2
´ 1 2 1
With a prefactor dx e− 2 αx 2π 2
´ − 12 αx2 +Jx 2π
α1 J2
With a source dx e α e2 2α
´ 1 2 +iJx 1
2πi 2 −iJ
2
dx e− 2 αx

Complex source e 2α
´ α 1
1 T (2π)d 2
d-dim. Gaussian dd x e− 2 x Ax
det A
´ 1 T
Ax+J T x

(2π)d
1
2 1 T A−1 J
d-dim with source dd x e− 2 x det A e2J
Table 2: Different
´ ´ +∞of the Gaussian integral. Note that all the integrals have limits x ∈
forms
(−∞, ∞), i.e. dx ≡ −∞ dx.
Now that we have derived the PI formalism, we must ask the question of what it is good for?
Conceptually, we have previously seen the most direct method of quantization is the canonical
quantization program. This closely mimics the development of QM in the sense that time is
singled out as a special coordinate and manifest Lorentz invariance is sacrificed. The advantage
of canonical quantization is that it quantizes only physical modes. Unitarity of the system is
thus obvious. The PI formalism on the other hand, makes the Lorentz symmetry manifest (if
it is there to start with). We will argue that it is interesting and very useful since it allows us
to calculate quantum observables via classical solutions- this is done using a technique called
saddle-point approximation we will explore in the next section. Another feature the PI possesses
is that it enables us to build approximation schemes via Monte-Carlo methods that helps us
simulate complex systems. For simple theories such as QED, the canonical quantization method
is not too difficult, but the canonical quantization of more complicated theories, such as non-
Abelian gauge theories, is often prohibitively tedious. This constitutes another reason to prefer
the PI formalism at times.
As we have seen, it is possible to easily go back and forth between many of the other quantization
programs to see the relationships between them.
1.4 The Classical limit

´
Restoring the reduced Planck’s constant ~, we observe that both the action S = dt L and ~
possess dimensions of [energy]×[time] thus making the exponential piece in the integral of the
VEV dimensionless (as it should be). We now have
ˆ
i
hΩ, ti |Ω, tf i = N DΦ(~x, t)e ~ S[Φ] . (19)
9
We can now use the method of stationary phase to evaluate the path integral in the limit ~ → 0.
Now we notice that under the limit of ~ → 0, the integral is highly oscillatory and is dominated
by a specific value of Φ for which S[Φ] has an extremum. This observation can be used as the
starting point to derive the classical limit of quantum mechanics. If S[Φ] ~ holds then almost
everywhere in function space the strongly oscillating contributions to the path integral made
by neighbouring paths will cancel each other. This cancellation is only avoided if the action
does not change much in the vicinity of a path, i.e. if the action is stationary with respect
to variations of Φ(~x, t) and ∂t Φ(~x, t). Note that this is precisely the condition that determines
the classically allowed trajectory Φcl (~x, t) according to Hamilton’s principle of least action that
determines the Euler-Lagrange equations which a classical field satisfies, i.e.
ˆ
δS = δ dt L(Φ, ∂t Φ) = 0. (20)
The path integral is dominated by this extremum and hence, the only contribution to the
amplitude is from the the classical solutions to the eom thus providing the name classical limit.
Thus the classical description of a system can be derived from quantum theory as the field
configuration in the stationary phase approximation.
Note that the stationary phase approximation is related to the saddle point approximation via
Wick rotation to Euclidean path integral, for which the spacetime has signature (0,4) instead of
(1,3), by the transformation t → −iτ . In this case, in the limit ~ → 0, the contribution to the
path integral of field configurations away from the saddle points falls off exponentially, since
i
´ 1
´
dtL[Φ,∂t Φ] dτ L[Φ,∂t Φ]
e~ = e~ .
1.5 Correlators in the path integral

Generalizing the vacuum amplitude discussed earlier, we now consider the
correlators. Again
focussing on one quantum field, we consider the n-point correlator hΩ|T φ̂(x1 ) . . . φ̂(xn ) |Ωi,
where T (. . .) is the time-ordered product (somtimes simply referred to as the T-product) defined
as
T φ̂(t1 )φ̂(t2 ) . . . φ̂(tn ) = φ̂(to(1) )φ̂(to(2) ) . . . φ̂(to(n) ) (21)
where the dependence on ~x has been suppressed, and o : {1, 2 . . . , n} → {1, 2 . . . , n} is the
time-ordering map satisfying
to(1) ≥ to(2) ≥ · · · ≥ to(n) .
For instance, for n = 2 we have

T φ̂(t1 )φ̂(t2 ) = Θ(t1 − t2 )φ̂(t1 )φ̂(t2 ) + Θ(t2 − t1 )φ̂(t2 )φ̂(t1 ). (22)
where Θ is the Heaviside theta function, whose derivative is given by the Dirac delta function :
d
Θ(x) = δ(x). (23)
dx
10
In what follows, we will prove
´
DΦ(x)eiS[Φ] Φ(x1 ) . . . Φ(xn )
hΩ|T φ̂(x1 ) . . . φ̂(xn ) |Ωi = ´ , (24)
DΦ(x)eiS[Φ]
which provides the bridge between the Hamiltonian and Lagrangian, or the canonical and path
integral, formulation of quantum field theories.
It is important to notice that the operators on the LHS of the n-point correlator 24, φ̂(x) are
Heisenberg operators, while the operators Φ(x) on the RHS are the real values of the classical
path at x = (~x, t). It is also to be noted that the ordering of Φ(x) in the RHS is irrelevant. With
the distinction understood, from now on we will often use the standard notation φ(x) instead
of Φ(x) for the classical fields being integrated over in the path integral. Lastly, we see that the
normalisation of the vacuum, hΩ|Ωi = 1, is manifest in 24.
1.6 Generating functional

It will turn out to be very convenient to define the generating functionals I[J], as they efficiently
capture the information of all correlators as defined by the RHS of 24. We let
ˆ ˆ
4
I[J] := Dφ exp iS[φ] + i d x J(x)φ(x) , (25)
where J(x) is a source that is coupled to the field φ(x). Notice that when J = 0, the generating
functional reduces to the vacuum amplitude. The source can be viewed as the map that takes
an input in 4-dimensional Euclidean space and spits out a real number, i.e. J : R4 → R. We
can express the source at a different position by integrating it at the original position against a
Dirac delta function as follows
ˆ
J(y) = d4 x δ (4) (x − y)J(x), (26)
and hence, it is natural to define the variational partial derivative as follows
δ
J(y) = δ (4) (x − y). (27)
δJ(x)
We will use 27 repeatedly in what follows. Acting on the generating functional with the varia-
tional operator yields an interesting result. First acting on I[J] with δJ yields
ˆ ˆ
δ
I[J] = Dφ exp iS[φ] + i d4 y J(y)φ(y) φ(y)δ (4) (x − y) (28)
iδJ(x)
Setting the source to zero leads to
11
´ ´
1 δ
Dφ exp iS[φ] + i d4 y J(y)φ(y) φ(y)δ (4) (x − y)

i δJ(y) I[J] J=0 =
J=0
(29)
´
= Dφ eiS[φ] φ(x),
which is nothing but the 1-point correlator up to a normalization constant! In particular,

according to 24 this is proportional to ∼ hΩ|φ̂(x)|Ωi. Similarly, by taking more functional
derivatives we are led to
´
1 1 δn Dφ eiS[φ] φ(x1 ) . . . φ(xn )
I[J] = ´ , (30)
I[0] in δJ(x1 ) . . . δJ(xn ) J=0 Dφ eiS[φ]

which equals hΩ|T φ̂(x1 . . . φ̂(xn ) |Ωi according to 24 . The generating functional is the QFT
analog of the partition function in statistical mechanics – it tells us everything we could possibly
want to know about a system. To solve the quantum theory in hand completely is to possess
an exact closed form expression for I[J]. In particular, to prove 24, it is sufficient to prove the
equivalence (up to a normalizing factor) between I[J] and the similar generating function for
canonical quantization
´
d4 x J(x)φ̂(x)
Ẑ[J] ≡ hΩ|T ei |Ωi. (31)
In what follows we will do this with the aid of a differential equation called the Schwinger-Dyson
(SD) equation. In other words, by showing that they both solve the SD equation and have the
same boundary condition, we will establish
I[J]
Ẑ[J] = =: Z[J], (32)
I[0]
and thereby 24.
1.7 Schwinger-Dyson Equation in Canonical Quantization

The SD differential operator is defined as follows

1 δ 0 1 δ
∇SD ≡ ∂x2 +V − J(x) , (33)
i δJ(x) i δJ(x)
where we use the notation 2 = ∂x2 = ∂t2 − ∂~x2 . We will show that ∇SD annihilates Ẑ[J] order-
∞
1 n
by-order in J. Using ex =
P
n! x , we write
n=0
∞ n ˆ Y
n
X i
Ẑ[J] = d4 xj J(x1 ) . . . J(xn )hΩ|T φ̂(x1 ) . . . φ̂(xn ) |Ωi. (34)
n!
n=0 j=1
Using 27 and shifting the sum over n by 1, it is not hard to see that
∞X in ˆ
1 δ
Ẑ[J] = d4 x1 . . . d4 xn J(x1 ) . . . J(xn )hΩ|T φ̂(x)φ̂(x1 ) . . . φ̂(xn ) |Ωi. (35)
i δJ(x) n!
n=0
12

The next step is to compute ∂x2 hΩ|T φ̂(x)φ̂(x1 ) . . . φ̂(xn ) |Ωi. First we note that

∂t hΩ|T φ̂(x)φ̂(x0 ) |Ωi = ∂t hΩ|φ̂(x)φ̂(x0 )Θ(t − t0 ) + φ̂(x0 )φ̂(x)Θ(t0 − t)|Ωi

= hΩ|T ∂t φ̂(x)φ̂(x0 ) |Ωi + hΩ|φ̂(x)φ̂(x0 )∂t Θ(t − t0 ) + φ̂(x0 )φ̂(x)∂t Θ(t0 − t)|Ωi

= hΩ|T ∂t φ̂(x)φ̂(x0 ) |Ωi + δ(t − t0 )hΩ|[φ̂(x), φ̂(x0 )]|Ωi
(36)
where we have used 23. The second term vanishes, since [φ̂(x), φ̂(x0 )] = 0 at equal time, which
is enforced by the delta function.
Moreover, we have

∂t2 hΩ|T φ̂(x)φ̂(x0 ) |Ωi = hΩ|T ∂t2 φ̂(x)φ̂(x0 ) |Ωi + δ(t − t0 )hΩ|[∂t φ̂(x), φ̂(x0 )]|Ωi
(37)
= hΩ|T ∂t2 φ̂(x)φ̂(x0 ) |Ωi − iδ (4) (x − x0 ).
h i
In the last line, we have used the equal time commutation relation φ̂(~x), π̂(~y ) = iδ (3) (~x − ~y ),
the form of the Hamiltonian density
1
Ĥ = π̂ 2 + V (φ̂), (38)
2
and i∂t φ(x) = [φ̂(x), Ĥ]. Since the ∂~x2 term does not lead to a similar contact term, we obtain

∂x2 hΩ|T φ̂(x)φ̂(x0 ) |Ωi = (∂t2 − ∂~x2 )hΩ|T φ̂(x)φ̂(x0 ) |Ωi
(39)
= hΩ|T ∂x2 φ̂(x)φ̂(x0 ) |Ωi − iδ (4) (x − x0 ).
A straightforward generalization yields

∂x2 hΩ|T φ̂(x)φ̂(x1 ) . . . φ̂(xn ) |Ωi = hΩ|T ∂x2 φ̂(x)φ̂(x1 ) . . . φ̂(xn ) |Ωi
n
δ (4) (x − xj )hΩ|T
P
−i φ̂(x1 ) . . . φ̂(xj−1 )φ̂(xj+1 ) . . . φ̂(xn ) |Ωi.
j=1
(40)
From this we see
h i
δ
∂x2 1i δJ(x) + V 0 1i δJ(x)
δ
Ẑ[J] = J(x)Ẑ[J]
∞ n ´
i
d4 x1 . . . d4 xn J(x1 ) . . . J(xn )
P
+ n! (41)
n=0
×hΩ|T (2φ̂ + V 0 [φ̂])(x)φ̂(x1 ) . . . φ̂(xn ) |Ωi
Since the second term vanishes, we see that
∇SD Ẑ[J] = 0 (42)
13
as claimed. Now we give some details about the vanishing of the first term:

hΩ|T (2φ̂ + V 0 [φ̂])(x)φ̂(x1 ) . . . φ̂(xn ) |Ωi = 0 (43)
The classical equation of motion states
δL[φ, φ̇]
∼ (2φ + V 0 )(x) = 0 (44)
δφ(x)
Quantum mechanically, the statement is that of the correlator
δL[φ, φ̇]
hΩ| . . . |Ωi = 0 (45)
δφ(x)
as can be shown by considering the intermediate amplitude 17 and recalling that Φj ’s are being
integrated over and the integral is independent of the shift Φj → Φj + δΦ. Note that the SD
equation 42 can viewed as capturing the quantum aspects of the correlator: the delta-function
term in 40 is present due to the non-vanishing commutator, and will be multiplied by a relative
factor of ~ if we restore the ~ dependence. We see that the quantum correction comes in the
form of a contact interaction, consistent with the claim that classical physics is a good decription
at large scale when such contact interactions plays no role.
1.8 Schwinger-Dyson Equation in Path Integral Formalism

In order to prove the equivalence between the path integral and the canonical formulations of
QFT, namely 32 and as a result 24, we now need to show that ∇SD also annihilates I[J].
To see that, note that a transformation φ(x) 7→ φ(x) + δφ(x) in the integrand of I[J] leaves I[J]
invariant, since φ(x) is integrated over in the functional integration. As a result, we must have
´ ´
δ
0= Dφ δφ(x) exp iS[φ] + i d4 y J(y)φ(y)
´
δS[φ]
´ ´
=i Dφ δφ(x) + d4 y J(y)δ (4) (x − y) exp iS[φ] + i d4 x J(y)φ(y) (46)
´
δS[φ]
´
=i Dφ δφ(x) + J(x) exp iS[φ] + i d4 x J(y)φ(y) .
Using the form of the Lagrangian L = 12 (∂x φ)2 − V [φ], we get that
ˆ ˆ !
δS[φ] 4
0= Dφ + J(x) exp iS[φ] + i d x J(y)φ(y)
δφ(x)
ˆ ´
0 iS[φ]+i d4 x J(y)φ(y)
−(∂x2 φ

= Dφ + V )(x) + J(x) e (47)
ˆ ´
iS[φ]+i d4 x J(y)φ(y)
= −∇SD Dφ e = −∇SD I[J]
which immediately implies that the SD operator also annihilates Z[J] := I[J]/I[0]. As Ẑ[J]
14
and Z[J] are both functionals of the auxiliary source field J(x) that are annihilated by the
differential operator ∇SD , and moreover have the same boundary value
Ẑ[J]|J=0 = Z[J]|J=0 = 1,
we conclude that they represent different ways to express the same function
Ẑ[J] = Z[J] (48)
and we have thereby proven the equivalence between the PI and canonical formulations of QFT.
Finally, let us make more comments on SD equation. First, note that this formalism is non-
perturbative, since it comes directly from a variational principle. By deriving it, we have used
the fact that the path integral measure is invariant under a shift of φ(x). This could fail and lead
to situations where there is an “anomalous symmetry”, a symmetry that is broken by quantum
effects.
1.9 Property of path integrals: time-ordering

We know that time-ordering of operators is an important element to take into account in the
canconical formalism. Now that we have proven the equivalence between canonical and path
integral formalism, one might wonder how this comes into play in the path integral formalism.
For instance, 24 states
´ i
Dφ e ~ S[φ] φ(x1 )φ(x2 )
hΩ|T φ̂(x1 )φ̂(x2 ) |Ωi = hΩ|Θ(t1 −t2 )φ̂(t1 )φ̂(t2 )+Θ(t2 −t1 )φ̂(t2 )φ̂(t1 )|Ωi = ´ i .
Dφ e ~ S[φ]
(49)
While time ordering is explicit on the LHS, it does not seem to be present on the RHS. In what
follows we will see that time-ordering in fact happens automatically in path integral. To see
this, we turn to the discretized version of the path integral
ˆ ˆ Y
n
i
0 0
Dφ e ~
S[φ]
φ(t, ~x)φ(t , ~x ) = lim Dφj (~x)hΩ|e−iδt Ĥ(tn )
|φn i . . . hφ1 |e−iδt Ĥ(t1 )
|Ωiφ(x)φ(x0 ).
n→∞
j=1
(50)
Let us consider tj := t0 + jδt = t, tj 0 = t0 . For the case of t > t0 , we have j > j 0 and
15
´ i
´ n
Dφ e ~ S[φ] φ(t, ~x)φ(t0 , ~x0 ) = lim Dφj (~x) hΩ|e−iδt Ĥ(tn ) |φ
Q
ni . . .
n→∞ j=1
×hφj+1 |e−iδt Ĥ(tj ) φ̂(~

x)|φ ji . . .
(51)
×hφj 0 +1 |e−iδt Ĥ(tj 0 )

φ̂(~x0 )|φj 0 i . . . hφ1 |e−iδtĤ(t1 ) |Ωi
= N hΩ|φ̂(x)φ̂(x0 )|Ωi.
Similarly, for the case of t < t0 , j < j 0 , we have

ˆ
i
Dφ e ~ S[φ] φ(t, ~x)φ(t0 , ~x0 ) = N hΩ|φ̂(x0 )φ̂(x)|Ωi. (52)
This is elucidated in figure 2.
Figure 2: The path from (ti , ~xi ) to (tf , ~xf ) is separated into three paths. We have to distinguish
between the cases t < t0 and t > t0 when drawing this picture.
Thus, we see that time-ordering comes for free in the path integral formalism. The same analysis
done above also holds for the n-point function hΩ|T (φ(x1 )φ(x2 ) . . . φ(xn )) |Ωi whose discretiza-
tion is shown in figure 3.
Figure 3: The path from (xi , ti ) to (xf , tf ) is separated into (n + 1) paths.
With this understanding of the time-ordering, from now on we will often use the shorthand
notation
h. . . i := hΩ|T (. . . ) |Ωi.
1.10 Example: solving the free theory

Let us consider the (real) free scalar theory where we can calculate the generating functional
exactly. The Lagrangian takes the following form
16
1
L0 = − φ(x) ∂x2 + m2 φ(x).

(53)
2
The generating functional can be readily written as follows

ˆ ˆ ˆ
4 1 2 2 4

I[J] = Dφ exp i d x φ(x) −∂x − m + iε φ(x) + i d x J(x)φ(x) . (54)
2
From where did the iε pop up? We placed this in by hand in order to ensure the convergence
of the otherwise oscillatory integral. With
A = i(∂x2 + m2 − iε),
we immediately realize using the list of Gaussian integrals in table 2 that the generating func-
tional yields the following result
ˆ ˆ 1
(2π)d 2 − 1 J T A−1 J

4 1 2
I[J] = Dφ exp d x − A(x)φ (x) + iJ(x)φ(x) = e 2 (55)
2 det A
It is easy to see what A−1 is since by the definition of a Green’s function, A−1 is nothing but
the (Feynman) propagator DF (x, y) satisfying the following equation
A(x)DF (x, y) = δ (4) (x − y). (56)
It is important to note here that DF (x, y) = DF (x − y) due to Lorentz invariance. This is a

standard Green’s function equation that is solved by the following choice of propagator
ˆ
d4 p i
DF (x − y) = e−ip·(x−y) (57)
(2π)4 p2 − m2 + iε
As a result, we can write 55 as
1
´
d4 xd4 yJ(x)DF (x−y)J(y)
I[J] = I[0]e− 2 (58)
using the Fourier transform of the Dirac delta function

ˆ
(4) d4 p −ip·(x−y)
δ (x − y) = e .
(2π)4
17
As a result, we obtain the following
1 1 δn
hΩ|T (φ(x1 ) . . . φ(xn )) |Ωi = I[0] (i)n δJ(x1 )...δJ(xn ) I[J] J=0
´
1 δn
= in δJ(x1 )...δJ(xn ) exp − 12 d4 x d4 y J(x)DF (x − y)J(y)
J=0
P Q
= DF (xi − xj ).
pairs i,j
(59)
when n is even, and equals 0 when n is odd. This is called the Wick theorem for free theory.
For instance, when n = 4, one can compute explicitly that
1 1 δn
hΩ|T (φ(x1 ) . . . φ(xn )) |Ωi = I[0] (i)n δJ(x1 )...δJ(xn ) I[J] J=0
(60)
= D12 D34 + D13 D24 + D14 D23 ,
where we used the shorthand notation Dij := DF (xi − xj ), as these are the only three terms
that remain after taking J = 0. Drawing out Feynman diagrams, the three contractions are
shown in figure 4.
Figure 4: The three contributions to the Feynman propagator.
1.11 Perturbation theory

Considering now an interacting theory with
L = L0 + gV [φ] (61)
(see 53), where the g is small. This theory is susceptible to the treatment of perturbation theory,
as the small dimensionless parameter g is responsible for the smooth turning-on or turning-off
of the interaction. As a result, we can express the physical observables as a power expansion in
g. First, by expanding the exponential we expand the generating function as
∞ ˆ ´ ˆ
X gk iS0 [φ]+i d4 x J(x)φ(x)
I[J] = Dφ e d4 x1 . . . d4 xk (V [φ(x1 )] . . . V [φ(xk )]) . (62)
k!
k=0
If the potential V [φ(x)] (which we will often write as V (x) for convenience) possesses a power
18
series expansion in the field φ(x):
X vn
V [φ] = φn ,
n
n!
then each piece in the sum takes the following form

ˆ ´
d4 x J(x)φ(x)
Dφ eiS0 [φ]+i φ(x1 )m1 . . . φ(x` )m` , (63)
with the location xi being integrated over. Using this it is easy to see that of the correlator of
the interacting theory is given by
1 1 δn
hφ(y1 ) . . . φ(y` )i = I[0] in δJ(y1 )...δJ(yn ) I[J] J=0
´
(64)
∞
gk ´ Q
k vm1 ...vmk Dφ eiS0 [φ] φ(y1 )...φ(y` )φm1 (x1 )...φmk (xk )
d4 x ´
P P
= k! j m1 !...mk ! DeiS0 [φ]
.
k=0 j=1 {m1 ,...,m` }
Note that the part in orange is the free theory propagator

XY
hφ(z1 ) . . . φ(z` ) × φ(z`+1 ) . . . φ(z`+m1 ) × . . . φ(z`+Pk mi )i = DF (zi − zj ) (65)
i=1
pairs i,j
In the above formula zi = yi for i = 1, . . . , `, representing the external insertions, and z`+1 =
· · · = z`+m1 = x1 , representing the φm1 interaction at location x1 , etc. Combining the above,
order-by-order in g we see that we have recovered the Feynman rule for perturbative QFT.
1.12 Connected generating functionals

Let us consider a QFT with action S[φ] and let us turn-on a set of classical currents J(x) that are
coupled to the fields φ(x) of the theory. We have seen that I[J] is also the generating functional
of the correlators of the theory. Note that I[J] is the sum of all VEVs or vacuum-vacuum
amplitudes in the presence of the classical current J(x) and this includes connected as well as
disconnected diagrams. It will be handy to have a generating functional W [J] of the connected
Green function. We claim that it is given by
iW [J] ≡ ln I[J]. (66)
To prove this, schematically we denote as C the set of all inequivalent connected diagrams, and
as zc the contribution of c ∈ C to the vacuum amplitude. Then we have3
∞
YX zcnc
I[J] = . (67)
nc !
c∈C nc =0
P
Note that the above is nothing but I[J] = e c zc , and hence
X
iW [J] = zc (68)
c
3 1
Revisit the free theory four-point function 60 if you need to convince yourself of the factor nc !
.
19
is the generating function for the connected diagrams.
We will now explicitly demonstrate this with the free field example, which has
ˆ
1
I0 [J] = N exp − d4 x d4 y J(x)DF (x − y)J(y) . (69)
2
Schematically, it is shown as in Figure 5.
Figure 5: There are n number of propagators that make up the sum with the prefactor −1/2
being suppressed.
As a result, we have ˆ
i
W0 [J] = d4 x d4 y J(x)DF (x − y)J(y) (70)
2
which is indeed the generating function for the connected diagrams:

δk iD (x − x )
F 1 2 if k = 2
W0 [J]|J=0 = (71)
δJ(x1 ) . . . δJ(xk ) 0 otherwise.
One might note a parallel to statistical physics. Turning to Euclidean signature after a Wick
rotation, we have
X
W [J] = Zc = − ln Z[J]. (72)
c
This is nothing but the Euclidean counterpart of the Lorentzian expression 66, W [J] corresponds
to the (Helmholtz) free energy while Z[J] is partition function of the system 4 .
1.13 1PI Green’s functions

Apart from the connected diagrams, it is also useful to work with the sum of all connected one-
particle irreducible (1PI) graphs instead of W [J]. A 1PI graph is one that cannot be disconnected
by cutting through an internal line as shown in figure 6.
4
In the presence of fermions, one should choose the anti-periodic boundary condition for the fermions for the
analogy to hold.
20
Figure 6: Diagrammatically, the 1PI Green’s functions are constructed as the sum of all diagrams
that cannot be cut into disconnected parts by cutting a single propagator line as show here.
1.13.1 Quantum effective action and its relation to W [J]
First, let us define an object, which is often called the quantum effective action, as
ˆ
4
Γ[φ] ≡ supJ W [J] − d x J(x)φ(x) . (73)
Now we show that it is equal to

ˆ ˆ
4 4
Γ[φ] ≡ supJ W [J] − d x J(x)φ(x) = W [J] − d x J(x)φ(x) , (74)
J=Jφ
δ
´
d4 x J(x)φ(x) = 0, or

where Jφ solves the extremization equation δJ W [J] −
δ 1 1 δ
φ(x) = W [J] = I[J] , (75)
δJ(x) J=Jφ I[J] i δJ(x) J=Jφ
where we have used the definition 66 of W [J] in the last equality. In other words, we view Jφ
as being determined by φ(x) through 75. As such, 73 defines Γ[φ] as the Legendre transform of
W [J], where we treat φ and J as conjugate variables. As a consequence, we could also in turn
consider W [J] as the inverse Legendre transform of Γ[φ]:
ˆ ˆ
4 4
W [J] ≡ supφ Γ[φ] + d x J(x)φ(x) = Γ[φ] + d x J(x)φ(x) , (76)
φ=φJ
where φ = φJ is similarly determined by the extremum condition

ˆ
δ 4
Γ[φ] + d x J(x)φ(x) = 0. (77)
δφ
Note that φ = φJ is precisely the φ(x) that solves
δ 1 1 δ
φ(x) = W [J] = I[J]. (78)
δJ(x) I[J] i δJ(x)
In other words, the same equation 75 and 78, can be used to either solve for φJ given J, or solve
for Jφ given φ. To prove the above statement, let φJ to be the solution to 78 given J, and we
21
will now show that it solves 77 as well:
ˆ ˆ ˆ
δ 4 δ 4 4
Γ[φ] + d x J(x)φ(x) = W [Jφ ] − d x Jφ (x)φ(x) + d x J(x)φ(x)
δφ φ=φJ δφ φ=φJ

δJφ δW
= − φ − Jφ + J =0
δφ δJ J=Jφ φ=φJ
(79)
δW
where we have used δJ J=J = φ from 75, and the fact that JφJ = J by 75 and 78.
φ
Note that 78 might be viewed as a generalisation of the VEV

1 1 δ
hΩ|φ̂(x)|Ωi = I[J] (80)
I[J] i δJ(x) J=0
to situations in the presence of the current J.
1.13.2 The role of Γ[φ]
Now that we have established a relation between the quantum effective action Γ[φ] and the
generating functional of all connected Green’s functions W [J], we can go ahead and investigate
the role of Γ[φ]. We claim that Γ[φ] has the role of the generating functional of all the 1PI
Green’s functions. More concretely, consider the expansion
∞ ˆ
X 1
Γ[φ] = d4 x1 . . . d4 xn Γn (x1 , . . . , xn )φ(x1 ) . . . φ(xn ). (81)
n!
n=0
We claim that Γn (x1 , . . . , xn ) is the so-called 1PI n-point Green’s function. Note that the above
expression is completely equivalent to
δn
Γ[φ] = Γn (x1 , . . . , xn ). (82)
δφ(x1 ) . . . δφ(xn )
To prove the claim, let us first explain what 1PI n-point Green’s functions are. The key prop-
erties include
1. They are connected as the name suggests.
2. They are irreducible on internal lines, i.e. cannot be disconnected by cutting a single
propagator line as shown in figure 6.
3. They are truncated, i.e. they need to be multiplied by W2 for each external leg as shown
in figure 7
Connected Green’s functions are constructed from tree diagrams with 1PI Green’s functions as
vertices and propagators W2 given by the connected 2-point functions as shown in figure 8.
22
Figure 7: Each external leg (LHS) of a 1PI n-point Green’s function should be thought of as
two legs with a propagator W2 placed in between (RHS).
Figure 8: The full connected Green’s functions is obtained by summing all tree diagrams with the
1PI Green’s functions as vertices, using the full connected two-point function as the propagators.
Before we prove this statement rigorously, let us make some heuristic arguments for why this
should hold. Since working with the 1PI Green’s functions as building blocks requires no loop
expansions (cf. Figure 8), this shows that classical calculations with Γ[φ] suffice to render
quantum results from the point of view of the original action S[φ]. The reasoning behind this
lies in the act of taking the Legendre transformation. For the sake of argument, let’s pretend
that Γ[φ] is our action instead of S[φ] so that the generating functional corresponding to this
new action, IΓ [J] takes the following form
ˆ ˆ
4
IΓ [J] = Dφ exp iΓ[φ] + i d x J(x)φ(x) . (83)
The eom of this action is nothing but
δ
Γ[φ] + J(x) = 0, (84)
δφ(x)
which implies that φ(x) = φJ (x) as we saw back in 77. Hence, let us perform a semi-classical
saddle-point approximation around the point φ(x) = φJ (x) so that the classical generating
functional takes on the following form
n ´ o
IΓcl [J] ≈ exp i Γ[φJ ] + d4 x J(x)φJ (x) = eiW [J] = I[J] (85)
where we used 76 and 66.

Hence, we see that working with Γ[φ] instead of S[φ] and working in the classical limit is
23
equivalent to working directly with S[φ] at a full quantum level. Hence, it’s no surprise that Γn
derived from Γ[φ] as in 82 plays the role of the 1PI Green’s functions. Another way to see this
δΓ
is to note that 77 implies δφ = 0 when J = 0, implying that the possible quantum averaged
values for in the absence of a source are just the extrema of Γ[φ], which are usually associated
with the classical solutions of the theory with action Γ[φ].
A more rigorous proof is to realize the diagrammatic expansion shown in figure 8 via induction.
Consider the case of n = 2 where we have
δ2

δ δ δ
Γ2 (x1 , x2 ) = Γ[φ] = Γ[φ] = − Jφ (x2 ), (86)
δφ(x1 )δφ(x2 ) δφ(x1 ) δφ(x2 ) δφ(x1 )
where we have used 84. Note that

ˆ
(4) δJ(x) δ δφ(z) δJ(x)
δ (x − y) = = Jφ [φ = φJ ] = d4 z , (87)
δJ(y) δJ(y) δJ(y) δφ(z)
where φ and J are thought of as the conjugate variables constrained by 77. As a result, we write
δφ(z)
δJ(y) = ( δJ(x)
δφ(z) )
−1 as a continuous generalisation of the matrix inverse. Put that in 86 we obtain
−1
δ2

δφJ (x1 ) −1
Γ2 (x1 , x2 ) = −( ) = − W [J] = −W2 (x1 , x2 )−1 , (88)
δJ(x2 ) δJ(x1 )δJ(x2 )
where we have used 78. More compactly, we can simply write
Γ2 = −W2−1
. This is a trivial case that corresponds to the definition of the truncated diagram property of
Γn (the third property of the 1PI Green’s function that we mentioned earlier). Now consider
n = 3 and play the same game to obtain
δ3 δ −1
W3 (x1 , x2 , x3 ) = W [J] =− Γ (x2 , x3 ), (89)
δJ1 δJ2 δJ3 J=0 δJ1 2
where Ji = J(xi ) and the functional derivatives are with respect to J since we are expressing
W [J] in terms of Γ[Φ] and hence make use of the variable J which is canonical to φ. Multiplying
and dividing by δφJ (y1 ), we obtain
ˆ
δφJ (y) δ
W3 (x1 , x2 , x3 ) = − d4 y Γ−1 (x2 , x3 ) . (90)
δJ(x1 ) δφ(y) 2 φ(y)=φJ (y)
δφJ (y1 ) δ2
Since δJ(x1 ) = δJ(y1 )δJ(x1 ) W [J] = W2 (x1 , y1 ), the above equation becomes
ˆ
4 δ −1
W3 (x1 , x2 , x3 ) = − d y1 W2 (x1 , y1 ) Γ (x2 , x3 ) . (91)
δφ(y1 ) 2 φ(y1 )=φJ (y1 )
Now, we can simplify δΓ−1

2 by making use of the matrix differentiation formula which reads
24
dK −1 = −K −1 dK K −1 . (92)
Using this, we finally obtain
´
δ
W3 (x1 , x2 , x3 ) = d4 y1 d4 y2 d4 y3 W2 (x1 , y1 )W2 (x2 , y2 ) δφ(y1 ) Γ2 (y2 , y3 ) W2 (x3 , y3 )
(93)
´
= d4 y1 d4 y2 d4 y3 W2 (x1 , y1 )W2 (x2 , y2 )W2 (x3 , y3 )Γ3 (y1 , y2 , y3 ),
where we have used 82.

This is diagrammatically visualized in figure 9.
Figure 9: The Legendre tree for the case when n = 3.
Similarly, we can now continue taking functional derivatives to obtain more terms and hence,
prove Wn by induction. The connected generating functional Wn is visualized for the case of
n = 4 in figure 10.
Figure 10: The Legendre tree for the case when n = 4.
25
1.14 Further Reading
References
[1] Schwartz, M. D. (2014). Quantum field theory and the standard model. Cambridge University
Press.
[2] Banks, T. (2008). Modern quantum field theory: a concise introduction. Cambridge Univer-
sity Press.
[3] Zee, A. (2010). Quantum field theory in a nutshell (Vol. 7). Princeton university press.
[4] Nakahara, M. (2018). Geometry, topology and physics. CRC press.
[5] Weinberg, S. (1995). The quantum theory of fields (Vol. 2). Cambridge university press.
In particular Schwartz Chapter 14, 7.1 and Banks Chapter 3.
26
Topic-II: Symmetries

Partially based on notes taken by Naveen Balaji Umasankar
27
2 Topic-II: Symmetries
2.1 Why symmetries?
As the pioneering physicist and Nobel prize winner Philip Anderson once said: it is only slightly
overstating the case to say that physics is the study of symmetry.
The importance of symmetries in modern physics cannot be overstated. The types of symmetries
common in physics include space-time (geometric) symmetries, global symmetries, and gauge
symmetries. The generalisations and exploitations of these ideas are an active area of research.
For instance: diffeomorphism symmetries lead to the key idea of general relativity; gauge theo-
ries dictate the dynamics of standard model particles; scaling symmetry is an earmark of critical
phenomena, etc. More speculatively, the idea of symmetries has led to the proposals of su-
persymmetry and the grand unified theory (GUT), which keep providing intuition about new
physics beyond the standard model. Moreover, from Goldstone theorem to Higgs mechanism,
how a symmetry is broken is also crucial in understanding modern physics. We aim to discuss
the symmetry aspects of quantum field theories in this topic.
2.2 Noether’s theorem

The Lagrangian formalism provides a natural framework to study symmetries of the quantum
system. This stems from the fact that one can easily study the variation of a Lagrangian and
thereby the effects of field transformations on the dynamics. Before we turn to the quantum
discussion, let us quickly remind ourselves of the classical Noether’s theorem.
To derive Noether’s theorem, first consider a symmetry transformation, in the form of a field
transformation
φ(x) 7→ φ(x) + δG φ(x) (94)
(or its integrated counterpart) that leaves the action

ˆ
S[φ, ∂µ φ] = d4 y L[φ(y), ∂µ φ(y)]
invariant. In the above, we have used the common notation of xµ = (t, ~x) for µ = 0, 1, 2, 3, and
∂
∂µ := ∂xµ .
In other words, consider
φ(x) 7→ φ(x) + δG φ(x) (95)
for a (space-time independent) constant , we have

∂L ∂L
δG L = (x)δG φ(x) + (x)∂µ (δG φ(x))
∂φ ∂(∂µ φ)
(96)
∂L ∂L
= (x) − ∂µ (x) δG φ(x) + total derivative
∂φ ∂(∂µ φ)
28
Assuming that (95) is a symmetry of the Lagrangian, we have δG L = 0 and

∂L ∂L
(x) − ∂µ (x) δG φ(x) = total derivative. (97)
∂φ ∂(∂µ φ)
Note that the above equation looks very similar to the equation of motion

∂L ∂L
(x) − ∂µ (x) = 0. (98)
∂φ ∂(∂µ φ)
The difference is that 97 holds for all φ(x), irrespective of whether it solves the equation of
motion. This is because the classical field configuration is the one that renders δS = 0 for any
small variation around it, while we are considering a symmetry which renders δG S = 0 under the
specific variation δG φ given by the symmetry transformation, and now for any field configuration.
Next, we consider a field transformation
φ(x) 7→ φ(x) + ε(x) δG φ(x), (99)
now with ε(x) being a (not necessarily constant) function on the space-time. Unlike 95, the
above transformation is no longer a symmetry due to the space-time variation of (x). Hence,
we expect δS to be vanishing only when ∂µ (x) = 0, and as a result
ˆ ˆ
4 µ µ
δS = d x ∂µ (x) JN (x) =− d4 x (x) ∂µ JN (x) + boundary term (100)
µ
for some JN (x). As mentioned earlier, the situation for classical solutions is different, and we
have δS|φ=φclas = 0 for arbitrary (x) even when ∂µ (x) 6= 0. As a result, we have
µ
∂µ J N (x)|φ=φclas = 0. (101)
Noether’s theorem is then the theorem that states : in the presence of a one-parameter family
of continuous symmetry of the Lagrangian, there is one current associated to it that is conserved
classically.
As usual, the fact that the current is conserved can be rephrased as

ˆ
µ
d3 xn̂µ JN (x) = 0 (102)
∂M
where the integral is over any 3-dimensional submanifold ∂M which bounds a region M in the
´
space-time. In particular, we see that the conserved charge j 0 := d3 ~xJN
0 (x), where the integral
is over an equal-time slice, is invariant, since

ˆ ˆ
d 0 3 0
j = d ~x ∂0 JN (x) =− d3 ~x ∂i JN
i
(x) = 0 (103)
dt
i (x) decays sufficiently fast in the asymp-
where the vanishing is conditioned on the fact that ∂i JN
totic region of the equal-time slice.
29
anoir r
m anoi
LLeem wellll
deM
de
axwe
Max
Figure 11: A visual representation of the charge conservation. The figure (left) shows the
presence of a charge source in the room, and by the figure (right) which shows the current as
charge moves towards the door. The change of the charge inside the room is given by the integral
of the current across the walls of the room, which is the RHS of 103 that does not vanish in this
case.
In the above, one can think of the one parameter as our , and it immediately follows that one
has n conserved currents in the presence of n-parameter continuous symmetries.
Finally, after the above somewhat abstract arguments, we will now derive the explicit form of
µ
the Noether current JN . For that, we repeat 96 for now a spacetime dependent (x). In this
case, we have
φ(x) 7→ φ(x) + ε(x) δG φ(x), ∂µ φ(x) 7→ ∂µ φ(x) + ∂µ (ε(x) δG φ(x)), (104)
and we get
ˆ
∂L ∂L
δS = d4 x (x)
(x)δG φ(x) + (x)∂µ ((x)δG φ(x))
∂φ ∂(∂µ φ)
ˆ ˆ
4 ∂L ∂L ∂L
= d x (x) (x) + (x)∂µ (δG φ(x)) + d4 x ∂µ (x) (x) δG φ(x) (105)
∂φ ∂(∂µ φ) ∂(∂µ φ)
ˆ
∂L
= d4 x ∂µ (x) (x) δG φ(x).
∂(∂µ φ)
where we have used 97, namely the fact that δG φ is a symmetry transformation, when going to
the final line. Comparing with 100, we obtain

µ ∂L
JN (x) = δG φ (x). (106)
∂(∂µ φ)
30
The above is how the theory, in terms of its Lagrangian density L, and its symmetry in terms
of the transformation δG φ(x), determines what the Noether current is. As mentioned before,
the generalisation is obvious when the theory has more than one field: one simlpy needs to sum
over the contribution to 106 from each field.
2.3 The Ward-Takahashi Identity

Noether’s theorem, as mentioned above, is an important result about the classical aspect of sym-
metries. Next we will look at how the symmetries manifest themselves quantum mechanically.
Using the same notation as in the last subsection, and assuming again that the measure is
invariant under the transformation, we have
φ(x) 7→ φ(x) + ε(x)δG φ(x)
´ ´
I[J] = Dφ exp iS[φ] + i d4 y J(y)φ(y)
´ ´
= Dφ exp iS[φ + εδG φ] + i d4 y J(y) (φ(y) + ε(y)δG φ(y))
´ ´ ´
= Dφ eiS[φ]+i d4 y J(y)φ(y) 1 + iδSε(x) + i d4 x J(x)ε(x)δG φ(x) + O(ε2 )
´ ´ ´
Dφ eiS[φ]+i d4 y J(y)φ(y) µ
d4 x ∂µ ε(x)JN (x) + ε(x)J(x)δG φ(x) + O(ε2 ) .

= 1+i
(107)
As a result, after an integration by part we have
ˆ ˆ
´
4 d4 y J(y)φ(y) µ
d x (x)∂µ Dφ eiS[φ]+i JN (x)
ˆ ˆ (108)
´
4 iS[φ]+i d4 y J(y)φ(y)
= d x (x) Dφ e J(x)δG φ(x)
1 δn
To obtain a relation among the correlators, we take in δJ(x1 )...δJ(xn ) (. . . ) J=0 on the integrands
on both sides and obtain
n
µ
X
i∂µ hJN (x)φ(x1 ) . . . φ(xn )i = δ (4) (x − xi )hδG φ(x)φ(x1 ) . . . φ(xi−1 )φ(xi+1 ) . . . φ(xn )i (109)
i=1
In the above, we have used the notation of the angular brackets
hOi := hΩ|T (O)|Ωi. (110)
Note that the Ward-Takahashi identity 109 in particular implies
µ
∂µ hJN (x)φ(x1 ) . . . φ(xn )i = 0, when x 6∈ {x1 , x2 , . . . , xn } (111)
31
when there are no coincident insertions. This is then indeed the direct analogue of Noether’s
theorem, and now in terms of the correlators– a distinctively quantum concept. The appearance
of the contact terms, embodied by the Dirac delta functions, is then again a typical manifestation
of the fact that the quantum effects appear at short distances.
2.4 The energy-momentum tensor

The object Tµν , or equivalently T µν , is often referred to as either the energy-momentum tensor
or the stress-energy tensor. The first name comes from the fact that the conserved charges 103
associated to Tµν are precisely the energy
ˆ
E := d3 ~x T 00 (x)
and the momenta ˆ

i
P := d3 ~x T 0i (x).
Moreover, T ij represents the flux of i-th component of linear momentum across the surface
perpendicular to the j-th direction, which is the stress. This is where the second name of T µν
comes from.
So where does this tensor come from? The 4 × 4 entries correspond to the four components of
the four conserved Noether currents, each of them associated to a shift symmetry
xµ 7→ xµ + aµ , (112)
while xν 7→ xν for µ 6= ν, where aµ is a constant and hence corresponds to a global symmetry. As

before, in order to compute the Noether currents, we promote this to a local transformation. We
write xµ 7→ xµ +f µ (x), where f µ (x) is a four-vector that is position-sensitive. The corresponding
transformation of the field φ(x) and its derivative are
φ(x) 7→ φ(x) + (f µ ∂µ φ) (x),

(113)
∂ν φ(x) 7→ ∂ν φ(x) + ∂ν (f µ ∂ µ φ) (x).
As usual, we assume that the Lagrangian density depends on the field and its derivative, i.e.
L = L[φ, ∂φ]. By definition, we have on-shell δS|φ=φclas = 0. Following the procedure when
deriving the Noether current, we compute the variation of the action under shift symmetry
xµ 7→ xµ + f µ :
32
´
∂L µ ∂L
δf µ S = d4 x ∂φ f ∂µ φ + µ
∂(∂ν φ) ∂ν (f ∂µ φ) (x)
 
´  ∂L ∂L

∂L
d4 x
 µ µ
= f ∂µ φ + ∂µ ∂ν φ +∂ν f ∂(∂ν φ) ∂µ φ (x)

 ∂φ ∂(∂ν φ) 
| {z } (114)
= ∂µ L[φ,∂φ]
´
= d4 x f µ ∂µ L + ∂ν f µ ∂(∂∂L
ν φ)
∂ µ φ (x)
´
∂L
= d4 x ∂ν f µ (x) −δ νµ L + ∂(∂ν φ) ∂µ φ (x) + boundary term.
As a result, we define the (canonical) energy-momentum tensor as follows
∂L
T (c)νµ ≡ −δ νµ L + ∂µ φ . (115)
∂(∂ν φ)
satisfying
ˆ ˆ
δf µ S = d4 x ∂ν f µ T (c)νµ = − d4 x f µ ∂ν T (c)νµ . (116)
Note that the idea of the calculation of the current is the same but the concrete steps are different
from (97). This is because the shift symmetry is not a symmetry realized at the level of the
Lagrangian density; the invariance arisis when one integrates L into the action.
Since δS = 0 on-shell (irrespective of whether the field variation is a symmetry of the theory
or not), we obtain the following expression that tells us that the energy-momentum tensor is
(classically) conserved,
∂ν T (c)νµ (x) = 0. (117)
Similarly, we can obtain the quantum analoge of the above, by evoking the Ward-Takahashi
ν =T (c)ν
identity 109, applied to the current JN µ associated to the shift symmetry xµ 7→ xµ + aµ ,
as
n
X
i∂µ hT µν (x)φ(x1 ) . . . φ(xn )i = δ (4) (x − xi )hφ(x1 ) . . . φ(xi−1 ) (∂ ν φ) (xi )φ(xi+1 ) . . . φ(xn )i.
i=1
(118)
(c)ν
Now, we would like to make two remarks regarding Tµ . First, we have
∂L ∂L
T (c)νµ = T (c)µν iff ∂ρ φ g µρ = ∂ρ φ g νρ .
∂(∂ν φ) ∂(∂µ φ)
33
This is the case when, for example, the only term in the Lagrangian involving ∂φ is proportional
to ∂µ φ∂ν φ g µν . In particular, there are theories for which T (c)νµ 6= T (c)µν . Second, there is some
ambiguity in the definition of the stress-energy tensor: the equation 116 which we used to define
the energy-momentum tensor is invariant under
T µν 7→ T 0µν = T µν + ∂λ M λµν
for any M λµν = −M µλν . In particular, using this additional freedom we can always make T µν a
symmetric tensor. In particular, the conservation of T µν is not affected the above change since
∂µ T µν = ∂µ T 0µν .
A systematic way to unambiguously define the stress-energy tensor and moreover to obtain a
symmetric tensor is to consider the quantum field theory on a general, fixed (curved) space-time
with metric gµν . By working with fixed spacetime, we will not take into account the back-
reaction of the quantum fields on the metric.
As a result, instead of using η µν to contract the indices, we will use the general inverse metric
g µν . In particular, the changes in the action include
√
measure : d4 x 7→ d4 x −g,
(119)
metric : η µν ∂ µ φ∂ν φ 7→ g µν ∂µ φ∂ν φ.
In this setup, the stress-energy tensor will be obtained by considering the variation of the action
with respect to a change of the metric
gµν (x) 7→ gµν (x) + δgµν (x).
To see why such a change of metric is related to the shift symmetry we discussed earlier, note
that under a shift xµ 7→ x̃µ = xµ − f µ (x), the metric changes as follows
∂xα ∂xβ
g̃µν (x̃) = gαβ (x), (120)
∂ x̃µ ∂ x̃ν
which, in the leading order of f µ , is equivalent to
g̃µν (x) = gµν (x) + 2∇(µ fν) . (121)
Hence, a shift, or diffeomorphism transformation, indeed leads to a change of metric. As a

result, varying the action with respect to a change of metric gives the answer of the change of
action when performing a shift transformation. That said, the transformation x 7→ x̃(x) can of
course be more general than just a shift. Rotations, for instance, are also fair games.
We now define the energy-momentum tensor in terms of the variation of the action functional
as
34
2 δS
T µν (x) ≡ p , (122)
|g| δgµν (x)
which is manifestly symmetric : T µν = T νµ .

Note that the definition 122 and 121 leads to
ˆ
δS = d4 x∇µ fν T µν (123)
which indeed reduces to 116 in flat spacetime.
Now making 122 more explicit, using

ˆ
√
S= d4 x −gL[φ, ∂φ] (124)
and the fact that √ √

δ −g −g µν
=− g ,
δgµν (x) 2
we obtain
δL
T µν = 2 − gµν L . (125)
δgµν
As mentioned above, energy-momentum tensor defined above captures the variation with respect
to shift symmetries, and much more. In particular, we consider the change of the metric of the
form of a local rescaling
δgµν (x) = s(x)gµν (x), (126)
often called a Weyl transformation. Note that such transformations preserve the angle.
Note that, above from the shift symmetries, diffeomorphisms that incur such Weyl transforma-
tions include the scaling transformation xµ 7→ x̃µ = λxµ , and the special conformal transforma-
tions, δxµ = bµ x2 − 2bν xν xµ , generating the conformal group SO(2, 4) of the Minkowski space.
Putting 126 and 122 together, we obtain that under a Weyl transformation
ˆ ˆ
1 √ 1 √
δS = d4 x −g s(x)Tµν g µν (x) = d4 x −g s(x)T µ µ . (127)
2 2
This gives us an all important result- a theory posseses Weyl symmetry (also referred to confor-
mal symmetry) when the energy-momentum tensor is traceless, i.e.
T µ µ (x) = 0. (128)
2.5 Spontaneous Symmetry Breaking

In QFT, by a symmetry we mean a symmetry of the action. Then spontaneous symmetry break-
ing refers to a situation where this symmetry is not preserved by the vacuum. Necessarily, in
this situation, there are more than one vacuum, and these degenerate vacua are related by the
symmetry transformation.
35
For the fun of it, we will describe two real life examples to illustrate the idea. Imagine a seesaw
in a playground. It is equally possible for an empty seesaw to tilt to one side or the other,
since a good seesaw is built in a left-right symmetric way. That said, most of the time we find
the seesaw in an empty playground to be tilting either one way or the other. This “state” of
the seesaw hence breaks the “intrinsic” symmetry it possesses. Another example: Halibut vs.
Flounder5 . The Halibut is a variety of flatfish that lives on the sea floors that start off with
symmetric eyes, i.e. one eye on either side of the head, but end up with both eyes on the right
side of the head. This migration of the eyes happens during the larval metamorphosis when the
fish is six months old. Natural selection has enabled the Halibut to be a right-eye fish in order
for it to camouflage itself with the sea or ocean floor to escape predators. Now, consider the
Flounder which is another type of flatfish that thrives on the ocean floor. Just like the larval
Halibut, the Flounder too is born with symmetric eyes but ends up with both eyes on the left
of the head during metamorphosis. Of course, it doesn’t matter which side of the head the eyes
migrate to since both the fish are able to camouflage just fine! What is important to notice is
that the choice of spontaneously choosing one orientation over the other is what differentiates
the species.
Next we turn to an important and familiar physical example, namely that of a a ferromagnet,
Figure 12: Halibut vs. Flounder: During the larval metamorphosis, the eyes of the Halibut
undergo SSB to end up on the right side of its head while that of the Flounder undergo SSB to
end up on the left side of its head, all while starting from a degenerate symmetric state.
such as iron. This system is described by the Heisenberg Hamiltonian
X X
H = −J S ~j + ~h ·
~i · S ~i ,
S (129)
<i,j> i
and this describes an N ferromagnetic Heisenberg spins {Si }, and ~h denotes the constant external
magnetic field. The parameter J is positive, the sum is over the nearest neighbor pairs hi, ji. It
is easy to write a partition function for this model as Z = Tr e−βH , where β = 1/T . The free
energy F is defined by exp(−βF ) = Z. Note that there is a rotation symmetry
~i 7→ R S
S ~i for all i,
where R represents any SO(3) rotation, in the absence of the magnatic field, i.e. ~h = 0. Above
the critical point T > Tc , the spins point at random directions and the magnetisation
5
inspired by the Arnold-Sommerfeld colloquium given by Hitoshi Murayama on symmetry breaking.
36
~ ≡ 1
X
M ~i i
hS (130)
N
i
~ is called an order
vanishes. In this phase, the SO(3) symmetry is not broken. The vector M
parameter for this phase transition, since it helps to describe the phase transition between the
ordered phase (T < Tc ) where M~ 6= 0 and the disordered phase (T > Tc ) where M ~ = 0. Indeed,
~ | 6= 0, and the precise value of it can
when T < Tc we have in the minimal energy state |M
be determined in terms of the temperature and J. However, the direction is not, and we get
~ =µ
a family of minimal energy states, with M ~ with different µ
~ with the same magnitude. In
other words, we have vacua that are labelled by points on S 2 . Note that the SO(3) symmetry
is (partially) spontaneously broken. To be more concrete, the SO(3) symmetry is broken to the
SO(2) symmetry the remains which corresponds to the rotations around the axis aligning in the
direction of µ
~ . For the sake of concreteness, we choose µ
~ = µẑ to be lying in the z-direction,
then the SO(2) symmetry that is unbroken is exactly the rotation symmetry around the z-axis.
Around such a minimal energy state, we can expand the spin as
~i = µ
S ~ + ~σi . (131)
In field theory language (in the continuum limit), the field ~σ (x), which apart from massive modes
has also zero modes corresponding to moving to another nearby minimal energy state. In the
following subsection we will analyze this phenomenon quantitatively.
2.5.1 Spontaneous breaking of discrete symmetries: a Z2 example
Let us consider now the φ4 -theory, where φ : R3,1 → R is a real scalar, which has as Lagrangian
1 1 λ
L = ∂µ φ∂ µ φ − m2 φ2 − φ4 , (132)
2 2 4!
where λ is a positive number. This Lagrangian possesses a Z2 symmetry, acting as φ 7→ −φ.

When m2 > 0, the classical configuration has φ = 0, which is invariant under the Z2 symmetry.
When m2 < 0, however, the classical configuration must satisfy
r
2 2 6|m2 |
φ(x) = v , v = (133)
λ
which leads two degenerate vacua |Ω+ i and |Ω− i, satisfying
Z
2
|Ω+ i ←→ |Ω− i, hΩ± |φ|Ω± i = ±v. (134)
Clearly, when m2 < 0, the symmetry is spontaneously broken by the vacua.
We can now make contact to the ferromagnet example encountered earlier. It was Ginzburg and
Landau who argued that the above model offers to an effective description of a ferromagnet (with
1d spins) near criticality, with coefficients m2 and λ being temperature-sensitive. In particular,
near the critical point we may identify m2 (T ) = c(T − TC ), where c is a positive constant.
37
At this point, one might worry that the theory is acausual, as a negative square mass implies
a spacelike momentum , i.e. m2 < 0 ⇒ p2 < 0, which corresponds to a particle moving at a
superluminal speed. Consequently, negative mass-squared particles are called tachyons which
came from to the word “fast” in Greek. But don’t worry! We will now argue that there is
no breaching of causality in this theory. What happens in the broken phase with m2 < 0 is
simply that φ = 0 is not a good place around which to expand the field, since it is far away
from either of the vacua |Ω± i. As a result, the expansion near φ = 0 does not really correspond
to a “particle”, let alone a superliminal particle. To see this explicitly, let us look at the same
Lagrangian, but now expanded near the chose vacuum |Ω+ i. To do this, write φ = φ̃ + v and
rewrite the Lagrangian in terms of φ̃:
L = 21 ∂µ (φ̃ + v)∂ µ (φ̃ + v) − 12 m2 (φ̃ + v)2 − λ

4! (φ̃ + v)4
q (135)
1 µ λ λ 4 3m4
= 2 ∂µ φ̃∂ φ̃ − |m2 |φ̃2 − 2 3
6 |m |φ̃ − 4! φ̃ + 2λ .
The mass-squared for the new field φ̃, which does represent a perturbation around the classical
√ p
solution, is now manifestly positive and takes the value mφ̃ = 2 |m2 | and hence, there are
indeed no tachyons! The same conclusion would hold if we pick φ = −v as the classical solution
to expand around. Note that it might appear that the Z2 symmetry is no longer present in the
Lagrangian 135, but this is not possible since all we have done is a change of variable φ = φ̃ + v.
In terms of φ̃, the Z2 symmetry acts as φ̃ 7→ −φ̃ − 2v. Hence, we conclude that the symmetry is
realized non-linearly when we expand near a vacuum which breaks the symmetry spontaneously.
This is a common feature of spontaneous symmetry breaking.
2.5.2 Spontaneous breaking of global symmetries and Goldstone’s theorem
Earlier, we have seen the spontaneous symmetry breaking in the ferromagnet. We claim that,
by expanding the spin near the value of the magnetisation, we obtain in the field theory limit
massless modes corresponding to moving towards a nearby vacuum. In this section we will see
how this comes about as a general phenomenon when a continuous global symmetry is sponta-
neously broken.
To see the appearance of a massless particle, we turn to the Ward-Takahashi identity 109 for
one field insertion
µ
i∂µ hJN (x)φ(x1 )i = δ (4) (x − x1 )hδG φ(x)i.
When integrated over spacetime, we have

ˆ ˆ
µ µ
i d4 x ∂µ hJN (x)φ(x1 )i = d3 x n̂µ hJN (x)φ(x1 )i = hδG φ(x1 )i, (136)
∂M
where ∂M indicates that the integration is over the boundary of the manifold M.
38
Note that the RHS does not vanish when the vacuum does not preserve the symmetry. In other
words, when the symmetry is spontaneously broken we have
ˆ
µ
d3 x n̂µ hJN 6 0.
(x)φ(x1 )i = ihδG φ(x1 )i = (137)
∂M
Is it important to note that this integral is non-vanishing even when all points of the boundary
x ∈ ∂M is extremely far away from the insertion location x1 . Such long range correlation is
only possible when there are massless particles (associated to the symmetry). We will call such
particles the Nambu-Goldstone (NG), or simply Goldstone, bosons.
Goldstone Theorem
Spontaneous breaking of a one-parameter continuous symmetry implies the existence of
a massless particle.
More generally, the spontaneous breaking of a group of symmetry with N parameters to a

subgroup with with N − n parameters leads to n massless particles. More concretely and more
mathematically (don’t panic if you don’t understand this sentence), if a theory has symmetry
group G while a chosen vacuum breaks it to a subgroup H ⊂ G, then the Goldstone bosons
parametrise the tangent space of the homogeneous space, the coset G/H. As a concrete example,
recall that in the ferromagnetic phase, the magnetization M~ =µ ~ breaks the SO(3) symmetry
to SO(2). Note that
SO(3)/SO(2) ∼
= S2,
~ | = |~
which also is the manifold parametizing all vacua, which are specified by |M µ|, we see that
there are two Goldstone bosons in this case, corresponding to moving towards a nearby vacuum.
2.5.3 The linear sigma model
The linear sigma model is the simplest relativistic model where there is a spontaneous breaking
of continuous symmetry. Consider a complex scalar field φ(x) with the Lagrangian
λ 4
L = (∂µ φ)(∂ µ φ∗ ) + m2 φφ∗ − |φ| (138)
4
where λ > 0. We focus on the broken phase, where m2 > 06 . The minima of the potential are
where φ satisfies r
2m2
|φ|2 = v 2 , where v := . (139)
λ
See Figure 13. Clearly, the theory possesses a U (1) (continuous) symmetry, acting by φ(x) 7→
eiθ φ(x) with constant θ, and so is the set of the vacua {φ| |φ|2 = v 2 } too. However, individual
choice of such a vacuum does break the symmetry: from U (1) to the trivial group (aka nothing)
in this case. To be more precise, we denote the one-parameter family of vacua by |Ωθ i, which
satisfies hΩθ |φ|Ωθ i = veiθ . Around such a vacuum, we reparametrize the complex scalar in terms
of the “phase” degree of freedom (π(x)) and the “radial” degree of freedom (σ(x))
6
Note that in the previous subsections we wrote the mass term of the Lagrangian as −m2 φ2 /2 and consider
2
m < 0. Of course this is completely equivalent.
39
π(x)
1 i +θ
φ(x) = v + √ σ(x) e Fπ (140)
2
where Fπ is a normalizing positive constant.
It is clear from the picture (see Figure 13) that while the radial excitation brings φ out of
the minima of the potential (and hence should be massive), the excitation of the phase degree
of freedom just moves the vacuum to an neighbouring vacuum, and should be expected to be
massless. In what follows we will demonstrate this explicitly. For concreteness, we choose the
Figure 13: the potential of the linear sigma model, often referred to as the Mexican hat potential.
The masses squared of particles are given by the second derivatives of the potential. Expanding
around the origin, there are two tachyonic (negative mass-squared) modes (long-dashed line).
Expanding around a minimum, there is one massive mode, corresponding to excitations along
the radial direction (σ(x)), and one massless mode (solid line), corresponding to excitations
along the symmetry direction where the potential is flat (π(x)).
π(x)
vacuum corresponding to θ = 0, and plug φ(x) = v + √1 σ(x)
2
ei Fπ into the action, and obtain
r !
2m2 1√ m4

1 1 1 1
L = (∂µ σ)2 + + √ σ(x) 2
(∂µ π) − 4
λσ + 3 2 2
λmσ + m σ − . (141)
2 λ 2 Fπ2 16 2 λ
√
We immediately see that the field σ(x) has a positive mass-squared, mσ = 2m, while the π(x)
fields are massless, m2π = 0.
There is a more instinctive way to prove that π must be massless, which we will now present.
The symmetry transformation φ(x) 7→ eiθ φ(x) is equivalent the transformation
π(x) 7→ π(x) + Fπ θ, (142)
while σ(x) remains invariant. Note that a non-vanishing mass term m2π π(x)2 is not invariant
under the above shift, hence we must have mπ = 0 in order to retain the symmetry of the
40
Lagrangian.
Finally, note that the above linear sigma model is equivalent to a theory with two real scalars,
through the parametrization φ(x) = φ1 (x) + iφ2 (x). This can be generalised to the theory with
N real scalars
N
X 1 2 λ 2 2
L= (∂µ φi ) + m2 φ2i − φ , (143)
2 4 i
i=1
which has an O(N ) global symmetry of rotations that gets spontaneously broken to O(N − 1)
by a vacuum. The number of massless Goldstone bosons is then equal to the dimension of the
∼ S N −1 .
coset O(N )/O(N − 1) =
2.6 The Higgs mechanism

The Goldstone theorem, in a nutshell, states that broken global symmetry implies massless spin
zero Goldstone bosons. Here, the word “global” is the keyword, as the same will not hold if
the symmetry was a local one, signified by the presence of an accompanying gauge field. The
spontaneous symmetry breaking in that case will see the would-be Goldstone boson disappears,
or “eaten” by the gauge field, and the gauge field on the other hand will acquire a mass in
this “eating” process. One way to understand it is to remind ourselves that gauge symmetries,
unlike global symmetries, are not real symmetries to start with, but rather signify a (convenient)
redundancy of the physical degrees of freedom. Hence, perhaps the right way to think about
it, when we rid ourselves of the historical context, is that there was no Goldstone boson to be
eaten to start with.
We will illustrate this using a simple example shortly.
2.6.1 The Abelian Higgs model
To illustrate the idea of the Higgs mechanism, let’s consider the so-called Abelian Higgs model,
with Lagrangian
1 λ
L = − Fµν F µν + (∂µ − ieAµ ) φ∗ (∂ µ + ieAµ ) φ + m2 φφ∗ − |φ|4 . (144)
4 4
This is nothing but the linear sigma model with an additional U (1) gauge field, under which the
complex scalar is charged. Note that the global U(1) symmetry has now been promoted to the
local symmetry:
1
φ(x) 7→ ei(x) φ(x), Aµ 7→ Aµ − ∂µ (x). (145)
e
To simplify the expressions, by adding an irrelevant constant to the Lagrangian7 we can rewrite
it as 2
v2

1 λ
L = − Fµν F µν + |Dµ φ|2 − 2
|φ| − (146)
4 4 2
7
Adding a constant to the Lagrangian changes nothing of the physics since it gets cancelled out in I[J]/I[0].
In particular, it clearly does not change the equations of motion.
41
where v 2 = 4m2 /λ. As before, without loss of generality we can expand around the minimum
φ = v of the Mexican potential:

v + σ(x) π(x)
φ(x) = √ ei Fπ . (147)
2
Expanding the Lagrangian in terms of σ(x) and π(x) as before, we obtain

v+σ(x)
2
∂µ π
√
L = − 14 Fµ F µ + 12 ∂µ σ∂ µ σ + √
2
( Fπ + eAµ )2 − m2 σ(x)2 + 21 λmσ(x)3 + 1
16 λσ(x)
4 .
(148)
where we have again discarded the irrelevant constant term.
√
As before, we obtain a massive scalar σ(x) with mass given by mσ = 2m. What used to be the
Goldstone boson, the π(x), on the other hand, can be made disappear by exploiting the gauge
symmetry 145 and choosing the unitary gauge:
1
∂µ π(x) 7→ 0, Aµ 7→ A0µ := Aµ + ∂µ π(x). (149)
Fπ e
In this gauge, the action simplifies into

2
1√

1 v + σ(x) 1 1
L = − Fµ F µ + e 2 √ A0µ A0µ µ 2 2
+ ∂µ σ∂ σ − m σ(x) + 3 4
λmσ(x) + λσ(x) .
4 2 2 2 16
(150)
Now we can explicitly see the Higgs mechanism : the would-be Goldstone boson in the global
case, π(x), has been “eaten” by the gauge field Aµ though the gauge transformation. At the
same time, the gauge field has acquired a mass mA = ev.
It might help to see this more clearly by working with the so-called decoupling limit, in which
√
m, λ → ∞ while keeping v = 2m/ λ fixed. In this limit, the field σ(x) is infinitely heavy and
decouples, and as a result we are left with the Lagrangian
1 e2 v 2 0 0µ
Ldecoupling = − Fµ F µ + Aµ A (151)
4 2
which simply describes a massive U (1) field.
2.7 Further Reading
References
[1] Schwartz, M. D. (2014). Quantum field theory and the standard model. Cambridge University
Press.
sity Press.
42
The most relevant parts are: Schwartz Chapter 28 (in particular 28.1-28.3), 3.3 and Banks
Chapter 7 (in particular 7.1-7.2).
43
Topic-III: Wilsonian Renormalization Group
(in Real Space)

44
3 Topic-III: Wilsonian Renormalization Group (in Real Space)
3.1 The Idea
The idea of RG relies on the so-called separation of scales in physics: high energy, microscopic
physics is not relevant for the study of long range physics. An example: to describe the flow of
the river we do not need to study the movement of the individual electrons in the atoms of the
water molecule. Standard model cannot replace fluid dynamics, and whether there is supersym-
metry at high energy or not is not directly relevant for the engineering of new materials; the
reason is that they describe physics at very different length scales.
More formally speaking, the reason why the study of renormalization group is possible and useful
is the assumption of locality in energy scales: the degrees of freedom at different scales interact
locally and we can use a theory that is correct up to the energy scale Λ to describe physics that
we are interested in at the energy scale E as long as E << Λ.
Such a theory, which is capable to describe the physics of energy scale up to Λ, is said to be a
effective field theory with UV cutoff Λ. Practically all quantum field theories in physics, from
theories describing critical phenomena to the standard model, are effective theories. For in-
stance, at present we know of no traditional quantum field theories that are totally valid up to
the Planck scale.
Renormalization group (RG), which is importantly not a group mathematically as it is not

invertible, is the procedure to obtain from a theory with UV cutoff Λ with Lagrangian L({g}; Λ),
another theory with UV cutoff Λ0 < Λ with Lagrangian L({g 0 }; Λ0 ). In particular, in generic
situations, the coupling constants g 0 have changed in this process. From this point of view, we
can view RG flows as flows in the space of all possible couplings, the {g}-space, described by the
beta function
dg
βg := Λ . (152)
dΛ
As always when one has such a system of differential equations as in 152, a point of particular
dg
interest is its fixed point , where Λ dΛ = 0 and a change of scale does not change the theory:
L({g∗ }; Λ) = L({g∗ }; Λ0 ). A dramatic illustration of the fact that RG is not a group is the
phenomenon of universality, which states that theories with very different behaviours at high
energy, say L({g1 }; Λ) and L({g2 }; Λ), can flow via RG group to the same low energy theory
L({g∗ }) which is the fixed point theory.
This fact is crucial in the application of QFT to study critical phenomena. When the sys-
tem approaches the critical point, the correlation length approaches infinity and the scale is no
longer a relevant quantity. As a result, the system at criticality is described by an RG fixed point.
Another practical application of RG flow is the following. The fact that a quantum field theory
45
is an effective theory valid only up to a certain energy scale is often reflected in the fact that a
naive application of the theory leads to infinities in the correlators. As we will see in Topic 4,
RG flow then provides a way to “remove the infinities” and makes it possible to extract finite
predictions that can then be compared with the experiments.
Below we will introduce the modern version of the RG developed by Kadanoff and Wilson.
In this topic we will first study the topic in the context of lattice theories, which, apart from
describing discrete systems such as spin systems can also be seen as a way to approximate (or
regularize) the infinite dimensional integrals in a quantum field theory. In this context, an RG
steps involves studying the change of the theory under a change of length scale, x 7→ x0 = x/b
with some b > 1, or, equivalently, under a rescaling of the lattice spacing a → a0 = ba, i.e.
a → a0 = ba ⇔ x → x0 = x/b for some b > 1. (153)
In particular, the correlation length transforms as
ξ 7→ ξ 0 = ξ/b. (154)
In momentum space, one can think about the above transformation as integrating out the degrees
of freedom with energy between Λ0 ∼ 1
ab and Λ ∼ a1 .
3.2 RG on Lattices
Here we describe a concrete lattice system, which serves as a simple example that you can recall
when studying RG on lattice systems more abstractly.
Consider a statistical system defined on a d-dimensional square lattice with Ld lattice sites of
lattice spacing a, with the degree of freedom attached to each lattice vertex i being the spin
si ∈ S = {±1} placed on its sites. Let H({si }, {g (k) }) be the Hamiltonian of the system, where
gk are the coupling constants of the various interactions among the spins si . The most general
Hamiltonian can be split into an odd and even sector, under the global transformation si 7→ −si
for all lattice sites i. The Hamiltonian in the even sector is simply written down as a linear
combination of all even-powered terms with a corresponding coupling constant as follows
X X
H({si }, {g (k) }) = g2,ij si sj + g4,ijk` si sj sk s` + . . . (155)
i,j i,j,k,`
Usually we consider Hamiltonians that are invariant under the symmetry of the lattice. For
instance one might require g2,ij to depend only on the distance |ij| between the vertices i and j,
when the periodic boundary condition is imposed. The partition function can be readily written
down as follows
X
Z({gk }) = e−H({si },{gk }) , (156)
{si }
46
d
where the sum is over S ⊗L , namely the set of all possible spin configurations. The inverse-
temperature, i.e. the factor β = 1/kB T , is included in the definition of the coupling constants.
With this prototype lattice theory in mind, we will now study in two steps how RG is performed
on lattices.
3.2.1 Step one: reducing the degrees of freedom
Starting with a finite lattice with Ld lattice sites, changing the lattice spacing a 7→ ab renders a
smaller lattice with L0d lattice sites, where L0 = L/b8 , which amounts to a reduction of degrees
of freedom. Denoting the spins by σi , we would like to find a “transformation map”
σk0 = fk ({σi }) (157)
mapping a configuration of the original spins to a configuration of the new spins on the lattice
with smaller number of lattice sites.
The first property of the spin transformation that it retains the locality property of the lattice.
In other words, the new spin σk0 depends only on a group of original spins σi ’s that are not too far
away from the site of σk0 (in the new, reduced lattice). To be more concrete, the transformation
a 7→ ba divides the original lattice into blocks, as shown in 14 denoted by Bk , each of them made
of bd spins. Each block leads to a single new spin σk0 . We then require that the value σk0 takes
only depends on the original spins {σi }i∈Bk , in the corresponding block Bk .
One can repeat the step many times. To summarize, we have the spins in the (n + 1)-th step
determined by the spins in the previous step by
(1)
σi = f ({si }), with i ∈ Bk . (158)
As examples, we now discuss two commonly used forms of the transformation function f –
decimation and the majority rule. Decimation9 refers to the RG transformation in which the
(n+1) (n)
spin σk is assigned the value of one of the spins σi in the block Bk , i.e.
(n+1) (n)
σk = σj , for some j ∈ Bk . (159)
(n+1) (n)
For instance, for b = 3 and d = 1, we can let σk = σ3k−1 . The other transformation law
(n+1)
is called the majority rule, and as the name suggests we assign the value of σk to be the
(n)
value of the majority of the spins σi in the block Bk . For instance, for a spin-1/2 system with
(n+1) (n) (n) (n)
b = 3 and d = 1, we assign σk = 1 if 2 or 3 out of the three spins σ3k−2 , σ3k−1 , σ3k have
(n+1)
value 1, and let σk = −1 otherwise. Once we have decided on a particular choice for the
transformation law, we can now introduce the so-called transformation operator which is defined
8
We assume b is a positive integer dividing L.
9
The term decimation is used to describe the elimination of a subset of degrees of freedom labeled by points
in space. It was used in ancient Rome as a punishment in which a tenth of an army unit, chosen by lots, was
sentenced to death, the sentence often to be carried out by the other nine-tenths.
47
Figure 14: A blocking transformation: from the original lattice, with lattice spacing a and
variables si , to a new lattice with a0 = ba and block spins σi .
as follows

(n+1) (n)


1, if σk = f ({σi })

(n+1) (n)
T σk , {σi }i∈Bk = . (160)



0, otherwise
More generally, the operator does not need to take value in {0, 1} but can give us the conditional
(n+1) (n)
probability of having σk taking a specific value for a given configuration {σi }, i ∈ Bk . In
particular, it takes value in [0, 1] and satisfies

(n+1) (n)
X
T σk , {σi }i∈Bk = 1. (161)
(n+1)
σk
What we have just done here is to replace the lattice with a coarser lattice with fewer vertices,
and defining the new spins according to the above transformation from of the original spins.
3.2.2 Step two: new effective Hamiltonian
Once we have made a choice for the transformation law, we can determine the effective Hamil-
(n+1) (n+1)
tonian H (n+1) ({σi }, {gk }) for the new spins. Since the transformation T depends only
(n) (n−1)
on the configurations of the spins σi and not on the spins σ` of the previous steps10 , the
(n) (n)
new Hamiltonian will be determined by the nth step Hamiltonian H (n) ({σi }, {gk }) as follows
10
This is a property that sometimes is referred to the “Markov process” property : the future depends on the
present and not on the past.
48
h i X Y (n+1) (n) h i
(n+1) (n+1) (n) (n)
exp −H (n+1) {σk }, {gk } = T σk , σi exp −H (n) {σi }, {gi } .
} blocks
(n)
{σi Bk
(162)
(n+1)
Through the above formula, we determine the form of H (n+1) as well as the values of gk , for
a given H (n) and T . The idea behind 162 is the conditional probability identity, i.e.
X
P (B) = P (B|Aj )P (Aj ), (163)
j
where P (B|Aj ) is the probability of B happening given that Aj has happened. In this case,
P (Aj ) is the Boltzmann distribution
(n) 1 h
(n)

(n) (n)
i
P ({σi }) = (n)
exp −H {σ i }, {gi } ,
Z
and the transformation operator T plays the role of the conditional probability P (B|Aj ). When
1
exp −H (n+1) , the new Hamiltonian

expressing P (B) in a form of Boltzmann distribution Z (n+1)
162 follows. In particular, it is now easy to see that the partition function is invariant
h i
(n+1) (n+1) (n+1)
X
Z (n+1) ({gk }) = exp −H (n+1) {σk }, {gk }
(n+1)
{σk }
 
h i
(n+1) (n) (n) (n)
X  X Y
(n)
= T σk , σi exp −H {σ }, {g }

  i i (164)
} blocks
(n) (n+1)
{σi } {σk B k
h i
(n) (n) (n)
X
(n)
= exp −H {σi }, {gi } = Z (n) ({gk }),
(n)
{σi }
where we have used 161.

(n)
It is easy to generalize this approach to observables. Consider an operator X (n) ({σk }) which
satisfies

(n+1) (n+1) (n) (n)
X Y
X (n+1) ({σi }) T σk , σi = X (n) ({σk }), (165)
(n+1)
{σi } blocks
Bk
then the equality holds for the expectation value of the observable, i.e.
h i
1 (n+1) (n+1) (n+1)
X (n+1) ({σk })exp −H (n+1) {σk
P
hXi = Z (n+1)
}, {gk }
(n+1)
{σi }
(166)
h i
1 (n) (n) (n)
X (n) ({σi })exp −H (n)
P
= Z (n)
{σi }, {gi } .
(n)
{σi }
49
In general, the complicated rule 162 generates new couplings at each step. For instance, starting
P
with the two-spin Hamiltonian H = g2,ij σi σj , it is generically impossible to keep this form of
i,j
the Hamiltonian, as terms such as g4,ijk` σi σj σk σ` , g6,ijk`mn σi σj σk σ` σm σn , etc. will be generated
upon applying 162. As a result, a priori we can associate to an RG trajectory a trajectory in
the space of coupling constants, parametrised by
n o
(n) (n) (n)
g (n) = g2 , g4 , g6 , . . . ,

(167)
an ensemble of all possible coupling constants that are compatible with the symmetry of the
model. In practise, we often truncate the manifold to a finite set, described by
(n) (n)
{g (n) } = {g2 , . . . , g(2N ) }, (168)
for some finite N . We can summarize the motion of the point {g} in one discrete RG step by
{g (n+1) } = R({g (n) }), (169)
where R is a complicated non-linear transformation. Starting fro a point {g (0) } and applying
169, the point of the system evolves in the sequence {g (1) }, {g (2) }, . . ., giving rise in this way to
(discrete version of) an RG trajectory like the ones shown in figure 15. The continuous version
is given by the beta function differential equation 152. All points on the RG trajectory describe
the same physics, but just at different length scales; all we are doing is just using a different
description of the same low-energy physics with a different UV cutoff.
Figure 15: Trajectories of the RG and fixed points: A is a repulsive fixed point, B is an attractive
fixed point, whereas C is a mixed fixed point.
50
3.3 RG fixed points
When studying the RG trajectories governed by 169, of special interests are the fixed points of
the RG transformation
{g∗ } = R{g∗ }. (170)
Recall that under this transformation, the correlation length transforms as 154, namely
ξ({g (n+1) }) = b−1 ξ({g (n) }), (171)
and at a fixed point, g = g∗ , the correlation length either diverges or vanishes, satisfying
ξ({g∗ }) = b−1 ξ({g∗ }). (172)
The fixed points where ξ → ∞ are called critical points, while those where ξ = 0 are called
trivial fixed points. It is called a trivial fixed point since zero correlation length means there is
no coupling between fields.
We are also interested in the more detailed properties of the critical points, in particular whether
they are attractive, replusive, or mixed. We will first explain what we mean using Figure 15,
where we see that in the immediate vicinity of point A, all lines emanate outward and hence,
A is called a repulsive fixed point. On the other hand, we notice that all the points in the
immediate vicinity of point B converge towards and hence, this is called the attractive point.
Finally, we notice a mixed behaviour for the point C where there are trajectories ending in and
emanating from it and hence the name mixed point.
We will now analyse this in terms of equations, by considering an infinitesimal displacement

from the fixed point
g = g∗ + δg
and studying its subsequent movement under RG transformations. In order to do this, let us
spell out 169 in components. Recall that {g} stands for a N -tuple of coupling constants, in the
example of 168. Using gα , gβ etc to denote the components (for instance the two-spin, four-spin
interaction coupling constants g2 , g4 , etc.), we write 169 in components as
gα(n+1) = Rα ({g (n) }) (173)
where the RHS depends on all the N -tuple of couplings. Starting from a point g = g∗ + δg near
a fixed point, we have, to the first order in δg,
(n+1) ∂Rβ
gβ = Rβ (g∗ + δg) = Rβ (g∗ ) + δgα = g∗,β + Rβα δgα . (174)
∂gα
51
where we have written
∂Rβ
Rβα := (175)
∂gα
From this it is clear that whether RG brings the point g∗ + δg further away or closer to the fixed
point is determined by the property of the matrix Rβα .
Denoting by v i and λi the left eigenvectors and the corresponding eigenvalues of the matrix11 :
X
v i,β Rβα = λi v i,α . (176)
β
Using the eigenvectors we can now define a near coordinates for the space of couplings, namely
X
ui := v i,β gβ . (177)
β
In terms of this new bases, the infinitesimal RG transformation 174 reads
0
ui = v i,β Rβ (g + δg) = v i,β g∗,β + v i,β Rβα δgα = ui∗ + λi δui .
P P P
(178)
β β α,β
With b as the rescaling parameter of the block spins a 7→ ab, it is common practise to param-
eterize the eigenvalues as λi = byi , or yi = log λi / log b. Putting this all together we see that
linearization yields the following
Rb (g∗ + δg) = ui∗ + ∆i ui = ui∗ + byi δui (179)
Assuming that y i is real12 , we have the following cases,
1. yi > 0: In this case the corresponding ui corresponds to a relevant deformation away from
the fixed point. Namely, a repeated application of the transformations moves its value
away from the critical point.
2. yi < 0: In this case the corresponding ui corresponds to a irrelevant deformation from the
fixed point. Starting sufficiently close to the fixed point, the iteration of the transformation
brings g back to the fixed point.
3. yi = 0: In this case ui corresponds to a marginal deformation from the fixed point. Iterating
the transformation, g remains where it is.
It should be noted that all of the above analysis holds only to the linearized level. Marginal
operators are most of the time either marginally relevant or marginally irrelevant, and rarely
11
We assume that there is no Jordan blocks.
12
When y i is a complex number, the RG trajectory spirally converges to g∗ when Re(yi ) < 0 and spirally
diverges from it when Re(yi ) > 0)
52
exactly marginal. Said differently, the higher order terms, corresponding to O((δg)2 ) in 174,
give yi one sign or the other. They can play a very important role in, for instance, fine tuning.
For this reason and to simplify the arguments, we will from now on assume that all directions
are either relevant or irrelevant. Going back to the figure 15, the attractive point B is one where
all yi < 0, the replusive point A is one where all yi > 0, and the mixed point C is one where
some yi > 0 and some yi < 0.
3.4 Critical surface
Figure 16: Starting with anywhere on the critical surface, the couplings flow to the fixed point,
and starting from anywhere near the critical surface, the path converges to that of the red line
(emanating from the critical point).
Classically, we can include operators with arbitrarily high mass dimension as in the action by
including more and more fields and derivatives. However, these high dimension operators do not
affect the low-energy physics. In terms of RG transformation, this means that all these theories
flow to the same IR fixed point. This is the key to the idea of universality.
Concretely, consider a fixed point g∗ around which there are n relevant directions and (N − n)
irrelevant directions in the N -dimensional space of coupling constants. We will say that the
(N − n)-dimensional subspace around g∗ spanned by the (N − n) irrelevant directions is the
critical surface of g∗ , and it is by definition the attractive basin for the fixed point g∗ .
The other n couplings, on the other hand, need to be fine-tuned in order for the phase transition
to happen. For instance, in our familiar examples of phase transitions, these directions can
correspond to temperature, pressure etc.
Moreover, the correlation length is infinite not just at the fixed point, but on the whole critical
surface. To see this, recall that starting with a point {g}, after n iterations, the correlation
length transforms as
ξ(g) = b ξ({g (1) }) = b2 ξ({g (2) }) = . . . = bn ξ({g (n) }). (180)
53
If the initial point {g} was on the critical surface , then in the limit of n → ∞ the sequence
of {g (n) } converges to {g∗ }, i.e. lim {g (n) } = {g∗ }. Since ξ({g∗ }) → ∞ and b > 1, we have
n→∞
ξ({g}) → ∞ for all points on the critical surface.
3.5 Further Reading
References
[1] Mussardo, G. (2010). Statistical field theory: an introduction to exactly solved models in
statistical physics. Oxford University Press.
54
Topic-IV: Wilsonian Renormalization Group
(in Momentum Space)

55
4 Topic-IV: Wilsonian Renormalization Group (in Momentum
Space)
In this topic we apply the idea of Wilsonian renormalization, akin to what we have seen in
the previous topic, to quantum field theories. In particular, the basic idea of renormalization,
namely to change the coupling constants as well as the UV cut-off the theory, in such a way that
the low-energy physics remains unchanged, is common in real space RG and in QFT Wilsonian
RG. The way to achieve this in QFT might be described as “the trick of doing the path integral
a little at a time”.
4.1 Integrating out degrees of freedom

Let us consider a QFT that is governed by the following action
ˆ !
1 X
SΛ0 [Φ] = dd x ∂µ Φ∂ µ Φ + Oi (x)Λd−d
0
i
gi0 , (181)
2
i
where Oi (x) are local operators of (classical) dimension di > 0, for instance a Lorentz-invariant
monomials comprising of some ni powers of the fields Φ and their derivatives, for example
ni (d−2)
Oi ∼ (∂Φ)ì Φni −ì (so the mass dimension is di = 2 + ì in this case), and Λ0 is some
energy scale parametrizing the couplings, chosen so as to ensure that the coupling constants gi0
themselves are dimensionless. Throughout this chapter, we shall work in Euclidean signature.
Now we will also “regularize” the path integral by introducing a cut-off Λ1 (which can be does
not need to be chosen to coincide with Λ0 ). Recall that a field configuration in path integral
is a map from the space-time M to the space in which the field takes value. For instance, a
configuration of a scalar field is a specif map Φ : M → R. If we require a field configuration to be
smooth (namely, infinitely differentiable), it amounts to Φ ∈ C ∞ (M) mathematically speaking,
where the superscript indicates the condition that all derivatives of Φ must exist.
Now, let us define the regularized partition function for some high-energy scale Λ1 (> Λ0 ) as
follows
ˆ
1
ZΛ1 (gi0 ) = DΦ e− ~ SΛ0 [Φ] , (182)
C ∞ (M)≤Λ1
where the integral is now defined over the space C ∞ (M)≤Λ1 of smooth functions on M whose
energy is at most Λ1 . We will shortly see what this precisely means in terms of the Fourier space
expressions. The energy scale Λ1 here can sometimes be (but not necessarily has to be) thought
of as some natural microscopic length scale, for example the lattice spacing a, Planck scale `p ,
string length scale `s , etc. Thanks to the UV cutoff, the partition function 182 is well-defined
and possesses no UV divergences, by construction.
With this cut-off theory as the starting point, we are interested in obtaining an effective theory
at a scale Λ(< Λ1 ). We do so by performing the path integral over the modes with energy
56
between Λ1 and Λ. To make this precise, we turn to the momentum space. In terms of its
Fourier modes Φ̃(p), the field Φ(x) can be expressed and split as follows
´ dd p ip·x
Φ(x) = (2π)d
e Φ̃(p)
|p|≤Λ1
´ ´ !
dd p ip·x (183)
= + (2π)d
e Φ̃(p)
|p|≤Λ Λ<|p|≤Λ1
= φ(x) + χ(x),
where φ ∈ C ∞ (M)≤Λ is the low-energy mode while χ ∈ C ∞ (M)(λ,Λ1 ] is the high-energy mode,
corresponding to momenta in the shell that sits between the d-dimensional spheres of radius Λ
and Λ1 . The path integral measure then factorizes into the measure of low- and the measure of
high-energy modes,
DΦ = DφDχ. (184)
Thus, we can write 182 as

ˆ ˆ
− ~1 SΛ0 [φ+χ] 1 eff [φ]
ZΛ1 (g0i ) = DφDχ e = Dφ e− ~ SΛ , (185)
C ∞ (M) ≤Λ1 C ∞ (M)≤Λ
which leads to the definition of the effective action as

 
ˆ
1
SΛeff [φ] = −~ ln  Dχ e− ~ SΛ0 [φ+χ]  . (186)
 
C ∞ (M)(Λ,Λ1 ]
This effective action defined with a cutoff scale Λ involves only the low-energy modes, namely
the modes with energy not more than Λ. This process of integrating out the modes is what
we mean by the Wilsonian RG. This process can be repeated to obtain an effective field theory
with a even lower cutoff Λ0 < Λ:
 
ˆ
1 eff [φ+χ]
SΛeff0 [φ] = −~ ln  Dχ e− ~ SΛ . (187)
 
C ∞ (M)(Λ0 ,Λ]
Note that equations 186 and 187 can be seen as counterparts of 164 we encountered in real space
Wilsonian RG.
57
4.2 Running couplings
Via 186, we see that RG induces a flow of the coupling constants : gi = gi (Λ). The partition
function corresponding to this coupling at energy scale Λ,
ˆ
1 eff [φ]
ZΛ (gi (Λ)) = Dφ e− ~ SΛ , (188)
C ∞ (M)≤Λ
is by construction identical to the partition function over the energy scale Λ0 , i.e.
ZΛ (gi (Λ)) = ZΛ0 (gi (Λ0 )). (189)
When considering an infinitesimal change of the cutoff, 189 leads to the differential equation
!
d ∂ ∂g(Λ) ∂
Λ ZΛ (gi (Λ)) = Λ +Λ ZΛ (gi (Λ)) = 0. (190)
dΛ ∂Λ g ∂Λ ∂g Λ
This RG equation of the partition function is called the Callan-Symanzik equation for the par-
tition function. This equation states that the coupling constants in the effective action SΛeff [φ]
must change in such a way that they balance out precisely the change in the cutoff scale to
render a constant partition function.
With a generic initial action, the effective action can be parametrized in terms of the general
form
ˆ "
ni
#
zΛ µ X
SΛeff [φ] = d
d x ∂ φ∂µ φ + Λ d−di 2
zΛ gi (Λ)Oi (x) , (191)
2
i
where zΛ is sometimes called the wavefunction renormnalization and accounts for the quantum
corrections to the kinetic term, di is the classical dimension of Oi , gi (Λ) is the dimensionless
coupling constant, and Oi (x) ∼ (∂φ)ì φni −ì . We define the beta-function βi to be the the
derivative of the running coupling with respect to the logarithm of the scale, i.e.
∂gi (Λ) ∂
βi (gj (Λ)) := Λ = gi (Λ) = (di − d)gi (Λ) + βiquant (gj ), (192)
∂Λ ∂ lnΛ
where βiquant represents the quantum effects of integrating out the high-energy modes. In par-
ticular, βiquant = 0 if and only of there is no effect of integrating out and Λd−di gi (Λ) is simply
independent of Λ.
4.3 Anomalous dimensions

The anomalous dimension of the field φ is defined as
1 ∂zΛ
γφ := − Λ . (193)
2 ∂Λ
58
In other words, The wavefunction renormalization scales as zΛ ∼ Λ−2γφ . This means that the
canonical field ϕ, defined to have as kinetic term 21 (∂ϕ)2 at cutoff scale Λ scales with Λ as
√
ϕ := zΛ φ ∼ Λ−γφ φ.
The effective action 191 can now be expressed in terms of the renormalized field as follows
ˆ " #
1 X
SΛeff [φ] = d µ
d x ∂µ ϕ∂ ϕ + Λ d−di
gi (Λ)Oi (x) , (194)
2
i
where Oi (x) takes the form of the monomial (∂ϕ)ì ϕni −ì . In the case of a more general theory
with more than one field, each of them will have its own anomalous dimension. Wavefunction
renormalization will play an important role in our subsequent study of the correlation functions.
Consider the n-point corelator,
´ eff [φ]
C ∞ (M)≤Λ Dφ e−SΛ φ(x1 ) . . . φ(xn ) −n
hφ(x1 ) . . . φ(xn )i = ´ eff = zΛ 2 hϕ(x1 ) . . . ϕ(xn )ig(Λ);Λ , (195)
C ∞ (M)≤Λ Dφ e−SΛ [φ]
where
´ eff [ϕ;g
C ∞ (M)≤Λ Dϕ e−SΛ i (Λ)] ϕ(xi ) . . . ϕ(xn )
hϕ(x1 ) . . . ϕ(xn )ig(Λ);Λ := ´ eff [ϕ;g (Λ)]
−SΛ
. (196)
C ∞ (M)≤Λ Dϕ e i
Suppose the field insertions involve only modes with energies Λ (so |xi − xj | 1/Λ), then
we should be able to compute the same correlator using an effective theory with a lower cutoff.
Consider therefore a theory with cutoff sΛ, where s < 1 plays the role similar to 1/b in real
space RG. We now have
−n
hφ(x1 ) . . . φ(xn )i = zΛ 2 hϕ(x1 ) . . . ϕ(xn )ig(Λ);Λ
(197)
−n
= zsΛ2 hϕ(x1 ) . . . ϕ(xn )ig(sΛ);sΛ ,
or equivalently,

d − n2 −n ∂ ∂
0=Λ zΛ hϕ(x1 ) . . . ϕ(xn )ig(Λ);Λ = zΛ 2 nγφ + Λ + βi hϕ(x1 ) . . . ϕ(xn )ig(Λ);Λ .
dΛ ∂Λ ∂gi
(198)
This is called the Callan-Symanzik equation for correlation functions, generalizing 190 which
holds for the partition function.
59
Independent of renormalizations, let us recall the following scaling transformation
x 7→ x0 = sx. (199)
Note that we have a symmetry

ˆ ˆ
d
d x (∂φ(x)) =2
dd x0 (∂φ0 (x0 ))2
if the field is transformed as
2−d
φ(x) 7→ φ0 (sx) = s 2 φ(x). (200)
After veryfing the invariance of the action, note that the cutoff scale needs to be rescaled as
Λ → Λ/s simultaneously to ensure the invariance of the physical quantities computed, as is clear
from 183. As a result, we have
−n (1) − n
zΛ 2 hϕ(x1 ) . . . ϕ(xn )ig(Λ);Λ = zsΛ2 hϕ(x1 ) . . . ϕ(xn )ig(sΛ);sΛ (201)
(2) − n
d−2 n
= zsΛ2 s 2 hϕ(sx1 ) . . . ϕ(sxn )ig(sΛ);Λ . (202)
Here, (1) is obtained by invariance under integrating out degrees of freedom, namely an RG step,
and (2) follows from invariance under trivial rescaling 199 and 200. Putting things together, we
obtain
hϕ(x1 ) . . . ϕ(xn )ig(Λ);Λ = sn∆φ hϕ(sx1 ) . . . ϕ(sxn )ig(sΛ);Λ (203)
where
d−2
∆φ = + γφ
2
d−2
is called the scaling dimension of the field φ, which is the sum of the classical dimension 2
and the anomalous dimension γφ .
The equality 203 offers a different interpretation of the flow of the coupling constants incurred
by RG. Note that both sides are evaluated in a path integral with the same cutoff, but with
a different action with a different set of coupling constants. At the same time, the distances
between the insertions have been rescaled when going from the LHS to the RHS. As s < 1, the
LHS represents the IR (infrared, referring to long-distance property) while the RHS represents
the UV degrees of freedom as the insertions are closer to each other. Intuitively, this makes
sense since long-distance physics (LHS) is governed by IR degrees of freedom (g = g(sΛ)) that
survive when we integrate out the high-energy modes.
Points of special interests in the space of coupling constants are the fixed point of the RG flow,
satisfying g∗ (Λ) = g∗ (sΛ) = g∗ . At such a fixed point, the above equality 203 implies the
60
following: At critical point, the correlation length diverges and one gets power law decay.
To prove this, note that at the critical point, for n = 2 we have
hϕ(x)ϕ(y)ig∗ ;Λ = s2∆φ hϕ(sx)ϕ(sy)ig∗ ;Λ (204)
which immediately suggests that
1
hϕ(x) . . . ϕ(y)ig∗ ;Λ ∼ . (205)
|x − y|2∆φ
d−2
In more details, we know on the one hand that the mass dimension of the field ϕ is 2 , and
on the other hand, the only scales we have in the quantity hϕ(x) . . . ϕ(y)ig∗ ;Λ is |x − y| and Λ,
whose product is dimensionless. As a result, we have
hϕ(x) . . . ϕ(y)ig∗ ;Λ = Λd−2 f (Λ|x − y|), (206)
c(g∗ )
and from 204 we know that f (x) = 2∆ , which leads to
(Λ|x−y|) φ
1 c(g∗ )
hϕ(x) . . . ϕ(y)ig∗ ;Λ = , (207)
Λ2γφ |x − y|2∆φ
where c(g∗ ) is a dimensionless constant.

Infinitesimally, the above can also be shown using the Callan-Symanzik equation 198 at gi = gi∗
(where βi = 0) for the two-point correlators.
4.4 Renormalization group trajectories

A main reason to study RG is to study its fixed points. In particular, we would like to know
how points near the fixed points evolve under RG flow. In the current setup, we would like to
study the QFT counterpart of 174:
∂
Λ gi = βi (g∗ + δg) = Bi j δgj + O(δg 2 ), (208)
∂Λ g∗ +δg
where
∂βi
Bi j := (g∗ )
∂gj
Let σI and (∆I − d) be the eigenvector and eigenvalue of Bij . We now have
∂
σIi gi = σIi Bi j δgj = (∆I − d)(σIi δgi ).

Λ (209)
∂Λ g∗ +δg
Classically, we expected a dimensionless coupling to scale with a power of Λ determined by the

explicit powers of Λ included in the action in 181, so classically we expect σIi 6= 0 for a group
of operators Oi with the same mass dimension di , and ∆I = di . The difference between ∆I
and di is then the analogue of the anomalous scaling dimension γφ of the field φ, which arises
through the quantum effects that renormalizes the kinetic term of φ. We will therefore call it
61
the anomalous dimension of the operator. Parametrizing the region in the coupling space near
the fixed point in terms of the eigenvector σI , we have
σI := σIi (gi + δgi ) = σ∗,I + δσI . (210)
Plugging this back into 209, we get
∂ ∂
Λ σI = Λ δσI = (∆I − d)∆σI , (211)
∂Λ ∂Λ
or equivalently
∆I −d
Λ
δσI (Λ) = δσI (Λ0 ). (212)
Λ0
We now have the following three distinct cases:
Figure 17: Theories living on the critical surface flow (dashed lines) to a critical point in the
IR. Turning on relevant operators drives the theory away from the critical surface (solid lines),
with flow lines focusing on the (red) renormalized trajectory emanating from the critical point.
1. ∆I < d: Relevant.
δσI grows in the IR, i.e. the RG flow will drive us away from the critical surface C as we
head into the IR.
2. ∆I > d: Irrelevant.
δσI becomes smaller as the scale Λ is lowered, or as we probe the theory in the IR.
3. ∆I = d: Marginal.
62
δσI has vanishing eigenvalues and hence, it neither increases nor decreases under RG flow.
Since every field or derivative adds to the dimension of the operator, in a fixed spacetime
dimension d there will only be a finite number of relevant operators. Thus, we can conclude
that the critical surface has finite co-dimension which is given by the number of independent
relevant operators with ∆i < d.
For instance, let us suppose that there is a non-trivial fixed point g ∗ around which there is just
one negative eigenvalue (i.e. ∆i − d < 0) that corresponds to a relevant operator. The set of
trajectories that are attracted into this fixed point for Λ → 0, therefore, forms a critical surface
of codimension-one; i.e. a surface defined by a single condition on the coupling (for instance
T = Tc ). This is shown in figure 17.
A generic QFT will have an action that involves all types of operators and so lies somewhere off
Dimension Relevant operators Marginal operators
d=2 φ2k , ∀k ≥ 0 (∂φ)2 , φ2k (∂φ)2 , ∀k ≥ 0

d=3 φ2k , k = 1, 2 (∂φ)2 , φ6
d=4 φ2 (∂φ)2 , φ4
d>4 φ2 (∂φ)2
Table 3: Relevant and marginal operators in a Lorentz invariant theory of a single scalar field in
various dimensions, near the Gaussian critical point where the classical dimensions of operators
are a good guide. Only reflection symmetric (φ → −φ = φ) operators are shown.
the critical surface C. Under the RG evolution, all the irrelevant operators are quickly suppressed,
while the relevant ones grow. The flow line of a generic theory thus strongly focusses onto the
renormalized trajectory, and so in the IR a generic QFT will closely resemble a theory emanating
from the critical point, where only relevant operators have been turned on. The fact that many
different high-energy theories will flow to look the same in the IR is known as universality.
It guarantees that the properties of the theory in the IR are insensitive to the infinite set of
irrelevant couplings {gi }. In other words, the theory has “forgotten where it came from”. We
say that theories whose RG flows focus onto the same trajectory emanating from a given critical
point are in the same universality class. Theories in a given universality class could look very
different microscopically, but will all end up looking the same at large distances.
4.5 Counterterms and continuous limits

In a finite theory with a UV cutoff Λ, the physics at energies E Λ is independent of the precise
value of Λ. Changing Λ changes the couplings in the theory so that observables remain the same.
This is the philosophy of Wilsonian renormalization. The continuum renormalization group, on
the other hand, tells us that observables are independent of the renormalization conditions, in
particular, of the scales {p0 } at which we choose to define our renormalized quantities. This
invariance holds after the theory is renormalized and the cutoff is removed (Λ = ∞, d = 4).
In dimensional regularization with the MS scheme, the scales {p0 } are replaced by µ, and the
continuum renormalization group comes from µ independence.
63
In Wilsonian RG, we started with a theory with a cutoff Λ0 described by an UV action S0 and
partition function ˆ
1
ZΛ0 (gi0 ) = DΦ e− ~ S0 [Φ] , (213)
C ∞ (M)≤Λ0
from which we obtain an effective theory described by the action 186. We can ask the following
question: Given a low-energy theory, one that is dictated by experiments for example, can we
remove the cutoff (sending Λ0 → ∞)? Is there a theory S0 satisfying
 
ˆ
1 1 eff[φ]
lim  Dχ e− ~ S0 [φ+χ]  = e− ~ SΛ (214)
 
Λ0 →∞
C ∞ (M)(Λ,Λ0 ]
eff[φ]
for the given IR theory with action SΛ and cutoff Λ? The limit of taking the cutoff to infinity
is called the continuum limit, since it corresponds to taking the lattice spacing to zero. The
answer to this question is that it depends. Let us analyze this case-by-case.
1. Case I: When {g(Λ)} ∈ C

When we have a theory described by S0 on the critical surface, we have some r number of
parameters that are tuned. In such a case, the continuum limit always exists, i.e.
 
ˆ
1 1 eff[φ]
lim  Dχ e− ~ SΛ0 [φ+χ]  ≡ lim e− ~ SΛ , (215)
 
Λ0 →∞ 0→∞
C ∞ (M)(Λ,Λ0 ]
as long as we take the limit after computing the path integral and is described by the
critical theory. The resulting scale-Λ effective theory will be a CFT that is scale (Λ) inde-
pendent.
eff[φ]
Conversely, if SΛ corresponds to a theory on the critical surface that is not at the critical
point, than it cannot be obtained via 215 from any UV-complete (namely Λ0 → ∞) theory.
This does not mean, of course, that there can be no theory describing the UV physics of
the system, but rather implies that new degrees of freedom (such as a new particle) must
exists at some energy scale that is higher than Λ.
An interesting case is given by the marginally irrelevant operators. Strictly speaking, just
like irrelevant operators, marginally irrelevant operators are arbitrarily suppressed as the
cut–off is removed. However, they typically decay only logarithmically as Λ0 is raised,
rather than as a power law. Such operators thus afford us a tiny glimpse of new physics
at exponentially high-energy scales, far beyond the range of current accelerators.
2. Case II: When {g} ∈

/C
The low-energy theory at Λ not being on C implies that the theory at the cutoff scale Λ0
won’t be either. Let µ be the energy scale at which the flow passes the point closest to
g∗ on the particular RG trajectory. In this case, to take the continuum limit Λ0 → ∞ we
64
must tune the initial couplings such that µ = Λ0 f (gi0 ) remains finite under Λ0 → ∞. This
tuning is achieved by the addition of counterterms, such that the following limit exists
 
ˆ
1 eff 1
e− ~ SΛ = lim  Dχ exp − S0 [φ + χ] − SCT [φ + χ, Λ0 ]  . (216)
 
Λ0 →∞ ~
C ∞ (M)(Λ,Λ0 ]
Notice again that the limit is taken after performing the path integral. Sending Λ0 → ∞
defines a continuum QFT with finite (or renormalized) relevant couplings at scale Λ. To
evaluate the path integral, we first compute quantum corrections to 1-loop order using
the original action S. Unfortunately, these loop corrections depend on the cutoff Λ0 and
typically diverge in the continuum limit. This reflects the fact that we lose control of the
original theory if the cutoff is removed naively.
4.6 Polchinski’s equation

Polchinski’s equation (also called Polchinski’s flow equation) describes a perturbative approach
to Wilsonian RG. We will soon see that this RG equation can be recast in the form of a heat
equation.
To derive it, first recall the calculation of the effective action from 182 to 187. Written together,
we have
ˆ ˆ ˆ ˆ
1 1 1 eff [φ]
ZΛ0 (gi0 ) = DΦ e− ~ S[Φ] = Dφ Dχ e− ~ S[φ+χ] ≡ Dφ e− ~ SΛ .
C ∞ (M)≤Λ0 C ∞ (M)≤Λ C ∞ (M)(Λ,Λ0 ] C ∞ (M)≤Λ
(217)
Expanding the action S[φ + χ], we can separate it into
S[φ + χ] = S 0 [φ] + S 0 [χ] + S int [φ, χ]. (218)
We can write the effective action as
SΛeff [φ] = S 0 [φ] + SΛint [φ] (219)
where we have defined

 
ˆ
1 0 [χ]−S int [φ,χ]
SΛint [φ] ≡ −~ ln  Dχ e− ~ (S )
. (220)

C ∞ (M)(Λ,Λ0 ]
The path integral over the action S 0 [χ] yields a φ-independent constant whose effect can be
ignored (as they will cancel out in any physical quantity). Now consider an infinitesimally thin
65
shell in the momentum space: Λ = Λ0 − δΛ. Then the position space propagator for the high-
energy field χ reads
ˆ ˆ
(χ) dd p eip·(x−y) 1 Λd−1 δΛ
D (x, y) = = dΩ eiΛp̂·(x−y) (221)
(2π)d p2 + m2 (2π)d Λ2 + m2 S d−1
Λ<|p|≤Λ0
up to terms of order O(δΛ2 ). Since the propagator D(χ) (x, y) comes with one power of δΛ, in
the leading order only diagrams with at most one χ propagator contribute. Since χ is summed
over, there are only two classes of diagrams: in which the χ propagator connects two separate
vertices in S int[φ,χ] , or in which it starts and ends a single vertex to itself. See Figure 18 for a
schematic representation: the beta function of the coupling gn corresponding to the operator
φn involves in the first category coupling constants gr+1 and gn−r+1 , and gn+2 in the second
category.
Figure 18: The dashed line is a χ-propagator that will be integrated out.
Polchinski wrote down an equation for the change in the effective action that captures the
information in the Feynman diagrams in figure 18. This was just the infinitesimal version of
Wilson’s RG equation 186 for the effective action. Define t ≡ ln Λ, then we have
ˆ
d 1 Λd
Ḋ(χ) (x, y) := D(χ) (x, y) = dΩ eiΛp̂·(x−y) , (222)
dt (2π)d Λ2 + m2 S d−1
in terms of which one obtains

ˆ " #
δ 2 −SΛint

d δSΛint (χ) δSΛint
−SΛint [φ] = d d (χ)

d xd y Ḋ (x, y) + Ḋ (x, y) (223)
dt δφ(x) δφ(y) δφ(x)δφ(y)
where the two summands in the integrand above correspond to the two types of diagrams in Fig
18. In particular, the second term is only non-vanishing when x = y since S int [φ] is local
It is illuminating to rewrite the Polchinski equation as follows

d −S int [φ] int Λ
e Λ = −∆ e−SΛ [φ] , t = ln (224)
dt Λ0
where we have defined the Laplacian-like operator

ˆ
δ2
∆≡ dd x dd y Ḋ(χ) (x, y) . (225)
δφ(x)δφ(y)
66
Notice that written in this way, the Polchinski equation in 224 has the form of a heat equation.In
terms of its eigenfunctions ∆uk ({g}) = λk uk ({g}), the solution to the Polchinski equation reads
int [φ]
X
e−SΛ = uk ({g})f˜k e−λk t . (226)
k
One can see that the summand on the RHS either grows or decays when Λ gets lower, depending
on the sign of the eigenvalue λk . This is a theme that we have seen from many viewpoints by
now: an operator either becomes more dominant, negligible, or stable in the IR.
4.7 Further Reading
References
[1] Lecture notes by Prof. David Skinner
http://www.damtp.cam.ac.uk/user/dbs26/AQFT/Wilsonchap.pdf
sity Press.
67
5 Topic-V: Conformal Field Theory
5.1 Why should we care about conformal field theories?
Figure 19: A topographical representation of where we are headed. A highly simplistic map is
particles7→fields conformal invariance
the following: QM −→ QFT −→ CFT. This picture can again be split into
a region of Lorentzian signature with spacetime Rd−1,1 and a region of Euclidean signature with
spacetime Rd , we shall reside in the latter region for the purposes of this chapter.
A conformal field theory (CFT) describes a quantum system that is invariant under angle-
preserving transformations of spacetime. One might be tempted to discard such systems as
physically irrelevant since they do not possess a length scale, but it turns out that conformal
invariance plays a crucial role in many areas of theoretical physics, all the way from statistical
mechanics to quantum gravity. As a soft introduction, let us consider a statistical system with
N number of particles, where N is a very large number. Under the thermodynamics limit, i.e.
N → ∞, the system becomes equivalent to a Euclidean QFT and under this continuum limit, we
can approximate states of the system with fields. To further the analogy, we note that the prob-
ability distribution over the states becomes a distribution over the space of field configurations,
also known as the path integral measure. The effective field-theoretic description with such a
measure is massive but at the phase transition, as the correlation length of the statistical system
becomes infinite, the resulting field theory enjoys scale invariance and with (mild) assumptions
also enjoys conformal invariance. Hence, we see that CFTs model continuum limits of critical
statistical systems. What is really amazing is that our Universe exhibits scale invariance also
on the largest scales. Zooming out beyond galaxies, clusters and further, the structure keeps
repeating itself. Hence, all structures we observe might only be present in a small range of scales
embedded in an ever self-repeating Universe!
A classic example of such a system is the Ising model which is described by the following
Hamiltonian
X X
H({sk }) = −J si sj − h si , (227)
hi,ji i
where spins sk ∈ Z2 ≡ {±1} live on a lattice Λ with each spin interacting with its nearest
neighbors while they all lie in an external magnetic field. The two parameters that appear
in the Hamiltonian are the interaction strength J and the magnetic field h. Tuning these
68
parameters result in different descriptions, for example, suppose J > 0, then the Hamiltonian
describes a ferromagnet. Consider now perhaps the simplest case of the Ising model, when we
have a vanishing magnetic field and a hypercubic lattice Λ = Rd . In such a case, the calculation
of the partition function
X
Z= e−βH({sk }) , (228)
{sk }
is incredibly difficult for d ≥ 3. It is not very useful to look at the one dimension case since
the model does not exhibit a phase transition. In two dimensions, however, the model had been
analytically solved by Onsager in the 40s. A highly nontrivial fact from statistical mechanics is
that phase transitions are universal. For example, as a liquid transitions into a gas, the heat
capacity diverges as a power law of temperature, c ∼ |T − Tc |−γ , where Tc is the critical temper-
ature and the universality aspect being encoded in the exponent γ since it turns out to be the
same for all gases. It turns out that universality helps us to make a stronger analogy between
this liquid-gas transition and certain critical systems that describe magnets that are similar to
the Ising model.
For the same reasons as in statistical mechanics, CFTs occupy a significant place in the space of
all QFTs, viewed in the Wilsonian paradigm. In this heuristic picture, there is a (RG) flow on
the space of all theories, provided by the integration of field modes. Fixed points of the flow are,
by definition, scale-invariant theories. When combined with other properties of a QFT, scale
invariance, in most cases, enhances to the full conformal invariance. Thus, conformal theories
may be used to organize the universality classes of QFTs. Quantum theories of a class are all
the ones that flow into a given fixed point. The mere ability to solve a certain CFT helps us
access the neighboring QFTs through perturbation theory! Also, CFT descriptions of systems
that do not exhibit conformal symmetry exist with a prime example being the AdS/CFT cor-
respondence, or the gauge-gravity duality. According to this correspondence, quantum theories
of gravity with asymptotically AdSd+1 metrics are equivalent to d-dimensional CFTs. Also, the
symmetries on the gauge side match that of the gravity side, i.e. the conformal group SO(d, 2)
is interpreted as the isometry group of AdSd+1 in the original gravitational theory.
A particular reason why CFTs is interesting is the fact that the expansion of operator products
of an algebra a of local operators is convergent rather than asymptotic as is in a QFT. We can
define a certain algebra of local operators with all the properties of the theory encoded in them
and we notice that the construction of a CFT reduces to that of the associated algebra. In this
algebraic language, the most constraining consistency condition that is to be satisfied turns out
to be associativity, i.e. (f1 f2 )f3 = f1 (f2 f3 ) with fi ∈ a. These can be formulated as the set
of crossing symmetry equations satisfied by four-point correlators. Note that the problem, now
recast in a new language, is still a very difficult one. Local operators are in on-to-one corre-
spondence with the irreducible representations of the conformal group that make up the Hilbert
space. It is possible to show that there are necessarily infinitely many irreducible components.
Therefore, there are infinitely many crossing equations. Matters become more interesting in two
69
dimensions since it is sometimes possible to organise the irreducible representations into a finite
number of modules of (two copies of) a certain algebra called the Virasoro algebra we will en-
counter later. Theories with this property are known as minimal models. Their construction and
subsequent identification with two-dimensional statistical systems was the major achievement
of the field in the 80s. CFTs also lie at the intersection of several non-perturbative approaches
to QFTs. Many of these apply to theories with supersymmetry and include localization, super-
conformal index calculations, and techniques coming from chiral algebras.
Besides the analysis of concrete models, the approach of constraining general CFTs using cross-
ing symmetry, known as the conformal bootstrap program, has seen a revival since its discovery
in 2008 that small subsets of the infinite system of crossing equations can be studied numerically
in their own right to produce interesting conclusions. The crowning achievement of the boot-
strap to date is a precise determination of critical exponents in the 3D Ising model. The study
of two-dimensional CFT, originating simultaneously in string theory and the theory of critical
phenomena, has been a very active field. Its applications have found their way to numerous
fields- ranging from spin chains to number theory. This chapter is dedicated to the study of the
basics of CFT in two dimensions.
5.2 Conformal transformations

As noted previously, at the critical point, the correlation length diverges and there is a scaling
symmetry, i.e.
x 7→ λx, Φi 7→ λ∆i Φi , (229)
where the exponent ∆i is not easy to determine in general, as we had previously seen in the
context of RG. The above global scaling transformation is a special case of a conformal trans-
formation, x 7→ λ(x)x. This motivates the study of CFTs.
Let us consider a two-dimensional Euclidean QFT that is described by the following PI

ˆ
Z= Dφ e−S[φ] , (230)
Assuming spacetime to be flat, i.e gµν = δµν , consider now the infinitesimal coordinate transfor-
mations
xµ 7→ xµ + µ (x), (231)
under which the action changes by the following
70
ˆ
d2 x
δS = − Tµν ∂ µ ν , (232)
2π
where Tµν is the symmetric stress-energy tensor. Translational invariance implies that the con-
servation law ∂ µ Tµν = 0 holds, and if we assume scale invariance, i.e. invariance under rescaling
xµ 7→ λxµ , then we can conclude that the stress-energy tensor is also traceless, T µµ = 0. This
property automatically implies invariance under the extended class of coordinate transforma-
tions that satisfy
∂ µ ν + ∂ ν µ = δ µν ∂α α , (233)
the so-called conformal transformations. These are exactly the transformations that transform
the metric with a local scale factor gµν 7→ Ω(x)gµν . This means that the physics of the theory
under consideration looks the same at all length scales. In other words, CFTs preserve angles
but not necessarily lengths. Let us take a closer look at these transformations and motivate the
above equation. Consider a transformation of the following form
x 7→ x0 (x) (234)
such that the metric transforms as follows
0 ∂xµ ∂xν
gµν (x) 7→ gµν (x0 ) = gµν (x) = Ω(x)gµν (x), (235)
∂x0r ∂x0ν
where the scale factor Ω(x) is called the conformal factor. Infinitesimally, this transformation
of the coordinate can be expressed as follows
x 7→ x + (x), (236)
and the corresponding transformation of the metric reads
gµν (x) 7→ gµν (x) + ∇µ ν + ∇ν µ . (237)
We make the following identification
∇µ ν + ∇ν µ = f (x)gµν , (238)
where f (x) is a proportionality factor that can be a function of spacetime but more importantly,
it is a scalar transformation. Notice that when f (x) = 0, the resulting equation is nothing but
the Killing equation. Taking the trace fixes
71
2 α
f (x) = ∇ α . (239)
d
Said differently, we can reformulate the covariant derivatives as follows
2
∇µ ν + ∇ν µ = (∇α α ) gµν (x). (240)
d
Notice that this is nothing but a conformal Killing equation. In flat Euclidean space, the
covariant derivative becomes the partial derivative and gµν = δµν . This yields
2
∂µ ν + ∂ν µ = (∂ · ) δµν , (241)
d
which is a generalization of 233 to d-dimensions. The claim now is that the divergence ∂ ·
is at most linear in x for d 6= 2. To prove this we first start with 241 and make the following
manipulations
2
∂ρ ∂ν (∂µ ν + ∂ν µ ) = d [(∂ · ) ∂ν ∂ρ δµν + δµν ∂ν ∂ρ (∂ · )]
⇒ ∂ρ ∂µ (∂ · ) + 2∂ρ µ = d2 ∂ρ ∂µ (∂ · ) (242)
2
∂ρ ∂µ (∂ · ) + 2∂ρ µ = 0.

⇒ 1− d
In the last line, we notice that since the first piece, ∂ρ ∂µ (∂ · ) is manifestly symmetric in ρ and
µ, the second piece also has to be. Thu, rewriting it as ∂(ρ µ) /2, we obtain

2 1
1− ∂ρ ∂µ (∂ · ) + δµρ 2 (∂ · ) = 0. (243)
d d
Contracting the indices yields

1
2 1− 2∂ · = 0 ⇐⇒ 2 (d − 1) ∂ 2 f (x) = 0. (244)
d
From this we immediately see that for the case of d > 1, we have 2∂ · = 0. Plugging this back
into 242 we get

2
1− ∂ρ ∂µ (∂ · ) = 0. (245)
d
Notice now that the case of d 6= 2, ∂ρ ∂µ (∂ · ) = 0 implies
∂ · = P1 (x), (246)
72
where P1 (x) ∼ x1 , i.e. is a degree 1 polynomial. In other words, we have ∂µ ∂ρ f (x) = 0 provided
that d > 2 which tells us that f (x) ∼ ∂ · can at most be linear in coordinates,
f (x) = A + B α xα , (247)
and µ at most quadratic,
µ = aµ + bµν xν + cµνα xν xα , (248)
with the restriction cµνα = cµαν . Plugging 248 into 238 for the case of gµν = δµν yields the
following conditions on aµ , bµν , and cµνα that read
1. aµ is unconstrained and is the generator of infinitesimal translations.
2. bµν = λδµν + ωµν with ωµν = −ωνµ is the generator of the Lorentz group and the trace
(the first piece) is an infinitesimal scale transformation.
3. cµνα = ηµα bν + ηµν bα − ηνα bµ with a constant vector bµ . These generate the so called
special conformal transformations (SCTs) and act on coordinates as follows
xµ 7→ x0µ = xµ + 2(x · b)xµ − bµ x2 . (249)
Locally, a conformal transformation acts as a combination of a translation, a rotation, and a

dilatation. So, all angles will be preserved. The finite version of these transformations are the
following
1. ∼ x0 :
In this case, we have a shift symmetry or translation, xµ 7→ xµ + aµ , where aµ is an
infinitesimal parameter.
Figure 20: We can view the complex plane as a projection from the Riemann sphere that lives a
dimension above. Translation on the 2D plane can be understood as translations of the sphere
as shown.
73
2. ∼ x1 :
(a) Global dilatation: µ = λxµ .

The metric transforms up to a scale factor, i.e. gµν 7→ Ω(x)gµν = λ2 gµν where λ is a
constant function.
Figure 21: Dialatations or simply dialations are zooming in and out in the complex plane which
can be understood to be projections from the Riemann sphere that traverses the z-axis.
(b) Rotation: µ = ω µν xν .
Here ω µν = −ω νµ is a constant tensor that does not depend on the spacetime and µ
satisfies the following equation
∂ν µ + ∂µ ν = ωµν + ωνµ = 0. (250)
Figure 22: Rotations in the complex plane can be understood to be projections from a rotating
Riemann sphere about the z-axis.
3. ∼ x2 :
When is quadratic, the resulting transformation is called a special conformal transfor-
mation (SCT) which take the following form
µ = bµ x2 − 2xµ b · x, (251)
where bµ is a constant. Manifestly, SCTs are nothing but the combination inversion13
13 xµ
this is defined as the following map, I : xµ 7→ x2
, with I 2 the identity element.
74
followed by a translation followed by another inversion, i.e.
Figure 23: Inversions on the complex plane can be understood to be projections from a hori-
zontally rotating Riemann sphere.
I x0µ T x0µ I
x0µ 7→ 02
7 → 02
7 x0µ .
+ bµ → (252)
x x
This is visualized in figure 24 and figure 25 to understand the SCT from the perspective
of the moving Riemann sphere..
Thus, the finite version of the SCT reads
Figure 24: The point xµ in the upper half plane within the circle |z| = 1 is first inverted to
xµ /x2 and then translated by an amount −bµ to x0µ /x02 , and then finally inverted again to the
point x0µ within the circle. The inverse transformation, i.e from x0µ to xµ involves a shift by
+bµ and is described by 252.
x0µ xµ
→
7 + bµ . (253)
x02 x2
Solving for x we obtain the following
x0µ /x02 xµ /x2 + bµ xµ /x2 + bµ xµ + bµ x2

x0µ = = = = . (254)
x02 /x04 (xν /x2 + bν )2 1
x2
+ 2b·x
x2
+ b2 1 + 2(b · x) + b2 x2
The finite versions of the transformation mentioned above are listed in table 4 and a combination
75
of conformal transformations are shown in 25.
More abstractly, we can think of the infinitesimal transformations as being generated by linear
Transformations Action on coordinate
Translations (shift transformations) x0µ = xµ + aµ

Dilatation (scale transformation) x0µ = λxµ
Rotation (Lorentz transformation) x0µ = ω µν xν
xµ +bµ x2
Special conformal transformations x0µ = 1+2(b·x)+b 2 x2
Table 4: The finite versions of the generators of the conformal symmetry group
Figure 25: A combination of conformal transformations: inversion + translation + inversion

(SCT).
operators. To motivate this let us first consider a local QFT. We now know that every local
QFT has a conserved stress-energy tensor, i.e. ∂µ T µν = 0. the associated conserved charge is
´
nothing but the four-momentum P ν = Σ dxµ T µν , where Σ is a codimension-one surface. For a
spacelike Σ, this reduces to the more familiar form given in ??. One can now generate conserved
charges if ∂µ (ν T µν ) = 0 and this gives rise to the Killing equation
∂µ ν + ∂ν µ = 0. (255)
Vector fields = µ ∂µ satisfying the Killing equation generate relevant symmetries which are
Pµ (translations) and Mµν (rotations). Imposing the traceless condition T µµ = 0 yields the
conformal Killing equation 240 and this is responsible for two more extra Killing vector fields,
D (dilatation) and Kµ (SCTs). We can conclude that field theories with traceless Tµν possess
conformal symmetry! We shall soon see that a conserved traceless stress tensor and other local
operators O, that possibly form an uncountable set, are what define a local unitary CFT. The
four linear operators mentioned above read
Pµ = −i∂µ
Mµν = 2ix[µ ∂ν]
(256)
D = −ixµ ∂µ
Kµ = −i 2xµ xν ∂ν − x2 ∂µ ,

where the factors of i are chosen as to ensure that the generators are Hermitian. These linear
operators generate the conformal algebra, which is locally isomorphic to SO(m + 1, n + 1). This
76
can be easily seen by some basic counting. First, we set m + n = d and then we see that there
are m + n generators for Pµ , (m + n)(m + n − 1)/2 for Mµν , m + n for Kµ , and 1 for D. Hence,
in total, the conformal algebra has a total of (d + 1)(d + 2)/2 generators. Explicitly, for d 6= 2,
we have
#generators = # translations + #SO(D) rotation + #SCT + #Dilatation
d(d−1)
=d+ 2 +d+1 (257)
(d+1)(d+2)
= 2 .
The conformal algebra, in its full glory, is defined by the following set of commutation relations
[D, Pµ ] = −iPµ , [D, Kµ ] = −iKµ , [Kµ , Pν ] = 2i (ηµν D − Mµν ) , [Kα , Mµν ] = i (ηαµ Kν − ηαν Kµ ) ,
[Pα , Mµν ] = i (ηαµ Pν − ηαν Pµ ) , [Mµν , Mαβ ] = i (ηνα Mµβ + ηµβ Mνα − ηµα Mνβ − ηνβ Mµα ) .
(258)
This can be easily verified using the generators in 257. Now, making the following redefinitions
1 1
Jµν = Mµν , J−1,µ = (Pµ − Kµ ) , J−1,0 = D, J0,µ = (Pµ + Kµ ) , (259)
2 2
with Jab = −Jba and a, b ∈ {−1, 0, 1, . . . , d}, we see that the generators satisfy the SO(d + 1, 1)
commutation relations which read
[Jab , Jcd ] = i (ηad Jbc + ηbc Jad − ηac Jbd − ηbd Jac ) . (260)
This establishes the isomorphism between the conformal group in d-dimensions and SO(d + 1, 1)
in (d + 2) dimensions mentioned previously.
5.3 Algebra of local fields

Thus far we have defined a QFT in terms of a Lagrangian. The definition of a CFT on the other
hand usually takes an algebraic route. There is an a priori assumption that there is a basis of
local fields, where the “fields” need not necessarily be the fundamental fields we usually meant
in previous chapters. Subsequently, we define the CFT in terms of the algebra of local fields
which is a completely non-perturbative definition as opposed to the perturbative definition of
QFT. In either way, the prime goal of a QFT (including CFT) is to compute observables which
here take the form of correlation functions of local operators O(x),
ˆ
1 1
hO1 (x) . . . On (x)ihΩ|T (O1 (x1 ) . . . On (xn )) |Ωi ≡ DΦ e− ~ S[Φ] O1 (x1 ) . . . On (xn ). (261)
Z
77
As a first example, consider the scalar field theory where local fields include φ(x), : φ2 (x) :,
: φ3 (x) :, . . ., : φn (x) : with scaling dimensions ∆φ , 2∆φ , 3∆φ , . . ., n∆φ respectively where
∆φ = (d − 2)/2 with γφ = 0 following its definition in ??. Here : φ2 (x) : and higher powers
is the normal-ordered product of the power of the bosonic field φ(x) which by definition is the
subtraction of the divergent term coming from the propagator of the product of two operators,
i.e.
: φ(x)φ(y) : ≡ φ(x)φ(y) − hφ(x)φ(y)i. (262)
Placing the operators at the same position yields the desired result,
: φ2 (x) : = lim : φ(x)φ(y) : . (263)

x→y
All other normal-ordered products can be iteratively constructed starting from this relation.
the set of fields {: φn (x) :}, to which we have to add the identity operator 1 ≡ φ0 , can be
used to express any other regular density A(x) of the scalar field. This is achieved by the series
expansion of A(x)
∞
X
A(x) = an : φn (x) :, (264)
n=0
where an ∈ C. In general, given a basis of fields O1 (x), O2 (x), . . ., we consider general operators
to be a linear combination of the fields, i.e.
∞
X
A(x) = an On (x). (265)
n=0
It is important to emphasize that the validity of the operatorial identity 265, and identities of
similar flavour, is to be read from the vantage point of correlation functions, i.e. suppose we are
to find the correlator hA1 (x1 )A2 (x2 ) . . . Aj (xj )i, then we make us of the identity 265 to express
any of these, say A2 (x2 ), arriving in this way using the expression
∞
X
hA1 (x1 )A2 (x2 ) . . . Aj (xj ) = an hA1 (x1 )On (x2 ) . . . Aj (xj )i. (266)
n=0
Suppose the basis {Oi (x)} has a scaling dimension14 ∆i which is defined by the condition

x 7→ x0 = λx,
(267)
O(x) 7→ O0 (x0 ) = O (λx) = λ−∆i O (x).
i i
14
Note: ∆i = ∆classical
i + γi , where γi is the anomalous dimension of the operator defined in ?? which is in
general quite hard to compute
78
A (more) physical way to understand this operator is via RG. Firstly, we know that a CFT
is invariant under a RG step, x 7→ x0 = λx. We also know that in scale/conformal invariant
theories the Hamiltonian is left invariant, Ĥ 7→ Ĥ 0 = Ĥ. Under a scaling transformation, the
kinetic term of the Lagrangian transforms as follows
2 2
dd x0 ∂x0 φ0 (x0 ) 7→ dd x λd−2 ∂x φ0 (x0 ) , (268)
d−2
and for invariance, we require φ0 (x0 ) 7→ λ 2 φ(x) ⇒ ∆(x) = (d − 2)/2. Quantum corrections
can modify this invariance requirement and result in O 7→ λ−∆ O. This operator identity is
understood, as stated previously, in terms of correlators and hence, we write
hOi (λx)i = λ−∆i hOi (x)i. (269)
From this we see that hOi (x)i = 0 whenever ∆i 6= 0 since there is no distinguished scale in a
CFT. Following a similar argument, the propagator of operators Oi (x) takes on the following
form at the critical point
Aij
Dij (x − y) ≡ hOi (x)Oj (y)i = , (270)
|xi − xj |∆i +∆j
where Aij is a numerical constant. It is also easy to check that the following holds
hOi (λx)Oj (λy)i = λ−∆i −∆j hOi (x)Oj (y)i. (271)
Thanks to the work of Wilson and Zimmerman we know that in QFT, products of operators
can be approximated with an asymptotic expansion over operators. In a CFT, this expansion
becomes convergent and is known as the operator product expansion or just as the OPE. Recall
from quantum mechanics that we have a notion of multiplying representations of the rotation
group together and then expressing the same as a sum over the representations. The OPE is
very similar in spirit in the sense that we consider the product of two operators and express this
product as a sum over operators inserted at a single point. The goal is to express the product of
operators A(x)B(y) as a sum over local operators inserted at a single point. Consider the limit
lim A(x)B(y), (272)

x→y
under which we have the expansion
X
A(x)B(y) = Cn (x, y)On (x), (273)
n
where Cn (x, y) is a function of |x − y| and this is entirely fixed by symmetry. In particular, the
79
OPE of two operators Oi (x) and Oj (y) reads
X
k
Oi (x)Oj (y) = Cij (x − y)Ok (x). (274)
n
k (x, y) = C k (x − y) = C k (|x − y|) with O(λx) = λ−∆ O(x) gives us the scaling trans-
Now, Cij ij ij
formation of the OPE
X
k
Oi (λx)Oj (λy) = Cij (x − y) Ok (λx) . (275)
| {z } n
| {z }
∝λ−∆i −∆j ∝λ−∆k
k must possess dimensions of ∆ − ∆ − ∆ and since it is a function

From this we see that, Cij k i j
of |x − y|, the most general expression we can write is
k
ckij
Cij = , (276)
|x − y|∆i +∆j −∆k
where ckij is the structure constants of the operator algebra that are also called the OPE coeffi-
cients. Putting this all together, we get
X ckij
Oi (x)Oj (y) = Ok (x). (277)
n
|x − y|∆i +∆j −∆k
The set of OPE coefficients and scaling dimensions, called the CFT data, completely specifies
the CFT,
data = {∆i , ckij }. (278)
When we consider fermionic operators or other operators with non-zero spin, the CFT data
would also include the spin {ì }.
5.4 Crossing symmetry

Let us now consider the mixed four-point correlator hO1 (x1 )O2 (x2 )O3 (x3 )O4 (x4 )i. Clearly,
there are two unique ways in which we can decompose this using the OPE. The first OPE
decomposition, which we shall call the s-channel for reasons that will soon be clear, we have
X
k `
hO1 O2 O3 O4 i = C12 (x1 − x2 )Dk` (x2 − x4 )C34 (x3 − x4 ). (279)
k,`
Note that the contractions here refer to the OPEs of operators and not wick contractions.
Similarly, the other unique OPE decomposition, which we shall refer to as the t-channel reads
0 0
X
k `
hO1 O2 O3 O4 i = C13 (x1 − x3 )Dk0 `0 (x2 − x3 )C24 (x2 − x4 ). (280)
k0 ,`0
80
The OPE decomposition of the four-point function is not unique, i.e.
hO1 O2 O3 O4 i = hO1 O2 O3 O4 i. (281)
Figure 26: Crossing symmetry represented diagrammatically.
Diagrammatically, this can be expressed as shown in 26. Clearly, 279 resembles the s-channel
and 280 resembles the t-channel. It is important to emphasize the fact that we could have also
considered the OPE between operators 14 and 23, which would have given us the u-channel.
Doing so would not give us any new information or in other words, does not fix any new
constraints on the CFT data. Hence, s − t crossing symmetry is the same as that of t − u and
s − u. Simplifying 281 yields the bootstrap equation. In spirit with the meaning of bootstrap
which is to advance oneself or accomplish something as in the common saying, “to pull oneself
up one’s bootstrap”, the conformal bootstrap program helps us obtain results sans any external
data. In a nutshell, the philosophy of the bootstrap program is as follows:
1. Assume the existence of an OPE.
2. Assume additionally that the OPE contains only physical operators.
3. Impose crossing symmetry on the OPE.
4. Use tools at hand to constrain the CFT data.
The tools referred to here are conformal symmetry, causality, crossing symmetry, and unitarity.
The implementation of these tools can either be numerical, analytical, or even a mixture of the
two in which case we have the numerical and the analytical bootstrap programs. The prime
aim of the bootstrap approach is to obtain meaning bounds on the CFT data, i.e. the scaling
dimensions and OPE coefficients by reducing the problem to the classification of solutions to (a
family of) functional equations. These bounds help us to better understand the theory landscape
and locate precisely where physical theories reside as shown in figures 27 and 28.
Numerical conformal bootstrap has been used for CFTs in various dimensions with both bosonic
and fermionic operators. In analytic bootstrap, the focus thus far has been on bosonic correlators
that do not have access to fermionic sectors. The analytic study of fermionic sectors are of
81
Figure 27: Numerical bootstrap enables us to derive some strong constraints on the scalar
spectrum of the Z2 -symmetric CFT of the 3D Ising model. The axes represent the scaling
dimensions of the fields that show up in the mixed correlator hσσi. The coloured outline
represents the allowed region following the analysis of four-point correlators. Experimentally
founds bounds were (∆σ , ∆ ) = (0.5181489, 1.412625) while the numerical bootstrap gives us
the bounds (∆σ , ∆ ) = (1/8, 1). Diagram from Kos, Poland, Simmons-Duffin, 0 14.
Figure 28: The O(N ) archipelago. The coloured outlines are the individual O(N ) archipelagos
which describe the physical bounds on the scaling dimensions and help us locate where the
theories live in the complete landscape. Diagram from Kos et al., 0 15.
interest for various reasons- critical points of non-gauged fermionic models such as the Gross-
Neveu model can be accesed, understanding supersymmetric CFTs such as the 3-dimensional
N = 1 super Ising model, etc.
5.5 Quasi-primary fields

If a field transforms under conformal transformations x 7→ x0 according to
∆
∂x0 d
O(x) = O(x0 ) (282)
∂x
it is called a quasi-primary field of conformal dimension. Here |∂x0 /∂x| is the Jacobian of the
mapping. For the case of rotations and translations, we have
∂x0
= 1, (283)
∂x
and hence, we have O(x) = O(x0 ). For the case of dialations and SCTs, we have
82
∂x0 ∂x0 1
= λd , = . (284)
∂x ∂x (1 + 2b · x + b2 x2 )d
In terms of correlation functions with a basis of quasi-primary fields, we have

 ∆i

n
Y ∂x0 d
 hO1 (x01 ) . . . On (x0n )i.
hO1 (x1 . . . On (xn )i =  (285)
∂x x=xi
i=1
This equation shall be very useful in helping us fix the form of correlators. For example, consider
the two-point correlator, hO1 (x1 )O2 (x2 )i. Invariance under translations and rotations implies
that it must be a function of the absolute value of the distance between the positions x1 and
x2 , i.e.
hO1 (x1 )O2 (x2 )i = f (|x1 − x2 |) . (286)
Next, invariance under dialations implies that there is a scale dependence and hence, we expect
the behaviour to be ∼ |x1 − x2 |−∆1 −∆2 . Fixing an appropriate normalization of the fields Oi ,
the most general expression we can write for the two-point correlator reads
c12
hO1 (x1 )O2 (x2 )i = . (287)
|x1 − x2 |−∆1 +∆2
For invariance under SCT, we first consider the transformation of the form
xµ 7→ x0µ = xµ + bµ x2 − 2xµ b · x + O(b2 ). (288)
The Jacobian corresponding to this transformation takes the following form



1 − 2b · x,
µ=ν
∂x0µ


µ µ µ
= δ ν (1 − 2b · x) + 2 (b xν − x bν ) = (289)
∂xν 

2 (bµ x − xµ b ) , µ 6= ν

ν ν
From this we see that the form of the Jacobian is
∂x0
= (1 − 2b · x)d + O(b2 ). (290)
∂x
Substituting this in 285, we get
83
hO1 (x1 )O2 (x2 )i = (1 − 2b · x1 )∆1 (1 − 2b · x2 )∆2 hO1 (x01 )O(x02 )i
≈ (1 − 2∆1 b · x1 ) (1 − 2∆2 b · x2 ) hO1 (x01 )O(x02 )i (291)
= 1 − 2∆1 b · x1 − 2∆2 b · x2 + O(b2 ) hO1 (x1 )O(x2 )i.

Ignoring the term of order b2 , we see that the infinitesimal behaviour of the two-point correlator
is as follows
δhO1 (x1 )O2 (x2 )i = −2 (∆1 b · x1 + ∆2 b · x2 ) hO1 (x01 )O(x02 )i. (292)
This implies that the two-point function is non-zero only when ∆1 = ∆2 and we have
hO1 (x1 )O2 (x2 )i = (1 − 2∆1 b(x1 + x2 )) hO1 (x1 )O(x2 )i. (293)
Thus, the two-point correlators of the quasi-primary fields satisfy an orthogonality condition.
This completely the fixes the two-point correlator as follows
c12 δ∆1 ∆2
hO1 (x1 )O2 (x2 )i = . (294)
|x1 − x2 |2∆1
In most theories, the matrix with entries cij are diagonal and we can always choose a basis of
local operators such that
δij
hOi (x)Oj (y)i = , (295)
|x − y|2∆i
where ∆i = ∆j .
5.6 Two-dimensional conformal transformations

We saw earlier that when d ≥ 3, ∂ · is a degree 1 polynomial as shown in 246. For the case of
d = 2, however, there are no such restrictions. In this case, something remarkable happens: in
Euclidean space the conformal Killing equation 240 becomes
∂µ ν + ∂ν µ = δµν ∂ · . (296)
This implies
∂1 1 = ∂2 2 , ∂1 2 = −∂2 1 (297)
which is nothing but the Cauchy-Riemann conditions! this implies that (z) is a holomorphic
84
function where we complexify our real space to the following
(z) = 1 + i2 , (z) = 1 − i2 , (298)
where we have introduced the new complex variable (R2 ∼

= C)
z = x1 + ix2 , z = x1 − ix2 . (299)
The corresponding metric takes on the following form
2
ds2 = dxi ⊗ dxi
P
i=1
1 (300)
= 2 (dz ⊗ dz + dz ⊗ dz)
= dz ⊗ dz,
with components gzz = gzz = 1/2 and gzz = gzz = 0. Defining
1 1
∂z ≡ (∂1 − i∂2 ) , ∂z ≡ (∂1 + i∂2 ) , (301)
2 2
we see that the action of these derivatives on z and z read
∂z z = 0, ∂z z = 0. (302)
Since (z) is holomorphic, so is z 7→ f (z) = z+(z) and from this we conclude that a holomorphic
function f (z) = z + (z) gives rise to an infinitesimal two-dimensional conformal transformation
z 7→ f (z). This implies that the metric tensor transforms as follows
∂f ∂f
ds2 = dz ⊗ dz 7→ dz ⊗ dz. (303)
∂z ∂z
From this we can immediately infer that the scale factor is |∂f /∂z|2 . Hence, z 7→ f (z) and
z 7→ f (z) imply Weyl rescalings which by definition, generate conformal transformations. A
two-dimensional conformal transformation is given by any (anti-)holomorphic function (possibly
with appropriate smoothness conditions)
z 7→ f (z) = z + (z), z 7→ f (z) = z + (z). (304)
Notice that this is an infinite-dimensional set of transformations, as opposed to what we observed

in the case of d > 2. The power of complex analysis aids us in simplifying CFT2 since all
85
correlators transform nicely under arbitrary holomorphic transformations. In order to find the
infinite-dimensional algebra, all we have to do is find a basis for the infinitesimal transformations
(z). What we do know is that in the complex plane, all functions can be expanded in a Laurent
expansion, and hence, we have
X
cn −z n+1 .

(z) = (305)
n∈Z
Hence, the transformations 304 are generated by the following generators or Killing fields cor-
responding to a particular n
`n = −z n+1 ∂z , `n = −z n+1 ∂z . (306)
The negative sign is just a convention which is related to the fact that the change of coordinates
that offsets a physical transformation must have opposite signs. It is important to note that
since n ∈ Z, the number of independent infinitesimal conformal transformations is infinite! The
power of complex analysis aids us in simplifying CFT2 since all correlators transform nicely
under arbitrary holomorphic transformations. The generators satisfy the so-called Witt algebra
w with commutation relations
d d
[`m , `n ] = −z n+1 dz , −z m+1 dz

d d2 d d2
= z n+1 (m + 1)z m dz + z m+1 dz 2 − z m+1 (n + 1)z n dz + z n+1 dz 2
d
= ((m + 1) − (n + 1)) z (m+n)+1 dz
(307)
= (m − n)`m+n ,

`m , `n = (m − n)`m+n ,

`n , `n = 0.
TThe first commutation relation defines one copy of w, and because of the other two relations,
there is a second copy that commutes with the first one. Hence, we conclude that conformal
symmetry is expressed as the sum of two isomorphic infinite-dimensional algebras. We pause
here to reflect on what we have just learned. We know from 257, that in d = 2, the dimension
of the conformal group is 6 which is clearly finite. This is at odds with the statement that
conformal symmetry is infinite-dimensional. Where did the infinite symmetries come from?
This is where we note that not all transformations we discussed earlier are well-behaved and
invertible everywhere, only those corresponding to “global conformal transformations”, that
form a six-dimensional group, are. This is rather easy to see. Consider transformations near
z = 0,
86
X ∞
X
v(z) = − an `n = an z n+1 ∂z . (308)
n∈Z n=−∞
This is well-defined at z = 0 if and only if n ≥ −1. Changing variables z 7→ τ = −1/z, the

transformations near z → ∞ (or equivalently, near τ → 0) read
∞
1 −n+1
X
v(z) = − an − ∂τ . (309)
n=−∞
τ
This is well-defined at τ = 0 if and only if n ≤ 1. For the conformal transformation v(z) to be

well-defined at both z = 0 and z → ∞, we require it to take the following form
v(z) = −a0 `0 − a1 `1 − a−1 `−1 , (310)
which is parameterized by three complex or six real parameters a−1 , a0 , a1 ∈ C. The generator
`−1 = −∂z generates translations z 7→ z + b, `0 = −z∂z generates dialations and rotations, and
`1 = −z 2 ∂z generates SCTs. This is quite easy to see. Consider z = reiφ , then
1 i 1 i
`0 = −z∂z = − r∂r + ∂φ , `0 = −z∂z = − ∂r + ∂φ . (311)
2 2 2 2
The following combinations generate dialations and rotations

`0 + `0 = −r∂r , i `0 − `0 = −∂φ . (312)
The set {`−1 , `0 , `1 } generates the Möbius group PSL(2, C). The three Killing fields `−1 , `0 , and
`1 generate a closed subalgebra inside the Witt algebra
[`1 , `−1 ] = 2`0 , [`0 , `−1 ] = `−1 , [`0 , `1 ] = −`1 . (313)
Notice that w is much like the angular momentum algebra but with the wrong signature,
SO(2, 1) ∼ SL(2, R). Hence, there are six global generators and put together with the anti-
holomorphic part this is nothing but the 2D conformal group, SO(2, 1) × SO(2, 1) ∼ SO(2, 2).
To see why the transformations v(z) are well defined for all z ∈ C, note that they generate
fractional linear or Möbius transformations, i.e.
az + b
z 7→ v(z) = , (a, b, c, d) ∼ λ(z, b, c, d), (314)
cz + d
where λ ∈ C, a, b, c, d ∈ C and, ad − bc = 1. These anomaly-free transformations are called the

global conformal transformations. In contrast, the local conformal transformations are in general
anomalous, controlled by something called a central charge c we shall encounter later. The set of
87
all these transformations is isomorphic to the group PSL2, C) = SL(2, C)/Z2 , where the quotient
with respect to Z2 comes from the invariance of the mapping 314 under the multiplication of
all the parameters for −1. As a final comment, we mention here that we shall soon see that
there exists a quantum mechanical deformation of the Witt algebra called a central extension
and this is the only deformation of the algebra compatible with the Jacobi identity. Upgrading
the differential operators `n to operators in a (quantum) Hilbert space Ln , i.e. `n → Ln , we get
the Virasoro algebra.
5.7 Primary fields and Ward identities

Primary operators behave as follows under holomorphic transformations
−∆ −∆
∂z 0 ∂z 0

0 0
O(z, z) 7→ O (z , z) = O(z, z), (315)
∂z ∂z
which is equivalent to
O(z, z)dz ∆ ⊗ dz ∆ = O0 (z 0 z 0 )dz ∆ ⊗ dz ∆ . (316)
Here (∆, ∆) are the holomorphic and anti-holomorphic conformal weights. The latter definition
316 also enables us to consider the primary operator, or just simply primary, O(z, z) as a tensor
of degree (∆, ∆) under conformal transformations. A primary is also a quasi-primary, but the
same does not hold vice-versa. More explicitly, any operators that transforms as 315 is called
a primary of conformal dimension (∆, ∆). If 315 holds only for z 0 , z 0 ∈ PSL(2, C), i.e. only for
global conformal transformations, then O(z, z) is called a quasi-primary. It is also important to
emphasize here that not all operators in a CFT are called primary or quasi-primary. The others
are just referred to as secondary operators. The definition 316 is can be used to constrain the
form of correlation functions of primaries. For example, consider in 316 the two-point function
hΦ1 (z1 )Φ2 (z2 )i. Let us consider a transformation of the form z 7→ z 0 = (z − z2 )/z12 , where
z12 = z1 − z2 . Notice that this is clearly a Möbius transformation when a = 1, b = −z1 , c = 0,
and d = z12 , i.e.
!
a b az + b z − z2
z0 = γ · z = ·z = = , γ ∈ PSL(2, C). (317)
c d cz + d z1 − z2
Hence, Φi are quasi-primaries of conformal weight ∆i . when z = z1 then z 0 = 1, when z = z2

then z 0 = 0, and similarly for the anti-holomorphic counterparts. From 285, the transformation
of two-point correlator reads
∆1 +∆2 ∆1 +∆2

∂z 0 ∂z 0

c12
hΦ1 (z1 , z 1 )Φ2 (z2 , z 2 )i = hΦ01 (1, 1)Φ02 (0, 0)i = . (318)
∂z ∂z ∆1 +∆2 ∆1 +∆2
z12 z 12
88
It is important to remark that both holomorphic and anti-holomorphic weights must match for
this two-point function to be non-vanishing. This is something that directly follows from SCTs.
Like in the case of QFT, correlations functions in CFT also satisfy a Ward identity. Recall that
the Ward-Takahashi identity reads
µ
X
−∂µ hJN (x)Φ(x1 ) . . . Φ(xn )i = δ (4) (x − xi )hδG Φ(x)Φ(x1 ) . . . Φ(xi−1 )Φ(xi+1 ) . . . Φ(xn )i,
i
(319)
where the negative sign on the LHS pops up since we are presently working in Euclidean space.
In particular, when the insertion x ∈
/ {x1 , . . . , xn } then we end up with the quantum analog of
Noether’s theorem,
µ
∂µ hJN Φ(x1 ) . . . Φ(xn )i = 0. (320)
In order to derive a Ward identity for CFT correlators, we first considered an infinitesimal
translation of the following form
xµ 7→ x0µ = xµ + µ (x). (321)
The energy-momentum tensor Tµν = Tνµ for Lorentz-invariant (Euclidean-invariant here) the-
ories is conserved, i.e, ∂ µ Tµν = 0. In a CFT, the conserved current is written in terms of the
energy-momentum tensor as follows
µ
JN = T µν ν . (322)
This is sometimes called the dilation current since the associated charge is the dialation op-
erator. Integrating the current over a surface Sigma gives us the known conserved charge,
´
Pµ = Σ dd x Tµν ν , which are nothing but momentum operators that are translationally sym-
metric. Scale invariance leads to the energy-momentum tensor being traceless, i.e.
µ
∂µ J N = 0 = T µµ . (323)
The tracelessness of the energy-momentum tensor holds in any CFT. In 2D, with coordinates
299 and derivatives 301, tracelessness translates to the following
T µµ = 0 =⇒ Tzz = Tzz = 0. (324)
89
This immediately implies the following conservation laws
∂µ T µν = 0 =⇒ ∂z Tzz = ∂z Tzz = 0. (325)
Stated differently, we have
0 = ∂ z Tzz + ∂ z Tz = 12 ∂Tzz
(326)
1
0= ∂ z Tzz + ∂ z Tzz = 2 ∂Tzz .
This tells us that Tzz is holomorphic and Tzz is anti-holomorphic. Tensors Tzz and Tzz are called
chiral and anti-chiral currents respectively. Let us define
T (z) ≡ Tzz (z, z), T (z) ≡ Tzz (z, z). (327)
We also note that it is a convention to include the factor of 2π in the definition of the energy-
momentum tensor in terms of the variation of the action,
2π δS
T µν = p . (328)
|g| δgµν
Consider now a local infinitesimal holomorphic coordinate transformation z 7→ z 0 = z + (z),

under which the current reads

0, a=z
Ja = (329)
 1 T z (z)z (z) = 2
2π z 2π T (z)(z), a = z,
where the factor of 2 in the numerator comes from the metric tensor g zz . We ignore the anti-
holomorphic coordinate transformation (z) for now, since this can be treated in a similar fashion.
For the CFT current J a and operator basis {Oi (zi }, we have the following identity
2
∂a hJ a O1 (z1 ) . . . On (zn )i = 2π ∂z hT (z)(z)O1 (z1 . . . On (zn )i
n
δ (2) (z − zi )hδOi (zi )O1 (z1 ) . . . Oi−1 (zi−1 )Oi+1 (zi+1 ) . . . On (zn )i.
P
=
i=1
(330)
This establishes the following
90
n
2 X
∂z hT (z)(z)O1 (z1 . . . On (zn )i = δ (2) (z−zi )hδOi (zi )O1 (z1 ) . . . Oi−1 (zi−1 )Oi+1 (zi+1 ) . . . On (zn )i.
2π
i=1
(331)
Integrating the LHS of the above expression over a disk D1 around the point z1 (see figure 29),
we get
´
2
LHS = − 2π D1
d2 σ ∂z hT (z)(z)O1 (z1 ) . . . O(zn )i
¸
dz
=− ∂D1 2πi
hT (z)(z)O1 (z1 ) . . . O(zn )i (332)
= − Res hT (z)(z)O1 (z1 ) . . . O(zn )i,

z→z1
where we have first made use of the divergence theorem followed by the residue theorem. Recall
Figure 29: Integration domain: over the disk D1 where z1 ∈ D1 and zi ∈

/ D1 when i 6= 1.
that the divergence or the Gauss-Ostrogradsky theorem states the following

ˆ ˆ
d
d x ∂a f (x) = dd−1 x n̂a f (x), (333)
M ∂M
where M is a (compact) manifold, ∂M is the boundary of that (compact) manifold, and n̂a is
the normal pointing outward. As a warm-up let us consider an example of integrating the test
function ∂a f = ∂z (1/z) over a disk D with coordinates (σ 1 , σ 2 ). This reads
ˆ ˆ ‰
1 2 1 1 1 2 1 1 1
dσ dσ ∂z = dσ dσ (∂1 + i∂2 ) = dz · , (334)
D1
z 2 z 2i z
where the integration over the contour is done counterclockwise. With z = ρeiθ , this integral
yields
91
˛ ˆ 2π
i 1
iθ 1
dθ ρe iθ
= dθ = π. (335)
2i ρe 2 0
In general, the following holds

ˆ ˛
2 1
d σ ∂z f (z) = dz f (z). (336)
2i
Returning back to 331, we now integrate the RHS to get
RHS = hδO1 (z1 )O2 (z2 ) . . . On (zn )i. (337)
From 332 and 337, we get
− Res hT (z)(z)O1 (z1 ) . . . O(zn )i = hδO1 (z1 )O2 (z2 ) . . . On (zn )i. (338)
z→z1
As can be seen from above, a problem of computing correlators now becomes purely a prob-
lem in complex analysis! We can get analogous expressions for z2 , . . . , zn by integrating over
other disks Di . Let us now apply this for the case when O1 (z) = O(z), a primary (note that
we are only considering the holomorphic part of the field since the computation for the anti-
holomorphic counterpart is entirely analogous). Firstly, we know that it transforms as follows
under a conformal transformation z 7→ z 0 = z + (z),
−∆
∂z 0

O(z) 7→ O0 (z 0 ) = O(z). (339)
∂z
Substituting for z 0 on either side, we get
O0 (z + ) = (1 + ∂z )−∆ O(z). (340)
Expanding the LHS about , we have
O0 (z + ) = O0 (z) + ∂z O0 (z) + O(2 ), (341)
where O is the big-O symbol not to be confused with the primary. The difference O0 (z) − O(z)
reads
δO(z) = O0 (z) − O(z) = − (∆∂z + ∂z ) O. (342)
Hence, for a primary field, we end up with the following identity
92
Res hT (z)(z)O1 (z1 ) . . . On (zn i = h((∆∂z + ∂z ) O1 ) (z1 )O2 (z2 ) . . . On (zn )i, (343)
z→z1
where (. . .)(z) is shorthand for stating that all quantities in the parenthesis are functions of z.
this above equivalence is valid for all holomorphic (z). For example, a particular choice of (z),
say (z) = µz, we have the following identity
Res hµzT (z)O1 (z1 ) . . . On (zn i = h((∆∂z µ + µz∂z ) O1 ) (z1 )O2 (z2 ) . . . On (zn )i. (344)
z→z1
Let us now consider the following case when we have just one primary
Res T (z)(z)O(w) = ((∆∂z + ∂z ) O) (w). (345)

z→w
The product of the energy-momentum tensor and the primary can be expressed as a sum over
a basis of operators as follows
X X Õn (z1 )
T (z)O(w) = (z − w)n Õn (z) = , (346)
(z − w)n
n∈Z n>0
where in the second equality we consider operators with n > 0 since we only care about the
singular terms while calculating the residue. The residue of 346 reads
∞ ∞

(k) (w) Õn (w)
w)m
P P
Res (z)T (z)O(w) = Res m! (z − (z−w)n
z→w z→w m=0 n=1
∞
∞ P

P 1 (m) (w)
m! Õn (w)
= Res (z−w)n−m
(347)
z→w m=0 n=1
∞
P (n−1) (w)
=− (n−1)! Õn (w),
n=1
where (`) denotes the `th derivative with respect to w. Putting this all together we have,
∞ (n−1)
X (w)
− Õn (w) = ∆(1) (w)O(w) + ∂w O(w). (348)
(n − 1)!
n=1
We can now read-off all non-vanishing fields Õ to be
Õ1 (w) = ∂w O(w), Õ2 (w) = ∆O(w), (349)
while all other fields Õn are null-valued for singular terms. Plugging this back into the operator
expansion 346, we get
93
∆O(w) ∂O(w)
T (z)O(w) = 2
+ + reg. (350)
(z − w) z−w
where ∂O(w) = ∂w O(w) and reg. stands for regular, non-singular terms. This is called the
conformal Ward identity. As noted earlier, we are only writing the holomorphic part of the
primary and suppressing the anti-holomorphic dependence. In reality, we actually have O(w, w)
but since T (z) is holomorphic, it only acts on the w component of the primary. Analogously,
for the anti-holomorphic counterparts, we have the following operator expansion
∆O(w, w) ∂O(w, w)
T (z)O(w, w) = + + reg. (351)
(z − w)2 z−w
The primary O(w, w) can be regarded to be comprised of two chiral primaries O(w) and O(w),
i.e. O(w, w) = O(w)O(w). Any operator that is a primary will have an operator expansion as
shown in 350. This is also called the operator product expansion (OPE) of the stress tensor and
the primary. From 350 and 351 we see that the OPE of the primary with T (z) and T (z) have at
most second-order poles, this is a characteristic of primary fields. The OPE of any other field of
the theory with the stress tensor will have poles of higher order. For example, we shall soon see
that the stress tensor itself is not a primary since the OPE with itself has at most fourth-order
poles but can be made a primary if something called central charge is null-valued. The main
takeaway here is that the primary O(w) is, in a sense, an eigenstate of the stress tensor.
5.8 The free boson

One of the most important examples of a 2D CFT is the free boson. This is described by the
following action in flat space
ˆ
α
S= d2 x ∂µ φ∂ µ φ, (352)
2π
where α is a coupling constant, the 1/2π is a convention, and ∂µ φ∂ µ φ = η µν ∂µ φ∂ν φ. In a general

background, we can write
ˆ
α p
S= d2 σ |g|g µν ∂µ φ∂ν φ, (353)
2π
where g = detgµν with gµν = diag(+1, +1), and φ(σ 0 , σ 1 ) is a real massless scalar field defined
on a cylinder given by σ 0 ∈ R and σ 1 ∈ R subject to the identification σ ! ' σ 1 + 2π. Using 328,
we obtain the following expression for the stress tensor

µν 1 µν
µ ν
T = α −∂ φ∂ + g ∂µ φ∂ν φ . (354)
2
In 2D, we know that the stress tensor is tracessless and hence,
94
2
T µµ = −∂µ φ∂ µ φ + ∂µ φ∂ µ φ = T11 + T22 = 2Tzz = 0. (355)
2
The non-zero components of the stress tensor reads
T zz (z) = −α∂ z φ∂ z φ, Tzz (z) = T (z) = −α∂z φ∂z φ. (356)
In order to study the action 353 in greater detail, we map the cylinder S 1 × R to the complex
plane C2 via the following transformation
z = exp σ 0 · exp iσ 1 .

(357)
With this change of variable and d2 z = (i/2)dzdz = d2 σ, we denote the new field as φ(z, z) and
the action 353 takes the following form
ˆ
α
S= d2 z 2∂φ∂φ, (358)
2π
where we have used the notation ∂ = ∂z and ∂ = ∂z . The equation of motion is obtained taking
a functional derivative of the action 358 with respect to the field as follows
δ
0= δφ S
´
α
d2 z 2 ∂δφ · ∂φ + ∂φ · ∂δφ

= 2π
´ (359)
α
d2 z

= 2π 2 ∂ δφ · ∂φ − δφ · ∂∂φ + ∂ (∂φ · δφ) − ∂∂φ · δφ
´
α
d2 z

= − 2π 4∂∂φ · δφ.
Hence, the equation of motion reads
2φ = 4∂∂φ = 0, (360)
The conservation of the stress tensor (holomorphic and anti-holomorphic parts) is
∂T (z) = 0 = ∂T (z). (361)
The quantum version of the stress tensor contains normal-ordered fields, i.e.
T (z) = −α : ∂φ∂φ : (z). (362)
95
what we wrote down in 356 is the classical stress tensor while the above normal-ordered ex-
pression is the quantum counterpart. It is to be noted that the VEV of the stress tensor is
null-valued, i.e. hT (z)i = 0. To compute the propagator, we can make use of the path integral.
Path integrals are independent of the field variable φ and are left invariant under a change of
variables. We can now make use of the fact that the integral of a total derivative is null values
as long as there are no boundary contributions to write
ˆ ˆ
δ α
e−S φ(σ 0 ) = Dφ e−S 2φ(σ)φ(σ 0 ) + δ (2) (σ − σ 0 ) ,

0= Dφ (363)
δφ)(σ) 2π
which gives us
4α
h∂∂φ(σ)φ(σ 0 )i = δ (2) (σ − σ 0 ). (364)
2π
Now, recall that ∂(1/z) = πδ (2) (σ) = ∂∂ ln z with |z|2 = zz and hence, we finally get the
propagator
1
hφ(z)φ(z 0 )i = − ln |z − z 0 |2 . (365)
2α
Notice that the RHS of the above equation is not single-valued which implies that φ(z) is not
entirely a nice operator. Suppose we want to now define the composite operator φ2 (z). We
consider the product φ(z)φ(z 0 ) and then take the limit z − z 0 → 0. However, in any correlator,
there would be a non-zero contribution from the Wick contraction between these two fields,
which is singular when the points become coincident. Hence, in order to obtain a well-defined
limit, we follow definition 262 to perform normal-ordering where we subtract this contraction as
follows
1
: φ(z)φ(z 0 ) : = φ(z)φ(z 0 ) + ln |z − z 0 |2 . (366)
2α
This normal-ordered product is now non-singular under the limit z → z 0 and this can be taken
advantage of to define composite operators. Since the product φ(z)φ(z 0 ) can be expressed as the
addition of the normal-ordered product and the correlator, from 365 and 366 we have
1
φ(z)φ(z 0 ) = : φ(z)φ(z 0 ) : +hφ(z)φ(z 0 )i = − ln |z − z 0 |. (367)
2α
A better operator than φ(z) is the holomorphic derivative of it, i.e. ∂φ(z). Taking the derivative
of 367, we have
1 1
∂φ(z)∂φ(z 0 ) = − + reg. (368)
2α (z − z 0 )2
96
Let us now study the structure of the OPE of different operators with the stress tensor.
1. Primary operator O(w) = ∂φ(w):

Consider the primary O(w) = ∂φ(w) with conformal weights (∆, ∆) = (1, 0). The OPE
between this operator and the holomorphic stress tensor T (z) reads
T (z)O(w) = −α : ∂φ(z)∂φ(z)O(w) : +T (z)O(w)
= −α : ∂φ∂φ : (z)O(w) + reg. (369)
2∂φ(z)
= −α − α1

(z−w)2
+ reg.
where we could have also considered the contraction between the other primary ∂φ(z) and
O(w). Both contractions are equivalent and yield the same result. Now, we expand ∂φ(z)
around z = w to get
∂φ(z) = ∂φ(w) + (z − w)∂ 2 φ(w) + O (z − w)2 .

(370)
Substituting 370 in 369 we get
∂φ(w) ∂ 2 φ(w) ∆O ∂O(w)

T (z)O(w) = − 2
+ + reg. = 2
+ + reg. (371)
(z − w) (z − w) (z − w) z−w
Here, ∆ = 1 and since the OPE has poles at most of order two, we conclude that O = ∂φ
is indeed a primary. It is straightforward to check that operators such as ∂ 2 φ(z), φ3 φ(z),
and so on are not primaries.
2. Vertex operator O(z) = Vp (z) ≡ : eipφ(z) :

µ (w)
Consider the vertex operator Vp (w) =: eipφ(w) := eipµ φ with conformal weights (∆, ∆) =
(p2 /4α, p2 /4α). We first take a look at the OPE of ∂φ and O
∞
in n : φn (w) :
P
∂φ(z)O(w) = n! ∂φ(z)p
n=0
∞ :φn−1 (w):
P in pn 1 (372)
= (n−1)! − 2α z−w + ...
n=0
ip O(w)
= − 2α z−w + reg.
Normal-ordering yields
1
: ∂φ(z)O(w) := ∂O(w) + . . . (373)
ip
97
where other terms are of the order O [(z − w)]. Now, the OPE of the stress tensor and the
vertex operator reads
T (z)O(w) = −α : ∂φ(z)∂φ(z)O(w) : +T (z)O(w)
= −α : ∂φ∂φ : (z)O(w) + reg.

ip 1 (374)
= −α − 2α ∂φ(z)O(w) z−w + reg.
p2 O(w) ∂O(w)
= 4α (z−w)2 + z−w + reg
∆O(w) ∂O(w)
= (z−w)2
+ z−w + reg.
Here, ∆ = p2 /4α and since the OPE has poles at most of order two, we conclude that the
vertex operator is a primary.
Central charge
As mentioned in the previous section, the stress tensor itself isn’t a primary and this can be
explicitly verified by looking at the OPE of the stress tensor with itself, which in the case of the
free boson reads
T (z)T (w) = α2 : ∂φ∂φ : (z)∂φ∂φ : (w)
2
1
= 2α2 − 2α 1
(z−w)2
+ 4α :∂φ(z)∂φ(w):
−2α(z−w)2
+ reg.
1/2
= (z−w)4
+ (−2α) :∂φ(z)∂φ(w):
(z−w)2
+ reg. (375)
1/2 2 φ∂φ:(w)
= (z−w)4
+ (−2α) :∂φ∂φ:(w)
(z−w)2
+ (−2α) :∂ z−w + reg.
1/2 2T (w) ∂T (w)

= (z−w)4
+ (z−w)2
+ z−w + reg.
Here, the term with a fourth-order pole is called the central term and the coefficient that sits in
the numerator is one-half of something called a central charge.
5.9 Energy-momentum tensor and the Virasoro algebra

For the free boson, the central charge is c = 1. More generally, the OPE of the stress tensor
with itself, or the TTOPE, reads
c/2 2T (w) ∂T (w)

T (z)T (w) = 4
+ 2
+ + reg. (376)
(z − w) (z − w) z−w
Also, it can be easily checked that T (z)T (w) = T (w)T (z). Notice that the stress tensor can
98
only be considered as a primary in theories where c = 0. For one free boson, we have a central
charge of c = 1 while a theory of D number of free bosons comes with a central charge of c = D.
A theory of one free fermion on the other hand has a central charge of c = 1/2. The central
charge can in principle be any real number, though positive for unitary theories and it is, in
some sense, a measure of the number of degrees of freedom of the theory. We shall soon see that
the central charge appears in the Virasoro algebra as the central extension. Lastly, the central
charge is related to the trace anomaly in curved spacetimes. For an arbitrary metric gµν the
trace of the stress tensor Tzz does not vanish but rather its expectation value is given by
c c
hTzz i = − R(g) ⇐⇒ hT µµ i = − R(g), (377)
48 12
where R(g) is the two-dimensional Ricci scalar. It is important to note that general covariance
dictates the form of the anomaly and it is the constant of proportionality that we are to estab-
lish.
Let us now consider an infinitesimal coordinate transformation of T (z). This is easily obtained
by taking the residue of the product of the OPE with the infinitesimal (z), i.e.
δT (w) = Res [(z)T (z)T (w)] . (378)
Expanding the infinitesimal around w, we obtain
(z − w)2 (2) (z − w)3 (3)

(z) = (w) + (z − w)(1) (w) + (w) + (w) + . . . , (379)
2! 3!
where (`) denotes the `th derivative with respect to w. Substituting this back into 378 and
focusing only on the simple poles and those of zeroth-order, we obtain (changing variables
w → z)
c 3
δT (z) = (2∂ + ∂) T (z) + ∂ (z), (380)
12
where the factor of 1/12 comes from the factor 1/2 associated with the central charge c in the
TTOPE and the factor 1/3! in the expansion of . We can now use Cauchy’s theorem to extract
the coefficient of this expansion as follows
˛
1 c 3
δhT (z)i = dz (z)hT (z)T (w)i = ∂ (z). (381)
2πi 12
There are two main properties of the stress tensor we will focus on
1. Transformation law under a conformal mapping.
99
2. Taylor-Laurent expansion in terms of an operator basis.
The first property leads us to introduce the Schwartz derivative or the Schwartzian,
2
η 000 (z) 3 η 00 (z)

{η, z} = − . (382)
η 0 (z) 2 η 0 (z)
We will soon see that the second property naturally paves a path to obtain the Virasoro algebra.
Recall that we had three modes `0 , `±1 that formed a finite subalgebra inside the Witt algebra
as shown in 313. Indeed, this is nothing but the angular momentum algebra so(3) but with a
different sign. What is important to note is that we want to find a function f (z) that is left
invariant under the transformations of this group,
af (η) + b
f (η) → , a, b, c, d ∈ R, ad − bc = 1. (383)
cf (η) + d
We now want the transformations to be left invariant under translations, scaling, and inversion.
Translations imply that we ought to look at derivatives f (η), f 0 (η), etc. Scalings imply that we
ought to consider ratios f (p) (η)/f (q) (η). Inversions help fix a constant (unimportant). The ideal
candidate that obeys the above form the Schwartz derivative,
2
f 000 (η) 3 f 00 (z)

{f, η} ≡ 00 − . (384)
f (η) 2 f 0 (z)
We can now write down the form of the T (z) under a (conformal) transformation
2
dη c
T (z) = T (η) + {η, z}. (385)
dz 12
It is important to note that the Schwartz derivative is null-valued if and only if η(z) is a Möbius
transformation. Under this case, we have
" #
(bc − ad)2 T (η)
T (z) = T (η) = . (386)
(d + cz)4 (d + cz)4
The stress tensor T (z) is nothing but a Taylor-Laurent expansion, which in operator basis
(around the origin) takes the form
∞
X Ln
T (z) = . (387)
n=−∞
z n+2
We can now use the Cauchy’s theorem to extract the coefficients of the poles as follows
˛
1
Ln = dz z n+1 T (z), (388)
2πi C
100
where C is a closed contour around the point where an operator is defined on which we may
choose to act Ln . The natural question to ask ourselves now is what exactly is the algebra of
the operators we have just defined. In order to obtain this algebra, we must first define the
commutators in the complex plane. The intuition is as follows: commutators of equal-time
operators that show up in time-ordered correlation functions are nothing but the evaluation of
operators at slightly different times. We can think of this to be analogous to the modulus in the
complex plane. Consider two operators O(z) and R(w). The commutator of these operators
read
[O(z), R(w)] = lim (O (e z) R(w) − R(w)O (e z)) . (389)

→0
This formula has a rather neat interpretation when one of the operators is a contour integral as
shown below
h¸ i ¸ ¸
C
O(z), R(w) = lim C+
O(z)R(w)dz − C−
O(z)R(w)dz
→0
(390)
¸
= w
O(z)R(w)dz.
Observe that the commutator became a contour integral around the point of insertion of the
operator R(w). Now, we know that Ln is nothing but a contour integral and with the newfound
knowledge of operator-contour relation we obtain
1
¸
z n+1 T (z)dz, O(w)

[Ln , O(w)] = 2πi C
(391)
1
¸
= 2πi w z n+1 T (z)O(w)dz.
The takeaway is that action of the operator Ln on a field O(w) filters the conformal field that
shows up in front of (z − w)−n in the OPE of T(z) and O(w). Now that we know the action of
Ln on a field, we are ready to compute the commutator of Ln ’s and hence, fix the algebra. The
commutator [Ln , Lm ] is
˛ ˛
dz dw dw dz
[Ln , Lm ] = − z n+1 T (z)wm+1 T (w). (392)
2πi 2πi 2πi 2πi
Let us consider the action of one part of the commutator on a field at z, i.e.
˛ ˛
1 00
Ln , Lm Φ(z) = dz dz 0 (z 00 − z)n+1 (z 0 − z)m+1 T (z 00 )T (z 0 )Φ(z). (393)
(2πi)2 C0 C
This is pictorially shown in figure 30. From the deformed contour in 30, we can obtain two
separate contours- one encircling z and one encircling z 0 as shown in 31.
After all this world-building, we are now ready to explicitly compute the commutator.
101
Figure 30: (Left) contour C encircles only z while contour C 0 encircles both z and z 0 . (Right)
contour C 0 , deformed.
Figure 31: (Left-to-right) The first diagram represents Ln Lm Φ(z), the second represents
Lm Ln Φ(z) and the third represents the result of the commutator (or equivalently, the deformed
contour)
˛ ˛
1 1
[Ln , Lm ] = dz 0 (z 0 − z)m+1 dz 00 (z 00 − z)n+1 T (z 00 )T (z 0 )Φ(z), (394)
(2πi) C
2πi C0
| {z }
I1
where in the integral over C 0 , points z 0 and z 00 can be taken arbitrarily close, and hence, we
replace T (z 00 )T (z 0 ) with it’s OPE counterpart. This yields
˛
2T (z 0 ) ∂T (z 0 )

1 00 00 n+1 c/2
I1 = dz (z − z) + + + ... (395)
2πi C0
(z 00 − z 0 )4 (z 00 − z 0 )2 (z 00 − z 0 )
We can now reduce the integral by taking residues as follows
¸ h
c/2 2T (z 0 ) ∂T (z 0 )
i
I1 = 1
2πi C0
dz 00 (z 00 − z)n+1 (z 00 −z 0 )4
+ (z 00 −z 0 )2
+ (z 00 −z 0 ) + ...
¸ ¸ ¸
c dz 00 (z 00 −z)n+1 dz 00 00
0 (z −z)
n+1
dz 00 00
0 (z −z)
n+1
= 2 C0 2πi (z 00 −z)4 +2 C0 2πi T (z ) (z 00 −z)2 + C0 2πi ∂T (z ) (z 00 −z) + ...
(396)
c (z 00 −z)n+1 00 00
h n+1 n+1
i
= 00lim 0 Res 2 (z 00 −z)4 + 2T (z 0 ) (z(z−z)
00 −z)4 + ∂T (z 0 ) (z(z−z)
00 −z)4
z →z
c 2 − 1)(z 0 − z)−2 + 2(n + 1)T (z 0 ) + (z 0 − z)∂T (z 0 ) (z 0 − z)n .

= 12 n(n
Lastly, all we are to do now is to perform the integral over C post which we find the commutator
to read
102
c 2 − 1)(z 0 − z)n+m−1 + 2(n + 1)T (z 0 )(z 0 − z)n+m+1 + (z 0 − z)n+m+2 ∂T (z 0 )

[Ln , Lm ] = 12 n(n
c 2
= 12 n(n − 1)δn+m,0 + 2(n + 1)Ln+m − (n + m + 2)Ln+m
c 2
= (n − m)Ln+m + 12 n(n − 1)δn+m,0 .
(397)
A similar computation holds for the anti-holomorphic part and it can be easily checked (expected
from Witt) that the commutator of a holomorphic and anti-holomorphic operator is null-valued.
The algebra of the basis operators of the stress tensor, in its full glory is presented below
c 2
[Ln , Lm ] = (n − m)Ln+m + 12 n(n − 1)δn+m,0 ,
c 2

Ln , Lm = (n − m)Ln+m + 12 n(n − 1)δn+m,0 , (398)

Ln , Lm = 0.
This is the Virasoro algebra. Notice that when we have a null central charge c = 0, this algebra
reduces to our old friend the Witt algebra w shown in 307. This is no coincidence since the
two algebras are indeed related. We should have almost expected the central charge to show up
because what we have basically done is called a central extension. Let v denote the Virasoro
algebra. Then, with x̃, ỹ ∈ v, x, y ∈ w, c ∈ C, the statement that the Witt algebra admits a
central extension can be expressed as follows
v = w ⊕ C,
(399)
[x̃, ỹ]v = [x, y]w + cp(x, y), [x̃, c]v = 0, [c, c]v = 0,
where p : w × w → C. In other words, this technology tells us that we ought to add a function
f (◦), that is proportional to the identity, to the algebra as follows
[Ln , Lm ] = (n − m)Ln+m + f (n, m) · 1. (400)
Now, we are to simply plug the algebra into the Jacobi identity which would give us constraints
on the function. The central charge reveals itself when these constraints are solved. Central
extensions are classified by the second cohomology group H 2 (G, C), where G is an arbitrary
group that acts on the Abelian group C. This cohomology is zero for any finite-dimensional semi-
simple Lie algebra. For example, su(2) will not have any central extensions. Hence, consistently
sl(2, C) subalgebra 313, generated by `1 , `−1 and `0 has no central extension and still correspond
to global conformal transformations. Lastly, a key question to address is why do we even want
to do this, to begin with. The reason is a bit subtle. We consider the central extension since we
103
don’t want unitary representations of the algebra but rather want those of the projective algebra
(remember that we are working with PSL(2, C)) and it is known that projective representations
must satisfy an algebra up to a central charge. The takeaway is that quantum mechanically,
there exists a deformed algebra (with its only deformation compatible with the Jacobi identity)
called the Virasoro algebra.
5.10 Ground state energy

We now turn to the Hamiltonian formalism and consider quantization on a cylinder S 1 × R, with
a periodic space coordinate θ and a Euclidean time coordinate τ , that can be combined into one
complex coordinate w = τ + iθ. We can then make use of the following transformation
z = ew = eτ eiθ , (401)
to map the cylinder to the punctured complex plane. This maps the infinite past to the complex
plane and the infinite future to the point at infinity. Notice that translations in time τ 7→ τ + a
are mapped to dialations z 7→ ea z, and translations in space θ 7→ θ + b are mapped to rotations
z 7→ eib z. This is pictorially depicted in figure 32.
Figure 32: The cylinder and the complex plane show a mapping between circles at fixed radius
and fixed time. States on the cylinder live on these circles while local operators on the plane
live at points. Circles of different radius are mapped onto different sections of the cylinder and
vice-versa via the map 401.
The energy (Hamiltonian) in of the cylindrical system can be obtained by integrating the time-
time component of the stress tensor over the spatial coordinate, i.e.
104
ˆ
H= dθ Tτ τ . (402)
As noted in the previous section, we know that the stress tensor is related to the Schwartzian
as shown in 385. Let us denote the Schwartzian by S(z, w) = {z, w}. For the conformal map
401, the Schwartzian reads
2
z 000 (w) 3 z 00 (w)

3 1
S(z, w) = 0 − = ew e−w − ew e−w = − . (403)
z (w) 2 z 0 (z) 2 2
Hence, the stress tensor of the complex plane is related to the stress tensor of the cylinder as
follows
c −2w
TC (z) = TS 1 ×R (w) + e . (404)
24
We can the compute the energy as follows

´ (S 1 ×R)
H= dθ Tτ τ
´
(S 1 ×R) (S 1 ×R)

= dθ Tww + Tww (405)
´
c+c
dθ e2w TC (z) + e2w T C (z) −

= 24 .
Now, we know that the VEV ofthe stress tensor is null-valued15 , i.e. hΩ|T (z)|Ωi = 0 =
hΩ|T (z)|Ωi. Hence, this tells us that the ground state energy of is
2π(c + c
E ≡ hHiΩ = − , (406)
24
where the factor 2π comes from integration over θ since θ ∈ [0, 2π). In terms of mode expansion,
we know that the stress tensor can be written as a Taylor-Laurent series expansion. The mode
expansion of the stress tensor of the cylinder would then be shifted by −c/24, i.e.
X c
TS 1 ×R (w) = e−nw Ln − . (407)
24
n∈Z
The exponent in the mode expansion can be simplified as follows
e−nw = e−n ln z = z −n (408)
Substituting 407 in 404, we obtain

15
This will be rigorously proved later.
105
∞
X X Ln
TC (z) = z −n z −2 Ln = . (409)
n=−∞
z n+2
n∈Z
This matches our initial ansatz in 387 and the reason for the shift by a factor of 2 can be traced
back to the transformation of the stress tensor.
5.11 Radial quantization

As mentioned previously, the transformation 401 maps the past infinity on the cylinder to the
origin of the complex plane while the future infinity on the cylinder is mapped to the infinity of
the complex plane. In terms of coordinates, this reads
z=eiw
τ → −∞ ⇐⇒ z = 0,
z=eiw
(410)
τ → +∞ ⇐⇒ z → ∞.
what this implies for operators is that time ordering on the cylinder is replaced by radial ordering
on the complex plane, i.e.

O1 (z)O2 (w), |z| < |w|



T (O1 (z)O(w)) = (411)


O (w)O (z), otherwise

2 1
In scale-invariant theories, it is natural to foliate spacetime with spheres centered around the
origin and consider evolving states from smaller spheres to the larger ones using the dialation
operator. This is called radial quantization. The states live on the spheres which in the case of
the complex plane is just the circle S 1 . Recall now that in the previous section we learned how
to relate the OPE to the commutators as established in 390. Consider now two analytic fields
γ(z) and β(z). Let Γ be the contour integral of the analytic field γ(z), i.e.
˛
Γ= dz γ(z). (412)
The commutator of Γ and β(z) reads
[Γ, β(w)] = Γβ(w) − β(w)Γ

¸ ¸
= C1
dz γ(z)β(w) − C2
dz β(w)γ(z) (413)
¸
= Cw
dz β(w)γ(z).
Now, from 391, we know what the commutator of the generator Ln and an operator O look like.
106
We rewrite this here.
˛
1
[Ln , O(w)] = dz (z − w)n+1 T (z)O(w). (414)
2πi w
Let us consider O(w) to be a primary. We can now simply substitute the OPE between the
stress tensor and the primary we computed back in 374 to obtain
1
¸ n+1

∆O(w) ∂O(w)

[Ln , O(w)] = 2πi w dz (z − w) (z−w)2
+ z−w
h i
1 d z n+1 ∆O(w) 1 z n+1 ∂O(w)
= Res (2−1)! dz (z−w)2
· (z − w)2 + 0! z−w · (z − w) (415)
z→w
= (n + 1)∆wn O(w) + wn+1 ∂O(w),
where we have used the residue formula,
dk−1

1 k
Res f (z) = lim (z − z0 ) f (z) . (416)
z→z0 (k − 1)! dz k−1
The commutator of generators Ln and Lm yield the Virasoro algebra as calculated in 397,
c
[Ln , Lm ] = (n − m)Ln+m + n(n2 − 1)δn+m,0 · 1. (417)
12
Notice that the central extension vanishes when n = 0, ±1. This implies that the global trans-
formations (Möbius) do not “see” the central charge.
5.12 State-operator correspondence

In QFT, the states and local operators are not one and the same. States are objects that live over
spatial slices while operators are those that are defined at and live at local points in spacetime.
In CFT, however, there exists an isomorphism between states and operators. We first define
|Oi ≡ lim O(z, z)|Ωi, (418)

z,z→0
where |Ωi is the vacuum state. The state |Oi is defined at z = 0 which is equivalent to insertion
of a so-called “in-state” at the past infinity slice τ → −∞ on the cylinder. The notion of an
“out-state” follows directly from this in the sense that we can simply replace z 7→ w = 1/z
so that the state is now defined at infinity on the complex plane or equivalently, at the future
infinity slice on the cylinder. If O is a quasi-primary (⊃ primary), we define

0 1 1 1 1
O (w, w) = 2∆ O , ≡ O† (w, w). (419)
w w2∆ w w
In terms of state, the bra hO|, reads
107
hO| = lim hΩ|O0 (w, w)
w,w→0
= lim hΩ|Ω† (w, w) (420)

w,w→0
= lim hΩ|z 2∆ z 2∆ O(z, z).

z,z→∞
This defines the final out-state. Let us consider the case when O(z, z) = T (z), the holomorphic
stress tensor with (∆, ∆) = (2, 0). From the Laurent series expansion in 409, we find the
Hermitian adjoint to read

† 1 1 1 X n+2
T (z) = 4 T = 4 z Ln , (421)
z z z
n∈Z
or said differently,
X L†m X Ln
= . (422)
z m+2 z −n+2
m∈Z n∈Z
From this, we observe that the replacing n 7→ −n, we obtain the following equivalence
L†n = L−n . (423)
We can now use all this information about how the generator to prove something important
about the stress tensor. To begin, let us consider the action of the (holomorphic) stress tensor
on the vacuum ket,
X Ln
T (z)|Ωi = |Ωi. (424)
z n+2
n∈Z
Demanding regularity at the origin, we arrive at conditions that identify this state
Ln |Ωi = 0, n ≥ 1, (425)
and analogously for the anti-holomorphic counterpart. The null-valuedness of the VEV hLn iΩ
∀n ∈ Z follows from this. The conditions L0,±1 |Ωi = 0 and L0,±1 |Ωi = 0 establish the fact that
|Ωi is a SL(2, C) invariant. With this newfound knowledge, we arrive at the following fact
hΩ|T (z)|Ωi = 0hΩ|T (z)|Ωi. (426)
But wait, there’s more. Going back to 415, the implication of this can be stated as follows
108
[Ln , O(0)] = 0, n > 0 (427)
Acting this commutator on the vacuum reveals
0 = [Ln , O(0)] |Ωi
= Ln O(0)iΩi − O(0)Ln |Ωi (428)
= Ln lim O(z)|Ωi,
z→0
i.e. the action of the generator Ln on the primary state is null valued
Ln |Oi = 0, n > 0. (429)
What about the cases when n = 0 and n = −1? To see these cases, again consider the action of
the commutator on the vacuum state
[Ln , O(0)] |Ω| = Ln |Oi
(415)
¸
1
= 2πi w
dz z n+1 lim T (z)O(z)|Ωi
w→0
¸
∆O(w) ∂O(w)

1
= lim dz z n+1 (z−w)2
+ |Ωi
w→0 2πi w z−w
 (430)

∆ lim O(w)|Ωi = ∆|Ωi, n = 0


 w→0





= |∂Oi, n = −1









0, n>0
5.13 Representation theory of conformal algebra

Succinctly put, the action of the generator on the state |O∆ i reads



0, n = 1, 2, . . .







Ln |O∆ i = ∆|O∆ i, n = 0 (431)








|∂O i, n < 0

∆
109
Here, |O∆ i is a primary of conformal weight ∆. blueSpecifically, |O∆ i is called a SL(2, R)
primary, i.e. it is killed by the generator L1 . The primary we have here is called a Virasoro
primary is on the other hand, is defined as follows
L0 |O∆ i = ∆|O∆ i
(432)
Ln |O∆ i = 0, n > 0.
All Virasoro primaries are SL(2, R) primaries, but the opposite is not true. The states |∂O∆ i
are called descendants of the primary state since these directly follow from the action of the
generator Ln on the primary state with n < 0. For example, when n = −1, we have the first
descendant
L−1 |O∆ i = |∂O∆ i. (433)
To label all the descendants, we write
|O∆ ; n1 , . . . , nk i ≡ L−n1 . . . L−nk |O∆ i

(434)
0 < n1 ≤ n1 ≤ . . . . . . ≤ nk .
Using the commutation relation of the Virasoro algebra in 397, we have
k
X
L0 |∆; n1 , . . . , nk i = (∆ + N ) |∆; n1 , . . . , nk i, N = ni . (435)
i=1
This tells us that the descendants are also eigenstates of L0 with an eigenvalue that is related
to their level N . The negative modes L−m of behave as creation operators of the quantum
harmonic oscillator with the only difference being that they move by m the eigenvalues of the
state that they act on. Given a primary |O∆ i, we can build a so-called Virasoro multiplet by
acting on this state with raising operators as shown in table 5.
L0 -eigenvalue State
∆ |O∆ i
∆+1 L−1 |O∆ i
L2−1 |O∆ i
∆+2
L−2 |O∆ i
L3−1 |O∆ i
∆+3 L−2 L−1 |O∆ i
L−3 |O∆ i
... ...
Table 5: Tower of Virasoro multiplets obtained by repeated action of the raising operator.
110
The descendant states are linearly independent and their set provides an irreducible represen-
tation of the Virasoro algebra. The space of states comprising of the primary and arbitrary
descendant states is called the Verma module or the conformal family. Say we now want to
count the number of descendants in a given level N . Let us call this number P (N ). The num-
bers can be fixed combinatorially with the first few given in table 6.
Level N Combinatorics P (N ) Generator(s) Young Tableaux
1+1 (L−1 )2
2
2 L−2
1+1+1 (L−1 )3
3
2+1 L−2 L−1
3 L−3
Table 6: Table shows the different combinatorics for a given level and the associated generators
and young tableaux.
The first few P (N ) read
P (0) = 1, P (1) = 1, P (2) = 2, P (3) = 3, P (4) = 5, . . . (436)
The task of writing a formula for general N simplifies to considering partitions of N , i.e. the
unique number of ways in which we can assemble N (= Z+ ) by summing other positive integers.
The general formula reads
∞
X 1
P (N )q N = ∞ . (437)
qn)
Q
N =0 (1 −
n=1
Expanding this expression we can read off the first few P (N ) stated in 6
P (q) = 1 + q + 2q 2 + 3q 3 + 5q 4 + . . . (438)
For the case of N being asymptotically large, the resulting partition of N is given by the Hardy-
Ramanujan formula which reads
q
exp π 2N3
P (N ) ' √ . (439)
4 3N
This is a special case of the Cardy formula which states that the number of states with L0 -
111
q
eigenvalue N grows exponentially like ∼ exp π 2N
3 c , where c is the central charge. This is
another way to see that the central charge indeed measures the number of degrees of freedom.
The space of conformal fields

The Virasoro algebra can also act on a space of conformal fields in a completely analogous fashion
to its action on quantum states. Given a conformal field O(z), by definition of the operators
Ln , we have
X 1
T (z)O(w) = (Ln O) (w). (440)
(z − w)n+2
n∈Z
Just as we had encountered before, for a primary field O, the action of Ln reads
(L0 O) (z) = ∆O(z)
(L−1 O) (z) = ∂O(z) (441)
(Ln O) (z) = 0, n ≥ 1.
The other generators Lm with negative index are responsible for the creation of descendants
fields
O(m) (z) ≡ (L−m O) (z). (442)
All the fields can be recovered by the following recurrence relation
O(n1 ,...,nk ) (z) ≡ (L−n1 . . . L−nk O) (z), (443)
with ordering n1 ≤ n2 ≤ . . . ≤ nk . Again, akin to 435, these fields are eigenvectors of the
generator L0 with eigenvalues given by
L0 O(n1 ,...,nk ) (z) = (∆ + N ) O(n1 ,...,nk ) (z), (444)
Pk
where N = i=1 ni . The conformal fields O(...) and the derivatives of the same form a conformal
family. The claim we now want to prove is the following:
Knowledge of the correlation functions of primaries is sufficient to obtain all the correlation
functions of descendants fields.
Consider for example,
112
hO1 (z1 ) . . . On−1 (zn−1 ) (L−n On ) (zn )i = L−k hO1 (z1 ) . . . On−1 (zn−1 On (zn )i, k > 0 (445)
where L−k is a linear differential operator defined as follows
n−1
X
(1 − k)∆i 1 ∂
L−k = + . (446)
(zi − z)k (zi − z)k−1 ∂zi
i=1
To go about proving the claim, we first use the conformal Ward identity to obtain
n
X (1 − k)∆i 1 ∂
hT (z)O1 (z1 ) . . . On (zn )i = + hO1 (z1 ) . . . On (zn )i. (447)
(zi − z)k (zi − z)k−1 ∂zi
i=1
Here, we can pull out the derivative ∂zi since two insertions are at coincident points. The next
step is contour manipulation as shown in figure 33.
Figure 33
Following the steps outlined in the figure, we have
¸
hO1 (z1 ) . . . On−1 (zn−1 ) (L−k On ) (zn )i = dz
C1 2πi
(z − zn )1−k hT (z)O1 (z1 ) . . . On (zn )i
(374)
¸ n h
(1−k)∆i
i
dz
(z − zn )1−k 1 ∂
P
= C1 2πi (zi −z)k
+ (zi −z)k−1 ∂zi
hT (z)O1 (z1 ) . . . On (zn )i
i=1
¸ n h
(1−k)∆i
i
dz
(z − zn )1−k 1 ∂
P
=− C2 2πi (zi −z)k
+ (zi −z)k−1 ∂zi
hT (z)O1 (z1 ) . . . On (zn )i
i=1
n−1 h i
Res (z − zn )1−k (1−k)∆ 1 ∂
P i
=− (z −z)k
+ (zi −z)k−1 ∂zi
hO1 (z1 ) . . . On (zn )i
j=1 z→zj
i
n−1
Ph i
(1−k)∆j 1 ∂
=− (zj −zn )k
+ (zj −zn )k−1 ∂zi
hO1 (z1 ) . . . On (zn )i,
j=1
(448)
113
where we have used the following expansion to obtain the penultimate line
(z − zn )1−k = ((z − zj ) + (zj − zn ))1−k = (zj − zn)1−k + (zj − zn )−k (z − zj ). (449)
This ends our proof.
5.14 Further Reading
References
[2] Francesco, P., Mathieu, P., & Sénéchal, D. (2012). Conformal field theory. Springer Science
& Business Media.
114

Main

Uploaded by

Copyright:

Available Formats

Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Main

Uploaded by

Copyright:

Available Formats

AQFT Lecture Notes February 15, 2023

3 Topic-III: Wilsonian Renormalization Group (in Real Space) 45

5 Topic-V: Conformal Field Theory 68

Instructor: Miranda C. N. Cheng

1.1 Idea of path integrals from quantum mechanics

1.2 Path integral in quantum mechanics

hf |ii = hxn |e−i(tn −t0 )Ĥ |xi i. (3)

= N eiδt L(xi ,ẋi ) ,

where N is a normalization constant we ignore and

This infinite-dimensional integral is known as a functional integral, and the representation of

Table 1: QM→QFT map

Type Gaussian form Result

1.4 The Classical limit

1.5 Correlators in the path integral

For instance, for n = 2 we have

1.6 Generating functional

and hence, it is natural to define the variational partial derivative as follows

Setting the source to zero leads to

which is nothing but the 1-point correlator up to a normalization constant! In particular,

and thereby 24.

1.7 Schwinger-Dyson Equation in Canonical Quantization

A straightforward generalization yields

Since the second term vanishes, we see that

∇SD Ẑ[J] = 0 (42)

The classical equation of motion states

Quantum mechanically, the statement is that of the correlator

1.8 Schwinger-Dyson Equation in Path Integral Formalism

Ẑ[J] = Z[J] (48)

1.9 Property of path integrals: time-ordering

×hφj+1 |e−iδt Ĥ(tj ) φ̂(~

×hφj 0 +1 |e−iδt Ĥ(tj 0 )

Similarly, for the case of t < t0 , j < j 0 , we have

This is elucidated in figure 2.

Figure 3: The path from (xi , ti ) to (xf , tf ) is separated into (n + 1) paths.

1.10 Example: solving the free theory

The generating functional can be readily written as follows

A(x)DF (x, y) = δ (4) (x − y). (56)

It is important to note here that DF (x, y) = DF (x − y) due to Lorentz invariance. This is a

As a result, we can write 55 as

using the Fourier transform of the Dirac delta function

Figure 4: The three contributions to the Feynman propagator.

1.11 Perturbation theory

then each piece in the sum takes the following form

Note that the part in orange is the free theory propagator

1.12 Connected generating functionals

iW [J] ≡ ln I[J]. (66)

Schematically, it is shown as in Figure 5.

1.13 1PI Green’s functions

1.13.1 Quantum effective action and its relation to W [J]

Now we show that it is equal to

where φ = φJ is similarly determined by the extremum condition

Note that φ = φJ is precisely the φ(x) that solves

to situations in the presence of the current J.

1.13.2 The role of Γ[φ]

1. They are connected as the name suggests.

The eom of this action is nothing but

where we used 76 and 66.

where we have used 84. Note that

where we have used 78. More compactly, we can simply write

for a (space-time independent) constant , we have