AI Paper Atlas

Research map / Learning theory and algorithms / Optimal transport

A Memory-Efficient Hierarchical Algorithm for Large-scale Optimal Transport Problems

Wenzhou Xia, Ya-Nan Zhu, Jingwei Liang, Xiaoqun Zhang

ICLR 2026 · Learning theory and algorithms › Optimal transport · approach: Sinkhorn and entropic OT algorithms

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add GPU PDHG solver cuPDLPx in place of Gurobi barrier (ablation)HALO · DOTmark r=256time (s): 159.06 -> 4.31Improved
add multiscale hierarchy (ablation)HALO · DOTmark r=64time (s): 128.4 -> 1.50Improved
add dual-violation correction to updateActive (ablation)HALO · DOTmark r=1024max time (s): 633.34 -> 154.20Improved
add dual-violation correction to updateActive (ablation)HALO · DOTmark r=1024avg time (s): 52.85 -> 27.73Improved
use constant stepsize via Pock-Chambolle rescaling instead of power iteration (ablation)HALO with cuPDLPx · DOTmark r=1024speedup: 1.97×Improved
raise shielding candidate budget Umax from 10 to 20 (ablation)HALO · ModelNet10 n=2^18time (s): 265.8 -> 229.7Improved
HALO versus M3S entropic solver at r=1024HALO · DOTmark r=1024time (s): 247.22 -> 27.73Improved
HALO versus HOT at r=512HALO · DOTmark r=512time (s): 78.43 -> 11.17Improved
HALO versus HiRef at n=2^18HALO · ModelNet10 n=2^18time (s): 422.7 -> 229.7Improved
HALO versus ShortCut at r=512HALO · DOTmark r=512time (s): 438.14 -> 11.17Improved
HALO versus ShortCut in memory at r=512HALO · DOTmark r=512peak memory (GB): 0.56 -> 2.07Hurt
sparser image classes (Shapes, Microscopy) versus dense ClassicImagesHALO · DOTmark r=1024time (s): 23.78 -> 44.39Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy