2012 Matthew Schwartz

Appendix B: Regularization
1 Integration parameters
Loop integrals in quantum eld theory often involve products of propagators, each of whose denominators are quadratic in loop momenta. Products of such propagators therefore involve denominators with powers of loop momenta at high order. To evaluate these it is helpful to introduce auxiliary integration parameters which simplify the evaluation of these products.

1.1 Feynman parameters

Feynman parameters are based on a number of easy to verify mathematical identities 1 AB
1

=
0 1 0

dx

1 = [A + (B A)x]2

1 0

dxdy(x + y 1)

1 [xA + yB]2 (1)

1 = AB n 1 = ABC

dxdy(x + y 1)
1

ny n1 [xA + yB]n+1 2 [xA + yB + zC]3

dxdydz(x + y + z 1)

The reason these are useful is that they let us complete the square in the denominator. 1 d4k 1 = 4 k 2 (k p)2 (2) = d4k (2)4 dk (2)4
4 1

dx
0 1 0

[k 2 + x( (k

1 dx [(k xp)2 ]2

1 p)2 k 2 )]2

(2)

where = p2x(1 x). Then we can shift k k + xp leaving an integral that only depends on k 2.

1.2 Schwinger parameters

Another useful set of integration parameters are called Schwinger parameters. They are based on the following mathematical identities which hold when Im(A) > 0 i = A i A
2

dsei sA
0

(3) (4)

=
0

sdseisA

Further identities can be derived by taking additional derivatives with respect to A. Also, Eq. (3) implies 1 = ds dteisA+it B (5) AB 0 0 when Im(A) > 0 and Im(B) > 0 (i.e. with Feynman propagators). These Schwinger parameters s and t have a nice physical interpretation: s and t are the proper times of the particles as they travels along their paths in the Feynman graph. This Schwinger proper-time interpretation is discussed in Lecture IV-8. 1

Section 2

Note that writing s + t = and x = s + t , or t = x and s = (1 x) , we get 1 = AB =

0 1

d
0 1 0

dxei (A+ (B A)x)

(6) (7)
s s+t

dx

1 [A + (B A)x]2

So the Feynman parameter x also has an interpretation, as the relative proper time two particles in the loop. Other useful related identities are (n + m) 1 = AnB m (n)(m) 1 = AB

of the

ds
0

sm1 (A + Bs)n+m

(8) (9)

ds
0

1 (A + Bs)2

Schwinger parameters are not as useful as Feynman parameters for Lorentz invariant integrals, but they do play an important role in eective theories, such as Heavy Quark Eective Theory (Lecture IV-10).

2 Wick Rotations
After introducing Feynman parameters and completing the square, one is often left with an integral over a loop momentum k in Minkowski space. Once the i factors are put in for Feynman propagators, 1-loop integrals often look like 1 d4k (2)4 (k 2 + i)n
2 2

(10)

Q Q gral has poles at k0 = k + i and k0 = k + + i, as shown in Figure 1. Since the poles are in the top left and bottom right quadrants of the k0 complex plane, the integral over the gure eight contour shown vanishes. Thus, the integral over the real axis and imaginary axis 2 2 Q2 are equal and opposite. Therefore, we can substitute k0 i k0 so that k 2 k0 k = kE , 2 2 Q2 where kE = k0 + k is the Euclidean momentum. This is known as a Wick rotation. After the Wick rotation, the i will no longer play a role and we can just set = 0.

Assuming > 0 (you can check that Wick rotation still works for < 0 in Problem 1), this inte-

?
Figure 1. Wick rotations. Poles in integrations over Feynman propagators often have poles at at k0 = Q k 2 + i. Integrating over the real axis is then equivalent to integrating over the imaginary axis.

Once Wick rotated, the integrals are evaluated in a straightforward way. We will need the formula for the surface area of the Euclidean 4-sphere: d4 = 2 2. Then 1 d4kE 2 f (kE ) = d4 16 4 (2)4
3 2 kEdkEf (kE ) = 0

1 8 2

3 2 kEdkEf (kE ) 0

(11)

Then, for example, Eq.(10) with n = 3 is evaluated as 1 d4k =i (2)4 (k 2 + i)3 1 d4kE 2 (2)4 (kE )3 k3 i dkE 2 E 3 =(1)3 2 8 0 (kE + ) i = 32 2

(12)

Regularization schemes

Other useful formulas following from Wick rotations are k2 i 1 d4k = 4 (k 2 + i)4 48 2 (2) d4k 1 (1)r 1 1 =i , r>2 4 (k 2 + i)r (2) (4)2 (r 1)(r 2) (r 2) 2 r 1 4 k (1) 2 1 d k =i , (4)2 (r 1)(r 2)(r 3) (r 3) (2)4 (k 2 + i)r

(13) r>3

and so on [cf Section 6.3 of Peskin and Schroeder]. Keep in mind that the Wick rotation is just a trick to do integrals. There is nothing physical about it. Also, the Wick rotation can only be justied if there are no new poles which invalidate the contour rotation. This caveat is only relevant for 2-loop and higher integrals which we will not encounter.

3 Regularization schemes
Here we will discuss some of the more popular regularization schemes.

3.1 Derivative method

A quick way to extract the ultraviolet divergence of an integral is by taking derivatives. Consider a logarithmically divergent integral, such as I() = 1 d4k = (2)4 (k 2 + i)2 2 i d4k = 4 (k 2 + i)3 16 2 (2) I() = i ln 16 2 2 (14)

If we take the derivative, the integral can be done d I() = d (15) (16)

So

where is an integration constant representing the UV cuto and is formally innite. Similarly, for a quadratically divergent integral, one could take the second derivative and then integrate twice to give k2 d4k i 1 = 6 d d 4 (k 2 + i)2 (2) 48 2 = i ln 2 + 2 2 2 8 1 (17)

for two integration constants 1 and 2. The derivative method is not an ideal regulator. Since the cuto appears as a constant of integration, there is no way to relate from one integral to from another. In particular, cancellations which we expect due to constraints like gauge invariance are not guaranteed to hold. Nevertheless, the derivative method is a quick way to check the coecient of the logarithms appearing in any particular integral.

3.2 Pauli-Villars regularization

A useful regulator was invented by Pauli and Villars [Pauli ..]. Their approach was for each particle of mass m to add a new unphysical ghost particle to the theory of mass with either the wrong statistics or the wrong-sign kinetic term. These new particles are designed to exactly cancel loop amplitudes with physical particles at large loop momentum. For example, one can write down a Pauli-Villars Lagrangian for QED which works at the 1-loop level as 1 2 1 2 1 2 LPV = F + i eA eA m + F 2A + i eA eA 4 4 2 (18)

Section 3

with A the ghost photon and the ghost electron and F = A A . We assume that both the ghost photon and ghost electron have bosonic statistics; the ghost photon has a wrongsign kinetic term. For example, this leads to a Feynman-gauge ghost-photon propagator of the form 0 T A (x)A (y) 0 = ig d4 p ip (x y) e 2 2 + i 4 p (2) (19)

Since this has the opposite sign from the photon propagator, it will cancel the photons contribution, for example, to the electron self-energy loop for loop momenta k (see Lecture III4). Note that the sign of the residue of the propagator is dictated by unitarity a physical particle with this sign would have negative norm, and violate conservation of probability. So this particle is purely ctitious, often called a Pauli-Villars ghost. The ghost electron propagator is same as the regular electron propagator; however, ghost electron loops do not get a factor of 1 (since they are bosonic) and so they cancel regular electron loops when k . In more detail, an amplitude with Pauli-Villars regularization will sum over the real particle, with mass m and the ghost particle, with xed large mass m. d4k 1 (2)4 (k 2 m2 + i)2 1 1 d4k (2)4 (k 2 m2 + i)2 (k 2 2 + i)2
1

(20)

For k , m both terms in the new integrand scale as k4 and so the integrand vanishes at least 1 as k6 making the integral convergent. We can now perform this integral by Wick rotation
3 i d4k 1 kE 1 k3 = 2 (1)2 dkE 2 2 E 22 2 4 (k 2 m2 + i)2 2 + i)2 2)2 8 (2) (k (kE m (kE ) 0 i m2 = ln 16 2 2

So that 1 i 2 d4k ln 2 16 2 m (2)4 (k 2 m2 + i)2 (21)

Note that the coecient of the logarithm is consistent with what we found using the derivative method, in Eq. (16). A useful trick in Pauli-Villars is to note that 1 1 = k 2 m2 k 2 2
2 m2

1 d (k 2 )2

(22)

which allows us to simply square the propagator and add another Feynman-like parameter . This can be useful for computing regulated integrals in practice. Also note that 1 d 1 (23) dm2 = dm2 2 2 k 2 m2 dm (k m2)2 So that, in a sense, Pauli-Villars can be viewed as a systematic implementation of the derivative method. Pauli-Villars was historically important. It also serves a useful pedagogical function. Indeed, the introduction of Pauli-Villars ghosts is much more clearly a deformation in the ultraviolet, relevant at energy scales of order the Pauli-Villars mass or larger, than analytically continuing to 4 dimensions. However, in modern applications Pauli-Villars is only occasionally useful. The problem is that for complicated multi-loop diagrams, you need to introduce many ctitious particles (one for each real particle will not do it; the Lagrangian LPV only works at 1-loop). Thus, Pauli-Villars quickly becomes impractical. Another problem is that it is not useful in nonAbelian gauge theories, since a massive gauge boson breaks gauge invariance. (In an Abelian theory Pauli-Villars does work, at least at 1-loop, as long as the gauge boson couples to a conserved current.)

Dimensional Regularization

3.3 Other regulators

In addition to Pauli-Villars and dimensional regularization (see below), there are a number of other regulators which are sometimes used: Hard cuto: kE < . This breaks translation invariance, and usually every symmetry in the theory. But it is perhaps the most intuitive regularization procedure. Point splitting (Dirac 1934). Divergences at k correspond to two elds approaching each other x1 x2. Point splitting puts a lower bound on this |x1 x2 | > | |. This also breaks translation invariance and is impractical for gauge theories, but is useful in theories with composite operators. Lattice regularization. Although a lattice breaks both translation invariance and Lorentz invariance, it is possible to construct a lattice such that translation and Lorentz invariance are restored in the continuum limit.

4 Dimensional Regularization
The most important regularization scheme for modern applications is dimensional regularization. Dimensional regularization was invented by t Hooft in 1973 [??]. The basic idea is that an integral like ddk 1 (24) (2)d (k 2 + i)2 is divergent only if d 4. If d < 4, then it will converge. If its convergent we can Wick rotate, and the answer comes from just analytically continuing all our formulas above to d dimensions.

4.1 Spinor algebra

In d-dimensions, the metric is g = diag(1, 1, 1, , 1) (25)

which means that there is exactly one timelike dimension in even non-integer d. This metric satises g g = d The Lorentz-invariant phase space is dLIPS (2)d dd1 p j 1 d (p) (2)d1 2E p j (27) (26)

nal states j

Spinor algebra can be assumed to work the same way in d = 4 dimensions as in d = 4. We assume there are d 4-dimensional -matrices satisfying { , } = 2g . The identity matrix in spinor space satises Tr1 = 4 as in 4 dimensions. In theories which involve 5 we also assume such a matrix exists satisfying {5, } = 0 (28)

The only place there can be subtleties with such a denition is in theories with anomalies. Anomalies are therefore best dealt with using a dierent regulator (see Lecture IV-8). An excellent discussion of spinors in various dimensions is given in the appendix of [Polchinski vol 2].

4.2 Scalar integrals

Since we will manipulate the expressions so they are only functions of the magnitude of k we write ddk = dd k d1 dk (29)

Section 4

where dd denotes the dierential solid angle of the d-dimensional unit sphere. Explicitly dd = sind2 (d1)sind3(d2) sin(2)d1 dd1 (30)

where i is the angle to the ith axis, with 0 1 < 2 and 0 i < for i > 1. For example, d2 = d. For d = 3, we normally write 1 = and 2 = giving d3 = d cosd (31)

which is the usual volume element of a 2-dimensional surface. Keep in mind, d is the dimension of the solid volume, not the surface, which has dimension d 1. The (d 1) dimensional surface areas of a ball of radius 1 in integer dimensions are: 2 = d2 = 2 (circle), d3 = 4 (sphere),
R

d4 = 2 2 (three-sphere),

(32)

The equivalent volumes are Vd = d

0 4 1

1 drr d1 = d Rd d

(33)

which are V2 = R2 , V3 = 3 R3, V4 = 2 2R4, etc. For non-integer dimensions, the surface area formula can be derived using the same trick we used for Gaussian integrals in Lecture II-7: ( )d = so that

dxe x

= d =

dd
0

2 1 d drr d1 e r = 2 2

dd

(34) (35)

dd =

2 d/2
d 2

Alternatively, one can just integrate Eq. (30):

d1 d1

d =2
n=2 0 2 2 3 2

dn sinn1 n = 2
n=2

n 2 n+1 2

=2 d/2

3 2 4 2

d1 2 d 2

= 2 d/2

(1)
d 2

and using (1) = 1 this reproduces Eq. (35). In these expressions (x) is the Gamma function, which is the analytic continuation of the factorial. For integer arguments, it evaluates to (1) = 1, (2) = 1, (3) = 2, (x) = (x 1)! (36)

(z) has simple poles at 0 and all the negative integers. We will often need to expand (x) around the pole at x = 0: 1 () = E + O() + (37) were E is the Euler-Mascheroni constant E 0.577. Sometimes relations like sin (x) = or the Euler -function (a, b) = (a)(b) = (a + b)
1 0

(1 x) , (x)(2 x)

cos(x) =

1 2x (1 x)(1 + x) 2x (2 2x)(2x)

(38)

dx(1 x)a1 xb1

(39)

allow us to simplify expressions. The integrals over Euclidean kE are straightforward dkE
a a+1 kE b = 2 2 b (kE + ) a+1 2

b 2(b)

a+1 2

(40)

Dimensional Regularization

Eqs. (29), (35) and (40) can be combined into a general formula a+ 2 ba 2 k 2a 1 1 ddk = i(1)ab d d (2)d (k 2 )b (4)d/2 ba 2 (b)( )
2 d d

(41)

Important special cases used in the text are ddk 4d 1 i 1 = 2 (2)d (k 2 + i)2 (4)d/2 2 d 2 ddk 2d k2 d i 1 = d (k 2 + i)2 2 2 (4)d/2 1 d (2) 2 k2 d i 1 4d ddk = d (k 2 + i)3 4 (4)d/2 2 d 2 (2) 2 6d 1 i 1 ddk = d (k 2 + i)3 d/2 3 d 2 (2) 2(4) 2 (42) (43) (44)

(45)

This last integral is convergent in d = 4; however, the d-dimensional form is important for loops with infrared divergences (see Lecture III-6). All dimensionally-regulated versions of divergent integrals will have poles at d = 4. Therefore we often expand as d = 4 and drop terms of order . Another common convention is d = 4 2, so if youre ever o by a factor of 2 in comparing to the literature, check the convention!

4.3 Field dimensions

Next, we should calculate the dimensions of all the elds and couplings in the Lagrangian. For the action to be dimensionless the Lagrangian density should have mass dimension d. For example, in QED, the Lagrangian is 1 LQED = ( A A )2 + (i m) e A 4 [A ] = M and also [e] = M tional to take
4d 2 d 2 2

(46) (47)

which implies

[] = M

d 1 2

[m] = M ,

. However, rather than have a non-integer dimensional coupling, it is convene

4d 2

(48)

where is an arbitrary parameter of mass dimension 1. Then e remains dimensionless. One usually only makes this change for the factors of e (or other gauge couplings) directly participating in a loop. If a loop graph is not one-particle irreducible, there may be other factors of e for which it is often simpler to leave 4 dimensional. This is just a convention. If all factors of e are modied as in Eq. (48), the answer will still be correct, but may contain awkward logarithms of dimensionful scales when expanded around d = 4. These awkward logarithms drop out of physical quantities, of course, but they can be avoided at intermediate steps as well by only adding factors of to coupling constants participating in the loop. The factors of coming from Eq. (48) modies loop integrals as d4k e2 4d (2)4 (k 2 + i)2 ddk e2 (2)d (k 2 + i)2 (49)

Keep in mind is not a large scale. It is not a UV cuto. The dimensional regularization is removed when d 4, not when . Thus is not like the Pauli-Villars mass M or a generic UV scale . In fact, we will often use as a proxy for a physical infrared scale associated with a renormalization group point. Nevertheless, there are two unphysical parameters in dimensional regularization, and ; both must drop out of physical predictions.

Section 4

Including this factor of , the logarithmically divergent integral becomes e2 ie2 4d d4k 4d 4 (k 2 + i)2 2 (2) (4)d/2 Now letting d = 4 we expand this around = 0 and get 4d ie2 4d d/2 2 (4) 1
2 2
d

2 2

(50)

ie2 16 2 ie2 = 16 2
ddk , k4

2 + E + ln4 + ln2 ln 4eE2 2 + ln + O()

+ O()

(51)

The E comes from the integral comes from the

4d

the 4 comes from the phase space

E 2

1 (2)d

and the (52) (53)

. This combination 4e

shows up a lot, so we give it a symbol

Thus

2 4eE2 d4k e2 ie2 2 2 + ln + O() 4 (k 2 + i)2 2 (2) 16

which has the same coecient of ln as found with Pauli-Villars in Eq.(21), with 2 = 2. Sometimes we will omit the twiddle and just write for Note that there is still a divergence . in this expression as 0. An important feature of dimensional regularization is that it helps characterize the degree to d4k which integrals diverge at high energy. For example, an integral like is logarithmi(k2 )2 cally divergent. In d dimensions, it has a simple pole at d = 4, and no other poles for d < 4. You 4d d4k can see this from the ( 2 ) term in Eq. (42). A quadratically divergent integral, like k2 has a stronger singularity at k = . With a hard cuto on Euclidean momentum, it would scale d4k 2 which is quadratically divergent. However, in non-integer dimensions, it is not like k2 divergent, it just has a funny dependence on d. If d = 2, the integral is logarithmically divergent, so there is a corresponding pole in the dimensionally regularized amplitude at d = 2. This shows 2d up in a ( 2 ) term, which has poles at d = 2 and d = 4, but no poles for d < 2. So dimensional regularization translates the degree of divergence into the singularity structure of the amplitudes in d dimensions. This can be really helpful for seeing the singularity structure of more complicated integrals. Dimensional regularization can also be used to regulate infrared divergent integrals. For 1 example ddk (k2 m2)k4 is infrared divergent for d < 4. We can evaluate this integral in d = 4 dimensions with < 0 instead of > 0. A nice feature of dimensional regularization as an infrared regulator is that it can be used both for virtual graphs and phase space integrals. Occasionally when using dimensional regularization we encounter an integral which is both ddk UV and IR divergent. For example, the scaleless integral . This integral is not convergent k4 for any d. Nevertheless, it is useful to be able to do such integrals. To progress, we can introduce an arbitrary scale to divide the UV and IR regions of Euclidean momenta: ddkE =d 4 kE
0 d5 dkE kE + d d5 dkE kE

=d ln

1 IR

+ d

1 UV

(54) ln

where we have written d = 4 IR for the rst integral, assuming IR < 0 and d = 4 UV for the second integral, assuming UV > 0. Rather than doing this split for every scaleless integral, since we know IR and UV must vanish from physical quantities, we often just set IR = UV = . When this is done, the integral is just 0. A simpler justication is that since there is no available d4k must vanish in d dimenquantity with non-zero mass dimension, scaleless integrals like k4 sions.

Problems

Often we are interested in just the UV divergence of an integral, which can be extracted from a scaleless integral as ddk 1 (2)d k 4 =i
UVdiv

2 d/2 1 i 1 d 1 =i = 2 d (d/2) d 8 UV (2) (2) UV UV

(55)

This is a very useful shortcut to extracting the ultraviolet divergence.

4.4 k integrals
We will often have integrals with factors of momenta, such as k k , in the numerator: F () = k k d4k 4 (k 2 )n (2) (56)

These can be simplied using a trick. Since the integral is a tensor under Lorentz transformations but only depends on the scalar , it must be proportional to the only tensor around, g . Then, just by dimensional analysis, we must get the same thing as in an integral with k k 1 replaced by ck 2 g for some number c. Contracting with g , we see that c = 4 or more gener1 ally c = d . d4k k k 1 ddk k2 g (57) 4 (k 2 )n d (k 2 )n (2) d (2) If there is just one factor of k in the numerator, for example in F (p2) = d4k (k p) (2)4 (k 2 p2)4 (58)

then the integrand is antisymmetric under k k. Since we are integrating over all k, the integral must vanish. So we will only need to keep terms with even powers of k in the numerator.

Problems
1. Show that the Wick rotation still works if < 0.