---
title: Viscosity-Scaled Mass-Matrix Preconditioners
url: https://www.emergentmind.com/topics/viscosity-scaled-mass-matrix-preconditioners
type: topic
---

# Viscosity-Scaled Mass-Matrix Preconditioners

A viscosity-scaled mass-matrix preconditioner is a matrix-based approximation technique used to accelerate the solution of saddle-point systems that arise in the finite element discretization of Stokes and generalized Stokes (e.g., p-Stokes) problems with variable or strongly heterogeneous viscosity. By incorporating variable viscosity as a weight in the pressure mass matrix, these preconditioners produce spectral properties and Krylov solver convergence that are robust to mesh refinement, high viscosity contrast, and regularization parameters. Viscosity-scaled approaches are now implemented in large-scale ice-sheet modeling and stochastic Stokes formulations, and have motivated more advanced preconditioning frameworks for high-contrast, non-Newtonian, and stochastic PDEs [2307.14688][1712.06472][1607.03936].

## 1. Mathematical Setting and Motivation

Finite element discretizations of Stokes-type flows with variable (often power-law) viscosity yield coupled block-linear systems of the form
\[
\begin{pmatrix}
A & B^\top \\
B & 0
\end{pmatrix}
\begin{pmatrix} u \\ p \end{pmatrix}
=
\begin{pmatrix} f \\ 0 \end{pmatrix}
\]
where $A$ is SPD (viscous block), $B$ the discrete divergence, $u$ the velocity, and $p$ the pressure. For variable viscosity $\nu(x)$ or $\mu(x)$, the stiffness matrix $A$ and the Schur complement $S = B A^{-1} B^\top$ can become severely ill-conditioned, especially as the minimum viscosity or regularization parameter tends to zero, leading to slow Krylov-method convergence.

A classical preconditioning strategy for saddle-point systems is to block-precondition the Schur complement with a pressure mass matrix. However, when viscosity is strongly variable, the standard (non-viscosity-scaled) mass matrix provides poor spectral equivalence, with small eigenvalues decaying with the regularization parameter and high viscosity contrast [2307.14688][1607.03936]. Viscosity-scaled mass-matrix preconditioners address this by replacing the mass matrix with one inversely weighted by the viscosity.

## 2. Construction of Viscosity-Scaled Mass-Matrix Preconditioners

The canonical viscosity-scaled pressure mass matrix for piecewise polynomial pressure bases $\{\psi_i\}$ is defined as
\[
(M_\nu)_{ij} = \int_\Omega \nu(x)^{-1} \psi_i(x) \psi_j(x) \, dx
\]
For the preconditioned system, one replaces the Schur complement $S$ or its inverse with $M_\nu$ or $M_\nu^{-1}$ respectively. The resulting block-diagonal preconditioner is
\[
P = \begin{pmatrix}
A & 0 \\
0 & M_\nu
\end{pmatrix}
\]
For stochastic or high-dimensional settings (e.g., lognormal random viscosity), the scaling is applied in the pressure block across all stochastic degrees of freedom [1712.06472].

Spectrally, the preconditioned Schur complement $M_\nu^{-1} S$ (or its generalized versions) exhibits eigenvalue bounds that depend only weakly on the minimum viscosity or regularization $\varepsilon$, as opposed to the severe dependence for non-scaled mass preconditioners [2307.14688].

## 3. Spectral Properties and Eigenvalue Bounds

Let $\lambda$ denote the eigenvalues of $M_\nu^{-1} S$. In generalized p-Stokes settings, for both Picard and Newton linearizations one obtains
\[
c_\nu^2 \leq \lambda \leq \frac{d}{1+\gamma(p-2)}
\]
where $d$ is domain dimension, $\gamma$ indicates the linearization type, and $c_\nu$ is an auxiliary inf-sup constant dependent only weakly on the minimum viscosity [2307.14688].

By contrast, non-viscosity-scaled preconditioning yields
\[
c_0^2 \varepsilon^{2-p} \leq \lambda \leq \frac{(\varepsilon^2+\|\mathbf{D}u\|^2_{L^\infty})^{(2-p)/2}}{\nu_0(1+\gamma(p-2))}
\]
which becomes arbitrarily ill-conditioned as $\varepsilon\to0$ for shear-thinning ($p<2$) fluids. For variable viscosity, classic (non-scaled) mass-matrix preconditioners also exhibit eigenvalue clustering that degrades with high viscosity contrast [1607.03936].

Eigenvalue clustering for viscosity-scaled preconditioning is robust:
- Both upper and lower bounds are independent of the regularization parameter.
- The ratio $\lambda_{\max}/\lambda_{\min}$ remains moderate ($<10$ for standard discretizations), ensuring rapid, mesh-independent Krylov convergence in practice.
- For high-contrast settings or random viscosity, spectral equivalence persists, with all eigenvalues in a compact interval determined only by global viscosity-norms [1712.06472][1607.03936].

## 4. Practical Implementation and Numerical Results

Viscosity-scaled mass-matrix preconditioners are implemented by assembling the pressure mass matrix with $1/\mu(x)$ or $1/\nu(x)$ as the weight. For standard piecewise-constant pressure bases, this diagonal form lends itself to efficient sparse and parallel computation across distributed-memory systems [2307.14688].

Numerical validation has been performed in several contexts:
- **Ice-sheet simulations:** Picard/Newton iterations for full-Stokes, power-law ice flow on complex geometries (e.g., Haut Glacier d’Arolla benchmark), using both Taylor–Hood (P2–P1) and MINI elements, with viscosity scaling yielding nearly constant condition numbers and iteration counts across several orders of magnitude in regularization [2307.14688].
- **Stochastic Stokes problems** (lognormal random viscosity): Krylov iteration counts are $h$-independent, robust to stochastic dimension, and approximately 30% fewer for block triangular (Bramble–Pasciak) variants employing viscosity-weighted pressure mass [1712.06472].
- **Highly heterogeneous viscosity:** For multi-sinker problems (up to $10^{10}$ viscosity contrast), standard Schur-comp mass-matrix preconditioners fail or require thousands of iterations, while viscosity-scaled (and especially weighted BFBT) strategies keep GMRES iterations nearly flat ($30$–$60$), independent of mesh refinement and viscosity contrast [1607.03936].
- **High-order FEM and parallel scalability:** Viscosity-scaled mass-matrix preconditioners, especially when combined with hybrid spectral–geometric–algebraic multigrid, retain excellent parallel efficiency and only mild $k$-dependence [1607.03936].

## 5. Advanced Variants and Connections

While inverse-viscosity-weighted mass matrices provide robust spectral equivalence in many non-Newtonian and random coefficient settings, certain pathologies (e.g., extreme inclusions or Dirichlet boundaries) can still degrade convergence. Weighted BFBT (w-BFBT) preconditioners, in which velocity mass matrices are lumped with $\sqrt{\mu(x)}$ weights, further improve the spectral proximity to the Schur complement and enhance robustness for complex viscosity landscapes [1607.03936]:
\[
\widetilde S^{-1}_{w\text{-}BFBT}
= (B\,C^{-1}B^\top)^{-1} (B\,C^{-1}A\,D^{-1}B^\top)(B\,D^{-1}B^\top)^{-1}
\]
with $C = D = $ mass-lumped matrices with $\sqrt{\mu}$ weighting. Boundary amplification of weights on $\partial\Omega$ allows mesh- and order-independence for Dirichlet problems.

Triangular block preconditioners and short-recurrence Krylov schemes (e.g., Bramble–Pasciak conjugate gradient, BPCG) further exploit the structure imposed by viscosity scaling, often yielding further reductions in iteration counts and sharper spectral inclusions [1712.06472].

## 6. Recommendations and Limitations

For the finite-element solution of variable viscosity Stokes or generalized Stokes (p-Stokes) systems:
- Use the viscosity-scaled mass matrix $(\nu^{-1}\psi_i, \psi_j)_\Omega$ to approximate the Schur complement.
- Employ Newton linearization when possible, as it tightens eigenvalue bounds and improves nonlinear convergence [2307.14688].
- Pair with robust multigrid or AMG solvers for the velocity block $A$ and efficient solvers for the scaled mass matrix.
- For high-order, inf-sup stable discretizations, consider advanced weighted commutator-based preconditioners such as w-BFBT for extreme contrast or boundary-dominated problems [1607.03936].
- Maintain high mesh quality, particularly for MINI elements, as mesh skewness can degrade the relevant inf-sup constants and thus spectral clustering [2307.14688].
- Viscosity-scaled mass matrices are particularly well suited to parallel implementation, as element-local computations dominate for standard discontinuous pressure spaces.

## 7. Broader Impact and Applications

Viscosity-scaled mass-matrix preconditioners have become the standard in large-scale ice-sheet simulation codes and stochastic flow solvers, providing a critical foundation for robust, efficiently preconditioned Krylov solves in uncertain, multiscale, and high-contrast regimes. Their development has directly motivated advances in multigrid-in-multiphysics methods, matrix-theoretic analysis of saddle-point spectra, and has facilitated state-of-the-art solver scalability on leadership-class computing hardware [2307.14688][1712.06472][1607.03936].

Their success demonstrates the effectiveness of physically informed scaling in preconditioner design and suggests further research into customized operator-weighted preconditioners for more general classes of PDE-constrained optimization and inverse problems.

Source: https://www.emergentmind.com/topics/viscosity-scaled-mass-matrix-preconditioners