{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:ERTVQSU4UEYSWIOKMI3MD4SGUX","short_pith_number":"pith:ERTVQSU4","schema_version":"1.0","canonical_sha256":"2467584a9ca1312b21ca6236c1f246a5c27df86b7791ffb4854780c84b5b9294","source":{"kind":"arxiv","id":"2505.19227","version":1},"attestation_state":"computed","paper":{"title":"Scaling Laws for Gradient Descent and Sign Descent for Linear Bigram Models under Zipf's Law","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["math.OC"],"primary_cat":"cs.LG","authors_text":"Francis Bach, Frederik Kunstner","submitted_at":"2025-05-25T16:43:51Z","abstract_excerpt":"Recent works have highlighted optimization difficulties faced by gradient descent in training the first and last layers of transformer-based language models, which are overcome by optimizers such as Adam. These works suggest that the difficulty is linked to the heavy-tailed distribution of words in text data, where the frequency of the $k$th most frequent word $\\pi_k$ is proportional to $1/k$, following Zipf's law. To better understand the impact of the data distribution on training performance, we study a linear bigram model for next-token prediction when the tokens follow a power law $\\pi_k "},"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":"2505.19227","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-25T16:43:51Z","cross_cats_sorted":["math.OC"],"title_canon_sha256":"f2f4e9cef2198697e0e22f5c18f3ebef5b32e3b6d19691a6e686222c3faec435","abstract_canon_sha256":"47ac04ab56cac7634661dd5a1d28f3ffd04ed865b015a61fa74ce2315cef1e6b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:09:26.729223Z","signature_b64":"rQtS54YpUGKRFyukpar9cFCkc+aC2ZCPbit9cfOEYmWUIuTBG2gLCis2kfN5IyurUrCVHijlVg9vJeuWjX/8BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2467584a9ca1312b21ca6236c1f246a5c27df86b7791ffb4854780c84b5b9294","last_reissued_at":"2026-07-05T11:09:26.728726Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:09:26.728726Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Scaling Laws for Gradient Descent and Sign Descent for Linear Bigram Models under Zipf's Law","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["math.OC"],"primary_cat":"cs.LG","authors_text":"Francis Bach, Frederik Kunstner","submitted_at":"2025-05-25T16:43:51Z","abstract_excerpt":"Recent works have highlighted optimization difficulties faced by gradient descent in training the first and last layers of transformer-based language models, which are overcome by optimizers such as Adam. These works suggest that the difficulty is linked to the heavy-tailed distribution of words in text data, where the frequency of the $k$th most frequent word $\\pi_k$ is proportional to $1/k$, following Zipf's law. To better understand the impact of the data distribution on training performance, we study a linear bigram model for next-token prediction when the tokens follow a power law $\\pi_k "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.19227","kind":"arxiv","version":1},"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/2505.19227/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":"2505.19227","created_at":"2026-07-05T11:09:26.728787+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.19227v1","created_at":"2026-07-05T11:09:26.728787+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.19227","created_at":"2026-07-05T11:09:26.728787+00:00"},{"alias_kind":"pith_short_12","alias_value":"ERTVQSU4UEYS","created_at":"2026-07-05T11:09:26.728787+00:00"},{"alias_kind":"pith_short_16","alias_value":"ERTVQSU4UEYSWIOK","created_at":"2026-07-05T11:09:26.728787+00:00"},{"alias_kind":"pith_short_8","alias_value":"ERTVQSU4","created_at":"2026-07-05T11:09:26.728787+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2603.26554","citing_title":"Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11170","citing_title":"Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data","ref_index":35,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX","json":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX.json","graph_json":"https://pith.science/api/pith-number/ERTVQSU4UEYSWIOKMI3MD4SGUX/graph.json","events_json":"https://pith.science/api/pith-number/ERTVQSU4UEYSWIOKMI3MD4SGUX/events.json","paper":"https://pith.science/paper/ERTVQSU4"},"agent_actions":{"view_html":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX","download_json":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX.json","view_paper":"https://pith.science/paper/ERTVQSU4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.19227&json=true","fetch_graph":"https://pith.science/api/pith-number/ERTVQSU4UEYSWIOKMI3MD4SGUX/graph.json","fetch_events":"https://pith.science/api/pith-number/ERTVQSU4UEYSWIOKMI3MD4SGUX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX/action/storage_attestation","attest_author":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX/action/author_attestation","sign_citation":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX/action/citation_signature","submit_replication":"https://pith.science/pith/ERTVQSU4UEYSWIOKMI3MD4SGUX/action/replication_record"}},"created_at":"2026-07-05T11:09:26.728787+00:00","updated_at":"2026-07-05T11:09:26.728787+00:00"}