DFPI, A unified framework for deflated linear solvers: bridging the gap between Krylov subspace methods and Fixed-Point Iterations

TL;DR

This paper introduces DFPI, a unified theoretical framework that combines various linear solver techniques, bridging the gap between Krylov methods and fixed-point iterations for improved convergence in PDE-based simulations.

Contribution

The paper presents a novel framework, DFPI, that unifies existing acceleration methods and provides new insights into the design of iterative solvers for linear systems.

Findings

DFPI unifies several acceleration techniques like RPM, BoostConv, and Anderson acceleration.

Convergence depends on the invariance of the projection space, with a minimization principle needed otherwise.

Numerical tests show DFPI's effectiveness on CFD problems.

Abstract

Iterative algorithms are instrumental in modern numerical simulation for solving systems arising from the discretization of PDEs. They face however significant challenges in industrial applications, such as slow convergence, limit cycle oscillations, or iterations blow-up. An ideal preconditioner is rarely available and naive approaches such as Richardson iterations often fail to converge on complex cases, calling for generic sophistications such as deflation techniques and/or Krylov subspaces approaches. However the quest for an optimal general linear solver is still open and a matter of active research. This paper introduces a new theoretical framework, called DFPI (Deflated Fixed Point Iterations) for the iterative solution of linear systems. It unifies several existing acceleration and stabilization techniques such as RPM, BoostConv and Anderson acceleration, and bridges the gap between Richardson iterations and Krylov subspace methods, including GMRES, PCG, BiCGStab and variants. DFPI is structured around two key building blocks : the choice of a projection operator, on the one hand and the trouble vectors recruitment strategy, on the other hand. A general convergence result will be presented, showing the choice of a specific projection operator has minimal impact as long as the projection space remains invariant by the iteration matrix. However when this is not guaranteed, a minimization principle becomes a must have. Finally numerical comparisons will be conducted on a variety of relevant CFD cases.