The Atiyah-Singer Index Theorem

⚡ Key Takeaways

The Atiyah-Singer Index Theorem asserts that for an elliptic differential operator $D$ on a compact smooth manifold, the analytic index (dimension of solutions minus dimension of cosolutions) equals a purely topological index computed from characteristic classes. In words: analysis = topology for elliptic operators. It applies to elliptic (pseudo)differential operators on closed manifolds and unifies the Gauss-Bonnet, Riemann-Roch-Hirzebruch, and Hirzebruch signature theorems. Proved by Michael Atiyah and Isadore Singer in 1963 (announcement) and published in a celebrated series 1968–1971.

💡 The Core Insight

An elliptic operator $D$ has a symbol $\sigma(D)$ , a bundle map that becomes invertible away from the zero section of $T^*M$ . This invertibility data defines a class in compactly supported K-theory $K^0(T^*M)$ . The analytic index factors through this K-theory class — only the symbol's homotopy type matters, not $D$ itself. K-theory carries two natural maps to $\mathbb{Z}$ : the analytic index and the topological index (defined via embeddings into Euclidean space and Bott periodicity). The theorem is the statement that these two maps coincide. Equivalently, the index is a cobordism invariant, and characteristic numbers detect cobordism classes.

DefinitionElliptic operator and symbol

Let $M$ be a closed (compact, boundaryless) smooth manifold of dimension $n$ , and let $E,F\to M$ be smooth complex vector bundles. A linear differential operator of order $m$ ,

1

\quad D=\sum_{|\alpha|\le m} a_\alpha(x)\,\partial^\alpha : \Gamma(E)\to\Gamma(F),

Derivation hint: $\alpha$ is a multi-index, $a_\alpha(x)\in\operatorname{Hom}(E_x,F_x)$ are smooth bundle homomorphisms, $\partial^\alpha=\partial^{\alpha_1}_{x_1}\cdots\partial^{\alpha_n}_{x_n}$ in local coordinates and trivializations. has principal symbol at $(x,\xi)\in T^*M$ given by

2

\quad \sigma_m(D)(x,\xi)=\sum_{|\alpha|=m} a_\alpha(x)\,(i\xi)^\alpha\;\in\;\operatorname{Hom}(E_x,F_x).

Derivation hint: keep only top-order terms and replace $\partial_{x_j}\mapsto i\xi_j$ ; this is the leading behavior under Fourier transform / the principal part of $D$ . $D$ is elliptic iff $\sigma_m(D)(x,\xi)$ is an isomorphism for all $x\in M$ and all $\xi\neq 0$ .

DefinitionAnalytic index

For an elliptic operator on a closed manifold, both $\ker D$ and $\operatorname{coker} D$ are finite-dimensional (elliptic regularity + Fredholm theory on Sobolev spaces). The analytic index is

3

\quad \operatorname{ind}_a(D)=\dim_{\mathbb C}\ker D-\dim_{\mathbb C}\operatorname{coker} D\;\in\;\mathbb Z.

Derivation hint: $D$ extends to a bounded Fredholm map $H^s(E)\to H^{s-m}(F)$ between Sobolev spaces; the index is independent of $s$ and stable under compact perturbations.

TheoremAtiyah-Singer Index Theorem — cohomological form

Let $D$ be an elliptic differential operator of order $m$ on a closed oriented manifold $M^n$ , with symbol class $[\sigma(D)]\in K^0(T^*M)$ . Then the analytic index equals the topological index:

4

\quad \operatorname{ind}_a(D)=\operatorname{ind}_t(D)=(-1)^n\int_{T^*M}\operatorname{ch}\big([\sigma(D)]\big)\,\pi^*\operatorname{Td}(TM\otimes\mathbb C).

Derivation hint: $\operatorname{ch}$ is the Chern character, $\operatorname{Td}$ the Todd class, $\pi:T^*M\to M$ the projection; the integral is over the total space $T^*M$ (an open manifold, with the integrand compactly supported by ellipticity). The sign $(-1)^n$ comes from orientation conventions on $T^*M$ . Equivalently, for the standard geometric operators one obtains a class on $M$ itself: if $D$ is the Dirac operator twisted by a bundle $E$ ,

5

\quad \operatorname{ind}(D_E)=\int_M \hat{A}(TM)\,\operatorname{ch}(E),

Derivation hint: specialize [4]; the $T^*M$ integral collapses to an $M$ integral by Thom isomorphism, the Todd class combines with the symbol's Bott class to produce the $\hat{A}$ -genus $\hat A(TM)=\prod_j \frac{x_j/2}{\sinh(x_j/2)}$ (Pontryagin roots $x_j$ ).

TheoremK-theoretic form

There are two homomorphisms $K^0(T^*M)\to\mathbb Z$ :

6

\quad \operatorname{ind}_a:\;[\sigma]\longmapsto \dim\ker D-\dim\operatorname{coker} D,\qquad \operatorname{ind}_t=p_!,

Derivation hint: $\operatorname{ind}_t$ is the pushforward (Gysin map) to a point, constructed by choosing an embedding $M\hookrightarrow \mathbb R^N$ , using the Thom isomorphism for the normal bundle and Bott periodicity $K^0(\mathbb R^{2k})\cong\mathbb Z$ . and the theorem is the operator-free identity $\operatorname{ind}_a=\operatorname{ind}_t$ as maps on $K^0(T^*M)$ .

ProofEmbedding/K-theory proof — structure

The original 1968 proof reduces the theorem to two functorial properties and a single normalization, following the strategy used by Hirzebruch–Grothendieck for Riemann-Roch.

Step 1 (Symbol class). Ellipticity of $D$ means $\sigma(D)$ is an isomorphism off the zero section, hence the triple $(\pi^*E,\pi^*F,\sigma)$ defines a difference class $[\sigma(D)]\in K^0(T^*M)=K^0_{\mathrm{cpt}}(T^*M)$ . The analytic index depends only on this class:

7

\quad \operatorname{ind}_a:K^0(T^*M)\to\mathbb Z\quad\text{is well-defined and a homomorphism.}

Derivation hint: homotopies of symbols through elliptic ones, and direct sums of operators, give the homomorphism property; lower-order terms do not affect the index (compact perturbation).

Step 2 (Axioms). Both $\operatorname{ind}_a$ and $\operatorname{ind}_t$ satisfy: (i) Normalization: on a point, $\operatorname{ind}=$ identity on $\mathbb Z$ . (ii) Excision / multiplicativity under embeddings $i:M\hookrightarrow M'$ , compatible with the induced K-theory pushforward $i_!$ .

Step 3 (Embedding into Euclidean space). Choose $M\hookrightarrow\mathbb R^N$ . The Gysin/pushforward $i_!:K^0(T^*M)\to K^0(T^*\mathbb R^N)$ commutes with $\operatorname{ind}_a$ (analytically: a tubular neighborhood theorem for operators) and with $\operatorname{ind}_t$ (topologically: by definition of $\operatorname{ind}_t$ ).

8

\quad \operatorname{ind}_a = \operatorname{ind}_a^{\,\mathbb R^N}\circ i_!,\qquad \operatorname{ind}_t = \operatorname{ind}_t^{\,\mathbb R^N}\circ i_!.

Derivation hint: the analytic invariance under $i_!$ is the technical heart — an elliptic operator on $M$ corresponds to one on $\mathbb R^N$ with the same index via the Thom class of the normal bundle.

Step 4 (Reduce to $\mathbb R^N$ and Bott). On $T^*\mathbb R^N\cong\mathbb R^{2N}$ Bott periodicity gives $K^0_{\mathrm{cpt}}(\mathbb R^{2N})\cong\mathbb Z$ generated by the Bott class, on which $\operatorname{ind}_a$ and $\operatorname{ind}_t$ each compute to $1$ .

9

\quad \operatorname{ind}_a^{\,\mathbb R^N}=\operatorname{ind}_t^{\,\mathbb R^N}:K^0_{\mathrm{cpt}}(\mathbb R^{2N})\xrightarrow{\ \cong\ }\mathbb Z.

Derivation hint: the generator is represented by the de Rham/Dolbeault-type complex on $\mathbb C^N$ whose index is $1$ ; Bott periodicity makes this the universal normalization.

Conclusion. By Steps 3–4, $\operatorname{ind}_a$ and $\operatorname{ind}_t$ agree on the image of $i_!$ , which is all of $K^0(T^*M)$ up to the isomorphism. Hence $\operatorname{ind}_a=\operatorname{ind}_t$ on $K^0(T^*M)$ , giving [4]. Translating $\operatorname{ind}_t$ via Chern character and the Thom isomorphism produces the cohomological formula. $\blacksquare$

∎

ProofHeat-kernel (local) proof — McKean–Singer

A second, more analytic proof yields [5] directly and explains the appearance of $\hat A$ .

Step 1 (McKean–Singer formula). For $D:\Gamma(E)\to\Gamma(F)$ form the Laplacians $D^*D$ , $DD^*$ . For all $t>0$ ,

10

\quad \operatorname{ind}(D)=\operatorname{Tr}\big(e^{-tD^*D}\big)-\operatorname{Tr}\big(e^{-tDD^*}\big).

Derivation hint: nonzero eigenvalues of $D^*D$ and $DD^*$ coincide with equal multiplicity (the map $D$ intertwines them), so they cancel; only the kernels survive, independent of $t$ .

Step 2 (Heat-kernel asymptotics). As $t\to 0^+$ the supertrace has an asymptotic expansion in $t$ with coefficients that are integrals of local invariants of the metric and connection:

11

\quad \operatorname{Tr}_s\big(e^{-tD^2}\big)\sim\sum_{k\ge 0} t^{(k-n)/2}\int_M a_k(x)\,dx.

Derivation hint: Minakshisundaram–Pleijel expansion of the heat kernel $K(t,x,x)$ ; the supertrace $\operatorname{Tr}_s=\operatorname{Tr}_E-\operatorname{Tr}_F$ .

Step 3 (Fantastic cancellation). Since the left side is the $t$ -independent integer $\operatorname{ind}(D)$ , all $t\to0$ singular terms vanish and only the $t^0$ coefficient survives. Getzler's rescaling shows this coefficient is exactly $\big[\hat A(TM)\operatorname{ch}(E)\big]_{\text{top}}$ .

12

\quad \operatorname{ind}(D)=\int_M \big[\hat A(TM)\,\operatorname{ch}(E)\big]_n.

Derivation hint: Getzler's symbol calculus rescales the Clifford variables and time so the heat operator limits to a harmonic-oscillator (Mehler) kernel whose trace is precisely the $\hat A$ characteristic form. $\blacksquare$

∎

Figure

Schematic of the index map: from elliptic symbol through K-theory to an integer

mathematician · Michael Atiyah & Isadore Singer | 1929–2019 / 1924–2021

Atiyah, an Oxford/Cambridge algebraic geometer and topologist (Fields Medal 1966), brought K-theory and characteristic-class machinery; Singer, an MIT analyst, brought operator theory and PDE. Their collaboration, beginning around 1962, fused two disciplines that had barely communicated. The pair received the Abel Prize in 2004 specifically for the index theorem. Atiyah's later work linked index theory to gauge theory and physics; Singer's pushed into eta-invariants and noncommutative geometry. The breadth of collaborators — Bott, Patodi, Hirzebruch, Donaldson, Witten — radiates from this single result.

🧠 Intuition

Think of $\operatorname{ind}(D)$ as a signed count that refuses to change under deformation. If you wiggle the coefficients of $D$ continuously, individual eigenvalues drift, kernels and cokernels can jump — but they jump in tandem: a vector that leaves $\ker D$ pairs with one entering $\operatorname{coker} D$ , so the difference is locked. This rigidity is exactly what a topological invariant should have. The miracle is that this rigidity is strong enough to be computed without solving any equation — the answer depends only on the symbol, the leading-order "shape" of $D$ , and through it only on the manifold's curvature integrated into characteristic classes. The simplest avatar is Gauss-Bonnet: the Euler characteristic (a count of "holes") equals $\frac{1}{2\pi}\int K\,dA$ (a curvature integral). The index theorem says every elliptic operator obeys a Gauss-Bonnet-type law, with $\hat A$ or $L$ or $\operatorname{Td}$ replacing curvature.

EW

Edward Witten

b. 1951

Witten reinterpreted the index theorem through supersymmetric quantum mechanics: the Dirac operator becomes a supercharge $Q$ , the index becomes the Witten index $\operatorname{Tr}(-1)^F e^{-\beta H}$ counting supersymmetric ground states, and the heat-kernel proof becomes a path-integral localization onto constant loops. This physical lens (1982) demystified the "fantastic cancellation" of the heat-kernel proof and seeded the field of index theory in QFT, anomalies, and the Atiyah-Bott fixed-point formula in physics. Anomaly cancellation in the Standard Model and string theory is, at root, an index computation.

ExampleGauss-Bonnet from the de Rham complex

Given: A closed oriented surface $M^2$ (so $n=2$ ), with the de Rham operator $D=d+d^*:\Omega^{\mathrm{even}}\to\Omega^{\mathrm{odd}}$ .

Steps: 1. The index of $d+d^*$ from even to odd forms equals the Euler characteristic:

13

\quad \operatorname{ind}(d+d^*)=\sum_k (-1)^k \dim H^k_{\mathrm{dR}}(M)=\chi(M).

Derivation hint: by Hodge theory $\ker(d+d^*)$ restricted to even/odd forms gives harmonic even/odd forms, i.e. $\sum b_{\text{even}}-\sum b_{\text{odd}}=\chi$ . 2. The index formula [5] specializes (Euler class form) to $\operatorname{ind}=\int_M e(TM)$ , the Euler class. For a surface $e(TM)=\frac{1}{2\pi}K\,dA$ .

14

\quad \chi(M)=\frac{1}{2\pi}\int_M K\,dA.

Derivation hint: $e(TM)$ in two dimensions is the Gaussian curvature 2-form divided by $2\pi$ ; this is the Chern-Gauss-Bonnet integrand. 3. For the round sphere of radius $R$ : $K=1/R^2$ , area $=4\pi R^2$ .

Answer: $\boxed{\chi(S^2)=\frac{1}{2\pi}\cdot\frac{1}{R^2}\cdot 4\pi R^2=2}$ (dimensionless integer), matching $\chi(S^2)=2$ .

ExampleRiemann-Roch for a line bundle on a curve via the Dolbeault operator

Given: A compact Riemann surface $X$ of genus $g$ , a holomorphic line bundle $L$ of degree $d=\deg L$ . Take $D=\bar\partial_L:\Omega^{0,0}(L)\to\Omega^{0,1}(L)$ .

Steps: 1. The analytic index is the holomorphic Euler characteristic:

15

\quad \operatorname{ind}(\bar\partial_L)=\dim H^0(X,L)-\dim H^1(X,L)=\chi(X,L).

Derivation hint: $\ker\bar\partial_L=H^0$ , $\operatorname{coker}\bar\partial_L\cong H^1$ by Dolbeault + Serre duality. 2. The topological index [4] specializes (complex dimension $1$ ) to the Hirzebruch-Riemann-Roch integrand $\operatorname{ch}(L)\operatorname{Td}(TX)$ with $\operatorname{Td}(TX)=1+\tfrac12 c_1(TX)$ and $\operatorname{ch}(L)=1+c_1(L)$ :

16

\quad \chi(X,L)=\int_X\Big(c_1(L)+\tfrac12 c_1(TX)\Big)=\deg L+\tfrac12\chi(X)=d+(1-g).

Derivation hint: keep the degree-2 (top) part; $\int_X c_1(L)=d$ and $\int_X c_1(TX)=\chi(X)=2-2g$ .

Answer: $\boxed{h^0(L)-h^1(L)=d+1-g}$ — the classical Riemann-Roch theorem, recovered as one line of the index theorem. For $L=\mathcal O$ (degree $0$ ): $h^0-h^1=1-g$ , i.e. $1-g$ , consistent with $h^0=1,\,h^1=g$ .

⚠️ Common mistake

The theorem does NOT compute $\dim\ker D$ or $\dim\operatorname{coker} D$ individually. Only their difference is a topological invariant; the separate dimensions depend on the metric, the connection, and lower-order terms, and can jump under deformation. Asking the index theorem for $\dim\ker D$ alone is a category error. Likewise, ellipticity and compactness of $M$ are both essential: on a noncompact manifold the operator need not be Fredholm and "the index" may be ill-defined without boundary or APS conditions.

⚠️ Common mistake

Orientation, complex vs. real, and sign conventions matter. The factor $(-1)^n$ in [4], the choice of $\operatorname{Td}(TM\otimes\mathbb C)$ vs $\hat A(TM)^2$ , and whether forms are taken with $2\pi i$ normalizations are sources of constant confusion. Always fix the convention $\operatorname{ch}=\sum \tfrac{1}{k!}\big(\tfrac{c}{?}\big)^k$ at the outset. Many "errors" in student computations are just normalization mismatches.

Counterexample

A non-elliptic operator has no index theorem. Consider $\partial/\partial x_1$ on the torus $T^2$ . Its symbol is $i\xi_1$ , which vanishes whenever $\xi_1=0,\ \xi_2\ne 0$ — not invertible off the zero section, hence not elliptic. The kernel is infinite-dimensional (any function of $x_2$ alone), so $\operatorname{ind}$ is not even defined as an integer. This shows ellipticity is not a cosmetic hypothesis but the load-bearing one: it is precisely what makes $D$ Fredholm and the symbol class live in K-theory.

connection

The index theorem sits at the crossroads of several theories. Via the symbol class it is fundamentally a statement in Topological K-Theory and uses Bott Periodicity for its normalization. The right-hand side lives in Characteristic Classes (Chern, Pontryagin, Euler), packaged into the Todd Class and Â-genus. The geometric operators it specializes to come from Clifford Algebras and Spin Geometry (the Dirac Operator). Its proofs use Hodge Theory, Pseudodifferential Operators, and (for the local version) Heat Kernel Methods. The boundary version, the Atiyah-Patodi-Singer Index Theorem, introduces the eta-invariant and connects to Spectral Flow. Through Witten's reinterpretation it links to Supersymmetry and Quantum Anomalies.

🔬 Where this appears

⏳ History

- 1827–1848: Gauss-Bonnet theorem links curvature to Euler characteristic — the prototype "analysis = topology" identity.

- 1953–1954: Hirzebruch proves the Signature Theorem and Riemann-Roch-Hirzebruch, expressing invariants via

L

-genus and Todd genus — the patterns the index theorem will generalize.

- 1957: Grothendieck reformulates Riemann-Roch functorially using

K

-theory of coherent sheaves, foreshadowing the K-theoretic proof.

- 1959–1961: Gelfand poses the general problem of computing the index of an elliptic operator topologically.

- 1963: Atiyah and Singer announce the index theorem (Bull. AMS), with the embedding/K-theory proof strategy.

- 1965–1966: Atiyah's K-theory book and Bott periodicity provide the tools; the result is presented in the Palais seminar notes (1965).

- 1968: The definitive series "The Index of Elliptic Operators I–III" (with Atiyah-Segal for the equivariant case) appears in Annals of Mathematics.

- 1971: Atiyah-Singer IV–V extend to families (the families index theorem, valued in

K^0

of the base).

- 1973–1975: Atiyah-Patodi-Singer treat manifolds with boundary, introducing the eta-invariant.

- 1982: Witten's supersymmetric proof; Getzler (1983–86) gives a short heat-kernel proof via rescaling.

- 2004: Atiyah and Singer awarded the Abel Prize.

open-problem

Several frontiers remain active. (1) Index theory on noncompact and singular spaces — operators on manifolds with corners, conical/cusp singularities, or stratified spaces require new analytic frameworks (Melrose's b-calculus, Cheeger's intersection cohomology); a fully general index formula is incomplete. (2) Higher index theory and the Baum-Connes / Novikov conjectures — for noncompact groups, the index lives in the K-theory of a group $C^*$ -algebra, and the equality with a topological assembly map is the Baum-Connes conjecture, open in general (with counterexamples for certain Gromov groups). (3) Index theory of foliations (Connes' noncommutative geometry) and the longitudinal index theorem extend the framework but raise measurability and traciality questions. (4) $L^2$ -index theory (Atiyah's $L^2$ -index theorem on covers) feeds the Atiyah conjecture on integrality of $L^2$ -Betti numbers, only partially resolved. (5) Making the families index and equivariant versions compatible with positive scalar curvature obstructions (Gromov-Lawson-Rosenberg conjecture) is ongoing.

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also