Bandwidth-Constrained MAP Estimation for Wireless Sensor Networks S. Faisal Ali Shah, Alejandro Ribeiro and Georgios B. Giannakis Dept. of ECE, University of Minnesota 200 Union Street S.E., Minneapolis, MN 55455 Abstract— We deal with distributed parameter estimation algorithms for use in wireless sensor networks (WSNs) with a fusion center when only quantized observations are available due to power/bandwidth constraints. The main goal of the paper is to design efcient estimators when the parameter can be modelled as random with a priori information. In particular, we develop maximum a posteriori (MAP) estimators for dis- tributed parameter estimation and formulate the problem under different scenarios. We show that the pertinent objective function is concave and hence, the corresponding MAP estimator can be obtained efciently through simple numerical maximization algorithms. I. I NTRODUCTION Characterized by small low-cost, low-power devices equipped with limited sensing, computational and commu- nication capabilities, wireless sensor networks (WSNs) are well motivated for environmental monitoring, industrial instru- mentation and surveillance applications [7]. The distributed topology of these networks along with their limited power budget and communication resources gave rise to the area of collaborative signal and information processing [3]. Bandwidth-constrained distributed estimation arises when deploying a WSN for monitoring, in which case the estimation of certain parameters of interest necessitates collection of different sensor estimates. As sensor observations have to be quantized, WSN-based parameter estimators must rely on (per- haps severely) quantized observations [4], [5]. Interestingly, it has been shown in [6] that when the noise variance is comparable to the parameter’s dynamic range, even quanti- zation to a single bit per observation leads to a small penalty in estimation variance when comparing maximum likelihood (ML) estimators based on quantized (binary) versus original (analog-amplitude) observations. A characteristic common to all these works is the assumption that some prior information is available, at least to bound the range of possible parameter values. This suggests naturally, a connection with Bayesian estimation. Building on this observation, the present paper addresses the problem of maximum a posteriori (MAP) estimation based on binary observations, and shows that the graceful Work in this paper was prepared through collaborative participation in the Communications and Networks Consortium sponsored by the U. S. Army Research Laboratory under the Collaborative Technology Alliance Program, Cooperative Agreement DAAD19-01-2-0011. The U. S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation thereon. Email at (sfaisal, aribeiro, georgios)@ece.umn.edu Physical Phenomenon /Process Estimator (Fusion Center) Sensor nodes Transducer Quantizer/ Encoder .. . .. . .. . BW/Power Constrained Links Fig. 1. Block diagram of wireless sensor network performance degradation of ML estimators extends to their MAP counterparts. Moreover, we establish that while MAP estimators cannot be expressed in closed form, they are found as the maximum of a concave function; thus ensuring the convergence of fast descent algorithms e.g., Newton’s method. II. PROBLEM FORMULATION Consider a physical phenomenon characterized by a set of P parameters that we lump in vector form as θ := [ θ 1 ,...,θ P ] T . From available a priori knowledge, θ is modelled as a random vector parameter with prior probability density function (pdf) p θ ( θ) and mean E( θ)= μ θ . For measuring θ, we deploy a WSN composed of N sensors {S n } N-1 n=0 , with each sensor observing θ through a linear transformation x( n)= Hθ + w( n) , (1) where x( n) := [ x 1 ( n) ,...,x K ( n)] T R K is the measure- ment vector at sensor S n , w( n) R K is zero-mean additive noise with pdf p w ( w) and the matrix H R K×P . We denote the vector formed by concatenating all the observa- tions {x( n) } N-1 n=0 of N sensors as x 0:N-1 . For simplicity of exposition, we assume that H and p w ( w) are constant across sensors, and we also assume that w( n 1 ) is independent of w( n 2 ) for n 1 = n 2 . A clairvoyant (CV) benchmark for estimators based on bi- nary observations corresponds to having all analog-amplitude observations x 0:N-1 be available at the fusion center. In this case, a possible approach to estimate θ is the MAP estimator [2] ˆ θ CV = arg max θ {p[ θ|x 0:N-1 ] } , (2) where p[ θ|x 0:N-1 ] is the conditional pdf of θ given x 0:N-1 . As discussed earlier, power and bandwidth constraints, dictate the need of a quantizer mapping the analog-amplitude observations x( n) to a nite set: b( n) := q( x( n)) , with q : R K → {−1 , 1 } K , (3) 215 1424401321/05/$20.00 ©2005 IEEE