Schur–Low-Rank (SLR) Preconditioners

• SLR preconditioners are techniques that combine domain decomposition with low-rank spectral corrections to approximate the inverse of large sparse systems.
• They employ a tractable block-diagonal Schur complement enhanced by low-rank updates via methods like Lanczos and the Sherman–Morrison–Woodbury formula.
• SLR methods support hierarchical, parallel, and GPU implementations, making them effective for both SPD and indefinite problems.

A Schur–Low-Rank (SLR) preconditioner combines domain decomposition and low-rank spectral corrections to produce scalable, robust preconditioners for large sparse linear systems. By leveraging the structure and spectrum of the Schur complement, SLR preconditioners efficiently approximate the inverse action on interface or coarse-level variables by supplementing a tractable solver (e.g., block-diagonal or local direct/ILU factorization) with a low-rank correction that targets the nonlocal or global coupling. SLR methods are applicable to both SPD and indefinite systems, support hierarchical and parallel construction, and are compatible with robust iterative solvers. Their development spans algebraic, hierarchical, and randomized techniques.

1. Mathematical and Algorithmic Foundations

SLR preconditioners arise from the observation that the Schur complement in a block-partitioned system contains strong global coupling but is often well-approximated by a simple local operator plus a correction with rapidly decaying spectrum. Let $A\in\mathbb{R}^{n\times n}$ be partitioned as

$$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$

where $I$ denotes interior variables and $B$ the interface. The Schur complement is

$S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$

The key SLR ansatz is to approximate $S^{-1}$ as

$S^{-1} \approx C^{-1} + U \Lambda U^T,$

where $C$ is a block-diagonal or otherwise tractable approximate Schur complement, and $U\Lambda U^T$ (or related structured terms) is a low-rank correction derived from leading eigenmodes of $S^{-1} - C^{-1}$ or underlying coupling matrices. The approximation rank $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$0 controls clustering of the spectrum and is chosen based on the decay of relevant eigenvalues or error criteria (Li et al., 2015).

The preconditioner is extended to the original system: $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$1 with $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$2 low-rank update, applied in block-triangular solves as in domain-decomposition preconditioning schemes.

Hierarchical and recursive variants (e.g., $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$3 and HSS-based) generalize this decomposition through nested partitioning and repeated Schur complements, enabling multilevel SLR constructions (Börm et al., 2014, Gatto et al., 2015).

2. Construction and Low-Rank Correction Strategies

Low-rank components are constructed via analytic, algebraic, or randomized algorithms, depending on the method and available structure:

• Eigenvalue Decomposition / Lanczos: The spectral properties of the correction $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$4 are explored using the Lanczos process to extract leading eigenpairs, revealing rapid decay and permitting efficient spectral truncation. In practice, matrix-vector products required for the correction employ subdomain factorizations and interface matvecs (Li et al., 2015, Li et al., 2015).
• Sherman–Morrison–Woodbury Formula: The low-rank update is frequently applied via the SMW formula, yielding efficient numerical routines where only a small dense system of size $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$5 is solved per preconditioner application (Li et al., 2015, Zheng et al., 2020).
• Hierarchically Semi-Separable (HSS) and $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$6 Matrices: For PDE applications, the Schur complements at each level of nested dissection can be compressed into HSS or $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$7 forms, allowing efficient rank-truncated updates and inversion (Börm et al., 2014, Gatto et al., 2015).
• Randomized Sketching and Nyström's Method: When explicit Schur formation is infeasible, randomized projections (e.g., Gaussian sketches) on $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$8 are used to extract range information, and the Nyström approximation yields a low-rank surrogate with spectral control (Daas et al., 2021).
• Neumann/Power Series and Correction: In some non-SPD and parallel regimes, a truncated Neumann series approximates $$A = \begin{pmatrix} A_{II} & A_{IB} \ A_{BI} & A_{BB} \end{pmatrix},$$9, while a low-rank SMW update targets the (potentially slowly decaying) series tail (Zheng et al., 2020).

3. Multilevel, Parallel, and Hierarchical Extensions

SLR preconditioners are designed for high parallel efficiency and scalability. ParGeMSLR implements a multilevel recursive SLR framework, partitioning the symmetrized adjacency graph at each level using $I$0-way vertex separators. Each level applies domain decomposition and SLR correction, resulting in purely local factorizations (block-diagonal ILUT or similar) and distributed low-rank Arnoldi procedures (Xu et al., 2022). Strong and weak scaling is ensured by balancing subdomains and distributing low-rank and interface work across MPI ranks. GPU acceleration is supported for both local solves and small dense low-rank operations.

Hierarchical SLR approaches, such as those based on HSS or $I$1 matrices, recursively build compressed Schur complements at each node of a partition tree. Local low-rank updates are recompressed at each recursive application, with storage and arithmetic complexity $I$2 and $I$3 for $I$4 degrees of freedom and correction rank $I$5 (Börm et al., 2014, Gatto et al., 2015).

4. Spectral Analysis and Quality Guarantees

The efficacy of SLR preconditioners is traced to the spectral properties of the Schur complement and the rapid decay in the eigenvalues of the correction operator. For symmetric systems, the eigenvalues of the preconditioned operator $I$6 are tightly clustered around 1 when the correction rank $I$7 is chosen so that the $I$8 eigenvalue of the relevant operator falls below a specified threshold (Li et al., 2015, Li et al., 2015).

Analytical results include:

• If every Schur update in a hierarchical factorization is approximated to relative accuracy $I$9, the preconditioner $B$0 satisfies $B$1, where $B$2 depends on the tree depth, and eigenvalues of $B$3 are clustered in $B$4 (Börm et al., 2014).
• For truncated spectral corrections, the largest preconditioned eigenvalue $B$5 is $B$6, and full uniform clustering up to the $B$7 mode is obtained (Li et al., 2015).
• In randomized Nyström-based approaches, the expected condition number is controlled by the $B$8 eigenvalue of $B$9, and oversampling offers additional robustness (Daas et al., 2021).

In application to KKT systems in interior-point methods, SLR-style low-rank updates of the Schur complement ensure that the preconditioned spectrum is contained in a small interval around 1, with explicit eigenvalue bounds and effective iteration count control (Bellavia et al., 2013).

5. Implementation Aspects and Practical Recommendations

The construction and application of SLR preconditioners have several key implementation considerations:

• Domain Partitioning: Effective DD-based SLR relies on partitioning such that interface size is minimized and subdomain balance is preserved. Edge-based or $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$0-way separators are standard (Li et al., 2015, Xu et al., 2022).
• Solver Selection: Block-diagonal solves (subdomain ILU/LDL) are performed in parallel. The solver on the interface (or the approximate Schur block) may itself be recursively preconditioned using SLR if large (Li et al., 2015, Xu et al., 2022).
• Eigenvalue Computations: Partial reorthogonalization in the Lanczos process provides high accuracy for moderate $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$1 (typically $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$2). Enriching the low-rank space requires only additional Lanczos iterations, and previous directions can be deflated (Li et al., 2015).
• Low-Rank Update Application: The SMW formula is used, typically requiring only small $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$3 dense solves and a small number of additional vector-matrix products per Krylov iteration (Li et al., 2015, Xu et al., 2022).
• Hierarchical and Multilevel Setup: In $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$4/HSS-based SLR, local updates and recompression steps scale as $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$5 where $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$6 and $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$7 are cluster sizes, and application costs are $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$8 (Börm et al., 2014).
• Parallel and GPU Capability: All local solves are independent, and most interface and low-rank operations can be parallelized over distributed ranks or nodes, including offloading dense and triangular solves to GPUs (Xu et al., 2022).

6. Performance, Robustness, and Applications

SLR preconditioners have been extensively validated on problems from finite element discretizations of elliptic, hyperbolic, and indefinite PDEs (Poisson, reaction–diffusion, Helmholtz), as well as on general large sparse matrices (e.g., UF collection, interior-point KKT systems). Key empirical findings include:

• Iteration counts with SLR-CG or SLR-GMRES scale favorably with problem size and are often independent of $S = A_{BB} - A_{BI} A_{II}^{-1} A_{IB}.$9 (mesh size), $S^{-1}$0 (polynomial order), or even spectral shift in indefinite problems (Li et al., 2015, Gatto et al., 2015, Xu et al., 2022).
• Build/setup cost is higher than pure ILU, but amortized rapidly for multiple right-hand sides (Li et al., 2015, Börm et al., 2014).
• SLR is robust where ILU, RAS, or multigrid preconditioners fail or stagnate, especially for indefinite or highly heterogeneous systems (Li et al., 2015, Xu et al., 2022).
• In hierarchical low-rank Schur frameworks, the number of iterations remains constant under mesh refinement, and off-diagonal HSS rank grows slowly with system size or complexity (Gatto et al., 2015).
• At scale, parGeMSLR demonstrates strong and weak scalability to thousands of MPI ranks, and GPU acceleration yields 2–5× speedup for the low-rank dominated solve phase (Xu et al., 2022).
• In interior-point optimization contexts, SLR-based preconditioners reduce overall solve times by 20–80% compared to periodic full refactorization, with minimal impact on Krylov iteration counts (Bellavia et al., 2013).

7. Comparison to Classical and Alternative Preconditioners

SLR preconditioners generalize and often outperform several established classes:

Preconditioner	Interface Solves	Parallelism	Robustness to Indefinite/Shifted Problems	SLR Feature
ILU/ICT	Triangular	Limited (sequential/serial)	Poor for indefinite/complex shift	SLR avoids triangular
Block-Jacobi/Additive Schwarz	Block-diagonal	High	No global coupling preconditioning	SLR adds low-rank global
Classical Schur	Global dense	Poor	High (if factorizable), but cost-prohibitive	SLR approximates Schur
AMG	Multilevel	High	Optimal for SPD, struggles with indefinite	SLR robust for all

SLR combines the high parallelism of DD and the spectral clustering typical of direct methods, with low-rank updates bridging the gap between local and global structure (Li et al., 2015, Li et al., 2015, Xu et al., 2022).

---

References:

(Börm et al., 2014, Li et al., 2015, Li et al., 2015, Gatto et al., 2015, Zheng et al., 2020, Daas et al., 2021, Bellavia et al., 2013, Xu et al., 2022)