From the flat Laplacian to a curved one
The Laplacian has been hiding in plain sight for three guides. In Guide 2 the minimal surface equation linearized to u_xx + u_yy = 0; in Guide 3 a harmonic map into R was just a function killed by the Laplacian, the critical point of Dirichlet energy. Now we put that operator at center stage and ask the obvious question: what IS the Laplacian on a curved manifold? The flat answer, Delta f = sum_i d^2 f / dx_i^2, is written in Cartesian coordinates that a general manifold does not have. We need a definition that uses only the metric g, and the cleanest route is variational — exactly the move that has organized this whole rung.
Start from the Dirichlet energy of a function, E(f) = (1/2) integral over M of |grad f|^2 dV, where grad f is the metric gradient and dV is the Riemannian volume form. Take the first variation in f exactly as Guide 1 took it in a curve: differentiate, integrate by parts using the divergence theorem on M, and the critical-point condition pops out as div(grad f) = 0. The operator div(grad) IS the Laplace-Beltrami operator, the manifold's intrinsic Laplacian. So a function is harmonic precisely when it is a critical point of Dirichlet energy — the same Euler-Lagrange logic that produced geodesics from energy and minimal surfaces from area, now producing harmonic functions from the Laplacian.
Spelled out in local coordinates, the coordinate-free div(grad) becomes Delta f = (1/sqrt(det g)) d_i ( sqrt(det g) g^{ij} d_j f ) — the metric enters twice, once in the inverse metric g^{ij} contracting the derivatives and once in the volume factor sqrt(det g) that makes the divergence the right one for the volume form dV. Set the metric to the identity and the sqrt(det g) factors are 1, recovering the flat Delta f = sum_i d^2 f / dx_i^2; so the curved Laplacian is the flat one with the metric woven through every slot. The lesson is that there is nothing exotic here: Delta is built mechanically out of g exactly as the gradient and divergence were.
Lifting the Laplacian to forms: the codifferential and the Hodge star
A function is a 0-form, and the exterior derivative d turns a k-form into a (k+1)-form. To build a Laplacian on forms we need a partner that goes the OTHER way, lowering degree by one — an adjoint of d. The metric supplies it. With g you can define an inner product on k-forms (integrate the pointwise inner product against dV), and then the codifferential delta is simply the formal adjoint of d with respect to that inner product: integral of <d alpha, beta> = integral of <alpha, delta beta> on a closed manifold. Where d raises degree k -> k+1, delta lowers it k -> k-1. That is the whole idea; everything else is bookkeeping for how delta is computed.
The concrete way to write delta uses the Hodge star operator, the one new piece of linear algebra this guide needs. On an oriented n-manifold the metric and orientation single out, for each k-form alpha, a complementary (n-k)-form *alpha (read 'star alpha') — the part of the volume form 'left over' after alpha. On R^3 the star is the gadget behind classical vector calculus: it identifies the 1-form dx with the 2-form dy ^ dz, which is exactly why the curl and the cross product feel like they live in the same world. With it, delta is just d conjugated by the star (up to a sign depending on k and n), and that single operator delta unifies the gradient, divergence, and curl adjoints into one clean object on forms.
Now define the Hodge Laplacian on k-forms by Delta = d delta + delta d. The two terms are forced: d alone cannot be a Laplacian because d^2 = 0 makes it nilpotent, and you need a second-order operator. The symmetric combination d delta + delta d is second order, self-adjoint, and nonnegative — exactly the structure you want. On 0-forms (functions) delta = 0, so Delta f = delta d f reduces to the Laplace-Beltrami operator from the last section, with the geometer's sign built in. So the Hodge Laplacian is not a new operator competing with Laplace-Beltrami; it is the SAME operator, finally extended to act on forms of every degree.
Harmonic forms and the Hodge decomposition
A k-form omega is harmonic if Delta omega = 0. On a compact manifold this innocent condition has a striking consequence. Pair Delta omega = 0 against omega and integrate: 0 = integral of <Delta omega, omega> = integral of <d omega, d omega> + integral of <delta omega, delta omega> = ||d omega||^2 + ||delta omega||^2. A sum of two squares is zero only if both vanish, so omega harmonic forces BOTH d omega = 0 AND delta omega = 0 simultaneously. A harmonic form is therefore both closed (d omega = 0) and co-closed (delta omega = 0) — the most balanced, symmetric kind of form there is. This little integration-by-parts argument is the seed of the entire theory.
Hold onto the picture: a harmonic form is the 'most relaxed' representative in its class. Recall from the earlier rung that de Rham cohomology H^k(M) measures closed forms modulo exact ones — a closed form omega and omega + d eta represent the same class for any eta. Among all these representatives of a fixed class, the harmonic one is the unique one of LEAST norm: adding any exact d eta to a harmonic form strictly increases its L^2 norm (the cross term vanishes precisely because the harmonic form is co-closed). So 'harmonic' is to a cohomology class what 'geodesic' is to a homotopy class of loops or 'minimal surface' is to a boundary — the energy-minimizing, most symmetric representative. The whole rung has been telling one story in four costumes.
The Hodge decomposition theorem assembles this into one of the most beautiful statements in geometry. On a compact oriented Riemannian manifold, the space of all k-forms splits as an orthogonal direct sum of three pieces: the exact forms (image of d), the co-exact forms (image of delta), and the harmonic forms (kernel of Delta). Every k-form omega is uniquely written omega = d alpha + delta beta + h with h harmonic, and the three summands are mutually orthogonal in the L^2 inner product. The harmonic part h is the irreducible core that neither d nor delta can produce — the genuinely topological residue once you have stripped away everything exact and everything co-exact.
Hodge decomposition on a COMPACT oriented Riemannian manifold (M, g):
Omega^k(M) = d( Omega^{k-1} ) (+) delta( Omega^{k+1} ) (+) Harmonic^k
exact co-exact ker Delta
the three summands are MUTUALLY ORTHOGONAL in the L^2 inner product
consequence (the Hodge theorem):
Harmonic^k(M, g) is isomorphic to H^k(M; R) (de Rham cohomology)
so dim Harmonic^k = b_k = the k-th Betti number (a TOPOLOGICAL number)
hypotheses you may NOT drop: M compact (no boundary), g a Riemannian metric.Why this is astonishing: one harmonic form per cohomology class
Trace what the decomposition does to a CLOSED form, because that is where cohomology lives. If omega is closed (d omega = 0), feed it into omega = d alpha + delta beta + h and apply d: the co-exact piece delta beta is the only one that can survive and it is forced to be zero, so a closed form is exactly an exact part plus a harmonic part, omega = d alpha + h. Read modulo exact forms — which is precisely passing to a cohomology class — and the d alpha disappears: every de Rham class is represented by its harmonic part h, and by the orthogonality, h is the UNIQUE harmonic form in that class. So each cohomology class contains exactly one harmonic representative, no more and no fewer.
This is the Hodge theorem, and it is genuinely astonishing in two directions at once. Topology -> analysis: the dimension of the space of harmonic k-forms, an analytic kernel dimension of the operator Delta, equals the k-th Betti number b_k, a purely topological count that knows nothing about the metric. Analysis -> topology: although the harmonic forms THEMSELVES depend on the metric g you chose, HOW MANY there are does not — change the metric and each harmonic form deforms, but their number is pinned by topology. The Laplacian, an operator built entirely from a metric, secretly counts holes.
What the harmonic representative buys you, and where it flows next
The payoff of a canonical representative is immediate and concrete. On a flat torus T^n, the harmonic forms are exactly the constant-coefficient forms (dx^I with constant coefficients), and there are 'n choose k' of them in degree k — which is precisely b_k of the torus, recovered by hand. On the round 2-sphere S^2 the only harmonic 0-forms and 2-forms are constants times 1 and the volume form, with no harmonic 1-forms at all, giving b_0 = b_2 = 1 and b_1 = 0 — the sphere has no holes for a 1-form to wrap, and the Laplacian sees that. These tiny computations, lead with them: they carry the theorem far better than the abstract three-way splitting stated in maximum generality.
- Compute the Betti number b_k as the dimension of the kernel of the Hodge Laplacian Delta = d delta + delta d on k-forms — an analytic eigenvalue-zero count standing in for a topological one.
- On a Kähler manifold the single Hodge Laplacian splits into matching holomorphic and anti-holomorphic Laplacians, forcing the Betti numbers to refine into the Hodge numbers and giving the symmetry b_k = sum of h^{p,q} — the bridge from this real theory to complex geometry and Dolbeault cohomology.
- Guide 5 confronts the SAME Hodge Laplacian head-on: the Bochner-Weitzenbock formula rewrites it as a rough Laplacian plus a curvature term, turning a positive curvature hypothesis directly into the vanishing of harmonic forms — and hence of Betti numbers — and then sets the whole eigenvalue spectrum into motion with the geometric heat flows.