Adaptive Composite Solvers

• Adaptive composite solvers are algorithmic frameworks that integrate heterogeneous solver components with adaptive mechanisms to tackle PDEs and composite optimization.
• They decompose complex problems into tractable subproblems that are recombined via additive, multiplicative, or hierarchical strategies to enhance robustness and parallel efficiency.
• Adaptive triggers—such as eigenvalue thresholds and online error indicators—dynamically tune solver configurations, improving convergence and reducing computational expense.

Searching arXiv for the cited papers to ground the encyclopedia entry with current paper metadata. Searching arXiv for "High-performance dune modules strongly anisotropic elliptic problems GenEO (Butler et al., 2019)" Searching arXiv for "fully adaptive high-order fast Poisson solver complex two-dimensional geometries (Fortunato et al., 29 Jan 2025)" Searching arXiv for "Adaptive Accelerated Composite Minimization (Baghbadorani et al., 2024)" Searching arXiv for "An Adaptive Proximal ADMM for Nonconvex Linearly Constrained Composite Programs (Maia et al., 2024)" Searching arXiv for "Characterization of the near-null error components utilized in composite adaptive AMG solvers (Nelson et al., 10 Sep 2025)" Searching arXiv for "Solver composition across the PDE linear algebra barrier (Kirby et al., 2017)" Adaptive composite solvers are solver configurations or algorithms that combine multiple computational components and adapt their interaction to problem structure, discretization, or runtime information. Across recent work, the term covers composable preconditioners for PDE discretizations, multilevel or domain-decomposed solvers with adaptive coarse spaces, adaptive proximal and accelerated methods for composite optimization, randomized block linear solvers with adaptive sampling, and discretization-aware schemes in which quadrature, time stepping, or stochastic path generation are themselves adapted during the solve (Kirby et al., 2017, Baghbadorani et al., 2024, Patel et al., 2022). This suggests a family resemblance rather than a single canonical algorithm: the common feature is not one update formula, but the deliberate composition of heterogeneous solver ingredients together with an explicit adaptation mechanism.

1. Conceptual structure and recurring principles

A recurring formulation is to decompose the computational task into solver subproblems that are individually tractable and then to re-couple them by additive, multiplicative, block, or hierarchical composition. In PDE settings, this appears as coarse-grid correction plus local subdomain solves, or as a bulk solver coupled to a boundary-strip solver through layer potentials. In optimization, it appears as a smooth part plus a prox-friendly nonsmooth part, or as block subproblems coupled through linear constraints. In algebraic solvers, it appears as row-action and column-action steps, or as multiple AMG components each specialized to a slow error mode (Butler et al., 2019, Fortunato et al., 29 Jan 2025, Maia et al., 2024, Nelson et al., 10 Sep 2025).

The adaptive mechanism differs by domain but has a stable role. Typical triggers are local generalized eigenvalues in overlapping subdomains, zero-order function evaluations in line search, stationarity residuals and penalty tests in ADMM, Gram-determinant conditions for randomized block solvers, or online indicators comparing a training quadrature to a richer reference quadrature (Butler et al., 2019, Baghbadorani et al., 2024, Maia et al., 2024, Badia et al., 1 May 2026). In all cases, adaptation is used to control a quantity that would otherwise be difficult to tune globally: anisotropy-induced low-energy error components, unknown smoothness constants, penalty scaling, search-space diversity, or quadrature distortion.

A second recurring principle is that composite structure is often introduced to preserve both robustness and computational throughput. Multiplicative multigrid or exact coarse solves are typically robust but hard to parallelize; additive variants or duplicated coarse solves increase concurrency but need damping or auxiliary corrections to suppress overcorrection (Murray et al., 2019, Butler et al., 2019). Likewise, fixed smoothness or penalty parameters simplify analysis but are fragile in practice; adaptive rules replace these constants by quantities inferred during the run (Baghbadorani et al., 2024, Maia et al., 2024).

2. Elliptic and multilevel PDE realizations

In large-scale elliptic PDEs, adaptive composite solvers are often built around multilevel preconditioning. For strongly anisotropic aerospace composites, dune-composites implements a two-level additive Schwarz preconditioner with a GenEO coarse space,

$$\mathbf{M}^{-1}_{AS,2} = \mathbf{R}_H^T\,\mathbf{A}_H^{-1}\,\mathbf{R}_H + \sum_{j=1}^N R_j^T\,\mathbf{A}_j^{-1}\,R_j,$$

where the adaptivity comes from local generalized eigenproblems in overlapping subdomains. The coarse basis retains all local eigenfunctions satisfying

$$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$

which yields a condition-number bound independent of coefficient contrast, anisotropy, mesh size, and number of subdomains once all modes below the threshold are included (Butler et al., 2019). The same framework reported weak scaling on Archer from $6.4\times 10^6$ to $2.0\times 10^8$ degrees of freedom across $480$ to $15{,}360$ cores, with total times approximately $734$–$845$ s, and solved an aerospace composite problem with over $200$ million degrees of freedom in a few minutes (Butler et al., 2019).

A distinct composite design appears in the 2D Poisson solver that partitions the domain into a bulk region and a variable-width boundary-conforming strip. The bulk particular solution is computed by a truncated free-space box code on an adaptive quadtree, while the strip problem is solved by spectral collocation on curvilinear elements and the two particular solutions are glued across the fictitious interface $\tilde{\Gamma}$ by Laplace single- and double-layer potentials,

$$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$0

This avoids function extension and cut-cell quadrature near the boundary, yields overall $$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$1 complexity for the adaptive discretization, and demonstrates typically $$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$2-digit accuracy (Fortunato et al., 29 Jan 2025).

Adaptive composite multigrid on spacetrees provides a third PDE realization. The AFACx-style additive multigrid variants adAFAC-Jac and adAFAC-PI solve all levels concurrently and add an auxiliary damping equation per level to suppress the additive overshoot that would otherwise arise from simultaneous corrections. The auxiliary damping is designed from the next coarser level, either through smoothed restriction,

$$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$3

or through prolongation-injection damping, and is executed with single-touch semantics on dynamically adaptive meshes (Murray et al., 2019). This places damping itself inside the composite solver rather than treating it as an external relaxation parameter.

Adaptive mesh refinement can also be part of the composite architecture. In the AMR+BDDC framework, p4est-based one-level hanging-node refinement is coupled to two- and three-level BDDC preconditioners. The local Schur complements are corrected by a coarse space built from edge and face averages, with corner constraints for elasticity. The reported largest adaptive run exceeded $$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$4 unknowns on $$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$5 cores, and the number of iterations remained acceptable despite refined meshes and disconnected subdomains (Kůs et al., 2017).

3. Runtime composition and solver software architectures

A central software development is the move from model-specific solver implementations to runtime-composable, PDE-aware solver stacks. In Firedrake and PETSc, this is expressed by crossing the “PDE/linear algebra barrier”: implicit matrices store UFL bilinear forms, function spaces, boundary conditions, and application context, while custom preconditioners can request auxiliary operators such as mass matrices, Laplacians, or pressure convection–diffusion operators that were not part of the original Jacobian (Kirby et al., 2017). This supports runtime selection of matrix-free application, block decompositions, Schur-complement approximations, AMG backends, and field splits without hard-wiring these choices into the model formulation.

For incompressible flow and Rayleigh–Bénard systems, the same framework composes FieldSplit block factorizations with PCD inside the pressure Schur complement and AMG on temperature blocks. The relevant approximation is

$$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$6

while the global operator may remain matrix-free and only selected blocks are assembled through AssembledPC for algebraic preconditioning (Kirby et al., 2017). The architectural point is that “composite” here means both mathematical block structure and software-level composition of KSP and PC objects.

DUNE-based implementations emphasize a different but related pattern. dune-composites encapsulates anisotropic elasticity operators, serendipity elements, stacking sequences, defects, and GenEO coarse-space assembly inside a C++ framework built on PDELab and ISTL (Butler et al., 2019). The implementation exposes configuration parameters such as overlap width, eigenvalue threshold, choice of partition of unity, and local direct solver, while keeping coarse-space assembly and MPI communication inside specialized modules such as geneo/partitionofunity.hh and geneo/multicommdatahandle.hh (Butler et al., 2019).

Spacetree-based multigrid brings the same compositional philosophy to dynamically adaptive meshes. The AFAC implementation combines FAS, HTMG, additive level updates, auxiliary damping equations, and BoxMG-style operator construction inside a single pipelined traversal, with partial residual exchange and vertical integration to avoid a conventional coarse-grid bottleneck (Murray et al., 2019). A plausible implication is that adaptive composite solvers have become as much an issue of software architecture and data movement as of asymptotic iteration counts.

4. Composite optimization methods

In convex composite minimization, the canonical model is

$$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$7

with $$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$8 smooth and $$\lambda_k^{(j)} \le \frac{\mathrm{diam}(\Omega_j)}{\mathrm{width}(\Omega_j^\circ)},$$9 proper, closed, convex, and prox-friendly. The adaptive accelerated composite method selects $6.4\times 10^6$0 by a zero-order linesearch using function evaluations at $6.4\times 10^6$1 and $6.4\times 10^6$2, and then performs the proximal step

$6.4\times 10^6$3

The method proves $6.4\times 10^6$4 without using $6.4\times 10^6$5 explicitly, and achieves $6.4\times 10^6$6 in the non-accelerated case and $6.4\times 10^6$7 in the accelerated case (Baghbadorani et al., 2024). The composite aspect lies in decoupling the adaptive rule, which uses only function evaluations of $6.4\times 10^6$8, from the nonsmooth handling, which remains in the proximal map.

For nonconvex linearly constrained composite programs,

$6.4\times 10^6$9

the adaptive proximal ADMM decomposes the problem into sequential block proximal subproblems and adaptive multiplier and penalty updates. The augmented Lagrangian is

$2.0\times 10^8$0

and each block subproblem is solved only to a certified inexactness level. The method is adaptive to weak convexity constants, Lipschitz constants, penalty scaling, and subproblem accuracy; it does not require “last-block” rank assumptions; and it attains a $2.0\times 10^8$1-stationary point with total primal iteration complexity

$2.0\times 10^8$2

matching the state of the art for proximal ADMM-type methods in this class (Maia et al., 2024).

A higher-order counterpart is provided by adaptive third-order methods for composite convex optimization. These minimize quartically regularized third-order Taylor models,

$2.0\times 10^8$3

and solve the auxiliary problem by a Bregman gradient method whose generator includes both $2.0\times 10^8$4 and the quartic term (Grapiglia et al., 2022). The regularization parameter is adapted according to the observed progress of the inner solver rather than solely by model backtracking. The abstract states that the basic method needs at most $2.0\times 10^8$5 inner iterations and the accelerated variant $2.0\times 10^8$6 (Grapiglia et al., 2022).

Adaptive smoothing provides a closely related homotopy view. For nonsmooth composite problems with

$2.0\times 10^8$7

Nesterov smoothing replaces $2.0\times 10^8$8 by

$2.0\times 10^8$9

and then decreases $480$0 automatically during accelerated proximal-gradient iterations (Tran-Dinh, 2015). The worst-case complexity remains $480$1, but the smoothing parameter is updated online instead of being fixed from a target tolerance.

5. Algebraic, randomized, and acceleration-based iterations

Composite adaptivity also appears at the purely algebraic level. In adaptive composite AMG, one starts from an SPD, $480$2-convergent solver $480$3, detects a slow-to-converge error component $480$4, constructs a new AMG component $480$5 whose interpolation contains that mode, and then forms the symmetrized composition

$480$6

The key theorem shows that if

$480$7

then $480$8 is near-null for $480$9, quantified by

$15{,}360$0

This justifies using the observed undamped component to drive modularity-based aggregation in successive AMG components (Nelson et al., 10 Sep 2025). Reported experiments on anisotropic diffusion, SPE10, and the AMD G3_circuit matrix show that stationary composite performance approaches Krylov-preconditioned performance as the number of components grows (Nelson et al., 10 Sep 2025).

Randomized block adaptive linear system solvers treat composition differently. They allow row-action or column-action projections, adaptive sketch or block selection, finite-memory state, and changing metrics or preconditioners. The theory reduces convergence to three criteria: Markovian finite-memory dependence, $15{,}360$1-Exploratory action selection, and Uniformly Nontrivial block geometry (Patel et al., 2022). Under these conditions, the methods achieve worst-case exponential convergence, even for under-determined, over-determined, inconsistent, or rank-deficient systems. The stopping-time bounds scale with $15{,}360$2, so exploration strength becomes an explicit design parameter (Patel et al., 2022).

Nonlinear fixed-point acceleration can also be composed adaptively. Composite Anderson acceleration combines stationary AA with different window sizes either additively,

$15{,}360$3

or multiplicatively, by applying sAA$15{,}360$4 in an outer loop and sAA$15{,}360$5 in an inner loop (Chen et al., 2022). When combined with optimized damping, the resulting non-stationary schemes often converge faster than stationary sAA and can reduce storage substantially. On the Bratu problem, AAoptD$15{,}360$6 was reported at $15{,}360$7 s, compared with $15{,}360$8 s for $15{,}360$9 and more than $734$0 s for AAoptD$734$1 (Chen et al., 2022).

At the smallest algebraic scale, structure detection itself becomes the adaptive composite mechanism. The Armadillo dense linear solver detects banded, triangular, and symmetric positive definite structure and then dispatches to matched LAPACK routines such as xGBTRF/xGBTRS, xTRTRS, xPOTRF/xPOTRS, or xGETRF/xGETRS, with rcond estimation and xGELSD fallback for poorly conditioned systems (Sanderson et al., 2020). The reported wall-clock reductions reached approximately $734$2–$734$3 for banded cases and $734$4–$734$5 for triangular cases, while overhead on dense unstructured problems remained negligible (Sanderson et al., 2020).

6. Adaptive discretization, stochastic drivers, and open limitations

Several recent contributions shift the adaptive composite idea from solver kernels to discretization and uncertainty control. In calibrated probabilistic ODE solvers, an IWP$734$6 Gauss–Markov prior is combined with EKF/EKS inference, quasi maximum-likelihood calibration of the diffusion $734$7, and adaptive step-size control based on the local posterior covariance (Bosch et al., 2020). Calibration makes the uncertainty magnitudes descriptive rather than merely rate-correct, and numerical experiments benchmark the approach against Dormand–Prince 4/5 while showing that EKS1 can handle stiff regimes such as Van der Pol with $734$8 (Bosch et al., 2020).

For adaptive high-order SDE solvers, the extended Virtual Brownian Tree generates Brownian motion together with its Gaussian time integrals from a single PRNG seed. The construction exactly matches the joint distribution of Brownian motion and the required integrals at arbitrary query times, provided the queries are at least $734$9 apart, and does so with logarithmic complexity in $845$0 and constant memory footprint (Jelinčič et al., 2024). Two applications are highlighted: in a high-volatility CIR model, adaptive solvers achieved more than twice the convergence order of constant stepping, and in underdamped Langevin MCMC the adaptive third-order solver outperformed NUTS while using only a tenth of its function evaluations (Jelinčič et al., 2024).

In neural PDE residual minimization, the anisotropic adaptive composite quadrature framework separates approximation, quadrature, and optimization errors through a nonlinear Strang-type estimate and then adaptively rebuilds the quadrature only when an online indicator exceeds a prescribed threshold (Badia et al., 1 May 2026). The quadrature is composite because the domain is partitioned into cells with local primal and reference rules, and adaptive because cells are bisected anisotropically according to directional fourth-difference surrogates. Numerical experiments report that this narrows the gap between training and reference losses and uses quadrature points more efficiently than non-adaptive strategies (Badia et al., 1 May 2026).

The surveyed literature also delineates a consistent set of limitations. Structured-grid support remains a bottleneck in some PDE implementations, with unstructured UGGrid explicitly described as future work in dune-composites (Butler et al., 2019). The 2D Poisson strip method assumes smooth boundaries and notes that corners require special handling or rounding (Fortunato et al., 29 Jan 2025). The adaptive accelerated proximal-gradient rule relies on accurate function evaluations and has no nonconvex convergence theory (Baghbadorani et al., 2024). The adaptive proximal ADMM assumes weak convexity of $845$1, convex block-separable $845$2, bounded product domains, and a Slater-like interior feasibility condition (Maia et al., 2024). Composite adaptive AMG can become memory-intensive as operator complexity grows or many components are added (Nelson et al., 10 Sep 2025). These limitations indicate that adaptive composite solvers are not a finished doctrine but an active design space in which composition, adaptivity, and implementation constraints remain tightly coupled.