no. 14 · Friday 13 November 2026
Emulating the Darcy Solution Operator by Heat-Semigroup Functional Calculus and Pointwise Products
Abstract
We approximate the Darcy solution operator $(a,f)\mapsto u$, where $-\operatorname{div}(a\nabla u)=f$, by operator networks built only from heat-semigroup evaluations $e^{-tL_0}$ of a single fixed reference operator and pointwise products. The key obstacle is the variable-coefficient coupling $\nabla a\cdot\nabla u$. We show that it is produced by the heat product defect $[P_t(bu)-P_tb\,P_tu]/(2t)$, which converges quantitatively to $\nabla b\cdot\nabla u$. Inserted into a Laplacian-preconditioned Richardson iteration, this yields an explicit scheme of depth $O(\log(1/\varepsilon)/\log(1/q))$ whose $\dot H^1$ error is at most $\varepsilon$, uniformly over the coefficient class and $L^2$ data, with all parameters explicit. Each block is controlled in continuum operator norm, without Galerkin truncation. On the torus, band-limited inputs admit an exact pseudo-spectral realization on a grid of $O(1/(\varepsilon a_{\min}))$ points per direction, up to logarithmic and class-dependent factors. The argument extends by carre du champ calculus to strongly local Dirichlet spaces with a spectral gap and a Bakry–Emery bound $\mathrm{BE}(-K,\infty)$, with dimension-independent constants after spectral-gap normalization. On the torus we also obtain a $C^2$-type coefficient class and a refinement to Wiener classes of regularity $1+θ$, with product-defect rate $t^{θ/2}$. The factor $1/(2t)$ makes the proved scheme ill-conditioned at small $\varepsilon$.
- comments
- 34 pages, 4 tables
- fields
- Numerical Analysis, Analysis of PDEs