{"id":"https://openalex.org/W4416138624","doi":"https://doi.org/10.1109/lsp.2026.3684818","title":"On the Convergence of Gradient Descent on Learning Transformers With Residual Connections","display_name":"On the Convergence of Gradient Descent on Learning Transformers With Residual Connections","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W4416138624","doi":"https://doi.org/10.1109/lsp.2026.3684818"},"language":"en","primary_location":{"id":"doi:10.1109/lsp.2026.3684818","is_oa":false,"landing_page_url":"https://doi.org/10.1109/lsp.2026.3684818","pdf_url":null,"source":{"id":"https://openalex.org/S120629676","display_name":"IEEE Signal Processing Letters","issn_l":"1070-9908","issn":["1070-9908","1558-2361"],"is_oa":false,"is_in_doaj":false,"is_core":true,"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"journal"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Signal Processing Letters","raw_type":"journal-article"},"type":"article","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2506.05249","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Zhen Qin","orcid":"https://orcid.org/0009-0000-0388-8840"},"institutions":[{"id":"https://openalex.org/I27837315","display_name":"University of Michigan","ror":"https://ror.org/00jmfr291","country_code":"US","type":"education","lineage":["https://openalex.org/I27837315"]},{"id":"https://openalex.org/I4210117557","display_name":"Michigan Center for Translational Pathology","ror":"https://ror.org/027t4kk23","country_code":"US","type":"facility","lineage":["https://openalex.org/I4210114445","https://openalex.org/I4210117557"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zhen Qin","raw_affiliation_strings":["Michigan Institute for Computational Discovery and Engineering, Department of Electrical Engineering and Computer Science, Department of Statistics, University of Michigan, Ann Arbor, MI, USA"],"raw_orcid":"https://orcid.org/0009-0000-0388-8840","affiliations":[{"raw_affiliation_string":"Michigan Institute for Computational Discovery and Engineering, Department of Electrical Engineering and Computer Science, Department of Statistics, University of Michigan, Ann Arbor, MI, USA","institution_ids":["https://openalex.org/I27837315","https://openalex.org/I4210117557"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5011391021","display_name":"Jin\u2010Xin Zhou","orcid":"https://orcid.org/0000-0002-8353-896X"},"institutions":[{"id":"https://openalex.org/I52357470","display_name":"The Ohio State University","ror":"https://ror.org/00rs6vg23","country_code":"US","type":"education","lineage":["https://openalex.org/I52357470"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Jinxin Zhou","raw_affiliation_strings":["Department of Computer Science and Engineering, the Ohio State University, Columbus, OH, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science and Engineering, the Ohio State University, Columbus, OH, USA","institution_ids":["https://openalex.org/I52357470"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Jiachen Jiang","orcid":null},"institutions":[{"id":"https://openalex.org/I52357470","display_name":"The Ohio State University","ror":"https://ror.org/00rs6vg23","country_code":"US","type":"education","lineage":["https://openalex.org/I52357470"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Jiachen Jiang","raw_affiliation_strings":["Department of Computer Science and Engineering, the Ohio State University, Columbus, OH, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science and Engineering, the Ohio State University, Columbus, OH, USA","institution_ids":["https://openalex.org/I52357470"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5011989964","display_name":"Zhihui Zhu","orcid":"https://orcid.org/0000-0002-3856-0375"},"institutions":[{"id":"https://openalex.org/I52357470","display_name":"The Ohio State University","ror":"https://ror.org/00rs6vg23","country_code":"US","type":"education","lineage":["https://openalex.org/I52357470"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zhihui Zhu","raw_affiliation_strings":["Department of Computer Science and Engineering, the Ohio State University, Columbus, OH, USA"],"raw_orcid":"https://orcid.org/0000-0002-3856-0375","affiliations":[{"raw_affiliation_string":"Department of Computer Science and Engineering, the Ohio State University, Columbus, OH, USA","institution_ids":["https://openalex.org/I52357470"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.01772366,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":"33","issue":null,"first_page":"1661","last_page":"1665"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T12611","display_name":"Neural Networks and Reservoir Computing","score":0.8758000135421753,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T12611","display_name":"Neural Networks and Reservoir Computing","score":0.8758000135421753,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.02879999950528145,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10502","display_name":"Advanced Memory and Neural Computing","score":0.01640000008046627,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/residual","display_name":"Residual","score":0.6917999982833862},{"id":"https://openalex.org/keywords/gradient-descent","display_name":"Gradient descent","score":0.5759999752044678},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.47589999437332153},{"id":"https://openalex.org/keywords/convergence","display_name":"Convergence (economics)","score":0.42239999771118164},{"id":"https://openalex.org/keywords/discretization","display_name":"Discretization","score":0.3894999921321869},{"id":"https://openalex.org/keywords/control-theory","display_name":"Control theory (sociology)","score":0.3682999908924103}],"concepts":[{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.6917999982833862},{"id":"https://openalex.org/C153258448","wikidata":"https://www.wikidata.org/wiki/Q1199743","display_name":"Gradient descent","level":3,"score":0.5759999752044678},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.4814999997615814},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.47589999437332153},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.461899995803833},{"id":"https://openalex.org/C2777303404","wikidata":"https://www.wikidata.org/wiki/Q759757","display_name":"Convergence (economics)","level":2,"score":0.42239999771118164},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4106999933719635},{"id":"https://openalex.org/C73000952","wikidata":"https://www.wikidata.org/wiki/Q17007827","display_name":"Discretization","level":2,"score":0.3894999921321869},{"id":"https://openalex.org/C47446073","wikidata":"https://www.wikidata.org/wiki/Q5165890","display_name":"Control theory (sociology)","level":3,"score":0.3682999908924103},{"id":"https://openalex.org/C57869625","wikidata":"https://www.wikidata.org/wiki/Q1783502","display_name":"Rate of convergence","level":3,"score":0.3675999939441681},{"id":"https://openalex.org/C38858127","wikidata":"https://www.wikidata.org/wiki/Q5441228","display_name":"Feed forward","level":2,"score":0.35530000925064087},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.34310001134872437},{"id":"https://openalex.org/C129844170","wikidata":"https://www.wikidata.org/wiki/Q41299","display_name":"Quadratic equation","level":2,"score":0.3095000088214874},{"id":"https://openalex.org/C192465680","wikidata":"https://www.wikidata.org/wiki/Q1413450","display_name":"Precondition","level":2,"score":0.3095000088214874},{"id":"https://openalex.org/C188441871","wikidata":"https://www.wikidata.org/wiki/Q7554146","display_name":"Softmax function","level":3,"score":0.2590999901294708}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.1109/lsp.2026.3684818","is_oa":false,"landing_page_url":"https://doi.org/10.1109/lsp.2026.3684818","pdf_url":null,"source":{"id":"https://openalex.org/S120629676","display_name":"IEEE Signal Processing Letters","issn_l":"1070-9908","issn":["1070-9908","1558-2361"],"is_oa":false,"is_in_doaj":false,"is_core":true,"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"journal"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Signal Processing Letters","raw_type":"journal-article"},{"id":"pmh:oai:arXiv.org:2506.05249","is_oa":true,"landing_page_url":"https://arxiv.org/abs/2506.05249","pdf_url":"https://arxiv.org/pdf/2506.05249","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2506.05249","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2506.05249","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2506.05249","is_oa":true,"landing_page_url":"https://arxiv.org/abs/2506.05249","pdf_url":"https://arxiv.org/pdf/2506.05249","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G2068745335","display_name":"Collaborative Research: Scalable, Robust, and Distributed Nonconvex Approaches for Structured Tensor Recovery","funder_award_id":"2409701","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G2674245822","display_name":null,"funder_award_id":"ECCS-2409701","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G2816040685","display_name":null,"funder_award_id":"2409701","funder_id":"https://openalex.org/F4320337392","funder_display_name":"Division of Electrical, Communications and Cyber Systems"},{"id":"https://openalex.org/G8406723238","display_name":"Collaborative Research: RI: Medium: Principled Approaches to Deep Learning for Low-dimensional Structures","funder_award_id":"2402952","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"}],"funders":[{"id":"https://openalex.org/F4320306076","display_name":"National Science Foundation","ror":"https://ror.org/021nxhr62"},{"id":"https://openalex.org/F4320337392","display_name":"Division of Electrical, Communications and Cyber Systems","ror":"https://ror.org/01krpsy48"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416138624.pdf","grobid_xml":"https://content.openalex.org/works/W4416138624.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Transformer":[0],"models":[1],"have":[2],"emerged":[3],"as":[4,45],"fundamental":[5],"tools":[6],"across":[7],"various":[8],"scientific":[9],"and":[10,48,90,115],"engineering":[11],"disciplines,":[12],"owing":[13],"to":[14,69],"their":[15,36],"outstanding":[16],"performance":[17],"in":[18,34,178],"diverse":[19],"applications.":[20],"Despite":[21],"this":[22,65,71],"empirical":[23],"success,":[24],"the":[25,53,75,107,113,120,124,143,146,153,158,172],"theoretical":[26,133,169],"foundations":[27],"of":[28,78,119,135,145,175],"Transformers":[29],"remain":[30],"relatively":[31],"underdeveloped,":[32],"particularly":[33],"understanding":[35],"training":[37],"dynamics.":[38],"Existing":[39],"research":[40],"predominantly":[41],"examines":[42],"isolated":[43],"components\u2013such":[44],"self-attention":[46],"mechanisms":[47],"feedforward":[49,88],"networks\u2013without":[50],"thoroughly":[51],"investigating":[52],"interdependencies":[54],"between":[55],"these":[56],"components,":[57],"especially":[58],"when":[59],"residual":[60,91,136,176],"connections":[61,137,177],"are":[62],"present.":[63],"In":[64],"paper,":[66],"we":[67],"aim":[68],"bridge":[70],"gap":[72],"by":[73,112,138,157],"analyzing":[74],"convergence":[76,104,108,129,180],"behavior":[77],"a":[79,87,102,132],"structurally":[80],"complete":[81],"yet":[82],"single-layer":[83],"Transformer,":[84],"comprising":[85],"self-attention,":[86],"network,":[89],"connections.":[92],"We":[93],"demonstrate":[94],"that,":[95],"under":[96],"appropriate":[97],"initialization,":[98],"gradient":[99],"descent":[100],"exhibits":[101],"linear":[103],"rate,":[105],"where":[106],"speed":[109],"is":[110],"determined":[111],"minimum":[114],"maximum":[116],"singular":[117],"values":[118],"output":[121,148],"matrix":[122],"from":[123,152],"attention":[125,147],"layer.":[126],"Moreover,":[127],"our":[128,168],"analysis":[130],"establishes":[131],"characterization":[134],"showing":[139],"that":[140],"they":[141],"alleviate":[142],"ill-conditioning":[144],"matrix,":[149],"which":[150],"arises":[151],"low-rank":[154],"structure":[155],"induced":[156],"softmax":[159],"operation,":[160],"thereby":[161],"improving":[162],"optimization":[163],"stability.":[164,181],"Empirical":[165],"results":[166],"corroborate":[167],"insights,":[170],"illustrating":[171],"beneficial":[173],"role":[174],"promoting":[179]},"counts_by_year":[],"updated_date":"2026-05-06T06:03:25.996018","created_date":"2025-10-10T00:00:00"}