78 Hadrons

2.5 U (1)A anomaly

We have seen that spontaneous chiral symmetry breaking should affect all axial sym-
metries including the flavor-singlet U (1)A . The fact that there is no good candidate for
a flavor-singlet (pseudo-) Goldstone boson in the spectrum is related to the anomalous
U (1)A breaking. Anomalies are symmetries of classical Lagrangians that are broken at
the quantum level. They arise when regularization destroys a symmetry and there is no
regulator choice that can preserve it. Since the symmetry is lost, there is no Goldstone
boson because the quantum corrections will generate a mass for that mode.
Anomalies are again a typical feature of axial symmetries. In contrast to spontaneous
symmetry breaking, where the symmetry is lost due to dynamical effects, anomalies
have their origin in short-distance singularities of the currents Aµa = ψ γ µ γ5 ta ψ and
Aµ = ψ γ µ γ5 ψ. These are composite operators at the same space-time point which
are potentially divergent and have to be regularized. In principle, the problem would
also affect vector currents, but in that case it is possible to find appropriate regulariza-
tion prescriptions that leave their symmetry intact. Vector symmetries are related to
conserved charges (color charge, electromagnetic charge, flavor charges, etc.). If they
were broken at the quantum level, we would not only lose charge conservation but also
gauge symmetry, and the theory would become nonrenormalizable and inconsistent.
Global axial symmetries are in that sense ’less important’ and the fact that they pro-
duce anomalies is not a serious problem for the theory.14 In the following we will see
that QCD leads only to an anomalous U (1)A breaking, whereas QED also induces an
anomalous SU (Nf )A breaking.

Anomalies from the path integral. We have anticipated in Eq. (2.29) that the
divergence of the axialvector singlet current picks up an anomalous contribution

∂µ Aµ = 2i ψ M γ5 ψ + Nf Q(x) , (2.159)

where Q(x) is the topological charge density that we encountered in Section 1.1:

g2 1
Q(x) := Tr {Feµν F µν }, Feµν = εµναβ Fαβ . (2.160)
8π 2 2
The derived relation (2.108) entails that the mass of the η0 does not vanish in the chiral
limit, so there is no flavor-singlet Goldstone boson. To see how this term comes about,
start with the massless Lagrangian L = ψ iD / ψ and perform an axial U (1)A rotation
with a group parameter ε(x):

L0 = ψ eiεγ5 iD
/ eiεγ5 ψ = ψ (iD
/ − ∂/ ε γ5 ) ψ = L − (∂µ ε) Aµ . (2.161)

If we put this in the path integral, we have

Z Z  Z 
0 0 iS[ψ 0 ,ψ 0 ] iS[ψ,ψ] 4 µ
Z[0] = D[ψ , ψ ] e = D[ψ, ψ] e 1 + i d x ε(x) ∂µ A (x) , (2.162)

14
Except when they are also promoted to gauge symmetries: if a gauge symmetry is broken anoma-
lously, then one needs anomaly cancellations between different sectors of the theory.
2.5 U (1)A anomaly 79

and hence h∂µ Aµ i = 0, i.e., current conservation holds inside the expectation value.15
As always we have assumed that the path integral measure remains invariant under
the transformation. However, for axial transformations this is not necessarily the case.
The origin of this behavior is the transformation of the Dirac spinors

ψ 0 (x) = e+iεγ5 ψ , ψ 0 (x) = ψ(x) e+iεγ5 , (2.163)

which leads to a Jacobian determinant of the transformation:

D[ψ 0 , ψ 0 ] = (det C)−2 D[ψ, ψ] . (2.164)

It turns out that this determinant is ill-defined and requires regularization, which in
turn breaks the U (1)A symmetry. The final result is exactly the anomalous term:
 Z 
−2 4
(det C) = exp −i d x ε(x) Nf Q(x) . (2.165)

Putting this back in Eq. (2.162) yields the anomalous correction to the PCAC relation:
Z  Z 
iS[ψ,ψ] 4 µ


Z[0] = D[ψ, ψ] e 1 + i d x ε(x) ∂µ A (x) − Nf Q(x) . (2.166)
| {z }
!
=0

Fujikawa’s method. In order to calculate this, we can expand the functional deter-
minant into eigenfunctions of the Dirac operator. The hermitian Dirac operator D / has
real eigenvalues λn and a set of orthonormal, complete eigenfunctions: 16

Z
d4 x ϕ†m,i (x) ϕn,j (x) = δmn δij ,
D/ ϕn (x) = λn ϕn (x) , P (2.167)
† 4 (x − y) δ ,
ϕ
n n,i (x) ϕ n,j (y) = δ ij

where i, j collect the remaining Dirac, color and flavor indices. Then we can expand the
spinors ψ, ψ into these eigenfunctions, where the coefficients an and b̄n are independent
Grassmann variables, and write down the path integral measure:
X X Y Y
ψ(x) = an ϕn (x) , ψ(x) = ϕ†n (x) b̄n , D[ψ, ψ] = dan db̄m . (2.168)
n n n m

The fermionic path integral can be written as the determinant of the Dirac operator:
Z Z Y Y
/
i d4 x ψ iD
R P
ψ
/ = D[ψ, ψ] e
det D = dan db̄n e− n b̄n λn an = λn . (2.169)
n n

15
The proper argument would start from the background field method in Section 2.1: add a source
term B · A with δBµ = ∂µ ε to the Lagrangian in the path integral exponential, with the purpose
to cancel the additional term in (2.162) from the local U (1)A gauge transformation. The resulting
partition function is locally gauge invariant, and one derives the WTI as in Eq. (2.64).
16
To ensure (anti-)hermiticity of the Dirac operator, we should actually do this in Euclidean space.
80 Hadrons

An axial transformation changes the coefficients an and b̄n to

Z XZ
a0n = d4 x ϕ†n (x) ψ 0 (x) = d4 x ϕ†n (x) eiε(x)γ5 ϕm (x) am (2.170)
m | {z }
=:Cnm

so that we have X X
a0n = Cnm am , b̄0m = Cnm b̄n . (2.171)
m n
Because the Grassmann measure transforms with the inverse determinant we arrive at
Eq. (2.164). Using det C = eTr ln C and expanding the logarithm, we obtain
Z !
X
−2 4 †
(det C) = exp −2i d x ε(x) ϕn (x) γ5 ϕn (x) , (2.172)
n

which involves the ’functional trace’ over γ5 . With the completeness relation in (2.167),
the sum becomes
X
ϕ†n (x) γ5 ϕn (x) = lim Tr {γ5 } δ 4 (x − y) , (2.173)
y→x
n

where the trace is now the usual trace over Dirac, color and flavor indices. (The color-
flavor trace would just produce a factor Nf NC .) Normally, this trace would be zero
and the determinant 1, but because of the short-distance singularity the expression is
not well-defined (0 · ∞) and must be regulated.
Fujikawa suggested to damp the contribution from the large eigenvalues by a Gaus-
sian cutoff with regulator mass M which is taken to infinity in the end:
X X
lim
2
ϕ†n (x) γ5 e−(λn /M ) ϕn (x) = lim / )2 ϕ (x)
ϕ†n (x) γ5 e−(D/M n
M →∞ M →∞
n n
n o (2.174)
/ )2 δ 4 (x − y).
= lim Tr γ5 e−(D/M
M →∞
y→x

The regularization is gauge-invariant because the covariant derivative appears in it;

/ 2 as
hence, it preserves the vector symmetry. To proceed, one can express D
1 1 ig
/ 2 = γ µ γ ν Dµ Dν =
D {γµ , γν } Dµ Dν + [γµ , γν ] Dµ Dν = D2 − [γµ , γν ] F µν . (2.175)
2 2 4
With the generic relation f (∂x ) eikx = eikx f (∂x + ik), where unsaturated derivatives in
the final expression vanish, we can take Eq. (2.174) to momentum space:
Z 4 Z 4
4 d k ik (x−y) d k
lim f (∂x ) δ (x − y) = lim f (∂x ) e = f (∂x + ik) , (2.176)
y→x y→x (2π)4 (2π)4
so one has to expand the exponential in (2.174) and replace ∂µ → ∂µ + ikµ . The lowest
nonvanishing contribution is the square of the commutator term ∼ F µν in Eq. (2.175)
because the trace with γ5 requires four γ−matrices to be nonzero:
i 
Tr γ5 γ µ γ ν γ α γ β = εµναβ . (2.177)
4
2.5 U (1)A anomaly 81

The resulting expansion has the form

 Z 4 
1 d k − k22 2 1
Eq. (2.174) = lim e M g Nf Tr {Feµν F } + 6 (. . . ) + . . . .
µν
M →∞ M 4 (2π)4 M
(2.178)
After sending M → ∞ and integrating out the momentum k, one obtains
X g 2 Nf
ϕ†n (x) γ5 ϕn (x) = lim Tr {γ5 } δ 4 (x − y) = Tr {Feµν F µν }, (2.179)
n
y→x 16π 2

where the trace over the SU (3)C generators in F µν = Faµν ta remains. Inserted in the
determinant (2.172), we arrive at the result in Eq. (2.165).

Remarks:

• Note that we did not perform an ‘additional renormalization’ because the theory
was already renormalized before. Renormalization means that the regulator re-
mains in the theory, but it is hidden in the renormalization constants which must
cancel each other in observables. Here we have merely cured a 0 · ∞ situation by
introducing a cutoff M that we sent to infinity at the end; however, the resulting
finite expression has the property that it breaks the U (1)A symmetry. While we
used exponential damping, one can show that this result is indeed independent
of the chosen regularization as long as it is gauge invariant.

• Since the topological charge is essentially the trace over γ5 , one can ask why the
non-Abelian global SU (NfP )A transformations do not lead to anomalies. Repeat-
ing the analysis with ε → a εa ta yields:

g 2 αβµν b c
∂µ Aµa = ε Fαβ Fµν TrF {ta } TrC {tb tc } , (2.180)
(4π)2

which vanishes in the flavor-octet case because Tr{ta } = 0. In other words, gluons
couple only to flavor-singlet currents, and the anomaly signals the breakdown of
the U (1)A symmetry in the presence of gluons.

• The topological charge density can be written as the divergence of a current, the
Chern-Simons current:
 
µ µ g 2 µναβ 2ig
Q(x) = ∂µ K , K = 2ε Tr Fαβ Aν + Aα Aβ Aν . (2.181)
8π 3

One could then conclude that the flavor-singlet PCAC relation (in the chiral limit)
still induces a conserved current, which leads back to the argument that there
should Rbe a flavor-singlet Goldstone boson. However, K µ and its corresponding
charge d3 x K 0 are not gauge invariant, so they cannot couple to physical states
and hence there is no conserved axial charge.
82 Hadrons

A A A V A A

V V A A V V A V

Figure 2.9: Anomalous 1-loop fermion diagrams.

Triangle diagrams. The axial anomaly will show up (and was originally derived) in
the calculation of Green functions that involve axialvector currents, e.g.

h0|T Aµ (x) V α (y) V β (z) |0i , h0|T Aµ (x) Aα (y) Aβ (z) |0i , etc. (2.182)

Take for example the axialvector and vector WTIs for an AV V correlator:

∂µx hAµ V α V β i = h (∂µ Aµ )V α V β i + δ(x0 − y 0 ) h[A0 , V α ] V β i

(2.183)
+ δ(x0 − z 0 ) hV α [A0 , V β ]i = 0 .

The last two terms on the right-hand side are zero because the commutators of the
singlet currents vanish, cf. (2.42). The first term produces the pseudoscalar density via
the PCAC relation. Repeating this for derivatives with respect to y and z, we arrive at

∂µx hAµ V α V β i = 2m hP V α V β i , ∂αy hAµ V α V β i = 0, ∂βz hAµ V α V β i = 0, (2.184)

without taking into account the anomaly. These diagrams are linearly divergent and
therefore not translationally invariant. If one calculates them explicitly to 1-loop order,
shifting integration variables by a different momentum routing will produce results that
differ by surface terms. The freedom in distributing these surface terms can be used
in the regularization procedure when getting rid of all infinite pieces. It turns out that
the relations (2.184) cannot be satisfied simultaneously, and in order to preserve the
vector symmetries the axialvector WTI must pick up the additional anomalous term.
A theorem by Adler and Bardeen states that the full structure of the anomaly is al-
ready contained in the perturbative one-loop fermion diagrams. Higher-loop corrections
do not renormalize the anomaly except for replacing the fields and coupling constants
by their renormalized values. For anomaly considerations it is therefore enough to
calculate the triangle and rectangle diagrams in Fig. 2.9. These are the superficially
divergent ones (in fact, pentagon diagrams should be included as well although they
are convergent), and they include an odd number of axial currents and thus an odd
number of γ5 matrices.
2.5 U (1)A anomaly 83

QED anomaly and π 0 → γγ decay. Anomalies have observable consequences. The

prime example are the η and η 0 masses, but in that case the anomalous contribution
is also difficult to quantify due to the explicit breaking of chiral symmetry and mixing
effects. A much cleaner system is the decay of the π 0 into two photons, which is almost
exclusively caused by the axial anomaly from QED effects.
So far we have considered the axial anomaly in QCD (the ’gluon anomaly’) which
is the relevant one for the η − η 0 problem. Quarks couple to gluons, and the quark’s
flavor-singlet axialvector current Aµ picks up an anomalous term containing the gluonic
field-strength tensor. On the other hand, quarks can also couple to photons which will
also produce an anomaly, although the related effects are much weaker (αQED  αQCD ).
If we repeat the derivation for the QED Lagrangian, replace F µν by the electromagnetic
field-strength tensor and the coupling g with e, we obtain the electromagnetic ’photon
anomaly’ (Adler-Bell-Jackiw or ABJ anomaly):
 
e2  1 2 0 0
∂µ Aµa = εαβµν Fαβ Fµν TrF ta Q2 TrC {1} , Q =  0 −1 0  , (2.185)
(4π)2 3 0 0 −1

stated here without the fermion mass term. Since fermions with different flavors have
different charges (expressed by the quark charge matrix Q), photons can also couple to
flavor-nonsinglet currents. Therefore, the electromagnetic anomaly produces additional
terms for the divergences of the axial currents Aµ and Aµa .
To extract the π 0 → γγ decay, consider the three-point function of an axialvector
current and two electromagnetic vector currents:

h0|T Aµa (x) Vem

α β
(x1 ) Vem (x2 )|0i . (2.186)

The electromagnetic current is proportional to the quark charges and thus given by
µ
Vem (x) = ψ(x) γ µ Q ψ(x) = V3µ (x) + √1
3
V8µ (x) . (2.187)

To lowest order perturbation theory, Eq. (2.186) is the AVV triangle diagram of Fig. 2.9
and diverges linearly. However, it has also a nonperturbative spectral representation
in terms of pseudoscalar bound-state poles. We can derive this in complete analogy to
Eqs. (2.151)–(2.153). First, write down its Ward-Takahashi identity by acting with the
derivative on the index µ:

β

β

∂µx h0|T Aµa (x) Vem
α z z α z z
2 Vem − 2 |0i − 2m h0|T Pa (x) Vem 2 Vem − 2 |0i = . . . (2.188)

We are interested in the π 0 with a = 3; in that case the commutators on the right-hand
side obtained from (2.52) vanish, because they contain the structure constants f338 = 0,
etc. Instead we have the contribution from the anomaly:

e2 D αβρσ α z

β z


··· = ε h0|T Fαβ (x) Fρσ (x) Vem 2 Vem −2 |0i , (2.189)
(4π)2

where the factor D = NC /6 comes from the flavor and color traces.
84 Hadrons
𝜒

∂�𝐴��𝟤𝑚𝑃 ∂�𝐴��𝟤𝑚𝑃

𝑓�

= =
𝛤���

𝛾 𝛾

Figure 2.10: π 0 → γγ decay in the chiral limit.

If we work out the time orderings on the left-hand side and insert the completeness
relation, we can again isolate the Feynman propagator. The pole residues are the two
decay constants from Eq. (2.78) and the π → γγ decay amplitude, defined via
Z
α z

β
αβ d4 q −iqz
z
ihλ|T Vem 2 Vem − 2 |0i =: Γλ (z, p) = e Γλ (q, p) εαβρσ q ρ pσ . (2.190)
(2π)4

Its expression in momentum space is due to Lorentz and parity invariance: p is the pion
momentum, q is the relative momentum between the photons, and the only negative-
parity Lorentz tensor that depends on two momenta is the one above. If we inte-
grate (2.188) over x and z, the poles drop out again and the analogue of Eq. (2.153)
becomes X
lim fλ Γαβ αβ
λ (q, p) = lim fπ Γπ (q, p) = 0 , (2.191)
p→0 p→0
q→0 λ q→0

as long as we discard the anomaly on the right-hand side. We have again removed the
sum over λ because the decay constants are zero for all excited states with mλ 6= 0.
Since the transition matrix elements are defined at p2 = m2π = 0, this is a chiral-
limit relation. Hence, the decay amplitude should be zero, which is known as the
Sutherland-Veltman theorem.
In order to take the anomaly into account, we would have to work out the right-
hand side of Eq. (2.189). However, since the anomaly is produced already in the lowest
order perturbation theory, it is sufficient to start again from Eq. (2.188) and work
out its perturbative 1-loop contributions, the AV V and P V V triangle diagrams. The
ambiguity in shifting integration variables produces just the anomalous term.17 The
result has the same structure in momentum space ∼ εαβρσ q ρ pσ , and the transition form
factor in Eq. (2.191) becomes

e2 D
Γπ (0, 0) = . (2.192)
2π 2 fπ
The calculated π → γγ decay width using this result is 7.862 eV; the experimental value
is 7.8 ± 0.9 eV. Therefore, the neutral pion decay doesn’t probe the nonperturbative
structure of QCD at all — it is practically completely determined by the axial anomaly.

17
The calculation can be found in most textbooks, for example Kaku, p.414ff.