Theoretical Analysis of Cross-Validation for Estimating the Risk of the k-nearest Neighbor Classifier

Size: px
Start display at page:

Download "Theoretical Analysis of Cross-Validation for Estimating the Risk of the k-nearest Neighbor Classifier"

Transcription

1 Joural of Machi Larig Rsarch Submittd 09/15; Rvisd 07/18; Publishd 11/18 Thortical Aalysis of Cross-Validatio for Estimatig th Ris of th -Narst Nighbor Classifir Alai Cliss Laboratoir d Mathématiqus UMR 854 CNRS-Uivrsité d Lill Iria Modal Projct-tam, Lill F Villuv d Ascq Cdx, Frac cliss@math.uiv-lill1.fr Trista Mary-Huard maryhuar@agroparistch.fr INRA, UMR 030 / UMR 810 Géétiqu Quatitativ t Évolutio L Moulo, F Gif-sur-Yvtt, Frac UMR MIA-Paris, AgroParisTch, INRA, Uivrsité Paris-Saclay F-75005, Paris, Frac Editor: Hui Zou Abstract Th prst wor aims at drivig thortical guaratis o th bhavior of som crossvalidatio procdurs applid to th -arst ighbors NN rul i th cotxt of biary classificatio. Hr w focus o th lav-p-out cross-validatio LpO usd to assss th prformac of th NN classifir. Rmarably this LpO stimator ca b fficitly computd i this cotxt usig closd-form formulas drivd by Cliss ad Mary-Huard 011. W dscrib a gral stratgy to driv momt ad xpotial coctratio iqualitis for th LpO stimator applid to th NN classifir. Such rsults ar obtaid first by xploitig th coctio btw th LpO stimator ad U-statistics, ad scod by maig a itsiv us of th gralizd Efro-Sti iquality applid to th L1O stimator. O othr importat cotributio is mad by drivig w quatificatios of th discrpacy btw th LpO stimator ad th classificatio rror/ris of th NN classifir. Th optimality of ths bouds is discussd by mas of svral lowr bouds as wll as simulatio xprimts. Kywords: Classificatio, Cross-validatio, Ris stimatio 1. Itroductio Th -arst ighbor NN algorithm Fix ad Hodgs, 1951 i biary classificatio is a popular prdictio algorithm basd o th ida that th prdictd valu at a w poit is basd o a majority vot from th arst labld ighbors of this poit. Although quit simpl, th NN classifir has b succssfully applid to may difficult classificatio tass Li t al., 004; Simard t al., 1998; Schirr ad Slay, 003. Efficit implmtatios hav b also dvlopd to allow dalig with larg datasts Idy ad Motwai, 1998; Adoi ad Idy, 006. Th thortical prformacs of th NN classifir hav b alrady xtsivly ivstigatd. I th cotxt of biary classificatio prlimiary thortical rsults dat bac to c 018 Alai Cliss ad Trista Mary-Huard. Lics: CC-BY 4.0, s Attributio rquirmts ar providd at

2 Cliss ad Mary-Huard Covr ad Hart 1967; Covr 1968; Györfi Th NN classifir has b provd to b waly uivrsally cosistt by Sto 1977 as log as = + ad / 0 as +. For th 1NN classifir, a asymptotic xpasio of th rror rat has b drivd by Psaltis t al Th sam stratgy has b succssfully applid to th NN classifir by Sapp ad Vatsh Hall t al. 008 study th ifluc of th paramtr o th ris of th NN classifir by mas of a asymptotic xpasio drivd from a Poisso or biomial modl for th traiig poits. Mor rctly, Caigs t al. 017 poitd out som limitatios suffrd by th classical NN classifir ad dducd a improvd vrsio basd o a local choic of i th smi-suprvisd cotxt. I cotrast to th aformtiod rsults, th wor by Chaudhuri ad Dasgupta 014 focuss o th fiit-sampl framwor. Thy typically provid uppr bouds with high probability o th ris of th NN classifir whr th bouds ar ot distributio-fr. Altrativly i th rgrssio sttig, Kulari ad Posr 1995 drivd a stratgy ladig to a fiit-sampl boud o th prformac of 1NN, which has b xtdd to th wightd NN rul 1 by Biau t al. 010a,b s also Brrtt t al., 016, whr a wightd NN stimator is dsigd for stimatig th tropy. W rfr itrstd radrs to Biau ad Dvroy 016 for a almost thorough prstatio of ow rsults o th NN algorithm i various cotxts. I umrous if ot all practical applicatios, computig th cross-validatio CV stimator Sto, 1974, 198 has b amog th most popular stratgis to valuat th prformac of th NN classifir Dvroy t al., 1996, Sctio 4.3. All CV procdurs shar a commo pricipl which cosists i splittig a sampl of poits ito two disjoit substs calld traiig ad tst sts with rspctiv cardialitis p ad p, for ay 1 p 1. Th p traiig st data srv to comput a classifir, whil its prformac is valuatd from th p lft out data of th tst st. For a complt ad comprhsiv rviw o cross-validatio procdurs, w rfr th itrstd radr to Arlot ad Cliss 010. I th prst wor, w focus o th lav-p-out LpO cross-validatio. Amog CV procdurs, it blogs to xhaustiv stratgis sic it cosidrs ad avrags ovr all th p possibl such splittigs of {1,..., } ito traiig ad tst sts. Usually th iducd computatio tim of th LpO is prohibitiv, which givs ris to its surrogat calld V fold cross-validatio V-FCV with V /p Gissr, Howvr, Stl 009; Cliss ad Mary-Huard 011 rctly drivd closd-form formulas rspctivly for th bootstrap ad th LpO procdurs applid to th NN classifir. Such formulas allow for a fficit computatio of th LpO stimator. Morovr sic th V-FCV stimator suffrs th sam bias but a largr variac tha th LpO o Cliss ad Robi, 008; Arlot ad Cliss, 010, LpO with p = /V strictly improvs upo V-FCV i th prst cotxt. Although big favord i practic for assssig th ris of th NN classifir, th us of CV coms with vry fw thortical guarats rgardig its prformac. Morovr probably for tchical rasos, most xistig rsults apply to Hold-out ad lav-o-out L1O, that is LpO with p = 1 Kars ad Ro, I this papr w rathr cosidr th gral LpO procdur for 1 p 1 usd to stimat th ris altrativly th classificatio rror rat of th NN classifir. Our mai purpos is th to provid distributio-fr thortical guarats o th bhavior of LpO with rspct to iflutial paramtrs such as p,, ad. For istac w aim at aswrig qustios such as: Dos

3 Prformac of CV to Estimat th Ris of NN thr xist ay rgim of p = p with p som fuctio of whr th LpO stimator is a cosistt stimat of th ris of th NN classifir?, or Is it possibl to dscrib th covrgc rat of th LpO stimator dpdig o p? Cotributios. Th mai cotributio of th prst wor is two-fold: i w dscrib a w gral stratgy to driv momt ad xpotial coctratio iqualitis for th LpO stimator applid to th NN biary classifir, ad ii ths iqualitis srv to driv th covrgc rat of th LpO stimator towards th ris of th NN classifir. This w stratgy rlis o svral stps. First xploitig th coctio btw th LpO stimator ad U-statistics Korolju ad Borovsich, 1994 ad th Rosthal iquality Ibragimov ad Sharahmtov, 00, w prov that uppr boudig th polyomial momts of th ctrd LpO stimator rducs to drivig such bouds for th simplr L1O stimator. Scod, w driv w uppr bouds o th momts of th L1O stimator usig th gralizd Efro-Sti iquality Bouchro t al., 005, 013, Thorm Third, combiig th two prvious stps provids som isight o th itrplay btw p/ ad i th coctratio rats masurd i trms of momts. This fially rsults i w xpotial coctratio iqualitis for th LpO stimator applyig whatvr th valu of th ratio p/ 0, 1. I particular whil th uppr bouds icras with 1 p / + 1, it is o logr th cas if p > / + 1. W also provid svral lowr bouds suggstig our uppr bouds caot b improvd i som ss i a distributio-fr sttig. Th rmaidr of th papr is orgaizd as follows. Th coctio btw th LpO stimator ad U-statistics is clarifid i Sctio, whr w also rcall th closd-form formula of th LpO stimator applid to th NN classifir Cliss ad Mary-Huard, 011. Ordr-q momts q of th LpO stimator ar th uppr boudd i trms of thos of th L1O stimator. This stp ca b applid to ay classificatio algorithm. Sctio 3 th spcifis th prvious uppr bouds i th cas of th NN classifir, which lads to th mai Thorm 3. charactrizig th coctratio bhavior of th LpO stimator with rspct to p,, ad i trms of polyomial momts. Drivig xpotial coctratio iqualitis for th LpO stimator is th mai cocr of Sctio 4 whr w highlight th strgth of our stratgy by comparig our mai iqualitis with coctratio iqualitis drivd with lss sophisticatd tools. Fially Sctio 5 xploits th prvious rsults to boud th gap btw th LpO stimator ad th classificatio rror of th NN classifir. Th optimality of ths uppr bouds is first provd i our distributio-fr framwor by stablishig svral w lowr bouds matchig th uppr os i som spcific sttigs. Scod, mpirical xprimts ar also rportd which support th abov coclusios.. U-statistics ad LpO stimator.1. Statistical framwor Classificatio W tacl th biary classificatio problm whr th goal is to prdict th uow labl Y {0, 1} of a obsrvatio X X R d. Th radom variabl X, Y has a uow joit distributio P X,Y dfid by P X,Y B = P X, Y B ] for ay Borlia st B X {0, 1}, whr P dots a rfrc probability distributio. I what follows o particular distributioal assumptio is mad rgardig X. To prdict th labl, o aims at buildig a classifir ˆf : X {0, 1} o th basis of a st of radom variabls 3

4 Cliss ad Mary-Huard D = {Z 1,..., Z } calld th traiig sampl, whr Z i = X i, Y i, 1 i rprst copis of X, Y draw idpdtly from P X,Y. I sttigs whr o cofusio is possibl, w will rplac D by D. Ay stratgy to build such a classifir is calld a classificatio algorithm, ad ca b formally dfid as a fuctio A : 1 {X {0, 1}} F that maps a traiig sampl D oto th corrspodig classifir A D = ˆf F, whr F is th st of all masurabl fuctios from X to {0, 1}. Numrous classifirs hav b cosidrd i th litratur ad it is out of th scop of th prst papr to rviw all of thm s Dvroy t al for may istacs. Hr w focus o th -arst ighbor rul NN iitially proposd by Fix ad Hodgs 1951 ad furthr studid for istac by Dvroy ad Wagr 1977; Rogrs ad Wagr Th NN algorithm For 1, th NN classificatio algorithm, dotd by A, cosists i classifyig ay w obsrvatio x usig a majority vot dcisio rul basd o th labls of th closst poits to x, dotd by X 1 x,..., X x, amog th traiig sampl X 1,..., X. I what follows ths arst ighbors ar chos accordig to th distac associatd with th usual Euclida orm i R d. Not that othr adaptiv mtrics hav b also cosidrd i th litratur s for istac Hasti t al., 001, Chap. 14. But such xampls ar out of th scop of th prst wor, that is our rfrc distac dos ot dpd o th traiig sampl at had. Lt us also mphasiz that possibl tis ar bro by usig th smallst idx amog tis, which is o possibl choic for th Sto lmma to hold tru Biau ad Dvroy, 016, Lmma 10.6, p.15. Formally, giv V x = { 1 i, X i { X 1 x,..., X x }} th st of idics of th arst ighbors of x amog X 1,..., X, th NN classifir is dfid by 1, if 1 A D ; x = f i V x Y i = 1 Y ix > 0.5 D ; x := 0, if 1 Y ix < 0.5,.1 B0.5, othrwis whr Y i x is th labl of th i-th arst ighbor of x for 1 i, ad B0.5 dots a Broulli radom variabl with paramtr 1/. Lav-p-out cross-validatio For a giv sampl D, th prformac of ay classifir ˆf = A D rspctivly of ay classificatio algorithm A is assssd by th classificatio rror L ˆf rspctivly th ris R ˆf dfid by L ˆf = P ˆfX Y D, ad R ˆf ] = E P ˆfX Y D. I this papr w focus o th stimatio of L ˆf ad its xpctatio R ˆf by us of th Lav-p-Out LpO cross-validatio for 1 p 1 Zhag, 1993; Cliss ad Robi, 008. LpO succssivly cosidrs all possibl splits of D ito a traiig st of cardiality p ad a tst st of cardiality p. Dotig by E p th st of all possibl substs of {1,..., } with cardiality p, ay E p dfis a split of D ito a traiig sampl D = {Z i i } ad a tst sampl Dē, whr ē = {1,..., } \. For a giv classificatio algorithm A, th fial LpO stimator of th prformac of A D = f is th avrag ovr all possibl splits of th classificatio rror stimatd o ach tst st, that is 1 1 R p A, D = 1 p p {A D X i Y i},. E p i ē 4

5 Prformac of CV to Estimat th Ris of NN whr A D is th classifir built from D. W rfr th radr to Arlot ad Cliss 010 for a dtaild dscriptio of LpO ad othr cross-validatio procdurs. I th squl, th lgthy otatio R p A, D is rplacd by R p, i sttigs whr o cofusio ca aris about th algorithm A or th traiig sampl D, ad by R p D if th traiig sampl has to b pt i mid. Exact LpO for th NN classificatio algorithm Usually du to its smigly prohibitiv computatioal cost, LpO is ot applid xcpt with p = 1 whr it rducs to th wll ow lav-o-out. Howvr i svral cotxts such as dsity stimatio Cliss ad Robi, 008; Cliss, 014 or rgrssio Cliss, 008, closd-form formulas hav b drivd for th LpO stimator wh applid with projctio ad rl stimators. Th NN classifir is aothr istac of such stimators for which fficitly computig th LpO stimator is possibl. Its computatio rquirs a tim complxity that is liar i p as prviously stablishd by Cliss ad Mary-Huard 011. Lt us brifly rcall th mai stps ladig to th closd-form formula. 1. From Eq.. th LpO stimator ca b xprssd as a sum ovr th obsrvatios of th complt sampl of probabilitis: p p {A D X i Y i} = p p {A D X i Y i} 1 {i/ } E p i/ E p = 1 P A D X i Y i i / P i /. p Hr P mas that th itgratio is mad with rspct to th radom variabl E p, which follows th uiform distributio ovr th p possibl substs i E p with cardiality p. For istac P i / = p/ sic it is th proportio of subsampls with cardiality p which do ot cotai a giv prscribd idx i, which quals 1 p / p. S also Lmma D.4 for furthr xampls of such calculatios.. For ay X i, lt X 1,..., X +p 1, X +p,..., X 1 b th ordrd squc of ighbors of X i. This list dpds o X i, that is X 1 should b otd X i,1. But this dpdcy is sippd hr for th sa of radability. Th y i th drivatio is to coditio with rspct to th radom variabl R i which dots th ra i th whol sampl D of th -th ighbor of X i i th D. For istac R i = j mas that X j is th -th ighbor of X i i D. Th P A D X i Y i i / = +p 1 j= P A D X i Y i R i = j, i / P R i = j i /, whr th sum ivolvs p trms sic oly X,..., X +p 1 ar cadidats for big th -th ighbor of X i i at last o traiig subst. 3. Obsrv that th rsultig probabilitis ca b asily computd s Lmma D.4: P i / = p P R i = j i / = j P U = j P A D X i Y i V i = j, i / = 1 Y j 1 F +1 ] H + Yj 1 FH 5 1 ],

6 Cliss ad Mary-Huard with U Hj, j 1, p 1, H HN j i, j N j i 1, 1, ad H HN j i 1, j N j i, 1, whr F H ad F H rspctivly dot th cumulativ distributio fuctios of H ad H, H dots th hyprgomtric distributio, ad N j i is th umbr of 1 s amog th j arst ighbors of X i i D. Th computatioal cost of LpO for th NN classifir is th sam as that of L1O for th + p 1NN classifir whatvr p, that is Op. This cotrasts with th usual p prohibitiv computatioal complxity smigly suffrd by LpO... U-statistics: Gral bouds o LpO momts Th purpos of th prst sctio is to dscrib a gral stratgy allowig to driv w uppr bouds o th polyomial momts of th LpO stimator. As a first stp of this stratgy, w stablish th coctio btw th LpO ris stimator ad U-statistics. Scod, w xploit this coctio to driv w uppr bouds o th ordr-q momts of th LpO stimator for q. Not that ths uppr bouds, which rlat momts of th LpO stimator to thos of th L1O stimator, hold tru with ay classifir. Lt us start by itroducig U-statistics ad rcallig som of thir basic proprtis that will srv our purposs. For a thorough prstatio, w rfr to th boos by Srflig 1980; Korolju ad Borovsich Th first stp is th dfiitio of a U-statistic of ordr m N as a avrag ovr all m-tupls of distict idics i {1,..., }. Dfiitio.1 Korolju ad Borovsich Lt h : X m R dot ay masurabl fuctio whr m 1 is a itgr. Lt us furthr assum h is a symmtric fuctio of its argumts. Th ay fuctio U : X R such that U x 1,..., x = U hx 1,..., x = whr m, is a U-statistic of ordr m ad rl h. 1 h x i1,..., x im m 1 i 1 <...<i m Bfor clarifyig th coctio btw LpO ad U-statistics, lt us itroduc th mai proprty of U-statistics our stratgy rlis o. It cosists i rprstig ay U-statistic as a avrag, ovr all prmutatios, of sums of idpdt variabls. Propositio.1 Eq. 5.5 i Hoffdig With th otatio of Dfiitio.1, lt us dfi W : X R by W x 1,..., x = 1 r r h x j 1m+1,..., x jm,.3 j=1 whr r = /m dots th itgr part of /m. Th U x 1,..., x = 1 W x! σ1,..., x σ, σ whr σ dots th summatio ovr all prmutatios σ of {1,..., }. 6

7 Prformac of CV to Estimat th Ris of NN W ar ow i positio to stat th y rmar of th papr. All th dvlopmts furthr xposd i th followig rsult from this coctio btw th LpO stimator dfid by Eq.. ad U-statistics. Thorm.1. For ay classificatio algorithm A ad ay 1 p 1 such that a classifir ca b computd from A o p traiig poits, th LpO stimator R p, is a U-statistic of ordr m = p + 1 with rl h m : X m R dfid by h m Z 1,..., Z m = 1 m m 1 { A Di m X i Y i }, whr D i m dots th sampl D m = Z 1,..., Z m with Z i withdraw. Not for istac that wh A = A dots th NN algorithm, th cardiality of D i m has to satisfy p, which implis that 1 p 1. Proof of Thorm.1. From Eq.., th LpO stimator of th prformac of ay classificatio algorithm A computd from D satisfis R p A, D = R p, = 1 1 p p E p = 1 1 p p E p 1 {A D X i Y i} i ē i ē v E p+1 1 {v= {i}} 1 {A D X i Y i}, sic thr is a uiqu st of idics v with cardiality p + 1 such that v = {i}. Th R p, = 1 1 p p v E p+1 1 {v= {i}} 1 {i ē} 1 { }. A Dv\{i} X i Y i E p Furthrmor for v ad i fixd, E p 1 {v= {i}} 1 {i ē} = 1 {i v} sic thr is a uiqu st of idics such that = v \ i. O gts R p, = 1 1 p p = 1 p+1 v E p+1 by oticig p p = p! p! p! =! p 1! p! 1 {i v} 1 { } A Dv\{i} X i Y i 1 p + 1 v E p+1 = p { }, A Dv\{i} X i Y i i v p+1. 7

8 Cliss ad Mary-Huard Th rl h m is a dtrmiistic ad symmtric fuctio of its argumts that dos oly dpd o m. Lt us also otic that h m Z 1,..., Z m rducs to th L1O stimator of th ris of th classifir A computd from Z 1,..., Z m, that is h m Z 1,..., Z m = R 1 A, D m = R 1, p+1..4 I th cotxt of tstig whthr two biary classifirs hav diffrt rror rats, this fact has alrady b poitd out by Fuchs t al W ow driv a gral uppr boud o th q-th momt q 1 of th LpO stimator that holds tru for ay classifir as log as th followig xpctatios ar wll dfid. Thorm.. For ay classifir A, lt A D ad A Dm b th corrspodig classifirs built from rspctivly D ad D m, whr m = p + 1. Th for vry 1 p 1 such that a classifir ca b computd from A o p traiig poits, ad for ay q 1, ] q] ] q ] E Rp, E Rp, E R1,m E R1,m..5 Furthrmor as log as p > / + 1, o also gts for q = E Rp, E Rp, ] ] ] ] E R1,m E R1,m m.6 for vry q > ] q] E Rp, E Rp, Bq, γ ] q R 1,m E R1,m q max q γ E m, m Var R1,m m,.7 whr γ > 0 is a umric costat ad Bq, γ dots th optimal costat dfid i th Rosthal iquality Propositio D.. Th proof is giv i Appdix A.1. Eq..5 ad.6 straightforwardly rsult from th Js iquality applid to th avrag ovr all prmutatios providd i Propositio.1. If p > / + 1, th itgr part /m bcoms largr tha 1 ad Eq..6 bcoms bttr tha Eq..5 for q =. As a cosquc of our stratgy of proof, th right-had sid of Eq..6 is qual to th classical uppr boud o th variac of U-statistics which suggsts it caot b improvd without addig furthr assumptios. Uli th abov os, Eq..7 is drivd from th Rosthal iquality, which abls us to uppr boud a sum r ξ i q of idpdt ad idtically ctrd radom variabls i trms of r ξ i q ad r Varξ i. Lt us rmar that, for q =, both trms of th right-had sid of Eq..7 ar of th sam ordr as Eq..6 up to costats. Furthrmor usig th Rosthal iquality allows taig advatag of th itgr part /m wh p > / + 1 uli what w gt by usig Eq..5 for q >. I particular it provids a w udrstadig of th bhavior of th LpO stimator wh p/ 1 as highlightd latr by Propositio 4.. 8

9 Prformac of CV to Estimat th Ris of NN 3. Nw bouds o LpO momts for th NN classifir Our goal is ow to spcify th gral uppr bouds providd by Thorm. i th cas of th NN algorithm A 1 itroducd by.1. Sic Thorm. xprsss th momts of th LpO stimator i trms of thos of th L1O stimator computd from D m with m = p + 1, th xt stp cosists i focusig o th L1O momts. Drivig uppr bouds o th momts of th L1O is achivd usig a gralizatio of th wll-ow Efro-Sti iquality s Thorm D.1 for Efro- Sti s iquality ad Thorm 15.5 i Bouchro t al. 013 for its gralizatio. For th sa of compltss, w first rcall a corollary of this gralizatio that is provd i Sctio D.1.4 s Corollary D.1. Propositio 3.1. Lt ξ 1,..., ξ dot idpdt Ξ-valud radom variabls ad ζ = fξ 1,..., ξ, whr f : Ξ R is ay masurabl fuctio. With ξ 1,..., ξ idpdt copis of th ξ i s, thr xists a uivrsal costat κ 1.71 such that for ay q, ζ Eζ q κq fξ 1,..., ξ i,..., ξ fξ 1,..., ξ i,..., ξ q/. Th applyig Propositio 3.1 with ζ = R 1 A, D m = R 1,m L1O stimator computd from D m with m = p + 1 ad Ξ = R d {0, 1} lads to th followig Thorm 3.1. It cotrols th ordr-q momts of th L1O stimator applid to th NN classifir. Thorm 3.1. For vry 1 1, lt A Dm m = p + 1 dot th NN classifir lart from D m ad R 1,m b th corrspodig L1O stimator giv by Eq... Th for q =, ] ] 3/ E R1,m E R1,m C 1 m ; 3.1 for vry q >, ] q ] E R1,m E R1,m C q q m q/, 3. with C 1 = + 16γ d ad C = 4γ d κ, whr γd is a costat arisig from Sto s lmma, s Lmma D.5 that grows xpotially with dimsio d, ad κ is dfid i Propositio 3.1. Its proof dtaild i Sctio A. rlis o Sto s lmma Lmma D.5. For a giv X i, it provs that th umbr of poits i D i havig X i amog thir arst ighbors is ot largr tha γ d. Th dpdc of our uppr bouds with rspct to γ d s xplicit costats C 1 ad C iducs thir strog dtrioratio as th dimsio d grows sic γ d 4.8 d 1. Thrfor th largr th dimsio d, th largr th rquird sampl siz for th uppr boud to b small at last smallr tha 1. Not also that th ti braig stratgy basd o th smallst idx i th prst wor is chos so that it surs Sto s lmma to hold tru. 9

10 Cliss ad Mary-Huard I Eq. 3.1, th asir cas q = abls to xploit xact calculatios ] of rathr tha uppr bouds o th variac of th L1O stimator. Sic E R1,m = R ris of A D p th NN classifir computd from D p, th rsultig 3/ /m rat is a strict improvmt upo th usual /m that is drivd from usig th sub-gaussia xpotial coctratio iquality provd by Thorm 4.4 i Dvroy t al By cotrast th largr q arisig i Eq. 3. rsults from th difficulty to driv a tight uppr boud for th xpctatio of } q with q >, whr D i 1{ A Di m X i A Di,j m X i rsp. D m i,j dots th sampl D m whr Z i has b rsp. Z i ad Z j hav b rmovd. W ar ow i positio to stat th mai rsult of this sctio. It follows from th combiatio of Thorm. coctig momts of th LpO ad L1O stimators ad Thorm 3.1 providig a uppr boud o th ordr-q momts of th L1O. Thorm 3.. For vry p, 1 such that p+, lt R p, dot th LpO ris stimator s. of th NN classifir A D dfid by.1. Th thr xist ow costats C 1, C > 0 such that for vry 1 p, for q =, ] ] 3/ E Rp, E Rp, C 1 p + 1 ; 3.3 m for vry q >, ] q ] E Rp, E Rp, C q q p+1 q/, 3.4 with C 1 = 18κγ d π ad C = 4γ d κ, whr γd dots th costat arisig from Sto s lmma Lmma D.5. Furthrmor i th particular sttig whr / + 1 < p, th for q =, for vry q >, ] ] 3/ E Rp, E Rp, C 1 p + 1 p+1, 3.5 ] q ] E Rp, E Rp, p + 1 Γ q 3/ max p + 1 p+1 q, p + 1 p+1 q 3 q/ 3.6, whr Γ = max C 1, C. 10

11 Prformac of CV to Estimat th Ris of NN Th straightforward proof is dtaild i Sctio A.3. Lt us start by oticig that both uppr bouds i Eq. 3.3 ad 3.4 dtriorat as p grows. This is o logr th cas for Eq. 3.5 ad 3.6, which ar spcifically dsigd to covr th stup whr p > / + 1, that is whr /m is o logr qual to 1. Thrfor uli Eq. 3.3 ad 3.4, ths last two iqualitis ar particularly rlvat i th stup whr p/ 1, as +, which has b ivstigatd by Shao 1993; Yag 006, 007; Cliss 014. Eq. 3.5 ad 3.6 lad to rspctiv covrgc rats at wors 3/ / for q = ad q / q 1 for q >. I particular this last rat bcoms approximatly qual to / q as q gts larg. O ca also mphasiz that, as a U-statistic of fixd ordr m = p + 1, th LpO stimator has a ow Gaussia limitig distributio, that is s Thorm A, Sctio Srflig, 1980 m Rp, E Rp, ] L N 0, σ + 1, whr σ 1 = Var gz 1 ], with gz = E h m z, Z,..., Z m ]. Thrfor th uppr boud giv by Eq. 3.5 is o-improvabl i som ss with rspct to th itrplay btw ad p sic o rcovrs th right magitud for th variac trm as log as m = p + 1 is assumd to b costat. Fially Eq. 3.6 has b drivd usig a spcific vrsio of th Rosthal iquality Ibragimov ad Sharahmtov, 00 statd with th optimal costat ad ivolvig a balacig factor. I particular this balacig factor has allowd us to optimiz th rlativ wight of th two trms btw bracts i Eq This lads us to claim that th dpdc of th uppr boud with rspct to q caot b improvd with this li of proof. Howvr w caot coclud that th trm i q 3 caot b improvd usig othr tchical argumts. 4. Expotial coctratio iqualitis This sctio provids xpotial coctratio iqualitis for th LpO stimator applid to th NN classifir. Our mai rsults havily rly o th momt iqualitis prviously drivd i Sctio 3, amly Thorm 3.. I ordr to mphasiz th gai allowd by this stratgy of proof, w start this sctio by succssivly provig two xpotial iqualitis obtaid with lss sophisticatd tools. W th discuss th strgth ad wass of ach of thm to justify th additioal rfimts w itroduc stp by stp alog th sctio. A first xpotial coctratio iquality for R p A, D = R p, ca b drivd by us of th boudd diffrc iquality followig th li of proof of Dvroy t al. 1996, Thorm 4.4 origially dvlopd for th L1O stimator. Propositio 4.1. For ay itgrs p, 1 such that p +, lt R p, dot th LpO stimator. of th classificatio rror of th NN classifir A D dfid by.1. Th for vry t > 0, P Rp, E Rp, > t t 8+p 1 γ d. 4.1 whr γ d dots th costat itroducd i Sto s lmma Lmma D.5. 11

12 Cliss ad Mary-Huard Th proof is giv i Appdix B.1. Th uppr boud of Eq. 4.1 strogly xploits th facts that: i for X j to b o of th arst ighbors of X i i at last o subsampl X, it rquirs X j to b o of th + p 1 arst ighbors of X i i th complt sampl, ad ii th umbr of poits for which X j may b o of th + p 1 arst ighbors caot b largr tha + p 1γ d by Sto s Lmma s Lmma D.5. This rasoig rsults i a rough uppr boud sic th domiator i th xpot xhibits a + p 1 factor whr ad p play th sam rol. Th raso is that w do ot distiguish btw poits for which X j is amog or abov th arst ighbors of X i i th whol sampl although ths two stups lad to highly diffrt probabilitis of big amog th arst ighbors i th traiig sampl. Cosqutly th dpdc of th covrgc rat o ad p i Propositio 4.1 ca b improvd, as cofirmd by forthcomig Thorms 4.1 ad 4.. Basd o th prvious commts, a sharpr quatificatio of th ifluc of ach ighbor amog th + p 1 os lads to th xt rsult. Thorm 4.1. For vry p, 1 such that p +, lt R p, dot th LpO stimator. of th classificatio rror of th NN classifir A D dfid by.1. Th thr xists a umric costat > 0 such that for vry t > 0, max P Rp, E Rp, > t, P E Rp, R t p, > t xp ], p p 1 1 with = 104κ1+γ d, whr γ d is itroducd i Lmma D.5 ad κ 1.71 is a uivrsal costat. Th proof is giv i Sctio B.. Uli Propositio 4.1, taig ito accout th ra of ach ighbor i th whol sampl abls us to cosidrably rduc th wight of p compard to that of i th domiator of th xpot. I particular, lttig p/ 0 as + with assumd to b fixd for istac mas th ifluc of th + p factor asymptotically gligibl. This would allow for rcovrig up to umric costats a similar uppr boud to that of Dvroy t al. 1996, Thorm 4.4, achivd with p = 1. Howvr th uppr boud of Thorm 4.1 dos ot rflct th right dpdcis with rspct to ad p compard with what has b provd for polyomial momts i Thorm 3.. I particular it dtriorats as p icrass uli th uppr bouds drivd for p > / + 1 i Thorm 3.. This drawbac is ovrcom by th followig rsult, which is our mai cotributio i th prst sctio. Thorm 4.. For vry p, 1 such that p+, lt R p, dot th LpO stimator of th classificatio rror of th NN classifir ˆf = A D dfid by.1. Th for vry t > 0, ] ] max P Rp, E Rp, > t, P E Rp, R t p, > t xp p + 1, 4. 1

13 Prformac of CV to Estimat th Ris of NN whr = 4 max C, C 1 with C1, C > 0 dfid i Thorm 3.1. Furthrmor i th particular sttig whr p > / + 1, it coms ] max P Rp, E Rp, xp 1 mi p + 1 p + 1 ] > t, P E Rp, R p, > t t 4Γ 3/, p + 1 p + 1 p + 1 1/3 t, 4Γ 4.3 whr Γ ariss i Eq. 3.6 ad γ d Lmma D.5. dots th costat itroducd i Sto s lmma Th proof has b postpod to Appdix B.3. It ivolvs diffrt argumts for drivig th two iqualitis 4. ad 4.3 dpdig o th rag of valus of p. Firstly for p /+1, a simpl argumt is applid to driv Iq. 4. from th two corrspodig momt iqualitis of Thorm 3. charactrizig th sub-gaussia bhavior of th LpO stimator i trms of its v momts s Lmma D.. Scodly for p > / + 1, w rathr xploit: i th appropriat uppr bouds o th momts of th LpO stimator giv by Thorm 3., combid with ii Propositio D.1 which stablishs xpotial coctratio iqualitis from gral momt uppr bouds. I accordac with th coclusios draw about Thorm 3., th uppr boud of Eq. 4. icrass as p grows uli that of Eq Th bst coctratio rat i Eq. 4.3 is achivd as p/ 1, whras Eq. 4. turs out to b uslss i that sttig. Howvr Eq. 4. rmais strictly bttr tha Thorm 4.1 as log as p/ δ 0, 1 as +. Not also that th costats Γ ad γ d ar th sam as i Thorm 3.1. Thrfor th sam commts rgardig thir dpdc with rspct to th dimsio d apply hr. I ordr to facilitat th itrprtatio of th last Iq. 4.3, w also driv th followig propositio provd i Appdix B.3 which focuss o th dscriptio of ach dviatio trm i th particular cas whr p > / + 1. Propositio 4.. With th sam otatio as Thorm 4., for ay p, 1 such that p +, p > / + 1, ad for vry t > 0 P R ] Γ p, E Rp, > p + 1 3/ p+1 whr Γ > 0 is th costat arisig from 3.6. t + p+1 t 3/ p + 1 Th prst iquality is vry similar to th wll-ow Brsti iquality Bouchro t al., 013, Thorm.10 xcpt th scod dviatio trm of ordr t 3/ istad of t for th Brsti iquality. With rspct to, th first dviatio trm is of ordr 3/ /, which is th sam as with th Brsti iquality. Th scod dviatio trm is of a somwhat diffrt ordr, that is p + 1/, as compard with th usual 1/ i th Brsti iquality. t, 13

14 Cliss ad Mary-Huard Nvrthlss w almost rcovr th / rat by choosig for istac p 1 log /, which lads to log /. Thrfor varyig p allows to itrpolat btw th / ad th / rats. Not also that th dpdc of th first sub-gaussia dviatio trm with rspct to is oly 3/, which improvs upo th usual rsultig from Iq. 4. i Thorm 4. for istac. Howvr this 3/ rmais crtaily too larg for big optimal v if this qustio rmais widly op at this stag i th litratur. Mor grally o strgth of our approach is its vrsatility. Idd th two abov dviatio trms dirctly rsult from th two uppr bouds o th momts of th L1O stablishd i Thorm 3.1. Thrfor ay improvmt of th lattr uppr bouds would immdiatly lad to hac th prst coctratio iquality without chagig th proof. 5. Assssig th gap btw LpO ad classificatio rror 5.1. Uppr bouds First, w driv w uppr bouds o diffrt masurs of th discrpacy btw R p, = R p A, D ad th classificatio rror L ˆf or th ris R ˆf = E L ˆf ]. Ths bouds o th LpO stimator ar compltly w for p > 1, som of thm big xtsios of formr os spcifically drivd for th L1O stimator applid to th NN classifir. Thorm 5.1. For vry p, 1 such that p ad +, lt R p, dot th LpO ris stimator s. of th NN classifir ˆf = A D dfid by.1. Th, ] E Rp, R ˆf 4 p π, 5.1 ad E Rp, R ˆf ] 18κγ d 3/ π p p π 5. Morovr, E Rp, L ˆf ] p π 5.3 I cotrast to th rsults i th prvious sctios, a w rstrictio o p ariss i Thorm 5.1, that is p. This coms from th us of Lmma D.6 provd by Dvroy ad Wagr 1979b, which givs a uppr boud o th L 1 stability of th NN classifir wh p obsrvatios ar rmovd from th traiig sampl D. Actually this uppr boud oly rmais maigful as log as 1 p. 14

15 Prformac of CV to Estimat th Ris of NN Proof of Thorm 5.1. Proof of 5.1: With ˆf = AD, Lmma D.6 immdiatly provids E Rp, L ˆf ] E = L ˆf ] E L ˆf ] 1{A E D X Y } 1 {A D = P A D X AD X 4 p π X Y } ] Proof of 5.: Th proof combis th prvious uppr boud with th o stablishd for th variac of th LpO stimator, that is Eq E Rp, E L ˆf ] ] ] ] ] = E Rp, E Rp, + E Rp, E L ˆf ] 18κγ d 3/ π p p, π which cocluds th proof. Th proof of Iq. 5.3 is mor itricat ad has b postpod to Appdix C.1. ] Kpig i mid that E Rp, = RA D p, th right-had sid of Iq. 5.1 is a uppr boud o th bias of th LpO stimator, that is o th diffrc btw th riss of th classifirs built from rspctivly p ad poits. Thrfor, th fact that this uppr boud icrass with p is rliabl sic th classifirs A D p+1 ad A D ca bcom mor ad mor diffrt from o aothr as p icrass. Mor prcisly, th uppr boud i Iq. 5.1 gos to 0 providd p / dos. With th additioal rstrictio p, this rducs to th usual coditio / 0 as + s Dvroy t al., 1996, Chap. 6.6 for istac, which is usd to prov th uivrsal cosistcy of th NN classifir Sto, Th mootoicity of this uppr uppr boud with rspct to ca sm somwhat uxpctd. O could thi that th two classifirs would bcom mor ad mor similar to ach othr as icrass ough. Howvr it ca b provd that, i som ss, this dpdc caot b improvd i th prst distributio-fr framwor s Propositio 5.1 ad Figur 1. Not that a uppr boud similar to that of Iq. 5. ca b asily drivd for ay ordr-q momt q at th pric of icrasig th costats by usig a + b q q 1 a q + b q, for vry a, b 0. W also mphasiz that Iq. 5. allows us to cotrol th discrpacy btw th LpO stimator ad th ris of th NN classifir, that is th xpctatio of its classificatio rror. Idally w would hav lid to rplac th ris R ˆf by th prdictio rror L ˆf. But with our stratgy of proof, this would rquir a additioal distributio-fr coctratio iquality o th prdictio rror of th NN classifir. To th bst of our owldg, such a coctratio iquality is ot availabl up to ow. Uppr boudig th squard diffrc btw th LpO stimator ad th prdictio rror is prcisly th purpos of Iq Provig th lattr iquality rquirs a compltly diffrt stratgy which ca b tracd bac to a arlir proof by Rogrs ad Wagr 15

16 Cliss ad Mary-Huard 1978, s th proof of Thorm.1 applyig to th L1O stimator. Lt us mtio that Iq. 5.3 combid with th Js iquality lad to a lss accurat uppr boud tha Iq Fially th appart diffrc btw th uppr bouds i Iq. 5. ad 5.3 rsults from th compltly diffrt schms of proof. Th first o allows us to driv gral uppr bouds for all ctrd momts of th LpO stimator, but xhibits a wors dpdc with rspct to. By cotrast th scod o is xclusivly ddicatd to uppr boudig th ma squard diffrc btw th prdictio rror ad th LpO stimator ad lads to a smallr. Howvr v if probably ot optimal, th uppr boud usd i Iq. 5. still abls to achiv miimax rats ovr som Höldr balls as provd by Propositio Lowr bouds Bias of th L1O stimator Th purpos of th xt rsult is to provid a coutr-xampl highlightig that th uppr boud of Eq. 5.1 caot b improvd i som ss. W cosidr th followig discrt sttig whr X = {0, 1} with π 0 = P X = 0 ], ad w dfi η 0 = P Y = 1 X = 0 ] ad η 1 = P Y = 1 X = 1 ]. I what follows this two-class grativ modl will b rfrrd to as th discrt sttig DS. Not that i th 3 paramtrs π 0, η 0 ad η 1 fully dscrib th joit distributio P X,Y, ad ii th distributio of DS satisfis th strog margi assumptio of Massart ad Nédélc 006 if both η 0 ad η 1 ar chos away from 1/. Howvr this favourabl sttig has o particular ffct o th forthcomig lowr boud xcpt a fw simplificatios alog th calculatios. Propositio 5.1. Lt us cosidr th DS sttig with π 0 = 1/, η 0 = 0 ad η 1 = 1, ad assum that is odd. Th thr xists a umric costat C > 1 idpdt of ad such that, for all / 1, th NN classifirs A D ad A D 1 satisfy E L A D L A D 1 ] C Th proof of Propositio 5.1 is providd i Appdix C.. Th rat / i th righthad sid of Eq. 5.1 is th achivd udr th grativ modl DS for ay /. As a cosquc this rat caot b improvd without ay additioal assumptio, for istac o th distributio of th X i s. S also Figur 1 blow ad rlatd commts. Empirical illustratio To furthr illustrat th rsult of Propositio 5.1, w simulatd data accordig to DS, for diffrt valus of ragig from 100 to 500 ad diffrt valus of ragig from 5 to 1. Figur 1 a displays th volutio of th absolut bias E L A D L A D 1 ] as a fuctio of, for svral valus of plai curvs. Th absolut bias is a odcrasig fuctio of, as suggstd by th uppr boud providd i Eq. 5.1 which is also plottd dashd lis to as th compariso. Th o-dcrasig bhavior of th absolut bias is ot always rstrictd to high valus of w.r.t., as illustratd i Figur 1 b which 16

17 Prformac of CV to Estimat th Ris of NN corrspods to DS with paramtr valus π 0, η 0, η 1 = 0., 0., 0.9. I particular th o-dcrasig bhavior of th absolut bias ow appars for a rag of valus of that ar smallr tha /. Not that a rough ida about th locatio of th pa, dotd by pa, ca b dducd as follows i th simpl cas whr η 0 = 0 ad η 1 = 1. For th pa to aris, th two classifirs basd o ad rspctivly 1 obsrvatios hav to disagr th most strogly. This rquirs o of th two classifirs say th first o to hav tis amog th arst ighbors of ach labl i at last o of th two cass X = 0 or X = 1. With π 0 < 0.5, th tis will most lily occur for th cas X = 0. Thrfor th discrpacy btw th two classifirs will b th highst at ay w obsrvatio x 0 = 0. For th ti situatio to aris at x 0, half of its ighbors hav to b 1. This oly occurs if i > 0 with 0 th umbr of obsrvatios such that X = 0 i th traiig st, ad ii 0 η η 1 = /, whr 0 rsp. 1 is th umbr of ighbors of x 0 such that X = 0 rsp. X = 1. Sic > 0, o has 0 = 0 ad th last xprssio boils dow to = 0η 1 η 0 η 1 1/ For larg valus of, o should hav 0 π 0, that is th pa should appar at pa π 0η 1 η 0 η 1 1/ I th sttig of Propositio 5.1, this rasoig rmarably yilds pa, whil it lads to pa 0.4 i th sttig of Figur 1 b, which is clos to th locatio of th obsrvd pas. This also suggsts that v smallr valus of pa ca aris by tuig th paramtr π 0 clos to 0. Lt us mtio that vry similar curvs hav b obtaid for a Gaussia mixtur modl with two disjoit classs ot rportd hr. O th o had this mpirically illustrats that th / rat is ot limitd to DS discrt sttig. O th othr had, all of this cofirms that this rat caot b improvd i th prst distributio-fr framwor. Lt us fially cosidr Figur 1 c, which displays th absolut bias as a fuctio of whr = Cof for diffrt valus of Cof, whr dots th itgr part. With this choic of, Propositio 5.1 implis that th absolut bias should dcras at a 1/ rat, which is supportd by th plottd curvs. By cotrast, pal d of Figur 1 illustrats that choosig smallr valus of, that is = Cof, lads to a fastr dcrasig rat Ma squard rror Followig a xampl dscribd by Dvroy ad Wagr 1979a, w ow provid a lowr boud o th miimal covrgc rat of th ma squard rror s also Dvroy t al., 1996, Chap. 4.4, p.415 for a similar argumt. 17

18 Cliss ad Mary-Huard Bias Bias a b Bias Bias Cof c Cof d Figur 1: a Evolutio of th absolut valu of th bias as a fuctio of, for diffrt valus of plai lis. Th dashd lis corrspod to th uppr boud obtaid i 5.1. b Sam as prvious, xcpt that data wr gratd accordig to th DS sttig with paramtrs π 0, η 0, η 1 = 0., 0., 0.9. Uppr bouds ar ot displayd i ordr to fit th scal of th absolut bias. c Evolutio of th absolut valu of th bias with rspct to, wh is chos such that = Cof dots th itgr part. Th diffrt colors corrspod to diffrt valus of Cof. d Sam as prvious, xcpt that is chos such that = Cof. 18

19 Prformac of CV to Estimat th Ris of NN Propositio 5.. Lt us assum is v, ad that P Y = 1 X = P Y = 1 = 1/ is idpdt of X. Th for = 1 odd, it rsults E R1, L ˆf ] = 1 0 t P R1, L ˆf ] > t dt 1 8 π 1 From th uppr boud of ordr / providd by Iq. 5.3 with p = 1, choosig = 1 lads to th sam 1/ rat as that of Propositio 5.. This suggsts that, at last for vry larg valus of, th / rat is of th right ordr ad caot b improvd i th distributio-fr framwor Miimax rats Lt us coclud this sctio with a corollary, which provids a fiit-sampl boud o th gap btw R p, ad R ˆf = E L ˆf ] with high probability. It is statd udr th sam rstrictio o p as th prvious Thorm 5.1 it is basd o, that is for p. Corollary 5.1. With th otatio of Thorms 4. ad 5.1, lt us assum p, 1 with p, +, ad p / + 1. Th for vry x > 0, thr xists a vt with probability at last 1 x such that R ˆf R p, 1 p 1 x + 4 π p, 5.4 whr ˆf = A D. Proof of Corollary 5.1. Iq. 5.4 rsults from combiig th xpotial coctratio rsult drivd for R p,, amly Iq. 4. from Thorm 4. ad th uppr boud o th bias, that is Iq R ˆf R ] ] R E p, ˆf E Rp, + Rp, R p, 4 π p + p + 1 x Not that th right-had sid of Iq. 5.4 could b usd to driv bouds o R ˆf that sm similar to cofidc bouds. Howvr w do ot rcommd doig this i practic for svral rasos. O th o had, Iq. 5.4 rsults from th rpatd us of coctratio iqualitis whr umric costats ar ot optimizd at all. This would lad to rquir a larg sampl siz for th dviatio trms to b small i practic. O th othr had, xplicit umric costats such as i Corollary 5.1 xhibit a dpdc o γ d 4.8 d 1, which bcoms xpotially larg as d icrass. Provig that this dpdc ca b wad or ot rmais a compltly op qustio at this stag. Nvrthlss o ca highlight that, for a giv, icrasig d will quicly ma th dviatio trm largr tha 1, whras both R ˆf ad R p, blog to 0, 1]. 19

20 Cliss ad Mary-Huard Th right-most trm of ordr / i Iq. 5.4 rsults from th bias. This is a cssary pric to pay which caot b improvd i th prst distributio-fr framwor accordig to Propositio 5.1. Bsids combiig th rstrictio p with th usual cosistcy costrait / = o1 lads to th coclusio that small valus of p w.r.t. hav almost o ffct o th covrgc rat of th LpO stimator. Waig th y rstrictio p would b cssary to pottially uac this coclusio. I ordr to highlight th itrst of th abov dviatio iquality, lt us dduc a optimality rsult i trms of miimax rat ovr Höldr balls Hτ, α dfid by { Hτ, α = g : R d R, gx gy τ x y α}, with α ]0, 1 ad τ > 0. I th followig statmt, Corollary 5.1 is usd to prov that, uiformly with rspct to, th LpO stimator R p, ad th ris R ˆf of th NN classifir rmai clos to ach othr with high probability. Propositio 5.3. With th sam otatio as Corollary 5.1, for vry C > 1 ad θ > 0, thr xists a vt of probability at last 1 C 1 o which, for ay p, 1 such that p, +, ad p /+1, th LpO stimator of th NN classifir satisfis 1 θ R ˆf ] L θ 1 C 4 log R ˆf 4 p L π R p A, D L 1 + θ R ˆf ] L + θ 1 C 4 log R ˆf + 4 p L π, 5.5 whr L dots th classificatio rror of th Bays classifir. Furthrmor if o assums th rgrssio fuctio η blogs to a Höldr ball Hτ, α for som α ]0, mid/4, 1 rcall that X i R d ad τ > 0, th choosig = = 0 α α+d lads to R p A, D L + R ˆf L, a.s Iq. 5.5 givs a uiform cotrol ovr of th gap btw th xcss ris R ˆf L ad th corrspodig LpO stimator R p ˆf L with high probability. Th dcrasig rat i C 1 of this probability is dirctly rlatd to th log factor i th lowr ad uppr bouds. This dcrasig rat could b mad fastr at th pric of icrasig th xpot of th log factor. I a similar way th umric costat θ has o prcis maig ad ca b chos as clos to 0 as w wat, ladig to icras o of th othr dviatio trms by a umric factor θ 1. For istac o could choos θ = 1/ log, which would rplac th log by a log. Th quivalc stablishd by Eq. 5.6 rsults from owig that this choic = mas th NN classifir achiv th miimax rat α α+d ovr Höldr balls Yag, This holds tru for α ]0, 1 as log as d 4. Howvr if d < 4 th miimax rat is oly achivd ovr ]0, d/4. This limitatio rsults from th dpdc of th dviatio trms with rspct to i Eq. 5.5, which is ot optimal ad should b furthr improvd. 0

21 Prformac of CV to Estimat th Ris of NN Proof of Propositio 5.3. Lt us dfi K as th maximum valu of ad assum x = C log for som costat C > 1 for ay 1 K. Lt us also itroduc th vt Ω = { Th P Ω c ] 1 C 1 1 K, K x = =1 R ˆf R p A, D x + 4 p π 0, as +, sic a uio boud lads to K =1 C log = K C log C 1 log = 1 C 1 Furthrmor combiig for a, b > 0 th iquality ab a θ + b θ /4 for vry θ > 0 with a + b a + b, it rsults that hc Iq x θ R ˆf L + θ 1 4 R ˆf L x θ R ˆf L + θ 1 4 R ˆf C log, L Lt us ow prov th xt quivalc, amly 5.6, by mas of th Borl-Catlli lmma. First Yag 1999 combid with Thorm 7 i Chaudhuri ad Dasgupta 014 provid that th miimax rat ovr th Höldr ball Hτ, α is achivd by th NN classifir with =, that is R ˆf L α α+d, whr a b mas thr xist umric costats l, u > 0 such that l b a u b. Morovr it is th asy to chc that D 1 := θ 1 4 R ˆf L C log Cθ d 3α α+d log = o + R ˆf L, D := p = 0 d α+d = o + R ˆf L. Bsids 1 C 1 P Rp A, D L R ˆf L Ωc ] P R ˆf L > θ + D 1 + D R ˆf L Rp A, D L = P R ˆf 1 L > θ + D 1 + D R ˆf. L }. 1

22 Cliss ad Mary-Huard Lt us ow choos ay ɛ > 0 ad itroduc th squc of vts {A ɛ} 1 such that Rp A, D L A ɛ = R ˆf 1 L > ɛ. Usig that D 1 ad D ar gligibl with rspct to R ˆf L as +, thr xists a itgr 0 = 0 ɛ > 0 such that, for all 0 ad with θ = ɛ/, θ + D 1 + D R ˆf L ɛ. Hc Rp A, D L P A ɛ ] P R ˆf 1 L > θ + D 1 + D R ˆf 1 L C 1 Fially, choosig ay C > lads to + =1 P A ɛ ] < +, which provids th xpctd coclusio by mas of th Borl-Catlli lmma. 6. Discussio Th prst wor provids svral w rsults quatifyig th prformac of th LpO stimator applid to th NN classifir. By xploitig th coxio btw LpO ad U- statistics Sctio, th polyomial ad xpotial iqualitis drivd i Sctios 3 ad 4 giv som w isight o th coctratio of th LpO stimator aroud its xpctatio for diffrt rgims of p/. I Sctio 5, ths rsults srv for istac to coclud to th cosistcy of th LpO stimator towards th ris or th classificatio rror rat of th NN classifir Thorm 5.1. Thy also allow us to stablish th asymptotic quivalc btw th LpO stimator shiftd by th Bays ris L ad th xcss ris ovr som Höldr class of rgrssio fuctios Propositio 5.3. It is worth mtioig that th uppr-bouds drivd i Sctios 4 ad 5 s for istac Thorm 5.1 ca b miimizd by choosig p = 1, suggstig that th L1O stimator is optimal i trms of ris stimatio wh applid to th NN classificatio algorithm. This obsrvatio corroborats th rsults of th simulatio study prstd i Cliss ad Mary-Huard 011, whr it is mpirically show that small valus of p ad i particular p = 1 lad to th bst stimatio of th ris for ay fixd, whatvr th lvl of ois i th data. Th suggstd optimality of L1O for ris stimatio is also cosistt with rsults by Burma 1989 ad Cliss 014, whr it is provd that L1O is asymptotically th bst cross-validatio procdur to prform ris stimatio i th cotxt of low-dimsioal rgrssio ad dsity stimatio rspctivly. Altrativly, th LpO stimator ca also b usd as a data-dpdt calibratio procdur to tu, by choosig th valu ˆ p which miimizs th LpO stimat. For istac

23 Prformac of CV to Estimat th Ris of NN i classificatio, LpO ca b usd to gt th valu of ladig to th bst prdictio prformac. I this cotxt th valu of p th splittig ratio ladig to th bst NN classifir ca b vry diffrt from p = 1. This is illustratd by th simulatio rsults summarizd by Figur i Cliss ad Mary-Huard 011 whr p has to b largr tha 1 as th ois lvl bcoms strog. This phomo is ot limitd to th NN classifir, but xtds to various stimatio/prdictio problms Brima ad Spctor, 199; Arlot ad Lrasl, 01; Cliss, 014. If w tur ow to th qustio of idtifyig th bst prdictor amog svral cadidats, choosig p = 1 also lads to poor slctio prformacs as provd by Shao 1993, Eq. 3.8 with th liar rgrssio modl. For th LpO, Shao 1997, Thorm 5 provs th modl slctio cosistcy if p/ 1 ad p + as +. For rcovrig th bst prdictor amog two cadidats, Yag 006, 007 provd th cosistcy of CV udr coditios rlatig th optimal splittig ratio p to th covrgc rats of th prdictors to b compard, ad furthr rquirig that mip, p + as +. Although th focus of th prst papr is diffrt, it is worth mtioig that th coctratio rsults stablishd i Sctio 4 ar a sigificat arly stp towards drivig thortical guarats o LpO as a modl slctio procdur. Idd, xpotial coctratio iqualitis hav b a y igrdit to assss modl slctio cosistcy or modl slctio fficicy i various cotxts s for istac Cliss 014 or Arlot ad Lrasl 01 i th dsity stimatio framwor. Still thortically ivstigatig th bhavior of ˆ p rquirs som furthr ddicatd dvlopmts. O first stp towards such rsults is to driv a tightr uppr boud o th bias btw th LpO stimator ad th ris. Th bst ow uppr boud currtly availabl is drivd from Dvroy ad Wagr 1980, s Lmma D.6 i th prst papr. Ufortuatly it dos ot fully captur th tru bhavior of th LpO stimator with rspct to p at last as p bcoms larg ad could b improvd i particular for p > as mphasizd i th commts followig Thorm 5.1. Aothr importat dirctio for studyig th modl slctio bhavior of th LpO procdur is to prov a coctratio iquality for th classificatio rror rat of th NN classifir aroud its xpctatio. Whil such coctratio rsults hav b stablishd for th NN algorithm i th fixd-dsig rgrssio framwor Arlot ad Bach, 009, drivig similar rsults i th classificatio cotxt rmais a challgig problm to th bst of our owldg. Acowldgmts Th authors would li to tha th associat ditor ad rviwrs for thir highlightful commts which gratly hlpd to improv th prstatio of th papr. This wor was partially fudd by th BFast PEPS CNRS. 3

24 Cliss ad Mary-Huard Appdix A. Proofs of polyomial momt uppr bouds A.1. Proof of Thorm. Th proof rlis o Propositio.1 that allows to rlat th LpO stimator to a sum of idpdt radom variabls. I th followig, w distiguish btw th two sttigs q = whr xact calculatios ca b carrid out, ad q > whr oly uppr bouds ca b drivd. Wh q >, our proof dals sparatly with th cass p / + 1 ad p > / + 1. I th first o, a straightforward us of Js s iquality lads to th rsult. I th scod sttig, o has to b mor cautious wh drivig uppr bouds. This is do by usig th mor sophisticatd Rosthal s iquality, amly Propositio D.. A.1.1. Exploitig Propositio.1 Accordig to th proof of Propositio.1, it ariss that th LpO stimator ca b xprssd as a U-statistic sic with R p, = 1 W Z! σ1,..., Z σ, σ W Z 1,..., Z = ad h m Z 1,..., Z m = 1 m m 1 m m a=1 h m Za 1m+1,..., Z am 1 { A Di m X i Y i } = R 1, p+1, with m = p + 1 whr A Di m. dots th classifir basd o sampl D i m = Z 1,..., Z i 1, Z i+1,..., Z m. Furthr ctrig th LpO stimator, it coms ] R p, E Rp, = 1! σ W Z σ1,..., Z σ, whr W Z 1,..., Z = W Z 1,..., Z E W Z 1,..., Z ]. Th with h m Z 1,..., Z m = h m Z 1,..., Z m E h m Z 1,..., Z m ], o gts ] q] E Rp, E Rp, E W Z1,..., Z q ] Js s iquality 1 m = E m q h m Zi 1m+1,..., Z im q q m = E h m Zi 1m+1,..., Z im. m A.1 4

25 Prformac of CV to Estimat th Ris of NN A.1.. Th sttig q = If q =, th by idpdc it coms ] q] m E Rp, E Rp, Var h m Zi 1m+1,..., Z im m which lads to th rsult. A.1.3. Th sttig q > = m m Var h m Zi 1m+1,..., Z im ] 1 = Var R1 A, Z 1, p+1, m If p / + 1: A straightforward us of Js s iquality from A.1 provids ] q] E Rp, E Rp, m 1 m = E R1, p+1 E E hm Zi 1m+1,..., Z im q] R1, p+1 ] q]. If p > / + 1: Lt us ow us Rosthal s iquality Propositio D. by itroducig symmtric radom variabls ζ 1,..., ζ /m such that 1 i /m, ζ i = h m Zi 1m+1,..., Z im hm Zi 1m+1,..., Z im, whr Z 1,..., Z ar i.i.d. copis of Z 1,..., Z. Th it coms for vry γ > 0 q m m E h m Zi 1m+1,..., Z im E ζ i q, which implis q m E h m Zi 1m+1,..., Z im Bq, γ max γ m E ζ i q m ], E ζi q ]. Th usig for vry i that E ζ i q ] q E hm Zi 1m+1,..., Z im q ], 5

26 Cliss ad Mary-Huard it coms q m E h m Zi 1m+1,..., Z im Bq, γ max q γ ] q ] E R1,m E R1,m, m q Var R1,m. m Hc, it rsults for vry q > ] q] E Rp, E Rp, q+1 ] Bq, γ max q q ] q/ γ E R1,m E R1,m, Var R1,m m m q, which cocluds th proof. A.. Proof of Thorm 3.1 Our stratgy of proof follows svral idas. Th first o cosists i usig Propositio 3.1 which says that, for vry q, hm Z 1,..., Z m q m κq h m Z 1,..., Z m h m Z 1,..., Z j m,..., Z, q/ j=1 whr h m Z 1,..., Z m = R 1,m by Eq..4, ad h m Z 1,..., Z m = h m Z 1,..., Z m E h m Z 1,..., Z m ]. Th scod ida cosists i drivig uppr bouds of j h m = h m Z 1,..., Z j,..., Z m h m Z 1,..., Z j,..., Z m by rpatd uss of Sto s lmma, that is Lmma D.5 which uppr bouds by γ d th maximum umbr of X i s that ca hav a giv X j amog thir arst ighbors. Fially, for tchical rasos w hav to distiguish th cas q = whr w gt tightr bouds ad q >. A..1. Uppr boudig j h m For th sa of radability lt us ow us th otatio D i = D m i s Thorm.1, ad lt D i j dot th st Z 1,..., Z j,..., Z whr th i-th coordiat has b rmovd. Th, j h m = h m Z 1,..., Z m h m Z 1,..., Z j,..., Z m is ow uppr boudd by j h m 1 m m { } 1 A i j Di X i Y { } i A Di j X i Y i 1 m m { } i j A Di X i A D i. A. j X i 6

Theoretical analysis of cross-validation for estimating the risk of the k-nearest Neighbor classifier

Theoretical analysis of cross-validation for estimating the risk of the k-nearest Neighbor classifier Thortical aalysis of cross-validatio for stimatig th ris of th -Narst Nighbor classifir Alai Cliss, Trista Mary-Huard To cit this vrsio: Alai Cliss, Trista Mary-Huard. Thortical aalysis of cross-validatio

More information

Statistics 3858 : Likelihood Ratio for Exponential Distribution

Statistics 3858 : Likelihood Ratio for Exponential Distribution Statistics 3858 : Liklihood Ratio for Expotial Distributio I ths two xampl th rjctio rjctio rgio is of th form {x : 2 log (Λ(x)) > c} for a appropriat costat c. For a siz α tst, usig Thorm 9.5A w obtai

More information

z 1+ 3 z = Π n =1 z f() z = n e - z = ( 1-z) e z e n z z 1- n = ( 1-z/2) 1+ 2n z e 2n e n -1 ( 1-z )/2 e 2n-1 1-2n -1 1 () z

z 1+ 3 z = Π n =1 z f() z = n e - z = ( 1-z) e z e n z z 1- n = ( 1-z/2) 1+ 2n z e 2n e n -1 ( 1-z )/2 e 2n-1 1-2n -1 1 () z Sris Expasio of Rciprocal of Gamma Fuctio. Fuctios with Itgrs as Roots Fuctio f with gativ itgrs as roots ca b dscribd as follows. f() Howvr, this ifiit product divrgs. That is, such a fuctio caot xist

More information

On the approximation of the constant of Napier

On the approximation of the constant of Napier Stud. Uiv. Babş-Bolyai Math. 560, No., 609 64 O th approximatio of th costat of Napir Adri Vrscu Abstract. Startig from som oldr idas of [] ad [6], w show w facts cocrig th approximatio of th costat of

More information

DTFT Properties. Example - Determine the DTFT Y ( e ) of n. Let. We can therefore write. From Table 3.1, the DTFT of x[n] is given by 1

DTFT Properties. Example - Determine the DTFT Y ( e ) of n. Let. We can therefore write. From Table 3.1, the DTFT of x[n] is given by 1 DTFT Proprtis Exampl - Dtrmi th DTFT Y of y α µ, α < Lt x α µ, α < W ca thrfor writ y x x From Tabl 3., th DTFT of x is giv by ω X ω α ω Copyright, S. K. Mitra Copyright, S. K. Mitra DTFT Proprtis DTFT

More information

Solution to 1223 The Evil Warden.

Solution to 1223 The Evil Warden. Solutio to 1 Th Evil Ward. This is o of thos vry rar PoWs (I caot thik of aothr cas) that o o solvd. About 10 of you submittd th basic approach, which givs a probability of 47%. I was shockd wh I foud

More information

APPENDIX: STATISTICAL TOOLS

APPENDIX: STATISTICAL TOOLS I. Nots o radom samplig Why do you d to sampl radomly? APPENDI: STATISTICAL TOOLS I ordr to masur som valu o a populatio of orgaisms, you usually caot masur all orgaisms, so you sampl a subst of th populatio.

More information

PURE MATHEMATICS A-LEVEL PAPER 1

PURE MATHEMATICS A-LEVEL PAPER 1 -AL P MATH PAPER HONG KONG EXAMINATIONS AUTHORITY HONG KONG ADVANCED LEVEL EXAMINATION PURE MATHEMATICS A-LEVEL PAPER 8 am am ( hours) This papr must b aswrd i Eglish This papr cosists of Sctio A ad Sctio

More information

Discrete Fourier Transform (DFT)

Discrete Fourier Transform (DFT) Discrt Fourir Trasorm DFT Major: All Egirig Majors Authors: Duc guy http://umricalmthods.g.us.du umrical Mthods or STEM udrgraduats 8/3/29 http://umricalmthods.g.us.du Discrt Fourir Trasorm Rcalld th xpotial

More information

Law of large numbers

Law of large numbers Law of larg umbrs Saya Mukhrj W rvisit th law of larg umbrs ad study i som dtail two typs of law of larg umbrs ( 0 = lim S ) p ε ε > 0, Wak law of larrg umbrs [ ] S = ω : lim = p, Strog law of larg umbrs

More information

INTRODUCTION TO SAMPLING DISTRIBUTIONS

INTRODUCTION TO SAMPLING DISTRIBUTIONS http://wiki.stat.ucla.du/socr/id.php/socr_courss_2008_thomso_econ261 INTRODUCTION TO SAMPLING DISTRIBUTIONS By Grac Thomso INTRODUCTION TO SAMPLING DISTRIBUTIONS Itro to Samplig 2 I this chaptr w will

More information

2617 Mark Scheme June 2005 Mark Scheme 2617 June 2005

2617 Mark Scheme June 2005 Mark Scheme 2617 June 2005 Mark Schm 67 Ju 5 GENERAL INSTRUCTIONS Marks i th mark schm ar plicitly dsigatd as M, A, B, E or G. M marks ("mthod" ar for a attmpt to us a corrct mthod (ot mrly for statig th mthod. A marks ("accuracy"

More information

A Simple Proof that e is Irrational

A Simple Proof that e is Irrational Two of th most bautiful ad sigificat umbrs i mathmatics ar π ad. π (approximatly qual to 3.459) rprsts th ratio of th circumfrc of a circl to its diamtr. (approximatly qual to.788) is th bas of th atural

More information

Chapter 2 Infinite Series Page 1 of 11. Chapter 2 : Infinite Series

Chapter 2 Infinite Series Page 1 of 11. Chapter 2 : Infinite Series Chatr Ifiit Sris Pag of Sctio F Itgral Tst Chatr : Ifiit Sris By th d of this sctio you will b abl to valuat imror itgrals tst a sris for covrgc by alyig th itgral tst aly th itgral tst to rov th -sris

More information

10. Joint Moments and Joint Characteristic Functions

10. Joint Moments and Joint Characteristic Functions 0 Joit Momts ad Joit Charactristic Fctios Followig sctio 6 i this sctio w shall itrodc varios paramtrs to compactly rprst th iformatio cotaid i th joit pdf of two rvs Giv two rvs ad ad a fctio g x y dfi

More information

Option 3. b) xe dx = and therefore the series is convergent. 12 a) Divergent b) Convergent Proof 15 For. p = 1 1so the series diverges.

Option 3. b) xe dx = and therefore the series is convergent. 12 a) Divergent b) Convergent Proof 15 For. p = 1 1so the series diverges. Optio Chaptr Ercis. Covrgs to Covrgs to Covrgs to Divrgs Covrgs to Covrgs to Divrgs 8 Divrgs Covrgs to Covrgs to Divrgs Covrgs to Covrgs to Covrgs to Covrgs to 8 Proof Covrgs to π l 8 l a b Divrgt π Divrgt

More information

Digital Signal Processing, Fall 2006

Digital Signal Processing, Fall 2006 Digital Sigal Procssig, Fall 6 Lctur 9: Th Discrt Fourir Trasfor Zhg-Hua Ta Dpartt of Elctroic Systs Aalborg Uivrsity, Dar zt@o.aau.d Digital Sigal Procssig, I, Zhg-Hua Ta, 6 Cours at a glac MM Discrt-ti

More information

Derivation of a Predictor of Combination #1 and the MSE for a Predictor of a Position in Two Stage Sampling with Response Error.

Derivation of a Predictor of Combination #1 and the MSE for a Predictor of a Position in Two Stage Sampling with Response Error. Drivatio of a Prdictor of Cobiatio # ad th SE for a Prdictor of a Positio i Two Stag Saplig with Rspos Error troductio Ed Stak W driv th prdictor ad its SE of a prdictor for a rado fuctio corrspodig to

More information

Restricted Factorial And A Remark On The Reduced Residue Classes

Restricted Factorial And A Remark On The Reduced Residue Classes Applid Mathmatics E-Nots, 162016, 244-250 c ISSN 1607-2510 Availabl fr at mirror sits of http://www.math.thu.du.tw/ am/ Rstrictd Factorial Ad A Rmark O Th Rducd Rsidu Classs Mhdi Hassai Rcivd 26 March

More information

H2 Mathematics Arithmetic & Geometric Series ( )

H2 Mathematics Arithmetic & Geometric Series ( ) H Mathmatics Arithmtic & Gomtric Sris (08 09) Basic Mastry Qustios Arithmtic Progrssio ad Sris. Th rth trm of a squc is 4r 7. (i) Stat th first four trms ad th 0th trm. (ii) Show that th squc is a arithmtic

More information

Session : Plasmas in Equilibrium

Session : Plasmas in Equilibrium Sssio : Plasmas i Equilibrium Ioizatio ad Coductio i a High-prssur Plasma A ormal gas at T < 3000 K is a good lctrical isulator, bcaus thr ar almost o fr lctros i it. For prssurs > 0.1 atm, collisio amog

More information

FORBIDDING RAINBOW-COLORED STARS

FORBIDDING RAINBOW-COLORED STARS FORBIDDING RAINBOW-COLORED STARS CARLOS HOPPEN, HANNO LEFMANN, KNUT ODERMANN, AND JULIANA SANCHES Abstract. W cosidr a xtrmal problm motivatd by a papr of Balogh [J. Balogh, A rmark o th umbr of dg colorigs

More information

A Note on Quantile Coupling Inequalities and Their Applications

A Note on Quantile Coupling Inequalities and Their Applications A Not o Quatil Couplig Iqualitis ad Thir Applicatios Harriso H. Zhou Dpartmt of Statistics, Yal Uivrsity, Nw Hav, CT 06520, USA. E-mail:huibi.zhou@yal.du Ju 2, 2006 Abstract A rlatioship btw th larg dviatio

More information

Chapter Five. More Dimensions. is simply the set of all ordered n-tuples of real numbers x = ( x 1

Chapter Five. More Dimensions. is simply the set of all ordered n-tuples of real numbers x = ( x 1 Chatr Fiv Mor Dimsios 51 Th Sac R W ar ow rard to mov o to sacs of dimsio gratr tha thr Ths sacs ar a straightforward gralizatio of our Euclida sac of thr dimsios Lt b a ositiv itgr Th -dimsioal Euclida

More information

1985 AP Calculus BC: Section I

1985 AP Calculus BC: Section I 985 AP Calculus BC: Sctio I 9 Miuts No Calculator Nots: () I this amiatio, l dots th atural logarithm of (that is, logarithm to th bas ). () Ulss othrwis spcifid, th domai of a fuctio f is assumd to b

More information

Triple Play: From De Morgan to Stirling To Euler to Maclaurin to Stirling

Triple Play: From De Morgan to Stirling To Euler to Maclaurin to Stirling Tripl Play: From D Morga to Stirlig To Eulr to Maclauri to Stirlig Augustus D Morga (186-1871) was a sigificat Victoria Mathmaticia who mad cotributios to Mathmatics History, Mathmatical Rcratios, Mathmatical

More information

Lectures 9 IIR Systems: First Order System

Lectures 9 IIR Systems: First Order System EE3054 Sigals ad Systms Lcturs 9 IIR Systms: First Ordr Systm Yao Wag Polytchic Uivrsity Som slids icludd ar xtractd from lctur prstatios prpard by McCllla ad Schafr Lics Ifo for SPFirst Slids This work

More information

NEW VERSION OF SZEGED INDEX AND ITS COMPUTATION FOR SOME NANOTUBES

NEW VERSION OF SZEGED INDEX AND ITS COMPUTATION FOR SOME NANOTUBES Digst Joural of Naomatrials ad Biostructurs Vol 4, No, March 009, p 67-76 NEW VERSION OF SZEGED INDEX AND ITS COMPUTATION FOR SOME NANOTUBES A IRANMANESH a*, O KHORMALI b, I NAJAFI KHALILSARAEE c, B SOLEIMANI

More information

Discrete Fourier Transform. Nuno Vasconcelos UCSD

Discrete Fourier Transform. Nuno Vasconcelos UCSD Discrt Fourir Trasform uo Vascoclos UCSD Liar Shift Ivariat (LSI) systms o of th most importat cocpts i liar systms thory is that of a LSI systm Dfiitio: a systm T that maps [ ito y[ is LSI if ad oly if

More information

An Introduction to Asymptotic Expansions

An Introduction to Asymptotic Expansions A Itroductio to Asmptotic Expasios R. Shaar Subramaia Asmptotic xpasios ar usd i aalsis to dscrib th bhavior of a fuctio i a limitig situatio. Wh a fuctio ( x, dpds o a small paramtr, ad th solutio of

More information

Chapter Taylor Theorem Revisited

Chapter Taylor Theorem Revisited Captr 0.07 Taylor Torm Rvisitd Atr radig tis captr, you sould b abl to. udrstad t basics o Taylor s torm,. writ trascdtal ad trigoomtric uctios as Taylor s polyomial,. us Taylor s torm to id t valus o

More information

Chapter (8) Estimation and Confedence Intervals Examples

Chapter (8) Estimation and Confedence Intervals Examples Chaptr (8) Estimatio ad Cofdc Itrvals Exampls Typs of stimatio: i. Poit stimatio: Exampl (1): Cosidr th sampl obsrvatios, 17,3,5,1,18,6,16,10 8 X i i1 17 3 5 118 6 16 10 116 X 14.5 8 8 8 14.5 is a poit

More information

ln x = n e = 20 (nearest integer)

ln x = n e = 20 (nearest integer) H JC Prlim Solutios 6 a + b y a + b / / dy a b 3/ d dy a b at, d Giv quatio of ormal at is y dy ad y wh. d a b () (,) is o th curv a+ b () y.9958 Qustio Solvig () ad (), w hav a, b. Qustio d.77 d d d.77

More information

Empirical Study in Finite Correlation Coefficient in Two Phase Estimation

Empirical Study in Finite Correlation Coefficient in Two Phase Estimation M. Khoshvisa Griffith Uivrsity Griffith Busiss School Australia F. Kaymarm Massachustts Istitut of Tchology Dpartmt of Mchaical girig USA H. P. Sigh R. Sigh Vikram Uivrsity Dpartmt of Mathmatics ad Statistics

More information

MONTGOMERY COLLEGE Department of Mathematics Rockville Campus. 6x dx a. b. cos 2x dx ( ) 7. arctan x dx e. cos 2x dx. 2 cos3x dx

MONTGOMERY COLLEGE Department of Mathematics Rockville Campus. 6x dx a. b. cos 2x dx ( ) 7. arctan x dx e. cos 2x dx. 2 cos3x dx MONTGOMERY COLLEGE Dpartmt of Mathmatics Rockvill Campus MATH 8 - REVIEW PROBLEMS. Stat whthr ach of th followig ca b itgratd by partial fractios (PF), itgratio by parts (PI), u-substitutio (U), or o of

More information

Technical Support Document Bias of the Minimum Statistic

Technical Support Document Bias of the Minimum Statistic Tchical Support Documt Bias o th Miimum Stattic Itroductio Th papr pla how to driv th bias o th miimum stattic i a radom sampl o siz rom dtributios with a shit paramtr (also kow as thrshold paramtr. Ths

More information

Worksheet: Taylor Series, Lagrange Error Bound ilearnmath.net

Worksheet: Taylor Series, Lagrange Error Bound ilearnmath.net Taylor s Thorm & Lagrag Error Bouds Actual Error This is th ral amout o rror, ot th rror boud (worst cas scario). It is th dirc btw th actual () ad th polyomial. Stps:. Plug -valu ito () to gt a valu.

More information

ELECTRONIC APPENDIX TO: ELASTIC-PLASTIC CONTACT OF A ROUGH SURFACE WITH WEIERSTRASS PROFILE. Yan-Fei Gao and A. F. Bower

ELECTRONIC APPENDIX TO: ELASTIC-PLASTIC CONTACT OF A ROUGH SURFACE WITH WEIERSTRASS PROFILE. Yan-Fei Gao and A. F. Bower ELECTRONIC APPENDIX TO: ELASTIC-PLASTIC CONTACT OF A ROUGH SURFACE WITH WEIERSTRASS PROFILE Ya-Fi Gao ad A. F. Bowr Divisio of Egirig, Brow Uivrsity, Providc, RI 9, USA Appdix A: Approximat xprssios for

More information

6. Comparison of NLMS-OCF with Existing Algorithms

6. Comparison of NLMS-OCF with Existing Algorithms 6. Compariso of NLMS-OCF with Eistig Algorithms I Chaptrs 5 w drivd th NLMS-OCF algorithm, aalyzd th covrgc ad trackig bhavior of NLMS-OCF, ad dvlopd a fast vrsio of th NLMS-OCF algorithm. W also mtiod

More information

2.29 Numerical Fluid Mechanics Spring 2015 Lecture 12

2.29 Numerical Fluid Mechanics Spring 2015 Lecture 12 REVIEW Lctur 11: Numrical Fluid Mchaics Sprig 2015 Lctur 12 Fiit Diffrcs basd Polyomial approximatios Obtai polyomial (i gral u-qually spacd), th diffrtiat as dd Nwto s itrpolatig polyomial formulas Triagular

More information

International Journal of Advanced and Applied Sciences

International Journal of Advanced and Applied Sciences Itratioal Joural of Advacd ad Applid Scics x(x) xxxx Pags: xx xx Cotts lists availabl at Scic Gat Itratioal Joural of Advacd ad Applid Scics Joural hompag: http://wwwscic gatcom/ijaashtml Symmtric Fuctios

More information

Journal of Modern Applied Statistical Methods

Journal of Modern Applied Statistical Methods Joural of Modr Applid Statistical Mthods Volum Issu Articl 6 --03 O Som Proprtis of a Htrogous Trasfr Fuctio Ivolvig Symmtric Saturatd Liar (SATLINS) with Hyprbolic Tagt (TANH) Trasfr Fuctios Christophr

More information

Comparison of Simple Indicator Kriging, DMPE, Full MV Approach for Categorical Random Variable Simulation

Comparison of Simple Indicator Kriging, DMPE, Full MV Approach for Categorical Random Variable Simulation Papr 17, CCG Aual Rport 11, 29 ( 29) Compariso of Simpl Idicator rigig, DMPE, Full MV Approach for Catgorical Radom Variabl Simulatio Yupg Li ad Clayto V. Dutsch Ifrc of coditioal probabilitis at usampld

More information

Bayesian Test for Lifetime Performance Index of Exponential Distribution under Symmetric Entropy Loss Function

Bayesian Test for Lifetime Performance Index of Exponential Distribution under Symmetric Entropy Loss Function Mathmatics ttrs 08; 4(): 0-4 http://www.scicpublishiggroup.com/j/ml doi: 0.648/j.ml.08040.5 ISSN: 575-503X (Prit); ISSN: 575-5056 (Oli) aysia Tst for iftim Prformac Idx of Expotial Distributio udr Symmtric

More information

UNIVERSALITY LIMITS INVOLVING ORTHOGONAL POLYNOMIALS ON AN ARC OF THE UNIT CIRCLE

UNIVERSALITY LIMITS INVOLVING ORTHOGONAL POLYNOMIALS ON AN ARC OF THE UNIT CIRCLE UNIVERSALITY LIMITS INVOLVING ORTHOGONAL POLYNOMIALS ON AN ARC OF THE UNIT CIRCLE DORON S. LUBINSKY AND VY NGUYEN A. W stablish uivrsality limits for masurs o a subarc of th uit circl. Assum that µ is

More information

Probability & Statistics,

Probability & Statistics, Probability & Statistics, BITS Pilai K K Birla Goa Campus Dr. Jajati Kshari Sahoo Dpartmt of Mathmatics BITS Pilai, K K Birla Goa Campus Poisso Distributio Poisso Distributio: A radom variabl X is said

More information

Figure 2-18 Thevenin Equivalent Circuit of a Noisy Resistor

Figure 2-18 Thevenin Equivalent Circuit of a Noisy Resistor .8 NOISE.8. Th Nyquist Nois Thorm W ow wat to tur our atttio to ois. W will start with th basic dfiitio of ois as usd i radar thory ad th discuss ois figur. Th typ of ois of itrst i radar thory is trmd

More information

Bipolar Junction Transistors

Bipolar Junction Transistors ipolar Juctio Trasistors ipolar juctio trasistors (JT) ar activ 3-trmial dvics with aras of applicatios: amplifirs, switch tc. high-powr circuits high-spd logic circuits for high-spd computrs. JT structur:

More information

A GENERALIZED RAMANUJAN-NAGELL EQUATION RELATED TO CERTAIN STRONGLY REGULAR GRAPHS

A GENERALIZED RAMANUJAN-NAGELL EQUATION RELATED TO CERTAIN STRONGLY REGULAR GRAPHS #A35 INTEGERS 4 (204) A GENERALIZED RAMANUJAN-NAGELL EQUATION RELATED TO CERTAIN STRONGLY REGULAR GRAPHS B d Wgr Faculty of Mathmatics ad Computr Scic, Eidhov Uivrsity of Tchology, Eidhov, Th Nthrlads

More information

On Deterministic Finite Automata and Syntactic Monoid Size, Continued

On Deterministic Finite Automata and Syntactic Monoid Size, Continued O Dtrmiistic Fiit Automata ad Sytactic Mooid Siz, Cotiud Markus Holzr ad Barbara Köig Istitut für Iformatik, Tchisch Uivrsität Müch, Boltzmastraß 3, D-85748 Garchig bi Müch, Grmay mail: {holzr,koigb}@iformatik.tu-much.d

More information

Linear Algebra Existence of the determinant. Expansion according to a row.

Linear Algebra Existence of the determinant. Expansion according to a row. Lir Algbr 2270 1 Existc of th dtrmit. Expsio ccordig to row. W dfi th dtrmit for 1 1 mtrics s dt([]) = (1) It is sy chck tht it stisfis D1)-D3). For y othr w dfi th dtrmit s follows. Assumig th dtrmit

More information

BOUNDS FOR THE COMPONENTWISE DISTANCE TO THE NEAREST SINGULAR MATRIX

BOUNDS FOR THE COMPONENTWISE DISTANCE TO THE NEAREST SINGULAR MATRIX SIAM J. Matrix Aal. Appl. (SIMAX), 8():83 03, 997 BOUNDS FOR THE COMPONENTWISE DISTANCE TO THE NEAREST SINGULAR MATRIX S. M. RUMP Abstract. Th ormwis distac of a matrix A to th arst sigular matrix is wll

More information

Discrete Fourier Transform. Discrete Fourier Transform. Discrete Fourier Transform. Discrete Fourier Transform. Discrete Fourier Transform

Discrete Fourier Transform. Discrete Fourier Transform. Discrete Fourier Transform. Discrete Fourier Transform. Discrete Fourier Transform Discrt Fourir Trasform Dfiitio - T simplst rlatio btw a lt- squc x dfid for ω ad its DTFT X ( ) is ω obtaid by uiformly sampli X ( ) o t ω-axis btw ω < at ω From t dfiitio of t DTFT w tus av X X( ω ) ω

More information

LECTURE 13 Filling the bands. Occupancy of Available Energy Levels

LECTURE 13 Filling the bands. Occupancy of Available Energy Levels LUR 3 illig th bads Occupacy o Availabl rgy Lvls W hav dtrmid ad a dsity o stats. W also d a way o dtrmiig i a stat is illd or ot at a giv tmpratur. h distributio o th rgis o a larg umbr o particls ad

More information

CDS 101: Lecture 5.1 Reachability and State Space Feedback

CDS 101: Lecture 5.1 Reachability and State Space Feedback CDS, Lctur 5. CDS : Lctur 5. Rachability ad Stat Spac Fdback Richard M. Murray ad Hido Mabuchi 5 Octobr 4 Goals: Di rachability o a cotrol systm Giv tsts or rachability o liar systms ad apply to ampls

More information

ONLINE SUPPLEMENT Optimal Markdown Pricing and Inventory Allocation for Retail Chains with Inventory Dependent Demand

ONLINE SUPPLEMENT Optimal Markdown Pricing and Inventory Allocation for Retail Chains with Inventory Dependent Demand Submittd to Maufacturig & Srvic Opratios Maagmt mauscript MSOM 5-4R2 ONLINE SUPPLEMENT Optimal Markdow Pricig ad Ivtory Allocatio for Rtail Chais with Ivtory Dpdt Dmad Stph A Smith Dpartmt of Opratios

More information

Washington State University

Washington State University he 3 Ktics ad Ractor Dsig Sprg, 00 Washgto Stat Uivrsity Dpartmt of hmical Egrg Richard L. Zollars Exam # You will hav o hour (60 muts) to complt this xam which cosists of four (4) problms. You may us

More information

On a problem of J. de Graaf connected with algebras of unbounded operators de Bruijn, N.G.

On a problem of J. de Graaf connected with algebras of unbounded operators de Bruijn, N.G. O a problm of J. d Graaf coctd with algbras of uboudd oprators d Bruij, N.G. Publishd: 01/01/1984 Documt Vrsio Publishr s PDF, also kow as Vrsio of Rcord (icluds fial pag, issu ad volum umbrs) Plas chck

More information

Minimax Rényi Redundancy

Minimax Rényi Redundancy Miimax Réyi Rdudacy Smih Yagli Pricto Uivrsity Yücl Altuğ Natra, Ic Srgio Vrdú Pricto Uivrsity Abstract Th rdudacy for uivrsal losslss comprssio i Campbll s sttig is charactrizd as a miimax Réyi divrgc,

More information

New Sixteenth-Order Derivative-Free Methods for Solving Nonlinear Equations

New Sixteenth-Order Derivative-Free Methods for Solving Nonlinear Equations Amrica Joural o Computatioal ad Applid Mathmatics 0 (: -8 DOI: 0.59/j.ajcam.000.08 Nw Sixtth-Ordr Drivativ-Fr Mthods or Solvig Noliar Equatios R. Thukral Padé Rsarch Ctr 9 Daswood Hill Lds Wst Yorkshir

More information

COMPUTING FOLRIER AND LAPLACE TRANSFORMS. Sven-Ake Gustafson. be a real-valued func'cion, defined for nonnegative arguments.

COMPUTING FOLRIER AND LAPLACE TRANSFORMS. Sven-Ake Gustafson. be a real-valued func'cion, defined for nonnegative arguments. 77 COMPUTNG FOLRER AND LAPLACE TRANSFORMS BY MEANS OF PmER SERES EVALU\TON Sv-Ak Gustafso 1. NOTATONS AND ASSUMPTONS Lt f b a ral-valud fuc'cio, dfid for ogativ argumts. W shall discuss som aspcts of th

More information

KISS: A Bit Too Simple. Greg Rose

KISS: A Bit Too Simple. Greg Rose KI: A Bit Too impl Grg Ros ggr@qualcomm.com Outli KI radom umbr grator ubgrators Efficit attack N KI ad attack oclusio PAGE 2 O approach to PRNG scurity "A radom umbr grator is lik sx: Wh it's good, its

More information

Learning objectives. Three models of aggregate supply. 1. The sticky-wage model 2. The imperfect-information model 3. The sticky-price model

Learning objectives. Three models of aggregate supply. 1. The sticky-wage model 2. The imperfect-information model 3. The sticky-price model Larig objctivs thr modls of aggrgat supply i which output dpds positivly o th pric lvl i th short ru th short-ru tradoff btw iflatio ad umploymt kow as th Phillips curv Aggrgat Supply slid 1 Thr modls

More information

MILLIKAN OIL DROP EXPERIMENT

MILLIKAN OIL DROP EXPERIMENT 11 Oct 18 Millika.1 MILLIKAN OIL DROP EXPERIMENT This xprimt is dsigd to show th quatizatio of lctric charg ad allow dtrmiatio of th lmtary charg,. As i Millika s origial xprimt, oil drops ar sprayd ito

More information

Chapter 4 - The Fourier Series

Chapter 4 - The Fourier Series M. J. Robrts - 8/8/4 Chaptr 4 - Th Fourir Sris Slctd Solutios (I this solutio maual, th symbol,, is usd for priodic covolutio bcaus th prfrrd symbol which appars i th txt is ot i th fot slctio of th word

More information

Motivation. We talk today for a more flexible approach for modeling the conditional probabilities.

Motivation. We talk today for a more flexible approach for modeling the conditional probabilities. Baysia Ntworks Motivatio Th coditioal idpdc assuptio ad by aïv Bays classifirs ay s too rigid spcially for classificatio probls i which th attributs ar sowhat corrlatd. W talk today for a or flibl approach

More information

STIRLING'S 1 FORMULA AND ITS APPLICATION

STIRLING'S 1 FORMULA AND ITS APPLICATION MAT-KOL (Baja Luka) XXIV ()(08) 57-64 http://wwwimviblorg/dmbl/dmblhtm DOI: 075/МК80057A ISSN 0354-6969 (o) ISSN 986-588 (o) STIRLING'S FORMULA AND ITS APPLICATION Šfkt Arslaagić Sarajvo B&H Abstract:

More information

Folding of Hyperbolic Manifolds

Folding of Hyperbolic Manifolds It. J. Cotmp. Math. Scics, Vol. 7, 0, o. 6, 79-799 Foldig of Hyprbolic Maifolds H. I. Attiya Basic Scic Dpartmt, Collg of Idustrial Educatio BANE - SUEF Uivrsity, Egypt hala_attiya005@yahoo.com Abstract

More information

Minimax Rényi Redundancy

Minimax Rényi Redundancy 07 IEEE Itratioal Symposium o Iformatio Thory ISIT Miimax Réyi Rdudacy Smih Yagli Pricto Uivrsity Yücl Altuğ Natra, Ic Srgio Vrdú Pricto Uivrsity Abstract Th rdudacy for uivrsal losslss comprssio i Campbll

More information

They must have different numbers of electrons orbiting their nuclei. They must have the same number of neutrons in their nuclei.

They must have different numbers of electrons orbiting their nuclei. They must have the same number of neutrons in their nuclei. 37 1 How may utros ar i a uclus of th uclid l? 20 37 54 2 crtai lmt has svral isotops. Which statmt about ths isotops is corrct? Thy must hav diffrt umbrs of lctros orbitig thir ucli. Thy must hav th sam

More information

Time Dependent Solutions: Propagators and Representations

Time Dependent Solutions: Propagators and Representations Tim Dpdt Solutios: Propagators ad Rprstatios Michal Fowlr, UVa 1/3/6 Itroductio W v spt most of th cours so far coctratig o th igstats of th amiltoia, stats whos tim dpdc is mrly a chagig phas W did mtio

More information

CDS 101: Lecture 5.1 Reachability and State Space Feedback

CDS 101: Lecture 5.1 Reachability and State Space Feedback CDS, Lctur 5. CDS : Lctur 5. Rachability ad Stat Spac Fdback Richard M. Murray 7 Octobr 3 Goals: Di rachability o a cotrol systm Giv tsts or rachability o liar systms ad apply to ampls Dscrib th dsig o

More information

Part B: Transform Methods. Professor E. Ambikairajah UNSW, Australia

Part B: Transform Methods. Professor E. Ambikairajah UNSW, Australia Part B: Trasform Mthods Chaptr 3: Discrt-Tim Fourir Trasform (DTFT) 3. Discrt Tim Fourir Trasform (DTFT) 3. Proprtis of DTFT 3.3 Discrt Fourir Trasform (DFT) 3.4 Paddig with Zros ad frqucy Rsolutio 3.5

More information

Problem Value Score Earned No/Wrong Rec -3 Total

Problem Value Score Earned No/Wrong Rec -3 Total GEORGIA INSTITUTE OF TECHNOLOGY SCHOOL of ELECTRICAL & COMPUTER ENGINEERING ECE6 Fall Quiz # Writt Eam Novmr, NAME: Solutio Kys GT Usram: LAST FIRST.g., gtiit Rcitatio Sctio: Circl t dat & tim w your Rcitatio

More information

macro Road map to this lecture Objectives Aggregate Supply and the Phillips Curve Three models of aggregate supply W = ω P The sticky-wage model

macro Road map to this lecture Objectives Aggregate Supply and the Phillips Curve Three models of aggregate supply W = ω P The sticky-wage model Road map to this lctur macro Aggrgat Supply ad th Phillips Curv W rlax th assumptio that th aggrgat supply curv is vrtical A vrsio of th aggrgat supply i trms of iflatio (rathr tha th pric lvl is calld

More information

THREE-WAY ROC ANALYSIS USING SAS SOFTWARE

THREE-WAY ROC ANALYSIS USING SAS SOFTWARE ACTA UNIVERSITATIS AGRICULTURAE ET SILVICULTURAE MENDELIANAE BRUNENSIS Volum LXI 54 Numbr 7, 03 http://d.doi.org/0.8/actau0360769 THREE-WAY ROC ANALYSIS USING SAS SOFTWARE Juraj Kapasý, Marti Řzáč Rcivd:

More information

Thomas J. Osler. 1. INTRODUCTION. This paper gives another proof for the remarkable simple

Thomas J. Osler. 1. INTRODUCTION. This paper gives another proof for the remarkable simple 5/24/5 A PROOF OF THE CONTINUED FRACTION EXPANSION OF / Thomas J Oslr INTRODUCTION This ar givs aothr roof for th rmarkabl siml cotiud fractio = 3 5 / Hr is ay ositiv umbr W us th otatio x= [ a; a, a2,

More information

EFFECT OF P-NORMS ON THE ACCURACY ORDER OF NUMERICAL SOLUTION ERRORS IN CFD

EFFECT OF P-NORMS ON THE ACCURACY ORDER OF NUMERICAL SOLUTION ERRORS IN CFD rocdigs of NIT 00 opyright 00 by ABM 3 th Brazilia ogrss of Thrmal Scics ad girig Dcmbr 05-0, 00, brladia, MG, Brazil T O -NORMS ON TH ARAY ORDR O NMRIAL SOLTION RRORS IN D arlos Hriqu Marchi, marchi@ufpr.br

More information

Frequency Measurement in Noise

Frequency Measurement in Noise Frqucy Masurmt i ois Porat Sctio 6.5 /4 Frqucy Mas. i ois Problm Wat to o look at th ct o ois o usig th DFT to masur th rqucy o a siusoid. Cosidr sigl complx siusoid cas: j y +, ssum Complx Whit ois Gaussia,

More information

Gaps in samples of geometric random variables

Gaps in samples of geometric random variables Discrt Mathmatics 37 7 871 89 Not Gaps i sampls of gomtric radom variabls William M.Y. Goh a, Pawl Hitczko b,1 a Dpartmt of Mathmatics, Drxl Uivrsity, Philadlphia, PA 1914, USA b Dpartmts of Mathmatics

More information

Introduction to Quantum Information Processing. Overview. A classical randomised algorithm. q 3,3 00 0,0. p 0,0. Lecture 10.

Introduction to Quantum Information Processing. Overview. A classical randomised algorithm. q 3,3 00 0,0. p 0,0. Lecture 10. Itroductio to Quatum Iformatio Procssig Lctur Michl Mosca Ovrviw! Classical Radomizd vs. Quatum Computig! Dutsch-Jozsa ad Brsti- Vazirai algorithms! Th quatum Fourir trasform ad phas stimatio A classical

More information

Review Exercises. 1. Evaluate using the definition of the definite integral as a Riemann Sum. Does the answer represent an area? 2

Review Exercises. 1. Evaluate using the definition of the definite integral as a Riemann Sum. Does the answer represent an area? 2 MATHEMATIS --RE Itgral alculus Marti Huard Witr 9 Rviw Erciss. Evaluat usig th dfiitio of th dfiit itgral as a Rima Sum. Dos th aswr rprst a ara? a ( d b ( d c ( ( d d ( d. Fid f ( usig th Fudamtal Thorm

More information

Hadamard Exponential Hankel Matrix, Its Eigenvalues and Some Norms

Hadamard Exponential Hankel Matrix, Its Eigenvalues and Some Norms Math Sci Ltt Vol No 8-87 (0) adamard Exotial al Matrix, Its Eigvalus ad Som Norms İ ad M bula Mathmatical Scics Lttrs Itratioal Joural @ 0 NSP Natural Scics Publishig Cor Dartmt of Mathmatics, aculty of

More information

Numerov-Cooley Method : 1-D Schr. Eq. Last time: Rydberg, Klein, Rees Method and Long-Range Model G(v), B(v) rotation-vibration constants.

Numerov-Cooley Method : 1-D Schr. Eq. Last time: Rydberg, Klein, Rees Method and Long-Range Model G(v), B(v) rotation-vibration constants. Numrov-Cooly Mthod : 1-D Schr. Eq. Last tim: Rydbrg, Kli, Rs Mthod ad Log-Rag Modl G(v), B(v) rotatio-vibratio costats 9-1 V J (x) pottial rgy curv x = R R Ev,J, v,j, all cocivabl xprimts wp( x, t) = ai

More information

Blackbody Radiation. All bodies at a temperature T emit and absorb thermal electromagnetic radiation. How is blackbody radiation absorbed and emitted?

Blackbody Radiation. All bodies at a temperature T emit and absorb thermal electromagnetic radiation. How is blackbody radiation absorbed and emitted? All bodis at a tmpratur T mit ad absorb thrmal lctromagtic radiatio Blackbody radiatio I thrmal quilibrium, th powr mittd quals th powr absorbd How is blackbody radiatio absorbd ad mittd? 1 2 A blackbody

More information

Outline. Ionizing Radiation. Introduction. Ionizing radiation

Outline. Ionizing Radiation. Introduction. Ionizing radiation Outli Ioizig Radiatio Chaptr F.A. Attix, Itroductio to Radiological Physics ad Radiatio Dosimtry Radiological physics ad radiatio dosimtry Typs ad sourcs of ioizig radiatio Dscriptio of ioizig radiatio

More information

A Propagating Wave Packet Group Velocity Dispersion

A Propagating Wave Packet Group Velocity Dispersion Lctur 8 Phys 375 A Propagating Wav Packt Group Vlocity Disprsion Ovrviw and Motivation: In th last lctur w lookd at a localizd solution t) to th 1D fr-particl Schrödingr quation (SE) that corrsponds to

More information

Estimation of Consumer Demand Functions When the Observed Prices Are the Same for All Sample Units

Estimation of Consumer Demand Functions When the Observed Prices Are the Same for All Sample Units Dpartmt of Agricultural ad Rsourc Ecoomics Uivrsity of Califoria, Davis Estimatio of Cosumr Dmad Fuctios Wh th Obsrvd Prics Ar th Sam for All Sampl Uits by Quirio Paris Workig Papr No. 03-004 Sptmbr 2003

More information

UNIT 2: MATHEMATICAL ENVIRONMENT

UNIT 2: MATHEMATICAL ENVIRONMENT UNIT : MATHEMATICAL ENVIRONMENT. Itroductio This uit itroducs som basic mathmatical cocpts ad rlats thm to th otatio usd i th cours. Wh ou hav workd through this uit ou should: apprciat that a mathmatical

More information

Supplemental Material for "Automated Estimation of Vector Error Correction Models"

Supplemental Material for Automated Estimation of Vector Error Correction Models Supplmtal Matrial for "Automatd Estimatio of Vctor Error Corrctio Modls" ipg Liao Ptr C. B. Pillips y Tis Vrsio: Sptmbr 23 Abstract Tis supplmt icluds two sctios. Sctio cotais t proofs of som auxiliary

More information

A Review of Complex Arithmetic

A Review of Complex Arithmetic /0/005 Rviw of omplx Arithmti.do /9 A Rviw of omplx Arithmti A omplx valu has both a ral ad imagiary ompot: { } ad Im{ } a R b so that w a xprss this omplx valu as: whr. a + b Just as a ral valu a b xprssd

More information

22/ Breakdown of the Born-Oppenheimer approximation. Selection rules for rotational-vibrational transitions. P, R branches.

22/ Breakdown of the Born-Oppenheimer approximation. Selection rules for rotational-vibrational transitions. P, R branches. Subjct Chmistry Papr No and Titl Modul No and Titl Modul Tag 8/ Physical Spctroscopy / Brakdown of th Born-Oppnhimr approximation. Slction ruls for rotational-vibrational transitions. P, R branchs. CHE_P8_M

More information

The Interplay between l-max, l-min, p-max and p-min Stable Distributions

The Interplay between l-max, l-min, p-max and p-min Stable Distributions DOI: 0.545/mjis.05.4006 Th Itrplay btw lma lmi pma ad pmi Stabl Distributios S Ravi ad TS Mavitha Dpartmt of Studis i Statistics Uivrsity of Mysor Maasagagotri Mysuru 570006 Idia. Email:ravi@statistics.uimysor.ac.i

More information

Ideal crystal : Regulary ordered point masses connected via harmonic springs

Ideal crystal : Regulary ordered point masses connected via harmonic springs Statistical thrmodyamics of crystals Mooatomic crystal Idal crystal : Rgulary ordrd poit masss coctd via harmoic sprigs Itratomic itractios Rprstd by th lattic forc-costat quivalt atom positios miima o

More information

ω (argument or phase)

ω (argument or phase) Imagiary uit: i ( i Complx umbr: z x+ i y Cartsia coordiats: x (ral part y (imagiary part Complx cougat: z x i y Absolut valu: r z x + y Polar coordiats: r (absolut valu or modulus ω (argumt or phas x

More information

SOME IDENTITIES FOR THE GENERALIZED POLY-GENOCCHI POLYNOMIALS WITH THE PARAMETERS A, B AND C

SOME IDENTITIES FOR THE GENERALIZED POLY-GENOCCHI POLYNOMIALS WITH THE PARAMETERS A, B AND C Joural of Mathatical Aalysis ISSN: 2217-3412, URL: www.ilirias.co/ja Volu 8 Issu 1 2017, Pags 156-163 SOME IDENTITIES FOR THE GENERALIZED POLY-GENOCCHI POLYNOMIALS WITH THE PARAMETERS A, B AND C BURAK

More information

8(4 m0) ( θ ) ( ) Solutions for HW 8. Chapter 25. Conceptual Questions

8(4 m0) ( θ ) ( ) Solutions for HW 8. Chapter 25. Conceptual Questions Solutios for HW 8 Captr 5 Cocptual Qustios 5.. θ dcrass. As t crystal is coprssd, t spacig d btw t plas of atos dcrass. For t first ordr diffractio =. T Bragg coditio is = d so as d dcrass, ust icras for

More information

NET/JRF, GATE, IIT JAM, JEST, TIFR

NET/JRF, GATE, IIT JAM, JEST, TIFR Istitut for NET/JRF, GATE, IIT JAM, JEST, TIFR ad GRE i PHYSICAL SCIENCES Mathmatical Physics JEST-6 Q. Giv th coditio φ, th solutio of th quatio ψ φ φ is giv by k. kφ kφ lφ kφ lφ (a) ψ (b) ψ kφ (c) ψ

More information

Class #24 Monday, April 16, φ φ φ

Class #24 Monday, April 16, φ φ φ lass #4 Moday, April 6, 08 haptr 3: Partial Diffrtial Equatios (PDE s First of all, this sctio is vry, vry difficult. But it s also supr cool. PDE s thr is mor tha o idpdt variabl. Exampl: φ φ φ φ = 0

More information

Partition Functions and Ideal Gases

Partition Functions and Ideal Gases Partitio Fuctios ad Idal Gass PFIG- You v lard about partitio fuctios ad som uss ow w ll xplor tm i mor dpt usig idal moatomic diatomic ad polyatomic gass! for w start rmmbr: Q( N ( N! N Wat ar N ad? W

More information