Stochastic Network Optimization with Non-Convex Utilities and Costs

Size: px
Start display at page:

Download "Stochastic Network Optimization with Non-Convex Utilities and Costs"

Transcription

1 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB Sochasic Nework Opiizaion wih Non-Convex Uiliies and Coss Michael J. Neely Absrac This work considers non-convex opiizaion of ie averages of nework aribues in a general sochasic nework. This includes axiizing a non-concave uiliy funcion of he ie average hroughpu vecor in a ie-varying wireless syse, subjec o nework sabiliy and o an addiional collecion of ie average penaly consrains. We develop a siple algorih ha ees all desired sabiliy and penaly consrains, and, subjec o a convergence assupion, yields a ie average vecor ha is a local opiu of he desired uiliy funcion. We also consider algorihs ha yield local near opial soluions, where he disance o a local opiu can be ade as sall as desired wih a corresponding radeoff in average delay. Our soluion uses Lyapunov opiizaion wih a cobinaion of sochasic dual and prial-dual echniques. We also discuss he relaive advanages and disadvanages of hese echniques. Index Ters Queueing analysis, opporunisic scheduling, flow conrol, wireless neworks I. INTRODUCTION We consider a queueing nework ha operaes in discree ie {0, 1, 2,...}. Le Q() = (Q 1 (),..., Q K ()) be he vecor of queue backlogs on slo. The arrival and service of he queues are deerined by a rando even ω() and a nework conrol acion α(), chosen in reacion o his even fro an absrac se of possible acions A ω() ha possibly depends on ω(). The α() and ω() values also affec wo aribue vecors x() = (x 1 (),..., x M ()), y() = (y 0 (), y 1 (),..., y L ()) according o general (possibly non-convex, disconinuous) funcions ˆx ( ), ŷ l ( ) for {1,..., M} and l {0, 1,..., L}: x () = ˆx (α(), ω()), y l () = ŷ l (α(), ω()) Define x = (x 1,..., x M ) as he liiing ie average expecaion of he aribue vecor x() under a paricular policy (eporarily assued o exis): x= 1 1 li E {x(τ)} Define ie average expecaions y l siilarly. The goal is o deerine an algorih for choosing acions α() A ω() Michael J. Neely is wih he Elecrical Engineering deparen a he Universiy of Souhern California, Los Angeles, CA. (web: hp://wwwrcf.usc.edu/ jneely). This aerial is suppored in par by one or ore of he following: he DARPA IT-MANET progra gran W911NF , he NSF Career gran CCF , and coninuing hrough paricipaion in he Nework Science Collaboraive Technology Alliance sponsored by he U.S. Ary Research Laboraory. Fig. 1. Uiliy(x) Aribue x (such as hroughpu) An exaple non-concave uiliy for hroughpu axiizaion. over ie o solve he following general sochasic nework opiizaion proble: Miniize: y 0 f(x) (1) Subjec o: y l g l (x) 0 l {1,..., L} (2) x X (3) All queues Q k () are ean rae sable (4) α() A ω() (5) where g l (x) are coninuous and convex funcions of x R M for each l {1,..., L}, f(x) is a coninuous bu possibly non-convex funcion of x R M, and X is a convex subse of R M. The definiion of ean rae sable will be given in Secion II. This non-convex proble has applicaions in nework uiliy axiizaion and in oher areas of nework science. For exaple, suppose we desire o axiize a non-concave uiliy funcion of hroughpu in a wireless nework. An exaple sigoidal-ype uiliy funcion is shown in Fig. 1, which has a fla area near zero, illusraing ha we earn sall uiliy unless hroughpu crosses a hreshold. Consrained opiizaion of a su of such uiliy funcions is a difficul proble ha can be shown, in soe cases, o be a leas as difficul as cobinaorial bin-packing. Hence, we do no aep o find a global soluion. Such non-convex probles are sudied for non-sochasic (saic) neworks in [1], where difficulies are discussed and a heurisic algorih is derived ha has opialiy properies in a lii of a large nuber of users. Furher reaen of non-convex opiizaion in a saic conex, including efficien heurisics, are provided in [2]. In his paper, we consider non-convex opiizaion for sochasic neworks via he general proble (1)-(5). We show ha, under soe convergence assupions, we can find a local opiu. In he special case when f(x) is convex, we provide a sronger resul ha shows he global soluion can

2 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB be approached o wihin any desired accuracy. Our echniques exend our prior drif-plus-penaly or dual-based approach o sochasic nework opiizaion in [3], which uses Lyapunov opiizaion, virual queues, and auxiliary variables. These echniques alone would allow one o iniize (1) in he case f(x) is convex, subjec o consrains (2), (4), (5). We inroduce an absrac se consrain (3), which we believe is new even in he convex case (we recenly also used his in a universal scheduling conex in [4]). However, if hese dualbased echniques were eployed for a non-convex funcion f(x), he algorih would approach a global opiu of he ie average of f(x()), which is no necessarily even a local opiu of f(x). In his paper, we achieve local opialiy by cobining hese echniques wih a prial-dual based approach siilar o ha of [5][6][7]. I is known ha, in he convex case, he dual approach is robus o syse changes [3][8][4] and provides direc bounds on uiliy, delay, and convergence ie [3][9][10][11]. Unforunaely, such properies are less clear wih he prialdual approach, paricularly for he non-convex case, and he proof we provide in his paper for he non-convex case does no specify he ie required for convergence. This illusraes soe of he challenges involved wih non-convex opiizaion, even if we only wan o find a local opiu. We parially address his issue in he special case of axiizing nonconvex uiliy funcions ha are enrywise non-decreasing. In his case, we briefly describe an alernaive 3-phase algorih ha explores he aribue space and ies all sochasic decisions direcly o he resul of a saic non-convex opiizaion. The advanage here is ha he sochasic phases are convex and have srong convergence guaranees, while he non-convex opiizaion is isolaed o a saic proble for which any saic solver can be eployed. The ehod of Lyapunov drif o solve queue sabiliy probles (i.e., probles ha involve only he consrain (4)) was pioneered by Tassiulas and Ephreides in [12][13], where i was shown ha greedy acions ha iniize a drif expression () every slo guaranee sabiliy whenever possible. In [9][3][10][11], we exended his resul o provide join sabiliy and perforance opiizaion (or penaly iniizaion ), using a greedy acion o iniize a drif-plus-penaly eric () V penaly() every slo, where V is a non-negaive paraeer ha affecs a perforance-delay radeoff. Choosing V = 0 reduces o he original Tassiulas-Ephreides approach, while increasing V ainains nework sabiliy while pushing he ie average penaly of he proble o wihin O(1/V ) of opial, wih a radeoff in average queue backlog ha is O(V ). This ehod can be applied o uli-hop neworks wih general consrains, and is known o be relaed o dual subgradien algorihs when applied o saic convex progras [3]. A relaed ehod, also of he dual ype, is presened for a wireless downlink in [14], alhough he fluid odel arguens here do no prove he [O(1/V ), O(V )] perforance-delay properies. A differen prial-dual approach is considered for he special case of uiliy-opial opporunisic scheduling in a wireless downlink in [5][6]. This work assues all users always have daa o send, and hence does no consider randoly arriving raffic or queue sabiliy issues. I also uses an infinie horizon ie-average o infor scheduling decisions, which, unlike he dual ehod, akes i difficul o apply o syses wih paraeers ha igh change. Join queue sabiliy and perforance opiizaion is reaed for uli-hop neworks wih his prial-dual approach in [7]. This work proves ha he uiliy of a relaed fluid nework is opial, and conjecures ha he uiliy of he acual nework will also be close o opial. Specifically, he infinie horizon ie average of [5][6] is replaced by an exponenially weighed average in [7], and i is conjecured ha uiliy approaches opialiy when he exponenial decay paraeer is scaled (see Secion 4.9 in [7]). In he nex secion we describe he syse odel and he firs approach o he non-convex sochasic opiizaion. Secion III analyzes he perforance of he algorih using he drif-plus-penaly echnique wih a penaly ha involves a parial derivaive (incorporaing he prial-dual approach). We also copare agains he perforance of a pure dual-based algorih in he special case of convex probles. Secion IV develops our second approach o he non-convex proble, for he case of opiizing non-concave uiliy funcions ha are enrywise non-decreasing. A. Model Assupions II. SOLVING THE PROBLEM The queueing dynaics for k {1,..., K} are given by: Q k ( 1) = ax[q k () b k (), 0] a k () (6) where b() = (b 1 (),..., b K ()), a() = (a 1 (),..., a K ()) are non-negaive service and arrival vecors, deerined as arbirary funcions of α() and ω(): a k () = â k (α(), ω()), b k () = ˆb k (α(), ω()) The srucure of a k () being ouside he ax[, 0] operaor in (6) also allows reaen of uli-hop neworks, where a k () can be a su of exogenous and endogenous arrivals [3][15]. We assue he rando even process ω() is i.i.d. over slos, wih soe (possibly unknown) disribuion. Second oens of he arrival, service, and aribue coponens are assued o be finie under all possible conrol acions. Specifically, for all (possibly randoized) choices of α A ω(), ade based on knowledge of ω(), here is a finie consan σ 2 > 0 such ha: E { â k (α(), ω()) 2} σ 2, E {ˆbk (α(), ω()) 2} σ 2 E {ˆx (α(), ω()) 2} σ 2, E { ŷ l (α(), ω()) 2} σ 2 These second oen bounds also iply he firs oens of nework aribues lie wihin a bounded region, even if he acual realizaions of hese aribues are unbounded. The se X in (3) is assued o be a copac and convex subse of R M. We assue he funcions g l (x) are convex (no necessarily differeniable) for all l {1,..., L}, and ha hey are Lipschiz coninuous, in he sense ha for any x = (x 1,..., x M ), γ = (γ 1,..., γ M ) we have: g l (x) g l (γ) ν l, x γ (7)

3 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB where ν l, are finie non-negaive consans. The funcion f(x) is possibly non-convex, bu is assued o be bounded, coninuously differeniable, and o have bounded parial derivaives, so ha: f(x) f(γ) β x γ (8) for soe finie consans β 0. Le f in, f ax be finie consans such ha for all we have f in f(x()) f ax. B. Queue Sabiliy We firs define wo ypes of queue sabiliy. Le Q() be a discree ie process defined over slos {0, 1, 2,...}, assued o ake (possibly negaive) real values. Negaive queue processes will be useful o ee cerain consrains. Definiion 1: Q() is ean rae sable if: E { Q() } li = 0 Definiion 2: Q() is srongly sable if: 1 1 li sup E { Q(τ) } < Under ild echnical assupions i can be shown ha srong sabiliy iplies ean rae sabiliy, and ha srong sabiliy is equivalen o he exisence of a saionary disribuion in cases when Q() is defined over a counably infinie sae Markov chain. C. Auxiliary Variables To enable he absrac se consrain (3) o be e and o allow opiizaion over he convex bu possibly non-linear funcions g l (x), we inroduce a vecor of auxiliary variables γ() = (γ 1 (),..., γ M ()). This is an addiional collecion of decision variables ha us be chosen every slo, subjec only o he consrain: γ() X We ransfor he proble (1)-(4) ino: Miniize: y 0 f(x) (9) Subjec o: y l g l (γ) 0 l {1,..., L} (10) where noaion g l (γ) is defined: γ = x {1,..., M} (11) All queues Q k () are ean rae sable (12) α() A ω(), γ() X (13) g l (γ) = 1 1 li E {g l (γ(τ))} The oivaion for his new proble is he observaion ha, assuing we (eporarily) resric o algorihs ha have well defined liis, he new proble is equivalen o (1)-(5). Indeed, suppose ha α () is a policy ha ees all consrains of he proble (1)-(5), and ha yields a cos eric in (1) of y 0 f(x ), where y l and x represen ie averages associaed wih his policy. Then his policy also saisfies he consrains of proble (9)-(13), and yields he sae cos value in (9), if we define consan auxiliary variable decisions γ () = x for all (hese saisfy he required consrains (13) because of (3)). I follows ha he infiu cos eric of he proble (1)-(5) over all policies ha ee he consrains is greaer han or equal o he infiu associaed wih proble (9)-(13). Conversely, by Jensen s inequaliy, i can be shown ha any algorih ha solves he consrains in he proble (9)-(13) us also saisfy he consrains of he proble (1)-(5), and produces a cos (1) ha is he sae. For sipliciy of exposiion, we have chosen o wrie he sochasic nework opiizaion probles (1)-(5) and (9)-(13) wih noaion ha iplicily assues liis are well defined. However, a soluion need no have well defined liis, and we can ore precisely define he proble using a li sup, as done in fuure secions (for exaple, he consrain (2) is wrien ore generally as (16)). D. Virual Queues To ensure he inequaliy consrains (10) and equaliy consrains (11) are e, we inroduce virual queues Z l () and H () for l {1,..., L}, {1,..., M}, wih updae equaion: Z l ( 1) = ax[z l () y l () g l (γ()), 0] (14) H ( 1) = H () γ () x () (15) The Z l () queues are always non-negaive. The H () queues have a differen srucure because hey enforce an equaliy consrain, and can possibly be negaive. Lea 1: If E {Z l (0)} < and E { H (0) } < for all l {1,..., L} and {1,..., M}, and if Z l () and H () are ean rae sable, hen: li sup [y l () g l (x())] 0 l {1,..., L} (16) li γ () x () = 0 {1,..., M} (17) li dis (x(), X ) = 0 (18) where for each > 0, x() and γ() are defined: x() = 1 1 E {x(τ)}, γ() = 1 1 E {γ(τ)} (19) where y l () is defined siilarly, and where dis (x(), X ) is he disance beween he vecor x() and he se X, being zero if and only if x() is in he (closed) se X. The proxiiy of he ie averages o heir required values for all ie > 0 is provided as a funcion of he virual queue sizes in (20) and (21) of he proof. Proof: (Lea 1) Fix > 0. Noe by suing (15) over τ {0,..., 1} we have: 1 1 H () H (0) = γ (τ) x (τ) Dividing boh sides by and aking expecaions yields: E {H ()} E {H (0)} = γ () x () (20)

4 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB The above holds for all > 0. Assuing ha H () is ean rae sable, we can ake a lii of (20) as o yield (17). 1 Now noe ha γ() X for all, and hence (because X is convex), γ() X for all > 0. This ogeher wih (17) iplies (18). Finally, fro (14) we have for all slos τ: Z l (τ 1) Z l (τ) y l (τ) g l (γ(τ)) Fixing > 0 and suing over τ {0,..., 1} yields: 1 1 Z l () Z l (0) y l (τ) g l (γ(τ)) Taking expecaions, dividing by and using Jensen s inequaliy yields: E {Z l ()} E {Z l(0)} y l () g l (γ()) y l () g l (x()) ν l, x () γ () (21) where he las inequaliy is due o (7). Assuing Z l () is ean rae sable, we can ake he li sup of boh sides of (21) and use (17) o yield (16). E. Lyapunov Opiizaion Define Θ() =[Q(); Z(); H(); x av ()] as he collecive queue vecor ogeher wih he curren running ie-average of aribue vecor x(τ): x av () = 1 1 x(τ) (22) Define he following non-negaive funcion, called a Lyapunov funcion: L(Θ()) = 1 Q k () 2 1 Z l () 2 1 H () 2 (23) Define he condiional Lyapunov drif (Θ()) by: 2 (Θ()) =E {L(Θ( 1)) L(Θ()) Θ()} (24) Lea 2: Under any algorih for choosing α() A ω() and γ() X for all, we have for all and all possible Θ(): where (Θ()) RHS(α(), γ(), Θ()) (25) RHS(α(), γ(), Θ()) = { B Q k ()E â k (α(), ω()) ˆb } k (α(), ω()) Θ() Z l ()E {ŷ l (α(), ω()) g l (γ()) Θ()} H ()E {γ () ˆx (α(), ω()) Θ()} 1 Noe ha if E { H () } / 0 hen E {H ()} / 0. 2 Sricly speaking, correc noaion should be (Θ(), ), as he drif ay be non-saionary, alhough we use he sipler noaion (Θ()) as a foral represenaion of he righ hand side of (24). where B is a finie consan ha saisfies he following for all and all possible Θ(): B E { b k () 2 a k () 2 Θ() } E { (y l () g l (γ())) 2 Θ() } E { (γ () x ()) 2 Θ() } (26) Such a finie consan B exiss by he syse boundedness assupions. Proof: The proof follows by squaring he equaions (6), (14), (15) and is oied for breviy (see [3]). Our algorih seeks o ake conrol acions α() A ω(), γ() X every slo in reacion o he observed ω() and o he curren queue saes and x av () values in Θ(), o iniize: RHS(α(), γ(), Θ()) V E {ŷ 0 (α(), ω()) Θ()} V f(x av ()) E {ˆx (α(), ω()) Θ()} (27) where x av () is defined in (22), and where we define x av (0) as an iniial saple ˆx(α( 1), ω( 1)) under any decision α( 1) A ω( 1) aken a soe preliinary slo. The paraeer V is chosen as a non-negaive consan ha will affec proxiiy o he opial soluion, wih a corresponding radeoff in average queue congesion, as in [3]. In he special case when here is no non-convex funcion, so ha f(x) = 0, we do no require he ie averages x av (). In his case he above algorih reduces o one ha is siilar o our work in [3], wih he excepion ha we are considering an absrac se consrain X and he convex funcions g l (x) are no necessarily onoonic. In he case when f(x) 0, our algorih includes a derivaive uliplied by a running ie average, as in he prial-dual algorihs ha do no consider queue sabiliy in [5][6]. This prial-dual coponen is also siilar o fluid lii works in [7][16], which do consider queue sabiliy, wih he excepion ha our algorih includes auxiliary variables and does no use an exponenial weighed average. 3 An advanage of our approach is ha i is copaible wih non-convex objecives, ensures all convex consrains are saisfied as, and provides explici bounds on consrain violaion for all > 0. The algorih is specified as follows: Every slo, observe ω(), Θ() (where Θ() includes all curren queue saes and he curren x av ()), and perfor he following: (Auxiliary Variables) Choose γ() X o iniize: Z l ()g l (γ()) H ()γ () 3 Using an exponenial weighed average would be advanagous for robusness o syse changes, bu i is no clear if he algorih would converge properly in his case.

5 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB (Choosing α()) Choose α() A ω() o iniize: Q k ()[â k (α(), ω()) ˆb k (α(), ω())] Z l ()ŷ l (α(), ω()) V ŷ 0 (α(), ω()) H ()ˆx (α(), ω()) V f(x av ()) ˆx (α(), ω()) (Queue Updaes) Updae he virual queues Z l () and H () according o (14), (15), and updae he acual queues Q k () via (6). Also updae x av () by (22). III. PERFORMANCE ANALYSIS A. Opialiy via ω-only Policies We say ha he proble (1)-(5) is feasible if here is an algorih for choosing α() A ω() over ie so ha all consrains (2)-(5) are saisfied. We assue hroughou ha he proble is feasible. Define y op 0 f op as he infiu value of (1) over all feasible policies. I urns ou ha his infiu can be characerized by ω-only policies. Specifically, we say ha a policy α () is ω-only if for all i chooses α () A ω() as a saionary and randoized funcion of he observed ω(), i.i.d. over all slos for which he sae ω() value is observed. By he law of large nubers, all ie averages resuling fro an ω-only policy are well defined and are equal o heir expecaions over one slo. In paricular, we have for all : y l = E {ŷ l (α (), ω())}, x = E {ˆx(α (), ω())} For a given ɛ > 0, define he se Y ɛ as he se of all vecors (y, x) = (y 0, y 1,..., y L, x 1,..., x M ) ha can be achieved as ie averages by soe ω-only policy α () A ω() ha saisfies: y l g l (x) ɛ l {1,..., L} dis(x, X ) ɛ { E â k (α (), ω()) ˆb } k (α (), ω()) ɛ The se Y ɛ is bounded by he boundedness assupions, and can be shown o be convex and o saisfy Y ɛ1 Y ɛ2 whenever ɛ 1 ɛ 2. Define Cl(Y ɛ ) as he closure of Y ɛ, and define Y as he lii of Cl(Y ɛ ) as ɛ 0: Y = i=1 Cl(Y 1/i ) I can be shown ha Y is non-epy whenever he proble (1)-(5) is feasible, and is bounded, closed, and convex. Theore 1: (Opialiy over he se Y) The se Y represens he se of all vecors (y, x) ha can be achieved as a lii poin of a rajecory (y(), x()) generaed by a feasible policy ipleened over {0, 1, 2,...}. Furher, he infiu value y op 0 f op is he soluion o he following proble: Miniize: y 0 f(x) Subjec o: (y 0, y 1,..., y L, x 1,..., x M ) Y Proof: Oied for breviy (see, for exaple, [9][10]). Theore 2: For any vecor (y, x ) Y and for any ɛ > 0, here is an ω-only policy α () such ha: E {ŷ l (α (), ω())} y l ɛ l {0, 1,..., L} (28) E {ˆx (α (), ω())} x ɛ {1,..., M} (29) E {ŷ l (α (), ω())} g l (E {ˆx(α (), ω())}) ɛ (30) dis (E {ˆx(α (), ω())}, X ) ɛ (31) } E {â k (α (), ω())} E {ˆbk (α (), ω()) ɛ k (32) Proof: The proof follows by observing ha any poin in Y can be achieved arbirarily closely by an ω-only policy ha saisfies he required consrains o wihin ɛ. B. The Perforance Theore For a given consan C 0, we say ha an algorih is a C-approxiaion if every slo, given he exising Θ() on slo, i chooses α() A ω(), γ() X o achieve a value in (27) ha is wihin C of he infiu over all possible decisions given Θ(). A 0-approxiaion achieves he exac infiu and is given by he algorih in Secion II. Using C > 0 allows for approxiae ipleenaions. Theore 3: Suppose V 0, ω() is i.i.d. over slos, he proble (1)-(5) is feasible, and any C-approxiae algorih is used every slo. For sipliciy, assue ha all queues are iniially zero, so ha Θ(0) = 0. Then: (a) All queues are ean rae sable, and hence all required consrains are saisfied. In paricular, he equaliies and inequaliies (16)-(18) are ensured. Bounds on expeced queue sizes (and hence, consrain violaions) for all > 0 are provided in inequaliy (35) of he proof. (b) For all > 0 and for any (y, x ) Y (including he one ha opiizes he infiu cos), we have: y 0 () 1 1 x (τ) f(x av (τ)) E y0 1 1 x E f(x av (τ)) B C V where B is defined by (26). (c) If all ie averages converge, so ha here are consan vecors x, y such ha x av () x wih probabiliy 1, x() x, and y () y, hen he achieved liiing poin is a near local opiu, in he sense ha for any (y, x ) Y (including any local or global opiu): x f(x) y 0 y x 0 f(x) B C V The resul of par (c) can be viewed as a near-local opiu resul because of he fudge-facor (B C)/V, which can be ade arbirarily sall as he V paraeer is increased. Indeed, he above resul shows ha he achieved cos, as easured by a linearized version of he f(x) funcion abou our achieved x vecor, is no ore han (B C)/V above he cos associaed wih any oher (y, x ) Y, when easured agains he sae linearized funcion. In paricular, suppose we consider oving fro our achieved (y, x) vecor owards any oher (y, x )

6 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB Y by aking a sall sep ɛ(y y, x x) in his direcion (for soe sall value ɛ > 0). If ɛ is sall, hen he cos differenial cos (ɛ) incurred saisfies: cos (ɛ) ɛ (y 0 y 0 ) (x x ) f(x) (B C) V where he inequaliy follows by par (c) of he above heore. I follows ha such a sep canno decrease cos by ore han approxiaely ɛ(bc)/v (where he approxiaion is due o he linearizaion), which can be ade arbirarily sall wih a suiably large choice of V. Hence, his is a near local opiu. The value of V affecs he coefficien in he decay of E { H () } /, E {Z l ()} /, E {Q k ()} / derived in (35) of he proof, which also affecs he decay in consrain violaion via (20), (21). I can be shown ha if a ild Slaer-ype condiion holds for he consrains (2)-(5) of he proble, hen all queues are srongly sable and have average backlog ha is O(V ) (see [3][10][9] for siilar resuls). Proof: (Theore 3 par (a)) By siply adding he sae hing o boh sides of (25) in Lea 2, we have: (Θ()) V E {y 0 () Θ()} V f(x av ()) E {x () Θ()} B Q k ()E {a k () b k () Θ()} Z l ()E {y l () g l (γ()) Θ()} H ()E {γ () x () Θ()} V E {y 0 () Θ()} V f(x av ()) E {x () Θ()} Because, given he exising Θ() on slo, our C- approxiaion iniizes he righ-hand-side of he above expression o wihin C of he infiu over all policies for choosing α() A ω(), γ() X, we have: (Θ()) V E {y 0 () Θ()} V f(x av ()) E {x () Θ()} B C Q k ()E {a k() b k() Θ()} Z l ()E {yl () g l (γ ()) Θ()} H ()E {γ() x () Θ()} V E {y 0() Θ()} V f(x av ()) E {x () Θ()} where α () A ω(), γ () X are any alernaive conrol acions, and for all k {1,..., K}, l {0, 1,..., L}: a k() =â k (α (), ω()), b k() =ˆb k (α (), ω()) y l () =ŷ l (α (), ω()) For any (y, x ) Y and any ɛ > 0, Theore 2 ensures here is an ω-only algorih α () ha saisfies (28)-(32). For sipliciy, assue ha (28)-(32) hold for ɛ = 0. 4 Now plug his algorih ino he righ hand side of he above drif expression, and noe ha, because i is ω-only, i akes decisions independen of backlog (so ha he condiional expecaion given Θ() is he sae as he uncondiional expecaion). Furher, plug decisions γ () = E {ˆx(α (), ω())} = x (his saisfies γ() X as required because of (31) wih ɛ = 0 and noing ha X is closed). This siplifies he righ-hand-side of he above drif bound o: (Θ()) V E {y 0 () Θ()} V f(x av ()) E {x () Θ()} B C V y 0 V x f(x av ()) (33) Because of he boundedness assupions on he se X, he firs and second oens of he aribue vecors, and he funcion f(x) and is derivaives, he above drif bound siplifies o: (Θ()) D (34) where D is any finie consan ha saisfies for all, Θ(): D B C V [y0 E {y 0 () Θ()}] V β [x E {x () Θ()}] Taking an expecaion of boh sides and using he law of ieraed expecaions yields: E {L(Θ( 1))} E {L(Θ())} D The above holds for all. Suing over τ {0,..., 1} and using he fac ha L(Θ(0)) = 0 yields: E {L(Θ())} D By definiion of L(Θ()) in (23) we have: E { Q k () 2}, E { Z l () 2}, E { H () 2} 2D Because E { X } 2 E { X 2} for any rando variable X, we have: E {Q k ()} 2, E {Z l ()} 2, E { H () } 2 2D Dividing by 2 and aking square roos shows ha for all > 0: E {Q k ()}, E {Z l()}, E { H () } 2D (35) Taking a lii as proves ean rae sabiliy of all queues. 4 The sae resul (33) can be derived wihou his assupion by aking a lii as ɛ 0.

7 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB Proof: (Theore 3 pars (b) and (c)) We sar wih he drif bound in (33) derived in he proof of par (a), which holds for all 0. Consider his inequaliy for soe slo τ. Taking expecaions of boh sides and using he law of ieraed expecaions and he definiion of (Θ(τ)) in (24), we have: E {L(Θ(τ 1))} E {L(Θ(τ))} V E {y 0 (τ)} V x (τ) f(x av (τ)) E V x B C V y0 E f(x av (τ)) (36) Fix any slo > 0. Suing he above over τ {0,..., 1} and using elescoping sus yields: 1 E {L(Θ())} E {L(Θ(0))} V E {y 0 (τ)} 1 V x (τ) f(x av (τ)) E 1 (B C V y0) V x E f(x av (τ)) Dividing by V and using he fac ha L(Θ(0)) = 0 and L(Θ()) 0 yields: y 0 () 1 1 x (τ) f(x av (τ)) E B C y0 1 1 x E f(x av (τ)) V This proves par (b). Par (c) is an iediae consequence, noing boundedness and coninuiy of derivaives of f(x). C. Variable V for Local Opialiy As in [17], we can use an algorih ha gradually increases he value of V while ainaining ean rae sabiliy. This eliinaes he fudge facor (BC)/V in Theore 3 par (c). A disadvanage is ha i precludes srong sabiliy of queues even when a Slaer condiion is saisfied. Theore 4: Suppose ha we use V () insead of V in he drif expression (27), and we use any C-approxiaion, i.e., any decisions ha coe wihin C of iniizing (27) under his variable V () seing on all slos. Suppose ω() is i.i.d. over slos, he proble (1)-(5) is feasible, and all queues are iniially epy. Furher assue ha we use he following increasing V () funcion: V () =V 0 (1 ) d where V 0 > 0 and d saisfies 0 < d < 1. Then: (a) All queues are ean rae sable as before, and hence he equaliies and inequaliies in (16)-(18) are saisfied. (b) If all ie averages converge, so ha here are vecors x, y such ha x av () x wih probabiliy 1, x() x, and y () y, hen he achieved poin is a local opiu, in he sense ha for any (y, x ) Y (including any local or global opiu): y 0 x f(x) y 0 x f(x) Tha is, he achieved vecor (y, x ) globally opiizes he linearized cos funcion. In paricular, if we sar a our achieved (y, x) and ove in he direcion of any oher (y, x ) Y by aking a sall sep ɛ(y y, x x), hen: cos (ɛ) li ɛ 0 ɛ = (y 0 y 0 ) (x x ) f(x) 0 Thus, such a sall sep canno iprove cos. While he heore holds for any V 0 > 0 and 0 < d < 1, hese values affec he anner in which he uiliy and virual queue values converge, as shown in he proof. Proof: (Theore 4) The proof of par (a) is siilar o [17] and is oied for breviy. Here we prove par (b). Following he proof of Theore 3 alos exacly, we find ha (36) ranslaes o he following for any slo τ 0: E {L(Θ(τ 1))} E {L(Θ(τ))} V (τ)e {y 0 (τ)} V (τ)x (τ) f(x av (τ)) E V (τ)x B C V (τ)y0 E f(x av (τ)) Dividing everyhing by V (τ) yields: E {L(Θ(τ 1))} E {L(Θ(τ))} E {y 0 (τ)} V (τ) V (τ) x (τ) f(x av (τ)) E B C x V (τ) y 0 E f(x av (τ)) Fix > 0, and su he above over τ {0,..., 1}. Collecing ers yields: E {L(Θ())} V ( 1) 1 1 τ=1 1 E {y 0 (τ)} 1 [ E {L(Θ(τ))} B C V (τ) y 0 1 V (τ 1) 1 V (τ) x (τ) f(x av (τ)) E 1 x E f(x av (τ)) Because V (τ) is non-decreasing, we have for all τ: 1 V (τ 1) 1 V (τ) 0 ]

8 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB Using his, dividing everyhing by, and using non-negaiviy of L( ) yields for all > 0: y 0 () 1 1 x (τ) f(x av (τ)) E 1 1 (B C) y0 1 1 x E f(x av (τ)) V (τ) However, we have: 1 1 (B C) (B C) 1 1 = ( 1) d = O( d ) 0 V (τ) V 0 Thus, aking liis of he above drif bound and using he assupion ha ie averages converge yields he resul. D. Convex Probles and he Pure Dual Approach A significan liiaion of he prial-dual coponen in he above algorihs is ha he running ie average x av () in (22) precludes adapaion when syse condiions change. An addiional liiaion is ha our heore requires vecors x av (), x(), y() o converge, bu lacks a proof of convergence and a easure of how long such convergence would ake. 5 For coparison, here we show ha hese liiaions can be solved by he drif-plus-penaly algorih wihou using parial derivaives, via he pure dual-based approach, when f(x) is convex. These advanages are exploied in an alernaive algorih for he non-convex proble in Secion IV. The analysis in his subsecion is siilar o [3][17], wih he excepion ha we include he absrac se consrain X. Assue here ha he funcion f(x) is coninuous and convex (possibly non-differeniable). Define finie bounds f in and f ax o saisfy for all and all (possibly randoized) decisions α() A ω() : f in f(e {ˆx(α(), ω())}) f ax The following algorih is oivaed by seeking o solve a odified version of he proble (9)-(13), wih he cos eric (9) changed o iniizing y 0 f(γ), which can be shown o sill be equivalen o he original proble (1)-(5) by Jensen s inequaliy and he fac ha f(γ) is convex. Using he sae virual queues and he sae Lyapunov funcion L(Θ()) as before (wih he excepion ha Θ() no longer includes x av () inforaion), we have by adding he sae hing o boh sides of he drif inequaliy of Lea 2: (Θ()) V E {y 0 () f(γ()) Θ()} B Q k ()E {a k () b k () Θ()} Z l ()E {y l () g l (γ()) Θ()} H ()E {γ () x () Θ()} V E {y 0 () f(γ()) Θ()} (37) 5 Under an addiional Slaer condiion, he policy is a Markov chain wih all queues srongly sable, and so convergence would follow if we had a counably infinie sae space. Our policy is now o observe Θ() and ω() for each slo, and o ake acions α() A ω(), γ() X o iniize he righ hand side of he above drif bound. For a given consan C 0, we define a C-approxiaion for his algorih o be one ha, every slo and given Θ(), chooses α() A ω(), γ() X o coe wihin C of he infiu of he righ hand side of (37). A 0-approxiaion would ake decisions every slo as follows: (Auxiliary Variables) Choose γ() X o iniize: V f(γ()) H ()γ () Z l ()g l (γ()) (Choosing α()) Choose α() A ω() o iniize: Q k ()[â k (α(), ω()) ˆb k (α(), ω())] Z l ()ŷ l (α(), ω()) H ()ˆx (α(), ω()) V ŷ 0 (α(), ω()) (Queue Updae) Updae virual queues Z l (), H () by (14), (15), and updae acual queues Q k () by (6). Theore 5: ( Pure Dual Perforance) Suppose proble (1)-(5) is feasible, ω() is i.i.d. over slos, and he funcion f(x) is convex. For sipliciy, assue all iniial queue values are 0. If we use any C-approxiaion on each slo, hen: (a) All queues are ean rae sable, and hence all required consrains are saisfied. In paricular, he equaliies and inequaliies (16)-(18) are ensured. Furher, bounds on he expeced queue size for all > 0 are provided in inequaliy (39) of he proof. (b) For all slos > 0, he achieved ie average expeced cos saisfies: y 0 () f(x()) y op 0 f op B C V β E { H () } where we recall ha x() is defined in (19), and y 0 () is defined siilarly. The er E { H () } / decays o zero a leas as fas as O(1/ ), as shown in (39) of he proof. (c) Suppose we use any C-approxiaion applied o a variable-v version of (37), wih V () = V 0 (1 ) d for any consans V 0 > 0 and 0 < d < 1. Then his ainains ean rae sabiliy of all queues, ensures all consrains (16)-(18), and yields exac cos opialiy: li [y 0() f(x())] = y op 0 f op We noe ha if an addiional Slaer condiion is saisfied, hen under he fixed V algorih all queues can be shown o be srongly sable wih averages O(V ) (see relaed resuls in [3][10][9]). In any flow conrol probles, he queues can be shown o be deerinisically bounded by a consan of size O(V ), even wihou he Slaer condiion [10][4]. In he variable-v scenario, queues will no be srongly sable even if he Slaer condiion is saisfied, and hence he cos of achieving exac opialiy is having infinie average backlog and delay.

9 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB Proof: (Theore 5) Because every slo τ he C- approxiaion coes wihin C of he infiu in he righ hand side of (37), we have: (Θ()) V E {y 0 () f(γ()) Θ()} B C Q k ()E {a k() b k() Θ()} Z l ()E {yl () g l (γ ()) Θ()} H ()E {γ() x () Θ()} V E {y 0() f(γ ()) Θ()} where a k (), b k (), y l (), γ () are associaed wih any alernaive decisions α () A ω(), γ () X. Fix (y op, x op ) Y as he opial soluion in Theore 1, yielding cos f(x op ) = f op. Plug he policy of (28)-(32), and for sipliciy again assue ɛ = 0. Also plug γ () = x op. Because hese decisions are independen of Θ(), we have: (Θ()) V E {y 0 () f(γ()) Θ()} B C V y op 0 V f op (38) Thus, (Θ()) ˆD, where ˆD is defined: ˆD =B C V (y op 0 y in 0 ) V (f op f in ) where we define y0 in as a lower bound on E {y 0 () Θ()} over all possible α(). This has he sae srucure as (34), and hence we obain for all > 0 (copare wih (35)): E {Q k ()}, E {Z l()}, E { H () } 2 ˆD (39) Taking a lii of (39) as shows ha all queues are ean rae sable, proving par (a). To prove (b), ake expecaions of boh sides of (38) and use ieraed expecaions o obain: E {L(Θ( 1))} E {L(Θ())} V E {y 0 () f(γ())} B C V (y op 0 f op ) Fix > 0, su he above over τ {0,..., 1}, and divide by o yield: E {L(Θ())} E {L(Θ(0))} V y 0 () V f(γ()) B C V (y op 0 f op ) where we have used Jensen s inequaliy in he convex funcion f(γ). Noing ha L(Θ(0)) = 0, L(Θ()) 0, and dividing by V, we have: y 0 () f(γ()) B C y op 0 f op (40) V However, noe fro (8) ha: f(x()) f(γ()) β x () γ () f(γ()) β E { H () } / (41) where (41) follows fro (20). Using (41) and (40) proves he resul of par (b). Par (c) is siilar o Theore 4. Reark 1: While Theores 3-5 assue ω() is i.i.d. over slos, his assupion is no crucial, and he sae policies can be shown o yield siilar resuls by using a T -slo Lyapunov drif arguen, under he assupion ha ω() is ergodic wih a decaying eory propery, so ha averages over T slos are close o he seady sae average [18][19][20][3]. Reark 2: We can reove auxiliary variables γ() and se H () = 0 for all if: (i) We use he prial-dual ehod, here is no absrac se consrain (3), and g l (x) = 0 for all l, or (ii) We use he pure dual ehod, here is no absrac se consrain (3), and g l (x) = f(x) = 0 for all l. IV. AN ALTERNATIVE SEARCH METHOD Here we provide an alernaive algorih for he non-convex proble in he special case of uiliy axiizaion wih enrywise non-decreasing uiliy funcions. I involves hree phases: Two phases are convex sochasic nework opiizaions for which we have sronger perforance guaranees as in Secion III-D. All non-convexiy is isolaed o he reaining phase ha involves a deerinisic non-convex proble for which any available solver can be eployed. To begin, le f(x) = φ(x), where φ(x) is a uiliy funcion o be axiized. Suppose ha φ(x) is possibly non-concave, bu is enrywise non-decreasing. For sipliciy, assue ŷ l ( ) = 0 for all l, so ha he se Y consiss of vecors (0, x). Le θ 1,..., θ N be a collecion of N differen search direcion vecors, being non-negaive M-diensional vecors. Assue for sipliciy ha for each θ n, he se Y conains a poin x such ha x = ηθ n for soe non-negaive scalar η (his is rivially rue if Y conains he origin). Define η n as he larges value of η for which his is rue, so ha x n = η n θ n is he larges value of x ha we can push in his direcion. Our new goal is o find a feasible policy ha yields a uiliy φ(x) ha is a local axiu over he se C defined: C =Conv({x 1,..., x N }) where Conv( ) denoes he convex hull. Because Y is convex, any poin in his convex hull is achievable. The value of each x n depends on he value of η n, which is unknown if he sochasics of he nework are unknown. However, suppose we know a bound on η n, so ha 0 η n ηn ax. If his is no a rue bound, hen we siply redefine x n= in[η n, ηn ax ]θ n. Our algorih has 3 phases: Phase 1: Deerine he vecors {x 1,..., x N } by solving N differen sochasic nework opiizaion probles, each wih auxiliary variables η n (): Maxiize: Subjec o: η n x = η n θ n x X g l (x) 0 l {1,..., L} All queues Q k () are ean rae sable α() A ω(), 0 η n () η ax n

10 PROC. INFORMATION THEORY AND APPLICATIONS WORKSHOP (ITA), SAN DIEGO, FEB Each of hese is a siple convex sochasic nework opiizaion proble, and we can run he pure dual approach o provide accurae esiaes η n of he opial η n value for each n. Noe ha his can be done using N differen virual neworks in parallel, reusing he sae ω() evens ha are observed. The equaliy consrains can eiher be reaed as addiional aribues h() = x() η n ()θ n wih an absrac se consrain h {0}, or can siply be reaed using virual queues of he ype (15). Phase 2: Define esiaes x n = η n θ n. Use any nonsochasic opiizer o find a local iniu of he following saic opiizaion: Maxize: Subjec o: φ(x) N n=1 p n x n = x p n 0 n {1,..., N} N n=1 p n = 1 This ype of proble can be solved, for exaple, via ehods in [21], including Newon-ype ehods ha do no resric o sall sep sizes. Phase 3: Given he opial x fro phase 2, run an algorih o solve he following sochasic nework opiizaion proble: Maxiize: Subjec o: η x X x = ηx g l (x) 0 l {1,..., L} All queues Q k () are ean rae sable α() A ω(), 0 η n () 2 This is again a basic convex sochasic nework opiizaion. I seeks o axiize he ie average aribue x in he direcion of he x vecor obained in phase 2 (noicing ha he resuling opiu igh push even pas he se C). Here we lii our search o doubling x, wih he inen of obaining a vecor x wih uiliy φ(x) a leas as good as φ(x ). The x vecor used in his phase igh be periodically changed as a resul of periodically running he oher wo phases in he background. V. CONCLUSIONS We have provided wo approaches o consrained opiizaion of non-convex funcions of ie average aribues in a sochasic nework. The firs approach cobines dual-based and prial-dual operaions. I ensures all desired (convex) consrains are saisfied, and, provided ha he algorih converges, i also provides eiher a local opiu or nearlocal opiu, depending on wheher we use a fixed V or variable V algorih. While he ehod provides guaranees on he expeced queue backlogs a any ie, i does no provide such guaranees for he uiliy, and i is no clear how uch ie is required for convergence o he local in. I also requires an infinie horizon ie average ha is no robus o syse changes. This is in conras o he pure dual approach o convex probles ha provide explici backlog and uiliy bounds for all ie, and which are known o adap o syse changes [4][8]. The second approach uses 3 phases, wo of which are sochasic and convex, while he reaining phase is purely deerinisic bu non-convex. These algorihs significanly exend our abiliy o opiize dynaic neworks. REFERENCES [1] J.W. Lee, R. R. Mazudar, and N. B. Shroff. Non-convex opiizaion and rae conrol for uli-class services in he inerne. IEEE/ACM Trans. on Neworking, vol. 13, no. 4, pp , Aug [2] M. Chiang. Nonconvex opiizaion of counicaion syses. Advances in Mechanics and Maheaics, Special volue on Srang s 70h Birhday, Springer, vol. 3, [3] L. Georgiadis, M. J. Neely, and L. Tassiulas. Resource allocaion and cross-layer conrol in wireless neworks. Foundaions and Trends in Neworking, vol. 1, no. 1, pp , [4] M. J. Neely. Universal scheduling for neworks wih arbirary raffic, channels, and obiliy. ArXiv echnical repor, arxiv: v1, Jan [5] R. Agrawal and V. Subraanian. Opialiy of cerain channel aware scheduling policies. Proc. 40h Annual Alleron Conference on Counicaion, Conrol, and Copuing, Monicello, IL, Oc [6] H. Kushner and P. Whiing. Asypoic properies of proporionalfair sharing algorihs. Proc. of 40h Annual Alleron Conf. on Counicaion, Conrol, and Copuing, [7] A. Solyar. Maxiizing queueing nework uiliy subjec o sabiliy: Greedy prial-dual algorih. Queueing Syses, vol. 50, pp , [8] M. J. Neely and R. Urgaonkar. Cross layer adapive conrol for wireless esh neworks. Ad Hoc Neworks (Elsevier), vol. 5, no. 6, pp , Augus [9] M. J. Neely. Dynaic Power Allocaion and Rouing for Saellie and Wireless Neworks wih Tie Varying Channels. PhD hesis, Massachuses Insiue of Technology, LIDS, [10] M. J. Neely. Energy opial conrol for ie varying wireless neworks. IEEE Transacions on Inforaion Theory, vol. 52, no. 7, pp , July [11] M. J. Neely, E. Modiano, and C. Li. Fairness and opial sochasic conrol for heerogeneous neworks. Proc. IEEE INFOCOM, March [12] L. Tassiulas and A. Ephreides. Sabiliy properies of consrained queueing syses and scheduling policies for axiu hroughpu in ulihop radio neworks. IEEE Transacaions on Auoaic Conrol, vol. 37, no. 12, pp , Dec [13] L. Tassiulas and A. Ephreides. Dynaic server allocaion o parallel queues wih randoly varying conneciviy. IEEE Transacions on Inforaion Theory, vol. 39, pp , March [14] A. Eryilaz and R. Srikan. Fair resource allocaion in wireless neworks using queue-lengh-based scheduling and congesion conrol. Proc. IEEE INFOCOM, March [15] M. J. Neely and R. Urgaonkar. Opporunis, backpressure, and sochasic opiizaion wih he wireless broadcas advanage. Asiloar Conference on Signals, Syses, and Copuers, Pacific Grove, CA, Oc [16] A. Solyar. Greedy prial-dual algorih for dynaic resource allocaion in coplex neworks. Queueing Syses, vol. 54, pp , [17] M. J. Neely. Max weigh learning algorihs wih applicaion o scheduling in unknown environens. arxiv: v1, Feb [18] L. Tassiulas. Scheduling and perforance liis of neworks wih consanly changing opology. IEEE Trans. on Inf. Theory, May [19] M. J. Neely, E. Modiano, and C. E Rohrs. Dynaic power allocaion and rouing for ie varying wireless neworks. IEEE Journal on Seleced Areas in Counicaions, vol. 23, no. 1, pp , January [20] R. Urgaonkar and M. J. Neely. Opporunisic scheduling wih reliabiliy guaranees in cogniive radio neworks. IEEE Transacions on Mobile Copuing, vol. 8, no. 6, pp , June [21] D. P. Bersekas. Nonlinear Prograing. Ahena Scienific, Belon, MA, 1995.

A Primal-Dual Type Algorithm with the O(1/t) Convergence Rate for Large Scale Constrained Convex Programs

A Primal-Dual Type Algorithm with the O(1/t) Convergence Rate for Large Scale Constrained Convex Programs PROC. IEEE CONFERENCE ON DECISION AND CONTROL, 06 A Primal-Dual Type Algorihm wih he O(/) Convergence Rae for Large Scale Consrained Convex Programs Hao Yu and Michael J. Neely Absrac This paper considers

More information

TIME DELAY BASEDUNKNOWN INPUT OBSERVER DESIGN FOR NETWORK CONTROL SYSTEM

TIME DELAY BASEDUNKNOWN INPUT OBSERVER DESIGN FOR NETWORK CONTROL SYSTEM TIME DELAY ASEDUNKNOWN INPUT OSERVER DESIGN FOR NETWORK CONTROL SYSTEM Siddhan Chopra J.S. Laher Elecrical Engineering Deparen NIT Kurukshera (India Elecrical Engineering Deparen NIT Kurukshera (India

More information

1 Widrow-Hoff Algorithm

1 Widrow-Hoff Algorithm COS 511: heoreical Machine Learning Lecurer: Rob Schapire Lecure # 18 Scribe: Shaoqing Yang April 10, 014 1 Widrow-Hoff Algorih Firs le s review he Widrow-Hoff algorih ha was covered fro las lecure: Algorih

More information

Fourier Series & The Fourier Transform. Joseph Fourier, our hero. Lord Kelvin on Fourier s theorem. What do we want from the Fourier Transform?

Fourier Series & The Fourier Transform. Joseph Fourier, our hero. Lord Kelvin on Fourier s theorem. What do we want from the Fourier Transform? ourier Series & The ourier Transfor Wha is he ourier Transfor? Wha do we wan fro he ourier Transfor? We desire a easure of he frequencies presen in a wave. This will lead o a definiion of he er, he specru.

More information

Problem set 2 for the course on. Markov chains and mixing times

Problem set 2 for the course on. Markov chains and mixing times J. Seif T. Hirscher Soluions o Proble se for he course on Markov chains and ixing ies February 7, 04 Exercise 7 (Reversible chains). (i) Assue ha we have a Markov chain wih ransiion arix P, such ha here

More information

Opportunism, Backpressure, and Stochastic Optimization with the Wireless Broadcast Advantage

Opportunism, Backpressure, and Stochastic Optimization with the Wireless Broadcast Advantage ASILOMAR CONFERENCE ON SIGNALS, SYSTEMS, AND COMPUTERS, PACIFIC GROVE, CA, OCTOBER 2008 1 Opporunism, Backpressure, and Sochasic Opimizaion wih he Wireless Broadcas Advanage Michael J. Neely, Rahul Urgaonkar

More information

Introduction to Numerical Analysis. In this lesson you will be taken through a pair of techniques that will be used to solve the equations of.

Introduction to Numerical Analysis. In this lesson you will be taken through a pair of techniques that will be used to solve the equations of. Inroducion o Nuerical Analysis oion In his lesson you will be aen hrough a pair of echniques ha will be used o solve he equaions of and v dx d a F d for siuaions in which F is well nown, and he iniial

More information

T L. t=1. Proof of Lemma 1. Using the marginal cost accounting in Equation(4) and standard arguments. t )+Π RB. t )+K 1(Q RB

T L. t=1. Proof of Lemma 1. Using the marginal cost accounting in Equation(4) and standard arguments. t )+Π RB. t )+K 1(Q RB Elecronic Companion EC.1. Proofs of Technical Lemmas and Theorems LEMMA 1. Le C(RB) be he oal cos incurred by he RB policy. Then we have, T L E[C(RB)] 3 E[Z RB ]. (EC.1) Proof of Lemma 1. Using he marginal

More information

Connectionist Classifier System Based on Accuracy in Autonomous Agent Control

Connectionist Classifier System Based on Accuracy in Autonomous Agent Control Connecionis Classifier Syse Based on Accuracy in Auonoous Agen Conrol A S Vasilyev Decision Suppor Syses Group Riga Technical Universiy /4 Meza sree Riga LV-48 Lavia E-ail: serven@apollolv Absrac In his

More information

Wave Mechanics. January 16, 2017

Wave Mechanics. January 16, 2017 Wave Mechanics January 6, 7 The ie-dependen Schrödinger equaion We have seen how he ie-dependen Schrodinger equaion, Ψ + Ψ i Ψ follows as a non-relaivisic version of he Klein-Gordon equaion. In wave echanics,

More information

THE FINITE HAUSDORFF AND FRACTAL DIMENSIONS OF THE GLOBAL ATTRACTOR FOR A CLASS KIRCHHOFF-TYPE EQUATIONS

THE FINITE HAUSDORFF AND FRACTAL DIMENSIONS OF THE GLOBAL ATTRACTOR FOR A CLASS KIRCHHOFF-TYPE EQUATIONS European Journal of Maheaics and Copuer Science Vol 4 No 7 ISSN 59-995 HE FINIE HAUSDORFF AND FRACAL DIMENSIONS OF HE GLOBAL ARACOR FOR A CLASS KIRCHHOFF-YPE EQUAIONS Guoguang Lin & Xiangshuang Xia Deparen

More information

On the approximation of particular solution of nonhomogeneous linear differential equation with Legendre series

On the approximation of particular solution of nonhomogeneous linear differential equation with Legendre series The Journal of Applied Science Vol. 5 No. : -9 [6] วารสารว ทยาศาสตร ประย กต doi:.446/j.appsci.6.8. ISSN 53-785 Prined in Thailand Research Aricle On he approxiaion of paricular soluion of nonhoogeneous

More information

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation Course Noes for EE7C Spring 018: Convex Opimizaion and Approximaion Insrucor: Moriz Hard Email: hard+ee7c@berkeley.edu Graduae Insrucor: Max Simchowiz Email: msimchow+ee7c@berkeley.edu Ocober 15, 018 3

More information

5. Stochastic processes (1)

5. Stochastic processes (1) Lec05.pp S-38.45 - Inroducion o Teleraffic Theory Spring 2005 Conens Basic conceps Poisson process 2 Sochasic processes () Consider some quaniy in a eleraffic (or any) sysem I ypically evolves in ime randomly

More information

Lecture 20: Riccati Equations and Least Squares Feedback Control

Lecture 20: Riccati Equations and Least Squares Feedback Control 34-5 LINEAR SYSTEMS Lecure : Riccai Equaions and Leas Squares Feedback Conrol 5.6.4 Sae Feedback via Riccai Equaions A recursive approach in generaing he marix-valued funcion W ( ) equaion for i for he

More information

Vehicle Arrival Models : Headway

Vehicle Arrival Models : Headway Chaper 12 Vehicle Arrival Models : Headway 12.1 Inroducion Modelling arrival of vehicle a secion of road is an imporan sep in raffic flow modelling. I has imporan applicaion in raffic flow simulaion where

More information

Opportunistic Scheduling with Reliability Guarantees in Cognitive Radio Networks

Opportunistic Scheduling with Reliability Guarantees in Cognitive Radio Networks PROC. IEEE INFOCO, PHOENIX, AZ, APRIL 008 Opporunisic Scheduling wih Reliabiliy Guaranees in Cogniive Radio Neworks Rahul Urgaonkar, ichael J. Neely Universiy of Souhern California, Los Angeles, CA 90089

More information

Intelligent Packet Dropping for Optimal Energy-Delay Tradeoffs in Wireless Downlinks

Intelligent Packet Dropping for Optimal Energy-Delay Tradeoffs in Wireless Downlinks IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 54, NO. 3, PP. 565-579, MARCH 2009. 1 Inelligen Packe Dropping for Opimal Energy-Delay Tradeoffs in Wireless Downlinks Michael J. Neely Universiy of Souhern

More information

Matrix Versions of Some Refinements of the Arithmetic-Geometric Mean Inequality

Matrix Versions of Some Refinements of the Arithmetic-Geometric Mean Inequality Marix Versions of Some Refinemens of he Arihmeic-Geomeric Mean Inequaliy Bao Qi Feng and Andrew Tonge Absrac. We esablish marix versions of refinemens due o Alzer ], Carwrigh and Field 4], and Mercer 5]

More information

Chapter 9 Sinusoidal Steady State Analysis

Chapter 9 Sinusoidal Steady State Analysis Chaper 9 Sinusoidal Seady Sae Analysis 9.-9. The Sinusoidal Source and Response 9.3 The Phasor 9.4 pedances of Passive Eleens 9.5-9.9 Circui Analysis Techniques in he Frequency Doain 9.0-9. The Transforer

More information

EXERCISES FOR SECTION 1.5

EXERCISES FOR SECTION 1.5 1.5 Exisence and Uniqueness of Soluions 43 20. 1 v c 21. 1 v c 1 2 4 6 8 10 1 2 2 4 6 8 10 Graph of approximae soluion obained using Euler s mehod wih = 0.1. Graph of approximae soluion obained using Euler

More information

Approximate Message Passing with Consistent Parameter Estimation and Applications to Sparse Learning

Approximate Message Passing with Consistent Parameter Estimation and Applications to Sparse Learning APPROXIMATE MESSAGE PASSING WITH CONSISTENT PARAMETER ESTIMATION Approxiae Message Passing wih Consisen Paraeer Esiaion and Applicaions o Sparse Learning Ulugbe S. Kailov, Suden Meber, IEEE, Sundeep Rangan,

More information

Hamilton- J acobi Equation: Explicit Formulas In this lecture we try to apply the method of characteristics to the Hamilton-Jacobi equation: u t

Hamilton- J acobi Equation: Explicit Formulas In this lecture we try to apply the method of characteristics to the Hamilton-Jacobi equation: u t M ah 5 2 7 Fall 2 0 0 9 L ecure 1 0 O c. 7, 2 0 0 9 Hamilon- J acobi Equaion: Explici Formulas In his lecure we ry o apply he mehod of characerisics o he Hamilon-Jacobi equaion: u + H D u, x = 0 in R n

More information

arxiv: v1 [math.fa] 12 Jul 2012

arxiv: v1 [math.fa] 12 Jul 2012 AN EXTENSION OF THE LÖWNER HEINZ INEQUALITY MOHAMMAD SAL MOSLEHIAN AND HAMED NAJAFI arxiv:27.2864v [ah.fa] 2 Jul 22 Absrac. We exend he celebraed Löwner Heinz inequaliy by showing ha if A, B are Hilber

More information

An introduction to the theory of SDDP algorithm

An introduction to the theory of SDDP algorithm An inroducion o he heory of SDDP algorihm V. Leclère (ENPC) Augus 1, 2014 V. Leclère Inroducion o SDDP Augus 1, 2014 1 / 21 Inroducion Large scale sochasic problem are hard o solve. Two ways of aacking

More information

Chapter 2. First Order Scalar Equations

Chapter 2. First Order Scalar Equations Chaper. Firs Order Scalar Equaions We sar our sudy of differenial equaions in he same way he pioneers in his field did. We show paricular echniques o solve paricular ypes of firs order differenial equaions.

More information

On Boundedness of Q-Learning Iterates for Stochastic Shortest Path Problems

On Boundedness of Q-Learning Iterates for Stochastic Shortest Path Problems MATHEMATICS OF OPERATIONS RESEARCH Vol. 38, No. 2, May 2013, pp. 209 227 ISSN 0364-765X (prin) ISSN 1526-5471 (online) hp://dx.doi.org/10.1287/moor.1120.0562 2013 INFORMS On Boundedness of Q-Learning Ieraes

More information

3.1.3 INTRODUCTION TO DYNAMIC OPTIMIZATION: DISCRETE TIME PROBLEMS. A. The Hamiltonian and First-Order Conditions in a Finite Time Horizon

3.1.3 INTRODUCTION TO DYNAMIC OPTIMIZATION: DISCRETE TIME PROBLEMS. A. The Hamiltonian and First-Order Conditions in a Finite Time Horizon 3..3 INRODUCION O DYNAMIC OPIMIZAION: DISCREE IME PROBLEMS A. he Hamilonian and Firs-Order Condiions in a Finie ime Horizon Define a new funcion, he Hamilonian funcion, H. H he change in he oal value of

More information

Optimal Stopping of Partially Observable Markov Processes: A Filtering-Based Duality Approach

Optimal Stopping of Partially Observable Markov Processes: A Filtering-Based Duality Approach 1 Opial Sopping of Parially Observable Marov Processes: A Filering-Based Dualiy Approach Fan Ye, and Enlu Zhou, Meber, IEEE Absrac In his noe we develop a nuerical approach o he proble of opial sopping

More information

Online Appendix to Solution Methods for Models with Rare Disasters

Online Appendix to Solution Methods for Models with Rare Disasters Online Appendix o Soluion Mehods for Models wih Rare Disasers Jesús Fernández-Villaverde and Oren Levinal In his Online Appendix, we presen he Euler condiions of he model, we develop he pricing Calvo block,

More information

RANDOM LAGRANGE MULTIPLIERS AND TRANSVERSALITY

RANDOM LAGRANGE MULTIPLIERS AND TRANSVERSALITY ECO 504 Spring 2006 Chris Sims RANDOM LAGRANGE MULTIPLIERS AND TRANSVERSALITY 1. INTRODUCTION Lagrange muliplier mehods are sandard fare in elemenary calculus courses, and hey play a cenral role in economic

More information

Lecture 4 Notes (Little s Theorem)

Lecture 4 Notes (Little s Theorem) Lecure 4 Noes (Lile s Theorem) This lecure concerns one of he mos imporan (and simples) heorems in Queuing Theory, Lile s Theorem. More informaion can be found in he course book, Bersekas & Gallagher,

More information

Optimality Conditions for Unconstrained Problems

Optimality Conditions for Unconstrained Problems 62 CHAPTER 6 Opimaliy Condiions for Unconsrained Problems 1 Unconsrained Opimizaion 11 Exisence Consider he problem of minimizing he funcion f : R n R where f is coninuous on all of R n : P min f(x) x

More information

Oscillation Properties of a Logistic Equation with Several Delays

Oscillation Properties of a Logistic Equation with Several Delays Journal of Maheaical Analysis and Applicaions 247, 11 125 Ž 2. doi:1.16 jaa.2.683, available online a hp: www.idealibrary.co on Oscillaion Properies of a Logisic Equaion wih Several Delays Leonid Berezansy

More information

SIGNALS AND SYSTEMS LABORATORY 8: State Variable Feedback Control Systems

SIGNALS AND SYSTEMS LABORATORY 8: State Variable Feedback Control Systems SIGNALS AND SYSTEMS LABORATORY 8: Sae Variable Feedback Conrol Syses INTRODUCTION Sae variable descripions for dynaical syses describe he evoluion of he sae vecor, as a funcion of he sae and he inpu. There

More information

Intelligent Packet Dropping for Optimal Energy-Delay Tradeoffs in Wireless Downlinks

Intelligent Packet Dropping for Optimal Energy-Delay Tradeoffs in Wireless Downlinks PROC. OF 4TH INT. SYMPOSIUM ON MODELING AND OPTIMIZATION IN MOBILE, AD HOC, AND WIRELESS NETWORKS (WIOPT), APRIL 2006 1 Inelligen Packe Dropping for Opimal Energy-Delay Tradeoffs in Wireless Downlinks

More information

2.1 Level, Weight, Nominator and Denominator of an Eta Product. By an eta product we understand any finite product of functions. f(z) = m.

2.1 Level, Weight, Nominator and Denominator of an Eta Product. By an eta product we understand any finite product of functions. f(z) = m. Ea Producs.1 Level, Weigh, Noinaor and Denoinaor of an Ea Produc By an ea produc we undersand any finie produc of funcions f(z = η(z a where runs hrough a finie se of posiive inegers and he exponens a

More information

Optimal Server Assignment in Multi-Server

Optimal Server Assignment in Multi-Server Opimal Server Assignmen in Muli-Server 1 Queueing Sysems wih Random Conneciviies Hassan Halabian, Suden Member, IEEE, Ioannis Lambadaris, Member, IEEE, arxiv:1112.1178v2 [mah.oc] 21 Jun 2013 Yannis Viniois,

More information

An Extension to the Tactical Planning Model for a Job Shop: Continuous-Time Control

An Extension to the Tactical Planning Model for a Job Shop: Continuous-Time Control An Exension o he Tacical Planning Model for a Job Shop: Coninuous-Tie Conrol Chee Chong. Teo, Rohi Bhanagar, and Sephen C. Graves Singapore-MIT Alliance, Nanyang Technological Univ., and Massachuses Insiue

More information

Lecture 18 GMM:IV, Nonlinear Models

Lecture 18 GMM:IV, Nonlinear Models Lecure 8 :IV, Nonlinear Models Le Z, be an rx funcion of a kx paraeer vecor, r > k, and a rando vecor Z, such ha he r populaion oen condiions also called esiain equaions EZ, hold for all, where is he rue

More information

14 Autoregressive Moving Average Models

14 Autoregressive Moving Average Models 14 Auoregressive Moving Average Models In his chaper an imporan parameric family of saionary ime series is inroduced, he family of he auoregressive moving average, or ARMA, processes. For a large class

More information

Application of Homotopy Analysis Method for Solving various types of Problems of Partial Differential Equations

Application of Homotopy Analysis Method for Solving various types of Problems of Partial Differential Equations Applicaion of Hooopy Analysis Mehod for olving various ypes of Probles of Parial Differenial Equaions V.P.Gohil, Dr. G. A. anabha,assisan Professor, Deparen of Maheaics, Governen Engineering College, Bhavnagar,

More information

Inventory Analysis and Management. Multi-Period Stochastic Models: Optimality of (s, S) Policy for K-Convex Objective Functions

Inventory Analysis and Management. Multi-Period Stochastic Models: Optimality of (s, S) Policy for K-Convex Objective Functions Muli-Period Sochasic Models: Opimali of (s, S) Polic for -Convex Objecive Funcions Consider a seing similar o he N-sage newsvendor problem excep ha now here is a fixed re-ordering cos (> 0) for each (re-)order.

More information

Lecture 2-1 Kinematics in One Dimension Displacement, Velocity and Acceleration Everything in the world is moving. Nothing stays still.

Lecture 2-1 Kinematics in One Dimension Displacement, Velocity and Acceleration Everything in the world is moving. Nothing stays still. Lecure - Kinemaics in One Dimension Displacemen, Velociy and Acceleraion Everyhing in he world is moving. Nohing says sill. Moion occurs a all scales of he universe, saring from he moion of elecrons in

More information

WEEK-3 Recitation PHYS 131. of the projectile s velocity remains constant throughout the motion, since the acceleration a x

WEEK-3 Recitation PHYS 131. of the projectile s velocity remains constant throughout the motion, since the acceleration a x WEEK-3 Reciaion PHYS 131 Ch. 3: FOC 1, 3, 4, 6, 14. Problems 9, 37, 41 & 71 and Ch. 4: FOC 1, 3, 5, 8. Problems 3, 5 & 16. Feb 8, 018 Ch. 3: FOC 1, 3, 4, 6, 14. 1. (a) The horizonal componen of he projecile

More information

Oscillation of an Euler Cauchy Dynamic Equation S. Huff, G. Olumolode, N. Pennington, and A. Peterson

Oscillation of an Euler Cauchy Dynamic Equation S. Huff, G. Olumolode, N. Pennington, and A. Peterson PROCEEDINGS OF THE FOURTH INTERNATIONAL CONFERENCE ON DYNAMICAL SYSTEMS AND DIFFERENTIAL EQUATIONS May 4 7, 00, Wilmingon, NC, USA pp 0 Oscillaion of an Euler Cauchy Dynamic Equaion S Huff, G Olumolode,

More information

23.2. Representing Periodic Functions by Fourier Series. Introduction. Prerequisites. Learning Outcomes

23.2. Representing Periodic Functions by Fourier Series. Introduction. Prerequisites. Learning Outcomes Represening Periodic Funcions by Fourier Series 3. Inroducion In his Secion we show how a periodic funcion can be expressed as a series of sines and cosines. We begin by obaining some sandard inegrals

More information

I-Optimal designs for third degree kronecker model mixture experiments

I-Optimal designs for third degree kronecker model mixture experiments Inernaional Journal of Saisics and Applied Maheaics 207 2(2): 5-40 ISSN: 2456-452 Mahs 207 2(2): 5-40 207 Sas & Mahs www.ahsjournal.co Received: 9-0-207 Acceped: 20-02-207 Cheruiyo Kipkoech Deparen of

More information

Physics 235 Chapter 2. Chapter 2 Newtonian Mechanics Single Particle

Physics 235 Chapter 2. Chapter 2 Newtonian Mechanics Single Particle Chaper 2 Newonian Mechanics Single Paricle In his Chaper we will review wha Newon s laws of mechanics ell us abou he moion of a single paricle. Newon s laws are only valid in suiable reference frames,

More information

Expert Advice for Amateurs

Expert Advice for Amateurs Exper Advice for Amaeurs Ernes K. Lai Online Appendix - Exisence of Equilibria The analysis in his secion is performed under more general payoff funcions. Wihou aking an explici form, he payoffs of he

More information

Decentralized Stochastic Control with Partial History Sharing: A Common Information Approach

Decentralized Stochastic Control with Partial History Sharing: A Common Information Approach 1 Decenralized Sochasic Conrol wih Parial Hisory Sharing: A Common Informaion Approach Ashuosh Nayyar, Adiya Mahajan and Demoshenis Tenekezis arxiv:1209.1695v1 [cs.sy] 8 Sep 2012 Absrac A general model

More information

Decision Tree Learning. Decision Tree Learning. Decision Trees. Decision Trees: Operation. Blue slides: Mitchell. Orange slides: Alpaydin Humidity

Decision Tree Learning. Decision Tree Learning. Decision Trees. Decision Trees: Operation. Blue slides: Mitchell. Orange slides: Alpaydin Humidity Decision Tree Learning Decision Tree Learning Blue slides: Michell Oulook Orange slides: Alpaydin Huidiy Sunny Overcas Rain ral Srong Learn o approxiae discree-valued arge funcions. Sep-by-sep decision

More information

Homework 2 Solutions

Homework 2 Solutions Mah 308 Differenial Equaions Fall 2002 & 2. See he las page. Hoework 2 Soluions 3a). Newon s secon law of oion says ha a = F, an we know a =, so we have = F. One par of he force is graviy, g. However,

More information

Thus the force is proportional but opposite to the displacement away from equilibrium.

Thus the force is proportional but opposite to the displacement away from equilibrium. Chaper 3 : Siple Haronic Moion Hooe s law saes ha he force (F) eered by an ideal spring is proporional o is elongaion l F= l where is he spring consan. Consider a ass hanging on a he spring. In equilibriu

More information

Appendix to Online l 1 -Dictionary Learning with Application to Novel Document Detection

Appendix to Online l 1 -Dictionary Learning with Application to Novel Document Detection Appendix o Online l -Dicionary Learning wih Applicaion o Novel Documen Deecion Shiva Prasad Kasiviswanahan Huahua Wang Arindam Banerjee Prem Melville A Background abou ADMM In his secion, we give a brief

More information

ON NONLINEAR CROSS-DIFFUSION SYSTEMS: AN OPTIMAL TRANSPORT APPROACH

ON NONLINEAR CROSS-DIFFUSION SYSTEMS: AN OPTIMAL TRANSPORT APPROACH ON NONLINEAR CROSS-DIFFUSION SYSTEMS: AN OPTIMAL TRANSPORT APPROACH INWON KIM AND ALPÁR RICHÁRD MÉSZÁROS Absrac. We sudy a nonlinear, degenerae cross-diffusion odel which involves wo densiies wih wo differen

More information

Delay Reduction via Lagrange Multipliers in Stochastic Network Optimization

Delay Reduction via Lagrange Multipliers in Stochastic Network Optimization IEEE TRANS. ON AUTOMATIC CONTROL, OL. 56, ISSUE 4, PP. 842-857, APRIL 2011 1 Delay Reducion via Lagrange Mulipliers in Sochasic Nework Opimizaion Longbo Huang, Michael J. Neely Absrac In his paper, we

More information

Section 3.5 Nonhomogeneous Equations; Method of Undetermined Coefficients

Section 3.5 Nonhomogeneous Equations; Method of Undetermined Coefficients Secion 3.5 Nonhomogeneous Equaions; Mehod of Undeermined Coefficiens Key Terms/Ideas: Linear Differenial operaor Nonlinear operaor Second order homogeneous DE Second order nonhomogeneous DE Soluion o homogeneous

More information

. Now define y j = log x j, and solve the iteration.

. Now define y j = log x j, and solve the iteration. Problem 1: (Disribued Resource Allocaion (ALOHA!)) (Adaped from M& U, Problem 5.11) In his problem, we sudy a simple disribued proocol for allocaing agens o shared resources, wherein agens conend for resources

More information

An Introduction to Malliavin calculus and its applications

An Introduction to Malliavin calculus and its applications An Inroducion o Malliavin calculus and is applicaions Lecure 5: Smoohness of he densiy and Hörmander s heorem David Nualar Deparmen of Mahemaics Kansas Universiy Universiy of Wyoming Summer School 214

More information

Boundedness and Exponential Asymptotic Stability in Dynamical Systems with Applications to Nonlinear Differential Equations with Unbounded Terms

Boundedness and Exponential Asymptotic Stability in Dynamical Systems with Applications to Nonlinear Differential Equations with Unbounded Terms Advances in Dynamical Sysems and Applicaions. ISSN 0973-531 Volume Number 1 007, pp. 107 11 Research India Publicaions hp://www.ripublicaion.com/adsa.hm Boundedness and Exponenial Asympoic Sabiliy in Dynamical

More information

STATE-SPACE MODELLING. A mass balance across the tank gives:

STATE-SPACE MODELLING. A mass balance across the tank gives: B. Lennox and N.F. Thornhill, 9, Sae Space Modelling, IChemE Process Managemen and Conrol Subjec Group Newsleer STE-SPACE MODELLING Inroducion: Over he pas decade or so here has been an ever increasing

More information

8. Basic RL and RC Circuits

8. Basic RL and RC Circuits 8. Basic L and C Circuis This chaper deals wih he soluions of he responses of L and C circuis The analysis of C and L circuis leads o a linear differenial equaion This chaper covers he following opics

More information

Practice Problems - Week #4 Higher-Order DEs, Applications Solutions

Practice Problems - Week #4 Higher-Order DEs, Applications Solutions Pracice Probles - Wee #4 Higher-Orer DEs, Applicaions Soluions 1. Solve he iniial value proble where y y = 0, y0 = 0, y 0 = 1, an y 0 =. r r = rr 1 = rr 1r + 1, so he general soluion is C 1 + C e x + C

More information

Higher Order Difference Schemes for Heat Equation

Higher Order Difference Schemes for Heat Equation Available a p://pvau.edu/aa Appl. Appl. Ma. ISSN: 9-966 Vol., Issue (Deceber 009), pp. 6 7 (Previously, Vol., No. ) Applicaions and Applied Maeaics: An Inernaional Journal (AAM) Higer Order Difference

More information

E β t log (C t ) + M t M t 1. = Y t + B t 1 P t. B t 0 (3) v t = P tc t M t Question 1. Find the FOC s for an optimum in the agent s problem.

E β t log (C t ) + M t M t 1. = Y t + B t 1 P t. B t 0 (3) v t = P tc t M t Question 1. Find the FOC s for an optimum in the agent s problem. Noes, M. Krause.. Problem Se 9: Exercise on FTPL Same model as in paper and lecure, only ha one-period govenmen bonds are replaced by consols, which are bonds ha pay one dollar forever. I has curren marke

More information

Boosting MIT Course Notes Cynthia Rudin

Boosting MIT Course Notes Cynthia Rudin Credi: Freund, Schapire, Daubechies Boosing MIT 5.097 Course Noes Cynhia Rudin Boosing sared wih a quesion of Michael Kearns, abou wheher a weak learning algorih can be ade ino a srong learning algorih.

More information

Stability and Bifurcation in a Neural Network Model with Two Delays

Stability and Bifurcation in a Neural Network Model with Two Delays Inernaional Mahemaical Forum, Vol. 6, 11, no. 35, 175-1731 Sabiliy and Bifurcaion in a Neural Nework Model wih Two Delays GuangPing Hu and XiaoLing Li School of Mahemaics and Physics, Nanjing Universiy

More information

Optima and Equilibria for Traffic Flow on a Network

Optima and Equilibria for Traffic Flow on a Network Opima and Equilibria for Traffic Flow on a Nework Albero Bressan Deparmen of Mahemaics, Penn Sae Universiy bressan@mah.psu.edu Albero Bressan (Penn Sae) Opima and equilibria for raffic flow 1 / 1 A Traffic

More information

Finish reading Chapter 2 of Spivak, rereading earlier sections as necessary. handout and fill in some missing details!

Finish reading Chapter 2 of Spivak, rereading earlier sections as necessary. handout and fill in some missing details! MAT 257, Handou 6: Ocober 7-2, 20. I. Assignmen. Finish reading Chaper 2 of Spiva, rereading earlier secions as necessary. handou and fill in some missing deails! II. Higher derivaives. Also, read his

More information

A Nonexistence Result to a Cauchy Problem in Nonlinear One Dimensional Thermoelasticity

A Nonexistence Result to a Cauchy Problem in Nonlinear One Dimensional Thermoelasticity Journal of Maheaical Analysis and Applicaions 54, 7186 1 doi:1.16jaa..73, available online a hp:www.idealibrary.co on A Nonexisence Resul o a Cauchy Proble in Nonlinear One Diensional Theroelasiciy Mokhar

More information

Conduction Equation. Consider an arbitrary volume V bounded by a surface S. Let any point on the surface be denoted r s

Conduction Equation. Consider an arbitrary volume V bounded by a surface S. Let any point on the surface be denoted r s Conducion Equaion Consider an arbirary volue V bounded by a surface S. Le any poin on he surface be denoed r s and n be he uni ouward noral vecor o he surface a ha poin. Assue no flow ino or ou of V across

More information

( ) ( ) if t = t. It must satisfy the identity. So, bulkiness of the unit impulse (hyper)function is equal to 1. The defining characteristic is

( ) ( ) if t = t. It must satisfy the identity. So, bulkiness of the unit impulse (hyper)function is equal to 1. The defining characteristic is UNIT IMPULSE RESPONSE, UNIT STEP RESPONSE, STABILITY. Uni impulse funcion (Dirac dela funcion, dela funcion) rigorously defined is no sricly a funcion, bu disribuion (or measure), precise reamen requires

More information

2. Nonlinear Conservation Law Equations

2. Nonlinear Conservation Law Equations . Nonlinear Conservaion Law Equaions One of he clear lessons learned over recen years in sudying nonlinear parial differenial equaions is ha i is generally no wise o ry o aack a general class of nonlinear

More information

Math 334 Fall 2011 Homework 11 Solutions

Math 334 Fall 2011 Homework 11 Solutions Dec. 2, 2 Mah 334 Fall 2 Homework Soluions Basic Problem. Transform he following iniial value problem ino an iniial value problem for a sysem: u + p()u + q() u g(), u() u, u () v. () Soluion. Le v u. Then

More information

Hamilton- J acobi Equation: Weak S olution We continue the study of the Hamilton-Jacobi equation:

Hamilton- J acobi Equation: Weak S olution We continue the study of the Hamilton-Jacobi equation: M ah 5 7 Fall 9 L ecure O c. 4, 9 ) Hamilon- J acobi Equaion: Weak S oluion We coninue he sudy of he Hamilon-Jacobi equaion: We have shown ha u + H D u) = R n, ) ; u = g R n { = }. ). In general we canno

More information

10. State Space Methods

10. State Space Methods . Sae Space Mehods. Inroducion Sae space modelling was briefly inroduced in chaper. Here more coverage is provided of sae space mehods before some of heir uses in conrol sysem design are covered in he

More information

1 Review of Zero-Sum Games

1 Review of Zero-Sum Games COS 5: heoreical Machine Learning Lecurer: Rob Schapire Lecure #23 Scribe: Eugene Brevdo April 30, 2008 Review of Zero-Sum Games Las ime we inroduced a mahemaical model for wo player zero-sum games. Any

More information

A Dynamic Model of Economic Fluctuations

A Dynamic Model of Economic Fluctuations CHAPTER 15 A Dynamic Model of Economic Flucuaions Modified for ECON 2204 by Bob Murphy 2016 Worh Publishers, all righs reserved IN THIS CHAPTER, OU WILL LEARN: how o incorporae dynamics ino he AD-AS model

More information

On the Stability Region of Multi-Queue Multi-Server Queueing Systems with Stationary Channel Distribution

On the Stability Region of Multi-Queue Multi-Server Queueing Systems with Stationary Channel Distribution On he Sabiliy Region of Muli-Queue Muli-Server Queueing Sysems wih Saionary Channel Disribuion Hassan Halabian, Ioannis Lambadaris, Chung-Horng Lung Deparmen of Sysems and Compuer Engineering, Carleon

More information

O Q L N. Discrete-Time Stochastic Dynamic Programming. I. Notation and basic assumptions. ε t : a px1 random vector of disturbances at time t.

O Q L N. Discrete-Time Stochastic Dynamic Programming. I. Notation and basic assumptions. ε t : a px1 random vector of disturbances at time t. Econ. 5b Spring 999 C. Sims Discree-Time Sochasic Dynamic Programming 995, 996 by Chrisopher Sims. This maerial may be freely reproduced for educaional and research purposes, so long as i is no alered,

More information

arxiv: v1 [math.oc] 11 Sep 2017

arxiv: v1 [math.oc] 11 Sep 2017 Online Learning in Weakly Coupled Markov Decision Processes: A Convergence ime Sudy Xiaohan Wei, Hao Yu and Michael J. Neely arxiv:709.03465v [mah.oc] Sep 07. Inroducion Absrac: We consider muliple parallel

More information

Notes for Lecture 17-18

Notes for Lecture 17-18 U.C. Berkeley CS278: Compuaional Complexiy Handou N7-8 Professor Luca Trevisan April 3-8, 2008 Noes for Lecure 7-8 In hese wo lecures we prove he firs half of he PCP Theorem, he Amplificaion Lemma, up

More information

arxiv: v1 [math.pr] 19 Feb 2011

arxiv: v1 [math.pr] 19 Feb 2011 A NOTE ON FELLER SEMIGROUPS AND RESOLVENTS VADIM KOSTRYKIN, JÜRGEN POTTHOFF, AND ROBERT SCHRADER ABSTRACT. Various equivalen condiions for a semigroup or a resolven generaed by a Markov process o be of

More information

Two Coupled Oscillators / Normal Modes

Two Coupled Oscillators / Normal Modes Lecure 3 Phys 3750 Two Coupled Oscillaors / Normal Modes Overview and Moivaion: Today we ake a small, bu significan, sep owards wave moion. We will no ye observe waves, bu his sep is imporan in is own

More information

Underwater vehicles: The minimum time problem

Underwater vehicles: The minimum time problem Underwaer vehicles: The iniu ie proble M. Chyba Deparen of Maheaics 565 McCarhy Mall Universiy of Hawaii, Honolulu, HI 968 Eail: chyba@ah.hawaii.edu H. Sussann Deparen of Maheaics Rugers Universiy, Piscaaway,

More information

Ordinary Differential Equations

Ordinary Differential Equations Ordinary Differenial Equaions 5. Examples of linear differenial equaions and heir applicaions We consider some examples of sysems of linear differenial equaions wih consan coefficiens y = a y +... + a

More information

OBJECTIVES OF TIME SERIES ANALYSIS

OBJECTIVES OF TIME SERIES ANALYSIS OBJECTIVES OF TIME SERIES ANALYSIS Undersanding he dynamic or imedependen srucure of he observaions of a single series (univariae analysis) Forecasing of fuure observaions Asceraining he leading, lagging

More information

2.1 Harmonic excitation of undamped systems

2.1 Harmonic excitation of undamped systems 2.1 Haronic exciaion of undaped syses Sanaoja 2_1.1 2.1 Haronic exciaion of undaped syses (Vaienaaoan syseein haroninen heräe) The following syse is sudied: y x F() Free-body diagra f x g x() N F() In

More information

An Introduction to Backward Stochastic Differential Equations (BSDEs) PIMS Summer School 2016 in Mathematical Finance.

An Introduction to Backward Stochastic Differential Equations (BSDEs) PIMS Summer School 2016 in Mathematical Finance. 1 An Inroducion o Backward Sochasic Differenial Equaions (BSDEs) PIMS Summer School 2016 in Mahemaical Finance June 25, 2016 Chrisoph Frei cfrei@ualbera.ca This inroducion is based on Touzi [14], Bouchard

More information

Problem Set 5. Graduate Macro II, Spring 2017 The University of Notre Dame Professor Sims

Problem Set 5. Graduate Macro II, Spring 2017 The University of Notre Dame Professor Sims Problem Se 5 Graduae Macro II, Spring 2017 The Universiy of Nore Dame Professor Sims Insrucions: You may consul wih oher members of he class, bu please make sure o urn in your own work. Where applicable,

More information

Mapping in Dynamic Environments

Mapping in Dynamic Environments Mapping in Dynaic Environens Wolfra Burgard Universiy of Freiburg, Gerany Mapping is a Key Technology for Mobile Robos Robos can robusly navigae when hey have a ap. Robos have been shown o being able o

More information

Least Favorable Compressed Sensing Problems for First Order Methods

Least Favorable Compressed Sensing Problems for First Order Methods eas Favorable Copressed Sensing Probles for Firs Order Mehods Arian Maleki Deparen of Elecrical and Copuer Engineering Rice Universiy Richard Baraniuk Deparen of Elecrical and Copuer Engineering Rice Universiy

More information

Diebold, Chapter 7. Francis X. Diebold, Elements of Forecasting, 4th Edition (Mason, Ohio: Cengage Learning, 2006). Chapter 7. Characterizing Cycles

Diebold, Chapter 7. Francis X. Diebold, Elements of Forecasting, 4th Edition (Mason, Ohio: Cengage Learning, 2006). Chapter 7. Characterizing Cycles Diebold, Chaper 7 Francis X. Diebold, Elemens of Forecasing, 4h Ediion (Mason, Ohio: Cengage Learning, 006). Chaper 7. Characerizing Cycles Afer compleing his reading you should be able o: Define covariance

More information

Heat kernel and Harnack inequality on Riemannian manifolds

Heat kernel and Harnack inequality on Riemannian manifolds Hea kernel and Harnack inequaliy on Riemannian manifolds Alexander Grigor yan UHK 11/02/2014 onens 1 Laplace operaor and hea kernel 1 2 Uniform Faber-Krahn inequaliy 3 3 Gaussian upper bounds 4 4 ean-value

More information

Comments on Window-Constrained Scheduling

Comments on Window-Constrained Scheduling Commens on Window-Consrained Scheduling Richard Wes Member, IEEE and Yuing Zhang Absrac This shor repor clarifies he behavior of DWCS wih respec o Theorem 3 in our previously published paper [1], and describes

More information

Riemann Hypothesis and Primorial Number. Choe Ryong Gil

Riemann Hypothesis and Primorial Number. Choe Ryong Gil Rieann Hyohesis Priorial Nuber Choe Ryong Gil Dearen of Maheaics Universiy of Sciences Gwahak- dong Unjong Disric Pyongyang DPRKorea Eail; ryonggilchoe@sar-conek Augus 8 5 Absrac; In his aer we consider

More information

arxiv: v3 [math.na] 9 Oct 2017

arxiv: v3 [math.na] 9 Oct 2017 Fas randoized ieraion: diffusion Mone Carlo hrough he lens of nuerical linear algebra Lek-Heng Li 1 and Jonahan Weare 1,2 arxiv:1508.06104v3 [ah.na] 9 Oc 2017 1 Deparen of Saisics, Universiy of Chicago

More information

SZG Macro 2011 Lecture 3: Dynamic Programming. SZG macro 2011 lecture 3 1

SZG Macro 2011 Lecture 3: Dynamic Programming. SZG macro 2011 lecture 3 1 SZG Macro 2011 Lecure 3: Dynamic Programming SZG macro 2011 lecure 3 1 Background Our previous discussion of opimal consumpion over ime and of opimal capial accumulaion sugges sudying he general decision

More information

Approximation Algorithms for Unique Games via Orthogonal Separators

Approximation Algorithms for Unique Games via Orthogonal Separators Approximaion Algorihms for Unique Games via Orhogonal Separaors Lecure noes by Konsanin Makarychev. Lecure noes are based on he papers [CMM06a, CMM06b, LM4]. Unique Games In hese lecure noes, we define

More information