{"id":"https://openalex.org/W7162000890","doi":"https://doi.org/10.48550/arxiv.2605.20271","title":"Multi-Head Attention as Ensemble Nadaraya-Watson Estimation: Variance Reduction, Decorrelation, and Optimal Head Diversity","display_name":"Multi-Head Attention as Ensemble Nadaraya-Watson Estimation: Variance Reduction, Decorrelation, and Optimal Head Diversity","publication_year":2026,"publication_date":"2026-05-18","ids":{"openalex":"https://openalex.org/W7162000890","doi":"https://doi.org/10.48550/arxiv.2605.20271"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.20271","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.20271","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.20271","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5136643930","display_name":"Ernest Fokou\u00e9","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Fokou\u00e9, Ernest","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5136643930"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10429","display_name":"EEG and Brain-Computer Interfaces","score":0.16110000014305115,"subfield":{"id":"https://openalex.org/subfields/2805","display_name":"Cognitive Neuroscience"},"field":{"id":"https://openalex.org/fields/28","display_name":"Neuroscience"},"domain":{"id":"https://openalex.org/domains/1","display_name":"Life Sciences"}},"topics":[{"id":"https://openalex.org/T10429","display_name":"EEG and Brain-Computer Interfaces","score":0.16110000014305115,"subfield":{"id":"https://openalex.org/subfields/2805","display_name":"Cognitive Neuroscience"},"field":{"id":"https://openalex.org/fields/28","display_name":"Neuroscience"},"domain":{"id":"https://openalex.org/domains/1","display_name":"Life Sciences"}},{"id":"https://openalex.org/T10241","display_name":"Functional Brain Connectivity Studies","score":0.13529999554157257,"subfield":{"id":"https://openalex.org/subfields/2805","display_name":"Cognitive Neuroscience"},"field":{"id":"https://openalex.org/fields/28","display_name":"Neuroscience"},"domain":{"id":"https://openalex.org/domains/1","display_name":"Life Sciences"}},{"id":"https://openalex.org/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.09359999746084213,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/projection","display_name":"Projection (relational algebra)","score":0.5498999953269958},{"id":"https://openalex.org/keywords/decorrelation","display_name":"Decorrelation","score":0.43549999594688416},{"id":"https://openalex.org/keywords/variance","display_name":"Variance (accounting)","score":0.3928999900817871},{"id":"https://openalex.org/keywords/dimension","display_name":"Dimension (graph theory)","score":0.3785000145435333},{"id":"https://openalex.org/keywords/measure","display_name":"Measure (data warehouse)","score":0.3772999942302704},{"id":"https://openalex.org/keywords/subspace-topology","display_name":"Subspace topology","score":0.36880001425743103},{"id":"https://openalex.org/keywords/parameterized-complexity","display_name":"Parameterized complexity","score":0.3605000078678131},{"id":"https://openalex.org/keywords/covariance","display_name":"Covariance","score":0.3564000129699707},{"id":"https://openalex.org/keywords/kernel","display_name":"Kernel (algebra)","score":0.35120001435279846}],"concepts":[{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.7610999941825867},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.5498999953269958},{"id":"https://openalex.org/C177860922","wikidata":"https://www.wikidata.org/wiki/Q788608","display_name":"Decorrelation","level":2,"score":0.43549999594688416},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.3928999900817871},{"id":"https://openalex.org/C33676613","wikidata":"https://www.wikidata.org/wiki/Q13415176","display_name":"Dimension (graph theory)","level":2,"score":0.3785000145435333},{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.3772999942302704},{"id":"https://openalex.org/C32834561","wikidata":"https://www.wikidata.org/wiki/Q660730","display_name":"Subspace topology","level":2,"score":0.36880001425743103},{"id":"https://openalex.org/C165464430","wikidata":"https://www.wikidata.org/wiki/Q1570441","display_name":"Parameterized complexity","level":2,"score":0.3605000078678131},{"id":"https://openalex.org/C178650346","wikidata":"https://www.wikidata.org/wiki/Q201984","display_name":"Covariance","level":2,"score":0.3564000129699707},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.35120001435279846},{"id":"https://openalex.org/C139945424","wikidata":"https://www.wikidata.org/wiki/Q1940696","display_name":"Mean squared error","level":2,"score":0.33649998903274536},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.33570000529289246},{"id":"https://openalex.org/C183115368","wikidata":"https://www.wikidata.org/wiki/Q856577","display_name":"Weighting","level":2,"score":0.3321000039577484},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.3312000036239624},{"id":"https://openalex.org/C83546350","wikidata":"https://www.wikidata.org/wiki/Q1139051","display_name":"Regression","level":2,"score":0.31769999861717224},{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.3140999972820282},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.3089999854564667},{"id":"https://openalex.org/C207390915","wikidata":"https://www.wikidata.org/wiki/Q1230525","display_name":"Divergence (linguistics)","level":2,"score":0.3052000105381012},{"id":"https://openalex.org/C2776036281","wikidata":"https://www.wikidata.org/wiki/Q48769818","display_name":"Constraint (computer-aided design)","level":2,"score":0.3034999966621399},{"id":"https://openalex.org/C62100291","wikidata":"https://www.wikidata.org/wiki/Q1936288","display_name":"Total variation","level":2,"score":0.3003000020980835},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.2919999957084656},{"id":"https://openalex.org/C185429906","wikidata":"https://www.wikidata.org/wiki/Q1130160","display_name":"Estimator","level":2,"score":0.2854999899864197},{"id":"https://openalex.org/C31170391","wikidata":"https://www.wikidata.org/wiki/Q188619","display_name":"Hierarchy","level":2,"score":0.27950000762939453},{"id":"https://openalex.org/C152877465","wikidata":"https://www.wikidata.org/wiki/Q208042","display_name":"Regression analysis","level":2,"score":0.27790001034736633},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.2777000069618225},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.2648000121116638},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.2628999948501587},{"id":"https://openalex.org/C2778023678","wikidata":"https://www.wikidata.org/wiki/Q554403","display_name":"Duality (order theory)","level":2,"score":0.26190000772476196},{"id":"https://openalex.org/C115311070","wikidata":"https://www.wikidata.org/wiki/Q5347255","display_name":"Effective dimension","level":3,"score":0.26010000705718994},{"id":"https://openalex.org/C39927690","wikidata":"https://www.wikidata.org/wiki/Q11197","display_name":"Logarithm","level":2,"score":0.250900000333786},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.25}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.20271","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.20271","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.20271","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.20271","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"We":[0,56,110],"develop":[1],"a":[2,37,47,117,152,183,249],"rigorous":[3,140],"statistical":[4],"theory":[5,222,229],"of":[6,13,40,52,62,78,86,121,148,202,217,223,248],"multi-head":[7],"attention":[8,27,149,243],"(MHA)":[9],"as":[10],"an":[11,58],"ensemble":[12,39,231],"Nadaraya-Watson":[14],"(NW)":[15],"kernel":[16],"regression":[17,178],"estimators.":[18],"Building":[19],"on":[20,75,83],"the":[21,29,53,63,76,84,93,112,138,144,163,169,188,199,208,220,226,234,245],"algebraic":[22],"identity":[23],"between":[24,96,237],"single-head":[25,224],"softmax":[26],"and":[28,124,177,233,239],"NW":[30,42,221],"estimator,":[31],"we":[32,161],"prove":[33,125],"that":[34,69,126],"MHA":[35,64,127],"is":[36,90,131,244],"structured":[38],"H":[41,80,158],"estimators,":[43],"each":[44],"operating":[45],"in":[46,134],"distinct":[48],"learned":[49,97],"projection":[50,98],"subspace":[51],"key":[54],"space.":[55],"derive":[57],"explicit":[59],"Bias-Variance-Covariance":[60],"decomposition":[61],"mean":[65,128],"squared":[66,129],"error,":[67],"showing":[68],"variance":[70,104],"reduction":[71],"depends":[72],"not":[73],"merely":[74],"number":[77,201],"heads":[79,203],"but":[81],"fundamentally":[82],"decorrelation":[85],"head":[87],"outputs.":[88],"Decorrelation":[89],"governed":[91],"by":[92],"principal":[94],"angles":[95],"subspaces:":[99],"orthogonal":[100],"projections":[101,107],"yield":[102,108],"maximum":[103],"reduction;":[105],"aligned":[106],"none.":[109],"introduce":[111],"Head":[113],"Diversity":[114],"Index":[115],"(HDI),":[116],"computable":[118],"spectral":[119],"measure":[120],"inter-head":[122],"decorrelation,":[123],"error":[130],"monotonically":[132],"decreasing":[133],"HDI.":[135],"This":[136],"provides":[137],"first":[139],"theoretical":[141],"explanation":[142],"for":[143,230],"empirically":[145],"observed":[146],"specialization":[147],"heads.":[150],"Under":[151],"fixed":[153],"total-dimension":[154],"budget":[155,210],"D":[156],"=":[157],"*":[159],"d_k,":[160],"solve":[162],"optimal":[164,189,200],"head-dimension":[165],"allocation":[166],"problem,":[167],"deriving":[168],"MSE-minimizing":[170],"pair":[171],"(H*,":[172],"d_k*)":[173],"from":[174],"data":[175],"distribution":[176],"smoothness.":[179],"The":[180],"solution":[181],"yields":[182,257],"new":[184],"architectural":[185],"scaling":[186],"law:":[187],"per-head":[190],"dimension":[191],"grows":[192,204],"logarithmically":[193],"with":[194,207],"training":[195],"set":[196],"size,":[197],"while":[198],"nearly":[205],"linearly":[206],"total":[209],"D.":[211],"Our":[212],"framework":[213],"unifies":[214],"three":[215],"strands":[216],"prior":[218],"work:":[219],"attention,":[225],"general":[227],"weighting":[228],"learning,":[232],"decorrelation-variance-reduction":[235],"isomorphism":[236],"biological":[238],"computational":[240],"ensembles.":[241],"Multi-head":[242],"Transformer's":[246],"instantiation":[247],"universal":[250],"principle:":[251],"identical":[252],"agents":[253],"plus":[254],"diversity-enforcing":[255],"mechanisms":[256],"emergent":[258],"optimality.":[259]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-05-22T00:00:00"}