Paperlane

no. 14 · Friday 13 November 2026

Iterative Refinement of A Modified Lavrentiev Regularization Method for De-convolution of the Discrete Helmholtz Type Differential Filter

Nathaniel Mays and Ming Zhong

math.NA · v2

Abstract

We analyze an iterated, modified Lavrentiev regularization method (Mitlar) for deconvolving the Helmholtz-type differential filter $G=(I-δ^2Δ)^{-1}$ used in large eddy simulation. The method replaces the Lavrentiev operator $G+αI$ by the convex combination $(1-α)G+αI$; each step costs one Helmholtz solve, exactly as in iterated Lavrentiev regularization. We show that this one-line change has a precise structural meaning: the regularized operator is again a Helmholtz filter, now with the smaller radius $\sqrtα\,δ$, and $J$ refinement steps coincide with van Cittert approximate deconvolution of order $J$ applied to that filter. Consequently the regularization error is governed by the product $αδ^2$ rather than by $α$ alone. We prove sharp error bounds of order $(αδ^2)^{s}$ on a Hilbert scale, $0\le s\le J+1$, with saturation exactly at $s=J+1$; show that for noise-free data the Mitlar error never exceeds the iterated Lavrentiev error, for every datum, every $α$ and every $J$; give a priori parameter choices for noisy data with explicit $δ$-dependence; and prove finite element error bounds whose constants grow only linearly in $J$. Finally we show that the iterates decrease a natural energy functional for every $α>0$, and derive a computable stopping rule that guarantees monotone decrease of a weighted error and always terminates. Numerical examples illustrate the theory and compare Mitlar with a family of Tikhonov–Lavrentiev methods.

fields
Numerical Analysis
pdf ↗ page arXiv:1801.08913 ↗