What an index is, and why it refuses to move
Take a linear differential operator D acting on sections of a vector bundle over a closed manifold M — think of the Laplacian, the Dirac operator, or the d-bar operator on a complex manifold. Two numbers describe its solvability: dim(ker D), the space of solutions of D u = 0, and dim(coker D), the obstructions to solving D u = f. Each one alone is wildly sensitive — perturb D a little and individual solutions appear or vanish. But their difference, the analytic index ind(D) = dim(ker D) - dim(coker D), is astonishingly rigid: as a solution drops out of the kernel it reappears in the cokernel, and the difference never budges. That stubbornness is the entire reason an index can equal a topological invariant.
The class of operators where this works is the elliptic operators — those whose top-order part, the symbol, is invertible in every nonzero cotangent direction. Ellipticity is what guarantees the kernel and cokernel are finite-dimensional in the first place, so the index is even a well-defined integer. A first-order intuition: replace each derivative d/dx^j by the covector xi_j, keep only the highest-order terms, and ask whether the resulting matrix-valued function of xi in T*M is invertible away from xi = 0. The Laplacian's symbol is |xi|^2 times the identity — invertible — so it is elliptic; the wave operator's symbol vanishes on the light cone, so it is not.
Two old miracles you already trust
Before the general theorem, meet two special cases you have already lived through — both are secretly index theorems. The first is the Gauss-Bonnet theorem. The de Rham complex d: even forms -> odd forms, with its Hodge Laplacian, is elliptic, and its index counts even Betti numbers minus odd Betti numbers — exactly the Euler characteristic chi(M). The topological side is the integral of the Euler-class curvature, the Euler class of TM. So 'curvature integrates to chi(M)' is the index of the de Rham operator. You proved this for surfaces a whole rung ago; the index theorem is its grown-up form in every even dimension.
The second is the Riemann-Roch theorem. On a compact complex manifold the d-bar operator running through the Dolbeault complex is elliptic; its index is the holomorphic Euler characteristic chi(O) = sum of (-1)^q dim H^q(M, O), an analytic count of global holomorphic data. Hirzebruch-Riemann-Roch evaluates it as the integral over M of the Chern character of the relevant bundle times the Todd class of the tangent bundle. For a Riemann surface of genus g this collapses to the classical Riemann-Roch you met among divisors: deg(L) + 1 - g. One analytic side, one purely topological side, joined by an equals sign.
Atiyah and Singer's insight was that these were not two coincidences but two instances of one law. Gauss-Bonnet, Riemann-Roch, the Hirzebruch signature theorem (analytic index of the signature operator = the L-genus, a Pontryagin-class integral), and the spin Dirac case all fit a single template: an analytic index on the left, an integral of characteristic classes on the right. The job of the theorem is to write that universal right-hand side once and for all.
The statement: analytic index equals topological index
Here is the theorem in one breath. For an elliptic operator D on a closed manifold M, the analytic index equals the topological index, a number built entirely from the homotopy class of D's symbol and the topology of M. The Atiyah-Singer index theorem computes that topological side by a cohomological formula: integrate, over M, the Chern character of the symbol class times the Todd class of the complexified tangent bundle. In symbols, ind(D) = integral over M of ch(symbol of D) . Td(TM ⊗ C), with the integrand cut down to the top-degree piece. Every classical case is this one formula with a different operator dropped in.
analytic index: ind(D) = dim ker D - dim coker D (D elliptic, M closed) Atiyah-Singer: ind(D) = integral over M of ch(sigma(D)) . Td(TM (x) C) [ top-degree part ] special cases (one formula, different D): de Rham operator -> ind = chi(M) = integral of e(TM) [ Gauss-Bonnet ] signature operator -> ind = sign(M) = integral of L(TM) [ Hirzebruch ] Dolbeault d-bar (x) E -> ind = chi(M, E) = integral of ch(E).Td(TM) [ Riemann-Roch ] spin Dirac operator -> ind = A-hat genus = integral of A-hat(TM) [ integer! ]
Two features deserve a pause. First, the right-hand side is a real-number integral of curvature, yet it must come out an integer, because the left-hand side counts dimensions — so the formula silently encodes deep integrality constraints (this is exactly why the A-roof genus of a spin manifold is an integer, not merely a rational number). Second, only the symbol's homotopy class enters: deform D continuously while keeping it elliptic and the index cannot change. That is the precise sense in which the index is topological, and it is the hook by which K-theory grabs the whole problem.
Why this rung was the runway: K-theory does the proof
Now the earlier guides pay off. The symbol of an elliptic operator is invertible away from the zero section of T*M, so it glues two bundles into a difference class supported on the cotangent bundle — an element of the compactly supported K-theory of T*M. The analytic index is a homomorphism from that K-group to the integers. The topological index is a second homomorphism, built geometrically: embed M in a big sphere, push the symbol class forward by the Thom isomorphism of the normal bundle, then collapse to a point. The theorem is the statement that these two homomorphisms agree.
Why can such a slippery thing even be proved? Because both homomorphisms are forced to be equal once you check they agree on enough generators and respect the same operations — and Bott periodicity is precisely what makes the generators manageable, reducing the whole of K-theory to a periodic, computable engine. The original Atiyah-Singer proof is a K-theory argument almost entirely; the Chern character, a ring isomorphism from K-theory to rational cohomology, is the dictionary that translates the abstract K-theoretic equality into the concrete curvature integral you saw above. Every named tool in this rung — Thom, Bott, the Chern character, classifying spaces — is load-bearing in this one proof.
The Dirac operator and the heat-kernel proof
There is a single operator from which all the classical cases descend: the Dirac operator. On a spin manifold there is a spinor bundle and a first-order operator D whose square is the Laplacian plus a curvature term — Dirac's original square root of the wave operator, transplanted to geometry. The de Rham, signature, and Dolbeault operators are all Dirac operators twisted by an auxiliary bundle. So the deep content of Atiyah-Singer is really the index of the twisted Dirac operator, equal to the integral of the A-roof genus times the Chern character of the twisting bundle; everything else is a corollary.
There is a second, gorgeously analytic proof — the heat-kernel method — that bypasses K-theory entirely. The idea: for any t > 0 the supertrace of the heat operator e^{-t D*D} minus e^{-t D D*} equals the index exactly, independent of t (the McKean-Singer formula), because nonzero eigenvalues cancel in pairs. Let t -> infinity and only the kernel and cokernel survive, giving the index. Let t -> 0 and the heat kernel localizes near the diagonal, where its short-time asymptotics are governed entirely by local curvature. The miraculous 'fantastic cancellation' (Getzler's rescaling makes it transparent) leaves precisely the A-roof and Chern-character densities — the topological side appears as a curvature integral, on the nose.
Be honest about what a guide can and cannot do here. We have stated the theorem, named its two proofs, and shown you the special cases it unifies — but a real proof, K-theoretic or heat-kernel, is a graduate course and a thick book (Atiyah-Bott-Patodi, or Berline-Getzler-Vergne), not a page. What you can carry away is genuine and not a slogan: an analytic count of solutions is a topological integral; the symbol's homotopy class is the only input that survives; and the bridge is woven from exactly the K-theory, Bott periodicity, Thom isomorphism, and Chern character you spent this rung building.
Where the bridge leads
Step back and see what was unified. A differential operator is an analytic object; its index lives in arithmetic; the answer is read off curvature and characteristic classes. That single equation retroactively explained why Gauss-Bonnet, Riemann-Roch, and the signature theorem all had the same flavour, and it became a workhorse far beyond geometry — fixed-point formulas (the Lefschetz and Atiyah-Bott versions), representation theory of Lie groups via equivariant indices, and the integrality theorems that constrain which manifolds can even carry a positive-scalar-curvature metric.
It also opened directly onto physics. The index of the Dirac operator counts chiral fermion zero modes, so anomaly cancellation in gauge theory is an index computation; the same circle of ideas runs through Donaldson and Seiberg-Witten theory on four-manifolds. And it left honest open territory: index theory on noncompact and singular spaces, the analysis on manifolds with boundary, and connections to the still-mysterious four-dimensional smooth world are active research, not closed books. The theorem is a summit of this rung, not the end of the climb.