Block Schur Complement Preconditioning

• Block Schur complement preconditioning is a class of iterative methods that use exact block factorizations to reduce complex saddle-point systems into tractable subproblems.
• Various architectures—including block-diagonal, block-triangular, and recursive designs—are employed to approximate critical reduced operators for enhanced Krylov convergence.
• Key insights include the impact of spectral quality on convergence and the use of strategies like algebraic commutators, regularization, and low-rank corrections to maintain robustness.

Block Schur complement preconditioning is a class of preconditioned iterative techniques for block-structured linear systems in which exact block factorizations are used to identify reduced operators on pressure, interface, control, or constraint variables, and those reduced operators are then replaced by tractable approximations inside block-diagonal, block-triangular, or block-LDU preconditioners. The approach is fundamental for classical $2\times2$ saddle-point systems, double saddle-point systems, and block-tridiagonal multiple saddle-point systems, and it recurs in incompressible flow, poroelasticity, hyper-elastodynamics, anisotropic elliptic problems, radiation diffusion, optimal control, and domain-decomposition Schur formulations. Its central premise is that Krylov convergence is governed less by the original coupled matrix than by the spectral quality of the Schur-complement approximation embedded in the preconditioner (Southworth et al., 2019, Greif, 26 Jan 2026, Pilotto et al., 5 Feb 2026).

1. Algebraic setting and nested Schur complements

The basic $2\times2$ setting starts from

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

with $A_{11}$ nonsingular and Schur complement

$S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$

Exact block factorization shows that once $A_{11}^{-1}$ is available, the remaining difficulty is the action of $S_{22}^{-1}$. In a block-lower-triangular preconditioner,

$L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$

so the preconditioned operator is explicitly controlled by $\widehat S_{22}^{-1}S_{22}$ (Southworth et al., 2019).

For double saddle-point systems, the structure is already nested. A representative form is

$$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$

with

$2\times2$0

The exact $2\times2$1 block factorization exposes $2\times2$2 as the trailing diagonal block. The algebraic problem is therefore not merely a Schur complement, but a Schur complement of a Schur complement (Greif, 26 Jan 2026).

For block-tridiagonal multiple saddle-point matrices, the same recursion persists. In finite-dimensional form one defines

$2\times2$3

while in the Hilbert-space formulation of block-tridiagonal Hessians the recurrence is

$2\times2$4

These recursions provide the exact block-diagonal or block-triangular factors used in ideal preconditioners and make precise the sense in which “multiple” saddle-point preconditioning is recursive Schur reduction (Sogn et al., 2017, Pilotto et al., 5 Feb 2026).

2. Preconditioner architectures

The main architectural variants differ by where the Schur approximation is placed and how aggressively the factorization is truncated.

Family	Representative form	Typical surrogate
Block-diagonal	$2\times2$5, $2\times2$6	pressure mass matrix, recursive $2\times2$7
Block-triangular / LDU	$2\times2$8 or exact/inexact block-LDU	sparse $2\times2$9, matrix-free Schur action
Nested / recursive	$A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$0, $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$1	inner Schur approximations with sign control
Deflated / corrected Schur	$A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$2, projectors $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$3, coarse $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$4	block-Jacobi plus coarse correction

In double saddle-point problems, one specialized construction is the BFB$A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$5 or “least-squares commutator” approximation. With $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$6, $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$7, and $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$8, one approximates

$A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$9

that is,

$A_{11}$0

and inserts $A_{11}$1 into the block-lower-triangular preconditioner

$A_{11}$2

This retains the exact outer algebraic form while only perturbing the terminal Schur block (Greif, 26 Jan 2026).

A second major pattern is nested block preconditioning. In incompressible Navier–Stokes and hyper-elastodynamics, the Schur operator is not formed explicitly; instead one applies

$A_{11}$3

matrix-free inside an intermediate Krylov solve, and that intermediate solve is itself embedded inside an outer FGMRES iteration for the original block system. This yields a three-level arrangement: outer Krylov on the full system, intermediate Krylov on the Schur system, and inner solves with $A_{11}$4 (Liu et al., 2019, Liu et al., 2018).

A third pattern augments Schur preconditioning by coarse-space correction. In discontinuous Poisson–Neumann formulations, the interface Schur system is right-preconditioned by a block-Jacobi matrix

$A_{11}$5

and deflated through

$A_{11}$6

The solve is then decomposed into coarse and fine components, combining local Schur solves with a global low-dimensional correction (Joshi et al., 2016).

Other architectures are based on permutation. For twofold or extended saddle-point systems, row and column permutations can expose a more favorable $A_{11}$7 form and an additive or triangular Schur complement, after which block-triangular preconditioners with approximate $A_{11}$8 become available (Cai et al., 2021, Wang et al., 2024, Wang, 2024).

3. Spectral theory and Krylov convergence

The exact Schur-complement setting gives the cleanest theory. For $A_{11}$9 systems, block-triangular or block-LDU preconditioners with the exact Schur complement guarantee at most two GMRES iterations in exact arithmetic. In the saddle-point case $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$0, block-diagonal preconditioners $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$1 yield GMRES convergence in at most three iterations. For general non-saddle-point matrices with $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$2, however, there is no uniform $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$3 bound on the number of GMRES iterations for block-diagonal preconditioning, even with the exact Schur complement (Southworth et al., 2020).

A sharper statement for nonsymmetric $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$4 block methods is that convergence of block-preconditioned GMRES is equivalent, up to modest constant factors, to convergence of GMRES on the preconditioned Schur-complement system $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$5. This establishes that an effective Schur-complement preconditioner is a necessary and sufficient condition for rapid convergence of $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$6 block-preconditioned GMRES, provided one diagonal block is inverted exactly (Southworth et al., 2019).

For double saddle-point systems, exact block-lower-triangular preconditioning with exact $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$7, $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$8, and $S_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}.$9 yields all eigenvalues equal to $A_{11}^{-1}$0, with minimal polynomial of degree $A_{11}^{-1}$1. If $A_{11}^{-1}$2 and $A_{11}^{-1}$3 remain exact but $A_{11}^{-1}$4 is replaced by an arbitrary SPD approximation $A_{11}^{-1}$5, then $A_{11}^{-1}$6 eigenvalues of $A_{11}^{-1}$7 are exactly $A_{11}^{-1}$8, and the remaining $A_{11}^{-1}$9 eigenvalues are precisely the generalized eigenvalues of

$S_{22}^{-1}$0

In the symmetric case $S_{22}^{-1}$1 SPD, $S_{22}^{-1}$2, $S_{22}^{-1}$3, the block-diagonal preconditioned spectrum consists of $S_{22}^{-1}$4 with algebraic multiplicity $S_{22}^{-1}$5, $S_{22}^{-1}$6 each with multiplicity $S_{22}^{-1}$7, and $S_{22}^{-1}$8 additional eigenvalues given by the roots of

$S_{22}^{-1}$9

For the BFB$L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$0 choice, one obtains $L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$1 in the SPD-$L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$2 case, which keeps the perturbed cubic roots uniformly bounded away from zero (Greif, 26 Jan 2026).

For multiple saddle-point systems, exact recursive Schur preconditioning yields operator-independent bounds. In the Hilbert-space block-tridiagonal setting, if $L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$3, then for $L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$4,

$L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$5

hence

$L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$6

These constants are sharp and can be expressed through roots of the difference of two Chebyshev polynomials of the second kind (Sogn et al., 2017).

The inexact multi-block theory generalizes this picture. For symmetric block-tridiagonal multiple saddle-point matrices with approximate Schur complements $L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$7, the spectrum of the block-diagonally preconditioned matrix is enclosed by intervals determined by roots of recursively defined monic polynomials $L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$8. When the blockwise indicators $L_{22}=\begin{pmatrix}A_{11}&0\A_{21}&\widehat S_{22}\end{pmatrix}, \qquad L_{22}^{-1}A= \begin{pmatrix} I & A_{11}^{-1}A_{12}\ 0 & \widehat S_{22}^{-1}S_{22} \end{pmatrix},$9 are all close to $\widehat S_{22}^{-1}S_{22}$0, the spectrum lies in two tight clusters around $\widehat S_{22}^{-1}S_{22}$1 and $\widehat S_{22}^{-1}S_{22}$2, which is the regime favorable to MINRES (Pilotto et al., 5 Feb 2026).

For non-symmetric discretizations, eigenvalue analysis is often replaced by field-of-values arguments. In the shifted-boundary Stokes problem, the full operator is written as

$\widehat S_{22}^{-1}S_{22}$3

and the field of values of $\widehat S_{22}^{-1}S_{22}$4 is shown to lie in a wedge away from the origin for sufficiently small $\widehat S_{22}^{-1}S_{22}$5. This yields a mesh-independent GMRES residual-reduction factor in the energy inner product induced by $\widehat S_{22}^{-1}S_{22}$6 (Wichrowski et al., 2 Jul 2026).

4. Inexactness, regularization, and approximation strategies

Practical block Schur-complement preconditioning is almost always inexact. The essential design problem is therefore not whether to approximate the Schur complement, but how to do so while preserving the spectral structure needed by the outer Krylov method.

One common strategy is to use physically motivated sparse surrogates. In weak-Galerkin Stokes, the Schur complement

$\widehat S_{22}^{-1}S_{22}$7

is approximated by the pressure mass matrix $\widehat S_{22}^{-1}S_{22}$8, leading to block-diagonal and block-triangular preconditioners for MINRES and GMRES. In shifted-boundary Stokes, the simplest approximation is again effective: $\widehat S_{22}^{-1}S_{22}$9 where $$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$0 is the pressure mass matrix on the surrogate domain (Huang et al., 2024, Wichrowski et al., 2 Jul 2026).

A second strategy is algebraic commutator or diagonal approximation. The BFB$$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$1 construction replaces the nested inverse $$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$2 by $$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$3 (Greif, 26 Jan 2026). In hemodynamic and hyper-elastodynamic nested SCR solvers, the explicit sparse Schur approximation

$$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$4

serves as a preconditioner for the matrix-free Schur Krylov solve (Liu et al., 2019, Liu et al., 2018). In anisotropic elliptic MMAP systems, the natural approximation

$$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$5

is accurate in one regime but tends to be singular for very small anisotropic parameters in the aligned case, whereas the algebraic approximation

$$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$6

is robust for $$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$7. Boundary-row-repaired approximations $$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$8, $$K= \begin{pmatrix} A & B^T & 0\ B & -D & C^T\ 0 & C & 0 \end{pmatrix},$$9, $2\times2$00, and the Robin variant $2\times2$01 were introduced precisely to repair the bad-row pathology (Li et al., 2021).

A third strategy addresses singularity or near-singularity by regularization. In nearly incompressible poroelasticity under pure Dirichlet displacement conditions, the system is reformulated as a three-field problem and stabilized by adding a rank-one term $2\times2$02. The resulting Schur block

$2\times2$03

is approximated by dropping the off-diagonal couplings. The block $2\times2$04 is then applied through the Sherman–Morrison–Woodbury formula because $2\times2$05 is diagonal (Huang et al., 23 Dec 2025). A closely related inherent regularization program for weak-Galerkin poroelasticity and elasticity adds the same type of rank-one correction to preserve the original solution while restoring nonsingularity, after which standard block Schur-complement preconditioning becomes parameter-free in the locking regime (Huang et al., 30 Jul 2025).

A fourth strategy augments local Schur approximations by coarse or low-rank correction. In Poisson–Neumann Schur solves, block-Jacobi is supplemented by deflation through $2\times2$06, $2\times2$07, and the projectors $2\times2$08 and $2\times2$09, which remove the slow coarse modes without changing the physical solution beyond the arbitrary constant (Joshi et al., 2015, Joshi et al., 2016). In general sparse systems reordered into interior/interface form, the Schur complement inverse is approximated by a truncated Neumann series around a block-diagonal anchor $2\times2$10,

$2\times2$11

and then corrected by a low-rank Arnoldi-based Sherman–Morrison–Woodbury update (Zheng et al., 2020).

A fifth strategy is blockwise AMG or hybrid AMG/Jacobi on approximate Schur factors. In multi-group radiation diffusion, Schur1 and Schur2 eliminate different physical blocks first, replace dense group Schur complements by diagonal block approximations such as $2\times2$12 or $2\times2$13, and apply a few AMG V-cycles or Jacobi sweeps depending on diagonal dominance and coupling strength (Yue et al., 2020).

5. Applications and empirical behavior

In incompressible flow, Schur-complement preconditioning is closely tied to pressure-Poisson structure. For Navier–Stokes coupled to reduced boundary models, the nested SCR preconditioner converged in $2\times2$14 outer iterations in the FDA nozzle benchmark for $2\times2$15–$2\times2$16 and $2\times2$17–$2\times2$18, while SIMPLE required $2\times2$19–$2\times2$20 iterations and failed completely at high $2\times2$21. Strong scaling on a $2\times2$22 M-unknown FDA mesh was nearly ideal $2\times2$23 up to $2\times2$24 ranks), and weak scaling from $2\times2$25 M–$2\times2$26 M unknowns required only $2\times2$27–$2\times2$28 iterations per linear solve (Liu et al., 2019). In hyper-elastodynamics, the analogous nested SCR preconditioner gave outer FGMRES convergence in $2\times2$29–$2\times2$30 iterations independent of stiffness in the isotropic cube-compression example, and weak scaling remained robust up to $2\times2$31 processors (Liu et al., 2018).

For Stokes and Stokes-coupled problems, the behavior depends on both algebra and geometry. In the double saddle-point Stokes–Darcy MAC discretization, eigenvalue plots show very strong clustering of eigenvalues of $2\times2$32 near $2\times2$33, and GMRES(20) iteration counts remain essentially mesh-independent and robust under variation of the physical parameters $2\times2$34 (Greif, 26 Jan 2026). In weak-Galerkin Stokes, representative $2\times2$35D iteration counts with $2\times2$36 were $2\times2$37 for MINRES with $2\times2$38 at $2\times2$39, and $2\times2$40 for GMRES with $2\times2$41 at $2\times2$42, with similar nearly mesh-independent and $2\times2$43-independent behavior in $2\times2$44D (Huang et al., 2024). In shifted-boundary Stokes, $2\times2$45 discretizations yielded GMRES counts of approximately $2\times2$46–$2\times2$47 in one arrangement and approximately $2\times2$48–$2\times2$49 after an initial coarse-mesh spike in a more complex arrangement, while $2\times2$50 pairs could be non-convergent on under-resolved meshes until the geometry was captured (Wichrowski et al., 2 Jul 2026).

In poroelasticity and nearly incompressible elasticity, block Schur-complement preconditioners are often coupled to structural regularization. For the three-field Bernardi–Raugel/weak-Galerkin formulation, numerical experiments in $2\times2$51D and $2\times2$52D show MINRES approximately twice GMRES and essentially constant counts as $2\times2$53 and $2\times2$54; in the spinal-cord simulation with three material layers and mesh sizes from $2\times2$55 up to $2\times2$56, GMRES required $2\times2$57 iterations for $2\times2$58 and $2\times2$59 for $2\times2$60 (Huang et al., 23 Dec 2025). In the related weak-Galerkin poroelasticity and elasticity framework, two-field poroelasticity in $2\times2$61D showed MINRES counts $2\times2$62–$2\times2$63 at $2\times2$64 and $2\times2$65 at $2\times2$66, while GMRES required $2\times2$67–$2\times2$68 at $2\times2$69 and $2\times2$70 at $2\times2$71; the three-field system remained $2\times2$72- and $2\times2$73-robust, though with higher counts (Huang et al., 30 Jul 2025).

In interface Schur problems for Poisson–Neumann systems, deflation changes the scaling regime. In the high-order discontinuous element discretization, block-Jacobi plus deflation makes GMRES iteration counts essentially independent of polynomial order $2\times2$74; on a $2\times2$75 grid with polynomial order $2\times2$76, the deflated solver needs only approximately $2\times2$77 GMRES iterations for a $2\times2$78 residual, and iteration counts remain within $2\times2$79 iterations as $2\times2$80 increases from $2\times2$81 to $2\times2$82 (Joshi et al., 2016). On long domains, the combined block-Jacobi-plus-deflation method achieves convergence independent of the grid size and requires half as many GMRES iterations and $2\times2$83 less wall-clock time than two-level non-overlapping additive Schwarz for a variety of grid sizes and domain aspect ratios (Joshi et al., 2015).

In block systems derived from optimal control and boundary control, permutation-based Schur preconditioning can sharply reduce iteration counts. For elliptic boundary optimal control with mixed boundary conditions, GMRES with the block-triangular preconditioner converged in $2\times2$84–$2\times2$85 iterations independent of $2\times2$86 on problems with over $2\times2$87 degrees of freedom (Wang, 2024). For the extended saddle-point system arising from Neumann boundary control, the Schur-complement triangular preconditioner produced iteration counts $2\times2$88 for $2\times2$89, $2\times2$90 for $2\times2$91, $2\times2$92 for $2\times2$93, and $2\times2$94 for $2\times2$95 as the degree of freedom count increased from $2\times2$96 to $2\times2$97, outperforming the compared MINRES and trivial-triangular alternatives (Wang et al., 2024).

In multi-physics diffusion, Schur complements can be organized by physical variable. For multi-group radiation diffusion, Schur1 and Schur2 took $2\times2$98–$2\times2$99 iterations on one-group systems and $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$00–$A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$01 on twenty-group systems, while adaptive variants delivered another $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$02–$A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$03 cpu reduction and scaled $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$04–$A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$05 from $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$06 to $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$07 cores (Yue et al., 2020).

6. Design principles, limitations, and recurrent misconceptions

A recurrent misconception is that access to the exact Schur complement automatically yields a uniformly optimal preconditioner regardless of block structure. That is false for general non-saddle $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$08 systems: block-diagonal preconditioning with an exact Schur complement does not necessarily converge in a fixed number of iterations, and there are explicit examples where it converges no faster than block-diagonal preconditioning with the original diagonal blocks (Southworth et al., 2020). The favorable fixed-iteration statements belong to more specific settings: exact block-triangular or block-LDU factorizations, classical saddle-point structure, or multiple saddle-point systems with the sign and symmetry hypotheses required by the theory.

A second misconception is that more elaborate symmetric or LDU-like variants necessarily improve over simple triangular forms. For nonsymmetric $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$09 block problems, approximate block-LDU or symmetric block-triangular preconditioners offer minimal reduction in iteration over block-triangular preconditioners, despite the additional computational cost, and block-Jacobi typically needs about twice as many Krylov steps (Southworth et al., 2019). This supports a widespread practical rule: once one diagonal block can be solved reliably, effort is usually better spent improving $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$10 than adding extra triangular sweeps.

A third issue is sign choice. In twofold and block-tridiagonal saddle-point systems, some exact Schur-based preconditioners are positively stable only for specific signs in front of the Schur blocks. The paper on twofold and block tridiagonal saddle point problems shows that positively stable preconditioners outperform other preconditioners if the Schur complements are further approximated inexactly (Cai et al., 2021). This is not a minor implementation detail; it determines whether the right-half-plane structure needed by GMRES is retained under approximation.

A fourth issue is that “natural” Schur approximations are often regime-dependent. In the MMAP anisotropic elliptic scheme, the natural continuous approximation $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$11 degrades as $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$12 in the aligned case, while the purely algebraic diagonal approximation $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$13 worsens as $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$14 grows. The robust behavior was recovered only after a boundary-row repair, via $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$15 or $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$16 (Li et al., 2021). A similar structural message appears in singular and nearly singular pressure systems: pinning one pressure degree of freedom, adding $A=\begin{pmatrix}A_{11}&A_{12}\A_{21}&A_{22}\end{pmatrix},$17, or projecting out a null space is not merely a numerical convenience but part of the block-Schur design itself (Huang et al., 2024, Huang et al., 23 Dec 2025, Huang et al., 30 Jul 2025).

Finally, mesh-independence claims are frequently asymptotic rather than unconditional. In the shifted-boundary Stokes problem, there is an explicitly identified coarse-mesh regime in which an under-resolved grid produces elevated iteration counts; the effect vanishes once the mesh captures the geometry and the boundary-shift perturbation becomes asymptotically small (Wichrowski et al., 2 Jul 2026). A plausible implication is that block Schur-complement preconditioning should be interpreted as a structural reduction framework whose success depends jointly on algebraic approximation quality, null-space handling, sign selection, and the discretization regime in which the reduced operator actually reflects the dominant physics.