{"paper":{"title":"Rescaled Asynchronous SGD: Optimal Distributed Optimization under Data and System Heterogeneity","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Rescaling worker stepsizes by computation time fixes bias in asynchronous SGD so it converges to the true global objective.","cross_cats":["cs.DC","math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Ammar Mahran, Artavazd Maranjyan, Peter Richt\\'arik","submitted_at":"2026-05-13T12:27:22Z","abstract_excerpt":"Asynchronous stochastic gradient descent (ASGD) is a standard way to exploit heterogeneous compute resources in distributed learning: instead of forcing fast workers to wait for slow ones, the server updates the model whenever a gradient arrives. Vanilla ASGD applies each arriving gradient with the same weight. When local data distributions are heterogeneous, this becomes problematic: faster workers contribute more updates, and we show theoretically that the method is biased toward a frequency-weighted average of the local objectives rather than the desired global objective. Existing remedies "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"we prove that the resulting method, Rescaled ASGD, converges to stationary points of the correct global objective in the fixed-computation model. Its time complexity matches the known lower bound in the leading term, while the effects of staleness and data heterogeneity appear only in lower-order terms.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"under smoothness and bounded heterogeneity assumptions","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Rescaled ASGD recovers convergence to the true global objective by rescaling worker stepsizes proportional to computation times, matching the known time lower bound in the leading term under non-convex smoothness and bounded heterogeneity.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Rescaling worker stepsizes by computation time fixes bias in asynchronous SGD so it converges to the true global objective.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"bed43d911b513c3ba7be443a98374a6b8d93807a454477bcce4e5a28607203ab"},"source":{"id":"2605.13434","kind":"arxiv","version":1},"verdict":{"id":"3df6c769-5d61-42f4-b85b-e2d8b0175c8e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:27:27.645909Z","strongest_claim":"we prove that the resulting method, Rescaled ASGD, converges to stationary points of the correct global objective in the fixed-computation model. Its time complexity matches the known lower bound in the leading term, while the effects of staleness and data heterogeneity appear only in lower-order terms.","one_line_summary":"Rescaled ASGD recovers convergence to the true global objective by rescaling worker stepsizes proportional to computation times, matching the known time lower bound in the leading term under non-convex smoothness and bounded heterogeneity.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"under smoothness and bounded heterogeneity assumptions","pith_extraction_headline":"Rescaling worker stepsizes by computation time fixes bias in asynchronous SGD so it converges to the true global objective."},"references":{"count":300,"sample":[{"doi":"10.1109/msp.2020.2975749","year":2020,"title":"Federated Learning: Chal- lenges, Methods, and Future Directions.IEEE Signal Processing Magazine, 37(3):50–60","work_id":"4916403b-b4d2-45a0-9053-7b8f137c878f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.48550/arxiv.1909.06335","year":1909,"title":"Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification","work_id":"1e8d2981-bd90-4c86-8676-223498b6d816","ref_index":2,"cited_arxiv_id":"1909.06335","is_internal_anchor":true},{"doi":"10.1109/tit.2017.2736066","year":2015,"title":"arXiv.org , author =","work_id":"42d7abc2-f71a-4d63-885a-c432425462ef","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.48550/arxiv.1812.01097","year":null,"title":"Leaf: A benchmark for federated settings","work_id":"94ec8be1-6cf1-4962-994c-e762a0b9c3ee","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2012,"title":"arXiv.org , author =","work_id":"a670dcca-8345-441d-9855-f6bc678e0fbd","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":300,"snapshot_sha256":"421548bf02e07daa14540f2b965d273433a08dbfcc992d1b513da02c06f42655","internal_anchors":34},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}