Lectures on Stochastic Stability. Sergey FOSS. Heriot-Watt University. Lecture 5. Monotonicity and Saturation Rule

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

Math 426: Probability MWF 1pm, Gasson 310 Homework 4 Selected Solutions

More metrics on cartesian products

Continuous Time Markov Chain

Appendix B. Criterion of Riemann-Stieltjes Integrability

Bounds on the bias terms for the Markov reward approach

Markov chains. Definition of a CTMC: [2, page 381] is a continuous time, discrete value random process such that for an infinitesimal

The Order Relation and Trace Inequalities for. Hermitian Operators

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

APPENDIX A Some Linear Algebra

First day August 1, Problems and Solutions

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

Applied Stochastic Processes

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

Analysis of Discrete Time Queues (Section 4.6)

TAIL BOUNDS FOR SUMS OF GEOMETRIC AND EXPONENTIAL VARIABLES

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

Lecture 4: September 12

Exercise Solutions to Real Analysis

Randić Energy and Randić Estrada Index of a Graph

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

REAL ANALYSIS I HOMEWORK 1

LARGEST WEIGHTED DELAY FIRST SCHEDULING: LARGE DEVIATIONS AND OPTIMALITY. By Alexander L. Stolyar and Kavita Ramanan Bell Labs

STEINHAUS PROPERTY IN BANACH LATTICES

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

EXPONENTIAL ERGODICITY FOR SINGLE-BIRTH PROCESSES

TCOM 501: Networking Theory & Fundamentals. Lecture 7 February 25, 2003 Prof. Yannis A. Korilis

Canonical transformations

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

AN EXTENDED CLASS OF TIME-CONTINUOUS BRANCHING PROCESSES. Rong-Rong Chen. ( University of Illinois at Urbana-Champaign)

SL n (F ) Equals its Own Derived Group

Equilibrium Analysis of the M/G/1 Queue

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

A combinatorial problem associated with nonograms

Convexity preserving interpolation by splines of arbitrary degree

Simultaneous Optimization of Berth Allocation, Quay Crane Assignment and Quay Crane Scheduling Problems in Container Terminals

arxiv: v1 [math.co] 1 Mar 2014

Case Study of Markov Chains Ray-Knight Compactification

Suggested solutions for the exam in SF2863 Systems Engineering. June 12,

a b a In case b 0, a being divisible by b is the same as to say that

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

Exercises of Chapter 2

Foundations of Arithmetic

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

Lecture 21: Numerical methods for pricing American type derivatives

2.3 Nilpotent endomorphisms

FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP

BOUNDEDNESS OF THE RIESZ TRANSFORM WITH MATRIX A 2 WEIGHTS

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

Modelli Clamfim Equazioni differenziali 7 ottobre 2013

Hidden Markov Models

Affine transformations and convexity

Edge Isoperimetric Inequalities

Randomness and Computation

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

Lecture 4: Universal Hash Functions/Streaming Cont d

6. Stochastic processes (2)

Short running title: A generating function approach A GENERATING FUNCTION APPROACH TO COUNTING THEOREMS FOR SQUARE-FREE POLYNOMIALS AND MAXIMAL TORI

ELASTIC WAVE PROPAGATION IN A CONTINUOUS MEDIUM

6. Stochastic processes (2)

Window Flow Control in FIFO Networks with Cross Traffic

Google PageRank with Stochastic Matrix

TAIL PROBABILITIES OF RANDOMLY WEIGHTED SUMS OF RANDOM VARIABLES WITH DOMINATED VARIATION

Min Cut, Fast Cut, Polynomial Identities

k t+1 + c t A t k t, t=0

Notes on Frequency Estimation in Data Streams

Complete subgraphs in multipartite graphs

Modelli Clamfim Equazione del Calore Lezione ottobre 2014

A Hybrid Variational Iteration Method for Blasius Equation

DIFFERENTIAL FORMS BRIAN OSSERMAN

Lecture 3: Probability Distributions

Math 217 Fall 2013 Homework 2 Solutions

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

Difference Equations

Queueing Networks II Network Performance

A new Approach for Solving Linear Ordinary Differential Equations

Application of Queuing Theory to Waiting Time of Out-Patients in Hospitals.

Asymptotics of the Solution of a Boundary Value. Problem for One-Characteristic Differential. Equation Degenerating into a Parabolic Equation

Lecture 12: Discrete Laplacian

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

Salmon: Lectures on partial differential equations. Consider the general linear, second-order PDE in the form. ,x 2

Appendix B. The Finite Difference Scheme

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE

9 Characteristic classes

Convergence of random processes

PHYS 705: Classical Mechanics. Canonical Transformation II

The equation of motion of a dynamical system is given by a set of differential equations. That is (1)

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

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that

Dirichlet s Theorem In Arithmetic Progressions

Another converse of Jensen s inequality

Maximizing the number of nonnegative subsets

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

Numerical Heat and Mass Transfer

Expected Value and Variance

A combinatorial proof of multiple angle formulas involving Fibonacci and Lucas numbers

MAT 578 Functional Analysis

Infinitely Split Nash Equilibrium Problems in Repeated Games

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

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

Transcription:

Lectures on Stochastc Stablty Sergey FOSS Herot-Watt Unversty Lecture 5 Monotoncty and Saturaton Rule Introducton The paper of Loynes [8] was the frst to consder a system (sngle server queue, and, later, queues n tandem wth statonary-ergodc drver. The classcal recurson X n+ = (X n + ξ n + studed by Loynes s monotone. There s a varety of monotone models (ncludng queues n tandem, mult-server queues, Jackson-type networks, etc. for whch one can develop a unfed approach for stablty study. Here we provde a short survey, see the references for more detals. 2 Sngle-server queue revsted Consder a sngle server queue wth nterarrval tmes σ n and servce tmes t n. Assume that a sequence {(σ n, t n } s statonary wth fnte means b = Eσ and a = Et and satsfes the SLLN,.e. n σ = b and n t = a a.s. ( A suffcent condton for ( to hold (but not necessary n general see, e.g., a dscusson n [6] s that a sequence {(σ n, t n } s ergodc (see, e.g., Lecture for defntons. Assume further that b < a. Assume also, for smplcty, that customer arrves n an empty queue. Let W n be a watng tme of customer n. Then W = 0 and Also, {W n } satsfy the followng recurson n W n = max (σ t. n = W n+ = max(0, W n + σ n t n. (2

Note that W n concdes n dstrbuton wth W n θ = max + 0 = (σ t. The latter sequence ncreases a.s. and couples wth an a.s. fnte lmt W 0 = sup 0 = (σ t due to the SLLN (. One can also defne, for any m, Then {W m } s a statonary sequence and W m = sup (σ t. m = W m+ = max(0, W m + σ m t m. Thus, ths s a statonary soluton to recursve equaton (2. Exercse. Show that ths s the only statonary soluton. 3 Tandem of two sngle-server queues Tandem queue wll be our toy example. Consder an open network wth two sngle-server statons n tandem. Customers arrve to the frst staton wth nterarrval tmes {t n } and form a queue there. In ths example, t s convenent to assume that t n s an nterarrval tme between customers n and n. A server serves them n order of arrval wth servce tmes {σ n ( }. Upon servce completon, customers go to the second staton where are served also n order of arrval wth servce tmes {σ n (2 }. Assume that customer arrves n an empty system. Denote by Z n a soourn tme of customer n. Then Z = σ ( + σ (2 and, more generally, for any n, Z n = max k m n Exercse 2. Prove formula (3. ( m k σ ( + n m σ (2 n t It s known that f a sequence {(σ n (, σ n (2, t n } s statonary ergodc and f max(b (, b (2 a k+ (3 < (4 where b ( = Eσ ( and a = Et, then a dstrbuton of Z n converges to a lmtng statonary dstrrbuon n the total varaton norm. 2

Exercse 3. Show that f there s the opposte strct nequalty n (4, then a sequence {Z n } tends to nfnty a.s. In order to ntroduce a recursve scheme, let Z n ( be a soourn tme of customer n n the frst system (.e. a sum of a watng tme and of servce tme σ n (. Then Z ( = σ ( and we get the followng recursve relatons: Z ( n+ = max(0, Z ( n t n+ + σ ( n+, Z n+ = max(z ( n+, Z n t n+ + σ (2 n+. In other words, ntroduce a functon f : R 5 R 2 as follows: f(x, x 2, y, y 2, y 3 = (max(0, x y 3 + y, max(max(0, x y 3 + y, x 2 y 3 + y 2. Then f s monotone non-decreasng n (x, x 2, and we get a recurson where ξ n+ = (σ ( n+, σ(2 n+, t n+. (Z ( n+, Z n+ = f((z ( n, Z n, ξ n+ 4 General statements on monotone and homogeneous recursons There are many applcatons, especally n queueng networks, where monotoncty n the dynamcs can be exploted to prove exstence and unqueness of statonary solutons. Although the theory can be presented n the very general setup of a partally ordered state space (see Brandt et al. [6] we wll only focus on the case where the state s R d. Consder then the SRS X n+ = f(x n, ξ n+ =: ϕ n+ (X n and assume that ϕ 0 : R d + R d + s ncreasng and rght-contnuous, where the orderng s the standard component-wse orderng on R d +. Let θ be statonary and ergodc flow on (Ω, F, P and assume that ϕ n = ϕ 0 θ n, n Z. In other words, {ϕ n } s a statonary-ergodc sequence of random elements of the space of rght-contnuous ncreasng functons on R d +. We frst explan Loynes method. Defne Φ n := ϕ n ϕ. Thus, Φ n (Y s the soluton of the SRS at n 0 when X 0 = Y, a.s. Snce 0 s the least element of (R d +,, we have Φ n (0 Φ n (Y, a.s., for any R d + valued r.v. Y. Next consder Φ m+n (0 θ m = ϕ n ϕ m+ (0, n m, and nterpret Φ m+n (0 as the soluton of the SRS at tme n m, startng wth 0 at tme m. Clearly, Φ m+n (0 ncreases as m ncreases, because: Φ (m++n (0 θ (m+ = ϕ n ϕ m+ ϕ (m+ (0 = ϕ n ϕ m+ (ϕ (m+ (0 ϕ n ϕ m+ (0 = Φ m+n (0 θ m. 3

Fnally defne X n := lm m Φ m+n(0 θ m, n Z. The r.v. X n s ether fnte a.s., or s nfnte a.s., by ergodcty. Assumng that the frst case holds, we further have X n+ = lm m Φ m+(n+(0 θ m = lm m ϕ m+n+ϕ m+n ϕ (0 θ m = lm m ϕ n+ϕ n ϕ m+ (0 = lm m ϕ n+(ϕ n ϕ m+ (0 = lm m ϕ n+(φ m+n (0 θ m+ = ϕ n+ ( X n. Provdes then that we have a method for provng P(X 0 < > 0, Loynes technque results n the constructon of a statonary-ergodc soluton { X n } of the SRS. For a tandem queue, we get X n = (Z n (, Z n θ = σ ( 0 + max + 0 (σ ( = t, max + k 0 ( k σ ( + k σ (2 + t and clearly ths sequence ncreases a.s. Here, by conventon, r... = 0 s > r. Consder another example of a multserver queue. Example. A multserver queue G/G/s wth s servers and FCFS servce dscplne. Customers arrve wth nterarrval tmes {t n } and have servce tmes {σ n } (servce tmes are assocated wth customers, not wth servers. Upon arrval to the system, a customer s mmedately sent to a server (one of servers wth a mnmal workload. At each server, customers are served n order of arrval. Denote by V n, a workload of server ust before arrval of nth customer. Let R be an operator that orders coordnates of a vector n the non-decreasng order. Let W n = R(V n,,... V n,s. Then vectors {W n } satsfy the followng recursve equaton ( Kefer-Wolfowtz : W n+ = R(W n + e σ n t n+ + Here x + = max(x, 0 (coordnatewse, and e = (, 0,..., 0 and = (,,...,. Exercse 4. Show the monotoncty of the latter recurson. Return to the general settng. Wthout further assumptons and structure, not much can be sad. Assume next that, n addton, ϕ 0 s homogeneous,.e., ϕ 0 (x + c = ϕ 0 (x + c, 4

for all x R d + and all c R. Such s the case, e.g., wth the usual Lndley functon ϕ 0 : R + R +, wth ϕ 0 (x = max(x+ξ 0, 0. The homogenety assumpton s qute frequent n queueng theory. It s easy to see that ϕ 0 (x ϕ 0 (y x y, where x := max( x,..., x d. Suppose then that {X n }, {Y n } are two statonary solutons of the SRS. Then X n+ Y n+ = ϕ n+ (X n ϕ n+ (Y n X n Y n, (5 for all n, a.s., and snce { X n Y n, n Z} s statonary and ergodc, ths a.s. monotoncty may only hold f X n Y n = r, for some constant r 0. Thus, a necessary and suffcent condton for the two solutons to concde s that, P( ϕ (X 0 ϕ (Y 0 < X 0 Y 0 > 0. (6 Remark. Condtons (5 and (6 form the basc for the so-called contracton approach. A classcal example where, n general, (5 holds but (6 fals s the G/G/s queue, that s, the s-server queue wth statonary-ergodc data. Let λ, µ be the arrval and servce rates, respectvely. Here, there s a mnmal and a maxmal statonary soluton whch, provded that λ < sµ, may not concde. For detals see Brandt et al [6]. But condton (6 holds f the drvng sequences are..d. (or satsfy a weaker mxng condton. Also condton (6 holds for the tandem queue wth statonary and ergodc drvng sequences. Ths s Exercse 5 to you. 5 The Monotone-Homogeneous-Separable (MHS framework Consder a recurson of the form W n+ = f(w n, ξ n+, τ n+, where ξ n are general marks, and τ n 0. The nterpretaton s that τ n s the nterarrval tme between the n -th and n-th customer, and W n s the state ust after the arrval of the n-th customer. We consder arrval epochs {T n } such that T n+ T n = τ n. We wrte W m,n for the soluton of the recurson at ndex n when we start wth a specfc state, say 0, at m n. Fnally we consder a functons of the form X [m,n] = f m+n (W m,n ; T m,..., T n ; ξ m+,..., ξ n, whch wll be thought of as epochs of last actvty n the system. For nstance, when we have an s-server queue, X[m, n] represents the departure tme of the last customer when the queue s fed only by customers wth ndces from m to n. Correspondngly, we defne the quantty Z [m,n] := X [m,n] T n, the tme elapsed between the arrval of the last customer and the departure of the last customer. The framework s formulated n terms of the X [m,n], Z [m,n] and ther dependence on the {T n }. For c R, let {T n } + c = {T n + c}. For c > 0, let c{t n } = {ct n }. Defne {T n } {T n} f T n T n for all n. We requre a set of four assumptons: 5

(A Z [m,n] 0 (A2 {T n } {T n} X [m,n] X [m,n]. The frst assumpton s natural. In the second one, X [m,n] are the varables obtaned by replacng each T n by T n; t says that delayng the arrval epochs results n delayng of the last actvty epochs. (A3 {T n} = {T n } + c X [m,n] = X [m,n] + c. Ths s a tme-homogenety assumpton. (A4 For m l < l + n, X [m,l] T l+ X [m,n] = X [l+,n]. If the premse X [m,l] T l+ of the last assumpton holds, we say that we have separablty at ndex l. It means that the last actvty due to customers wth ndces n [m, l] happens pror to the arrval of the l + -th customer, and so the last actvty due to customers wth ndces n [m, n] s not nfluenced by those customers wth ndces n [m, l]. Basc consequences of the above assumptons are summarzed n: Lemma. ( The response Z [m,n] depends on T m,..., T n only through the dfferences τ m,..., τ n. ( Let a b be ntegers. Let T n = T n + Z [a,b] (n > b, T n = T n Z [a,b] (n b. And let X [m,n], X [m,n] be the correspondng last actvty epochs. Then both of them exhbt separablty at ndex b. ( The varables X [m,n], Z [m,n] ncrease when m decreases. (v For a b < b + c, Z [a,c] Z [a,b] + Z [b+,c]. Proof. ( Follows from the defnton Z [m,n] = X [m,n] T n and the homogenety assumpton (A3. ( Obvously, Z [a,b] τ b + Z [a,b], and so X [a,b] T b τ b + Z [a,b], whch mples X [a,b] T b+ + Z [a,b]. The rght-hand sde s T b+, by defnton. The left-hand sde s equal to X [a,b] because T n = T n for n b. So X [a,b] T b+ and ths s separablty at ndex b. Smlarly for the other varable. ( Let a = b = m n (. Snce we have separablty at ndex m, we conclude that X [m,n] = X [m+,n]. But T k = T k for k [m +, n] and so X [m+,n] = X [m+,n]. On the other hand, {T k } {T k} and so, by (A2, X [m,n] X [m,n]. Thus X [m,n] X [m+,n]. And so Z [m,n] Z [m+,n] also. (v Apply ( agan. Snce {T k } {T k }, (A2 gves X [a,c] X [a,c]. By separablty at ndex b, as proved n (, we have X [a,c] = X [b+,c]. Because T k = T k + Z [a,b] for all k [b +, c], we have, by (A3, X [b+,c] = X [b+,c] + Z [a,b]. Thus, X [a,c] X [b+,c] + Z [a,b]. Subtractng T c from both sdes gves the desred. Introduce next the usual statonary-ergodc assumptons. Namely, consder (Ω, F, P and a statonary-ergodc flow θ. Let ξ n = ξ 0 θ n, τ n = τ 0 θ n, set T 0 = 0, and suppose Eτ 0 = λ (0,, EZ 0,0 <. Stablty of the orgnal system can, n specfc but mportant cases, be translated n a stablty statement for Z [m,n]. Hence we shall focus on t. Note that Z [m,n] θ k = Z [m+k,n+k] for all k Z. For any c 0, ntroduce the epochs 6

c{t n } = {ct n } and let X [m,n] (c, Z [m,n] (c be the quanttes of nterest. The subaddtve ergodc theorem gves that γ(c := lm n n Z [, ](c = lm n n EZ [, ](c s a nonnegatve, fnte constant. The prevous lemma mples that γ(c γ(c when c > c. Smlarly, lm n X [,n] (c = γ(c + λ c, and the latter quantty ncreases as c ncreases. Monotoncty mples that Z [, ] (c ncreases as n ncreases, and let Z(c be the lmt. Ergodcty mples that P( Z(c < {0, }. Put Z = Z(. The stablty theorem s: Theorem. If λγ(0 < then P( Z < =. If λγ(0 > then P( Z < = 0. Proof. Assume frst that λγ(0 >. Fx n. Defne T k = T for all k Z. Hence X [,0] ( X [,0]( = Z [,0] (, by (A2. On the other hand, by (A3, X [,0] ( = X [,0] (0 + T = Z [,0] (0 + T. Thus, n Z [,0] ( n Z [,0] (0 + n T, and, takng lmts as n, we conclude lm nf n Z [,0] ( γ(0 λ > 0, a.s. Assume next that λγ(0 <. Let γ n (0 := EZ [+,0] (0/n. Snce γ(0 = lm n γ n (0 = nf n γ n (0, we can fnd an nteger K such that λγ K (0 <. Consder next an auxlary sngle server queue wth servce tmes σn := Z [ Kn+, K(n ] (0 and nterarrval tmes t n := K(n = Kn+ t. Notce that {(t n, s n, n Z} s a statonary sequence whch satsfes the SLLN. Consder the watng tme W n of ths auxlary system: W n+ = (W n +s n t n +. Snce Es n = γ K < λ = Et n, the auxlary queue s stable. Snce the separablty property holds, we have the followng domnaton: Z [K+,0] ( W n θ + s 0, a.s., where W n here s the watng tme of the n-th customer f the queue starts empty. By the Loynes scheme, W n θ converges (ncreases to an a.s. fnte random varable. Hence Z = lm n Z [K+,0] ( s also a.s. fnte. Example Consder a tandem queue and fnd γ(0. Let b = max(b (, b (2. Then ( m γ(0 = lm n n max σ ( 0 m n + σ (2 m = b. lm n max ( σ (, σ (2 From the other sde, assume that, say, b = b ( > b (2. Then ( γ(0 = lm σ ( + max (σ (2 n 0 m n σ ( lm n Ths s known as the saturaton rule σ ( + n sup m 0 m m (σ (2 σ ( 7

where the supremum n the RHS s fnte a.s. and does not depend on n. Therefore, the second term tends to 0 a.s. Thus, γ(0 = b. The same concluson holds f b ( < b (2 (by the symmetry and f b ( = b (2 (Why?? ths s Exercse 6 for you! Exercse 7. Usng Theorem, fnd stablty condtons for G/G/s queue. 6 Saturaton rule for large devatons The proposed constructon of an upper sngle-server queue may be of use not only for stablty study, but also for study of large devatons. We consder here only an example of a tandem queue. In ths secton, we assume that three drvng sequences {σ n ( }, {σ n (2 }, and {t n } are mutually ndependent and each of them conssts of..d.r.v. s. We assume also that the stablty condton a > b holds. We are nterested n the asymptotcs for P(Z > x as x. For the smplcty, we consder only the case when both dstrbutons of random varables σ ( and σ (2 are lght-taled, and we study only the logarthmc asymptotcs: log P(Z > x.... For =, 2, let ϕ ( (u = E exp(uσ ( and ϕ τ (u = E exp(ut. Let γ ( = sup{u : ϕ ( (uϕ τ ( u }. Theorem 2. If both γ ( and γ (2 are postve, then where γ = mn(γ (, γ (2. Sketch of Proof. Snce Z Z (, lm sup x log P(Z > x γx log P(Z > x x γ (. Smlarly, consder an auxlary system where servce tmes n the frst queue are replaced by zeros. Then we get a sngle server queue wth servce tmes {σ n (2 } and nterarrval tmes {t n }. If we denote by Z (2 a statonary servce tme n ths system, then Z (2 Z and, therefore, log P(Z > x lm sup γ (2. x x Thus, log P(Z > x lm sup γ. x To obtan the lower bound, we take a suffcently large K (see the proof of Theorem and an auxlary upper sngle server queue. If we let γ = sup{u : ϕ σ (uϕ t ( u }, then one can show that log P(Z>x (alm nf x γ, (b one can choose γ as close to γ as possble. 8

Exercse 8. Complete the proof of Theorem 2. Remark. The exact asymptotcs for P(Z > x n the heavy tal case have been found n [3]. References [] Baccell, F. and Brémaud, P. (2003 Elements of Queueng Theory. Sprnger, Berln. [2] Baccell, F. and Foss, S. (995 On the Saturaton Rule for the Stablty of Queues. J. Appl. Probab. 23, n.2, 494-507. [3] Baccell F, and Foss, S. (2004 Moments and tals n monotone-separable stochastc networks. Ann. of Appl. Probab. 4, 62-650. [4] Borovkov, A.A. (998 Ergodcty and Stablty of Stochastc Processes. Wley, New York. [5] Borovkov, A.A. and Foss, S.G. (992 Stochastcally recursve sequences and ther generalzatons. Sb. Adv. Math. 2, n., 6-8. [6] Brandt, A., Franken, P. and Lsek, B. (992 Statonary Stochastc Models. Wley, New York. [7] Dacons, P. and Freedman, P. (999 Iterated random functons. SIAM Revew 4 45-76. [8] Loynes, R.M. (962 The stablty of queues wth non-ndependent nterarrval and servce tmes. Proc. Cambrdge Phl. Soc. 58, 497-520. [9] Meyn, S. and Tweede, R. (993 Markov Chans and Stochastc Stablty. Sprnger, New York. [0] Whtt, W. (2002 Stochastc-Process Lmts. Sprnger, New York. 9