[論文レビュー] A BFBt preconditioner for Double Saddle-Point Systems

Chen Greif|arXiv (Cornell University)|Jan 26, 2026

Matrix Theory and Algorithms被引用数 0

ひとこと要約

この論文は Elman の BFBt 前処置を二重サドルポイント系へ拡張し、非完全な S2 を用い、得られる固有値分布を分析し、Stokes-Darcy 方程式の MAC離散化で手法を検証します。

ABSTRACT

We consider block preconditioners for double saddle-point systems, and investigate the effect of approximating the nested Schur complement associated with the trailing diagonal block on the eigenvalue distribution of the preconditioned matrix. We develop a variant of Elman's BFBt method and adapt it to this family of linear systems. Our findings are illustrated on a Marker-and-Cell discretization of the Stokes-Darcy equations.

研究の動機と目的

• Motivate and study block preconditioners for double saddle-point systems and how approximating the nested Schur complement S2 affects eigenvalues and convergence.
• Adapt and analyze a BFBt-style preconditioner to the double saddle-point setting.
• Demonstrate the approach with a MAC discretization of Stokes-Darcy equations and study practical approximations of S1 and S2.

提案手法

• Formulate double saddle-point system K with blocks A, B, C, D and defined Schur complements S1 = D + B A^{-1} B^T and S2 = C S1^{-1} C^T.
• Propose inexact S2 preconditioning and analyze eigenvalues of M_LT^{-1} K and M_D^{-1} K under inexact S2.
• Develop a BFBt-inspired approximation for S2^{-1} and derive spectral consequences via generalized eigenvalue problems.
• Express and analyze the preconditioned system using block-LDU and block-diagonal preconditioners with inexact Schur complements.
• Use SVD-based transformations to relate the spectrum of S2^{-1} S2 to reduced matrices and derive eigenvalue structure (including 1, golden ratio ±(1±√5)/2, and cubic roots).
• Apply to Stokes-Darcy MAC discretization to illustrate spectral effects and convergence.

実験結果

リサーチクエスチョン

• RQ1 How does approximating the nested Schur complement S2 affect the eigenvalue distribution of the preconditioned double saddle-point system?
• RQ2 Can a BFBt-style preconditioner be effectively adapted to double saddle-point systems with inexact S2 while maintaining robust convergence?
• RQ3 What is the spectral structure of the preconditioned matrix under inexact S2 in the common cases D ≠ 0 and A symmetric positive definite?
• RQ4 How do practical approximations of S1 and S2 influence iterative solver performance on Stokes-Darcy discretizations?

主な発見

• Eigenvalues of the inexact-M_D preconditioned system exhibit a fixed set of values (including 1, (1±√5)/2, and roots of a cubic depending on μ) as characterized in the analysis.
• When S2 is approximated but A and S1 are solved exactly, a subset of eigenvalues cluster near 1, with remaining eigenvalues governed by generalized eigenvalues μ from S2 z = μ Ŝ2 z.
• For the symmetric case with D = 0 and A SPD, the preconditioned spectrum consists of six distinct eigenvalues plus cubic-root terms, implying potential six-iteration MINRES convergence in the exact-solver idealization.
• Numerical experiments on Stokes-Darcy show eigenvalues of the LT-preconditioned system cluster in the right-half plane near 1, with complex eigenvalues present due to nonsymmetry but with fast convergence in practice.
• Practical Schur-complement approximations (S1̃, S2̃) and inner solvers (multigrid for CC^T) yield robust convergence under mesh refinement, though slow-downs appear at very fine meshes or extreme parameter choices.