Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

Renormalization in QED
Divergences, regularisation and renormalization
As we will see, in calculating loop Feynman diagrams in QED one finds integrals which diverge.
Such apparent pathologies are dealt with by the process of renormalization. This procedure can be
viewed in several ways. From one perspective it is a formal manipulation, part of the definition
of the quantum field theory, which allows one to calculate finite, testable expectation values and
scattering amplitudes. From a more physical perspective, one starts by noting that the key divergences
come from the high energy limits of the momentum integration. However, experimentally, we do not
now what physics describes high energy processes. For instance, there maybe be new very heavy
particles which contribute to the loop diagrams (as virtual particles in the loop) and would change the
divergence of the integral. From this perspective, renormalization is a procedure which allows us to
sensibly calculate the effects of the low-energy physics, independent of how it is corrected at high
energies.
Concretely renormalization proceeds as follows. As an example, suppose we are considering the
electron self-energy. We define the full electron propagator GF (p) as including order-by-order in the


perturbation theory the corrections to the Feynman Green function. Writing e0 and m0 for the “bare”
charge and mass parameters which enter the QED action we have

GF (p) = + + O(e40 )

i (1)
= + GF (p) + O(e40 ).
/ − m0 + i
p
(1)
The term GF (p) is divergent. The three steps in renormalization are as follows.
(1)
(1) regularisation: Introduce a new finite integral GF (p, Λ) which depends on a parameter Λ,
sometimes known as the “cut-off scale”. This has the property
(1) Λ→∞ (1)
GF (p, Λ) −→ GF (p)

We can then split the new function into divergent and finite pieces Adiv and Ac respectively
(1)
GF (p, Λ) = Adiv (p, Λ) + Ac (p, Λ)

The finite part Ac (Λ) leads to physically measurable effects and is known as a radiative correc-
tion.

(2) renormalization: If the theory is renormalizable the divergent part Ac (p, Λ) can be combined
with the tree-level propagator SF (p), so that
i
GF (p, Λ) = + Adiv (p, Λ) + Ac (p, Λ) + O(e40 )
/ − m0 + i
p
iZ2 (Λ)
= + Ac (p, Λ) + O(e40 )
/ − m(Λ) + i
p

1
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

Thus the divergent terms can be incorporated as a modification of the tree-level propagator by a
rescaling or “wavefunction renormalization” Z2 (Λ) and a renormalised mass parameter m(Λ).

(3) removing Λ-dependence: Finally we consider taking the limit Λ → ∞. In doing we let the
original bare parameters m0 and e0 become singular so that the new physical parameters m and
e are finite. The perturbation expansion is then really a series in e rather than e0 .

While this procedure may seem a sleight of hand, it is well-defined and gives definite (and testable)
physical results independent of the particular choice of regularisation. In general we say

a theory is renormalizable if all divergences can be removed by renormalization of a finite

number of couplings in the Lagrangian.

In what follows, we will discuss the structure of the renormalization procedure for the photon and
electron self-energy graphs as well as the vertex correction.

Photon self-energy
We define the full photon propagator with contributions from tree- and loop-level terms

 
hΩ| T Aµ (k)Aν (k 0 ) |Ωi ≡ (2π)4 δ 4 (k + k 0 )GFµν (k)
p+k
←k ←k ←k
= µ ν +µ α β ν + O(e40 )

= (2π)4 δ 4 (k + k 0 )Dµν
F
(k) + (2π)4 δ 4 (k + k 0 )GFµν(1) (k) + O(e40 )

where O(e40 ) denotes the contribution from higher-order Feynman diagrams. Evaluating the second
Feynman diagram we have

d4 p
Z h i
β
GFµν(1) (k) = F
Dµα (k) · (−) ie γ α
S
0 ab F
bc
(p + k) ie γ S
0 cd F
da
(p) F
· Dβν (k)
(2π)4
F
≡ Dµα (k) · ie20 Παβ (k) · Dβν
F
(k)

Substituting for the SF Feynman Green functions we have

d4 p
Z  
αβ α 1 β 1
Π (k) = i Tr γ γ
(2π)4 p/ + k/ − m0 + i p/ − m0 + i

This integral is quadratically divergent. To see this write p = E(1, v). Then d4 p ∼ E 3 dEd3 v,
extracting the overall dependence on E we have

E 3 dE
Z Z
large E
Π ∼ ∼ EdE
E2

2
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

which diverges as E 2 . We say this is a ultra-violet divergence since it diverges due to the high-energy
behaviour of the integral.
To remove the divergence we cold introduce a cut-off at some energy scale Λ to simulate the effect
of unknown physics. However this would break Lorentz invariance and is not very elegant. A simple
alternative is to remove the divergence by regulating the integral. This is done by introducing a new
function dependent on a parameter Λ, such that Παβ (k, Λ) → Παβ (k) as Λ → ∞. For instance one
can write
d4 p Λ4
Z  
1 1
Παβ (k, Λ) = i Tr γ α
γ β
(2π)4 p/ + k/ − m0 + i p/ − m0 + i (Λ2 − p2 )2
At finite Λ, as p → Λ this is now convergent since Π ∼ dE/E 3 at large E 1 . As Λ → ∞ the
R

new function Παβ (k, Λ) tends to the original integral Παβ (k). It is easy to show that the leading Λ
behaviour in an expansion of Παβ (k, Λ) is Λ2 , reflecting the divergence of Παβ (k). Thus analysing
the Λ dependence of Παβ (k, Λ) we can separate out the divergent parts from the finite contributions
which will correspond to those of the full theory as we approach the limit Λ → ∞ . Note that this
type of regulator breaks gauge invariance but more complicated regulators can preserve it.
One can explicitly calculate the value of this integral, taking care to deal correctly with the i
defining the poles in the fermion Green functions. However, here we will simply discuss the structure
of the result of the integration. First we note that by Lorentz invariance Παβ must have the form

Παβ (k, Λ) = −g αβ A(k 2 , Λ) + k α k β B(k 2 , Λ)

Concentrating on A(k 2 , Λ) and expanding in k 2 gives

A(k 2 , Λ) = A0 (Λ) + k 2 A1 (Λ) + k 2 Πc (k 2 , Λ)

where Πc (k 2 , Λ) ∼ k 2 for small k 2 . From the original expression for Παβ (k) we have, keeping the
k dependence, Π ∼ EdEf (k/E). Thus expanding f (k/E), we see that in the k 2 expansion each
R

successive term introduces an extra power of E −2 . Thus the leading term diverges as EdE ∼ E 2 ,
R

the next as dE/E ∼ log E and the third dE/E 3 is finite. This implies
R R

A0 (Λ) ∼ Λ2 A1 (Λ) ∼ log Λ Πc (k 2 , Λ) is finite

Given this expansion the full propagator has the form

igµν 1 1
GFµν (k, Λ) = − + igµν 2 e20 A(k 2 , Λ) 2 + O(e40 )
k2
+ i k + i k + i
e20 A0 (Λ) ie20 gµν 
 
igµν
A1 (Λ) + Πc (k 2 , Λ) + O(e40 )

=− 2 1− 2
+ 2
k + i k k + i
Writing 1 − e20 A0 (Λ)/k 2 = [1 + e20 A0 (Λ)/k 2 ]−1 + O(e40 ) we have, keeping only terms of order e20 ,
igµν 1 − e20 A1 (Λ)
 
F ie20 gµν
Gµν (k, Λ) = − 2 + Πc (k 2 , Λ) + O(e40 )
k + e20 A0 (Λ) + i k 2 + i
(0.1)
iZ3 (Λ)gµν ie20 gµν 2 4
≡− 2 + Πc (k , Λ) + O(e0 )
k − m2γ (Λ) + i k 2 + i
1
To see this expand the regulator in the limit Λ2 p2  1 and show by power counting.

3
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

where

Z3 (Λ) = 1 − e20 A1 (Λ) + O(e40 )

m2γ (Λ) = −e20 A0 (Λ) + O(e40 )

Comparing with the general Källen–Lehmann expression for the two-point function in an inter-
acting scalar theory
∞
dM 2
Z
iZ i
GF (k) = + ρ(M 2 ) 2
k 2 − m2 + i ∼4m2 2π p − M 2 − i

we see that Z3 (Λ) corresponds to a wavefunction renormalization Z3 (Λ), while we also introduce a
new renormalised mass term m2γ (Λ). Physically, we measure that the photon is massless. Thus we
expect
m2γ (Λ) = 0

One finds that this is indeed the case provided the regularization preserves the gauge symmetry. (Note
that the regularization mentioned above, does not preserve the gauge symmetry. Such regularizations
can be dealt with, but the analysis is a bit more complicated.) By fixing the gauge one can also ensure
that the B(Λ) function vanishes.
We can incorporate the effect of the wavefunction renormalization by defining a rescaled field

−1/2
wavefunction renormalization: Aµ → Aph
µ = Z3 (Λ) Aµ .

such that
ph 0 0 −igµν
hΩ| T Aph 4 4
µ (k)Aν (k ) |Ωi = (2π) δ (k + k ) + ...
k 2 − i
1/2
Physically this means that the actual measured field Aph
µ is Z3 times the field which appears in the
“bare” Lagrangian. Taking the Λ → ∞ limit, we hold the rescaled physical field Aph
µ fixed, while the
bare field Aµ becomes singular.

Electron self-energy

This calculation is analogous to the photon self-energy just discussed. We define the full electron
propagator with contributions from tree- and loop-level terms

 
hΩ| T ψ(p)ψ̄(p0 ) |Ωi ≡ (2π)4 δ 4 (p + p0 )GF (p)
←k

p p p
= + + O(e40 )
α p−k β

(1)
= (2π)4 δ 4 (p + p0 )SF (p) + (2π)4 δ 4 (p + p0 )GF (p) + O(e40 )

4
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

Evaluating the second Feynman diagram we have

d4 k h
Z i
(1) α β F
GF (p) = SF (p) · ie0 γ iS F (p − k) ie 0 γ iD αβ (k) · SF (p)
(2π)4
≡ SF (p) · ie20 Σ(p) · SF (p)

F Feynman Green functions we have

Substituting for the SF and Dαβ

d4 k α
Z
1 1
Σ(p) = i 4
γ γα 2
(2π) p/ − k/ − m0 + i k + i

Again the integral is ultraviolet divergent. Writing p = E(1, v) we have

E 3 dE
Z Z
large E
Σ ∼ ∼ dE
E3
which diverges as E. (Note that the integral also diverges at small E. This is an infrared divergence.
It is actually cancelled by a related tree-level divergence and we will not consider it further.)
Again there are various ways to regulate the integral. For instance, we cam replace the photon
Feynman Green function by a Λ dependent function
1 1
DF (k) → DF (k, Λ) = − 2
k2 + i k − Λ2 + i
This gives a finite function Σ(p, Λ) which diverges linearly as Λ → ∞.
Recall that Σ(p, Λ) is combination of gamma matrices. The only Lorentz invariant combinations
of gamma matrices we can form are the identity 1 and powers of p/. Since p/p/ = p2 1 we thus can
choose to expand

Σ(p, Λ) = Σ0 (Λ) + Σ1 (Λ)(p/ − m0 ) + Σc (p2 , Λ)(p/ − m0 )

where Σ0 , Σ1 and Σc are scalars and Σc (p2 , Λ) → 0 as p2 → m20 . Note that this expansion implies
that, acting on the free-particle spinor u(p), we have Σ(p, Λ)u(p) = Σ0 (Λ)u(p). Expanding the
integral defining Σ(p) in terms of p/E gives

Σ0 (Λ) ∼ Λ Σ1 (Λ) ∼ log Λ Σc (p2 , Λ) is finite

The full propagator can then be written as

i i 1
GF (p, Λ) = − e2 Σ0 (Λ)
/ − m0 + i p
p / − m0 + i 0 p/ − m0 + i
i i
− e20 Σ1 (Λ) − e2 Σc (p2 , Λ) + O(e40 )
/ − m0 + i
p p/ − m0 + i 0

For matrices A and B it is easy show that

(A − B)−1 = A−1 + A−1 BA−1 + O(B 2 )

5
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

/ − m0 + i and B = e20 Σ0 (Λ)1 we thus can rewrite, working to order O(e40 ),

Taking A = p
i i i
GF (p, Λ) = − e2 Σ1 (Λ) − e2 Σc (p2 , Λ) + O(e40 )
/ − m0 + e20 Σ0 (Λ) + i p/ − m0 + i 0
p p/ − m0 + i 0
i[1 − e20 Σ1 (Λ)] i
= − e2 Σc (p2 , Λ) + O(e40 )
/ − m0 + e0 Σ0 (Λ) + i p/ − m0 + i 0
p 2

iZ2 (Λ) i
≡ − e2 Σc (p2 , Λ) + O(e40 )
/ − m(Λ) + i p
p / − m0 + i 0
(0.2)
where we have defined

Z2 (Λ) = 1 − e20 Σ1 (Λ) + O(e40 )

m = m0 − e20 Σ0 (Λ) + O(e40 )

We see that Z2 corresponds to a wavefunction renormalization and we also have a renormalization

of the mass m0 ,

wavefunction renormalization: ψ → ψ ph = Z2 (Λ)−1/2 ψ

mass renormalization: m0 → m = m0 + δm(Λ)

where δm(Λ) = −e20 Σ0 (Λ) + O(e40 ). Thus

i
hΩ| T ψ ph (p)ψ̄ ph (p0 ) |Ωi = (2π)4 δ 4 (p + p0 ) + ...
p/ − m + i

Vertex modification
Finally we consider renormalization of the vertex. We define the full vertex with contributions from
tree- and loop-level terms

 

hΩ|T Aµ (p3 )ψ(p2 )ψ̄(−p1 ) |Ωi
 
 
= + + + +  + O(e50 )
 
 

p2


← p3
=
p1

= (2π)4 δ 4 (−p1 + p2 + p3 )GFµα (p3 )GF (p2 ) · V α (p1 , p2 ) · GF (p1 )

(0.3)

6
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

This case is rather more complicated than the previous cases since more diagrams contribute. Note
that the first three diagrams in brackets correspond to self-energy corrections to the electron and pho-
ton propagators GF (p1 ), GF (p2 ) and GFµα (p3 ), while the last diagram is a correction to the vertex
V α (p1 , p2 ). In the following line these different corrections are represented by the solid blobs. Ex-
panding the propagators to order e20 then gives the diagrams in the previous line provided the vertex is
given by, with p3 = p1 − p2 ,

p2


p2 p2 − k
← p3 ← p3 α
µ
V (p1 , p2 ) = µ + µ ↑k + O(e50 )
β
p1 p1 − k
p1
µ
= ie0 γ + ie30 Λµ (p1 , p2 ) + O(e50 ).
Evaluating the second Feynman diagram we have
d4 k α
Z
µ 1 1 1
Λ (p1 , p2 ) = −i 4
γ γµ γα 2
(2π) p/2 − k/ − m0 + i p/1 − k/ − m0 + i k + i
Again the integral is ultra-violet divergent. Writing p = E(1, v) we have
E 3 dE
Z Z
large E dE
Λ ∼ 4
∼
E E
which diverges as log E.
Again we can regulate the integral introducing a cut-off scale Λ and then expand in terms of
momenta p1 and p2 . We write

Λµ (p1 , p2 , Λ) = Λµ0 (Λ) + Λµc (p1 , p2 , Λ)

where (compare with expansion of Σ(p, Λ)) we define the expansion by requiring
ū(p2 )Λµc (p1 , p2 , Λ)u(p1 ) = 0. This is equivalent to Λµc (p1 , p2 , Λ) = Aµ (p1 , p2 , Λ)(p/1 −
m0 ) + (p/2 − m0 )B µ (p1 , p2 , Λ) for some matrices Aµ and B µ . We then have

Λµ0 (Λ) ∼ log Λ Λµc (p1 , p2 , Λ) is finite

By Lorentz covariance we have

Λµ0 (Λ) = L(Λ)γ µ
and hence
V µ (p1 , p2 , Λ) = ie0 γ µ + ie30 L(Λ)γ µ + ie30 Λµc (p1 , p2 , Λ) + O(e50 )
(0.4)
≡ iZ1 (Λ)−1 e0 γ µ + ie30 Λµc (p1 , p2 , Λ) + O(e50 )
where we have defined Z1 (Λ)−1 = 1 + e20 L(Λ) + O(e40 ) so

Z1 (Λ) = 1 − e20 L(Λ) + O(e40 )

7
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

This leads to a renormalization of the electric charge e. To see exactly how we must be careful


about how we define e. In the free theory we have

h0| T Aµ (p3 )ψ(−p1 )ψ̄(p2 ) |0i =

i i i
= (2π)4 δ 4 (−p1 + p2 + p3 ) · ie0 γµ ·
p23 + i p/2 + m0 + i p/1 + m0 + i

From this perspective, the free electric charge is given by the coefficient of the correlation function as
p21 → m20 , p22 → m20 . Similarly the measured physical charge e should be given by the residue of the
physical three point function. That is we define e by the interacting three-point correlation function

hΩ| T Aph ph ph
µ (p3 )ψ (p2 )ψ̄ (−p1 ) |Ωi
i i i
= (2π)4 δ 4 (−p1 + p2 + p3 ) · ieγµ · + ...
p23 + i p/2 + m + i p/1 + m + i

where m is now the physical mass and the dots represent terms which do not diverge as p21 → m2 and
p22 → m2 .
Given definitions of the physical fields, the relation in the last line of (0.3), the definition (0.4) of
Z1 and the expressions (0.1) and (0.2) for GFµν (p) and GF (p) we see that

1/2
Z2 Z3
charge renormalization: e0 → e = e0 .
Z1

Ward identity

We have seen that there are three contributions to the renormalization of the charge e: from the photon
and fermion wavefunctions and from the vertex modification. Thus far we considered only one type
of fermion. In fact in QED we have three types: electrons, muons and taus. Each have different mass
and so a priori each lead to different Z1 and Z2 renormalizations. Thus, labelling the type of particle
by the index (i) we have

(i) 1/2
Z2 Z3
e0 → e(i) = (i)
e0 for (i) = e, µ, τ
Z1

Note that the photon wavefunction renormalization is the same for each particle, though it now has
three contributions arising from integrating over a loop of electrons, muons or taus. Generically we
see that each particle gets a different renormalized charge. But this is contrary to observation (and
the principle of minimal coupling) that all particles couple to the same basic unit of electric charge.
Equivalently, it would imply that the quantum loop corrections violate the original gauge invariance
of the bare action.

8
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

In fact, there is an relation, known as the Ward identity which ensures that all the e(i) are equal. It
states simply that for any given fermion

Z1 = Z2 Ward identity

which ensures that the renormalized theory is still gauge invariant.

It is easy to see that this relation holds for our (formally divergent) one-loop unregulated expres-
sions. Recall that
d4 k α
Z
Σ(p) = −i γ SF (p − k)γα DF (k)
(2π)4
Taking a derivative with respect to pµ gives

d4 k α ∂SF (p − k)
Z
∂Σ(p)
= −i γ γα DF (k)
∂pµ (2π)4 ∂pµ

Now by definition SF−1 (q) = i (q/ − m0 + i), so since SF SF−1 = 1, we have

∂SF (q) ∂SF−1 (q)

= −S F (q) SF (q) = −iSF (q)γ µ SF (q)
∂q µ ∂q µ

This gives, comparing with Λµ (p1 , p2 ),

d4 k α
Z
∂Σ(p)
= γ SF (p − k)γ µ SF (p − k)γα DF (k)
∂pµ (2π)4
= Λµ (p, p)

Assuming these relations still hold after regularization, we have, by definition, that the divergent
pieces of each expression are given by

∂Σ(p, Λ)
= Σ1 (Λ)γ µ + . . .
∂pµ
Λµ (p, p, Λ) = L(Λ)γ µ + . . .

where both Σ1 and L diverge as log Λ. Thus we have that Σ1 (Λ) = L(Λ) and hence

Z1 (Λ) = Z2 (Λ)

proving the Ward identity to this order. In fact it is possible to show that the identity holds at all orders
in perturbation theory. It is necessary for the quantum theory to be gauge invariant.

Renormalizability of QED

Finally let us briefly sketch how one shows that all potential divergences in QED, to all orders in
perturbation theory can be absorbed by renormalization, that is, that the theory is renormalizable.
First one has to classify the divergent diagrams. The initial step is to identify “proper” diagrams

9
Quantum Electrodynamics (QED), Winter 2015/16 Renormalization in QED

which cannot be reduced to two diagrams simply by cutting an internal line. The first diagram below
is not proper; the second is.

 
One then removes all the self-energy and vertex corrections from graphs, since these we know are
renormalizable. These give the “skeleton graphs”.

  −→

For the remaining (infinite set) of graphs one can then count the expected order of divergence from
the powers of E in the propagators and integrals. One finds that there are only five possible divergent

  
graphs

 
Z3 Z2 , m Z1

zero finite
The first three are the familiar diagrams we have already considered. Of the last two, one vanishes
and one is finite essentially as a result of charge conjugation symmetry and gauge invariance. Thus
we see the theory is indeed renormalizable.

10