{"id":"https://openalex.org/W7127616444","doi":"https://doi.org/10.48550/arxiv.2602.02986","title":"Why Some Models Resist Unlearning: A Linear Stability Perspective","display_name":"Why Some Models Resist Unlearning: A Linear Stability Perspective","publication_year":2026,"publication_date":"2026-02-03","ids":{"openalex":"https://openalex.org/W7127616444","doi":"https://doi.org/10.48550/arxiv.2602.02986"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2602.02986","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2602.02986","pdf_url":"https://arxiv.org/pdf/2602.02986","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,"listed_in":[],"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"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2602.02986","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5026060582","display_name":"Wei-I Chang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I219193219","display_name":"Purdue University West Lafayette","ror":"https://ror.org/02dqehb95","country_code":"US","type":"education","lineage":["https://openalex.org/I219193219"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Chang, Wei-Kai","raw_affiliation_strings":["Purdue University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Purdue University","institution_ids":["https://openalex.org/I219193219"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5068930801","display_name":"Rajiv Khanna","orcid":"https://orcid.org/0000-0003-2241-0353","observed_orcids":["https://orcid.org/0000-0003-2241-0353"]},"institutions":[{"id":"https://openalex.org/I219193219","display_name":"Purdue University West Lafayette","ror":"https://ror.org/02dqehb95","country_code":"US","type":"education","lineage":["https://openalex.org/I219193219"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Khanna, Rajiv","raw_affiliation_strings":["Purdue University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Purdue University","institution_ids":["https://openalex.org/I219193219"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I219193219"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":3,"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/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.9839000105857849,"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/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.9839000105857849,"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":0.012799999676644802,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.000699999975040555,"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/machine-unlearning","display_name":"machine unlearning","score":0.9940000176429749},{"id":"https://openalex.org/keywords/linear-stability","display_name":"linear stability","score":0.8450000286102295},{"id":"https://openalex.org/keywords/data-coherence","display_name":"data coherence","score":0.7990000247955322},{"id":"https://openalex.org/keywords/memorization","display_name":"memorization","score":0.6620000004768372},{"id":"https://openalex.org/keywords/signal-to-noise-ratio","display_name":"signal-to-noise ratio","score":0.7760000228881836},{"id":"https://openalex.org/keywords/random-matrix-theory","display_name":"random matrix theory","score":0.7519999742507935},{"id":"https://openalex.org/keywords/data-privacy","display_name":"data privacy","score":0.4180000126361847},{"id":"https://openalex.org/keywords/stability-threshold","display_name":"stability threshold","score":0.3869999945163727}],"concepts":[{"id":"https://openalex.org/C2781181686","wikidata":"https://www.wikidata.org/wiki/Q4226068","display_name":"Coherence (philosophical gambling strategy)","level":2,"score":0.5669000148773193},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.566100001335144},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.478300005197525},{"id":"https://openalex.org/C64812099","wikidata":"https://www.wikidata.org/wiki/Q176604","display_name":"Random matrix","level":3,"score":0.46470001339912415},{"id":"https://openalex.org/C203616005","wikidata":"https://www.wikidata.org/wiki/Q620495","display_name":"Hessian matrix","level":2,"score":0.4510999917984009},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4374000132083893},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3700999915599823},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.3456999957561493},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.33959999680519104},{"id":"https://openalex.org/C163175372","wikidata":"https://www.wikidata.org/wiki/Q3339222","display_name":"Linear model","level":2,"score":0.33799999952316284},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.3327000141143799},{"id":"https://openalex.org/C7149132","wikidata":"https://www.wikidata.org/wiki/Q1377840","display_name":"Forgetting","level":2,"score":0.32820001244544983},{"id":"https://openalex.org/C6802819","wikidata":"https://www.wikidata.org/wiki/Q1072174","display_name":"Linear system","level":2,"score":0.32499998807907104},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.305400013923645},{"id":"https://openalex.org/C39891107","wikidata":"https://www.wikidata.org/wiki/Q5767098","display_name":"Hinge loss","level":3,"score":0.301800012588501},{"id":"https://openalex.org/C198931094","wikidata":"https://www.wikidata.org/wiki/Q2610096","display_name":"Hurwitz matrix","level":3,"score":0.272599995136261},{"id":"https://openalex.org/C174348530","wikidata":"https://www.wikidata.org/wiki/Q188635","display_name":"Bridging (networking)","level":2,"score":0.2702000141143799},{"id":"https://openalex.org/C120936955","wikidata":"https://www.wikidata.org/wiki/Q2155640","display_name":"Empirical research","level":2,"score":0.26910001039505005},{"id":"https://openalex.org/C143170015","wikidata":"https://www.wikidata.org/wiki/Q17007850","display_name":"Stability conditions","level":3,"score":0.2646999955177307},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.2639000117778778},{"id":"https://openalex.org/C104267543","wikidata":"https://www.wikidata.org/wiki/Q208163","display_name":"Signal processing","level":3,"score":0.2574000060558319}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2602.02986","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2602.02986","pdf_url":"https://arxiv.org/pdf/2602.02986","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,"listed_in":[],"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.2602.02986","doi":"https://doi.org/10.48550/arxiv.2602.02986","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.02986","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,"listed_in":[],"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":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2602.02986","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2602.02986","pdf_url":"https://arxiv.org/pdf/2602.02986","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,"listed_in":[],"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":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7127616444.pdf","grobid_xml":"https://content.openalex.org/works/W7127616444.grobid-xml"},"referenced_works_count":30,"referenced_works":["https://openalex.org/W2155894447","https://openalex.org/W2912713668","https://openalex.org/W2950181023","https://openalex.org/W2961073278","https://openalex.org/W2962933129","https://openalex.org/W2985940692","https://openalex.org/W2994747787","https://openalex.org/W3006064982","https://openalex.org/W3021616083","https://openalex.org/W3091401866","https://openalex.org/W3094049233","https://openalex.org/W3118039497","https://openalex.org/W3135378441","https://openalex.org/W3158390351","https://openalex.org/W3193974325","https://openalex.org/W3214709531","https://openalex.org/W4321392443","https://openalex.org/W4323651343","https://openalex.org/W4387838615","https://openalex.org/W4391871165","https://openalex.org/W4393147294","https://openalex.org/W4394717673","https://openalex.org/W4400065270","https://openalex.org/W4405956499","https://openalex.org/W4415109331","https://openalex.org/W4415338567","https://openalex.org/W4417100113","https://openalex.org/W7110933161","https://openalex.org/W7133217935","https://openalex.org/W7135101949"],"related_works":["https://openalex.org/W7137882798","https://openalex.org/W4414849787","https://openalex.org/W7092288443","https://openalex.org/W4415798088","https://openalex.org/W7163140489","https://openalex.org/W4393929290","https://openalex.org/W4403345195","https://openalex.org/W4414920171","https://openalex.org/W7204908097","https://openalex.org/W7128610365"],"abstract_inverted_index":{"Machine":[0],"unlearning,":[1],"the":[2,6,49,57,76,85,94,98,143,184,188,215,220],"ability":[3],"to":[4,55,119,145],"erase":[5],"effect":[7],"of":[8,35,51,80,191,198,219],"specific":[9],"training":[10],"samples":[11],"without":[12],"retraining":[13],"from":[14,112],"scratch,":[15],"is":[16,71,75,149,154,206],"critical":[17],"for":[18],"privacy,":[19],"regulation,":[20],"and":[21,37,62,101,104,134,158,178,201,213,226],"efficiency.":[22],"However,":[23],"most":[24],"progress":[25],"in":[26,68,208],"unlearning":[27,39,47,160,194],"has":[28],"been":[29],"empirical,":[30],"with":[31,183],"little":[32],"theoretical":[33],"understanding":[34],"when":[36,142],"why":[38],"works.":[40],"We":[41,87],"tackle":[42],"this":[43],"gap":[44],"by":[45],"framing":[46],"through":[48],"lens":[50],"asymptotic":[52],"linear":[53],"stability":[54,107,189],"capture":[56],"interaction":[58],"between":[59,102,223],"optimization":[60],"dynamics":[61],"data":[63,72,117],"geometry.":[64],"The":[65],"key":[66],"quantity":[67],"our":[69],"analysis":[70,205],"coherence":[73,89,157],"which":[74],"cross":[77,151],"sample":[78,152],"alignment":[79,153],"loss":[81],"surface":[82],"directions":[83],"near":[84],"optimum.":[86],"decompose":[88],"along":[90],"three":[91],"axes:":[92],"within":[93,97],"retain":[95],"set,":[96,100],"forget":[99],"them,":[103],"prove":[105],"tight":[106],"thresholds":[108],"that":[109,136,175],"separate":[110],"convergence":[111],"divergence.":[113],"To":[114],"further":[115],"link":[116],"properties":[118],"forgettability,":[120],"we":[121,173],"study":[122],"a":[123,129,196],"two":[124],"layer":[125],"ReLU":[126],"CNN":[127,179],"under":[128],"signal":[130,144],"plus":[131],"noise":[132,146],"model":[133],"show":[135,174],"stronger":[137],"memorization":[138],"makes":[139],"forgetting":[140],"easier:":[141],"ratio":[147],"(SNR)":[148],"lower,":[150],"weaker,":[155],"reducing":[156],"making":[159],"easier;":[161],"conversely,":[162],"high":[163],"SNR,":[164],"highly":[165],"aligned":[166],"models":[167],"resist":[168],"unlearning.":[169,227],"For":[170],"empirical":[171],"verification,":[172],"Hessian":[176],"tests":[177],"heatmaps":[180],"align":[181],"closely":[182],"predicted":[185],"boundary,":[186],"mapping":[187],"frontier":[190],"gradient":[192],"based":[193],"as":[195],"function":[197],"batching,":[199],"mixing,":[200],"data/model":[202],"alignment.":[203],"Our":[204],"grounded":[207],"random":[209],"matrix":[210],"theory":[211],"tools":[212],"provides":[214],"first":[216],"principled":[217],"account":[218],"trade":[221],"offs":[222],"memorization,":[224],"coherence,":[225]},"counts_by_year":[{"year":2026,"cited_by_count":3}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-02-06T00:00:00"}