Hierarchically Low-Rank Schur Preconditioners

• Hierarchically Low-Rank Schur Preconditioners are algebraic techniques that approximate Schur complements using hierarchical low-rank formats.
• They efficiently solve large-scale linear systems from PDE, BIE, and graph problems through recursive block partitioning and structured low-rank compression.
• They offer near-linear computational complexity and robust spectral properties while supporting scalable, parallel implementations.

A hierarchically low-rank Schur preconditioner is an algebraic preconditioning framework for large-scale linear systems, particularly those with kernel or sparse structure, in which the system is recursively partitioned so that at each level the Schur complement associated with a separator/interface is approximated by a matrix in a hierarchical low-rank format. Fundamental to their performance is the observation that for many PDE, BIE, and graph-based problems, submatrices or Schur complements created during Gaussian elimination or nested dissection admit efficient low-rank representations. This property enables the construction of scalable and robust preconditioners with rigorous spectral bounds, near-linear complexity, and application to both direct and iterative solution schemes.

1. Block Schur-Complement Decomposition and Hierarchical Recursion

The central algebraic operation in this class of preconditioners is the recursive block-partitioning of a matrix $A$ into "interior" and "interface" or "separator" variables. Given $A\in\mathbb{R}^{n\times n}$ permuted into block form,

$$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$

elimination of the interior block yields the Schur complement $S = A_{22} - A_{21}A_{11}^{-1}A_{12}$. For spatially sparse, PDE-derived, or integral-equation matrices, repeated application of this procedure—via nested dissection or a multilevel domain decomposition—yields a block-tree hierarchy, with Schur complements defined on ever-smaller separators (Li et al., 2015, Chávez et al., 2016, Pouransari et al., 2015).

At each level, the block-tridiagonal or two-by-two structure persists (e.g., for block-cyclic reduction or Stochastic Galerkin systems (Sousedík et al., 2012, Chávez et al., 2016)), and the preconditioner can be assembled recursively: the Schur complement at a given level is approximated in a hierarchical low-rank format, and the process recurses down to either small enough blocks (solved exactly) or a fixed depth $L$. This approach generalizes to multilevel domain decomposition, algebraic graph partitioning, and arbitrary separator hierarchies (Xu et al., 2022).

2. Hierarchical Low-Rank Approximation of Schur Complements

The key assumption for efficiency is that, after suitable row/column clustering, the off-diagonal blocks in the Schur complements are numerically low-rank. Applicable representations include:

• $\mathcal{H}$-matrix and $\mathcal{H}^2$-matrix formats: block cluster trees with admissibility criteria (weak or strong) are constructed so off-diagonal blocks admit approximations $A_{ij}\approx U_{ij}V_{ij}^T$ of fixed rank (Börm et al., 2014, Chávez et al., 2017, Chávez et al., 2016).
• HODLR/HSS formats: two-by-two or multilevel binary clustering with rank constraints on off-diagonal blocks and their children (Gatto et al., 2015, Chen et al., 2022).
• Randomized and Nyström methods: low-rank factors of the Schur complement constructed via randomized matrix-vector sampling (Daas et al., 2021).
• Extended sparsification: well-separated fill-in blocks are compressed and represented hierarchically via auxiliary variables, enforcing sparsity at fine levels (Pouransari et al., 2015).
• Accumulated update techniques: reductions in the number of low-rank updates by deferring and aggregating block updates for improved setup efficiency (Börm, 2017).

Low-rank approximation steps typically use truncated SVD, ACA, RRQR, or randomized algorithms, with truncation prescribed by a blockwise relative tolerance $\epsilon$.

3. Preconditioning Strategies and Algorithmic Structure

The Schur complement preconditioner at each level is typically obtained in one of three forms:

• Direct inversion of a low-rank compressed Schur complement, yielding $M^{-1} \approx S^{-1}$, either in H-matrix/HSS/HODLR format (Gatto et al., 2015, Chen et al., 2022, Börm et al., 2014).
• Low-rank correction to an approximate inverse: $A\in\mathbb{R}^{n\times n}$0 or, equivalently, a Woodbury-type update to the (block) interface matrix (Li et al., 2015, Xu et al., 2022, Daas et al., 2021).
• Reverse-Schur or extended system approach: embedding the problem into a higher-dimensional sparse system in which $A\in\mathbb{R}^{n\times n}$1 itself appears as a Schur complement; an approximate solver for the extended system is used within iterative outer iterations for the original system (Sushnikova et al., 2014).

Practical algorithmic steps include:

• Recursive block partitioning and cluster tree construction;
• Assembling low-rank block representations using suitable tolerances;
• Hierarchical Cholesky or LDL factorizations with low-rank updates at each Schur step;
• Preconditioner application as either two-level or multilevel V-cycle solves, involving direct, sparse, or Krylov-based inner solvers on the bulk or mean-value blocks, and low-rank corrections on the interface (Sousedík et al., 2012, Klockiewicz et al., 2020).

Parallel and distributed implementations partition blocks across processes and require (in the case of Arnoldi or Lanczos for low-rank basis computation) global reductions for orthogonalization and low-rank projections (Xu et al., 2022).

4. Complexity, Spectral Properties, and Error Analysis

Hierarchically low-rank Schur preconditioners achieve favorable computational complexity and conditioning under precise spectral and algebraic assumptions:

• Setup: For $A\in\mathbb{R}^{n\times n}$2 and HODLR approximations, setup is $A\in\mathbb{R}^{n\times n}$3 or $A\in\mathbb{R}^{n\times n}$4, where $A\in\mathbb{R}^{n\times n}$5 is the maximal block rank and $A\in\mathbb{R}^{n\times n}$6 is the order of the matrix (Börm et al., 2014, Chen et al., 2022).
• Application: Each preconditioner apply (solve) is $A\in\mathbb{R}^{n\times n}$7 or $A\in\mathbb{R}^{n\times n}$8, with secondary dependences on depth or number of blocks (Gatto et al., 2015, Börm, 2017).
• Spectral clustering: With truncation tolerance $A\in\mathbb{R}^{n\times n}$9, the spectrum of the preconditioned operator $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$0 is clustered in $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$1 or, for second-order corrections, in $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$2 (Börm et al., 2014, Klockiewicz et al., 2020, Li et al., 2015, Xu et al., 2022).
• Condition number bounds: For recursively constructed preconditioners, $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$3, with each $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$4 a local Schur complement equivalence constant (Sousedík et al., 2012).
• Extension to indefinite problems: Clustered spectra persist empirically for a wide range of Helmholtz, advection-diffusion, and elasticity systems, provided Schur complements possess rapidly decaying eigenvalues (Pouransari et al., 2015, Xu et al., 2022).

5. Implementation, Parameter Tuning, and Software

The construction and application of these preconditioners require careful selection of several algorithmic parameters:

Parameter	Role	Recommended Range/Strategy
Cluster depth ($$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$5)	Hierarchy depth	$$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$6 for binary splits
Block rank ($$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$7)	Max. rank in low-rank blocks	Enforce $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$8 so truncation error $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{21} & A_{22} \end{pmatrix},$$9
Tolerance ($S = A_{22} - A_{21}A_{11}^{-1}A_{12}$0)	Relative error in block truncation	$S = A_{22} - A_{21}A_{11}^{-1}A_{12}$1 for balance of accuracy/cost
Leaf size ($S = A_{22} - A_{21}A_{11}^{-1}A_{12}$2)	Minimum dense block	$S = A_{22} - A_{21}A_{11}^{-1}A_{12}$3 unknowns per cluster
Krylov iterations	# steps in inner solves	3-5 for reverse-Schur in BIE; 2–10 for block CG

Parallel implementations, such as parGeMSLR, exploit domain decomposition and multilevel graph separators, and support high concurrency on distributed-memory and GPU-enabled architectures (Xu et al., 2022, Chen et al., 2022). Setup cost and preconditioner robustness can be controlled by tuning low-rank thresholds ($S = A_{22} - A_{21}A_{11}^{-1}A_{12}$4), separator sizes, and stopping criteria for iterative inner solvers.

6. Applications and Numerical Results

Hierarchically low-rank Schur preconditioners have been successfully applied to:

• Sparse linear systems from finite difference/finite element PDE discretizations, including unstructured and high-contrast media (Pouransari et al., 2015, Li et al., 2015, Klockiewicz et al., 2020).
• Dense matrices from boundary integral equations, with the underlying operator approximated via $S = A_{22} - A_{21}A_{11}^{-1}A_{12}$5, HODLR, or H-matrix schemes (Sushnikova et al., 2014, Chen et al., 2022).
• Stochastic Galerkin systems with recursive block structure induced by polynomial chaos expansions (Sousedík et al., 2012).
• Large-scale domain decomposition and parallel-in-time solvers (Xu et al., 2022, Chávez et al., 2017).
• Kernel machine learning methods (large kernel matrices are amenable to HODLR and H-matrix compression) (Chen et al., 2022).

Benchmark studies demonstrate iteration counts with GMRES/CG nearly independent of mesh size or PDE coefficients (provided block ranks are properly controlled), memory footprint scaling as $S = A_{22} - A_{21}A_{11}^{-1}A_{12}$6, and total solution times several times faster than standard ILU/AMG preconditioners in high-dimensional or indefinite problems. In particular, for 3D elliptic PDEs, preconditioned Krylov methods have convergence rates essentially independent of spatial discretization $S = A_{22} - A_{21}A_{11}^{-1}A_{12}$7, polynomial order $S = A_{22} - A_{21}A_{11}^{-1}A_{12}$8, or stochastic dimension, provided the off-diagonal ranks are bounded (Sousedík et al., 2012, Pouransari et al., 2015, Xu et al., 2022).

7. Extensions, Current Directions, and Performance Enhancements

Key recent advancements include:

• Second-order accurate hierarchical sparsification, which reduces approximation error from $S = A_{22} - A_{21}A_{11}^{-1}A_{12}$9 to $L$0 and halves CG iterations without increasing asymptotic cost (Klockiewicz et al., 2020).
• Reverse-Schur preconditioning via extended sparse forms for $L$1 matrices, providing memory-efficient preconditioners with near-linear cost in large BIE systems (Sushnikova et al., 2014).
• Randomized low-rank (Nyström) approximations for Schur complements, enabling efficient algebraic two-level or multilevel preconditioning with explicit spectral bounds (Daas et al., 2021).
• Accumulated updates in H-matrix arithmetic, which reduce the number of expensive rank-revealing factorizations in H-LU/LDL^T steps, significantly cutting setup times while preserving preconditioner quality (Börm, 2017).
• Hybrid algebraic-geometric partitioning and GPU acceleration for very large, distributed, or accelerated environments (Xu et al., 2022, Chen et al., 2022).

These enhancements further increase the scalability, robustness, and efficiency of hierarchically low-rank Schur preconditioners, enabling their use in modern PDE/BIE solvers and data science applications at scale.