A Simple Memoryless Proof of the Capacity of the Exponential Server Timing Channel

Similar documents
An Alternative Proof of the Rate-Distortion Function of Poisson Processes with a Queueing Distortion Measure

The Rate-Distortion Function of a Poisson Process with a Queueing Distortion Measure

Shannon and Poisson. sergio verdú

THE capacity of the exponential server timing channel

SEVERAL results in classical queueing theory state that

Time Reversibility and Burke s Theorem

LECTURE 15. Last time: Feedback channel: setting up the problem. Lecture outline. Joint source and channel coding theorem

Lecture 20: Reversible Processes and Queues

Statistics 150: Spring 2007

Lecture 22: Final Review

MARKOV PROCESSES. Valerio Di Valerio

Lecture 4 Channel Coding

Stochastic process. X, a series of random variables indexed by t

THE SYNCHRONIZATION OF POISSON PROCESSES AND QUEUEING NETWORKS WITH SERVICE AND SYNCHRONIZATION NODES

Lecture 4 Noisy Channel Coding

Data analysis and stochastic modeling

Performance Evaluation of Queuing Systems

On the Throughput, Capacity and Stability Regions of Random Multiple Access over Standard Multi-Packet Reception Channels

ELEC546 Review of Information Theory

Probability Models in Electrical and Computer Engineering Mathematical models as tools in analysis and design Deterministic models Probability models

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 22 12/09/2013. Skorokhod Mapping Theorem. Reflected Brownian Motion

Optimal Natural Encoding Scheme for Discrete Multiplicative Degraded Broadcast Channels

Multiaccess Channels with State Known to One Encoder: A Case of Degraded Message Sets

Introduction to Markov Chains, Queuing Theory, and Network Performance

Positive Harris Recurrence and Diffusion Scale Analysis of a Push Pull Queueing Network. Haifa Statistics Seminar May 5, 2008

Inequality Comparisons and Traffic Smoothing in Multi-Stage ATM Multiplexers

Figure 10.1: Recording when the event E occurs

Refined Bounds on the Empirical Distribution of Good Channel Codes via Concentration Inequalities

TCOM 501: Networking Theory & Fundamentals. Lecture 6 February 19, 2003 Prof. Yannis A. Korilis

Markov Chains. X(t) is a Markov Process if, for arbitrary times t 1 < t 2 <... < t k < t k+1. If X(t) is discrete-valued. If X(t) is continuous-valued

GI/M/1 and GI/M/m queuing systems

Intro Refresher Reversibility Open networks Closed networks Multiclass networks Other networks. Queuing Networks. Florence Perronnin

Capacity of the Discrete Memoryless Energy Harvesting Channel with Side Information

The Timing Capacity of Single-Server Queues with Multiple Flows

Part I Stochastic variables and Markov chains

The Poisson Channel with Side Information

Networking = Plumbing. Queueing Analysis: I. Last Lecture. Lecture Outline. Jeremiah Deng. 29 July 2013

Queuing Networks: Burke s Theorem, Kleinrock s Approximation, and Jackson s Theorem. Wade Trappe

Channel Polarization and Blackwell Measures

Operations Research Letters. Instability of FIFO in a simple queueing system with arbitrarily low loads

Solutions to Homework Discrete Stochastic Processes MIT, Spring 2011

Recap. Probability, stochastic processes, Markov chains. ELEC-C7210 Modeling and analysis of communication networks

EE 4TM4: Digital Communications II. Channel Capacity

Chapter 4. Data Transmission and Channel Capacity. Po-Ning Chen, Professor. Department of Communications Engineering. National Chiao Tung University

Fundamental Limits of Invisible Flow Fingerprinting

Control Over Noisy Channels

Upper Bounds on the Capacity of Binary Intermittent Communication

Intermittent Communication

Capacity Bounds on Timing Channels with Bounded Service Times

Queuing Analysis of Markovian Queue Having Two Heterogeneous Servers with Catastrophes using Matrix Geometric Technique

Network coding for multicast relation to compression and generalization of Slepian-Wolf

Estimating a linear process using phone calls

10.2 For the system in 10.1, find the following statistics for population 1 and 2. For populations 2, find: Lq, Ls, L, Wq, Ws, W, Wq 0 and SL.

Midterm Exam Information Theory Fall Midterm Exam. Time: 09:10 12:10 11/23, 2016

Stochastic Network Calculus

Lecture 7: Simulation of Markov Processes. Pasi Lassila Department of Communications and Networking

Equivalent Models and Analysis for Multi-Stage Tree Networks of Deterministic Service Time Queues

Quality of Real-Time Streaming in Wireless Cellular Networks : Stochastic Modeling and Analysis

Lecture 6 I. CHANNEL CODING. X n (m) P Y X

Stability and Rare Events in Stochastic Models Sergey Foss Heriot-Watt University, Edinburgh and Institute of Mathematics, Novosibirsk

LECTURE #6 BIRTH-DEATH PROCESS

Lecture 5 Channel Coding over Continuous Channels

Chapter 5. Continuous-Time Markov Chains. Prof. Shun-Ren Yang Department of Computer Science, National Tsing Hua University, Taiwan

Interactions of Information Theory and Estimation in Single- and Multi-user Communications

On Scalable Coding in the Presence of Decoder Side Information

Arimoto Channel Coding Converse and Rényi Divergence

Fluid Models of Parallel Service Systems under FCFS

Lecture 11: Polar codes construction

EE5139R: Problem Set 7 Assigned: 30/09/15, Due: 07/10/15

IEOR 6711, HMWK 5, Professor Sigman

LECTURE 10. Last time: Lecture outline

Queueing Theory I Summary! Little s Law! Queueing System Notation! Stationary Analysis of Elementary Queueing Systems " M/M/1 " M/M/m " M/M/1/K "

Streaming Algorithms for Optimal Generation of Random Bits

Chapter 3 Balance equations, birth-death processes, continuous Markov Chains

A Study on Performance Analysis of Queuing System with Multiple Heterogeneous Servers

Error Exponent Region for Gaussian Broadcast Channels

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Department of Electrical Engineering and Computer Science Transmission of Information Spring 2006

DISCRETE STOCHASTIC PROCESSES Draft of 2nd Edition

Introduction to queuing theory

[4] T. I. Seidman, \\First Come First Serve" is Unstable!," tech. rep., University of Maryland Baltimore County, 1993.

Lecture 5: Asymptotic Equipartition Property

LECTURE 13. Last time: Lecture outline

Lecture 3: Channel Capacity

Lecture 10: Broadcast Channel and Superposition Coding

CHAPTER 3 STOCHASTIC MODEL OF A GENERAL FEED BACK QUEUE NETWORK

Readings: Finish Section 5.2

Link Models for Packet Switching

Technical Appendix for: When Promotions Meet Operations: Cross-Selling and Its Effect on Call-Center Performance

Homework Set #2 Data Compression, Huffman code and AEP

5 Lecture 5: Fluid Models

Queueing Theory II. Summary. ! M/M/1 Output process. ! Networks of Queue! Method of Stages. ! General Distributions

Feedback Capacity of a Class of Symmetric Finite-State Markov Channels

A new converse in rate-distortion theory

Queues and Queueing Networks

EE376A - Information Theory Final, Monday March 14th 2016 Solutions. Please start answering each question on a new page of the answer booklet.

CS 798: Homework Assignment 3 (Queueing Theory)

IEOR 6711: Stochastic Models I, Fall 2003, Professor Whitt. Solutions to Final Exam: Thursday, December 18.

Markov processes and queueing networks

National University of Singapore Department of Electrical & Computer Engineering. Examination for

Efficient Nonlinear Optimizations of Queuing Systems

Transcription:

A Simple Memoryless Proof of the Capacity of the Exponential Server iming Channel odd P. Coleman ECE Department Coordinated Science Laboratory University of Illinois colemant@illinois.edu Abstract his paper provides a conceptually simple, memoryless-style proof to the capacity of the Anantharam and Verdu s exponential server timing channel (ESC). he approach is inspired by Rubin s approach for characterizing the ratedistortion of a Poisson process with structured distortion measures. his approach obviates the need for using the information density to prove achievability, by exploiting: ) the ESC channel on [, ] law is a product of [, ] ESC channel laws given intermediate queue states; 2) for Poisson inputs, the queue states form a Harris-recurrent Markov process. Achievability is subsequently shown by first demonstrating via the law of large numbers for Markov chains that an AEP holds, followed by standard random coding arguments. We extend our methodology to demonstrate achievability with Poisson inputs for any point process channel where the conditional intensity at any time is only a function of the queue state. Lastly, we demonstrate achievability with Poisson inputs for the tandem queue. I. INRODUCION he landmark Bits through Queues (BQ) paper by Anantharam & Verdú [] characterized the fundamental limits of coding through the timing of packets in queuing systems. A special case of FCFS queuing timing channel is the exponential server timing channel (ESC) where service times are i.i.d. and exponentially distributed of rate µ. For an arrival process of rate λ, it was shown in [] that the First- Come First-Serve (FCFS) /M/ continuous-time queue with service rate µ > λ > has a capacity C(λ, µ) given by C(λ, µ) λ log µ, λ < µ nats/s. () λ hen the capacity of the FCFS /M/ continuous-time queue with service rate µ is given by the maximum of C(λ, µ) over all possible arrival rates λ < µ, namely, C(µ) e µ nats/s, (2) where the maximum corresponding to (2) is achieved in () at λ e µ. m Fig.. encoder ESC (µ) Q decoder Conveying information through packet timings in a queueing system. ˆm A. Related Work on Proving C(λ, µ) he original BQ work [] first proved C(λ, µ) by considering the probabilistic dynamics relating n packet arrivals to n packet departures of an ESC. hey describe these probabilistic dynamics and then use a standard Fano s inequality converse by controlling the idling times. Achievability with a Poisson input process was shown using the General formula for capacity [2] information density calculus - because the channel is not memoryless. In [3], Bedekar and Azizoglu considered the discrete-time memoryless queuing system. Using similar techniques to [] involving information density arguments pertaining to the probabilistic dynamics relating n packet arrivals to n departures, they proved the analogous capacity of the discrete-time memoryless queue. In [4], Prabhakar and Gallager discussed the discrete-time memoryless queuing system from the n packet arrivals to n packet departures, but from a different viewpoint. hey upper bounded the limit of maximal normalized inputoutput mutual information by noting that the service times form a bijection with the output times given the input times. hat being said, as they themselves mention (in [4, Sec IV, footnote 2 ]), this technically is not sufficient to prove capacity, because this is not a memoryless channel. In [5], Sundaresan and Verdú re-visited the ESC, this time from a point processes viewpoint, by studying probabilistic dynamics between the arrival and departure point processes on [, ]. hey exploited the intimate relationship between the ESC and the Poisson channel, and were able to use capacity results of Poisson channels to provide a converse for the ESC, but still used information density arguments to illustrate achievability. B. Methodology of Our Approach his paper, similar to [5], considers all input/output processes in continuous time on the [, ] interval. Ours differs because we illustrate that one can reason about this channel coding problem completely from a traditional memoryless channel perspective using random coding and the AEP, by exploiting a few key properties: he ESC channel on [, ] law is a product of [, ] ESC channel laws given intermediate queue states;

For Poisson inputs, the queue states form a Harrisrecurrent Markov process. he channel law can be completely characterized in terms of the output process and the queue states he law of large numbers for Markov chains [4] holds to demonstrate that an AEP holds, Standard random coding arguments apply to arrive at achievability. We extend our methodology to demonstrate (via Burke s theorem) achievability with Poisson inputs for any point process channel where the conditional intensity at any time is only a function of the queue state. Lastly, we demonstrate achievability with Poisson inputs for the tandem queue. Since many converses for the ESC have been established, we focus here completely on proving achievability and showing such channels are all information-stable. he reasoning of the proof techniques for the ESC in this is paper is in some sense dual to that of lossy compression scheme for compression Poisson processes [7] with queuing distortion measures, developed in [8], [9]. II. PRELIMINARIES A. Notation on Point Processes Let (Ω, F) to be a measurable space. Denote ( t ) t [, ] as a stochastic process on [, ], and define σ ( s ) s [,t] F t Define to be the set of functions x : [, ] Z + that are non-decreasing, right-continuous, and x. In other words, is the set of point processes on [, ]. Define the input space as ( ), F and the output space as ( ), F with. With a probability space (Ω, F, P ), a stochastic point process ( t ) t [, ] on the space ( ), F, adapted to (F t ) t [, ] has the conditional intensity process λ (λ t ) t [, ] with respect to (F t ) t [, ] if P ( t+ t F t ) λ t lim. (3) Define P to be the probability measure on ( ), F s.t. is a point process with unit-rate conditional intensity with respect to ( ) Ft t [, ]. B. he Exponential Server iming Channel Let us now consider an ESC on [, ], where we denote the arrival process as on ( ), F, and the departure process as on ( ), F. Define the random variable Q as the queue state at time. Since the service times are i.i.d. and exponentially distributed (and thus memoryless), for a fixed x and q Z +, the probability measure P,Q ( x, q ) on ( ), F for the ESC has the rate process λ t µ Qt > (4) Q t q + x t t. (5) with respect to the filtration (H t ) t [, ] : H t σ ( t ) t [,t]. Define the Radon-Nikodym derivative of P,Q ( x, q ) with respect to P as [], [5]: f,q (y x, q ) dp,q ( x, q ) dp (y) (6a) exp log λ t dy t + [ λ t ]dt (6b) exp log ρ(q t )dy t + [ ρ(q t )]dt µ y e exp ρ(q) µ q> ρ(q t )dt (6c) (6d) Define P ( ) as the probability measure for the stochastic process on ( ), F. Denote the density of P ( ) with respect to P is given by [3], [] f (x) dp dp (x). Given f (x), a marginal distribution on Q, and f,q (y x, q ), the probability measure P ( ) on ( ), F follows. Similarly, denote the density of P ( ) with respect to P as f (y) dp dp (y). he mutual information I (; ) is given as I (; Q ) E log f,q (, Q ) f. Q ( Q ) C. Encoder Constraints, Achievability and Capacity M equally likely messages are mapped to codewords, each of which a member of. For λ < µ, we constrain the encoders to satisfy [ ] E λ. (7) he decoder observes the departure process on [, ], y, and decodes to one of the M messages. Such a code with an average error probability of P e is defined a (, M, ɛ, P e ) code. A rate R is achievable if there exists a sequence of ( k, 2 k R, Pe,k, λ) codes with lim k k and lim k P e,k. he λ-capacity is the supremum of all achievable rates.

III. ACHIEVABILI FOR HE ESC USING HE AEP Consider an arbitrary > and a positive integer n. For x and y, define x x n ( x,..., x n ), x i ( ) x t x (i ) t [(i ),i ) y ỹ n (ỹ,..., ỹ n ), ỹ i ( ) y t y (i ) t [(i ),i ) q n ( q,..., q n ), (8a) (8b) q i (q t ) t [(i ),i ) (8c) Note that x i and ỹ i. See Figure 2 as an example of the relationship between x and x n ( x,..., x n ). 9 8 7 6 5 4 3 2 3 2 x x 2 3 x x 2 x 3 Fig. 2. op: a realization of x, with n 3. Note that x 9 - so at time t, 9 packets have arrived. Bottom: x can be equivalently expressed in terms of n 3 counting functions x ( x, x 2, x 3 ) with each x i, given by (8a). he simple yet crucial observation we make is that For any t [(i ), i ): q t q + x t y t q + x (i ) y (i ) + [x t x (i ) ] [y t y (i ) ] q (i ) + x i,t ỹ i,t. It follows from (6) that the channel law is memoryless conditioned upon intermediate queue states: f,q (y x, q ) exp log ρ(q t )dy t + [ ρ(q t )]dt i exp t(i ) f,q (ỹ i x i, q i, ) where f,q (, ) is given by (6). log ρ(q t )dy t + [ ρ(q t )]dt (9a) (9b) We develop an achievability theorem for the ESC with Poisson inputs by proving the existence of an asymptotic equipartition property (AEP) [2]. Define P ( ) as the probability measure for the stochastic process on ( ), F corresponding to a Poisson process of rate λ. As such, the density of P ( ) with respect to P is given by [3], [] f (x) dp dp (x) λ x e (λ ) () Draw Q according to the steady-state distribution of a (λ, µ) birth-death Markov chain: ( P Q (k) λ ) ( ) k λ, k. µ µ hen from Burke s theorem [3], [], the output will also be a Poisson process of rate λ, and so f (y) dp dp (y) λ y e (λ ). () With these observations, we can develop the following AEP with Poisson inputs: Marginal on : log f () λ n i, log λ a.s. λ( log λ) (2) where (2) follows because of the independent increments property of the Poisson Process. Marginal on : from Burke s theorem, the output is Poisson. So from (2): log f ( ) a.s. λ( log λ) (3) Joint on and : Define: ν ( q i ) ρ( q i,t )dt (4) Note that since the queue states (Q t ) t [, ) form a Harris-recurrent continuous-time Markov process, the ( Q i : i n) form a discrete-time Harris-recurrent Markov chain [4]. Note from (9),(6b), (4) that: log f,q (, Q ) Ỹ i, n log µ + ν ( n Q i ), a.s. λ( log µ) by the strong law of large numbers (SLLN) and the SLLN for Markov chains [4]. As a consequence, the AEP holds, and from standard random coding arguments [2], any rate R < I (; ) λ log ( ) µ λ is achievable.

IV. ACHIEVABILI HEOREM FOR AN CHANNEL WHERE HE CONDIIONAL INENSI ONL DEPENDS ON HE QUEUE SAE We now extend the aforementioned approach to general timing channels where the conditional intensity is only a function of the current queue state. An example of which is the /M/m queue, where an achievabiilty theorem with Poisson inpus was developed in [5]. We illustrate that for such general channels, the mutual information corresponding to Poisson inputs is an achievable rate. Our assumption is that for a fixed x and q Z +, the probability measure P,Q ( x, q ) on ( ), F has the rate process λ t ρ (Q t ) (5) with respect to (H t ) t [, ], with H t σ ( t ) t [,t] and Q t given by (5). hus, by (6b), for x and y : f,q (y x, q ) exp log ρ (q t ) dy t + [ ρ (q t )]dt f ),Q (ỹi x i, q (i ) φ(ỹ i, q i ) (6) ) where (6) is explicitly denoting how f,q (ỹi x i, q (i ) is a measurable function of (ỹ i, q i ). heorem 4.: he rate R I (; Z) pertaining to P ( ) on ( ), F as a Poisson process of rate λ is achievable on any point process channel satisfying the assumption in (5) for which the induced queue process is recurrent. Proof: Define a b (a τ : τ [, b]). Let f (x) be the density of a Poisson process of rate λ on ( ), F. hen the process (Q t ) t [, ] is a birth-death Markov process. Draw Q P Q ( ) with P Q ( ) the steady-state probability distribution on the birth-death chain. Note the joint process (Q t, t ) t [, ] is also a Markov process: In infinitesimal time, the queuing dynamics satisfy Q t Q t ( t t ) ( t t ). So: P ( Q t Q t σ t, Q t ) P ( t t t t σ t, Q t ) P ( t t t t ) (7) λ t t where (7) follows because of the independentincrements nature of the Poisson process. In infinitesimal time (small ), from (5): P ( t+ t σ t, Q t ) ρ (Q t ) (8) Define π as the steady-state distribution on the Harris-recurrent process (Ỹi, Q i ) i n. hus it follows that the AEP holds: Marginal on : from (2) log f () a.s. λ( log λ) (9) Marginal on : via reversibility on the birth-death process, Burke s theorem [3], [] holds and so is also Poisson log f ( ) a.s. λ( log λ) (2) Joint on and : from (6), log f,q (, Q ) n log φ(ỹi, Q i ) a.s. E π log φ(ỹi, Q i ) (2) where (2) follows because of the SLLN for Markov chains applied to the Markov chain (Ỹi, Q i ). V. ACHIEVABILI FOR HE ANDEM QUEUE Now consider communication in tandem over two ESCs, of parameter µ and µ 2 (see Figure 3), with the same constraint (7). Sundaresan and Verdú studied this channel when µ µ 2 and simulated the mutual information rate corresponding to Poisson inputs, but were unable to demonstrate achievability [5, Sec IV.B]. Here, we demonstrate that the mutual information rate corresponding to Poisson inputs is indeed achievable, by again demonstrating the existence of an AEP using the LLN numbers for Markov chains applied to the joint queue states. Define Z and denote Z as the stochastic process on ( ) Z, F Z pertaining to the output of the second queue. Assume the initial states of both queues are and define: Q,t x t t, Q 2,t t Z t (22) Q k (Q k,t ) t [, ], k, 2 (23) From the total probability theorem, for x, the probability measure PZ ( x) on ( ) Z, F Z has Radon-Nikodym derivative with respect to P as: fz (z x) dp Z ( x) dp (z) fz,q 2, (z y, ) f,q, (y x, ) P (dy) µ z 2 e ν (y z) µ y e ν (x y) P (dy) (24) (µ µ 2 ) z µ y z e ν (y z)+ν (x y) P (dy) (25) ( ) with fz,q 2, (, ) f,q, (, ) given by (6) with parameter µ (µ 2 ) respectively. We state the following theorem:

ESC (µ ) ESC (µ 2 ) Q Q 2 Fig. 3. he tandem queue heorem 5.: he rate R I (; Z) pertaining to P ( ) on ( ), F as a Poisson process of rate λ < min(µ, µ 2 ) is achievable on the tandem queue. Proof: Define the abelian group G as the set of all functions g : [, ] Z such that g is right-continuous, with ( t ) t [, ] and g g (g t g t) t [, ] for any g, g G. Define the following two functions β : G R, β 2 : G R as β (g) µ g e ν (g), β 2 (g) e ν (g). (26) Note that G and so we can define P (G \ ) PZ (G \ x) to extend P ( ) and PZ ( ) to probability measures on (G, F G ). hus: fz (z x) (µ µ 2 ) z β (y z)β 2 (x y)p (dy) (27) G (µ µ 2 ) z β β 2 (x z) (28) (µ µ 2 ) z β β 2 (q + q 2 ) (29) where (27) follows from (25) and (26); (28) follows from the definition of convolution on abelian groups; and (29) follows from (22) and (23). So by defining, q i,k (q k,t ) t [(i ),i ], k, 2 (3) ( ) z i zt z (i ), (3) t [(i ),i ) for z Z, x, we have: fz (z x) (µ µ 2 ) z β (y z)β 2 (x y)p (dy) G (µ µ 2 ) z µ y z e ν (y z)+ν (x z) P G (µ µ 2 ) n z i, n µ ỹi, z i, G n exp ν (ỹ i z i ) + ν ( x i z i ) Z (dy) P (dy) (32) (µ µ ) z i, 2 β β 2 ( q,i + q 2,i ) (33) where (32) follows because of (8), (4), and (3); and (33) follows from (3). Let f (x) be the density of a Poisson process of rate λ on ( ), F. Note from Burke s theorem, asymptotically, Z is Poisson. Lastly, note that the queue states (Q,t, Q 2,t ) form a Harris-recurrent Markov process. hus, from (33) and the SLLN for Markov chains, the AEP holds: Marginal on : from (2) log f () a.s. λ( log λ) (34) Marginal on Z: via Burke s theorem holds and so Z is asymptotically Poisson: log fz (Z) a.s. λ( log λ) (35) Joint on and Z: Define π as the steady-state distribution on the Harris-recurrent Markov process ( Q,i, Q 2,i ) i n. From (33), log fz (Z ) n n z i, log(µ µ 2 ) log β β 2 ( Q,i + Q 2,i ) a.s. λ log(µ µ 2 ) E π log β β 2 ( Q,i + Q 2,i ) by the SLLN for Markov chains. hus, a random coding argument along with jointly typical decoding suffices to achieve the rate I (; Z). ACKNOWLEDGEMENS he author would like to thank Negar Kiyavash, Maxim Raginsky, and Vijay Subramanian for useful discussions. REFERENCES [] V. Anantharam and S. Verdú, Bits through queues, IEEE ransactions on Information heory, vol. 42, no., pp. 4 8, 996. [2] S. Verdú, A general formula for channel capacity, Information heory, IEEE ransactions on, vol. 4, no. 4, pp. 47 57, 994. [3] A. S. Bedekar and M. Azizoglu, he information-theoretic capacity of discrete-time queues, IEEE ransactions on Information heory, vol. 44, no. 2, pp. 446 46, 998. [4] B. Prabhakar and R. Gallager, Entropy and the timing capacity of discrete queues, IEEE ransactions on Information heory, vol. 49, no. 2, pp. 357 37, February 23. [5] R. Sundaresan and S. Verdú, Capacity of queues via point-process channels, IEEE ransactions on Information heory, vol. 52, no. 6, pp. 2697 279, June 26. [6] B. Prabhakar and N. Bambos, he entropy and delay of traffic processes in AM networks, in Proceedings of the Conference on Information Science and Systems (CISS), Baltimore, Maryland, 995, pp. 448 453. [7] S. Verdú, he exponential distribution in information theory, Problems of Information ransmission, vol. 32, no., pp. 86 95, Jan-Mar 996. [8] I. Rubin, Information rates and data-compression schemes for Poisson processes, IEEE ransactions on Information heory, vol. 2, no. 2, pp. 2 2, 974. [9]. P. Coleman, N. Kiyavash, and V. Subramanian, he rate-distortion function of a Poisson process with a queuing distortion measure, IEEE ransactions on Information heory, submitted May 28. [] J. McFadden, he entropy of a point process, SIAM Journal of Applied Mathematics, vol. 3, pp. 988 994, 965. [] P. Bremaud, Point Processes and Queues: Martingale Dynamics. New ork: Springer-Verlag, 98. [2]. M. Cover and J. homas, Elements of Information heory, 2nd ed. New ork: Wiley, 26. [3] R. Gallager, Discrete Stochastic Processes. Boston, MA: Kluwer, 996. [4] S. Meyn and R. weedie, Markov chains and stochastic stability. Springer-Verlag, 993. [5] R. Sundaresan and S. Verdú, Sequential decoding for the exponential server timing channel, IEEE ransactions on Information heory, vol. 46, no. 2, pp. 75 79, March 2.