Princeton University Department of Physics

The Renormalization Group - Lecture Notes

(Condensed)

Jan Tuzlić Offermann

January 22, 2018

1. Renormalization Group Scaling

Ising model recap:

X X
E ({si }) = −B si − J si sj , (1.1)
i hiji

where {si } denotes some particular configuration of spins. The partition function is then
given by
X
Z= exp (−βE[si ]) (1.2)
{si }
X X
= exp (−β)E[si ] (1.3)
m {si }|m
X
= exp (−βF (m)) , (1.4)
m

where we’ve defined the effective free energy F (m). Keep in mind that we’ve already
performed some kind of coarse-graining – let there already be some UV cutoff in place,
given by Λ ∼ 1/a.

1
 Conditions on F (m):

• Locality of Interactions: The Ising model is defined so that each spin only interacts
directly with its nearest neighbors, those on adjacent lattice sites. Non-neighboring
spins may only interact through intermediate ones. Thus, we can write
Z
F [φ(x)] = dd xf [φ(x)] ,

where f [φ(x)] is a local function that may depend on φ, ∇φ and higher derivatives.

• Translational/Rotational Invariance: The Ising model has a discrete translation

symmetry. Certain lattice types (e.g. the standard square lattice) also have dis-
crete rotational symmetries. For distances much larger than our lattice scale, the
continuum version of these symmetries should emerge (and thus appear in F ).

• Z2 symmetry: The Ising model exhibits symmetry under the combined transfor-
mation B → −B, si → −si , and so this should be reflected in F with symmetry
under transformation B → −B, φ(x) → −φ(x).

• Analyticity: We assume that F is analytic in φ(x), so that we may Taylor expand

F [φ(x)] and consider low powers of φ. Also, we assume that φ(x) varies appreciably
in x only over distances  a (our coarse-graining is coming into play). Considering
a gradient expansion of f [φ(x)] in dimensionless quantity a∇, the terms with lower-
order derivatives will dominate over higher-order ones.

From the above assumptions, we can write the general form of F [φ(x)] as
Z  
d 1 1 2 2 4
F [φ] = d x ∇φ · ∇φ + µ φ + gφ + . . . . (1.5)
2 2

For the Ising model, the temperature dependence sits in µ2 ∼ T − Tc . The renormalized
form looks basically the same, as:
Z  
0 d 1 0 0 1 2 02 04
Fζ [φ ] = d x ∇φ · ∇φ + µ (ζ)φ + g(ζ)φ + . . . . (1.6)
2 2

2
1.1. The "Scaling Dimension"
At a fixed point of a renormalisation group, any scale must be "washed away". This is
1 1 This "washing away" of scale suggests that the 2-point correlation
because ξ ∼ |T −T c|
ν.

function should take the form of a power law. By convention (see footnote #1), we write
it as
1
hφ(x)φ(0)i ∼ . (1.7)
rd−2+η
Now, consider how our variables are rescaled under a renormalization group transforma-
tion. We have our lengths rescale as x → x0 = x/ζ, so Equation 3.7 suggests rescaling φ
as φ(x) → φ0 (x0 ) = ζ ∆φ φ(x) with

d−2+η
∆φ = . (1.8)
2

If you’ve kept in mind the form of F [φ(x)], you might notice something funny going on.
Let’s think about inverse-mass dimensions, and do some "naive" dimensional analysis:

[x] = −1 , (1.9)
[∂/∂x] = +1 . (1.10)
Z 
d 1
[F ] = d x ∇φ · ∇φ = 0 (1.11)
2
d−2
⇒ [φ] = . (1.12)
2
But something seems off here, because looking at both sides of Equation 3.7, this gives
us

[hφ(x)φ(0)i] = d − 2 , (1.13)
 
1
=d−2+η (1.14)
rd−2+η
⇒η=0. (1.15)
1
Our definition of ξ stems from that of the correlation function – ξ defines two regimes for the function,
for different distance scales. For the mean-field theory example, where ν = 1/2, see Equation 2.22
on page 41 (and onwards) in Tong’s notes – in this case we define ξ ∼ 1/µ2 with µ2 ∼ |T − Tc |, and
the correlation function looks like
(
1
d−2 rξ ,
hφ(x)φ(y)i = rexp −r/ξ
r (d−1)/2
r ξ .

The first case may also have the denominator written more generally as rd−2+η , with η = 0 for
mean-field theory. For reasons explained in Tong’s notes (that we may not get to), mean-field theory
results for critical exponents like η and ν are only good for d ≥ 4. However, ν > 0 for the d = {2, 3}
cases as well, so we still find that ξ diverges as we approach the critical point in the familiar 2D Ising
model.

3
That’s not the general result – we know from experiments that ν 6= 0 for the 2D Ising
model. The η/2 difference between ∆p hi and Eq. 1.12 arises from the third step of the
renormalization group procedure. Rescaling φ is effectively coarse-graining it over blocks
of larger and larger size a. This dresses φ with a, 2 resulting in the correlation function
aη
hφ(x)φ(0)i ∼ . (1.16)
rd−2+η
This form is nicer for those used to using "engineering dimensions", but we commonly
absorb this a into our definition of φ, in favor of a correlation function like Equation 3.7.

1.2. Relevance of Interactions

Let’s see how the analysis of scaling dimensions can help us determine if a given interac-
tion term in our free energy is relevant, irrelevant or marginal. As a quick reminder on no-
tation, for some general operator O, we define scaling dimension ∆O so that O → ζ ∆O O
under the renormalization group.
Let’s consider some interaction term O(x) in our free energy F ,
Z
Fint [φ] ∼ dd xgO O(x) . (1.17)

O will be some product of φ and its derivatives. O is referred to as an operator. We’re

interested in operators that have a well-defined scaling dimension,3 so that under the
renormalization group we have

O(x) → ζ ∆O O(x) . (1.18)

From Equation 3.17, we see that

∆ gO = d − ∆ O (1.19)

Therefore, gO will

• diverge for ∆O < d (O is relevant)

• vanish for ∆O > d (O is irrelevant)

For the case ∆O = d, we say that O is marginal.

2
This short comment on "dressing φ with a" is based on Tong’s notes. For a fuller picture, see pages
49-52 in Cardy’s "Scaling and Renormalization in Statistical Physics", on the scaling of the spin-spin
2-point correlation function in the Ising model.
3
According to Tong (page 66), the difficult part is finding an O that exhibits this property – these
operators are typically complicated linear combinations of φ and its derivatives (and their products).

4
1.3. Critical Exponents – Demystified
We’re going to investigate how the critical exponents, which govern the power-law re-
lations followed by all the thermodynamic variables with respect to the reduced tem-
perature, t = |T − Tc |/Tc , arise from scale invariance. These power laws can be simply
derived for the case of Landau theory – this is more general. For fixed B = 0, we have
heat capacity (c), magnetization (φ) and magnetic susceptibility (χ) given by

c ∼ c± t−α , (1.20)
β
φ∼t , (1.21)
 γ
1
χ∼ , (1.22)
t

where c± reminds us of the discontinuity in c at t = 0. For fixed t, we have

φ ∼ B 1/δ . (1.23)

We’ve already seen that the correlation length ξ follows a power law given by

ξ ∼ t−ν . (1.24)

Thinking about scaling dimensions, we simply see that ∆ξ = −1 since ξ is just a length
scale, so it transforms like ξ → ξ/ζ. It immediately follows that

1
∆t = . (1.25)
ν
We can get at the other critical exponents by considering the expression for the thermo-
dynamic free energy,
Z
Fthermo (t) = dd x f (t) . (1.26)

F is scale-invariant at the fixed point. So by Equation 3.26, f (t) must have ∆f = d.

In fact, since we want F → F , we can employ Equation 3.25 (and the fact that dd x →
ζ −d dd x) to find that

f (t) ∼ tdv . (1.27)

∂2f
From this expression, we can find the per-spin heat capacity c, if we recall4 that c ∼ ∂t2
.
It then follows that

c ∼ tdν−2 (1.28)
⇒ α = 2 − dν . (1.29)
2
4 ∂ f
Recall that by definition, c = −T ∂T 2 . (See "Thermal Physics" by Kittel & Kroemer, page 63.)

5
Now let’s look at φ in Equation 3.21. We have

∆φ = β∆t (1.30)
⇒ β = ν∆φ
ν(d − 2 + η)
= . (1.31)
2
Now, we’re going to add a term of the form dd xBφ to F . This represents turning on
R

the magnetic field B – we could have done this earlier in our discussion, but it will only
∂φ
be helpful now since χ = ∂B |T , and Equation 3.23 features variations in B. Looking at
our new term in F , we see that

∆B = d − ∆φ
d+2−η
= . (1.32)
2

Looking at Equation 3.22, and using the derivative form of χ above, we see that

−γ
∆φ − ∆B = (1.33)
ν
⇒ γ = ν(2 − η) . (1.34)

Lastly, Equation 3.23 gives us

∆B
δ=
∆φ
d+2−η
= . (1.35)
d−2+η
So if we know η and ν, we can calculate critical exponents α, β, γ and δ.

For T > Tc :

∂f 1 ∂f
T = (t + 1)Tc , = ,
∂T Tc ∂t
1 ∂2f ∂2f
⇒ c = −(t + 1)Tc 2 2 ∼ (for small t) .
Tc ∂t ∂t2

For T < Tc :

∂f 1 ∂f
T = (−t + 1)Tc , =−
∂T Tc ∂t
1 ∂2f ∂2f
⇒ c = −(−t + 1)Tc 2 2 ∼ .
Tc ∂t ∂t2
(I may have a minus sign error somewhere – but the proportionality should be clear.)