Source-linked AI summary
A unified variational framework for augmented Lagrangian preconditioning of saddle point systems: applications to mixed-dimensional and fictitious domain methods
Michele Benzi, Marco Feder, Luca Heltai, Federica Mugnaioni
TL;DR
Large saddle point systems can be difficult to precondition as meshes are refined, including in mixed-dimensional and fictitious-domain settings. The paper develops continuous-level augmented Lagrangian preconditioners and analyzes their spectra. Numerical experiments report low, mesh-robust iteration counts across several settings, including coefficient jumps.
Problem
Large sparse saddle point systems may have conditioning that deteriorates under mesh refinement and physical or model parameter variation, including in mixed-dimensional and fictitious-domain problems.
Source: §1 · pp. 1, 3Method
The framework augments the variational formulation continuously before discretization and develops a spectral analysis of the resulting preconditioners.
Source: §5.1 · p. 10, §6.4 · p. 24Results
The experiments report low, mesh-robust FGMRES iteration counts, including for coefficient jumps and a three-dimensional case with up to 13 × 10^6 degrees of freedom.
Source: §6.3 · pp. 21–22Takeaways & Limitations
The framework offers practical checks for assessing its efficiency for constrained formulations while retaining the spectral advantages of algebraic augmentation.
Source: §7 · p. 24, §6.4 · p. 24Continuous augmentation may alter the discrete solution, and the resulting error can be nontrivial to control in some cases.
Source: §1.1 · p. 4
Abstract
from arXiv · showhide
We present a general framework to derive augmented Lagrangian preconditioners for saddle point systems arising from mixed finite element discretizations. Our approach is based on augmenting the variational formulation at the continuous level prior to discretization, providing a unified strategy for designing efficient preconditioners that are robust under mesh refinement for a broad class of problems leading to symmetric saddle point systems. The robustness of the proposed preconditioners is established through a detailed spectral analysis. This framework is applied to different model problems, focusing on mixed-dimensional and fictitious domain formulations involving Lagrange multipliers, with some configurations featuring large coefficient jumps. Extensive numerical experiments on different test cases and geometries are presented, validating the effectiveness and robustness of our preconditioners when used in conjunction with flexible GMRES.
1. Introduction.
Large saddle point systems can become poorly conditioned under mesh refinement and parameter variation, and mixed-dimensional and fictitious-domain formulations add difficult coupling structures. The paper introduces continuous-level augmentation as a unified preconditioning framework, while noting that it can alter discrete solutions.
1. Introduction.
- Large sparse saddle point systems may become poorly conditioned under mesh refinement or parameter variation, motivating augmented Lagrangian preconditioning.Mixed-dimensional constraints couple different-dimensional domains, while fictitious-domain methods use weak multiplier constraints and can require coupling across non-matching, overlapping meshes.Source: §1 · pp. 1–3
1.1. Motivation and main contributions.
- Continuous augmentation with mesh-dependent scaling gives a directly assemblable augmented form that avoids explicitly forming C^T W^-1 C and preserves local assembly.The framework is analyzed for symmetric positive-semidefinite bilinear forms and identifies conditions on the augmentation operator for favorable spectral properties.Source: §1.1 · pp. 3–4
- The framework covers multiplier-based Dirichlet boundary conditions, boundary and distributed fictitious-domain formulations, and immersed linear elasticity.The authors note that continuous augmentation may alter the discrete solution; the resulting error may be nontrivial to control in some cases.Source: §1.1 · p. 4
2. Notation.
Section 2 establishes notation for function spaces, pairings, spectra, and discretization-independent constants.
2. Notation.
- L2(D) denotes square-integrable functions on an open bounded domain D, with norm induced by the inner product (·,·)_D.The norm is written ∥·∥_0,D.Source: §2 · p. 4
- Sobolev spaces use W^s,p(D), with H^s(D)=W^s,2(D) and associated norm and seminorm denoted ∥·∥_s,D and |·|_s,D..Source: §2 · p. 4
- The notation section uses Spec(·) for a matrix spectrum, λ for the weak-form Lagrange multiplier, and c for a discretization-independent constant.The value of c may vary between formulas.Source: §2 · p. 4
3. Augmented Lagrangian formulation.
The formulation adds a constraint-residual penalty to the variational problem, then uses a scaled H-inner product for a practical discrete augmentation. Its well-posedness and scaling require conditions that depend on the application.
3. Augmented Lagrangian formulation.
- Adding γ(Lu−g,Lv)_Λ to the first equation leaves the continuous solution unchanged because the constraint gives Lu=g.The discrete scheme adds the corresponding penalty to the first equation.Source: §3 · pp. 4–5
- For positive-semidefinite a0 and a well-posed original problem, the augmented bilinear form satisfies coercivity for every γ>0, leaving the discrete inf-sup condition to verify.Source: §3 · p. 5
- Because computing the Λ-inner product may be impractical, the discrete augmentation instead uses γh^-α(Lu_h−g,Lv_h)_H with α≥0.The spectral analysis determines the choice of α.Source: §3 · p. 5
- Using the weaker H-norm requires an additional uniform coercivity condition, which must be verified separately for each application.Source: §3 · p. 5
3.1. Algebraic problem.
- The algebraic scheme represents the augmented bilinear form through matrices for the augmented block, constraint operator, and adjusted right-hand side.The augmented block includes the scaled H-inner-product term.Source: §3.1 · p. 5
- The H-inner-product augmentation has the same algebraic properties as C^T M^-1 C, where M is the multiplier-space mass matrix.This extends the earlier Grad-Div stabilization result discussed in.Source: §3.1 · p. 6
4. Connection between continuous and algebraic augmentation.
The continuous augmentation decomposes into an algebraic term matching the multiplier-mass matrix product and a projection-based stabilizing term. The latter controls when the continuous and algebraic formulations agree or converge toward one another.
4. Connection between continuous and algebraic augmentation.
- The H-inner-product augmentation is generally not equivalent to algebraically augmenting the discrete saddle point system, but decomposes into an algebraic term and a projection-type stabilizing term.The algebraic term coincides with C^T W^-1 C when W=M.Source: §4 · p. 6
- The algebraic term is represented using the multiplier-space mass matrix and constraint matrix, yielding the entries of C^T M^-1 C.The projection of each Lϕ_i onto Λ_h supplies the coefficient vector used in this matrix representation.Source: §4 · pp. 6–7
- The stabilizing term’s effect on the discrete solution is problem-dependent, and its bound depends on the residual associated with the continuous augmentation.The projection relation gives π_H(Lu_h−g)=0, which is used with the decomposition to obtain the bound.Source: §4 · p. 7
- When the stabilizing term vanishes under mesh refinement, continuous augmentation asymptotically retains algebraic augmentation’s properties without explicitly forming C^T M^-1 C.The paper cites Taylor–Hood Grad-Div stabilization and divergence-free Scott–Vogelius pairs as examples or relevant cases.Source: §4 · p. 7
5. Preconditioning and spectral analysis.
The section analyzes the augmented preconditioner’s spectrum and conditions for mesh-independent spectral bounds, including a boundary-Lagrange-multiplier example examining sparsity and assembly-time effects of continuous augmentation. It also considers how the choice α = 1 and stabilization weighting relate to approximation accuracy and iteration counts.
5. Preconditioning and spectral analysis.
- Under the stated symmetry, positive-semidefiniteness, and full-row-rank assumptions, the preconditioned eigenvalues are positive and real, and 1 is an eigenvalue of algebraic multiplicity at least n.When the augmentation lacks the stabilizing term, the other eigenvalues are below 1 and cluster toward 1 as γ increases.Source: §5 · pp. 8–9
- If the saddle-point matrix is singular, as can occur for enclosed Stokes flows with rank-deficient C, a zero eigenvalue need not prevent preconditioned GMRES convergence when the system is consistent.The spectral analysis also anticipates using a practical approximation to the ideal preconditioner in numerical experiments.Source: §5 · pp. 8, 10
5.1. Mesh-independence of the lower bound.
- For α = 2s, the lower spectral bound η stays away from zero as h → 0 when cκ is uniformly bounded; for algebraic augmentation, its cκ term is absent and increasing γ drives η toward 1.Continuous augmentation generally needs additional problem-dependent assumptions on cκ to establish a mesh-independent lower bound.Source: §5.1 · pp. 10–11
5.2. Example: boundary conditions through Lagrange multiplier.
- For weak boundary imposition with Lagrange multipliers, the analysis selects α = 1, and the continuous-augmentation consistency residual tends to zero under the stated mesh and regularity assumptions.The example uses multiplier space Λh ⊂ L2(Γ) and comparable mesh sizes hΓ ≃ hΩ for the residual argument.Source: §5.2 · pp. 12–14
- The continuous augmentation preserves the original matrix’s sparsity pattern, while algebraic augmentation is more than 16 times denser and continuous augmentation assembles 90× faster on the finest mesh.For the reported three-dimensional example, the global (1,1)-block size is 3,817; later tests directly assemble the augmented bilinear form rather than forming a triple matrix product.Source: §5.2 · p. 14, Figure 1 · p. 14, Figure 2 · p. 15
- The preconditioned eigenvalues are real and positive, the smallest remains uniformly bounded under mesh refinement, and increasing γ from 10 to 1000 clusters the spectrum near 1.The refinement tests vary γ across 10, 100, and 1000; the smallest eigenvalue stays bounded at each fixed γ.Source: §5.2 · p. 15
5.2.1. Interplay between h−α, convergence rates, and iteration counts.
- Both tested powers retain optimal convergence rates and essentially unchanged errors, but iteration counts remain bounded only for f_W = h_ΓM.The preconditioning-driven choice α = 1 is not motivated by accuracy; the experiment checks that it does not degrade the discrete solution.Source: §5.2.1 · p. 16
6. Numerical experiments.
The experiments evaluate augmented Lagrangian preconditioners for mixed finite element systems across weak boundary enforcement and fictitious-domain formulations, including immersed elasticity. Across the tested refinements and geometries, the reported solver iteration counts remain robust; the strong-scaling study also identifies communication-intensive assembly phases.
6. Numerical experiments.
- The experiments assess preconditioners for mixed-dimensional and fictitious-domain formulations with Lagrange multipliers, using flexible GMRES and inexact block solves.The suite includes several geometries and configurations with coefficient jumps.Source: §6 · pp. 16–17
6.1. Weak imposition of boundary conditions with Lagrange multipli-
- For weakly imposed boundary conditions, the AL preconditioner has constant, very low iteration counts across refinement in two dimensions and remains robust on a challenging three-dimensional geometry.In the three-dimensional test, the average inner PCG count for approximately inverting the augmented block never exceeds nine.Source: §6.1 · p. 18, Table 4 · p. 18, Table 3 · p. 18
6.2. Fictitious domain with boundary supported Lagrange multipliers.
- For a three-dimensional popcorn-like immersed interface with boundary-supported multipliers, iteration counts remain robust under refinement; robust counts are also reported for a simpler two-dimensional test.The surface mesh contains 12,178 triangles.Source: §6.2 · p. 19, Table 5 · p. 19, Figure 4 · p. 19
6.3. Fictitious domain with distributed Lagrange multipliers.
- For distributed-multiplier fictitious domains, the modified AL preconditioner remains robust across coefficient jumps and mesh refinement, including immersed elasticity with nearly four orders of magnitude in material contrast and up to 13 million degrees of freedom.The two-dimensional elliptic-interface test reports robust FGMRES counts as β2 varies; the three-dimensional elasticity test reports low, robust counts under simultaneous refinement.Source: §6.3 · pp. 21–22, Table 6 · p. 21
6.4. Strong scaling study.
- FGMRES iteration counts remain independent of MPI process count, while linear-solver and classical-assembly phases scale well until communication overhead grows; particle detection and augmented-term assembly dominate assembly cost.The communication-heavy particle-detection phase locates immersed-mesh quadrature points in the independently partitioned background mesh.Source: §6.4 · p. 23, Figure 7 · p. 24
7. Conclusions, practical guidelines and outlook.
Numerical experiments validate the framework on mixed-dimensional formulations, varied discretizations, and complex geometries, while its abstract analysis covers a broader class of constrained formulations. Four checks guide whether the framework is likely to be efficient for a new problem; a projected augmentation is a possible alternative when a check fails.
7. Conclusions, practical guidelines and outlook.
- Experiments with mixed-dimensional formulations, different discretizations, and complex geometries validate the theoretical findings, while the analysis applies to broader constrained formulations.The authors present these conditions as a practical scope for the framework, not as a claim that every constrained problem meets them.Source: §7 · p. 24
- Assess a new application through coercivity, an inverse inequality, vanishing augmentation residual under mesh refinement, and compatible mesh-dependent scaling.These checks are conditions for assessing whether the framework provides an efficient alternative to algebraic augmentation while retaining its spectral advantages.Source: §7 · p. 24
- If a check fails, projected augmentation is a possible alternative that still avoids explicitly computing the triple matrix product C^T W^-1 C.Source: §7 · p. 24
- Future work will extend the approach to fluid-structure interaction and develop tailored multigrid solvers for approximate inversion of augmented blocks.Source: §7 · p. 24