On stochastic gradient Langevin dynamics with dependent data streams in the logconcave case

We study the problem of sampling from a probability distribution pi on R-d which has a density w.r.t. the Lebesgue measure known up to a normalization factor x bar right arrow e(-U(x)) / f(R)d e (-U(y)) dy.. We analyze a sampling method based on the Euler discretization of the Langevin stochastic di...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Bernoulli : official journal of the Bernoulli Society for Mathematical Statistics and Probability 2021-02, Vol.27 (1), p.1-33
Hauptverfasser: Barkhagen, M., Chau, N. H., Moulines, E., Rasonyi, M., Sabanis, S., Zhang, Y.
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page 33
container_issue 1
container_start_page 1
container_title Bernoulli : official journal of the Bernoulli Society for Mathematical Statistics and Probability
container_volume 27
creator Barkhagen, M.
Chau, N. H.
Moulines, E.
Rasonyi, M.
Sabanis, S.
Zhang, Y.
description We study the problem of sampling from a probability distribution pi on R-d which has a density w.r.t. the Lebesgue measure known up to a normalization factor x bar right arrow e(-U(x)) / f(R)d e (-U(y)) dy.. We analyze a sampling method based on the Euler discretization of the Langevin stochastic differential equations under the assumptions that the potential U is continuously differentiable, del U is Lipschitz, and U is strongly concave. We focus on the case where the gradient of the log-density cannot be directly computed but unbiased estimates of the gradient from possibly dependent observations are available. This setting can be seen as a combination of a stochastic approximation (here stochastic gradient) type algorithms with discretized Langevin dynamics. We obtain an upper bound of the Wasserstein-2 distance between the law of the iterates of this algorithm and the target distribution pi with constants depending explicitly on the Lipschitz and strong convexity constants of the potential and the dimension of the space. Finally, under weaker assumptions on U and its gradient but in the presence of independent observations, we obtain analogous results in Wasserstein-2 distance.
doi_str_mv 10.3150/19-BEJ1187
format Article
fullrecord <record><control><sourceid>hal_cross</sourceid><recordid>TN_cdi_crossref_primary_10_3150_19_BEJ1187</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>oai_HAL_hal_03529653v1</sourcerecordid><originalsourceid>FETCH-LOGICAL-c301t-de993fed393dd1fe394b7db50d65999c34cde586a1cdb4fecd97495784751cdd3</originalsourceid><addsrcrecordid>eNqNkLFOwzAURS0EEqWw8AVeAQXsOI7jsUSFgiJ1gYkhcuyX1qi1q9i06t_jqqgz07u6OucNF6FbSh4Z5eSJyux5-k5pJc7QiHLBMim4PE-ZcZKJvOSX6CqEb0JoUZZkhL7mDofo9VKFaDVeDMpYcBE3yi1gax02e6fWVge8s3GJDWzAmQNgVFTJHECtA05cXAJe-YX2TqstYK0CXKOLXq0C3PzdMfp8mX7Us6yZv77VkybTjNCYGZCS9WCYZMbQHpgsOmE6TkzJpZSaFdoAr0pFtemKHrSRopBcVIXgqTJsjO6Of5dq1W4Gu1bDvvXKtrNJ0x46wnguS862NLH3R1YPPoQB-pNASXuYsKWy_ZswwdUR3kHn-6DTMhpOAiGEy7ws8jwlQmsbVbTe1f7HxaQ-_F9lv7PPhSE</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>On stochastic gradient Langevin dynamics with dependent data streams in the logconcave case</title><source>EZB-FREE-00999 freely available EZB journals</source><source>Project Euclid Complete</source><creator>Barkhagen, M. ; Chau, N. H. ; Moulines, E. ; Rasonyi, M. ; Sabanis, S. ; Zhang, Y.</creator><creatorcontrib>Barkhagen, M. ; Chau, N. H. ; Moulines, E. ; Rasonyi, M. ; Sabanis, S. ; Zhang, Y.</creatorcontrib><description>We study the problem of sampling from a probability distribution pi on R-d which has a density w.r.t. the Lebesgue measure known up to a normalization factor x bar right arrow e(-U(x)) / f(R)d e (-U(y)) dy.. We analyze a sampling method based on the Euler discretization of the Langevin stochastic differential equations under the assumptions that the potential U is continuously differentiable, del U is Lipschitz, and U is strongly concave. We focus on the case where the gradient of the log-density cannot be directly computed but unbiased estimates of the gradient from possibly dependent observations are available. This setting can be seen as a combination of a stochastic approximation (here stochastic gradient) type algorithms with discretized Langevin dynamics. We obtain an upper bound of the Wasserstein-2 distance between the law of the iterates of this algorithm and the target distribution pi with constants depending explicitly on the Lipschitz and strong convexity constants of the potential and the dimension of the space. Finally, under weaker assumptions on U and its gradient but in the presence of independent observations, we obtain analogous results in Wasserstein-2 distance.</description><identifier>ISSN: 1350-7265</identifier><identifier>EISSN: 1573-9759</identifier><identifier>DOI: 10.3150/19-BEJ1187</identifier><language>eng</language><publisher>VOORBURG: Int Statistical Inst</publisher><subject>Mathematics ; Physical Sciences ; Science &amp; Technology ; Statistics &amp; Probability</subject><ispartof>Bernoulli : official journal of the Bernoulli Society for Mathematical Statistics and Probability, 2021-02, Vol.27 (1), p.1-33</ispartof><rights>Distributed under a Creative Commons Attribution 4.0 International License</rights><lds50>peer_reviewed</lds50><oa>free_for_read</oa><woscitedreferencessubscribed>true</woscitedreferencessubscribed><woscitedreferencescount>20</woscitedreferencescount><woscitedreferencesoriginalsourcerecordid>wos000592642200001</woscitedreferencesoriginalsourcerecordid><citedby>FETCH-LOGICAL-c301t-de993fed393dd1fe394b7db50d65999c34cde586a1cdb4fecd97495784751cdd3</citedby><cites>FETCH-LOGICAL-c301t-de993fed393dd1fe394b7db50d65999c34cde586a1cdb4fecd97495784751cdd3</cites><orcidid>0000-0002-3991-362X ; 0000-0001-8198-3392 ; 0000-0002-2058-0693</orcidid></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>230,315,781,785,886,27928,27929</link.rule.ids><backlink>$$Uhttps://inria.hal.science/hal-03529653$$DView record in HAL$$Hfree_for_read</backlink></links><search><creatorcontrib>Barkhagen, M.</creatorcontrib><creatorcontrib>Chau, N. H.</creatorcontrib><creatorcontrib>Moulines, E.</creatorcontrib><creatorcontrib>Rasonyi, M.</creatorcontrib><creatorcontrib>Sabanis, S.</creatorcontrib><creatorcontrib>Zhang, Y.</creatorcontrib><title>On stochastic gradient Langevin dynamics with dependent data streams in the logconcave case</title><title>Bernoulli : official journal of the Bernoulli Society for Mathematical Statistics and Probability</title><addtitle>BERNOULLI</addtitle><description>We study the problem of sampling from a probability distribution pi on R-d which has a density w.r.t. the Lebesgue measure known up to a normalization factor x bar right arrow e(-U(x)) / f(R)d e (-U(y)) dy.. We analyze a sampling method based on the Euler discretization of the Langevin stochastic differential equations under the assumptions that the potential U is continuously differentiable, del U is Lipschitz, and U is strongly concave. We focus on the case where the gradient of the log-density cannot be directly computed but unbiased estimates of the gradient from possibly dependent observations are available. This setting can be seen as a combination of a stochastic approximation (here stochastic gradient) type algorithms with discretized Langevin dynamics. We obtain an upper bound of the Wasserstein-2 distance between the law of the iterates of this algorithm and the target distribution pi with constants depending explicitly on the Lipschitz and strong convexity constants of the potential and the dimension of the space. Finally, under weaker assumptions on U and its gradient but in the presence of independent observations, we obtain analogous results in Wasserstein-2 distance.</description><subject>Mathematics</subject><subject>Physical Sciences</subject><subject>Science &amp; Technology</subject><subject>Statistics &amp; Probability</subject><issn>1350-7265</issn><issn>1573-9759</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2021</creationdate><recordtype>article</recordtype><sourceid>HGBXW</sourceid><recordid>eNqNkLFOwzAURS0EEqWw8AVeAQXsOI7jsUSFgiJ1gYkhcuyX1qi1q9i06t_jqqgz07u6OucNF6FbSh4Z5eSJyux5-k5pJc7QiHLBMim4PE-ZcZKJvOSX6CqEb0JoUZZkhL7mDofo9VKFaDVeDMpYcBE3yi1gax02e6fWVge8s3GJDWzAmQNgVFTJHECtA05cXAJe-YX2TqstYK0CXKOLXq0C3PzdMfp8mX7Us6yZv77VkybTjNCYGZCS9WCYZMbQHpgsOmE6TkzJpZSaFdoAr0pFtemKHrSRopBcVIXgqTJsjO6Of5dq1W4Gu1bDvvXKtrNJ0x46wnguS862NLH3R1YPPoQB-pNASXuYsKWy_ZswwdUR3kHn-6DTMhpOAiGEy7ws8jwlQmsbVbTe1f7HxaQ-_F9lv7PPhSE</recordid><startdate>20210201</startdate><enddate>20210201</enddate><creator>Barkhagen, M.</creator><creator>Chau, N. H.</creator><creator>Moulines, E.</creator><creator>Rasonyi, M.</creator><creator>Sabanis, S.</creator><creator>Zhang, Y.</creator><general>Int Statistical Inst</general><general>Bernoulli Society for Mathematical Statistics and Probability</general><scope>BLEPL</scope><scope>DTL</scope><scope>HGBXW</scope><scope>AAYXX</scope><scope>CITATION</scope><scope>1XC</scope><orcidid>https://orcid.org/0000-0002-3991-362X</orcidid><orcidid>https://orcid.org/0000-0001-8198-3392</orcidid><orcidid>https://orcid.org/0000-0002-2058-0693</orcidid></search><sort><creationdate>20210201</creationdate><title>On stochastic gradient Langevin dynamics with dependent data streams in the logconcave case</title><author>Barkhagen, M. ; Chau, N. H. ; Moulines, E. ; Rasonyi, M. ; Sabanis, S. ; Zhang, Y.</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-c301t-de993fed393dd1fe394b7db50d65999c34cde586a1cdb4fecd97495784751cdd3</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2021</creationdate><topic>Mathematics</topic><topic>Physical Sciences</topic><topic>Science &amp; Technology</topic><topic>Statistics &amp; Probability</topic><toplevel>peer_reviewed</toplevel><toplevel>online_resources</toplevel><creatorcontrib>Barkhagen, M.</creatorcontrib><creatorcontrib>Chau, N. H.</creatorcontrib><creatorcontrib>Moulines, E.</creatorcontrib><creatorcontrib>Rasonyi, M.</creatorcontrib><creatorcontrib>Sabanis, S.</creatorcontrib><creatorcontrib>Zhang, Y.</creatorcontrib><collection>Web of Science Core Collection</collection><collection>Science Citation Index Expanded</collection><collection>Web of Science - Science Citation Index Expanded - 2021</collection><collection>CrossRef</collection><collection>Hyper Article en Ligne (HAL)</collection><jtitle>Bernoulli : official journal of the Bernoulli Society for Mathematical Statistics and Probability</jtitle></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Barkhagen, M.</au><au>Chau, N. H.</au><au>Moulines, E.</au><au>Rasonyi, M.</au><au>Sabanis, S.</au><au>Zhang, Y.</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>On stochastic gradient Langevin dynamics with dependent data streams in the logconcave case</atitle><jtitle>Bernoulli : official journal of the Bernoulli Society for Mathematical Statistics and Probability</jtitle><stitle>BERNOULLI</stitle><date>2021-02-01</date><risdate>2021</risdate><volume>27</volume><issue>1</issue><spage>1</spage><epage>33</epage><pages>1-33</pages><issn>1350-7265</issn><eissn>1573-9759</eissn><abstract>We study the problem of sampling from a probability distribution pi on R-d which has a density w.r.t. the Lebesgue measure known up to a normalization factor x bar right arrow e(-U(x)) / f(R)d e (-U(y)) dy.. We analyze a sampling method based on the Euler discretization of the Langevin stochastic differential equations under the assumptions that the potential U is continuously differentiable, del U is Lipschitz, and U is strongly concave. We focus on the case where the gradient of the log-density cannot be directly computed but unbiased estimates of the gradient from possibly dependent observations are available. This setting can be seen as a combination of a stochastic approximation (here stochastic gradient) type algorithms with discretized Langevin dynamics. We obtain an upper bound of the Wasserstein-2 distance between the law of the iterates of this algorithm and the target distribution pi with constants depending explicitly on the Lipschitz and strong convexity constants of the potential and the dimension of the space. Finally, under weaker assumptions on U and its gradient but in the presence of independent observations, we obtain analogous results in Wasserstein-2 distance.</abstract><cop>VOORBURG</cop><pub>Int Statistical Inst</pub><doi>10.3150/19-BEJ1187</doi><tpages>33</tpages><orcidid>https://orcid.org/0000-0002-3991-362X</orcidid><orcidid>https://orcid.org/0000-0001-8198-3392</orcidid><orcidid>https://orcid.org/0000-0002-2058-0693</orcidid><oa>free_for_read</oa></addata></record>
fulltext fulltext
identifier ISSN: 1350-7265
ispartof Bernoulli : official journal of the Bernoulli Society for Mathematical Statistics and Probability, 2021-02, Vol.27 (1), p.1-33
issn 1350-7265
1573-9759
language eng
recordid cdi_crossref_primary_10_3150_19_BEJ1187
source EZB-FREE-00999 freely available EZB journals; Project Euclid Complete
subjects Mathematics
Physical Sciences
Science & Technology
Statistics & Probability
title On stochastic gradient Langevin dynamics with dependent data streams in the logconcave case
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-17T06%3A35%3A19IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-hal_cross&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=On%20stochastic%20gradient%20Langevin%20dynamics%20with%20dependent%20data%20streams%20in%20the%20logconcave%20case&rft.jtitle=Bernoulli%20:%20official%20journal%20of%20the%20Bernoulli%20Society%20for%20Mathematical%20Statistics%20and%20Probability&rft.au=Barkhagen,%20M.&rft.date=2021-02-01&rft.volume=27&rft.issue=1&rft.spage=1&rft.epage=33&rft.pages=1-33&rft.issn=1350-7265&rft.eissn=1573-9759&rft_id=info:doi/10.3150/19-BEJ1187&rft_dat=%3Chal_cross%3Eoai_HAL_hal_03529653v1%3C/hal_cross%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true