|国家预印本平台
首页|On the choice of optimization norm for Anderson acceleration of the Picard iteration for Navier-Stokes equations

On the choice of optimization norm for Anderson acceleration of the Picard iteration for Navier-Stokes equations

On the choice of optimization norm for Anderson acceleration of the Picard iteration for Navier-Stokes equations

来源:Arxiv_logoArxiv
英文摘要

Recently developed convergence theory for Anderson acceleration (AA) assumes that the AA optimization norm matches the norm of the Hilbert space that the fixed point function is defined on. While this seems a natural assumption, it may not be the optimal choice in terms of convergence of the iteration or computational efficiency. For the Picard iteration for the Navier-Stokes equations (NSE), the associated Hilbert space norm is $H^1_0(\Omega),$ which is inefficient to implement in a large scale HPC setting since it requires multiplication of global coefficient vectors by the stiffness matrix. Motivated by recent numerical tests that show using the $\ell^2$ norm produces similar convergence behavior as $H^1_0$ does, we revisit the convergence theory of [Pollock et al, {\it SINUM} 2019] and find that i) it can be improved with a sharper treatment of the nonlinear terms; and ii) in the case that the AA optimization norm is changed to $L^2$ (and by extension $\ell^2$ or $L^2$ using a diagonally lumped mass matrix), a new convergence theory is developed that provides an essentially equivalent estimate as the $H^1_0$ case. Several numerical tests illustrate the new theory, and the theory and tests reveal that one can interchangeably use the norms $H^1_0$, $L^2$, $\ell^2$ or $L^2$ with diagonally lumped mass matrix for the AA optimization problem without significantly affecting the overall convergence behavior. Thus, one is justified to use $\ell^2$ or diagonally lumped $L^2$ for the AA optimization norm in Anderson accelerated Picard iterations for large scale NSE problems.

Elizabeth Hawkins、Leo Rebholz

数学力学

Elizabeth Hawkins,Leo Rebholz.On the choice of optimization norm for Anderson acceleration of the Picard iteration for Navier-Stokes equations[EB/OL].(2025-05-12)[2025-07-17].https://arxiv.org/abs/2505.07650.点此复制

评论