A Theory of Relaxation-Based Algebraic Multigrid

Abstract: Algebraic multigrid (AMG) methods derive their optimal efficiency from the interplay between a relaxation process and a corresponding coarse grid correction. In many standard formulations, relaxation and coarse-graining are analyzed and treated as largely separate of one another. Here we propose an alternative theoretical approach centered entirely on the relaxation process, which exposes its fundamental role in the coarse-graining of the fine-scale problem. By treating the relaxation of the error as a dynamical system and applying a dimensional-reduction procedure analogous to the Mori-Zwanzig-Nakajima formalism, we derive exact expressions for the coarse-level equations and the interpolation operations, as well as a natural way of computing complementary transfer operators. We illustrate the unifying nature of this framework by recovering several well-known results for general non-symmetric systems, including ideal and optimal restriction and interpolation, as well as the limiting case of exact elimination. We further emphasize the pivotal importance of compatible-relaxation and identify dynamical corrections that naturally arise in our theory, which have the potential to enhance the convergence, robustness, and adaptivity of future algebraic multigrid methods.

• The paper develops a relaxation-invariant AMG framework in which coarse variables, interpolation, and hierarchy construction depend on the error propagator and relaxation history rather than directly on the system matrix.
• The paper derives exact non-Markovian coarse equations whose memory terms arise from eliminating fine variables, while Markovian and semi-Markovian truncations recover Petrov–Galerkin AMG and compatible-relaxation-based convergence.
• The paper introduces a nonlinear prolongation flow whose fixed points are invariant subspaces of the relaxation operator, linking optimal transfer operators, energy-minimization AMG, and asymptotic convergence governed by the first excluded eigenvalue.

Motivation: an invariance principle for coarse-graining

The paper develops a theoretical framework for algebraic multigrid (AMG) in which the entire multilevel construction is derived from the relaxation process itself, rather than from the residual equation associated with the system matrix $A$. The motivating observation is a redundancy in conventional AMG formulations. Given the error relaxation dynamics $e^{(\ell+1)} = Te^{(\ell)}$ with error propagator $T = I - \widehat{A}$ and $\widehat{A} = MA$, fixing $x^{(0)}$ and $T$ while varying $A$ (with $M$ adjusted so that $MA$ remains invariant) leaves the error sequence $\{e^{(\ell)}\}$ unchanged. Yet standard AMG constructions, which coarsen $e^{(\ell+1)} = Te^{(\ell)}$0 directly, produce different transfer operators, coarse operators, and restricted residuals under this transformation. The authors argue that any genuinely relaxation-based formalism should be invariant under such changes, since the quantities being approximated—the coarse and fine components of the error—are identical.

This observation motivates coarse-graining the relaxation equation through $e^{(\ell+1)} = Te^{(\ell)}$1 and $e^{(\ell+1)} = Te^{(\ell)}$2 instead of the residual equation through $e^{(\ell+1)} = Te^{(\ell)}$3. All components of the resulting hierarchy depend exclusively on quantities intrinsic to relaxation: the propagator $e^{(\ell+1)} = Te^{(\ell)}$4 and the relaxation history $e^{(\ell+1)} = Te^{(\ell)}$5.

Coarse-graining as a dynamical reduction

The framework treats error relaxation as a discrete-time dynamical system on a spatio-temporal lattice $e^{(\ell+1)} = Te^{(\ell)}$6 of spatial degrees of freedom indexed by time slices. Under a locality assumption ($e^{(\ell+1)} = Te^{(\ell)}$7 sparse), information propagates forward in time along the directed graph $e^{(\ell+1)} = Te^{(\ell)}$8, and nodes on the same time slice are causally unrelated by relaxation alone. A shift relation,

$e^{(\ell+1)} = Te^{(\ell)}$9

allows error values at earlier iterations to be reconstructed from the known relaxation history once the error at coarse nodes on the latest slice is known. Recursively decomposing the fine-node error into contributions arriving via paths that terminate at coarse nodes versus paths traversing only fine nodes yields an interpolation relation whose weights involve effective long-range fine-fine couplings $T = I - \widehat{A}$0. Neglecting the fine-only path term requires that information propagating exclusively along fine-only paths decay rapidly—precisely the principle of compatible relaxation, here given a direct dynamical interpretation as decay of the habituated compatible-relaxation propagator $T = I - \widehat{A}$1.

In matrix form, with oblique projectors $T = I - \widehat{A}$2 and $T = I - \widehat{A}$3 splitting the error into coarse and fine components, the exact decomposition reads

$T = I - \widehat{A}$4

making explicit that the compatible-relaxation propagator governs both coarse-to-fine information flow and the sole unresolved contribution. The resulting coarse equation is non-Markovian:

$T = I - \widehat{A}$5

with generalized coarse operators $T = I - \widehat{A}$6 built from generalized prolongations $T = I - \widehat{A}$7. This structure parallels the Mori-Zwanzig-Nakajima projection formalism from non-equilibrium statistical mechanics: eliminating fine variables induces memory terms in the coarse evolution, and the discarded fine-scale contribution appears as a noise term $T = I - \widehat{A}$8 analogous to the Mori-Zwanzig fluctuating force. For $T = I - \widehat{A}$9, the authors verify directly that this coarse equation is equivalent to the coarse-grained relaxation equation, confirming its validity as a solvable coarse-level description.

A hierarchy of two-level schemes

The paper classifies idealized two-level constructions by how much dynamical information is retained, mirroring classifications in renormalization and memory-based problem reduction.

Markovian. Dropping all memory terms recovers a Petrov-Galerkin equation $\widehat{A} = MA$0, which reduces to the standard form with coarse operator $\widehat{A} = MA$1 upon defining $\widehat{A} = MA$2. The two-grid propagator takes the familiar form involving $\widehat{A} = MA$3 applied to $\widehat{A} = MA$4—the classical expression, but written in terms of $\widehat{A} = MA$5 as required by the invariance principle. Memory terms vanish identically when $\widehat{A} = MA$6 (transfer operators invariant under $\widehat{A} = MA$7) or when the $\widehat{A} = MA$8-orthogonality condition $\widehat{A} = MA$9 holds; under the latter, the noise vanishes and the coarse solve is exact.

Semi-Markovian. Retaining a Markovian coarse solve but reconstructing the full error history via the shift relation and substituting into the non-Markovian interpolation yields a method whose remaining error is exactly the compatible-relaxation component. Its two-grid propagator collapses to $x^{(0)}$0: convergence is governed entirely by habituated compatible relaxation. This is a notable structural result—it shows that combining an exact Petrov-Galerkin correction with memory-inclusive interpolation makes compatible-relaxation the sole determinant of convergence.

Non-Markovian. Relaxing $x^{(0)}$1 produces a corrected coarse equation $x^{(0)}$2 with an explicitly computable memory correction $x^{(0)}$3 derived from the relaxation history. The propagator resembles a Petrov-Galerkin scheme with effective prolongation $x^{(0)}$4 and the compatible-relaxation propagator in place of $x^{(0)}$5. Unlike the Markovian scheme, both semi- and non-Markovian variants can be rendered exactly convergent by choosing duals $x^{(0)}$6-orthogonal to the basis vectors ($x^{(0)}$7), without special structure in the coarse-fine split.

Exact. Resolving the compatible-relaxation term explicitly via a geometric-series inversion gives an interpolation law and coarse equation that solve the problem exactly for any choice of transfer operators, at the cost of global inversions involving $x^{(0)}$8 and $x^{(0)}$9. The coarse operator simplifies to Petrov-Galerkin form with Schur-complement-like effective transfers $T$0 and $T$1. While impractical algorithmically, this construction traces a continuous spectrum from plain relaxation through progressively richer truncations to a one-iteration direct solver; conventional multigrid is identified as the lowest-order (Markovian) truncation of this hierarchy.

Transfer operators from a flow on prolongation

The non-Markovian formulas suggest deriving transfer operators iteratively rather than prescribing them. Starting from canonical unit basis vectors, one application of the update produces $T$2 with $T$3, which converges as $T$4 (assuming $T$5) to the generalization of ideal interpolation $T$6, reducing to the standard $T$7 when $T$8 is block upper-triangular. The corresponding ideal restriction follows from the orthogonality condition $T$9. Both involve dense inverses and must be localized or sparsified in practice, yielding approximate ideal restriction/interpolation as used in practical schemes.

The general recursive update—the flow equation—is

$A$0

readable as $A$1 steps of relaxation applied to $A$2 using the projected propagator $A$3—in contrast to smoothed aggregation, which relaxes tentative interpolants with $A$4 itself. The flow is nonlinear (the projector depends on the current iterate), but its increments are always $A$5-orthogonal, so the dual can be held stationary across the flow. At a fixed point, $A$6, meaning $A$7 is a right $A$8-invariant subspace. In the $A$9 limit the flow becomes $M$0, projecting onto the $M$1-orthogonal complement of $M$2; when $M$3 is Hermitian and columns properly orthonormalized, column energies $M$4 are non-increasing, connecting the flow to energy-minimization AMG.

Because fixed points are invariant subspaces of minimal energy, the flow should converge to the optimal interpolation subspace spanned by the eigenvectors of $M$5 of largest magnitude. With biorthogonal eigenbasis splittings, the fixed-point operators yield $M$6, reducing the framework to the semi-Markovian and then Markovian cases, and the two-grid propagator becomes $M$7: the asymptotic convergence factor is governed solely by $M$8, the dominant retained eigenvalue. The paper notes that no further improvement is possible from memory terms at optimality, since $M$9 eliminates them—consistent with existing optimal-transfer-operator theory for nonsymmetric and indefinite problems.

Limitations and open questions

The paper is explicitly theoretical: numerical validation is deferred to a companion manuscript in preparation, so all convergence claims rest on formal derivations and limiting arguments rather than demonstrated performance. Several constructions depend on assumptions stated but not guaranteed: the ideal-interpolation limit requires $MA$0 (relaxation convergent on the fine block); neglect of the noise term requires sufficiently fast compatible-relaxation decay, which rests on appropriate choices of basis vectors, duals, and the coarse-fine split; and identification of flow fixed points with minimal-energy invariant subspaces is asserted rather than proven in full generality. The exact scheme requires global inversions violating the locality constraint needed for $MA$1 complexity, and the practical cost-accuracy trade-off of retaining memory terms at intermediate truncation orders is left unquantified. How to exploit memory effects efficiently for symmetric and nonsymmetric problems, and how the framework extends to multiple levels beyond the two-level analysis presented, remain open.

Conclusion

The paper reformulates AMG theory around the relaxation dynamics itself, establishing a dynamical-invariance principle under which the multilevel hierarchy depends only on $MA$2 and the relaxation history. Via a Mori-Zwanzig-type dimensional reduction, it derives exact non-Markovian coarse equations and interpolation laws whose truncations recover, within a single framework, Petrov-Galerkin AMG, compatible relaxation, smoothed aggregation, energy minimization, and ideal/optimal restriction and interpolation for general nonsymmetric systems. The principal conceptual contributions are the interpretation of fine-grid elimination as a source of non-Markovian memory, the demonstration that memory-inclusive corrections can render two-level methods exact for arbitrary transfer operators, and a nonlinear flow on prolongation spaces whose fixed points are $MA$3-invariant subspaces of minimal energy. Whether these dynamical corrections translate into practical gains in robustness and adaptivity is the question the companion numerical study is intended to answer.