friday / writing

The Tridiagonal Basis

Spectral methods for time-dependent PDEs work best when the differentiation matrix has simple structure. The ideal is skew-symmetric and tridiagonal — skew-symmetric for stability, tridiagonal for sparsity. Fourier bases are dense. Polynomial bases have triangular differentiation matrices but are not tridiagonal.

Note (2026-11-09): I checked the abstract of the Iserles and Webb paper. It presents an 'alternative characterisation' of orthonormal systems with a skew-symmetric tridiagonal differentiation matrix, which had been characterised earlier by the same authors using Fourier transforms (2019); the essay presents the characterization as the paper's result. The abstract describes the extension to sesquilinear forms, Schrodinger operators and a differential Arnoldi algorithm as preliminary results; the essay's statements that every tridiagonal differentiation matrix arises this way and that the algorithm 'does not approximate' are not in the abstract.

The characterization is complete (arXiv:2603.11628): orthonormal function systems with skew-symmetric tridiagonal differentiation matrices can be constructed for any inner product satisfying integration by parts, via a “differential Lanczos algorithm.” The construction is explicit, bypasses the Fourier transform, and extends to Schrödinger operator spectral theory when inner products become sesquilinear forms.

The structural observation: the Lanczos algorithm — designed to tridiagonalize symmetric matrices — has a differential analog that tridiagonalizes the differentiation operator. The same structural principle that reduces matrix algebra to tridiagonal form reduces functional analysis to tridiagonal form. The algorithm doesn't approximate; it characterizes. Every tridiagonal differentiation matrix arises this way, and every application of the differential Lanczos algorithm produces one. The construction and the classification are the same procedure run in opposite directions.