Loading…
SUPPORT UNION RECOVERY IN HIGH-DIMENSIONAL MULTIVARIATE REGRESSION
In multivariate regression, a K -dimensional response vector is regressed upon a common set of p covariates, with a matrix B* ∈ R p×K of regression coefficients. We study the behavior of the multivariate group Lasso, in which block regularization based on the l1/l2 norm is used for support union rec...
Saved in:
Published in: | The Annals of statistics 2011-02, Vol.39 (1), p.1-47 |
---|---|
Main Authors: | , , |
Format: | Article |
Language: | English |
Subjects: | |
Citations: | Items that this one cites Items that cite this one |
Online Access: | Get full text |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
cited_by | cdi_FETCH-LOGICAL-c506t-ca1264b502cd6ccee80dcdbaa2dbc49f304d1741ce53037de3119f23bf0916b43 |
---|---|
cites | cdi_FETCH-LOGICAL-c506t-ca1264b502cd6ccee80dcdbaa2dbc49f304d1741ce53037de3119f23bf0916b43 |
container_end_page | 47 |
container_issue | 1 |
container_start_page | 1 |
container_title | The Annals of statistics |
container_volume | 39 |
creator | Obozinski, Guillaume Wainwright, Martin J. Jordan, Michael I. |
description | In multivariate regression, a K -dimensional response vector is regressed upon a common set of p covariates, with a matrix B* ∈ R p×K of regression coefficients. We study the behavior of the multivariate group Lasso, in which block regularization based on the l1/l2 norm is used for support union recovery, or recovery of the set of s rows for which B* is nonzero. Under high-dimensional scaling, we show that the multivariate group Lasso exhibits a threshold for the recovery of the exact row pattern with high probability over the random design and noise that is specified by the sample complexity parameter θ(n, p, s):=n/[2ψ(B*) log(p − s)]. Here n is the sample size, and ψ(B*) is a sparsity-overlap function measuring a combination of the sparsities and overlaps of the K -regression coefficient vectors that constitute the model. We prove that the multivariate group Lasso succeeds for problem sequences (n, p, s) such that θ(n, p, s) exceeds a critical level θ u , and fails for sequences such that θ(n, p, s) lies below a critical level θ l . For the special case of the standard Gaussian ensemble, we show that θ l = θ u so that the characterization is sharp. The sparsity-overlap function ψ(B*) reveals that, if the design is uncorrelated on the active rows, l1/l2 regularization for multivariate regression never harms performance relative to an ordinary Lasso approach and can yield substantial improvements in sample complexity (up to a factor of K) when the coefficient vectors are suitably orthogonal. For more general designs, it is possible for the ordinary Lasso to outperform the multivariate group Lasso. We complement our analysis with simulations that demonstrate the sharpness of our theoretical results, even for relatively small problems. |
doi_str_mv | 10.1214/09-aos776 |
format | article |
fullrecord | <record><control><sourceid>jstor_proje</sourceid><recordid>TN_cdi_projecteuclid_primary_oai_CULeuclid_euclid_aos_1291388368</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><jstor_id>29783630</jstor_id><sourcerecordid>29783630</sourcerecordid><originalsourceid>FETCH-LOGICAL-c506t-ca1264b502cd6ccee80dcdbaa2dbc49f304d1741ce53037de3119f23bf0916b43</originalsourceid><addsrcrecordid>eNpFkE1Lw0AQhhdRsFYP_gAhCB48RPcju8nejDW2gbQp-Sh4WjabDTTUpmbbg__eLQ31NDDvM-_MvADcI_iCMPJeIXdlZ3yfXYARRixwA87YJRhByKFLCfOuwY0xLYSQco-MwHteLpdpVjjlIk4XThZN0lWUfTnxwpnF05n7Ec-jRW6lMHHmZVLEqzCLwyKy5DSL8qNyC64auTH6bqhjUH5GxWTmJuk0noSJqyhke1dJhJlXUYhVzZTSOoC1qispcV0pjzcEejXyPaQ0JZD4tSYI8QaTqoEcscojY_B28t31XavVXh_UZl2LXb_-lv2v6ORaTMpk6A7FZiEQ5ogEAWGBtXg8W_wctNmLtjv0W3u1CKhPMaEEW-j5BKm-M6bXzXkFguIYsoBchGluQ7bs02AojZKbppdbtTbnAUxsyCSglns4ca3Zd_2_zn17l333DxhGgA0</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>857523532</pqid></control><display><type>article</type><title>SUPPORT UNION RECOVERY IN HIGH-DIMENSIONAL MULTIVARIATE REGRESSION</title><source>JSTOR Archival Journals and Primary Sources Collection【Remote access available】</source><source>Project Euclid Journals</source><creator>Obozinski, Guillaume ; Wainwright, Martin J. ; Jordan, Michael I.</creator><creatorcontrib>Obozinski, Guillaume ; Wainwright, Martin J. ; Jordan, Michael I.</creatorcontrib><description>In multivariate regression, a K -dimensional response vector is regressed upon a common set of p covariates, with a matrix B* ∈ R p×K of regression coefficients. We study the behavior of the multivariate group Lasso, in which block regularization based on the l1/l2 norm is used for support union recovery, or recovery of the set of s rows for which B* is nonzero. Under high-dimensional scaling, we show that the multivariate group Lasso exhibits a threshold for the recovery of the exact row pattern with high probability over the random design and noise that is specified by the sample complexity parameter θ(n, p, s):=n/[2ψ(B*) log(p − s)]. Here n is the sample size, and ψ(B*) is a sparsity-overlap function measuring a combination of the sparsities and overlaps of the K -regression coefficient vectors that constitute the model. We prove that the multivariate group Lasso succeeds for problem sequences (n, p, s) such that θ(n, p, s) exceeds a critical level θ u , and fails for sequences such that θ(n, p, s) lies below a critical level θ l . For the special case of the standard Gaussian ensemble, we show that θ l = θ u so that the characterization is sharp. The sparsity-overlap function ψ(B*) reveals that, if the design is uncorrelated on the active rows, l1/l2 regularization for multivariate regression never harms performance relative to an ordinary Lasso approach and can yield substantial improvements in sample complexity (up to a factor of K) when the coefficient vectors are suitably orthogonal. For more general designs, it is possible for the ordinary Lasso to outperform the multivariate group Lasso. We complement our analysis with simulations that demonstrate the sharpness of our theoretical results, even for relatively small problems.</description><identifier>ISSN: 0090-5364</identifier><identifier>EISSN: 2168-8966</identifier><identifier>DOI: 10.1214/09-aos776</identifier><identifier>CODEN: ASTSC7</identifier><language>eng</language><publisher>Cleveland, OH: Institute of Mathematical Statistics</publisher><subject>62F07 ; 62J07 ; block-norm ; Covariance matrices ; Eigenvalues ; Exact sciences and technology ; General topics ; group Lasso ; high-dimensional scaling ; Infinity ; Jordan matrices ; LASSO ; Linear regression ; Mathematics ; Matrices ; Modeling ; Multivariate analysis ; multivariate regression ; Normal distribution ; Numerical analysis ; Numerical analysis in abstract spaces ; Numerical analysis. Scientific computation ; Numerical linear algebra ; Probability and statistics ; Regression analysis ; Regression coefficients ; Sample size ; Sciences and techniques of general use ; second-order cone program ; Simulation ; simultaneous Lasso ; sparsity ; Statistics ; Studies ; variable selection</subject><ispartof>The Annals of statistics, 2011-02, Vol.39 (1), p.1-47</ispartof><rights>Copyright © 2011 The Institute of Mathematical Statistics</rights><rights>2015 INIST-CNRS</rights><rights>Copyright Institute of Mathematical Statistics Feb 2011</rights><rights>Copyright 2011 Institute of Mathematical Statistics</rights><lds50>peer_reviewed</lds50><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed><citedby>FETCH-LOGICAL-c506t-ca1264b502cd6ccee80dcdbaa2dbc49f304d1741ce53037de3119f23bf0916b43</citedby><cites>FETCH-LOGICAL-c506t-ca1264b502cd6ccee80dcdbaa2dbc49f304d1741ce53037de3119f23bf0916b43</cites></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><linktopdf>$$Uhttps://www.jstor.org/stable/pdf/29783630$$EPDF$$P50$$Gjstor$$H</linktopdf><linktohtml>$$Uhttps://www.jstor.org/stable/29783630$$EHTML$$P50$$Gjstor$$H</linktohtml><link.rule.ids>230,315,786,790,891,932,27957,27958,58593,58826</link.rule.ids><backlink>$$Uhttp://pascal-francis.inist.fr/vibad/index.php?action=getRecordDetail&idt=23943385$$DView record in Pascal Francis$$Hfree_for_read</backlink></links><search><creatorcontrib>Obozinski, Guillaume</creatorcontrib><creatorcontrib>Wainwright, Martin J.</creatorcontrib><creatorcontrib>Jordan, Michael I.</creatorcontrib><title>SUPPORT UNION RECOVERY IN HIGH-DIMENSIONAL MULTIVARIATE REGRESSION</title><title>The Annals of statistics</title><description>In multivariate regression, a K -dimensional response vector is regressed upon a common set of p covariates, with a matrix B* ∈ R p×K of regression coefficients. We study the behavior of the multivariate group Lasso, in which block regularization based on the l1/l2 norm is used for support union recovery, or recovery of the set of s rows for which B* is nonzero. Under high-dimensional scaling, we show that the multivariate group Lasso exhibits a threshold for the recovery of the exact row pattern with high probability over the random design and noise that is specified by the sample complexity parameter θ(n, p, s):=n/[2ψ(B*) log(p − s)]. Here n is the sample size, and ψ(B*) is a sparsity-overlap function measuring a combination of the sparsities and overlaps of the K -regression coefficient vectors that constitute the model. We prove that the multivariate group Lasso succeeds for problem sequences (n, p, s) such that θ(n, p, s) exceeds a critical level θ u , and fails for sequences such that θ(n, p, s) lies below a critical level θ l . For the special case of the standard Gaussian ensemble, we show that θ l = θ u so that the characterization is sharp. The sparsity-overlap function ψ(B*) reveals that, if the design is uncorrelated on the active rows, l1/l2 regularization for multivariate regression never harms performance relative to an ordinary Lasso approach and can yield substantial improvements in sample complexity (up to a factor of K) when the coefficient vectors are suitably orthogonal. For more general designs, it is possible for the ordinary Lasso to outperform the multivariate group Lasso. We complement our analysis with simulations that demonstrate the sharpness of our theoretical results, even for relatively small problems.</description><subject>62F07</subject><subject>62J07</subject><subject>block-norm</subject><subject>Covariance matrices</subject><subject>Eigenvalues</subject><subject>Exact sciences and technology</subject><subject>General topics</subject><subject>group Lasso</subject><subject>high-dimensional scaling</subject><subject>Infinity</subject><subject>Jordan matrices</subject><subject>LASSO</subject><subject>Linear regression</subject><subject>Mathematics</subject><subject>Matrices</subject><subject>Modeling</subject><subject>Multivariate analysis</subject><subject>multivariate regression</subject><subject>Normal distribution</subject><subject>Numerical analysis</subject><subject>Numerical analysis in abstract spaces</subject><subject>Numerical analysis. Scientific computation</subject><subject>Numerical linear algebra</subject><subject>Probability and statistics</subject><subject>Regression analysis</subject><subject>Regression coefficients</subject><subject>Sample size</subject><subject>Sciences and techniques of general use</subject><subject>second-order cone program</subject><subject>Simulation</subject><subject>simultaneous Lasso</subject><subject>sparsity</subject><subject>Statistics</subject><subject>Studies</subject><subject>variable selection</subject><issn>0090-5364</issn><issn>2168-8966</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2011</creationdate><recordtype>article</recordtype><recordid>eNpFkE1Lw0AQhhdRsFYP_gAhCB48RPcju8nejDW2gbQp-Sh4WjabDTTUpmbbg__eLQ31NDDvM-_MvADcI_iCMPJeIXdlZ3yfXYARRixwA87YJRhByKFLCfOuwY0xLYSQco-MwHteLpdpVjjlIk4XThZN0lWUfTnxwpnF05n7Ec-jRW6lMHHmZVLEqzCLwyKy5DSL8qNyC64auTH6bqhjUH5GxWTmJuk0noSJqyhke1dJhJlXUYhVzZTSOoC1qispcV0pjzcEejXyPaQ0JZD4tSYI8QaTqoEcscojY_B28t31XavVXh_UZl2LXb_-lv2v6ORaTMpk6A7FZiEQ5ogEAWGBtXg8W_wctNmLtjv0W3u1CKhPMaEEW-j5BKm-M6bXzXkFguIYsoBchGluQ7bs02AojZKbppdbtTbnAUxsyCSglns4ca3Zd_2_zn17l333DxhGgA0</recordid><startdate>20110201</startdate><enddate>20110201</enddate><creator>Obozinski, Guillaume</creator><creator>Wainwright, Martin J.</creator><creator>Jordan, Michael I.</creator><general>Institute of Mathematical Statistics</general><general>The Institute of Mathematical Statistics</general><scope>IQODW</scope><scope>AAYXX</scope><scope>CITATION</scope><scope>JQ2</scope></search><sort><creationdate>20110201</creationdate><title>SUPPORT UNION RECOVERY IN HIGH-DIMENSIONAL MULTIVARIATE REGRESSION</title><author>Obozinski, Guillaume ; Wainwright, Martin J. ; Jordan, Michael I.</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-c506t-ca1264b502cd6ccee80dcdbaa2dbc49f304d1741ce53037de3119f23bf0916b43</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2011</creationdate><topic>62F07</topic><topic>62J07</topic><topic>block-norm</topic><topic>Covariance matrices</topic><topic>Eigenvalues</topic><topic>Exact sciences and technology</topic><topic>General topics</topic><topic>group Lasso</topic><topic>high-dimensional scaling</topic><topic>Infinity</topic><topic>Jordan matrices</topic><topic>LASSO</topic><topic>Linear regression</topic><topic>Mathematics</topic><topic>Matrices</topic><topic>Modeling</topic><topic>Multivariate analysis</topic><topic>multivariate regression</topic><topic>Normal distribution</topic><topic>Numerical analysis</topic><topic>Numerical analysis in abstract spaces</topic><topic>Numerical analysis. Scientific computation</topic><topic>Numerical linear algebra</topic><topic>Probability and statistics</topic><topic>Regression analysis</topic><topic>Regression coefficients</topic><topic>Sample size</topic><topic>Sciences and techniques of general use</topic><topic>second-order cone program</topic><topic>Simulation</topic><topic>simultaneous Lasso</topic><topic>sparsity</topic><topic>Statistics</topic><topic>Studies</topic><topic>variable selection</topic><toplevel>peer_reviewed</toplevel><toplevel>online_resources</toplevel><creatorcontrib>Obozinski, Guillaume</creatorcontrib><creatorcontrib>Wainwright, Martin J.</creatorcontrib><creatorcontrib>Jordan, Michael I.</creatorcontrib><collection>Pascal-Francis</collection><collection>CrossRef</collection><collection>ProQuest Computer Science Collection</collection><jtitle>The Annals of statistics</jtitle></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Obozinski, Guillaume</au><au>Wainwright, Martin J.</au><au>Jordan, Michael I.</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>SUPPORT UNION RECOVERY IN HIGH-DIMENSIONAL MULTIVARIATE REGRESSION</atitle><jtitle>The Annals of statistics</jtitle><date>2011-02-01</date><risdate>2011</risdate><volume>39</volume><issue>1</issue><spage>1</spage><epage>47</epage><pages>1-47</pages><issn>0090-5364</issn><eissn>2168-8966</eissn><coden>ASTSC7</coden><abstract>In multivariate regression, a K -dimensional response vector is regressed upon a common set of p covariates, with a matrix B* ∈ R p×K of regression coefficients. We study the behavior of the multivariate group Lasso, in which block regularization based on the l1/l2 norm is used for support union recovery, or recovery of the set of s rows for which B* is nonzero. Under high-dimensional scaling, we show that the multivariate group Lasso exhibits a threshold for the recovery of the exact row pattern with high probability over the random design and noise that is specified by the sample complexity parameter θ(n, p, s):=n/[2ψ(B*) log(p − s)]. Here n is the sample size, and ψ(B*) is a sparsity-overlap function measuring a combination of the sparsities and overlaps of the K -regression coefficient vectors that constitute the model. We prove that the multivariate group Lasso succeeds for problem sequences (n, p, s) such that θ(n, p, s) exceeds a critical level θ u , and fails for sequences such that θ(n, p, s) lies below a critical level θ l . For the special case of the standard Gaussian ensemble, we show that θ l = θ u so that the characterization is sharp. The sparsity-overlap function ψ(B*) reveals that, if the design is uncorrelated on the active rows, l1/l2 regularization for multivariate regression never harms performance relative to an ordinary Lasso approach and can yield substantial improvements in sample complexity (up to a factor of K) when the coefficient vectors are suitably orthogonal. For more general designs, it is possible for the ordinary Lasso to outperform the multivariate group Lasso. We complement our analysis with simulations that demonstrate the sharpness of our theoretical results, even for relatively small problems.</abstract><cop>Cleveland, OH</cop><pub>Institute of Mathematical Statistics</pub><doi>10.1214/09-aos776</doi><tpages>47</tpages><oa>free_for_read</oa></addata></record> |
fulltext | fulltext |
identifier | ISSN: 0090-5364 |
ispartof | The Annals of statistics, 2011-02, Vol.39 (1), p.1-47 |
issn | 0090-5364 2168-8966 |
language | eng |
recordid | cdi_projecteuclid_primary_oai_CULeuclid_euclid_aos_1291388368 |
source | JSTOR Archival Journals and Primary Sources Collection【Remote access available】; Project Euclid Journals |
subjects | 62F07 62J07 block-norm Covariance matrices Eigenvalues Exact sciences and technology General topics group Lasso high-dimensional scaling Infinity Jordan matrices LASSO Linear regression Mathematics Matrices Modeling Multivariate analysis multivariate regression Normal distribution Numerical analysis Numerical analysis in abstract spaces Numerical analysis. Scientific computation Numerical linear algebra Probability and statistics Regression analysis Regression coefficients Sample size Sciences and techniques of general use second-order cone program Simulation simultaneous Lasso sparsity Statistics Studies variable selection |
title | SUPPORT UNION RECOVERY IN HIGH-DIMENSIONAL MULTIVARIATE REGRESSION |
url | http://sfxeu10.hosted.exlibrisgroup.com/loughborough?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-09-23T07%3A33%3A09IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-jstor_proje&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=SUPPORT%20UNION%20RECOVERY%20IN%20HIGH-DIMENSIONAL%20MULTIVARIATE%20REGRESSION&rft.jtitle=The%20Annals%20of%20statistics&rft.au=Obozinski,%20Guillaume&rft.date=2011-02-01&rft.volume=39&rft.issue=1&rft.spage=1&rft.epage=47&rft.pages=1-47&rft.issn=0090-5364&rft.eissn=2168-8966&rft.coden=ASTSC7&rft_id=info:doi/10.1214/09-aos776&rft_dat=%3Cjstor_proje%3E29783630%3C/jstor_proje%3E%3Cgrp_id%3Ecdi_FETCH-LOGICAL-c506t-ca1264b502cd6ccee80dcdbaa2dbc49f304d1741ce53037de3119f23bf0916b43%3C/grp_id%3E%3Coa%3E%3C/oa%3E%3Curl%3E%3C/url%3E&rft_id=info:oai/&rft_pqid=857523532&rft_id=info:pmid/&rft_jstor_id=29783630&rfr_iscdi=true |