Where we are: the pieces are finally assembled
Across this rung you built a complete machine. Guide 1 gave singular homology H_k(X); guides 2 and 3 made it computable with exact sequences, Mayer-Vietoris and cellular homology, reading off Betti numbers and the Euler characteristic; guide 4 dualized everything into cohomology H^k(X), explained the universal coefficient theorem, and introduced the cup product that turns the cohomology of a space into a ring. We now have homology, cohomology, and a multiplication. This final guide is about the deep symmetry that ties them together the moment the space is a manifold — and the two great theorems that symmetry unlocks.
Here is the headline. For a general topological space, H_k and H^k are merely dual-ish, related only by the universal coefficient theorem and otherwise independent. But on a nice manifold a far stronger thing is true: the two are literally the same data read in mirror image. Poincaré duality is that mirror. It is not a computational trick; it is a statement that a manifold is a fundamentally self-dual object, and almost every special fact about manifolds — the symmetry of Betti numbers, the existence of an intersection pairing, the vanishing of odd-dimensional Euler characteristics — is a shadow of it.
Poincaré duality: state the hypotheses, not the slogan
The clean statement. Let M be a closed (compact, without boundary) oriented manifold of dimension n. Then for every k there is an isomorphism H_k(M) is isomorphic to H^{n-k}(M). Degree-k homology equals degree-(n-k) cohomology; the dimension n acts as a hinge that reflects one into the other. With field coefficients this immediately gives the symmetry of Betti numbers b_k(M) = b_{n-k}(M) — the kth and (n-k)th holes always come in equal numbers on a closed oriented manifold.
What is the actual map? This is where guide 4's last construction earns its keep. The orientation singles out a fundamental class [M] in H_n(M) — the n-cycle that is the whole manifold, oriented. Poincaré duality is then capping against it: the isomorphism H^k(M) -> H_{n-k}(M) sends a cohomology class alpha to the cap product alpha cap [M]. That cap product is the pairing between cochains and chains you met at the end of guide 4; here it is upgraded into the very engine of the duality. So 'a manifold is self-dual' unpacks concretely to 'capping with the fundamental class is an isomorphism', and the whole theorem rests on M having one well-defined top class, which is exactly what orientability provides.
The intersection form: geometry hidden inside the cup product
Now combine duality with the cup product and something geometric falls out. On a closed oriented M of dimension n, define a pairing on cohomology: take alpha in H^k and beta in H^{n-k}, multiply them with the cup product to land in the top group H^n(M) which is isomorphic to ℤ, and read off the integer by evaluating against the fundamental class. This pairing alpha cup beta evaluated on [M] is the intersection form, and Poincaré duality is exactly the statement that it is a perfect pairing — non-degenerate, so each class is determined by how it pairs with everything else.
Why call it 'intersection'? Because the cup product is the algebraic shadow of a transverse geometric meeting. If a class is dual to a submanifold A and another to a submanifold B, their cup product is dual to the intersection A cap B, and the integer the form spits out is the signed count of points where A and B cross — orientations giving each crossing a plus or minus. On a surface of genus g, the standard loops a_i and b_i pair to 1 (they cross once) and a_i with a_j pair to 0 (they can be slid apart); the form is the standard symplectic matrix [0, I; -I, 0]. You are literally counting how curves hit each other, and the cup product was secretly doing geometry all along.
alpha cup beta : H^k(M) x H^{n-k}(M) -> H^n(M) = Z, (alpha, beta) |-> < alpha cup beta, [M] >
middle dimension n = 2m: Q(alpha, beta) = < alpha cup beta, [M] > on H^m(M)
n = 4m: Q is symmetric -> signature(M) = (#positive) - (#negative) eigenvalues
n = 4m+2: Q is skew-symmetric -> standard symplectic form [0, I; -I, 0]The middle dimension is where this becomes spectacular. When n = 4m the form on H^{2m}(M) is symmetric, so it has a signature — the number of positive minus negative eigenvalues. For a closed oriented 4-manifold this signature, together with the intersection form's full congruence class over ℤ, is an astonishingly powerful invariant: by deep theorems (Whitehead, Milnor, and later Freedman) the intersection form essentially classifies simply connected closed topological 4-manifolds. But be honest about the boundary: the smooth classification is wide open — distinct smooth manifolds can share an intersection form, the smooth 4-dimensional Poincaré conjecture is unsolved, and exotic smooth structures on R^4 exist. The algebra is clean; the smooth geometry in dimension four is one of the hardest open territories in mathematics.
The Lefschetz fixed-point theorem: counting fixed points with traces
The second great payoff turns homology into a fixed-point counter. Take a continuous self-map f: X -> X of a nice compact space. It induces linear maps f_* on each homology group H_k(X; ℚ). Define the Lefschetz number L(f) as the alternating sum of the traces of these maps: trace on H_0, minus trace on H_1, plus trace on H_2, and so on. The Lefschetz fixed-point theorem then says: if L(f) is not zero, then f has a fixed point. A single rational number, computed from the action on homology, forces the existence of a point with f(p) = p.
Why is this believable, and where does duality enter? Geometrically, fixed points of f are exactly the intersections of the graph of f with the diagonal inside X x X. Poincaré duality and the cup product convert that geometric intersection number into the algebraic alternating trace — the cap product and the Künneth description of H_*(X x X) do the bookkeeping. So the Lefschetz number is, at heart, an intersection number computed by the very duality of the previous section. The theorem is duality wearing a different hat: count where the graph meets the diagonal, but do it with traces instead of pictures.
- Special case: if f is homotopic to the identity, then f_* is the identity on every H_k, so each trace is just the dimension b_k. The alternating sum of dimensions is the Euler characteristic: L(identity) = chi(X). So Lefschetz contains, as its simplest case, the statement that a self-map homotopic to the identity has a fixed point whenever chi(X) is nonzero.
- Apply this to the sphere S^2, where chi = 2, nonzero. Then any map homotopic to the identity — in particular the time-one flow of any continuous vector field — has a fixed point. This is the Poincaré-Hopf / 'hairy ball' theorem: you cannot comb a hairy sphere flat, because a nowhere-zero tangent field would generate a fixed-point-free flow, contradicting chi(S^2) = 2.
- Now a case where the count is the whole point: the torus T^2 has chi = 0, so the identity has Lefschetz number zero — and indeed a rotation of the torus is homotopic to the identity yet has no fixed point. The theorem is honest: L(f) = 0 gives no conclusion. But take a hyperbolic toral automorphism, the map induced by a matrix like [2, 1; 1, 1]; its trace on H_1 is the matrix trace 3, so L(f) = 1 - 3 + 1 = -1, nonzero, and the map must have a fixed point (the origin), even though it has no homotopy to the identity.
Stepping back: what this rung was really about
Look at what duality did. Homology counts holes; cohomology, with its cup product, counts holes and multiplies them; and on a manifold Poincaré duality says these are the same information, related by capping with the fundamental class. From that one symmetry came the intersection form (geometry, signatures, the four-manifold drama) and the Lefschetz theorem (fixed points, the hairy ball, the Euler characteristic as a special case). A surprising number of 'obvious' facts about manifolds are not obvious at all — they are this duality, viewed from different angles.
Be honest about scope and about the roads ahead, in the spirit of the whole tower. First, Poincaré duality has many faces beyond the version here — there is a hard refinement, the Hard Lefschetz theorem, that holds on Kähler manifolds and constrains Betti numbers far more tightly; and in algebraic geometry the analogue is Serre duality, a sibling you will meet on the algebraic-geometry track, not a competitor. These are different incarnations of one idea, each with its own hypotheses (Kähler, projective, coherent), and the slogan-free habit you practised on excision and Bonnet-Myers applies here too: never quote the duality without its conditions.
Second, resist the urge to oversell. These theorems are computational and structural triumphs, but they do not by themselves classify manifolds — the smooth 4-dimensional Poincaré conjecture remains open, exotic spheres exist in high dimensions, and homology and cohomology together are still coarser than the full homotopy type. What you genuinely own now is a coherent algebraic-topology toolkit: chains and cycles, exact sequences, cellular computation, cohomology rings, degree theory, duality, and fixed-point counting. That toolkit is the working language of the differential-topology, Riemannian, complex-geometry, and algebraic-geometry towers you can now climb — homology was never the destination, it is the alphabet the rest of geometry is written in.