Abstract:Graph-Laplacian systems $L\phi=b$ underlie spectral clustering, semi-supervised learning, finite-element analysis, and network-flow solvers. We present LAMG+, a lean, parameter-free, empirically linear-time algebraic multigrid solver: a Julia re-derivation of Lean Algebraic Multigrid (LAMG) with two targeted refinements. We establish three facts. (1) Benchmarking against approximate-Cholesky (AC) and four other solvers (BoomerAMG, PETSc GAMG, pyAMG, CMG): LAMG+ and AC are complementary peers -- AC is faster on social/citation graphs; LAMG+ is faster on finite-element/structural matrices (fastest robust solver, most memory-frugal, $2.2\times$ faster than the robust AC variant on large graphs). Only LAMG+ and AC converge across all 13 test classes; the others fail or slow by an order of magnitude off their home turf. (2) Linear scaling: LAMG+ is empirically $O(m)$ with $m$ nonzeros over the full 1,711-graph SuiteSparse set (100% converged, median 4 cycles, log-log slope 1.01), verified up to $2.4\times 10^8$ nonzeros. (3) Robustness: prior benchmarking reported LAMG non-convergent on certain families; running the unmodified LAMG 2.2.1 under identical conditions establishes full convergence, indicating an evaluation artifact. A Local Fourier Analysis proves a strict interpolation-order deficit on grid-aligned anisotropy. Two lean local refinements -- a strength-of-connection aggregation veto and selective caliber-2 interpolation -- resolve LAMG's anisotropy failure (convergence factor $\approx 0.99 \to 0.11$) with negligible overhead.
From: Oren Livne [view email]
[v1]
Tue, 23 Jun 2026 16:46:38 UTC (188 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。