Decentralized First-Order Methods Under Hölder Continuous Gradients: A Numerical Study on a Reaction-Diffusion Equation and Multi-Slice Sparse Reconstruction

Fixed-step gradient methods for objectives whose gradient is Hölder but not Lipschitz continuous stagnate at a step-dependent error level, and the standard analyses of exact decentralized methods such as gradient tracking assume Lipschitz gradients. This paper is a numerical study of what decentralized first-order methods do in that regime, on two small test problems: a two-dimensional steady reaction-diffusion equation with a fractional reaction order and a dead core (121 unknowns and 4 workers in the main runs, up to 2209 unknowns and 16 workers in a scaling study), and multi-slice sparse reconstruction with a shared/private image model (four 64×64 synthetic slices, plus eight 128×128 slices of a real brain template). On the reaction-diffusion problem, gradient tracking needs 828 iterations to reach a relative error of 10−3 against 67 for forward Euler and 12 for backward Euler, and, once the backward-Euler steps are solved accurately, it is also far less accurate than backward Euler. Forward Euler distributed by strips needs only neighbour exchange as well, so on this problem gradient tracking offers no advantage; the problem serves as a controlled test of its behaviour under Hölder gradients. In reconstruction, the decentralized solution differs from the pooled-data solution by 0.050±0.006 dB over 25 synthetic instances, and the shared/private model gains 2.46±0.09 dB over per-slice reconstruction; the gain survives smooth anatomical drift and moderate intensity variation but is lost under a one-pixel misregistration. With a fixed step, the disagreement between workers does not vanish: it settles on a floor that persists without the private variables and with them frozen, disappears at the Lipschitz exponent and under a smoothed penalty, and follows a power of the step size whose fitted exponent (1.61±0.05 with the private step fixed, 1.50 with it proportional) depends on the step protocol. A proximal form of gradient tracking removes the floor.

Authors

Institutions

Publication Details

Journal
Algorithms
Published
2026-11-18
DOI
https://doi.org/10.3390/a19111015
Primary Topic
Distributed Control Multi-Agent Systems
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Decentralized First-Order Methods Under Hölder Continuous Gradients: A Numerical Study on a Reaction-Diffusion Equation and Multi-Slice Sparse Reconstruction

Yuxuan Song
1 links
Algorithms
Distributed Control Multi-Agent Systems
article

Decentralized First-Order Methods Under Hölder Continuous Gradients: A Numerical Study on a Reaction-Diffusion Equation and Multi-Slice Sparse Reconstruction

Yuxuan Song
article en

Abstract

Fixed-step gradient methods for objectives whose gradient is Hölder but not Lipschitz continuous stagnate at a step-dependent error level, and the standard analyses of exact decentralized methods such as gradient tracking assume Lipschitz gradients. This paper is a numerical study of what decentralized first-order methods do in that regime, on two small test problems: a two-dimensional steady reaction-diffusion equation with a fractional reaction order and a dead core (121 unknowns and 4 workers in the main runs, up to 2209 unknowns and 16 workers in a scaling study), and multi-slice sparse reconstruction with a shared/private image model (four 64×64 synthetic slices, plus eight 128×128 slices of a real brain template). On the reaction-diffusion problem, gradient tracking needs 828 iterations to reach a relative error of 10−3 against 67 for forward Euler and 12 for backward Euler, and, once the backward-Euler steps are solved accurately, it is also far less accurate than backward Euler. Forward Euler distributed by strips needs only neighbour exchange as well, so on this problem gradient tracking offers no advantage; the problem serves as a controlled test of its behaviour under Hölder gradients. In reconstruction, the decentralized solution differs from the pooled-data solution by 0.050±0.006 dB over 25 synthetic instances, and the shared/private model gains 2.46±0.09 dB over per-slice reconstruction; the gain survives smooth anatomical drift and moderate intensity variation but is lost under a one-pixel misregistration. With a fixed step, the disagreement between workers does not vanish: it settles on a floor that persists without the private variables and with them frozen, disappears at the Lipschitz exponent and under a smoothed penalty, and follows a power of the step size whose fitted exponent (1.61±0.05 with the private step fixed, 1.50 with it proportional) depends on the step protocol. A proximal form of gradient tracking removes the floor.

AlgorithmsVol. 19(11)
Hong Kong Polytechnic University (HK)
Distributed Control Multi-Agent Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon