Proc. Nadl. Acad. Sci.

USA
Vol. 81, pp. 2597-2600, April 1984
Mathematics

Dirac operators coupled to vector potentials

(elliptic operators/index theory/characteristic classes/anomalies/gauge fields)

M. F. ATIYAHt

AND I.

M. SINGER:

tMathematical Institute, University of Oxford, Oxford, England; and tDepartment of Mathematics, University of California, Berkeley, CA 94720

Contributed by I. M. Singer, January 6, 1984

ABSTRACT
Characteristic classes for the index of the Dirac family OA are computed In terms of differential forms on
the orbit space of vector potentials under gauge transformations. They represent obstructions to the existence of a covariant Dirac propagator. The first obstruction is related to a chiral anomaly.
In this note we study the null spaces (zero frequency modes)
of OA, the massless Dirac operator coupled to a vector potential, as the potential A varies. We are interested in the null
spaces of positive chirality as opposed to those of negative
chirality. Their formal difference is a virtual bundle, Ind $;
we apply the index theorem for families of operators and
some infinite dimensional geometry to compute the characteristic classes of Ind $ explicitly in terms of differential
forms.
The formulas obtained may be of interest in quantum
chromodynamics. The path integral formulation uses gauge
invariant functionals of the propagator for WA. To define the
propagator $A' requires some consistent identification of the
null spaces of positive and negative chirality. The nonvanishing of the characteristic classes are obstructions to a consistent covariant identification of these null spaces-i.e.,
obstructions to the existence of a covariant propagator. The
first such obstruction is related to a chiral anomaly, as discussed below. We ask whether the higher obstructions have
physical significance as well.
Let M be a compact oriented Riemannian spin manifold of
dimension 2n, and P a principal bundle over M with group G.
Let W be the set of connections or vector potentials on P,
with '6 the group of gauge transformations of P. We denote
the action of 4 E T on A E W by frA. Let p be a representation of G on CN giving the associated vector bundle E = P Gx
CN. Each A E 2I gives a Dirac operator OA: C'(S+ 0 E) -3
C'(S- 0 E) where S+ are the spin bundles over M of positive and negative chirality, respectively. In local coordinates

WA

The analytic family indexed by W/'9 can be defined directly in terms of the Hilbert bundles We = W XL2(S' 0 E) over
%/'S. Covariance means {4AA}IEE gives an elliptic operator
0%.A mapping the fiber EV'A to Xi-A. The analytic index of
this family is Ind $ above.
When M = S4 and $ is the group of gauge transformations
leaving the north pole fixed, the index for the Dirac family
#sm is computed topologically in ref. 1. The index theorem
implies that the following two maps are homotopically equivalent. The first is given by the Dirac family

{q$vA}-I' EA
mapping W/'9 into Fredholm operators. For the second, we
have the composition of maps
a4
a2
Ad8 a,
al W
' ;
Q(G) )2 fl 3 (U(N)) a3 W
Q(U(00))
The map a, (which is a homotopy equivalence) is parallel
transport by means of A around closed curves parameterized
by the equator S3. (Follow a fixed geodesic from the north
pole to the south pole and follow a variable geodesic back.)
The map a2 is induced by the representation p: G SU(N),
and a3 by the injection of U(N)
U(oo). Finally, a4 is a
homotopy equivalence (Bott periodicity, twice).
Thus, the characteristic classes of Ind $ can be obtained
by pulling back the cohomology generators in 9; via the second map. For example, if G = U(N) and p is the identity, one
obtains nonzero characteristic classes, up to degree 2N - 4.
In general, to compute the characteristic classes of Ind
$ in terms of forms, we introduce a "universal" bundle with
connection. % acts on P x W by (p, A) -* (+O(p), SEA). This
action has no fixed points and gives a principal bundle

(PxW

sP x

Since the group action of G on P x 2I commutes with that of

'8, the group G acts on 9. If G acts without fixed points, one
obtains a principal bundle 9. with group G and base 9,/G = M
x %/'9. That occurs when one either restricts W to the space
of irreducible connections or restricts '9 to be gauge transformations leaving a point of P fixed. We assume the latter. The
principal G-bundle 9. has a natural connection w, obtained as
follows. The space P x 2f has a Riemannian metric invariant
under G x (6. At (p, A), the metric on T(P, p) is given by the
metrics of G, M and the connection A; while the metric on
T(W, A) is the usual metric on C(A1 0 g). The metric on P x
W descends to a metric on 9. invariant under G. The orthogonal complement to orbits of G gives the connection w.
(9, w) is universal in the following sense. Suppose Q is a
principal G-bundle over M x X, X compact and QIMXX P
for each x E X. Suppose, moreover, that Q has a fiber co.inection; that is, a choice of connection on QMXX continuous
for x E X. Then there is a map A: Q - inducing the fiber
connection from w. Conversely, any map 83 of X -+/'
leads to a fiber connection by pulling back (9., w) via I x 83:

Iy- 7(am + Fm + AM) (1; Y5)

where rM is the Riemannian connection and acts on spinorial

indices, while A,, acts on the scalar indices 1, ..., N. We have
the covariance $frA = ' 1EA'k
The analytic index of the Dirac family {A}AE', which we
denote by pees is the formal difference {ker $A}AC6I - {ker
$A}Aeu. Each term is not a vector bundle over 2I because the
dimensions of ker $A and ker WA can jump (the same amount)
as A varies over W. Nevertheless, the formal difference is
well defined as an element of K(W1). Moreover, because of
the covariance of OA, ker AO.A = (ker YA), and the formal
difference is an element of K(W1) equivariant under 'S. In our
case it descends to an element of K(2/'6) which we denote
by Ind $.
The publication costs of this article were defrayed in part by page charge
payment. This article must therefore be hereby marked "advertisement"
in accordance with 18 U.S.C. 1734 solely to indicate this fact.

M XX ->M X
2597

%/(a.

2598

Proc. NatL Acad Sci. USA 81

Mathematics: Atiyah and Singer

The curvature 3Z; of w is easily computed. It is a horizontal

2-form with values in g, the Lie algebra of G, and has components of type (2, 0), (1, 1), and (0, 2) reflecting the product
base space M x W/h. The formulas for 9 at (p, A) are as
follows: (1) 9;l,"2 = F, ,J2(A) for ti, t2 E T(M, v (p)); (2) it,
(t) for t E T(M, ir (p)) and T E T(W1/', {A}) (so that T E
= Gb*(o() where G =
CG(A1 g) and DA* = 0); and (3)

(DADA)-1 and bT: AO 0

Al 0

--

g is

given byf-* [i, fi.

In

local coordinates, the (2, 0) component of the field is F,v,,. The

(1, 1) component is 8AM, while the (0, 2) component is
(DADA)f1[8AM, 8B,](x) with 8A and 5B in the background
gauge. Here, iT is the projection of P onto M.
If we apply the index formula for a family (2) to $va one
obtains Theorem 1.
THEOREM 1. ch(Ind $) = fMd(M) ch(f) where W8 = a XCN
a vector bundle over M x A/%9, a is the usual characteristic
class associated with the spinor index and ch is the Chern
character.
The curvature formulas above give explicit formulas for
the characteristic classes of % in terms of differential forms
(3). For example, suppose M = S2n, G = SU(N), and p is the
identity representation. Then, a(M) = 1 and the Chern character of Ind is expressed in terms of the Chern classes of W
integrated over M. These Chern classes are the invariant
polynomials kj(9;) where

lkj(T)tNj = det(tIN

+ 2

T).

The invariant polynomials kj(T) are also expressible in terms

of tr(Tk), and there is some simplification for SU(N) since
tr(T) = 0, So,
ko=

(1984)

as a 2 form on A/%6 evaluated on the pair of tangent vectors

a, Oa.
When p + Id, the above formulas hold with p 9 replacing
9; and trp (the trace in the p-representation) replacing tr.
Suppose G = SU(N), M = S2n, and s is the group of gauge
transformations leaving a point fixed. Then % is the group of
the principal bundle with base 21/' and total space 21, which
is topologically trivial. The Chern classes d2j = fs2. kj+n
(9)2n,2j, which are 2j forms on 21/%, can be lifted to forms on
W that are exact on A: d2j = df32j_1. Moreover, 132j-llOrbit =
t2j-1 is a closed 2j - 1 form on 9 representing a generator of
H2j- (8, R) (N 1j + n), modulo products of lower order.
Although 32j-l are determined only up to an exact differential, secondary characteristic classes give explicit formulas for P2j-1 and t2j-1 in terms of differential forms. Lift kj+n
(S;) from M x A/% to 9, where it equals dQa2j+2,l, with
a2j+2n-1 the secondary characteristic class (formula 73 in
ref. 3). That is, a2j+2n-1 = a2j+2n-1 (w) = (j + n). f j kj+n(w,
9i;t, ..., 9,)dt with i;, = t9; + 1/2(t - t2)[w, w]. Lift 2j+2n-1 to
P x W and denote it by &2j+2n-1. For simplicity, assume P =
M x G (the k = 0 sector) so that M x 21 C P x W. Let I32j-j
= fM a2j+2n-l a 2j - 1 form on 2, and let tz,_1 be the restriction of 132j-1 to an orbit '6 A.
THEOREM 2. d32j-l = d2j. When p = Id, then t2jj represents a primitive element in H2i-l((, R)j + n N-i.e., t2j_
represents a generator modulo products of lower order.
The nonproduct case is slightly more complicated. The Gconnection w on a comes from a G-connection wi' on P x W.
Choose a connection B on P and extend it to P x 21. The
form d2j+2n-1 used above is replaced by a where
-

kj+,(Ov)- kj+n(FB)

= da

and a is given by formula 70 in ref. 3-i.e.,

1,

k1(T)

2 tr(T)

k2(T)

k3(T)
k4(T)

a =

0,

and

87r2t(tr(T2),

i~

tr(T3),

-2-26

-61f4

(tr(T)4

kj+n(fv- B,-9;t, ..., 9;t)dt

(j + n)

9B+t(rv-B),

It should be remarked that the characteristic classes

-

(tr(T)4) +

d2j are not local, for they involve the Green's operator
(D*DA)-1 in the curvature i. However, the closed forms

(tr(T)2)2)

k22(T)/2, etc.

COROLLARY 1.1. Let M = S2'. The Chern classes ofInd

expressible in terms of d2j = fS2n kj+n(9;)2n,2j forms of
degree 2j on 21/', where kj 1(P;)2n,2i stands for the (2n, 2j)
component of kj+n(9).
For example, when M = S4, the 0th Chern class is
are

t2j-1 on IS are local and Theorem 2 implies they are directly

expressible in terms of the Chern-Simons secondary classes. That is, suppose fl, ..., f2j-1 are elements in the Lie
algebra of '6-i.e,, in C-(AO 0 SU(N)); because % acts on P,
the f s can be viewed as vertical vector fields on P = M x G.
They are also left invariant vector fields on '6.
Let i(f) denote interior product by the vector field f and

iffl,

...,

f2j-1)

i(f2j-1)

...

iff1).

Then, at 4 E 'S.

k2(9;)40=

tr(5;2)4,0=

tr(F2),

i*
**, f2jI iff 1
t2j-.(ff,
f2j-l)a2j+2n-1((A).

the usual Pontrjagin index. While the first Chern class cl of

Ind equals

For example, for M = S4 andj = 1, we obtain the 1-form

t1(f)

t4k3(9;)4,2=-434
24ir3 4 tr(93 4,2
=

= f
+

i3

ry tr{ ~Fa ,Fy

FajGb*(oa)Fys

b*(oj

Fa,(TyoTA

oATy)}

'/2(t

i(f)a5(4A) =-24
-

3 f i(f) f

tr(4A(tFPA

t2)[+A, 4OA]) (tFOA + '/2(t - t2)[kA, kA])dt.

This formula for t1 is the formula for a nonabelian chiral

anomaly (4-6). See also refs. 7-9 for a self contained account

Mathematics:

Atiyah and Singer

Proc. NatL Acad. Sci. USA 81 (1984)

of the relationship between anomalies in all dimensions and

secondary characteristic classes.
One interpretation for this anomaly involves determinants. Consider the operator To = Z4,A: CN(S 0 E) -C'(S+ 0 E), when $A and #B have no zero frequency modes.
The operator To, is a Laplacian plus lower-order term. It has
pure point spectrum {AX}, and all but a finite number of eigenvalues lie inside a wedge about the positive real axis. Hence,
Y.7S-' makes sense except for a finite number of eigenvalues
lying on the negative real axis.
When T has positive eigenvalues one can define log det T
as
d
ds s=o tr(T-s).
We extend this definition by letting I-P denote projection on
a finite dimensional space spanned by the eigenfunctions
having eigenvalues XI, ..., Xk, including those eigenvalues in

[-oo, 0]. Let

det

T<1, = elog det (PTO).Fl

j=1

Xi,

which is well defined. Moreover, --- det T4, is a smooth

nonvanishing complex valued function on A. Since log det T4,
may not be definable, det T4, can give a nontrivial element in
H1(9, Z). A direct computation using ; function regularization gives Theorem 3.
THEOREM 3. t, = - d(det TO,)/det T4, + df; that is, t,
and 1 d(det TO,)/det2i7riTO, represent the same element of H'

(W/%, R).
As explained above, the 1-form t, on IS comes from the
first Chern class d2 of Ind #
a 2-form on 21/' equaling f54
k3(9;)4,2. The first Chern class of Ind $ is the Chem class of

the determinant line bundle of Ind $, and it has the following

physical interpretation. Consider the fermionic path integral

Jr(A) = f e*V(y,)q4(x)u(y2)1(x2)

...

4(Y,)(2

which equals

E(-1) r(det 0*0,A) 12{EA(y<1), xi)EA(Y*2), x2)

...

EA(Yirr), XX)}

is a permutation and EA is the propagator for

EA in terms of the eigenvectors 4f of $A*A and 4i
ir

MA$A obtaining

E1j(Y) 0

$A.
=

Expand

$Aqij/Xj of

41( W

E>1
0J(y)4f(x) (ll()
=

(x

on A. The expression makes sense

B
when there are no zero frequency modes. Suppose A
with XA
0. The expression 51(A) approaches det/Xl

all quantities depending

+14q(y)

--

However, i1, is determined only up to

a phase and a consistent choice must be made.
For r > 1, it is easy to see that because of the exclusion
principle, Jr(B) is indeterminate only when there are exactly
r zero frequency modes for $B. Moreover, the indetermin0

4j(x) with $B4q,

0.

ancy depends only on a phase, the choice of a generator in

Ar(ker OB), the 1-dimensional space of skewsymmetric r tensors of ker

EBB

THEOREM 4. A gauge covariant Sir(A) smooth in A exists if

and only if the determinant line bundle of Ind $ is triviali.e., d2 = 0 in H2(21/,, Z) or t1 = 0 in H1('8, Z).

The characteristic forms d2jEH2j(21f/, Z) are obstructions

to the existence of a covariant propagator for #We*. We ask
the question: Do the higher obstructions have physical sig-

nificance?
Using our earlier discussion of the topological index, one
can show, for M = S2n and G = SU(N), Theorem 5.
THEOREM 5. If p is the identity representation, then d2J E
H2W(W/'/, R) and t2pi E H2- '(%, R) do not vanish forj . N

- n.
Gravitational anomalies are the subject of a recent preprint (10), especially for the Dirac operator, the RaritaSchwinger operator, and the signature operator. These operators are dependent on the metric and are covariant under
diffeomorphisms. The formulas obtained in ref. 10 by perturbative calculations at the one-loop level can also be obtained
by the methods described in this paper, using the families
index and secondary characteristic classes (unpublished result; 0. Alvarez and B. Zumino, personal communication).
Specifically, W is replaced by the space of all metrics V of
the manifold M. 'S is replaced by the group of diffeomorphisms of M leaving a basis at one point fixed (Diffo(M)).
Each metric p E V gives a Dirac operator Op (and other geometric operators) with the covariance 4,,p = 0-10p4 for X E
Diffo(M). Thus, ?/Diffo(M) is the parameter space for the
family {$p}.
The space P x S is replaced by a sub-bundle of B x V
where B is the bundle of bases of M. The sub-bundle is the
set of all frames relative to each metric p E V. The group
Diffo(M) acts on the sub-bundle and gives a quotient Q
which is a principal O(n) bundle over a base space, itself a
fiber space over 3/Diffo(M) with fiber M. The first Chern
class of the family can be promoted to a 1-form on Diffo(M),
which is directly expressible in terms of secondary characteristic classes. Since only Pontrdagin classes are involved,
nonzero results are obtained only in dimensions n = 4k + 2.
It is our pleasure to record our gratitude to many physicists who
have helped us understand anomalies: 0. Alvarez, D. Friedan, J.
Goldstone, R. Jackiw, R. Stora, E. Witten, and B. Zumino. We are
especially indebted to D. Quillen for reading and correcting an error
in our original manuscript. He has independently investigated the
determinant line bundle, its Chern class, and the corresponding
anomaly for the a family on Riemannian surfaces. I.M.S. was supported in part by National Science Foundation Grant MCS80-23356
and in part by the Miller Foundation.
Note Added in Proof. An exposition of the first obstruction and its
relation to the chiral anomaly, intended primarily for physicists, can
be found in ref. 11.
1. Atiyah, M. F. & Jones, J. D. S. (1978) Commun. Math. Phys.
61, 97-118.
2. Atiyah, M. F. & Singer, I. M. (1971) Ann. Math. 93, 119-138.
3. Chern, S. S. (1972) Geometry of Characteristic Classes, Proceedings of the 13th Biennial Seminar (Canadian Mathematical
Congress), pp. 1-40; reprinted in Chem, S. S. (1979) Complex

In particular
e

2599

Manifolds Without Potential Theory (Springer, New York).

4. Bardeen, W. A. (1969) Phys. Rev. 184, 1848-1859.
5. Gross, D. J. & Jackiw, R. (1972) Phys. Rev. D 6, 477-493.
6. Bonora, L. & Cotta-Ramusino, P. (1983) Commun. Math.

Phys. 87, 589-600.

7. Zumino, B. (1984) in ChiralAnomalies and Differential Geometry, Relativity, Groups, and Topology, eds. De Witt, B. &
Stora, R. (North-Holland, Amsterdam), in press.
8. Zumino, B., Yang-Shi, W. & Zee, A. (1984) Nucl. Phys. B, in
press.

2600

Mathematics: Atiyah and Singer

9. Stora, R. (1984) in Algebraic Structure and Topological Origin

of Anomalies: Progress in Gauge Field Theories, eds. Hooft,
G., Jaffe, A., Lehmann, H., Mitter, P. K., Singer, I. M. &
Stora, R. (Plenum, New York), in press.

Proc. Nadl. Acad. Sci. USA 81 (1984)

10. Alvarez-Gaumd, L. & Witten, E., Gravitational Anomalies,
preprint HUTP-83/A039.
11. Alvarez-Gaum6, L. & Ginsparg, P. (1984) Nucl. Phys. B, in
press.