CHAPTER 7.

REGULARIZATION METHODS 201

with an action for the φ field plus a new field Φ of mass Λ:

Z    
1 1 
− dd x ∂µ φ∂ µ φ + m2 φ2 − ∂µ Φ∂ µ Φ + Λ2 Φ2
2 2
Then, the two-point correlation function is given by
Z  
d4 p 1 1
h(φ + Φ) (x) (φ + Φ) (y)i = −ih̄ 4
exp(−ip · (x − y)) 2 2
− 2
(2π) p + m − i p + Λ2 − i

Furthermore, we replace potentials V (φ) with V (φ + Φ), e.g.

λφ4 7→ λ (φ + Φ)4

and compute correlation functions of φ + Φ rather than just φ.

The Pauli-Villars regularization method is a technical improvement on the primitive cutoff

regularization scheme we have used so far. For example, since momentum integrals now run over
all of momentum space, Dirac delta functions are no longer Λ-regulated, and so momentum
conservation should be conserved exactly, not just approximately.

On the other hand, the Pauli-Villars turns out not to be nearly so efficient at handling more
complicated quantum field theories. For example, Pauli-Villars regularization breaks gauge
symmetries in theories such as electromagnetism. To understand this, recall that giving a mass to
the gauge field breaks the gauge symmetry. Since Pauli-Villars adds a propagator term with a
nonzero mass (Λ), a Pauli-Villars regularization necessarily breaks the gauge symmetry. That, by
itself, does not necessarily mean Pauli-Villars is less useful than other regularizations, but it does
turn out to be one reason why Pauli-Villars regularization is often inconvenient and inefficient.

7.3 Dimensional regularization

Most modern workers do not typically use either a cutoff regularization or Pauli-Villars, but
instead use a different method, known as dimensional regularization. It will turn out that this
method is computationally far more efficient in nontrivial theories than either cutoff regularization
or Pauli-Villars, and is also compatible with far more symmetries. The disadvantage to
dimensional regularization will be a lack of concrete intuition for why precisely it works.

In dimensional regularization, we formally evaluate integrals in d dimensions instead of 4

dimensions, where d need not be an integer. For example, the integral
Z
d4 p
(p2 + m2 )n

is replaced by an integral we shall write

Z
dd p
(p2 + m2 )n
CHAPTER 7. REGULARIZATION METHODS 202

We shall see that integrals which diverge for integer d, will give finite results for noninteger d. In
effect, dimensional regularization consists of performing a sort of analytic continuation in the
dimension, thus the name.

Integrals in d dimensions are defined to obey the following axioms:

1. Linearity: For any complex numbers a, b,

Z Z Z
dd p [af (p) + bg(p)] = a dd p f (p) + b dd p g(p)

2. Scaling: For any complex number s,

Z Z
dd p f (sp) = s−d dd p f (p)

3. Translation invariance: For any vector q,

Z Z
d
d p f (p + q) = dd p f (p)

Furthermore, when d is an integer, the integrals above should match ordinary integrals. Note in
particular translation invariance of integrals in dimensional regularization. As we observed in
section **** CITE ***, translation invariance is not a property of all integrals in cutoff
regularization, specifically, it is not a property of linearly divergent integrals, and this is essential
to understand anomalies in cutoff regularization. We shall see in section **** CITE **** that
anomalies appear in dimensional regularization in a very different fashion.

Given the axioms above, it is possible to systematically develop a general theory of integration in
d dimensions, see for example [Collins][chapter 4]. However, we will adopt a different approach to
dimensionally-regularized integrals here. We shall first evaluate examples of divergent loop
integrals in arbitrary integral dimension d, then rewrite those formulas in such a way as to make
sense for nonintegral d.

Now, let us work through a particular example. Consider the integral

Z
dd p
Id (q, n) =
(p2 + 2p · q + m2 )n

(in Minkowski metric) for some q and some integer n. Let us work out how dimensional
regularization applies to this integral. The basic idea will be to rewrite this as an integral that
only depends on p2 , removing the angular dependence, so that the integral over momentum space
consists of an angular piece (which can be factored out) and an integral over the magnitude of p,
which we can turn into an integral that formally does not require d to be integral.

First, let us Wick rotate the integral from Minkowski space to Euclidean space, to get
Z
dd p
Id (q, n) = i
(p2 + 2p · q + m2 )n
CHAPTER 7. REGULARIZATION METHODS 203

where we have used p0M = +ip0E .

First, let us complete the square in the denominator.

p2 + 2p · q + m2 = (p + q)2 + (m2 − q 2 )
Then, shift the momentum by an amount q: p0 = p + q. The integral becomes
Z
dd p
Id (q, n) = i
(p2 + (m2 − q 2 ))n
Note that at this point in the evaluation, we are using the translation-invariance axiom of
dimensionally-regularized integrals. (As we saw in section *** CITE ***, in cutoff regularization
not all integrals are translation-invariant.)

Next, we need to evaluate this integral in d dimensions. Since the integrand only depends upon
the magnitude of p, we can write
Z Z Z ∞
dd p = dΩd pd−1 dp
0
where Ωd is the d-dimensional solid angle. Since the integrand only depends upon the magnitude,
it is independent of Ωd , so we can evaluate it separately. To do so, we can use a trick:
Z d
√
d ∞
−x2
π = e
−∞
Z d
!
X
d
= d x exp − x2i
i=1
Z Z ∞ 2
= dΩd dx xd−1 e−x
0
We can evaluate the integral over x’s using the gamma function discussed in appendix **** CITE
*****:
Z ∞ Z
2 1 ∞ 2
dx xd−1 e−x = d(x2 )(x2 )d/2−1 e−x
0 2 0
1
= Γ(d/2)
2
To briefly review, the gamma function Γ(x) is a function which for complex numbers x such that
Re x > 0 is given by the expression (due to Euler)
Z ∞
Γ(x) = e−t tx−1 dt
0
More generally, the gamma function is defined almost everywhere in the complex plane, though
for Re x ≤ 0 one must work a little harder to define it. It diverges for x an integer less than or
equal to zero, i.e. x = 0, −1, −2, −3, · · ·, but otherwise Properties of the gamma function include:
√
Γ(1/2) = π
Γ(1) = 1
Γ(2) = 1
Γ(3) = 2
Γ(x + 1) = xΓ(x)
CHAPTER 7. REGULARIZATION METHODS 204

from which one finds that Γ(n) = (n − 1)! for n a positive integer. Thus, the gamma function is a
generalization of the factorial function – the factorial is only defined over the integers, whereas
the gamma function is defined for complex numbers.

****** NOTE TO SELF: At this point in my notes, there was an aside on the gamma function,
[Link] p 3 – what to include ?? I’ve put a little bit above, I’m just wondering if maybe more
should be included.

Thus, we see that Z 

√
d 1
π = dΩd Γ(d/2)
2
and so Z
2π d/2
dΩd =
Γ(d/2)
In particular, the expression above makes sense for non-integer d as well as integer d, so we can
use it to “analytically continue” to non-integer dimensions d.

Let us briefly check that the expression above for the solid angle is sensible. For d = 2, Ωd should
be the number of radians in a circle, and indeed

2π 2/2
= 2π
Γ(2/2)

Similarly, for d = 3, Ωd should count the solid angles in a sphere, and indeed

2π 3/2 2π 3/2 4π 3/2

= = √ = 4π
Γ(3/2) (1/2)Γ(1/2) π

Returning to our integral Id (q, n), we have

Z
dd p
Id (q, n) = i
(p2 + (m2 − q 2 ))n
Z Z ∞
pd−1 dp
= i dΩd
0 (p2 + (m2 − q 2 ))n
Z
2π d/2 1 ∞ (p2 )d/2−1 d(p2 )
= i
Γ(d/2) 2 0 (p2 + (m2 − q 2 ))n

Define
m2 − q 2
x ≡
p 2 + m2 − q 2
so
m2 − q 2
p 2 + m2 − q 2 =
x
m 2 − q2
d(p2 ) = − dx
x2
CHAPTER 7. REGULARIZATION METHODS 205

Thus,
Z !  d/2−1 !−n
π d/2 1 m2 − q 2   1 m2 − q 2
2 2 d/2−1
Id (q, n) = i dx m − q − 1
Γ(d/2) 0 x2 x x
Z
π d/2  2 d/2−n 1
= i m − q2 dx xn−d/2−1 (1 − x)d/2−1
Γ(d/2) 0

In appendix**** CITE ****, the beta function B(a, b) is defined to be a ratio of gamma functions
Γ(a)Γ(b)
B(a, b) =
Γ(a + b)
and it is shown there that Z 1
B(a, b) = dx xa−1 (1 − x)b−1
0
Thus, we see that
Z
dd p
Id (q, n) = (Minkowski)
(p2 + 2p · q + m2 )n
π d/2  2 d/2−n
= i m − q2 B(n − d/2, d/2)
Γ(d/2)
π d/2  2 d/2−n Γ(n − d/2)Γ(d/2)
= i m − q2
Γ(d/2) Γ(n)
 d/2−n Γ(n − d/2)
= iπ d/2 m2 − q 2 (7.2)
Γ(n)

Note that although the integral above was originally defined for integer d, the result of the
computation above is an expression that makes sense for arbitrary, not-necessarily-integer d.

Aside: The careful reader will complain that when d/2 − n is irrational, then the
expression above does not make sense, since in that case
 
(m2 − q 2 )d/2−n ≡ exp (d/2 − n) ln(m2 − q 2 )

and m2 − q 2 is not unitless, so we cannot compute the ln. However, when we apply
dimensional regularization to specific theories, we shall introduce an additional mass
parameter Λ, and that mass parameter will appear in such a way so as to solve this
problem – we will only ever compute ln’s of unitless quantities such as (m2 − q 2 )/Λ2 .

As a consistency check, let us compare singularities. The original integral, for large |p|, is
proportional to
Z ∞ d−1
r dr
r 2n
and so has a divergence at the upper end of the integral when d − 1 − 2n ≥ −1, or equivalently
n ≤ d/2. We derived a closed-form expression for the integral, proportional to Γ(n − d/2), and
CHAPTER 7. REGULARIZATION METHODS 206

since gamma functions diverge at integers less than or equal to zero, we see that for n − d/2 an
integer, the divergences of the original integral and our expression match. However, our
expression is not only defined more generally, but when n − d/2 is not an integer, our expression
converges, regardless of whether n ≤ d/2.

We originally advertised dimensional regularization as a simplification, but the reader may

complain that we have had to do a great deal of work so far. However, the work above only has to
be done once – once one has a table of integrals of the form above, the workload drops greatly. To
that end, here is a collection of commonly-used results:
Z  n−d/2
dd p 1 1 Γ(n − d/2) 1
= i (7.3)
(2π) (p + m2 )n
d 2 (4π) d/2 Γ(n) m2
Z  n−d/2−k
dd p (p2 )k 1 Γ(n − d/2 − k)Γ(d/2 + k) 1
= i (7.4)
(2π)d (p2 + m2 )n (4π) d/2 Γ(d/2)Γ(n) m2
Z  n−d/2−1
dd p pµ pν 1 gµν Γ(n − d/2 − 1) 1
= i (7.5)
(2π)d (p2 + m2 )n (4π)d/2 2 Γ(n) m2
Z  
dd p pµ pν pρ pσ 1 Γ(n − d/2 − 2) 1 n−d/2−2
= i
(2π)d (p2 + m2 )n (4π)d/2 Γ(n) m2
1
· (gµν gρσ + gµρ gνσ + gµσ gνρ ) (7.6)
4

Next, let us further simplify the expression for Id (q, n). To this end, we will use the fact that
!
1  π2
Γ() = + ψ(1) − ψ (1) (1) − ψ(1)2 − + O(2 )
 2 3

and more generally,

! !
(−)n 1  π2
Γ(−n + ) = + ψ(n + 1) − ψ (1) (n + 1) − ψ(n + 1)2 − + O(2 )
n!  2 3

for n a positive integer, where ψ(x) is the digamma function, defined by

d
ψ(x) = ln Γ(x).
dx

and ψ (1) (x) is a polygamma function, defined as the first derivative of the digamma function. The
Euler-Mascheroni constant γ is defined to be γ = −ψ(1). Derivations of the expansions above and
other information on the digamma and polygamma functions can be found in appendix *** FILL
IN ****.

Let us apply this to simplify our expression for Id (q, n). Suppose that the dimension d is “close
to” 4, and that n is an integer such that 2 − n > 0. Define  = 4 − d, then
 
(−)2−n 2
Γ(n − d/2) = Γ(n − 2 + /2) = + ψ(3 − n) + O()
(2 − n)! 
CHAPTER 7. REGULARIZATION METHODS 207

From this we find

Z
dd p
Id (q, n) =
(p2 + 2p · q + m2 )n
 d/2−n  
d/2 2 2 1 (−)2−n 2
= iπ m − q + ψ(3 − n) + O()
Γ(n) (2 − n)! 

Next, let us apply this to some specific examples. First, however, with a bit of foresight, we
should slightly modify the action so as to get more sensible units. We shall rewrite the interaction
term in λφ4 theory so that the coupling constant has its d = 4 mass dimensions for all d; in
particular, for λφ4 near four dimensions, this will insure that λ is always dimensionless. We will
see later that this has the effect of sanitizing units in our later expressions.

In addition, it is also sometimes convenient to arrange for the coupling constant to

be dimensionless. In λφ4 theory in four dimensions, λ is automatically dimensionless,
but this need not be true in other cases. Therefore, in such other cases, sometimes one
adds factors of Λ so as to manufacture a dimensionless constant, e.g. g = Λn g where g
is the original coupling constant and g is dimensionless.

We argued previously in section *** CITE *** that if we demand that the action be
scale-invariant, then the real scalar field φ has units of mass dimension d/2 − 1, so classically φ4
has mass dimension r(d/2 − 1). Thus, if we add a scale factor Λ so that the action becomes
Z  
d 1 
d x − ∂µ φ∂ µ φ + m2 φ2 − λΛ4−d φ4
2
then we see that the coupling constant has classical mass dimension zero. In other words, we have
made the mass scale explicit. We will see that this will have the effect of sanitizing the units in
the expressions we shall derive.

Now, in λφ4 theory “near” d = 4, consider the diagram

! Z
λΛ4−d
= −i (4)(3) dd zDF (x1 , z)DF (z, x2 )DF (z, z)
1 2 h̄

The DF (z, z) = DF (0, 0) factor is divergent. It is given by

Z
dd p 1
Λ4−d DF (z, z) = −ih̄Λ4−d
(2π)d p2 + m2
ih̄
= − Λ4−d Id (0, 1)
(2π)d
ih̄  d/2−1 Γ(1 − d/2)
4−d d/2
= − Λ (i)π m2
(2π)d Γ(1)
 d/2−1
= h̄ (4π)−d/2 m2 Λ4−d Γ(1 − d/2) (7.7)
CHAPTER 7. REGULARIZATION METHODS 208

where to clean up the units we have included the Λ4−d factor. For d “near” 4,  = 4 − d, we have
 d/2−1  
2
Λ4−d DF (z, z) = −h̄ (4π)−d/2 m2 Λ4−d
+ ψ(2) + O()

! !
m2 2 Λ2
= −h̄ + ψ(2) + ln 4π 2 + O() (7.8)
(4π)2  m

using the fact that

d/2−1
1−/2
m2 m2
Λ4−d = Λ2(2−d/2)
(4π)d/2 (4π)2−/2
!/2
m2 Λ2
= 4π 2
(4π)2 m
! !
m2  Λ2
= 1 + ln 4π 2 + O(2 )
(4π)2 2 m

Note that the divergent factor above could be obtained more simply from the amputated diagram

! Z
λΛ4−d
= −i (4)(3) dd zδd (x1 − z)δd (z − x2 )DF (z, z)
h̄
1××2 !
λΛ4−d
= −i (4)(3)δd (x1 − x2 )DF (0, 0)
h̄

As a related exercise, let us compute the two-loop “double scoop” diagram

!2 Z
λΛ4−d
= (144) −i dd z1 dd z2 DF (x, z1 )DF (z1 , y)DF (z1 , z2 )2 DF (z2 , z2 )
x y h̄
!2 Z Z
λΛ4−d 5 d d dd p1 dd p5
= (144) −i (−ih̄) d z1 d z2 · · ·
h̄ (2π)d (2π)d
e−ip1 ·(x−z1 ) e−ip2 ·(y−z1 ) e−i(p3 +p4 )·(z1 −z2 )
·
(p21 + m2 ) · · · (p25 + m2 )
!2 Z
λΛ4−d dd p1 dd p5
= (144) −i (−ih̄)5 · · · (2π)d δd (−p1 − p2 + p3 + p4 )
h̄ (2π)d (2π)d
e−ip1 ·x e−ip2 ·y
·(2π)d δd (p3 + p4 )
(p21 + m2 ) · · · (p25 + m2 )
!2 Z
λΛ4−d dd p1 e−ip1 ·(x−y)
= (144) −i (−ih̄)5
h̄ (2π)d (p21 + m2 )2
CHAPTER 7. REGULARIZATION METHODS 209

Z
dd p3 1 1
· 2 2
(2π) (p3 + m ) (p3 + m2 )
d 2
Z
dd p5 1
· d 2
(2π) p5 + m2
The second divergent integral we evaluated in equation (7.8). The first is given by
Z
dd p3 1 Λ4−d
Λ4−d = Id (0, 2)
(2π)d (p23 + m2 )2 (2π)d
!d/2−2
1 m2 Γ(2 − d/2)
= i
(4π)d/2 Λ2 Γ(2)
!/2  
i Λ2 1 2
= 4π 2 + ψ(1) + O()
(4π)2 m Γ(2) 
! !
i 1 2 Λ2
= 2
+ ψ(1) + ln 4π 2 + O()
(4π) Γ(2)  m
Putting this together, we see that the double scoop diagram is given by
!2 Z
λΛ4−d
(144) −i dd z1 dd z2 DF (x, z1 )DF (z1 , y)DF (z1 , z2 )2 DF (z2 , z2 )
h̄
 2 Z
λ 4 dd p1 e−ip1 ·(x−y)
= (144) −i (−ih̄)
h̄ (2π)d (p21 + m2 )2
! !
i 1 2 Λ2
· 2
+ ψ(1) + ln 4π 2 + O()
(4π) Γ(2)  m
! !
m2 2 Λ2
·(−h̄) + ψ(2) + ln 4π 2 + O()
(4π)2  m
 2 Z
λ 4 dd p1 e−ip1 ·(x−y) i 1 m2
= (144) −i (−ih̄) (−h̄)
h̄ (2π)d (p21 + m2 )2 (4π)2 Γ(2) (4π)2
"    !! #
2
2 2 Λ2
· + ψ(1) + ψ(2) + 2 ln 4π 2 + O(1) (7.9)
  m

Next, let us apply dimensional regularization to the diagram

!2
1 3 = 1 λΛ4−d
−i (4!)2
2 4 2! h̄
Z
· dd z1 dd z2 DF (x1 , z1 )DF (x2 , z1 )DF (x3 , z2 )DF (x4 , z2 )DF (z1 , z2 )2

In fact, to be more efficient, let us apply dimensional regularization to the amputated diagram

!2
1 λΛ4−d
1× ×3 = −i (4!)2
2× ×4 2! h̄
CHAPTER 7. REGULARIZATION METHODS 210
Z
· dd z1 dd z2 δd (x1 − z1 )δd (x2 − z1 )δd (x3 − z2 )δd (x4 − z2 )DF (z1 , z2 )2

Note that the Dirac delta functions above are not Λ-regulated Dirac delta functions, but rather
are ordinary Dirac delta functions, since we are not using momentum cutoff regularization. We
compute this diagram to be
!2
1 λΛ4−d
−i (4!)2 δd (x1 − x2 )δd (x3 − x4 )DF (x1 , x3 )2
2! h̄
!2
1 λΛ4−d
= −i (4!)2 δd (x1 − x2 )δd (x3 − x4 )
2! h̄
Z   
dd p1 dd p2 −ih̄ −ih̄
· e−i(p1 +p2 )·(x1 −x3 )
(2π)d (2π)d p21 + m2 p22 + m2
Define p = p1 , q = p1 + p2 , then this becomes
!2
1 λΛ4−d
−i (4!)2 δd (x1 − x2 )δd (x3 − x4 )DF (x1 , x3 )2
2! h̄
!2
1 λΛ4−d
= −i (4!)2 δd (x1 − x2 )δd (x3 − x4 )
2! h̄
Z   
dd p dd q −ih̄ −ih̄
· e−iq·(x1 −x3 )
(2π)d (2π)d p + m2
2 (q − p)2 + m2
!2 Z
1 λΛ4−d dd q −iq·(x1 −x3 )
= −i (4!)2 δd (x1 − x2 )δd (x3 − x4 ) e
2! h̄ (2π)d
Z   
dd p −ih̄ −ih̄
·
(2π)d p 2 + m2 (q − p)2 + m2
The p integral is logarithmically divergent, so let us apply dimensional regularization.

To do so, we would like to apply the results for the integral Id (q, α) that we computed previously,
but this integral is not yet of the desired form. In order to put it in the desired form, we will use
a trick due to Feynman. Feynman’s trick is to use the identity:
Z 1
1 dz
=
ab 0 (az + b(1 − z))2
To derive this identity, note that

b dx  Z
1 1 1 1 1
= − =
ab b−a a b b − a a x2
and then define z by x = az + b(1 − z). Applying this to the present case,
Z 1
1 1 dz
=
p + m (p − q)2 + m2
2 2
0 ((p2 + m2 )z + ((p − q)2 + m2 )(1 − z))2
Z 1 dz
=
0 (p2 z + (p − q)2 (1 − z) + m2 )2
Z 1 dz
=
0 (p + m − 2p · q(1 − z) + q 2 (1 − z))2
2 2
CHAPTER 7. REGULARIZATION METHODS 211

Define p0 = p − q(1 − z), then the denominator in the expression above becomes
 
(p0 )2 + m2 + q 2 (1 − z) − (1 − z)2 = (p0 )2 + m2 + q 2 z(1 − z)

Thus, after shifting the p integral, applying our earlier result for Id (q, α), and adding a factor of
Λ4−d to clean up the units, we can write
Z Z Z
4−d d 1 1 4−d
1 dd p
Λ d p 2 = Λ dz
p + m (p − q)2 + m2
2
0 (p2 + m2 + q 2 z(1 − z))
Z !d/2−2
1
d/2 m2 + q 2 z(1 − z) Γ(2 − d/2)
= dz (+i)π
0 Λ2 Γ(2)
Z !d/2−2
Γ(2 − d/2) 1 m2 + q 2 z(1 − z)
= +i dz π d/2
Γ(2) 0 Λ2

In the present case, let us assume d is “near” 4, and define  = 4 − d, as previously. In this case,
d/2 − 2 = −/2, so
Z !d/2−2
1 m2 + q 2 z(1 − z)
dz π d/2
0 Λ2
Z !!
1
2  m2 + q 2 z(1 − z)
= dz π exp − ln π
0 2 Λ2
Z !
 1 m2 + q 2 z(1 − z)
= π − π2 2
dz ln π + O(2 ) (7.10)
2 0 Λ2

using the fact that A = exp ( ln A). If we only keep track of leading 1/ effects, then we see that
Z  
4−d d 1 1 2
Λ d p 2 2 2 2
= iπ 2 + O(1) (7.11)
p + m (p − q) + m 
However, we can also extract the next order if we work harder, using the integral
Z 1   √ !
4 √ 1+a + 1
dz ln 1 + z(1 − z) = −2 + 1 + a ln √ , a>0 (7.12)
0 a 1+a − 1
This can be proven by factoring
  √  √ 
ln a + 4z − 4z 2 = ln 1 + a + (2z − 1) + ln 1 + a − (2z − 1)

and integrating each term separately. Thus,

Z !
1 m2 + q 2 z(1 − z)
dz ln π
0 Λ2
! Z !
m2 1 q2
= ln π 2 + dz ln 1 + 2 z(1 − z)
Λ 0 m
! s p p !
m2 4m2 q 2 + 4m2 + q 2
= ln π 2 − 2 + 1 + 2 ln p p
Λ q q 2 + 4m2 − q 2
CHAPTER 7. REGULARIZATION METHODS 212

Thus,
Z
1 1
Λ4−d dd p
p2 + m2 (p − q)2 + m2
" !
Γ(2 − d/2) 2  m2
= +i π − π 2 ln π 2 − 2 (7.13)
Γ(2) 2 Λ
s p p ! 
4m2 + q2+ 4m2
 + O(2 )
q2
+ 1 + ln p 2 p
q2 2
q + 4m − q 2

**** NOTE TO SELF: there’s a potential problem with the expression above – it’s only valid for
positive q 2 , but at the end we integrate over q’s. I’d say it’s still valid in momentum space
onshell, except that in my conventions, on-shell means q 2 is negative not positive. *Maybe* I can
analytically continue to get results valid more generally? (Or maybe not.)

Finally, putting this all together (and truncating the O(1) terms for brevity), we see that the
amputated diagram is given by
!2
1 λΛ4−d
−i (4!)2 δd (x1 − x2 )δd (x3 − x4 )DF (x1 , x3 )2
2! h̄
 2 Z
1 λ 2 d d d4 q −iq·(x1 −x3 )
= −i (4!) δ (x1 − x2 )δ (x3 − x4 ) e
2! h̄ (2π)4
 
π2 2
·(+i) 4
(−ih̄)2 + O(1)
(2π) 
 
(4!)2 λ2 d 2
= (i) 2
δ (x1 − x2 )δd (x3 − x4 )δd (x1 − x3 ) + O(1)
2 16π 

Then, for later reference, it is straightforward to show – either by repeating the computation
above or by multiplying in the external leg propagators – that the regularized unamputated
diagram is given by

2 2  Z    
1 3 = (i) (4!) λ 2 d4 p1 d4 p4 −ih̄ −ih̄
+ O(1) · · · 2 ··· 2
2 4 2 16π 2  (2π)4 (2π)4 p1 + m 2 p 4 + m2
·e−ip1 ·x1 · · · e−ip4 ·x4 (2π)4 δ4 (p1 + p2 + p3 + p4 )

**** NOTE TO SELF: What sort of Dirac delta function should I be using, to be consistent with
dimensional regularization? Do I want
Z Z
4 d4 p −ip·x dd p −ip·x
δ (x) = e or δd (x) = e
(2π)4 (2π)d

COMMENT explicitly on this choice – it only affects subleading ’s, but those are very important
for anyone who goes past leading order.
CHAPTER 7. REGULARIZATION METHODS 213

A note on notation, before proceeding. When we introduced cutoff regularization, we denoted a

cutoff-regularized field φ by φΛ , to distinguish it from a classical field. To be consistent, here we
should probably denote dimensionally-regularized fields by φΛ, . However, for reasons of brevity,
in the rest of this text dimensional regularization will always be assumed explicitly (unless
otherwise stated), and so we will simply denote all (dimensionally-regularized) fields by φ, the
same as the classical field.

7.4 Dimensional regularization, gamma matrices, and chiral

anomalies

In d dimensions, gµν obeys gµν gµν = d, so when deriving expressions for integrals, one should
replace
1
pµ pν 7→ p2 gµν
d

We think of the gamma matrices in dimensional regularization as a set of d matrices obeying

{γ µ , γ ν } = −2gµν I

and
Tr I = 4
(for dimensional regularization “near” four dimensions). Furthermore, we assume that the traces
are meromorphic in d, that they still possess the same cyclic property as in d = 4, and match
classical results in the special case that d = 4.

Also note that this means that although we are varying d, the gamma matrices are
tied to four dimensions – if we were to dimensionally regularize a theory in two
dimensions, for example, then we would impose a different constraint on Tr I, etc. If
we were to set d = 2 in this theory, we would not get gamma matrices for a
two-dimensional theory. This is unlike the previous section, where we considered
dimensional regularization of bosonic integrals. There, integer values of d would
correspond to classical theories in those dimensions. By contrast, when setting up
gamma matrices in dimensional regularization, the regularized gamma matrices are
intrinsically tied to the dimension we started with.

Some of the standard identities involving gamma matrices are modified:

γ α γ µ γα = (2 − )γ µ
γ α γ µ γ ν γα = 4gµν I + γ µ γ ν
γ α γ µ γ ν γ λ γα = 2γ λ γ ν γ µ − γ µ γ ν γ λ

where  = 4 − d.
CHAPTER 7. REGULARIZATION METHODS 214

Trace identities are largely unmodified. For example,

Tr (γ µ γ ν ) = −4gµν (7.14)
   
Tr γ α γ β γ λ γ δ = 4 gαβ gλδ − gαλ gβδ + gαδ gβλ (7.15)

 
Tr γ α γ β γ λ γ δ γ ρ γ σ

= 4 −gαβ gλδ gρσ + gαβ gλρ gδσ − gαβ gλσ gδρ + gλδ gαρ gβσ − gλδ gασ gβρ
+ gρσ gαλ gβδ − gρσ gαδ gβλ + gαρ gβλ gδσ − gβρ gαλ gδσ − gαρ gβδ gλσ

− gασ gβλ gδρ + gβρ gαδ gλσ + gασ gβδ gλρ + gβσ gαλ gδρ − gβσ gαδ gλρ (7.16)

in all d. Each of the identities above can be derived by first using the cyclic property of the trace
to move a gamma matrix from one side to the other, then systematically using the Dirac algebra
to move it back into its original position.

Furthermore, we need to generalize γ 5 to d dimensions. Doing so requires a certain amount of

care. In particular, we previously defined γ 5 = iγ 0 γ 1 γ 2 γ 3 , but in d dimensions this is no longer
the product of all the gamma matrices. One way to try to proceed would be to define γ 5 to be a
matrix with the property that
{γ 5 , γ µ } = 0
for all µ in all dimensions d, but we will see in exercise *** CITE **** that this leads to a
contradiction.

Instead, we will proceed as follows. We define

γ 5 ≡ iγ 0 γ 1 γ 2 γ 3

which formally looks the same as our previous definition in four dimensions in section *** CITE
***, but which now has the consequence that although γ 5 anticommutes with γ µ for µ = 0, 1, 2, 3,
it commutes with γ µ for other values of µ. (See [Collins][section 13.2] and references therein for
alternate approaches to this problem.)

Applying the Dirac algebra to the definition above, one immediate consequence is that
 2
γ5 = 1

for all d, just as in d = 4.

Another consequence of this definition is as follows. Write γ µ = γkµ + γ⊥ µ

, or equivalently
6 p =6 pk + 6 p⊥ , where k indicates components in the original four dimensions, and ⊥ indicates all
0,1,2,3
other components, so that, for example, γ⊥ = 0, γ 0,1,2,3 = γk0,1,2,3 , and

6 pγ 5 = − γ 5 6 pk + γ 5 6 p⊥
CHAPTER 7. REGULARIZATION METHODS 215

Then, for a trace to be nonzero, it must contain an even number of k components:

   
Tr γkµ1 · · · γkµm γ⊥
ν1 νn
· · · γ⊥ = Tr (γ 5 )2 γkµ1 · · · (γ 5 )2 γkµm (γ 5 )2 γ⊥
ν1 νn
· · · (γ 5 )2 γ⊥
      
= Tr γ 5 γkµ1 γ 5 · · · γ 5 γkµm γ 5 ν1 5
γ 5 γ⊥ νn 5
γ · · · γ 5 γ⊥ γ
 
= (−)m Tr γkµ1 · · · γkµm γ⊥
ν1 νn
· · · γ⊥

and so we see that if m is odd, then

 
Tr γkµ1 · · · γkµm γ⊥
ν1 νn
· · · γ⊥ = 0

Another useful identity is

Tr γ 5 = 0 (7.17)
in all d. We can prove this as follows. Let p, q be four-vectors, then
   
Tr γ 5 6 pk 6 qk = −Tr γ 5 6 qk 6 pk using cyclic property of trace
  
= −Tr γ 5 −2qk · pk − 6 pk 6 qk
 
= Tr γ 5 6 pk 6 qk + 2qk · pk Tr γ 5

Taking, for example, p = q spacelike, so that pk · qk 6= 0, we find the desired result.

One can also show  

Tr γ 5 γ µ γ ν = 0 (7.18)
for all µ, ν, in all d. Some other handy identities are
 
µ1 µ2m ν1 ν2
Tr γ 5 γ⊥ · · · γ⊥ γk γk = 0 (7.19)

for m ≥ 0 in all d,  
µ1 µ2m
Tr γ 5 γ⊥ · · · γ⊥ = 0 (7.20)
for m ≥ 0 in all d, and
   
α β ν1
Tr γ⊥ γ⊥ γk · · · γkν2n = −gαβ Tr γkν1 · · · γkν2n (7.21)

for n ≥ 0 in all d.

Some useful consequences of the identities above include, for example,

   
Tr γ 5 γ α γ β γ λ γ δ = Tr γ 5 γkα γkβ γkλ γkδ
(
−4iαβλδ α, β, λ, δ ∈ {0, 1, 2, 3}
= (7.22)
0 otherwise
CHAPTER 7. REGULARIZATION METHODS 216

and
   
Tr γ 5 γ α γ β γ λ γ δ γ ρ γ σ = Tr γ 5 γkα γkβ γkλ γkδ γkρ γkσ
 
α β λ δ ρ σ
+ Tr γ 5 γ⊥ γ⊥ γk γk γk γk + permutations
 
= Tr γ 5 γkα γkβ γkλ γkδ γkρ γkσ
 
− gαβ Tr γ 5 γkλ γkδ γkρ γkσ − permutations (7.23)

Now, let us apply this to the study of the chiral anomaly in a dimensionally-regularized theory.
Let us assume the same Lagrangian density as in section **** CITE ***, where we studied
anomalies in theories with a cutoff regularization.

Briefly, in that section we learned that the Ward-Takahashi identities for the vector and axial
symmetries could not all simultaneously be satisfied; due to an ambiguity caused by lack of
translation invariance in the linearly divergent integrals, one must work harder to uniquely define
the effect of the cutoff regularization, and we made the choice of preserving the vector symmetries
while letting the axial symmetry be violated.

The argument for the anomaly in section *** CITE *** revolved around lack of
translation-invariance in linearly divergent integrals, in cutoff regularization. In dimensional
regularization, on the other hand, all integrals are automatically translation invariant, including
the linearly divergent ones. Instead, the anomaly will turn out to arise from the difficulties in
defining γ 5 in a dimensionally-regularized theory.

Let us work through the details of the anomaly computation in the dimensionally-regularized
theory. In section *** CITE ****, the relevant part of a triangle Feynman diagram was labelled
S µλν (q1 , q2 ) in (5.44). For completeness, we repeat it here, in dimensional regularization:
Z  h i
dd p 1 1 1
S µνλ (q1 , q2 ) ≡ Tr (6 p− 6 q 1 )γ µ 5
γ (6 p+ 6 q 2 )γ λ
6 pγ ν
(2π)d (p − q1 )2 (p + q2 )2 p2
The quantity T µλν (q1 , q2 ) was defined in (5.45) to be the sum of two sets of S’s corresponding to
two Feynman diagram contributions:
T µλν (q1 , q2 ) ≡ S µλν (q1 , q2 ) + S µνλ (q2 , q1 )
If both the vector and axial anomalies were anomaly-free, then the Ward-Takahashi
identities (5.46, 5.47, 5.48) would be satisfied. For completeness, we repeat them here. The Ward
identity for the axial current, expressing ∂µ hJ5µ i = 0, is
(q1µ + q2µ ) T µλν (q1 , q2 ) = 0
The Ward identity for the vector current, expressing ∂µ hJ µ i = 0, can also be checked using the
same Feynman diagrams, as the triangle diagram is roughly a contribution to the correlation
function hJ5µ J λ J ν i. The statement that the two copies of J be divergence-free is
q1ν T µλν (q1 , q2 ) = 0
q2λ T µλν (q1 , q2 ) = 0
CHAPTER 7. REGULARIZATION METHODS 217

In section *** CITE ***, it was argued that

(q1µ + q2µ ) S µλν (q1 , q2 ) = 0

by using the identity (5.49), namely

(6 q1 + 6 q2 ) γ 5 = − (6 p− 6 q1 ) γ 5 − γ 5 (6 p+ 6 q2 )

to simplify the integrand. In dimensional regularization, identity (5.49) is replaced by

(6 q1 + 6 q2 ) γ 5 = − (6 p− 6 q1 ) γ 5 − γ 5 (6 p+ 6 q2 ) + 2γ 5 6 p⊥

The first two terms are the same as before; the third term is new, and reflects the technical issues
in defining γ 5 in dimensional regularization. The contribution of the first two terms to
(q1µ + q2µ )S µλν vanishes, by essentially the same argument2 as in section *** CITE ****.

Thus, we see that in dimensional regularization,

(q1µ + q2µ ) S µλν (q1 , q2 )

Z  h i
dd p 1 1 1 5 λ ν
= Tr (6 p− 6 q1 )(2γ 6 p⊥ )(6 p+ 6 q2 )γ 6 pγ
(2π)d (p − q1 )2 (p + q2 )2 p2

Next, let us combine denominators with Feynman’s trick:

Z 1 Z 1−x h i−3
1 1 1
= 2 dx dy x(p − q1 )2 + y(p + q2 )2 + p2 (1 − x − y)
(p − q1 )2 (p + q2 )2 p2 0 0
Z 1 Z 1−x h i−3
= 2 dx dy p2 − 2xp · q1 + 2yp · q2 + xq12 + yq22
0 0

Next, let us use translation invariance of the (dimensionally-regulated) integral to replace

p 7→ p0 = p + yq2 − xq1

so that we can write

(q1µ + q2µ ) S µλν (q1 , q2 )

Z Z Z
1 1−x dd p h 2 2 2 2
i−3
= 2 dx dy p + xq 1 + yq 2 − (yq 2 − xq 1 )
0 0 (2π)d
h i
·Tr (6 p − y 6 q2 + x 6 q1 − 6 q1 )(2γ 5 6 p⊥ )(6 p − y 6 q2 + x 6 q1 + 6 q2 )γ λ (6 p − y 6 q2 + x 6 q1 )γ ν

In writing the above, we have assumed, without meaningful loss of generality, that q1,2 are
ordinary four-vectors, i.e. 6 q1⊥ = 0 =6 q2⊥ . Furthermore, we shall also assume that λ and ν are
also ordinary four-vector indices, so that γ⊥ λ = 0 = γν .
⊥
2
The only point in the argument that is a concern is the trace identities; however, we see from identity (7.22) that
the trace identities used are unchanged in dimensional regularization.
CHAPTER 7. REGULARIZATION METHODS 218

Next, we shall simplify the trace above. In light of identity (7.23), since there is already one 6 p⊥
factor, the only nonzero contribution will arise from terms involving precisely one additional ⊥
factor, with all other factors k. Thus, we find
h i
Tr (6 p − y 6 q2 + x 6 q1 − 6 q1 )(2γ 5 6 p⊥ )(6 p − y 6 q2 + x 6 q1 + 6 q2 )γ λ (6 p − y 6 q2 + x 6 q1 )γ ν
h i
= Tr (6 p⊥ )(2γ 5 6 p⊥ )(6 pk − y 6 q2 + x 6 q1 + 6 q2 )γ λ (6 pk − y 6 q2 + x 6 q1 )γ ν
h i
+ Tr (6 pk − y 6 q2 + x 6 q1 − 6 q1 )(2γ 5 6 p⊥ )(6 p⊥ )γ λ (6 pk − y 6 q2 + x 6 q1 )γ ν
h i
+ Tr (6 pk − y 6 q2 + x 6 q1 − 6 q1 )(2γ 5 6 p⊥ )(6 pk − y 6 q2 + x 6 q1 + 6 q2 )γ λ (6 p⊥ )γ ν
= 2(p⊥ )2 (4i)(pk − yq2 + xq1 + q2 )α (pk − yq2 + xq1 )β αλβν
+ 2(p⊥ )2 (4i)(pk − yq2 + xq1 )α (pk − yq2 + xq1 − q1 )β λανβ
+ 2(p⊥ )2 (4i)(pk − yq2 + xq1 + q2 )α (pk − yq2 + xq1 − q1 )β αλνβ

= 8i(p⊥ )2 q2α (pk − yq2 + xq1 )β αλβν − q1β (pk − yq2 + xq1 )α λανβ

+ q2α (pk − yq2 + xq1 )β αλνβ − q1β (pk − yq2 + xq1 )α αλνβ − q2α q1β αλνβ
= −8i(p⊥ )2 q2α q1β αλνβ

***** NOTE: I need to include some Λ factors, look at the expression below. Go back and
include them.

Putting this all together, we find

(q1µ + q2µ ) S µλν (q1 , q2 )

Z Z Z h i−3
αλνβ
1 1−x dd p 2 2 2 2 2
= −16iq2α q1β  dx dy (p ⊥ ) p + xq 1 + yq 2 − (yq 2 − xq 1 )
0 0 (2π)d
Inside the symmetric integration, we can replace
d−4 2
(p⊥ )2 7→ p (7.24)
d
so that the integral becomes

(q1µ + q2µ ) S µλν (q1 , q2 )

Z Z Z
(d − 4) 1 1−x dd p 2 h 2 i−3
= −16iq2α q1β αλνβ dx dy p p + xq 2
1 + yq 2
2 − (yq 2 − xq 1 )2
d 0 0 (2π)d
Z 1 Z 1−x
(d − 4) i Γ(3 − d/2 − 1)Γ(d/2 + 1)
= −16iq2α q1β αλνβ dx dy d
d 0 0 (4π) Γ(d/2)Γ(3)
 d/2+1−3
· xq12 + yq22 − (yq2 − xq1 )2
16 Γ(3 − d/2)Γ(d/2 + 1)
= − d
q2α q1β αλνβ
(4π) Γ(d/2 + 1)Γ(3)
Z 1 Z 1−x  d/2−2
· dx dy xq12 + yq22 − (yq2 − xq1 )2
0 0
CHAPTER 7. REGULARIZATION METHODS 219

where in the next-to-last line we evaluated the integral using equation **** CITE ****.

The expression above has the property that it has a finite limit as d → 4; taking that limit, we find

16 Γ(3 − 2) 1
(q1µ + q2µ ) S µλν (q1 , q2 ) = − q2α q1β αλνβ ·
(4π)4 Γ(3) 2
1
= − q2α q1β αλνβ
64π 4
and hence
1 1
(q1µ + q2µ ) T µλν (q1 , q2 ) = − 4
q2α q1β αλνβ − q1α q2β ανλβ
64π 64π 4
1
= − q1β q2α αβλν
32π 4

****** ALMOST matches earlier result, but not quite – off by a sign, and a factor of 16π 2 .

***** Modulo factors******, this matches the result (5.50), and so repeating the result of the
analysis from that section we can immediately read off that
h̄
∂µ hJ5µ (x)i = − ∂µ Aν (x)∂ρ Aλ (x)ρµλν
32π 4

***** I’ve modified the factors above to be consistent with my work in this section.

**** NEED to track down this difference in factors. Where is it coming from? Is there an error in
dim’l reg’ computations, or in the earlier computation of chiral anomaly?

Let us also check whether the vector Ward identities are satisfied, beginning with the statement

q1ν T µλν (q1 , q2 ) = 0

It is straightforward to check that the analysis of section *** CITE *** for this case can be
repeated without any significant modification, modulo the fact that all integrals in dimensional
regularization are translation-invariant. Back in section *** CITE ****, the only nonzero
contribution to q1ν S µλν (q1 , q2 ) arose from the lack of translation-invariance in cutoff-regularized
linearly divergent integrals; here, translation invariance implies that

q1ν S µλν (q1 , q2 ) = 0

and hence the vector Ward identity above is satisfied. Similarly, the remaining vector Ward
identity is also satisfied.

Now, let us review what we have found, and compare to our earlier results in section *** CITE
**** for the same anomaly, in cutoff regularization. There, there was an ambiguity in the
specification of the loop integrals, arising because of a lack of translation invariance in linearly
divergent integrals in cutoff regularization. Because of that ambiguity, one could choose whether
to satisfy axial or vector current conservation – both could not simultaneously hold, but which
CHAPTER 7. REGULARIZATION METHODS 220

held was determined by how the ambiguity was resolved. By contrast, all integrals are
translation-invariant in dimensional regularization, even the linearly divergent ones. The anomaly
arises here because of difficulties in defining γ 5 in dimensional regularization. Taking that into
account, we find that vector Ward identities are automatically satisfied in dimensional
regularization, and the axial Ward identity is broken. There is no longer a choice of which
symmetry must be broken (reflecting the fact that dimensional regularization does not break
gauge symmetries, a fact we shall return to later in section *** CITE ****).

Ultimately, in cutoff regularization and dimensional regularization, we follow two different

computational methods, but arrive at the same final result: the axial Ward identity cannot hold,
and axial current conservation is violated as

∂µ hJ5µ (x)i ∝ Fµν Fρσ µνρσ

****** Also mention relation to path integrals, if in the fermion section I discussed Fujikawa’s
perspective.

7.5 An alternative method: zeta function regularization

So far in this text we have seen several regularization schemes: for most of the text so far we have
used cutoff regularization, and this chapter has introduced Pauli-Villars regularzation and
dimensional regularization. There exist additional regularization schemes beyond these. One
further regularization method is called lattice regularization and will be a powerful tool for
understanding the gauge theories that will be discussed later. We shall discuss lattice
regularization in chapter **** FILL IN ****.

In this section, we shall outline yet another regularization method, known as zeta function
regularization. Zeta function regularization is not used as commonly as the other methods we
have discussed, but it does pop up sufficiently often to warrant specific mention.

**** For more material, see the discussion in Berline-Getzler-Vergne, Heat kernels and Dirac
operators, section 9.6, pp 300-, “zeta function of a laplacian.”

**** Also see Nakahara section 1.4, 1.5 for some discussion, examples.

When encountering infinite products (as in operator determinants) and infinite sums, a powerful
technique called zeta function regularization is often used to make sense of them. The upshot is
relatively simple, though I’m going to take the time to explain the background behind the method.
P
At its most basic level, the idea is that given divergent sums like n>0 n, for example, we’re going
to use properties of zeta functions to formally make sense of them. [Such prescriptions aren’t
unique, since the sum is divergent; zeta functions will give a prescription.] In principle the idea
can be applied to operator determinants. Suppose the eigenfunctions and eigenvalues of an
operator A are given by
Afn (x) = an fn (x)
CHAPTER 7. REGULARIZATION METHODS 221

Construct
X 1
ζA (s) =
n asn
the zeta function associated to operator A. It can be shown that
Y
0
det A = an = exp −ζA (0)
n

so if one knows that infinite sum ζA (s), and can make sense of it, then you can get a finite
expression for det A. (See Ramond’s book for further comments on this, and also P. di Francesco,
P. Mathieu, D. Senechal, Conformal Field Theory, section 6.4.)

So far I’ve briefly outlined why being able to make divergent sums finite might be a good thing,
but how exactly do we do it? First, recall the ordinary Riemann zeta function is defined as
X
ζ(s) = n−s
n>0

(see appendix **** CITE ****) but this only converges for s ≥ 2. Unfortunately, we sometimes
P
run across sums such as n>0 n, which formally looks like ζ(−1), a divergent sum. Now, the trick
is we’re going to replace the ordinary Riemann zeta function with the Hurwitz zeta function (also
called the generalized Riemann zeta function). That by itself still will not converge at the desired
places, but we can analytically continue it to something that does, and we’ll use that last
something to formally make sense of divergent sums. It sounds complicated, but the final result is
actually very simple.

The Hurwitz zeta function, or generalized Riemann zeta function, is defined by

∞
X
ζ(s, q) = (q + n)−s
n=0

so that ζ(s, 1) is the same as the ordinary Riemann zeta function ζ(s). (Also, we now think of s
as a complex number, not necessarily an integer.) This converges so long as <s > 1 and q is not a
negative integer or zero, but unfortunately we will typically want to understand cases in which
<s < 0.

The next step is to analytically continue the Hurwitz zeta function, to a function we shall denote
ζ̃(s, q). This is discussed in appendix B.3.2; the pertinent result is that
Bn+1 (q)
ζ̃(−n, q) = −
n+1
where n is a positive integer, and Bn is the nth Bernoulli polynomial.

In any event, now we finally have a sophisticated way of making sense of divergent sums. For
example:
X
n = ζ̃(−1, 1)
n>0
X
(n − 1/2) = ζ̃(−1, 1/2)
n>0
CHAPTER 7. REGULARIZATION METHODS 222

and furthermore the results are finite numbers. For example, using B2 (x) = x2 − x + 1/6, you can
calculate using the expression above that
X
(n − 1/2) = 1/24
n>0

(Note as always that strictly speaking the expression on the left diverges, it’s only because of this
choice of regularization that we can get a finite number out of it.)

For more information on zeta function regularization, see for example

• E. Elizalde, Ten physical applications of spectral zeta functions, Springer-Verlag, Berlin,

1995.

**** More refs?

7.6 Further reading

*** In 7650 at Utah, regularization and renormalization were one long section spread over 2
weeks. Reading: PS sections 8, 10.1-10.2, 10.4-10.5; Ryder 9.1-9.3, 9.A, 9.B. Skim PS 6.3, 7.1, 7.5.
Math track: read Munkres 2-9, 2-10, 3-1, 3-5 over the course of the two weeks.

7.7 Notes
1. Several weeks ago, when discussing path integrals, I said that it was misleading to believe
that a path integral weights paths by the value of the classical action. At the time, I
pointed out that (a) this sounds like putting the cart before the horse, since we’re supposed
to get classical physics from quantum physics not vice-versa, and (b) the fields one sums
over are typically not differentiable, so any action phrased in terms of derivatives of fields
should be viewed with a bit of suspicion. At the time, I said that later we would learn that
talking about a classical action is misleading, that the classical action would eventually have
to be replaced by a ‘regularized’ action. Well, we’ve now gotten to the point that I can
explain what I meant.

2. To “regularize” a theory means to replace classical expressions with alternative expressions

in which the divergences of Feynman diagrams are rendered finite. In general, there are
many different ways to regularize theories. This week, you’ll learn about Pauli-Villars
regularization and dimensional regularization. Later, we’ll learn more. (For example, when
discussing nonabelian gauge theories, we’ll see that replacing spacetime with a lattice yields
another regularization, known as a lattice regularization, which turns out to be important
for computational work.)

3. The idea behind Pauli-Villars regularization is that since the divergences you’ve seen so far
arise from momenta going to infinity (so-called “ultraviolet” or “UV” divergences), one
CHAPTER 7. REGULARIZATION METHODS 223

should impose a cutoff on those momenta. Intuitively, this is akin to saying there’s a
shortest distance in spacetime, and nothing smaller. Pauli-Villars is comparatively intuitive,
and we will typically use the language of “imposing a cutoff,” but as a practical
computational matter Pauli-Villars is often very cumbersome. (For the simple examples
we’ll see this week they’ll be comparable, but when discussing nonabelian gauge theories
Pauli-Villars becomes grotesque.) This effectively truncates the infinitely-jagged fields
counted in the classical path integral.

4. Dimensional regularization is, computationally, much more efficient than Pauli-Villars in

general. The drawback is that it is far less intuitive. The idea is a bit more abstract: we
replace the divergent integrals with other integrals in which we’ve formally allowed the
dimension to stop being integral. (More properly, we work in spherical coordinates, split off
the angular part, then rethink the radial integral with gamma functions, which allows us to
play some games that formally resemble making the dimension non-integral.) How literally
should one interpret the idea of changing the dimension? (ie, is the dimension really a
fractal dimension or some such?) Nobody knows; everyone treats it as an efficient
computational technique, the deeper understanding, if there is one, does not yet exist.

5. In condensed matter systems there is often already a natural cutoff, provided by the lattice
size in the crystal. So, one might naively think that much of this discussion is irrelevant for
condensed matter physics. However, we saw last week that renormalization is teaching us
that physics is scale-dependent (the “renormalization group”), and that lesson is very
applicable to condensed matter physics.

6. What’s the difference between “regularization” and “renormalization”? There’s really a

two-step process here. First, we must alter the action in some way – impose a cutoff, play
formal games with the dimension, whatever – so as to make the infinities in Feynman
diagrams finite (reflecting the fact that the path integral can probably only be rigorously
defined after such regularization), then we must alter the coupling constants, masses, and
wavefunction scales so as to insure that physical quantities remain finite when we remove
the regularization. The first step is “regularization,” the second step is “renormalization.”
Neither step is unique – there are several different regularizations, and there are also several
ways to renormalize. One way is to alter the original (“bare”) masses, coupling constants,
and wavefunction scales so as correlation functions match certain fixed target values. If
we’re renormalizing QED, then, well, we can measure the electron mass and couplings, so
we should renormalize in such a way as to insure that they pop out. Another way is to
merely subtract off whatever is becoming divergent. If, for example, adding δλ = −1 to λ
will remove a pole from a loop amplitude, then that’s one thing we could do. But, if that
makes the amplitude finite, then so will δλ = −1 + 1. The choice of what to subtract off is
called a “subtraction scheme,” and the obvious minimal choice is known as “minimal
subtraction.” Subtraction schemes differ from one another in what finite parts to subtract,
known as “finite renormalizations.”

7. A practical aside. Peskin-Schroeder and Ryder take two slightly different approaches to
dimensional regularization: PS Wick rotates to Euclidean space and calculates integrals
there (as we do here), whereas Ryder works in Minkowski space.
CHAPTER 7. REGULARIZATION METHODS 224

8. Overlapping divergences. If you think through the details of how to carry out
renormalization to higher loops, you’ll quickly see there’s a potential rough spot. Some
higher-loop diagrams have divergent bits that are easily recognized as being cancelled by
lower-loop counterterms – for ex, if in the middle of some complicated diagram there’s a
one-loop propagator correction, then, the corresponding one-loop counterterm will make
that part of the complicated higher-loop diagram finite. However, it’s possible to have
situations in which, for example, a 2-loop diagram looks like a pair of divergent one-loop
diagrams, except that there’s a common propagator between them. These are called
overlapping divergences. In the limit that either of the loop momenta become large, the
corresponding subdiagram shrinks to zero size, and a one-loop counterterm can help cancel
the divergence. Residual divergences not taken care of by the lower-loop counterterms
should be cancelled by a higher-order counterterm. ***** ADD MATERIAL ON
OVERLAPPING DIVERGENCES *****

7.8 Exercises

1. In this problem we shall generalize Feynman’s trick, namely the identity

Z 1
1
= dx (ax + b(1 − x))−2
ab 0

as discussed in the text.

(a) Show that

Z 1 Z 1−x
1
= 2 dx dy (ax + by + c(1 − x − y))−3
abc 0 0

(b) Use induction to show that

Z 1 Z 1−x1 Z 1−x1 −···−xn−2
1
= (n − 1)! dx1 dx2 · · · dxn−1
a1 a2 · · · an 0 0 0
(a1 x1 + a2 x2 + an−1 xn−1 + an (1 − x1 − x2 − · · · − xn−1 ))−n

2. In this problem we shall study another generalization of Feynman’s trick.

(a) For real numbers α, β, show that

Z
1 Γ(α + β) 1 xα−1 (1 − x)β−1
= dx
α
a bβ Γ(α)Γ(β) 0 (ax + b(1 − x))α+β

Hint: make the change of variables

ax
z =
ax + b(1 − x)
CHAPTER 7. REGULARIZATION METHODS 225

(b) For real numbers α1 , · · · , αn , show that

Z
1 Γ(α1 + · · · + αn )
α1 = dx1 · · · dxn δ(1 − x1 − · · · − xn )
a1 · · · aαnn Γ(α1 ) · · · Γ(αn )
x1α1 −1 x2α2 −1 · · · xnαn −1
·
(a1 x1 + a2 x2 + · · · + an xn )α1 +···+αn

Hint: use the identity Z

Γ(α) ∞
= dttα−1 e−at
aα 0
to get an expression for
Γ(α1 ) · · · Γ(αn )
aα1 1 · · · aαnn
then multiply the other side by
Z ∞  X 
1 = ds δ s − ti
0

make the change of variables ti = sxi , and do the s integral.

**** The further generalization above is proven in Nayak’s lecture notes, p 60 – compare
this with that. Above based on Sredniki (14.1). Seems like the first two problems should be
truncated to something more efficient.

3. (**** OMIT for fall 2008 ****) Pauli-Villars regularization The basic idea behind
Pauli-Villars regularization of scalar field theory is to replace the propagator
Z
d4 p 1
DF (x − y) = −ih̄ exp(−ip · (x − y))
(2π)4 p2 + m2

by the propagator
Z  
d4 p 1 1
−ih̄ − 2 exp(−ip · (x − y))
(2π)4 p 2 + m2 p + Λ2

(where Λ2 is assumed to be much larger than m2 ). When Λ2 → ∞, the second propagator

reduces to the former.

(a) We can interpret the procedure above as adding a second scalar field, of mass Λ and all
same couplings as the first, but with a wrong-sign kinetic term. Comment on the
meaning of that sign.
(b) Show that
1 1 1 Λ2 − m2
− =
p 2 + m2 p2 + Λ2 p2 + m2 p2 + Λ2
Thus, by increasing the number of powers of p in the denominator, we make divergent
loop integrals more nearly convergent.
CHAPTER 7. REGULARIZATION METHODS 226

(c) The one-loop correction to the propagator in λφ4 theory in four dimensions has a
divergent loop integral which for large momenta takes the form
Z
d4 p
∼
p2
The Pauli-Villars regulator discussed so far improves that to
Z
d4 p
∼
p4
but that’s still log-divergent.
In such cases, we instead make the replacement

1 1 1 Λ22 − m2 1 m2 − Λ21
7→ − −
p + m2
2 p 2 + m2 p2 + Λ21 Λ22 − Λ21 p2 − Λ22 Λ22 − Λ21
1 Λ21 − m2 Λ22 − m2
= 2
p + m2 p2 + Λ21 p2 + Λ22

which makes the integral convergent. Compute the mass renormalization of λφ4 in four
dimensions, to first order in λ, using minimal subtraction and the Pauli-Villars
regulator above. After performing the substitution in the second form, for simplicity
take Λ1 = Λ2 . You will find that calculations are greatly simplified if you use the
Feynman parameter trick discussed above, and also feel free to use equation (7.2).
(**** PROBLEM: result converges as Λ → ∞, instead of diverging *****)

Although Pauli-Villars appears more intuitive than dimensional regularization (it imposes a
cutoff on momenta, which seems very natural), as a practical computational matter it can
be much more cumbersome in general. For example in gauge theories, Pauli-Villars
regularization automatically destroys gauge invariance, as the second “fake” photon one
adds is massive. Although the language of Pauli-Villars regularization is still sometimes
used, as a practical matter nowadays everyone uses dimensional regularization, so we shall
stick to that from now on.

4. Show that no quadratic, quartic, etc divergences are encountered in dimensional

regularization, in the sense that
Z
dd p(p2 )N = 0 for N > −d/2
R
by taking the limit m → 0 of the integral dd p(p2 + m2 )N , evaluated with dimensional
regularization.

5. (a) Show that

Z
dd p (p2 )k (m2 )d/2+k−n Γ(n − d/2 − k)Γ(d/2 + k)
= i
(2π)d (p2 + m2 )n (4π)d/2 Γ(d/2)Γ(n)
CHAPTER 7. REGULARIZATION METHODS 227

(b) Show that

Z
dd p pµ pν (m2 )d/2+1−n gµν Γ(n − d/2 − 1)
= i
(2π)d (p2 + m2 )n (4π)d/2 2 Γ(n)
Hint: use Lorentz invariance to argue that
Z Z
dd p pµ pν dd p Agµν p2
=
(2π)d (p2 + m2 )n (2π)d (p2 + m2 )n
for some constant A, then solve for A and use previous results to evaluate the integral.
(c) Show that
Z  
dd p pµ pν pρ pσ (m2 )d/2+2−n Γ(n − d/2 − 2) 1
= i (gµν gρσ + gµρ gνσ + gµσ gνρ )
(2π)d (p2 + m2 )n (4π)d/2 Γ(n) 4
Hint: use Lorentz invariance to argue that
Z Z
dd p pµ pν pρ pσ dd p A(p2 )2 (gµν gρσ + gµρ gνσ + gµσ gνρ )
=
(2π)d (p2 + m2 )n (2π)d (p2 + m2 )n
for some constant A, then solve for A and use previous results to evaluate the integral.

6. Gamma matrices in dimensional regularization In a dimensionally-regularized theory

‘near’ four dimensions, show that the gamma matrices have the properties below. Take
d = 4 − .
(a)

γ α γ µ γα = (2 − )γ µ

(b)

γ α γ µ γ ν γα = 4gµν I + γ µ γ ν
(c)

γ α γ µ γ ν γ λ γα = 2γ λ γ ν γ µ − γ µ γ ν γ λ

7. Properties of γ 5 in dimensional regularization In this problem, assume that for all d,

for all γ µ ,
{γ 5 , γ µ } = 0 (7.25)
We will show that this leads to a contradiction.

(a) Starting from

dTr γ 5 γ µ γ ν = −Tr γ 5 γ µ γ ν γλ γ λ = Tr γ 5 γ λ γ µ γ ν γλ
use the identity (7.25) and the Dirac algebra in d dimensions to show that

(d − 2)Tr γ 5 γ µ γ ν = 0

This implies that Tr γ 5 γ µ γ ν = 0 for d 6= 2, and as we assume the trace is meromorphic

in d, that implies the trace must vanish for all d.
CHAPTER 7. REGULARIZATION METHODS 228

(b) Starting from

dTr γ 5 γ α γ β γ µ γ ν = −Tr γ 5 γ α γ β γ µ γ ν γλ γ λ = +Tr γ 5 γ λ γ α γ β γ µ γ ν γλ

use the identity (7.25) and the Dirac algebra in d dimensions to show that

(d − 4)Tr γ 5 γ α γ β γ µ γ ν = 0

Thus, the trace must vanish for d 6= 4, and as the trace is assumed to be meromorphic
in d, the trace must vanish for all d. But this contradicts the results of section ***
CITE ***, where we saw explicitly that in four dimensions, the trace above is nonzero.
Thus, we have a contradiction, and identity (7.25) cannot hold for all γ µ in all
dimensions d.

8. Trace identities in dimensional regularization Using gamma matrices in dimensional

regularization, show the following identities.

(a) Show that

Tr (γ µ γ ν ) = −4gµν
in all d. (Hint: This problem and the next two can be solved by first cyclically
permuting the last factor through the trace to the front, then using the Dirac algebra
to move it to its original position.)
(b) Show that    
Tr γ α γ β γ λ γ δ = 4 gαβ gλδ − gαλ gβδ + gαδ gβλ
in all d.
(c) Show that
 
Tr γ α γ β γ λ γ δ γ ρ γ σ

= 4 −gαβ gλδ gρσ + gαβ gλρ gδσ − gαβ gλσ gδρ + gλδ gαρ gβσ − gλδ gασ gβρ
+ gρσ gαλ gβδ − gρσ gαδ gβλ + gαρ gβλ gδσ − gβρ gαλ gδσ − gαρ gβδ gλσ

− gασ gβλ gδρ + gβρ gαδ gλσ + gασ gβδ gλρ + gβσ gαλ gδρ − gβσ gαδ gλρ

in all d.
(d) Show that  
Tr γ 5 γ µ γ ν = 0
for all µ, ν, in all d.
(e) Show that  
µ1 µ2m ν1 ν2
Tr γ 5 γ⊥ · · · γ⊥ γk γk = 0
for m ≥ 0, in all d, by induction on m.
(f) Show that  
µ1 µ2m
Tr γ 5 γ⊥ · · · γ⊥ = 0
for m ≥ 0 in all d, by induction on m.
CHAPTER 7. REGULARIZATION METHODS 229

(g) Show that    

α β ν1
Tr γ⊥ γ⊥ γk · · · γkν2n = −gαβ Tr γkν1 · · · γkν2n
for n ≥ 0 in all d.

9. Casimir effect In derivingR

the Hamiltonian in canonical quantization, we would often get
an infinite constant, ∼ dd p(1), which we have so far ignored. Sometimes, however, we can
extract some physics from that constant. We will consider, for simplicity, a free massless
real scalar field in 1 + 1 dimensions between parallel plates, and we’ll see that quantum
effects cause there to be a force on the plates. (This Casimir effect has been measured
experimentally. *** REFS? ***)
**** See Itzykson-Zuber section 3-2-4 pp 138-142 for a more detailed discussion, that should
be translated into this problem here. In add’n, Radovanovic’s book section 9 page 53 has a
good problem.
Consider three parallel plates, at positions x = 0, d, L, with the property that the scalar
field is forced to vanish at the position of each plate. One can show that the energies of field
modes between plates separated by distance d is ∼ nπ/d. The constant term appearing in
the Hamiltonian is then
∞
X h̄ nπ
∼ f (d) =
n=1
2 d
so the total contribution from the regions between all the plates is

E = f (d) + f (L − d)

Both terms in the expression above are infinite, but by now you should have an idea how to
handle infinities: we regularize. Define
∞
X h̄ nπ
f˜(d) = exp(−anπ/d)
n=1
2 d

where a is a small positive number. Note f˜ → f in the limit a → 0.

(a) Show that

h̄π exp(aπ/d)
f˜(d) =
2d (exp(aπ/d) − 1)2

(b) Estimate f˜(d) for small a; compute the leading two terms.
(c) The force between the plates is −∂E/∂d. Calculate it, to leading orders. You should
find that the result is well-defined (and nonzero) in the limit we remove the regulator,
i.e. the limit a → 0.
(d) Now, let us repeat that analysis with zeta function regularization. Use zeta function
regularization to calculate E and −∂E/∂d, and compare your result to the result you
just obtained in the previous part using a different method.
CHAPTER 7. REGULARIZATION METHODS 230

10. When phonons are described in terms of vibrations of ions sitting at lattice points, there is a
natural momentum cutoff: momenta are only integrated over a single Brillouin zone, not all
of R3 . When phonons are described with a continuum model, that fact becomes more
obscure, and a UV cutoff must be imposed by hand. In this problem you will compute an
estimate for such a cutoff.
Consider a cubical Bravais lattice in three dimensions, with lattice spacing a in each
direction. Compute
Z
d3 k
3
(1)
B (2π)

for this case. Now, let’s approximate the Brillouin zone by a solid sphere, of radius ωD /c
(the cutoff frequency, known in this context as the Debye frequency). Compute
Z ωD /c d3 k
(1)
(2π)3

over this volume. By setting the two integrals equal, derive an approximate expression for
ωD in terms of the volume of a cell of the Bravais lattice. (See Fetter-Walecka section 44 for
more information.)