欢迎来到天天文库
浏览记录
ID:40961735
大小:197.82 KB
页数:12页
时间:2019-08-12
《Stability and instance optimality for Gaussian》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、StabilityandinstanceoptimalityforGaussianmeasurementsincompressedsensingP.WojtaszczykyInstitutofAppliedMathematicsUniversityofWarsawul.Banacha2;02-097WarszawaPolandFebruary14,2008AbstractIncompressedsensingweseektogaininformationaboutvectorx2NRfromd<2、velinearmeasurements.Candes,Donoho,Taoet.al.(seee.g.[2,4,8])proposedtoseekgoodapproximationtoxvia`1minimisation.InthispaperweshowthatinthecaseofGaussianmeasurementsitrecoversthesignalwellfrominacuratemeasurements,thusimprovingresultfrom[4].Wealsoshowthatwithbigp3、robabilityitgivesinformationcomparablewithbestktermapproximationineuclideannorm,kd=lnN.Thisprovidestherstnumericallyfriendlyalgorithmtodoso,see[7].1IntroductionCompressedsensingisanewschemewhichshowsthatsomesignalscanbere-constructedfromfewermeasurementsthatpr4、eviouslywereconsiderednecessary.Themathematicalformulationisthefollowing.Oursignalisavectorx2RN.WehaveaNdmatrixcalledmeasurementmatrixandourmeasurementsarerepresentedbyy=(x)2Rd.Wealsoneedadecoder(whichmaybenon-linear)whichproduces(y)2RNwhichshouldbeanapprox5、imationtox.Themainpointincompressedsensingasexpressedinrecentpapersisthatitisactuallypossibletorecovertheessentialinformationaboutxfromrelativelyfewnon-adaptivemeasurementsd<6、thorwouldliketoexpresshisgratitudetoPiotrMankiewiczforansweringquestionsaboutconvexgeometryandtoAlbertCohen,WolfgangDahmenandRonDeVoreforteachinghimcompressedsensing.ThanksarealsoduetoRachelWardforinterestingdiscussionaboutstabilityresults.Thisresearchwasmadepos7、siblebyECMarieCurieToKprogramSPADE-2atIMPAN1recentyearsinunderstandingtheperformanceofvariousmeasurementmatricesanddecoders.Generallywehavealsoanintegerkdwhichmeasurestheamountofinformationwewishtorecover.Thestandardinitialrequirementisthatforeveryk{sparsevec8、tor(i.e.x2k)wehave((x))=x.Thisclearlyforcesjktobeonetoone.Butfoetobenumericallyfriendlywemusthavethecorrespondingsystemsofequationswellconditioned.Thisleadstoth
2、velinearmeasurements.Candes,Donoho,Taoet.al.(seee.g.[2,4,8])proposedtoseekgoodapproximationtoxvia`1minimisation.InthispaperweshowthatinthecaseofGaussianmeasurementsitrecoversthesignalwellfrominacuratemeasurements,thusimprovingresultfrom[4].Wealsoshowthatwithbigp
3、robabilityitgivesinformationcomparablewithbestktermapproximationineuclideannorm,kd=lnN.Thisprovidestherstnumericallyfriendlyalgorithmtodoso,see[7].1IntroductionCompressedsensingisanewschemewhichshowsthatsomesignalscanbere-constructedfromfewermeasurementsthatpr
4、eviouslywereconsiderednecessary.Themathematicalformulationisthefollowing.Oursignalisavectorx2RN.WehaveaNdmatrixcalledmeasurementmatrixandourmeasurementsarerepresentedbyy=(x)2Rd.Wealsoneedadecoder(whichmaybenon-linear)whichproduces(y)2RNwhichshouldbeanapprox
5、imationtox.Themainpointincompressedsensingasexpressedinrecentpapersisthatitisactuallypossibletorecovertheessentialinformationaboutxfromrelativelyfewnon-adaptivemeasurementsd<6、thorwouldliketoexpresshisgratitudetoPiotrMankiewiczforansweringquestionsaboutconvexgeometryandtoAlbertCohen,WolfgangDahmenandRonDeVoreforteachinghimcompressedsensing.ThanksarealsoduetoRachelWardforinterestingdiscussionaboutstabilityresults.Thisresearchwasmadepos7、siblebyECMarieCurieToKprogramSPADE-2atIMPAN1recentyearsinunderstandingtheperformanceofvariousmeasurementmatricesanddecoders.Generallywehavealsoanintegerkdwhichmeasurestheamountofinformationwewishtorecover.Thestandardinitialrequirementisthatforeveryk{sparsevec8、tor(i.e.x2k)wehave((x))=x.Thisclearlyforcesjktobeonetoone.Butfoetobenumericallyfriendlywemusthavethecorrespondingsystemsofequationswellconditioned.Thisleadstoth
6、thorwouldliketoexpresshisgratitudetoPiotrMankiewiczforansweringquestionsaboutconvexgeometryandtoAlbertCohen,WolfgangDahmenandRonDeVoreforteachinghimcompressedsensing.ThanksarealsoduetoRachelWardforinterestingdiscussionaboutstabilityresults.Thisresearchwasmadepos
7、siblebyECMarieCurieToKprogramSPADE-2atIMPAN1recentyearsinunderstandingtheperformanceofvariousmeasurementmatricesanddecoders.Generallywehavealsoanintegerkdwhichmeasurestheamountofinformationwewishtorecover.Thestandardinitialrequirementisthatforeveryk{sparsevec
8、tor(i.e.x2k)wehave((x))=x.Thisclearlyforcesjktobeonetoone.Butfoetobenumericallyfriendlywemusthavethecorrespondingsystemsofequationswellconditioned.Thisleadstoth
此文档下载收益归作者所有