2012

Matthew Schwartz

III-3: The anomalous magnetic moment

1 Introduction
In the non-relativistic limit, the Dirac equation in the presence of an external magnetic field
produces a Hamiltonian
H=
Qp 2 + V (r) + e BQ · (LQ + g SQ ) (1)
2m 2m
acting on electron doublets |ψi where S Q = 2 Qσ . This was derived in Problem ?? of Lecture ??.
1

The coupling g is the g-factor of the electron, representing the relative strength of its intrinsic
magnetic dipole moment to the strength of the spin-orbit coupling. From the point of view of
the Schrödinger equation, g is a free parameter and could be anything. However, the Dirac
equation implies that g = 2 which was an historically important postdiction in excellent agree-
ment with data when Dirac presented his equation in 1932. A natural question is then, is g = 2
exactly, or does g receive quantum corrections? The answer should not be obvious. For
example, the charge of the electron is exactly opposite the charge of the proton, receiving no
radiative corrections (we will prove this in Lecture III-5), so perhaps the magnetic moment is
exact as well. By the late 1940s there was experimental data which could be partially explained
by the electron having an anomalous magnetic moment, that is, one different from 2. The
calculation of this anomalous moment by Schwinger, Feynman and Tomonaga in 1948, and its
agreement with data, was a triumph of quantum field theory.

2 Extracting the moment

We would like a way to extract the radiative corrections to g without having to take the non-rel-
ativistic limit. To see how to do this recall from Lecture II-3 how the electron’s magnetic dipole
moment was derived from the Dirac equation. Charged spinors satisfy (iD − m)ψ = 0. Multi-
2

plying this by (iD + m) shows that charged spinors also satisfy D + m2 ψ = 0. We then use
the operator relation (cf. Eq. (109??) of Lecture II-3)
2 e
D = D2µ + F µνσ µν (2)
2
i e e
where σ µν = 2 [γ µ , γν ] to find D 2µ + m2 + 2 F µνσ µν ψ = 0. The 2 F µνσ µν in this equation there-

fore encodes the difference between the way a scalar field, obeying (D2µ + m2)φ = 0, and a spinor
field interact with an electromagnetic field. In particular, in the Weyl representation

Q Q Q
!
e (B + iE )σ
F µνσ µν = −e
2 Q − iEQ )σQ
(B
(3)

2

Going to momentum space D + m2 ψ = 0 implies (cf. Eq. (112??) of Lecture II-3),

Q Q
2
!
(H − eA0)2 m p −eA e Q Q e Q Q
ψ= + −2 B ·S ±i E ·S ψ (4)
2m 2 2m 2m m

which can be compared directly to Eq. (1) to read off the strength of the magnetic dipole inter-
action geB Q · SQ .1 Since SQ = σQ2 for spin 12 , we find again that g = 2. If Eq. (2) had g ′ 4e F µνσ µν in it,
we would have found g = g ′ instead. Thus a general and relativistic way to extract corrections to
g is to look for loops which have the same effect as an additional F µνσ µν term.

1. The EQ ·S
Q term is not an electric dipole moment since it has an imaginary coefficient. Instead, it is Lorentz
invariant completion of the magnetic moment.

1


2 Section 2

A generally useful way to think about corrections to the way photons interact with spinors,
such as corrections to g, is to consider off-shell S-matrix elements. The Feynman rules for off-
shell S-matrix elements are the same as for on-shell S-matrix elements, except that p2i = m2i for
the various external states is not enforced. In this case, the relevant process is e−(q1)A µ(p) →
e−(q2) with polarization vector ǫµ(p) and two spinor states ū (q2) and u(q1). At tree level, the
matrix element is just ǫ µM0µ where

p
iM0µ = = −ieū (q2)γ µu(q1) (5)
q1 q2

with the photon momentum constrained by momentum conservation to be p µ = q2µ − q1µ. This
result actually contains g = 2 in it, although it is hard to see in this form. We expect something
equivalent to an F µνσ µν term which should look like ū (q2)pνσ µνu(q1) in momentum space. To
see where F µνσ µν is hiding, we need to massage the result a little.
For the magnetic moment, we only have to allow for the photon, which corresponds to an
unconstrained external magnetic field, to be off-shell; the spinors can be on-shell, which helps
simplify things. For example, we can use the Gordon identity, which you derived in
Problem ??, and holds for on-shell spinors:

ū (q2)(q1µ + q2µ)u(q1) = (2m) ū (q2)γ µu(q1) + i ū (q2)σ µν (q1ν − q2ν )u(q1) (6)
Therefore

q1µ + q2µ
 
e
M0µ = −e ū (q2)u(q1) − i ū (q2)pν σ µνu(q1) (7)
2m 2m

The first term is an interaction just like the scalar QED interaction: the photon couples to the
momentum of the field, as in the D2µ term in the Klein-Gordon equation. The q 1µ and q 2µ in this
first term are just the momentum factors which appear in the scalar QED Feynman rule. The
second term in Eq. (7) is spin-dependent and gives the magnetic moment. So we can identify g
4m
as e times the coefficient of i pνūσ µνu. Therefore, to calculate corrections to g we need to find
how the coefficient of ūpν σ µνu is modified at loop level.
The correction to the magnetic moment must come from graphs involving the photon and
the electron, to contribute corrections to the process in Eq. (5). We can parameterize the most
general possible result, at any-loop order, as

p
iM µ = q1 = ū (q2)(f1 γ µ + f2 pµ + f3 q1µ + f4 q2µ)u(q1) (8)

q2

Here we have included all Lorentz vectors which might possibly appear, with the fi their
unknown Lorentz scalar coefficients. The fi can depend in general on contractions of momenta,
like p · q or p2, or on contractions with γ-matrices like p. (In more general theories, they could
also depend on γ5, but QED is parity invariant so γ5 cannot appear.) For the magnetic moment
application, we can assume the external spinors are on-shell, but the photon, representing an
unconstrained external magnetic field, must still be off-shell. (Or, if you prefer, imagine this dia-
gram is embedded in a larger Coulomb-scattering diagram with an off-shell intermediate photon
and on-shell external spinors.)
The fi are not all independent. Using momentum conservation, p µ = q2µ − q1µ, we can set f2 =
0 and substitute away all the pµ dependence. Then, if there are factors of q1 or q2 in the fi,
they can be removed by using the Dirac equation q1u(q1) = m u(q1) and ū (q2)q2 = mū (q2). So,
we can safely assume the fi are real functions which can only depend on q1 · q2 and m, or more
conventionally on p2 = 2m2 − 2q1 · q2 and m2. Moreover, we can fix the relative dependence by
p2
dimensional analysis so the fi are functions of m2 .
Evaluating the graphs 3

Next, the Ward identity (which we showed in Lecture II-7 holds even if the photon is off-
shell) implies
0 = p µū (f1 γ µ + f3 q1µ + f4 q2µ)u
= f1ū p u+(p · q1)f3ūu + (p · q2)f4ūu (9)
= (p · q1)f3ūu + (p · q2)f4ūu

We then use p · q1 = q2 · q1 − m2 = −p · q2 to get f3 = f4. Thus there are only two independent
form factors. We can then use the Gordon identity, Eq. (6), to rewrite the q1µ and q2µ depen-
dence in terms of σ µν , leading to
  2  2 
µ p µ iσ µν p
iM = (−ie)ū (q2) F1 2
γ + pνF2 u(q1) (10)
m 2m m2

which is our final form. This parametrization holds to all orders in perturbation theory. The
functions F1 and F2 are known as form factors. The leading graph, Eq. (5) gives
F1 = 1, F2 = 0 (11)

Loops will give contributions to F1 and F2 at order α and higher.

Which of these two form factors could give an electron magnetic moment? F1 modifies the
original e A µψ̄γ µψ coupling. This renormalizes the electric charge, as we saw from the vacuum-
polarization diagram. In fact, the entire effect of this form-factor is to give scale-dependence to
the electric charge, so no other effect, like an anomalous magnetic moment, can come from it.
F2, on the other hand, has precisely the structure of a magnetic moment (which is, of course,
why we put it in this form with the Gordon identity). Using  that such a term without the F2
p2
factor gives g = 2, as in Eq. (7), we conclude that F2 m2 modifies the moment at the scale
 2
p
associated with p2 by g → 2 + 2F2 m2 . Since the actual magnetic moment is measured at non-
Q | ≪ m, the moment which can be compared to data is
relativistic energies with |p
g = 2 + 2F2(0) (12)

Thus, we have reduced the problem to calculating F2(0).

3 Evaluating the graphs

There are 4 possible 1-loop graphs which could contribute to M µ. Three of them,

a) b) c)

can only give terms proportional to γ µ. This is easy to see because these graphs just correct the
propagators for the corresponding particles. Thus these graphs can only contribute to F1 and
have no effect on the magnetic moment. The fourth graph is

k p+k
iM2µ =
q2
q1

k − q1

with p µ = q2µ − q1µ. This is the only graph we have to consider for g − 2.
4 Section 3

Employing the Feynman rules, this graph is

d4k −igνα i(p + k + m)

Z
µ 3 i(k + m)
iM2 = (−ie) ū (q2)γ ν γµ γ αu(q1) (13)
(2π)4 (k − q1)2 + iε (p + k)2 − m2 + iε k 2 − m2 + iε

d4k γ ν (p + k + m)γ µ(k + m)γν

Z
=−e3ū (q2) u(q1) (14)
4
(2π) [(k − q1) + iε][(p + k)2 − m2 + iε][k 2 − m2 + iε]
2

To simplify this, we start by combining denominators and completing the square. The denomi-
nator has 3 terms and can be simplified with the identity
Z 1
1 1
=2 dxdydzδ(x + y + z − 1) (15)
ABC 0 [xA + yB + zC]3
In this case

A= k 2 − m2 + iε
B = (p + k)2 − m2 + iε
C = (k − q1)2 + iε

The new denominator is the cube of

xA + yB + zC = k 2 + 2k(yp −zq1) + yp2 + zq12 − (x + y)m2 + iε

= (k µ + ypµ −zq1µ)2 − ∆ + iε

with
∆ = −xyp2 + (1 − z)2m2 (16)
Thus we want to shift k µ → k µ − yp µ + zq1µ to make the denominator (k 2 − ∆)3.
The numerator in Eq. (14) is

N µ = ū (q2)γ ν (p + k + m)γ µ(k + m)γνu(q1)

=−2ū (q2)[kγ µp + kγ µk + m2 γ µ − 2m(2k µ + p µ)]u(q1)

Shifting k µ → k µ − ypµ + zq1µ then gives

1
− N µ = ū (q2)[(k − yp + z q1)γ µp + (k − yp + z q1)γ µ(k − yp + z q1)]u(q1) (17)
2
+ ū (q2)[m2 γ µ − 2m(2k µ − 2yp µ + 2zq1µ + p µ)]u(q1) (18)
1
Using k µk ν = 4 g µνk 2, the Gordon identity, x + y + z = 1 and a fair amount of algebra, this sim-
plifies to  
1 µ 1 2 2 2 2
− N = − k + (1 − x)(1 − y)p + (1 − 4z + z )m ū (q2)γ µu(q1)
2 2
+imz(1 − z)pνū (q2)σ µνu(q1) (19)

+m(z − 2)(x − y)p µū (q2)u(q1)

We have found 3 independent terms instead of 2 since we have not used the Ward identity.
Indeed the Ward identity should fall out of the calculation automatically. To see that it does,
note that the p µ term gives a contribution to M2µ of the form
Z 1
d4k pµ
Z
µ 3
iM2 = 4e dxdydzδ(x + y + z − 1)m(z − 2)(x − y) ū (q2)u(q1) (20)
0
4
(2π) (k − ∆ + iε)3
2

Next, note that both ∆ in Eq.(16) and the integral measure are symmetric in x ↔ y, but the
integrand is antisymmetric. Thus this term is zero.
For the magnetic moment calculation we only need the σ µν term. Thus
 Z 1 
d4k
Z
z(1 − z)
iM2µ = pνū (q2)σ µνu(q1) 4ie3m dxdydzδ(x + y + z − 1) + (21)
0 (2π)4 (k 2 − ∆ + iε)3
Evaluating the graphs 5

where the do not contribute to the moment. Recalling that F2(p2) was defined as the coeffi-
2m
cient of this operator, normalized by e , we have
Z 1
d4k
Z
2m z(1 − z)
F2(p2) = (4ie3m) dxdydzδ(x + y + z − 1) + O(e4) (22)
e 0 (2π)4 (k 2 − ∆ + iε)3

For completeness, the other form factor is F1(p2) = 1 + f (p2) + O(e4) where
Z 1 4
d k k 2 − 2(1 − x)(1 − y)p2 − 2(1 − 4z + z 2)m2
f (p2) = −2ie2 4
dxdydzδ(x + y + z − 1) (23)
0 (2π) [k 2 − (m2(1 − z)2 − xyp2)]3
We will come back and evaluate f (p2) when we need to (in Lecture III-5).
To evaluate F2, we use the identity from Appendix B
d4k
Z
1 −i
= (24)
(2π)4 (k 2 − ∆ + iε)3 32π 2∆

to get that, up to terms of order α2,

Z 1
α z(1 − z)
F2(p2) = m2 dxdydzδ(x + y + z − 1) 2m2 − xyp2
(25)
π 0 (1 − z)
At p2 = 0 this integral is finite. Explicitly
α 1
Z Z 1 Z 1
z
F2(0) = dz dy dxδ(x + y + z − 1)
π 0 0 0 (1 − z)
α 1
Z Z 1−z
z
= dz dy
π 0 0 (1 − z)
α
=
2π
Thus
α
g = 2 + = 2.00232 (26)
π

with the next correction of order α2.

As a historical note, this result was first announced at the APS meeting in January 1948, by
Schwinger. Feynman and Tomonaga had both calculated the same result independently at the
same time. Schwinger actually found a different value for g − 2 for an electron bound in an atom
and a free electron, while Feynman found they were the same. Feynman’s result was the correct
one, and it was relativistically invariant, while Schwinger’s was not. The discrepancy was
quickly resolved. Tomonaga was the first to correctly present the full 1-loop formula for the
Lamb shift.
Unfortunately, it is not easy to measure g directly. Schwinger was able to check his calcula-
tion indirectly as giving part of the contribution to various hyperfine splittings in Hydrogen,
such as the Lamb shift. In order to make the comparison, he needed also to be able to get finite
predictions out of the divergent integrals, such as the contributions to F1 in addition to the
finite g − 2 integral. The comparison with data really required a full understanding of all the 1-
loop corrections in QED. For this reason, the simplicity of the finite g − 2 calculation we have
just done was not immediately appreciated. Nevertheless, this calculation and the Lamb shift
calculation more generally was critically important historically for convincing people that loops
in quantum field theory had physical consequences.
The current best measurement is g = 2.0023193043617 ± (3 × 10−13 ). The theory calculation
has been performed up to 4-loop level. One cannot compare theory to experiment directly, since
the theory is expressed as a function of α which cannot be measured more precisely any other
way. Therefore g − 2 is now used to define the renormalized value of the fine structure constant,
which comes out to α = 137.035999070±9.8 × 10−10 .