Distortion-Delay Tradeoff for a Gaussian Source Transmitted over a Fading Channel Qiang Li and C. N. Georghiades Electrical and Computer Engineering Department, Texas A&M University College Station, TX 77843-3128 E-mail: {qiangli,georghiades}@ece.tamu.edu Abstract— We study the end-to-end distortion-delay tradeoff for a Gaussian source transmitted over a fading channel. The analog source is quantized and stored in a buffer until it is transmitted. There are two extreme cases as far as buffer delay is concerned: no delay and infinite delay. We observe that there is a significant power gain by introducing a buffer delay. Our goal is to investigate the situation between these two extremes. Using the recently proposed effective capacity concept, we derive a closed- form expression for this tradeoff. In order to characterize the convergence behavior, we derive an asymptotically tight upper bound for our tradeoff curve, which approaches the infinite delay lower bound polynomially. Numerical results demonstrate that introduction of a small amount of delay can save significant transmission power. I. I NTRODUCTION Quality-of-Service (QoS) is a critical design objective for next-generation packet cellular networks. End-to-End distor- tion and transmission delay are two fundamental QoS metrics. Such QoS requirements pose a challenge for the system design due to the unreliability and time varying nature of the wireless link. Quantizer Buffer Analog Source Adaptive Transmitter Ideal CSI feedback Fading Channel Receiver Fig. 1. System model In this paper we consider transmission of a Gaussian source over a wireless time-varying fading channel. Our goal is to optimize the end-to-end distortion given a delay constraint. More precisely, we derive the distortion and delay tradeoff for the wireless fading channel. To this end, we adopt a cross-layer approach as shown in Figure 1. At this point, for simplicity we assume an independent and identically distributed (i.i.d.) block fading channel model. Such a model is suitable for serval practical communication scenarios, e.g., time hopping in TDMA, frequency hopping in FDMA and multicarrier systems. Extension to the more practical time-correlated case will be discussed later. Throughout this paper, we always assume channel state information (CSI) is perfectly known at the transmitter. We consider an i.i.d., real, Gaussian N (0, 1) source, which is quantized and then fed into a buffer. Since the channel is time-varying, the transmitter adjusts the transmission rate to the current channel state. The relevant performance criteria are end-to-end quadratic distortion and the buffer delay. We aim to find the relationship between the distortion and delay for some average transmission power. For this problem, there are two extreme cases: 1) There is no buffer, i.e. no delay); 2) we have an infinite buffer size, i.e., we allow an infinite transmission delay. For the first case, we adaptively quantize the Gaussian source according to the CSI. Assuming perfect transmission, we can approximate the average achievable quadratic distortion by: D 0 (ρ)= E exp(2 log(1 + α P N 0 W ) , (1) where P denotes the transmission power and W and N 0 the bandwidth and noise variance; α is the channel gain, a ran- dom variable with unit variance following a certain statistical distribution. Here, we have used the Gaussian distortion-rate function expressed as D(R s ) = exp(2R s ) with C(ρ)= log(1 + αρ) as the AWGN channel capacity-cost function and ρ = P N0W . For the infinite delay case, the average transmission rate can achieve the ergodic capacity of a fading channel and the quantizer can simply adopt a constant output rate. The average distortion is given by: D (ρ) = exp 2 E[log(1 + α P N 0 W )] . (2) The function exp(2(·)) is a convex function. By Jensen’s inequality then, the distortion D 0 is lower-bounded by D , i.e., D 0 ≥D . The two distortion functions are plotted in Figure 2 for a Rayleigh fading channel. Notice that there is a significant dB gap between the no-delay and infinite delay curves. We can call this transmission power gap the “Jensen’s gain”. So introducing a buffer at the transmitter to match the source rate with the instantaneous quality of the channel can save significantly in transmission power to meet some distortion requirement. Also, we have simplified the quantization step by allowing a constant rate. A natural question is therefore: if we only allow a finite delay or buffer, how much gain can we achieve? How fast does the distortion curve converge to the infinite-delay lower bound as the delay increases? The main result of this paper is a clear characterization of the tradeoff between end-to-end quadratic distortion and delay, which provides insight to the impact of the buffer delay on the achieved distortion function of the Gaussian source transmitted over a wireless fading channel. ISIT 2006, Seattle, USA, July 9  14, 2006 21 1424405041/06/$20.00 ©2006 IEEE