EURASIP Journal on Wireless Communications and Networking 2005:4, 473–482

c 2005 Alexey Krasnopeev et al.

Minimum Energy Decentralized Estimation in a Wireless

Sensor Network with Correlated Sensor Noises

Alexey Krasnopeev
Department of Electrical and Computer Engineering, University of Minnesota, 200 Union Street SE, Minneapolis, MN 55455, USA
Email: kras0053@umn.edu

Jin-Jun Xiao
Department of Electrical and Computer Engineering, University of Minnesota, 200 Union Street SE, Minneapolis, MN 55455, USA
Email: xiao0029@umn.edu

Zhi-Quan Luo
Department of Electrical and Computer Engineering, University of Minnesota, 200 Union Street SE, Minneapolis, MN 55455, USA
Email: luozq@umn.edu

Received 25 November 2004; Revised 20 May 2005

Consider the problem of estimating an unknown parameter by a sensor network with a fusion center (FC). Sensor observations
are corrupted by additive noises with an arbitrary spatial correlation. Due to bandwidth and energy limitation, each sensor is
only able to transmit a finite number of bits to the FC, while the latter must combine the received bits to estimate the unknown
parameter. We require the decentralized estimator to have a mean-squared error (MSE) that is within a constant factor to that
of the best linear unbiased estimator (BLUE). We minimize the total sensor transmitted energy by selecting sensor quantization
levels using the knowledge of noise covariance matrix while meeting the target MSE requirement. Computer simulations show
that our designs can achieve energy savings up to 70% when compared to the uniform quantization strategy whereby each sensor
generates the same number of bits, irrespective of the quality of its observation and the condition of its channel to the FC.
Keywords and phrases: wireless sensor networks, decentralized estimation, power control, energy efficiency.

1. INTRODUCTION to the FC. Instead, a more practical decentralized estima-

tion scheme is to let each sensor quantize its real-valued local
measurement to an appropriate length and send the result-
Wireless sensor networks (WSNs) are ideal for environmen-
ing discrete message (typically short) to the FC, while the
tal monitoring applications because of their low implemen-
latter combines all the received messages to produce a final
tation cost, agility, and robustness to sensor failures. A pop-
estimate of the unknown parameter. Naturally, the message
ular WSN architecture consists of a fusion center (FC) and a
lengths are dictated by the power and bandwidth limitations,
large number of spatially distributed sensors. The FC can be
sensor noise characteristics as well as the desired final esti-
either a standard base station or a mobile access point such
mation accuracy.
as an unmanned aerial vehicle hovering over the sensor field.
Recently, several decentralized estimation schemes (DES)
Each sensor in a WSN is responsible for local data collection
[1, 2, 3, 4] have been proposed for parameter estimation
as well as occasional transmission of a summary of its ob-
in the presence of additive sensor noise. These DESs re-
servations to the FC via a wireless link. In a practical WSN,
quire each sensor to send only a few bits to the fusion cen-
each sensor has only limited computation and communica-
ter, with the message length determined by the sensor’s lo-
tion capabilities due to various design considerations such as
cal SNR. Performance of the resulting estimator is shown to
small size battery, bandwidth, and cost. As a result, it is diffi-
be within a constant factor of the best linear unbiased esti-
cult for sensors to send their entire real-valued observations
mator (BLUE) performance. While the designs suggested by
[1, 2, 3, 4] give a guaranteed estimation performance with
This is an open access article distributed under the Creative Commons
low bandwidth requirement, the effect of wireless channel
Attribution License, which permits unrestricted use, distribution, and distortion and the important issue of total sensor energy
reproduction in any medium, provided the original work is properly cited. minimization were not directly modelled.
474 EURASIP Journal on Wireless Communications and Networking

In a practical WSN, the wireless links from sensors to the S1 m1

FC may have different qualities, depending on the sensor lo-
x1
cations relative to the FC. Intuitively, local message length d1
should depend not only on the quality of sensor’s observa- S2
m2 θ̄
tion (i.e., local SNR), but also on the quality of its wireless θ x2 d2 FC
.
link to the FC. In particular, even if a sensor has a high- ..
dN
quality observation, it should not perform any local quan- xN
tization or transmission when its wireless link to the FC is SN mN
weak, in order to conserve sensor energy. In general, min-
imizing the total sensor energy consumption for a decen-
Figure 1: Decentralized estimation scheme.
tralized estimation task is essential to ensure long lifespan
of a WSN. Motivated by these considerations, the authors
of [5, 6] proposed optimal coded and uncoded transmis-
sion strategies for sensor networks which can minimize the solution in closed form. The performance of our energy-
required energy per transmitted bit, although no consider- efficient design is analyzed in Section 4 by numerical simu-
ation was given to the quantization effect and the accuracy lation. Section 5 contains an extension of the work where we
of final estimation. In the recent work of [7, 8], the authors formulate an alternative problem of minimizing maximal in-
considered the problem of optimal energy scheduling for de- dividual sensor energy and present an analytic solution. Final
centralized estimation where sensor measurements are cor- remarks are given in Section 6.
rupted by additive noises, while communication links from Throughout, we use the following notations. Matrices
sensors to the fusion center differ in quality. In particular, and vectors are denoted by boldface letters, capital and small
[7] used an adaptive modulation scheme with an exponen- correspondingly, whereas same regular letters with indices
tial dependence of energy on the transmitted message size, denote their elements. Diagonal matrix with nonzero ele-
and then derived optimal sensor power and quantization lev- ments a1 , . . . , aN is denoted by diag(a1 , . . . , aN ). Logarithms
els via convex optimization. denoted by log(·) are taken to the base 2; for natural loga-
The aforementioned results all require an important as- rithms notation ln(·) is used. For any real number x ∈ R,
sumption that sensor observation noises are spatially uncor- we use x to denote the smallest integer greater or equal to
related. Unfortunately, this assumption can be restrictive in a x. For any random variable R, we use Ex R to denote the ex-
practical WSN, especially when sensors are densely deployed. pected value of R taken with respect to random variable x,
In this paper, we consider distributed parameter estimation while Ex| y R denotes the expected value of R with respect to
in situations where sensor observations are corrupted by cor- x given y. Finally, var R denotes the variance of random vari-
related additive noises. Assuming a standard energy model able R.
[5, 6], uniform quantization at sensors, and the knowledge
of sensor noise correlation matrix, we use convex optimiza-
2. PROBLEM FORMULATION
tion techniques to derive a nearly-optimal (modulo a minor
relaxation) energy scheduling strategy with a mean-squared Consider the problem of estimating an unknown parameter
error performance guaranteed to be within a constant factor θ by a sensor network consisting of N sensors. Measurement
to that of the centralized BLUE estimator. Computer simula- of each sensor xi is corrupted by additive noise ni so that
tions show that our designs can achieve energy savings up to
70% when compared to the uniform bit allocation strategy
xi = θ + ni , i = 1, . . . , N. (1)
whereby each sensor generates the same number of bits.
Our sensor energy scheduling strategy is suitable for di-
rect application when the sensor noise correlation matrix is We assume that both θ and ni have finite range, so that all
available at the FC. In practice, the sensor noise correlation xi belong to a common finite interval [−U, U], with U >
matrix may have to be determined in the sensor network cal- 0 a known constant. The noises ni are assumed to be zero
ibration phase, possibly with the help of training signals. In mean and correlated across sensors with covariance matrix
the absence of this knowledge, our scheme is also useful as C, but otherwise unknown. We assume C is known at the
it provides an upper bound on the performance of all other FC. Measurements xi are quantized to produce messages mi
energy scheduling schemes, both centralized and distributed. to be passed on to the fusion center; the latter then combines
In fact, our scheme gives an estimate of the amount of energy received messages in order to estimate θ, see Figure 1. The
“wasted” due to the lack of sensor noise correlation knowl- exact form of mi will be detailed later.
edge. The power schedules generated by our design also give We assume that each sensor sends messages to FC using
insight into the design of distributed energy scheduling algo- a separate channel. This can be achieved by using a multi-
rithms. ple access technique such as TDMA or FDMA. Each channel
Our paper is organized as follows. In Section 2 we de- is corrupted by additive white Gaussian noise (AWGN) with
scribe the DES and formulate total energy minimization power spectral density N0 /2:
problem. In Section 3 we present a convex relaxation of
the energy minimization problem and give a nearly-optimal m̂i = di−κ/2 mi + vi , (2)
Minimum Energy DES in a WSN with Correlated Sensor Noises 475

where m̂i is the received message at FC and vi is the AWGN. [a(i) (i) (i) (i)
k , ak+1 ). We denote r = ak+1 − xi and pi = (xi − ak )/∆i ∈
The signal power received at the FC is assumed to be inversely [0, 1]. Then, we have
proportional to diκ where di is the distance between sensor i

2
and the FC, and κ is the path loss exponent. Suppose that var mi = E pi mi − xi
message mi has length bi bits. We will assume that energy Wi

2

required for transmission of mi is proportional to the num- = ∆i − r 1 − p i + r 2 p i
ber of bits in the message. This is the case, for example, if 
(6)
2

sensors use M-QAM or M-PSK modulation to transmit mes- = ∆2i 1 − pi pi + pi2 1 − pi
sages. For example, if M-QAM is used, Wi can be found as

follows [5, 6]: = ∆2i pi 1 − pi .


 Thus, the maximum variance of mi is equal to ∆2i /4 and is

2
 4 1 − 2−s bi
Wi = N f N0 G0 diκ 2s − 1 ln ≡ wi bi , (3) achieved when the observation xi falls in the middle of quan-
3 sPb s
tization interval [a(i) (i)
k , ak+1 ).

where s = log M is the number of bits per symbol, N f is the 2.2. A linear fusion rule
receiver noise figure, Pb is the required bit error probability,
and G0 is the system constant defined as in [5]. The classical best linear unbiased estimator (BLUE) for θ is
given by [9]
2.1. Quantization strategy
1T C−1 x
Suppose that sensor observation xi is bounded to a finite in- θ̂ = , (7)
1T C−1 1
terval [−U, U]. Suppose further that we wish to quantize xi
in such a way that resulting message mi has length bi bits, where x = (x1 , . . . , xN )T and 1 is the vector of all ones. Es-
where bi is to be determined later. We therefore have Ki = 2bi timation performance is characterized by the variance of the
quantization points {a(i)j ∈ [−U, U], j = 1, . . . , Ki }. These estimator
points are uniformly spaced so that a(i) (i)
1 = −U < a2 < · · · <
(i) (i) (i)
−1
aKi = U and ak+1 −ak = ∆i for every k. Since end points {aij } var θ̂ = 1T C−1 1 . (8)
divide the observation range into Ki − 1 intervals, it follows
that ∆i = 2U/(Ki − 1). Quantization is done in the following
To implement BLUE exactly in a WSN setup, we must
probabilistic manner. Suppose that xi ∈ [a(i) (i)
k , ak+1 ). Then xi have mi = xi (i.e., real-valued message) and assume that the
is quantized to either a(i) (i)
k+1 or ak according to channel is distortion-less, both of which are unrealistic in
practice. Nonetheless, BLUE estimator serves as a good per-
  xi − a(i) formance benchmark for the DES to be designed. Motivated
P mi = a(i)
k+1 =
k
,
∆i by the centralized BLUE, we adopt the following fusion rule:
(4) upon receiving sensor messages mi , the FC combines them
  a(i) − xi into an estimator θ̄ given by
P mi = a(i) = k+1 .
k
∆i
1T C−1 m
This probabilistic quantization produces a message mi whose θ̄ = , (9)
1T C−1 1
expected value equals the observation itself:
    where m = (m1 , . . . , mN )T . Equation (5) gives us an impor-
E pi mi = a(i) (i) (i) (i)
k+1 Pr mi = ak+1 + ak Pr mi = ak tant property of θ̄: it is an unbiased estimator for θ. Indeed,
    we have
xi − a(i)
k a(i)
k+1 − xi
= a(i) + a(i) (5) 1T C−1 m 1T C−1 E p m 1T C−1 x
k+1
∆i k
∆i E p,x = Ex = Ex T −1 = θ, (10)
T −
1 C 1 1 T −
1 C 1 1 1 C 1
a(i)
k+1
(i)
− ak
= xi = xi ,
∆i where E p denotes expectation taken with respect to all sensor
quantization noises, and the last step is due to Ex x = θ1. The
where the expectation E pi is taken with respect to the proba- mean-squared error (MSE) of θ̄ can be expanded as follows:
bilistic quantization noise model (4).
Next, we consider any fixed observation value of xi , and MSE(θ̄) = E(θ̄ − θ)2 = E(θ̄ − θ̂ + θ̂ − θ)2
bound the variance var mi (taken with respect to the quan- (11)
tization noise) as follows. Suppose xi falls in the interval = E(θ̄ − θ̂)2 + E(θ̂ − θ)2 + 2 E(θ̄ − θ̂)(θ̂ − θ).
476 EURASIP Journal on Wireless Communications and Networking

Consider the third term in the last expression. We have To complete the formulation, we need to make explicit
the dependence of Q on bi . The unbiasedness of our quanti-

 
Em,x (θ̄ − θ̂)(θ̂ − θ) = Ex Em|x (θ̄ − θ̂)(θ̂ − θ) zation strategy leads to the following important property on
  (12) the quantization noise correlation matrix Q.
= Ex (θ̂ − θ)Em|x (θ̄ − θ̂) = 0,
Lemma 1. The quantization noise matrix Q is diagonal.
where the second step is due to the fact that θ̂ is independent
of m for any fixed x, and the last step follows from (10). Thus, Proof. Consider any (i, j)th element of the matrix Q, with
we can write i = j. We have



MSE(θ̄) Qi j = E mi − xi m j − x j


 
= E(θ̄ − θ̂)2 + E(θ̂ − θ)2
= Exi ,x j E p ,p xi ,x j
mi − xi m j − x j xi , x j
 2 i j

1T C−1 (m − x) 

 
=E + var θ̂ = Exi ,x j Ep mi − xi E p m j − x j |xi , x j = 0.
1T C−1 1 i xi j xj
2 (13) (18)
1
= 1T C−1 E(m − x)(m − x)T C−1 1
1T C−1 1 Here we use the fact that random variables mi and m j are
1 conditionally independent given corresponding observations
+ T −1
1 C 1 xi and x j , which together with (5) gives the desired result.
 
1T C−1 QC−1 1
= + 1 var θ̂,
1T C−1 1 Lemma 1 states that all the off-diagonal entries of Q must
be zero. Let Qii be the ith diagonal element of Q. Recalling
where (6), we obtain the following important bound on the diago-
nal entries of Q:
Q = E(m − x)(m − x)T (14)
U2
is the quantization noise correlation matrix. Qii = var mi ≤
2 , (19)
In our formulation, we seek an energy-efficient DES 2bi − 1
which can deliver an MSE performance that is comparable
to that of the centralized BLUE estimator. Specifically, we where bi is the number of bits in mi . This bound will be use-
will minimize the transmission energy while maintaining the ful in our final formulation of the energy minimization prob-
MSE(θ̄) to be within a constant factor of the BLUE perfor- lem.
mance, that is, MSE(θ̄) ≤ (1 + α) var θ̂ for some constant
α > 0. Therefore, the following condition must hold: 2.3. Total energy minimization
We introduce the notation c = C−1 1 and β = α/ var θ̂. Since
1T C−1 QC−1 1
≤ α. (15) var θ̂ = 1/1T C−1 1, we can rewrite the MSE condition (15) as
1T C−1 1

The total sensor transmission energy is equal to cT Qc ≤ β. (20)

N N
W= Wi = w i bi , (16) This constraint ensures that the MSE performance of the
i=1 i=1 DES is within a factor of α to the BLUE performance. Since
the distribution of x is unknown in general, we enforce a
where wi is the energy required for transmission of a single stronger condition, namely
bit from sensor i to the FC; see (3). Therefore, the minimum
energy DES design problem becomes max cT Qc ≤ β. (21)
x,p
N
minimize W = w i bi Recalling that Q is diagonal (cf. Lemma 1), we can use the
i=1
(17) bound (19) to rewrite the above condition as
1T C−1 QC−1 1
subject to ≤ α, bi ∈ N,
1T C−1 1 N N
U 2 ci2
max cT Qc = max Qii ci2 =
2 ≤ β. (22)
where N denotes the set of nonnegative integers. x,p x,p
i=1 i=1 2bi − 1
Minimum Energy DES in a WSN with Correlated Sensor Noises 477

Now we can reformulate the original energy minimization nonnegativity constraints (since the Lagrangian multipliers
problem (17) explicitly as follows: associated with these constraints will be zero). Associating a
multiplier λ with the MSE constraint, we can write the La-
N grangian for the problem (24) as follows:
minimize w i bi
i=1 N  N 

 ci2 β
N
ci2 β L bi , λ = wi b i + λ
2 − . (25)
subject to ≤ , bi ∈ N, i = 1, . . . , N. i=1 i =1 2bi − 1 U2
i=1
(2bi − 1)2 U2
(23) At the point of optimum we must have ∂L/∂bi = 0 for i =
1, . . . , N, yielding the following set of conditions:
To relate this formulation to physical parameters, we note
that the wireless channel conditions, the choice of modula- ∂L 2bi ci2
tions/BER, and so forth will determine the values of weight- = wi − 2λ ln 2
3 = 0, (26)
∂bi 2bi − 1
ing factors wi , as shown in (3). The values of ci are deter-
mined by the noise correlation matrix C. Without loss of or alternatively
generality we assume ci = 0 for all i. In case ci = 0 for some
sensors, we can exclude corresponding mi from fusion con-
2bi wi λ 
sideration, as it does not contribute to the fusion estimate θ̄.
3 = , (27)
2bi − 1 ci2

3. CONVEX RELAXATION WITH A CLOSED-FORM where λ = 1/2λ ln 2. Also, the main MSE constraint holds
NEARLY-OPTIMAL SOLUTION with equality at optimum point (as noted above), yielding
Since bi can only take integer values, problem (23) is actually N
a nonlinear integer program whose computational complex- ci2 β

2 = . (28)
ity is typically NP-hard. To make this problem computation- i =1 2bi − 1 U2
ally tractable, we relax the integer constraints on bi to allow
them to take real nonnegative values: The optimal solutions {bi , λ } can be found from the non-
linear equations (27) and (28) which unfortunately cannot
N be solved in the closed form. To facilitate a closed-form so-
minimize w i bi lution, we consider a slightly modified system in variables
i=1 {bi∗ , λ∗ }:
(24)
N
ci2
β
subject to bi − 1)2
≤ 2, bi ≥ 0, i = 1, . . . , N. N
ci2 β
i=1
(2 U
2 = , (29)
∗
i=1 2bi −1 U2
The relaxed problem (24) has a linear objective function and ∗
2bi − 1 1 wi λ∗
convex inequality constraints. Therefore, solution to prob-
 3 =
2 = . (30)
∗
2bi − 1
∗
2bi − 1 ci2
lem (24) can be efficiently found by the fusion center using
convex optimization techniques such as the interior point
methods [10]. Once the optimal bi ’s are found, the fusion The above system is almost identical to the original Karush-
center can round this solution to the nearest greater integer Kuhn-Tucker (KKT) system (27) and (28) except for the
and broadcast it to the sensors for power adjustment. small change in the numerators of the left-hand sides of (30)
In what follows, we will present an approximately- and (27). Simple algebraic manipulation shows that (29) and
optimal solution to the problem (24) in closed form. Such a (30) can be solved analytically, yielding
closed-form solution not only simplifies the energy schedul-   −1
N
ing process, but also provides valuable insight into the opti- β
λ∗ = wi . (31)
mal power-scheduling scheme. To begin, we first note that, U2 i =1
by a simple monotonicity argument, the main MSE con-
straint will be active (i.e., holds with equality) at any opti- Substituting this λ∗ into (30) gives the following feasible so-
mum point,1 while the remaining nonnegativity constraints lution to the original energy scheduling problem (24):
on bi will be inactive since bi = 0 for some i would vio-
late the main MSE constraint. Therefore, we can ignore the ci
bi∗ = log 1 + . (32)
λ∗ wi
1 Indeed, the left-hand side of MSE constraint is monotonically decreas-

ing in terms of bi function. Therefore, if at the optimum the inequality is It remains to quantify the performance of this particular en-
strict, we could change bk in the optimal solution to b̃k < bk for some k to ergy scheduling strategy. This is the content of next two lem-
decrease the objective function. mas.
478 EURASIP Journal on Wireless Communications and Networking

Lemma 2. Let {bi , λ } be the optimal solution to the problem Lemma 3 implies that |bi − bi∗ | < 1. Thus, rounded opti-
(24) such that bi ≥ 1 for all i, and let {bi∗ , λ∗ } be its approxi- mal solution bi  is at most one bit away from bi∗ . We can
mation defined by (29) and (30). Then interpret this result as follows: in situation when bi are suf-
ficiently large, for example, when high estimation precision
λ∗ ≤ λ ≤ 2λ∗ . (33) is required, the optimal solution behaves approximately as
log(1 + |ci |/ λ∗ wi ). Notice that ci = eTi C−1 1 (ei denotes the
Proof. Since ith unit vector), so ci signifies the inverse of “noisiness” of
signal xi in relation to the other sensor observations. Recall-
2bi 1 1 ing the definition of λ∗ we note that product ∗
 λ wi is propor-

3 =
2 +
3 , (34) tional to the relative energy per bit wi / w j and the value
2bi −1 2bi −1 2bi −1
of 1/ λ∗ wi can be interpreted as being proportional to the
an upper bound on λ can be found using (27) as follows: relative quality of wireless link between sensor i and the FC.
Thus, the local message length bi∗ can be intuitively inter-
 N  N N preted as being proportional to the logarithm of the product
2bi ci2 ci2 β
λ wi =
3 ≥
 = 2, (35) of signal quality and channel quality at sensor i.
bi − 1 bi − 1 2 U
i=1 i=1 2 i=1 2 We now consider a special case when the use of {bi∗ } is
especially appealing. Suppose that covariance matrix C has a
and we conclude that λ∗ ≤ λ . On the other hand, if all bi ≥ 1 block-diagonal structure
we can write
 
C1 0 · · · 0
1 2bi 2  0 C ··· 0 

2 ≤
3 ≤
2 , (36)  2 
2bi − 1 2bi − 1 2bi − 1 C=
 .. .. . . .. 
. (42)
 . . . . 
therefore λ ≤ 2λ∗ , and the result of the lemma follows. 0 0 · · · Cn

We now bound the difference |bi − bi∗ |. This situation may occur when sensors in the network are
partitioned into several clusters in such a way that sensors
Lemma 3. Under the conditions of Lemma 2, within each group are placed relatively close to each other
and far from the rest of the sensors. Thus, sensor observa-
1 1
bi∗ − < bi < bi∗ + ∀ i = 1, 2, . . . , N. (37) tions are uncorrelated unless they are generated from the
2 2 same cluster. In this case matrix C−1 is also block-diagonal:
Proof. Using left-hand side of (36) and right-hand side of  
C−1 1 0 · · · 0
(33) we can write  0 C−1 · · · 0 
 2 
C −1 = 
 .. .. .. ..  . (43)
1 wi ∗ 2bi  . . . . 

2 ≤
3 ≤ 2 2λ , (38)
2 −1
b i 2 −1
b i ci 0 0 · · · C−n 1

which gives the lower bound on bi : We assume further that sensors within each group can co-
operate to learn the corresponding covariance submatrix C j .
   
ci √ ci 1 1 Value of λ∗ can be computed by the fusion center and broad-
bi ≥ log 1 + = log 2+ ∗ − > bi∗ − . casted back to the sensors. Thus, each sensor can easily com-
2λ∗ wi λ wi 2 2
pute ci = [C−j 1 1]i and independently find its own quantiza-
(39)
tion level bi∗ . The advantage of this method is that the fusion
center needs to broadcast only one universal message for all
By analogy, from right-hand side of (36) and left-hand side
sensors.
of (33) we have
To conclude this section we observe that our strategy can
be applied even if sensor noises have infinite range. Indeed,
2 2bi wi ∗

2 ≥
3 ≥ λ , (40) with an appropriate choice of U, that is, if tails of the noise
2bi −1 2bi −1 ci2 pdf are negligible, the pdf can be approximated by a finite
support function. However, the estimator (9) will no longer
which further implies
be unbiased and cross terms E(θ̄ − θ̂)(θ̂ − θ) in the MSE ex-
 √    pression will no longer be zero. Thus, inequality (15) only
2 ci 1 ci 1 ∗ 1 defines a lower bound on estimation performance for some
bi ≤ log 1 + = log √ + + <b + .
λ∗ wi 2 λ∗ wi 2 i 2 α, and the gap between left-hand side of (15) and actual MSE
(41) is determined by the noise pdf. Therefore, the full pdf knowl-
edge will be required in order to specify constants U and α
This completes the proof. and quantify the estimation bias.
Minimum Energy DES in a WSN with Correlated Sensor Noises 479

4. NUMERICAL SIMULATIONS Recall that we have relaxed bi to take real values to make
In this section, we present numerical simulations to compare the problem convex. Therefore, the optimal energy obtained
the transmission energy requirement for two energy schedul- by allowing bi to take on real values is a lower bound on the
ing strategies: (i) quantization using the closed-form approx- actual optimal energy. If we round bi up to the closest integer
bi , we can obtain an upper bound (denoted by W opt ) on
imate solution (32); (ii) uniform bit allocation when all sen-
sors quantize their observations to the same number of bits the actual energy. Even though we use bi∗  to approximate
to achieve the same MSE. We denote by b the number of bits the actual optimal solution, significant energy can be saved
used in case of uniform bit allocation. We can find the mini- when compared with the uniform bit allocation strategy in
mum of b from the MSE constraint order to achieve the same target distortion. The percentage
of saving is defined as
N
ci2 β

2 ≤ , (44) Wuniform − W opt
i=1 2b − 1 U2 × 100. (47)
Wuniform
which gives
For a positive random variable R we define

 

 2 N √
 U  var R
b ≥ log 1 +  ci2 . (45) normalized deviation of R = , (48)
β i=1 ER

which will be used as a measure of the absolute heterogeneity

The number of bits can only take integer values, so the total
of R. The sensor noise variances {σi2 } are taken to be σi2 =
minimal energy is given by
1 + a2 Zi , where Zi are i.i.d. random variables with Zi ∼ χ12 (z).
    As can be easily verified, {σi2 } are also i.i.d. with σi ∼ χ12 ((x −

 2 N N 1)/a2 ). We control heterogeneity of sensor noise variances by
  U 
Wuniform =  log 1 +  ci2  wi . (46) varying the parameter a. In Figure 2a, we suppose that sensor
 β 
 i=1  i=1
noises have tri-diagonal correlation matrix

 
1 ρ ··· 0 0
ρ 1 ··· 0 0
 

 . .. . . .. .. 

C = diag σ1 , σ2 , . . . , σN 
 .. . . .

. diag σ1 , σ2 , . . . , σN , (49)
 
0 0 · · · 1 ρ
0 0 ··· ρ 1

where ρ = 0.2. In Figure 2b, we suppose that sensor noises supposing that sensors have same observation noise vari-
have correlation matrix ances with tri-diagonal structure as in (49) where σi2 = 1
for all i, and ρ = 0.2. Here we suppose that all sensors are
C

 
 uniformly distributed inside a unitary disk whose center is at
= diag σ1 , σ2 , . . . , σN (1 − ρ)I+ρ 11T diag σ1 , σ2 , . . . , σN . the FC. It is easy to show that in this case normalized devi-
(50) ation of wi depends only on κ (cf. (51)). In our simulation,
In all simulations, the total number of sensors N = 200. Since we choose 1 ≤ κ ≤ 8. We observe that percentage of saving
all coefficients wi are scaled by a common factor, in our sim- depends more on the heterogeneity of sensor noise variances
ulation, {wi } are taken to be channel path losses than that of channel gains. This can be understood regarding
expression (32) for bi∗ , where in the logarithm, the quantity
wi = diκ . (51) depends on the distribution of ci , but only on the distribu-
√
tion of 1/ wi .
Assume that the target estimation performance is fixed.
From Figure 2 we can see that the amount of energy sav-
ing becomes significant when the local noise variances be- 5. AN EXTENSION: MINIMAX FORMULATION
come more and more heterogeneous, assuming that all sen- Minimizing total transmission energy results in sensors hav-
sors have identical wi . In Figure 3, we plot the percentage ing different lifetimes. This may induce frequent changes in
of energy savings versus the heterogeneity of channel gains, the network topology. An alternative approach is to minimize
480 EURASIP Journal on Wireless Communications and Networking

70 70

60 60

Energy saving in percentage

Energy saving in percentage

50 50

40 40

30 30

20 20

10 10

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1 1.2 1.4
Normalized deviation of sensor noise variances Normalized deviation of sensor noise variances
(a) (b)

Figure 2: Percentage of energy saving increases when sensor noise variances become more heterogeneous.

maximal energy Wi which leads to maximum network life- which give

time. Relaxing {bi } as in (24), we can state the problem as
follows: N
µi = 1, (56)
minimize max wi bi i=1
i
2bi ci2
N
ci2 β λ µi =
3 , (57)
subject to
2 ≤ , bi ≥ 0, i = 1, . . . , N, 2bi − 1 wi
i=1 2bi − 1 U2
(52) where as before λ = 1/2λ ln 2. Taking sum of (57) over all i
we obtain
or alternatively
N
ci2 2bi
minimize max t λ =
3 . (58)
w
i=1 i 2 i − 1
b
N
ci2 β
subject to wi bi ≤ t
2 ≤ , bi ≥ 0, i = 1, . . . , N. Since each term in the right-hand side sum in (58) is positive,
i=1 2bi − 1 U2
we conclude that λ > 0, therefore µi > 0, and complimentary
(53) slackness condition gives
As in Section 3, we assume that ci = 0 for all i and ignore the N
nonnegativity constraints bi ≥ 0 (which must be inactive at ci2 β

2 = ,
optimum). The Lagrangian for problem (53) is found to be i =1 2bi −1 U2 (59)

N  N  wi bi = t.

 β
 ci2
L t, bi , µi , λ = t+ µi wi bi − t +λ
2 − 2 .
i=1 i=1 2 i − 1
b U Thus, the optimal value topt can be found as a solution to the
(54) following equation:

Differentiating L with respect to primal variables we obtain N

ci2 β
the following conditions:
2 = . (60)
i=1 2t/wi −1 U2
N
∂L The solution topt is unique due to the monotonicity of the
=1− µi = 0,
∂t i=1 left-hand side function in (60). The FC can solve (60) and
(55) broadcast topt to the sensors, which in turn can determine
∂L 2bi ci2 their quantization levels locally. In this case sensor lifetime is
= −2λ ln 2
3 + µi wi = 0,
∂bi 2bi − 1 not affected by transmitted power.
Minimum Energy DES in a WSN with Correlated Sensor Noises 481

[2] Z.-Q. Luo, “An isotropic universal decentralized estimation

20 scheme for a bandwidth constrained ad hoc sensor network,”
IEEE J. Select. Areas Commun., vol. 23, no. 4, pp. 735–744,
2005.
[3] Z.-Q. Luo and J.-J. Xiao, “Universal decentralized estimation
Energy saving in percentage

15 in an inhomogeneous environment,” to appear in IEEE Trans.

Inform. Theory.
[4] Z.-Q. Luo and J.-J. Xiao, “Decentralized estimation in an
inhomogeneous environment,” in Proc. IEEE International
10 Symposium on Information Theory (ISIT ’04), pp. 517–517,
Chicago, Ill, USA, June–July 2004.
[5] S. Cui, A. J. Goldsmith, and A. Bahai, “Energy-constrained
modulation optimization,” to appear in IEEE Transaction on
5 Wireless Communications.
[6] S. Cui, A. J. Goldsmith, and A. Bahai, “Joint modulation and
multiple access optimization under energy constraints,” in
Proc. IEEE Global Telecommunications Conference (GLOBE-
0 COM ’04), pp. 151–155, Dallas, Tex, USA, November–
0.4 0.6 0.8 1 1.2 1.4 December 2004.
Normalized deviation of channel gains [7] J.-J. Xiao, S. Cui, Z.-Q. Luo, and A. J. Goldsmith, “Power
scheduling of universal decentralized estimation in sensor
networks,” to appear in IEEE Trans. Signal Processing.
Figure 3: Percentage of energy saving increases when channel gains [8] X. Luo and G. B. Giannakis, “Energy-constrained optimal
become more heterogeneous. quantization for wireless sensor networks,” in Proc. 1st IEEE
Annual Communications Society Conference on Sensor and Ad
Hoc Communications and Networks (SECON ’04), pp. 272–
278, Santa Clara, Calif, USA, October 2004.
[9] S. M. Kay, Fundamentals of Statistical Signal Processing: Esti-
6. CONCLUSION mation Theory, Prentice-Hall, Upper Saddle River, NJ, USA,
1993.
In this paper we have shown that total energy consumption [10] S. Boyd and L. Vandenberghe, Convex Optimization, Cam-
required for transmission in a sensor network can be min- bridge University Press, Cambridge, UK, 2003.
imized if number of quantization levels for each sensor is
determined jointly by the fusion center using information
Alexey Krasnopeev received the B.S. de-
about correlation of sensor observations. We have also pre- gree in applied mathematics and physics in
sented a nearly-optimal solution in closed form to the energy 1999, and the M.S. degree in applied math-
minimization problem which can achieve the same target es- ematics and physics in 2001, both from the
timation performance as the optimal solution. It is shown by Moscow Institute of Physics and Technol-
numerical simulations that to attain the same MSE perfor- ogy, Moscow, Russia. He is currently pursu-
mance our energy-efficient quantization scheme can achieve ing his M.S. degree in electrical engineering
energy saving up to 70% when compared to simple uniform at the University of Minnesota. His research
bit allocation scheme. We plan to consider various exten- interests include wireless sensor networks,
sions of this work in our future work. These include joint information theory, and algebraic coding
theory.
estimation of a common vector signal by a WSN, and dis-
tributed least squares and target tracking for dynamic tar-
Jin-Jun Xiao received the B.S. degree in
gets. applied mathematics from Jilin University,
China, in 1997, and the M.S. degree in
mathematics from the University of Min-
ACKNOWLEDGMENTS nesota, in 2003. He is currently pursuing the
Ph.D. degree in electrical engineering at the
Authors would like to thank the anonymous reviewers for University of Minnesota. His research inter-
their valuable comments that helped improve the quality of ests are in wireless sensor networks, infor-
this paper. This research is supported in part by the Natural mation theory, and optimization.
Sciences and Engineering Research Council of Canada, Grant
no. OPG0090391, by the Canada Research Chair Program, Zhi-Quan Luo received the B.S. degree
and by the National Science Foundation, Grant no. DMS- in mathematics from Peking University,
0312416. China, in 1984. During the academic year
of 1984/1985, he was with Nankai Insti-
tute of Mathematics, Tianjin, China. From
REFERENCES 1985 to 1989, he studied at the Department
[1] Z.-Q. Luo, “Universal decentralized estimation in a band- of Electrical Engineering and Computer
width constrained sensor network,” to appear in IEEE Trans. Science, Massachusetts Institute of Tech-
Inform. Theory. nology, where he received a Ph.D. degree
482 EURASIP Journal on Wireless Communications and Networking

in operations research. In 1989, he joined the Department of Elec-

trical and Computer Engineering, McMaster University, Hamilton,
Canada, where he became a Professor in 1998 and held the Canada
Research Chair in information processing since 2001. Starting April
2003, he has been a Professor in the Department of Electrical and
Computer Engineering at the University of Minnesota, and holds
an ADC Chair in digital technology. His research interests lie in
the union of large-scale optimization, signal processing, data com-
munications, and information theory. He is a Member of SIAM
and MPS. He is presently serving as an Associate Editor for sev-
eral international journals including SIAM Journal on Optimiza-
tion, Mathematical Programming, Mathematics of Computation,
and Mathematics of Operations Research.