---
title: Universal Split-Preconditioner
url: https://www.emergentmind.com/topics/universal-split-preconditioner
type: topic
---

# Universal Split-Preconditioner

Searching arXiv for relevant papers on universal split preconditioners and related splitting-based preconditioning.
Universal split-preconditioner denotes a family of preconditioning constructions in which the decisive operation is built from a splitting of the underlying operator, matrix, or field variables rather than from a single left or right approximate inverse. In the cited literature, the term appears in several nonidentical but related senses: a unique matrix-free preconditioner for accretive splittings $A=L+V$ with $\|V\|<1$; a split-relaxation framework for generic coupled block operators; a field-split preconditioner for monolithic fluid-structure interaction systems; and a learned unitary split preconditioner for adaptive filtering and linear solvers. Across these settings, the common mechanism is to choose a decomposition that makes the induced fixed-point or Krylov iteration more contractive, more stable, or better conditioned than the original formulation [2207.14222][2212.11111][1909.13201][1812.04570].

## 1. Range of meanings in the literature

The literature does not use *universal split-preconditioner* in a single canonical sense. Instead, “universal” refers to a broad admissible class within a given framework: all accretive splittings with bounded discrepancy, generic $2\times2$ coupled operators, multiphysics systems admitting a physical-field decomposition, or arbitrary SPD/autocorrelation matrices under a learned transform.

| Variant | Core construction | Scope |
|---|---|---|
| Matrix-free universal split preconditioner | $P=(L+I)(I-V)^{-1}$ | accretive systems with $A=L+V$, $\|V\|<1$ |
| Split-preconditioner iteration | $\mathcal A=\mathcal M-\mathcal N$, $\mathcal R=\mathcal M^{-1}$ | coupled differential equations |
| Field-split preconditioner | $M_{FS}=\mathrm{diag}(M_1,M_2)$ | monolithic FSI systems |
| Unitary split preconditioner | $U_1^{-1}AU_2^{-T}$ | LMS and linear systems |

A central terminological distinction appears in the transform-domain setting. A preconditioner $\zeta$ is called a left preconditioner if $\kappa(\zeta^{-1}A)<\kappa(A)$, a right preconditioner if $A\zeta^{-1}$ has improved $\kappa$, and a split-preconditioner if $\zeta=U_1U_2^T$ and the transformed operator is $U_1^{-1}AU_2^{-T}$. When $U_1=U_2=U$ is unitary, $\zeta=UU^T=I$, but the split action $U^TAU$ can still improve conditioning; this is the unitary split-preconditioner of PrecoG [1812.04570].

## 2. Algebraic core of split preconditioning

At the operator level, split preconditioning is most explicitly formulated by rewriting a coupled problem as
$$
\mathcal A w=f,\qquad \mathcal A=\mathcal M-\mathcal N,
$$
with $\mathcal M$ chosen so that it is easy to invert blockwise. Setting the relaxation or preconditioning operator $\mathcal R:=\mathcal M^{-1}$ yields the stationary iteration
$$
w^{k+1}=\mathcal R\mathcal N w^k+\mathcal R f
      = (I-\mathcal R\mathcal A)w^k+\mathcal R f.
$$
This is exactly a preconditioned Richardson iteration for the monolithic operator $\mathcal A$, with iteration operator $T:=I-\mathcal R\mathcal A$ [2212.11111].

The basic convergence criterion is $\rho(T)<1$. In the same framework, a practical sufficient condition is obtained from the bound
$$
\|T\|=\|\mathcal R\mathcal N\|\le R_0N_0<1,
$$
where $R_0=\|\mathcal M^{-1}\|$ and $N_0=\|\mathcal N\|$. This places split preconditioning and sequential coupling schemes inside the same algebraic template: the splitting is not merely a decomposition of equations, but the source of the preconditioner itself [2212.11111].

A related point, emphasized by the unitary-split literature, is that split preconditioning need not coincide with any nontrivial matrix factor $\zeta$ in the usual left/right sense. In the unitary case, the matrix identity $\zeta=I$ coexists with a nontrivial basis rotation $U^TAU$. This corrects the common misconception that a split preconditioner must appear as an explicit approximate inverse in the original coordinates [1812.04570].

## 3. Universal matrix-free construction for accretive operators

For a linear problem $Ax=y$ on a Hilbert space, the matrix-free universal split preconditioner assumes an accretive splitting
$$
A=L+V,\qquad \|V\|<1,\qquad \operatorname{Re}A\ge 0,
$$
after scaling. The construction is restricted to preconditioners that are first-order in the single operator $(L+I)^{-1}$, allowing exactly one application of $(L+I)^{-1}$ per iteration and no nested solves. Under the requirement that $\|I-P^{-1}A\|<1$ hold for all such accretive systems, the unique admissible form is
$$
P^{-1}=B(L+I)^{-1}=(I-V)(L+I)^{-1},
$$
equivalently
$$
P=(L+I)(I-V)^{-1}.
$$
With $M:=I-P^{-1}A$, one obtains
$$
M=B\,[I-(L+I)^{-1}B],
$$
and the fixed-point iteration converges monotonically by a Neumann-series argument [2207.14222].

The same work gives quantitative estimates. The condition number of $-P^{-1}A$ satisfies
$$
\kappa(-P^{-1}A)\le (1+\sqrt{2S})^2,\qquad S:=\|A^{-1}\|\,\|V\|,
$$
and with a step size $\omega\in(0,1]$ the worst-case convergence factor obeys
$$
\|I-\omega P^{-1}A\|\le \sqrt{1-c/S^2},\qquad c\approx 1/8,
$$
so that choosing $\omega\approx 1$ is near-optimal. In the Hermitian-positive-definite case, the sharper bound $\|I-P^{-1}A\|\le 1-1/S$ follows [2207.14222].

A major implementation feature is the elimination of the forward operator from the preconditioned action. The residual-based outer iteration is
$$
z=y-Ax_k,\qquad
w=z-Vz,\qquad
(L+I)u=w,\qquad
x_{k+1}=x_k+u.
$$
Thus each step requires one $(L+I)$-solve and one $V$-application; in many PDE settings $L$ is chosen to be translation-invariant so that $(L+I)^{-1}$ is a single FFT solve. The paper states that this often halves the time required per iteration. Matrix-free evaluation of $(I-V)^{-1}$ is available through the Neumann expansion
$$
(I-V)^{-1}=\sum_{j=0}^{\infty}V^j,
$$
and 3–5 inner steps suffice in practice if $\|V\|$ is well below $1$ [2207.14222].

The reported application classes are wave problems, diffusion problems, pantograph delay differential equations, and eigenvalue problems. For non-accretive systems, the extension uses antisymmetrisation through
$$
\widehat A
=
c^{-1}
\begin{bmatrix}
0 & -A^*\\
A & 0
\end{bmatrix},
$$
which is skew-Hermitian and hence accretive [2207.14222].

## 4. Schur and field splits for coupled PDE systems

For coupled systems in product space $W=U\times V$ with block operator
$$
\mathcal A=
\begin{bmatrix}
A & B\\
C & D
\end{bmatrix},
$$
the universal split-preconditioner framework is developed by introducing an implicit/explicit splitting $\mathcal A=\mathcal M-\mathcal N$ in which $\mathcal M$ is block-triangular. The exact Schur complement on the $(2,2)$ block is
$$
S=D-CA^{-1}B,
$$
but in the Schur-based Partial Jacobi construction one replaces $A^{-1}$ with $\widetilde A^{-1}:=\operatorname{diag}(A)^{-1}$ and defines
$$
\widetilde S=D-C\widetilde A^{-1}B,\qquad
L:=C\widetilde A^{-1}B=D-\widetilde S.
$$
The compact split iteration then becomes
$$
\underbrace{\begin{pmatrix}A&0\\ C&D+L\end{pmatrix}}_{\mathcal M}
\begin{pmatrix}u^{k+1}\\ v^{k+1}\end{pmatrix}
=
\underbrace{\begin{pmatrix}0&B\\ 0&L\end{pmatrix}}_{\mathcal N}
\begin{pmatrix}u^k\\ v^k\end{pmatrix}
+
\begin{pmatrix}f_1\\ f_2\end{pmatrix}.
$$
The paper describes this as universal because the only ingredients are the diagonal sub-blocks $A,D$ and the coupling $B,C$; it also states that the same template extends to nonlinear systems by using the corresponding Jacobians [2212.11111].

Representative numerical tests are given for the Dual-Porosity model and a Quad-Laplacian operator. In a 1D manufactured-solution Dual-Porosity test with 128 cells and varying coupling $\beta$, the unrelaxed $BGS$ scheme plateaus or even diverges for large $\beta$, whereas the SPJ-alternate scheme converges in 20–30 steps uniformly in $\beta$. In 2D Dual-Porosity on $50^2$, $100^2$, and $200^2$ grids, Block-GS requires $59$, $60$, and $57$ iterations, while $\mathrm{SPJ}_a$ requires $28$, $40$, and $50$. For the 2D Quad-Laplacian problem, Block-GS diverges, whereas $\mathrm{S2PJ}_a$ converges in approximately $15$, $13$, and $13$ steps independently of mesh size [2212.11111].

A specialized multiphysics realization appears in monolithic fluid-structure interaction. There the unknowns are ordered as
$$
x=(d^s,d^f,p^s,u^s,u^f,p^f),
$$
and the field-split preconditioner separates the Jacobian into two physical blocks:
Block-1 $(d^s,d^f,p^s)$ and Block-2 $(u^s,u^f,p^f)$. The additive field-split preconditioner is $M_{FS}=\mathrm{diag}(M_1,M_2)$, used as the level smoother inside a V-cycle geometric multigrid preconditioner for right-preconditioned GMRES; each field block is further preconditioned by a locally multiplicative ASM (Vanka) strategy with one-element-layer overlap [1909.13201].

The reported performance is problem dependent but consistently favorable against a pure domain-decomposition preconditioner. In the 2D aneurysm case, the average GMRES iteration count is $N\approx 9.0$ with convergence rate $\rho\approx 0.16$, and the level-5 CPU time decreases from $555\,\mathrm{s}$ for AS to $454\,\mathrm{s}$ for FS. In the 3D aneurysm case on 4 processes, AS requires $11\,662\,\mathrm{s}$ while FS requires $4939\,\mathrm{s}$. The summary states that FS is consistently $\sim20\%$ faster in 2D and $\sim50\%$ faster in 3D than the pure AS preconditioner, while both methods show weak mesh-independence [1909.13201].

## 5. Learned transforms and algebraic Schwarz splittings

In PrecoG, universality is attached to a data-driven unitary split preconditioner. The method learns a unitary transform $T$ and a diagonal power-normalization
$$
\Lambda=\operatorname{diag}(d_1^{-1/2},\dots,d_n^{-1/2}),
$$
where $d_i$ are the diagonal entries of $\widetilde R=T^TRT$, and forms
$$
P=T\Lambda T^T.
$$
When solving $Ax=b$, one applies $x\leftarrow PAx$ and $b\leftarrow Pb$; in transform-domain LMS, one first computes $v=\Lambda^{1/2}T^Tx$. The transform is obtained from the eigenvectors of a graph Laplacian $L(w)=BW B^T$, with edge weights learned by gradient descent on
$$
E_N(w)=E(p)+\beta(\|w\|_2^2-1),
$$
subject to $w\ge 0$ and $L(w)\succeq 0$ [1812.04570].

The conditioning target is explicit. If the eigenvalues of $\Lambda^{1/2}T^TAT\Lambda^{1/2}$ lie in $[1-\epsilon_2,1+\epsilon_1]$, then
$$
\kappa(PA)=\frac{1+\epsilon_1}{1-\epsilon_2}.
$$
For a first-order Markov covariance with parameter $\rho$, the unpreconditioned condition number scales as $\bigl(\frac{1+\rho}{1-\rho}\bigr)^2$, DFT preconditioning gives $\frac{1+\rho}{1-\rho}$, DCT gives $1+\rho$, whereas the learned PrecoG transform can push $\epsilon_1,\epsilon_2\to \rho^2$ so that $\kappa\approx 1+O(\rho^2)$. The reported computational cost is $O(n^4)$ per gradient step in the worst case for a fully connected graph, with $O(n^2)$ storage for $U$ and $w$ [1812.04570].

The empirical comparisons are likewise explicit. For regularized Hilbert matrices, PrecoG reduces $\kappa$ by 1–2 orders of magnitude relative to Jacobi, GS, DCT, and DFT. For random Gaussian SPD matrices, $\kappa(PA)$ is on average 10–100$\times$ smaller than with off-the-shelf transforms. In an AR(1) example with $n=4$ taps and $\rho=0.95$, $\kappa(\mathrm{DCT}\,R\,\mathrm{DCT}^T)=1.15$ whereas $\kappa(\mathrm{PrecoG}\,R\,\mathrm{PrecoG}^T)=1.125$, and the PrecoG-TDLMS filter reaches within $1\%$ of the true taps in approximately $1500$ iterations versus approximately $4500$ for DCT-TDLMS [1812.04570].

A related but distinct algebraic direction appears in sparse normal equations. For $C=A^TA$, an algebraic local SPSD splitting is constructed on overlapping subdomains, and the corresponding generalized eigenproblems select coarse modes with $\lambda_{i,k}>1/\tau$. The resulting two-level additive Schwarz preconditioner is
$$
P(\tau)^{-1}
=
R_0(R_0CR_0^T)^{-1}R_0^T
+
\sum_{i=1}^N R_i^T(R_iCR_i^T)^{-1}R_i,
$$
with condition-number bound
$$
\kappa(P(\tau)^{-1}C)
\le
(k_c+1)\Bigl[2+(2k_c+1)\frac{k_m}{\tau}\Bigr].
$$
The paper emphasizes that this bound is independent of $n$ and $N$, and its conclusion states independence of $n$, $m$, and $N$; it also states that the method is a fully black-box, matrix-free two-level Schwarz preconditioner for $A^TA$ [2107.09006].

The numerical summary reports ten large sparse least-squares matrices from SuiteSparse, with dimensions up to $8\,\mathrm{M}\times 7.5\,\mathrm{M}$ and $\operatorname{condest}(C)$ up to $10^8$. Two-level Schwarz solved all ten problems with LSQR in $<1\,\mathrm{K}$ iterations, always converged in $50$–$200$ iterations, and was fastest in wall-clock time on $256$ MPI ranks. On the “watson_2” matrix, increasing $\tau$ from $0.01275$ to $0.6$ increases the coarse dimension from $2400$ to $28\,700$ and reduces the iteration count from $740$ to $142$; the summary states that $\tau\approx 0.6$ yields robust performance across many matrices [2107.09006].

## 6. Scope, robustness, and obstructions

The strongest limitation on universal split preconditioning arises for indefinite systems. For a Hermitian nonsingular matrix split as
$$
A=M-N,\qquad G:=M^{-1}N=I-M^{-1}A,
$$
contractivity of the stationary iteration is equivalent to $\rho(G)<1$, which in turn is equivalent to
$$
\operatorname{spec}(M^{-1}A)\subset B(1,1):=\{z\in\mathbb C:\ |z-1|<1\}.
$$
If $A$ and $M$ are Hermitian and invertible but have different inertia, then $M^{-1}A$ has at least one real negative eigenvalue, hence $|1-\lambda|\ge 1$ for that eigenvalue and the iteration cannot be contractive. The resulting theorem states that a necessary condition for $\rho(M^{-1}N)<1$ is exact inertia preservation:
$$
i_+(M)=i_+(A),\qquad i_-(M)=i_-(A),\qquad i_0(M)=i_0(A).
$$
The paper summarizes the consequence succinctly: no splitting matrix can lead to a contractive stationary iteration unless the inertia is exactly preserved [2412.01554].

This has direct implications for claims of universality. One cannot use a single fixed definite $M$ and expect contractive stationary behavior for every indefinite system. In multigrid smoothing and block-Jacobi-type preconditioning, successful constructions must reproduce the negative modes of the fine-grid operator; the discussion cites Vanka-type or patch-based smoothers for Stokes flow as examples where local blocks carry the same number of negative directions as the global operator [2412.01554].

A broader synthesis of the literature suggests that universality is class-dependent rather than absolute. In one setting it means *every accretive splitting* with $\|V\|<1$ [2207.14222]; in another it means *any $2\times2$ block PDE or algebraic system* with access to diagonal blocks and couplings [2212.11111]; in multiphysics it means *any coupled PDE system with a physical-field decomposition* such as thermo-elasticity, MHD, or poroelasticity [1909.13201]; and in transform learning it means a learned $U$ and $\Lambda$ can be applied “blindly” to arbitrary autocorrelation matrices or SPD systems [1812.04570]. This suggests that universal split-preconditioning is best understood not as a single algorithm, but as a design principle: construct the iteration around a splitting whose inverse action, approximate Schur complement, or transformed basis captures the dominant coupling while remaining computationally cheaper than a full solve.

Source: https://www.emergentmind.com/topics/universal-split-preconditioner