Asymptotic Properties of a Generalized Cross Entropy Optimization Algorithm

Size: px
Start display at page:

Download "Asymptotic Properties of a Generalized Cross Entropy Optimization Algorithm"

Transcription

1 1 Asymptotic Properties of a Generaized Cross Entropy Optimization Agorithm Zijun Wu, Michae Koonko, Institute for Appied Stochastics and Operations Research, Caustha Technica University Abstract The discrete cross entropy optimization agorithm iterativey sampes soutions according to a probabiity density on the soution space The density is adapted to the good soutions observed in the present sampe before producing the next sampe The adaptation is controed by a so-caed smoothing parameter We generaize this mode by introducing a fexibe concept of feasibiity and desirabiity into the samping process In this way, our mode covers severa other optimization procedures, in particuar the ant based agorithms The focus of this paper is on some theoretica properties of these agorithms We examine the first hitting time τ of an optima soution and give conditions on the smoothing parameter for τ to be finite with probabiity one For a simpe test case, we show that runtime can be poynomiay bounded in the probem size with a probabiity converging to 1 We then investigate the convergence of the underying density and of the samping process We show in particuar, that a constant smoothing parameter, as it is often used, makes the sampe process converge in finite time, freezing the optimization at a singe soution which need not be optima Moreover, we define a smoothing sequence that makes the density converge without freezing the sampe process and that sti guarantees the reachabiity of optima soutions in finite time This settes an open question from iterature Index Terms Evoutionary computation, cross entropy optimization, ant coony optimization, heuristic optimization, discrete optimization, mode based optimization I INTRODUCTION CROSS entropy CE) optimization is a widey used too for heuristic optimization in particuar for discrete probems, see 1 for an overview As was noted before see eg 2), it aso has much in common with ant agorithms ACOs) In this paper we use a generaized version of the CE agorithm that incudes most of the features of ACOs We concentrate on theoretica properties of the underying processes, that hod under very genera conditions and therefore appy to a types of agorithms covered by our mode The paper is inspired by 3 where for a particuar CE mode convergence properties were proven In a CE mode, we are considering a cost function f on a finite set S of soutions A soution is a tupe of fixed ength Manuscript received 28 Nov 2013; revised 16 Apr 2014 and 21 Jun 2014 The work of Z Wu was supported by China Schoarship Counci, No The authors are with Institut für Angewandte Stochastik und Operations Research, TU Caustha, Erzstr1, D Caustha-Zeerfed, Germany emai: zwu10@tu-causthade, koonko@mathtu-causthade) Copyright c) 2012 IEEE Persona use of this materia is permitted However, permission to use this materia for any other purposes must be obtained from the IEEE by sending a request to pubs-permissions@ieeeorg We use a probabiity distribution or density on S as a mode for producing good soutions This density is iterativey adapted with the goa to give an increasing weight to optima soutions Roughy, the method proceeds in two steps: Samping: We take a sampe of fixed size N from S, based on the present distribution Π t, starting with a given density Π 0 for t = 0 eg the uniform distribution) Adaptation: We evauate the soutions in the sampe with the cost function f and determine the reative frequencies of components in the best soutions We define Π t+1 by adapting Π t to these frequencies Here, a so-caed smoothing parameter ϱ t contros the reative weight of the sampe during the adaptation process In this way, Π t is expected to give increasing weight to the better part of the soution space In 3, the impact of the smoothing parameter on the convergence of the process Π t ) t 0 is investigated The authors give conditions on ϱ t under which an optima soution is samped with probabiity one after finitey many iterations reachabiity) In particuar, they found that for a constant smoothing parameter ϱ t ϱ, the distribution Π t wi converge to a one point mass convergence of density), but that one has to make ϱ arbitrariy sma to increase the probabiity of reachabiity Soutions are restricted to 0-1-tupes in 3 We extend these resuts in severa directions We aow soutions to be strings of arbitrary eements and do not require the optima soution to be unique Most importanty, we introduce a notion of feasibiity and desirabiity to restrict the set of soutions and to bias the samping The sampes in the Samping step above are actuay drawn according to a mixture of the present distribution Π t and a given feasibiity measure This aso aows to incude a greedy aspect into the construction of the soutions, as it is common in ACOs under the name of visibiity However, this additiona feature makes the mathematica mode much more compex compared to that in 3 Sti, we are abe to show that the resuts from 3 hod for this generaized mode We investigate τ, the number of iterations unti an optima soution is samped for the first time We show that, if the smoothing parameter ϱ t is a constant, then τ has a stricty positive probabiity to be infinite, and therefore the expected runtime is To get more insight into the finite time behavior of the agorithm, we consider a standard test probem LeadingOne, see 4) We show that for this probem, even if ϱ t is constant, the runtime, ie the number Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

2 2 of soutions samped unti an optima one is found, is bounded by a poynomia in the probem size with a probabiity that is converging fast against one In genetic agorithms, the phenomenon of genetic drift is we-known, see eg 5, it describes the oss of a variation in the soutions A simiar thing may aso happen in our mode We show that, for smoothing parameters ϱ t bounded away from 0, the distribution Π t as we as the samped soutions converge Thus, for fixed ϱ t ϱ > 0, the sampe process is absorbed into a fixed point after finitey many iterations, and this fix point need not be an optima soution This may be the theoretica proof for the stagnation in samping that has been reported in 6, where CE was appied to a Maxima Cut probem In CE, it is of great importance whether amost sure reachabiity of optima soutions and convergence of the underying density are compatibe This important question remained open in 3 We sove this question by giving smoothing sequences that show both, reachabiity of the optima soution and convergence of the density with probabiity one In this sense, these sequences baance what is sometimes caed expoitation and exporation However, at present we are not abe to show that the imiting density is concentrated on optima soutions Our resut aso shows the difference between the two types of convergence: convergence of the density does not necessariy mean that the sampe process is absorbed after finitey many iterations Appied to ACOs, our resuts compement the convergence resuts of 7 and 8 to the case where the update of pheromones ony uses current soutions They aso show that, with this update, ACOs wi end in suboptima soutions with a positive probabiity if a constant evaporation rate is used See Section IV for detais on ACOs The paper starts with an exact definition of the encoding of the soutions, the CE agorithm and the stochastic mode describing its evoution in Section II In Section III we state the main resuts and discuss their impications In Section IV, we discuss some extensions of our mode and how it appies to the ACO agorithms The very compex and technica proofs together with some auxiiary resuts are coected in Section V Section VI contains a concusion and an outook on further research A Probem Encoding II THE MATHEMATICAL MODEL We are considering a probem of discrete optimization with a finite set of feasibe soutions S and a cost function f : S R Let S := {s S fs) = min s S fs )} be the set of optima soutions To excude trivia cases, we assume that S > 1 and that S S We assume further that S has a particuar structure: each feasibe soution s S can be written as a finite string s = s 1,, s L ) of symbos from a finite aphabet A := {a 1,, a K } L is the fixed ength of the strings Often, CE modes eg in 3) use an unconstrained soution space S = A L We generaize this mode introducing a feasibiity function C i y, a) that assigns a weight to each a A for every possibe partia soution y of ength i Thus, if C i y, a) = 0, the partia soution y cannot be continued by adding the symbo a The arger the vaue of C i y, a), the more desirabe it seems, from a greedy point of view, to continue with symbo a We assume that these vaues are normed, an exampe is given after the forma definition beow Our CE method, described in Subsection II-B beow, constructs soution strings stepwise by adding new feasibe symbos to the right end of a partia soution unti it is compete More formay, we define the set R i of feasibe partia soutions of ength i recursivey as foows: Let denote the empty string over A We assume that we are given C 0, ) : A 0, 1, C 0, a) = 1, a A expressing the feasibiity or desirabiity of a symbo at the first position of a soution We then define R 1 := {a A C 0, a) > 0} as the set of feasibe partia soutions of ength 1 Assume now that we have defined R i for some i {0,, L 1} and that we are given C i y, ) : A 0, 1, C i y, a) = 1 for each y R i a A Let y, a) denote the concatenation of symbo a A to the right end of the partia soution string of symbos) y R i Then we define R i+1 := { y, a) y R i, a A, C i y, a) > 0 } and put S := R L For y R i, i {0,, L 1}, et C i y) := {a A C i y, a) > 0} be the support of C i y, ) We use the abbreviation I := {0,, L 1} in the seque As an exampe, we ook at a Traveing Saesman Probem TSP), where the aim is to find a tour of minima ength through K cities {a 1,, a K } Here, C 0, a) = 0 coud be used to prevent tours from starting in city a If y R i is a partia tour of ength i, then C i y) contains a cities that coud be visited next C i y, a) woud be zero if a has been visited before or cannot be reached from y 0 < C i y, a) coud be chosen as the normed distance from the end of) y to a Here, the distance has to be normed by the sum of distances from y to a a C i y) This feasibiity concept aows to incude the case where there is ony a set C i y) A of feasibe continuations of the partia soution y R i and no desirabiity information In this case, C i y, ) is chosen to be the uniform distribution on C i y) In particuar, if there are no constraints at a, we have C i ) A for a i {0,, L 1} and put C i y, a) 1 A, for a y R i, a A II1) We sha refer to II1) as the unrestricted case Athough the construction of S = R L seems to be very particuar, it imposes no restrictions on the optimization probems Formay, we can incude any finite set S of feasibe Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

3 3 soutions into our mode by choosing A := S, L := 1 and C 0, s) > 0 for a s S Thus we can force any probem into our framework, but the efficiency of the method which is not discussed in this paper) may not be high It is more reasonabe to use our mode in cases where A S B The Generaized CE Agorithm The cross entropy agorithm essentiay evoves a distribution on the set S = R L of a feasibe soutions with the aim to give a high probabiity to the optima soutions from S Let PA) denote the set of a probabiity measures on the set A Then p PA) L, p = p1),, pl)) is a product probabiity measure on A L, that describes the seection of a soution s = s 1,, s L ) A L, where the L symbos s 1,, s L are chosen independenty of each other Here, pi) = pa; i) ) PA) is the distribution for the symbo a A on the i-th position of the string Our CE agorithm takes the foowing items as input: the feasibiity distributions C i, ), i I; a sequence of smoothing parameters ϱ t ) t 1 with ϱ t 0, 1) ; a sampe size N N and a subsampe size N; a starting distribution p 0 PA) L Starting: For time-step t = 0, put p := p 0, then run iterativey through the foowing steps for t = 1, 2, unti some stopping criterion is fufied Samping: If the present distribution is p PA) L, a soution s = s 1,, s L ) S is drawn according to the probabiity L Q p s) := Q p s 1 ; 1, ) Q p s i ; i, s 1,, s i 1 )), i=2 II2) where pa, i)c i 1 y, a) Q p a; i, y) := a A pa, i)c i 1 y, a II3) ) is the probabiity that the feasibe symbo a A is added at position i to the feasibe partia soution y R i 1 We use the convention 0 0 = 0 throughout this paper In this way, the CE agorithm draws N soutions s 1),, s N) independenty and identicay distributed iid) Evauation: This sampe x := s 1),, s N) ) is ordered according to its cost vaues fs n1) ) fs n2) ) fs n N ) ) and the better soutions are seected: := {s n1), s n2),, s n N ) b } Then we define the reative frequency of symbo a at position i {1,, L} in the seected part of the sampe: wa; i, x) := 1 1 {a} s i ) II4) s and coect these frequencies for a a A to form wi, x) := wa; i, x) ) a A and wx) := w1, x),, wl, x) ) II5) Then wx) is a product probabiity measure from PA) L, that gives the reative frequencies of symbos in the better part of the sampe x drawn with Q p Update: We update the present distribution p as a convex combination of p and the reative frequencies wx): p := 1 ϱ t+1 )p + ϱ t+1 wx) II6) Next, the counter t is increased by 1 and the step Samping is performed with the new p Some extensions to this mode and a comparison to other approaches can be found in Section IV beow C The Soution Process Appying the above agorithm iterativey resuts in a sequence of probabiity measures and sampes that form a stochastic process Π t ; X t ) t=0,1, where Π t = Π t 1),, Π t L) ) is a random variabe taking on vaues in PA) L, that describes the distribution underying the samping in the t-th iteration Π t is caed the density of the agorithm with Π 0 = p 0 X t takes on vaues in S N and is the sampe of N soutions produced in the t-th Samping step of the agorithm, using Q Πt as defined in II2) In the seque, we put X t = X 1) t,, X N) ) t and X n) t = X n) t 1),, X n) t L) ) Then X n) t i) denotes the symbo at position i, 1 i L, drawn in the n-th soution in iteration t Simiary, X n) t 1,, i) denotes the partia soution up to the i-th position in that soution The process Π t, X t ) t 0 is we defined on a joint probabiity space Ω, P ) It can be shown, using II2) and II6), that ) Π t, X t as we as the margina process Π t 0 t )t 0 are Markov processes Due to the deterministic nature of the update mechanism, we may aso write Π t+1 = 1 ϱ t+1 )Π t + ϱ t+1 wx t ) on a vector eve and on the more detaied eve Π t+1 a ; i) = 1 ϱ t+1 )Π t a ; i) + ϱ t+1 wa ; i, X t ) II7) for each a A, i = 1,, L We sha refer to II7) as the basic recursion, most of our resuts are based on it D Genera assumptions Throughout this paper we assume that the starting distribution Π 0 = p 0 is given such that any item from A has a positive probabiity at any position i : p 0 a; i) > 0 for a a A, i = 1,, L II8) Aso, without oss in generaity we may assume that there are at east two soutions s = s 1,, s L ), s = s 1,, s L) with s 1 s 1 II9) Otherwise, a soutions woud start with the same symbo, which coud then be dropped from the encoding of the soutions Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

4 4 III MAIN RESULTS In this Section we coect the main resuts of this paper and discuss their meaning The quite compicated proofs are given in Section V A Resuts on the Reachabiity of Optima Soutions Let X t := {X 1) t,, X N) t } be the set of soutions samped in the t-th iteration, then τ := min{t 0 X t S } III1) denotes the first iteration, in which an optima soution is samped We now investigate whether we can guarantee to reach an optima soution in finite time, ie whether Pτ < ) = 1 Theorem 1 shows, that the resuts of 3 hod in our more genera mode Theorem 1 a) If t 1 ϱ m) L =, then Pτ < ) = 1 t 1 ϱ m) = b) If Pτ < ) = 1, then c) If ϱ t ϱ > 0 is a constant, then we have Pτ < ) < 1, but Pτ < ) 1, if either N or ϱ 0 The proof of this Theorem aong the ines of 3 is given in Section V-B beow Theorem 1 b) shows in particuar that for a constant smoothing parameter, ie ϱ t ϱ, it cannot be guaranteed that the optima soution is reached in finite time, as we have t 1 ϱ) = 1/ϱ 1 < But then aso the expected runtime, ie the expected number of soutions samped before an optima soution is reached, wi be infinite for arbitrary optimization probems For a particuar probem however, namey the LeadingOne probem, see eg 4, we are abe to show that the runtime is bounded by a poynomia in the probem size with a probabiity converging to 1, even if ϱ t is constant In the LeadingOne probem we consider the unrestricted case with A = {0, 1}, S = {0, 1} L and fs) := L L i=1 s i for s = s 1,, s L ) III2) Minimizing fs) then means to maximize the number of consecutive 1s counted from the eft of the soution Theorem 2 extends the runtime resuts of 4, who consider a specia case of the CE agorithm with ϱ t 1 on a LeadingOne probem The Theorem essentiay states that if we take ϱ t ϱ 0, 1) and et the sampe size grow as N = L 2+ɛ for some ɛ > 0, then we can reach the optima soution in L iterations with a probabiity converging to 1, ie the runtime is OL 3+ɛ ) in a stochastic sense Theorem 2 Let ϱ t ϱ, sampe size N = L 2+ɛ for some ɛ > 0 and = βn for some 0 < β < 1 ) 3e 1 1 ϱ) m Let Π 0 1, i) 1 2, ie we start with the uniform distribution Then for a LeadingOne probem, defined in III2), we have Pτ < L) 1 as L For a proof of Theorem 2, see Subsection V-C beow Theorem 1 b) and the ower bound in Lemma 10 b) beow show that a constant smoothing parameter reduces the goba exporation of the search space Theorem 2 however shows that we can compensate for that by increasing the oca expoitation in terms of a growing sampe size, at east in specific probems B Resuts on Convergence of Density and Sampes An important question, aso addressed in 3, is whether the density Π t wi converge to a density concentrated on one point, and whether this point is an optima soution For the unrestricted case, 3 showed that the agorithm with ϱ t ϱ has a convergent density in the foowing sense: Definition 3 We say that the agorithm has convergent densities, if Π t converges amost surey against a product )t 0 of one-point measures, ie ) P i = 1,, L a A im Π t a ; i) {0, 1} = 1 A convergent density means that asymptoticay, we ony sampe one identica soution But this may aso happen after finitey many iterations: Definition 4 We say that the agorithm has convergent sampes, if X t converges against some fixed sampe )t 0 x = s,, s) S N of identica soutions, amost surey, ie if the foowing event has probabiity one: s S T N m 1 n = 1,, N X n) T +m = s In this case, the samping freezes to a singe soution after the finite random time T and no more progress is possibe after that time Theorem 5 shows that this does happen, if a constant smoothing parameter is used Theorem 5 a) If ϱ t ϱ for some ϱ > 0, then the agorithm has convergent sampes b) For S = 1, convergent sampes impy Pτ < ) < 1, hence convergence of sampes and Pτ < ) = 1 are mutuay excusive in this case c) Convergent sampes impy a convergent density d) If the density converges, then ϱ t = The quite compex proof of Theorem 5 is given in Section V-D beow Theorem 5 a) shows that the phenomenon of genetic drift may aso occur in this generaized CE provided a constant smoothing parameter is empoyed From Theorem 5 b), we see that in this case an optima soution may not be found if it is unique C Reachabiity and Convergent Density are Compatibe In 3, the question was raised whether convergence of densities and reachabiity of optima soutions are mutuay excusive, and this remained an open question In Theorem 5, a partia answer was given with respect to the stronger concept of convergent sampes At east for the unrestricted case, we can give a compete answer: there are smoothing sequences such that the agorithm has Pτ < ) = 1, it has convergent densities but no convergent sampes This aso shows that the opposite direction in Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

5 5 Theorem 5 c) does not hod in genera, densities may converge without freezing the sampes in finite time Theorem 6 In the unrestricted case the foowing hods: Assume that ϱ 0, 1) and c k N for k = 0, 1,, with c 0 = 1, are chosen such that c k 1 ϱ) kl = k=1 Define e k := k i=1 c i 1, k 1, and et x k 0, 1) be any sequence such that k=1 x k < We may now define a smoothing sequence ϱ if t = e k for some k 1 ϱ t := 1 1 x k ) 1 c k 1 if e k < t < e k+1 for some k 1 III3) for t 1 Then the agorithm has convergent densities, it has Pτ < ) = 1, and its sampes do not converge The proof of Theorem 6 is again given in Section V-E As an exampe for the vaues in Theorem 6, take an arbitrary ϱ > 0, then one may choose c k = 1 ϱ) kl to obtain c k 1 ϱ) kl = 1 = k=1 k=1 For ϱ < 1/2, one coud, for exampe, use c k := 2 kl Theorem 6 shows that a goba exporation of the search space that impies reachabiity) and the abiity for thorough oca search, that incudes some kind of convergence, can be baanced in this type of agorithm It aso shows that the two types of convergence are different, convergence in the continuous space of densities does not impy convergence in the finite space of sampes We must admit, however, that we are currenty not abe to give a smoothing sequence under which the imiting density is concentrated on some optima soution IV EXTENSIONS AND COMPARISON TO ACOS The evauation step in II4) simpy cacuates the reative frequencies in the better part of the sampe A thorough inspection of the proofs beow shows that the resuts of the above Theorems except Theorem 2) sti hod, if we repace the reative frequencies wa; i, x) by some other measure 0 wa; i, x) 1 not depending on t with the foowing properties: i) a A wa; i, x) = 1, ie w ; i, x) is a probabiity measure on A; ii) wa; i, x) = 1 if a soutions in the sampe x have symbo a at their i th position; iii) wa; i, x) = 0 if none of the soutions in the sampe x has symbo a at its i th position For exampe, w coud be the reative frequencies of a randomy seected subset N of x, or the weighted reative frequencies where the weights refect the cost fs) of the soution as in s N wa; i, x) := 1 {a}s i )gfs)) s N gfs)), IV1) here g is some decreasing function w may even incude some kind of memory, as ong as properties i)-iii) hod In the proofs of Section V beow, we wi stick to the origina frequency mode as used in 3 With these extensions, our mode covers many heuristic optimization agorithms of the mode-based type Therefore the resuts given above wi appy to a these agorithms In particuar, they appy to ACO agorithms as described in 9 and 10 Here soutions are maxima oop-free paths in a so-caed construction graph The arcs k, ) of the path are seected according to pheromones τ k and a visibiity η k on the arcs Taking arcs as symbos a := k, ), pheromones can be expressed by the density in our mode, and the visibiity together with the oop-freeness can go into our feasibiity measure C i y, a) The evaporation rate for the update of pheromones is identica to the smoothing parameter, hence soutions in our CE mode are samped with the same probabiity as they are constructed by ants, if we choose IV1) as evauation of the soutions observed Note that our mode requires density and feasibiity to be normed probabiity measures, but this does not affect the samping probabiities as given, for exampe, by 9 Convergence resuts of ACOs can be found in 10, 7 and 8 They concentrate on ACOs with an update of pheromones that ony uses the best soution found so far and restricted pheromone vaues in 7, whereas our mode has no restrictions on the density and uses the present sampe for update With these differences, our resuts on reachabiity compement those found in 10 and 8 In particuar, the assumptions made in 7 fufi the sufficient condition of reachabiity in Theorem 1 a) The resuts on absorption for constant evaporation rate and the possibe baancing in Theorems 5 and 6 aso carry over to ACOs Our runtime resut in Theorem 2 compements findings in 11, 12 and 13, who inspect the runtime of ACOs with restricted pheromone vaues max-min ant system) on the so-caed OneMax probem Using best-so-far update, as in the ACO modes cited above, introduces a monotonicity that aows to show convergence to an optima soution once it has been samped, see 7 Currenty, we are not abe to show such a resut for our agorithm A Some Auxiiary Resuts V PROOFS OF THE THEOREMS We start with the basic recursion in II7), which is crucia for our work For a fixed pair a; i), this recursion fufis the condition V1) of the foowing Lemma 7, and therefore the concusions V2) - V4) hod This wi be used throughout this paper We assume as usua k i=m Lemma 7 Let 0 < r t < 1 for t = 1, 2, a) r t = 1 r t ) = 0 1 for m > k 1 cr t ) = 0 for any 0 < c < 1 Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

6 6 b) t r m i=m+1 1 r i ) = 1 1 r m ) for any t 1 c) For a given sequence w t 0, 1, t = 0, 1,, and q 0 0, 1), the recursion q t+1 = 1 r t+1 )q t + r t+1 w t, t 0, V1) has the unique soution q t = q 0 1 r m ) + with t r m w m 1 i=m+1 1 r i ) V2) 0 < q 0 1 r m ) q t V3) 1 1 q 0 ) 1 r m ) < 1, t 0 If, in particuar, w m w 0, 1 for m = 0,, t 1, then q t = w w q 0 ) 1 r m ) V4) Proof: see 3) a) The first part is a standard resut, the second foows as i=0 r i = i=0 cr i = b) foows, if r m = 1 1 r m ) is used on the eft hand side c) V2) can be proven using induction on t V4) foows from V2) using b) Now, V3) appied to II7) and II8) shows that 0 < Π t a; i) < 1 V5) for a t 0, i {1,, L} and a A We then aso have 0 < C i y, a)π t a; i+1) < 1 V6) for a t 0, i I, y R i and a C i y) In the unrestricted case as in 3, we have Q Πt a; i+1, y) = Π t a; i+1), ie the probabiity to continue a feasibe partia soution y R i with a symbo a in the t-th iteration coincides with the density Πa; i+1) We may therefore directy use the basic recursion II7) to derive the asymptotic behavior of the procedure from Lemma 7 In the genera constrained case this is not possibe, as Q Πt does not fufi an equation as II7) But we can find a bounding probabiity Q Π t for Q Πt, for which a simiar recursion hods under certain conditions see Lemma 11 beow) For p PA) L, i I, y R i and a A define { pa;i+1) Q pa; i+1, y) := a C i y) pa ;i+1) if a C i y), 0 otherwise, V7) with 0 0 = 0 This is simiar to Q p except that the feasibiity distribution C i y, a) has been repaced by a constant on its support C i y) Note, that for the unrestricted case of II1), Q p, Q p and p a coincide To bound Q p a; i+1, y) with the hep of Q pa; i+1, y), we first need bounds on C i y, a) Let η := max { C i y, a) y R i, a A, i I } V8) and λ := min { C i y, a) > 0 y R i, a A, i I } V9) Obviousy, 0 < λ η 1 Now define the foowing two bounding functions for x 0, 1 hx) := ηx λ + η λ)x, x) := λx η η λ)x V10) Then x) x hx) and in the unrestricted case hx) = x) = x Lemma 8 coects some properties of h and Lemma 8 Let h, be as defined in V10) a) h and are stricty increasing and continuous taking vaues in 0, 1 h is concave, is convex with x) x hx), x 0, 1 b) For η = λ we have hx) = x) = x for x 0, 1, for η > λ we have hx) = x = x) x { 0, 1 } c) hx) = 1 1 x) and x) = 1 h1 x) d) Let 0 x n 1, n N, be a convergent sequence, then for any 0 < c < 1, the foowing series are either a convergent or a divergent x n, hx n ), x n ), n=1 n=1 hcx n ), n=1 n=1 cx n ) n=1 Proof: Assertions a) - c) are straightforward to show Assertion d) foows immediatey from the we-known imit comparison test Lemma 9 beow coects some properties of Q Πt that are needed throughout the paper Reca that is the empty string that we use as a starting vaue for the recursions Note that some assertions ony hod if C i y) > 1 This excudes the case that there is just one possibe continuation of y, which then must have probabiity one Lemma 9 Let i I, y R i and a C i y) Then the foowing hods a) Q Π t a; i + 1, y) ) Q Πt a; i + 1, y) h Q Π t a; i + 1, y) ) for a t N b) If C i y) = 1 then Q Πt a; i + 1, y) = Q Π t a; i + 1, y) = 1 If C i y) > 1, then 0 < Q Π t a; i + 1, y) < 1 and 0 < Q Πt a; i + 1, y) < 1 for any t 0 Proof: a) In view of Lemma 8 b) we ony have to consider the case C i y) > 1 Π t a; i + 1)C i y, a) Q p a; i + 1, y) = a A Π ta ; i + 1)C i y, a ) ηπ t a; i + 1) ηπ t a; i + 1) + λ a C Π iy),a a ta ; i + 1) = h Q Π t a; i + 1, y) ), where we use V5) Simiary, the eft-hand inequaity of a) is derived b) From V5) we have Q Π t a; i+1, y) > 0 If C i y) > 1, we see that Q Π t a; i+1, y) < 1 for any a C i y) and t N, this impies the concusion using Lemma 8 b) and part a) Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

7 7 B Proof of Theorem 1 We first give ower bounds on the probabiities of a) reaching the set S of optima soutions and b) of staying forever with one possiby non-optima) soution Lemma 10 a) For s = s 1,, s L ) S and p 0 PA) L et L ) cs) := C 0, s 1 ) C i 1 s1,, s i 1 ), s i and i=2 δs) := p 0 s) cs) Then for τ as defined in III1) Pτ < ) 1 min 1 δs ) s S t=0 V11) 1 ϱ m ) L) N b) Let s S, then, with h as defined in V10), PX n) t = s for n = 1,, N and t = 0, 1, ) Q Π0 s) N 1 h t 1 ϱ m ) )) LN Note that the bound in part a) hods for the first hitting time of any subset of soutions, so it is in fact a bound for the probabiity that arbitrary soutions wi be visited in finite time Proof of Lemma 10: a) cf 3) We fix an arbitrary optima soution s = s 1,, sl) S, then we have ) Pτ = ) = P S X t = P t=0 t=0 ) s / X t = Ps / X 0 ) P s / X t s / X m, m = 0,, t 1 V12) We now derive an upper bound for the factors in V12) First we have P s / X t s / X m, m = 0,, t 1 V13) = E P s / X t Π t s / X m, m = 0,, t 1 From the definition II3) of Q p and V3) of Lemma 7 appied with q 0 = Π 0 a; i) we have Q Πt a; i, y) Π t a; i)c i 1 y, a) Π 0 a; i)c i 1 y, a) 1 ϱ m ) for a a A, y R i 1, i {1,, L} and t 0 As the soutions are samped iid, we may now concude P s / X t Π t V14) = P s X 1) ) N t Π t = 1 P s = X 1) ) N t Π t 1 δs ) 1 ϱ m ) L) N For the first factor in V12), we have from V14) for t = 0 N Ps / X 0 ) 1 δs )) Combining these resuts, we obtain Pτ < ) = 1 Pτ = ) 1 1 δs ) 1 ϱ t ) L) N t=0 V15) As this hods for a s S, the assertion foows b) Let s S We write X ) t s for X n) t = s, n = 1,, N and S for the event X m ) s for a m = 0,, t 1 Then, as sampes are iid, P X ) t s, t = 0, 1, = P X ) 0 s V16) P X ) t s X ) m s for m = 0,, t 1 = P X 1) 0 = s N P X 1) t = s S N We have PX 1) 0 = s) = Q Π0 s) for the first factor Using Lemma 9 a), we obtain for the other factors in V16) with s i := s 1,, s i ) P X 1) t = s S = P X 1) t 1) = s 1 S L i=2 P X 1) t i) = s i X 1) t 1,, i 1) = s i 1, S = E Q Πt s 1 ; 1, ) S L E Q Πt s i ; i, s i 1 ) X 1) t 1,, i 1) = s i 1, S i=2 L E Q Π t s i ; i, s i 1 ) ) S, i=1 with s 0 = Under the condition S, the reative frequencies ws i ; i, X m ) are a equa to 1 for a m = 0,, t 1, hence we may use Lemma 11 d) beow see the remark foowing Lemma 11 and V36)) to obtain Q Π t s i ; i, s i 1 ) 1 1 ϱ m ) for t 1 and i = 1,, L Now, from V16), we obtain using Lemma 8 c) P X ) t s, t = 0, 1, V17) Q Π0 s) N 1 h t 1 ϱ m ) )) LN Proof of Theorem 1: a) From Lemma 7 a), we see that with r t := t 1 ϱ m) L and c := δs ), the assumption t 1 ϱ m) L = impies 1 δs ) 1 ϱ m ) L) = 0 Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

8 8 Now Pτ < ) = 1 foows from Lemma 10 b) Let s S S with Q Π0 s) > 0 Note that such an s must exist, otherwise we coud ony sampe optima soutions If Pτ < ) = 1, the sampes cannot be identica to s for a times, hence from Lemma 10 b) we must have 0 = PX ) t s, t = 0, 1, ) Q Π0 s) N 1 h t 1 ϱ m ) )) LN, and 1 h t 1 ϱ m) )) = 0 must hod From Lemma 7 a) we see that this impies h ) 1 ϱ m ) = The assertion now foows from Lemma 8 d) c) Assume ϱ t ϱ > 0, we have t 1 ϱ) <, hence by part b) of this Theorem, Pτ < ) < 1 Note that the infinite product in V15) in this case is smaer than 1 and continuous in ϱ, so V15) approaches 1 for either N or ϱ 0 C Proof of Theorem 2 Proof: With A = {0, 1}, we write π t i) := Π t 1; i) As we consider the unrestricted case here, the sampe X t = X 1) t,, X N) t ) can be viewed as a random matrix with mutuay independent entries and soution X n) t as n-th row We denote by X t = X 1 t,, X N ) t this matrix with rows ordered according to increasing cost function vaues : fx 1 t ) fx N t ) Then is the sub-matrix of the first rows of X and the empirica distributions from may be written as w1; i, X t ) = 1 Nb n=1 Xn t i) =: wi, X t ) and wx t ) := w1, X t ),, wl, X t )) For the proof, we et wx t ), t = 0, 1,, move over a sequence of increasing eves wt These eves are defined in such a way that fx t ) is stricty decreasing and at the same time the update of π t can be controed such that we are abe to give a ower bound on the probabiity for this to happen For t = 0,, L 1 et w t := w t 1),, w t L) ) := 1,, 1, α t,, α t ) V18) with t + 1 entries 1 and some α t 0, 1) defined in V21) We write wx t ) wt if and ony if wi, X t ) = 1 for i = 1,, t+1 and wi, X t ) α t for i = t+2,, L For t = L 1, wx t ) wt means that consists of the optima soution s = 1,, 1) ony Hence we have Pτ < L) PwX L 1 ) wl 1) P wx 0 ) w0,, wx L 1 ) wl 1 ) = PwX 0 ) w0) L 1 V19) P wx t ) w t wx m ) w m, m = 0,, t 1 We show beow that there are constants a, b, c > 0 such that for a L arge enough and for a N P wx t ) w t wx m ) w m, m = 0,, t 1 1 e an ) 1 e b N L 2 +c) L t 1 V20) With N = L 2+ɛ for some ɛ > 0 we may then concude that Pτ < L) 1 e al2+ɛ ) L 1 e blɛ +c ) L 2 L)/2 Now this ast bound can be shown to converge to 1 for L using standard methods from anaysis Hence the proof of Theorem 2 is compete, once we have shown V20) for t = 0,, L 1 In the first step we show how the eves w t infuence π t We abbreviate the event wx m ) w m, m = 0,, t 1 by W t, W 0 indicating the empty condition Let α t := L )t+1, t = 0,, L 1, and α 1 := 1 2 V21) Conditioned on W t, { 1 1 ϱ) π t i) t i+1 for 1 i t for t < i L α t 1 V22) for t = 0,, L 1 The proof is by induction on t using the basic recursion II7) and the property wi, X m ) = 1 for a m = i 1, i,, t 1, if i t, wi, X m ) α m for a m = 0, 1,, t 1, if i > t, which foows from the definition of wt under W t Let v t+1 = P X n) t 1) = = X n) t t+1) = 1 W t be the probabiity to sampe a soution that has t+1 eading 1s, then from V22) we obtain t+1 v t+1 = π t i) ϱ) m ) =: κϱ) V23) 3e i=1 for L arge enough, as then α t 1/3e) Now we want to determine simpe conditions on X t that impy wx t ) wt To do so, we ook at the matrix X t coumnwise observing the independence of its entries Let M t+1) be the number of rows with at east t+1 eading 1s, then wi, X t ) = 1, i = 1,, t+1, requires that M t+1) These M t+1) rows are aso the first rows in X t Next, wi, X t ) α t requires for the number of 1s in coumn i = t + 2 Y i) := M i 1) n=1 X n t i) α t V24) Here we have to restrict the number of rows to the present candidate rows 1,, M i 1) from which the set is seected After ooking at coumn i in this way, we define M i) := max{, Y i) } V25) and repeat V24), V25) for i = t + 3,, L We then obtain P wx t ) w t W t V26) P M t+1), Y i) α t, i = t + 2,, L W t = P M t+1) W t L i=t+2 P Y i) α t Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

9 9 Y ) α t, = t+2,, i 1, M t+1), W t To derive the desired ower bounds for these expressions we need the Chernoff bound see eg 14, Theorem 42) in the foowing form: et Z 1,, Z m be iid 0-1-distributed with success probabiity p, then for any 0 < r < mp we have P m i=1 Z i r ) e r mp )2 mp V27) Conditioned on W t, M t+1) is distributed as the number of successes in a row of N iid experiments, each with success probabiity p := v t+1 We obtain for β < κϱ) v t+1 that = βn βn < v t+1 N Hence for t = 0,, L 1 P M t+1) W t = 1 P M t+1) < W t 1 E e v t+1 N )2 v t+1n Wt 1 e β κϱ) )2κϱ)N, V28) where we used V27) Hence, in V20) we may define a := β κϱ) )2 κϱ) Simiary, Y t+2) is distributed as the number of 1s in M t+1) iid trias each with success probabiity π t t + 2) From V22) and the definition of α t we see that under the condition used in V26) we have π t t + 2)M t+1) α t 1 M t+1) > α t Using the Chernoff bound we therefore obtain, for L arge enough, P Y t+2) α t W t, M t+1) V29) = 1 P Y t+2) < α t W t, M t+1) 1 e α t α ) 2 t 1 3e = 1 e 6eL 2 1 e βn 1 6eL 2, where we used αt α t 1 = 1 1 L A competey anaogous derivation hods for the other factors in V26) with i = t+3,, L Hence, with b := β 6e, c := 1 6e, we see from V29), V28) and V26) that V20) hods for L arge enough D Proof of Theorem 5 The proof of Theorem 5 about convergent sampes uses an inductive argument The induction hypothesis assumes, that the first i positions of a sampes have aready converged against a fixed vaue y R i and we show that then aso the samped vaue on the i + 1st position wi become constant in finite time More precisey, we assume as induction hypothesis that for a fixed i I the foowing hods with probabiity one im Xn) t 1,, i) = y for a n = 1,, N and some y R i As R i, the set of possibe vaues of X m n) 1,, i) is finite, this condition is equivaent to requiring that there are random variabes y and T taking on vaues in R i and N, defined on the common probabiity space, such that the foowing event has probabiity one: m 0 n = 1,, N X n) T +m 1,, i) = y V30) Lemmas 11 and 12 beow contain the crucia induction step to prove convergence of sampes Lemma 11 shows that under condition V30) we may appy the recursion of Lemma 7 c) to Q Π t We use the foowing definitions for p PA) L, i I and y R i G i y, p) := pa ; i + 1) and V31) ϱ y t := a C iy) ϱ t G i y, Π t ) for t 1 From V5) we see that aways G i y, Π t ) > 0 V32) Lemma 11 Let i I be fixed and assume that V30) hods with probabiity one for random variabes y, T as above Then the foowing hods amost surey for a m 1 a) G i y, Π T +m ) = 1 1 G i y, Π T ) ) m 1 ϱ T + ), and G i y, Π T +m ) is an increasing function of m b) 0 < ϱ T +m ϱ y T +m < 1 c) For a a C i y) we have Q Π T +m a; i + 1, y) = 1 ϱ y T +m )Q Π T +m 1 a; i + 1, y) + ϱ y T +m wa; i + 1, X T +m 1) V33) Thus the impications of Lemma 7 c) hod with q t := Q Π T +t a; i + 1, y), r t := ϱ y T +t and w t := wa; i + 1, X T +t ) with the restriction that the strict right hand bound in V3) ony hods if C i y) > 1 d) If wa; i+1, X T + ) = w 0, 1 for a = 0,, m 1, then Q Π T +m a; i + 1, y) = w w Q Π T a; i+1, y) ) m 1 ϱ y ) V34) w w Q Π T a; i+1, y) ) m 1 ϱ ) V35) In fact, the foowing proof shows that, for the assertions of Lemma 11 to hod for a fixed m 1, instead of V30) the foowing weaker condition is sufficient = 0,, m 1 n = 1,, N X n) T + 1,, i) = y V36) Proof of Lemma 11: a) From the basic recursion II7), we have for any m 1 G i y, Π T +m ) = Π T +m a ; i + 1) V37) a C iy) = 1 ϱ T +m )G i y, Π T +m 1 ) + ϱ T +m, as a C iy) wa ; i + 1, X T +m 1 ) = 1 Hence, q t := G i y, Π T +t ) fufis the condition V1) of Lemma 7 with w m 1 Now V4) shows that G i y, Π T +m ) = 1 1 G i y, Π T ) ) m 1 ϱ T + ) Aso, from V37) we have G i y, Π T +m+1 ) G i y, Π T +m ) b) We have 0 < ϱ T +m ϱ y T +m and G iy, Π T +m 1 ) > 0 by V5) From V37) we now see ϱ T +m < G i y, Π T +m ), hence ϱ y t < 1 c) From V37) we obtain 1 ϱ T +m )G i y, Π T +m 1 ) = G i y, Π T +m ) ϱ T +m Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

10 10 This together with the basic recursion for Π t fufis the recursion Q Π T +m shows that ConvW) There is a random variabe) a 0 as above such that amost surey Q Π T +m a; i + 1, y) = Π T +ma; i + 1) G i y, Π T +m ) = 1 ϱ T +m)π T +m 1 a; i + 1) G i y, Π T +m ) + ϱ T +m wa; i + 1, X T +m 1 ) G i y, Π T +m ) = 1 ϱ y T +m )Q Π T +m 1 a; i + 1, y) + ϱ y T +m wa; i + 1, X T +m 1) For the appication of Lemma 7 c), we ony have to check that 0 < r t = ϱ y t < 1, which was shown in part b), and 0 < q 0 = Q ΠT a; i + 1, y) < 1 for C i y) > 1, which was shown in Lemma 9 b) d) From part c) and Lemma 7 V4) with q 0 = Q Π T a; i+ 1, y) we obtain Q Π T +m a; i+1, y) = w w Q Π T a; i+1, y) ) m 1 ϱ y ) such that the assertion foows from part b) Lemma 12 contains the induction step for the proof of Theorem 5 Lemma 12 Assume that for some ϱ > 0 we have ϱ t ϱ > 0 for a t 1, V38) and assume, as in Lemma 11, that there are i I and random variabes y, T such that V30) hods with probabiity one Then the foowing hods amost surey: there are T 0 and a 0 A with X n) T +m i + 1) = a 0 for a n = 1,, N and a m 0 V39) This Lemma says that, if a sampes have a common eading partia soution y of ength i after a finite number of iterations, then the sampes wi finay aso coincide on the next position i + 1, if ϱ t is bounded away from 0 Note that the condition is aways fufied for i := 0, y := Thus iterated appication of this Lemma aows to prove Theorem 5 beow Proof of Lemma 12: Part of this proof extends the approach used in 3 to prove convergence of densities and coses a sma gap in that proof As the proof is quite ong, we separate it into severa abeed steps Let i I be fixed and et y and T be such that V30) hods with probabiity one Note that if C i y) = 1, the concusion triviay hods, so we may assume C i y) > 1 To show the fina assertion of the Lemma we have to prove: ConvQ) There is a random variabe) a 0 taking on vaues in A such that amost surey im Q Π t a 0 ; i + 1, y) = im Q Π t a 0 ; i + 1, y) = 1 ConvQ) wi foow, after we have shown convergence of the reative frequencies: im wa; i + 1, X t) = { 1 if a = a0 0 if a a 0 for a a A ConvW) in turn foows, if we have shown that wa; i+1, X t ) finay becomes monotone: MonW) For a a C i y), it hods amost surey that Za; m) := wa; i + 1, X T +m ) wa; i + 1, X T +m 1 ) has ony finitey many sign changes as function of m We now prove these three steps in reverse order Reca that under the condition of the Lemma, amost a soutions from X T +m and X T +m 1 coincide in their first i positions To prove MonW) et M k a) be the k-th m such that Za; m + 1)Za; m) < 0 As A is finite, MonW) hods if for any fixed a C i y) and M k := M k a) Now, observe that P k N M k < ) = PM 1 < ) P k N M k < ) = 0 V40) P M k < M k 1 < k=2 1 P Mk = M k 1 < ) k=2 V41) We are going to show that P M k = M k 1 < κ > 0 for a constant ower bound κ > 0, this proves V40) We have P M k = M k 1 < = P M k = M k 1 = d P M k 1 = d M k 1 < d=k 1 If M k 1 = d then Za; d + 1) 0, hence P M k = M k 1 = d = P M k = M k 1 = d, Za; d + 1) > 0 P Za; d + 1) > 0 M k 1 = d + P M k = M k 1 = d, Za; d + 1) < 0 P Za; d + 1) < 0 M k 1 = d V42) We use the abbreviation W m := wa; i + 1, X T +m ), M for the event M k 1 = d, Za; d + 1) > 0, and W m for W = 1, = d + 2,, m 1 Conditioned on M, the event M k = is impied by W m = 1, m d+2 We can therefore give the foowing rough bound for the first term in V42) P M k = M k 1 = d, Za; d + 1) > 0 P W m = 1, m d + 2 M = P W d+2 = 1 M m=d+3 QΠT E a; i + 1, y) N +d+2 M P W m = 1 W m, M V43) Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

11 11 m=d+3 E QΠT +m a; i + 1, y) N Wm, M From Lemma 9 a) and V44) beow, we get for the first factor in V43) QΠT E a; i + 1, y) N +d+2 M E Q a; i + 1, y)) N ΠT +d+2 M ϱ ) N, where we used the fact that M impies Za; d + 1) > 0 Therefore wa; i + 1, X T +d+1 ) > wa; i + 1, X T +d ) 0 and wa; i + 1, X T +d+1 ) 1 Lemma 11 c) now shows Q Π T +d+2 a; i + 1, y) ϱy T +d+2 ϱ > 0 V44) For the second factor in V43), we see from Lemma 11 d) that, under the condition W = 1, = d + 2,, m 1, we have Q Π T +m a; i + 1, y) 1 1 ϱ) m d 2 Hence, in the second term of V43) the integrand can be bounded for a m d + 3 in the foowing way QΠT +m a; i + 1, y) N Q ΠT +m a; i + 1, y)) N 1 1 ϱ) m d 2) N = 1 h 1 ϱ) m d 2) N Coecting the equations above, we now obtain for a k, d N P M k = M k 1 = d, Za; d + 1) > 0 V45) ϱ ) N 1 h 1 ϱ) m d 2) N = ϱ ) N m=d+3 1 h 1 ϱ) m) N =: κ From Lemma 8 d), we see that 1 ϱ)m < impies h 1 ϱ) m) <, hence κ > 0 In a competey anaogous manner, using the events W m = 0, m d + 2, we can aso show that P M k = M k 1 = d, Za; d + 1) < 0 κ, which now proves V40) and MonW) We are now going to prove ConvW) From MonW) we know that m wa; i + 1, X T +m ) must eventuay be monotonic and therefore converge to one of the vaues in 1 W := {0,,, 1, 1} remember that w ) is a reative frequency from a sampe of size ) Hence, there is a random variabe T T, and for each a A a random variabe V a taking on vaues in W, such that with probabiity 1, wa; i + 1, X T +m) = V a for a m 0 >From the recursion V33) and Lemma 7 c) and b), we see that, under our conditions for a m 0 and a C i y), Q Π T +m a; i + 1, y) = Q Π T a; i + 1, y) V46) 1 ϱ y T + ) + V m a 1 1 ϱ y T + )) From V38) and ϱ y T + ϱ T + > ϱ, we have ϱy T + = and hence 1 ϱy T + ) = 0 Therefore, V46) impies that, amost surey, im Q Π t a; i + 1, y) = V a for a a C i y) V47) The bounds from Lemma 9 a) and the continuity of the bounding functions h and see Lemma 8 a)) ead us to concude that for a a C i y) V a ) im inf Q Π t a; i + 1, y) V48) im sup Q Πt a; i + 1, y) hv a ) We want to show next, that the imit V a can take on vaues in {0, 1} ony, more precisey: a 0 C i y) V a0 = 1 and a C i y) {a 0 } ) V a = 0 V49) hods amost surey To prove V49), we first show that for a a C i y) we have PV a 0, 1)) = 0 We again use the abbreviation W m = wa; i + 1, X T +m ) and W 0 := W 0, 1) = { 1, 2,, 1 1 } We then have PV a 0, 1)) = P im W k W 0 ) V50) = im P ) W k W 0 P W m W 0 W W 0, = k,, m 1 ) im m=k+1 m=k+1 P W m W 0 W W 0, = k,, m 1 Writing W for W W 0, = k,, m 1, we have for m > k PW m W 0 W W 0, = k,, m 1 = 1 E 1 Q ΠT +m a; i + 1, y) ) N W 1 1 h 1 ϱ) m k ) ) N, V51) where we used the fact that under W we have W 1 1, = k,, m 1 Hence we can see as in V46), that for m > k Q Π T +m a; i + 1, y) Q Π T +k a; i + 1, y) V52) 1 ϱ y T + ) ) 1 1 ϱ y T + )) =k+1 =k+1 ϱ =k+11 y T + ) ϱ) m k But then, for m arge enough, the upper bound of the probabiity V51) is smaer than 1, and the infinite product in V50) vanishes Therefore, we have shown P V a {0, 1} for a a C i y) ) = 1 We aso know that, P-amost surey, 1 = W m = wa; i + 1, X T +m ) a C iy) a C iy) V53) Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

12 12 for a m 0 Hence, this must aso hod for the imits: P a C iy) V a = 1 ) = 1 This together with V53) proves V49) and hence ConvW) Now ConvQ) immediatey foows: from V47) we see that there is a random variabe a 0 such that P im Q Π t a 0 ; i + 1, y) = 1 ) = 1 V54) From h1) = 1) = 1 see Lemma 8 b)) and V48) we see that the same must hod for Q Πt a 0 ; i + 1, y) Thus we have proved ConvQ) and we are now going to show that this more precisey V54)) impies the assertion of the Lemma, ie X m n) i + 1) = a 0, n = 1,, N, for arge enough m, P-amost surey In other words, the next position i + 1 aso becomes identica in a sampes after finitey many iterations We write X ) t a 0, n = 1,, N that a sampes have the same i + 1-st i+1) a 0 for the event X n) t i+1) = component We then have ) P k N m k X m ) i+1) a 0 = im P X ) k i+1) a 0) m=k+1 V55) P X ) m i+1) a 0 X ) i+1) a 0, = k,, m 1 ) Now from V54), using bounded convergence, we see that for the first factor im P X ) k i+1) a 0) = im E Q Πk a 0 ; i+1, y) N) = E QΠk a 0 ; i+1, y) ) ) N = 1 V56) im For the second expression in V55), we obtain for m k T P X m ) i+1) a 0 ) X i+1) a 0, = k,, m 1 QΠm = E a 0 ; i+1, y) ) N X ) i+1) a 0, = k,, m 1 V57) We want to bound this ast expression from beow using Q Π m The condition X ) i+1) a 0, = k,, m 1, impies wa 0, i+1, X ) = 1 for = k,, m 1 Therefore, we have from Lemma 11 d) Q Π m a 0 ; i+1, y) 1 1 Q Π k a 0 ; i+1, y))1 ϱ) m k Using the bounds from Lemma 9 a) and Lemma 8 for m k T, we arrive at QΠm a 0 ; i+1, y) N Q Πm a 0 ; i+1, y) ) N 1 1 h Q a Πk 0; i+1, y) ) 1 ϱ) m k) N =: 1 H k,m k) N V58) where we introduced the abbreviation H k,m k) Note that H k,m k) is independent of the condition X ) i+1) a 0, = k,, m 1 Combining V57) and V58) we may bound the second term in V55) from beow by im E 1 H k,m k) ) N ) V59) m=k+1 im = ) N 1 EH k,m) 1 E im H k,m) ) ) N = 1, as im H k,m) = 0 amost surey by V54) To see that im and product may be interchanged, we use ogarithms and then bounded convergence with 0 n1 EH k,m ) n1 h1 ϱ) m ) for a k From V59), V56) and V55), we now see that there is a random variabe a 0 such that with probabiity one k N m k X ) m i+1) a 0, but this is the assertion of the Lemma We can now prove Theorem 5 about the convergence of the sampes Proof of Theorem 5: a) For i I, et Γ i denote the event y R i n = 1,, N and et γ i be the event a 0 A n = 1,, N im X n) t 1,, i) = y, im X t i) n) = a 0 In Lemma 12, we have shown that, if PΓ i ) = 1, then aso Pγ i+1 ) = 1 But, as Γ i γ i+1 Γ i+1, we see that PΓ i ) = 1 impies PΓ i+1 ) = 1 As a soutions have the empty substring as partia soution, we have PΓ 0 ) = 1, as was noted before But then we see from repeated appication of Lemma 12 that PΓ L ) = 1, ie convergence of the compete sampes hods with probabiity one b) We first show that convergence of sampes impies 1 ϱ ) < V60) Assume that convergence of sampes hods, then there must be at east one s = s 1,, s L ) S with 0 < P im X n) t = s for n = 1,, N ) P im X n) t 1) = s 1 for n = 1,, N ) = im P X ) k 1) s 1) m=k+1 P X m ) 1) s 1 X ) 1) s 1, = k,, m 1 im E h Q Π m s 1 ; 1, ) )) N m=k+1 ws 1 ; 1, X ) = 1, = k,, m 1 where we used the bound of Lemma 9 a) From Lemma 11 c) and V3) of Lemma 7, we see that E h Q Π m s 1 ; 1, ) )) N ws1 ; 1, X ) = 1, = k,, m 1 Copyright c) 2014 IEEE Persona use is permitted For any other purposes, permission must be obtained from the IEEE by emaiing pubs-permissions@ieeeorg

A Brief Introduction to Markov Chains and Hidden Markov Models

A Brief Introduction to Markov Chains and Hidden Markov Models A Brief Introduction to Markov Chains and Hidden Markov Modes Aen B MacKenzie Notes for December 1, 3, &8, 2015 Discrete-Time Markov Chains You may reca that when we first introduced random processes,

More information

MARKOV CHAINS AND MARKOV DECISION THEORY. Contents

MARKOV CHAINS AND MARKOV DECISION THEORY. Contents MARKOV CHAINS AND MARKOV DECISION THEORY ARINDRIMA DATTA Abstract. In this paper, we begin with a forma introduction to probabiity and expain the concept of random variabes and stochastic processes. After

More information

Problem set 6 The Perron Frobenius theorem.

Problem set 6 The Perron Frobenius theorem. Probem set 6 The Perron Frobenius theorem. Math 22a4 Oct 2 204, Due Oct.28 In a future probem set I want to discuss some criteria which aow us to concude that that the ground state of a sef-adjoint operator

More information

CS229 Lecture notes. Andrew Ng

CS229 Lecture notes. Andrew Ng CS229 Lecture notes Andrew Ng Part IX The EM agorithm In the previous set of notes, we taked about the EM agorithm as appied to fitting a mixture of Gaussians. In this set of notes, we give a broader view

More information

Efficiently Generating Random Bits from Finite State Markov Chains

Efficiently Generating Random Bits from Finite State Markov Chains 1 Efficienty Generating Random Bits from Finite State Markov Chains Hongchao Zhou and Jehoshua Bruck, Feow, IEEE Abstract The probem of random number generation from an uncorreated random source (of unknown

More information

XSAT of linear CNF formulas

XSAT of linear CNF formulas XSAT of inear CN formuas Bernd R. Schuh Dr. Bernd Schuh, D-50968 Kön, Germany; bernd.schuh@netcoogne.de eywords: compexity, XSAT, exact inear formua, -reguarity, -uniformity, NPcompeteness Abstract. Open

More information

Stochastic Variational Inference with Gradient Linearization

Stochastic Variational Inference with Gradient Linearization Stochastic Variationa Inference with Gradient Linearization Suppementa Materia Tobias Pötz * Anne S Wannenwetsch Stefan Roth Department of Computer Science, TU Darmstadt Preface In this suppementa materia,

More information

Efficient Generation of Random Bits from Finite State Markov Chains

Efficient Generation of Random Bits from Finite State Markov Chains Efficient Generation of Random Bits from Finite State Markov Chains Hongchao Zhou and Jehoshua Bruck, Feow, IEEE Abstract The probem of random number generation from an uncorreated random source (of unknown

More information

u(x) s.t. px w x 0 Denote the solution to this problem by ˆx(p, x). In order to obtain ˆx we may simply solve the standard problem max x 0

u(x) s.t. px w x 0 Denote the solution to this problem by ˆx(p, x). In order to obtain ˆx we may simply solve the standard problem max x 0 Bocconi University PhD in Economics - Microeconomics I Prof M Messner Probem Set 4 - Soution Probem : If an individua has an endowment instead of a monetary income his weath depends on price eves In particuar,

More information

Asynchronous Control for Coupled Markov Decision Systems

Asynchronous Control for Coupled Markov Decision Systems INFORMATION THEORY WORKSHOP (ITW) 22 Asynchronous Contro for Couped Marov Decision Systems Michae J. Neey University of Southern Caifornia Abstract This paper considers optima contro for a coection of

More information

C. Fourier Sine Series Overview

C. Fourier Sine Series Overview 12 PHILIP D. LOEWEN C. Fourier Sine Series Overview Let some constant > be given. The symboic form of the FSS Eigenvaue probem combines an ordinary differentia equation (ODE) on the interva (, ) with a

More information

A NOTE ON QUASI-STATIONARY DISTRIBUTIONS OF BIRTH-DEATH PROCESSES AND THE SIS LOGISTIC EPIDEMIC

A NOTE ON QUASI-STATIONARY DISTRIBUTIONS OF BIRTH-DEATH PROCESSES AND THE SIS LOGISTIC EPIDEMIC (January 8, 2003) A NOTE ON QUASI-STATIONARY DISTRIBUTIONS OF BIRTH-DEATH PROCESSES AND THE SIS LOGISTIC EPIDEMIC DAMIAN CLANCY, University of Liverpoo PHILIP K. POLLETT, University of Queensand Abstract

More information

Stochastic Complement Analysis of Multi-Server Threshold Queues. with Hysteresis. Abstract

Stochastic Complement Analysis of Multi-Server Threshold Queues. with Hysteresis. Abstract Stochastic Compement Anaysis of Muti-Server Threshod Queues with Hysteresis John C.S. Lui The Dept. of Computer Science & Engineering The Chinese University of Hong Kong Leana Goubchik Dept. of Computer

More information

NOISE-INDUCED STABILIZATION OF STOCHASTIC DIFFERENTIAL EQUATIONS

NOISE-INDUCED STABILIZATION OF STOCHASTIC DIFFERENTIAL EQUATIONS NOISE-INDUCED STABILIZATION OF STOCHASTIC DIFFERENTIAL EQUATIONS TONY ALLEN, EMILY GEBHARDT, AND ADAM KLUBALL 3 ADVISOR: DR. TIFFANY KOLBA 4 Abstract. The phenomenon of noise-induced stabiization occurs

More information

Explicit overall risk minimization transductive bound

Explicit overall risk minimization transductive bound 1 Expicit overa risk minimization transductive bound Sergio Decherchi, Paoo Gastado, Sandro Ridea, Rodofo Zunino Dept. of Biophysica and Eectronic Engineering (DIBE), Genoa University Via Opera Pia 11a,

More information

Cryptanalysis of PKP: A New Approach

Cryptanalysis of PKP: A New Approach Cryptanaysis of PKP: A New Approach Éiane Jaumes and Antoine Joux DCSSI 18, rue du Dr. Zamenhoff F-92131 Issy-es-Mx Cedex France eiane.jaumes@wanadoo.fr Antoine.Joux@ens.fr Abstract. Quite recenty, in

More information

SUPPLEMENTARY MATERIAL TO INNOVATED SCALABLE EFFICIENT ESTIMATION IN ULTRA-LARGE GAUSSIAN GRAPHICAL MODELS

SUPPLEMENTARY MATERIAL TO INNOVATED SCALABLE EFFICIENT ESTIMATION IN ULTRA-LARGE GAUSSIAN GRAPHICAL MODELS ISEE 1 SUPPLEMENTARY MATERIAL TO INNOVATED SCALABLE EFFICIENT ESTIMATION IN ULTRA-LARGE GAUSSIAN GRAPHICAL MODELS By Yingying Fan and Jinchi Lv University of Southern Caifornia This Suppementary Materia

More information

Approximate Bandwidth Allocation for Fixed-Priority-Scheduled Periodic Resources (WSU-CS Technical Report Version)

Approximate Bandwidth Allocation for Fixed-Priority-Scheduled Periodic Resources (WSU-CS Technical Report Version) Approximate Bandwidth Aocation for Fixed-Priority-Schedued Periodic Resources WSU-CS Technica Report Version) Farhana Dewan Nathan Fisher Abstract Recent research in compositiona rea-time systems has focused

More information

Homework 5 Solutions

Homework 5 Solutions Stat 310B/Math 230B Theory of Probabiity Homework 5 Soutions Andrea Montanari Due on 2/19/2014 Exercise [5.3.20] 1. We caim that n 2 [ E[h F n ] = 2 n i=1 A i,n h(u)du ] I Ai,n (t). (1) Indeed, integrabiity

More information

4 Separation of Variables

4 Separation of Variables 4 Separation of Variabes In this chapter we describe a cassica technique for constructing forma soutions to inear boundary vaue probems. The soution of three cassica (paraboic, hyperboic and eiptic) PDE

More information

Iterative Decoding Performance Bounds for LDPC Codes on Noisy Channels

Iterative Decoding Performance Bounds for LDPC Codes on Noisy Channels Iterative Decoding Performance Bounds for LDPC Codes on Noisy Channes arxiv:cs/060700v1 [cs.it] 6 Ju 006 Chun-Hao Hsu and Achieas Anastasopouos Eectrica Engineering and Computer Science Department University

More information

MATH 172: MOTIVATION FOR FOURIER SERIES: SEPARATION OF VARIABLES

MATH 172: MOTIVATION FOR FOURIER SERIES: SEPARATION OF VARIABLES MATH 172: MOTIVATION FOR FOURIER SERIES: SEPARATION OF VARIABLES Separation of variabes is a method to sove certain PDEs which have a warped product structure. First, on R n, a inear PDE of order m is

More information

THINKING IN PYRAMIDS

THINKING IN PYRAMIDS ECS 178 Course Notes THINKING IN PYRAMIDS Kenneth I. Joy Institute for Data Anaysis and Visuaization Department of Computer Science University of Caifornia, Davis Overview It is frequenty usefu to think

More information

A. Distribution of the test statistic

A. Distribution of the test statistic A. Distribution of the test statistic In the sequentia test, we first compute the test statistic from a mini-batch of size m. If a decision cannot be made with this statistic, we keep increasing the mini-batch

More information

arxiv: v1 [math.co] 17 Dec 2018

arxiv: v1 [math.co] 17 Dec 2018 On the Extrema Maximum Agreement Subtree Probem arxiv:1812.06951v1 [math.o] 17 Dec 2018 Aexey Markin Department of omputer Science, Iowa State University, USA amarkin@iastate.edu Abstract Given two phyogenetic

More information

Uniprocessor Feasibility of Sporadic Tasks with Constrained Deadlines is Strongly conp-complete

Uniprocessor Feasibility of Sporadic Tasks with Constrained Deadlines is Strongly conp-complete Uniprocessor Feasibiity of Sporadic Tasks with Constrained Deadines is Strongy conp-compete Pontus Ekberg and Wang Yi Uppsaa University, Sweden Emai: {pontus.ekberg yi}@it.uu.se Abstract Deciding the feasibiity

More information

Akaike Information Criterion for ANOVA Model with a Simple Order Restriction

Akaike Information Criterion for ANOVA Model with a Simple Order Restriction Akaike Information Criterion for ANOVA Mode with a Simpe Order Restriction Yu Inatsu * Department of Mathematics, Graduate Schoo of Science, Hiroshima University ABSTRACT In this paper, we consider Akaike

More information

On the Goal Value of a Boolean Function

On the Goal Value of a Boolean Function On the Goa Vaue of a Booean Function Eric Bach Dept. of CS University of Wisconsin 1210 W. Dayton St. Madison, WI 53706 Lisa Heerstein Dept of CSE NYU Schoo of Engineering 2 Metrotech Center, 10th Foor

More information

Approximated MLC shape matrix decomposition with interleaf collision constraint

Approximated MLC shape matrix decomposition with interleaf collision constraint Approximated MLC shape matrix decomposition with intereaf coision constraint Thomas Kainowski Antje Kiese Abstract Shape matrix decomposition is a subprobem in radiation therapy panning. A given fuence

More information

An Extension of Almost Sure Central Limit Theorem for Order Statistics

An Extension of Almost Sure Central Limit Theorem for Order Statistics An Extension of Amost Sure Centra Limit Theorem for Order Statistics T. Bin, P. Zuoxiang & S. Nadarajah First version: 6 December 2007 Research Report No. 9, 2007, Probabiity Statistics Group Schoo of

More information

THE REACHABILITY CONES OF ESSENTIALLY NONNEGATIVE MATRICES

THE REACHABILITY CONES OF ESSENTIALLY NONNEGATIVE MATRICES THE REACHABILITY CONES OF ESSENTIALLY NONNEGATIVE MATRICES by Michae Neumann Department of Mathematics, University of Connecticut, Storrs, CT 06269 3009 and Ronad J. Stern Department of Mathematics, Concordia

More information

DIGITAL FILTER DESIGN OF IIR FILTERS USING REAL VALUED GENETIC ALGORITHM

DIGITAL FILTER DESIGN OF IIR FILTERS USING REAL VALUED GENETIC ALGORITHM DIGITAL FILTER DESIGN OF IIR FILTERS USING REAL VALUED GENETIC ALGORITHM MIKAEL NILSSON, MATTIAS DAHL AND INGVAR CLAESSON Bekinge Institute of Technoogy Department of Teecommunications and Signa Processing

More information

Mat 1501 lecture notes, penultimate installment

Mat 1501 lecture notes, penultimate installment Mat 1501 ecture notes, penutimate instament 1. bounded variation: functions of a singe variabe optiona) I beieve that we wi not actuay use the materia in this section the point is mainy to motivate the

More information

MONTE CARLO SIMULATIONS

MONTE CARLO SIMULATIONS MONTE CARLO SIMULATIONS Current physics research 1) Theoretica 2) Experimenta 3) Computationa Monte Caro (MC) Method (1953) used to study 1) Discrete spin systems 2) Fuids 3) Poymers, membranes, soft matter

More information

First-Order Corrections to Gutzwiller s Trace Formula for Systems with Discrete Symmetries

First-Order Corrections to Gutzwiller s Trace Formula for Systems with Discrete Symmetries c 26 Noninear Phenomena in Compex Systems First-Order Corrections to Gutzwier s Trace Formua for Systems with Discrete Symmetries Hoger Cartarius, Jörg Main, and Günter Wunner Institut für Theoretische

More information

Approximated MLC shape matrix decomposition with interleaf collision constraint

Approximated MLC shape matrix decomposition with interleaf collision constraint Agorithmic Operations Research Vo.4 (29) 49 57 Approximated MLC shape matrix decomposition with intereaf coision constraint Antje Kiese and Thomas Kainowski Institut für Mathematik, Universität Rostock,

More information

More Scattering: the Partial Wave Expansion

More Scattering: the Partial Wave Expansion More Scattering: the Partia Wave Expansion Michae Fower /7/8 Pane Waves and Partia Waves We are considering the soution to Schrödinger s equation for scattering of an incoming pane wave in the z-direction

More information

Primal and dual active-set methods for convex quadratic programming

Primal and dual active-set methods for convex quadratic programming Math. Program., Ser. A 216) 159:469 58 DOI 1.17/s117-15-966-2 FULL LENGTH PAPER Prima and dua active-set methods for convex quadratic programming Anders Forsgren 1 Phiip E. Gi 2 Eizabeth Wong 2 Received:

More information

The EM Algorithm applied to determining new limit points of Mahler measures

The EM Algorithm applied to determining new limit points of Mahler measures Contro and Cybernetics vo. 39 (2010) No. 4 The EM Agorithm appied to determining new imit points of Maher measures by Souad E Otmani, Georges Rhin and Jean-Marc Sac-Épée Université Pau Veraine-Metz, LMAM,

More information

6 Wave Equation on an Interval: Separation of Variables

6 Wave Equation on an Interval: Separation of Variables 6 Wave Equation on an Interva: Separation of Variabes 6.1 Dirichet Boundary Conditions Ref: Strauss, Chapter 4 We now use the separation of variabes technique to study the wave equation on a finite interva.

More information

Bayesian Learning. You hear a which which could equally be Thanks or Tanks, which would you go with?

Bayesian Learning. You hear a which which could equally be Thanks or Tanks, which would you go with? Bayesian Learning A powerfu and growing approach in machine earning We use it in our own decision making a the time You hear a which which coud equay be Thanks or Tanks, which woud you go with? Combine

More information

K a,k minors in graphs of bounded tree-width *

K a,k minors in graphs of bounded tree-width * K a,k minors in graphs of bounded tree-width * Thomas Böhme Institut für Mathematik Technische Universität Imenau Imenau, Germany E-mai: tboehme@theoinf.tu-imenau.de and John Maharry Department of Mathematics

More information

Sequential Decoding of Polar Codes with Arbitrary Binary Kernel

Sequential Decoding of Polar Codes with Arbitrary Binary Kernel Sequentia Decoding of Poar Codes with Arbitrary Binary Kerne Vera Miosavskaya, Peter Trifonov Saint-Petersburg State Poytechnic University Emai: veram,petert}@dcn.icc.spbstu.ru Abstract The probem of efficient

More information

Unconditional security of differential phase shift quantum key distribution

Unconditional security of differential phase shift quantum key distribution Unconditiona security of differentia phase shift quantum key distribution Kai Wen, Yoshihisa Yamamoto Ginzton Lab and Dept of Eectrica Engineering Stanford University Basic idea of DPS-QKD Protoco. Aice

More information

Bourgain s Theorem. Computational and Metric Geometry. Instructor: Yury Makarychev. d(s 1, s 2 ).

Bourgain s Theorem. Computational and Metric Geometry. Instructor: Yury Makarychev. d(s 1, s 2 ). Bourgain s Theorem Computationa and Metric Geometry Instructor: Yury Makarychev 1 Notation Given a metric space (X, d) and S X, the distance from x X to S equas d(x, S) = inf d(x, s). s S The distance

More information

Recursive Constructions of Parallel FIFO and LIFO Queues with Switched Delay Lines

Recursive Constructions of Parallel FIFO and LIFO Queues with Switched Delay Lines Recursive Constructions of Parae FIFO and LIFO Queues with Switched Deay Lines Po-Kai Huang, Cheng-Shang Chang, Feow, IEEE, Jay Cheng, Member, IEEE, and Duan-Shin Lee, Senior Member, IEEE Abstract One

More information

Lower Bounds for the Relative Greedy Algorithm for Approximating Steiner Trees

Lower Bounds for the Relative Greedy Algorithm for Approximating Steiner Trees This paper appeared in: Networks 47:2 (2006), -5 Lower Bounds for the Reative Greed Agorithm for Approimating Steiner Trees Stefan Hougard Stefan Kirchner Humbodt-Universität zu Berin Institut für Informatik

More information

SydU STAT3014 (2015) Second semester Dr. J. Chan 18

SydU STAT3014 (2015) Second semester Dr. J. Chan 18 STAT3014/3914 Appied Stat.-Samping C-Stratified rand. sampe Stratified Random Samping.1 Introduction Description The popuation of size N is divided into mutuay excusive and exhaustive subpopuations caed

More information

Schedulability Analysis of Deferrable Scheduling Algorithms for Maintaining Real-Time Data Freshness

Schedulability Analysis of Deferrable Scheduling Algorithms for Maintaining Real-Time Data Freshness 1 Scheduabiity Anaysis of Deferrabe Scheduing Agorithms for Maintaining Rea- Data Freshness Song Han, Deji Chen, Ming Xiong, Kam-yiu Lam, Aoysius K. Mok, Krithi Ramamritham UT Austin, Emerson Process Management,

More information

Throughput Optimal Scheduling for Wireless Downlinks with Reconfiguration Delay

Throughput Optimal Scheduling for Wireless Downlinks with Reconfiguration Delay Throughput Optima Scheduing for Wireess Downinks with Reconfiguration Deay Vineeth Baa Sukumaran vineethbs@gmai.com Department of Avionics Indian Institute of Space Science and Technoogy. Abstract We consider

More information

Schedulability Analysis of Deferrable Scheduling Algorithms for Maintaining Real-Time Data Freshness

Schedulability Analysis of Deferrable Scheduling Algorithms for Maintaining Real-Time Data Freshness 1 Scheduabiity Anaysis of Deferrabe Scheduing Agorithms for Maintaining Rea-Time Data Freshness Song Han, Deji Chen, Ming Xiong, Kam-yiu Lam, Aoysius K. Mok, Krithi Ramamritham UT Austin, Emerson Process

More information

Course 2BA1, Section 11: Periodic Functions and Fourier Series

Course 2BA1, Section 11: Periodic Functions and Fourier Series Course BA, 8 9 Section : Periodic Functions and Fourier Series David R. Wikins Copyright c David R. Wikins 9 Contents Periodic Functions and Fourier Series 74. Fourier Series of Even and Odd Functions...........

More information

Separation of Variables and a Spherical Shell with Surface Charge

Separation of Variables and a Spherical Shell with Surface Charge Separation of Variabes and a Spherica She with Surface Charge In cass we worked out the eectrostatic potentia due to a spherica she of radius R with a surface charge density σθ = σ cos θ. This cacuation

More information

Turbo Codes. Coding and Communication Laboratory. Dept. of Electrical Engineering, National Chung Hsing University

Turbo Codes. Coding and Communication Laboratory. Dept. of Electrical Engineering, National Chung Hsing University Turbo Codes Coding and Communication Laboratory Dept. of Eectrica Engineering, Nationa Chung Hsing University Turbo codes 1 Chapter 12: Turbo Codes 1. Introduction 2. Turbo code encoder 3. Design of intereaver

More information

Rate-Distortion Theory of Finite Point Processes

Rate-Distortion Theory of Finite Point Processes Rate-Distortion Theory of Finite Point Processes Günther Koiander, Dominic Schuhmacher, and Franz Hawatsch, Feow, IEEE Abstract We study the compression of data in the case where the usefu information

More information

STA 216 Project: Spline Approach to Discrete Survival Analysis

STA 216 Project: Spline Approach to Discrete Survival Analysis : Spine Approach to Discrete Surviva Anaysis November 4, 005 1 Introduction Athough continuous surviva anaysis differs much from the discrete surviva anaysis, there is certain ink between the two modeing

More information

A proposed nonparametric mixture density estimation using B-spline functions

A proposed nonparametric mixture density estimation using B-spline functions A proposed nonparametric mixture density estimation using B-spine functions Atizez Hadrich a,b, Mourad Zribi a, Afif Masmoudi b a Laboratoire d Informatique Signa et Image de a Côte d Opae (LISIC-EA 4491),

More information

Pairwise RNA Edit Distance

Pairwise RNA Edit Distance Pairwise RNA Edit Distance In the foowing: Sequences S 1 and S 2 associated structures P 1 and P 2 scoring of aignment: different edit operations arc atering arc removing 1) ACGUUGACUGACAACAC..(((...)))...

More information

Gauss Law. 2. Gauss s Law: connects charge and field 3. Applications of Gauss s Law

Gauss Law. 2. Gauss s Law: connects charge and field 3. Applications of Gauss s Law Gauss Law 1. Review on 1) Couomb s Law (charge and force) 2) Eectric Fied (fied and force) 2. Gauss s Law: connects charge and fied 3. Appications of Gauss s Law Couomb s Law and Eectric Fied Couomb s

More information

Analysis of Emerson s Multiple Model Interpolation Estimation Algorithms: The MIMO Case

Analysis of Emerson s Multiple Model Interpolation Estimation Algorithms: The MIMO Case Technica Report PC-04-00 Anaysis of Emerson s Mutipe Mode Interpoation Estimation Agorithms: The MIMO Case João P. Hespanha Dae E. Seborg University of Caifornia, Santa Barbara February 0, 004 Anaysis

More information

Lecture Note 3: Stationary Iterative Methods

Lecture Note 3: Stationary Iterative Methods MATH 5330: Computationa Methods of Linear Agebra Lecture Note 3: Stationary Iterative Methods Xianyi Zeng Department of Mathematica Sciences, UTEP Stationary Iterative Methods The Gaussian eimination (or

More information

The Group Structure on a Smooth Tropical Cubic

The Group Structure on a Smooth Tropical Cubic The Group Structure on a Smooth Tropica Cubic Ethan Lake Apri 20, 2015 Abstract Just as in in cassica agebraic geometry, it is possibe to define a group aw on a smooth tropica cubic curve. In this note,

More information

6.434J/16.391J Statistics for Engineers and Scientists May 4 MIT, Spring 2006 Handout #17. Solution 7

6.434J/16.391J Statistics for Engineers and Scientists May 4 MIT, Spring 2006 Handout #17. Solution 7 6.434J/16.391J Statistics for Engineers and Scientists May 4 MIT, Spring 2006 Handout #17 Soution 7 Probem 1: Generating Random Variabes Each part of this probem requires impementation in MATLAB. For the

More information

Integrality ratio for Group Steiner Trees and Directed Steiner Trees

Integrality ratio for Group Steiner Trees and Directed Steiner Trees Integraity ratio for Group Steiner Trees and Directed Steiner Trees Eran Haperin Guy Kortsarz Robert Krauthgamer Aravind Srinivasan Nan Wang Abstract The natura reaxation for the Group Steiner Tree probem,

More information

THE THREE POINT STEINER PROBLEM ON THE FLAT TORUS: THE MINIMAL LUNE CASE

THE THREE POINT STEINER PROBLEM ON THE FLAT TORUS: THE MINIMAL LUNE CASE THE THREE POINT STEINER PROBLEM ON THE FLAT TORUS: THE MINIMAL LUNE CASE KATIE L. MAY AND MELISSA A. MITCHELL Abstract. We show how to identify the minima path network connecting three fixed points on

More information

Math 124B January 17, 2012

Math 124B January 17, 2012 Math 124B January 17, 212 Viktor Grigoryan 3 Fu Fourier series We saw in previous ectures how the Dirichet and Neumann boundary conditions ead to respectivey sine and cosine Fourier series of the initia

More information

Power Control and Transmission Scheduling for Network Utility Maximization in Wireless Networks

Power Control and Transmission Scheduling for Network Utility Maximization in Wireless Networks ower Contro and Transmission Scheduing for Network Utiity Maximization in Wireess Networks Min Cao, Vivek Raghunathan, Stephen Hany, Vinod Sharma and. R. Kumar Abstract We consider a joint power contro

More information

Lecture 6: Moderately Large Deflection Theory of Beams

Lecture 6: Moderately Large Deflection Theory of Beams Structura Mechanics 2.8 Lecture 6 Semester Yr Lecture 6: Moderatey Large Defection Theory of Beams 6.1 Genera Formuation Compare to the cassica theory of beams with infinitesima deformation, the moderatey

More information

Expectation-Maximization for Estimating Parameters for a Mixture of Poissons

Expectation-Maximization for Estimating Parameters for a Mixture of Poissons Expectation-Maximization for Estimating Parameters for a Mixture of Poissons Brandon Maone Department of Computer Science University of Hesini February 18, 2014 Abstract This document derives, in excrutiating

More information

Some Measures for Asymmetry of Distributions

Some Measures for Asymmetry of Distributions Some Measures for Asymmetry of Distributions Georgi N. Boshnakov First version: 31 January 2006 Research Report No. 5, 2006, Probabiity and Statistics Group Schoo of Mathematics, The University of Manchester

More information

Reichenbachian Common Cause Systems

Reichenbachian Common Cause Systems Reichenbachian Common Cause Systems G. Hofer-Szabó Department of Phiosophy Technica University of Budapest e-mai: gszabo@hps.ete.hu Mikós Rédei Department of History and Phiosophy of Science Eötvös University,

More information

Wave Equation Dirichlet Boundary Conditions

Wave Equation Dirichlet Boundary Conditions Wave Equation Dirichet Boundary Conditions u tt x, t = c u xx x, t, < x 1 u, t =, u, t = ux, = fx u t x, = gx Look for simpe soutions in the form ux, t = XxT t Substituting into 13 and dividing

More information

Lecture 9. Stability of Elastic Structures. Lecture 10. Advanced Topic in Column Buckling

Lecture 9. Stability of Elastic Structures. Lecture 10. Advanced Topic in Column Buckling Lecture 9 Stabiity of Eastic Structures Lecture 1 Advanced Topic in Coumn Bucking robem 9-1: A camped-free coumn is oaded at its tip by a oad. The issue here is to find the itica bucking oad. a) Suggest

More information

Melodic contour estimation with B-spline models using a MDL criterion

Melodic contour estimation with B-spline models using a MDL criterion Meodic contour estimation with B-spine modes using a MDL criterion Damien Loive, Ney Barbot, Oivier Boeffard IRISA / University of Rennes 1 - ENSSAT 6 rue de Kerampont, B.P. 80518, F-305 Lannion Cedex

More information

T.C. Banwell, S. Galli. {bct, Telcordia Technologies, Inc., 445 South Street, Morristown, NJ 07960, USA

T.C. Banwell, S. Galli. {bct, Telcordia Technologies, Inc., 445 South Street, Morristown, NJ 07960, USA ON THE SYMMETRY OF THE POWER INE CHANNE T.C. Banwe, S. Gai {bct, sgai}@research.tecordia.com Tecordia Technoogies, Inc., 445 South Street, Morristown, NJ 07960, USA Abstract The indoor power ine network

More information

BALANCING REGULAR MATRIX PENCILS

BALANCING REGULAR MATRIX PENCILS BALANCING REGULAR MATRIX PENCILS DAMIEN LEMONNIER AND PAUL VAN DOOREN Abstract. In this paper we present a new diagona baancing technique for reguar matrix pencis λb A, which aims at reducing the sensitivity

More information

JENSEN S OPERATOR INEQUALITY FOR FUNCTIONS OF SEVERAL VARIABLES

JENSEN S OPERATOR INEQUALITY FOR FUNCTIONS OF SEVERAL VARIABLES PROCEEDINGS OF THE AMERICAN MATHEMATICAL SOCIETY Voume 128, Number 7, Pages 2075 2084 S 0002-99390005371-5 Artice eectronicay pubished on February 16, 2000 JENSEN S OPERATOR INEQUALITY FOR FUNCTIONS OF

More information

Optimal Control of Assembly Systems with Multiple Stages and Multiple Demand Classes 1

Optimal Control of Assembly Systems with Multiple Stages and Multiple Demand Classes 1 Optima Contro of Assemby Systems with Mutipe Stages and Mutipe Demand Casses Saif Benjaafar Mohsen EHafsi 2 Chung-Yee Lee 3 Weihua Zhou 3 Industria & Systems Engineering, Department of Mechanica Engineering,

More information

Symbolic models for nonlinear control systems using approximate bisimulation

Symbolic models for nonlinear control systems using approximate bisimulation Symboic modes for noninear contro systems using approximate bisimuation Giordano Poa, Antoine Girard and Pauo Tabuada Abstract Contro systems are usuay modeed by differentia equations describing how physica

More information

The distribution of the number of nodes in the relative interior of the typical I-segment in homogeneous planar anisotropic STIT Tessellations

The distribution of the number of nodes in the relative interior of the typical I-segment in homogeneous planar anisotropic STIT Tessellations Comment.Math.Univ.Caroin. 51,3(21) 53 512 53 The distribution of the number of nodes in the reative interior of the typica I-segment in homogeneous panar anisotropic STIT Tesseations Christoph Thäe Abstract.

More information

8 APPENDIX. E[m M] = (n S )(1 exp( exp(s min + c M))) (19) E[m M] n exp(s min + c M) (20) 8.1 EMPIRICAL EVALUATION OF SAMPLING

8 APPENDIX. E[m M] = (n S )(1 exp( exp(s min + c M))) (19) E[m M] n exp(s min + c M) (20) 8.1 EMPIRICAL EVALUATION OF SAMPLING 8 APPENDIX 8.1 EMPIRICAL EVALUATION OF SAMPLING We wish to evauate the empirica accuracy of our samping technique on concrete exampes. We do this in two ways. First, we can sort the eements by probabiity

More information

12.2. Maxima and Minima. Introduction. Prerequisites. Learning Outcomes

12.2. Maxima and Minima. Introduction. Prerequisites. Learning Outcomes Maima and Minima 1. Introduction In this Section we anayse curves in the oca neighbourhood of a stationary point and, from this anaysis, deduce necessary conditions satisfied by oca maima and oca minima.

More information

Statistical Learning Theory: A Primer

Statistical Learning Theory: A Primer Internationa Journa of Computer Vision 38(), 9 3, 2000 c 2000 uwer Academic Pubishers. Manufactured in The Netherands. Statistica Learning Theory: A Primer THEODOROS EVGENIOU, MASSIMILIANO PONTIL AND TOMASO

More information

PHYS 110B - HW #1 Fall 2005, Solutions by David Pace Equations referenced as Eq. # are from Griffiths Problem statements are paraphrased

PHYS 110B - HW #1 Fall 2005, Solutions by David Pace Equations referenced as Eq. # are from Griffiths Problem statements are paraphrased PHYS 110B - HW #1 Fa 2005, Soutions by David Pace Equations referenced as Eq. # are from Griffiths Probem statements are paraphrased [1.] Probem 6.8 from Griffiths A ong cyinder has radius R and a magnetization

More information

8 Digifl'.11 Cth:uits and devices

8 Digifl'.11 Cth:uits and devices 8 Digif'. Cth:uits and devices 8. Introduction In anaog eectronics, votage is a continuous variabe. This is usefu because most physica quantities we encounter are continuous: sound eves, ight intensity,

More information

<C 2 2. λ 2 l. λ 1 l 1 < C 1

<C 2 2. λ 2 l. λ 1 l 1 < C 1 Teecommunication Network Contro and Management (EE E694) Prof. A. A. Lazar Notes for the ecture of 7/Feb/95 by Huayan Wang (this document was ast LaT E X-ed on May 9,995) Queueing Primer for Muticass Optima

More information

$, (2.1) n="# #. (2.2)

$, (2.1) n=# #. (2.2) Chapter. Eectrostatic II Notes: Most of the materia presented in this chapter is taken from Jackson, Chap.,, and 4, and Di Bartoo, Chap... Mathematica Considerations.. The Fourier series and the Fourier

More information

A CLUSTERING LAW FOR SOME DISCRETE ORDER STATISTICS

A CLUSTERING LAW FOR SOME DISCRETE ORDER STATISTICS J App Prob 40, 226 241 (2003) Printed in Israe Appied Probabiity Trust 2003 A CLUSTERING LAW FOR SOME DISCRETE ORDER STATISTICS SUNDER SETHURAMAN, Iowa State University Abstract Let X 1,X 2,,X n be a sequence

More information

STABLE GRAPHS BENJAMIN OYE

STABLE GRAPHS BENJAMIN OYE STABLE GRAPHS BENJAMIN OYE Abstract. In Reguarity Lemmas for Stabe Graphs [1] Maiaris and Sheah appy toos from mode theory to obtain stronger forms of Ramsey's theorem and Szemeredi's reguarity emma for

More information

Determining The Degree of Generalization Using An Incremental Learning Algorithm

Determining The Degree of Generalization Using An Incremental Learning Algorithm Determining The Degree of Generaization Using An Incrementa Learning Agorithm Pabo Zegers Facutad de Ingeniería, Universidad de os Andes San Caros de Apoquindo 22, Las Condes, Santiago, Chie pzegers@uandes.c

More information

Maximum likelihood decoding of trellis codes in fading channels with no receiver CSI is a polynomial-complexity problem

Maximum likelihood decoding of trellis codes in fading channels with no receiver CSI is a polynomial-complexity problem 1 Maximum ikeihood decoding of treis codes in fading channes with no receiver CSI is a poynomia-compexity probem Chun-Hao Hsu and Achieas Anastasopouos Eectrica Engineering and Computer Science Department

More information

Research Article On the Lower Bound for the Number of Real Roots of a Random Algebraic Equation

Research Article On the Lower Bound for the Number of Real Roots of a Random Algebraic Equation Appied Mathematics and Stochastic Anaysis Voume 007, Artice ID 74191, 8 pages doi:10.1155/007/74191 Research Artice On the Lower Bound for the Number of Rea Roots of a Random Agebraic Equation Takashi

More information

Componentwise Determination of the Interval Hull Solution for Linear Interval Parameter Systems

Componentwise Determination of the Interval Hull Solution for Linear Interval Parameter Systems Componentwise Determination of the Interva Hu Soution for Linear Interva Parameter Systems L. V. Koev Dept. of Theoretica Eectrotechnics, Facuty of Automatics, Technica University of Sofia, 1000 Sofia,

More information

Convergence Property of the Iri-Imai Algorithm for Some Smooth Convex Programming Problems

Convergence Property of the Iri-Imai Algorithm for Some Smooth Convex Programming Problems Convergence Property of the Iri-Imai Agorithm for Some Smooth Convex Programming Probems S. Zhang Communicated by Z.Q. Luo Assistant Professor, Department of Econometrics, University of Groningen, Groningen,

More information

1D Heat Propagation Problems

1D Heat Propagation Problems Chapter 1 1D Heat Propagation Probems If the ambient space of the heat conduction has ony one dimension, the Fourier equation reduces to the foowing for an homogeneous body cρ T t = T λ 2 + Q, 1.1) x2

More information

Partial permutation decoding for MacDonald codes

Partial permutation decoding for MacDonald codes Partia permutation decoding for MacDonad codes J.D. Key Department of Mathematics and Appied Mathematics University of the Western Cape 7535 Bevie, South Africa P. Seneviratne Department of Mathematics

More information

A Novel Learning Method for Elman Neural Network Using Local Search

A Novel Learning Method for Elman Neural Network Using Local Search Neura Information Processing Letters and Reviews Vo. 11, No. 8, August 2007 LETTER A Nove Learning Method for Eman Neura Networ Using Loca Search Facuty of Engineering, Toyama University, Gofuu 3190 Toyama

More information

arxiv: v1 [math.fa] 23 Aug 2018

arxiv: v1 [math.fa] 23 Aug 2018 An Exact Upper Bound on the L p Lebesgue Constant and The -Rényi Entropy Power Inequaity for Integer Vaued Random Variabes arxiv:808.0773v [math.fa] 3 Aug 08 Peng Xu, Mokshay Madiman, James Mebourne Abstract

More information

4 1-D Boundary Value Problems Heat Equation

4 1-D Boundary Value Problems Heat Equation 4 -D Boundary Vaue Probems Heat Equation The main purpose of this chapter is to study boundary vaue probems for the heat equation on a finite rod a x b. u t (x, t = ku xx (x, t, a < x < b, t > u(x, = ϕ(x

More information

FRIEZE GROUPS IN R 2

FRIEZE GROUPS IN R 2 FRIEZE GROUPS IN R 2 MAXWELL STOLARSKI Abstract. Focusing on the Eucidean pane under the Pythagorean Metric, our goa is to cassify the frieze groups, discrete subgroups of the set of isometries of the

More information