You are on page 1of 1

ISIT 2003, Yokohama, Japan, June 29 July 4, 2003

Transmission of Bursty Data over Wideband Fading Channels


Won S. Yoon
LIDS, M.I.T. Cambridge, MA 02139 e-mail: won@mit.edu

Thierry E. Klein
Bell Laboratories, Lucent Technologies Murray Hill, NJ 07974 e-mail: tek@lucent.com Using Bellmans equations, we nd that the optimal policy is to transmit only in a particular set of channel states, denoted by S (N ) () S, where N is the number of time steps until the deadline. We now present some properties of S (N ) (). It is easily observed that S (N ) () always includes the best channel state. However, the optimal policy may not be threshold-based. There are examples where it is optimal to transmit in a worse channel state and not transmit in a better channel state, due to the Markovian dynamics of the channel. In certain special cases (which include memoryless channels), the optimal policy does behave like a threshold rule. Theorem 1 (Symmetric Markov Channels) If the channel states all have the same duration and if the channel transition probabilities are independent of the current state, then the optimal policy is to transmit whenever the channel gain exceeds some threshold: gc t(N ) (). Theorem 2 (Monotonicity in N ) S (N 1) () S (N ) () lim S (N ) () = S ()
N

I. Introduction
The problem of transmitting bursty, delay-sensitive data over time-varying channels is an important issue in nextgeneration wireless networks. There is a fundamental tradeo of conserving a mobile users energy while providing a desired level of quality of service. Recently, there have been numerous studies which try to optimize this tradeo (e.g., [1, 2, 3]). We present the rst known results in this context for fading channels with innite bandwidth and investigate the tradeo between average packet delay and average power consumption.

II. Problem Formulation and Results


We rst consider a single user scenario in which packets are randomly generated from a data source according to a Poisson process and stored in an innite buer at the transmitter. The communication channel is modeled as an innite bandwidth, at fading, additive white Gaussian noise channel, in which the fading state behaves according to a nite-state Markov process. The Gaussian noise has unit variance and state c S = {1, . . . , |S|} corresponds to a constant channel gain gc , for which the capacity is known to be C = gc P bits/sec, and the minimum energy required to transmit a packet of L bits is Ec = L ln 2/gc Joules/packet [4, 5]. Due to the linear rate-power function, spreading the energy of a codeword over an arbitrarily large range of frequencies allows us to conne the codeword to an innitesimally small interval of time, while maintaining the same energy per bit. Assuming no peak power constraints, any nite number of packets can be transmitted in innitesimal time with energy Ec in channel state c. Designing a transmission strategy therefore simplies to a decision of which channel state to transmit every packet. There are two extreme cases that illustrate this fundamental tradeo. At one extreme, zero delay is achievable by instantly transmitting every packet as soon as it arrives, at the expense of using large energy in poor channel states (small channel gain gc ). At the other extreme, we can minimize energy by holding packets and only transmit in the best channel state (largest gc ) at the expense of increased delay. In general, there is a tradeo of using less energy by waiting for a better channel state or transmitting with smaller delay in a worse channel state. We formulate this problem using dynamic programming with a nite time horizon. The nite horizon corresponds to placing a deadline constraint (or a maximum delay constraint) on every packet. The cost is a Lagrangian function of the average delay per packet and the average transmission power: L() = E[delay] + E[power], where is the Lagrange multiplier.

Thus, as we approach the deadline it is optimal to transmit in more channel states to empty the queue. In the limit N , S (N ) () converges to some optimal set of transmission states with no deadline constraints, S (), which we can nd using standard iterative dynamic programming algorithms. Theorem 3 (Monotonicity in ) S (N ) () S (N ) ( ) for , N

This says that as the cost of power, , increases it is optimal to transmit in fewer channel states. Finally we have shown that, for a xed deadline constraint N , the optimal powerdelay tradeo curve is a decreasing, convex and piecewise linear function. The results presented here are easily extended to a system with multiple users by showing that it can be decoupled into independent single-user systems.

Acknowledgments
The authors would like to thank Professor R.G. Gallager for motivating this problem and for his invaluable assistance.

References
[1] R.A. Berry and R.G. Gallager, Communication over Fading Channels with Delay Constraints, in IEEE Trans. Info. Theory, vol 48, no. 5, pp. 1135 - 1149, May 2002. [2] E. Uysal, B. Prabhakar and A. El Gamal, Energy-ecient Transmission over a Wireless Link via Lazy Packet Scheduling, in Proc. INFOCOM 2001. [3] W.S. Yoon and T.E. Klein, Delay-Optimal Power Control for Wireless Data Users with Average Power Constraints, in Proc. ISIT 2002, Lausanne. [4] S. Verdu, Spectral Eciency in the Wideband Regime, in IEEE Trans. Info. Theory, vol. 48, no. 6, 2002. [5] R. G. Gallager, Information Theory and Reliable Communication, Wiley, 1968.

You might also like