Research Article. Almost Sure Convergence of Random Projected Proximal and Subgradient Algorithms for Distributed Nonsmooth Convex Optimization

Size: px
Start display at page:

Download "Research Article. Almost Sure Convergence of Random Projected Proximal and Subgradient Algorithms for Distributed Nonsmooth Convex Optimization"

Transcription

1 To appear n Optmzaton Vol. 00, No. 00, Month 20XX, 1 27 Research Artcle Almost Sure Convergence of Random Projected Proxmal and Subgradent Algorthms for Dstrbuted Nonsmooth Convex Optmzaton Hdea Idua a a Department of Computer Scence, Mej Unversty, Hgashmta, Tama-u, Kawasa-sh, Kanagawa Japan dua@cs.mej.ac.jp (Receved 00 Month 20XX; accepted 00 Month 20XX) Two dstrbuted algorthms are descrbed that enable all users connected over a networ to cooperatvely solve the problem of mnmzng the sum of all users objectve functons over the ntersecton of all users constrant sets, where each user has ts own prvate nonsmooth convex objectve functon and closed convex constrant set, whch s the ntersecton of a number of smple, closed convex sets. One algorthm enables each user to adjust ts estmate by usng the proxmty operator of ts objectve functon and the metrc projecton onto one constrant set randomly selected from a number of smple, closed convex sets. The other determnes each user s estmate by usng the subdfferental of ts objectve functon nstead of the proxmty operator. Investgaton of the two algorthms convergence propertes for a dmnshng step-sze rule revealed that, under certan assumptons, the sequences of all users generated by each of the two algorthms converge almost surely to the same soluton. It also showed that the rate of convergence depends on the step sze and that a smaller step sze results n qucer convergence. The results of numercal evaluaton usng a nonsmooth convex optmzaton problem support the convergence analyss and demonstrate the effectveness of the two algorthms. Keywords: almost sure convergence; dstrbuted nonsmooth convex optmzaton; metrc projecton; proxmty operator; random projecton algorthm; subgradent AMS Subject Classfcaton: 90C15; 90C25; 90C30 1. Introducton Future networ models have attracted a great deal of attenton. The concept of the networ model consdered here dffers completely from that of a conventonal clent-server networ model. Whle a conventonal clent-server networ model explctly dstngushes hosts provdng servces (servers) from hosts recevng servces (clents), the networ model consdered here does not assgn fxed roles to hosts. Hosts composng the networ, referred to here as users, can be both servers and clents. Hence, the networ can functon as an autonomous, dstrbuted, and cooperatve system. Although there are several forms of networs n whch some operatons are ntentonally centralzed (e.g., hybrd peer-to-peer networs), here the focus s on networs that do not have centralzed operatons. Therefore, dstrbuted mechansms need to be used that can wor n cooperaton wth each user and neghborng users to control the networ. Ths wor was supported by the Japan Socety for the Promoton of Scence through a Grant-n-Ad for Scentfc Research (C) (15K04763). 1

2 Dstrbuted optmzaton (see, e.g., [2], [4], [9], [11, Part II], [12, PART II], [27], [28], [33] and references theren) plays a crucal role n mang future networs [10, 14, 23, 30], such as wreless, sensor, and peer-to-peer networs, stable and hghly relable. One way [23, Chapter 5], [25], [33, Chapter 2] to acheve ths optmzaton s to model the utlty and strategy of each user respectvely as a concave utlty functon and convex constrant set and solve the problem of maxmzng the overall utlty of all users over the ntersecton of all users constrant sets. Ths paper focuses on a constraned convex mnmzaton problem n whch each user has ts own nonsmooth convex objectve functon (.e., a mnus nonsmooth concave utlty functon) and closed convex constrant set that s the ntersecton of a number of smple, closed convex sets (e.g., the ntersecton of affne subspaces, halfspaces, and hyperslabs). The constraned nonsmooth convex optmzaton problem covers the mportant stuatons n whch each user s objectve functon s dfferentable wth a non-lpschtz contnuous gradent or not dfferentable (e.g., the L 1 -norm) and ncludes, for nstance, the problem of mnmzng the total varaton of a sgnal over a convex set, Tyhonov-le problems wth L 1 -norms [13, I. Introducton], the classfer ensemble problem wth sparsty and dversty learnng [39, Subsecton 2.2.3], [40, Subsecton 3.2.4], whch s expressed as L 1 -norm mnmzaton, and the mnmal antenna-subset selecton problem [38, Subsecton 17.4]. The man objectve of the present paper s to show that the constraned nonsmooth convex optmzaton problem for many real-world applcatons can be solved by usng dstrbuted optmzaton technques. Two dstrbuted optmzaton algorthms are presented for solvng the constraned convex mnmzaton problem descrbed above. At each teraton of the frst algorthm, each user calculates the weghted average of ts estmate and the estmates receved from ts neghborng users and then updates ts estmate by usng the weghted average, the proxmty operator of ts own prvate nonsmooth convex functon, and the metrc projecton onto one constrant set randomly selected from a number of smple, closed convex sets. The second algorthm s obtaned by replacng the proxmty operator n the frst algorthm wth the subdfferental of each user s nonsmooth convex functon. The two algorthms are performed on the bass of a framewor [24, (2a)] for local user communcatons and random observatons [24, (2b)], [26, (2)] of the local constrants, whch ensures that each user can observe one smple, closed convex set onto whch the metrc projecton can be effcently calculated. Accordngly, the two algorthms can be appled to two complcated cases. In Case 1, each user does not now the full form of ts prvate constrant set n advance and can observe only one smple, closed convex set at each nstance. In Case 2, each user nows the full form of ts prvate constrant set n advance, and the constrant set s the ntersecton of a huge number of smple, closed convex sets, whch means that a metrc projecton onto the constrant set cannot be calculated easly. See [24, Secton I] and [26, Secton 1] and references theren for detals on applcatons of the two cases, ncludng collaboratve flterng for recommender systems and text classfcaton problems. Proxmal pont methods and subgradent methods wth randomzed order [4, Secton 4], [27, Secton 3] are useful for constraned nonsmooth convex optmzaton. These methods are mplemented under the condton that each user uses a randomly chosen component functon at each teraton whle the proposed algorthms are performed under the condton that each user uses a randomly chosen closed convex set onto whch the metrc projecton can be effcently calculated (Subsectons 3.2 and 4.2 explctly compare the methods n [4, Secton 4] and [27, 2

3 Secton 3] wth the proposed algorthms). Related random projecton algorthms have been proposed for convex optmzaton over the ntersecton of a number of closed convex sets. The algorthm most relevant to the wor reported here s the frst dstrbuted random projected gradent algorthm [24] that was proposed for solvng a constraned smooth convex mnmzaton problem when each user s objectve functon s convex wth a Lpschtz contnuous gradent. The centralzed random projected gradent and subgradent algorthms [26] were proposed for mnmzng a sngle objectve functon over the ntersecton of an arbtrary famly of convex nequaltes. The ncremental constrant projecton-proxmal algorthm [36] uses both random subgradent updates and random constrant updates. Whle there have been no reports on dstrbuted random projecton algorthms for nonsmooth convex optmzaton, thans to the useful deas n [24, 36], a dstrbuted random projected proxmal algorthm (Algorthm 3.1) can be devsed. Moreover, on the bass of [24, 26], a dstrbuted random projected subgradent algorthm (Algorthm 4.1) can be devsed that s a generalzaton of the frst dstrbuted random projected gradent algorthm [24, (2a), (2b)]. Furthermore, there has been much wor [35] on random projecton of vectors n a hgher dmensonal space to a randomly chosen lower dmensonal subspace, and convex programmng problems [29] and affne varatonal nequaltes [31] have been analyzed usng random projectons defned n the same way [35]. Snce, wth the proposed algorthms, each user s assumed to use a randomly chosen projecton at each teraton that projects a vector n the whole space to a closed convex subset of the whole space, the defnton of random projecton n [29, 31, 35] clearly dffers from the one used here. Ths paper maes two contrbutons that buld on prevously reported results. Frst, t presents two novel dstrbuted random projecton algorthms for constraned nonsmooth convex optmzaton that are based on each user s local communcatons. Ths means that they can be mplemented ndependently of the networ topology and that each user can calculate the weghted average of ts estmate and the neghborng users estmates. The algorthms proceed by performng a proxmal or subgradent step for each user s objectve functon at the weghted average and projectng onto one smple, closed convex set that s randomly selected from each user s local constrant sets. Snce the metrc projecton s a specal case of nonexpansve mappng, the algorthms are related to prevous fxed pont optmzaton algorthms [15, 16, 18, 19, 21] for convex optmzaton over fxed pont sets of nonexpansve mappngs. Furthermore, t presents a gradent method [18] that accelerates prevously reported optmzaton algorthms for mnmzng one smooth convex objectve functon over a fxed pont set of a nonexpansve mappng. These algorthms [16, 21] are synchronous decentralzed algorthms that can be appled to smooth convex optmzaton over fxed pont sets of nonexpansve mappngs. Snce these algorthms [16, 18, 21] wor for only smooth convex optmzaton, they cannot be appled to the constraned nonsmooth convex optmzaton problem consdered here. Parallel and ncremental subgradent algorthms [15] have been proposed to solve the problem of mnmzng the sum of nonsmooth, convex objectve functons over fxed pont sets, and ncremental proxmal pont algorthms [19] have been proposed for solvng the problem. These algorthms [15, 19] wor for nonsmooth convex optmzaton over the ntersecton of convex constrant sets that are not always smple. However, snce they wor only when each user maes the best use of ts own prvate nformaton, they cannot be appled to the case n whch each user does not now the explct form of ts constrant set n advance. Moreover, snce these algorthms [15, 19] can be appled only to determnstc optmzaton, they cannot be appled to the case n whch each user uses one constrant set selected 3

4 randomly at each teraton. In contrast, the proposed algorthms wor even when each user randomly sets one projecton selected from many projectons (see Case 1 descrpton above). Therefore, the proposed algorthms have wder applcaton than prevous fxed pont optmzaton algorthms [15, 16, 18, 19, 21]. The second contrbuton s an analyss of the proposed proxmal and subgradent algorthms. In contrast to the convergence analyss of the frst dstrbuted random projected gradent algorthm [24], smooth convex analyss, whch has tractable propertes due to the use of Lpschtz contnuous gradents, cannot be appled to the convergence analyses of the two proposed algorthms, whch optmze nonsmooth convex functons. However, convergence analyses of the two algorthms can be performed by usng useful propertes [1, Propostons 12.16, 12.27, and 16.14] (Proposton 2.1) of the proxmty operators and the subgradents of nonsmooth convex objectve functons. Thans to the supermartngale convergence theorem [5, Proposton ] (Proposton 2.2) and the portmanteau lemma [6, Theorem 16], [34, Lemma 2.2] (Proposton 2.3), t s guaranteed that, under certan assumptons, the sequences of all users generated by each of the two algorthms converge almost surely to the same soluton to the constraned nonsmooth convex optmzaton problem consdered n ths paper (Theorems 3.1 and 4.1). Moreover, the rates of convergence of the two algorthms (Proposton 3.1, (11), Proposton 4.1, and (22)) are provded to llustrate ther algorthm effcences. The convergence rate analyss leads to selecton of a step sze such that the two algorthms converge qucly. Numercal results for the two algorthms are provded to support the convergence and convergence rate analyses. Ths paper s organzed as follows. Secton 2 gves the mathematcal prelmnares and states the man problem. Secton 3 presents the proposed random projected proxmal algorthm for solvng the man problem and descrbes ts convergence propertes for a dmnshng step sze. Secton 4 presents the proposed random projected subgradent algorthm for solvng the man problem and descrbes ts convergence propertes for a dmnshng step sze. Secton 5 presents a numercal evaluaton usng a nonsmooth convex optmzaton problem and compares the behavors of the two algorthms. Secton 6 concludes the paper wth a bref summary and mentons future drectons for mprovng the proposed algorthms. 2. Prelmnares 2.1. Defntons and propostons Let R d be a d-dmensonal Eucldean space wth nner product, and ts nduced norm, and let R d + := {(x 1, x 2,..., x d ) R d : x 0 ( = 1, 2,..., d)}. Let [W ] j and W denote the (, j)th entry and the transpose of a matrx W. Let Pr{X} and E[X] denote the probablty and the expectaton of a random varable X. The probablty space consdered here s denoted by (Ω, F, Pr). The metrc projecton onto a nonempty, closed convex set C R d s denoted by P C, and t s defned for all x R d by P C (x) C and x P C (x) = d(x, C) := nf{ x y : y C}. Mappng P C satsfes the frm nonexpansvty condton [1, Proposton 4.8];.e., P C (x) P C (y) 2 + (x P C (x)) (y P C (y)) 2 x y 2 (x, y R d ). Ths means that P C (x) P C (y) x y (x, y R d );.e., P C s nonexpansve. The subdfferental [1, Defnton 16.1] of f : R d R s the set-valued operator defned for all x R d by f(x) := {u R d : f(y) f(x) + y x, u (y R d )}. The proxmty operator [1, Defnton 12.23] of a convex functon f : R d R, denoted by prox f, maps every x R d to the unque mnmzer of f( )+(1/2) x 2 ; 4

5 .e., prox f (x) Argmn y R d[f(y) + (1/2) x y 2 ] (x R d ). The unqueness and exstence of prox f (x) are guaranteed for all x R d [1, Defnton 12.23]. Proposton 2.1 Let f : R d R be convex. Then the followng hold: () f(x) s nonempty for all x R d. () Let x, p R d. p = prox f (x) f and only f x p f(p) (.e., y p, x p + f(p) f(y) for all y R d ). () prox f s frmly nonexpansve. (v) Let L > 0. Then f s Lpschtz contnuous wth a Lpschtz constant L f and only f u L for all x R d and for all u f(x). Proof: () () Propostons 16.14() and n [1] lead to () and () whle Proposton n [1] mples (). (v) Theorem 6.2.2, Corollary 6.1.2, and Exercse 6.1.9(c) n [7] lead to (v). The followng propostons are needed to prove the man theorems. Proposton 2.2 [The supermartngale convergence theorem [5, Proposton ]] Let (Y ) 0, (Z ) 0, and (W ) 0 be sequences of nonnegatve random varables, and let F ( 0) denote the collecton Y 0, Y 1,..., Y, Z 0, Z 1,..., Z, and W 0, W 1,..., W. Suppose that =0 W < almost surely and that almost surely, for all 0, E[Y +1 F ] Y Z + W. Then =0 Z < almost surely and (Y ) 0 converges almost surely to a nonnegatve random varable Y. Proposton 2.3 [The portmanteau lemma [6, Theorem 16], [34, Lemma 2.2]] Let (Y ) 0 be a sequence of random varables that converges n law to a random varable Y. Then lm sup Pr{Y F } Pr{Y F } for every closed set F. A drected graph G := (V, E) s a fnte nonempty set V of nodes (users) and a collecton E of ordered pars of dstnct nodes from V [3, p. 394]. A drected graph s sad to be strongly connected f, for each par of nodes and l, there exsts a drected path from to l [3, p. 394] Man problem and assumptons Let us consder a constraned nonsmooth convex optmzaton problem that s dstrbuted over a networ of m users, ndexed by V := {1, 2,..., m}. User ( V ) has ts own prvate functon f : R d R and constrant set X R d. On the bass of [24, Secton II], let us defne the constrant set X := m X for the whole networ. Suppose that X s the ntersecton of n closed convex sets. Let I := {1, 2,..., n}, and let I ( V ) be the partton of I such that I = m I and I I l = for all, l V wth l. Then X can be defned by the ntersecton of closed convex sets X j (j I );.e., X := j I X j ( V ) and X := V X = m The followng s assumed hereafter. Assumpton 2.1 Suppose that j I X j. (A1) X j Rd ( V, j I ) s a closed convex set onto whch the metrc projecton can be effcently computed, and X := V X ; P X j (A2) f : R d R ( V ) s convex; (A3) For all V, there exsts M R such that sup{ g : x X, g f (x)} M. Assumpton (A3) s satsfed f f ( V ) s polyhedral on X or X ( V ) s 5

6 bounded [5, p. 471]. Proposton 2.1(v) guarantees that, f f ( V ) s Lpschtz contnuous on X, Assumpton (A3) holds. The followng s the man problem dscussed here. Problem 2.1 Under Assumpton 2.1, mnmze f (x) := V f (x) subject to x X := X, where one assumes that Problem 2.1 has a soluton. The soluton set of Problem 2.1 s denoted by X := {x X : f(x ) = f := nf x X f(x)}. The condton X holds, for example, when one of X j s s bounded [1, Corollary 8.31, Proposton 11.14]. The man objectve here s to present dstrbuted optmzaton algorthms that enable each user to solve Problem 2.1 wthout usng other users prvate nformaton. Ths goal s addressed by assumng that each user and ts neghborng users form a networ n whch each user can transmt ts estmate to ts neghborng users. The networ topology at tme s expressed as a drected graph G() := (V, E()), where E() V V, and (, j) E() stands for a ln such that user receves nformaton from user j at tme. Let N () V be the set of users that send nformaton to user ;.e., N () := {j V : (, j) E()} and N () ( V, 0). To consder Problem 2.1, the followng assumptons are needed [24, Assumptons 4 and 5], whch leads to Lemma 3.3 (see [32, Theorem 4.2]) used to prove the man theorems. Assumpton 2.2 There exsts Q 1 such that the graph (V, Q 1 l=0 E( + l)) s strongly connected for all 0. Assumpton 2.3 For 0, user ( V ) has the weghted parameters w j () (j V ) satsfyng the followng: () w j () := [W ()] j 0 for all j V, and w j () = 0 when j / N (); () There exsts w (0, 1) such that w j () w for all j N (). () j V [W ()] j = 1 for all V and V [W ()] j = 1 for all j V. Moreover, user ( V ) has the step sze (α ) 0 =0 α = and (C2) =0 α2 <. V (0, ) satsfyng (C1) The matrx W () ( 0) satsfyng Assumpton 2.3() () s sad to be doubly stochastc [24, Assumpton 5]. Theorem 1 n [37] ndcates that the spectral radus ρ of the doubly stochastc matrx W () ( 0) s less than 1 f and only f, for all 0, lm W s ()s = ee m, (1) where e R d stands for the column vector n whch all entres are equal to 1. Moreover, Theorem n [8] and the remar n [37, Theorem 1] show that (1) means the networ s strongly connected. Accordngly, Assumpton 2.2 holds f ρ < 1 (e.g., ρ < 1 holds when one s a smple egenvalue of W () and all other egenvalues are strctly less than one n magntude [37, p. 67]). See [17, Examples 2.6 and 2.7] for examples of W () satsfyng Assumpton 2.3() (). Assumptons (A1) and (A2) mply that, for all 0, user ( V ) can determne 6

7 ts estmate by usng a subdfferental or proxmty operator of f and the metrc projecton onto a certan constrant set selected from ts own constrant sets X j (j I ). Here t s assumed that user ( V ) forms the metrc projecton on the bass of random observatons of the local constrants;.e., user observes a local constrant set at tme, X Ω(), where Ω () I s a random varable, and thus uses P Ω X (). Convergence analyses are performed by assumng the followng [24, Assumptons 2 and 3]: Assumpton 2.4 The random sequences (Ω ()) 0 ( V ) are ndependent and dentcally dstrbuted and ndependent of the ntal ponts x (0) ( V ) n the algorthms presented here. Moreover, Pr{Ω () = j} > 0 holds for V and for j I. Assumpton 2.5 There exsts c > 0 such that, for all V, for all x R d, and for all 0, d(x, X) 2 ce[d(x, X Ω () ) 2 ]. The sequences (Ω ()) 0 ( V ) satsfy Assumpton 2.4 when the sample constrants are generated through state transtons of a Marov chan [36, Subsecton 4.4]. Assumpton 2.4 s satsfed when the constrants are sampled n a cyclc manner by random shufflng or n accordance wth a determnstc cyclc order [36, Subsecton 4.3]. Other examples of (Ω ()) 0 satsfyng Assumpton 2.4 are descrbed n Subsectons 4.1 and 4.2 n [36]. Assumpton 2.5 holds when X j ( V, j I ) s a lnear nequalty or an equalty constrant, or when X has an nteror pont (see [24, p. 223] and references theren). Even f X j ( V, j I ) s a restrcted constrant such as a lnear nequalty or equalty constrant, X := j I X j ( V ) s complcated n the sense that metrc projecton onto X s not necessarly easy. Accordngly, the metrc projecton onto the whole constrant set X := V X cannot be effcently computed, and hence, X := V j I X j has a complcated form even when X j s smple enough to have a closed form expresson of the metrc projecton onto X j. The proposed algorthms (Algorthms 3.1 and 4.1) usng the metrc projecton onto X j can be appled to Problem 2.1 wth such a complcated constrant X. In contrast, the prevously reported algorthms [4, Secton 4], [27, Secton 3] need to use metrc projecton onto the whole constrant set X (see Subsectons 3.2 and 4.2 for detaled comparsons of the proposed algorthms wth the prevous ones [4, Secton 4], [27, Secton 3]). Let x () R d be the estmate of user at tme (see (4) and (16) n Algorthms 3.1 and 4.1 for detals of the defnton of (x ()) 0 ( V )). The proposed algorthms are analyzed by usng the expectaton taen wth respect to the past hstory of the algorthms defned as follows [24, p. 224]. Let F be the σ-algebra generated by the entre hstory of the algorthms up to tme ( 1) nclusvely;.e., F 0 := σ(x (0) ( V )), and for all 1, F := σ (x (0) ( V ), Ω (l) (l [0, 1], V )). (2) 3. Dstrbuted random projected proxmal algorthm Ths secton presents the proposed proxmal algorthm usng random projectons for solvng Problem 2.1 under Assumptons Algorthm 3.1 7

8 Step 0. User ( V ) sets x (0) arbtrarly. Step 1. User ( V ) receves x j () from ts neghborng users j N () and computes the weghted average v () := j N () User updates ts estmate x ( + 1) by usng x ( + 1) := P X Ω () The algorthm sets := + 1 and returns to Step 1. w j () x j (). (3) ( proxα f (v ()) ). (4) Defnton (2) mples that, gven F ( 0), the collecton x (0), x (1),..., x () and v (0), v (1),..., v () generated by Algorthm 3.1 s fully determned. Algorthm 3.1 enables user ( V ) to determne x ( + 1) by usng ts proxmty operator for the weghted average v () of the receved x j () (j N ()) and the metrc projecton onto a constrant set X Ω () randomly selected from X j. The convergence analyss of Algorthm 3.1 s as follows. Theorem 3.1 Under Assumptons , the sequence (x ()) 0 ( V ) generated by Algorthm 3.1 converges almost surely to a random pont x X Proof of Theorem 3.1 The followng lemma can be proven by usng Proposton 2.1 and the frm nonexpansvty condton of P Ω X () and by referrng to the proof of [26, Lemma 2]. Hence, the proof of Lemma 3.1 can be omtted. Lemma 3.1 Suppose that Assumpton 2.1 holds. Then for all x X, for all V, for all z X, for all 0, and for all τ, η, µ > 0, x ( + 1) x 2 v () x 2 2α (f (z) f (x)) + N (τ, η) α 2 + τ v () z 2 + (η + µ 2) v () prox αf (v ()) 2 ( 1 1 ) x ( + 1) v () 2, µ where N(τ, η) := (1/τ + 1/η) max V M 2 (A3). < and M ( V ) s as n Assumpton The followng lemma s shown. Lemma 3.2 Suppose that Assumptons 2.1, 2.3, and 2.5 hold. Then =0 v () prox α f (v ()) 2 <, and =0 d(v (), X) 2 < almost surely for all V. Proof: The defnton of d means that d(v (), X) = v () P X (v ()) and d(x ( + 1), X) x ( + 1) P X (v ()) ( V, 0). The condton x ( + 1) X Ω () ( V, 0) means that d(v (), X Ω () ) v () x ( + 1) ( V, 0). Then Lemma 3.1 wth x = z := P X (v ()) ( V, 0) guarantees 8

9 that, for all V, for all 0, for all τ, η > 0, and for all µ > 1, d (x ( + 1), X) 2 d (v (), X) 2 + τd (v (), X) 2 + (η + µ 2) v () prox αf (v ()) 2 + N (τ, η) α 2 ( 1 1 ) ( ) d v (), X Ω() 2. µ By tang the expectaton n ths nequalty condtoned on F defned n (2), Assumpton 2.5 leads to the fndng that, almost surely, for all V, for all 0, for all τ, η > 0, and for all µ > 1, E [d ] (x ( + 1), X) 2 F d (v (), X) 2 + (η + µ 2) v () prox αf (v ()) 2 { + τ 1 ( 1 1 )} d (v (), X) 2 + N (τ, η) α 2 c µ. Let us tae τ := 1/(6c), η := 1/3, and µ := 3/2. From η + µ 2 = 1/6, τ (1/c)(1 1/µ) = 1/(6c), and the convexty of d(, X) 2, almost surely for all V and for all 0, E [d ] (x ( + 1), X) 2 F m [W ()] j d (x j (), X) 2 1 v () prox 6 α f (v ()) 2 j=1 1 6c d (v (), X) 2 + N ( 1 6c, 1 ) α 2 3, where N(1/(6c), 1/3) < s guaranteed from Assumpton (A3). Hence, Assumpton 2.3 ensures that, almost surely, for all 0, [ m ] E d (x ( + 1), X) 2 F m d (x j (), X) j=1 1 6c m v () prox α f (v ()) 2 m d (v (), X) 2 + mn ( 1 6c, 1 ) α 2 3. (5) Proposton 2.2 and (C2) lead to m =0 v () prox αf (v ()) 2 < and m =0 d(v (), X) 2 < almost surely. Ths completes the proof. The followng lemma can be proven by usng Lemma 3.2 and by referrng to the proof of [24, Lemma 7]. Lemma 3.3 Suppose that Assumptons 2.1, 2.2, and 2.3 hold and defne v() := (1/m) m l=1 v l() for all 0. Then =0 e () 2 := =0 x ( +1) v () 2 < and =0 α v () v() < almost surely for all V. Next, let us show the followng lemma. Lemma 3.4 Suppose that the assumptons n Theorem 3.1 are satsfed, and defne z () := P X (v ()) for all V and for all 0 and z() := (1/m) m z () for all 0. Then the sequence ( x () x ) 0 converges almost surely for all V and for all x X and lm nf f( z()) = f almost surely. 9

10 Proof: Choose x X arbtrarly. The convexty of 2 and Assumpton 2.3 mply that m v () x 2 m j=1 x j() x 2 ( 0). Lemma 3.1 mples that, for all 0 and for all τ, η, µ > 0, m x ( + 1) x 2 m x () x 2 2α + τ m m v () z () 2 + (η + µ 2) (f (z ()) f (x )) ( 1 1 µ ) m x ( + 1) v () 2 m v () prox α f (v ()) 2 + mn (τ, η) α 2. From z () X ( V, 0), the convexty of X ensures that z() X X ( V ). Accordngly, (A3) means that ḡ () M for all ḡ () f ( z()) ( V, 0). The defnton of f ( V ) and the Cauchy-Schwarz nequalty thus guarantee that, for all V and for all 0, f (z ()) f (x ) = f (z ()) f ( z()) + f ( z()) f (x ) z () z(), ḡ () + f ( z()) f (x ) M z () z() + f ( z()) f (x ), where M := max V M <. Moreover, the convexty of and the nonexpansvty of P X mply that, for all V and for all 0, z () z() = (1/m) m l=1 (P X(v ()) z l ()) (1/m) m l=1 P X(v ()) P X (v l ()) (1/m) m l=1 v () v l (), whch, together wth the trangle nequalty, mples that, for all V and for all 0, z () z() v () v() + (1/m) m l=1 v l() v(), where v() := (1/m) m l=1 v l() ( 0). Accordngly, for all V and for all 0, f (z ()) f (x ) M v () v () M m + f ( z ()) f (x ). m v l () v () l=1 (6) Hence, the defntons of f and f mply that, for all 0 and for all τ, η, µ > 0, m x ( + 1) x 2 m x () x Mα m v () v () + 2 Mα m v l () v () 2α (f ( z ()) f ) + τ l=1 m v () z () 2 + (η + µ 2) ( 1 1 ) m x ( + 1) v () 2 µ m v () prox α f (v ()) 2 + mn (τ, η) α 2, whch, together wth d(v (), X) = v () z () and d(v (), X Ω () ) v () x ( + 1) ( V, 0), mples that, for all 0, for all τ, η > 0, and for all 10

11 µ > 1, m x ( + 1) x 2 m x () x Mα 2α (f ( z ()) f ) + τ m v () v () m d (v (), X) 2 ( 1 1 ) m ( ) d v (), X Ω() 2 µ m + (η + µ 2) v () prox α f (v ()) 2 + mn (τ, η) α 2. Accordngly, Assumpton 2.5 guarantees that, almost surely, for all 0, for all τ, η > 0, and for all µ > 1, [ m ] E x ( + 1) x 2 F m x () x Mα m v () v () 2α (f ( z ()) f ) { + τ 1 ( 1 1 )} m d (v (), X) 2 c µ + (η + µ 2) m v () prox αf (v ()) 2 + mn (τ, η) α 2. (7) Settng τ := 1/(6c), η := 1/3, and µ := 3/2 (also see proof of Lemma 3.2) n the nequalty above leads to the fndng that, almost surely, for all 0, [ m ] E x ( + 1) x 2 F m x () x 2 2α (f ( z ()) f ) + 4 Mα m v () v () + mn ( 1 6c, 1 ) α 2 3. Therefore, snce z() X mples f( z()) f 0 ( 0), Proposton 2.2, (C2), and Lemma 3.3 ensure that ( x () x ) 0 converges almost surely for all V. Moreover, (8) α (f ( z ()) f ) < (9) =0 s almost surely satsfed. Now, let us assume that lm nf f( z()) f does not hold almost surely. Then for all Ω F, Pr{ Ω} = 1, and there exsts ω Ω such that lm nf f( z()(ω)) f > 0. Hence, 1 > 0 and γ > 0 can be chosen 11

12 such that f( z()(ω)) f γ for all 1. Accordngly, (9) and (C1) mean that = γ = 1 α = 1 α (f ( z () (ω)) f ) <, whch s a contradcton. Therefore, lm nf f( z()) f almost surely. From f( z()) f 0 ( 0), lm nf f( z()) = f almost surely. Ths completes the proof. It s now possble to prove Theorem 3.1. Proof: Lemma 3.4 guarantees the almost sure convergence of (x ()) 0 ( V ). From (3), (v ()) 0 ( V ) also converges almost surely. The defnton of v() ( 0) mples the almost sure convergence of ( v()) 0. Moreover, Lemma 3.2 mples that, for all V, =0 d(v (), X) 2 = =0 v () z () 2 < almost surely;.e., lm v () z () = 0 almost surely. Accordngly, (z ()) 0 ( V ) converges almost surely. Ths mples that ( z()) 0 converges almost surely to x ;.e., ( z()) 0 converges n law to x. Hence, the closedness of X and Proposton 2.3 guarantee that lm sup Pr{ z() X} Pr{x X}. Snce the defnton of z() ( 0) mples that Pr{ z() X} = 1 ( 0), Pr{x X} = 1. Moreover, Lemma 3.4 and the contnuty of f ensure that, almost surely, f (x ) = lm f ( z()) = lm nf f ( z()) = f ;.e., x X. The defntons of v() and z() ( 0) mean that, for all 0, v() z() (1/m) m z () v (), whch, together wth lm v () z () = 0 almost surely for all V, mples that lm v() z() = 0 almost surely. Accordngly, the almost sure convergence of ( z()) 0 to x X guarantees that ( v()) 0 also converges almost surely to the same x X. Snce Lemma 3.3 mples that =0 α v () v() < almost surely for all V, a dscusson smlar to the one for obtanng lm nf f( z()) f almost surely (see proof of Lemma 3.4) and (C1) guarantee that lm nf v () v() = 0 almost surely for all V. Moreover, the trangle nequalty mples that v () x v () v() + v() x ( V, 0). Hence, from lm v() x = 0 and lm nf v () v() = 0 ( V ) almost surely, lm nf v () x = 0 almost surely for all V. Therefore, the almost sure convergence of (v ()) 0 ( V ) leads to the fndng that, for all V, lm v () x = 0 almost surely. (10) Snce Lemma 3.3 ensures that, for all V, lm e () 2 = lm x (+1) v () 2 = 0 almost surely, (x ()) 0 ( V ) converges almost surely to x X. Ths completes the proof Convergence rate analyss for Algorthm 3.1 The dscusson n Subsecton 3.1 leads to the fndng that the sequence of the feasblty error (d(x (), X) 2 ) 0 and the sequence of the teraton error ( x () x 2 ) 0 are stochastcally decreasng n the sense of the nequaltes n the followng proposton. 12

13 Proposton 3.1 Suppose that the assumptons n Theorem 3.1 hold, that x X s a soluton to Problem 2.1, and that (x ()) 0 ( V ) s the sequence generated by Algorthm 3.1. Then there exst β (j) > 0 (j = 1, 2, 3, 4, 5) such that, almost surely, for all 0, [ m ] E d (x ( + 1), X) 2 F [ m ] E x ( + 1) x 2 F where γ (1) α 2, γ(4) m d (x (), X) 2 m x () x 2 := m d(v (), X) 2, γ (2) := α m v () v(), and γ (5) =0 γ(j) < (j = 1, 2, 3, 4, 5). j=1,2 j=1,2,5 β (j) γ (j) + β (3) γ (3), β (j) γ (j) + j=3,4 β (j) γ (j), := m v () prox αf (v ()) 2, γ (3) := := α (f( z()) f ) ( 0) satsfy Proof: Lemma 3.2 guarantees that γ (1) := m d(v (), X) 2 and γ (2) := m v () prox αf (v ()) 2 ( 0) satsfy =0 γ(j) < almost surely for j = 1, 2. Condton (C2) mples that γ (3) := α 2 ( 0) satsfes =0 γ(3) <. From := α (f( z()) f ) ( 0) also satsfy =0 γ(j) < almost surely for j = 4, 5. Set τ := 1/(6c), η := 1/3, and µ := 3/2, and put β (1) := (τ (1/c)(1 1/µ)) = 1/(6c), β (2) := 2 η µ = 1/6, β (3) := mn(τ, η), β (4) := 4 M, and β (5) := 2, where β (3), β (4) < hold from (A3) and M := max V M. Accordngly, (5) and (7) ensure that Proposton 3.1 holds. Lemma 3.3 and (9), γ (4) := α m v () v() and γ (5) From j=1,2 β(j) γ (j) + β (3) γ (3) β (3) γ (3) = β (3) α 2 and j=1,2,5 β(j) γ (j) + j=3,4 β(j) γ (j) = (β (3) α + β (4) m v () v() )α ( 0), j=3,4 β(j) γ (j) Proposton 3.1 and Theorem 3.1 (see (10) for the almost sure boundedness of (v ()) 0 ( V )) ndcate that (x ()) 0 ( V ) n Algorthm 3.1 converges almost surely to a soluton to Problem 2.1 under the followng convergence rate condtons: for all 0, [ m ] m E d (x ( + 1), X) 2 F d (x (), X) 2 + O ( α) 2, [ m ] E x ( + 1) x 2 F m x () x 2 + O (α ). (11) Moreover, under the condton that x ( + 1) x x () x holds for all V and for a large enough, (8) means that, almost surely, f ( z()) f + 2 M m v () v() + O (α ), (12) where t s guaranteed from Theorem 3.1 that ( v () v() ) 0 ( V ) converges almost surely to 0. From Proposton 3.1, the convergence rate of Algorthm 3.1 depends on β (j) and (γ (j) ) 0 (j = 1, 2, 3, 4, 5);.e., the number of users m and the step sze (α ) 0 13

14 (see (11) and (12)). When m s fxed, t s desrable to set (α ) 0 so that, for all 0, E[ m d(x ( + 1), X) 2 F ] < m d(x (), X) 2 and E[ m x ( + 1) x 2 F ] < m x () x 2 are almost surely satsfed. Accordngly, Algorthm 3.1converges qucly f (α ) 0 s chosen so as to satsfy j=1,2 β (j) γ (j) + β (3) γ (3) < 0 and j=1,2,5 β (j) γ (j) + j=3,4 β (j) γ (j) < 0 ( 0). Hence, t would be desrable to set (α ) 0 so as to satsfy j=3,4 β(j) γ (j) = β (3) α 2 + β(4) α m v () v() 0 as much as possble, e.g., to set α := α/( + 1) wth a small postve constant α. Secton 5 gves numercal examples showng that Algorthm 3.1 wth α := 10 3 /( + 1) converges more qucly than Algorthm 3.1 wth α := 1/( + 1). Here, let us compare the convergence analyss of Algorthm 3.1 (Theorem 3.1 and Proposton 3.1) wth the convergence analyss of the ncremental subgradentproxmal methods n [4, Secton 4]. The problem consdered n [4] was mnmze F (x) = V F (x) := V (f (x) + h (x)) subject to x X, (13) where f, h : R d R ( V ) are convex and X R d s nonempty, closed, and convex. Under the condtons that f ( V ) are sutable for a proxmal teraton and the subgradents of h ( V ) are effcently computed, the followng method [4, Secton 4, (4.2)] was proposed for solvng Problem (13): z() := prox α f ω (x()), g ω h ω (z()), x( + 1) := P X (z() α g ω ), (14) where (ω ) 0 V := {1, 2,..., m} s a sequence of random varables. Algorthm (14) s an ncremental optmzaton algorthm that uses the metrc projecton onto the whole constrant set X and the proxmty operator and subdfferental of one functon selected randomly from {F } V whle Algorthm 3.1 s a dstrbuted optmzaton algorthm that uses the proxmty operator of each user s objectve functon and the metrc projecton onto one closed convex set selected randomly from each user s constrant sets. Proposton 4.3 n [4] shows that, under certan assumptons, Algorthm (14) wth (α ) 0 satsfyng (C1) and (C2) converges almost surely to some random pont n the soluton set of Problem (13). Theorem 3.1 guarantees the almost sure convergence of Algorthm 3.1 wth the condtons (C1) and (C2) to a random pont n X. Proposton 4.2 n [4] ndcates that, under certan assumptons, Algorthm (14) wth α := α > 0 ( 0) satsfes that, for all ϵ > 0, almost surely mn F (x()) F + 5αmc2 + ϵ, (15) 0 N 2 where F stands for the optmal value of Problem (13), c R s a constant, X s the soluton set of Problem (13), and N s a random varable wth E[N] md(x(0), X ) 2 /(αϵ). From (15) and Proposton 3.1, the convergence rates of Algorthms 3.1 and (14) depend on the number of elements n V and the step sze (α ) 0. 14

15 4. Dstrbuted random projected subgradent algorthm Ths secton presents the subgradent algorthm wth random projectons for solvng Problem 2.1. Algorthm 4.1 Step 0. User ( V ) sets x (0) arbtrarly. Step 1. User ( V ) receves x j () from ts neghborng users j N () and computes the weghted average v () defned as n (3) and the subgradent g () f (v ()). User updates ts estmate x ( + 1) by x ( + 1) := P X Ω () (v () α g ()). (16) The algorthm sets := + 1 and returns to Step 1. In ths secton, Assumpton (A3) s replaced wth (A3) For all V, there exsts M R such that sup{ g : x X, g f (x)} M. For all V, there exsts C R such that sup{ g () : g () f (v ()), 0} C. It s obvous that Assumpton (A3) s stronger than Assumpton (A3). Assumpton (A3) holds when f ( V ) s Lpschtz contnuous on R d (Proposton 2.1(v)). The boundedness of (g ()) 0 ( V ) s needed to show that Algorthm 4.1 satsfes =0 d(v (), X) 2 < almost surely for all V, whch s the essental part of the convergence analyss of Algorthm 4.1 (see Lemmas 4.1 and 4.2 for detals). Let us do a convergence analyss of Algorthm 4.1. Theorem 4.1 Under Assumptons , the sequence (x ()) 0 ( V ) generated by Algorthm 4.1 converges almost surely to a random pont x X. Let us compare the dstrbuted random projected gradent algorthm [24, (2a) and (2b)] wth Algorthm 4.1. The algorthm [24, (2a) and (2b)] s the poneerng dstrbuted optmzaton algorthm that s based on local communcatons of users estmates n a networ and a gradent descent wth random projectons. It can be appled to Problem 2.1 when f ( V ) s convex and dfferentable wth the Lpschtz gradent f [24, Assumpton 1 c)]. The algorthm [24, (2a) and (2b)] s x ( + 1) := P X Ω () (v () α f (v ())), (17) where v () s defned as n (3). Proposton 1 n [24] ndcates that, under the assumptons n Theorem 3.1, the sequence (x ()) 0 ( V ) generated by Algorthm (17) converges almost surely to x X. In contrast to Algorthm (17), Algorthm 4.1 can be appled to nonsmooth convex optmzaton (see Assumpton 2.1(A2)) by usng the subgradents g () f (v ()) ( V, 0) and enables all users to arrve at the same soluton to Problem 2.1 under the assumptons n Theorem 4.1 that are stronger than the ones n Theorem 3.1. In Algorthm 3.1, each user sets x ( + 1) by usng ts proxmty operator (see (4) for defnton of x () ( 0) n Algorthm 3.1) and can solve Problem 2.1 under the assumptons n Theorem 3.1 (see Theorem 3.1). 15

16 4.1. Proof of Theorem 4.1 The proof starts wth the followng lemma, whch can be proven by referrng to the proof of [26, Lemma 2]. Lemma 4.1 Suppose that Assumpton 2.1 holds. Then for all x X, for all V, for all z X, for all 0, and for all τ, η > 0, x ( + 1) x 2 v () x 2 2α (f (z) f (x)) + M (τ, η) α 2 + τ v () z 2 + (η 1) v () x ( + 1) 2, where M(τ, η) := max V (M 2 /τ + C2 /η) <, and M and C ( V ) are as n Assumpton (A3). The followng lemma can also be proven by referrng to the proof of Lemma 3.2 and by usng Lemma 4.1 (see [20] for the detals of the proof of Lemma 4.2). Lemma 4.2 Suppose that Assumptons 2.1, 2.5, and 2.3 hold. Then =0 d(v (), X) 2 < almost surely for all V. Proof: A dscusson smlar to the one for obtanng (5) mples that, almost surely, for all 0, [ m ] m E d (x ( + 1), X) 2 F d (x j (), X) 2 1 m d (v (), X) 2 4c j=1 (18) ( 1 + mm 2c, 1 ) α 2 4. Proposton 2.2 and (C2) lead to =0 m d(v (), X) 2 < almost surely;.e., =0 d(v (), X) 2 < almost surely for all V. Ths completes the proof. A dscusson smlar to the one for provng Lemma 3.3 leads to the followng (The proof of Lemma 4.3 s also descrbed n [20]). Lemma 4.3 Suppose that Assumptons 2.1, 2.2, and 2.3 hold, and defne v() := (1/m) m l=1 v l() and e () := x ( +1) v () for all 0 and for all V. Then =0 e () 2 < and =0 α v () v() < almost surely for all V. The same dscusson as for the proof of Lemma 3.4 leads to the followng (see [20] for the detals of the proof of Lemma 4.4). Lemma 4.4 Suppose that the assumptons n Theorem 4.1 are satsfed and defne z () := P X (v ()) for all V and for all 0 and defne z() := (1/m) m z () for all 0. Then the sequence ( x () x ) 0 converges almost surely for all V and for all x X, and lm nf f( z()) = f almost surely. Proof: Choose x X arbtrarly. It can be observed that (6) holds for Algorthm 4.1 because the defntons of z() and v() ( 0) n the proof of Theorem 3.1 are the same as the defntons of z() and v() ( 0) n Lemma 4.4. A dscusson smlar to the one for provng (7) means that, almost surely for all 0, for all 16

17 τ > 0, and for all η (0, 1), [ m ] E x ( + 1) x 2 F m x () x Mα 2α (f ( z ()) f ) + τ + η 1 c m v () v () m d (v (), X) 2 m d (v (), X) 2 + mm (τ, η) α 2, (19) whch, together wth τ := 1/(2c), η := 1/4, and τ + (η 1)/c = 1/(4c), mples that, almost surely, for all 0, [ m ] E x ( + 1) x 2 F m x () x 2 2α (f ( z ()) f ) + 4 Mα m v () v () + mm ( 1 2c, 1 ) α 2 4, where M(1/(2c), 1/4) < holds from (A3). Therefore, Proposton 2.2, (C2), and Lemma 4.3 ensure that ( m x () x 2 ) 0 converges almost surely;.e., ( x () x ) 0 converges almost surely for all V. Moreover, snce z() X mples f( z()) f 0 ( 0), there s also another fndng: almost surely (20) α (f ( z ()) f ) <. (21) =0 Hence, a dscusson smlar to the one for provng Lemma 3.4 leads to lm nf f( z()) = f almost surely. Ths completes the proof. Theorem 4.1 can be proven by referrng to the proof of Theorem 3.1. Proof: By referrng to the proof of Theorem 3.1, the almost sure convergence of (x ()) 0 ( V ) and Lemma 4.2 lead to the concluson that (v ()) 0, (z ()) 0 ( V ), ( v()) 0, and ( z()) 0 converge almost surely. Moreover, Lemma 4.4 and the contnuty of f ensure that ( v()) 0 and ( z()) 0 converge almost surely to x X. By referrng to the proof of Theorem 3.1, Lemma 4.3, (C1), and the trangle nequalty guarantee that lm v () x = 0 almost surely for all V. Snce Lemma 4.3 ensures that, for all V, lm e () 2 = lm x ( + 1) v () 2 = 0 almost surely, (x ()) 0 ( V ) converges almost surely to x X. Ths completes the proof Convergence rate analyss for Algorthm 4.1 The followng proposton s proven by referrng to the dscusson n Subsecton 4.1. Proposton 4.1 Suppose that the assumptons n Theorem 4.1 hold, that x X s a soluton to Problem 2.1, and that (x ()) 0 ( V ) s the sequence generated by Algorthm 4.1. Then there exst β (j) > 0 (j = 1, 2, 3, 4) such that, almost surely, 17

18 for all 0, [ m ] E d (x ( + 1), X) 2 F [ m ] E x ( + 1) x 2 F where γ (1) γ (4) m d (x (), X) 2 β (1) γ (1) + β (2) γ (2), m x () x 2 := m d(v (), X) 2, γ (2) := α 2, γ(3) j=1,4 β (j) γ (j) + j=2,3 β (j) γ (j), := α m v () v(), and := α (f( z()) f ) ( 0) satsfy =0 γ(j) < (j = 1, 2, 3, 4). Proof: From Lemma 4.2 and (C2), γ (1) ( 0) satsfy =0 γ(1) := m d(v (), X) 2 and γ (2) := α 2 < almost surely and =0 γ(2) <. Lemma 4.3 and (21) ensure that γ (3) := α m v () v() and γ (4) := α (f( z()) f ) ( 0) also satsfy =0 γ(j) < almost surely for j = 3, 4. Set τ := 1/(2c) and η := 1/4, β (1) := (τ + (η 1)/c) = 1/(4c), β (2) := mm(τ, η), β (3) := 4 M, and β (4) := 2, where β (2), β (3) < hold from (A3) and M := max V M. Accordngly, (18) and (19) ensure that Proposton 4.1 holds. A dscusson smlar to the one for obtanng (11), Proposton 4.1, and Theorem 4.1 ndcates that (x ()) 0 ( V ) n Algorthm 4.1 converges almost surely to a soluton to Problem 2.1 for the followng convergence rates: for all 0, [ m ] m E d (x ( + 1), X) 2 F d (x (), X) 2 + O ( α) 2, [ m ] E x ( + 1) x 2 F m x () x 2 + O (α ). (22) Under the condton that x ( + 1) x x () x for all V and for a large enough, (20) mples that, almost surely, f ( z()) f + 2 M m v () v() + O(α ), (23) where ( v () v() ) 0 s the almost sure convergent sequence (see proof of Theorem 4.1). It can be observed from (11), (12), (22), and (23) that Algorthms 3.1 and 4.1 have almost the same convergence rate. Secton 5 gves numercal examples showng that Algorthms 3.1 and 4.1 have almost the same convergence rate when they have the same step sze. Next, let us consder the case n whch f ( V ) s convex and dfferentable and f ( V ) satsfes the Lpschtz contnuty condton [24, Assumpton 1 c)]. Then Algorthm 4.1 concdes wth the frst dstrbuted random projecton algorthm [24, (2a) and (2b)] (see (17)) defned as follows for all 0 and for all V : x ( + 1) := P X Ω () (v () α f (v ())), (24) where v () ( V, 0) s as n (3) and (α ) 0 satsfes (C1) and (C2). The proof of Lemma 5 and (18) n [24] ndcate that algorthm (24), under (A3), almost 18

19 surely satsfes the followng convergence rate condtons for a large enough. [ m ] E d (x ( + 1), X) 2 F ( 1 + O ( α 2 )) m d (x (), X) 2 + O ( α) 2, [ m ] E x ( + 1) x 2 F ( 1 + O ( α 2 )) m x () x 2 + O (α ). (25) Proposton 4.1 mples that, f the stronger assumpton (A3) s satsfed, Algorthm (24) satsfes (22) that are better propertes for convergence rate than (25). Let us compare the convergence analyss used here (Theorem 4.1 and Proposton 4.1) wth that of the ncremental subgradent methods n [27]. The followng ncremental subgradent method wth randomzaton can optmze the sum of nonsmooth, convex functons V f over a nonempty, closed convex set X [27, (3.1)]: g ω f ω (x()), x( + 1) := P X (x() α g ω ), (26) where (ω ) 0 V s a sequence of random varables. Algorthm (26) s an ncremental optmzaton algorthm that uses the metrc projecton onto the whole constrant set X and the subdfferental of one functon selected randomly from {f } V whle Algorthm 4.1 s a dstrbuted optmzaton algorthm that uses the subdfferental of each user s objectve functon and the metrc projecton onto one closed convex set selected randomly from each user s constrant sets. Under certan assumptons, Algorthm (26) wth (α ) 0 satsfyng (C1) and (C2) converges almost surely to some mnmzer of V f over X [27, Proposton 3.2]. Theorem 4.1 wth condtons (C1) and (C2) ndcates the almost sure convergence of Algorthm 4.1 to a random pont n X. Proposton 3.1 n [27] ndcates that, under certan assumptons, Algorthm (26) wth α := α > 0 ( 0) means that the followng relatonshp s almost surely satsfed. nf f (x()) 0 V V f (x ) + αmc2, (27) 2 where x X represents the mnmzer of V f over X and c R s a constant. From (27) and Proposton 4.1, the convergence rates of Algorthms 4.1 and (26) depend on the number of elements n V and the step sze (α ) 0, as seen n Algorthms 3.1 and (14) (see Subsecton 3.2). 5. Numercal evaluaton Let us apply Algorthms 3.1 and 4.1 to Problem 2.1 wth f : R d R, and X R d ( V := {1, 2,..., m}) defned by f (x) := d a j x j b j and X := j=1 d j=1 {x R d : x c j r j }, 19

20 where a := (a j ) d j=1, b := (b j ) d j=1, r := (r j ) d j=1 Rd + ( V ), and c j R d ( V, j = 1, 2,..., d), meanng that ths s the problem of mnmzng the sum of the weghted L 1 -norms f(x) = m f (x) over the ntersecton of closed balls X = m X. The metrc projecton onto X j := {x R d : x c j r j } ( V, j = 1, 2,..., d) can be computed wthn a fnte number of arthmetc operatons [1, Chapter 28]. Functon f ( V ) satsfes the Lpschtz contnuty condton. Hence, Assumptons 2.1 and 2.5 hold. The set X Ω () ( V, 0) used n the numercal evaluaton was chosen randomly from the sets X j so as to satsfy Assumpton 2.4. The subdfferental f and the proxmty operator prox αf ( V, α > 0) can be calculated explctly [13, Lemma 10, (30), (35)] Fgure 1.: Networ model used n numercal evaluaton, where users wthn each mared area can communcate wth each other In the evaluaton, a networ wth 48 users (.e., m := 48) and 16 subnetwors was used, as llustrated n Fgure 1. It was assumed that users wthn each mared area (.e., all users n each subnetwor) could communcate wth each other. For example, user 2 could communcate wth users 1, 3, and 4 (.e., N 2 () = {1, 2, 3, 4}) whle user 1 could communcate wth not only users 2, 3, and 4 but also users 46, 47, and 48 (.e., N 1 () = {1, 2, 3, 4, 46, 47, 48}). The weghted parameters w j () ( V, j N ()) were set to satsfy Assumpton 2.3 (e.g., user 2 had w 2,j (j N 2 ()) such that w 2,j () = w j,2 () = 3/8 (j = 2, 3) and w 2,j () = w j,2 () = 1/8 (j = 1, 4), and user 1 had w 1,j (j N 1 ()) such that w 1,1 () = 2/8 and w 1,j () = w j,1 = 1/8 (j = 2, 3, 4, 46, 47, 48)). The pont v () ( V ) defned n (3) (e.g., v 2 () = (3/8)(x 2 () + x 3 ()) + (1/8)(x 1 () + x 4 ())) was computed by passng along x j () (j N ()) n a prearranged cyclc order. The computer used n the evaluaton had two Intel Xeon E v3 (2.60 GHz) CPUs. Each had 8 physcal cores and 16 threads;.e., the total number of cores was 16, and the total number of threads was 32. The computer had 64 GB DDR4 memory and ran the Ubuntu (Lnux ernel: generc, 64 bt) operatng system. The evaluaton programs were run n Python 3.4.0; Numpy was used to compute the lnear algebra operatons. We set d := 100 and m := 48 and used a := (a j ) d j=1 (0, 1]d, b := (b j ) d j=1 [0, 1)d, r := (r j ) d j=1 [3, 4)d ( V ), and c j [ (3/4)d, (3/4)d) d ( V, j = 1, 2,..., d) to satsfy X generated randomly by numpy.random. One hundred samplngs were performed, each startng from dfferent random ntal ponts x (0) ( V ) n the range [ 2, 2] d, and the results were averaged. From the dscusson n Subsectons 3.2 and 4.2, t can be expected that Algorthms 3.1 and 4.1 wth small step szes converge qucly. To see how the step sze affects ther convergence rates, we compared ther rates for α = 1/(+1) wth those for α = 10 3 /( + 1). Step sze α = 1/( + 1) s the smplest sequence satsfyng (C1) and (C2) n Assumpton 2.3. The selecton of step sze α = 10 3 /( + 1) was 20

21 based on the numercal results n [16, 22], whch presented fxed pont optmzaton algorthms that mnmze the sum of smooth, convex functons over the ntersecton of smple, closed convex sets. The prevous results [16, 22] ndcated that fxed pont optmzaton algorthms wth a small step sze such as α = 10 3 /( + 1) converge more qucly than ones wth the standard step sze α = 1/( + 1). Accordngly, the numercal evaluaton descrbed here used step szes α = 1/( + 1) and 10 3 /( + 1) and nvestgated whch one resulted n qucer convergence for Algorthms 3.1 and 4.1. Let us defne two performance measures for each V and for all 0, D () := x () d j=1 P X j (x ()) and F () := f (x ()), (28) and observe the behavors of D () and F () for Algorthms 3.1 and 4.1 wth α = 1/( + 1) and 10 3 /( + 1). If (D ()) 0 ( V ) converges to 0, (x ()) 0 converges to a fxed pont of d j=1 P X,.e., to a pont n X j = d j=1 Xj [1, Corollary 4.37]. Frst, let us observe the behavors of (x 1 ()) 1000 =0 calculated for user 1, whch belongs to two subnetwors. Fgures 2 and 3 show that Algorthm 3.1 converged more qucly to a pont n X 1 wth α = 10 3 /( + 1) than wth α = 1/( + 1). From Fgures 2 and 3, t can be seen that Algorthm 3.1 wth α = 10 3 /( + 1) optmzes f 1 over X 1 n the early stages. (a) D 1 () vs. no. of teratons (b) D 1 () vs. elapsed tme Fgure 2.: Behavors of D 1 () for Algorthm 3.1 wth α = 1/(+1) and 10 3 /(+1) (a) F 1 () vs. no. of teratons (b) F 1 () vs. elapsed tme Fgure 3.: Behavors of F 1 () for Algorthm 3.1 wth α = 1/(+1) and 10 3 /(+1) 21

General viscosity iterative method for a sequence of quasi-nonexpansive mappings

General viscosity iterative method for a sequence of quasi-nonexpansive mappings Avalable onlne at www.tjnsa.com J. Nonlnear Sc. Appl. 9 (2016), 5672 5682 Research Artcle General vscosty teratve method for a sequence of quas-nonexpansve mappngs Cuje Zhang, Ynan Wang College of Scence,

More information

Feature Selection: Part 1

Feature Selection: Part 1 CSE 546: Machne Learnng Lecture 5 Feature Selecton: Part 1 Instructor: Sham Kakade 1 Regresson n the hgh dmensonal settng How do we learn when the number of features d s greater than the sample sze n?

More information

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space. Lnear, affne, and convex sets and hulls In the sequel, unless otherwse specfed, X wll denote a real vector space. Lnes and segments. Gven two ponts x, y X, we defne xy = {x + t(y x) : t R} = {(1 t)x +

More information

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg prnceton unv. F 17 cos 521: Advanced Algorthm Desgn Lecture 7: LP Dualty Lecturer: Matt Wenberg Scrbe: LP Dualty s an extremely useful tool for analyzng structural propertes of lnear programs. Whle there

More information

More metrics on cartesian products

More metrics on cartesian products More metrcs on cartesan products If (X, d ) are metrc spaces for 1 n, then n Secton II4 of the lecture notes we defned three metrcs on X whose underlyng topologes are the product topology The purpose of

More information

Lecture Notes on Linear Regression

Lecture Notes on Linear Regression Lecture Notes on Lnear Regresson Feng L fl@sdueducn Shandong Unversty, Chna Lnear Regresson Problem In regresson problem, we am at predct a contnuous target value gven an nput feature vector We assume

More information

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011 Stanford Unversty CS359G: Graph Parttonng and Expanders Handout 4 Luca Trevsan January 3, 0 Lecture 4 In whch we prove the dffcult drecton of Cheeger s nequalty. As n the past lectures, consder an undrected

More information

The Order Relation and Trace Inequalities for. Hermitian Operators

The Order Relation and Trace Inequalities for. Hermitian Operators Internatonal Mathematcal Forum, Vol 3, 08, no, 507-57 HIKARI Ltd, wwwm-hkarcom https://doorg/0988/mf088055 The Order Relaton and Trace Inequaltes for Hermtan Operators Y Huang School of Informaton Scence

More information

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem.

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem. prnceton u. sp 02 cos 598B: algorthms and complexty Lecture 20: Lft and Project, SDP Dualty Lecturer: Sanjeev Arora Scrbe:Yury Makarychev Today we wll study the Lft and Project method. Then we wll prove

More information

Perron Vectors of an Irreducible Nonnegative Interval Matrix

Perron Vectors of an Irreducible Nonnegative Interval Matrix Perron Vectors of an Irreducble Nonnegatve Interval Matrx Jr Rohn August 4 2005 Abstract As s well known an rreducble nonnegatve matrx possesses a unquely determned Perron vector. As the man result of

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 12 10/21/2013. Martingale Concentration Inequalities and Applications

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 12 10/21/2013. Martingale Concentration Inequalities and Applications MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.65/15.070J Fall 013 Lecture 1 10/1/013 Martngale Concentraton Inequaltes and Applcatons Content. 1. Exponental concentraton for martngales wth bounded ncrements.

More information

MMA and GCMMA two methods for nonlinear optimization

MMA and GCMMA two methods for nonlinear optimization MMA and GCMMA two methods for nonlnear optmzaton Krster Svanberg Optmzaton and Systems Theory, KTH, Stockholm, Sweden. krlle@math.kth.se Ths note descrbes the algorthms used n the author s 2007 mplementatons

More information

10-801: Advanced Optimization and Randomized Methods Lecture 2: Convex functions (Jan 15, 2014)

10-801: Advanced Optimization and Randomized Methods Lecture 2: Convex functions (Jan 15, 2014) 0-80: Advanced Optmzaton and Randomzed Methods Lecture : Convex functons (Jan 5, 04) Lecturer: Suvrt Sra Addr: Carnege Mellon Unversty, Sprng 04 Scrbes: Avnava Dubey, Ahmed Hefny Dsclamer: These notes

More information

Lecture 20: November 7

Lecture 20: November 7 0-725/36-725: Convex Optmzaton Fall 205 Lecturer: Ryan Tbshran Lecture 20: November 7 Scrbes: Varsha Chnnaobreddy, Joon Sk Km, Lngyao Zhang Note: LaTeX template courtesy of UC Berkeley EECS dept. Dsclamer:

More information

Notes on Frequency Estimation in Data Streams

Notes on Frequency Estimation in Data Streams Notes on Frequency Estmaton n Data Streams In (one of) the data streamng model(s), the data s a sequence of arrvals a 1, a 2,..., a m of the form a j = (, v) where s the dentty of the tem and belongs to

More information

Maximizing the number of nonnegative subsets

Maximizing the number of nonnegative subsets Maxmzng the number of nonnegatve subsets Noga Alon Hao Huang December 1, 213 Abstract Gven a set of n real numbers, f the sum of elements of every subset of sze larger than k s negatve, what s the maxmum

More information

COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS

COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS Avalable onlne at http://sck.org J. Math. Comput. Sc. 3 (3), No., 6-3 ISSN: 97-537 COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS

More information

The Minimum Universal Cost Flow in an Infeasible Flow Network

The Minimum Universal Cost Flow in an Infeasible Flow Network Journal of Scences, Islamc Republc of Iran 17(2): 175-180 (2006) Unversty of Tehran, ISSN 1016-1104 http://jscencesutacr The Mnmum Unversal Cost Flow n an Infeasble Flow Network H Saleh Fathabad * M Bagheran

More information

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur Module 3 LOSSY IMAGE COMPRESSION SYSTEMS Verson ECE IIT, Kharagpur Lesson 6 Theory of Quantzaton Verson ECE IIT, Kharagpur Instructonal Objectves At the end of ths lesson, the students should be able to:

More information

Solutions to exam in SF1811 Optimization, Jan 14, 2015

Solutions to exam in SF1811 Optimization, Jan 14, 2015 Solutons to exam n SF8 Optmzaton, Jan 4, 25 3 3 O------O -4 \ / \ / The network: \/ where all lnks go from left to rght. /\ / \ / \ 6 O------O -5 2 4.(a) Let x = ( x 3, x 4, x 23, x 24 ) T, where the varable

More information

Random Projection Algorithms for Convex Set Intersection Problems

Random Projection Algorithms for Convex Set Intersection Problems Random Projecton Algorthms for Convex Set Intersecton Problems A. Nedć Department of Industral and Enterprse Systems Engneerng Unversty of Illnos, Urbana, IL 61801 angela@llnos.edu Abstract The focus of

More information

Generalized Linear Methods

Generalized Linear Methods Generalzed Lnear Methods 1 Introducton In the Ensemble Methods the general dea s that usng a combnaton of several weak learner one could make a better learner. More formally, assume that we have a set

More information

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0 MODULE 2 Topcs: Lnear ndependence, bass and dmenson We have seen that f n a set of vectors one vector s a lnear combnaton of the remanng vectors n the set then the span of the set s unchanged f that vector

More information

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009 College of Computer & Informaton Scence Fall 2009 Northeastern Unversty 20 October 2009 CS7880: Algorthmc Power Tools Scrbe: Jan Wen and Laura Poplawsk Lecture Outlne: Prmal-dual schema Network Desgn:

More information

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction ECONOMICS 5* -- NOTE (Summary) ECON 5* -- NOTE The Multple Classcal Lnear Regresson Model (CLRM): Specfcaton and Assumptons. Introducton CLRM stands for the Classcal Lnear Regresson Model. The CLRM s also

More information

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017 U.C. Berkeley CS94: Beyond Worst-Case Analyss Handout 4s Luca Trevsan September 5, 07 Summary of Lecture 4 In whch we ntroduce semdefnte programmng and apply t to Max Cut. Semdefnte Programmng Recall that

More information

Problem Set 9 Solutions

Problem Set 9 Solutions Desgn and Analyss of Algorthms May 4, 2015 Massachusetts Insttute of Technology 6.046J/18.410J Profs. Erk Demane, Srn Devadas, and Nancy Lynch Problem Set 9 Solutons Problem Set 9 Solutons Ths problem

More information

Appendix B. Criterion of Riemann-Stieltjes Integrability

Appendix B. Criterion of Riemann-Stieltjes Integrability Appendx B. Crteron of Remann-Steltes Integrablty Ths note s complementary to [R, Ch. 6] and [T, Sec. 3.5]. The man result of ths note s Theorem B.3, whch provdes the necessary and suffcent condtons for

More information

APPENDIX A Some Linear Algebra

APPENDIX A Some Linear Algebra APPENDIX A Some Lnear Algebra The collecton of m, n matrces A.1 Matrces a 1,1,..., a 1,n A = a m,1,..., a m,n wth real elements a,j s denoted by R m,n. If n = 1 then A s called a column vector. Smlarly,

More information

Finding Dense Subgraphs in G(n, 1/2)

Finding Dense Subgraphs in G(n, 1/2) Fndng Dense Subgraphs n Gn, 1/ Atsh Das Sarma 1, Amt Deshpande, and Rav Kannan 1 Georga Insttute of Technology,atsh@cc.gatech.edu Mcrosoft Research-Bangalore,amtdesh,annan@mcrosoft.com Abstract. Fndng

More information

Perfect Competition and the Nash Bargaining Solution

Perfect Competition and the Nash Bargaining Solution Perfect Competton and the Nash Barganng Soluton Renhard John Department of Economcs Unversty of Bonn Adenauerallee 24-42 53113 Bonn, Germany emal: rohn@un-bonn.de May 2005 Abstract For a lnear exchange

More information

Complete subgraphs in multipartite graphs

Complete subgraphs in multipartite graphs Complete subgraphs n multpartte graphs FLORIAN PFENDER Unverstät Rostock, Insttut für Mathematk D-18057 Rostock, Germany Floran.Pfender@un-rostock.de Abstract Turán s Theorem states that every graph G

More information

Structure and Drive Paul A. Jensen Copyright July 20, 2003

Structure and Drive Paul A. Jensen Copyright July 20, 2003 Structure and Drve Paul A. Jensen Copyrght July 20, 2003 A system s made up of several operatons wth flow passng between them. The structure of the system descrbes the flow paths from nputs to outputs.

More information

FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP

FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP C O L L O Q U I U M M A T H E M A T I C U M VOL. 80 1999 NO. 1 FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP BY FLORIAN K A I N R A T H (GRAZ) Abstract. Let H be a Krull monod wth nfnte class

More information

1 Convex Optimization

1 Convex Optimization Convex Optmzaton We wll consder convex optmzaton problems. Namely, mnmzaton problems where the objectve s convex (we assume no constrants for now). Such problems often arse n machne learnng. For example,

More information

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X Statstcs 1: Probablty Theory II 37 3 EPECTATION OF SEVERAL RANDOM VARIABLES As n Probablty Theory I, the nterest n most stuatons les not on the actual dstrbuton of a random vector, but rather on a number

More information

A PROBABILITY-DRIVEN SEARCH ALGORITHM FOR SOLVING MULTI-OBJECTIVE OPTIMIZATION PROBLEMS

A PROBABILITY-DRIVEN SEARCH ALGORITHM FOR SOLVING MULTI-OBJECTIVE OPTIMIZATION PROBLEMS HCMC Unversty of Pedagogy Thong Nguyen Huu et al. A PROBABILITY-DRIVEN SEARCH ALGORITHM FOR SOLVING MULTI-OBJECTIVE OPTIMIZATION PROBLEMS Thong Nguyen Huu and Hao Tran Van Department of mathematcs-nformaton,

More information

EEE 241: Linear Systems

EEE 241: Linear Systems EEE : Lnear Systems Summary #: Backpropagaton BACKPROPAGATION The perceptron rule as well as the Wdrow Hoff learnng were desgned to tran sngle layer networks. They suffer from the same dsadvantage: they

More information

A note on almost sure behavior of randomly weighted sums of φ-mixing random variables with φ-mixing weights

A note on almost sure behavior of randomly weighted sums of φ-mixing random variables with φ-mixing weights ACTA ET COMMENTATIONES UNIVERSITATIS TARTUENSIS DE MATHEMATICA Volume 7, Number 2, December 203 Avalable onlne at http://acutm.math.ut.ee A note on almost sure behavor of randomly weghted sums of φ-mxng

More information

ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EQUATION

ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EQUATION Advanced Mathematcal Models & Applcatons Vol.3, No.3, 2018, pp.215-222 ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EUATION

More information

A 2D Bounded Linear Program (H,c) 2D Linear Programming

A 2D Bounded Linear Program (H,c) 2D Linear Programming A 2D Bounded Lnear Program (H,c) h 3 v h 8 h 5 c h 4 h h 6 h 7 h 2 2D Lnear Programmng C s a polygonal regon, the ntersecton of n halfplanes. (H, c) s nfeasble, as C s empty. Feasble regon C s unbounded

More information

Supplement: Proofs and Technical Details for The Solution Path of the Generalized Lasso

Supplement: Proofs and Technical Details for The Solution Path of the Generalized Lasso Supplement: Proofs and Techncal Detals for The Soluton Path of the Generalzed Lasso Ryan J. Tbshran Jonathan Taylor In ths document we gve supplementary detals to the paper The Soluton Path of the Generalzed

More information

Math 217 Fall 2013 Homework 2 Solutions

Math 217 Fall 2013 Homework 2 Solutions Math 17 Fall 013 Homework Solutons Due Thursday Sept. 6, 013 5pm Ths homework conssts of 6 problems of 5 ponts each. The total s 30. You need to fully justfy your answer prove that your functon ndeed has

More information

Lecture 4: November 17, Part 1 Single Buffer Management

Lecture 4: November 17, Part 1 Single Buffer Management Lecturer: Ad Rosén Algorthms for the anagement of Networs Fall 2003-2004 Lecture 4: November 7, 2003 Scrbe: Guy Grebla Part Sngle Buffer anagement In the prevous lecture we taled about the Combned Input

More information

Matrix Approximation via Sampling, Subspace Embedding. 1 Solving Linear Systems Using SVD

Matrix Approximation via Sampling, Subspace Embedding. 1 Solving Linear Systems Using SVD Matrx Approxmaton va Samplng, Subspace Embeddng Lecturer: Anup Rao Scrbe: Rashth Sharma, Peng Zhang 0/01/016 1 Solvng Lnear Systems Usng SVD Two applcatons of SVD have been covered so far. Today we loo

More information

The lower and upper bounds on Perron root of nonnegative irreducible matrices

The lower and upper bounds on Perron root of nonnegative irreducible matrices Journal of Computatonal Appled Mathematcs 217 (2008) 259 267 wwwelsevercom/locate/cam The lower upper bounds on Perron root of nonnegatve rreducble matrces Guang-Xn Huang a,, Feng Yn b,keguo a a College

More information

Economics 101. Lecture 4 - Equilibrium and Efficiency

Economics 101. Lecture 4 - Equilibrium and Efficiency Economcs 0 Lecture 4 - Equlbrum and Effcency Intro As dscussed n the prevous lecture, we wll now move from an envronment where we looed at consumers mang decsons n solaton to analyzng economes full of

More information

Some modelling aspects for the Matlab implementation of MMA

Some modelling aspects for the Matlab implementation of MMA Some modellng aspects for the Matlab mplementaton of MMA Krster Svanberg krlle@math.kth.se Optmzaton and Systems Theory Department of Mathematcs KTH, SE 10044 Stockholm September 2004 1. Consdered optmzaton

More information

NP-Completeness : Proofs

NP-Completeness : Proofs NP-Completeness : Proofs Proof Methods A method to show a decson problem Π NP-complete s as follows. (1) Show Π NP. (2) Choose an NP-complete problem Π. (3) Show Π Π. A method to show an optmzaton problem

More information

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems Numercal Analyss by Dr. Anta Pal Assstant Professor Department of Mathematcs Natonal Insttute of Technology Durgapur Durgapur-713209 emal: anta.bue@gmal.com 1 . Chapter 5 Soluton of System of Lnear Equatons

More information

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification E395 - Pattern Recognton Solutons to Introducton to Pattern Recognton, Chapter : Bayesan pattern classfcaton Preface Ths document s a soluton manual for selected exercses from Introducton to Pattern Recognton

More information

On the Multicriteria Integer Network Flow Problem

On the Multicriteria Integer Network Flow Problem BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 5, No 2 Sofa 2005 On the Multcrtera Integer Network Flow Problem Vassl Vasslev, Marana Nkolova, Maryana Vassleva Insttute of

More information

Inexact Alternating Minimization Algorithm for Distributed Optimization with an Application to Distributed MPC

Inexact Alternating Minimization Algorithm for Distributed Optimization with an Application to Distributed MPC Inexact Alternatng Mnmzaton Algorthm for Dstrbuted Optmzaton wth an Applcaton to Dstrbuted MPC Ye Pu, Coln N. Jones and Melane N. Zelnger arxv:608.0043v [math.oc] Aug 206 Abstract In ths paper, we propose

More information

find (x): given element x, return the canonical element of the set containing x;

find (x): given element x, return the canonical element of the set containing x; COS 43 Sprng, 009 Dsjont Set Unon Problem: Mantan a collecton of dsjont sets. Two operatons: fnd the set contanng a gven element; unte two sets nto one (destructvely). Approach: Canoncal element method:

More information

18.1 Introduction and Recap

18.1 Introduction and Recap CS787: Advanced Algorthms Scrbe: Pryananda Shenoy and Shjn Kong Lecturer: Shuch Chawla Topc: Streamng Algorthmscontnued) Date: 0/26/2007 We contnue talng about streamng algorthms n ths lecture, ncludng

More information

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix Lectures - Week 4 Matrx norms, Condtonng, Vector Spaces, Lnear Independence, Spannng sets and Bass, Null space and Range of a Matrx Matrx Norms Now we turn to assocatng a number to each matrx. We could

More information

REAL ANALYSIS I HOMEWORK 1

REAL ANALYSIS I HOMEWORK 1 REAL ANALYSIS I HOMEWORK CİHAN BAHRAN The questons are from Tao s text. Exercse 0.0.. If (x α ) α A s a collecton of numbers x α [0, + ] such that x α

More information

Edge Isoperimetric Inequalities

Edge Isoperimetric Inequalities November 7, 2005 Ross M. Rchardson Edge Isopermetrc Inequaltes 1 Four Questons Recall that n the last lecture we looked at the problem of sopermetrc nequaltes n the hypercube, Q n. Our noton of boundary

More information

Global Sensitivity. Tuesday 20 th February, 2018

Global Sensitivity. Tuesday 20 th February, 2018 Global Senstvty Tuesday 2 th February, 28 ) Local Senstvty Most senstvty analyses [] are based on local estmates of senstvty, typcally by expandng the response n a Taylor seres about some specfc values

More information

CSCI B609: Foundations of Data Science

CSCI B609: Foundations of Data Science CSCI B609: Foundatons of Data Scence Lecture 13/14: Gradent Descent, Boostng and Learnng from Experts Sldes at http://grgory.us/data-scence-class.html Grgory Yaroslavtsev http://grgory.us Constraned Convex

More information

Errors for Linear Systems

Errors for Linear Systems Errors for Lnear Systems When we solve a lnear system Ax b we often do not know A and b exactly, but have only approxmatons  and ˆb avalable. Then the best thng we can do s to solve ˆx ˆb exactly whch

More information

On the Interval Zoro Symmetric Single-step Procedure for Simultaneous Finding of Polynomial Zeros

On the Interval Zoro Symmetric Single-step Procedure for Simultaneous Finding of Polynomial Zeros Appled Mathematcal Scences, Vol. 5, 2011, no. 75, 3693-3706 On the Interval Zoro Symmetrc Sngle-step Procedure for Smultaneous Fndng of Polynomal Zeros S. F. M. Rusl, M. Mons, M. A. Hassan and W. J. Leong

More information

COS 521: Advanced Algorithms Game Theory and Linear Programming

COS 521: Advanced Algorithms Game Theory and Linear Programming COS 521: Advanced Algorthms Game Theory and Lnear Programmng Moses Charkar February 27, 2013 In these notes, we ntroduce some basc concepts n game theory and lnear programmng (LP). We show a connecton

More information

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens THE CHINESE REMAINDER THEOREM KEITH CONRAD We should thank the Chnese for ther wonderful remander theorem. Glenn Stevens 1. Introducton The Chnese remander theorem says we can unquely solve any par of

More information

Singular Value Decomposition: Theory and Applications

Singular Value Decomposition: Theory and Applications Sngular Value Decomposton: Theory and Applcatons Danel Khashab Sprng 2015 Last Update: March 2, 2015 1 Introducton A = UDV where columns of U and V are orthonormal and matrx D s dagonal wth postve real

More information

Convexity preserving interpolation by splines of arbitrary degree

Convexity preserving interpolation by splines of arbitrary degree Computer Scence Journal of Moldova, vol.18, no.1(52), 2010 Convexty preservng nterpolaton by splnes of arbtrary degree Igor Verlan Abstract In the present paper an algorthm of C 2 nterpolaton of dscrete

More information

VQ widely used in coding speech, image, and video

VQ widely used in coding speech, image, and video at Scalar quantzers are specal cases of vector quantzers (VQ): they are constraned to look at one sample at a tme (memoryless) VQ does not have such constrant better RD perfomance expected Source codng

More information

MATH 829: Introduction to Data Mining and Analysis The EM algorithm (part 2)

MATH 829: Introduction to Data Mining and Analysis The EM algorithm (part 2) 1/16 MATH 829: Introducton to Data Mnng and Analyss The EM algorthm (part 2) Domnque Gullot Departments of Mathematcal Scences Unversty of Delaware Aprl 20, 2016 Recall 2/16 We are gven ndependent observatons

More information

Affine transformations and convexity

Affine transformations and convexity Affne transformatons and convexty The purpose of ths document s to prove some basc propertes of affne transformatons nvolvng convex sets. Here are a few onlne references for background nformaton: http://math.ucr.edu/

More information

Vector Norms. Chapter 7 Iterative Techniques in Matrix Algebra. Cauchy-Bunyakovsky-Schwarz Inequality for Sums. Distances. Convergence.

Vector Norms. Chapter 7 Iterative Techniques in Matrix Algebra. Cauchy-Bunyakovsky-Schwarz Inequality for Sums. Distances. Convergence. Vector Norms Chapter 7 Iteratve Technques n Matrx Algebra Per-Olof Persson persson@berkeley.edu Department of Mathematcs Unversty of Calforna, Berkeley Math 128B Numercal Analyss Defnton A vector norm

More information

6.854J / J Advanced Algorithms Fall 2008

6.854J / J Advanced Algorithms Fall 2008 MIT OpenCourseWare http://ocw.mt.edu 6.854J / 18.415J Advanced Algorthms Fall 2008 For nformaton about ctng these materals or our Terms of Use, vst: http://ocw.mt.edu/terms. 18.415/6.854 Advanced Algorthms

More information

Exercise Solutions to Real Analysis

Exercise Solutions to Real Analysis xercse Solutons to Real Analyss Note: References refer to H. L. Royden, Real Analyss xersze 1. Gven any set A any ɛ > 0, there s an open set O such that A O m O m A + ɛ. Soluton 1. If m A =, then there

More information

Infinitely Split Nash Equilibrium Problems in Repeated Games

Infinitely Split Nash Equilibrium Problems in Repeated Games Infntely Splt ash Equlbrum Problems n Repeated Games Jnlu L Department of Mathematcs Shawnee State Unversty Portsmouth, Oho 4566 USA Abstract In ths paper, we ntroduce the concept of nfntely splt ash equlbrum

More information

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS SECTION 8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS 493 8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS All the vector spaces you have studed thus far n the text are real vector spaces because the scalars

More information

Grover s Algorithm + Quantum Zeno Effect + Vaidman

Grover s Algorithm + Quantum Zeno Effect + Vaidman Grover s Algorthm + Quantum Zeno Effect + Vadman CS 294-2 Bomb 10/12/04 Fall 2004 Lecture 11 Grover s algorthm Recall that Grover s algorthm for searchng over a space of sze wors as follows: consder the

More information

Markov Chain Monte Carlo Lecture 6

Markov Chain Monte Carlo Lecture 6 where (x 1,..., x N ) X N, N s called the populaton sze, f(x) f (x) for at least one {1, 2,..., N}, and those dfferent from f(x) are called the tral dstrbutons n terms of mportance samplng. Dfferent ways

More information

n ). This is tight for all admissible values of t, k and n. k t + + n t

n ). This is tight for all admissible values of t, k and n. k t + + n t MAXIMIZING THE NUMBER OF NONNEGATIVE SUBSETS NOGA ALON, HAROUT AYDINIAN, AND HAO HUANG Abstract. Gven a set of n real numbers, f the sum of elements of every subset of sze larger than k s negatve, what

More information

SELECTED SOLUTIONS, SECTION (Weak duality) Prove that the primal and dual values p and d defined by equations (4.3.2) and (4.3.3) satisfy p d.

SELECTED SOLUTIONS, SECTION (Weak duality) Prove that the primal and dual values p and d defined by equations (4.3.2) and (4.3.3) satisfy p d. SELECTED SOLUTIONS, SECTION 4.3 1. Weak dualty Prove that the prmal and dual values p and d defned by equatons 4.3. and 4.3.3 satsfy p d. We consder an optmzaton problem of the form The Lagrangan for ths

More information

STEINHAUS PROPERTY IN BANACH LATTICES

STEINHAUS PROPERTY IN BANACH LATTICES DEPARTMENT OF MATHEMATICS TECHNICAL REPORT STEINHAUS PROPERTY IN BANACH LATTICES DAMIAN KUBIAK AND DAVID TIDWELL SPRING 2015 No. 2015-1 TENNESSEE TECHNOLOGICAL UNIVERSITY Cookevlle, TN 38505 STEINHAUS

More information

On a direct solver for linear least squares problems

On a direct solver for linear least squares problems ISSN 2066-6594 Ann. Acad. Rom. Sc. Ser. Math. Appl. Vol. 8, No. 2/2016 On a drect solver for lnear least squares problems Constantn Popa Abstract The Null Space (NS) algorthm s a drect solver for lnear

More information

SUCCESSIVE MINIMA AND LATTICE POINTS (AFTER HENK, GILLET AND SOULÉ) M(B) := # ( B Z N)

SUCCESSIVE MINIMA AND LATTICE POINTS (AFTER HENK, GILLET AND SOULÉ) M(B) := # ( B Z N) SUCCESSIVE MINIMA AND LATTICE POINTS (AFTER HENK, GILLET AND SOULÉ) S.BOUCKSOM Abstract. The goal of ths note s to present a remarably smple proof, due to Hen, of a result prevously obtaned by Gllet-Soulé,

More information

princeton univ. F 13 cos 521: Advanced Algorithm Design Lecture 3: Large deviations bounds and applications Lecturer: Sanjeev Arora

princeton univ. F 13 cos 521: Advanced Algorithm Design Lecture 3: Large deviations bounds and applications Lecturer: Sanjeev Arora prnceton unv. F 13 cos 521: Advanced Algorthm Desgn Lecture 3: Large devatons bounds and applcatons Lecturer: Sanjeev Arora Scrbe: Today s topc s devaton bounds: what s the probablty that a random varable

More information

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud Resource Allocaton wth a Budget Constrant for Computng Independent Tasks n the Cloud Wemng Sh and Bo Hong School of Electrcal and Computer Engneerng Georga Insttute of Technology, USA 2nd IEEE Internatonal

More information

Distributed Non-Autonomous Power Control through Distributed Convex Optimization

Distributed Non-Autonomous Power Control through Distributed Convex Optimization Dstrbuted Non-Autonomous Power Control through Dstrbuted Convex Optmzaton S. Sundhar Ram and V. V. Veeravall ECE Department and Coordnated Scence Lab Unversty of Illnos at Urbana-Champagn Emal: {ssrnv5,vvv}@llnos.edu

More information

Appendix B: Resampling Algorithms

Appendix B: Resampling Algorithms 407 Appendx B: Resamplng Algorthms A common problem of all partcle flters s the degeneracy of weghts, whch conssts of the unbounded ncrease of the varance of the mportance weghts ω [ ] of the partcles

More information

FREQUENCY DISTRIBUTIONS Page 1 of The idea of a frequency distribution for sets of observations will be introduced,

FREQUENCY DISTRIBUTIONS Page 1 of The idea of a frequency distribution for sets of observations will be introduced, FREQUENCY DISTRIBUTIONS Page 1 of 6 I. Introducton 1. The dea of a frequency dstrbuton for sets of observatons wll be ntroduced, together wth some of the mechancs for constructng dstrbutons of data. Then

More information

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016 U.C. Berkeley CS94: Spectral Methods and Expanders Handout 8 Luca Trevsan February 7, 06 Lecture 8: Spectral Algorthms Wrap-up In whch we talk about even more generalzatons of Cheeger s nequaltes, and

More information

MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS

MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS These are nformal notes whch cover some of the materal whch s not n the course book. The man purpose s to gve a number of nontrval examples

More information

Week 2. This week, we covered operations on sets and cardinality.

Week 2. This week, we covered operations on sets and cardinality. Week 2 Ths week, we covered operatons on sets and cardnalty. Defnton 0.1 (Correspondence). A correspondence between two sets A and B s a set S contaned n A B = {(a, b) a A, b B}. A correspondence from

More information

Linear Regression Analysis: Terminology and Notation

Linear Regression Analysis: Terminology and Notation ECON 35* -- Secton : Basc Concepts of Regresson Analyss (Page ) Lnear Regresson Analyss: Termnology and Notaton Consder the generc verson of the smple (two-varable) lnear regresson model. It s represented

More information

Neural networks. Nuno Vasconcelos ECE Department, UCSD

Neural networks. Nuno Vasconcelos ECE Department, UCSD Neural networs Nuno Vasconcelos ECE Department, UCSD Classfcaton a classfcaton problem has two types of varables e.g. X - vector of observatons (features) n the world Y - state (class) of the world x X

More information

System of implicit nonconvex variationl inequality problems: A projection method approach

System of implicit nonconvex variationl inequality problems: A projection method approach Avalable onlne at www.tjnsa.com J. Nonlnear Sc. Appl. 6 (203), 70 80 Research Artcle System of mplct nonconvex varatonl nequalty problems: A projecton method approach K.R. Kazm a,, N. Ahmad b, S.H. Rzv

More information

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty Addtonal Codes usng Fnte Dfference Method Benamn Moll 1 HJB Equaton for Consumpton-Savng Problem Wthout Uncertanty Before consderng the case wth stochastc ncome n http://www.prnceton.edu/~moll/ HACTproect/HACT_Numercal_Appendx.pdf,

More information

Feb 14: Spatial analysis of data fields

Feb 14: Spatial analysis of data fields Feb 4: Spatal analyss of data felds Mappng rregularly sampled data onto a regular grd Many analyss technques for geophyscal data requre the data be located at regular ntervals n space and/or tme. hs s

More information

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal Inner Product Defnton 1 () A Eucldean space s a fnte-dmensonal vector space over the reals R, wth an nner product,. Defnton 2 (Inner Product) An nner product, on a real vector space X s a symmetrc, blnear,

More information

Communication-Efficient Algorithms for Decentralized and Stochastic Optimization

Communication-Efficient Algorithms for Decentralized and Stochastic Optimization oname manuscrpt o. (wll be nserted by the edtor) Communcaton-Effcent Algorthms for Decentralzed and Stochastc Optmzaton Guanghu Lan Soomn Lee Y Zhou the date of recept and acceptance should be nserted

More information

Google PageRank with Stochastic Matrix

Google PageRank with Stochastic Matrix Google PageRank wth Stochastc Matrx Md. Sharq, Puranjt Sanyal, Samk Mtra (M.Sc. Applcatons of Mathematcs) Dscrete Tme Markov Chan Let S be a countable set (usually S s a subset of Z or Z d or R or R d

More information

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography CSc 6974 and ECSE 6966 Math. Tech. for Vson, Graphcs and Robotcs Lecture 21, Aprl 17, 2006 Estmatng A Plane Homography Overvew We contnue wth a dscusson of the major ssues, usng estmaton of plane projectve

More information

CHAPTER III Neural Networks as Associative Memory

CHAPTER III Neural Networks as Associative Memory CHAPTER III Neural Networs as Assocatve Memory Introducton One of the prmary functons of the bran s assocatve memory. We assocate the faces wth names, letters wth sounds, or we can recognze the people

More information

Supplementary material: Margin based PU Learning. Matrix Concentration Inequalities

Supplementary material: Margin based PU Learning. Matrix Concentration Inequalities Supplementary materal: Margn based PU Learnng We gve the complete proofs of Theorem and n Secton We frst ntroduce the well-known concentraton nequalty, so the covarance estmator can be bounded Then we

More information