Chapter 04

Face Signals & the Hodge Story

2-cochains, the face Laplacian L₂, Betti numbers as topological invariants, and the complete discrete de Rham complex — all on our running example.

In this chapter
  1. 07Signals on Faces: 2-Cochains
  2. 08The Complete Hodge Diamond

07Signals on Faces: 2-Cochains

We've now put data on vertices (0-cochains, Chapter 1) and edges (1-cochains, Chapter 2). The next level up: 2-cochains, which assign a value to each face. Physically, these represent fluxes, area densities, or — in the thermal context — effective material properties over a tile region.

Our complex has only two faces ($\sigma_1$, $\sigma_2$), so a 2-cochain is just a vector in $\mathbb{R}^2$. But the algebra is the same at any scale.

The Laplacian on Faces: $\mathbf{L}_2$

Since our complex has no rank-3 cells (no tetrahedra), there is no $\mathbf{B}_{2,3}$ matrix and hence no upper Laplacian. The face Laplacian is purely lower:

Hodge Laplacian on faces
$$\mathbf{L}_2 = \mathbf{B}_{1,2}^\top \mathbf{B}_{1,2} + \underbrace{\mathbf{B}_{2,3}\mathbf{B}_{2,3}^\top}_{= \mathbf{0}\text{ (no 3-cells)}} = \mathbf{B}_{1,2}^\top \mathbf{B}_{1,2}$$
Computing L₂ = B₁₂ᵀB₁₂ B₁₂ᵀ (2×5) · B₁₂ (5×2) = L₂ (2×2) σ₁·σ₁: 1+1+1+0+0 = 3 σ₁·σ₂: 0−1+0+0+0 = −1 σ₂·σ₁: 0−1+0+0+0 = −1 σ₂·σ₂: 0+1+0+1+1 = 3 L₂ σ₁ σ₂ σ₁ [ 3 −1 ] σ₂ [ −1 3 ]
Figure 7.1. The 2×2 Hodge Laplacian on faces. Diagonal entries (3) count the number of boundary edges of each face. The off-diagonal entry (−1) reflects the single shared edge $b$ between the two faces — exactly analogous to how the graph Laplacian had −1 for adjacent vertices.

This is beautifully parallel to $\mathbf{L}_0$. The graph Laplacian $\mathbf{L}_0 = \mathbf{D} - \mathbf{A}$ has degree on the diagonal and $-1$ for adjacent vertices. The face Laplacian $\mathbf{L}_2$ has "boundary edge count" on the diagonal and $\pm 1$ for faces sharing a boundary edge. The same structure repeats at every rank.

Eigenstructure of $\mathbf{L}_2$

For a 2×2 matrix, the eigenvalues are immediate from the characteristic polynomial $\lambda^2 - 6\lambda + 8 = (\lambda - 2)(\lambda - 4) = 0$:

Eigendecomposition of L₂
$$\lambda_0 = 2: \quad \mathbf{v}_0 = \frac{1}{\sqrt{2}}\begin{pmatrix} 1 \\ 1 \end{pmatrix} \qquad \lambda_1 = 4: \quad \mathbf{v}_1 = \frac{1}{\sqrt{2}}\begin{pmatrix} 1 \\ -1 \end{pmatrix}$$

No zero eigenvalue — meaning $\beta_2 = 0$: the complex encloses no voids. The low-frequency mode $\mathbf{v}_0 = (1,1)$ assigns the same value to both faces (smooth). The high-frequency mode $\mathbf{v}_1 = (1,-1)$ assigns opposite values (oscillating). This is the "Fourier basis" for face signals on our tiny complex.


08The Complete Hodge Diamond

Pulling together all three Laplacians, we can now state the full spectral picture of our simplicial complex:

RankObjectsCountLaplacianEigenvaluesβ_k
0Vertices4$\mathbf{L}_0 = \mathbf{B}_1\mathbf{B}_1^\top$0, 2, 4, 41
1Edges5$\mathbf{L}_1 = \mathbf{B}_1^\top\mathbf{B}_1 + \mathbf{B}_{1,2}\mathbf{B}_{1,2}^\top$2, 2, 4, 4, 40
2Faces2$\mathbf{L}_2 = \mathbf{B}_{1,2}^\top\mathbf{B}_{1,2}$2, 40
The Betti Numbers — Topological Invariants

The $k$-th Betti number $\beta_k = \dim\ker\mathbf{L}_k$ counts the number of zero eigenvalues of the Hodge Laplacian at rank $k$. It measures a topological feature of the complex:

$\beta_0 = 1$ — one connected component
$\beta_1 = 0$ — no independent loops (every cycle bounds a face)
$\beta_2 = 0$ — no enclosed voids (no "hollow" regions)

These are topological invariants: they don't change if you deform the complex continuously, and they are computable purely from the combinatorial structure (the incidence matrices).

The Euler characteristic check: $\chi = \beta_0 - \beta_1 + \beta_2 = 1 - 0 + 0 = 1$. Independently: $\chi = |V| - |E| + |F| = 4 - 5 + 2 = 1$ ✓. This is the Euler–Poincaré theorem — the alternating sum of cell counts equals the alternating sum of Betti numbers.

The Complete Discrete de Rham Complex

We can now draw the full picture with all the operators we've built:

The Complete Chain for Our Simplicial Complex ℝ⁴ vertex signals ℝ⁵ edge signals ℝ² face signals B₁ᵀ (grad) B₁₂ᵀ (curl) B₁ (div*) B₁₂ (curl*) L₀ = B₁B₁ᵀ λ: 0, 2, 4, 4 β₀ = 1 L₁ = B₁ᵀB₁+B₁₂B₁₂ᵀ λ: 2, 2, 4, 4, 4 β₁ = 0 L₂ = B₁₂ᵀB₁₂ λ: 2, 4 β₂ = 0 B₁₂ᵀB₁ᵀ = 0 (curl ∘ grad = 0)
Figure 8.1. The complete discrete de Rham complex for our simplicial complex with all operators, Laplacians, eigenvalues, and Betti numbers. Forward operators (coboundary) go right: gradient takes vertex potentials to edge differences, curl takes edge flows to face circulations. Backward operators (boundary) go left. Each rank has its own Hodge Laplacian governing diffusion at that scale.

This is the foundation everything else builds on. Chapters 4–7 will generalize this structure: combinatorial complexes relax the simplicial constraints; higher-order message passing replaces the Laplacian with learned operators; and the Hodge decomposition tells us exactly what information each rank can capture. But the core algebraic machinery — incidence matrices, boundary operators, Hodge Laplacians, Betti numbers — is everything you've now seen worked out by hand.