April 4, 2014

INTRODUCTION TO THE
RENORMALIZATION GROUP

Antti Kupiainen

1 Ising Model

We discuss first a concrete example of a spin system, the Ising model. This is a simple
model for ferromagnetism, i.e. the phenomenon that certain materials (e.g. iron) are can
stay permanently magnetized even in the absence of an external magnetic field. We can
think of such materials consisting of elementary magnetic moments, residing in atoms
located at a crystal lattice. A magnetic moment m ∈ R3 interacts with a magnetic field
B ∈ R3 with energy
E = −m · B
(· is the scalar product). Energy is minimized by having m parallel to B.
In Ising model one simplifies by letting each m take only two values, parallel or an-
B
tiparallel to the field: m = µ |B| σ, σ ∈ {1, −1} and µ > 0 a constant. The variable
σ is called "spin" where the terminology comes from the fact that the atomic magnetic
moments often come from the spin degree of freedom of electrons.
The crystal lattice is modeled by a regular lattice which for definiteness we take to be
Z i.e. x = (x1 , · · · , xd ) ∈ Zd with xi ∈ Z. To each x ∈ Zd , we associate a spin variable
d

σx ∈ {−1, 1}. We call σ = {{σx } | x ∈ Zd } a spin configuration on Zd . The set of all spin
d
configurations is denoted by Ω = {−1, 1}Z .
Physical lattices are finite and so we consider also spin configurations on finite subsets
Λ ⊂ Zd , |Λ| < ∞ where |Λ| be the number of elements in Λ. . We denote by ΩΛ = {−1, 1}Λ
the spin configurations in Λ. In practice |Λ| is very large (> 1023 ) so we need to study
the |Λ| → ∞ limit, so-called thermodynamic limit. For simplicity we mostly take Λ a
cube centered at origin: for L ∈ N let ΛL := {x ∈ Zd | |xi | ≤ L, i = 1, · · · , d} “cube of
side 2L + 1”.

1
Definition 1.1 Let |Λ| < ∞. The Ising Hamiltonian in Λ is HΛ : ΩΛ → R given by
X X
HΛ (σ) = −J σx σy − h σx (1.1)
{x,y}∈BΛ x∈Λ

where J > 0, h ∈ R and BΛ denotes the set of nearest neighbor pairs1

BΛ = {{x, y} | x, y ∈ Λ, |x − y| = 1}. (1.2)

Remarks. 1. The constant h equals µ|B| in the previous discussion.

2. The nearest neighbor interaction favors spins being parallel: if h = 0, H takes its
smallest value when σx = 1 for all x or σx = −1 for all x. This models ferromagnetism.
h
If h 6= 0 then H has a unique minimum at σx = sign h ≡ |h| .
Since the spins interact with their neighbors it will be necessary to discuss boundary
conditions for the model. Indeed, (1.1) is the so called free boundary condition Hamiltonian
where only σx with x ∈ Λ enter. Important examples of other boundary conditions are:
1. + boundary condition. Take H as above, but demand that x or y is in Λ. For x 6∈ Λ
let σx = 1.

σ = +1

2. − boundary condition is similar with σx = −1 for x 6∈ Λ.

3. Periodic boundary condition. Take Λ = (ZL )d where ZL = {0, 1, · · · , L − 1} and let BΛ
be as in (1.2) where we replace |x − y| = 1 by dL (x, y) = 1 where dL (x, y) is the periodic
distance
X
dL (x, y)2 = |xi − yi mod L|2 (1.3)
i

i.e. we view ZL as the cyclic group of order L. We can also consider ZL1 × ZL2 × · · · × ZLd .

More generally we define:

d
Definition 1.2 Let σ̄ ∈ {1, −1}Z be arbitrary. The Ising Hamiltonian with σ̄ as boundary
condition is defined as
X X
HΛσ̄ (σ) = −J σx σy − h σx (1.4)
{x,y}∈B̄Λ x∈Λ
Pd
1
We use the norm |x| = ( i=1 x2i )1/2 .

2
where

B̄Λ = {{x, y} | x or y ∈ Λ, |x − y| = 1} (1.5)

and it is understood in the sum that σx = σ̄x if x ∈

/ Λ (resp. y).

Λσ σ̄ Λc

Note that the ± boundary conditions are special cases where σ̄x = ±1, ∀x. Also note
that Hσ̄ depends on σ̄x only for x ∈ ∂Λ = {y ∈ Λc |dist(y, Λ) = 1}. Indeed,
X X X
σx σy = σx σy + σx σ̄y .
{x,y}∈B̄Λ {x,y}∈BΛ x∈Λ,y∈Λc ,|x−y|=1

Denote by HΛb.c. (h, σ) the Ising Hamiltonian in Λ with given b.c. as above. We also set
J = 1 since it will play no role below.
In statistical mechanics we view the the spin configurations σ ∈ ΩΛ random variables
whose probability distribution is determined by the Hamiltonian.
Remark. Recall some definitions from probability theory. A measure µ be of total mass
one on a σ-algebra Σ of subsets of some set M is a probability measure. A ∈ Σ is called
an event and µ(A) is the probability of A. Let f : M → R be measurable, where R is
equipped with the Borel σ-algebra, the smallest σ-algebra containing open sets in R. We
say that f is a random variable. The distribution of f is the probability measure ν on R
with ν(B) = µ(f −1 (B)). For the mean of f we use the following notations
Z
f dµ ≡ hf i ≡ E f
R R
and the variance of f is f 2 dµ − ( f dµ)2 .
Definition 1.3 The Ising measure on ΩΛ is the measure (note : ΩΛ is a finite set,
|ΩΛ | = 2|Λ| )
1 b.c. (h,σ)
µb.c.
β,h,Λ (σ) = b.c.
e−βHΛ
Z(β,h,Λ)
b.c.
where Z(β,h,Λ) is the partition function:
b.c. (h,σ)
X
b.c.
Z(β,h,Λ) = e−βHΛ
σ∈ΩΛ
X
so µ is a probability measure, µ(ΩΛ ) = µ(σ) = 1.
σ∈ΩΛ

3
 1
Remark. In physics, β = kT where k = Boltzman’s constant and T = temperature : β
is the “inverse temperature”. Note that for β → ∞ (T → 0) the minima of H dominate :
at low temperatures we expect magnetism.
In the real world |Λ| ∼ 1023 , so we inquire
1◦ What happens as Λ → Zd ? Is there a limit measure µ = lim µΛ on Ω ? This is called
Λ→Zd
the thermodynamic limit.
2◦ Does µ depend on b.c. ? How does it depend on β, h ?
We will study µΛ via its correlation functions :
Definition 1.4 Let A ⊂ Zd , |A| < ∞. Denote
Y
σA = σx .
x∈A

The correlation function of spins in A is (let A ⊂ Λ)

X 1 X −βHΛb.c. (h,σ)
hσA ib.c.
β,h,Λ := σA µb.c.
β,h,Λ (σ) = e σA .
σ∈ΩΛ
Z σ∈Ω
Λ

We often drop β, h in the notation and write hσA ib.c.

Λ .

Examples. For A = {x}, hσx ib.c. b.c.

Λ is the magnetization at x. For A = {x, y}, hσx σy iΛ is
the pair correlation or 2-point function.

2 Infinite volume limit

Suppose we succeed proving that limΛ↑Zd hσA ib.c. d

β,h,Λ exists for all finite sets A ⊂ Z where
the limit is taken e.g. along the sequence of cubes ΛL . It is then natural to inquire whether
these numbers actually are correlation functions of some probability measure µb.c. β,h on (some
sigma-algebra on) Ω. Unlike ΩΛ , Ω is not a finite set (indeed, it is uncountable). Thus we
need some measure theory to describe it. Intuitively the "density" of µb.c. β,h is proportional
b.c. b.c.
to exp(−βHZd (h, σ)) but this factor is ill defined since HΛ (h, σ) is proportional to |Λ|.
We will later explain how to make sense of this intuition, but for the time being we will
just show that the limits of correlation functions indeed are moments of some measure.
Let us drop b.c., β and h from the notation and suppose

lim hσA iΛL (2.1)

L→∞

exists for all A ⊂ Zd , |A| < ∞. Let

C0 (Ω) = {f : Ω → R | f depends on finitely many σx }.

4
Every f ∈ C0 (Ω) is a linear combination of the functions σA with |A| < ∞ with the
convention σ∅ = 1. Indeed, if g : {1, −1} → R we may write g(σ) = 21 (g(1) + g(−1)) +
 
1
2
σ(g(1) − g(−1)). Letting Px± f := 12 (f σx =1 ± f σx =−1 ) the desired representation follows
by expanding the product in
Y
f= (Px+ + σx Px− )f
x∈A

where A is the set of x s.t. f depends on σx . From this and (2.1) we conclude

`(f ) := lim hf iΛL (2.2)

L→∞

exists for all f ∈ C0 (Ω). C0 (Ω) is a vector space and ` defines a linear map C0 (Ω) → R.
Clearly we have

`(1) = 1 (2.3)

and

`(f ) ≥ 0 for all f ∈ C0 (Ω), f ≥ 0. (2.4)

Moreover, since |hf iΛ | ≤ supσ∈Ω |f (σ)| (where we take Λ to be large enough to include
the support A of f ; the sup is actually a max) we conclude

|`(f )| ≤ sup |f (σ)| := kf k (2.5)

σ∈Ω

where the RHS defines a norm in the vector space C0 (Ω).

Generally, let M be a compact metric space, and C(M ) = {f : M → R| f continuous}.
C(M ) is a Banach space (= vector space with norm which is complete i.e. Cauchy
sequences have limits), with the norm:

kf k = sup |f (x)|.
x∈M

A linear map ` : C(M ) → R is called a state on C(M ) if (2.3)-(2.5) hold. To get into this
setup we need to discuss the topology of Ω.
d
Ω is a compact metric space: {−1, 1} is compact, so Ω = {−1, 1}Z is also compact
in the product topology by the Tychonov theorem (see [35] chap. 4 or [25] chap. 5). A
metric compatible with this topology is e.g.
X
d(σ, σ 0 ) = 2−|x| |σx − σx0 |
x∈Zd

i.e. two spin configurations are close if they agree in a big box around origin. Using the
Stone-Weierstrass theorem (see [35] chap. 4) one then shows that C0 (Ω) is dense in C(Ω).
Thus our linear functional ` extends to C(Ω) and defines a state there.

5
Homework : Check all this !
We may now use the basic real analysis to get our infinite volume measure. Recall
that on a compact metric space we may consider the σ-algebra of Borel sets of M i.e.
the smallest σ-algebra containing the open sets and Borel measures which are measures
defined on this algebra. Given a Borel probability measure µ on M the linear map
Z
`µ (f ) = f dµ

on C(M ) satisfies clearly the conditions (2.3)-(2.5) and defines a a state on C(M ). Con-
versely:
Theorem 2.1 (Riesz Representation Theorem) Given a state ` on C(M ) there is
a Borel probability measure µ on M such that
Z
`(f ) = f dµ

Proof See Rudin, Real and Complex Analysis, chap 2 [37] or Reed-Simon, vol. 1, chap. 4
[35] 
Exercise Prove this directly for M = ΩΛ !
We have thus obtained the following
Corollary 2.2 Suppose that for all finite A ⊂ Zd the limit limL→∞ hσA iσ̄ΛL exists. Then
there is a probability measure µσ̄ such that this limit equals
Z
hσA i := σA dµσ̄ .
σ̄

We call h−iσ̄ (or µσ̄ ) infinite volume state and the question of thermodynamic limit
is : find all infinite-volume states by taking different b.c. We’ll see that for all b.c. the
Λ → Zd limit exists (at least, through subsequences), but there can be many different µ0 s.

3 d = 1 Ising Model : Transfer Matrix

Consider the Ising model above for d = 1. Let first h = 0 and consider Hσ̄
L+1
X
βHLσ̄ := −β σi−1 σi
i=−L

where σ−L−1 = σ̄−L−1 ≡ σ − and σL+1 = σ̄L+1 ≡ σ + (note that these may vary with L).
Consider first the partition function:
X L+1
Y
ZLσ̄ := eβσi−1 σi .
σ−L ,··· ,σL =±1 i=−L

6
Let T = (Tσσ0 )σ,σ0 =±1 be the 2 × 2 matrix:
!
0 eβ e−β
(Tσσ0 ) = (eβσσ ) = .
e−β eβ

Then,
X
ZLσ̄ = eβσ−L−1 σ−L (T 2L )σ−L σL eβσL σL+1 = (f − , T 2L f + )
σ−L ,σL =±1

± ±
where we use the scalar product (f, g) ≡ σ fσ gσ and f ± are the vectors (eβσ , e−βσ ).
P

T is a symmetric matrix with has eigenvalues the roots of (eβ − λ)2 − e−2β i.e.
2 cosh β := λ1 and 2 sinh β := λ2 . The corresponding orthonormal eigenvectors are
e1 = √12 (1, 1)T , e2 = √12 (1, −1)T and introducing the corresponding orthogonal projec-
tions ! !
1 1 1 1 1 −1
P1 = , P2 =
2 1 1 2 −1 1
we may write
T = λ1 P1 + λ2 P2 .
Since P1 P2 = 0 and Pi2 = Pi

T 2L = (2 cosh β)2L (P1 + (tanh β)2L P2 )

and

ZLσ̄ = (2 cosh β)2L (f − , P1 f + ) + (tanh β)2L (f − , P2 f + ) .

 
(3.1)

We have (f − , P1 f + ) = (f − , e1 )(f + , e1 ). Since

1
(f ± , e1 ) = √ (eβ + e−β ) > 0
2
for all β > 0 the leading term is non vanishing and we get
   
σ̄ 1 −αL
ZL = exp 2L log(2 cosh β) + O + O(e )
L
where α = −2 log(tanh β) > 0 for all β.
Definition 3.1 The free energy in volume Λ is
1
FΛσ̄ = − log ZΛb.c. .
β|Λ|
We get for the 1d Ising and Λ = [−L, L]:
1 1
FLσ̄ = − log(2 cosh β) + O( ) + O(e−αL ).
β L

7
It has the L → ∞ limit
1
F σ̄ = − log(2 cosh β)
β
which is independent on σ̄. Note that log ZL is extensive i.e., once divided by |Λ|, it has
the thermodynamic limit.
Let us next consider the magnetization:
1 X −βHLσ̄ (σ)
hσx iσ̄L = e σx .
ZLσ̄ σ

Proceeding as with the partition function we have

σ̄
X X
e−βHL (σ) σx = fσ−−L (T L+x )σ−L σ σ(T L−x )σσL fσ+L
σ σ−L ,σ,σL

= (2 cosh β)2L (f − , P1 σP1 f + ) + O(e−α(L+x) ) + O(e−α(L−x) )

where in the last formula σ denotes the diagonal matrix σδσσ0 . But P1 σP1 = 0 so the first
term in the numerator vanishes and combining with (3.1) we obtain

|hσx iσ̄L | ≤ C(e−α(L+x) + e−α(L−x) ) → 0 as L → ∞.

Hence the magnetization vanishes.

For the 2-point function with x < y we get in the same way
1 − L+x y−x L−y +
hσx σy iσ̄L = (f , T σT σT f )
ZL+
(f − , (P1 + e−α(L+x) P2 )σ(P1 + e−α(y−x) P2 )σ(P1 + e−α(L−y) P2 )f + )
= . (3.2)
(f − , P1 f + ) + O(e−αL )
Using again P1 σP1 = 0 and P2 σP2 = 0, we get
(f − , P1 σP2 σP1 f + ) −α(y−x)
hσx σy iσ̄L = e + O(e−α(L+x) ) + O(e−α(L−y) ) + O(e−αL )
(f − , P1 f + )

Since σP2 σ = P1 and P13 = P1 we obtain

lim hσx σy iσ̄L ≡ hσx σy i = e−α|y−x| α = − log tanh β > 0 β < ∞.

L→∞

The limit is again independent on σ̄.

We thus obtained

hσx σy i − hσx ihσy i = e−α|x−y] −→ 0

|x−y|→∞

i.e. 2-point correlation function decays exponentially (here we had hσx i = 0). ξ = α1 is
called the correlation length. Note that ξ → ∞ as β → ∞. At low temperatures, the 1d

8
model gets more correlated.
It is easy to show now (exercise) : let x1 < x2 < · · · < x2n . Then
2n n
!
Y X
h σxi iσ̄L L→∞
−→ exp −α (x2i − x2i−1 )
i=1 i=1

i.e. all the correlation functions have a limit which is independent on σ̄ and equals
2n
Y n
Y
h σ xi i = hσx2i−1 σx2i i. (3.3)
i=1 i=1

We get
Theorem 3.2 For the 1d Ising model the thermodynamic limit exists and is independent
on b.c. (for all β ∈ [0, ∞)). The correlation functions of the ∞-volume state are given by
(3.3) with
hσx σy i = (tanh β)|x−y|
1
Exercise. Calculate ZLb.c. (β, h) and F (β, h) = − limL→∞ 2βL
log ZLb.c. (β, h). Show the
limit is independent of b.c. and analytic in β, h.
Remark. T is called the transfer matrix. What was important above was that
(f σ̄ , e1 ) > 0. This is a general fact, following from the Perron-Frobenius theorem (see e.g.
[39] and [21], Theorem 3.3.2, for an extension to compact operators).
Exercise. Consider the d-dimensional Ising model in the cylinder [−L, L] × Zd−1 ` i.e.
we have periodic boundary conditions in d − 1 dimensions with period ` and σ̄ boundary
conditions in the two ends of the cylinder. Write the partition function and the correlations
in terms of a transfer matrix and using the Perron-Frobenius theorem show the limits exist
as L → ∞ (but ` fixed) and are independent on σ̄. We’ll see later the last statement is
not true in low temperatures if we took ` = L.
Remark : What is shown above means that the 1d Ising model has no phase transition
i.e. there is a unique infinite volume state. We’ll discuss this later.

4 d ≥ 2. High and Low Temperature Expansions

Ising model in two dimensions. Onsager succeeded calculating the free energy F (β)
in closed form in 1944 [32]. The result is not an analytic function of β. At the positive
√
real axis F (β) has a point of non analyticity at β = βc given by tanh(2βc ) = 1/ 2
where the second derivative of F (the specific heat) diverges logarithmically ∂β2 F (β) ∼
const. log |β − βc |. Subsequently he showed the existence of phase transition (see below)

9
by computing the magnetization m(β) = hσ0 i+ in closed form. It vanishes if β ≤ β0 and
is nonzero for β > β0 . Furthermore as β ↓ βc m(β) ∼ (β −βc )1/8 . He also calculated the 2-
point correlation function which has exponential decay for β 6= βc and at βc hσ0 σx i ∼ |x|1/4
as x → ∞. These results were revolutionary. It was the first proof of phase transitions
from first principles. The critical exponents (1/8 and 1/4 above and others) were different
form naive expectations (mean field theory) and called for explanation. This was finally
achieved with Wilson’s renormalization group theory 30 years later [44].
Other exactly solvable models exist in d = 2 (see Baxter’s book [4]). However, this
is very exceptional and, in d > 2, there are practically none so other methods are called
for. We will first develop methods to study hσA i for β small (high temperature) and β
large (low temperature). These methods work for all d and for much more general models
than the Ising model. The high and low temperature expansions are still the best ways
to numerically to study the critical point.
d
Recall our setup: Ω = {−1, 1}Z and the boundary condition on the boundary of
Λ ⊂ Zd , |Λ| < ∞, is given by a configuration σ̄ ∈ Ω.
Note that to say σ̄ = +1 means that we have on the boundary of Λ a positive magnetic
field that tends to force σx to be +1. The question is : can this result in a positive hσx i
as Λ ↑ Zd ?
Theorem 4.1 (a) There exists β0 > 0 such that if β < β0 then hσA iσ̄β,h,Λ has a limit as
Λ → Zd (via cubes say), independent on σ̄.
(b) Let h = 0, d ≥ 2. There exists β1 < ∞ such that if β > β1 then there are at least two
different infinite-volume (pure) states h−i+ 6= h−i− .
Remarks. 1. (a) means uniqueness of the Gibbs state, see below.
2. (b) means phase transition. See next section.
3. In d = 2 there are only 2 (pure) Gibbs states [1], in d ≥ 3 there are many more [12].
4. Finally, it can be proved, see [3] that β0 = β1 .
We prove (a) in this section and (b) in section 4.3. The result is due to Peierls,
Dobrushin and Griffiths [33, 9, 22].

4.1 High Temperature Expansion

We use an expansion due to M. Fischer [16]. When J = 0, in (1.4) (and β = 1), µb.c.
Λ is a
product measure
Y ehσx Y
µb.c.
Λ (σ) = ≡ νh (σx ).
x∈Λ
eh + e−h x∈Λ

10
Thus !|A|
X
hσA ib.c.
Λ = σνh (σ) = (tanh h)|A|
σ=±1

factorizes and is Λ-independent. We wish to show that for small β our measure approxi-
mately factorizes. Let h = 0 for simplicity. Write (since σx σy ∈ {−1, 1})

eβσx σy = cosh β + σx σy sinh β = cosh β(1 + σx σy tanh β).

Note that for β small tanh β is small. Consider e.g.

X
σx σy e−βH
σ
hσx σy iΛ = X .
e−βH
σ

Notation. A bond is a nearest neighbour pair {x, y}, x, y ∈ Zd , |x − y| = 1. We may

picture a bond by a line between the points x y
. It is a special subset of Zd .
Let us first discuss the free b.c. case. With this notation
X
HΛf ree (σ) = − σb
b∈BΛ

where we recall BΛ denotes the set of all bonds b ⊂ Λ. Thus (with |BΛ | the cardinality of
BΛ )
f ree Y
e−βHΛ (σ) = (cosh β)|BΛ | (1 + σb tanh β).
b∈BΛ

Insert this in hσx σy iΛ , cancel (cosh β)|BΛ | in numerator and denominator and expand the
product over b: X X Y
σx σy σb (tanh β)|B|
σ∈ΩΛ B⊂BΛ b∈B
hσx σy ifΛree = X X Y .
σb (tanh β)|B|
σ∈ΩΛ B⊂BΛ b∈B

Definition. Given B ⊂ BΛ , a family of bonds, let ∂B denote the set of sites x ∈ Λ that
occur an odd number of times in the bonds in B.

11
Example.
y

∂B = {x, y, z, w}
x w
Z
B
X
Since σ n = 0 if n is odd, we get
σ=±1

X
(tanh β)|B|
B:∂B={x,y}
hσx σy ifΛree = X .
(tanh β)|B|
B:∂B=∅

This is the starting point of the high temperature expansion. We use it to prove the
existence of a finite correlation length.

Theorem 4.2 There exists β0 > 0 such that for β < β0

|hσx σy ifΛree | ≤ Ce−|x−y|/ξ

uniformly in Λ (i.e. C, ξ < ∞ are Λ-independent).

Proof Each B with ∂B = {x, y} contains a subset of bonds P = {bi }ni=1 bi =

{xi , xi+1 }, x1 = x, xn+1 = y (if not, ∂B 6= {x, y}, show this !). We call P a (connected)
path joining x and y.

Thus every B in the numerator can be decomposed as B = P B 0 where P is some

S

set as above and ∂B 0 = ∅. Given B, P is not unique, but let us choose it arbitarily, and
call the choice P (B).
We may then rewrite the numerator as
0
X X
(tanh β)|P | (tanh β)|B |
P B 0 ∈B(P )

where the sum over P runs through the paths joining x and y and B(P ) consists of sets
B 0 ⊂ BΛ such that ∂B 0 = ∅ and P (P B 0 ) = P . Since B(P ) ⊂ {B ∈ BΛ |∂B = ∅} each
S

12
term in this sum occurs in the denominator and we have
|B 0 |
P
B 0 ∈B(P ) (tanh β)
P |B|
≤1
B:∂B=∅ (tanh β)

and so
X
0 ≤ hσx σy iΛ ≤ (tanh β)|P | . (4.1)
P

To control the sum over P write as above P = {(xi , xi+1 )}ni=1 , x1 = x, xn+1 = y. Given
xi there are at most 2d nearest neighbours xi+1 and n is at least |x − y|. Thus
X
(4.1) ≤ (2d tanh β)n ≤ Ce−|x−y|/ξ
n≥|x−y|

if 2d tanh β < 1 .
Remarks.
1. We could have chosen P self-avoiding and so could have had only 2d − 1 choices :
xi+1 6= xi−1 . Thus (2d − 1) tanh β < 1 suffices.
√
2. In d = 2 , tanh βc = 2 − 1 = 0, 4142.... We got tanh β < 31 which is not that bad. As
d → ∞, our estimate tanh βc ∼ 1/(2d − 1) becomes exact by combining [16] and [18], see
[6].

We now slightly modify the previous argument to prove part (a) of Theorem 4.1.

Proof of Theorem 4.1 (a) We prove uniqueness (independence on b.c) of the limit.
Existence is analogous. Consider σ̄, σ̄ 0 ∈ Ω and
σ̄
X
e−βHΛ (σ) σA
0 σ
hσA iσ̄Λ − hσA iσ̄Λ = X σ̄
− (same with primes)
e−βHΛ (σ)
σ
σ̄ σ̄ 0 0
X
e−βHΛ (σ)−βHΛ (σ ) (σA − σA0 )
σ,σ 0
= X σ̄ σ̄ 0 0
. (4.2)
e−βHΛ (σ)−βHΛ (σ )
σ,σ 0

We expand slightly differently (to understand why, see Remark after the Proof).
0 0
eβ(σx σy +σx σy ) = e−2β (1 + fxy )
0 0
fxy = eβ(σx σy +σx σy +2) − 1.
Note that

0 ≤ fxy ≤ 4β(1 + fxy ) (4.3)

13
using 0 ≤ ez − 1 ≤ zez for z ≥ 0, and 0 ≤ σx σy + σx0 σy0 + 2 ≤ 4. Expand in powers of
f as before, but only in the numerator : let B̄Λ be the set of bonds intersecting Λ (i.e.
b = {x, y}, either x or y is in Λ), then
X X Y
(σA − σA0 ) fb
0 B⊂B̄Λ σσ 0 b∈B
hσA iσ̄Λ − hσA iσ̄Λ = XY (4.4)
(1 + fb )
σσ 0 b∈B̄Λ

Let us say that b ∈ B is connected in B to A if ∃{bi }ni=1 , bi ∈ B i = 1, · · · , n such that

bi = {xi , xi+1 }, b1 = b and xn+1 ∈ A, see the picture.

b1 b2 b3 bn

Let B0 ⊂ B be the set of b in B connected to A. Suppose that no bond in B0

intersects Λc (i.e. all are in Λ i.e. all fb depend on σ, σ 0 and not on σ̄, σ̄ 0 ). Then that
term in (4.4) vanishes. Indeed, the sum over σx , σx0 with x ∈ ∪b∈B0 b ∪ A factorizes out and
vanishes, since b∈B0 fb is symmetric under the interchange of σ and σ 0 whereas σA − σA0
Q

is antisymmetric.
Hence, the nonvanishing terms in (4.4) have B’s such that B includes a connected path P
of bonds joining A to Λc .
Bound the numerator as
Y Y Y Y Y
|(σA − σA0 ) fb | ≤ 2 fb fb ≤ 2(4β)|P | (1 + fb ) fb
b∈B b∈P b∈B\P b∈P b∈B\P

using (4.3) in the last inequality (Recall that fb ≥ 0 !!); and so, picking, as in the proof
of Theorem 4.2, for each B, a choice P (B) of P ,
XX Y X X X Y Y
| (σA − σA0 ) fb | ≤ 2 (4β)|P | (1 + fb ) fb (4.5)
B σ,σ 0 b⊂B P σσ 0 B 0 ∈B(P ) b∈P b∈B 0

where B(P ) is the set of B 0 ⊂ B̄Λ \ P satisfying P (B 0 ∪ P ) = P . Now,

X Y X Y Y
fb ≤ fb = (1 + fb )
B 0 ∈B(P ) b∈B 0 B 0 ⊂B̄Λ \P b∈B̄Λ \P

so
X XY
(4.5) ≤ 2 (4β)|P | (1 + fb ). (4.6)
P σσ 0 b∈B̄Λ

14
 P Q
The denominator in (4.2) equals σσ 0 b∈B̄Λ (1 + fb ), so combining (4.2) and (4.6) we
obtain
0
X XX X
| hσA iσ̄Λ − hσA iσ̄Λ | ≤ 2 (4β)|P | ≤ 2 (4β)|P | . (4.7)
P x∈A y∈∂Λ P :x→y

Taking 8dβ < 1 the sum over paths is bounded by C(8dβ)|x−y| and then
0
| hσA iσ̄Λ − hσA iσ̄Λ | ≤ C|A| |∂Λ|(8dβ)dist(A,∂Λ) . (4.8)

where ∂Λ is the boundary of Λ. For Λ = ΛL a cube of side L |∂Λ| ∝ Ld−1 , dist(A, ∂Λ) ∝ L
and thus (4.8) is bounded by ≤ C|A|Ld−1 δ L with δ < 1. This tends to zero as L → ∞. .

Remark Note that the positivity of fxy was crucial in (4.5). This is why we did not use
the previous expansion. This method generalizes to a very general class of Hamiltonians,
see [7].

Exercises.
1. Estimate |hσA iσ̄Λ − hσA iσ̄Λ0 | in for Λ ⊂ Λ0 the same way:
0
X X
e−βHΛ (σ) e−βHΛ0 (σ ) (σA − σA0 )
σ∈ΩΛ σ 0 ∈ΩΛ0
hσA iσ̄Λ − hσA iσ̄Λ0 = X X 0
.
e−βHΛ (σ)−βHΛ0 (σ )
σ∈ΩΛ σ 0 ∈ΩΛ0

0
Expand e−βHΛ (σ)−βHΛ0 (σ ) as above and note that only B 0 s connecting A to Λc contribute.
Use this to show that the limit as Λ % Zd exists.

2. Similarily, write
1 X −βH(σ)−βH(σ0 ) 0
X 0
|hσX σY iσ̄Λ − hσX iσ̄Λ hσY iσ̄Λ | = e [σX − σX ][σY − σY0 ]/ e−βH(σ)−βH(σ ) ,
2 σσ0 σσ 0
0
expand e−βH(σ)−βH(σ ) as above and prove:

|hσX σY iσ̄Λ − hσX iσ̄Λ hσY iσ̄Λ | ≤ Ce−αdist(X,Y )

since here only B 0 s connecting X and Y will contribute i.e. all correlations functions
decay exponentially.

4.2 Low Temperature Expansions

X +
Consider HΛ+ (σ) = − σb . For β → ∞, e−βH reaches its maximum when all
b
σb = σx σy = 1 i.e. if all σx = +1 (because of the + bounadry condition). Consider,

15
say in d = 2, the configuration σx = 1 ∀x 6= x0 , σx0 = −1:

+ + +
+ − +
+ + +
This has energy βH = βHminimum +8β (4dβ in general dimension). Consider a connected
region R of −:
+ + + + +
+ − − − +
+ − − − +
+ − − − +
+ + + + +
This has H = Hmin + 2|∂R| where |∂R| = # bonds {x, y} with x ∈ R, y ∈ Rc . We say
that there is a contour around R. Let us formalize this. Given σ ∈ ΩΛ , let

C(σ) = {b ∈ B̄Λ | σb = −1}

i.e. all bonds in Λ or joining Λ to Λc where σx 6= σy .

The dual lattice of Zd is the lattice Zd + ( 21 , 12 , . . . , 12 ) = {(x1 , . . . , xd )| xi − 12 ∈ Z,
i = 1, . . . , d} where we consider both imbedded in Rd . Let us visualize bonds b = {x, y}
as the closed line segments in Rd from x to y x y
.
u z

Similarly is a 2-cell {x, y, z, u} a 3-cell etc. 2-cells are often called

x y

plaquettes.
Now a bond b in Zd defines a unique (d − 1)-cell b∗ "orthogonal" to it in the dual
lattice:

b*
b*

b b

d=3
d=2

Let B̄Λ∗ = {b∗ | b ∈ B̄Λ }. Let us denote

C ∗ (σ) := {b∗ ∈ B̄Λ∗ | b ∈ C(σ)}

i.e. C ∗ (σ) is a set of bonds (d=2) or plaquettes (d=3) in the dual lattice.

16
 We say two bonds b 6= b0 are connected if |b ∩ b0 | = 1 (i.e. they share one site) and two
plaquettes p 6= p0 are connected if |p ∩ p0 | = 2 (i.e. they share a bond). Given a set B of
bonds (plaquettes) consider the graph G(B) with vertex set B and edges {b, b0 } if b and b0
are connected. Let Gα be the connected components of G(B) and Bα the vertices of Gα .
We call Bα the connected components of B. We say B is connected if G(B) is connected.
By some abuse we say B and B 0 are disjoint if no b ∈ B, b0 ∈ B 0 are connected.
We will now characterize the connected components of C ∗ (σ).
Definition 4.3 A contour γ is a connected set γ ⊂ B̄Λ∗ , such that (d=2) each site x of
the dual lattice belongs to an even number of bonds b∗ , with b∗ ∈ γ or (d=3) each bond
of the dual lattice belongs to an even number of plaquettes b∗ ∈ γ. Hence contours are
closed paths (d=2) or closed surfaces (d=3). We say a family Γ of contours is compatible
if all γ ∈ Γ are connected and all γ, γ 0 ∈ Γ are disjoint.
Example.

Allowed :

Not allowed :

Lemma 4.4. The family Γ(σ) of connected components of C ∗ (σ) is a compatible family
of contours. Conversely, given a compatible family Γ of contours there is a unique σ such
that Γ = Γ(σ).
Proof. C ∗ (σ) has the property that x belongs to an even number of bonds (see figure).
Hence its connected components have this property.
Conversely, given Γ, let x0 ∈ Λc (so σx0 = +1) and let P be any path of bonds from x0
to x. Let N (P ) be the number of b ∈ P with b ∈ C(σ), i.e. the number of times P crosses
contours (see figure). Put σx = (−1)N (P ) . This is well defined since if P 0 is another path
then N (P ) − N (P 0 ) is even (show!). 

b*1

b1
u v
b2 We have since σ 2 = 1, σu σv σv σw σw σy
b*4
σy σu = 1 = σb1 σb2 σb3 σb4 ⇒ even num-
x b*2
ber of σbi = −1
b4
y w
b3
b*3

17
 Λ
- +
+ +
+

Let us denote by GΛ the set of compatible families Γ of contours γ ∈ B̄Λ∗ . The Lemma
implies:
X
−βHΛ+ (σ) = βNΛ − 2β |γ|
γ∈Γ(σ)

and
X Y
ZΛ+ = eβNΛ e−2β|γ| (4.9)
Γ∈GΛ γ∈Γ

where we sum over the set of compatible families of contours Γ. Also,

X Y Y
e−2β|γ| (−1)Nx (Γ)
Γ∈GΛ γ∈Γ x∈A
hσA i+
Λ = X Y (4.10)
e−2β|γ|
Γ∈GΛ γ∈Γ

Nx (Γ) = # of γ ∈ Γ surrounding x.

4.3 Magnetization
We prove Theorem 4.1. (b) by proving
Theorem 4.3. (Peierls argument). There exists β1 < ∞, δ > 0 such that for β > β1
−
hσ0 i+
Λ = −hσ0 iΛ ≥ δ

for all Λ.
Proof. We have

hσ0 i+
Λ = P(σ0 = 1) − P(σ0 = −1) = 1 − 2 P(σ0 = −1).

Now P(σ0 = −1) equals the probability that there are an odd number of contours sur-
rounding origin which in turn is bounded from above by the probability that there exists

18
a contour surrounding origin. Denote by G0Λ ⊂ GΛ those compatible families of contours
for which there exists γ0 ∈ G0Λ surrounding origin. Hence
X Y X Y
P(σ0 = −1) ≤ e−2β|γ| / e−2β|γ| .
Γ∈G0Λ γ∈Γ Γ∈GΛ γ∈Γ

This we may write as

X X Y X Y
= e−2β|γ| e−2β|γ| / e−2β|γ| .
γ surrounds 0 Γ:{γ}∪Γ∈GΛ γ∈Γ Γ∈GΛ γ∈Γ

Since each term in the sum in the numerator occurs in the denominator, we get
X ∞
X
−2β|γ|
P(σ0 = −1) ≤ e = e−2βn #{γ : γ surrounds 0, |γ| = n}. (4.11)
γ surrounds 0 n=2d

To estimate the number of contours of γ surrounding origin with |γ| = n we pick a

b∗ ∈ B̄Λ∗ and count the number of connected subsets of B̄Λ∗ of cardinality n containing b∗ .
By Lemma 4.4. below this number is bounded by c2n d with c2 = 4 and c3 = 12. Since the
distance of b to origin is bounded by n the number of choices of b∗ is bounded by (2n)d
∗

and so
#{γ : γ surrounds 0, |γ| = n} ≤ (2n)d c2n
d .

Hence, for β > log cd the series in (4.11) converges, uniformly in Λ and tends to zero as
β → ∞. The claim follows. Obviously hσ0 i− +
Λ = −hσ0 iΛ . 
Lemma 4.4. . The number of connected subsets of B̄Λ∗ of cardinality n containing b∗ is
bounded by c2n
d with c2 = 4 and c3 = 12.

Proof. A set B ∗ ⊂ B̄Λ∗ is connected if and only if the graph G(B ∗ ) is. By Lemma 4.5.
any such graph is covered by doing a walk on B̄Λ∗ with starting point b∗ , length 2n and
jumps between connected vertices. In 2d each b∗ has 4 such neighbors and in 3d 12. The
number of such walks is c2n
d . The claim follows. 
Lemma 4.5 (Köningsberg bridge Lemma ) Let G be a finite connected graph and let α0
be a vertex of G. Then there is a path starting and ending at α0 which includes each line
of G only twice.
Proof. Induction in the cardinality of the vertex set v(G). Let v(G) = n and {α1 , α2 } be
an edge of G. Consider the graph G0 = G\{α1 , α2 }. If G0 is connected, by induction there
is walk w : α0 → α0 visiting all its edges twice. The walk w0 : α0 → α1 → α2 → α1 → α0
where the first α1 is the first visit of w to α1 is the desired walk.
If G0 is disconnected and α0 and α1 are in the same connected component induction
provides two walks α0 → α0 and α2 → α2 . Then the desired walk is α0 → α1 → α2 →
α2 → α1 → α0 . 

19
Remark 1 hσ0 i 6= 0 is called spontaneous symmetry breaking : HΛ (σ) is invariant (if
h = 0) under σ → −σ except for the b.c. (HΛf ree and HΛper are). As Λ % Zd the b.c. have a
finite effect ! Another way to state this : Let h 6= 0. Construct limΛ%Zd h−iσ̄Λ,h ≡ h−ih . It
will be independent on σ̄ (see below). Then limh↓0 h−ih = h−i+ , limh↑0 h−i− . In particular,
let m(h) = hσ0 ih . Then m(h) is discontinuous at h = 0. This is called a first order phase
transition.

Remark 2 The existence of the limit Λ % Zd can be proved along the same lines as for
β small above.

4.4 High and Low Temperature Expansions

One can go further and get identities and not only inequalities. Let us return to high
temperature expansion X
(tanhβ)|B|
∂B={x,y}
hσx σy ifΛree = X
(tanhβ)|B|
∂B=∅

Numerator : x

Denominator :

Try to cancel the so-called “vacuum graphs”, i.e. the ones not involving x or y. Decompose
S
B into connected components: B = α Bα , Bα with Bα connected sets of bonds (as defined
in Section 4.3). Denote by B a family of connected, mutually disjoint (i.e. disconnected)
sets of bonds and by B the set of such families. In the numerator, we have one component,
say B1 , such that ∂B1 = {x, y}. So
X X Y XY
hσx σy ifΛree = ρ(B1 ) ρ(B)/ ρ(B).
B1 connected ∂B1 ={x,y} B:{B1 }∪B∈B B∈B B∈B B∈B

We have defined
ρ(B) = (tanh β)|B| .
The algebra we will perform does not depend on the explicit expression of ρ. The cancel-

20
lation is performed by the following trick. Consider the partition function
∞ n
XY X 1 X Y
Z = ρ(B) = ρ(Bα ) (4.12)
B∈B B∈B n=0
n! α=1
(B1 , B2 , . . . , Bn )
Bi ∩ Bj = ∅

∞ n
X 1 X Y Y
= ρ(Bα ) χ(Bα , Bβ ) (4.13)
n=0
n! α=1 α<β
(B1 , . . . , Bn )
( no constraint)

where the n! comes from writing the sum over the unordered sets B = {B1 , . . . , Bn } as
one over ordered ones (B1 , B2 , . . . , Bn ) and on the second line we introduced
(
0 Bα ∩ Bβ 6= ∅
χ(Bα , Bβ ) =
1 Bα ∩ Bβ = ∅

so that the last product imposes the constraint on the first line.
Note that if we didn’t have the constraint in (4.13) we could do the sum
∞ n
X 1 X Y X
ρ(Bα ) = exp( ρ(B)). (4.14)
n=0
n! α=1 B⊂BΛ
(B1 , . . . , Bn )
no constraint)
(

We will now expand the constraint to derive a generalization of (4.14). To achieve this
we write χ(Bα , Bβ ) = 1 − η(Bα , Bβ ) with
(
0 Bα ∩ Bβ = ∅
η(Bα , Bβ ) = .
1 Bα ∩ Bβ 6= ∅

Then
Y X Y
(1 − η(Bα , Bβ )) = (−η(Bα , Bβ )) (4.15)
α<β G (α,β)∈G

where G is a collection of pairs {α, β} α, β = 1, · · · , n, α 6= β. G can be identified with a

graph on the vertex set {1, · · · , n} with edges {α, β} that connect vertices. G is connected
if any two vertices α, β can be connected by paths in G. Decompose G into connected
components
k
[
G= Gi , Gi connected. (4.16)
i=1

Note that this forces the corresponding Bα ’s to form a conneted network since η 6= 0 only
[
if Bα ∩ Bβ 6= ∅ . Let Ci = Bα .
α∈v(Gα )

21
Reorganize the sum (4.13) with (4.15),(4.18) inserted :
sum over k and a sum over k connected sets (C1 , . . . , Ck ).
sum over n ≥ k
sum over all families of k subsets of {1, . . . , n} (i.e. v(Gi ) above)
sum over Bα ’s
sum over connected graphs on those subsets.
So, one gets:
∞ X 1
X X X n! 1
Z = (∗)
k=0 (C1 ,··· ,Ck ) n≥k
n! n1 ! · · · nk ! k!
(n1 , · · · , nk )
X
ni = n
i
 
 
Yk  X X ni
Y  Y
 
· 
 ρ(Bα ) (−η(Bα , Bβ ))

i=1  (B1 , · · · , Bni ) G α=1 (α,β)∈G 
ni
 [
Bj = Ci
connected graph on 
{1, · · · , ni }
j=1

The combinatorial factor (*) is the number of ways to choose a family of k subsets of
{1, . . . , n} of sizes {n1 , . . . , nk }. Indeed, this number equals
n−1
 
! ! ! X
n n − n1 n − n1 − n2  n− ni  1
··· i=1 k!

n1 n2 n3
nk
n! (n − n1 )! 1 n! 1
= · · · = Qk
n1 !(n − n1 )! n2 !(n − n1 − n2 )! k! i=1 ni !
k!
!
n
where comes from choosing the n1 elements of set 1 etc. and we divide by k!
n1
because order of the sets does not matter.
Thus
∞ k ni
X 1 X X Y 1 X X Y Y
Z = ρ(Bα ) (−η(Bα , Bβ ))
k=0
k! n!
i=1 i α=1
(C1 ,··· ,Ck ) (n1 ,··· ,nk ) (B1 , · · · , Bni ) G (α,β)∈G
ni
[
Bj = Ci
connected on
{1, · · · , n}
j=1

∞
!k
X 1 X X
= f (C) = exp f (C)
k=0
k! C

with
n
X X 1 X Y Y
f (C) = ρ(Bα ) (−η(Bα , Bβ )) (4.17)
n
n! α=1
(B1 , · · · , Bn ) G (α,β)∈G
connected on
[
Bj = C
{1, · · · , n}

22
We have achieved our goal. With our assumptions, the exponent is O(Λ). The main
estimate that guarantees this is

Lemma 4.6 Let |ρ(B)| ≤ |B| . There exists 0 s.t  < 0 ⇒ |f (C)| ≤ (2)|C| .
(2 is arbitrary here). (Recall, B is always connected here).

Proof Omitted, see Simon’s book [42] or Seiler [38]. 

Remark 3 Note that this is non-trivial. We can not estimate in (4.17) | − η| ≤ 1. Indeed,
2
the number of graphs on {1, · · · , n} is the subsets of {1, · · · , n} × {1, · · · , n} i.e. 2n and
2
the number of connected graphs is also ≥ ean a > 0 which is much bigger than n! ≤ nn so
the sum does not converge absolutely. One needs to account for cancellations. Note that
P Q Q
all G (−η) = χ(Bα , Bβ ) which is ≤ 1. Connectedness of G makes estimates harder.
As an example consider the trivial case of Ising model in a box of side 2. There is only
one nontrivial term in the high temperature expansion for the partition function:
(−1)n
ρ(B)n
P
Z = 1 + ρ(B) = elog(1+ρ(B) = e− n n

with ρ(B) = (tanh β)4 . On the other hand, in the expression for f (B) we have Bi = B
for all i and so
X ρ(B)n X Y
f (B) = (−1)G
n
n! G {α,β}∈G

where G runs through connected graphs on {1, . . . , n}. Comparing these two expressions
n
we see that this sum equals (−1)
n
.
Given the Lemma, write
X X X 1 X
f (C) = f (C) ≡ fb,Λ
C⊂BΛ b∈BΛ C⊂BΛ :b∈C
|C| b∈B Λ

where we noted that in the second sum each C ⊂ BΛ is counted |C| times. Then
∞
X 1
|fb,Λ | ≤ (2)n Sn
n=1
n

where Sn is the number of connected sets C, |C| = n, containing a given bond b. By

Lemma 4.4. |Sn | ≤ C n . Thus, for  small |fb,Λ | < C < ∞ uniformly in Λ and free energy
1
− |βΛ| log ZΛ is uniformly bounded.
Actually, one easily proves that
1
lim − log ZΛ = −d/β lim fb,Λ := f
Λ%Zd β|Λ| Λ%Zd

independently of b (there are d bonds per site in Zd ).

23
 !
Also, we get : X X
ρ(B) exp f (C)
∂B={x,y} C∩B=∅
hσx σy ifree
Λ = X
exp( f (C))
C
X X
= ρ(B) exp f (C) ≡ ρ̃(B) (4.18)
∂B={x,y} C:C∩B6=∅

and |ρ̃(B)| ≤ |B| eC|B| ≤ (2)|B| . (4.18) is a version of the high temperature expansion.
It can be further processed but we won’t do it here.

4.5 Ising model-summary

One expects the following phase diagram for the Ising model.
• If h 6= 0 the limit of h−iσ̄Λ is independent of the boundary condition σ̄.
• If h = 0 and β ∈ (βc , ∞) there are two translation invariant states states h−i± and if
β ≤ βc the limit of h−iσ̄Λ is independent of the boundary condition σ̄.
• The correlation length ξ(h, β) < ∞ except at h = 0, β = βc . I.e.

|(σA σB i − hσA ihσB i| ≤ C(A, B)e−d(A,B)/ξ .

• At h = 0, β = βc one expects
c
hσx σy i ∼|x−y|→∞
|x − y|a

At d = 2, a = 1/4 (exact)
d=3 a = 1.0364 (numerical, see wikipedia Ising critical exponents)
d≥4 a=d−2 (proven for d > 4)
Because of the behaviour for d ≥ 4, one usually writes a = d − 2 + η, and one calls η the
“anomalous exponent”. We will explain the exponent d − 2 below.
Rigorous results include : βc is unique, ξ < ∞ up to βc , Λ % Zd limit exists for all β and
h, and much else. The deepest is η = 0 for d > 4.
There are two different kinds of phase transitions in the Ising model:
1. First order transition For h 6= 0 let m(h, β) = hσx i where the expectation is in the
unique state (this state is translation invariant so hσx i is independent of x). If β > βc
(low temperatures) then limh↓0 m(h, β) = − lim m(h, β) > 0 i.e. m is discontinuous at
h↑0
h = 0 (if β ≤ βc , m is continuous).

24
 h

unique state

-1 -1
βc β
unique state

2 states

2. Second order transition Let h = 0, consider

 
+
hσ0 iβ = limhσ0 iβ,h ≡ m(β).
h↓0

Then m(β) > 0 for β > βc , m(β) = 0 for β < βc . What about m(βc )? m is continuous at
βc : limT % Tc m(T ) = 0 = m(Tc ) (we use T = β −1 as a parameter).

However, | dm
dT
|→ ∞ as T % Tc . One has m(T ) ∼ |Tc − T |β where β (not equal to T −1
here!) is a critical exponent for the magnetization. In d = 2, β = 18 (exact), in d = 3,
β = 0.3265 (numerical), in d ≥ 4 one expects β = 12 (this is proven for d ≥ 5). The
continuity or discontinuity of the order parameter (m here) is refered to as second order
and first order transition. Note also that ξ remains finite at a first order transition but
becomes infinite at second order transition.

25
Remark 1. The states above are translation invariant i.e.

hσA i = hσA+x i ∀A ⊂ Zd , ∀x ∈ Zd

where A + x = {a + x|a ∈ A}. Thus Ising model has two translation invariant states in
low temperature and one in high temperature or nonzero field.
Ising model has also non-translation invariant ones. One considers so called Dobrushin
boundary conditions, see the figure.

These force a contour (γ in the figure) joining the points where the boundary condition
changes. It turns out that in d = 2 one does not get a new state this way [1]. In a box
of side the contour fluctuates strongly: the expected deviation from the horizontal line is
1
L 2 and an observer sitting at origin will see either a plus state or a minus state: in the
L → ∞ limit the state is a convex combination:

h−i = ph−i+ + (1 − p)h−i− .

with some p ∈ (0, 1). However, in d = 3 at low temperatures the fluctuations are O(1) in
L and one does get a new state [12] with an interface : hσx i > 0 for x far above, hσx i < 0
for x far below. However it is believed that there is a βr > βc , the roughening transition
point so that the interface state disappears for β < βr due to large fluctuations of the
interface.
Remark 2 Why is h 6= 0 unique ? Let e.g. h > 0 Let us try to construct – a state:

26
 P P
The weight of σ = −1 i.e. the configuration with no contours is e−β(− σx σy −h σx ) =
eβ|B̄Λ | e−βh|Λ| . The weight of the configuration where we have a contour at the boundary
i.e. γ = ∂Λ :

2
equals eβ|B̄Λ | e−βh( h |∂Λ|−|Λ|) . For Λ a L−box |∂Λ| ∼ Ld−1 , |Λ| ∼ Ld so the contour is more
probable! Hence, we expect that the Λ → Zd limit is the same as that of the + boundary
condition state. One can make a proof along these lines, see [20].
Remark 3 What about h which is not constant :
X X
βH = β(− σx σy + hx σx ).
x∈Λ

Let us consider hx ’s random : let each hx be a random variable with zero mean : hx = 0
and different hx , hy are independent, identically distributed. Thus we pick a configuration
of h = {hx } randomly from such ensemble and consider the Gibbs state with this h. What
do we expect ? Consider low temperature and say + boundary conditions. A contour γ
costs now an energy and it weight is
X
exp(−β[2|γ| + 2 hx ])
x∈Intγ

+
–
Int γ

X
Let γ be an L-cube, so |γ| ∝ Ld−1 , how about hx ? This is a sum of Ld independent
x∈Int γ
random variables, therefore of size ∝ Ld/2 since its variance is
X X X X
E( hx )2 = Ehx hy = Eh2x + Ehx Ehy = Ld Eh2
x,y x x6=y

27
Thus, if d − 1 > d2 i.e. if d > 2, the magnetic field is unlikely to suppress contours and we
expect 2 states. For d ≤ 2 it seems to have a chance.

Theorem 4.8 For β large, Eh2 small, d ≥ 3 our model has 2 states: one with hσx i > 0
and one with hσx i < 0. For d ≤ 2 there is only one Gibbs state (this holds with Probability
one in h).

Proof See [5] for d ≥ 3 and [2] for d ≤ 2.

4.6 Bounded spin models

Let us say a few words about the general formalism of classical statistical mechanics on
a lattice.
1. The “spins” σx can take more general values than ±1. By a bounded spin model one
means σx ∈ M where M is “bounded”, generally a compact metric space. Let us list a
few examples :
a) q-state Potts model. σx ∈ {1, 2, · · · , q} and the Hamiltonian with free boundary
conditions is X
H(σ) = − δσx ,σy
{x,y}∈BΛ

(δ is Kronecker delta). Thus to minimize energy nearest neighbours want to be the same.
For q = 2 this is Ising model: δσσ0 = 21 (1 + σσ 0 ). The model has Sq symmetry (= group
of permutations on q objects): Let π ∈ Sn (so π is a bijection from {1, · · · , q} into itself),
and Tπ : ΩΛ → ΩΛ : (Tπ σ)x = πσx . Then H(Tπ σ) = H(σ). At low temperatures we have
q phases:
Exercise. Devise a Peierls argument to construct a state h−ip with boundary condition
σ̄x = p so that as β → ∞, P(σ0 = p) → 1.
At high temperatures there is a unique infinite volume limit which can be constructed
with a high temperature expansion. We can copy our arguments for the Ising model by
writing
e−βδσx ,σy = e−β (1 + fxy )
where fxy = eβ(1−δσx ,σy ) − 1 = δσx ,σy (eβ − 1). Then
Y X Y
eβ|BΛ | e−βHΛ (σ) = (1 + δσx ,σy (eβ − 1)) = δσx ,σy (eβ − 1) (4.19)
{x,y}∈BΛ B⊂BΛ {x,y}∈B

We have 0 ≤ fxy ≤ β(1 + fxy ) and the thermodynamic limit and decay of correlations can
be done as in the Ising case
Exercise. Check this!

28
 The high temperature expansion has actually an interesting structure. Let us consider
the partition function i.e. (up to a trivial multiplicative factor) the sum of (4.19) over σ.
The set of bonds B defines a graph G with vertex set v(G) = Λ and edge set e(G) = B.
Let Gα be the connected components of G. The Kronecker deltas force all the spins to
be equal in each v(Gα ) and the spin sum factorizes over the v(Gα ). We obtain then
X X
eβ|BΛ | e−βHΛ (σ) = (eβ − 1)|e(G)| q n(G) (4.20)
σ G

where n(G) is the number of connected components of G (note that each x ∈ Λ not
belonging to any of the bonds in B is a connected component Gα with v(Gα ) = {x} and
e(Gα ) = ∅). Eq. (4.20) is called the Fortyuin-Kasteleyn representation for the partition
function of Potts model. It has an interpretation of a percolation model, see [23].
As the Ising model Potts model as a critical (inverse) temperature βc separating the
high temperature phase from the low temperature phase. In d = 2 and q ≤ 4 the
correlation length ξ is infinite at βc and finite elsewhere and the transition at βc is second
order. For q > 4 the transition is 1st order and ξ < ∞. This should be the case in d > 2
and q > 2 as well. This is proven for q large enough in all dimensions, see [28].
b) O(N )-models. Here σx ∈ S N −1 = sphere in RN i.e. σx is a vector ~σx ∈ RN , k~σx k2 = 1.
We take X
H=− ~σx .~σy .
|x−y|=1

This has O(N ) symmetry: Let R ∈ O(N ) i.e. R = N × N orthogonal matrix, RT R = 1|.
Then H(R~σ ) = H(~σ ) where (R~σ )x = R~σx . The N = 2 case is called the XY -model,
N = 3 the classical Heisenberg model. These are very interesting. First, there is no
symmetry breaking when d = 2 : there are no states with h~σx i =6 0. This is called the
Mermin-Wagner theorem [31, 14]. We’ll prove it in section 5 below.

In d ≥ 3 the O(N ) models have symmetry breaking :

for β ≤ βc , h~σx i = 0, for β > βc h~σx i = m(β)n̂, n̂ ∈ S N −1 .

There are infinitely many low-temperature states, parametrized by unit vectors n̂ [18].
For d = 2, N = 2 and N ≥ 3 behave qualitatively differently. For N = 2 (XY -model
or plane rotator), there is a critical temperature βc where ξ becomes ∞: For β < βc
ξ(β) < ∞ and for β ≥ βc , ξ(β) = ∞. The correlation function has a peculiar decay if
β > βc : hσx σy i ∼ |x − y|−a(β) where a(β) depends on β [27, 19].
For d = 2, N ≥ 3, one expects ξ(β) diverge as ecβ when β → ∞ (which is proven as a
lower bound).

29
c) Gauge theories Consider bonds b with orientation b = (x, y) 6= (y, x) and denote
(y, x) = b−1 .
b
x y
The spins are indexed by oriented bonds b and take values σb ∈ SU (N ) (N × N unitary
matrices, det σb = 1). Moreover we let σb−1 = (σb )−1 . Hence the state space is ΩΛ =
(SU (N ))BΛ .
b1
b4 b2
Let p be a plaquette and choose arbitarily an orientation :

b3
Put :
1
Sp = Re Tr σb1 σb2 σb3 σb4 = (Tr σb1 σb2 σb3 σb4 + Tr σb1 · · · σb4 )
2
1
= (Tr σb1 · · · σb4 + Tr (σb1 · · · σb4 )−1 )
2
(since Tr U = Tr U + = Tr U −1 for U ∈ SU (N )) and
X
HΛ = Sp .
p⊂Λ

This Hamiltonian has a huge symmetry, the gauge symmetry : Let GΛ = ×x∈Λ SU (N )
i.e. g ∈ GΛ is a map Λ 3 x → gx ∈ SU (N ) i.e. at each point x ∈ Λ choose a matrix
gx ∈ SU (N ). GΛ is the gauge group in Λ. It acts on spins as follows. For σ ∈ ΩΛ , g ∈ GΛ ,
let
(U (g)σ)b = gx σb gy−1
where b = (x, y). Then we have

Sp (U (g)σ) = ReTr [gx1 σb1 gx−1 g σ g −1 · · · gx−1

2 x2 b 2 x3
g σ g −1 ] = Sp (σ)
4 x4 b4 x1

so HΛ (U (g)σ) = HΛ (σ).
x1 x2

x4 x3
This is a big symmetry and in all dimensions it remains unbroken. The reason being
that a local symmetry such as U (g), g ∈ GΛ does not change the Hamiltonian, even with
arbitrary boundary conditions, provided those b.c. are on the boundary of Λ0 , with Λ

30
strictly included in Λ0 . Then, the fact that the gauge symmetry remains unbroken is an
immediate consequence of the DLR equation (5.4) below.
2. The last example had a 4-spin interaction. Let us now generalize to give the general
formalism of bounded spin models. Thus, let M be a compact metric space, define as
d
before the infinite configuration space Ω = MZ , and, in a finite volume ΩΛ = MΛ . Ω is
still a compact metric space (by Tychonov), with metric
X
D(σ, σ 0 ) = 2−|x| d(σx , σx0 )
x∈Zd

with d the metric on M. The measures on Ω we are interested in are characterized by

two data, the Hamiltonian and an a priori measure.
Hamiltonian. We want to have general interactions:
Definition 4.9 A potential Φ is a collection of continuous functions

ΦX : MX → R for X ⊂ Zd , |X| < ∞.

(
−σx σy if X = {x, y}, |x − y| = 1
Example In the Ising model we have ΦX (σ) =
0 otherwise
We say the potential Φ is translation invariant if ΦX = ΦX+a for all a ∈ Zd where
X + a = {x + a|x ∈ X}. More explicitly this means ΦX (τa σ) = ΦX+a (σ) where (τa σ)x =
σx+a . We will assume our potentials are translation invariant.
Given a potential the Hamiltonian in Λ , |Λ| < ∞ is given by
X
HΛ (σ) = ΦX (σ).
X⊂Λ

Actually this is the free boundary condition Hamiltonian. (Note that e.g. in gauge the-
ories we rather consider spins on bonds, not sites so these definitions should be changed
appropriately).
A priori measure : Let ν be a Borel probability measure on M. Thus, in Ising model
we had ν(−1) = ν(1) = 21 , in O(N )-model ν is the uniform measure on S N −1 and in
SU (N )-Gauge theories : ν is the Haar measure on SU (N ). The a priori measure in MΛ
is the product measure
Y
dνΛ (σ) = dν(σx ) (4.21)
x∈Λ

We can the define

Definition 4.10 The Gibbs measure in volume Λ (with free boundary conditions) is
1 −βHΛ (σ)
dµΛ (σ) = e dνΛ (σ) (4.22)
ZΛ

31
where ZΛ = ΩΛ e−βHΛ (σ) dνΛ (σ). For more general boundary conditions we pick as before
R

σ̄ ∈ Ω and let, for σ ∈ ΩΛ ,

X
HΛσ̄ (σ) = ΦX (σ ∨ σ̄) (4.23)
X:X∩Λ6=∅
(
σx x ∈ Λ
where we use the notation : σ ∨ σ̄ ∈ Ω : (σ ∨ σ̄)x = . The corresponding
σ̄x x ∈ Λc
Gibbs measure is then
1 −βHΛσ̄ (σ)
dµσ̄Λ (σ) = e dνΛ (σ) (4.24)
ZΛσ̄
Note that (4.23) is an infinite sum so we need to address the issue of convergence.
Definition 4.11 Φ is finite range if there exists R < ∞ such that ΦX = 0 for all X
such that diameter of X is d(X) > R. Here d(X) := maxx,y∈X |x − y|. For finite range Φ,
(4.23) is finite trivially, since continuity of ΦX in σ implies

kΦX k = sup |ΦX (σ)| < ∞ (4.25)

σ∈MX

by compactness of MX .
Exercise. Let d = 1 and Φ finite range R. Show that for all β the correlation functions
hF (σ)iσ̄Λ have thermodynamic limit which is independent on σ̄. Here F depends on finitely
many spins. Moreover show the correlations decay exponentially. Hint: proceed as in 1d
Ising model by introducing a transfer matrix Tσ,σ0 where σ, σ 0 ∈ MR . For simplicity you
may take M a finite set.
Sometimes finite range is not general enough. There are various classes of potentials,
that allow for the Λ % Zd limit of the free energy, the convergence of high temperature
expansions etc.

Our first class of potentials, B0 , consists of Φ such that

X
kΦk0 = kΦX k < ∞,
X ⊂ Zd
0∈X

using the definition (4.25) i.e. the interaction energy of σ0 with all other spins is finite.
X
Note that translation invariance implies kΦk0 = kΦX k for any y ∈ Zd .
X⊂Zd
y∈X

B0 is a vector space, k k0 is a norm and B0 is complete in this norm, i.e. if kΦ(n) −

Φ(m) k0 n,m→∞
−→ 0 then ∃Φ ∈ B0 such that kΦ − Φ(n) k0 → 0 as n → ∞. So B0 is a Banach
space.
The Gibbs measure is well defined:

32
Lemma 4.12 Let Φ ∈ B0 . Then

|HΛσ̄ (σ)| ≤ |Λ| kΦk0 .

Proof We have
 X  X X 1  X X
ΦX (σ) =  ΦX (σ) ≤ kΦX k = |Λ| kΦk0 
   
|X ∩ Λ|

X:X∩Λ6=∅ y∈Λ X ⊂ Zd y∈Λ d
X ⊂Z
y∈X y∈X

We have then
Theorem 4.13 Let Φ ∈ B0 . Then the free energy
1
F = − lim log ZΛσ̄L (ΛL = L − box)
L→∞ Ld β

exists and is independent of σ̄.

Proof. a) Independence of σ̄. Let  > 0. Consider
0
X
|HΛσ̄ (σ) − HΛσ̄ (σ)| ≤ |ΦX (σ ∨ σ̄) − ΦX (σ ∨ σ̄ 0 )|
X∩Λ6=∅,X∩Λc 6=∅

(since terms with X ⊂ Λ cancel)

X X 1
= |ΦX (σ ∨ σ̄) − ΦX (σ ∨ σ̄ 0 )|
y∈Λ
|X ∩ Λ|
X as above
y∈X
X X 1
≤ 2 kΦX k (4.26)
y∈Λ
|X ∩ Λ|
X as above
y∈X

Divide the sum over y into two parts:

Λ
1◦ dist (y, Λc ) ≤ L0
2◦ dist (y, Λc ) ≥ L0
L0

L
Then the contribution of 1◦ to (4.26) is bounded by
1
CL0 Ld−1 kΦk0 ≤ 2 Ld (4.27)

if we take CL0 /L ≤ 12 .

33
 The second sum with the constraint 2◦ is bounded by
X
2Ld kΦX k ≡ Ld δ(L0 ).
X:0∈X, d(X)≥L0
P
Since the sum X:0∈X kΦX k converges, δ(L0 ) → 0 as L0 → ∞. Hence pick L0 so that
δ(L0 ) ≤ 21  and then L so that (4.27) holds. Then we have

(4.26) ≤ Ld

and thus
d 0 0 d
e−βL ZΛσ̄ ≤ ZΛσ̄ ≤ ZΛσ̄ eβL .
So, for L > L()
0
L−d | log ZΛσ̄L − log ZΛσ̄L | ≤ β.
Since this holds for all  > 0
0
lim L−d | log ZΛσ̄L − log ZΛσ̄L | = 0.
L→∞
0
Obviously we may replace ZΛσ̄L by the free boundary condition version above.
b) Existence. It suffices to consider the free boundary condition theory. Let  > 0. Then
∃R such that kΦ − Φ(R) k ≤  where
(
(R) ΦX diam (X) < R
ΦX = .
0 diam (X) ≥ R
So
|HΛ,Φ − HΛ,Φ(R) | ≤ |Λ|
and thus  1 1 
log ZΛ,Φ − log ZΛ,Φ(R)  ≤ β.
 
|Λ| |Λ|


Thus, it suffices to consider Φ of finite range, say R. Let L2 > L1 > R. We will compare
L−d −d
2 log ZΛL2 to L1 log ZΛL1 . Write L2 = nL1 + L with L < L1 and n ∈ N. Then
ΛL2 is a union of nd disjoint L1 -boxes Λ(i) with i = 1, . . . , nd and a region Λ of volume
|Λ| ≤ CL1 Ld−1 (i)
2 . Let σ|Λ(i) ≡ σ . We get

 nd
X 
0 d −1 −1
HΛL (σ) − HΛ(i) (σ ) ≤ CRLd−1
(i)  d d−1
1 n kΦk0 + CL1 L2 kΦk0 ≤ C L2 (L1 + L1 L2 )

i=1

where HΛ(i) is the free boundary condition Hamiltonian. The first term in the middle
expression bounds the contribution of ΦX with X ∩ Λ(i) 6= ∅, X ∩ (Λ(i) )c 6= 0 for some i
and the second one the ΦX with X ∩ Λ 6= ∅. Thus,
R −β P H (σ(i) )
0 d −1 −1 e Λ(i) dνΛL2 0 d −1 −1
e−C βL2 (L1 +L1 L2 ) ≤ R −βH (σ) ≤ eC βL2 (L1 +L1 L2 ) . (4.28)
e ΛL
dνΛL2

34
 Qnd (i)
Now using dνΛL2 (σ) = dνΛ (σΛ ) i=1 dνΛL(i) (σ ) we get
Z
HΛ(i) (σ (i) ) d
P
e−β dνΛL2 = (ZΛL1 )n

we get
1 nd 
 d log ZΛL2 − d log ZΛL1  ≤ C 0 β(L−1 −1
1 + L1 L2 )
 
L2 L2
Since | Ln2 − 1
L1
| ≤ 1
L2
this implies
1 1 
 d log ZΛL2 − d log ZΛL1  ≤ C 00 β(L−1 −1
1 + L1 L2 )
 
L2 L1

and therefore taking L2 → ∞

 1 1 
 lim sup d log ZΛL2 − d log ZΛL1  ≤ C 00 βL−1
 
1
L2 L1

and then L1 → ∞

 lim sup 1 log ZΛ − lim inf 1 log ZΛ  = 0

 
Ld2 L2
Ld1 L1

proving the claim. 

Remarks. 1. Idea in both parts was that boundary energy/volume → 0 as L → ∞. Note
also that we have not proved that h−iσ̄ is σ̄-independent nor that these states converge.
We have only shown that the free energy is independent of the boundary conditions.
2. The space B0 includes Ising type systems
X
H= Jxy σx σy (4.29)
x,y∈Λ

C
provided Jxy has enough decay as |x − y| → ∞ : e.g. |Jxy | ≤ [|x−y|+1] d+ will do.

There is a very important system where Φ 6∈ B0 : the Coulomb gas. It is like (4.29), σx
= charge at x i.e. + or − charges and
X 1
HΛ (σ) = σx σy .
|x − y|
x, y ∈ Λ
x 6= y

(
1
|x−y|
if X = (x, y) X X
1
Hence ΦX = and ΦX = |y|
= ∞. Nevertheless, the Λ % Zd
0 otherwise 0∈X y6=0
−βH
limit exists. The secret is in signs : note that e gets very small contribution from σx =
+1 ∀x or σx = −1 ∀x. The nonzero contribution comes from alternating configurations
(neutral ones : σ is charge). See Simon [42], p. 121.

35
Let us next consider High-temperature uniqueness. Let us proceed as in the Ising
model case. We first write
Y Y Y
e−βHΛ = e−βΦX = e−βkΦX k eβ Φ̃X
X⊂Λ X⊂Λ X⊂Λ
(4.30)
where we note
Φ̃X = kΦX k − ΦX ≥ 0.
Disregarding the σ-independent constant we then expand
Y Y XY
eβ Φ̃X = (1 + fX ) = fXα
X⊂Λ X⊂Λ {Xα } α

where the sum is over all families {Xα } of subsets of Λ and fX = eβ Φ̃X − 1 satisfies
0 ≤ fX ≤ β Φ̃X (1 + fX ).
We can now proceed as in the proof of Theorem 4.1.(a) to study the σ̄ or Λ dependence
of hFA (σ)iσ̄Λ where FA is a continuous function on MA . For instance eq. (4.7) will be
replaced essentially by
XY n
βkΦXi k
P i=1
where P = {X1 , . . . , Xn } is a connected path of sets from A to Λc i.e. A ∩ X1 =
6 ∅,
c
Xi ∩ Xi+1 6= ∅ and Xn ∩ Λ 6= ∅. To control this sum we need to assume a bit more from
the potential, namely that
X
kΦk1 := |X|kΦX k < ∞.
0∈X

Exercise. Show that this condition allows to bound the above sum if β is small enough
and show it tends to zero as d(A, Λc ) → ∞.
Indeed one can prove
Proposition 4.14. Let kΦk1 < ∞. Then, for β small enough the correlation functions
hFA (σ)iΛ converge as Λ ↑ Zd to a σ̄-independent limit and
|hFA GB i − hFA ihGB i| → 0
as dist (A, B) → ∞.
See [5] for more details. Assuming more on the decay of ΦX allows one to get infor-
mation about the decay of correlations. Eg. if we assume that for some α > 0
X
kΦk2 := eα|X| kΦX k < ∞
0∈X

then the correlation function deals exponentially with distance

0

0
|hFA GB i − hFA ihGB i ≤ Ce−α d(A,B)
with 0 < α < α.

36
5 Gibbs States and DLR Equations
Finally, let us define what one means by Gibbs states in the general framework. Recall
the discussion in Section 2 of limits of finite volume correlations. That can be repeated
in the general framework. Thus if we knew that for all finite X ⊂ Zd and all f ∈ C(MX )

`(f ) = lim hf iσ̄Λn

n→∞

exists where Λn is some sequence of boxes converging to Zd . Then we conclude as before

R d
that `(f ) = f dµ for some Borel measure on MZ . We now want to characterize these
limit measures as Gibbs measures.
Consider first the finite volume Gibbs measure dµσ̄Λn (σ) defined in (4.24) for a potential
Φ ∈ B0 :
1 −βHΛσ̄ (σ)
dµσ̄Λn (σ) = e n dνΛn (σ) (5.1)
ZΛσ̄n
Let Λ ⊂ Λn and f ∈ C(MΛ ). We use in this section the notation σX to denote the
configuration in X i.e. σX := {σx |x ∈ X}. Also, for brevity let σ̄ denote σ̄Λcn i.e we have
σ̄ configuration outside Λn . Then we have the

Lemma. Eσ̄Λn (f ) = Eσ̄Λn (EσΛΛc ∨σ̄ (f )).

Proof. We have
Z Z
1 σ̄
Eσ̄Λn (f ) = f dµσ̄Λn = σ̄ f (σΛ )e−βHΛn (σ) dνΛn (σ) (5.2)
ZΛn

Λn
Λ

Write
X X
HΛσ̄ n (σ) = ΦX (σ ∨ σ̄) = HΛσ∨σ̄ (σ) + ΦX (σ ∨ σ̄) ≡ HΛσ∨σ̄ + H̃
X∩Λn 6=∅ X ⊂ Λc
X ∩ Λn 6= ∅

where H̃ is independent of σΛ . Thus

Z Z
1 σ∨σ̄
σ̄
EΛn (f ) = dνΛ (σ)f (σΛ ) dνΛn \Λ (σ) σ̄ e−β(HΛ +H̃)
ZΛn
Z Z
1
= dνΛn \Λ (σ)( dµΛσ∨σ̄ (σ)f (σΛ )) σ̄ ZΛσ∨σ̄ e−β H̃ . (5.3)
ZΛn

37
But Z Z
σ∨σ̄ −β H̃ σ̄
ZΛσ∨σ̄ e−β H̃ = dνΛ (σ)e −βHΛ
e = dνΛ (σ)e−βHΛn (σ)

so (5.3) becomes Z Z
0 0
Eσ̄Λn (f ) = dµσ̄Λn (σ) dµσ∨σ̄
Λ (σ )f (σΛ )

as claimed. 
Another way to say this is the following. The measure µσ̄Λn projects to a measure
(µσ̄Λn )Λ on MΛ by the formula
Z Z
f dµΛn = f d(µσ̄Λn )Λ
σ̄

for all f ∈ C(MΛ ). Then the Lemma says

Z
(µΛn )Λ = dµσ̄Λn (σ)µΛσ∨σ̄
σ̄

i.e. the projection of a Gibbs state in Λn to a sub volume Λ is a convex combination

of Gibbs states in Λ with different boundary conditions. This simple fact motivates the
following definition:
d
Definition. A Borel measure µ in MZ is a Gibbs measure with potential Φ ∈ B0 if
c
for all finite Λ ⊂ Zd there exists a probability measure µ̃Λc on MΛ such that for all
f ∈ C(MΛ )
Z Z Z
f (σ)dµ(σ) = ( f (σΛ )dµσΛΛc (σ)) dµ̃Λc (σ) (DLR) (5.4)

We denote the set of Gibbs measures of the potential Φ by GΦ .

Remark. The first integral is the integral of f w.r.t. the Gibbs measure in finite volume,
with b.c. σΛc , and (5.4) is some average over boundary conditions. (DLR) is the DLR-
equation (Dobrushin [10, 11], Lanford, Ruelle [29]).
What is the connection of this definition to the thermodynamic limits of finite volume
Gibbs states? Let µσ̄Λ be a Gibbs measure on MΛ . We can extend µσ̄Λ to a measure µσ̄,Λ
d
in the infinite volume configuration space MZ by setting

µσ̄,Λ = µσ̄Λ × δΛσ̄ c

c
where δΛσ̄ c is the Dirac measure on MΛ i.e.
Z
f (τ )dδΛσ̄ c (τ ) = f (σ̄Λc )

38
 c
for all continuous f on MΛ (here σ̄Λc means just the restriction of the configuration σ̄
to Λc , where f is defined). Thus, if g ∈ C(MΛ ), then
Z Z
gdµ = gdµσ̄Λ ≡ hgiσ̄Λ
σ̄,Λ

our original state on C(MΛ ).

Consider then a sequence of boxes Λn tending to Zd and consider the measures µσ̄,Λn .
We need the following fact from measure theory:
Let Ω be a compact metric space and let B(Ω) = { Borel probability measures on Ω}.
R
We say µn ∈ B(Ω) converge if f dµn converge for all f ∈ C(Ω).Then B(Ω) is compact
in this (weak∗ -) topology i.e. if µn ∈ B(Ω), n ∈ N then there is a convergent subsequence
µni and limit µ ∈ B(Ω): µni → µ.
Hence, there is a sequence of boxes Λni % Zd such that µσ̄,Λni converges to some
measure µ.
Theorem 5.1 µ ∈ GΦ .
Proof. Let f ∈ C(Λ). By the Lemma
Z Z Z Z
σ̄ σ∨σ̄ σ̄
f dµΛn = ( f dµΛ )dµΛn = g(σΛc )dµσ̄,Λni (σ)

where g(σΛc ) = f dµσΛΛc is easily seen to be a continuous function (check!). Hence the
R

claim follows from the convergence of the sequence µσ̄,Λni . 

Remarks. In the general theory of Gibbs states one can now show the following.
Let Φ ∈ B0 and GΦ be the set of Gibbs states of Φ. Then :

a) GΦ is a convex set : µ1 , µ2 ∈ GΦ ⇒

sµ1 + (1 − s)µ2 ∈ GΦ , s ∈ [0, 1].

b) GΦ is also compact (in the weak∗ topology).

c) There is a set of extremal elements µα ∈ GΦ α ∈ A (index set), (µα is extremal, if

it cannot be written as a convex combination of two distinct µ’s) such that every
µ ∈ GΦ is uniquely a convex combination of µα (i.e., if |A| < ∞, A = {αi }N i=1 then
PN PN
∃ si ∈ [0, 1], i=1 si = 1 and µ = i=1 si µαi , in general, there exists a measure ν on
R
A, ν(A) = 1 and µ = µα dν(α)). The measures µα are called pure phases.

Example. Ising model : the translation invariant pure phases are µ+ and µ− ,
general translation invariant Gibbs state µ = sµ+ + (1 − s)µ− . For example µf ree =
µperiodic = 21 (µ+ + µ− ) (Why ?).

39
 d) Pure phases have correlations that tend to 0 at infinity:

hf (σA ); g(σB )iT := hf (σA )g(σB )i − hf (σA )ihg(σB )i → 0 as dist(A, B) → ∞.

Pure phases are physical, mixtures of them reflect our ignorance.

Let us finish this section by proving the Mermin-Wagner theorem for the XY -model.
This generalizes to general systems with compact Lie group symmetries. See Simon’s
book [42].

Theorem 5.2 Let µ be a Gibbs measure for the XY -model. Then µ is invariant under
a global rotation of the spins. I.e. for any f ∈ C(Ω) and R ∈ SO(2),
Z Z
f (σ)dµ(σ) = f (Rσ)dµ(σ)

where (Rσ)x ≡ Rσx .

Proof 1. The idea is as follows (we follow here [26], see [42], p. 296). Consider say f (σ)!=
cos ϕ sin ϕ
F (σ0 ), depends on σ0 only. Let R(ϕ) be rotation by angle ϕ, R(ϕ) = .
− sin ϕ cos ϕ
!
cos θx X
Go to angular variables σx = , so H = − cos(θx − θy ), and the a priori
sin θx hxyi
dθx
measure is on [0, 2π].
2π
We want to prove :
hF (θ0 + ϕ)i = hF (θ0 )i ∀φ,
D d  E
it suffices to prove  F (θ0 + ϕ) = 0 , ∀F smooth. Now h−i is approximatively

dϕ ϕ=0
θ̄
h−iΛ some large Λ, some b.c. θ̄ because Gibbs measures are limits of such measures. Thus
we could change variables
D Eθ̄ D Eθ̄+ϕ
F (θ0 + ϕ) = F (θ0 )
Λ Λ

which follows from HΛθ̄ (θ − ϕ) = HΛθ̄+ϕ (θ)

 X X 
θ̄
HΛ (θ) = − cos(θx − θy ) − cos(θx − θ̄y )
x,y∈Λ x∈Λ
y ∈ Λc

So the ϕ dependence is now in the b.c. This is tricky to control, so let us try to change
variables by slowly rotating the spins : rotate θ0 + ϕ → θ0 and the rest a bit less such

40
that on ∂Λ there is no rotation : Let g : Λ → R g(0) = 1, g(x) = 0 for x near ∂Λ. Then,
let (τϕ θ)x = θx + g(x)ϕ
D Eσ̄ D Eσ̄
F (θ0 + ϕ) = F ((τϕ θ)0 )
Λ Λ
Z
1 h i Y dθ
x
= σ̄ F (θ0 ) exp −βHΛσ̄ (τ−ϕ θ)
ZΛ 2π
↑
θx → θx − g(x)φ = (τ−ϕ θ)x
  σ̄
d  σ̄ d  σ̄

⇒ h  F (θ0 + φ)iΛ = −β  H (τ−ϕ θ) F (θ0 )
dϕ ϕ=0 dϕ ϕ=0 Λ Λ

Thus, by Schwartz inequality :

D d  E  
d  2 1/2
2 1/2 σ̄
 F  ≤ βhF i  H (τ−ϕ θ
  
dϕ ϕ=0 dϕ ϕ=0 Λ


writing
d σ̄ σ̄ d σ̄
β HΛ (τ−ϕ θ)e−βHΛ (τ−ϕ θ) = − e−βHΛ (τ−ϕ θ)
dϕ dϕ
and integrating by parts, we get :
D d   2 E D d2  E
σ̄ σ̄
β  H =  H (τ−ϕ θ) .
 
dϕ ϕ=0 Λ dϕ2 ϕ=0 Λ
But
d2 σ̄ d2 X   
H (τ −ϕ θ) = − cos θx − θy − ϕ g(x) − g(y)
dϕ2 Λ dϕ2
hxyix∈Λ,y∈Λ

where we sum only over x, y ∈ Λ since g(x) = 0 near ∂Λ.

So
 2   
 d    X 
σ̄ 
2 
 dϕ2 ϕ=0 HΛ  =  g(x) − g(y) cos(θx − θy )
 
hxyi in Λ
X  2
≤ g(x) − g(y) .
hxyi in Λ

Thus,
D  E
 d  X  2
 dϕ ϕ=0 F  ≤ C g(x) − g(y) (5.5)
 
hxyi
x, y ∈ Λ

where C is g-independent. (5.5) holds for all g : Λ → R, g(0) = 1, g = 0 near ∂Λ. Also
(5.5) holds for any Λ (C is Λ-independent).
To make the above argument rigorous, we use the DLR-equation to write
Z Z Z 
dF dF σ̄
dµ = dµ (σ) dµ̃Λc (σ̄)
dϕ dϕ Λ

41
i.e. an average over σ̄ of what we did. Then (5.5) follows as above.

2. Second step. The point now is that

X 2
inf inf g(x) − g(y) = 0.
Λ g
hxyi

Reason : we show that this equals

Z
inf (∇h(u))2 d2 u (5.6)
h∈C0∞ ,h(0)=1

and this equals zero. Thus, let h ∈ C0∞ , h(0) = 1 (smooth function h : R2 → R with
compact support). Let the support of h be contained in ΛR (= cube, center 0, side R).
Put gL : Z2 → R , gL (x) = h( R
L
x) (x ∈ Z2 ). Then gL (0) = 1, gL = 0 near ∂ΛL and
2    2
X
2
XX R  R
(gL (x) − gL (y)) = h x −h (x − ei )
2 i=1
L L
x∈Z

(where e1 = (1, 0), e2 = (0, 1))

" #2
X X R 2 (h( R x) − h( R (x − ei ))
L L
= R
x i
L L

d u. The infimum of this over h ∈ C0∞ , h(0) = 1 equals zero : let

R 2 2
−→
L→∞ (∇h(u))
(
1 |x| < 1
h (x) = −
. This is piecewise C 1 and not 0 on boundary, but one can
|x| |x| ≥ 1
approximate this function by a function in C0∞ (show !). Now,

~x
∇h (x) = − 2+ for |x| ≥ 1 and = 0, |x| < 1, ⇒
|x|
Z Z ∞
2 2
(∇h(u)) d u = 2π2
r−2−2 rdr = π → 0,
1

as  → 0 

6 The Ginzburg-Landau Model

We now consider models where σx ∈ R i.e. is unbounded.
Motivation :
1. Later (see Section 12) we see that our previous models naturally give rise to this by a
process called coarse-graining.
2. These are related to quantum fields.

42
 We will denote the spin by φ(x) = φx i.e. the spin configuration is φ : Zd → R. Let
us consider a simple Hamiltonian
X X X
HΛ (φ) = − φ(x)φ(y) + a φ2 (x) + λ φ4 (x) (6.1)
{x,y}∈BΛ x∈Λ x∈Λ

which again has the φ → −φ symmetry, and consider the probability measure
1 −βHΛ (φ) Y
e dφ(x) (6.2)
ZΛ x∈Λ

on R|Λ| . The normalization (i.e. partition function) is given by ZΛ = e−βHΛ (φ)

R Q
RΛ x∈Λ dφx .
This integral exists if λ > 0. This is easy to see as follows:
1 X X X 1 X
HΛ (φ) = (φx − φy )2 + (−d + a) φ2x + λ φ4x + nx φ2x
2 x∈Λ x∈Λ
2 x∈∂Λ
{x,y}∈BΛ
X X
≥ (a − d) φ2x + λ φ4x
x∈Λ x∈Λ

where, for x ∈ ∂Λ nx is the number of nearest neighbors y of x s.t. y ∈

/ Λ. Thus
Z ∞ |Λ|
−β[(a−d)φ2 +λφ4 ]
ZΛ ≤ e dφ < ∞ if λ > 0 or λ = 0 and a − d > 0.
−∞

r
Call a − d = 2
and so

1 X rX 2 X 1 X
HΛ (φ) = (φx − φy )2 + φx + λ φ4x + nx φ2x . (6.3)
2 2 2 x∈∂Λ
{x,y}∈BΛ

It will be much more convenient to work with the periodic boundary conditions instead
of the free ones above. In that case the last term is missing from (6.3) and BΛ contains
also the bonds joining opposite faces of the cube ΛL .
Remark. The Ising model is a limit of this model : take
r r 2
= −2λ so φ + λφ4x = λ(φ2x − 1)2 − λ
2 2 x
q
λ −λ(φ2 −1)2
and π
e −→
λ→∞ δ(φ2 − 1) (in the sense of distributions).

7 Gaussian Integrals
7.1 Definitions and elementary properties
Consider first the case λ = 0. Then our measure is Gaussian.

43
Definition 7.1 Let A be a real symmetric n×n matrix which is (strictly) positive definite
(i.e. all eigenvalues > 0 i.e. (φ, Aφ) > 0 ∀φ ∈ Rn , φ 6= 0). The Gaussian measure on Rn
with covariance A−1 and mean 0 is the probability measure
1 − 1 (φ,Aφ)
e 2 dµ(φ) =
Dφ
Z
Q X
where φ = (φ1 , · · · , φn ), Dφ = i dφi , (φ, Aφ) = φi Aij φj .
ij
n
Thus, each φi : R → R given by φi (φ) = φi is a random variable, with mean 0 and
variance hφ2i i = (A−1 )ii . Also, (φ, f ) is a random variable, with variance (f, A−1 f ).
Some calculations
1. Consider the partition function
Z
1
Z= e− 2 (φ,Aφ) Dφ.
Rn

A can be diagonalized by an orthogonal matrix S ∈ SO(n): S T AS = diag(λ1 , . . . , λn ).

So | det S| = 1 and changing variables φ = SΨ we have Dφ = | det S|DΨ = DΨ and thus
Z n Z ∞ n  1/2 i−1/2
− 21
P
λi Ψ2i
Y
− 12 λi Ψ2i
Y 2π h
Z = e i DΨ = e dΨi = = det(A/2π)
i=1 −∞ i=1
λi

2. The correlations functions h kα=1 φiα i where iα ∈ {i, . . . , n} (note: several iα may be
Q

the same) can be computed using the generating function : Let f ∈ Rn . Define
Z
1 1
S(f ) = he(φ,f )
i= e− 2 (φ,Aφ)+(φ,f ) Dφ,
Z

with (φ, f ) = ni=1 φi fi . Then S(f ) is smooth in f and

P

k k
Y Y ∂ 
h φiα i =  S(f ).
α=1 α=1
∂fiα f =0

But
Z
1 1 −1 −1 1 −1
S(f ) = e− 2 (φ−A f,A(φ−A f ))+ 2 (f,A f ) Dφ
Z
Z
1
(f,A−1 f ) 1 1 1 −1
= e 2 e− 2 (Ψ,AΨ) DΨ = e 2 (f,A f )
Z

using the change of variables Ψ = φ − A−1 f . Hence, say the 2-point function,

∂ 2  1 (f,A−1 f )
hφi φj i =  e2 = (A−1 )ij .
∂fi ∂fj 0

44
In general,

k
Y  0 k odd   

m
h φiα i = Y ∂  1 1 X Y
−1
(f, A f ) = (A−1 )iα iβ k = 2m
∂fiα 0 2 m!
 
α=1 
P {α,β}∈P

where P is a pairing of the set {1, 2, · · · , 2m} i.e. a partition into m sets of size 2. The
sum runs through all such pairings.
Example For n = 4 there are three pairings {{1, 2}, {3, 4}}, {{1, 3}, {2, 4}} and {{1, 4},
{2, 3}}. In general, the number of pairings is 2(2m)!
m m! (show !).

Thus
2m
Y X Y
h φ iα i = hφiα φiβ i
α=1 P {α,β}∈P

a sum of products of 2-point functions.

7.2 The Gaussian Ginzburg-Landau model

Let us now specialize to the λ = 0 Ginzburg-Landau model with periodic boundary
conditions on ΛL = ZdL . This gives rise to a Gaussian measure on R|ΛL | with density
− 12 ( {x,y}∈B (φx −φy )2 +r φ2x )
P P
Z −1 e ΛL
(7.1)

− φy )2 defines a matrix ∆per (lattice-Laplacean with

P
The quadratic form {x,y}∈BΛL (φx
periodic b.c.):
X
(φx − φy )2 = −(φ, ∆per φ).
{x,y}∈BΛ

Concretely
X
−(∆per φ)x = (φx − φx+u ) (7.2)
|u|=1

where the addition is modulo L. For example in d = 1 we have (−∆φ)x = 2φx − φx−1 −
φx+1 .
d
(7.2) defines as well an operator ∆ in infinite volume i.e. for φ ∈ RZ . Then ∆ is self
d
adjoint on `2 (Zd ) ⊂ RZ
X
`2 (Zd ) = {φ : Zd → C | |φx |2 < ∞}.
x∈Zd
P
Indeed, ∆ is bounded: k∆k ≤ 2d (show!) and in the scalar product (φ, ψ) = x∈Zd φ̄x ψx
we have (φ, ∆ψ) = (∆φ, ψ).

45
 d
The periodic Laplacean can be defined in RZ as well. For this consider the set
d
FLper ⊂ RZ of periodic functions : [we use φx or φ(x) below !]

φ ∈ FLper : φ(x + Ln) = φ(x) ∀n ∈ Zd .

Clearly FLper can be identified with RΛL , namely φ is determined by φ|ΛL , ΛL = {x|xi ∈
{0, . . . , L − 1}}. Clearly ∆ : FLper → FLper so the periodic Laplacean ∆per is the corre-
sponding matrix in RΛL .
d
We will now diagonalize ∆per and ∆ by Fourier series. Consider p ∈ [−π, π]d and φp ∈ RZ
φp (x) = eipx (where px := α pα xα ) . Then,
P

X X
∆φp (x) = (eip(x+u) − eipx ) = (eipu − 1)eipx = −µ(p)φp (x)
|u|=1 |u|=1

where
d
X
µ(p) = 2 (1 − cos pµ ).
µ=1

Let, for L odd,

2π
BL := {p ∈ Rd | pi = ni , ni ∈ Z, |ni | < L/2}. (7.3)
L
Then
{φp | p ∈ BL }
is a basis for FLper (since they are independent and there are Ld of them). We have
X
eip(x−y) = Ld δxy for x, y ∈ ΛL (7.4)
p∈BL

and
X
ei(p−q)x = Ld δpq for p, q ∈ BL (7.5)
x∈ΛL

so {L−d/2 φp }p∈BL is an orthonormal basis of eigenvectors of −∆per and from (7.4), we

conclude
X 1
(−∆per + r)−1
xy = L
−d
eip(x−y) ≡ GL (x − y) (7.6)
p∈B
µ(p) + r
L

Thus, the correlations of the λ = 0 GL-model are given in terms of the (periodic b.c.)
Green’s function GL .
How about L → ∞ ? In (7.6) we have a Riemann sum over cells of size ( 2π
L
)d . So, in
the limit,
dd p eip(x−y)
Z
lim GL (x − y) = d
≡ G(x − y). (7.7)
L→∞ [−π,π]d (2π) µ(p) + r

46
This equals of course the kernel of the operator (−∆ + r)−1 defined on `2 (Zd ). Indeed, on
`2 (Zd ) the operator −∆ has spectrum {µ(p)|p ∈ [−π, π]d } and we can diagonalize it by
Fourier series. If f ∈ `2 (Zd ), let
X
fˆ(p) = e−ipx f (x).
x∈Zd

Then fˆ ∈ L2 (B), B = [−π, π]d and

dd p ˆ
X Z
(f, g) ≡ f (x)g(x) = f (p)ĝ(p)
B (2π)d
x∈Zd

and we have the inverse formula

dd p
Z
f (x) = eipx fˆ(p) .
B (2π)d
Thus,
dd p
Z
−1 ipx 1 ˆ
X
((−∆ + r) f )(x) = e f (p) = G(x − y)f (y)
µ(p) + r (2π)d y

i.e. (−∆ + r)−1

xy = G(x − y). We have obtained

Proposition 7.5 Let µL be the Gaussian measure on RΛL , with covariance (−∆per +r)−1 .
Then
Z Y XY
−→
φ(xα )dµL (φ) L→∞ G(xα − xβ ). (7.8)
P hαβi

8 Measures on spaces of distributions

d
Can we write the RHS of (7.8) in terms of a measure on RZ ? Note that this space is not
locally compact not to mention compact as we had in the case of bounded spin systems.
Hence we can not take the route via Riesz representation theorem to reconstruct a measure
out of correlation functions. The way out is to consider instead the characteristic function
of the measure.

8.1 Bochner’s Theorem

Let µ be a probability (Borel) measure on Rn . Recall that we defined the generating
R
function of µ as S(f ) = e(φ,f ) dµ(φ) if e(φ,f ) ∈ L1 (µ). It is actually more convenient to
consider S(if ), f ∈ Rn i.e. the characteristic function i.e. the Fourier transform of µ:
Z
W (f ) = ei(φ,f ) dµ(φ) f ∈ Rn .

47
which always exists by dominated convergence theorem.
W (f ) has some obvious and less obvious properties:
R
a) |W (f )| ≤ 1dµ = 1.

b) W (0) = 1.

c) W is continuous. Indeed, by the dominated convergence theorem, if fn → f then,

since |ei(φ,fn ) | ≤ 1, W (fn ) → W (f ).

d) Let zα ∈ C, fα ∈ Rn , α = 1, · · · , N . Then,
X Z X Z X
i(φ,fα ) −i(φ,fβ )
zα z̄β W (fα − fβ ) = zα z̄β e e = | zα ei(φ,fα ) |2 dµ ≥ 0.
α,β

Definition 7.3 W : Rn → C is a function of of positive type if a) - d) hold.

Theorem 7.4. (Bochner’s theorem) W (f ) is of positive type if and only if there is a
R
Borel probability measure on Rn such that W (f ) = ei(φ,f ) dµ(φ).
Proof. “⇐” is done above.
“⇒”. One uses an idea that is useful in other contexts, namely we use W to construct a
scalar product. Let

H = {ψ : Rn → C|ψ(x) 6= 0 only for finitely many x}.

For φ, ψ ∈ H put X
(φ, ψ) = φ̄(x)ψ(y)W (x − y)
x,y∈Rn

(this is a finite sum). ( , )is an inner product except that (φ, φ) = 0 does not imply φ = 0.
Let H0 = {φ|(φ, φ) = 0}; H0 is a subspace of H. Put H e = H/H0 = equivalence classes
[φ] = [φ + φ0 ], φ0 ∈ H0 . Then (H, e (·, ·)) is an inner product space which can be completed
into a Hilbert space.
Let for t ∈ Rn , Ut : H → H (Ut ψ)(x) = ψ(x + t). Clearly, Ut : H0 → H0 so one
defines Uet : H e→H e by Uet [φ] = [Ut φ]. (Uet φ, U
et ψ) = (φ, ψ) so U
et is an isometry. We have
U
et+s = U et U e0 = 1|. Also t → U
es , U et is strongly continuous (i.e. ∀ψ ∈ H e t→U et ψ ∈ H
e is
continuous from Rn → H) e because W is continuous:

kU es ψk2 = 2(ψ, ψ) − (U
et ψ − U es ψ) − (U
et ψ, U et ψ) = [2(ψ, ψ) − (ψ, U
es ψ, U et−s ψ) − (ψ, U
es−t ψ)]

This equals
X
ψ̄(x)ψ(y)[2W (x − y) − W (x − y + t − s) − W (x − y + s − t)]
x,y

48
which tends to 0 as t → s. We have now verified the assumptions of

Stone’s Theorem. Let Rn 3 t → Ut unitary in Hilbert space be strongly continu-

ous and Ut Us = Ut+s , ∀s, t, U0 = 1|. Then for each φ ∈ H, kφk = 1 ∃ a probability Borel
R
measure µφ on Rn such that (φ, Ut φ) = ei(λ,t) dµφ (λ).

Proof See Reed-Simon [35], vol. 1, Th. VIII.12. [The heuristics behind the proof
is that, if H were finite dimensional, we could easily prove that Ut is differentiable in t,
∂Ut P
|t=0 = iAj , Aj Hermitean, [Ai , Aj ] = 0 and Ut = ei tj Aj . Diagonalize all Aj simulta-
∂tj
 (j) 
λ1 0 n
.
X
neously ⇒ Aj = 
 . . ~
 ⇒ dµφ (λ) =

δ(~λ(i) − ~λ)|φi |2 ].
(j) i=1
0 λn
To finish the proof of the Theorem, using Stone’s Theorem, take φ̃ ∈ H̃
(
1 x=0
φ̃ = [φ], φ(x) =
0 x 6= 0
X Z
W (t) = δ(x)δ(y + t)W (x − y) = (φ̃, Ũt φ̃) = ei(λ,t) dµφ̃ (λ) 
x,y∈Rn

Suppose now we have measures µk on Rn and want to prove µk converge as k → ∞.

The properties 1,2 and 4 of Wk usually carry to the limit so the main issue to check is 2
i.e. continuity. In our case WL however live in spaces whose dimensions tend to infinity as
L → infinity. We aim at conditions 1-4 in such a setup and a generalization of Bochner’s
theorem there.

8.2 Cylinder measures

Suppose we had a measure µ on some σ-algebra A of subsets of RN such that sets of the
form A × RN\n ∈ A where A ∈ An = some σ-algebra of subsets of Rn and RN\n denotes
×∞ ∞
i=n+1 R (and R ≡ ×i=1 R). Then we get a measure µn on An by “integrating out” the
N

other variables :

µn (A) = µ(A × RN\n ). (8.9)

This resulting set of measures is consistent i.e.

µn (A) = µn+m (A × Rm ) (8.10)

provided A × Rm ∈ An+m if A ∈ An . Thus, conversely, let {µn }∞ n=1 be a family of

n
measures, µn a Borel probability measure on R . We say µn are consistent if (8.10) holds

49
for all Borel sets A ⊂ Rn . A natural σ-algebra on RN is the σ-algebra generated by the
cylinder sets. Denote by B Borel sigma-algebra on R and by B n the one on Rn .

Definition 8.1 A cylinder set on RN is a set of the form SB1 ,··· ,Bn = {x ∈ RN |xi ∈ Bi , i =
1, · · · , n, Bi ∈ B, n < ∞}. The σ-algebra of subsets of RN generated by cylinder sets is
denoded by B N . A measure µ defined on B N is called a cylinder measure. We have then
the
Kolmogorov’s extension theorem. Let {µn }∞ n=1 be a consistent family of measures on
n
B . Then there is a unique cylinder measure µ on B N such that

µ(SB1 ,··· ,Bn ) = µn (B1 × B2 × · · · × Bn ). (8.11)

Proof. Main problem is countable additivity. We can do this with Riesz theorem by the
following trick.
Let Ṙ be the one-point compactification of R (i.e. Ṙ = R ∪ {∞} and open sets are
open sets in R and sets of the form A ∪ {∞} where A ⊂ R is open and Ac ⊂ [−n, n] some
n < ∞, Ṙ is easily seen to be compact). Let

M = ×∞ N
i=1 Ṙ = Ṙ .

Then M is compact in the product topology. Let C(M) = {f : M → R|f continuous }

and C0 (M) = {f : M → R|, f is continuous and depends on finitely many xi }. Define,
for f ∈ C0 (M), Z
`(f ) = f dµn

where n is large enough, and we extend µn to Ṙn by putting µn (B1 × · · · ×

{∞} × · · · × Bn ) = 0. By consistency, ` is n independent if n is large enough. So ` defines
a positive linear functional ` : C0 (M) → R, `(1) = 1, |`(f )| ≤ kf k∞ . The last one implies
that ` extends to C(M) since C0 (M) is dense in C(M). Hence by Riesz theorem there
R
exists unique Borel-measure µ on M such that `(f ) = f dµ. Recall that M = ṘN . But
∞
X ∞
X
µ({x ∈ X|xn = ∞ some n) ≤ µ({x|xn = ∞}) = µn ({x|xn = ∞}) = 0.
n=1 n=1

Hence µ is supported on RN . By construction (8.11) holds (B N ⊂ Borel sets of ṘN ) 2 ). 

As it turns out, we can embed all interesting function (and distribution) spaces into
RN . Let us start with two subsets :
2
Borel sets of ṘN = smallest σ-algebra containing open sets and cylinder sets are generated by open
sets.

50
8.3 Minlos Theorem : Measures on s0
Let ∞
X
sm = {x ∈ R | N
n2m |xn |2 ≡ kxk2m < ∞}
n=1

for m ∈ Z. Let \ [
s= sm s0 = sm
m∈Z m∈Z

(note : sm ⊃ sm+1 ⊃ · · · ). s is the set of sequences which decay as n → ∞ faster than

any power, s0 the ones with at most polynomial growth. Give s the topology generated
by the neighbourhoods of 0 of the form

N (m, ) = {x| kxkm < }

(so neighbourhoods of y ∈ s are y + N (m, )).

Lemma 8.2 The space s with the above topology is a complete metrizable space, i.e. s
has a complete metric d such that d gives the same topology as above.

Proof. Define the metric

∞
X kx − ykm
d(x, y) = 2−m .
m=1
1 + kx − ykm

Then (Exercise) : Show that this distance defines the same topology as above and that
s is complete (Cauchy sequences converge). 
Remark. s is an example of a Frechet space i.e. a complete locally convex metric space,
and which is not a Banach space.
Lemma 8.3 s0 is the dual space of s i.e. the space of continuous linear functionals on s.
Proof Let y ∈ s0 . y defines a continuous linear functional `y on s :
∞
X
`y (x) ≡ (y, x) ≡ yn xn .
n=1

Indeed, this converges since |yn | ≤ Cnm some m < ∞ and |xn | ≤ Cm nm for all m. To
prove continuity, let d(x(k) , x) → 0 as k → ∞. Then, by Schwartz’ inequality,
∞
(k)
X X 1
|`y (x ) − `y (x)| ≤ Cnm |xn(k) − xn | = C nm+1 |x(k)
n − xn |
n=1
n
X 1 1/2
≤ C kx(k) − xkm+1
n2
≤ C 0 2m+1 d(x(k) , x) → 0 as k → ∞.

51
Conversely, let ` : s → R be a continuous linear map. Then there exists C < ∞, m < ∞
such that
|`(z)| ≤ Ckzkm (8.12)

for all z. Indeed, by continuity there exists a neighbourhood U of zero in s such that
|`(x)| < 1, for x ∈ U . On the other hand U contains a set {z| kzkm < δ} some m, δ.
Hence, given a z ∈ s we have δ kzkz m ∈ U and so |`(δ kzkz m )| < 1 i.e. |`(z)| ≤ δ −1 kzkm ).
(k)
Let now e(k) ∈ s be given by en = δnk . Put yn = `(e(n) ). By (8.12)

|yn | = |`(e(n) )| ≤ Cke(n) km = Cnm

so y ∈ s0 . Given x ∈ s, let
n
X
x(n) = e(k) xk .
k=1

Then x(n) → x in s (prove !) and so

`(x(n) ) → `(x).
n
X
But `(x(n) ) = yk xk . Hence `(x) = (y, x) 
k=1

The reason we introduced s and s0 is the following. Let µ be a cylinder (probability)

measure on B N . Note that s0 is a µ-measurable set (prove !). Suppose µ(s0 ) = 1 i.e.
µ(B N \s0 ) = 0 i.e. µ is supported in s0 . Then,
Exercise. Let x ∈ s. Then the function y → (y, x) from s0 to R is µ-measurable.
Definition 8.4 Let µ be a cylinder probability measure on s0 . The characteristic function
W : s → C of µ is Z
W (x) = ei(y,x) dµ(y).
s0

Lemma 8.5 W satisfies

a) W (0) = 1, |W (x)| ≤ 1
b) W is positive definite : given zi ∈ C, x(i) ∈ s i = 1, · · · , n we have
n
X
zi zj W (x(i) − x(j) ) ≥ 0
i,j=1
c) W : s → R is continuous.

Proof As before, for c) use again the dominated convergence theorem. 

Finally, our goal:
Minlos’ Theorem A necessary and sufficient condition for a function W on s to be a
characteristic function of a probability measure on s0 is that it obeys a)-c) above.

52
Proof. Let W be given. Let u ∈ Rn , put x = (u1 , · · · , un , 0, 0, · · · ) ∈ s. Then W (x) ≡
fn (u) is a function of positive type in Rn ⇒ ∃ probability measure µn on Rn such that
W
Z
Wn (u) =
f ei(y,u) dµn (u)
Rn

{µn }∞
n=1 are consistent since Wn+m ((u1 · · · un , 0, · · · , 0)) = Wn (u) and Wn uniquely deter-
f f f
mines µn (Why ?). Hence ∃µ on RN such that
Z
W (x) = ei(y,x) dµ(y) (8.13)
RN

for all x ∈ s with only finitely many non zero xi . We need to show : µ(s0 ) = 1. In that
case the integral can be restricted to s0 and (8.13) holds for all x ∈ s (because both sides
are continuous in x and the set of x with a finite number of nonzero xi is dense in s).
[∞
µ(s0 ) = 1 : Recall that s0 = sm with
m=−∞

∞
X
sm = {y| n2m yn2 ≡ kyk2m < ∞}.
n=1

We show: given  > 0, there exists an m such that

µ(sm ) ≥ 1 − . (8.14)

Since µ(s0 ) ≥ µ(sm ) we get µ(s0 ) ≥ 1 −  ∀ i.e. the claim.

To prove (8.14), we use first the monotone convergence theorem to get a more mana-
geable expression :
Z N
! Z
α X 2m 2 α 2
lim lim exp − n yn dµ = lim e− 2 kykm dµ = µ(sm )
α↓0 N →∞ 2 n=1 α↓0

(
−α kyk2m 1 y ∈ sm
since e 2 % χ(y ∈ sm ) ≡
0 y∈6 sm , i.e. kykm = ∞
Z N
!
α X 2m 2
We consider exp − n yn dµ. Choose m:
2 n=1

Continuity of W (x) ⇒ ∃ m, δ such that

|W (x) − 1| ≤  if kxk2−m−1 ≤ δ (8.15)

(recall that W (0) = 1). (Note that m here tends to be < 0 !).
Then, for all x ∈ s,
2
Re W (x) ≥ 1 −  − kxk2−m−1 (8.16)
δ
53
since if kxk2−m−1 < δ, (8.16) holds by (8.15) and if kxk2−m−1 > δ, (8.16) holds because
|W (x)| ≤ W (0) = 1 implies Re W (x) ≥ −1.
Now write ! Z
N
α X 2m 2
exp − n yn = ei(y,x) dν(x)
2 n=1 R N

N
Y
2m −1/2
 x2n 
dν(x) = (2παn ) exp − dxn
n=1
2αn2m
to get
Z N
! Z Z N
! !
α X 2m 2 X
exp − n yn dµ = exp i yi xi dµ(y) dν(x)
2 n=1 RN i=1
Z Z
= W (x)dν(x) = Re W (x)dν(x)
RN

since the left hand side is real.

Hence by (8.16),
Z N
! Z
α X 2m 2 2
exp − n yn dµ ≥ 1 −  − kxk2−m−1 dν(x).
2 δ RN

But
∞
Z N Z N
!
X X X 1
kxk2−m−1 dν = n−2m−2 x2n dν(x) = n−2m−2 αn2m ≤ α.
RN n−1 n=1 n=1
n2

Thus !
Z N
α X 2m 2 cα
lim exp − n yn dµ ≥ 1 −  −
N →∞ 2 n=1 δ
and our claim follows 
Example. Gaussian measures. Let C be a matrix Cij i, j = 1, 2, · · · with |Cij | ≤ aim · j m
X
some a < ∞, m < ∞ and (x, Cx) = xi xj Cij > 0, ∀x 6= 0, for x ∈ s (note (x, Cx)
−n
makes sense since |xi | < bn i all n). Then the function
1
WC (x) = e− 2 (x,Cx)
is of positive type.

Proof W is continuous since |(x, Cx)| ≤ Akxk2n for some n. Hence WC (x) = lim WC (x(k) ),
( k→∞

(k) x i i ≤ k
xi = . But
0 i>k
k
! Z k
!
1 X X
WC (x(k) ) = exp − xi xj Cij = exp i xi yi dµk (y)
2 ij=1 Rk i=1

54
where µk is the Gaussian measure on Rk with covariance matrix Cij , i, j = 1, · · · , k. By
assumption, this is a positive k × k matrix. Hence, let WCk (x) = WC (x(k) ) x ∈ s so WCk is
of positive type and WCk (x) → WC (x) ∀x ∈ s → WC (x) is of positive type. 

We call the measure µC corresponding to WC the Gaussian measure in s0 with cova-

riance C.
Example. Order Zd arbitarily such that Zd = {xk |k = 1, 2, · · · } and |xk+1 | ≥ |xk |. Then
φ : Zd → R becomes φ̃ : N → R φ̃i = φxi . Thus, if Cxy x, y ∈ Zd is a positive matrix with
|Cxy | ≤ a(|x| + 1)m (|y|
X + 1)m we can define the Gaussian measure with covariance C on
s0 (Zd ) = {φ : Zd → R (|x| + 1)2m |φx |2 < ∞ for some m ∈ Z}. Thus C = (−∆ + r)−1 is

x∈Zd
an example (it is positive and |Cxy | ≤ constant; this holds : r ≥ 0, d ≥ 3 and r > 0, d ≥ 1).

8.4 Measures on Spaces of Distributions

Let
 dn f 
∞ m
S(R) = {f : R → R|f is C and sup(1 + |x| ) n < ∞ for all n, m ≥ 0} (8.17)
 
x∈R dx

i.e. f and its derivatives decay faster than any power. Let us define (“creation and
annihilation operators”)
1  d + 1  d
a= √ x+ , a = √ x−
2 dx 2 dx
2
and N = a+ a = − 21 dx
d 1 2 1
2 + 2 x − 2 . For f ∈ S(R), let

Z 1/2
m 2
kf km = |(N + 1) f | dx .
R

This is finite due to (8.17)3 .

We give S the topology where neighbourhoods of 0 are

N (m, ) = {f ∈ S| kf km < }

Proposition 8.6 S is homeomorphic to s.

Proof Consider the functions φn ∈ S :

1 − 21 x2 1
φ0 = e , φn = √ (a+ )n φ0
π 1/4 n!
3
See Remark 1 below.

55
These satisfy : Z ∞
(φn , φm ) ≡ φn φm dx = δnm
−∞

Proof aφ0 = 0, aa+ = a+ a + 1 imply

1 1 h i
aφn = √ a(a+ )n φ0 = √ (a+ )n−1 φ0 + a+ a(a+ )n−1 φ0
n! n!
1 h + n−1 i n n!
= √ 2(a ) φ2 + (a+ )2 aa+n−2 φ0 = . . . = √ (a+ )n−1 φ0 ⇒ an φn = √ φ0
n! n! n!
But also, for φ, ψ ∈ S, (a+ φ, ψ) = (φ, aψ) so (φn , φn ) = √1n! (φ0 , an φn ) = (φ0 , φ0 ) = 1.
n 6= m gives 0 since aφ0 = 0.
The functions φn are thus orthonormal in L2 (R). They are a basis (they are the Hermite
functions). Now our map S 7→ s is f 7→ x = (x1 , x2 , · · · ), with
xn = (f, φn−1 ). Note that
N φn = a+ aφn = nφn
so (N + 1)m φ 7→ x0 with x0n = nm xn .
Thus
Z X ∞  2 X∞
2
kf km = |(N + 1)m f |2 dx = (N + 1) m
f, φ = n2m |xn |2 = kxk2m
 
 k 
R k=0 n=1

i.e. the map f → x maps the neighbourhoods to each other. To conclude, we need to
show that this map is a bijection.

A) Injection : {φn } is a basis of L2 .

N
X
B) Surjection, or the map is onto : Let x ∈ s. Then we need to show that xn φn−1 con-
n=1
verges as N → ∞ to an element f of S with xn = (f, φn−1 ). The sequence N
P
n=1 xn φn−1
N
X N
X
converges in L2 to f ∈ L2 . (N + 1)m xn φn−1 = xn nm φn−1 converges in L2 ⇒ f ∈
n=1 n=1
C ∞ . And kf km = kxkm so f ∈ S. 

Definition 8.7 The space of tempered distributions on R is the set S 0 (R) of continuous
linear functionals on S(R).

Let ϕ ∈ S 0 (R). Put yn = ϕ(φn−1 ). Then ϕ continuous means4 : ∃C, m : |ϕ(f )| ≤ Ckf km
4
ϕ continuous : ∀ ∃ N(m, δ) : f ∈ N (m, δ) ⇒ |ϕ(f )| <  i.e. kf km < δ ⇒ |ϕ(f )| < . Thus for any
f ∈ S, |ϕ(f )| = |ϕ kffkm δ | kfδkm ≤ δ kf km = Ckf km .

56
∀f ∈ S ⇒ |yn | = |ϕ(φn−1 )| ≤ Ckφn−1 km = Cnm . Hence y ∈ s0 . Conversely, y ∈ s0 gives
∞
X
0
rise to ϕy ∈ S (R) by ϕy (f ) = yn (f, φn−1 ). Indeed, from above, xn = (f, φn ) is in s so
n=1
|ϕy (f )| ≤ Ckxkm for some m and this = Ckf km .

2
Remark 1 We see from above that  k · km is a norm on S : kfkm = 0 ⇒ (f, φn ) = 0∀n P
⇒
2 + m + m + 2m
f = 0. Moreover, since kf km = (a a + 1) f, (a a + 1) f = f, (a a + 1) f =
 
d
of f, (a+ a)k f and (a+ a)k f = (Polynomial in x and dx ) f , we get
X dβ
kf km ≤ Cm sup |xα f (x)| ≡ Cm kf k(m)
α,β≤2m
x dxβ

and conversely (show !). So, ϕ ∈ S 0 ⇐⇒ ∃C, N : |ϕ(f )| ≤ Ckf k(N ) . This is in practice
useful, see Examples 1,2 below.

Example 1 Let ϕ be a polynomially bounded function : |ϕ(x)| ≤ C(1 + |x|)m . Then it

defines a distribution ϕ ∈ S 0 by
Z
ϕ(f ) = ϕ(x)f (x)dx

since
Z  Z
 
 ϕ(x)f (x)dx ≤ C (1 + |x|)m |f (x)|dx ≤
 
Z
1
C (1 + |x|)m+2 |f (x)| dx ≤ C 0 sup(1 + |x|)m+2 |f (x)| ≤ C 00 kf k(m+2)
[1 + |x|]2 x

so ϕ ∈ S 0 .

Example 2 ϕ can be a “delta function” or its derivatives : let

dk f
ϕ(f ) = (−1)k (x0 ).
dxk
Then,
|ϕ(f )| ≤ kDk f k∞
dk
so ϕ ∈ S 0 . We denote ϕ = dxk
δ(x = x0 ).

Remark This works in Rn too :

S(Rn ) = {f ∈ C ∞ (Rn )| kxα Dβ f k∞ ≤ Cαβ ∀α, β}

n n
α
Y
αi β
Y ∂ βi
where α = (α1 , · · · , αn ), β = (β1 , · · · , βn ), x = xi , D = βi
. Basis {ϕk }k∈Nn
i=1 i=1 ∂xi
Yn X Y 2m
Nn
ϕk = (n)
ϕki (xi ) and we define s = {x ∈ R |x = (xk )k∈N n ki |xk |2 < ∞
i=1 k
∀m}. Do the rest !

57
Definition 8.8 1. The cylinder set σ-algebra in S 0 (R) is the smallest σ-algebra containing
all sets
A(f, B) = {ϕ ∈ S 0 (R)|ϕ(f ) ∈ B}
for f ∈ S, B Borel in R. It is just the image of our old cylinder algebra on s0 to S 0 by the
map constructed above. Indeed, the cylinder σ-algebra C in s0 is the smallest σ-algebra
of subsets of s0 that contains the sets σn (B) = {y ∈ s0 |yn ∈ B} where B is Borel in R.
I.e. it is the smallest σ-algebra such that functions

pn : s0 → R , pn (y) = yn

are measurable (i.e. p−1

n (B) ∈ C ∀B Borel i.e. σn (B) ∈ C ∀B). Let C̃ be the smallest
σ-algebra in s such that the functions πx : s0 → R πx (y) = (y, x) are measurable for all
0
∞
X
(n)
x ∈ s. Since pn = πe(n) , em = δnm we see that C ⊂ C̃. But, also, πx = xn pn so πx is
n=1
C-measurable. Hence C̃ ⊂ C. Thus C = C̃. Clearly the σ-algebra defined above in S 0 (R)
is the image of C̃ under the map y ∈ s0 → ϕy ∈ S 0 . This proves the claim.

Definition 8.9 2. A cylinder measure on S 0 (R) is a measure on this σ-algebra.

Translating from s, s0 to S, S 0 we get :

Theorem (Minlos) A necessary and sufficient condition for a function W on S(Rn ) to

be the characteristic function of a cylinder probability measure µ on S 0 (Rn ),
Z
W (f ) = eiϕ(f ) dµ(ϕ)
S0

is that W (0) = 1, W be of positive type and continuous.

Example Gaussian measures. Here W (f ) = exp − 21 C(f, f ) where C(f, g) is a continu-

 

ous bilinear function on S(Rn ) × S(Rn ), with C(f, f ) > 0 for f 6= 0. Important examples
are given by integral kernels
Z Z
C(f, g) = d x dn yC(x, y)f (x)g(y)
n

R
here |C(x, y)|[(1 + |x|)(1 + |y|)]m dn xdn y < ∞ for some m suffices. We are interested in
translation-invariant C’s i.e. C(x, y) = G(x − y). These are most conveniently expressed
in terms of Fourier transform : If f ∈ S(Rn ) we put
Z
fˆ(k) = f (x)e−ikx dn x

and one gets fˆ ∈ S(Rn ). [This is because (xα Dβ f )∧ (k) = iα+β Dα k β fˆ(k), do the details !]
Inverse :
dn k
Z
ikx ˆ
f (x) = e f (k) .
(2π)n

58
Then we have
dn k
Z Z
2 n
|f (x)| d x = |fˆ(k)|2
(2π)n
Z
(f ∗ g)(x) ≡ f (x − y)g(y)dn y

(f ∗ g)∧ (k) = fˆ(k)ĝ(k)

and
dn k
Z Z
f (x)G(x − y)f (y)d xd y = n n
|fˆ(k)|2 G(k)
b
(2π)n
which holds for f, G ∈ S to start with, but extends to much more general G’s.

Example The free Euclidean field with mass m is the Gaussian measure µG where
1
G(k)
b = (k ∈ Rn ).
k2 + m2
dn k
|fˆ(k)|2
R
Clearly S 3 f 7→ is continuous, so
(k 2 + m2 )(2π)n
" Z ˆ 2 n #
1 |f (k)| d k
W (f ) = exp −
2 k 2 + m2 (2π)n

determines a measure on S 0 (Rn ), by Minlos’ theorem. Strictly speaking, G(x) is defined as

a distribution5 , by the Fourier transform G
b but actually is a locally integrable function :

e−m|x|
|G(x)| ≤ C n−1 . (8.18)
|x| 2

To see this, let us consider a ultraviolet cutoff : let Λ > 0 and define

1 Λ2
G
bΛ (k) = .
k 2 + m2 k 2 + Λ2
bΛ (k) −→ G(k)
This is called Pauli-Villars cutoff. Note that G b pointwise, and as elements
Λ→∞
of S 0 .

Note S 0 is given a topology such that ϕi → ϕ as i → ∞ if ϕi (f ) → ϕ(f ) ∀f ∈ S.

For n < 4, G bΛ ∈ L1 (Rn ) and so GΛ (x) is a continuous function (for n ≥ 4 : replace
2
Λ  Λ2 k
by ).
p2 + Λ 2 p2 + Λ 2
5
Fourier transform ϕ̂ of a distribution ϕ ∈ S 0 is a distribution, ϕ̂ ∈ S 0 , defined by

ϕ̂(fˆ) = (2π)n ϕ(f˜) f˜(x) = f (−x)

n
ϕ̂(k)fˆ(k) (2π)
d k
ϕ(x)f (−x)dn x.
R R
which is the usual (see example 1 above) if ϕ ∈ S : n =

59
So, by rotational symmetry,

dn k eik1 |x| Λ2
Z
GΛ (x) =
(2π)n k 2 + m2 k 2 + Λ2
So the k1 integral :
q q
k + m = (k1 + i ~k 2 + m2 )(k1 − i ~k 2 + m2 ),
2 2
q q
k + Λ = (k1 + i k + Λ )(k1 − i ~k 2 + Λ2 ),
2 2 ~ 2 2

~k = (k2 , k3 , · · · ).

By Cauchy:
" √ √ #
~2 ~2
dn−1~k e−|x| k +m
2 2
Λ2 e−|x| k +Λ
Z
1
GΛ (x) = p + p
(2π)n−1 2 ~k 2 + m2 m2 + Λ2 2 ~k 2 + Λ2 m2 + Λ2

which is smooth and converges as Λ → ∞ to

√
~2
dn−1~k e−|x| k +m
Z 2

G(x) = p ;
(2π)n−1 2 ~k 2 + m2
p
this is a L1 function satisfying (8.18). To see this, expand ~k 2 + m2 = m + O(~k 2 ) for
small ~k, and change variables |x|~k = ~k 0 .
p

Note that, a priori, for ϕ ∈ S 0 , ϕ(x) is not defined, only ϕ(f ), f ∈ S. These are mea-
surable functions in S 0 and actually Gaussian random variables on the probability space
(S 0 (Rn ), µ):
1 2
heit(ϕ,f ) i = e− 2 t (f,Gf )
so the variance of (ϕ, f ) is (f, Gf ). Let fi ∈ S, i = 1, · · · , N . Then xi ≡ ϕ(fi ) are jointly
Gaussian :
P P 1 P
hei ti xi i = heiϕ( ti fi ) i = e− 2 ij ti tj (fi ,Gfj )
i.e. x = (x1 , · · · , xN ) ∈ RN is Gaussian with covariance matrix A−1 ij = (fi , Gfj ). Hence
e.g.
Z Z
d  d  −i[tϕ(f )+sϕ(g)]
ϕ(f )ϕ(g)dµG (ϕ) = −   he i = (f, Gg) = dn xdn yf (x)G(x − y)g(y)
dt 0 ds 0
and, more generally,
2N
Z Y X Y
ϕ(fi )dµG (ϕ) = (fi , Gfj )
i=1 P hiji∈P
Z X Y
= dn x1 · · · dn x2N f1 (x1 ) · · · fN (xN ) G(xi − xj )
P hiji∈P

60
Hence we may use the notation
Z X Y
ϕ(x1 ) · · · ϕ(x2N )dµG (ϕ) = G(xi − xj )
P hiji∈P

even if strictly speaking ϕ(x) is not a well defined random variable.

Remark Indeed, let us try to see what ϕ(x)2 would be. Formally
Z Z
2
ϕ(x) dµG (ϕ) = lim ϕ(x)ϕ(y)dµG = lim G(x − y) = G(0) = ∞.
x→y x→y

This is a first instance of an ultraviolet divergence in quantum field theory. Our G is such
that ϕ(x) is not a nice random variable (it is a “distribution valued random variable”).
We’ll return to this.

9 The Gaussian Ginzburg-Landau Model : Critical Point

and Continuum Limit
Let us return to the lattice and the Ginzburg-Landau-model. The λ = 0 case is the
d
Gaussian measure µC on RZ (or s0 (Zd )) with covariance

eip(x−y) dd p
Z Z
C(x − y) = φ(x)φ(y)dµC (φ) =
[−π,π]d µ(p) + r (2π)d
d
where x, y ∈ Zd , φ ∈ RZ .
We have
0 ≤ C(x − y) ≤ Ce−|x−y|/ξ ξ > 0 , ξ → ∞ as r → 0.
Proof Let R = max |xi − yi |. We may assume R = |x1 − y1 | = x1 − y1 .
i=1,··· ,d

!
π
dd−1 p~
Z Z
1 dp 1
C(x − y) = eip1 R ei~p·(~x−~y) .
−π 2(1 − cos p1 ) + µ(~p) + r 2π (2π)d−1

Use Cauchy for the p1 integral

61
 B Im p1

A  C
Re p1
−π π

A and C cancel since cos p1 , eip1 R are periodic

!
π
dd−1 p~
Z Z
1 dp 1
= e−R eip1 R ei~p(~x−~y)
−π 2(1 − cos(p1 + i)) + µ(~p) + r 2π (2π)d−1

Choose  so that : Re2(1 − cos(p1 + i)) + r > 0 i.e. 2(1 − cos p1 cosh ) + r > 0 i.e.
√
2 cosh  < 2 + r or  < cosh−1 (1 + r/2). [So, for r small, one can choose  ∼ r ]. So,
since µ(p) ≥ 0,

C(x − y) ≤ Ce−R 

Thus r > 0 is non critical. For r = 0, we have

x
eip |x| dd p
Z
2−d
C(x) = |x| ≡ |x|2−d h(x̂, |x|)
[−π|x|,π|x|]d x2 µ(p/|x|) (2π)d

where x̂ = x/|x|, provided d > 2 .

Homework Prove that

eip1 dd p
Z
h(x̂, |x|) |x|→∞
−→ 2 d
Rd p (2π)
√ Z ∞
2
e− p~ dd−1 p~ Γ(d − 2)Ad−1
Z
Ad−1 −r d−3
= p = e r dr = ≡ Cd < ∞
Rd 2 p~ 2 (2π)d−1 2(2π)d−1 0 2(2π)d−1

for d > 2, where Ad−1 = volume of S d−2 .

We can state the previous calculation for r = 0 in two ways :

a) We showed C(x) ∼ G(x) as |x| → ∞. Here G(x) = 2-point function of the massless
free field, G(p)
b b for d > 2, defines a continuous function e− 12 (f,Gf )
= p12 . Note that G,
on S : We have
Z
Cd
(f, Gg) = dd xdd y f (x)g(y) ≤ Ckf km kgkm for m large enough.
Rd ×Rd |x − y|d−2

62
 b) Consider the function

Ld−2 C(Lx − Ly) = Ld−2 hφ(Lx)φ(Ly)i

eip(Lx−Ly) dd p eip(x−y) dd p
Z Z
d−2
=L = .
[−π,π]d µ(p) (2π)d 2
[−Lπ,Lπ]d L µ(p/L) (2π)
d

Formally, this converges

eip(x−y) dd p
Z
−→
L→∞ = G(x − y),
Rd p2 (2π)d
since L2 µ(p/L) → p2 as L → ∞.

We call this the scaling limit : take the statistical mechanics at the critical point, look
d−2
at long distances, scale φ (by L 2 above) ⇒ get quantum field theory. In this Gaussian
d−2
model we get : let xi 6= xj i 6= j, and ϕL (x) = L 2 φ(Lx) (for x ∈ (L−1 Z)d ).
Then
Z Y 2N Z Y 2N
−→
ϕL (xi )dµC (φ) L→∞ ϕ(xi )dµG (ϕ).
s0 i=1 S 0 i=1

Here the LHS is with a measure defined on variables φ(x), x ∈ Zd , a lattice model and
the RHS with a measure defined on variables ϕ(x) x ∈ Rd , a continuum model (to make
this precise, observe that (L−1 Z)d ) becomes “dense” in Rd as L → ∞).

The massive case

How about r > 0 ? Consider again the Ginzburg-Landau model, Cr = (−∆ + r)−1 on
`2 (Zd ). So,
x
eip |x| dd p
Z
2−d
Cr (x) = |x| 2 2 d
.
[−|x|π,|x|π]d |x| µ(p/|x|) + |x| r (2π)

For |x| < r−1/2 (let r be small now)

Cr (x) ∼ |x|2−d

and for |x| > r−1/2 scale differently, p → r1/2 p,

1/2
eipr x dd p
Z
d−2
Cr (x) = r 2
−1 1/2 p) + 1 (2π)d
[− πr , πr ]d r µ(r
d−2 1/2 |x|
∼ r 2 e−r .

Stated differently, let us put r = L−2 m2 and consider, after the change of variable pm → p,
eipx dd p
Z
d−2
L Cm2 /L2 (Lx) = 2 2 d
[−Lπ,Lπ]d L µ(p/L) + m (2π)
eipx dd p
Z
−→
L→∞ 2 2 d
≡ Gm2 (x).
Rd p + m (2π)

63
Hence
2N
Z Y Z Y
d−2
L 2 φ(Lxi )dµCm2 /L2 (φ) → ϕ(xi )dµGm2 (ϕ)
s0 i=1 S0

i.e. if we approach the critical point r → 0 in a suitable way and scale distances and φ
we get the continuum theory at nonzero m i.e. correlation length < ∞.

Remark All the scalings are completely natural :

Formally
1 1 Y
dµCr (φ) = e− 2 (φ,(−∆+r)φ) dφ(x)
Z d x∈Z
2−d
and φ(x) = L 2ϕL (x/L) gives
X
(φ, (−∆ + r)φ)) = φ(x)((−∆ + r)φ)(x))
x∈Zd
 
X X
= φ(x)  (φ(x) − φ(x + u)) + rφ(x)
x∈Zd |u|=1
 
X x x X x u x
= L2−d ϕL ( )  (ϕL ( ) − ϕL ( + )) + rϕL ( )
L L L L L
x∈Zd |u|=1
 
X X ϕL (y) − ϕL (y + u )
= L−d ϕL (y)  −2
L
+ L2 rϕL (y)
1
L
d
y∈( L Z) |u|=1

which, if ϕL went to a smooth function as L → ∞,

Z
−→ ϕ(y)(−∆ + m2 )ϕ(y)dd y = (ϕ, G−1
m2 ϕ),
L→∞ Rd
r=m2 /L2

10 Non-Gaussian Theory on the lattice

10.1 Measures
Let us now consider the λ 6= 0 Ginzburg-Landau model.
There are various finite volume theories that we can consider : Let Λ ⊂ Zd be an L-cube.

A) The measure !
1 X
exp −λ φ(x)4 dµCΛ (φ)
ZeΛ x∈Λ

on R where µCΛ is Gaussian with covariance CΛ = (−∆Λ +r)−1 and ∆Λ is the Laplacean
|Λ|

in Zd with some boundary conditions on ∂Λ (we have considered periodic and free above;

64
one can consider Dirichlet or Neumann on the lattice too). ZeΛ normalizes the total mea-
sure to 1.

B) The measure !
1 X
exp −λ φ(x)4 dµC (φ)
ZΛ x∈Λ

on s0 (Zd ) when µC is Gaussian with covariance C = (−∆ + r)−1 . Thus µC is in infinite

volume and the φ!4 perturbation is in finite volume. Note that, whereas in A), ZeΛ =
Z X
exp −λ φ4x dµCΛ is strictly > 0, this is not immediately obvious in B). However,
x∈Λ
recall Jensen’s inequality :
If ν is a probability measure in R, and f : R → R is convex: f (tx + (1 − t)y) ≤
tf (x) + (1 − t)f (y), t ∈ (0, 1), then
Z Z 
f (x)dν(x) ≥ f xdν(x) .

X
Apply it to the random variable x = λ φ4y to get
y∈Λ

Z !
X X
exp −λ φ4y dµC ≥ exp[−λ hφ4y i]
y∈Λ y∈Λ

where 2
dd p
Z Z
1
hφ4y i ≡ dµC (φ)φ4y = 3C(0) = 3 2
µ(p) + r (2π)d
from our rules for Gaussian integrals. Also, since hφ4y i > 0, we get

exp[−3λ|Λ|C(0)2 ] ≤ ZΛ ≤ 1. (10.1)

Note how the extensivity of ZΛ (= eO(Λ) ) is visible here. Also, note that the free energy :
1
−3λC(0)2 ≤ |Λ| log ZΛ ≤ 0, uniformly in Λ. Similar inequalities hold in the case A) above.

10.2 Perturbation Theory

It is possible to prove quite generally, using Ising-model approximations and various corre-
lation inequalities that the Λ → Zd limit of the above measures exist. These are, however,
special tricks and we want to understand these issues eventually using the renormalization
group. But let us for now proceed more heuristically and study the correlation functions

65
perturbatively in λ. Consider e.g. the pair correlation
!
R X
φ(x)φ(y) exp −λ φ(x)4 dµC (φ)
x∈Λ
hφ(x)φ(y)iΛ ≡ ! ≡ G2 (x, y). (10.2)
R X
exp −λ φ(x)4 dµC (φ)
x∈Λ

It is not hard to see that, for |Λ| < ∞, this is C ∞ in λ. Indeed consider e.g. the
X
φ(x)4 and set Fk (λ) := V k e−λV dµC (φ). Since
R
denominator and let V =
x∈Λ

|Fn−1 (λ + ) − Fn−1 (λ)| ≤ V n−1 |e−V − 1|e−λV ≤ ||V n e−(λ−||)V ≤ ||n!δ −n e−(λ−||−δ)V

and e−(λ−||−δ)V dµC (φ) < ∞ for λ − || − δ ≥ 0 we see that Fn−1 (λ) is differentiable with
R
0
Fn−1 (λ) = Fn (λ) and the latter equals the n:th derivative of the denominator. Proceeding
similarly with the numerator one shows that both numerator and denominator are C ∞
and from (10.1) above, ZΛ > 0. Hence G2 is smooth (of course this way we get terrible
bounds for the derivatives as Λ gets large).
Thus, we have Taylor’s expansion for G2 :
N
X
G2 (x, y) = G2,n (x, y)λn + RN
n=0

where lim λ−N RN (λ) = 0. The Taylor coefficients G2,n have a nice graphical representa-
λ→0
tion which we now derive.
Let us start with small n. Denote (10.2) by N
D
.
R
n = 0. N = φ(x)φ(y)dµC + O(λ), D = 1 + O(λ) so G2,0 = C(x − y) our Gaussian
covariance.
n = 1.
XZ
N = C(x − y) − λ φ(x)φ(y)φ(z)4 dµC
z∈Λ

and XZ
D =1−λ φ(z)4 dµC .
z∈Λ

From our rules of Gaussian integrals we get:

Z
φ(z)4 dµC = 3C(0)2

from the three pairings. The integral

Z
φ(x)φ(y)φ(z)4 dµC

66
 6!
has 23 ·3!
= 15 pairings and it equals

3C(x − y)C(0)2 + 12C(x − z)C(y − z)C(z − z).

So, X
N = C(x − y)(1 − 3λ|Λ|C(0)2 ) − 12λ C(x − z)C(y − z)C(0)
z∈Λ

and
D = 1 − 3λ|Λ|C(0)2
Hence altogether X
G2,1 = −12C(0) C(x − z)C(z − y).
z∈Λ

To proceed further, we need to introduce some notation. Clearly, we need to calculate

expressions like
Z YN N
Z Y
4
φ(x)φ(y) φ(zi ) dµC and φ(zi )4 dµC .
i=1 i=1
P Q
Both are given as sums of products of pairings P hαβi C(uα − uβ ) where {uα } are the
points x, y, zi , however such that each zi occurs 4 times. More precisely, let u4(i−1)+j = zi
for i = 1, . . . , N , j = 1, 2, 3, 4 and u4N +1 = x, u4N +2 = y. Then the sum is over pairings
of the set 1, . . . , 4N + 2.
To each pairing P we associate a graph G(P ) as follows : G = (V, L) consists of a set
V of vertices and a set L of lines. Vertices are the set {x, y, z1 , · · · , zN } = V (for the 1st
case). Lines ` join vertices : ` = {u`1 , u`2 } where u`1 , u`2 ∈ V and u`1 = u`2 is allowed.
Let Γ2 (V )be the set of such graphs G such that each vertex zi belongs to four lines (where
we count lines {zi , zi } twice) and each vertex x, y belongs to one line.
Given a G satisfying these conditions, there may be several P ’s such that G(P ) = G.
Call the number |{P |G(P ) = G}| := n(G).
Then
Z Y N X Y
φ(zi )4 φ(x)φ(y)dµC = n(G) C(u`1 − u`2 ). (10.3)
i=1 G∈Γ2 (V ) `∈G

Example. In hφ(z)4 φ(x)φ(y)i we had 2 graphs, G1 has lines {z, z}, {z, z}, {x, y} and G2
lines {x, z}, {y, z}, {z, z}. n(G1 ) = 3, n(G2 ) = 12.
Now, once we sum (10.3) over the points z1 , . . . , zN the order of the points zi will not
anymore matter. We get finally the following graphical rules which we state for arbitrary
correlations :

Definition 10.1 A φ4 - graph with 2m external legs, labelled 1, · · · , 2m, and N unlabelled
vertices is a graph on N points (vertices) such that each vertex has 4 lines attached (one

67
line can start and end at a vertex and is counted as two here), 2m lines have one end with
no other lines attached.

Examples.

2m = 2, N = 1

2m = 2, N = 2

2m = 2, N = 3

2m = 4, N = 1

2m = 4, N = 2

More : etc.
, ,

With this notation, we have

2m
!
dn 
Z Y X X
4
φ(xi ) exp −λ φ(x) dµC (φ) = (−1)n n(G)A(G)
dλn λ=0

i=1 x∈Λ G

where G run through (2m, n)-graphs, n(G) is the number of pairings giving that G and
the amplitude A(G) corresponding to the graph G is the following expression:
1. Label the n vertices of G as z1 , · · · , zn
2. To each line ` = {u, v} of G put C` = C(u − v).
X Y
A(G) = C` . (10.4)
z1 ,··· ,zn ∈Λ `

 
We denote for short A by .

68
Example. We get to O(λ) :

−λ·4·3 − λ · 3(− ) + O(λ2 )

G2 =
1−λ·3 + O(λ2 )
   
= − 12λ − 3λ − + 3λ − + O(λ2 )

= C(x − y) − 12λ + O(λ2 )

X
= C(x − y) − 12λ C(x − z1 )C(z1 − z1 )C(z1 − y) + O(λ2 )
z1
X
= C(x − y) − 12λC(0) C(x − z1 )C(z1 − y) + O(λ2 )
z1 ∈Λ

Remark. Note the presence of disconnected graphs like that are cancelled upon
normalization.
Homework : Prove :

G2 = − 12λ + λ2 (α +β +γ ) + O(λ3 )

and find α, β, γ

Theorem 10.2
Z
4
P
n φ(x)φ(y)e−λ φ (x) dµC
d  n
X
c
Z = (−) n(G)A(G)
dλn λ=0
 P 4
e−λ φ (x) dµC G

X
c
where the sum runs over connected graphs.

Remark. In the numerator the graphs are of the type · ·

| {z }
connected

In the denominator · ·

69
I.e. connected components have either both x and y or neither.

Note that disconnected graphs are here proportional to a power of the volume :

X
= C(x − y) C(0)2 = |Λ|C(0)2 C(x − y)
z∈Λ

= C(x − y)C(0)4 |Λ|2 etc

Proof of the Theorem. Notation :

4
P
F (φ)e−λ φ dµC
R
hF iλ = R
e−λ φ4 dµC
P

R
so hF i0 = F dµC .
Then,
d X X
hF iλ = − hφ(x)4 F iλ + hφ(x)4 iλ hF iλ
dλ x

d X
i.e. hF iλ = − hφ(x)4 ; F iλ where we denote by
dλ x

hF ; Giλ = hF Giλ − hF iλ hGiλ

the so called truncated (or connected) correlation functions.

Y Y
Lemma 10.3 Let F = φ(xα ), G = φ(xβ ).
X Yα∈A β∈B

Then hF ; Gi0 = C(xγ − xδ ) where each pairing P has at least one pair {γ, δ}
P {γ,δ}∈P
such that γ ∈ A, δ ∈ B.

Proof Obvious, since hF ihGi has those pairings where no such pairs occur and hF Gi
has all pairings. 
Now (Prove !)

dn  n
X
n
hF iλ = (−1) hF ; φ(x1 )4 ; φ(x2 )4 ; · · · ; φ(xn )4 i0
dλ λ=0

x ···x 1 n

where
X Y Y
hF1 ; F2 ; · · · ; FN i0 = (−1)|π|−1 (|π| − 1)! h Fi i0
π α i∈πα
|π|
where π = {πα }α=1 is a partition of {1, · · · , N } into |π| subsets.

70
Example. hF1 ; F2 ; F3 i = hF1 F2 F3 i−hF1 F2 ihF3 i−hF1 F3 ihF2 i−hF2 F3 ihF1 i+2hF1 ihF2 ihF3 i

Now the thing to check is

Y
Lemma 10.4 Let Fi = φ(xiα ). Then
α∈Ai
X Y
hF1 ; · · · ; FN i0 = C(xγ − xδ )
P {γ,δ}∈P

−1
where for each partition into pairs P , there is a {{γi , δi }}N
i=1 ⊂ P that forms a connected
path, connecting the Ai ’s, i.e. there exists a permutation ρ of the Ai ’s such that

γ1 ∈ Aρ(1) , δ1 ∈ Aρ(2) , γ2 ∈ Aρ(2) , δ2 ∈ Aρ(3) · · · δN −1 ∈ Aρ(N ) .

A4
A2
A3

A1

Example

The Lemma yields our theorem since there cannot be a connected graph with only one
leg. 
4
Y
Problems. 1. Let G4 (x1 · · · x4 ) = h φ(xi )i and define the connected four point function
i=1
Gc4 (x1 · · · x4 ) = hφ(x1 ); . . . ; φ(x4 )i. Prove that

G4 (x1 · · · x4 ) = G2 (x1 − x2 )G2 (x3 − x4 ) + G2 (x1 − x3 )G2 (x2 − x4 ) + G2 (x1 − x4 )

G2 (x2 − x3 ) + Gc4 (x1 , · · · , x4 )
N
X
Show that Gc4 has an expansion λn Gc4,n + O(λN +1 ) with Gc4,n consisting of connected
n=1
graphs with 4 legs.
2. Generalize 1. to the connected N-point function

GcN (x1 · · · xN ) = hφ(x1 ); . . . ; φ(xN )i

by showing that XY
GN (x1 · · · xN ) = Gc|I| (xI )
π I∈π

71
and that the perturbation expansion of GcN consists of connected graphs with N legs.
3. Let
Z(J) := he(φ,J) i.
Hence !
N N
Y Y ∂ 
GN (x1 , . . . , xN ) := h φ(xi )i = Z(J).

∂J(xi )

J=0
i=1 i=1

Define F (J) := log Z(J). Show

N
!
Y ∂ 
GcN (x1 , . . . , xN ) = F (J).

∂J(xi )

J=0
i=1


dn 
4. Prove that dλ n  log Z is sum of connected vacuum graphs with n vertices. (a vacuum
0
graph is a graph with no legs).

How to calculate the amplitude A(G) of a given graph G? First, let us observe that
in our model G2m,n has a Λ → Zd limit : Clearly it suffices to consider a connected graph
which is an expression (let say m = 1)
X
A(x1 , x2 , z1 · · · zn )
z1 ···zn ∈Λ

and A is a product of C(yi − yj ) where the y’s are x’s or z’s. Use |C(x)| ≤ const. e−|x|/ξ ,
and note that, since G is connected, there exists a connected tree graph in G containing
all the vertices and all the legs (a tree graph has no loops).
Example.

z3
z2

−→ x z1 y

X 
Now do the zi sums using  C(zi − zj )≤ const (independent on Λ) starting at ends of
 
zj ∈Λ
branches (e.g. z3 above).

72
10.3 Momentum space representation
Actual calculations are easier in Fourier transform. In order not to worry about analysis
let us first work in finite volume Λ = ZdL . Recalling from Section 7.2. we have for p ∈ BL
(defined in (7.3)) X
fˆ(p) = e−ipx f (x).
x∈ΛL

and the inverse formula

X Z
f (x) = L −d
eipx
fˆ(p) ≡ dpeipx fˆ(p)
p∈BL L

R
where dp is a convenient shorthand for the Riemann sum, converging as L → ∞ to
L
dp where dp is the normalized Lebesgue measure i dp
R Q i
[−π,π]d 2π
. We have, see (7.4, 7.5),
Z
dpeip(x−y) = δxy for x, y ∈ ΛL (10.5)
L

and
X
ei(p−q)x = Ld δpq ≡ (2π)d δL (p − q) p, q ∈ BL . (10.6)
x∈ΛL

where we defined the discrete delta function δL (p − q). Products work out as
Z
ˆ
(f g)ˆ(p) = f ∗ ĝ := dq fˆ(p − q)ĝ(q)
L

and
(f ∗ g)ˆ = fˆĝ.

Example.
X
= C(x − z1 )C(z1 − z2 )3 C(z2 − y) = (C ∗ C 3 ∗ C)(x − y)
z1 z2

Hence using above rules

Z
ip(x−y)
= G(x − y) = dpG(p)e
b
L

with Z Z 
G(p)
b = dq drC(q)
b C(r)
b C(p b 2
b − q − r) C(p)
L L

or pictorially

73
 q
r
p p
p-q-r
A more useful way to derive the same result is to insert for each C its Fourier represen-
tation:
XZ 5
Y
i[p1 (x−z1 )+p2 (z1 −z2 )+p3 (z1 −z2 )+p4 (z1 −z2 )+p5 (z2 −y)]
= e C(p
b i )dpi
z1 z2 L i=1
Z 5
Y
d d ip1 x−ip5 y
= (2π) δL (−p1 + p2 + p3 + p4 )(2π) δL (−p2 − p3 − p4 + p5 )e C(p
b i )dpi
L i=1

The first delta-function is due to the sum over z1 and the second to the sum over z2 . Note
“momentum conservation” at vertices.

p2 p2
p1 4
X 4
X
p3 p3 p5 p1 = pi , p 5 = pi
i=2 i=2
p4 p4

Solving for the constraints we get again

Z
= dp1 dp2 dp3 C(p
b 1 )C(p
b 2 )C(p
b 3 )C(p b 1 )eip1 (x−y)
b 1 − p2 − p3 )C(p
L

using p5 = p1 , p4 = p1 − p2 − p3 . These expressions now have the obvious limits as L → ∞

R R
with L replaced by .
The general graph is now obvious :
pn
p1
1) But momentum pi to legs, such that
P
pi = 0
p2 Reason : G(x1 , · · · , xn ) is translation invari-
ant :
G(x1 , · · · , xn ) = G(x1 + y, x2 + y, · · · , xn + y)

74
for all y so
X P
b 1 , · · · , pn ) =
G(p G(x1 , x2 , · · · , xn )e−i i p i xi

x1 ,...,xn
X P
= G(0, x2 − x1 , · · · , xn − x1 )e−i i pi xi

x1 ,...,xn
X Pn P
= G(0, x2 , · · · , xn )e−i i=2 pi xi −ix1
e i pi

x1 ,...,xn
X
= ĝ(p2 , . . . , pn )(2π)d δL ( pi )

where g(x2 , . . . , xn ) = G(0, x2 · · · , xn ).

2) Satisfy momentum conservation at each vertex (coming from the delta functions).

3) Integrate over each independent internal momentum.

Example.

p1 -p-p-p
1 2 3
q1
C(p 1 − p2 − p3 )C(p2 )C(p3 )
b 1 )C(−p
b b b
Z Y3

q2 · dqi C(q
b 1 )C(q
b 2 )C(q
b 3)
i=1
p-q-q q +q -p-p-p
1 2
1 1 2 q3 1 2 3
·C(p
b 1 − q1 − q2 )C(q
b 1 + q2 − p 1 − p2 − p3 )
·C(p
b 1 + p2 − q1 − q2 − q3 ) ≡ Γ(p1 , p2 , p3 )

p+p-q-q-q
1 2 1 2 3
p3 and in x space the graph is
p2

Z 3
i(p1 (x1 −x4 )+p2 (x2 −x4 )+p3 (x3 −x4 ))
Y dpi
G(x1 , · · · , x4 ) = e Γ(p1 , p2 , p3 )
i=1
(2π)d

Remark 1. We have seen so far that :

(Λ)
a) If |Λ| < ∞, the functions G2m are C ∞ in λ
(Λ)
b) The Taylor coefficients G2m,n have a Λ → Zd limit.
(Λ)
It can be shown that if r > 0 and λ is small enough the functions G2m have a Λ → Zd
(Λ)
limit, they are C ∞ in λ and their Taylor coefficients are the lim G2m,n .
Λ→Zd

Remark 2 The expansion is not convergent. Consider e.g. a lattice consisting of one
point i.e. the integral Z ∞
4 r 2
F (λ) = e−λϕ − 2 ϕ dϕ
−∞

75
Then
dn F  n
Z
4n − r2 ϕ2 −2n n
√ (4n)!
αn = n
= (−) ϕ e dϕ = r (−1) 2π 2n ∼ C n (2n)!
dλ λ=0 2 (2n)!


P 1
as n → ∞. Thus the Taylor series α λn has αn!n ∼ C n n! i.e. is very badly divergent.
n! n

Homework. Prove that F (λ) is analytic in the region C\{λ ∈ R, λ ≤ 0}

and has an essential singularity at λ = 0.

11 Infrared and Ultraviolet Divergencies

11.1 Infrared
Recall that for λ = 0, the critical point of our model is at r = 0. What happens for
λ > 0 ? The following is expected to be true :

For λ > 0, there exists rc (λ) such that for r > rc (λ)

0 ≤ hφ(x)φ(y)i ≤ Ae−|x|/ξ , ξ < ∞ , ξ(r)−→

r↓rc (λ) ∞.

For r < rc (λ) there are two phases, where hφ(x)i = ± m 6= 0 and

|hφ(x); φ(y)i |≤ Ae−|x|/ξ ξ < ∞,

with ξ(r) → ∞ as r ↑ rc (λ).

For r = rc (λ),
lim hφ(x)φ(y)i|x − y|d−2+η 6= 0
|x−y|→∞

where η is independent of λ > 0, it depends on the dimension d, η = 0 if d ≥ 4, and it

equals the η of the Ising model. These claims are proven, for small λ and d ≥ 4. The rest
remains a conjecture, although with plenty of theoretical and numerical evidence.

Let us see how these facts are reflected in the behavior of the perturbation theory when
r = 0. Let us consider the momentum space expressions for the graphs at r = 0. Since
all integrals are over a bounded region [−π, π]d the only problem could come from p = 0
Xd
where µ(p) = 2 (1 − cos pi ) = p2 + O(p4 ) vanishes. Consider the following graph
i=1

76
 p = p 1 + p2

Now
q
dd q 1
Z
1
= d
≡ I(p)
[−π,π]d (2π) µ(q) µ(p − q)
p-q
This is integrable in d > 2 if p 6= 0 and behaves as

d−4
 |p|
 d<4
I(p) ∼p→0 log |p| d=4

constant d>4


n n(d−4)
Thus | {z } = I(p) ∼ |p| which is not integrable if
n
n(4 − d) ≥ d (i.e. for d = 3 if n ≥ 3). Hence e.g. the graph

p p–q

has amplitude
I(q)n dd q
Z

µ(p − q) (2π)d
which is ill-defined for d < 4 and n large.

Conclusion. The individual terms of our perturbation expansion

X
Gm = λn Gm,n

Gm,n = Gm,n (x1 · · · xm , r) have no limit as r → 0 if d < 4 (at least some of them). For
d ≥ 4 it may be shown that this limit exists. Hence we expect to need radically new ideas
for d < 4. The p ∼ 0 divergencies are called IR (infrared) divergencies.

77
11.2 Ultraviolet
Consider now field theory. Let dµG (ϕ) be the Gaussian measure on S 0 (Rd ) with covariance
G = (−∆+m2 )−1 . We would like to consider the continuum limit of the Ginzburg-Landau
model i.e. to start with finite volume Λ ⊂ Zd we would consider the measure

ϕ(x)4 dd x
R
e−λ Λ dµG (ϕ).

We immediately run into trouble: ϕ(x)4 , as we saw, is not a well defined random variable.
R R
Indeed, ϕ(x)4 dµG = ∞. Actually, we only know ϕ(f ) ≡ ϕ(x)f (x)dx for f ∈ S(Rd ) is
a measurable function on S 0 and has finite moments ϕ(f )n dµ < ∞, ∀n.
R
R
One way to proceed is to regularize the theory. Recall that ϕ(x)4 dµG = 3G(0)2 and
dd p 1
R
G(0) = Rd G(p)
b
(2π)d
. For us, G(p)
b = p2 +m 2 so the divergence is due to insufficient decay

of G(p)
b as |p| → ∞ i.e. UV (ultraviolet) divergence. Thus, let us replace G(p)b by

G
b (p) = G(p)χ(p)
b ≡ G(p)χ
b  (p)

2
where χ ∈ S(Rd ) cuts of large |p|. We demand χ(0) = 1 and χ ≥ 0. E.g. χ(p) = e−p
is a good choise. As  → 0, G
b (p) → G(p)
b pointwise. Now ϕ(x) is a well defined random
variable :
dd p
Z Z
2n (2n)! n
ϕ(x) dµG = n G (0) , G (0) = G(p)χ(p)b <∞
2 n! (2π)d

[To be pedantic : a priori we know only that ϕ(f ) is measurable, for f ∈ S, so consider
(x−y)2
e− 2
ϕ(f,x ), f,x (y) = . This is a measurable function ` : S 0 → R : ϕ 7→ ϕ(f,x ).
(2π)d/2
Thus lim→0 ` also is measurable and this is what we call ϕ(x) (formally : ϕ(f,x ) =
R R
ϕ(y)f,x (y)dy−→ →0 ϕ(y)δ(x − y)dy = ϕ(x))].
Hence, consider the measure
1 − R [aϕ(x)2 +λϕ(x)4 ]dd x
e Λ dµG (ϕ) (11.1)
ZΛ,

where we added for later purpose also a quadratic term to the Hamiltonian. This measure
is well defined provided a ∈ R, λ > 0. Indeed, as before by Jensen’s inequality we get

ZΛ, ≥ exp −|Λ|(aG (0) + 3λG (0)2 )

 

2 a2
a
and since aϕ2 + λϕ4 ≥ − 4λ , ZΛ, ≤ e|Λ| 4λ .
The correlation functions are again C ∞ in λ and a, and we have the expansion
N
X
G2k = λn ap G2k,n,p + RN RN = O(λN +1 , aN +1 )
n,p=0

78
given in terms of graphs. The coefficients, G2k,n,p for m2 > 0, have a |Λ| → Rd limit but
not an  → 0 limit, as we saw above.
The question of continuum limit will not arise in statistical mechanics where the lattice
spacing  is a fixed nonzero quantity. However, in quantum field theory this question arises.
In this case there is no fundamental length and one would like to have a quantum theory
of continuum fields. Moreover these fields depend also on time and are defined on the four
dimensional space time R4 . Brushing aside for the moment the fact that this space time
carries the Minkowski metric instead of the Euclidean one we have been considering and
that in quantum mechanics the correlation functions of fields are replaced by expectation
values of field operators in Hilbert space, it turns out that in case of a scalar field theory
one needs to address precisely the question of  → 0 limit as above.
The problem of divergences of the coefficients of perturbation theory arose in the
1930’s when physicists were studying Quantum Electrodynamics (QED), the quantum
theory of electromagnetic field interacting with electrically charged matter. In this theory
there are two sorts of fields, the electromagnetic field A(x) which resembles our φ(x) but
instead takes values in A(x) ∈ R4 . A(x) is the four dimensional electromagnetic vector
potential. It is a Gaussian field with zero mass, thus resembling our r = 0 free field.
The second field describes electrons and has non zero mass. The interaction term in the
Hamiltonian makes the theory non Gaussian; the analogue of our parameter λ is played
by the electron charge e. One can then proceed to derive a perturbation series as above
in powers of e and as above the individual terms are ill defined, diverging due to small
scale (large momentum) behavior of the integrands.
One would like to think about the electron charge (and its mass) as given physical
constants that enter the Hamiltonian describing the dynamics. The divergence of the
perturbation theory indicates that this point of view is incorrect. Rather one should
think about these parameters depending on scale. Thus physicists introduced in the 50’s
the idea of renormalization. We should think about the parameters entering -cutoff
theory describing the charge and mass of the electron in that scale. These can very well
be different from the ones we measure that are interpreted as the charge and mass of
the electron in the measurement scale. Once we have introduced the renormalization
group we will make this picture more precise. For the time being we just remark that the
parameters a and λ entering the measure (11.1) should be allowed to depend on the scale
. Then the  → 0 limit question can be posed as follows.
Question. Find three functions Z(), a(), λ() such that
2m
Y
Z()m h ϕ(xi )i
i=1

79
 2m
Y
0 2md m
converge, as  → 0, to elements of S (R ) (or Z() h ϕ(fi )i , converge for all fi ∈
i=1
S(Rd )). Here h i is the expectation w.r.t. the limit of the measure (11.1), as Λ → Rd ,
a = a() and λ = λ().

Remarks. 1. We allowed (it turns out this often is necessary) for a third parameter that
1
renormalizes the value of the field: ϕ(xi ) is replaced by Z() 2 ϕ(xi ).
Obviously λ() = 0, a() = 0, Z = 1 is a solution. We ask if there are any limits that are
non-Gaussian. The answer is the following:
For d = 2, Z = 1, a = −6λG (0), λ() = λ works.
For d = 3, Z = 1, a = −6λG (0) + αλ2 log  (α explicit), λ() = λ works.
For d ≥ 4 there is no non-Gaussian limit.
Our objective is to explain these claims.

1
Remark. χ is a regulator. It cuts off momentum |p| > ∼  very effectively. We could also
use a lattice-cutoff. Thus consider the following Hamiltonian on fields defined on (Z)d :
Let ϕ : (Z)d → R and set

1
X (ϕ(x) − ϕ(y))2 r X d X
H= 2
d +  ϕ(x) 2
+ d (a()ϕ(x)2 + λ()ϕ(x)4 ). (11.2)
2 2 x x
|x−y|=

Consider as above the renormalized correlations

2m
Y
G (x1 , . . . , x2m ) := Z()m h ϕ(xi )i .
i=1

As we saw before, setting

d−2
φ(x) =  2 ϕ(x) , x ∈ Zd (11.3)

we have

G (x1 , . . . , x2m ) = Z()m (2−d)m G() (x1 /, . . . , x2m /) (11.4)

where
2m
Y
G() (y1 , . . . , y2m ) = h φ(yi )i()
i=1
()
and h−i is the expectation w.r.t. the measure
" #
1 X
lim exp − 2 a()φ(x)2 + 4−d λ()φ(x)4 dµC (φ)
Λ↑Z d ZΛ,
x∈Λ

80
where dµC is the Gaussian measure on unit lattice fields φ ∈ s0 (Zd ) with covariance
C = (−∆ + 2 r)−1 .
Thus our UV problem  → 0 is the same thing as staying on a fixed lattice, taking
distances to ∞ (xi /), and going to critical point (2 r → 0). It should now be obvious
that both IR and UV problem have something to do with Statistical Mechanics at the
critical point.

Remark. The -lattice theory can be interpreted as a particular cutoff of the continuum
theory. Namely, the measure with H as in (11.2) with a = λ = 0 is the Gaussian measure
on s0 ((Z)d ) with covariance
X C = (−∆ + r)−1 where the latter is an operator on `2 (Zd )
i.e. (−∆ ϕ)(x) = [ϕ(x) − ϕ(x + i)]−2 for ϕ : (Z)d → R. Concretely
|i|=1

eip(x−y) dd p
Z
1
(x, y) =
−∆ + r [− π , π ]d −2 µ(p) + r (2π)d

which shows that p is cut-off to |pi | ≤ π/. Similarily, instead of the lattice model we may
consider the Statistical Mechanics where ϕ is on Rd but has the cutoff χ, i.e. the measure
1 −λ R ϕ(x)4 dx
lim e Λ dµG1 (ϕ)
Λ→Zd ZΛ
1
with G1 (p) = p2 +r
χ(p).

12 The Renormalization Group

12.1 The Block-spin Transformation
Suppose that we have a Statistical Mechanics model at its critical point such that
A B
| hφ(x)φ(y)i − a
|≤ (12.5)
|x − y| |x − y|a+

(as |x − y| > 0) where  > 0. We are after a theory that explains the leading term in
this asymptotics and the fact that this term (or the exponent a) is universal i.e. stays
the same when the Hamiltonian is changed (at least under some changes). Thus suppose
there is a class of Hamiltonians that satisfy (12.5). Then the details of these Hamiltonians
affect the sub leading asymptotics i.e. the RHS of (12.5). We can get a special theory
where the scale invariance satisfied by the leading term is exact by taking the scaling limit.
Recall the relations (11.3) and (11.4) between cutoff quantum field theory and statistical
mechanics. Let, for L > 0
a
ϕL (x) = L 2 φ(Lx). (12.6)

81
Then if (12.5) is satisfied we get

A
lim hϕL (x)ϕL (y)i = .
L→∞ |x − y|a

On the other hand ϕL can be seen as a cutoff 1/L field as it depends on x ∈ (L−1 Z)d .
Hence if the LHS is given by an expectation hϕ(x)ϕ(y)i the field ϕ lives on Rd . Thus we
expect the limits, if they exist,
Y
lim h ϕL (xi )i
L→∞
i

to be the correlation functions of a field theory. This field theory is called the scaling
limit of our model. By definition it is scale invariant, e.g.

La hϕ(Lx)ϕ(Ly)i = hϕ(x)ϕ(y)i.

Universality then would mean that several Hamiltonians give rise to the same scaling
limit.
The drawback of this formulation is that, first, the scaling limit is a different object
than the one we started with. The latter one is a fixed (unit) lattice model whereas the
former one is a continuum object. More importantly, we have not given any constructive
approach to the study of the scaling limit: we still need to understand the statistical
mechanics at the critical point, in particular we need to show (12.5). The Renormalization
group addresses these two problems. First, it supplements scaling by another operation,
coarse graining, that allows one to stay in the category of fixed lattice spacing theories.
Second, it provides actually a tool to study the critical theory.
We replace the scale transformation (12.6) by
a
φL (x) = L 2 φaverage (Lx) (12.7)

where φaverage (Lx) is the average of φ in a L-sided cube centered at Lx:

X
φaverage (Lx) = L−d φ(Lx + y) (12.8)
y:|yi |< L
2

d d
(12.7) and (12.8) define a map φ ∈ RZ → φL ∈ RZ , i.e. the scaled field is also defined
on Zd . We take for convenience L > 1 odd integer.

• • •
Here L = 3
• • Lx •

• • • Lx + y

82
We see that φ → φL involves :
a) “Coarse graining” : average over details of φ for scales ≤ L.
b) Scaling : φL (x) depends on φ near Lx and we multiply φ by La/2
φ → φL is called the Block-spin transformation. (12.7) and (12.8) define a linear map
d d
CL : RZ → RZ , concretely,
(
a 1 |(Lx − y)i |< L/2
(CL )xy = L 2 −d
0 otherwise

Note that
CL2 = CL2 .
Thus, the iteration of (12.8) n times is the same as doing it once with Ln :

φLn (x) = (CLn φ)(x) = (CLn φ)(x)

a
X
= Ln( 2 −d) φ(Ln x + y)
|yi |<Ln /2

Now observe :

Proposition 12.1 Suppose that (12.5) holds. Then, for maxi |xi − yj | > 1,
∗
h(CLn φ)(x)(CLn φ)(y)i−→
n→∞AG (x, y)

with Z Z
∗
G (x, y) = du dv|x − y + u − v|−a
 

and this holds also for all x, y, if a < d. Here  = [− 21 , 12 ]d .

Remark. G∗ decays as 1
|x−y|a
as |x − y| → ∞.
Proof. We get, for x 6= y,
X
| hCLn φ(x)CLn φ(y)i − AL(a−2d)n |Ln (x − y) + u − v|−a |
u,v
|vi |,|ui |<Ln /2
X
≤ BL(a−2d)n |Ln (x − y) + u − v|−a−
u,v
X
= BL−n L−2dn |x − y + L−n (u − v)|−a−
u,v
Z Z
−→ 0
n→∞ · du dv|x − y + u − v|−a− = 0
 

and the A-term → A  du  dv|x − y + u − v|−a . We used max |xi − yi | > 1, |(u − v)i | <
R R

1 ⇒ |x − y + u − v| > 0, so integrals converge. The rest is similar. 

83
 1
Example For the Gaussian Ginzburg-Landau model with covariance −∆ we take a = d−2
and get n n
n n n(d−2) −2nd
XZ dd p eip(L x−L y+u−v)
hCL φ(x)CL φ(y)i = L L d
d (2π) µ(p)
u,v [−π,π]
−n
dd p eip(x−y+L (u−v))
X Z
−2nd −2n
=L L
u,v [−Ln π,Ln π]d (2π)
d µ(L−n p)
Z Z  
1
−→
n→∞ du dv (x − y + u − v),
  −∆Rd

using L2n µ(L−n p) → p2 , as n → ∞, and where −∆Rd is the usual −∆ on L2 (Rd ). Now,
1 const
(x − y) = .
−∆Rd |x − y|d−2

1
Remark. In this example, CLn φ(x) are Gaussian, with covariance CLn −∆ (CLn )T . T denotes
the transpose, this is just what we write above, i.e.
 
n 1 1
X
n T n n
CL (C ) (x, y) = (CL )xz (CL )yw .
−∆ L z,w
−∆ zw

How about generally ?

12.2 Transformations on measures

d
Suppose µ is a cylinder measure in RZ and consider the random variables CL φ. Their
generating function is
Z Z
i(CL φ,f ) T
WL (f ) = e dµ(φ) = ei(φ,CL f ) dµ(φ)

= W (CLT f )

which is obviously of positive type, so there is a measure µL such that

Z
WL (f ) = ei(Ψ,f ) dµL (Ψ).

d d
µL is the probability distribution of the block spins CL φ. Thus CL : RZ → RZ induces
a map
CL∗ : B → B
where B are the cylinder measures on s0 (Zd ).
CL∗ is called the Renormalization Group Transformation on measures.

84
12.3 Transformations on Hamiltonians
Concretely, let us work in finite volume ΛN and take ΛN = LN -box centered at the origin.
Thus φ ∈ RΛN , CL φ ∈ RΛN −1 and CLn φ ∈ RΛN −n . Let µ be a probability measure
1 −H(φ) ΛN
dµ(φ) = e d φ
Z
when H is some function H : RΛN → R such that e−H is integrable. Then,

dµL (Ψ) = F (Ψ)dΛN −1 Ψ

with Z
1 Y
F (Ψ) = e−H(φ) δ(Ψ(x) − (CL φ)(x))dΛN φ
Z x∈ΛN −1
R
(Z normalizes F dΛN −1 Ψ = 1).
a
Let us write this more explicitely : Ψ(x) equals L 2 times the average of φ in the L-cube
centered at Lx. Thus, if we fix Ψ(x), ∀x ∈ ΛN −1 , we need to integrate out all fluctuations
around this average. In other words, let φ0 be a configuration that is constant in each
a
L-cube, and equals the average of φ there. I.e. φ0 (Lx + y) = L− 2 Ψ(x) for all x ∈ ΛN −1 ,
all y with |yi | < L/2. Using our previous notation this equals
a a
φ0 = L− 2 L− 2 +d CLT Ψ = Ld−a CLT Ψ
a
(recall, (CLT )xy = L 2 −d if x ∈ L-cube at Ly, and equals 0 otherwise). Thus φ = φ0 + Z
with Z having 0 average over L-cubes i.e. CL Z = 0. Thus
Z !
1 d−a T
Y
F (Ψ) = e−H(L CL Ψ+Z) δ (CL Z)(x) dΛN Z
Z x∈Λ N −1

Even more concretely, take z ∈ RΛN \LΛN −1 i.e. z(x) ∈ R, ∀x ∈ ΛN where x 6= Lw, w ∈ Zd .
Put 
 z(x) x ∈ ΛN \LΛN −1
X
Z(x) = − z(Lw + y) x = Lw

|yi |<L/2

Then CL Z = 0 and so
Z
1 d−a C T Ψ+Z)
F (Ψ) = e−H(L dΛN \LΛN −1 z
Z
a concrete integral over “fluctuation variables” z(x).
0
Let F = e−H . The map RL : H → H0 is called the RG Transformation on Hamiltonians,
i.e.
Z
d−a T
RL H = − log e−H(L CL Ψ+Z) dz (12.9)

85
Here H : RΛN → R and RL H : RΛN −1 → R. Note how the volume contracted from an
LN -box to an LN −1 -box.
Remarks. 1. It was trivial to define the RG on measures and in infinite volume : CL∗ µ
is automatically a measure. On Hamiltonians, we need to
a) work in finite volume,
b) make some assumptions on H so that RL H in (12.9) is well defined.
2. RG is called a “group” because CLn = CL n i.e. (CL∗ )n = CL∗ n and (RL )n = RLn . Actually,
it is a semigroup : R1 = id but RL−1 is not defined.
Our Proposition 12.1 says that
Z
−→ G∗ (x, y)
(CLn φ)(x)(CLn φ)(y)dµ n→∞

(absorbing the constant A in the definition of G∗ ), i.e.

Z
Ψ(x)Ψ(y)dµLn → G∗ (x, y)

Thus, we might hope that the measures µLn converge to some measure µ∗ such that
Z
Ψ(x)Ψ(y)dµ∗ (Ψ) = G∗ (x, y)

and µ∗ is a fixed point of CL∗ :

CL∗ µ∗ = µ∗
Equivalently, provided we make sense of H in infinite volume, we might expect

−→ H∗ ,
(RL )n H n→∞ RL H∗ = H∗

in some topology in a space of H’s.

Remarks. 1. It is actually quite difficult to set up nice spaces of Hamiltonians for
unbounded spins : things are very sensitive to the dependence of H on large values of
φ(x)’s : we need the φ(x) integral to converge at ±∞. This problem did not occur for
bounded spins where we could define H via potentials ΦX and talk about the Banach space
B of Hamiltonians (i.e. of ΦX ’s) : Given Φ ∈ B, we could define the set of Gibbs measures
µ in infinite volume corresponding to Φ. Moreover, above the critical temperature, we
expected Φ to uniquely determine a µ. For unbounded spins no similar formalism exists.
2. For bounded spins σ the RG, as defined above, will in general change the range of
values that σ takes : e.g. for Ising spins σ(x) ∈ {±1} CL σ takes values in the set
a a a
{σ ∈ L 2 −d Z||σ| ≤ L 2 } i.e. for a > 0, larger values. Upon iteration, CLn σ(x) ∈ L( 2 −d)n Z,
an
|CLn σ(x)| ≤ L 2 . Thus, e.g. at d = 3, where we expect a = d − 2 + η with η ≈ 0.1,
as n → ∞ the block spins become unbounded and continuous! Anyway, for n < ∞,

86
we have a bounded spin model and can ask whether RL actually maps Hamiltonians to
Hamiltonians. I.e. suppose µ is a Φ-Gibbs measure, Φ ∈ B. Is CL∗ µ a Gibbs measure for
some Φ0 ? If so, define RL Φ = Φ0 . This is expected to be true in Ising model, but proven
only for small β or large β (where one needs to enlarge B, surprisingly !).
3. For unbounded spins, RL and its iterations have been rigorously studied for the
Ginzburg-Landau model. It turns out that one can, to a certain extent set up a space
of Hamiltonians where RL acts; however, one has to supplement this with a different
representation of µ for large values of φ’s.

For the rest of the time, we will not discuss these problems. Rather we

a) Pretend that RL is defined in some space of H’s and see what implications this could
have (Section 13).

b) Carry out a perturbative analysis of RLn H for H = Ginzburg-Landau model and un-
derstand the IR and UV problem this way (Sections 14-16).

13 RL near a Fixed Point : Critical Exponents

13.1 General Framework
We assume the following setup
1. We have some space K of “Hamiltonians” H such that H ∈ K determines a unique
d
measure µH on RZ . We denote hF iH = F dµH . Here µH is the limit of Z1Λ e−HΛ (φ) dΛ φ as
R

Λ % Zd and HΛ is some finite volume version of H (or more properly, H = {HΛ }Λ⊂Zd ).
2. The RG is defined in K : RL : K → K i.e.
CL∗ µH = µRL H
We also suppose that there is a metric in K and if Hn → H, then µHn → µH .
Let us see what kind of picture of critical phenomena emerges from such assumptions.
Let H∗ be a fixed point of RL : RL H∗ = H∗ .
Definition 13.1 The stable manifold of H∗ in K is
Ms = {H ∈ K | RLn H → H∗ }
What can we say about the decay of correlations of H ∈ Ms ? Let us define
Definition 13.2 H ∈ Ms is critical if
sup |hφ(x)φ(y)iH |e|x−y|/ξ = ∞
x,y

87
for all ξ < ∞.
Hence for critical H there are no A > 0, ξ < ∞ such that |hφ(x)φ(y)iH | ≤ Ae−|x−y|/ξ for
all x, y. X
a
Recall a in the definition of CL : (CL φ)(x) = L 2 L−d φ(Lx + y). We have
|yi |< L
2

Proposition 13.2 a) Suppose a < d. Then all H ∈ Ms are critical.

b) If a = d and H ∈ Ms is not critical, then hφ(x)φ(y)iH∗ = 0 for x 6= y.

Proof a) Suppose ∃A, ξ. Then,

0 6= |hφ(x0 )φ(y0 )iH∗ | = | lim hφ(x0 )φ(y0 )iRLn H |

n→∞
n n
= | lim hCL φ(x0 )CL φ(y0 )iH |
n→∞
X X
= | lim Ln(a−2d) hφ(Ln x0 + u)φ(Ln y0 + v)iH |
n→∞
|ui |<Ln /2 |vi |<Ln /2
X X
≤ A lim Ln(a−2d) exp[−Ln |x0 − y0 + L−n (u − v)|]/ξ
n→∞
|ui |<Ln /2 |vi |<Ln /2
Z Z
n |x
= A lim L na
du dve−L 0 −y0 +u−v|/ξ
n→∞  
n(a−d)
≤ AC lim L = 0 ⇒ contradiction
n→∞

b) Now, if x0 6= y0 this is ≤ AC lim Ln(a−d−1) = 0. 

n
Remark. Also, if hφ(x0 )φ(y0 )iH∗ 6= 0 for |x0 − y0 | > 1 get above ≤ lim Lna e−L = 0.
Thus, if a < d, both H and H∗ are critical.
If a = d our block spins are
d
X
φL (x) = L− 2 φ(Lx + y)
i.e. we normalize like independent random variables. The calculation above just showed
that if φ(x)’s are (exponentially) weakly dependent then φLn (x) become independent as
n → ∞ (central limit theorem).
Thus a < d is the interesting case.
[By the way, it can be proven quite generally that if hφ(x)φ(y)i decays as |x − y|−a , a > d
then it decays exponentially].
Identical argument shows that it can not be the case that H ∈ Ms hφ(x)φ(y)iH ≤
0
A|x − y|−a with a0 > a. Thus it is reasonable to expect:
Summary. Ms is a critical surface consisting of H ∈ K, all having the same critical
exponent for the 2-point function.

88
Next, consider
Mξ = {H|H has correlation length ξ}
For this we assume that all H ∈ K have a well defined ξ, i.e. the limit
1
lim − loghφ(0)φ(x)i = ξ −1 exists
|x|→∞ |x|

Then a) Ms = M∞

b) RL : Mξ → Mξ/L

b) follows since RL scales by L:

X
hφ(0)φ(x)iRL H = La hφ(u)φ(Lx + v)iH L−2d ∼ e−L|x|/ξ .
u,v

Thus, we have a picture of K :

R2 H0

RH0
Mξ1 /L2
H0 Mξ1 /L
Mξ 1
H∗ R2 H
RH H
Ms

RL takes :
– critical H to H∗ upon iteration
– noncritical H away from H∗ upon iteration.

13.2 Linear Analysis around Fixed Points

Let us study RL near H∗ . We assume that the usual dynamical systems ideas are appli-
cable i.e. that RL is, say, a smooth map in K6 . Thus, let L = DRH∗ be the derivative of
R at H∗ . This means that, for any H ∈ K,

R(H∗ + H) = H∗ + LH + O(2 ).

6
If K is a Banach space, the derivative of R at H is defined as usual as the linear map DRH : K → K :
1
DRH H1 = lim [R(H + H1 ) − R(H)].
→0 

89
The spectrum of L plays now a crucial rule in the analysis of R near H∗ .
In the known examples, the spectrum of L has the following structure :

a) It is discrete, consisting of, real, positive, eigenvalues λi of finite multiplicity.

b) There is a finite number of λi with λi ≥ 1.

Let Hi be an eigenvector : LHi = λi Hi . We say that :

1◦ Hi is relevant if λi > 1
2◦ Hi is marginal if λi = 1
3◦ Hi is irrelevant if λi < 1
n
Note that DRH n
∗ = D(R ◦ R ◦ R ◦ · · · ◦ R)H∗ = (DRH∗ ) = L
n
(since R(H∗ ) = H∗ ), so

Rn (H∗ + Hi ) = H∗ + λni Hi + . . .

i.e. the Hi perturbation increases (λi > 1) or decreases (λi < 1) exponentially in n or
stays constant (λi = 1). Let us interpret this. Suppose first the case where there is only
one λ > 1, with multiplicity 1, say λ1 . Then for |x − y| >> 1

hφ(Ln x)φ(Ln y)iH∗ +H1 ≈ L−na hCLn φ(x)CLn φ(y)iH∗ +H1

= L−na hφ(x)φ(y)iRn (H∗ +H) = L−na hφ(x)φ(y)iH∗ +λn1 H1 +O((λn1 )2 ) .

Take n such that λn1 = O(1). Recall that we had one direction in K where we depart
from H∗ upon iteration, namely the one parametrized by ξ. Thus, it is natural to assume
that H∗ + H1 has ξ < ∞ and H∗ + O(1)H1 has ξ = O(1). From above, then

ξ(H∗ + H1 ) = Ln ξ(H∗ + O(1)H1 )

= Ln O(1).

Write λ1 = Lα (so α > 0). Then  ∼ L−nα ⇒ ξ(H∗ + H1 ) ∼ −1/α .

Let us identify  = 0 with T = Tc (T = temperature). Thus, our Hamiltonian H∗ could be
say Ising model βc HIsing and H∗ + H1 would be βHIsing with (1 + )β ∗ = β or  ∼ T − Tc .
More generally, we could start with Ising and consider some effective long distance model
(see below for precise definitions) for its correlations. Then again  ∼ T − Tc . So then

ξ ∼ (T − Tc )−1/α

and −1/α is the so called correlation length critical exponent describing how ξ diverges
as T → Tc 7 .
|x−y|
7 1 −
Note that we got hφ(x)φ(y)iT ∼ |x−y|a e
ξ(T ) for all x, y.

90
Universality. Consider now H = H∗ + Hi , λi < 1. As before,

hCLn φ(x)CLn φ(y)iH = hφ(x)φ(y)iH∗ +λni Hi

= hφ(x)φ(y)iH∗ + O(λni )

so, if λi = L−αi , αi > 0, then, since |u−v| ∼ Ln ⇒ L−αi n ∼ |u−v|αi . Thus these directions
give subleading corrections to the decay determined by H∗ . We have universality : the
N
Y YN
leading assymptotics of h φ(L xi )iH is h φ(Ln xi )iH∗ if H ∈ Ms (as n → ∞ and xi
n

i=1 i=1
are apart from each other) and the details of H are only seen in the O(L−αi n ) corrections.
We say that all H ∈ Ms have the same critical behaviour.
Also, all the exponents αi are determined by H∗ , which appear not to depend on which
H we have (in Ms ).

14 The Gaussian Fixed Point

14.1 Definition of RG in Momentum Space
1
As we saw above, the RG drives upon iteration the Gaussian measure µC where C = −∆
(on Zd ) to the fixed point
1 dd p
Z Z Z
∗
C (x − y) = du dv eip(x−y) eip(u−v) 2 d
x, y ∈ Zd
  Rd p (2π)
i.e.
lim hCLn φ(x)CLn φ(y)iµC = C ∗ (x − y).
n→∞
Since
Z d Z 1/2 d
ipu
Y
ipµ uµ
Y 2 sin pµ /2
due = e duµ = ,
 µ=1 −1/2 µ=1
pµ
we have,
Z d  2 d
∗ 1 Y 2 sin pµ /2
ip(x−y) d p
C (x − y) = e 2
Rd p µ=1 pµ (2π)d
R XZ
(note ! this converges absolutely !) and writing R dpµ f (pµ ) = dpµ f (pµ + 2πn),
n∈Z [−π,π]
we have
dd p
Z
∗ ∗ ip(x−y)
C (x − y) = e C (p)
b
[−π,π]d (2π)d
d  2
∗
X 1 Y 2 sin pµ /2
C (p) =
b .
d
(p + 2πn)2 µ=1 pµ + 2πnµ
n∈Z

91
This is the Gaussian fixed point of the block spin RG. We should now like to perform
the linear analysis near C ∗ . We will do this in a slightly different model, where it is less
messy (everything can be done in the case of block spin with identical conclusions).

We will consider, instead of a spin model on Zd (lattice cutoff), a “spin model” on Rd

where we use a momentum space cutoff. A priori this is as good from the physics point
of view.
Thus, let dµG be the Gaussian measure on S 0 (Rd ) with covariance
χ(p) dd p
Z
G(x − y) = eip(x−y) 2
Rd p (2π)d
where d > 2, χ(p) is a cutoff, say
2
χ(p) = e−p
but anything with χ(0) = 1 and fast decay at ∞ will do (say χ ∈ S(Rd )). A priori we
know that only smeared fields ϕ(f ), f ∈ S(Rd ) are integrable w.r.t. µG and
Z
1
eiϕ(f ) dµG = e− 2 (f,Gf )
R
so that e.g. hϕ(f )ϕ(g)i = (f, Gg) = f (x)g(y)G(x − y)dxdy. However, now ϕ(x) itself
1 2
makes sense (defined as lim ϕ(f ) where lim f = δx (e.g. f (y) = (2π)−d/2 e− 2 (x−y) ) and
R →0 →0
we have ϕ(x)ϕ(y)dµG = G(x − y) well defined, ∀x, y. (Technically, µG is supported on
distributions that are C ∞ functions, see Homework).
Thus, let us consider the finite volume measures
1 −VΛ (ϕ)
dνΛ (ϕ) = e dµG (ϕ)
ZΛ
R
where ZΛ normalizes dνΛ = 1, VΛ is a function of ϕ’s localized in Λ, e.g., VΛ =
R
Λ
dxP (ϕ(x)) P e.g. a polynomial.
We now describe the RG in this setup. We need to split ϕ into “local” and “global” or
“high momentum” and “low momentum” parts as in the block spin case. We do this by
first splitting G : write, for L > 1,
1 1 1
G(p)
b = 2 χ(p) = 2 χ(Lp) + 2 (χ(p) − χ(Lp)).
p p p
Or
dd p
Z Z
ip(x−y)1 ip(x−y) 1
G(x − y) = e χ(Lp) + e (χ(p) − χ(Lp))
Rd p2 (2π)d Rd p2
dd p
Z Z
2−d ip x−y 1 1
= L e L
2
χ(p) d
+ eip(x−y) 2 (χ(p) − χ(Lp))
Rd p (2π) Rd p
x − y
= L2−d G( ) + Γ(x − y) (14.1)
L
92
Here Γ(x−y) is the “high momentum” or “fluctuation within blocks” part : χ(p)−χ(Lp) =
2 2 2
e−p − e−L p = O(p2 ) for p ≈ 0 and it gets its main contribution if |p| ∈ [ L1 , 1]. In
particular, for χ analytic (prove !) |Γ(x − y)| ≤ Ae−|x−y|/L , i.e. it has correlation length
< ∞.

Lemma 14.1 Suppose µGi , i = 1, 2 are two Gaussian measures on S 0 (Rd ) and G = G1 +G2 .
Then for all F ∈ L∞ (µG ),
Z Z
F (ϕ)dµG (ϕ) = F (ϕ1 + ϕ2 )dµG1 (ϕ1 )dµG2 (ϕ2 ).

Remark. Hence ϕ is the sum of independent Gaussians.

1
Proof 1◦ . Holds for F (ϕ) = eiϕ(f ) , f ∈ S(Rd ) : e− 2 (f,Gf ) = eiϕ(f ) dµG and
R

Z Z Z
1 1
e i(ϕ1 (f )+ϕ2 (f ))
dµG1 dµG2 = e iϕ1 (f )
dµG1 eiϕ2 (f ) dµG2 = e− 2 [(f,G1 f )+(f,G2 f )] = e− 2 (f,Gf ) .

2◦ . Suffices to check for F a cylinder-function i.e. recall that cylinder sets

C(f1 , · · · , fn , A1 , · · · , An ) = {ϕ ∈ S 0 (Rd )|ϕ(fi ) ∈ Ai , i = 1, · · · , n}

where fi ∈ S(Rd ), Ai Borel sets in R. Let χC be the characteristic function of this C.

Then our claim follows (why ?) if we prove it for g = χC for all cylinder sets C. But
ϕ(f1 ), · · · , ϕ(fn ) are Gaussian with covariance

hϕ(fi )ϕ(fj )i = (fi , Gfj ) ≡ Aij

i.e. for g : Rn → R Borel measurable,

Z Z
g(ϕ(f1 ), · · · , ϕ(fn ))dµG (ϕ) = g(x1 , · · · , xn )dµA (x) (14.2)

and similarity for ϕα (fi ) α = 1, 2 :

Z
g(ϕ1 (f1 ) + ϕ2 (f1 ), · · · , ϕ1 (fn ) + ϕ2 (fn )dµG1 dµG2
Z
= g(y1 + z1 , · · · , yn + zn )dµA1 (y)dµA2 (z). (14.3)
R
But (14.2) = (14.3) if we can show it for g(x) = ei(λ,x) for all λ ∈ Rn i.e. ei(λ,x) dµA (x) =
dµA1 (y)dµA2 (z) which is a special case of 1◦ .
R i(λ,y+z)
e 

d−2
Let us apply the Lemma to (14.1). We note that the covariance of L− 2 ϕ( Lx ) is L2−d G( x−y
L
)
if ϕ has covariance G. Hence :
·
Z Z
d−2
F (ϕ)dµG (ϕ) = F (L− 2 ϕ( ) + Z)dµG (ϕ)dµΓ (Z).
L

93
This gives our RG :

F (ϕ)e−VΛ (ϕ) dµG (ϕ)

Z R
hF iνΛ = F (ϕ)dνΛ (ϕ) = R
e−VΛ (ϕ) dµG (ϕ)
− d−2
d−2
ϕ( L· )+Z)
F (L− ϕ( L· ) + Z)e−VΛ (L
R 2
2 dµG (ϕ)dµΓ (Z)
= d−2
−VΛ (L− 2 ϕ( L· )+Z)
R
e dµG (ϕ)dµΓ (Z)
= hFeiν̃

with
Z
1 −Ve (ϕ)
dν̃(ϕ) = e dµG (ϕ) Z = e−V dµG
e
Z
Z
− d−2 ·
e−V (ϕ) = e−VΛ (L 2 ϕ( L )+Z) dµΓ (Z)
e

− d−2
d−2
ϕ( L· )+Z)
F (L− ϕ( L· ) + Z)e−VΛ (L
R 2
2 dµΓ (Z)
Fe(ϕ) = − d−2
.
ϕ( L· )+Z)
e−VΛ (L
R 2
dµΓ (Z)
Hence, we got a renormalized measure ν̃ which is again of same form as ν and we can
define the RG as a map of V ’s by
Z
− d−2 ·
RL V (ϕ) = V (ϕ) = − log e−V (L 2 ϕ( L )+Z) dµΓ (Z).
e (14.4)

Compare with block-spin ! The latter can actually, with a suitable definition of ϕ on the
RHS, be written exactly as (14.4).
R
Remark. If V = Λ P (ϕ(x))dx then
·
Z
− d−2 d−2 x
V (L 2 ϕ( ) + Z) = P (L− 2 ϕ( ) + Z(x))dx
L Λ L
Z
d−2
= Ld P (L− 2 ϕ(y) + Z(Ly))dy
L−1 Λ

i.e. it depends on ϕ in L−1 Λ. So we may call Ve by VeL−1 Λ if we wish. Of course this will
not matter in the limit Λ % Zd .

We have now a fixed point V = 0, i.e. dµG → dµG under our RG. Call this the Gaussian
fixed point.

14.2 Compute DRV =0 = L

From (14.4) we get immediately
·
Z
d−2
(LV )(ϕ) = V (L− 2 ϕ( ) + Z)dµΓ (Z).
L

94
This starts to look not too bad ! We will study the spectrum of this linear map. Let us
first see some examples.
R
Example 1. Let V = Λ ϕ(x)2 dx. Then
Z Z  x 2
− d−2
LV (ϕ) = L 2 ϕ + Z(x) dxdµΓ (Z) =
L
Z  Λ 2 Z Z
x
= L2−d ϕ dx + dx Z(x)2 dµΓ
L
Z Λ Λ
 Z 
2 2
=L ϕ(x) dx + |Λ|Γ(0) since Z(x)2 dµΓ = Γ(0)
Λ/L
Z  
= L2 ϕ(x)2 + Ld−2 Γ(0) dx.
Λ/L

Hence, since, from (14.1), G(0) = L2−d G(0) + Γ(0), we get Ld−2 Γ(0) = Ld−2 G(0) − G(0),
and Z Z
2 2
L (ϕ(x) − G(0))dx = L (ϕ(x)2 − G(0))dx
Λ Λ/L
R
i.e. we found an eigenfunction (as Λ → Rd ) (ϕ(x)2 − G(0))dx with eigenvalue L2 i.e.
with exponent α = 2. Note that this corresponds to a mass term. In the lattice cutoff
case we have
1 − m2 Px∈Λ ϕ(x)2
lim e 2 dµG (ϕ) = dµGm2 (ϕ)
Λ→Rd ZΛ
Z
2 1
Gm2 (x − y) = eip(x−y) e−p .
µ(p) + m2
(Can you figure out what happens in the momentum space cutoff case?).This is a relevant
eigenvector.
R  2  
x
Example 2. Consider V (ϕ) = Λ ∇ϕ(x) dx. In the same way, using ∇ ϕ( L ) =
R  2
−1 x
L (∇ϕ)( L ) we get : LV = Λ/L ∇ϕ(x) dx − |Λ|∇2 Γ(0) and
Z h 2 i  Z  2 
2 2
L ∇ϕ(x) + ∇ G(0) dx = ∇ϕ(x) + ∇ G(0) dx
Λ Λ/L

so this is a marginal eigenvector : note that it corresponds to a change of the fixed point
R
which is corresponding to H = (∇ϕ)2 : in fact we have a one-parameter family of fixed
points dµαG α > 0.
R R
Example 3. Consider V (ϕ) = Λ ϕ(x)4 dx. Now (using dµΓ Z 2m+1 = 0)
Z   x 4  x 2 
2(2−d) 2−d 2 4
dµΓ (Z) L ϕ + 6L ϕ Z(x) + Z(x)
L L
 x 4  x 2
= L2(2−d) ϕ + 6L2−d ϕ Γ(0) + 3Γ(0)2
L L
95
so Z Z
4−d 4 2
LV = L ϕ(x) dx + 6L Γ(0) ϕ(x)2 dx + 3Γ(0)2 |Λ|
Λ/L Λ/L

so, putting Z
V4 = (ϕ(x)4 − 6G(0)ϕ(x)2 + 3G(0)2 )dx
Λ
4−d
we get LV4 = L V4 , i.e. V4 is relevant if d < 4, marginal if d = 4 and irrelevant if d > 4.
It is obvious now that we have eigenvectors Vn , polynomial of degree n, with eigenvalues
2−d
Ld Ln 2 . A simple way to get the formula for these eigenvectors is to consider
2−d R
Z
ϕ(f ) L 2 x
ϕ( L )f (x)dx ˜ 1
Le =e dµΓ eZ(f ) = eϕ(f )+ 2 (f,Γf )

2−d 2+d
where f˜(x) = L 2 Ld f (Lx) = L 2 f (Lx). Since
Z Z x − y 
(f˜, Gf˜) = L2+d
f (Lx)G(x − y)f (Ly)dxdy = L2−d
f (x)G f (y)dxdy
L
we have, see (14.1), (f, Γf ) = (f, Gf ) − (f˜, Gf˜) and so
1 1 ˜ ˜ ˜
Le− 2 (f,Gf )+ϕ(f ) = e− 2 (f ,Gf )+ϕ(f ) .

This is a kind of generating functional of eigenvectors. Let

dn  1 2
Hn (t, a) = n  e− 2 aλ +λt
dλ λ=0
be the nth Hermite polynomial (polynomial of degree n in t, homogeneous of degree n in
√
t, a). Then, let
  dn  1 2
Vn (ϕ, f ) = Hn ϕ(f ), (f, Gf ) = n  e− 2 (f,Gf )λ +λϕ(f )
dλ λ=0
so,
LVn (ϕ, f ) = Vn (ϕ, f˜).
Taking the limit ϕ(f ) → ϕ(x) (i.e. letting f tend to a delta function), (f, Gf ) → G(x −
x) = G(0), Vn (ϕ, f ) becomes a polynomial Pn (ϕ(x) in ϕ(x), Pn (ϕ(x) = Hn (ϕ(x), G(0)) =
ϕ(x)n + a1 ϕ(x)n−2 G(0) + · · · and
Z   Z  
d+n 2−d
L Pn ϕ(x) dx = L 2 Pn ϕ(x) dx.
Λ Λ/L

where Pn are Hermite polynomials.

A local V is an integral of a polynomial in ϕ(x) and its derivatives. Clearly we get

eigenvectors
Z ` 
Y nk
n k
ϕ(x) ∇ ϕ(x) dx + lower order polynomial
Λ k=1

96
with eigenvalue
2−d P P
Ld L 2
(n+ nk )
L−( knk )

i.e. the more there are ϕ’s or derivatives, the more V is irrelevant. In the class of even
local V ’s the relevant ones are, for d ≥ 3 :
R 2
ϕ L2 always
R 4
ϕ L4−d d < 4

the marginal ones are :

R
(∇ϕ)2 always
R 6
ϕ d=3
R 4
ϕ d=4
R d+2
In case of odd V , we have ϕ with eigenvalue L 2 always relevant (this corresponds
R 6−d R
to a magnetic field), ϕ3 with eigenvalue L 2 relevant for d < 6, ϕ5 with eigenvalue
3d R R
L5− 2 , relevant 3d < 10. (note that ∇ϕ, ∇ϕϕ2 are boundary terms).

14.3 The space K for L

We can set up a space K so that L : K → K and study the spectrum of L.
Basically, we want to consider V ’s of the form
Z n
Y
dx1 · · · dxn K(x1 , · · · , xn ) ϕ(xi )
i=1

and specify the K’s. For this, generalize a bit what we found above. Define the normal
ordered product :
n
Y X |I| Y Y
: ϕ(xi ) := (−1) 2 ϕ(xj )h ϕ(xi )i.
i=1 I⊂{1,··· ,n} j6∈I i∈I

Homework.
Z n n n
Y Y Y ∂  h 1 i
: ϕ(xi ) : f (xi )dxi = exp − (f, Gf ) + ϕ(f )
∂λi λi =0 2

i=1 i=1 i=1
n
X
f= λi fi , fi ∈ S(Rd ).
i=1

This implies :
n n x 
2−d
Y Y i
n
L: ϕ(xi ) := L 2 : ϕ :.
i=1 i=1
L

97
The space Kγ consists of finite sums of the form. (Let ϕ ∈ S(Rd ))

XZ N
Y N
Y
ni mi
V (ϕ) = K∼
n m (x · · · xN ) :
∼ 1
ϕ(xi ) ∇ϕ(xi ) : dxi (14.5)
n m
∼ ∼ i=1 i=1

d
Y
where n∼ = (n1 , · · · , nN ), m d
∼ = (m1 , · · · , mN ), ni ∈ N, mi ∈ N , (∇ϕ)
mi
= (∇α ϕ)miα ,
α=1
d
mi = (mi1 , · · · , mid ) and K∼n∼m(x1 · · · xN ) = K∼ n∼m(x1 + x, x2 + x, · · · , xN + x) ∀x ∈ R ,
(translation invariance) and
Z
γL(0,x2 ,··· ,xN )
|K∼ m(0, x2 , · · · , xN )|e
n∼ ≡ kK∼
n∼mkγ < ∞

where L(x1 , · · · , xN ) ≡ d(x∼) is the length of the shortest connected path Γ in Rd such
that all xi belong to Γ; here γ > 0.
Then LV has kernels K 0 :
2−d PN P
Kn0 m(x∼) = L 2 i=1 (ni +mi ) L− mi
LN d K∼m(Lx
n∼ ∼)
∼∼

and
Z
2−d PN
− N 0
P
kKn0 mkγ 0 = L 2 i=1 (ni +mi ) L i=1 mi L Nd
|K∼m(Lx
n∼ ∼)|e
γ d(x
∼) dx2 · · · dxn
∼∼
Z
2−d PN
− N 0 −1 x )
P
= L 2 i=1 (ni +mi ) L i=1 mi L d
|K∼
n∼m(x
∼
)|eγ d(L ∼ dx2 · · · dxn .

But d(L−1 x∼) = L−1 d(x∼), so

2−d PN PN
kKn0 mkLγ = Ld L 2 i=1 (ni +mi ) L− i=1 mi kK∼mkγ
n∼
∼∼

Hence LV ∈ KLγ and thus also LV ∈ Kγ .

Thus only (for V even)
Z Z Z
K(x − y)ϕ(x)ϕ(y), Kαβ (x − y)∇α ϕ(x)∇β ϕ(y), K(x − y)ϕ(x)∆ϕ(y)
Z Z
K(x1 · · · x4 )ϕ(x1 ) · · · ϕ(x4 ), K(x − y)ϕ(x)2 ϕ(y)2
Z Z
K(x1 x2 x3 )ϕ(x1 )ϕ(x2 )ϕ(x3 ) , K(x − y)ϕ(x)ϕ(y)3
2
(14.6)
R R R
are not irrelevant in addition to ϕ2 , ϕ4 , (∇ϕ)2 . These can be split into relevant/margi-
nal and irrelevant as follows. Consider e.g. the first term in (14.6). Let K(p)
b be the Fourier
R γ|x|
transform of K. Since |K(x)|e dx < ∞ we get K(p) b is analytic in |Imp| < γ. For
simplicity we consider only rotation invariant V ’s i.e. K(x) = K(|x|).

98
 R
Lemma 14.2 Let K(x) be rotation invariant with kKkγ = |K(x)|eγ|x| dx < ∞. Then
for all ϕ ∈ S(Rd ))
Z Z
K(x − y)ϕ(x)ϕ(y)dxdy = α ϕ(x)2 + β(∇ϕ(x))2 + Ve (ϕ)

where |α|, |β| ≤ CkKkγ and Ve ∈ Kγ/2 is irrelevant with kernel ≤ CkKkγ .
R
Proof. Since |K(x)|eγ|x| dx < ∞ we get
Z Z
|K(p)| = | e K(x)dx| ≤ e|Imp||x| |K(x)|dx
b ipx

is analytic in |Imp| < γ. Taylor expand

K(p)
b = K(0)
b + βp2 + R(p)
b (14.7)

(we used the rotation symmetry to get the second term of the form βp2 ) where

1 1 d3 b
Z
R(p) =
b dt(1 − t)2 3 K(tp)
2 dt
X0
= pα pβ pγ H
b αβγ (p)
αβγ

we have
|α| = |K(0)|
b ≤ kKkγ
and using Cauchy’s theorem

|β| = |∂ 2 K(0)|
b ≤ γ −2 kKkγ

It remains to study the remainder. We have

d3 b X
K(tp) = pα pβ pγ (∂α ∂β ∂γ K)(tp)
b
dt3 αβγ

and so
Z 1
H
b αβγ (p) = dt(1 − t)2 (∂α ∂β ∂γ K)(tp)
b
0
Z 1
3
= i dt(1 − t)2 (xα x\
β xγ K(x))(tp).
0

Taking inverse Fourier transform and changing variables we get

Z 1 Z
3 2 −d
Hαβγ (x) = i dt(1 − t) t eipx/t (xα x\
β xγ K(x))(p)dp
0
Z 1
= i3
dt(1 − t)2 t−d−3 xα xβ xγ K(x/t).
0

99
Hence
Z 1 Z
−d
kHαβγ kγ/2 ≤ dtt dx(|x|/t)3 |K(x/t)|eγ|x|/2
Z0 1 Z
= dt dx|x|3 |K(x)|eγt|x|/2 .
0

48 γ|x|/2
Since |x|3 ≤ γ3
e we get finally

48
kHαβγ kγ/2 ≤ kKkγ .
γ3
Thus Z
3
Ṽ = i Hαβγ (x − y)∂α φ(x)∂α ∂β φ(y)dxdy

is irrelevant and satisfies the claim. 

The other terms can be analyzed similarly, e.g.
Z Y Z
K(x1 , · · · , x4 ) ϕ(xi ) = a ϕ4 + Ve

R
where a = K(0, x2 , x3 , x4 )dx2 dx3 dx4 and Ve is irrelevant. Thus every V ∈ Kγ can be
written as
Z Z Z
2 2
V =r : ϕ : +z : (∇ϕ) : +λ : ϕ4 : +Ve , Ve ∈ Kγ/2

and then Z Z Z
LV = L r 2 2
: ϕ : +z 2
: (∇ϕ) : +L 4−d
λ : ϕ4 : +Ve 0

with Ve 0 ∈ KLγ/2 , kVe 0 kLγ/2 ≤ L−2 kVe k. These are our “coordinates” in K.

15 Perturbative analysis of RV
15.1 General formalism
Consider now the full R:
·
Z h i
2−d
RV (ϕ) = − log dµΓ (Z) exp −V (L ϕ( ) + Z) .
2 (15.1)
L
Suppose V is in K as above. There are several problems :

1. ϕ in (15.1) lives on Rd and is not in S(Rd ), rather it is in S 0 ∩C ∞ i.e. it may increase

as x → ∞. This is the IR-problem. We used Λ for that, so we may replace the
kernels K∼n m∼
above by K∼ n m
∼
χ where χ restricts xi ∈ Λ.

100
 2. More serious problem is the convergence of (15.1) as Z becomes large. This is the
stability problem. Say for V = λϕ4 we need at least Re λ > 0. For a V as in the
previous section, this positivity property is not easy to state. Instead of trying to
address the stability problem, we will calculate RV perturbatively in λ if V = λϕ4 .
With a lot of work, RV can actually be analyzed rigorously. One does the following :

a) For |ϕ(x)| < C one uses the representation in terms of K∼

n m
∼
, actually an
infinite series.
b) For |ϕ(x)| > C one uses another representation where positivity is explicit.

The problem is to combine a) and b) : ϕ(x) can be small and ϕ(y) large nearby.
One uses expansion ideas from statistical mechanics to decouple these problems.
R
Thus, let us start with V (ϕ) = λ Λ ϕ(x)4 dx and see what RV looks like. Expand in
powers of λ:
X N
(RV )(ϕ) = λn (RV )n (ϕ) + Remainder
n=0
1 dn  1 dn   Z 2−d 
−V (L 2 ϕ( L· )+Z)
(RV )n (ϕ) =  (RV )(ϕ) = − log e dµΓ (Z) .
n! dλn λ=0 n! dλn λ=0

Hence,
Z Z  x 4
2−d
λ(RV )1 (ϕ) = λ L ϕ 2 + Z(x) dxdµΓ = LV (ϕ)
Λ L
Z Z Z
2 1h 2
i
λ (RV )2 (ϕ) = − V dµΓ − V dµΓ V dµΓ
2
1 2−d · 2−d ·
= − hV (L 2 ϕ( ) + Z); V (L 2 ϕ( ) + Z)iΓ
2 L L
where we use the notation of truncated expectations.
In general
1
λn (RV)n (ϕ) = (−)n−1 hV ; V ; · · · ; V iΓ .
n!
2−d
Consider (RV)2 : Denote L 2 ϕ( Lx ) ≡ Ψ(x). So
D 4  4 E D 4 E D 4 E
Ψ(x) + Z(x) Ψ(y) + Z(y) − Ψ(x) + Z(x) Ψ(y) + Z(y)
µ µ µ
 
2 2 2 3 3
= 36 · 2 · Ψ(x) Ψ(y) Γ(x − y) + 16 · 3 Ψ(x)Ψ(y) + Ψ(x) Ψ(y) Γ(x − y)Γ(0)
!2
4
+4 ! Γ(x − y)4 + · 2Γ(x − y)2 Γ(0)2
2
+16 · 3 ! Ψ(x)Ψ(y)Γ(x − y)3 + 6 · 4 · 3(Ψ(x)2 + Ψ(y)2 )Γ(x − y)2 Γ(0)
+16Ψ(x)3 Ψ(y)3 Γ(x − y) + 16 · 3 · 3Ψ(x)Ψ(y)Γ(x − y)Γ(0)2

or graphically

101
 + + + +

+ + +

where we have same rules as before except that :

- edges are Γ(x − y)
2−d
- external edges have Ψ(x) = L 2 ϕ( Lx ).
Thus we end up with the representation
X
(RV )n (ϕ) = WG (ϕ)
G

where the sum runs through all connected graphs G with n + m vertices x1 ,. . . , xn , y1 ,. . . ,
ym . Vertices yi have four edges ({z, z} allowed) attached and vertices xj have 4 − nj edges
attached where nj > 0. The WG (ϕ) is equals
 
Z Z Y m
Y Yn Z n
Y
0 n
n(G)  Γ(z − z ) dyj  Ψ(xi ) dxi ≡ KG (x1 , . . . , xn )
i
Ψ(xi )ni dxi
{z,z 0 }∈e(G) i=1 i=1 i=1

Recalling that |Γ(x − y)| ≤ Ce−α|x−y| , for some α and the fact that the graph is connected
we get
Exercise. kKG kγ < ∞, for some γ > 0.
N
X
Thus λn (RV )n (ϕ) ∈ Kγ for some γ. It is more convenient to consider V (ϕ) =
R n=0
λ Λ
: ϕ(x)4 : dx where, recall, : ϕ4 (x) := ϕ(x)4 − 6ϕ(x)2 G(0) + 3G(0)2 . Then
Z
4−d
λ (RV)1 (ϕ) = LV = λL : ϕ(x)4 :
Λ/L

Exercise. Prove this !

Thus, e.g. (RV)2 = + + +

From our analysis in the previous section, we get (for d = 4)

Z Z
2 3 4 2 3
(RV )(ϕ) = (λ + aλ + O(λ )) : ϕ(x) : +(bλ + O(λ )) : ϕ(x)2 :
Λ/L Λ/L
Z
+(cλ2 + O(λ3 )) (∇ϕ(x))2 + Ve
Λ/L

102
where Ve is irrelevant i.e. LVe contracts and Ve = O(λ2 ). Here a comes from

, a = − 72 Γ(x)2 dx+ 16.9 Γ(x−y)L2−d C( Lx − Ly ) (see computation

R R
and 2 2

for d = 4 below); b, c from and and

(see computation for d = 4 below).

15.2 Iteration
R R R
Suppose now V = λ : ϕ4 : +r : ϕ2 : +z : (∇ϕ)2 : +Ve where Ve ∈ K is irrelevant,
kLVe kγ ≤ L−α kVe kγ for α > 0. Calculate perturvatively RV:
N
X 1
(RV )(ϕ) = (−1)n+1 ( )hV ; V ; · · · ; V iΓ + Remainder
n=0
n!
(15.2)
d−2
where V = V (L− 2 ϕ( L· ) + Z)
N
X
≡ (RV)n (ϕ) + Remainder
n=0

The point is now that RVn ∈ Kγ again. Indeed, we get graphs as before and also from
r, z and Ve . Pictorially, the Ve ones are

where are K∼n m

∼
and lines are Γ(x − y).
It is not hard to show that the exponential decay of K∼ n m∼
and Γ give rise to exponential
decay of these graphs. Thus the claim.
DR 4
Y 2 E
Example E.g. K(x1 , · · · , x4 ) ϕ(xi ) gives for instance
i=1

103
 4
Z Y 4
x  Z Y
2−d i
= dxi L 2 ϕ dyi K(x1 , x2 , y1 , y2 )Γ(y1 − y3 )Γ(y2 − y4 )K(y3 , y4 , x3 , x4 )
i=1
L i=1
Z Y
≡ e 1 , · · · , x4 )
K(x ϕ(xi )dxi

where
4
Z Y
e ) = L4−2d L4d
K(x dyi K(Lx1 , Lx2 , y1 , y2 )Γ(y1 − y3 )Γ(y2 − y4 )K(y3 , y4 , Lx3 , Lx4 )
∼
i=1

and
Z
eγL(0,x2 ,x3 ,x4 ) |K(0,
e · · · , x4 )|dx2 · · · dx4
Z
−3d −1
=L e L−1 x2 , L−1 x3 , L−1 x4 )|dx2 · · · dx4
eL γL(0,x2 ,x3 ,x4 ) |K(0, (15.3)

Use L(0, x2 , x3 , x4 ) ≤ L(0, x2 , y1 , y2 ) + |y1 − y3 | + |y2 − y4 | + L(y3 , y4 , x3 , x4 ) (since the RHS

is the length of a particular path Γ such that 0, x2 , x3 , x4 ∈ Γ).
So,
4
Z Y 4
Y
4−d
(15.3) ≤ L dyi dxi K̄(0, x2 , y1 y2 )Γ̄(y1 − y3 )Γ̄(y2 − y4 )K̄(y3 , · · · , x4 )
i=1 i=2

−1 −1
with K̄ = eL γL K, Γ̄(x) = eL γ|x| Γ(x).
Use |Γ̄(y2 − y4 )| ≤ C, integrate over y4 , x3 , x4 :
3
Z Y
4−d
(15.3) ≤ L dyi dx2 |K̄(0, x2 , y1 , y2 )Γ̄(y1 − y3 )|kKkL−1 γ .
i=1

|Γ̄(y1 − y3 )|dy3 < C 0 and then the rest :

R
Integrate y3 :

(15.3) ≤ CL4−d kKk2L−1 γ ≤ CL4−d kKk2γ .

Summary. Perturbatively, the Rn V retains the form :

Z Z Z
n 2 2
Vn ≡ R V (ϕ) = zn : (∇ϕ) : +rn : ϕ : +λn : ϕ4 : +Ven
L−n Λ L−n Λ L−n Λ

104
where Ven ∈ K and kLVen k ≤ L−α kVn k, α > 0.
We get the recursion (denote gn = (zn , rn , Ven ))

zn+1 = zn + O(Vn2 )
rn+1 = L2 rn + O(Vn2 )
λn+1 = L4−d λn + aλ2n + O(λ3n , λn gn , gn2 )
Ven+1 = LVen + O(Vn2 ).

Let us see how these behave as n → ∞.

15.3 d>4
We want to determine the critical point, i.e. take
Z Z Z
V = z : (∇φ) : +r : ϕ : +λ : ϕ4 :
2 2

and find r(λ) such that V ∈ Ms . Since 4 − d < 0 we expect the relevant fixed point to
R
be the Gaussian V = z̃ : (∇ϕ)2 :, for some constant z̃ 6= z, noted z∞ below.
Hence, the problem is to find r0 = r(λ) such that

rn , λn , Ven → 0 as n → ∞.

Since rn+1 = L2 rn + · · · tends to increase, we need to be careful. Let us construct

r0 inductively. Given λ (small), take r0 ∈ [−Aλ2 , Aλ2 ] ≡ I0 , A chosen below. Then
|r1 − L2 r0 | ≤ Cλ2 + C(A)λ3 where C is A-independent. Consider r1 as a function of r0 .
For A large enough, r1 maps our interval
L2 2 L2 2
 
r1 (I0 ) ⊃ − Aλ , Aλ
2 2
i.e., since λ1 ≤ L4−d λ + O(λ2 ) < λ, we can, by continuity, find an interval

I1 ⊂ I0

such that
r1 (I1 ) = [−Aλ21 , Aλ21 ].
Now, keep on iterating. We find intervals In ⊂ In−1 ⊂ · · · I0 such that rn as a function of
r0 satisfies
rn (In ) = [−Aλ2n , Aλ2n ]
and λn ≤ L(4−d−)n λ, for  > 0, kVen k → 0.
∞
\ ∞
\
Thus, In 6= ∅ (since In closed ⊂ In−1 ) and, for r0 ∈ In , we have
n=0 n=0

Ven , λn , rn → 0 as n → ∞.

105
 ∞
\
Actually it is readily seen that |In | n→∞
−→ 0 so In = one point ≡ r(λ), the critical r value.
n=0
Clearly r(λ) = O(λ2 ).
n−1
X
How about zn ? Since zn+1 = zn + O(λ2n ) and z0 = 0 we have zn = O(λ2m ) =
m=0
n−1
X
O((L(4−d)m λ)2 ) n→∞
−→ z∞ = O(λ2 ). So
m=0
Z
n
−→ z∞
R V n→∞ : (∇ϕ)2 :

2
R
Hence, we have the one-parameter family of Gaussian fixed points e−z : (∇ϕ) : dµG and
2 4 −z∞ : (∇ϕ)2 :
R R R
e−r(λ) : ϕ : −λ : ϕ : dµG −→
R n e dµG .
The Gaussian line has one unstable direction : ϕ2 : and, once that is fixed to the critical
value, we flow to the line :
r
z

Ms
(r(λ), λ)

15.4 d=4
Since λn+1 = λn to leading order we need to go to O(λ2 ). The sign of a in λn+1 =
λn + aλ2n + · · · is very important so let us calculate it.
We will prove that λn → 0 and rn , zn , Ven = O(λ2n ). Thus a gets contribution from the
term Z
1 D  2−d  x  4  2−d  y  4 E
− : L ϕ2 + Z(x) : : L ϕ 2 + Z(y) : dxdy.
2 L L
This gives graph a) which contributes directly and b) which
contributes indirectly.

106
a) Equals
Z
72 4−2d  x 2  y 2
− L dxdyϕ ϕ Γ(x − y)2 dxdy
2 L L
Z  x 4 Z
72 4−2d 
= − L dxϕ dyΓ(y)2 + irrelevant (goes into Ve ).
2 L

b) We have to normal order since Ven+1 is normal ordered. This graph equals

Z
1 2−d
 x 3  y 3
− 42 L−( 2 )6 :ϕ : Γ(x − y) : ϕ :.
2 L L
Use
 x 3 y 3  x 3  y 3
:ϕ ::ϕ :=: ϕ ϕ :
L L L L
 x − y   x 2  y 2
+ 9G :ϕ ϕ : + quadratic term + constant
L L L
and write
x − y 
Z  x 2  y 2
G Γ(x − y) : ϕ ϕ : dxdy
L L L
Z  x 4 Z y
= dx : ϕ : dy G Γ(y) + irrelevant.
L L
Thus a) and b) together give
Z Z  y  Z
4−d 4 2 2−d
−36L dx : ϕ(x) : Γ(y) + 2L G Γ(y) = −b(L) : ϕ4 : dx
L

(using L4−d = 1, Γ(y)2 + 2L2−d G( Ly )Γ(y) = G(y)2 − (L2−d G( Ly ))2 ), where

Z h y i
b(L) = 36 dy G(y)2 − (L−2 G )2 . (15.4)
L
Therefore our recursion now is

zn+1 = zn + O(λ2n )
rn+1 = L2 rn + O(λ2n )
λn+1 = λn − b(L)λ2n + O(λ3n ) (15.5)
Ven+1 = LVen + O(λ2n )
kLVen k ≤ L−2 kVen k

How does (15.5) behave? Consider first the differential equation

dλ
= −βλ2
dn
107
so
dλ 1 1 λ
2
= −βdn, = + βn, λn = →0
λ λn λ 1 + nβλ
1
i.e. for large n, λn ∼ nβ
.
Homework. Prove that (15.5) ⇒ λn < nc .
Actually there is some interest in calculating b(L). We have
Z Z log L
d −2s
b(L) = −36 4
dy ds (e G(e−s ~y ))2
0 ds
but
d  −2s 2 
~ y ) e−2s G(e−s ~y )
2
e G(e−s ~y ) = (−4 − ~y · ∇
ds
where we used the fact that G(~x ) depends on |~x| : G(~x) = G(r)
e r = |~x| and

d e −s e0 (e−s r) = −r d G(e ~ x G(e−s~x ).

G(e r) = −e−s rG e −s r) = −~x · ∇
ds dr
R
Thus, (with A = S 3 dΩ)
Z log L Z ∞ Z
d  
b(L) = +36 ds dr r 4 + r (e−2s G(e−s r))2 3
dΩ
0 0 dr S3
Z log L Z ∞
d 4 −2s
 
= 36A ds dr r (e G(e−s r))2
0 0 dr
Z log L  2
4 −2s −s
= 36A ds lim r e G(e r) .
0 r→∞

But
d4 p eip1 r −p2
Z
G(r) = e (15.6)
(2π)4 p2
√
Z 4 ip1 Z 3 − p2
dp e 2 2 dp e
= r−2 4 2
e−p /r r→∞
−→ r−2 p .
(2π) p (2π)3 2 p2

So 2 2
Z ∞ 
4

−2s −s
2 dρ −ρ 1
lim r e G(e r) = ρe = .
r→∞ 0 2π 2 2π 2
Hence
b(L) = 9Aπ −4 log L ≡ β2 log L
A = area of unit sphere in R4 .
2
Remark. β2 is universal, it did not depend on our cutoff e−p . We could have used any
cutoff χ(p) ! This is a deep fact, see below.

108
Summary. For d = 4, again we find a r0 (λ) such that
Z Z
R (r0 : ϕ : +λ : ϕ4 :) = Vn
n 2

has λn → 0 like n1 , rn = O(λ2n ) = O( n12 ) → 0 Ven = O(λ2n ) → 0 and

∞
X
zn → z∞ = O(λ2n ) < ∞
n=0
P 1
since n2
converges. We go to the Gaussian fixed point, but only logarithmically in the
scale : n ∼ log Ln .

15.5 d<4
Now λn+1 = L4−d λn + O(λ2n ) increases, so sooner or later λn is big enough so we can not
trust our perturbative analysis that holds only if λn is small enough.
It is very instructive to pretend that d is a continuous variable and take d = 4 − , 
small. Then
λn+1 = L λn − b(L)λ2n + O(λ3n ).
Let us look for a fixed point λ∗ = a + O(2 ).  We get 1 = L − b(L)a + O(2 ) or
a = log L
= β2−1 above (we used b(L) = b(L) +O() = β2 log L + O()). Thus a
 
b(L)
d=4− d=4
fixed point would be
λ∗ = β2−1  + O(2 ).
However, now zn+1 = zn + O(λ2n ) = zn + O(λ∗2 ) increases. What went wrong ? We have
indeed a new fixed point λ∗ 6= 0 and we need a different scaling of ϕ in our RG :
d−2 d−2+η
L− 2 → L− 2 .

Let us proceed formally. Now

1 −1 ϕ)−V 1
(∇ϕ)2 −V
R
e−H = e− 2 (ϕ,G ≈ e− 2
η
−η 1 (∇ϕ)2 −RV (L− 2 ϕ)
R
→ e−L 2

1 −η +O(λ2 )) (∇ϕ)2 −r1 :ϕ2 :−λ1 :ϕ4 :−Ve

R R R
= e− 2 (L .

Thus, 1 + z0 = 1 (z0 = 0)

1 + zn+1 = L−η (1 + zn ) + O(λ2n )

⇒ η = O(2 ) and

λn+1 = L−2η λn − β2 log Lλ2n + O(λ3n )

= L λn + O(2 )λn − β2 log Lλ2n + O(λ3n )

109
so λ∗ is still λ∗ = β2−1  + O(2 ).
We got also rn+1 = L2−η rn + O(λ2n ), kLVen k ≤ L−2−η kVen k.
Summary d = 4 − 
Now we have a new fixed point
Z Z Z
∗ 1 ∗ ∗
H = 2
(∇ϕ) + r 2
: ϕ : +λ : ϕ4 : +Ve ∗
2
with λ∗ = β2  + O(2 ), r∗ = O(2 ), Ve ∗ = O(2 ).
Also D E 1
ϕ(x)ϕ(y) ∼
H∗ |x − y|d−2+η
η = O(2 ).
Remark d = 3 is  = 1 i.e. is non-perturbative. We expect, say for block-spin RG, that
for all λ there is r(λ) such that H = (ϕ(x) − ϕ(y))2 + r(λ) ϕ(x)2 + λ ϕ(x)4 is
P P P

critical and ∃η such that

Rn H → H∗
d−2+η
where Rn has Cϕ(y) = L+ 2 L1d ϕ(Lx + y) (d = 3) and η, H∗ are λ-independent. H∗
P

has only one relevant direction, namely the correlation length and no marginal ones (for
small , one sees that the marginal (∇ϕ)2 becomes irrelevant).

16 The Continuum Limit

16.1 Effective Field Theories
We finally consider the QFT problem from the RG point of view. Suppose we have a
scalar QFT i.e. a measure ν on S 0 (Rd ). Let σin be L−n cube centered at L−n i, i ∈ Zd .
Suppose L+nd σn ϕ(x)dx ≡ ϕn (L−n i) are well defined random variables (for the free field
R
i R R
they are, the covariance is just L2dn σn dx σn dyG(x − y)), i.e. given a f ∈ s(Zd ) suppose
i j
Z P
−n −nd
Sn (f ) = ei fj ϕn (L j)L dν(ϕ)
R
is a positive definite function (this means that S(f ) = eiϕ(f ) dν extends from f ∈ S(Rd )
fj χσjn , χ characteristic function). Then Sn defines a measure on s0 (L−n Zd ), call
P
to f =
it νn which is the joint distribution of the variables ϕn (L−n j). We call νn the effective
QFT on scale L−n corresponding to ν. Thus, the correlation functions of the effective
theory are just averages of those of ν over cubes σin :
N
DY E N Z
Y N
DY E
−n N dn
ϕn (L iα ) =L dxiα ϕ(xiα ) . (16.1)
νn σinα ν
α=1 α=1 α=1

110
It is convenient to relate all of these effective theories to a fixed lattice Zd . For this, define
a Zd field φn (i) = L−an ϕn (L−n i) where a is a parameter to be fixed later, and let µn be
the distribution of φn , i.e.
DY E DY E
−an −n
φn (iα ) = L ϕn (L iα ) . (16.2)
µn νn
α α

How are the different φn ’s related ? Easy :

Z
−a(n−1) −n+1 −a(n−1) (n−1)d
φn−1 (i) = L ϕn−1 (L i) = L L ϕ(x)dx
σin−1
X Z X
= L−a(n−1) L−d L nd
ϕ(x)dx = La−d φn (Li + j)
n
σLi+j
|jα |<L/2 j

= (Cφn )(i) (16.3)

where C is the block spin operator. Thus

µn−1 = C ∗ µn .

Suppose µn are given by Hamiltonians Hn . Then

Hn−m = Rm Hn ∀n, m.

Definition 16.1 Let RH∗ = H∗ . The unstable manifold Mu of H∗ is

n  o
Mu = H∃ Hn n→∞ −→ H∗ , Rn Hn = H .


Note that R : Mu → Mu . Thus, we are led to the conjecture :

There exists a ∈ R and H∗ such that RH∗ = H∗ and Hn n→∞ −→ H∗ . Then all Hn are
on the unstable manifold of H∗ . Hence we got a nice picture: critical phenomena are on
MS , QFT on Mu !

1
Example ν = Gaussian measure, covariance G(p) b = p2 +m 2 (free field, mass m). Then

νn has covariance
Z  Z 2 dd p
−n −n ipL−n (i−j) b  nd
Gn (L i, L j) = e G(p)L eipx dx .

R d n
σ0 (2π)d
 Z 2
 nd
Let ρn (p) = L eipx dx and, for a = d−2 , µn has covariance Cn :

2
σ0n

dd p
Z
(2−d)n −n 1
Cn (i, j) = L eipL (i−j) 2 ρn (p)
p + m2 (2π)d
dd p
Z
ip(i−j) 1 n
= e ρn (L p)
p2 + L−2n m2 (2π)d

111
where
d 
2 sin pµ/2 2
Z 2 Y 
n ipx
ρn (L p) =  e dx = ≡ ρ(p)
 
σ00 µ=1
p µ

is n-independent. Here Cn are covariances that parametrize the Gaussian part of the
unstable manifold of the Gaussian fixed point Cn=∞ of the block spin RG. They have
correlation length Ln m−1 .

Suppose now that we have a RG R : K → K and a fixed point H∗ . How would we

get a QFT from these data ?
Easiest way is the scaling limit : For xi 6= xj i 6= j, consider the limit
N
DY E
N an n
lim L φ(L xi ) ≡ lim G(n) (x1 , · · · , xN )
n→∞ H∗ n→∞
i=1
≡ G(x1 , · · · , xN ). (16.4)

(to be more precise, suppose xi ∈ (L−M Z)d for some M ).

We expect this limit to exist. We expect that for |yi − yi | large ∀i 6= j, we have for
Q
C(y1 , · · · , yN ) ≡ h φ(yi )iH∗ :
 
C(y1 + z, y2 , · · · , yN ) − C(y1 , y2 , · · · , yN )
 
 
≈ (min |y1 − yi |)− C(y1 · · · yN )
 
i>1

 > 0, if |z| < O(1) (say L). Then

N
DY E N
DY E
n
φ(L xi ) = n
φav (L xi ) (1 + O(L−n ))
H∗ H∗
i=1 i=1
X
where φav (Lx) = L−d φ(Lx + y). Thus,
|yα |< L
2

DY E
G(n) (x1 , · · · , xN ) = LN an φav (L(Ln−1 xi )) (1 + O(L−n ))
H∗
i

= G (n−1)
(x1 , · · · , xN )(1 + O(L−n ))

since (Cφ)(x) = La φav (Lx) and RH∗ = H∗ . Thus, one expects the limit in (16.4) to exist.
The distributions G(x1 , · · · , xN ) are correlations of a QFT (xi ∈ Rd now). What are the
effective field theories νm or µm or Hm of this QFT ?

Answer All Hm are just H∗ !

112
Proof This is just bookkeeping. By definition (16.1), (16.2),
N
DY E Z
φ(iα ) = dx1 · · · dxN χ(xα ∈ σimα )G(x1 , · · · , xN )LN (d−a)m
µm
α=1

where σimα = L−m cube at L−m iα . Write the integral as a Riemann sum and use the
definition (16.4) of G:
Z Y X N
DY E
−N dn
dxα χG = lim L χ(xα ∈ σimα )LN (d−a)m LN an φ(Ln xα ) .
n→∞ H∗
{xα ∈L−n Zd } α=1

Write xα = L−n yα . Then yα is summed over yα ∈ Zd , yα ∈ Ln−m -cube at Ln−m iα (take

n > m). Thus, using (16.3),
N
DY E DY E
n−m
φ(iα ) = lim (C φ)(iα )
µm n→∞ H∗
α=1 α
N
DY E
= φ(iα ) . 
H∗
α=1

The scaling limit is a scale-invariant, massless, field theory :

GN (λx1 , · · · , λxN ) = λ−N a G(x1 · · · xN ).

In particular (assuming translation invariance):

const
G2 (x1 , x2 ) = .
|x1 − x2 |a

16.2 Non-Scale Invariant Theories

How to get a QFT with effective theories Hn 6= H∗ ? E.g. we would like to get a massive
theory. We would like to pose this constructively : find (lattice) cutoff-L−n theories that
converge as n → ∞ to a QFT.
Example Look at free field again. Consider Ginzburg-Landau model on Zd with corre-
lation length ∼ Ln :

1 X 2 L−2n m2 X
Hn = φ(x) − φ(y) + φ(x)2
2 2 x
hxyi

i.e.

e−Hn Dφ = dµCn (φ)

Cn = (−∆ + L−2n m2 )−1

113
on Zd . We know that
 
2n( d−2 )
D
n n
E 1
L 2 φ(L x)φ(L y) −→
n→∞ (x, y).
Cn −∆Rd + m2
From RG we find
1
Rn−` Hn n→∞−→ (φ, G−1 ` φ) = H`
∗
2
dd p
Z
1
G` (x, y) = eip(x−y) 2 ρ(p) (x, y ∈ Zd )
Rd p + L−2` m2 (2π)d
which are on the unstable manifold : RH`∗ = H`−1
∗
. Pictorially,

So, in general, we should look for a 1-parameter family of Hamiltonians HLn such that
Rn−` HLn converge as n → ∞ for all m. It is easy to see that HLn must tend to MS as
n → ∞ and Rn−` HLn to Mu .
As an example we consider the ϕ4 QFT in various dimensions.

16.3 ϕ4d
We consider a momentum cutoff for simplicity and let
1 −z R :(∇ϕ)2 :−ρ R :ϕ2 :−g R :ϕ4 :
dν (ϕ) =
e Λ Λ Λ dµG .
Z
b (p) = 2 1 2 χ(p) and
R R R
Let z Λ : (∇ϕ)2 : +ρ Λ : ϕ2 : +g Λ : ϕ4 := W (ϕ); here8 G p +m
−p2
let, e.g., χ = e . We ask : are there functions of  z , ρ , g such that ν converges to a
8
We put m2 into the covariance for simplicity : we are concerned with the UV-problem, not the IR.

114
measure ν on S 0 as  → 0 ? We know that λ = 0 is a solution, and ask if a non-Gaussian
ν is possible. The effective measures are defined as above, only easier : Write, for ` > 
1 1 1  
Gb (p) = χ(p) = χ(`p) + χ(p) − χ(`p)
p2 + m 2 p2 + m 2 p2 + m2
≡ G
b` (p) + Γ
b` (p)

and define Z 
() −W (ϕ+z)
dν` (ϕ) = e dµΓ` (z) dµG` (ϕ)

which has cutoff `. To connect to Statistical Mechanics and RG, we need to scale as in
(16.2) :
(d−2)
φ (x) =  2 ϕ(x)
d−2
(we take a = 2
to anticipate what comes later). Then ν becomes µ :
2−d
2 φ( · ))
dµ (φ) = e−W (  dµC (φ) ≡ e−V (φ) dµC
1
where C
b (p) =
p2 +2 m2
χ(p), has UV-cutoff 1, and is Gaussian. (Note the IR cutoff 2 m2 ).
Here Z Z Z
2 2 2 4−d
V (φ) = z : (∇ϕ) : + ρ : ϕ : + g : ϕ4 :

note the powers of  (so called canonical dimensions of mass, coupling). Then we have
()
V` ≡ R`/ V

where RL is as before :
Z 2−d
2 φ( · )+Z)
RL V = − log e−V (L L dµΓ (Z)

1
Γ(p)
b = p2
(χ(p) − χ(Lp)). Let us put

r = 2 ρ , λ = 4−d g .

As before, we could study the iteration rn , zn , λn → rn+1 , zn+1 , λn+1 but here we can
also do it infinitesimally, i.e. as a differential equation. We put Ln = es and get, for
Res V = z (s), r (s), λ (s), Ve (s) ≡ V (s):

dz
= O(V 2 )
ds
dr
= 2r + O(V 2 )
ds
dλ  
= (4 − d)λ − β2 λ2 + O λ3 , λ(z + r + Ve ), (z + r + Ve )2
ds
dVe
= MVe + O(V 2 )
ds
115
where Z
d  2−d
MV (φ) = 
e Ve (e 2 s φ(e−s ·) + Z)dµΓ (Z).
ds 0
We want to choose r (0), z (0), λ (0) such that lim Ve−t (t + s) ≡ V eff (s) exists, ∀s ∈ R.
t→∞

d=2. Take r = 0, λ = 2 g, z = 0.

Put
λ (s) = 2 e2s gt (s), r (s) = 2 e2s ρt (s),  = e−t .
We get
dg dρ
= O(e−2(t−s) g 2 ), = O(e−2(t−s) g 2 ).
ds ds
Hence the limits lim gt (t + s), limt→∞ rt (t + s) exist. This is a super renormalizable the-
t→∞
ory : the “bare” coupling g can be taken -independent and z , ρ also.

d=3. Now λ (s) = es gt (s) and we have a “resonance” in the ρ-equation:

dg
= O(e−(t−s) g 2 )
ds
dρ
= O(g 2 ).
ds
dr
More precisely, write ds
= 2r + γλ2 + O(λ3 ), then,

dρ
= γg 2 + O(e−(t−s) g 3 ).
ds
Since g stays almost constant, we get

ρt (s) = ρt (0) + sγg 2 + bounded.

Thus, we need to take ρt (0) = −tγg 2 i.e.

g = g, ρ = γg 2 log 

this is a “mass counterterm” of O(g 2 ), coming from the graph that di-
verges logarithmically in UV.

d=4. Now we stay in the (r, λ) picture. λ decreases logarithmically :

dλ
= −β2 λ2 + O(λ3 )
ds
λ(0)
λ(s) ∼ .
1 + β2 sλ (0)

116
If λ (0) stays bounded in , as we have to assume in our perturbative analysis, then
−→ 0. Thus the only perturbative theory (= the one where the coupling constant
λ(s) s→∞
remains small on all scales) is the Gaussian. Same holds for

dλ
d > 4. ds
= −(d − 4)λ + O(λ2 ) i.e. λ(s) → 0.

One might hope that, by taking λ (0) large enough, things might change. E.g. there
might be another fixed point λ∗ for the RG, i.e. the flow in λ would be

λ=0 λ∗

Unfortunately, this is unlikely to happen : in d > 4 it is rigorously proven that for all
z(), r(), λ() the only  → 0 limits are Gaussian. For d = 4 there are strong theoretical
and numerical arguments supporting the same conjecture.

d<4: nontrivial fixed points.

Consider the d = 3 QFT constructed above. This has mass ∼ m2 , coupling λ. What
about m → 0 ? Above we could have taken m = 0 and started with r , λ to produce an
effective Hamiltonian on scale 1 of the form
Z Z Z
z : ∇ϕ : +r : ϕ : +λ : ϕ4 : +Ve
2 2

with various values of r. To solve for the IR we would need to keep iterating the RG. Now
λ increases and eventually, if we really choose r = r(λ) = critical point, we would tend
to the nontrivial fixed point H∗ . Thus, UV is controlled by the Gaussian fixed point, IR
by the non-Gaussian.
We could construct a scaling limit at H∗ : this would be a scale-invariant QFT with non-
Gaussian IR and UV. H∗ has (at least) one relevant direction : the correlation length.
We could construct, as above, a massive QFT corresponding to the unstable manifold of
H∗ . We expect the unstable manifold of the Gaussian fixed point H0∗ to be 2-dimensional
(r, λ), and the unstable manifold of H∗ to be one dimensional. So the flow is

117
 r
Mu (H0∗ )

H0∗ λ

H∗

Mu (H∗ ) Ms (H∗ )

Thus, a “typical” QFT is a typical point of the r, λ plane with Gaussian UV and finite
correlation length. The ones on Ms (H∗ ) have non-Gaussian IR, Gaussian UV and the
ones on Mu (H∗ ) have non-Gaussian UV, ξ < ∞. (We should really include also the
r = ∞ (ξ = 0) “high-temperature” fixed point).

16.4 Asymptotic Freedom

What would have happened if β2 above had been negative at d = 4 ? Now

dλ λ (s)
= |β2 |λ2 + O(λ3 ) ⇒ λ (0) =
ds 1 + |β2 |sλ (s)

i.e. if  = e−t , s = t, and we want to fix λe−t (t) = g, then the “bare coupling” λ is
g
λ = −→ 0
1 + |β2 | log −1 g →0

i.e. to have a non-Gaussian theory, we need to let λ → 0 as  → 0. This is obvious,

since λ is now unstable. Actually, it is instructive to calculate exactly λ : we need the
λ-equation to O(λ3 ) (i.e. again, we need to consider z, r, Ve too, to O(λ3 )). So consider

dλ
= β(λ) = β2 λ2 + β3 λ3 + β̃(λ), β̃ = O(λ4 ).
ds

118
So
Z g
dλ
= log −1
λ β(λ)
Z g Z g  
1 1 1 β3
= dλ = dλ − + O(1)
λ β2 λ2 1 + β3
β2
λ + O(λ2)
λ β2 λ2 β22 λ

1 1 β3 g
= − − 2 log + O(g) = log −1
β2 λ β2 g β2 λ
g
⇒ λ = gβ3
.
1 + β2 g log  − β2 log log −1 + O(1)
−1

This is how the bare coupling should be chosen to have a continuum limit.
Remark. We have seen that β2 is universal (independent on regularization). It can be
shown that β3 is also. (β4 is not). They are intrinsic, depend only on the continuum limit
(which does not depend on the regularization). An example of asympotically free theory
is the Ginzburg-Landau model with λ negative. This is not stable, but can be rigorously
defined by analytic continuation. Since λ = −g, g > 0 we get dg ds
= β2 g 2 + · · · β2 > 0.
More interesting one is the SU (N ) gauge theory. There one has two couplings z and λ,
both marginal, with
dz
= γλ2 + · · ·
ds
dλ
= β3 λ3 + · · ·
ds
2 Rt
so λ(s)2 ≈ 1+β3 gg2 (t−s) where  = e−t , λ2 (t) = g 2 and z (t) − z (0) = γ 0
λ(s)2 ∼ γ
β3
log t so
z (0) must diverge as − βγ3 log log −1 .
Remarks.

1. Thus, we expect to find QFT’s at fixed points and on their unstable manifolds. For
Ginzburg-Landau model, d < 4 we find several non-Gaussian ones. For d ≥ 4 the
fixed points and unstable manifolds are Gaussian.

2. Asymptotically free theories have Gaussian fixed points with marginally unstable
non-Gaussian direction. Here we have a non-Gaussian continuum limit.

Examples : Non Abelian gauge theories with not too many fermions (QCD is one).
Non-linear σ-models in d = 2 (e.g. O(N ), N > 2 spin systems we discussed earlier).
Certain d = 2 fermionic models (Gross-Neveu model).

3. IR behaviour of asymptotically free models is interesting. Now λ(s) % as s in-

creases. E.g. in non-Abelian gauge and σ-models. These theories have no relevant

119
 variable like r in Ginzburg-Landau model that would provide a mass (correlation
length). Nevertheless, they are massive (ξ < ∞) : ξ with be ∼ e−const/λ (so called
2

dimensional transmutation). There is no other critical point than λ = 0. (λ2 is the

temperature in Heisenberg model).

4. Theories that are expected to be Gaussian (trivial) are : Ginzburg-Landau model,

QED, Standard Model of weak and EM interactions. This sounds paradoxical since
QED has a very successful renormalized perturbation theory and is the most ac-
curate known physical theory. The point is that, as long as we do not take  to
0, we can keep λ > 0 and so λ (s) > 0. Now if one calculates a correlation func-
tion in perturbation theory, the graphs that are convergent integrals of the form
R
|p|<−1
I(p)dp depend on the cutoff like  to some power. But we may take (recall,
λ = 1−β2 logλ
−1 λ
has to be small so ((log −1 )−1 ∼ λ)

 ∼ e−const/λ .

Hence, the dependence on cutoff is ∼ e−const/λ , very small for λ small ( λ1 = 137 for
QED). We say that all the above theories are good effective theories for distances
>>  (in practice, energies << 1015 GeV). For smaller distances one needs to find
a new theory (say string theory).

References
[1] Aizenman M., Translation invariance and instability of phase coexistence in the
two-dimensional Ising system, Commun. Math. Phys. 73 (1) (1980) 83–94.

[2] Aizenman M. and Wehr J., Rounding effects of quenched randomness on first-order
phase transitions, Commun. Math. Phys. 130 (3) (1990) 489–528.

[3] Aizenman M., Barsky D., and Fernández R., The Phase Transition in a General
Class of Ising-Type Models is Sharp. J. Stat. Phys. 47 (1987) 343-374.

[4] Baxter Rodney J., Exactly solved models in statistical mechanics, London, Academic
Press Inc. [Harcourt Brace Jovanovich Publishers] (1982).

[5] Bricmont J. and Kupiainen A., Phase Transition in the 3d Random Field Ising
model, Commun. Math. Phys. 116 (1988) 539–572.

[6] Bricmont J. and Fontaine J.R., Perturbation about the mean field critical point,
Commun. Math. Phys. 86 (1982) 337–362.

120
 [7] Bricmont J. and Kupiainen A., High temperature expansions and dynamical systems,
Commun. Math. Phys. 178 (1996) 703–732.

[8] Brydges D.C., “A short course on cluster expansions” in Phénomènes critiques,

systèmes aléatoires, théories de jauge, Les Houches, 1984 (Elsevier/North Holland,
Amsterdam, 1986) 129–183.

[9] Dobrushin R.L., Existence of a phase transition in the two-dimensional and three-
dimensional Ising models, Soviet Physics Doklady 10 (1965) 111–113.

[10] Dobrushin R.L., The description of a random field by means of conditional proba-
bilities and conditions of its regularity, Theory Prob. Appl. 13 (1968) 197–224.

[11] Dobrushin R.L., Prescribing a system of random variables by conditional distribu-

tions, Theory Prob. Appl. 15 (3) (1970) 458–86.

[12] Dobrushin R.L., Gibbs state describing coexistence of phases for a three-dimensional
Ising model, Theory Prob. Appl. 17 (4) (1972) 582–600.

[13] Dobrushin R.L., “ Gaussian Random Fields – Gibbsian Point of View, Multicom-
ponent Random Systems”, in “Advances in Probability and Related Topics”, 6, New
York, Dekker (1980) 119–51.

[14] Dobrushin R.L. and Shlosman S.B., “Absence of breakdown of continuous symmetry
in two-dimensional models of statistical physics”, Commun. Math. Phys. 42 (1975)
31.

[15] Fernández R., Fröhlich J. and Sokal A.D., Random Walks, Critical Phenomena, and
Triviality in Quantum Field Theory, New York, Springer Verlag (1992).

[16] Fisher M.E., Critical temperatures of anisotropic Ising lattices, II. General upper
bounds, Phys. Rev. 162 (3) (1967) 480–85.

[17] Fröhlich J. and Pfister C.E., On the absence of spontaneous symmetry breaking
and of crystalline ordering in two-dimensional systems, Commun. Math. Phys. 81
(2) (1981) 277–98.

[18] Fröhlich J., Simon B. and Spencer T., Infrared bounds, phase transitions and con-
tinuous symmetry breaking, Commun. Math. Phys. 50 (1) (1976) 79–95.

[19] Fröhlich J. and Spencer T., The Kosterlitz-Thouless transition in two-dimensional

abelian spin systems and the Coulomb gas, Commun. Math. Phys. 81 (1981), 527–
602.

121
[20] Gawedzki K., Kotecký R. and Kupiainen A., Coarse-Graining Approach to First-
OrderPhase Transitions, J. Stat. Phys. 47 (1988) 701–724.

[21] Glimm J. and Jaffe A., Quantum Physics, A Functional Integral Point of View, New
York, Springer (1981).

[22] Griffiths R.B., Peierls proof of spontaneous magnetization in a two-dimensional Ising

ferromagnet, Physical Review A 136 (1964) 437–439.

[23] Grimmett G., Percolation, New York, Springer-Verlag (1989).

[24] Itzykson C. and Drouffe J.-M., Statistical field theory, Volume 1 : From Brownian
motion to renormalization and lattice gauge theory, Cambridge University Press,
(1989) ISBN 0521408059.

[25] Kelley J.L., General Topology, Graduate Texts in Mathematics, New York, Springer
Verlag (1955).

[26] Klein A., Landau L.J. and Shucker D.S., On the absence of spontaneous breakdown
of continuous symmetry for equilibrium states in two dimensions, J. Stat. Phys. 26
(3) (1981) 505–12.

[27] Kosterlitz J.M. and Thouless D.J., “Ordering, metastability and phase transitions
in two-dimensional systems”, Journal of Physics C : Solid State Physics 6 (1973)
1181–1203.

[28] Kotecky R. and Shlosman R.B., First-order transitions in large entropy lattice mo-
dels, Commun. Math. Phys. 83 (1982) 493–515.

[29] Lanford III O.E. and Ruelle D., Observables at infinity and states with short range
correlations in statistical mechanics, Commun. Math. Phys. 13 (1969) 194–215.

[30] McCoy Barry M., Advanced Statistics Mechanics, International series of monographs
on physics 146, Oxford University Press (2010).

[31] Mermin N.D. and Wagner H., Absence of Ferromagnetism or Antiferromagnetism in

One- or Two-Dimensional Isotropic Heisenberg Models, Phys. Rev. Lett. 17 (1966)
1133–1136.

[32] Onsager L., Crystal Statistics. I. A Two-Dimensional Model with an Order-Disorder

Transition, Phys. Rev. 65 (1966) 117-149.

[33] Peierls R., On the Ising model of ferromagnetism, Proceedings of the Cambridge
Philosophical Society 32 (1936) 477–481.

122
[34] Pirogov S. and Sinai Ya. G., Phase diagrams of classical lattice systems, Theoretical
and Mathematical Physics 25 (1975) 1185–1192 ; 26 (1976) 39–49.

[35] Reed M. and Simon B., Methods of Modern Mathematical Physics 1, Functional
Analysis, New York (N.Y.), Academic Press (1980).

[36] Reed M. and Simon B., Methods of Modern Mathematical Physics 2, Fourier Ana-
lysis, Self-adjointness, New York, Academic Press (1975).

[37] Rudin W., Real and Complex Analyis, New York, McGraw-Hill (1987).

[38] Seiler E., Gauge Theories as a Problem of Constructive Quantum Field Theory and
Statistical Mechanics, Lecture Notes in Physics 159, Berlin, Springer (1982).

[39] Seneta E., Non-negative Matrices and Markov Chains, 2nd ed., New York, Springer
Verlag (1981).

[40] Sinai Y.G., Theory of Phase Transitions : Rigorous results, Pergamon Press, Oxford
(1982).

[41] Simon B., Functional Integration and Quantum Physics, New York, Academic Press
(1979).

[42] Simon B., The Statistical Mechanics of Lattice Gases, Volume 1, Princeton, Prince-
ton University Press, (1993).

[43] Van Enter A., Fernández R., Sokal A.D., Regularity properties and pathologies
of position-space renormalization-group transformations : Scope and limitations of
Gibbsian theory, J. Stat. Phys. 72 (1993) 879–1167.

[44] Wilson, K. (1975). The renormalization group: Critical phenomena and the Kondo
problem, Reviews of Modern Physics 47 (1975) 773-840.

[45] Yang C.N., “The spontaneous magnetization of a two-dimensional Ising model”,

Physical Rev. (2) 85 (5) (1952) 808–816.

123