The Atiyah-Singer Index Theorem, Dirac Operators & Spin Geometry

The shape of the miracle: analytic index equals topological index

Guide 1 watched Perelman steer a metric by Ricci flow until topology fell out of the geometry. This guide chases a different bridge between analysis and topology, but the punchline has the same flavor: a quantity that looks purely analytic turns out to be a topological invariant, immovable under any deformation. The hero is the index of an elliptic operator. Take a linear differential operator D acting between sections of two vector bundles over a compact manifold M; if D is elliptic, both its kernel (solutions of D s = 0) and the cokernel (obstructions to solving D s = f) are finite-dimensional. Their difference, index(D) = dim ker D - dim coker D, is the analytic index.

Here is why the index is special, and it is worth feeling before any formula. The dimensions dim ker D and dim coker D each jump wildly as you wiggle D — perturb the coefficients and a solution can appear or vanish. But they jump TOGETHER: every time a kernel dimension is lost, a cokernel dimension is lost in lockstep, so their difference never moves. The index is therefore a homotopy invariant of D, deaf to continuous deformation. The Atiyah-Singer index theorem (Atiyah and Singer, 1963) answers the question this stability forces upon you: if the index does not depend on the analytic details, it must be computable from topology alone — and indeed it equals a topological index, an integral over M of explicit characteristic classes built from the symbol of D.

Warm-up you already own: Gauss-Bonnet is an index theorem

Do not take the index theorem on faith — you have met a case of it already and can verify the pattern by hand. Recall the de Rham complex from the Forms rung: the exterior derivative d maps k-forms to (k+1)-forms with d-squared = 0. Bundle the even-degree forms together and the odd-degree forms together, and form the single operator D = d + d-star going from even forms to odd forms (d-star is the formal adjoint built from the Hodge star). This D is elliptic, and a short computation with Hodge theory shows its kernel is the even-degree cohomology and its cokernel the odd-degree cohomology.

So index(D) = (sum of even Betti numbers) - (sum of odd Betti numbers) = chi(M), the Euler characteristic. That is the analytic index. The topological index for THIS D is the integral over M of the Euler class — the Pfaffian of the curvature — which on a surface is exactly (1 / 2 pi) integral over M of the Gaussian curvature K. Setting them equal recovers the Gauss-Bonnet theorem you proved long ago: integral of K dA = 2 pi chi(M). Stare at that: Gauss-Bonnet is the Atiyah-Singer index theorem for the de Rham operator. The general theorem simply says EVERY elliptic operator obeys a law of this exact shape.

ONE THEOREM, MANY FACES   (analytic index  =  topological index)

  operator D            analytic index = dim ker - dim coker      topological side
  ------------------    -----------------------------------       -------------------------
  d + d*  (de Rham)     Euler characteristic  chi(M)               integral of Euler class
  signature operator    signature  sigma(M)                       integral of L-class (Hirzebruch)
  Dolbeault  dbar+dbar* arithmetic genus / chi(O)                  integral of Todd class (Riemann-Roch)
  Dirac operator  D     A-hat genus (index of Dirac)               integral of A-hat class

  Gauss-Bonnet, Hirzebruch signature, Hirzebruch-Riemann-Roch  =  three special cases of ONE line.

The Dirac operator: a square root of the Laplacian

Of all elliptic operators the index theorem governs, one is fundamental: every classical case above is, in the right sense, a twisted version of it. This is the Dirac operator, born when Paul Dirac asked for a first-order operator D whose square is the Laplacian — D-squared = Delta. You cannot do this with scalar coefficients; the cross terms refuse to cancel. Dirac's trick was to let the coefficients be MATRICES gamma_i obeying gamma_i gamma_j + gamma_j gamma_i = -2 delta_ij. Then D = sum gamma_i nabla_i squares to the Laplacian precisely because those anticommutation relations kill every cross term. Those relations define a Clifford algebra, and the matrices act on an auxiliary space of spinors — that is the whole algebraic seed of spin geometry.

To globalize Dirac's local matrices into an operator on a manifold M, you must consistently choose the spinor bundle chart by chart. The obstruction to doing so is topological and is the content of spin geometry: M admits a spin structure exactly when its second Stiefel-Whitney class w_2 vanishes (orientability handles w_1 first). The orientable surfaces all qualify; the real projective plane RP^2 does not. This is an honest hypothesis you must check, not a formality — when w_2 is nonzero there is literally no globally defined Dirac operator, only a spin-c variant that buys existence by adding a line bundle, the door into Seiberg-Witten theory in Guide 3.

Curvature speaks: the Weitzenbock formula and Lichnerowicz

The Dirac operator does not just square to the rough Laplacian — it squares to the Laplacian PLUS a curvature term, and that one extra term is where geometry talks to topology directly. The Lichnerowicz-Weitzenbock formula says D-squared = nabla-star nabla + S/4, where nabla-star nabla is the (nonnegative) connection Laplacian and S is the scalar curvature. Read it as a sum of a manifestly nonnegative operator and a pure pointwise multiple of curvature. The Bochner technique you met in the Comparison rung is exactly this style of argument, now run on spinors instead of forms.

Now squeeze a genuine theorem out of it, and watch index theory and curvature collide. The index of the Dirac operator equals a topological number called the A-hat genus, an integral of characteristic classes over M. Suppose M is a closed spin manifold carrying a metric of strictly positive scalar curvature S > 0. Then the curvature term S/4 is strictly positive, so D-squared = nabla-star nabla + S/4 is strictly positive, hence has no kernel — meaning D itself has no harmonic spinors, so index(D) = 0. But index(D) = A-hat(M) is a fixed topological number. The conclusion (Lichnerowicz) is stark: a closed spin manifold with nonzero A-hat genus CANNOT admit any metric of positive scalar curvature. Pure topology forbids a geometric possibility — an obstruction you could never see by staring at the metric.

Proving it with heat: the McKean-Singer cancellation

Among the proofs, one is so geometric it belongs in any first encounter: the heat-kernel proof. Run the heat equation for D-squared on each spinor bundle and watch the supertrace of the heat operator, str(exp(-t D-squared)) = trace on plus minus trace on minus. The McKean-Singer miracle is that this supertrace is INDEPENDENT of t and equals index(D) for every t > 0 — because every nonzero eigenvalue of D-squared pairs a plus-eigenstate with a minus-eigenstate that cancel in the supertrace, leaving only the zero modes, which are exactly ker and coker.

1. Write index(D) = str(exp(-t D-squared)) for ALL t > 0 — the McKean-Singer identity. The left side is the integer you want; the right side is now free to be evaluated at whatever t is convenient.
2. Send t toward 0. The heat kernel near the diagonal has a universal short-time asymptotic expansion in t whose coefficients are local polynomials in the curvature of the metric and the bundle connection — pure differential geometry, computable pointwise.
3. Integrate the t-to-0 limit over M. The famous fantastic cancellation (Getzler's rescaling makes it transparent) collapses the messy expansion to exactly one surviving term: the A-hat class times the Chern character of the twisting bundle.
4. Equate the two evaluations. The t-independent integer (analytic index) equals the integral over M of those characteristic classes (topological index). That single equation IS the Atiyah-Singer index theorem.

The same engine prints the classical theorems on demand. Feed it the de Rham operator and the surviving class is the Euler class, giving Gauss-Bonnet. Feed it the Dolbeault operator dbar + dbar-star on a complex manifold and the surviving class is the Todd class, giving Hirzebruch's form of the Riemann-Roch theorem you met on Riemann surfaces — now valid in every dimension. Feed it the signature operator and you get Hirzebruch's signature theorem. One operator, one cancellation, and the great theorems of the century drop out as corollaries.

What you have built, and where it flows next

Assemble the arc. An elliptic operator D on a compact M has an integer index that is stubbornly topological because ker and coker shed dimensions together. The Dirac operator — Dirac's matrix square root of the Laplacian, existing only when spin geometry permits via w_2 = 0 — is the universal case, and the Weitzenbock formula turns its curvature term into a hard obstruction (Lichnerowicz: nonzero A-hat genus forbids positive scalar curvature). The heat-kernel proof then prints the whole family — Gauss-Bonnet, Riemann-Roch, signature — from one t-independent supertrace.

1. Guide 3 cashes spin geometry into low dimensions: the Dirac operator twisted by a line bundle, plus the spin-c structures that rescue manifolds with w_2 nonzero, are the raw material of Seiberg-Witten theory and Donaldson's invariants of smooth 4-manifolds.
2. The characteristic-class integrals on the topological side are exactly the Chern-Weil story from the bundles rung — the A-hat, Todd, and L classes are specific polynomials in the curvature form, so revisit that machinery if any of these classes felt like a black box.
3. The index theorem also reaches into physics, where the Dirac operator is literally the operator of relativistic fermions and the A-hat term is an anomaly; that bridge to gauge theory and the Yang-Mills functional reappears throughout the rest of this rung.