Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces operator networks that emulate the solution operator of $-\operatorname{div}(a\nabla u)=f$ using only heat-semigroup functions of a fixed reference operator and pointwise products. An emulation theorem is established for strongly local Dirichlet spaces with a spectral gap and Bakry--Émery bound $\mathrm{BE}(-K,\infty)$, achieving an explicit $\dot H^1$-error $\varepsilon$ uniformly over coefficients and $L^2$ data, with parameters dependent only on coefficient bounds, accuracy, and $K$. On the torus, this is refined to Wiener classes of fractional regularity $A^{1+θ}$, with an explicit rate achieved by the scheme's product-defect block, effectively providing approximation theorems for continuum operators.
A novel operator network architecture, built solely from heat-semigroup functions and pointwise products, precisely emulates elliptic PDE solution operators without Galerkin truncation.
We approximate the solution operator $(a,f)\mapsto u$ of $-\operatorname{div}(a\nabla u)=f$ by operator networks whose only primitives are heat-semigroup functions of a fixed reference operator and pointwise products. The main result is an emulation theorem on strongly local Dirichlet spaces with a spectral gap and a Bakry--Émery bound $\mathrm{BE}(-K,\infty)$: an explicit scheme reaches $\dot H^1$-error $\varepsilon$ uniformly over coefficients with bounded carré du champ and bounded generator image and over $L^2$ data, with parameters depending only on the coefficient bounds, the accuracy and $K$; on the torus this gives a $C^2$-type coefficient class with constants independent of the dimension. The torus version refines this to Wiener classes of fractional regularity $A^{1+θ}$, with an explicit rate attained by the scheme's product-defect block. The scheme unrolls the Laplacian-preconditioned Richardson iteration, controls each inexact step in the operator norm of $\dot H^1$ without Galerkin truncation, and synthesises the coupling $\nabla a\cdot\nabla u$ from the product defect of one heat layer. All parameters are explicit; the results are approximation theorems for continuum operators.