Online Geometric Optimization in the Bandit Setting Against an Adaptive Adversary

Size: px
Start display at page:

Download "Online Geometric Optimization in the Bandit Setting Against an Adaptive Adversary"

Transcription

1 lie Geomeric pimizaio i he Badi Seig Agais a Adapive Adversary H. Breda McMaha ad Avrim Blum Caregie Mello Uiversiy, Pisburgh, PA, 15213, Absrac. We give a algorihm for he badi versio of a very geeral olie opimizaio problem cosidered by Kalai ad Vempala [1], for he case of a adapive adversary. I his problem we are give a bouded se S of feasible pois. A each ime sep, he olie algorihm mus selec a poi x S while simulaeously a adversary selecs a cos vecor c. The algorihm he icurs cos c x. Kalai ad Vempala show ha eve if S is expoeially large (or ifiie), so log as we have a efficie algorihm for he offlie problem (give c, fid x S o miimize c x) ad so log as he cos vecors are bouded, oe ca efficiely solve he olie problem of performig early as well as he bes fixed x S i hidsigh. The Kalai-Vempala algorihm assumes ha he cos vecors c are give o he algorihm afer each ime sep. I he badi versio of he problem, he algorihm oly observes is cos, c x. Awerbuch ad Kleiberg [2] give a algorihm for he badi versio for he case of a oblivious adversary, ad a algorihm ha works agais a adapive adversary for he special case of he shores pah problem. They leave ope he problem of hadlig a adapive adversary i he geeral case. I his paper, we solve his ope problem, givig a simple olie algorihm for he badi problem i he geeral case i he presece of a adapive adversary. Igorig a (polyomial) depedece o, we achieve a regre boud of T 3 4 l T. 1 Iroducio Kalai ad Vempala [1] give a elega, efficie algorihm for a broad class of olie opimizaio problems. I heir seig, we have a arbirary (bouded) se S of feasible pois. A each ime sep, a olie algorihm mus selec a poi x S ad simulaeously a adversary selecs a cos vecor c (hroughou he paper we use superscrips o idex ieraios). The algorihm he observes c ad icurs cos c x. Kalai ad Vempala show ha so log as we have a efficie algorihm for he offlie problem (give c fid x S o miimize c x) ad so log as he cos vecors are bouded, we ca efficiely solve he olie problem of performig early as well as he bes fixed x S i hidsigh. This geeralizes he classic exper advice problem, because we do o require he se S o be represeed explicily: we jus eed a efficie oracle for selecig he bes x S i hidsigh. Furher, i decouples he umber of expers from he uderlyig dimesioaliy of he decisio se, uder he assumpio he cos of a decisio is a liear fucio of feaures of he decisio. The sadard expers seig ca be recovered by leig S e 1 e, he colums of he ideiy marix.

2 A problem ha fis aurally io his framework is a olie shores pah problem where we repeaedly ravel bewee wo pois a ad b i some graph whose edge coss chage each day (say, due o raffic). I his case, we ca view he se of pahs as a se S of pois i a space of dimesio equal o he umber of edges i he graph, ad c is simply he vecor of edge coss o day. Eve hough he umber of pahs i a graph ca be expoeial i he umber of edges (i.e., he se S is of expoeial size), sice we ca solve he shores pah problem for ay give se of edge leghs, we ca apply he Kalai-Vempala algorihm. (Noe ha a differe algorihm for he special case of he olie shores pah problem is give by Takimoo ad Warmuh [3].) A aural geeralizaio of he above problem, cosidered by Awerbuch ad Kleiberg [2], is o imagie ha raher ha beig give he eire cos vecor c, he algorihm is simply old he cos icurred c x. For example, i he case of shores pahs, raher ha beig old he leghs of all edges a ime, his would correspod o jus beig old he oal ime ake o reach he desiaio. Thus, his is he badi versio of he Kalai-Vempala seig. Awerbuch ad Kleiberg prese wo resuls: a algorihm for he geeral problem i he presece of a oblivious adversary, ad a algorihm for he special case of he shores pah problem ha works i he presece of a adapive adversary. The differece bewee he wo adversaries is ha a oblivious adversary mus commi o he eire sequece of cos vecors i advace, whereas a adapive adversary may deermie he ex cos vecor based o he olie algorihm s play (ad hece, he iformaio he algorihm received) i he previous ime seps. Thus, a adapive adversary is i essece playig a repeaed game. They leave ope he quesio of achievig good regre guaraees for a adapive adversary i he geeral seig. I his paper we solve he ope quesio of [2], givig a algorihm for he geeral badi seig i he presece of a adapive adversary. Moreover, our mehod is sigificaly simpler ha he special-purpose algorihm of Awerbuch ad Kleiberg for shores pahs. ur bouds are somewha worse: we achieve regre bouds of T 3 4 lt! compared o he " T 2 3! bouds of [2]. We believe improveme i his direcio may be possible, ad prese some discussio of his issue a he ed of he paper. The basic idea of our approach is as follows. We begi by oicig ha he oly hisory iformaio used by he Kalai-Vempala algorihm i deermiig is acio a ime is he sum c 1:# 1 # τ$ 1 1 cτ of all cos vecors received so far (we use his abbreviaed oaio for sums over ieraio idexes hroughou he paper). Furhermore, he way his is used i he algorihm is by addig radom oise µ o his vecor, ad he callig he offlie oracle o fid he x S ha miimizes c 1:# 1 % µ! x. So, if we ca desig a badi algorihm ha produces a esimae ĉ 1:# 1 of c 1:# 1, ad show ha wih high probabiliy eve a adapive adversary will o cause ĉ 1:# 1 o differ oo subsaially from c 1:# 1, we ca he argue ha he disribuio ĉ 1:# 1 % µ is close eough o c 1:# 1 % µ for he Kalai-Vempala aalysis o apply. I fac, o make our aalysis a bi more geeral, so ha we could poeially use oher algorihms as subrouies, we will argue a lile differely. Le PT c!& mi x' S c x!. We will show ha wih high probabiliy, PT ĉ 1:T! is close o PT c 1:T! ad ĉ 1:T saisfies codiios eeded for he subrouie o achieve low regre o ĉ 1:T. This meas ha our subrouie, which believes i has see ĉ 1:T, will achieve performace o ĉ 1:T close o PT c 1:T!. We he fiish off by arguig ha our performace o c 1:T is close o is performace o ĉ 1:T.

3 The behavior of he badi algorihm will i fac be fairly simple. We begi by choosig a basis B of (a mos) pois i S o use for samplig (we address he issue of how B is chose whe we describe our algorihm i deail). The, a each ime sep, wih probabiliy γ we explore by playig a radom basis eleme, ad oherwise (wih probabiliy 1 ( γ) we exploi by playig accordig o he Kalai-Vempala algorihm. For each basis eleme b j, we use our cos icurred while explorig wih ha basis eleme, scaled by ) γ, as a esimae of c 1:# 1 b j. Usig marigale ail iequaliies, we argue ha eve a adapive adversary cao make our esimae differ oo wildly from he rue value of c 1:# 1 b j, ad use his o show ha afer marix iversio, our esimae ĉ 1:# 1 is close o is correc value wih high probabiliy. 2 Problem Formalizaio We ca ow fully formalize he problem. Firs, however, we esablish a few oaioal coveios. As meioed previously, we use superscrips o idex ieraios (or rouds) of our algorihm, ad use he abbreviaed summaio oaio c 1: whe summig variables over ieraios. Vecors quaiies are idicaed i bold, ad subscrips idex io vecors or ses. Has (such as ĉ ) deoe esimaes of he correspodig acual quaiies. The variables ad cosas used i he paper are summarized i Table (1). As meioed above, we cosider he seig of [1] i which we have a arbirary (bouded) se S * of feasible pois. A each ime sep, he olie algorihm mus selec a poi x S ad simulaeously a adversary selecs a cos vecor c. The algorihm he icurs cos c x. Ulike [1], however, raher ha beig old c, he algorihm simply lears is cos c x. For simpliciy, we assume a fixed adapive adversary + ad ime horizo T for he duraio of his paper. Sice our choice of algorihm parameers depeds o T, we assume 1 T is kow o he algorihm. We refer o he sequece of decisios made by he algorihm so far as a decisio hisory, which ca be wrie h, x 1 x -. Le H. be he se of all possible decisio hisories of legh 0 hrough T ( 1. Wihou loss of geeraliy (e.g., see [5]), we assume our adapive adversary is deermiisic, as specified by a fucio + : H.0/1, a mappig from decisio hisories o cos vecors. Thus, +2 h # 1!3 c is he cos vecor for imesep. We ca view our olie decisio problem as a game, where o each ieraio he adversary + selecs a ew cos vecor c based o h # 1, ad he olie algorihm selecs a decisio x S based o is pas plays ad observaios, ad possibly addiioal hidde sae or radomess. The, pays c x ad observes his cos. For our aalysis, we assume a L 1 boud o S, amely 4 x4 1 5 D) 2 for all x S, so 4 x ( y4 1 5 D for all x y S. We also assume ha 6 c x6 5 M for all x S ad all c played by +. We also assume S is full rak, if i is o we simply projec o a lower-dimesioal represeaio. Some of hese assumpios ca be lifed or modified, bu his se of assumpios simplifies he aalysis. For a fixed decisio hisory h T ad cos hisory k T 7 c 1 8 c T!, we defie loss h T k T!9 $ T 1 c x!. For a radomized algorihm ad adversary +, we defie he radom 1 e ca remove his requireme by guessig T, ad doublig he guess each ime we play loger ha expeced (see, for example, Theorem 6.4 from [4]).

4 C variable loss: +;! o be loss h T k T!, where h T is draw from he disribuio over hisories defied by ad +, ad k T <=+2 h 0! 8 +2 h T# 1!!. Whe i is clear from coex, we will omi he depedece o +, wriig oly loss>?!. ur goal is o defie a olie algorihm wih low regre. Tha is, we wa a guaraee ha he oal loss icurred will, i expecaio, o be much larger ha he opimal sraegy i hidsigh agais he cos sequece we acually faced. To formalize his, firs defie a : / S ha solves he offlie opimizaio c! argmi x' S c x!. We he defie PT k T! c c 1:T!. Similarly, PT> +;! is he radom variable PT k T! whe k T is geeraed by playig + agais. We agai drop he depedece o + ad whe i is clear from coex. Formally, we defie expeced regre as E, loss> +2!B( PT> +;! - E, loss: +;! - ( E mi x' S T c x! (1) $ 1 Noe ha he E, PT: +;! - erm correspods o applyig he mi operaor separaely o each possible cos hisory o fid he bes fixed decisio wih respec o ha paricular cos hisory, ad he akig he expecaio wih respec o hese hisories. I [5], a aleraive weaker defiiio of regre is give. We discuss relaioships bewee he defiiios i Appedix B. 3 Algorihm Choose parameers γ ad ε, where ε is a parameer of GEX C 1 Fix a basis B C7D b 1EFFF>E b G S while playig do Le χ C 1 wih probabiliy γ ad χ C 0 oherwise if χ C 0 he Selec x from he disribuio GEX ĉ 1 EFFFHE ĉ I 1 Icur cos z C c x ĉ C 0 else Draw j uiformly a radom from D 1EFFF>E G x C b j Icur cos J ad observe z C c x Defie ˆ J by ˆ i C 0 for i K J j ad ˆ j C L γ z ĉ C B I 1ˆ J ed if ĉ 1: C ĉ 1:I 1 M ĉ C M 1 ed while Algorihm 1: BGA

5 We iroduce a algorihm we call BGA, sadig for Badi-syle Geomeric decisio algorihm agais a Adapive adversary. The algorihm aleraes bewee playig decisios from a fixed basis o ge ubiased esimaes of coss, ad playig (hopefully) good decisios based o hose esimaes. I order o deermie he good decisios o play, i uses some olie geomeric opimizaio algorihm for he full observaio problem. We deoe his algorihm by GEX (Geomeric Expers algorihm). The implemeaio of GEX we aalyze is based o he FPL algorihm of Kalai ad Vempala [1]; we deail his implemeaio ad aalysis i Appedix A. However, oher algorihms could be used, for example he algorihm of Zikevich [6] whe S is covex. We view GEX as a fucio from he sequece of previous cos vecors ĉ 1 8 ĉ # 1! o disribuios over decisios. Pseudocode for our algorihm is give i Algorihm (1). each imesep, we make decisio x. Wih probabiliy 1 ( γ!, BGA plays a recommedaio x x S from GEX. Wih probabiliy γ, we igore x ad play a basis decisio, x b i uiformly a radom from a samplig basis B N b 1 b. The idicaor variable χ is 1 o exploraio ieraios ad 0 oherwise. ur samplig basis B is a marix wih colums b i S, so we ca wrie x Bw for ay x ad weighs w. For a give cos vecor c, le P B c (he superscrip idicaes raspose). This is he vecor of decisio coss for he basis decisios, so i c b i. We defie ˆ, a esimae of, as follows: Le ˆ 0 o exploiaio ieraios. If o a exploraio ieraio we play b j, he ˆ is he vecor where ˆ i 0 for i Q j ad ˆ j γ c b j!. Noe ha c b j is he observed quaiy, he cos of basis decisio b j. each ieraio, we esimae c by ĉ R B! # 1ˆ. I is sraighforward o show ha ˆ is a ubiased esimae of basis decisio coss ad ha ĉ is a ubiased esimae of c o each imesep. The choice of he samplig basis plays a impora role i he aalysis of our algorihm. I paricular, we use a bariceric spaer, iroduced i [2]. A bariceric spaer B S b 1 b is a basis for S such ha b i S ad for all x S we ca wrie x Bw wih coefficies w i,t( I may o be easy o fid exac bariceric spaers i all cases, bu [2] proves hey always exis ad gives a algorihm for fidig 2-approximae bariceric spaers (where he weighs w i,u( 2 2 -!, which is sufficie for our purposes. 4 Aalysis 4.1 Prelimiaries A each ime sep, BGA eiher (wih probabiliy 1 ( γ) plays he recommedaio x from GEX, or else (wih probabiliy γ) plays a radom basis vecor from B. For purposes of aalysis, however, i will be coveie o imagie ha we reques a recommedaio x from GEX o every ieraio, ad also ha we radomly pick a basis o explore, b b 1 b, o each ieraio. We he decide o play eiher x or b based o he oucome χ of a coi of bias γ. Thus, he complee hisory of he algorihm is specified by he algorihm hisory G # 1 *, χ 1 x 1 b 1 χ 2 x 2 b 2 χ # 1 x # 1 b # 1-, which ecodes all previous radom choices. The sample space for all probabiliies ad expecaios is

6 ] ` S V D Table 1. Summary of oaio se of decisios, a compac subse of L 1 boud o diameer of S, W xe y SEYX x Z yx 1 [ D N dimesio of decisio space h decisio hisory, h C x 1 EFFF>E x H\ se of possible decisio hisories : H\_^ adversary, fucio from decisio hisories o cos vecors a olie opimizaio algorihm G I 1 hisory of BGA radomess for imeseps 1 hrough Z 1 c cos vecor o ime ĉ BGA s esimae of he cos vecor o ime M 0a boud o sigle-ieraio cos, X c x X [ M B S samplig basis B CbD b 1EFFF>E b G β marix max orm o B I 1 J dc Z ME Me J vecor, i C c b i for b i B Jˆ J BGA s esimae of T f ed of ime, idex of fial ieraio x S x S BGA s decisio o ime decisio recommeded by GEX o ime idicaor, χ C 1 if BGA explores o, 0 oherwise he probabiliy BGA explores o each imesep χ D 0E 1G γ dc 0E 1e z dc Z ME Me BGA s loss o ieraio, z C c x, ẑ dc Z RE Re loss of GEX, ẑ C ĉ x he se of all possible algorihm hisories of legh T. Thus, for a give adversary +, he various radom variables ad vecors we cosider, such as x c ĉ x, ad ohers, ca all be viewed as fucios o he se of possible algorihm hisories. Uless oherwise saed, our expecaios ad probabiliies are wih respec o he disribuio over hese hisories. A parial hisory G # 1 ca be viewed a subse of he sample space (a eve) cosisig of all complee hisories ha have G # 1 as a prefix. We frequely cosider codiioal disribuios ad correspodig expecaios wih respec o parial algorihm hisories. For isace, if we codiio o a hisory G # 1, he radom variables c 1 c, 1 1, ˆ ˆ# 1, ĉ 1 ĉ # 1, x 1 x # 1, ad χ 1 χ # 1 are fully deermied. We ow oulie he geeral srucure of our argume. Le ẑ ĉ x be he loss perceived by he GEX o ieraio. I keepig wih earlier defiiios, loss BGA!& z 1:T ad loss GEX!0 ẑ 1:T. We also le PT PT BGA +;!0 c c 1:T!, he performace of he bes pos-hoc decisio, ad similarly PT PT ĉ 1 ĉ T!P ĉ ĉ 1:!. The base of our aalysis is a boud o he loss of GEX wih respec o he cos vecors ĉ of he form E, loss GEX! - 5 E, PT -8% erms! (2)

7 Such a resul is give i Appedix A, ad follows from a adapaio of he aalysis from [1]. We he prove saemes havig he geeral form E, loss BGA! - 5 E, loss GEX! -8% erms! (3) ad E, PT - 5 E, PT -g% erms! (4) These saemes coec our real loss o he imagiary loss of GEX, ad similarly coec he loss of he bes decisio i GEX s imagied world wih he loss of he bes decisio i he real world. Combiig he resuls correspodig o Equaios (2), (3), ad (4) leads o a overall boud o he regre of BGA. 4.2 High Probabiliy Bouds o Esimaes We prove a boud o he accuracy of BGA s esimaes ˆ, ad use his o show a relaioship bewee PT ad PT of he form i Equaio 4. Defie radom variables e 0 0 ad e ( ˆ. We are really ieresed i he correspodig sums e 1:, where e 1: is he oal error i our esimae of c 1: b i. We ow boud 6 e 1: i 6. Theorem 1. For λ h 0, i Pr e 1: i i λ M γ 5 2e # λ2 2 Proof. I is sufficie o show he sequece e 0 e 1 e 1:2 e 1:3 e 1:T of radom variables is a bouded marigale sequece wih respec o he filer G 0 G 1 G T ; ha is, E, e 1: i 6 G # 1- e 1:# 1. The resul he follows from Azuma s Iequaliy (see, for example,[7]). Firs, observe ha e 1: i i j i ( ˆ % i e 1:# i 1 we kow G # 1, ad so i is also fixed. Thus, accouig for he γ a paricular basis decisio b i, we have E e 1: i 6 G # 1 γ i ( γ i % e 1:# i 1. Furher, he cos vecor c is deermied if probabiliy we explore % 1 ( γ, i ( 0 % e 1:# i 1 - e 1:# i 1 ad so we coclude ha he e 1: i forms a marigale sequece. Noice ha 6 e 1: i ( e 1:# i 1 6 i ( ˆ i 6. If we do sample, ˆ i 0 ad so 6 e 1: i ( e 1:# i M. If we do sample, we have ˆ i γ i, ad so 6 e 1: i ( e 1:# i 1 M 6 5 γ. This boud is worse, so i holds i boh cases. The resul ow follows from Azuma s iequaliy. lm Le β *4= B! # 1 4, a marix L -orm o B! # 1, so ha for ay w, 4 B! # 1 w4 5 β 4 w4. Corollary 1. For δ 0 1 -, ad all from 1 o T, where J δ γ!o 1 γ M 2l 2) δ!. Pr 4 ĉ 1: ( c 1: 4 i β J δ γ! 5 δ 6k

8 ml # Proof. Solvig δ) 2e λ2 2 yields λ 2l 2) δ!, ad he usig his value i Theorem (1) gives for all i 1 2. The, Pr 6 e 1: i 6 i J δ γ! 5 δ) Pr 4 e 1: 4 i J δ γ! 5 5 δ i$ 1 Pr 6 e 1: i 6 i J δ γ! 1: by he uio boud. Now, oice ha we ca relae ˆ ad ĉ 1: by B! # 1ˆ 1: B! # 1 τ τ$ 1 B! # 1 τ τ$ 1 τ$ 1 ĉ τ ĉ 1: ad similarly for 1: ad c 1:. The Pr 4 ĉ 1: ( c 1: 4 i β J δ γ! Pr 4 B! # 1 ˆ 1: 1: (?!p4 i β J δ γ! 5 Pr β 4 e 1: 4 i β J δ γ! Pr 4 e 1: 4 i J δ γ! 5 δ We ca ow prove our mai resul for he secio, a saeme of he form of Equaio (4) relaig PT ad PT: Theorem 2. If we play + agais BGA for T imeseps, E, PT - 5 E, PT -g% 1 ( δ! 3 2 Dβ J δ γ! T % δmt Proof. Le Φ ĉ 1:T ( c 1:T. By ĉ 1:T! ĉ 1:T c 1:T! ĉ 1:T or c 1:T % Φ! c 1:T % Φ! c 1:T! c 1:T % Φ!, ad so by expadig ad rearragig we have c 1:T % Φ! c 1:T (;@A c 1:T! c 1:T 5 >@A c 1:T!q(;@7 c 1:T % Φ!! Φ 5 D4 Φ4 (5) 6 PT ( c 1:T! c 1:T (2@A c 1:T % Φ! c 1:T % Φ!6 5 6s>@A c 1:T!q(2@A c 1:T % Φ!! c 1:T 6 % 6@A c 1:T % Φ! Φ6 5 D % D) 2!p4 Φ4

9 D) 2 for where we have used Equaio (5). Recall from Secio (2), we assume 4 x4 1 5 all x S, so 4 x ( y4 1 5 D for all x y S. The heorem follows by applyig he boud o Φ give by Corollary (1), ad he observig ha he above relaioship holds for a leas a 1 ( δ fracio of he possible algorihm hisories. For he oher δ fracio, he differece migh be as much as δmt. Wriig he overall expecaio as he sum of wo expecaios codiioed o wheher or o he boud holds gives he resul. lm 4.3 Relaig he Loss of BGA ad is GEX Subrouie Now we prove a saeme like Equaio (3), relaig loss BGA! o loss GEX!. Theorem 3. If we ru BGA wih parameer γ agais + for T imeseps, E, loss BGA! ( γ! E, loss GEX! -8% γmt Proof. For a give adversary +, G # 1 fully deermies he sequece of cos vecors give o algorihm GEX. So, we ca view GEX as a fucio from G # 1 o probabiliy disribuios over S. If we prese a cos vecor ĉ o GEX, he he expeced cos o GEX give hisory G # 1 is x' S Pr x 6 G # 1!& ĉ x!. If we defie x x' S Pr x 6 G # 1! x, we ca re-wrie he expeced loss of GEX agais ĉ as ĉ x ; ha is, we ca view GEX as icurrig he cos of some covex combiaio of he possible decisios i expecaio. Le ˆu j be ˆ give ha we explore by playig basis vecor b j o ime, ad similarly le ĉ u j R B! # 1ˆu j. bserve ha ˆu j i γ i for j i ad 0 oherwise, ad so j$ 1 ˆ u j γ γ B c (6) Now, we ca wrie E, ẑ 6 G # 1- S 1 ( γ! 0 % γ j$ 1 γ j$ 1 ĉu 1 j γ B! # 1 c x j$ 1 x ˆ 1 x ' S Pr x 6 G # 1!v ĉ u j x! u j x ad usig Equaio (6), Now, we cosider he codiioal expecaio of z ad see ha E, z 6 G # 1- S 1 ( γ! c x! % γ i$ 1 c b i! 5 1 ( γ! E, ẑ 6 G # 1-8% γm (7) 1

10 The we have, E, z - E E, z 6 G # 1-5 E 1 ( γ! E, ẑ 6 G # 1-g% γm w 1 ( γ! E E, ẑ 6 G # 1- % γm w 1 ( γ! E, ẑ -g% γm (8) by usig he iequaliy from Equaio (7). The heorem follows by summig he iequaliy (8) over from 1 o T ad applyig lieariy of expecaio. lm 4.4 A Boud o he Expeced Regre of BGA Theorem 4. If we ru BGA wih parameer γ usig subrouie GEX wih parameer ε (as defied i Appedix A), he for all δ 0 1 -, E, loss BGA! - 5 E, PT -g% D 1 γ M 2l 2) δ! T % δmt % ε γ 2 3 M 2 T % % γmt ε Proof. I Appedix A, we show a algorihm o plug i for GEX, based o he FPL algorihm of [1] ad give bouds o regre agais a deermiisic adapive adversary. We firs show how o apply ha aalysis o GEX ruig as a subrouie o BGA. Bw Firs, we eed o boud 6 ĉ x6. By defiiio, for ay x S, we ca wrie x for weighs w wih w i,u( (or,t( if i is a approximae bariceric spaer). Noe ha 4 ˆ γ! M, ad for ay x S, we ca wrie x as Bw where w i,u( Thus, Le R 6 ĉ x686s B! # 1ˆ Bw6qx6y ˆ! B # 1 Bw686 ˆ w6 5 4 ˆ w4 5 2M γ 2M) γ. Suppose a he begiig of ime we fix he radom decisios of BGA ha are o made by GEX, ha is, we fix a sequece X z, χ 1 b 1 χ T b T-. Fixig his radomess ogeher wih + ˆ deermies a ew deermiisic adapive adversary + ha GEX is effecively playig agais. To see his, le h # 1 z, x 1 x # 1-. If we combie h # 1 wih he iformaio i X, i fully deermies a parial hisory G # 1. If we le h # 1, x 1 x # 1- be he parial decisio hisory ha ca be recovered from G # 1, he +2 ˆ h # 1!o χ d +2 h# γ 1!. Thus, whe GEX is ru as a subrouie of BGA, we ca apply Lemma (3) from he Appedix ad coclude E, loss GEX!{6 X - 5 E, PT 6 X -g% ε 4 % 2! R 2 T % 4 ε For he remaider of his proof, we use big-h oaio o simplify he preseaio. Now, akig he expecaio of boh sides of Equaio (9), E, loss GEX! - 5 E, PT -g% εr 2 T % ε (9)

11 Applyig Theorem (3), E, loss BGA! - ad he usig Theorem (2) we have E, loss BGA! ( γ! E, PT -g% εr 2 T % % γmt ε % γmt 5 1 ( γ! E, PT -g% J δ γ! D T % δmt % εr 2 T % ε 5 E, PT -8% D 1 γ M 2l 2) δ! T % δmt % ε γ 2 3 M 2 T % % γmt ε For he las lie, oe ha while E, PT - could be egaive, i is sill bouded by MT, ad so his jus adds aoher γmt erm, which is capured i he big-h erm. lm Igorig he depedece o, M, ad D ad simplifyig, we see BGA s expeced regre is bouded by E, regre BGA! - z T l 1) δ! γ % δt % εt % 1 % γ 2 γt ε Seig γ δ T # 1 4 ad ε T # 3 4, we ge a boud o our loss of order " T 3 4 lt!. 5 Coclusios ad pe Problems We have preseed a geeral algorihm for olie opimizaio over a arbirary se of decisios S R, ad proved regre bouds for our algorihm ha hold agais a adapive adversary. A umber of quesios are raised by his work. I he fla badis problem, bouds of he form " T! are possible agais a adapive adversary [4]. Agais a oblivious adversary i he geomeric case, a boud of " T 2 3! is achieved i [2]. We achieve a boud of " T 3 4 lt! for his problem agais a adapive adversary. I [4], lower bouds are give showig ha he T! resul is igh, bu o such bouds are kow for he geomeric decisio-space problem. Ca he T 3 4 lt! ad possibly he " T 2 3! bouds be igheed o " T!? A relaed issue is he use of iformaio received by he algorihm; our algorihm ad he algorihm of [2] oly use a γ fracio of he feedback hey receive, which is iuiively uappealig. I seems plausible ha a algorihm ca be foud ha uses all of he feedback, possibly achievig igher bouds. Ackowledgmes The auhors wish o hak Adam Kalai, Geoff Gordo, Bobby Kleiberg, Tom Hayes, ad Varsha Dai for useful coversaios ad correspodece. Fudig provided by NSF gras CCR , NSF-ITR CCR , ad NSF-ITR IIS

12 Refereces 1. Kalai, A., Vempala, S.: Efficie algorihms for o-lie opimizaio. I: Proceedigs of he The 16h Aual Coferece o Learig Theory. (2003) 2. Awerbuch, B., Kleiberg, R.: Adapive rouig wih ed-o-ed feedback: Disribued learig ad geomeric approaches. I: Proceedigs of he 36h ACM Symposium o Theory of Compuig. (2004) To appear. 3. Takimoo, E., Warmuh, M.K.: Pah kerels ad muliplicaive updaes. I: Proceedigs of he 15h Aual Coferece o Compuaioal Learig Theory. Lecure Noes i Arificial Ielligece, Spriger (2002) 4. Auer, P., Cesa-Biachi, N., Freud, Y., Schapire, R.E.: The osochasic muliarmed badi problem. SIAM Joural o Compuig 32 (2002) Auer, P., Cesa-Biachi, N., Freud, Y., Schapire, R.E.: Gamblig i a rigged casio: he adversarial muli-armed badi problem. I: Proceedigs of he 36h Aual Symposium o Foudaios of Compuer Sciece, IEEE Compuer Sociey Press, Los Alamios, CA (1995) Zikevich, M.: lie covex programmig ad geeralized ifiiesimal gradie asce. I: Proceedigs of he Tweieh Ieraioal Coferece o Machie Learig. (2003) 7. Mowai, R., Raghava, P.: Radomized algorihms. Cambridge Uiversiy Press (1995)

13 A Specificaio of a Geomeric Expers Algorihm I his secio we poi ou how he FPL algorihm ad aalysis of [1] ca be adaped o our seig o use as he GEX subrouie, ad prove he correspodig boud eeded for Theorem (4). I paricular, we eed a boud for a arbirary S ad arbirary cos vecors, requirig oly ha o each imesep, 6 c x6 5 R. Furher, he boud mus hold agais a adapive adversary. FPL solves he olie opimizaio problem whe he eire cos vecor c is observed a each imesep. I maiais he sum c 1:# 1, ad o each imesep plays decisio c 1:# 1 % µ!, where µ is chose uiformly a radom from, 0 1) ε -, give ε, a parameer of he algorihm. The aalysis of FPL i [1] assumes posiive cos vecors c saisfyig 4 c4 1 5 A, ad posiive decisio vecors from S wih 4 x ( y4 1 5 D for all x y S ad 6 c x ( c y6 5 R for all cos vecors c ad x y S. Furher, he bouds proved are wih respec o a fixed series of cos vecors, o a adapive adversary. We ow show how o bridge he gap from hese assumpios o our assumpios. Firs, we adap a argume from [2], showig ha by usig our bariceric spaer basis, we ca rasform our problem io oe where he assumpios of FPL are me. We he argue ha a correspodig boud holds agais a adapive adversary. Lemma 1. Le S j be a se of (o ecessarily posiive) decisios, ad k 7, c 1 c T- a se of cos vecors o hose decisios, such ha 6 c x6 5 R for all x S ad c k. The, here is a algorihm? ε! ha achieves E, loss:? ε! k! - 5 PT k! % ε 4 % 2! R 2 T % 4 ε Proof. This a adapaio of he argumes of Appedix A of [2]. Fix a bariceric spaer B N b 1 b for S. The, for each x S, le x Bw ad defie f x!},t( i$ 1 w i w 1 w -. Le f S!_ S~. For each cos vecor c defie g c!_x, R R % c bs 1 R % c b -. I is sraighforward o verify ha c x g c! f x!, ad furher g c! i 0, 4 g c!p % 1! R, ad he differece i cos of ay wo decisios agais a fixed g c! is a mos 2R. By defiiio of a bariceric spaer, w i,t( ad so he L 1 diameer of S~ is a mos 4. Noe he assumpio of posiive decisio vecors i Theorem 1 of [1] ca easily be lifed by addiively shifig he space of decisio vecors uil i is posiive. This chages he loss of he algorihm ad of he bes decisio by he same amou, so addiive regre bouds are uchaged. The resul of his lemma he follows from he boud of Theorem 1 from [1]. lm We ow eed o exed he above boud o adapive adversaries. The key poi here is ha he algorihm is self-oblivious. A self-oblivious algorihm always plays a decisio from some disribuio ha depeds oly o he cos hisory so far ad o he oucome of is previous probabilisic choices. Thus, a self-oblivious algorihm ca be viewed as a fucio from cos hisories o disribuios over decisios. For such algorihms, for ay (possibly adapive) adversary + here always exiss a oblivious adversary ha causes a leas as much regre. The idea for he proof below is due o Adam Kalai. 2 2 We hak Tom Hayes ad Varsha Dai for poiig ou a bug i he proof we had i he origial versio of his paper.

14 Lemma 2. Fix T, le H. be he se of decisio hisories of legh 0 o T ( 1, ad le K. be he se of all cos hisories of legh 0 o T ( 1. The, fix a decisio algorihm : K.&/ S!, where S! is he se of probabiliy disribuios o he se S of possible decisios. Defie Le + R> +2!o E u T T c x ( mi $ 1 x' S $ 1 be a arbirary adversary. The, here exiss a oblivious adversary +V~ such ha R: + ~! i R> +;! c x Proof. A adversary is -oblivious if is firs coss are chose obliviously; oe all adversaries are 1-oblivious. Le + be a arbirary adversary, ad suppose i is k-oblivious. If k T, we are doe. herwise, le c 1 o c k o be he firs k (obliviously chose) coss seleced by +. Expecaios are over he radom variables x 1 x T ad c 1 c T whe + plays agais, hough i his case c 1 c k are fully deermied. Le K T c 1 c T, he radom vecor correspodig o he cos hisory. Le g K T! mi x' S $ T 1 c x. Usig lieariy of expecaio, we ca spli he expeced regre R> +;! io 3 erms: k T E, c ox -g% E, c k 1 x k 1-8% E, $ 1 $ c x ( g K T! - k 2 Sice ad c 1 c k are fixed, E, x k 1- E,? c 1 o c k o! - x is also kow. Sice + is oly k-oblivious, i ges o pick c k 1 wih kowledge of x 1 x k. We have Pr c k 1!o x 1 u u x k Pr x1 x k!ƒ, +; x 1 x k!o c k 1- where is a idicaor fucio, reurig 1 if +2 x 1 x k!y c k 1 ad zero oherwise. The probabiliy Pr x 1 x k! is well defied because + ad are fixed. Imporaly, oe ha he disribuio over c k 1 is idepede of he disribuio over x k 1 ; his follows from he assumpio ha is self-oblivious, ha is, i picks is disribuios based oly o he pas cos vecors, o o is ow acios. Thus, leig L k E, $ k 1 c ox - we ca wrie T R: +;!3 L k % xe, c k 1-8% E, $ c x ( g K T! - (10) k 2 T L k % c k Pr ck 1! c k 1 x % E, 1 $ c x ( g K T!36 c k 1- dc k 1 k 2 T 5 L k % sup c k c k 1 x % E, 1 $ c x ( g K T!36 c k 1- k 2 where he sup is over all c k 1 wih Pr c k 1!&h (11) (12) 0. bserve ha he quaiy iside he supremum is well defied before ay coss or decisios are seleced, ad so + could do

15 a leas as well by selecig c k 1 obliviously o be some c ha achieves he supremum. Thus, here is a k % 1! -oblivious adversary ha causes a leas as much regre as +. Exedig his resul iducively, we coclude here is a fully oblivious (T -oblivious) adversary + ~ such ha R: + ~! i R> +2!. lm Lemma 3. The regre boud from Lemma 1 applies eve if he adversary is adapive. Proof. Firs, observe ha as log as FPL re-radomizes a each imesep, i is selfoblivious, ad so Lemma 2 applies. Suppose some adapive adversary + causes regre ha exceeds he boud i Lemma 1. We ca apply Lemma 2 o + ad cosruc a oblivious +V~ ha also exceeds he boud, a coradicio. Thus, we ca use? ε! as our GEX subrouie for full-observaio olie geomeric opimizaio.

16 B Noios of Regre I [5], a aleraive defiiio of regre is give, amely, E, loss 0u h T T! - ( mi x' E S c x (13) $ 1 This defiiio is equivale o ours i he case of a oblivious adversary, bu agais a adapive adversary he bes decisio for his defiiio is o he bes decisio for a paricular decisio hisory, bu he bes decisio if he decisio mus be chose before a cos hisory is seleced accordig o he disribuio over such hisories. I paricular, E mi x' S T $ 1 c T x 5 mi x' E S c x $ 1 ad so a boud o Equaio (1) is a leas as srog as a boud o Equaio (13). I fac, bouds o Equaio (13) ca be very poor whe he adversary is adapive. There are aural examples where he sroger defiiio (1) gives regre " T! while he weaker defiiio (13) idicaes o regre. Adapig a example from [5], le S < e 1 e (he fla badi seig) ad cosider he algorihm ha plays uiformly a radom from S. The adversary + gives c 1 0, ad if he plays e i o he firs ieraio, hereafer he adversary plays he cos vecor c where c i 0 ad c j 1 for j Q i. The expeced loss of is # 1T. For regre as defied by Equaio (13), mi x' S E, c 1:T x - # 1 T, idicaig o regre, while E, mi x' S c 1:T x! - 0, ad so he sroger defiiio idicaes " T! regre. Uforuaely, his implies like he proof echiques for bouds o expeced weak regre like hose i [4] ad [2] cao be used o ge bouds o regre as defied by Equaio (1). The problem is ha eve if we have ubiased esimaes of he coss, hese cao be used o evaluae he erm E, mi x' S $ T 1 c x! - i (1) because mi is a oliear operaor. We surmou his problem by provig high-probabiliy bouds o our esimaes of c, which allows us o use a uio boud o evaluae he expecaio over he mi operaor. Noe ha he high probabiliy bouds proved i [4] ad [2] ca be see as correspodig o our defiiio of expeced regre.

BEST LINEAR FORECASTS VS. BEST POSSIBLE FORECASTS

BEST LINEAR FORECASTS VS. BEST POSSIBLE FORECASTS BEST LINEAR FORECASTS VS. BEST POSSIBLE FORECASTS Opimal ear Forecasig Alhough we have o meioed hem explicily so far i he course, here are geeral saisical priciples for derivig he bes liear forecas, ad

More information

Supplement for SADAGRAD: Strongly Adaptive Stochastic Gradient Methods"

Supplement for SADAGRAD: Strongly Adaptive Stochastic Gradient Methods Suppleme for SADAGRAD: Srogly Adapive Sochasic Gradie Mehods" Zaiyi Che * 1 Yi Xu * Ehog Che 1 iabao Yag 1. Proof of Proposiio 1 Proposiio 1. Le ɛ > 0 be fixed, H 0 γi, γ g, EF (w 1 ) F (w ) ɛ 0 ad ieraio

More information

Extremal graph theory II: K t and K t,t

Extremal graph theory II: K t and K t,t Exremal graph heory II: K ad K, Lecure Graph Theory 06 EPFL Frak de Zeeuw I his lecure, we geeralize he wo mai heorems from he las lecure, from riagles K 3 o complee graphs K, ad from squares K, o complee

More information

1 Notes on Little s Law (l = λw)

1 Notes on Little s Law (l = λw) Copyrigh c 26 by Karl Sigma Noes o Lile s Law (l λw) We cosider here a famous ad very useful law i queueig heory called Lile s Law, also kow as l λw, which assers ha he ime average umber of cusomers i

More information

Lecture 9: Polynomial Approximations

Lecture 9: Polynomial Approximations CS 70: Complexiy Theory /6/009 Lecure 9: Polyomial Approximaios Isrucor: Dieer va Melkebeek Scribe: Phil Rydzewski & Piramaayagam Arumuga Naiar Las ime, we proved ha o cosa deph circui ca evaluae he pariy

More information

High-Probability Regret Bounds for Bandit Online Linear Optimization

High-Probability Regret Bounds for Bandit Online Linear Optimization High-Probabiliy Regre Bouds for Badi Olie Liear Opimizaio Peer L. Barle UC Berkeley barle@cs.berkeley.edu Varsha Dai Uiversiy of Chicago varsha@cs.uchicago.edu Aleader Rakhli UC Berkeley rakhli@cs.berkeley.edu

More information

COS 522: Complexity Theory : Boaz Barak Handout 10: Parallel Repetition Lemma

COS 522: Complexity Theory : Boaz Barak Handout 10: Parallel Repetition Lemma COS 522: Complexiy Theory : Boaz Barak Hadou 0: Parallel Repeiio Lemma Readig: () A Parallel Repeiio Theorem / Ra Raz (available o his websie) (2) Parallel Repeiio: Simplificaios ad he No-Sigallig Case

More information

ODEs II, Supplement to Lectures 6 & 7: The Jordan Normal Form: Solving Autonomous, Homogeneous Linear Systems. April 2, 2003

ODEs II, Supplement to Lectures 6 & 7: The Jordan Normal Form: Solving Autonomous, Homogeneous Linear Systems. April 2, 2003 ODEs II, Suppleme o Lecures 6 & 7: The Jorda Normal Form: Solvig Auoomous, Homogeeous Liear Sysems April 2, 23 I his oe, we describe he Jorda ormal form of a marix ad use i o solve a geeral homogeeous

More information

Big O Notation for Time Complexity of Algorithms

Big O Notation for Time Complexity of Algorithms BRONX COMMUNITY COLLEGE of he Ciy Uiversiy of New York DEPARTMENT OF MATHEMATICS AND COMPUTER SCIENCE CSI 33 Secio E01 Hadou 1 Fall 2014 Sepember 3, 2014 Big O Noaio for Time Complexiy of Algorihms Time

More information

Math 6710, Fall 2016 Final Exam Solutions

Math 6710, Fall 2016 Final Exam Solutions Mah 67, Fall 6 Fial Exam Soluios. Firs, a sude poied ou a suble hig: if P (X i p >, he X + + X (X + + X / ( evaluaes o / wih probabiliy p >. This is roublesome because a radom variable is supposed o be

More information

STK4080/9080 Survival and event history analysis

STK4080/9080 Survival and event history analysis STK48/98 Survival ad eve hisory aalysis Marigales i discree ime Cosider a sochasic process The process M is a marigale if Lecure 3: Marigales ad oher sochasic processes i discree ime (recap) where (formally

More information

Lecture 15 First Properties of the Brownian Motion

Lecture 15 First Properties of the Brownian Motion Lecure 15: Firs Properies 1 of 8 Course: Theory of Probabiliy II Term: Sprig 2015 Isrucor: Gorda Zikovic Lecure 15 Firs Properies of he Browia Moio This lecure deals wih some of he more immediae properies

More information

λiv Av = 0 or ( λi Av ) = 0. In order for a vector v to be an eigenvector, it must be in the kernel of λi

λiv Av = 0 or ( λi Av ) = 0. In order for a vector v to be an eigenvector, it must be in the kernel of λi Liear lgebra Lecure #9 Noes This week s lecure focuses o wha migh be called he srucural aalysis of liear rasformaios Wha are he irisic properies of a liear rasformaio? re here ay fixed direcios? The discussio

More information

L-functions and Class Numbers

L-functions and Class Numbers L-fucios ad Class Numbers Sude Number Theory Semiar S. M.-C. 4 Sepember 05 We follow Romyar Sharifi s Noes o Iwasawa Theory, wih some help from Neukirch s Algebraic Number Theory. L-fucios of Dirichle

More information

Solution. 1 Solutions of Homework 6. Sangchul Lee. April 28, Problem 1.1 [Dur10, Exercise ]

Solution. 1 Solutions of Homework 6. Sangchul Lee. April 28, Problem 1.1 [Dur10, Exercise ] Soluio Sagchul Lee April 28, 28 Soluios of Homework 6 Problem. [Dur, Exercise 2.3.2] Le A be a sequece of idepede eves wih PA < for all. Show ha P A = implies PA i.o. =. Proof. Noice ha = P A c = P A c

More information

CSE 241 Algorithms and Data Structures 10/14/2015. Skip Lists

CSE 241 Algorithms and Data Structures 10/14/2015. Skip Lists CSE 41 Algorihms ad Daa Srucures 10/14/015 Skip Liss This hadou gives he skip lis mehods ha we discussed i class. A skip lis is a ordered, doublyliked lis wih some exra poiers ha allow us o jump over muliple

More information

FIXED FUZZY POINT THEOREMS IN FUZZY METRIC SPACE

FIXED FUZZY POINT THEOREMS IN FUZZY METRIC SPACE Mohia & Samaa, Vol. 1, No. II, December, 016, pp 34-49. ORIGINAL RESEARCH ARTICLE OPEN ACCESS FIED FUZZY POINT THEOREMS IN FUZZY METRIC SPACE 1 Mohia S. *, Samaa T. K. 1 Deparme of Mahemaics, Sudhir Memorial

More information

Online Supplement to Reactive Tabu Search in a Team-Learning Problem

Online Supplement to Reactive Tabu Search in a Team-Learning Problem Olie Suppleme o Reacive abu Search i a eam-learig Problem Yueli She School of Ieraioal Busiess Admiisraio, Shaghai Uiversiy of Fiace ad Ecoomics, Shaghai 00433, People s Republic of Chia, she.yueli@mail.shufe.edu.c

More information

Lecture 8 April 18, 2018

Lecture 8 April 18, 2018 Sas 300C: Theory of Saisics Sprig 2018 Lecure 8 April 18, 2018 Prof Emmauel Cades Scribe: Emmauel Cades Oulie Ageda: Muliple Tesig Problems 1 Empirical Process Viewpoi of BHq 2 Empirical Process Viewpoi

More information

Comparison between Fourier and Corrected Fourier Series Methods

Comparison between Fourier and Corrected Fourier Series Methods Malaysia Joural of Mahemaical Scieces 7(): 73-8 (13) MALAYSIAN JOURNAL OF MATHEMATICAL SCIENCES Joural homepage: hp://eispem.upm.edu.my/oural Compariso bewee Fourier ad Correced Fourier Series Mehods 1

More information

Section 8 Convolution and Deconvolution

Section 8 Convolution and Deconvolution APPLICATIONS IN SIGNAL PROCESSING Secio 8 Covoluio ad Decovoluio This docume illusraes several echiques for carryig ou covoluio ad decovoluio i Mahcad. There are several operaors available for hese fucios:

More information

F D D D D F. smoothed value of the data including Y t the most recent data.

F D D D D F. smoothed value of the data including Y t the most recent data. Module 2 Forecasig 1. Wha is forecasig? Forecasig is defied as esimaig he fuure value ha a parameer will ake. Mos scieific forecasig mehods forecas he fuure value usig pas daa. I Operaios Maageme forecasig

More information

An interesting result about subset sums. Nitu Kitchloo. Lior Pachter. November 27, Abstract

An interesting result about subset sums. Nitu Kitchloo. Lior Pachter. November 27, Abstract A ieresig resul abou subse sums Niu Kichloo Lior Pacher November 27, 1993 Absrac We cosider he problem of deermiig he umber of subses B f1; 2; : : :; g such ha P b2b b k mod, where k is a residue class

More information

OLS bias for econometric models with errors-in-variables. The Lucas-critique Supplementary note to Lecture 17

OLS bias for econometric models with errors-in-variables. The Lucas-critique Supplementary note to Lecture 17 OLS bias for ecoomeric models wih errors-i-variables. The Lucas-criique Supplemeary oe o Lecure 7 RNy May 6, 03 Properies of OLS i RE models I Lecure 7 we discussed he followig example of a raioal expecaios

More information

Ideal Amplifier/Attenuator. Memoryless. where k is some real constant. Integrator. System with memory

Ideal Amplifier/Attenuator. Memoryless. where k is some real constant. Integrator. System with memory Liear Time-Ivaria Sysems (LTI Sysems) Oulie Basic Sysem Properies Memoryless ad sysems wih memory (saic or dyamic) Causal ad o-causal sysems (Causaliy) Liear ad o-liear sysems (Lieariy) Sable ad o-sable

More information

Outline. simplest HMM (1) simple HMMs? simplest HMM (2) Parameter estimation for discrete hidden Markov models

Outline. simplest HMM (1) simple HMMs? simplest HMM (2) Parameter estimation for discrete hidden Markov models Oulie Parameer esimaio for discree idde Markov models Juko Murakami () ad Tomas Taylor (2). Vicoria Uiversiy of Welligo 2. Arizoa Sae Uiversiy Descripio of simple idde Markov models Maximum likeliood esimae

More information

10.3 Autocorrelation Function of Ergodic RP 10.4 Power Spectral Density of Ergodic RP 10.5 Normal RP (Gaussian RP)

10.3 Autocorrelation Function of Ergodic RP 10.4 Power Spectral Density of Ergodic RP 10.5 Normal RP (Gaussian RP) ENGG450 Probabiliy ad Saisics for Egieers Iroducio 3 Probabiliy 4 Probabiliy disribuios 5 Probabiliy Desiies Orgaizaio ad descripio of daa 6 Samplig disribuios 7 Ifereces cocerig a mea 8 Comparig wo reames

More information

The Central Limit Theorem

The Central Limit Theorem The Ceral Limi Theorem The ceral i heorem is oe of he mos impora heorems i probabiliy heory. While here a variey of forms of he ceral i heorem, he mos geeral form saes ha give a sufficiely large umber,

More information

1. Solve by the method of undetermined coefficients and by the method of variation of parameters. (4)

1. Solve by the method of undetermined coefficients and by the method of variation of parameters. (4) 7 Differeial equaios Review Solve by he mehod of udeermied coefficies ad by he mehod of variaio of parameers (4) y y = si Soluio; we firs solve he homogeeous equaio (4) y y = 4 The correspodig characerisic

More information

LINEAR APPROXIMATION OF THE BASELINE RBC MODEL JANUARY 29, 2013

LINEAR APPROXIMATION OF THE BASELINE RBC MODEL JANUARY 29, 2013 LINEAR APPROXIMATION OF THE BASELINE RBC MODEL JANUARY 29, 203 Iroducio LINEARIZATION OF THE RBC MODEL For f( x, y, z ) = 0, mulivariable Taylor liear expasio aroud f( x, y, z) f( x, y, z) + f ( x, y,

More information

Moment Generating Function

Moment Generating Function 1 Mome Geeraig Fucio m h mome m m m E[ ] x f ( x) dx m h ceral mome m m m E[( ) ] ( ) ( x ) f ( x) dx Mome Geeraig Fucio For a real, M () E[ e ] e k x k e p ( x ) discree x k e f ( x) dx coiuous Example

More information

th m m m m central moment : E[( X X) ] ( X X) ( x X) f ( x)

th m m m m central moment : E[( X X) ] ( X X) ( x X) f ( x) 1 Trasform Techiques h m m m m mome : E[ ] x f ( x) dx h m m m m ceral mome : E[( ) ] ( ) ( x) f ( x) dx A coveie wa of fidig he momes of a radom variable is he mome geeraig fucio (MGF). Oher rasform echiques

More information

Exercise 3 Stochastic Models of Manufacturing Systems 4T400, 6 May

Exercise 3 Stochastic Models of Manufacturing Systems 4T400, 6 May Exercise 3 Sochasic Models of Maufacurig Sysems 4T4, 6 May. Each week a very popular loery i Adorra pris 4 ickes. Each ickes has wo 4-digi umbers o i, oe visible ad he oher covered. The umbers are radomly

More information

Dynamic h-index: the Hirsch index in function of time

Dynamic h-index: the Hirsch index in function of time Dyamic h-idex: he Hirsch idex i fucio of ime by L. Egghe Uiversiei Hassel (UHassel), Campus Diepebeek, Agoralaa, B-3590 Diepebeek, Belgium ad Uiversiei Awerpe (UA), Campus Drie Eike, Uiversieisplei, B-260

More information

12 Getting Started With Fourier Analysis

12 Getting Started With Fourier Analysis Commuicaios Egieerig MSc - Prelimiary Readig Geig Sared Wih Fourier Aalysis Fourier aalysis is cocered wih he represeaio of sigals i erms of he sums of sie, cosie or complex oscillaio waveforms. We ll

More information

Additional Tables of Simulation Results

Additional Tables of Simulation Results Saisica Siica: Suppleme REGULARIZING LASSO: A CONSISTENT VARIABLE SELECTION METHOD Quefeg Li ad Ju Shao Uiversiy of Wiscosi, Madiso, Eas Chia Normal Uiversiy ad Uiversiy of Wiscosi, Madiso Supplemeary

More information

CLOSED FORM EVALUATION OF RESTRICTED SUMS CONTAINING SQUARES OF FIBONOMIAL COEFFICIENTS

CLOSED FORM EVALUATION OF RESTRICTED SUMS CONTAINING SQUARES OF FIBONOMIAL COEFFICIENTS PB Sci Bull, Series A, Vol 78, Iss 4, 2016 ISSN 1223-7027 CLOSED FORM EVALATION OF RESTRICTED SMS CONTAINING SQARES OF FIBONOMIAL COEFFICIENTS Emrah Kılıc 1, Helmu Prodiger 2 We give a sysemaic approach

More information

Using Linnik's Identity to Approximate the Prime Counting Function with the Logarithmic Integral

Using Linnik's Identity to Approximate the Prime Counting Function with the Logarithmic Integral Usig Lii's Ideiy o Approimae he Prime Couig Fucio wih he Logarihmic Iegral Naha McKezie /26/2 aha@icecreambreafas.com Summary:This paper will show ha summig Lii's ideiy from 2 o ad arragig erms i a cerai

More information

LINEAR APPROXIMATION OF THE BASELINE RBC MODEL SEPTEMBER 17, 2013

LINEAR APPROXIMATION OF THE BASELINE RBC MODEL SEPTEMBER 17, 2013 LINEAR APPROXIMATION OF THE BASELINE RBC MODEL SEPTEMBER 7, 203 Iroducio LINEARIZATION OF THE RBC MODEL For f( xyz,, ) = 0, mulivariable Taylor liear expasio aroud f( xyz,, ) f( xyz,, ) + f( xyz,, )( x

More information

Four equations describe the dynamic solution to RBC model. Consumption-leisure efficiency condition. Consumption-investment efficiency condition

Four equations describe the dynamic solution to RBC model. Consumption-leisure efficiency condition. Consumption-investment efficiency condition LINEARIZING AND APPROXIMATING THE RBC MODEL SEPTEMBER 7, 200 For f( x, y, z ), mulivariable Taylor liear expasio aroud ( x, yz, ) f ( x, y, z) f( x, y, z) + f ( x, y, z)( x x) + f ( x, y, z)( y y) + f

More information

A note on deviation inequalities on {0, 1} n. by Julio Bernués*

A note on deviation inequalities on {0, 1} n. by Julio Bernués* A oe o deviaio iequaliies o {0, 1}. by Julio Berués* Deparameo de Maemáicas. Faculad de Ciecias Uiversidad de Zaragoza 50009-Zaragoza (Spai) I. Iroducio. Le f: (Ω, Σ, ) IR be a radom variable. Roughly

More information

Procedia - Social and Behavioral Sciences 230 ( 2016 ) Joint Probability Distribution and the Minimum of a Set of Normalized Random Variables

Procedia - Social and Behavioral Sciences 230 ( 2016 ) Joint Probability Distribution and the Minimum of a Set of Normalized Random Variables Available olie a wwwsciecedireccom ScieceDirec Procedia - Social ad Behavioral Scieces 30 ( 016 ) 35 39 3 rd Ieraioal Coferece o New Challeges i Maageme ad Orgaizaio: Orgaizaio ad Leadership, May 016,

More information

Fresnel Dragging Explained

Fresnel Dragging Explained Fresel Draggig Explaied 07/05/008 Decla Traill Decla@espace.e.au The Fresel Draggig Coefficie required o explai he resul of he Fizeau experime ca be easily explaied by usig he priciples of Eergy Field

More information

MATH 507a ASSIGNMENT 4 SOLUTIONS FALL 2018 Prof. Alexander. g (x) dx = g(b) g(0) = g(b),

MATH 507a ASSIGNMENT 4 SOLUTIONS FALL 2018 Prof. Alexander. g (x) dx = g(b) g(0) = g(b), MATH 57a ASSIGNMENT 4 SOLUTIONS FALL 28 Prof. Alexader (2.3.8)(a) Le g(x) = x/( + x) for x. The g (x) = /( + x) 2 is decreasig, so for a, b, g(a + b) g(a) = a+b a g (x) dx b so g(a + b) g(a) + g(b). Sice

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 4 9/16/2013. Applications of the large deviation technique

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 4 9/16/2013. Applications of the large deviation technique MASSACHUSETTS ISTITUTE OF TECHOLOGY 6.265/5.070J Fall 203 Lecure 4 9/6/203 Applicaios of he large deviaio echique Coe.. Isurace problem 2. Queueig problem 3. Buffer overflow probabiliy Safey capial for

More information

David Randall. ( )e ikx. k = u x,t. u( x,t)e ikx dx L. x L /2. Recall that the proof of (1) and (2) involves use of the orthogonality condition.

David Randall. ( )e ikx. k = u x,t. u( x,t)e ikx dx L. x L /2. Recall that the proof of (1) and (2) involves use of the orthogonality condition. ! Revised April 21, 2010 1:27 P! 1 Fourier Series David Radall Assume ha u( x,) is real ad iegrable If he domai is periodic, wih period L, we ca express u( x,) exacly by a Fourier series expasio: ( ) =

More information

N! AND THE GAMMA FUNCTION

N! AND THE GAMMA FUNCTION N! AND THE GAMMA FUNCTION Cosider he produc of he firs posiive iegers- 3 4 5 6 (-) =! Oe calls his produc he facorial ad has ha produc of he firs five iegers equals 5!=0. Direcly relaed o he discree! fucio

More information

B. Maddah INDE 504 Simulation 09/02/17

B. Maddah INDE 504 Simulation 09/02/17 B. Maddah INDE 54 Simulaio 9/2/7 Queueig Primer Wha is a queueig sysem? A queueig sysem cosiss of servers (resources) ha provide service o cusomers (eiies). A Cusomer requesig service will sar service

More information

Mean Square Convergent Finite Difference Scheme for Stochastic Parabolic PDEs

Mean Square Convergent Finite Difference Scheme for Stochastic Parabolic PDEs America Joural of Compuaioal Mahemaics, 04, 4, 80-88 Published Olie Sepember 04 i SciRes. hp://www.scirp.org/joural/ajcm hp://dx.doi.org/0.436/ajcm.04.4404 Mea Square Coverge Fiie Differece Scheme for

More information

A Note on Random k-sat for Moderately Growing k

A Note on Random k-sat for Moderately Growing k A Noe o Radom k-sat for Moderaely Growig k Ju Liu LMIB ad School of Mahemaics ad Sysems Sciece, Beihag Uiversiy, Beijig, 100191, P.R. Chia juliu@smss.buaa.edu.c Zogsheg Gao LMIB ad School of Mahemaics

More information

Conditional Probability and Conditional Expectation

Conditional Probability and Conditional Expectation Hadou #8 for B902308 prig 2002 lecure dae: 3/06/2002 Codiioal Probabiliy ad Codiioal Epecaio uppose X ad Y are wo radom variables The codiioal probabiliy of Y y give X is } { }, { } { X P X y Y P X y Y

More information

The analysis of the method on the one variable function s limit Ke Wu

The analysis of the method on the one variable function s limit Ke Wu Ieraioal Coferece o Advaces i Mechaical Egieerig ad Idusrial Iformaics (AMEII 5) The aalysis of he mehod o he oe variable fucio s i Ke Wu Deparme of Mahemaics ad Saisics Zaozhuag Uiversiy Zaozhuag 776

More information

Review Answers for E&CE 700T02

Review Answers for E&CE 700T02 Review Aswers for E&CE 700T0 . Deermie he curre soluio, all possible direcios, ad sepsizes wheher improvig or o for he simple able below: 4 b ma c 0 0 0-4 6 0 - B N B N ^0 0 0 curre sol =, = Ch for - -

More information

A TAUBERIAN THEOREM FOR THE WEIGHTED MEAN METHOD OF SUMMABILITY

A TAUBERIAN THEOREM FOR THE WEIGHTED MEAN METHOD OF SUMMABILITY U.P.B. Sci. Bull., Series A, Vol. 78, Iss. 2, 206 ISSN 223-7027 A TAUBERIAN THEOREM FOR THE WEIGHTED MEAN METHOD OF SUMMABILITY İbrahim Çaak I his paper we obai a Tauberia codiio i erms of he weighed classical

More information

Electrical Engineering Department Network Lab.

Electrical Engineering Department Network Lab. Par:- Elecrical Egieerig Deparme Nework Lab. Deermiaio of differe parameers of -por eworks ad verificaio of heir ierrelaio ships. Objecive: - To deermie Y, ad ABD parameers of sigle ad cascaded wo Por

More information

11. Adaptive Control in the Presence of Bounded Disturbances Consider MIMO systems in the form,

11. Adaptive Control in the Presence of Bounded Disturbances Consider MIMO systems in the form, Lecure 6. Adapive Corol i he Presece of Bouded Disurbaces Cosider MIMO sysems i he form, x Aref xbu x Bref ycmd (.) y Cref x operaig i he presece of a bouded ime-depede disurbace R. All he assumpios ad

More information

TAKA KUSANO. laculty of Science Hrosh tlnlersty 1982) (n-l) + + Pn(t)x 0, (n-l) + + Pn(t)Y f(t,y), XR R are continuous functions.

TAKA KUSANO. laculty of Science Hrosh tlnlersty 1982) (n-l) + + Pn(t)x 0, (n-l) + + Pn(t)Y f(t,y), XR R are continuous functions. Iera. J. Mah. & Mah. Si. Vol. 6 No. 3 (1983) 559-566 559 ASYMPTOTIC RELATIOHIPS BETWEEN TWO HIGHER ORDER ORDINARY DIFFERENTIAL EQUATIONS TAKA KUSANO laculy of Sciece Hrosh llersy 1982) ABSTRACT. Some asympoic

More information

A Note on Prediction with Misspecified Models

A Note on Prediction with Misspecified Models ITB J. Sci., Vol. 44 A, No. 3,, 7-9 7 A Noe o Predicio wih Misspecified Models Khresha Syuhada Saisics Research Divisio, Faculy of Mahemaics ad Naural Scieces, Isiu Tekologi Badug, Jala Gaesa Badug, Jawa

More information

Some Properties of Semi-E-Convex Function and Semi-E-Convex Programming*

Some Properties of Semi-E-Convex Function and Semi-E-Convex Programming* The Eighh Ieraioal Symposium o Operaios esearch ad Is Applicaios (ISOA 9) Zhagjiajie Chia Sepember 2 22 29 Copyrigh 29 OSC & APOC pp 33 39 Some Properies of Semi-E-Covex Fucio ad Semi-E-Covex Programmig*

More information

Fermat Numbers in Multinomial Coefficients

Fermat Numbers in Multinomial Coefficients 1 3 47 6 3 11 Joural of Ieger Sequeces, Vol. 17 (014, Aricle 14.3. Ferma Numbers i Muliomial Coefficies Shae Cher Deparme of Mahemaics Zhejiag Uiversiy Hagzhou, 31007 Chia chexiaohag9@gmail.com Absrac

More information

Actuarial Society of India

Actuarial Society of India Acuarial Sociey of Idia EXAMINAIONS Jue 5 C4 (3) Models oal Marks - 5 Idicaive Soluio Q. (i) a) Le U deoe he process described by 3 ad V deoe he process described by 4. he 5 e 5 PU [ ] PV [ ] ( e ).538!

More information

LIMITS OF FUNCTIONS (I)

LIMITS OF FUNCTIONS (I) LIMITS OF FUNCTIO (I ELEMENTARY FUNCTIO: (Elemeary fucios are NOT piecewise fucios Cosa Fucios: f(x k, where k R Polyomials: f(x a + a x + a x + a x + + a x, where a, a,..., a R Raioal Fucios: f(x P (x,

More information

INTEGER INTERVAL VALUE OF NEWTON DIVIDED DIFFERENCE AND FORWARD AND BACKWARD INTERPOLATION FORMULA

INTEGER INTERVAL VALUE OF NEWTON DIVIDED DIFFERENCE AND FORWARD AND BACKWARD INTERPOLATION FORMULA Volume 8 No. 8, 45-54 ISSN: 34-3395 (o-lie versio) url: hp://www.ijpam.eu ijpam.eu INTEGER INTERVAL VALUE OF NEWTON DIVIDED DIFFERENCE AND FORWARD AND BACKWARD INTERPOLATION FORMULA A.Arul dass M.Dhaapal

More information

Research Article A Generalized Nonlinear Sum-Difference Inequality of Product Form

Research Article A Generalized Nonlinear Sum-Difference Inequality of Product Form Joural of Applied Mahemaics Volume 03, Aricle ID 47585, 7 pages hp://dx.doi.org/0.55/03/47585 Research Aricle A Geeralized Noliear Sum-Differece Iequaliy of Produc Form YogZhou Qi ad Wu-Sheg Wag School

More information

Notes 03 largely plagiarized by %khc

Notes 03 largely plagiarized by %khc 1 1 Discree-Time Covoluio Noes 03 largely plagiarized by %khc Le s begi our discussio of covoluio i discree-ime, sice life is somewha easier i ha domai. We sar wih a sigal x[] ha will be he ipu io our

More information

FORBIDDING HAMILTON CYCLES IN UNIFORM HYPERGRAPHS

FORBIDDING HAMILTON CYCLES IN UNIFORM HYPERGRAPHS FORBIDDING HAMILTON CYCLES IN UNIFORM HYPERGRAPHS JIE HAN AND YI ZHAO Absrac For d l

More information

CSE 202: Design and Analysis of Algorithms Lecture 16

CSE 202: Design and Analysis of Algorithms Lecture 16 CSE 202: Desig ad Aalysis of Algorihms Lecure 16 Isrucor: Kamalia Chaudhuri Iequaliy 1: Marov s Iequaliy Pr(X=x) Pr(X >= a) 0 x a If X is a radom variable which aes o-egaive values, ad a > 0, he Pr[X a]

More information

Four equations describe the dynamic solution to RBC model. Consumption-leisure efficiency condition. Consumption-investment efficiency condition

Four equations describe the dynamic solution to RBC model. Consumption-leisure efficiency condition. Consumption-investment efficiency condition LINEAR APPROXIMATION OF THE BASELINE RBC MODEL FEBRUARY, 202 Iroducio For f(, y, z ), mulivariable Taylor liear epasio aroud (, yz, ) f (, y, z) f(, y, z) + f (, y, z)( ) + f (, y, z)( y y) + f (, y, z)(

More information

NEWTON METHOD FOR DETERMINING THE OPTIMAL REPLENISHMENT POLICY FOR EPQ MODEL WITH PRESENT VALUE

NEWTON METHOD FOR DETERMINING THE OPTIMAL REPLENISHMENT POLICY FOR EPQ MODEL WITH PRESENT VALUE Yugoslav Joural of Operaios Research 8 (2008, Number, 53-6 DOI: 02298/YUJOR080053W NEWTON METHOD FOR DETERMINING THE OPTIMAL REPLENISHMENT POLICY FOR EPQ MODEL WITH PRESENT VALUE Jeff Kuo-Jug WU, Hsui-Li

More information

Calculus Limits. Limit of a function.. 1. One-Sided Limits...1. Infinite limits 2. Vertical Asymptotes...3. Calculating Limits Using the Limit Laws.

Calculus Limits. Limit of a function.. 1. One-Sided Limits...1. Infinite limits 2. Vertical Asymptotes...3. Calculating Limits Using the Limit Laws. Limi of a fucio.. Oe-Sided..... Ifiie limis Verical Asympoes... Calculaig Usig he Limi Laws.5 The Squeeze Theorem.6 The Precise Defiiio of a Limi......7 Coiuiy.8 Iermediae Value Theorem..9 Refereces..

More information

K3 p K2 p Kp 0 p 2 p 3 p

K3 p K2 p Kp 0 p 2 p 3 p Mah 80-00 Mo Ar 0 Chaer 9 Fourier Series ad alicaios o differeial equaios (ad arial differeial equaios) 9.-9. Fourier series defiiio ad covergece. The idea of Fourier series is relaed o he liear algebra

More information

Available online at J. Math. Comput. Sci. 4 (2014), No. 4, ISSN:

Available online at   J. Math. Comput. Sci. 4 (2014), No. 4, ISSN: Available olie a hp://sci.org J. Mah. Compu. Sci. 4 (2014), No. 4, 716-727 ISSN: 1927-5307 ON ITERATIVE TECHNIQUES FOR NUMERICAL SOLUTIONS OF LINEAR AND NONLINEAR DIFFERENTIAL EQUATIONS S.O. EDEKI *, A.A.

More information

ECE-314 Fall 2012 Review Questions

ECE-314 Fall 2012 Review Questions ECE-34 Fall 0 Review Quesios. A liear ime-ivaria sysem has he ipu-oupu characerisics show i he firs row of he diagram below. Deermie he oupu for he ipu show o he secod row of he diagram. Jusify your aswer.

More information

Review Exercises for Chapter 9

Review Exercises for Chapter 9 0_090R.qd //0 : PM Page 88 88 CHAPTER 9 Ifiie Series I Eercises ad, wrie a epressio for he h erm of he sequece..,., 5, 0,,,, 0,... 7,... I Eercises, mach he sequece wih is graph. [The graphs are labeled

More information

Approximately Quasi Inner Generalized Dynamics on Modules. { } t t R

Approximately Quasi Inner Generalized Dynamics on Modules. { } t t R Joural of Scieces, Islamic epublic of Ira 23(3): 245-25 (22) Uiversiy of Tehra, ISSN 6-4 hp://jscieces.u.ac.ir Approximaely Quasi Ier Geeralized Dyamics o Modules M. Mosadeq, M. Hassai, ad A. Nikam Deparme

More information

xp (X = x) = P (X = 1) = θ. Hence, the method of moments estimator of θ is

xp (X = x) = P (X = 1) = θ. Hence, the method of moments estimator of θ is Exercise 7 / page 356 Noe ha X i are ii from Beroulli(θ where 0 θ a Meho of momes: Sice here is oly oe parameer o be esimae we ee oly oe equaio where we equae he rs sample mome wih he rs populaio mome,

More information

INVESTMENT PROJECT EFFICIENCY EVALUATION

INVESTMENT PROJECT EFFICIENCY EVALUATION 368 Miljeko Crjac Domiika Crjac INVESTMENT PROJECT EFFICIENCY EVALUATION Miljeko Crjac Professor Faculy of Ecoomics Drsc Domiika Crjac Faculy of Elecrical Egieerig Osijek Summary Fiacial efficiecy of ivesme

More information

Solutions to Problems 3, Level 4

Solutions to Problems 3, Level 4 Soluios o Problems 3, Level 4 23 Improve he resul of Quesio 3 whe l. i Use log log o prove ha for real >, log ( {}log + 2 d log+ P ( + P ( d 2. Here P ( is defied i Quesio, ad parial iegraio has bee used.

More information

Economics 8723 Macroeconomic Theory Problem Set 2 Professor Sanjay Chugh Spring 2017

Economics 8723 Macroeconomic Theory Problem Set 2 Professor Sanjay Chugh Spring 2017 Deparme of Ecoomics The Ohio Sae Uiversiy Ecoomics 8723 Macroecoomic Theory Problem Se 2 Professor Sajay Chugh Sprig 207 Labor Icome Taxes, Nash-Bargaied Wages, ad Proporioally-Bargaied Wages. I a ecoomy

More information

Convergence theorems. Chapter Sampling

Convergence theorems. Chapter Sampling Chaper Covergece heorems We ve already discussed he difficuly i defiig he probabiliy measure i erms of a experimeal frequecy measureme. The hear of he problem lies i he defiiio of he limi, ad his was se

More information

If boundary values are necessary, they are called mixed initial-boundary value problems. Again, the simplest prototypes of these IV problems are:

If boundary values are necessary, they are called mixed initial-boundary value problems. Again, the simplest prototypes of these IV problems are: 3. Iiial value problems: umerical soluio Fiie differeces - Trucaio errors, cosisecy, sabiliy ad covergece Crieria for compuaioal sabiliy Explici ad implici ime schemes Table of ime schemes Hyperbolic ad

More information

Department of Mathematical and Statistical Sciences University of Alberta

Department of Mathematical and Statistical Sciences University of Alberta MATH 4 (R) Wier 008 Iermediae Calculus I Soluios o Problem Se # Due: Friday Jauary 8, 008 Deparme of Mahemaical ad Saisical Scieces Uiversiy of Albera Quesio. [Sec.., #] Fid a formula for he geeral erm

More information

6/10/2014. Definition. Time series Data. Time series Graph. Components of time series. Time series Seasonal. Time series Trend

6/10/2014. Definition. Time series Data. Time series Graph. Components of time series. Time series Seasonal. Time series Trend 6//4 Defiiio Time series Daa A ime series Measures he same pheomeo a equal iervals of ime Time series Graph Compoes of ime series 5 5 5-5 7 Q 7 Q 7 Q 3 7 Q 4 8 Q 8 Q 8 Q 3 8 Q 4 9 Q 9 Q 9 Q 3 9 Q 4 Q Q

More information

Toward Learning Gaussian Mixtures with Arbitrary Separation

Toward Learning Gaussian Mixtures with Arbitrary Separation Toward Learig Gaussia Mixures wih Arbirary Separaio Mikhail Belki Ohio Sae Uiversiy Columbus, Ohio mbelki@cse.ohio-sae.edu Kaushik Siha Ohio Sae Uiversiy Columbus, Ohio sihak@cse.ohio-sae.edu Absrac I

More information

Samuel Sindayigaya 1, Nyongesa L. Kennedy 2, Adu A.M. Wasike 3

Samuel Sindayigaya 1, Nyongesa L. Kennedy 2, Adu A.M. Wasike 3 Ieraioal Joural of Saisics ad Aalysis. ISSN 48-9959 Volume 6, Number (6, pp. -8 Research Idia Publicaios hp://www.ripublicaio.com The Populaio Mea ad is Variace i he Presece of Geocide for a Simple Birh-Deah-

More information

C(p, ) 13 N. Nuclear reactions generate energy create new isotopes and elements. Notation for stellar rates: p 12

C(p, ) 13 N. Nuclear reactions generate energy create new isotopes and elements. Notation for stellar rates: p 12 Iroducio o sellar reacio raes Nuclear reacios geerae eergy creae ew isoopes ad elemes Noaio for sellar raes: p C 3 N C(p,) 3 N The heavier arge ucleus (Lab: arge) he ligher icomig projecile (Lab: beam)

More information

Time Dependent Queuing

Time Dependent Queuing Time Depede Queuig Mark S. Daski Deparme of IE/MS, Norhweser Uiversiy Evaso, IL 628 Sprig, 26 Oulie Will look a M/M/s sysem Numerically iegraio of Chapma- Kolmogorov equaios Iroducio o Time Depede Queue

More information

State and Parameter Estimation of The Lorenz System In Existence of Colored Noise

State and Parameter Estimation of The Lorenz System In Existence of Colored Noise Sae ad Parameer Esimaio of he Lorez Sysem I Eisece of Colored Noise Mozhga Mombeii a Hamid Khaloozadeh b a Elecrical Corol ad Sysem Egieerig Researcher of Isiue for Research i Fudameal Scieces (IPM ehra

More information

Development of Kalman Filter and Analogs Schemes to Improve Numerical Weather Predictions

Development of Kalman Filter and Analogs Schemes to Improve Numerical Weather Predictions Developme of Kalma Filer ad Aalogs Schemes o Improve Numerical Weaher Predicios Luca Delle Moache *, Aimé Fourier, Yubao Liu, Gregory Roux, ad Thomas Warer (NCAR) Thomas Nipe, ad Rolad Sull (UBC) Wid Eergy

More information

SUMMATION OF INFINITE SERIES REVISITED

SUMMATION OF INFINITE SERIES REVISITED SUMMATION OF INFINITE SERIES REVISITED I several aricles over he las decade o his web page we have show how o sum cerai iiie series icludig he geomeric series. We wa here o eed his discussio o he geeral

More information

Homotopy Analysis Method for Solving Fractional Sturm-Liouville Problems

Homotopy Analysis Method for Solving Fractional Sturm-Liouville Problems Ausralia Joural of Basic ad Applied Scieces, 4(1): 518-57, 1 ISSN 1991-8178 Homoopy Aalysis Mehod for Solvig Fracioal Surm-Liouville Problems 1 A Neamay, R Darzi, A Dabbaghia 1 Deparme of Mahemaics, Uiversiy

More information

Comparisons Between RV, ARV and WRV

Comparisons Between RV, ARV and WRV Comparisos Bewee RV, ARV ad WRV Cao Gag,Guo Migyua School of Maageme ad Ecoomics, Tiaji Uiversiy, Tiaji,30007 Absrac: Realized Volailiy (RV) have bee widely used sice i was pu forward by Aderso ad Bollerslev

More information

Extended Laguerre Polynomials

Extended Laguerre Polynomials I J Coemp Mah Scieces, Vol 7, 1, o, 189 194 Exeded Laguerre Polyomials Ada Kha Naioal College of Busiess Admiisraio ad Ecoomics Gulberg-III, Lahore, Pakisa adakhaariq@gmailcom G M Habibullah Naioal College

More information

Minimizing the Total Late Work on an Unbounded Batch Machine

Minimizing the Total Late Work on an Unbounded Batch Machine The 7h Ieraioal Symposium o Operaios Research ad Is Applicaios (ISORA 08) Lijiag, Chia, Ocober 31 Novemver 3, 2008 Copyrigh 2008 ORSC & APORC, pp. 74 81 Miimizig he Toal Lae Work o a Ubouded Bach Machie

More information

Optimally Sparse SVMs

Optimally Sparse SVMs A. Proof of Lemma 3. We here prove a lower boud o the umber of support vectors to achieve geeralizatio bouds of the form which we cosider. Importatly, this result holds ot oly for liear classifiers, but

More information

Regret of Multi-Channel Bandit Game in Cognitive Radio Networks

Regret of Multi-Channel Bandit Game in Cognitive Radio Networks MAEC Web of Cofereces 56, DOI: 10.1051/ maeccof/20165605002 Regre of Muli-Chael Badi Game i Cogiive Radio Neworks Ju Ma ad Yoghog Zhag School of Elecroic Egieerig, Uiversiy of Elecroic Sciece ad echology

More information

CS623: Introduction to Computing with Neural Nets (lecture-10) Pushpak Bhattacharyya Computer Science and Engineering Department IIT Bombay

CS623: Introduction to Computing with Neural Nets (lecture-10) Pushpak Bhattacharyya Computer Science and Engineering Department IIT Bombay CS6: Iroducio o Compuig ih Neural Nes lecure- Pushpak Bhaacharyya Compuer Sciece ad Egieerig Deparme IIT Bombay Tilig Algorihm repea A kid of divide ad coquer sraegy Give he classes i he daa, ru he percepro

More information

BRIDGE ESTIMATOR AS AN ALTERNATIVE TO DICKEY- PANTULA UNIT ROOT TEST

BRIDGE ESTIMATOR AS AN ALTERNATIVE TO DICKEY- PANTULA UNIT ROOT TEST The 0 h Ieraioal Days of Saisics ad Ecoomics Prague Sepember 8-0 06 BRIDGE ESTIMATOR AS AN ALTERNATIVE TO DICKEY- PANTULA UNIT ROOT TEST Hüseyi Güler Yeliz Yalҫi Çiğdem Koşar Absrac Ecoomic series may

More information

A Generalized Cost Malmquist Index to the Productivities of Units with Negative Data in DEA

A Generalized Cost Malmquist Index to the Productivities of Units with Negative Data in DEA Proceedigs of he 202 Ieraioal Coferece o Idusrial Egieerig ad Operaios Maageme Isabul, urey, July 3 6, 202 A eeralized Cos Malmquis Ide o he Produciviies of Uis wih Negaive Daa i DEA Shabam Razavya Deparme

More information

Problems and Solutions for Section 3.2 (3.15 through 3.25)

Problems and Solutions for Section 3.2 (3.15 through 3.25) 3-7 Problems ad Soluios for Secio 3 35 hrough 35 35 Calculae he respose of a overdamped sigle-degree-of-freedom sysem o a arbirary o-periodic exciaio Soluio: From Equaio 3: x = # F! h "! d! For a overdamped

More information