{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6AYKRCMXVUEZOSOVLFDHYBL3DU","short_pith_number":"pith:6AYKRCMX","schema_version":"1.0","canonical_sha256":"f030a88997ad099749d559467c057b1d0dedaf5811a02693d0305a2d45fbccdd","source":{"kind":"arxiv","id":"2503.12645","version":2},"attestation_state":"computed","paper":{"title":"Understanding Gradient Orthogonalization for Deep Learning via Non-Euclidean Trust-Region Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Dmitry Kovalev","submitted_at":"2025-03-16T20:49:34Z","abstract_excerpt":"Optimization with matrix gradient orthogonalization has recently demonstrated impressive results in the training of deep neural networks (Jordan et al., 2024; Liu et al., 2025). In this paper, we provide a theoretical analysis of this approach. In particular, we show that the orthogonalized gradient method can be seen as a first-order trust-region optimization method, where the trust-region is defined in terms of the matrix spectral norm. Motivated by this observation, we develop the stochastic non-Euclidean trust-region gradient method with momentum, which recovers the Muon optimizer (Jordan "},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2503.12645","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-03-16T20:49:34Z","cross_cats_sorted":["math.OC","stat.ML"],"title_canon_sha256":"02f556139f0ea6494085dee48aed5eb68d8a163f387d77b3ce8c83eccc61c55f","abstract_canon_sha256":"31402d2ae3ca535d26562b263de46ed4211d596da4d05f7cda2dc4d783f213d9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:46:02.376960Z","signature_b64":"8ySP3u1h7d42c7A2pizo8mIPZBNqtoFo/nr4zcPRlN4VtEvaN2nxWRU0Vvrf4wXdJLcBI2doBrXZML2nNGzsDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f030a88997ad099749d559467c057b1d0dedaf5811a02693d0305a2d45fbccdd","last_reissued_at":"2026-07-05T10:46:02.376500Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:46:02.376500Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding Gradient Orthogonalization for Deep Learning via Non-Euclidean Trust-Region Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Dmitry Kovalev","submitted_at":"2025-03-16T20:49:34Z","abstract_excerpt":"Optimization with matrix gradient orthogonalization has recently demonstrated impressive results in the training of deep neural networks (Jordan et al., 2024; Liu et al., 2025). In this paper, we provide a theoretical analysis of this approach. In particular, we show that the orthogonalized gradient method can be seen as a first-order trust-region optimization method, where the trust-region is defined in terms of the matrix spectral norm. Motivated by this observation, we develop the stochastic non-Euclidean trust-region gradient method with momentum, which recovers the Muon optimizer (Jordan "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.12645","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2503.12645/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2503.12645","created_at":"2026-07-05T10:46:02.376558+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.12645v2","created_at":"2026-07-05T10:46:02.376558+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.12645","created_at":"2026-07-05T10:46:02.376558+00:00"},{"alias_kind":"pith_short_12","alias_value":"6AYKRCMXVUEZ","created_at":"2026-07-05T10:46:02.376558+00:00"},{"alias_kind":"pith_short_16","alias_value":"6AYKRCMXVUEZOSOV","created_at":"2026-07-05T10:46:02.376558+00:00"},{"alias_kind":"pith_short_8","alias_value":"6AYKRCMX","created_at":"2026-07-05T10:46:02.376558+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":27,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.21581","citing_title":"Convergence Analysis of Muon-type Methods with Inexact LMO in the Degenerate Case","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2606.21354","citing_title":"Restart and Adaptive Acceleration in Stochastic Gradient Methods","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01487","citing_title":"How to Allocate Your Tokens? Scaling Laws with Training Steps and Batch Size","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09658","citing_title":"Muon Learns More Robust and Transferable Features than Adam","ref_index":80,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01124","citing_title":"Muon as a Residual Connection","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06418","citing_title":"Double Preconditioning (DoPr): Optimization for Test-Time Performance, not Validation Loss","ref_index":174,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04662","citing_title":"Why Muon Outperforms Adam: A Curvature Perspective","ref_index":158,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03899","citing_title":"Denoise First, Orthogonalize Later: Understanding Momentum in Muon via Spectral Filtering","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18528","citing_title":"Scale-Invariant Neural Network Optimization: Norm Geometry and Heavy-Tailed Noise","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18106","citing_title":"Symmetry-Compatible Principle for Optimizer Design: Embeddings, LM Heads, SwiGLU MLPs, and MoE Routers","ref_index":85,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30634","citing_title":"One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23871","citing_title":"Move on Muon : A Hamiltonian probability gradient flow perspective of Muon optimizer","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20866","citing_title":"LOSCAR-SGD: Local SGD with Communication-Computation Overlap and Delay-Corrected Sparse Model Averaging","ref_index":292,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18174","citing_title":"Ringmaster LMO: Asynchronous Linear Minimization Oracle Momentum Method","ref_index":291,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18106","citing_title":"Symmetry-Compatible Principle for Optimizer Design: Embeddings, LM Heads, SwiGLU MLPs, and MoE Routers","ref_index":83,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19619","citing_title":"MiMuon: Mixed Muon Optimizer with Improved Generalization for Large Models","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2505.23737","citing_title":"On the Convergence Analysis of Muon","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2509.11983","citing_title":"Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2602.21545","citing_title":"MUON+: Towards More Effective Muon via One Additional Normalization Step for LLM Pre-training","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.13079","citing_title":"Spectral Flattening Is All Muon Needs: How Orthogonalization Controls Learning Rate and Convergence","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02505","citing_title":"Optimal Projection-Free Adaptive SGD for Matrix Optimization","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11850","citing_title":"Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11181","citing_title":"Muon is Not That Special: Random or Inverted Spectra Work Just as Well","ref_index":40,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10468","citing_title":"Can Muon Fine-tune Adam-Pretrained Models?","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09331","citing_title":"Dimension-Free Saddle-Point Escape in Muon","ref_index":19,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU","json":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU.json","graph_json":"https://pith.science/api/pith-number/6AYKRCMXVUEZOSOVLFDHYBL3DU/graph.json","events_json":"https://pith.science/api/pith-number/6AYKRCMXVUEZOSOVLFDHYBL3DU/events.json","paper":"https://pith.science/paper/6AYKRCMX"},"agent_actions":{"view_html":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU","download_json":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU.json","view_paper":"https://pith.science/paper/6AYKRCMX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.12645&json=true","fetch_graph":"https://pith.science/api/pith-number/6AYKRCMXVUEZOSOVLFDHYBL3DU/graph.json","fetch_events":"https://pith.science/api/pith-number/6AYKRCMXVUEZOSOVLFDHYBL3DU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU/action/storage_attestation","attest_author":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU/action/author_attestation","sign_citation":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU/action/citation_signature","submit_replication":"https://pith.science/pith/6AYKRCMXVUEZOSOVLFDHYBL3DU/action/replication_record"}},"created_at":"2026-07-05T10:46:02.376558+00:00","updated_at":"2026-07-05T10:46:02.376558+00:00"}