Hindawi Publishing Corporation

EURASIP Journal on Applied Signal Processing

Volume 2006, Article ID 20858, Pages 1–23
DOI 10.1155/ASP/2006/20858

A Constrained Least Squares Approach to Mobile Positioning:

Algorithms and Optimality

K. W. Cheung,1 H. C. So,1 W.-K. Ma,2 and Y. T. Chan3

1 Department of Electronic Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong
2 Department of Electrical Engineering, National Tsing Hua University, Hsinchu 30013, Taiwan
3 Department of Electrical & Computer Engineering, Royal Military College of Canada, Kingston, ON, Canada K7K 7B4

Received 20 May 2005; Revised 25 November 2005; Accepted 8 December 2005

The problem of locating a mobile terminal has received significant attention in the field of wireless communications. Time-of-
arrival (TOA), received signal strength (RSS), time-difference-of-arrival (TDOA), and angle-of-arrival (AOA) are commonly used
measurements for estimating the position of the mobile station. In this paper, we present a constrained weighted least squares
(CWLS) mobile positioning approach that encompasses all the above described measurement cases. The advantages of CWLS in-
clude performance optimality and capability of extension to hybrid measurement cases (e.g., mobile positioning using TDOA and
AOA measurements jointly). Assuming zero-mean uncorrelated measurement errors, we show by mean and variance analysis that
all the developed CWLS location estimators achieve zero bias and the Cramér-Rao lower bound approximately when measurement
error variances are small. The asymptotic optimum performance is also confirmed by simulation results.

Copyright © 2006 Hindawi Publishing Corporation. All rights reserved.

1. INTRODUCTION In the TOA method, the distance between the MS and BS

is determined from the measured one-way propagation time
Accurate positioning of a mobile station (MS) will be one of the signal traveling between them. For two-dimensional
of the essential features that assists third generation (3G) (2D) positioning, this provides a circle centered at the BS
wireless systems in gaining a wide acceptance and trigger- on which the MS must lie. By using at least three BSs to re-
ing a large number of innovative applications. Although the solve ambiguities arising from multiple crossings of the lines
main driver of location services is the requirement of lo- of position, the MS location estimate is determined by the
cating Emergency 911 (E-911) callers within a specified ac- intersection of circles. The RSS approach employs the same
curacy in the United States [1], mobile position informa- trilateration concept where the propagation path losses from
tion will also be useful in monitoring of the mentally im- the MS to the BSs are measured to give their distances. In the
paired (e.g., the elderly with Alzheimer’s disease), young TDOA method, the differences in arrival times of the MS sig-
children and parolees, intelligent transport systems, location nal at multiple pairs of BSs are measured. Each TDOA mea-
billing, interactive map consultation and location-dependent surement defines a hyperbolic locus on which the MS must
e-commerce [2–6]. Global positioning system (GPS) could lie and the position estimate is given by the intersection of
be used to provide mobile location, however, it would be two or more hyperbolas. Finally, the AOA method necessi-
expensive to be adopted in the mobile phone network be- tates the BSs to have multielement antenna arrays for mea-
cause additional hardware is required in the MS. Alterna- suring the arrival angles of the transmitted signal from the
tively, utilizing the base stations (BSs) in the existing net- MS at the BSs. From each AOA estimate, a line of bearing
work for mobile location is preferable and is more cost effec- (LOB) from the BS to the MS can be drawn and the position
tive for the consumer. The basic principle of this software- of the MS is calculated from the intersection of a minimum
based solution is to use two or more BSs to intercept of two LOBs. In general, the MS position is not determined
the MS signal, and common approaches [6–8] are based geometrically but is estimated from a set of nonlinear equa-
on time-of-arrival (TOA), received signal strength (RSS), tions constructed from the TOA, RSS, TDOA, or AOA mea-
time-difference-of-arrival (TDOA), and/or angle-of-arrival surements, with knowledge of the BS geometry.
(AOA) measurements determined from the MS signal re- Basically, there are two approaches for solving the non-
ceived at the BSs. linear equations. The first approach [9–12] is to solve them
2 EURASIP Journal on Applied Signal Processing

directly in a nonlinear least squares (NLS) or weighted least Table 1: List of abbreviations and symbols.
squares (WLS) framework. Although optimum estimation
performance can be attained, it requires sufficiently precise AOA Angle-of-arrival
initial estimates for global convergence because the corre- CWLS Constrained weighted least squares
sponding cost functions are multimodal. The second ap- CRLB Cramér-Rao lower bound
proach [13–17] is to reorganize the nonlinear equations into NLS Nonlinear least squares
a set of linear equations so that real-time implementation is
RSS Received signal strength
allowed and global convergence is ensured. In this paper, the
latter approach is adopted, and we will focus on a unified de- TOA Time-of-arrival
velopment of accurate location algorithms, given the TOA, TDOA Time-difference-of-arrival
RSS, TDOA, and/or AOA measurements. AT Transpose of matrix A
For TDOA-based location systems, it is well known that A−1 Inverse of matrix A
for sufficiently small noise conditions, the corresponding Ao Optimum matrix of A
nonlinear equations can be reorganized into a set of linear
equations by introducing an intermediate variable, which is σ2 Noise variance
a function of the source position, and this technique is com- Cn Noise covariance matrix
monly called spherical interpolation (SI) [13]. However, the I(x) Fisher information matrix for parameter vector x
SI estimator solves the linear equations via standard least x Optimization variable vector for x
squares (LS) without using the known relation between the
x Estimate of x
intermediate variable and the position coordinate. To im-
prove the location accuracy of the SI approach, Chan and diag(x) Diagonal matrix formed from vector x
Ho have proposed [14] to use a two-stage WLS to solve IM M × M identity matrix
for the source position by exploiting this relation implic- 1M M × 1 column vector with all ones
itly via a relaxation procedure, while [15] incorporates the 0M M × 1 column vector with all zeros
relation explicitly by minimizing a constrained LS function
OM ×N M × N matrix with all zeros
based on the technique of Lagrange multipliers. According
to [15], these two modified algorithms are referred to as the  Element-by-element multiplication
quadratic correction least squares (QCLS) and linear correc-
tion least squares (LCLS), respectively. Recently, we have im-
proved [18] the performance of the LCLS estimator by in- is studied in Section 4. Simulation results are presented in
troducing a weighting matrix in the optimization, which can Section 5 to evaluate the location estimation performance of
be regarded as a hybrid version of the QCLS and LCLS algo- the proposed estimators and verify our theoretical findings.
rithms. The idea of this constrained weighted least squares Finally, conclusions are drawn in Section 6. A list of abbre-
(CWLS) technique has also been extended to the RSS [19] viations and symbols that are used in the paper is given in
and TOA [20] measurements. Using a different way of con- Table 1.
verting nonlinear equations to linear equations without in-
troducing dummy variables, Pages-Zamora et al. [16] have
developed a simple LS AOA-based location algorithm. In this 2. MEASUREMENT MODELS
work, our contributions include (i) development of a unified
In this section, the models and assumptions for the TOA,
approach for mobile location which allows utilizing different
TDOA, RSS, and AOA measurements are described. Let x =
combinations of TOA, RSS, TDOA, and AOA measurements
[x, y]T be the MS position to be determined and let the
via generalizing [18–20] and improving [16] with the use
known coordinates of the ith BS be xi = [xi , yi ]T , i = 1, 2,
of WLS; and (ii) derivation of bias and variance expressions
. . . , M, where the superscript T denotes the transpose opera-
for all the proposed algorithms. In particular, we prove that
tion and M is the total number of receiving BSs. The distance
the performance of all the proposed estimation methods can
between the MS and the ith BS, denoted by di , is given by
achieve zero bias and the Cramér-Rao lower bound (CRLB)
[21] approximately when the measurement errors are uncor-  2  2
related and small in magnitude. di = x − xi + y − yi , i = 1, 2, . . . , M. (1)
The rest of this paper is organized as follows. In Section 2,
we formulate the models for the TOA, TDOA, RSS, and
AOA measurements and state our assumptions. In Section 3,
2.1. TOA measurement
three CWLS location algorithms using TDOA, RSS, and TOA
measurements, respectively, are first reviewed, and a WLS The TOA is the one-way propagation time taken for the sig-
AOA-based location algorithm is then devised via modi- nal to travel from the MS to a BS. In the absence of distur-
fying [16]. Mobile location using various combinations of bance, the TOA measured at the ith BS, denoted by ti , is
TOA, TDOA, RSS, and AOA measurements is also examined.
In particular, a TDOA-AOA hybrid algorithm is presented
di
in detail. The performance of all the developed algorithms ti = , i = 1, 2, . . . , M, (2)
c
K. W. Cheung et al. 3

where c is the speed of light. The range measurement based where

on ti in the presence of disturbance, denoted by rTOA,i , is  T
modeled as rTDOA = rTDOA,2 rTDOA,3 · · · rTDOA,M ,

 T
rTOA,i = di + nTOA,i nTDOA = nTDOA,2 nTDOA,3 · · · nTDOA,M ,
 2  2 ⎡  2  2  2  2 ⎤
= x − xi + y − yi + nTOA,i , i = 1, 2, . . . , M, x − x2 + y − y 2 − x − x1 + y − y 1
⎢ ⎥
⎢ ⎥
(3) ⎢  2  2  2  2 ⎥
⎢ x − x + y − y − x − x + y − y ⎥
⎢ 3 3 1 1 ⎥
fTDOA (x) = ⎢
⎢ .. ⎥.
⎥
⎢ . ⎥
where nTOA,i is the range error in rTOA,i . Equation (3) can also ⎢ ⎥
⎣ ⎦
be expressed in vector form as  2  2  2  2
x − xM + y − y M − x − x1 + y − y 1
(8)
rTOA = fTOA (x) + nTOA , (4)
2.3. RSS measurement
where Without measurement error, the RSS or received power at
the ith BS, denoted by Pir , can be modeled as [22]
 T
rTOA = rTOA,1 rTOA,2 · · · rTOA,M , Pit
Pir = Ki , i = 1, 2, . . . , M, (9)
dia
 T
nTOA = nTOA,1 nTOA,2 · · · nTOA,M , where Pit is the transmitted power, Ki accounts for all other
factors which affect the received power, including the an-
⎡  2  2 ⎤ tenna height and antenna gain, and a is the propagation con-
x − x1 + y − y 1 (5)
⎢ ⎥ stant. Note that the propagation parameter a can be obtained
⎢  2  2 ⎥
⎢ ⎥ via finding the path loss slope by measurement [22]. In free
⎢ x − x2 + y − y 2 ⎥
fTOA (x) = ⎢
⎢ ..
⎥.
⎥ space, a is equal to 2, but in some urban and suburban areas,
⎢ . ⎥ a can vary from 3 to 6. From (9), the range measurements
⎢ ⎥
⎣ ⎦ based on the RSS data with the use of the known {Pit } and
 2  2
x − xM + y − y M {Ki }, denoted by {rRSS,i }, are determined as

Pit
rRSS,i = Ki + nRSS,i
Pir
2.2. TDOA measurement  2  2 a/2
= x − xi + y − yi + nRSS,i , i = 1, 2, . . . , M,
The TDOA is the difference in TOAs of the MS signal at a pair (10)
of BSs. Assigning the first BS as the reference, it can be easily
deduced that the range measurements based on the TDOAs where nRSS,i is the range error in rRSS,i . It is noteworthy that
are of the form if a = 1, then (10) will be of the same form as (3). Equation
(10) can also be expressed in vector form as
 
rTDOA,i = di − d1 + nTDOA,i rRSS = fRSS (x) + nRSS , (11)

  where
2  2 2  2  T
= x − xi + y − yi − x − x1 + y − y1
rRSS = rRSS,1 rRSS,2 · · · rRSS,M ,
+ nTDOA,i , i = 2, 3, . . . , M,  T
(6) nRSS = nRSS,1 nRSS,2 · · · nRSS,M ,
⎡ ⎤
 2  2 a/2
where nTDOA,i is the range error in rTDOA,i . Notice that if the ⎢ x − x 1 + y − y 1 ⎥
⎢ ⎥ (12)
TDOA measurements are directly obtained from the TOA ⎢      ⎥
⎢ 2 2 a/2 ⎥
data, then nTDOA,i = nTOA,i − nTOA,1 , i = 2, 3, . . . , M. In vector ⎢ x − x2 + y − y 2 ⎥
⎢ ⎥
fRSS (x) = ⎢
⎢ . ⎥.
⎥
form, (6) becomes ⎢ . ⎥
⎢ . ⎥
⎢ ⎥
⎣  2  2  a/2 ⎦
rTDOA = fTDOA (x) + nTDOA , (7) x − xM + y − y M
4 EURASIP Journal on Applied Signal Processing

2.4. AOA measurement 3. ALGORITHM DEVELOPMENT

The AOA of the transmitted signal from the MS at the ith BS, This section describes our development of the CWLS/WLS
denoted by φi , is related to x and xi by mobile positioning approach for the cases of TDOA, RSS,
TOA, and AOA measurements. We also discuss how the
  y − yi proposed methods can be extended to hybrid measurement
tan φi = , i = 1, 2, . . . , M. (13)
x − xi cases, such as the TDOA-AOA.

Geometrically, φi is the angle between the LOB from the ith 3.1. TDOA [18]
BS to the MS and the x-axis. The AOA measurements in the
presence of angle errors, denoted by {rAOA,i }, are modeled as Without disturbance, (6) becomes
   
−1 y − yi 2  2 2  2
rAOA,i = φi +nAOA,i = tan +nAOA,i , i = 1, 2, . . . , M, rTDOA,i = x − xi + y − yi − x − x1 + y − y1
x − xi
(14)  2  2
=⇒ rTDOA,i + x − x1 + y − y1
where nAOA,i is the noise in rAOA,i . Equation (14) can also be
expressed in vector form as  2  2
= x − xi + y − yi , i = 2, 3, . . . , M.
rAOA = fAOA (x) + nAOA , (15) (17)

where Squaring both sides of (17) and introducing an intermediate

variable, R1 , which has the form
 T
rAOA = rAOA,1 rAOA,2 · · · rAOA,M ,
 2  2
 T R1 = d1 = x − x1 + y − y1 , (18)
nAOA = nAOA,1 nAOA,2 · · · nAOA,M ,

⎡   we obtain the following set of linear equations [13]

−1
y − y1 ⎤
⎢ tan x − x1 ⎥
⎢ ⎥ (16)      
⎢  ⎥ x − x1 xi − x1 + y − y1 yi − y1 + rTDOA,i R1
⎢ y − y 2 ⎥
⎢ tan−1 ⎥
⎢
fAOA (x) = ⎢ x − x2 ⎥⎥.
⎢ ⎥ 1  2  2 2 
⎢ .. ⎥ = xi − x1 + yi − y1 − rTDOA,i , i = 2, 3, . . . , M.
⎢ . ⎥ 2
⎢ ⎥
⎢  ⎥ (19)
⎣ y − y M ⎦
tan−1
x − xM
Writing (19) in matrix form gives
To facilitate the development and analysis of the pro-
posed location algorithms, we make the following assump-
Gϑ = h, (20)
tions for the TOA, TDOA, RSS, and AOA measurements.
(A1) All measurement errors, namely, {nTOA,i }, {nTDOA,i },
where
{nRSS,i }, and {nAOA,i } are sufficiently small and are
modeled as zero-mean Gaussian random variables ⎡ ⎤
with known covariance matrices, denoted by Cn,TOA , x2 − x1 y2 − y1 rTDOA,2
⎢ ⎥
Cn,TDOA , Cn,RSS , and Cn,AOA , respectively. The zero- ⎢ .. .. .. ⎥
mean error assumption implies that multipath and G=⎢
⎢ . . . ⎥,
⎥
⎣ ⎦
non-line-of-sight (NLOS) errors have been mitigated, xM − x1 yM − y1 rTDOA,M
which can be done by considering the techniques in
[23–27]. Nevertheless, the effect of NLOS propaga- ⎡  2  2 ⎤ (21)
2
tion will be studied in Section 5 for the TOA measure- x2 − x1 + y2 − y1 − rTDOA,2
⎢ ⎥
ments. 1⎢
⎢ .
⎥
⎥
h= ⎢ .. ⎥,
(A2) For RSS-based location, the propagation parameter a 2⎢
⎣
⎥
⎦
is known and has a constant value for all RSS measure- 2  2 2
xM − x1 + yM − y1 − rTDOA,M
ments.
(A3) The numbers of BSs for location using the TOA,
TDOA, RSS, and AOA measurements are at least 3, 4, and the parameter vector ϑ = [x − x1 , y − y1 , R1 ]T consists of
3, and 2, respectively. the MS location as well as R1 .
K. W. Cheung et al. 5

In the presence of measurement errors, the SI technique Similar to [15], the CWLS problem is solved by using the
determines the MS position by simply solving (20) via stan- technique of Lagrange multipliers and the Lagrangian to be
dard LS, and the location estimate is found from [13] minimized is

ϑ = arg min(Gϑ̆ − h)T (Gϑ̆ − h)

T
LTDOA (ϑ̆, η) = (Gϑ̆ − h)T Υ−1 (Gϑ̆ − h) + ηϑ̆ Σϑ̆, (27)
ϑ̆
(22)
 −1
= GT G GT h, where η is the Lagrange multiplier to be determined. The es-
timate of ϑ is obtained by differentiating LTDOA (ϑ̆, η) with
where ϑ̆ = [x̆ − x1 , y̆ − y1 , R˘1 ]T is an optimization variable respect to ϑ̆ and then equating the results to zero (see Appen-
vector and −1 represents the matrix inverse, without utilizing dix A.1):
the known relationship between x̆, y̆, and R˘1 .
 −1
An improvement to the SI estimator is the LCLS method ϑ = GT Υ−1 G + ηΣ GT Υ−1 h, (28)
[15], which solves the LS cost function in (22) subject to the
constraint of (x̆ − x1 )2 + ( y̆ − y1 )2 = R˘21 , or equivalently, where η is found from the following 4-root equation:
T
ϑ̆ Σϑ̆ = 0, (23) 
3
αi βi
 2 = 0 (29)
where Σ = diag(1, 1, −1). i=1 η + ζi
On the other hand, Chan and Ho [14] have improved
the SI estimator through two stages. In the first stage of the and {αi }, {βi }, and {ζi }, i = 1, 2, 3, have been defined in Ap-
QCLS estimator, a coarse estimate is computed by minimiz- pendix A.1. The procedure for CWLS TDOA-based location
ing a WLS function is summarized as follows.
(i) Set Υ = IM −1 , where IM −1 denotes the identity matrix
(Gϑ̆ − h)T Υ−1 (Gϑ̆ − h), (24) of dimension (M − 1).
(ii) Find all roots of (29) by using a standard root finding
where Υ is a symmetric weighting matrix, which is a function algorithm. Then take only the real roots into consider-
of the estimate of R1 , denoted by R1 . A better estimate of ϑ is ation as the Lagrange multiplier is always real for a real
then obtained in the second stage via minimizing (x̆ − x1 )2 + optimization problem.
( y̆ − y1 )2 − R˘21 according to another WLS procedure. Since (iii) Put the real η’s back to (28) and obtain subestimates of
R1 is not available at the beginning, normally a few iterations  Then choose the solution ϑ
ϑ.  from those subestimates
between the two stages are required to attain the best solution which makes the expression (Gϑ̆ − h)T Υ−1 (Gϑ̆ − h)
[15]. minimum.
The idea of our CWLS estimator is to combine the key (iv) Construct Υ according to (25) using the obtained R1 in
principles in the CWLS and LCLS methods, that is, the MS step (iii). Then, repeat steps (ii) and (iii) until ϑ con-
position estimate is determined by minimizing (24) subject verges.
to (23). For sufficiently small measurement errors, the in-
verse of the optimum weighting matrix Υ−1 for the CWLS 3.2. RSS [19]
algorithm is found using the best linear unbiased estimator
(BLUE) [21] as in [14]: Without measurement errors, (10) becomes

Υo = s1 sT1  Cn,TDOA , (25)  2  2 a/2

rRSS,i = x − xi + y − yi , i = 1, 2, . . . , M. (30)
where
⎡ ⎤ Extending the SI technique and taking power 2/a on both
d2 ⎡ ⎤
d2 − d1 + R1 sides of (30) yields
⎢ ⎥ ⎢d −d +R ⎥
⎢d ⎥ ⎢ 3 1⎥
⎢ 3⎥ ⎢ 1
⎥  
⎢ ⎥ ⎢ ⎥ 2/a
s1 = ⎢ ⎥= .. (26) rRSS,i = R22 − 2xxi − 2y yi + xi2 + yi2
⎢ .. ⎥ ⎢ ⎥
⎢ . ⎥ ⎢ . ⎥
⎣ ⎦ ⎣ ⎦
dM − d1 + R1 =⇒ xi x + yi y − 0.5R22 (31)
dM
1 2 
= xi + yi2 − ri2/a , i = 1, 2, . . . , M,
and  denotes element-by-element multiplication. Since Υ 2
contains the unknown {di }, we express di = di − d1 + R1
and approximate di − d1 by rTDOA,i and thus an approximate where
version of Υo , namely, s1sT1  Cn,TDOA with s1 = [rTDOA,2 + 
R1 · · · rTDOA,M + R1 ]T is employed in practice. R2 = x2 + y 2 (32)
6 EURASIP Journal on Applied Signal Processing

is the introduced intermediate variable in order to linearize The covariance matrix of the disturbance, which leads to the
(30) in terms of x, y, and R22 . Similar to the TDOA measure- optimum weighting matrix, is thus of the form
ments, (31) can be expressed in matrix-vector form:
 
Ψo = E εεT = s2 sT2  Cn,RSS , (41)
Aθ = b, (33)

where where
⎡ ⎤ ⎡ ⎤
x1 y1 −0.5 x  T
⎢ ⎥ ⎢ ⎥ 1  2−a 1  2−a 1  2−a
⎢ . .. .. ⎥ ⎢ ⎥ s2 = d1 d2 ··· dM . (42)
A=⎢
⎢ .. .
⎥,
. ⎥ θ = ⎢ y ⎥, a a a
⎣ ⎦ ⎣ ⎦
xM yM −0.5 R22
(34) Since s2 depends on the unknowns {di }, we use {ri1/a } instead
⎡ 2/a ⎤ of {di } to form an estimate of s2 , denoted by s2 , which is
x12 + y12 − rRSS,1
⎢ ⎥
1⎢
⎢ .. ⎥
⎥.
b=  T
2⎢
⎣
. ⎥
⎦ 1 2/a−1 1 2/a−1 1 2/a−1
2 2 2/a
s2 = rRSS,1 rRSS,2 · · · rRSS,M . (43)
xM + yM − rRSS,M a a a

The CWLS estimate of θ is obtained by minimizing Minimizing (35) subject to (36) is equivalent to minimizing
the Lagrangian
(Aθ̆ − b)T Ψ−1 (Aθ̆ − b), (35)
 T 
where Ψ−1 is the corresponding weighting matrix, subject to LRSS (θ̆, λ) = (Aθ̆ − b)T Ψ−1 (Aθ̆ − b) + λ qT θ̆ + θ̆ Pθ̆ ,
(44)
T
qT θ̆ + θ̆ Pθ̆ = 0 (36)
where λ is the corresponding Lagrange multiplier. The CWLS
such that solution using the RSS measurements is given by (see Appen-
dix A.2)
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
x̆ 1 0 0 0
⎢ ⎥ ⎢ ⎥ ⎢ ⎥  
⎢ ⎥ ⎢ ⎥ ⎢ ⎥  −1 λ
θ̆ = ⎢ y̆ ⎥ , P = ⎢0 1 0⎥ , q = ⎢ 0 ⎥. (37) θ = AT Ψ−1 A + λP AT Ψ−1 b − q , (45)
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
2
R̆2 0 0 0 −1

Here, (36) is a matrix characterization of the relation in (32). where λ is determined from the 5-root equation:
The optimum value of Ψ is also determined based on the
BLUE as follows. For sufficiently small measurement errors,
2/a λ 2
ci f i λ  c i gi
2 2
ei f i γ i
the value of rRSS,i can be approximated as c3 f 3 − c 3 g 3 + − +  2
2 i=1
1 + λγ i 2 i=1
1 + λγ i i=1 1 + λγi
2/a  2/a
rRSS,i = dia + nRSS,i
λ  e i gi γ i λ  ci f i γ i λ2  ci gi γi
2 2 2
(38) −  2 −  2 +   = 0.
2  2−a 2 i=1 1+λγi 2 i=1 1+λγi 4 i=1 1+λγi 2
≈ di2 + di nRSS,i , i = 1, 2, . . . , M.
a (46)

As a result, the disturbance between the true and estimate of

the squared distances is The {ci }, {ei }, { fi }, and {gi }, i = 1, 2, 3, have been defined in
Appendix A.2. The CWLS solution using the RSS measure-
2/a 2  2−a ments is found by the following procedure.
εi = rRSS,i − di2 ≈ di nRSS,i , i = 1, 2, . . . , M. (39)
a (i) Obtain the real roots of (46) using a root finding algo-
rithm.
In vector form, {εi } is expressed as (ii) Put the real λ’s back to (45) and obtain subestimates of

θ.
 
2  2−a 2  2−a 2  2−a T
(iii) The subestimate that yields the smallest objective value
ε= d1 nRSS,1 , d2 nRSS,2 , . . . , dM nRSS,M . of (Aθ̆ − b)T Ψ−1 (Aθ̆ − b) is taken as the globally opti-
a a a
(40) mal CWLS solution.
K. W. Cheung et al. 7

3.3. TOA [20] It is noteworthy that (52) is similar to the Taylor series lin-
earization based on a geometrical viewpoint [17], although
Since the models of the TOA and RSS will have the same form the latter considers only one AOA measurement with the cor-
if the propagation constant is equal to unity, putting a = 1 in responding BS locates at the origin. By expanding sin(φi +
Section 3.2 yields the algorithm of the CWLS estimator using nAOA,i ) and cos(φi +nAOA,i ), and considering sufficiently small
the TOA data. angle errors such that sin(nAOA,i ) ≈ nAOA,i and cos(nAOA,i ) ≈
1, we obtain the residual error in rAOA,i as
3.4. AOA
       
In the absence of noise, (13) becomes δi = nAOA,i x − xi cos φi + y − yi sin φi ,
(53)
  i = 1, 2, . . . , M.
  sin rAOA,i
tan rAOA,i =  
cos rAOA,i
(47) In vector form, {δi } is expressed as
y − yi
= , i = 1, 2, . . . , M. ⎡         ⎤
x − xi nAOA,1 x − x1 cos φ1 + y − y1 sin φ1
⎢         ⎥
⎢ n ⎥
By cross-multiplying and rearranging (47), a set of linear ⎢ AOA,2 x − x2 cos φ2 + y − y2 sin φ2 ⎥
⎢ ⎥
equations in x and y for the AOA measurements is obtained δ=⎢
⎢ . ⎥.
⎥
as ⎢ .
. ⎥
⎣ ⎦
           
x sin rAOA,i − y cos rAOA,i nAOA,M x − xM cos φM + y − yM sin φM
(54)
   
= xi sin rAOA,i − yi cos rAOA,i , i = 1, 2, . . . , M.
(48) Thus the inverse of the optimum weighting matrix, Ωo , is

Expressing (48) in matrix form, we have [16]  

Ωo = E δδ T = s3 sT3  Cn,AOA , (55)
Hx = k, (49)
where
where
⎡   ⎤ ⎡         ⎤ ⎡ ⎤
− cos rAOA,1
sin rAOA,1 x − x1 cos φ1 + y − y1 sin φ1 d1
⎢ ⎥ ⎢         ⎥ ⎢ ⎥
⎢ ⎥ ⎢ x − x2 cos φ2 + y − y2 sin φ2 ⎥ ⎢ d2 ⎥
H=⎢ .. .. ⎥, ⎢ ⎥ ⎢ ⎥
⎢ . . ⎥ s3 = ⎢ ⎥=⎢ ⎥
⎣  ⎢ ⎢
⎥ ⎢ . ⎥
⎦
.. ⎥
  ⎢ ⎥ .
sin rAOA,M − cos rAOA,M ⎣ . ⎦ ⎢
⎣ . ⎦
⎥
       
(50) x − xM cos φM + y − yM sin φM dM
⎡     ⎤
x1 sin rAOA,1 − y1 cos rAOA,1 (56)
⎢ ⎥
⎢ .. ⎥
k=⎢
⎢ .
⎥.
⎥
⎣ because cos(φi ) = (x − xi )/di and sin(φi ) = (y − yi )/di . Again,
   ⎦
xM sin rAOA,M − yM cos rAOA,M since s3 involves the unknown parameters x and {φi }, they
will be approximated as x and {rAOA,i }, respectively, in the
To improve the performance of the LS estimator of [16], we actual implementation. In summary, the WLS procedure for
propose to use WLS to estimate the MS location x and the AOA-based location is
solution is
(i) set Ω = IM ;
x = arg min(Hx̆ − k)T Ω−1 (Hx̆ − k) (ii) use (51) to determine the estimate of x;
x̆
(iii) construct Ω based on (55) using the computed x in
(51)
 −1 step (ii) and repeat step (ii) until parameter conver-
−1 −1
= H Ω H T
H Ω k,
T
gence.
where Ω−1 is the corresponding weighting matrix and x̆ = It is noteworthy that since H also consists of noise, we
[x̆, y̆]T . Again, we use the BLUE technique to determine the have already attempted to introduce constraints in the WLS
optimum Ω as follows. In the presence of measurement er- solution in order to remove the bias due to the noisy com-
rors, (48) becomes ponents, but improvement over the WLS estimator has not
    been observed. As a result, it is believed that the noise in
x sin φi + nAOA,i − y cos φi + nAOA,i H can be ignored for sufficiently high signal-to-noise ratio
    (SNR) conditions. In fact, Pages-Zamora et al. [16] have sim-
= xi sin φi + nAOA,i − yi cos φi + nAOA,i , i = 1, 2, . . . , M. ilarly observed that the LS estimator performs even better
(52) than its total least squares counterpart.
8 EURASIP Journal on Applied Signal Processing

3.5. TDOA-AOA hybrid For TDOA-TOA or TDOA-RSS hybrid positioning, a

simple and effective way is to convert the TOA and RSS,
It is apparent that combining different types of the mea- respectively, into TDOA measurements and then apply the
surements, if available, can improve location performance CWLS TDOA-based location algorithm. Finally, it is straight-
and/or reduce the number of receiving BSs. Among various forward to combine TOA and RSS measurements via con-
hybrid schemes, the most popular one is to use the TDOA verting the former to the latter or vice versa. Localization
and AOA measurements simultaneously [17]. To perform with more than two types of measurements can be extended
TDOA-AOA mobile positioning, (48) is now rewritten by easily in a similar manner.
adding y1 cos(rAOA,i ) − x1 sin(rAOA,i ) on both sides:
        4. PERFORMANCE ANALYSIS
x − x1 sin rAOA,i − y − y1 cos rAOA,i
        As briefly mentioned in Section 1, the CWLS and WLS es-
= xi − x1 sin rAOA,i − yi − y1 cos rAOA,i ,
timators in Section 3 can achieve zero bias and the CRLB
i = 1, 2, . . . , M. approximately when the noise is uncorrelated and small in
power. In the following subsections we provide the proofs of
(57)
this desirable property for each measurement case.

Combining (19) and (57) into a single matrix-vector form

yields 4.1. Mean and variance analysis for generic
unconstrained minimization problems
Bϑ = w, (58)
The idea behind the performance analysis here is to recast the
CWLS estimators to unconstrained minimization problems,
where and then to use the analysis technique for unconstrained
⎡ ⎤ ⎡ ⎤ problems [28] to find out the mean and covariance of the
G h estimators. To describe the latter, consider a generic uncon-
B=⎣ ⎦, w = ⎣ ⎦ , strained estimation problem as follows:
H 0M k
⎡ ⎤ y = arg min J(y̆), (63)
0 y̆
⎢   ⎥
⎢ x − x  sin r     ⎥
⎢ 2 1 AOA,2 − y2 − y1 cos rAOA,2 ⎥
 ⎢ ⎥ where J(y̆) is a function continuous in y̆. Given that y is the
k =⎢
⎢ . ⎥
⎥
⎢ .
. ⎥ true value of the estimated parameter, it is shown [28] that
⎣ ⎦
       
xM − x1 sin rAOA,M − yM − y1 cos rAOA,M  −1  
∂2 J ∂J  ,
(59) bias(y) ≈ −E E (64)
∂y̆∂y̆T ∂y̆ y̆=y
with 0M is an M × 1 column vector with all zeros. Then ϑ is  −1   T   −1 
∂2 J ∂J ∂J ∂2 J 
solved by minimizing Cy ≈ E E E  ,
∂y̆∂y̆T ∂y̆ ∂y̆ ∂y̆∂y̆T 
y̆=y

(Bϑ̆ − w)T W−1 (Bϑ̆ − w) (60) (65)

subject to where bias(y) and Cy represent the bias and the covariance
matrix associated with y, respectively. The approximations
T
in (64) and (65) are based on the assumption that noise
ϑ̆ Σϑ̆ = 0. (61) variances are sufficiently small. In the following, we will ap-
ply (64) and (65) to show that all the developed algorithms
The optimum weighting matrix, denoted by Wo −1 , is deter- are approximately unbiased and to produce their theoretical
mined from the inverse of variances.

Wo = s4 sT4  Cn,TDOA-AOA , (62) 4.2. TDOA

Although the CWLS problem of (24) subject to (23) consists

where s4 = [s1 s3 ]T and Cn,TDOA-AOA is the covariance ma- of a parameter vector ϑ̆ with 3 variables, namely, x̆ −x1 , y̆ − y1 ,
trix of the TDOA and AOA measurement errors. By follow- and R̆1 , it can be reduced to a 2-variable optimization prob-
T
ing the estimation procedure in Section 3.1, the parameter lem using the relation of (18), that is, setting R̆1 = (ϑ̆1 ϑ̆1 )1/2
vector ϑ is determined. Similarly, mobile location algorithms where ϑ̆1 = [x̆ − x1 y̆ − y1 ]T . In so doing, the CWLS po-
using AOA and RSS or TOA measurements can be deduced. sition estimate using the TDOA measurements is equivalent
K. W. Cheung et al. 9

to where 1M −1 is denoted as an (M − 1) × 1 column vector with

  all ones. Equation (69) shows that the estimator is approx-
ϑ1 = arg min JTDOA ϑ̆1 , (66) imately unbiased, while the two diagonal elements in (70)
ϑ̆1
correspond to the variance of the position estimate x. Now
where we are going to compute Cx particularly when all the mea-
surement errors are uncorrelated. This implies that the co-
    T 1/2 T
JTDOA ϑ̆1 = Sϑ̆1 + ϑ̆1 ϑ̆1 rTDOA − h variance matrix for the TDOA measurement errors has the
form of
(67)
  T 1/2 
−1
×Υ Sϑ̆1 + ϑ̆1 ϑ̆1 rTDOA −h ⎡ ⎤
2
σTDOA,2 0 ··· 0
⎢ ⎥
which is the cost function of the CWLS algorithm using ⎢ 2
··· ⎥
⎢ 0 σTDOA,3 0 ⎥
TDOA measurements in terms of ϑ̆1 with Cn,TDOA
⎢
=⎢
⎥
⎥. (71)
⎢ .. .. .. .. ⎥
⎡ ⎤ ⎢ . . . . ⎥
x2 − x1 y 2 − y 1 ⎣ ⎦
⎢ ⎥ 0 0 2
· · · σTDOA,M
⎢ x3 − x1 y 3 − y 1 ⎥
⎢ ⎥
S=⎢
⎢
⎥. (68)
⎢
.
.. .. ⎥
. ⎥
⎣ ⎦
Considering sufficiently small error conditions such that Υ ≈
xM − x1 y M − y 1 Υo , we have
T
The values of E[∂JTDOA (ϑ̆1 )/∂ϑ̆1 ], E[∂2 JTDOA (ϑ̆1 )/∂ϑ̆1 ∂ϑ̆1 ],
and E[(∂JTDOA (ϑ̆1 )/∂ϑ̆1 )(∂JTDOA (ϑ̆1 )/∂ϑ̆1 )T ] at ϑ̆1 = ϑ1 are Υ ≈ s1 sT1  Cn,TDOA
calculated in Appendix B.1. Using (64) and (65) with J = ⎡ ⎤
JTDOA (ϑ̆1 ), the mean and the covariance matrix of the MS po- d22 σTDOA,2
2
0 ··· 0
⎢ ⎥
sition estimated by the CWLS algorithm are ⎢ d32 σTDOA,3
2
··· ⎥
⎢ 0 0 ⎥ (72)
⎢ ⎥
=⎢ .. .. .. ⎥.
E[x] ≈ x, (69) ⎢ .. ⎥
⎢ . . . . ⎥
⎣ ⎦
    2 2
Cx ≈ ST + d1−1 x − x1 sT1 − d1 1TM −1 , 0 0 · · · dM σTDOA,M
 (70)
  T −1
× Υ−1 S + d1−1 s1 − d1 1M −1 x − x1 ,
We also note that

   d − d1 ⎡    dM − d1 ⎤
⎢ x2 − x1 + x − x1 2 · · · xM − x1 + x − x1
 T    ⎢ d1 d1 ⎥ ⎥
S + d1−1 x − x1 sT1 − d1 1TM −1 = ⎢ ⎥ (73)
⎣    d2 − d1     dM − d1 ⎦
y2 − y1 + y − y1 · · · yM − y1 + y − y1
d1 d1

and [S + d1−1 (s1 − d1 1M −1 )(x − x1 )T ] is given by the transpose Substituting (72) and (73) into (70), the inverse of co-
of (73). variance matrix Cx is calculated as

⎡  2   ⎤

M
1 x − xi x − x1 
M
1 x − xi x − x1 y − yi y − y1
⎢ − − − ⎥
⎢ 2
σTDOA,i di d1 2
σTDOA,i di d1 di d1 ⎥
⎢ i=2 i=2 ⎥
Cx−1
≈⎢
⎢M
⎥.
⎥ (74)
⎢    
M  2 ⎥
⎣ 1 x − xi x − x1 y − yi y − y1 1 y − yi y − y1 ⎦
2 − − 2 −
i=2
σTDOA,i di d1 di d1 i=2
σTDOA,i di d1
10 EURASIP Journal on Applied Signal Processing

On the other hand, the Fisher information matrix (FIM) for measurement errors is computed in Appendix C as shown
the TDOA-based mobile location problem with uncorrelated below

⎡  2   ⎤

M
1 x − xi x − x1 
M
1 x − xi x − x1 y − yi y − y1
⎢ − − − ⎥
⎢ 2
σTDOA,i di d1 2
σTDOA,i di d1 di d1 ⎥
⎢ i=2 i=2 ⎥
⎢ ⎥
ITDOA (x) = ⎢
⎢M
⎥
⎥
(75)
⎢    
M  2 ⎥
⎣ 1 x − xi x − x1 y − yi y − y1 1 y − yi y − y1 ⎦
2 − − 2 −
i=2 σTDOA,i di d1 di d1 i=2 σTDOA,i di d1

which implies Cx−1 ≈ ITDOA (x). As a result, the performance Again, the unbiasedness of the algorithm is illustrated in
of the TDOA-based mobile positioning algorithm via the use (79). For uncorrelated measurement errors, we have
of CWLS achieves the CRLB for uncorrelated measurement
errors. It is also expected that the optimality still holds when
the TDOA measurement errors are correlated. ⎡ ⎤
2
σRSS,1 0 ··· 0
⎢ ⎥
⎢ 0 2
··· ⎥
4.3. RSS ⎢ σRSS,2 0 ⎥
⎢ ⎥
Cn,RSS =⎢ . .. .. ⎥. (81)
Similar to Section 4.1, R̆2 in θ̆ is substituted by xT x so the ⎢ . .. ⎥
⎢ . . . . ⎥
⎣ ⎦
CWLS solution using the RSS measurements is equivalent to 2
0 0 · · · σRSS,M

x = arg min JRSS (x̆), (76)

x̆
Assuming ideal weighting matrix as in the previous analysis,
the inverse of Ψ−1 for the RSS-based algorithm is
where

   T
JRSS (x̆) = XBS x̆ − 0.5 x̆T x̆ 1M − b Ψ ≈ s2 sT2  Cn,RSS
(77) ⎡ ⎤
    1 2(2−a) 2
× Ψ−1 XBS x̆ − 0.5 x̆T x̆ 1M − b ⎢ a2 d1 σRSS,1 0 ··· 0 ⎥
⎢ ⎥
⎢ 1 ⎥
⎢
⎢ 0 d2(2−a) σRSS,2
2 2
2
··· 0 ⎥
⎥
⎢
=⎢
a ⎥.
with .. .. .. .. ⎥
⎢ . ⎥
⎢ . . . ⎥
⎢ ⎥
⎡ ⎤ ⎣ 1 2(2−a) 2 ⎦
x1 y1 0 0 ··· d σRSS,M
⎢ ⎥ a2 M
⎢x y2 ⎥ (82)
⎢ 2 ⎥
⎢ ⎥
XBS =⎢ . .. ⎥. (78)
⎢ . ⎥
⎢ . . ⎥
⎣ ⎦
xM y M It is also noted that

The required values of the derivatives have been computed ⎡ ⎤

in Appendix B.2. Putting them into (64) and (65) with J = x1 − x x2 − x · · · xM − x
JRSS (x̆) gives
T
XBS − x1TM =⎣ ⎦ (83)
y1 − y y2 − y · · · yM − y

E[x] ≈ x, (79)

   −1 and (XBS − 1M xT ) is the transpose of (83). Hence the inverse

−1
Cx ≈ T
XBS − x1TM Ψ XBS − 1M x T
. (80) of the covariance matrix is
K. W. Cheung et al. 11

M 2 2 ⎡ M 2   ⎤
 a x − xi di2(a−2)  a x − xi y − yi di2(a−2)
⎢ ⎥
⎢ 2
σRSS,i 2
σRSS,i ⎥
 T    ⎢ i=1 i=1 ⎥
Cx−1 ≈ XBS − x1TM Ψ−1 XBS − 1M xT = ⎢
⎢M     
⎥.
⎥ (84)
⎢ a2 x − x y − y d 2(a −2) M
a y − yi di
2 2 2(a −2) ⎥
⎣ i i i ⎦
2 2
i=1
σRSS,i i=1
σRSS,i

From Appendix C, the FIM for RSS-based mobile location Considering sufficiently small noise conditions, we have
with uncorrelated measurement errors can be computed,
which is given by
Ω ≈ s3 sT3  Cn,AOA
IRSS (x)
⎡ ⎤ ⎡ ⎤
M 2 2 M 2   d12 σAOA,1
2
···
 a x − xi di2(a−2)  a x − xi y − yi di2(a−2) 0 0
⎢ ⎥ ⎢ ⎥
⎢ 2
σRSS,i 2
σRSS,i ⎥ ⎢ 0 d22 σAOA,2
2
··· 0 ⎥
⎢ i=1 i=1 ⎥ ⎢ ⎥
⎢
=⎢ ⎥ ⎢ ⎥
⎥ =⎢ ⎥,
⎢M 2   2(a−2)  
 a2 y − yi di
M 2 2(a −2) ⎥ ⎢ .. .. .. .. ⎥
⎣ a x − xi y − yi di ⎦ ⎢
⎣ . . . . ⎥
⎦
2 2
i=1 σRSS,i i=1 σRSS,i 0 0 2 2
· · · dM σAOA,M
(85) (90)
⎡ ⎤
y − y1 x − x1
which means IRSS (x) ≈ Cx−1 , and thus the optimality of ⎢ d −
⎢ 1 d1 ⎥ ⎥
the RSS-based location algorithm for white disturbance is ⎢ ⎥
⎢ y−y ⎥
proved. ⎢ 2 x − x2 ⎥
⎢ − ⎥
H≈⎢
⎢ d 2 d2 ⎥.
⎥
⎢ .. .. ⎥
4.4. TOA ⎢ . . ⎥
⎢ ⎥
⎢ ⎥
⎣ y − yM x − xM ⎦
By putting a = 1 in Section 4.2, the bias and variance ex- −
pressions for the position estimate using the TOA data are dM dM
obtained. Nevertheless, we have already shown that its esti-
mation performance attains the CRLB in uncorrelated mea-
surement errors in [20]. Putting (90) into (88) yields

4.5. AOA
Cx−1 ≈ HT Ω−1 H
From Section 3.4, the WLS cost function for AOA-based mo- ⎡ M  2 M   ⎤
bile positioning is  y − yi  x − xi y − y i
⎢ − ⎥
⎢ σ 2 d4 2
σAOA,i di4 ⎥
⎢ i=1 AOA,i i i=1 ⎥
⎢ ⎥
JAOA (x̆) = (Hx̆ − k)T Ω−1 (Hx̆ − k). (86) ≈⎢ ⎥.
⎢ M      ⎥
⎢  x−x y− y M
x − xi
2 ⎥
⎣ i i ⎦
In Appendix B.3, the mean and the covariance matrix of the − 2 4 2 4
i=1 σAOA,i di i=1 σAOA,i di
MS position estimate are calculated as
(91)
E[x] ≈ x, (87)

 −1 On the other hand, the FIM for AOA-based mobile loca-
Cx ≈ HT Ω−1 H . (88) tion with uncorrelated measurement errors is computed in
Appendix C as
In particular, for uncorrelated measurement errors, Cn,AOA is
of the form
⎡ M  2 M   ⎤
⎡ ⎤  y − yi  x − xi y − y i
2
σAOA,1 0 ··· 0 ⎢ − ⎥
⎢ ⎥ ⎢ σ 2 d4 2
σAOA,i di4 ⎥
⎢ ⎥ ⎢ i=1 AOA,i i i=1 ⎥
⎢ 0 2
σAOA,2 ··· 0 ⎥ ⎢ ⎥
⎢ ⎥ IAOA (x) = ⎢
⎢ M 
⎥
⎥
Cn,AOA =⎢ ⎥. (89) ⎢  x−x y− y     ⎥
⎢ .. .. .. .. ⎥ ⎣ i i M
x − xi
2
⎦
⎢ . . . . ⎥ −
⎣ ⎦ 2 4
σAOA,i di 2 4
σAOA,i di
2 i=1 i=1
0 0 · · · σAOA,M (92)
12 EURASIP Journal on Applied Signal Processing

which implies IAOA (x) ≈ Cx−1 . As a result, the performance In particular, for uncorrelated measurement errors, we have
of using the WLS estimator for AOA-based mobile loca-
⎡ ⎤
tion with uncorrelated measurement errors is optimal under Cn,TDOA O(M −1)×M
small noise conditions. Cn,TDOA-AOA = ⎣ ⎦, (98)
OM ×(M −1) Cn,AOA
4.6. TDOA-AOA hybrid
where
Similar to Section 4.1, the CWLS position estimate using
2
 2 2

both TDOA and AOA measurements is equivalent to Cn,TDOA = diag σTDOA,2 , σTDOA,3 , . . . , σTDOA,M ,
(99)
 
  2
Cn,AOA = diag σAOA,1 2
, σAOA,2 2
, . . . , σAOA,M ,
ϑ1 = arg min JTDOA-AOA ϑ̆1 , (93)
ϑ̆1
and O(M −1)×M is denoted as an (M − 1) × M matrix with all
where zeros. Using the ideal weighting matrix, we get

  W = s4 sT4  Cn,TDOA-AOA
JTDOA-AOA ϑ̆1
⎡ ⎤
s1 sT1  Cn,TDOA O(M −1)×M
= (Bϑ̆ − w)T W−1 (Bϑ̆ − w) =⎣ ⎦
OM ×(M −1) s3 sT3  Cn,AOA (100)
⎡⎡ ⎤ ⎡ ⎤ ⎤T
S  T 1/2 rTDOA (94) ⎡ ⎤
= ⎣⎣ ⎦ ϑ̆1 + ϑ̆1 ϑ̆1 ⎣ ⎦ − w⎦ Υ O(M −1)×M
H 0M =⎣ ⎦
OM ×(M −1) Ω
⎡  ⎡ ⎤ ⎤
S  T 1/2 rTDOA
× W−1 ⎣ ϑ̆1 + ϑ̆ ϑ̆1 ⎣ ⎦ − w⎦
1 which is a diagonal matrix. Substituting (100) into (97) yields
H 0M
   
Cx−1 ≈ ST + d1−1 x − x1 sT1 − d1 1TM −1
with
   T 
× Υ−1 S + d1−1 s1 − d1 1M −1 x − x1 (101)
 
S rTDOA
B= . (95)
H 0M + HT Ω−1 H.

In Appendix B.4, we have shown that In Appendix C, the FIM for the TDOA-AOA hybrid mobile
positioning problem with uncorrelated errors can be com-
puted as
E[x] ≈ x (96)
ITDOA-AOA (x) = ITDOA (x) + IAOA (x). (102)
which indicates its unbiasedness and
From the results of (74), (75), (91), and (92), it is noted that
⎧⎡ ⎛ ⎡ ⎤ ⎡ ⎤⎞T ⎤ Cx−1 ≈ ITDOA-AOA . As a result, it is proved that the perfor-
⎪
⎨   
⎢  s 1 mance of the TDOA-AOA hybrid mobile positioning algo-
⎦⎠ ⎥
1 M −1
Cx ≈ ⎪⎣ ST HT + x − x1 ⎝d1−1 ⎣ ⎦ − ⎣ ⎦ rithm achieves the CRLB for sufficiently small uncorrelated
⎩ 0M 0M
noise conditions.
⎡⎡ ⎤ ⎛ ⎡ ⎤ ⎡ ⎤⎞ ⎤⎫−1
S s1 1M −1  T ⎬ 5. SIMULATION RESULTS
×W −1 ⎣⎣ ⎦ + ⎝ d −1 ⎣ ⎦−⎣ ⎦⎠ x − x1 ⎦
H
1
0M 0M ⎭
Computer simulation using MATLAB had been conducted
* to evaluate the performance of the proposed TOA-based,
    
= ST + d1−1 x − x1 sT1 − d1 1TM −1 HT TDOA-based, RSS-based, AOA-based, and TDOA-AOA hy-
brid mobile positioning algorithms. Comparisons with
the NLS approach as well as corresponding CRLBs were
⎡   T ⎤⎫−1 also made. We considered a 5-BS geometry with coordi-
S + d1−1 s1 − d1 1M −1 x − x1 ⎬ √ √
× W−1 ⎣ ⎦ . nates [0, 0] m, [3000√3, 3000] m, [0, 6000] m, [−3000 3,
⎭
H 3000] m, and [−3000 3, −3000] m, while the MS position
(97) was fixed at [x, y] = [1000, 2000] m. The value of a was set
K. W. Cheung et al. 13

No. of BS = 5, MS at [1000, 2000] m methods when the TOA noise power was less than 35 dBm2 ,
120
but its effect became negligible for larger power of nTOA,i , par-
ticularly for the CWLS estimator.
100 Figures 2, 3, and 4 show the MSREs of the RSS-based,
Mean square range error (dBm2 )

TDOA-based, and AOA-based positioning algorithms, re-

80 spectively, as well as the corresponding CRLBs, versus power
of measurement errors. The disturbances in the RSS and
60
AOA measurements were white Gaussian processes with
identical variances as in the TOA measurements. As the units
2
of the σRSS,i 2
and σAOA,i were m2a and rad2 , they became dBm2a
40
and dBrad2 when represented in dB scales. While the TDOA
measurements were Gaussian with covariance matrix of the
20 form
⎡ ⎤
2 1 ··· 1
0 ⎢ ⎥
10 20 30 40 50 60 70 80 90 100 ⎢1 2 ··· 1⎥
σ2 ⎢ ⎥
TOA noise power (dBm2 ) Cn,TDOA = TDOA ⎢. .. ⎥
⎢. .. . . ⎥. (104)
2 ⎢. . . .⎥
Proposed NLS with NOLS ⎣ ⎦
NLS CRLB 1 1 ··· 2
Proposed with NLOS
From the figures, we observe that the performance of all the
proposed methods approached the corresponding CRLBs for
Figure 1: Mean square range errors for TOA measurements in un-
sufficiently small measurement errors, which verified their
correlated noise.
optimality at sufficiently high SNRs. Moreover, the superi-
ority of the CWLS approach over the NLS scheme was again
demonstrated for larger disturbance environments.
to be 2 in all RSS measurements. For the proposed approach,
Figure 5 shows the MSREs with TDOA-AOA hybrid mea-
steps (ii) and (iii) of the TDOA-based and TDOA-AOA hy-
surements, where the disturbances in the same type of mea-
brid algorithms and step (ii) of the AOA-based algorithm
surements had identical power with zero mean, and they
were only repeated once because no obvious improvement
were uncorrelated with each other. It can be observed that
was observed for more iterations. On the other hand, we used
the variances of the CWLS estimator approached the corre-
the Newton-Raphson iterative procedure in the NLS imple-
sponding CRLB for all cases while the NLS scheme failed to
mentation with three iterations. For TDOA, TOA, and RSS
produce optimum performance particularly when the AOA
measurements, NLS initialization was given by (28) and (45)
noise power was −10 dBrad2 . This illustrated that the CWLS
with setting the values of the Lagrange multipliers to zero. As
estimator for TDOA-AOA hybrid mobile positioning was op-
for AOA measurements, (51) was employed to initialize the
timum for uncorrelated TDOA and AOA measurements and
NLS estimator with Ω = IM . All results were averages of 1000
was more robust than the NLS method.
independent runs.
The computational complexity of the CWLS and NLS
Figure 1 shows the mean square range errors (MSREs) of
methods was also compared using the average number of
the TOA-based CWLS and NLS estimators as well as CRLB
floating point operations (FLOPS) provided by MATLAB,
versus power of distance error based on the TOA measure-
and the results are given in Table 2. It is seen that for AOA
ments. For simplicity, we assumed that the disturbances in
measurements, the proposed method required fewer FLOPS
the TOA measurements, namely, {nTOA,i }, were white Gaus-
than the NLS while it needed more FLOPS for RSS and TOA
sian processes with identical variances. The MSRE was de-
measurements. For TDOA and TDOA-AOA hybrid measure-
fined as E[(x − x)2 + (y − y)2 ] and its unit was m2 , which be-
ments, both methods had comparable complexity. It is note-
came dBm2 in dB scale. We observe that the performance of
worthy to mention that the computational requirements of
the proposed and NLS methods met the CRLB when the TOA
the CWLS approach can be significantly reduced if we only
noise power was less than 75 dBm2 and 60 dBm2 , respectively,
solve for the Lagrange multiplier whose value is closest to
which indicated that the former had a larger optimum oper-
zero as in the LCLS method [15].
ation range. The effect of positive mean TOA errors, which
corresponded to NLOS propagation, was also illustrated in
the same figure. Here the range measurements were modeled 6. CONCLUSIONS
as
This paper considers a unified constrained weighted least
rTOA,i = di + nTOA,i + Nui , (103) squares (CWLS)/weighted least squares (WLS) mobile lo-
cation approach for time-of-arrival (TOA), received sig-
where N = 100 m was the maximum error introduced by nal strength (RSS), time-difference-of-arrival (TDOA), and
NLOS and ui , i = 1, 2, . . . , M, were independent uniformly angle-of-arrival (AOA) measurements. The basic idea is to
distributed random numbers ranged from 0 to 1. It is seen reorganize the nonlinear equations obtained from the mea-
that the nonzero mean errors introduced biases in both surements into linear equations. These linear equations are
14 EURASIP Journal on Applied Signal Processing

No. of BS = 5, MS at [1000, 2000] m No. of BS = 5, MS at [1000, 2000] m

100 150
90
Mean square range error (dBm2 )

Mean square range error (dBm2 )

80
70
100
60
50
40
50
30
20
10
0 0
90 100 110 120 130 140 150 160 170 180 −70 −60 −50 −40 −30 −20 −10 0 10 20
RSS noise power (dBm2a ) AOA noise power (dBrad2 )

Proposed CRLB Proposed CRLB

NLS NLS

Figure 2: Mean square range errors for RSS measurements in un- Figure 4: Mean square range errors for AOA measurements in un-
correlated noise. correlated noise.

No. of BS = 5, MS at [1000, 2000] m

No. of BS = 5, MS at [1000, 2000] m 140
140
120
120
Mean square range error (dBm2 )
Mean square range error (dBm2 )

100
100

80
80

60 AOA noise power = −10 dBrad2

60

40 40 AOA noise power = −40 dBrad2

20 20
AOA noise power = −70 dBrad2
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70
TDOA noise power (dBm2 ) TDOA noise power (dBm2 )

Proposed CRLB Proposed CRLB

NLS
NLS

Figure 3: Mean square range errors for TDOA measurements in Figure 5: Mean square range errors for using both TDOA and AOA
correlated noise. measurements.

then solved in an optimum manner with the use of weighted bound (CRLB) approximately. Simulation results indicate
least squares and/or method of Lagrange multipliers. The that these theoretical approximation results are accurate, in
proposed approach is quite flexible in that it can be easily that the simulated mean square error performance of the de-
extended to hybrid measurement cases such as the TDOA- veloped algorithms closely approaches the CRLBs when the
AOA. We have proved that for small uncorrelated noise dis- noise variance is small. It is also shown that the proposed
turbances, the performance of all the proposed CWLS and approach outperforms the nonlinear least squares scheme in
WLS algorithms attains zero bias and the Cramér-Rao lower terms of larger optimum operation range.
K. W. Cheung et al. 15

Table 2: Computational complexity of proposed and NLS methods The solution to (A.6) is
in terms of FLOPS.

Proposed NLS  
 −1 λ
θ = AT Ψ−1 A + λP AT Ψ−1 b − q , (A.7)
TOA 7125 1978 2
RSS 6991 1393
TDOA 9892 8058 where λ is not determined yet. To find λ, we substitute (A.7)
AOA 1075 2667 into the equality constraint of (36):
TDOA-AOA 11464 11994
   
 −1 λ λ
qT AT Ψ−1 A + λP AT Ψ−1 b − q + bT Ψ−1 A − qT
2 2
APPENDICES  −1  −1
× AT Ψ−1 A + λP P AT Ψ−1 A + λP
A.  
λ
A.1. TDOA × AT Ψ−1 b − q = 0.
2
(A.8)
Following [15], we differentiate (27) and equate the expres-
sion to zero:
Note that the matrix (AT Ψ−1 A)−1 P can be diagonalized as
∂LTDOA (ϑ̆, η)  
= 2 GT Υ−1 G + ηΣ ϑ̆ − 2GT Υ−1 h = 0.
∂ϑ̆  −1
(A.1) AT Ψ−1 A P = UΛU−1 , (A.9)

The solution to (A.1) is

where Λ = diag(γ1 , γ2 , γ3 ), and γi , i = 1, 2, 3, are the eigen-
 −1 values of the matrix (AT Ψ−1 A)−1 P. Substituting (A.9) into
ϑ = GT Υ−1 G + ηΣ GT Υ−1 h, (A.2)
(AT Ψ−1 A + λP)−1 gives
where η is not yet determined. The Lagrange multiplier is
then found by substituting (A.2) into the constraint (23):  −1  −1  −1
AT Ψ−1 A + λP = U I3 + λΛ U−1 AT Ψ−1 A .
 −1  −1 (A.10)
hT Υ−1 G GT Υ−1 G + ηΣ Σ GT Υ−1 G + ηΣ GT Υh = 0.
(A.3)
Putting (A.10) into (A.8), we get
Using eigenvalue factorization, the matrix GT Υ−1 GΣ can be
diagonalized as  −1 λ  −1
cT I3 + λΛ f − cT I3 + λΛ g
−1 −1 2
G Υ GΣ = SDS ,
T
(A.4)
 −1  −1
+ eT I3 + λΛ Λ I3 + λΛ f
where D = diag(ζ1 , ζ2 , ζ3 ) and ζi , i = 1, 2, 3, are the eigenval-
ues of the matrix GT Υ−1 GΣ. Substituting (A.4) into (A.3), λ T −1  −1
− e I3 + λΛ Λ I3 + λΛ g (A.11)
the constraint can be rewritten as 2
λ T −1  −1
 −2 − c I3 + λΛ Λ I3 + λΛ f
αT D + ηI3 β = 0, (A.5) 2
λ2 T  −1 
where α = ST ΣGT Υ−1 h = [α1 , α2 , α3 ]T and β = S−1 GT Υ−1 h = + c I3 + λΛ Λ I3 + λΛ)−1 g = 0,
4
[β1 , β2 , β3 ]T . Simplifying (A.5) gives (29).

A.2. RSS where

The minimum of (44) is obtained by differentiating LRSS (θ̆,  

cT = qT U = c1 , c2 , c3 ,
λ) with respect to θ̆ and then equating the resultant expres-
sions to zero:  −1  T
g = U−1 AT Ψ−1 A q = g 1 , g2 , g3 ,
  (A.12)
∂LRSS (θ̆, λ)   eT = bT Ψ−1 AU = e1 , e2 , e3 ,
= 2 AT Ψ−1 A + λP θ̆ − 2AT Ψ−1 b + λq = 0.
∂θ̆  −1  T
(A.6) f = U−1 AT Ψ−1 A AT Ψ−1 b = f1 , f2 , f3 .
16 EURASIP Journal on Applied Signal Processing

Since the matrix (AT Ψ−1 A)−1 P is of rank 2, one of its eigen- Then differentiating (B.1) with respect to x̆, one of the vari-
values, say, γ3 , must be zero. After expanding (A.11) and ables in ϑ̆1 , by using product rule [29], we get
putting γ3 = 0, (A.11) can be simplified to (46).
    T −1/2 
∂ ∂JTDOA
B. = 2 ST + ϑ̆1 ϑ̆1 ϑ̆1 rTTDOA
∂x̆ ∂ϑ̆1
For notation convenience, JTDOA (ϑ̆1 ), JRSS (x̆), JAOA (x̆), and ∂  T 1/2 
× Υ−1 Sϑ̆1 + ϑ̆1 ϑ̆1 rTDOA − h
JTDOA-AOA (ϑ̆1 ) are written as JTDOA , JRSS , JAOA , and JTDOA-AOA , ∂x̆
respectively. ∂  T  T −1/2 T 
+2 S + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA
∂x̆
B.1. TDOA    T 1/2 
× Υ−1 Sϑ̆1 + ϑ̆1 ϑ̆1 rTDOA − h
Differentiate (67) with respect to ϑ̆1 :
  T −1/2 
∂JTDOA T  −1/2  = 2 ST + ϑ̆1 ϑ̆1 ϑ̆1 rTTDOA
= 2 ST + ϑ̆1 ϑ̆1 ϑ̆1 rTTDOA ⎡ ⎡ ⎤ ⎤
∂ϑ̆1  −1/2 
(B.1) 
  T 1/2  ×Υ −1 ⎣S ⎣1⎦ + rTDOA ϑ̆T ϑ̆1 x̆ − x1 ⎦
× Υ−1 Sϑ̆1 + ϑ̆1 ϑ̆1 rTDOA − h . 0
1

If the derivative of JTDOA is located at the true source posi- ∂  T  T −1/2 T 

+2 S + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA
tion ϑ1 , assuming that the disturbance to the TDOA mea- ∂x̆
surements is relatively small so that {n2TDOA,i } can be ignored,    T 1/2 
then (B.1) becomes × Υ−1 Sϑ̆1 + ϑ̆1 ϑ̆1 rTDOA − h .
(B.6)

∂JTDOA
  T   
 ≈ 2 ST +d1−1 ϑ1 s1 − d1 1M −1 Υ−1
s1  nTDOA
∂ϑ̆1  ϑ̆1 =ϑ1 By substituting the true source location ϑ1 into (B.6) and ig-
     noring the square of the measurement errors {n2TDOA,i }, we
= 2 ST + ϑ1 d1−1 sT1 − 1TM −1 Υ −1
s1  nTDOA . obtain
(B.2)
 
∂ ∂JTDOA  
Taking the expected value on both sides of (B.2) and then 
∂x̆ ∂ϑ̆1 ϑ̆1 =ϑ1
applying the fact that E[nTDOA ] = 0M −1 gives
 
  = 2 ST + d1−1 ϑ1 rTTDOA
∂J 
E TDOA 
 ⎡ ⎡ ⎤ ⎤
∂ϑ̆1 ϑ̆1 =ϑ1 1  
      × Υ−1 ⎣S ⎣ ⎦ + d1−1 rTDOA x − x1 ⎦
≈ 2 ST + ϑ1 d1−1 sT1 − 1TM −1 Υ−1 s1  E nTDOA 0
= 02 . ∂  T  T −1/2 T 

(B.3) +2 S + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA 
∂x̆ ϑ̆1 =ϑ1
 
Substituting (B.3) into (64) yields × Υ−1 s1  nTDOA
  
  ≈ 2 ST + d1−1 ϑ1 sT1 − d1 1TM −1
E ϑ1 ≈ ϑ1 (B.4) ⎡ ⎡ ⎤ ⎤
−1 ⎣ ⎣1⎦
  
which indicates that the estimator is unbiased for sufficiently ×Υ S + d1 s1 − d1 1M −1 x − x1 ⎦
−1

small measurement errors. 0

Multiplying (B.2) by its transpose and then taking the ex-     
+ 2 ST +d1−1 ϑ1 sT1 − d1 1TM −1 d1−1 Υ−1 nTDOA x − x1
pected value yields
  T  + 2d1−1 ϑ1 nTTDOA Υ−1

∂JTDOA ∂JTDOA  ⎡ ⎡ ⎤ ⎤
E 
∂ϑ̆1 ∂ϑ̆1  1⎦   
ϑ̆1 =ϑ1 ⎣
× S ⎣ + d1 s1 − d1 1M −1 x − x1 ⎦
−1
   0
≈ 4 ST + ϑ1 d1−1 sT1 − 1TM −1
     T ∂  T  T −1/2 T 

× Υ−1 s1 sT1  Cn,TDOA Υ−1 S + d1−1 s1 − 1M −1 ϑ1 +2 S + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA 
∂x̆ ϑ̆1 =ϑ1
      T  
= 4 ST +ϑ1 d1−1 sT1 − 1TM −1 Υ−1 S+ d1−1 s1 − 1M −1 ϑ1 . × Υ−1 s1  nTDOA .
(B.5) (B.7)
K. W. Cheung et al. 17

Taking the expected value on both sides of (B.7) and applying B.2. RSS
the fact that E[nTDOA ] = 0M −1 gives
Differentiate (77) with respect to x̆,
  
∂ ∂JTDOA 
E   T     
 ∂JRSS
∂x̆ ∂ϑ̆1 ϑ̆1 =ϑ1 = 2 XBS − x̆1TM Ψ−1 XBS x̆ − 0.5 x̆T x̆ 1M − b .
∂x̆
   (B.13)
≈ 2 ST + d1−1 ϑ1 sT1 − d1 1TM −1
⎡ ⎡ ⎤ ⎤
  
×Υ −1 ⎣S ⎣1⎦ + d −1 s1 − d1 1M −1 x − x1 ⎦ . Assuming that the disturbances due to the RSS measure-
1
0 ments are sufficiently small such that {n2RSS,i } can be ignored,
(B.8) the derivative of JRSS evaluated at the true MS position x be-
comes

Similarly, repeating the derivation in (B.6), (B.7), and (B.8) 

∂JRSS 

 T   
with the variable y̆,  ≈ 2 XBS − x1TM Ψ−1 s2  nRSS . (B.14)
∂x̆ x̆=x
  
∂ ∂JTDOA 
E  Take the expected value on both sides of (B.14) and then ap-

∂ y̆ ∂ϑ̆1 ϑ̆1 =ϑ1 ply the fact that E[nRSS ] = 0M , we get
  
≈ 2 ST + d1−1 ϑ1 sT1 − d1 1TM −1
 
⎡ ⎡ ⎤ ⎤ ∂JRSS 

 T    
E ≈ 2 XBS − x1TM Ψ−1 s2  E nRSS = 02 .
0    ∂x̆ x̆=x
× Υ−1 ⎣S ⎣ ⎦ + d1−1 s1 − d1 1M −1 y − y1 ⎦ .
1 (B.15)
(B.9)
Substituting (B.15) into (64) yields (79).
Multiplying (B.14) by its transpose and then taking the
We also get
expected value yields
 
∂2 JTDOA 
   T 

E  ∂JRSS ∂JRSS 
T  E 
∂ϑ̆1 ∂ϑ̆1 ϑ̆1 =ϑ1 ∂x̆ ∂x̆ 
x̆=x
            
∂ ∂JTDOA   ∂ ∂JTDOA   T
≈ 4 XBS − x̆1TM Ψ−1 s2 sT2  Cn,RSS Ψ−1 XBS − 1M x̆T
= E  E  .
∂x̆ ∂ϑ̆1 ϑ̆1 =ϑ1 ∂ y̆ ∂ϑ̆1 ϑ̆1 =ϑ1  T   
(B.10) = 4 XBS − x̆1TM Ψ−1 XBS − 1M x̆T .
(B.16)
Hence substituting (B.8) and (B.9) into (B.10) yields
On the other hand, differentiating (B.13) with respect to x̆,
  the first variable in x̆, and with the use of product rule [29],
∂2 JTDOA 
    we get
E T  ≈ 2 ST + d1−1 ϑ1 sT1 − d1 1TM −1

∂ϑ̆1 ∂ϑ̆1 ϑ̆1 =ϑ1
   T  
∂ ∂JRSS
× Υ−1 S + d1−1 s1 − d1 1M −1 ϑ1 .
∂x̆ ∂x̆
(B.11)
 T  ∂   
= 2 XBS − x̆1TM Ψ−1 XBS x̆ − 0.5 x̆T x̆ 1M − b
∂x̆
Then by substituting (B.5) and (B.11) into (65), the covari-  
ance matrix for the MS position estimate ϑ1 is obtained as ∂ T     
+2 XBS − x̆1TM Ψ−1 XBS x̆ − 0.5 x̆T x̆ 1M − b
∂x̆
   ⎡ ⎡ ⎤ ⎤
Cϑ1 ≈ ST + d1−1 ϑ1 sT1 − d1 1TM −1  
⎢ ⎥ ⎢1⎥
   −1 (B.12) T
= 2 XBS − x̆1TM Ψ−1 ⎣XBS ⎣ ⎦ − 1M x̆⎦
−1 −1
×Υ S + d1 s1 − d1 1M −1 ϑT1 . 0
 
∂ T     
Substituting x − x1 back to ϑ1 in (B.4) and (B.12) and apply- +2 X − x̆1TM Ψ−1 XBS x̆ − 0.5 x̆T x̆ 1M − b .
∂x̆ BS
ing the fact that Cx = Cϑ1 gives (69) and (70). (B.17)
18 EURASIP Journal on Applied Signal Processing

Ignoring the terms of {n2RSS,i } again, the value of (B.17) com- Assuming that the disturbances due to the AOA measure-
puted at x is ments are sufficiently small such that {n2AOA,i } can be ignored,
the derivative of JAOA evaluated at the true value of x becomes
⎡ ⎡ ⎤ ⎤
  
∂ ∂JRSS 

 T  ⎢ ⎢1⎥ ⎥ ∂JAOA   
= 2 XBS − x1TM Ψ−1 ⎣XBS ⎣ ⎦ − 1M x⎦  ≈ 2HT Ω−1 s3  nAOA .
∂x̆ ∂x̆ x̆=x 0 ∂x̆ x̆=x
(B.24)

 
∂ T    
+2 XBS − x̆1TM  Ψ−1 s2  nRSS . Taking the expected value on both sides of (B.24) and then
∂x̆ x̆=x applying the fact that E[nAOA ] = 0M , we obtain
(B.18)
 
∂JAOA 

  
E ≈ 2HT Ω−1 s3  E nAOA = 02 . (B.25)
Taking the expected value on both sides of (B.18) and apply- ∂x̆ x̆=x
ing E[nRSS ] = 0M gives

⎡ ⎡ ⎤ ⎤ Substituting (B.25) into (64) gives (87).

   Multiplying (B.24) by its transpose and then taking the
∂ ∂JRSS    1
E  T
≈ 2 XBS − x1TM Ψ−1 ⎣XBS ⎣ ⎦ − 1M x⎦ .
∂x̆ ∂x̆  expected value yields
x̆=x 0
(B.19)
  T 
  
∂JAOA ∂JAOA 
E  ≈ 4HT Ω−1 s3 sT3  Cn,AOA Ω−1 H
∂x̆ ∂x̆ 
Similarly, repeating the derivations in (B.17)–(B.19) with the x̆=x
second variable y̆, we obtain
= 4HT Ω−1 H.
⎡ ⎡ ⎤ ⎤ (B.26)
  
∂ ∂JRSS    0
E  T
≈ 2 XBS − x1TM Ψ−1 ⎣XBS ⎣ ⎦ − 1M y ⎦ .
∂ y̆ ∂x̆ 
x̆=x 1 Differentiating (B.23) with respect to x̆, we get
(B.20)

∂2 JAOA
We also have = 2HT Ω−1 H. (B.27)
∂x̆∂x̆T

 
∂2 JRSS 
 Since (B.27) does not contain x and nAOA , taking the expected
E
∂x̆∂x̆T x̆=x value on both sides of (B.27) yields
       
∂ ∂JRSS  ∂ ∂JRSS   
= E  E  . ∂2 JAOA 
  
∂x̆ ∂x̆ ∂ y̆ ∂x̆ E = 2HT Ω−1 H. (B.28)
∂x̆∂x̆T x̆=x
x̆=x x̆=x
(B.21)

Substituting (B.26) and (B.28) into (65) gives (88).

Substituting (B.19) and (B.20) into (B.21) yields

  B.4. TDOA-AOA hybrid

∂2 JRSS 

 T   
E ≈ 2 XBS − x1TM Ψ−1 XBS − 1M xT .
∂x̆∂x̆T x̆=x Differentiate (94) with respect to ϑ̆1
(B.22)
∂JTDOA-AOA
Then substituting (B.16) and (B.22) into (65) gives (80). ∂ϑ̆1
   T −1/2  
B.3. AOA = 2 ST HT + ϑ̆1 ϑ̆1 ϑ̆1 rTTDOA 0TM

Differentiating (86) with respect to x̆, we get ⎡⎡ ⎤ ⎡ ⎤ ⎤

S  T 1/2 ⎢rTDOA ⎥
−1 ⎢⎢ ⎥ ⎥
×W ⎣⎣ ⎦ ϑ̆1 + ϑ̆1 ϑ̆1 ⎣ ⎦ − w⎦ .
∂JAOA H 0M
= 2HT Ω−1 (Hx̆ − k). (B.23) (B.29)
∂x̆
K. W. Cheung et al. 19

If the derivative of JTDOA-AOA is located at the true source po- Then differentiating (B.29) with respect to x̆, one of the vari-
sition ϑ1 , assuming that the disturbances are relatively small ables in ϑ̆1 , by using product rule [29], we get
so that {n2TDOA,i } and {n2AOA,i } can be ignored, then (B.29)
becomes
  
∂JTDOA-AOA 
 ∂ ∂JTDOA-AOA

∂ϑ̆1 ϑ̆1 =ϑ1 ∂x̆ ∂ϑ̆1
⎡ ⎛⎡ ⎤ ⎡ ⎤⎞T ⎤    T −1/2  
 
⎢ ⎝ ⎣ s1 ⎦ 1M −1 ⎦⎠ ⎥ = 2 ST HT + ϑ̆1 ϑ̆1 ϑ̆1 rTTDOA 0TM
≈ 2 ⎣ ST HT + d1−1 ϑ1 − d1 ⎣ ⎦
0M 0M
⎡⎡ ⎤ ⎡ ⎤ ⎤
  ∂ ⎣⎣ S ⎦  T 1/2 rTDOA
−1
× W s4  nTDOA-AOA ×W −1
ϑ̆1 + ϑ̆1 ϑ̆1 ⎣ ⎦ − w⎦
∂x̆ H 0M
⎡ ⎛ ⎡ ⎤ ⎡ ⎤⎞T ⎤
 
⎢ s 1M −1 ⎦⎠ ⎥
= 2 ⎣ ST HT + ϑ1 ⎝d1−1 ⎣ ⎦ − ⎣ ∂   T T   T −1/2  T 
1
⎦
0M 0M +2 S H + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA 0TM
∂x̆
  ⎧ ⎡⎡ ⎤ ⎡ ⎤ ⎤⎫
× W−1 s4  nTDOA-AOA . ⎨ ⎬
S  T 1/2 rTDOA
(B.30) × W−1 ⎣⎣ ⎦ ϑ̆1 + ϑ̆1 ϑ̆1 ⎣ ⎦ − w⎦
⎩ H 0M ⎭
Taking the expected value on both sides of (B.30) and then
applying the fact that E[nTDOA ] = 0M −1 and E[nAOA ] = 0M    T −1/2  
= 2 ST HT + ϑ̆1 ϑ̆1 ϑ̆1 rTTDOA 0TM
gives
  ⎡⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎤
∂JTDOA-AOA 
    
E  ×W −1 ⎣⎣ S ⎦ ⎣1⎦ + ⎣rTDOA ⎦ ϑ̆T ϑ̆1 −1/2 x̆ − x1 ⎦
∂ϑ̆1 ϑ̆1 =ϑ1 1
⎡
H 0 0TM
⎛ ⎡ ⎤ ⎡ ⎤⎞T ⎤
 
⎢ s1 1M −1 ⎦⎠ ⎥
= 2 ⎣ ST HT + ϑ1 ⎝d1−1 ⎣ ⎦ − ⎣ ⎦ ∂   T T   T −1/2  T 
0M 0M +2 S H + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA 0TM
∂x̆
  ⎧ ⎡⎡ ⎤ ⎡ ⎤ ⎤⎫
× W−1 s4  nTDOA-AOA = 02 ⎨ ⎬
 
(B.31) × W −1 ⎣⎣ S ⎦ ϑ̆1 + ϑ̆T ϑ̆1 1/2 ⎣rTDOA ⎦ − w⎦ .
⎩ H
1
0M ⎭
which results in (96) and indicates that the estimator is un- (B.33)
biased for sufficiently small measurement errors.
Multiplying (B.29) by its transpose and then taking the
expected value yields
By substituting the true source location ϑ1 into (B.33) and
    ignoring the square of the measurement errors {n2TDOA,i } and
∂JTDOA-AOA ∂JTDOA-AOA T 

E  {n2AOA,i }, we obtain
∂ϑ̆1 ∂ϑ̆1 
ϑ̆1 =ϑ1
⎡ ⎛ ⎡ ⎤ ⎡ ⎤⎞T ⎤
 
⎢ s 1M −1 ⎦⎠ ⎥  
≈ 4 ⎣ ST HT + ϑ1 ⎝d1−1 ⎣ ⎦ − ⎣
1
⎦ ∂ ∂JTDOA-AOA 

0M 0M 
∂x̆ ∂ϑ̆1 ϑ̆1 =ϑ1
 
× W−1 s4 sT4  Cn,TDOA-AOA    
= 2 ST HT + d1−1 ϑ1 rTTDOA 0TM
⎡⎡ ⎤ ⎛ ⎡ ⎤ ⎡ ⎤⎞ ⎤
⎜ s1 ⎥ ⎡⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎤
⎢ S ⎢
× W−1 ⎣⎣ ⎦ + ⎝d1−1 ⎣ ⎦ − ⎣
1M −1 ⎦⎟ T⎥
⎠ ϑ1 ⎦  
H 0M 0M ×W −1 ⎣⎣ S ⎦ ⎣1⎦ + d1−1 ⎣rTDOA ⎦ x − x1 ⎦
H 0 0M
⎡ ⎛ ⎡ ⎤ ⎡ ⎤⎞T ⎤
 
⎢ s 1 ⎥ ∂   T T   T −1/2  T  
= 4 ⎣ ST HT + ϑ1 ⎝d1−1 ⎣ 1 ⎦ − ⎣ M −1 ⎦⎠ ⎦ 
0M 0M +2 S H + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA 0TM 
ϑ̆1 =ϑ1
∂x̆
⎡⎡ ⎤ ⎛ ⎡ ⎤ ⎡ ⎤⎞ ⎤  
S s1 1M −1 ⎦⎠ T ⎦ × W−1 s4  nTDOA-AOA
× W−1 ⎣⎣ ⎦ + ⎝d1−1 ⎣ ⎦ − ⎣ ϑ1 .
H 0M 0M       
(B.32) ≈ 2 ST HT + d1−1 ϑ1 sT1 0TM − d1 1TM −1 0TM
20 EURASIP Journal on Applied Signal Processing

⎡⎡ ⎤ ⎡ ⎤ ⎛⎡ ⎤ ⎡ ⎤⎞ ⎤
Hence substituting (B.35) and (B.36) into (B.37), we get
−1 ⎣⎣ S ⎦ ⎣1⎦ −1 ⎝⎣ s1 ⎦ 1M −1 ⎦⎠  
×W +d1 − d1 ⎣ x − x1 ⎦
H 0 0M 0M
 
       ∂2 JTDOA-AOA 

+2 ST HT
−1
+ d1 ϑ1 sT1 0TM − d1 1TM −1 0TM E T 
∂ϑ̆1 ∂ϑ̆1 ϑ̆1 =ϑ1

  ⎡
× d1−1 W−1 nTDOA-AOA x − x1 + 2d1−1 ϑ1 nTTDOA W−1 ⎛ ⎡ ⎤ ⎡ ⎤⎞T ⎤
 
⎢ T s 1 ⎦⎠ ⎥
≈ 2 ⎣ S HT + ϑ1 ⎝d1−1 ⎣ ⎦ − ⎣
1 M −1
⎡⎡ ⎤ ⎡ ⎤ ⎛⎡ ⎤ ⎡ ⎤⎞ ⎤ ⎦
0M 0M
⎣⎣ S ⎦ ⎣1⎦ −1 ⎝⎣ s1 ⎦ 1M −1 ⎦⎠  
× + d1 − d1 ⎣ x − x1 ⎦
H 0 0M 0M ⎡⎡ ⎤ ⎛ ⎡ ⎤ ⎡ ⎤⎞ ⎤
×W −1 ⎣⎣ S ⎦ + ⎝d −1 ⎣ s1 ⎦ − ⎣1M −1 ⎦⎠ ϑT ⎦ .
∂   T T   T −1/2  T   1 1
 H 0M 0M
+2 S H + ϑ̆1 ϑ̆1 ϑ̆1 rTDOA 0TM 
ϑ̆1 =ϑ1
∂x̆ (B.38)
 
× W−1 s4  nTDOA-AOA .
(B.34) Then by substituting (B.32) and (B.38) into (65), the covari-
ance matrix for the MS position estimate ϑ1 is obtained as
Taking the expected value on both sides of (B.34) and apply-
ing the fact that E[nTDOA-AOA ] = 02M −1 gives ⎧⎡ ⎛ ⎡ ⎤ ⎡ ⎤⎞T ⎤
⎪
⎨  
⎢ s 1M −1 ⎦⎠ ⎥
≈ ⎣ ST HT + ϑ1 ⎝d1−1 ⎣ ⎦ − ⎣
1
   Cϑ1 ⎦
∂ ∂JTDOA-AOA  ⎪
⎩
E  0M 0M

∂x̆ ∂ϑ̆1 ϑ̆1 =ϑ1

       ⎡⎡ ⎤ ⎛ ⎡ ⎤ ⎡ ⎤⎞ ⎤⎫−1
⎪
≈ 2 ST HT + d1−1 ϑ1 sT1 0TM − d1 1TM −1 0TM S s1 1M −1 ⎦⎠ T ⎦⎬
× W−1 ⎣⎣ ⎦ + ⎝d1−1 ⎣ ⎦ − ⎣ ϑ1 ⎪ .
H 0M 0M ⎭
⎡⎡ ⎤ ⎡ ⎤ ⎛⎡ ⎤ ⎡ ⎤⎞ ⎤
−1 ⎣⎣S ⎦ ⎣1⎦ −1 ⎝⎣ s1 ⎦
  (B.39)
×W +d1 − d1 ⎣1M −1 ⎦⎠ x − x1 ⎦ .
H 0 0M 0M
(B.35)
Substituting x − x1 back to ϑ1 in (B.32) and (B.39) and ap-
plying the fact that Cx = Cϑ1 gives (96) and (97).
Similarly, repeating the derivation in (B.33), (B.34), and
(B.35) with the variable y̆ gives
C.
  
∂ ∂JTDOA-AOA 
E  The Cramér-Rao lower bound (CRLB) gives a lower bound
∂ y̆ ∂ϑ̆1 ϑ̆1 =ϑ1 on variance attainable by any unbiased estimators and thus
       it can serve as a benchmark for the mean square posi-
≈ 2 ST HT + d1−1 ϑ1 sT1 0TM − d1 1TM −1 0TM tion errors (MSPEs) of the positioning algorithms. To de-
termine it, the key step is to construct the Fisher infor-
⎡⎡ ⎤ ⎡ ⎤ ⎛⎡ ⎤ ⎡ ⎤⎞ ⎤ mation matrix (FIM) using the probability density func-
 
×W −1 ⎣⎣S ⎦ ⎣0⎦ +d1 −1 ⎝⎣ s1 ⎦ − d1 ⎣1M −1 ⎦⎠ y − y1 ⎦ . tion of the measurements parameterized by the MS posi-
H 1 0M 0M tion, and the standard procedure for obtaining the CRLB
(B.36) can be found in [21]. When the measurement errors are
Gaussian distributed, the FIM for mobile positioning us-
ing TDOA measurements, denoted by ITDOA (x), is given by
We also have [14, 15]
 
∂2 JTDOA-AOA 

E T   T  
∂ϑ̆1 ∂ϑ̆1 ϑ̆1 =ϑ1 ∂fTDOA −1 ∂fTDOA 
 ,
ITDOA (x) = Cn,TDOA  (C.1)
   
∂x̆ ∂x̆ x̆=x
   
∂ ∂JTDOA-AOA  ∂ ∂JTDOA-AOA 
= E  E  .
∂x̆ ∂ϑ̆1 ϑ̆1 =ϑ1 ∂ y̆ ∂ϑ̆1 ϑ̆1 =ϑ1
(B.37) where
K. W. Cheung et al. 21

⎡      
⎤  
x̆ − x2 x̆ − x1 y̆ − y2 y̆ − y1
⎢  2  2 −   2  2   2  2 −   2  2 ⎥
⎢ x̆ − x2 + y̆ − y2 x̆ − x1 + y̆ − y1 x̆ − x2 + y̆ − y2 x̆ − x1 + y̆ − y1 ⎥
⎢ ⎥
⎢ ⎥
∂fTDOA ⎢ .. .. ⎥
=⎢
⎢ . . ⎥.
⎥
∂x̆ ⎢         ⎥
⎢ ⎥
⎢ x̆ − x x̆ − x y̆ − yM y̆ − y1 ⎥
⎣ M
 1
  ⎦
 2  2 −  2  2  2  2 −  2  2
x̆ − xM + y̆ − yM x̆ − x1 + y̆ − y1 x̆ − xM + y̆ − yM x̆ − x1 + y̆ − y1
(C.2)

 T  
Similarly, the FIMs for RSS, AOA, and TDOA-AOA hybrid ∂fAOA −1 ∂fAOA 
 ,
IAOA (x) = Cn,AOA (C.4)
based mobile positioning, denoted by IRSS (x), IAOA (x), and ∂x̆ ∂x̆ x̆=x
ITDOA-AOA (x), respectively, are given by  T  
∂f ∂fTDOA-AOA 
 ,
 T ITDOA-AOA (x) = TDOA-AOA −1
Cn,TDOA-AOA 
∂fRSS ∂x̆ ∂x̆ x̆=x
IRSS (x) =
∂x̆ (C.5)
  (C.3)
−1 ∂fRSS 
 ,
× Cn,RSS
∂x̆ x̆=x where

⎡   2  2 a/2−1   2  2 a/2−1 ⎤

⎢ a x̆ − x1 x̆ − x1 + y̆ − y1 a y̆ − y1 x̆ − x1 + y̆ − y1 ⎥
⎢ ⎥
∂fRSS ⎢
⎢ .. .. ⎥
⎥
=⎢ . . ⎥,
∂x̆ ⎢ ⎥
⎢ ⎥
⎣   2  2 a/2−1   2  2 a/2−1 ⎦

a x̆ − xM x̆ − xM + y̆ − yM a y̆ − yM x̆ − xM + y̆ − yM

⎡     ⎤
y̆ − y1 x̆ − x1
⎢ − 2  2 ⎥ 2  2
⎢ x̆ − x1 + y̆ − y1 ⎥ x̆ − x1 + y̆ − y1
⎢ ⎥
⎢ . ⎥ .. (C.6)
∂fAOA ⎢ .. ⎥
=⎢ ⎥, .
∂x̆ ⎢ ⎥
⎢     ⎥
⎢ y̆ − y x̆ − x ⎥
⎣ M M ⎦
− 2  2  2  2
x̆ − xM + y̆ − yM x̆ − xM + y̆ − yM
⎡
∂fTDOA ⎤
⎢ ∂x̆ ⎥
∂fTDOA-AOA ⎢ ⎥
=⎢ ⎥.
∂x̆ ⎣ ∂f ⎦
AOA
∂x̆

It is noted that ITOA (x) can be computed from IRSS (x) REFERENCES
in (C.3) by putting a = 1. Then the CRLBs, namely,
CRLBTDOA (x), CRLBRSS (x), CRLBAOA (x), CRLBTDOA-AOA (x), [1] CC Docket no. 94-102, “Revision of the Commissions Rules to
Ensure Compatibility with Enhanced 911 Emergency Calling
and CRLBTOA (x) are obtained from the diagonal elements of
Systems, RM-8143,” July, 1996.
the inverses of the corresponding FIMs. [2] C. Drane, M. Macnaughtan, and C. Scott, “Positioning GSM
telephones,” IEEE Communications Magazine, vol. 36, no. 4,
ACKNOWLEDGMENTS pp. 46–54, 59, 1998.
[3] H. Koshima and J. Hoshen, “Personal locator services emerge,”
The authors thank Mr. K. W. Chan for his help in develop- IEEE Spectrum, vol. 37, no. 2, pp. 41–48, 2000.
ing the nonlinear least squares approach. This work was sup- [4] Y. Zhao, “Mobile phone location determination and its im-
ported by a grant from the Research Grants Council of the pact on intelligent transportation systems,” IEEE Transactions
Hong Kong Special Administrative Region, China (Project on Intelligent Transportation Systems, vol. 1, no. 1, pp. 55–64,
No. CityU 1119/01E). 2000.
22 EURASIP Journal on Applied Signal Processing

[5] D. Porcino, “Performance of a OTDOA-IPDL positioning re- [23] J. Vidal, M. Najar, and R. Jativa, “High resolution time-of-
ceiver for 3GPP-FDD mode,” in Proceedings of the IEE 2nd In- arrival detection for wireless positioning systems,” in Proceed-
ternational Conference on 3G Mobile Communication Technolo- ings of 56th IEEE Vehicular Technology Conference (VTC ’02),
gies (3G ’01), pp. 221–225, London, UK, March 2001. vol. 4, pp. 2283–2287, Vancouver, BC, Canada, September
[6] J. J. Caffery Jr., Wireless Location in CDMA Cellular Radio Sys- 2002.
tems, Kluwer Academic, Boston, Mass, USA, 2000. [24] J. Riba and A. Urruela, “A robust multipath mitigation tech-
[7] J. C. Liberti and T. S. Rappaport, Smart Antennas for Wireless nique for time-of-arrival estimation,” in Proceedings of 56th
Communications: IS-95 and Third Generation CDMA Applica- IEEE Vehicular Technology Conference (VTC ’02), vol. 4, pp.
tions, Prentice-Hall, Upper Saddle River, NJ, USA, 1999. 2263–2267, Vancouver, BC, Canada, September 2002.
[8] M. McGuire and K. N. Plataniotis, “A comparison of radi- [25] S. Al-Jazzar, J. J. Caffery Jr., and H.-R. You, “A scattering model
olocation for mobile terminals by distance measurements,” in based approach to NLOS mitigation in TOA location systems,”
Proceedings of International Conference on Wireless Communi- in Proceedings of 55th IEEE Vehicular Technology Conference
cations, pp. 1356–1359, 2000. (VTC ’02), vol. 2, pp. 861–865, Birmingham, Ala, USA, May
[9] J. J. Caffery Jr. and G. L. Stuber, “Subscriber location in CDMA 2002.
cellular networks,” IEEE Transactions on Vehicular Technology, [26] M. P. Wylie-Green and S. S. Wang, “Robust range estimation
vol. 47, no. 2, pp. 406–416, 1998. in the presence of the non-line-of-sight error,” in Proceedings
[10] M. A. Spirito, “On the accuracy of cellular mobile station lo- of 54th IEEE Vehicular Technology Conference (VTC ’01), vol. 1,
cation estimation,” IEEE Transactions on Vehicular Technology, pp. 101–105, Atlantic City, NJ, USA, September 2001.
vol. 50, no. 3, pp. 674–685, 2001. [27] N. Patwari, A. O. Hero III, M. Perkins, N. S. Correal, and R.
[11] W. H. Foy, “Position-location solutions by Taylor-series esti- J. O’Dea, “Relative location estimation in wireless sensor net-
mation,” IEEE Transactions on Aerospace and Electronic Sys- works,” IEEE Transactions on Signal Processing, vol. 51, no. 8,
tems, vol. 12, no. 2, pp. 187–194, 1976. pp. 2137–2148, 2003.
[12] D. J. Torrieri, “Statistical theory of passive location systems,” [28] K. W. Cheung and H. C. So, “A multidimensional scaling
IEEE Transactions on Aerospace and Electronic Systems, vol. 20, framework for mobile location using time-of-arrival measure-
pp. 183–197, 1984. ments,” IEEE Transactions on Signal Processing, vol. 53, no. 2,
[13] J. O. Smith and J. S. Abel, “Closed-form least-squares source pp. 460–470, 2005.
location estimation from range-difference measurements,” [29] T. K. Moon and W. C. Stirling, Mathematical Methods and
IEEE Transactions on Acoustics, Speech, and Signal Processing, Algorithms for Signal Processing, Prentice-Hall, Upper Saddle
vol. 35, no. 12, pp. 1661–1669, 1987. River, NJ, USA, 2000.
[14] Y. T. Chan and K. C. Ho, “A simple and efficient estimator for
hyperbolic location,” IEEE Transactions on Signal Processing,
vol. 42, no. 8, pp. 1905–1915, 1994. K. W. Cheung was born in Hong Kong.
[15] Y. Huang, J. Benesty, G. W. Elko, and R. M. Mersereati, “Real- He received the B.Eng. degree with first
time passive source localization: a practical linear-correction class honors in electrical and electronic en-
least-squares approach,” IEEE Transactions on Speech and Au- gineering from Imperial College of Sci-
dio Processing, vol. 9, no. 8, pp. 943–956, 2001. ence, Technology & Medicine, University of
[16] A. Pages-Zamora, J. Vidal, and D. R. Brooks, “Closed-form London, in 2001 and the M.Phil. degree
solution for positioning based on angle of arrival measure- in computer engineering and information
ments,” in Proceedings of the 13th IEEE International Sympo- technology from the City University of
sium on Personal, Indoor and Mobile Radio Communications Hong Kong in 2004. From October to
(PIMRC ’02), vol. 4, pp. 1522–1526, Lisbon, Portugal, Septem- November 2001, he was a Research Assistant
ber 2002. in the Department of Computer Engineering & Information Tech-
[17] L. Cong and W. Zhuang, “Hybrid TDOA/AOA mobile user lo- nology at the City University of Hong Kong. He is currently work-
cation for wideband CDMA cellular systems,” IEEE Transac- ing in Hong Kong Science & Technology Parks. His research inter-
tions on Wireless Communications, vol. 1, no. 3, pp. 439–447, ests are in array signal processing, and developing efficient methods
2002. in radiolocation for mobile terminals. Mr. Cheung is an Associate
[18] H. C. So and S. P. Hui, “Constrained location algorithm using Member of Institution of Electrical Engineers in UK and the Hong
TDOA measurements,” IEICE Transactions on Fundamentals of Kong Institution of Engineers.
Electronics, Communications and Computer Sciences, vol. E86-
A, no. 12, pp. 3291–3293, 2003. H. C. So was born in Hong Kong. He ob-
[19] K. W. Cheung, H. C. So, W.-K. Ma, and Y. T. Chan, “Re- tained the B.Eng. degree from City Uni-
ceived signal strength based mobile positioning via con- versity of Hong Kong and the Ph.D. de-
strained weighted least squares,” in Proceedings of the IEEE In- gree from The Chinese University of Hong
ternational Conference on Acoustic, Speech and Signal Process- Kong, both in electronic engineering, in
ing (ICASSP ’03), vol. 5, pp. 137–140, Hong Kong, April 2003. 1990 and 1995, respectively. From 1990 to
[20] K. W. Cheung, H. C. So, W.-K. Ma, and Y. T. Chan, “Least 1991, he was an electronic engineer at the
squares algorithms for time-of-arrival-based mobile location,” Research & Development Division of Everex
IEEE Transactions on Signal Processing, vol. 52, no. 4, pp. 1121– Systems Engineering Ltd., Hong Kong. Dur-
1130, 2004. ing 1995-1996, he worked as a postdoctoral
[21] S. M. Kay, Fundamentals of Statistical Signal Processing: Estima- fellow at The Chinese University of Hong Kong. From 1996 to 1999,
tion Theory, Prentice-Hall, Englewood Cliffs, NJ, USA, 1993. he was a Research Assistant Professor at the Department of Elec-
[22] H.-L. Song, “Automatic vehicle location in cellular commu- tronic Engineering, City University of Hong Kong. Currently he is
nications systems,” IEEE Transactions on Vehicular Technology, an Associate Professor in the Department of Electronic Engineer-
vol. 43, no. 4, pp. 902–908, 1994. ing at City University of Hong Kong. His research interests include
K. W. Cheung et al. 23

adaptive filter theory, detection and estimation, wavelet transform,

and signal processing for communications and multimedia.

W.-K. Ma obtained the B.Eng. (with

first class honors) in electrical and elec-
tronic engineering from the University of
Portsmouth, Portsmouth, UK, in 1995. He
received the M.Phil. and Ph.D. degrees,
both in electronic engineering, from The
Chinese University of Hong Kong (CUHK),
Hong Kong, in 1997 and 2001, respectively.
Since August 2005, he has been an Assistant
Professor in the Department of Electrical
Engineering and the Institute of Communications Engineering,
the National Tsing Hua University, Taiwan. Previously he has held
research positions at McMaster University, Canada, CUHK, Hong
Kong, and the University of Melbourne, Australia. His research
interests are in signal processing for communications and statistical
signal processing. Dr. Ma’s Ph.D. dissertation was commended to
be “of very high quality and well-deserved honorary mentioning”
by the Faculty of Engineering, CUHK, in 2001.

Y. T. Chan was born in Hong Kong, and re-

ceived his electrical engineering education
in Canada. His Bachelor’s and Master’s de-
grees are from Queen’s University, and his
Ph.D. degree from the University of New
Brunswick. He was an engineer with Nor-
tel Networks and has been a Professor in the
Electrical and Computer Engineering De-
partment at the Royal Military College of
Canada, serving as Head of the department
from 1994 to 2000. From 2002 to 2005, he was a Visiting Professor
at the Electronic Engineering Department of The Chinese Univer-
sity of Hong Kong. Presently he is an Adjunct Professor at the Royal
Military College. His research interests are in detection, estimation,
localization, and tracking. Kluwer Academic Publishers published
his text Wavelet Basics in 1994. Dr. Chan was an Associate Editor of
the IEEE Transactions on Signal Processing, the Technical Chair of
ICASSP-84, General Chair of ICASSP-91, Vice Chair of ICASSP-03,
and Social Chair of ICASSP-04. He directed a NATO ASI in 1988.