{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2024:SYOJYXXJFRRYQWXZNVGQ4UDIVU","short_pith_number":"pith:SYOJYXXJ","canonical_record":{"source":{"id":"2409.20325","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-09-30T14:26:12Z","cross_cats_sorted":["math.OC"],"title_canon_sha256":"39abec0fc200fdc70fc498152835562a3c6f30b4dd3976d634a4756f4cd422dc","abstract_canon_sha256":"759bece6d884e94788ed87b467f6aaa1261bd538a1ddd102c6e57b729487309b"},"schema_version":"1.0"},"canonical_sha256":"961c9c5ee92c63885af96d4d0e5068ad3f1c72f9a17404056aa0b45be74b0b2e","source":{"kind":"arxiv","id":"2409.20325","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2409.20325","created_at":"2026-05-17T23:38:48Z"},{"alias_kind":"arxiv_version","alias_value":"2409.20325v2","created_at":"2026-05-17T23:38:48Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2409.20325","created_at":"2026-05-17T23:38:48Z"},{"alias_kind":"pith_short_12","alias_value":"SYOJYXXJFRRY","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"SYOJYXXJFRRYQWXZ","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"SYOJYXXJ","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2024:SYOJYXXJFRRYQWXZNVGQ4UDIVU","target":"record","payload":{"canonical_record":{"source":{"id":"2409.20325","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-09-30T14:26:12Z","cross_cats_sorted":["math.OC"],"title_canon_sha256":"39abec0fc200fdc70fc498152835562a3c6f30b4dd3976d634a4756f4cd422dc","abstract_canon_sha256":"759bece6d884e94788ed87b467f6aaa1261bd538a1ddd102c6e57b729487309b"},"schema_version":"1.0"},"canonical_sha256":"961c9c5ee92c63885af96d4d0e5068ad3f1c72f9a17404056aa0b45be74b0b2e","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:48.693124Z","signature_b64":"yJZ609WtGY9IG+Hfxy/JfnChz0hKYwfMyIcBzD3xXrHxodhpVbcRIEJyvZmDdls1cuG2cV/EFWpJDwUOVzdJAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"961c9c5ee92c63885af96d4d0e5068ad3f1c72f9a17404056aa0b45be74b0b2e","last_reissued_at":"2026-05-17T23:38:48.692561Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:48.692561Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2409.20325","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:48Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"kH1snxTsIeWRIJbM1g+qa31h3W6o3TAdOXkgDAZo9isA/UrP7aghjn8TKRxygdH6IzIc5+PTIyv5swMgm6XgBA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T13:56:04.451795Z"},"content_sha256":"998de4a0cbf692f84c1428ad6735832692852d68c3f214f681b4c4b536b6d291","schema_version":"1.0","event_id":"sha256:998de4a0cbf692f84c1428ad6735832692852d68c3f214f681b4c4b536b6d291"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2024:SYOJYXXJFRRYQWXZNVGQ4UDIVU","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Old Optimizer, New Norm: An Anthology","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Deep learning optimizers like Adam are equivalent to steepest descent under particular norms once exponential moving averages are disabled.","cross_cats":["math.OC"],"primary_cat":"cs.LG","authors_text":"Jeremy Bernstein, Laker Newhouse","submitted_at":"2024-09-30T14:26:12Z","abstract_excerpt":"Deep learning optimizers are often motivated through a mix of convex and approximate second-order theory. We select three such methods -- Adam, Shampoo and Prodigy -- and argue that each method can instead be understood as a squarely first-order method without convexity assumptions. In fact, after switching off exponential moving averages, each method is equivalent to steepest descent under a particular norm. By generalizing this observation, we chart a new design space for training algorithms. Different operator norms should be assigned to different tensors based on the role that the tensor p"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"after switching off exponential moving averages, each method is equivalent to steepest descent under a particular norm","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that disabling exponential moving averages preserves the essential behavior and that assigning role-specific norms will produce measurable gains in stability or speed without further assumptions on the loss surface","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Optimizers like Adam reduce to steepest descent under particular norms, opening a design space of norm assignments tailored to layer roles.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Deep learning optimizers like Adam are equivalent to steepest descent under particular norms once exponential moving averages are disabled.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"67bd05920e2854583f3b72ef19f96521f68f7a1b5fa4021f0ca943879404338e"},"source":{"id":"2409.20325","kind":"arxiv","version":2},"verdict":{"id":"482e00af-a874-40e2-afe1-cd1cfd799d1c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T07:23:30.832679Z","strongest_claim":"after switching off exponential moving averages, each method is equivalent to steepest descent under a particular norm","one_line_summary":"Optimizers like Adam reduce to steepest descent under particular norms, opening a design space of norm assignments tailored to layer roles.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that disabling exponential moving averages preserves the essential behavior and that assigning role-specific norms will produce measurable gains in stability or speed without further assumptions on the loss surface","pith_extraction_headline":"Deep learning optimizers like Adam are equivalent to steepest descent under particular norms once exponential moving averages are disabled."},"references":{"count":38,"sample":[{"doi":"","year":2002,"title":"Scalable second order optimization for deep learning","work_id":"9c7f025f-ec0e-4000-b1d0-db6e212787c5","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1966,"title":"Minimization of functions having Lipschitz continuous first partial derivatives","work_id":"3b1df2b4-f7e4-44ed-87ec-7b0f603ebd80","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Dissecting A dam: The sign, magnitude and variance of stochastic gradients","work_id":"7ff26b1a-62bb-4a8c-814a-3161e9c55ad1","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"sign SGD : Compressed optimisation for non-convex problems","work_id":"7a596119-b115-47ca-8c3d-0b1d5ed0c546","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"arXiv preprint arXiv:2304.05187 , year =","work_id":"df4c9a56-689f-4cd3-95f1-a8683f6cb5ce","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":38,"snapshot_sha256":"fc3047fa577390a1439c93ad4f2df7e30e911aca7d0a24b0bd1ba6d0353f534f","internal_anchors":1},"formal_canon":{"evidence_count":2,"snapshot_sha256":"281438f169f085058353704cc9100be78fe47066b3f564ee466401a2d949c30d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"482e00af-a874-40e2-afe1-cd1cfd799d1c"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:48Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"bir1Tm4uliAXEyBHVnE1g3rgnQD9wsFeUACIQsFBtraozFx3YNBZjtgUmL/UFIjN7Jj68CLQK5HmyPEuLwJFDg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T13:56:04.452907Z"},"content_sha256":"39b1bc49055d72e0537e53091bda4527d3d281d79e79c033607a5cfd60513a19","schema_version":"1.0","event_id":"sha256:39b1bc49055d72e0537e53091bda4527d3d281d79e79c033607a5cfd60513a19"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/SYOJYXXJFRRYQWXZNVGQ4UDIVU/bundle.json","state_url":"https://pith.science/pith/SYOJYXXJFRRYQWXZNVGQ4UDIVU/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/SYOJYXXJFRRYQWXZNVGQ4UDIVU/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-10T13:56:04Z","links":{"resolver":"https://pith.science/pith/SYOJYXXJFRRYQWXZNVGQ4UDIVU","bundle":"https://pith.science/pith/SYOJYXXJFRRYQWXZNVGQ4UDIVU/bundle.json","state":"https://pith.science/pith/SYOJYXXJFRRYQWXZNVGQ4UDIVU/state.json","well_known_bundle":"https://pith.science/.well-known/pith/SYOJYXXJFRRYQWXZNVGQ4UDIVU/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2024:SYOJYXXJFRRYQWXZNVGQ4UDIVU","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"759bece6d884e94788ed87b467f6aaa1261bd538a1ddd102c6e57b729487309b","cross_cats_sorted":["math.OC"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-09-30T14:26:12Z","title_canon_sha256":"39abec0fc200fdc70fc498152835562a3c6f30b4dd3976d634a4756f4cd422dc"},"schema_version":"1.0","source":{"id":"2409.20325","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2409.20325","created_at":"2026-05-17T23:38:48Z"},{"alias_kind":"arxiv_version","alias_value":"2409.20325v2","created_at":"2026-05-17T23:38:48Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2409.20325","created_at":"2026-05-17T23:38:48Z"},{"alias_kind":"pith_short_12","alias_value":"SYOJYXXJFRRY","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"SYOJYXXJFRRYQWXZ","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"SYOJYXXJ","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:39b1bc49055d72e0537e53091bda4527d3d281d79e79c033607a5cfd60513a19","target":"graph","created_at":"2026-05-17T23:38:48Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"after switching off exponential moving averages, each method is equivalent to steepest descent under a particular norm"},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"that disabling exponential moving averages preserves the essential behavior and that assigning role-specific norms will produce measurable gains in stability or speed without further assumptions on the loss surface"},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Optimizers like Adam reduce to steepest descent under particular norms, opening a design space of norm assignments tailored to layer roles."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Deep learning optimizers like Adam are equivalent to steepest descent under particular norms once exponential moving averages are disabled."}],"snapshot_sha256":"67bd05920e2854583f3b72ef19f96521f68f7a1b5fa4021f0ca943879404338e"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"281438f169f085058353704cc9100be78fe47066b3f564ee466401a2d949c30d"},"paper":{"abstract_excerpt":"Deep learning optimizers are often motivated through a mix of convex and approximate second-order theory. We select three such methods -- Adam, Shampoo and Prodigy -- and argue that each method can instead be understood as a squarely first-order method without convexity assumptions. In fact, after switching off exponential moving averages, each method is equivalent to steepest descent under a particular norm. By generalizing this observation, we chart a new design space for training algorithms. Different operator norms should be assigned to different tensors based on the role that the tensor p","authors_text":"Jeremy Bernstein, Laker Newhouse","cross_cats":["math.OC"],"headline":"Deep learning optimizers like Adam are equivalent to steepest descent under particular norms once exponential moving averages are disabled.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-09-30T14:26:12Z","title":"Old Optimizer, New Norm: An Anthology"},"references":{"count":38,"internal_anchors":1,"resolved_work":38,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Scalable second order optimization for deep learning","work_id":"9c7f025f-ec0e-4000-b1d0-db6e212787c5","year":2002},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Minimization of functions having Lipschitz continuous first partial derivatives","work_id":"3b1df2b4-f7e4-44ed-87ec-7b0f603ebd80","year":1966},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Dissecting A dam: The sign, magnitude and variance of stochastic gradients","work_id":"7ff26b1a-62bb-4a8c-814a-3161e9c55ad1","year":2018},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"sign SGD : Compressed optimisation for non-convex problems","work_id":"7a596119-b115-47ca-8c3d-0b1d5ed0c546","year":2018},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"arXiv preprint arXiv:2304.05187 , year =","work_id":"df4c9a56-689f-4cd3-95f1-a8683f6cb5ce","year":2023}],"snapshot_sha256":"fc3047fa577390a1439c93ad4f2df7e30e911aca7d0a24b0bd1ba6d0353f534f"},"source":{"id":"2409.20325","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-16T07:23:30.832679Z","id":"482e00af-a874-40e2-afe1-cd1cfd799d1c","model_set":{"reader":"grok-4.3"},"one_line_summary":"Optimizers like Adam reduce to steepest descent under particular norms, opening a design space of norm assignments tailored to layer roles.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Deep learning optimizers like Adam are equivalent to steepest descent under particular norms once exponential moving averages are disabled.","strongest_claim":"after switching off exponential moving averages, each method is equivalent to steepest descent under a particular norm","weakest_assumption":"that disabling exponential moving averages preserves the essential behavior and that assigning role-specific norms will produce measurable gains in stability or speed without further assumptions on the loss surface"}},"verdict_id":"482e00af-a874-40e2-afe1-cd1cfd799d1c"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:998de4a0cbf692f84c1428ad6735832692852d68c3f214f681b4c4b536b6d291","target":"record","created_at":"2026-05-17T23:38:48Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"759bece6d884e94788ed87b467f6aaa1261bd538a1ddd102c6e57b729487309b","cross_cats_sorted":["math.OC"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-09-30T14:26:12Z","title_canon_sha256":"39abec0fc200fdc70fc498152835562a3c6f30b4dd3976d634a4756f4cd422dc"},"schema_version":"1.0","source":{"id":"2409.20325","kind":"arxiv","version":2}},"canonical_sha256":"961c9c5ee92c63885af96d4d0e5068ad3f1c72f9a17404056aa0b45be74b0b2e","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"961c9c5ee92c63885af96d4d0e5068ad3f1c72f9a17404056aa0b45be74b0b2e","first_computed_at":"2026-05-17T23:38:48.692561Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:48.692561Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"yJZ609WtGY9IG+Hfxy/JfnChz0hKYwfMyIcBzD3xXrHxodhpVbcRIEJyvZmDdls1cuG2cV/EFWpJDwUOVzdJAw==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:48.693124Z","signed_message":"canonical_sha256_bytes"},"source_id":"2409.20325","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:998de4a0cbf692f84c1428ad6735832692852d68c3f214f681b4c4b536b6d291","sha256:39b1bc49055d72e0537e53091bda4527d3d281d79e79c033607a5cfd60513a19"],"state_sha256":"71ec1d025a7bb8a699e58417b36f6e278fb5a78d0435a84e92150b409a475e87"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"xIykyRJcxKlCwYzInVH7SwImhiH/0j3LMnzIcGzUTuZWQTju2KSgwjjEx7Xm6LZV2OTIGWmDa3ajfjDM1V4CDA==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-10T13:56:04.459002Z","bundle_sha256":"5b1290803aa5ff8bdd2dc42a57e24224479ac0f91a8b510e88cfb46c3bb0d4e3"}}