Digital Signal Processing 19 (2009) 650–659

Contents lists available at ScienceDirect

Digital Signal Processing

www.elsevier.com/locate/dsp

A study of two-dimensional sensor placement using

time-difference-of-arrival measurements
Kenneth W.K. Lui, H.C. So ∗
Department of Electronic Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong

a r t i c l e i n f o a b s t r a c t

Article history: Finding the position of a radiative source based on time-difference-of-arrival (TDOA)
Available online 31 January 2009 measurements from spatially separated receivers has important applications in sonar,
radar, mobile communications and sensor networks. Each TDOA defines a hyperbolic locus
Keywords:
on which the source must lie and the position estimate can then be determined with
Time-difference-of-arrival
Optimum array geometry
the knowledge of the sensor array geometry. While extensive research works have been
Source localization performed on algorithm development for TDOA estimation and TDOA-based localization,
limited attention has been paid in sensor array geometry design. In this paper, an optimum
two-dimensional sensor placement strategy is derived with the use of optimum TDOA
measurements, assuming that each sensor receives a white signal source in the presence of
additive white noise. The minimum achievable Cramér–Rao lower bound is also produced.
© 2009 Elsevier Inc. All rights reserved.

1. Introduction

Passive source localization using measurements from an array of spatially separated sensors is an important problem
in radar, sonar, mobile communications and wireless sensor networks. The time-difference-of-arrival (TDOA) method is a
popular strategy for source localization and it usually proceeds in a two-step fashion as follows. TDOA measurements of
the source signal received at the sensor array are first obtained. In the second step, the TDOA information is utilized to
construct a set of hyperbolic equations that are highly nonlinear, from which the source position can be determined with
the knowledge of the sensor array geometry.
Although extensive research has been performed in TDOA estimation [1,2] as well as TDOA-based localization [3–6], most
of them do not consider the impact of the sensor array geometry on the localization accuracy. Nevertheless, Yang et al. [7–9]
have recently pioneered the theoretical study for sensor array placement strategies. In [7], the properties of the Cramér–Rao
lower bound (CRLB) for TDOA-based positioning [3,4] and optimum sensor arrays are derived. It is proved that for two-
dimensional (2D) source localization, uniform angular arrays (UAAs) and its superpositions can attain the minimum CRLB.
Works in [8] and [9] are extensions to [7]. The former studies the performance loss for UAAs with reduced angular apertures
while the latter develops the relationship between several sensor placement schemes. In spite of the elegant and rigorous
mathematical treatment, it is invalid to assume uncorrelated TDOA measurements as they should be correlated [10]. In this
paper, we will study the array geometry based on the correlated TDOA estimates, which are optimally computed using the
sensor outputs where each of them receives an uncorrelated signal source in the presence of additive white Gaussian noise.
The rest of the paper is organized as follows. In Section 2, we first develop the Fisher information matrices (FIMs) for
the TDOA and 2D position estimates in the case of white signal source and noise. By minimizing the trace of the CRLB
for positioning, it is shown that UAAs and its superpositions correspond to an optimum sensor placement and this is in

* Corresponding author.
E-mail address: [email protected] (H.C. So).

1051-2004/$ – see front matter © 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.dsp.2009.01.002
 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659 651

agreement with [7]. The minimum achievable CRLB for positioning is also derived. Numerical examples are provided in
Section 3 to validate our research findings. Finally, conclusions are drawn in Section 4.

2. Development of optimum sensor placement

Suppose there are L  3 sensors and the signal received at the lth sensor is modeled as

zl (n) = s(n − D l ) + ql (n), l = 1, 2, . . . , L , n = 0, 1, . . . , N − 1, (1)

where s(n) is the white source signal, ql (n) and D l are the additive white Gaussian noise and signal propagation delay,
respectively, at the lth sensor, and N is the number of samples available at each sensor. Without loss of generality, we
assign the first sensor as the reference and define the TDOA parameter vector as d = [d21 , d31 , . . . d L1 ] T where T denotes the
transpose operator and di1 = D i − D 1 , i = 2, 3, . . . , L. The FIM for d, denoted by F(d), is [10]:

π
N S 2 (ω)    
F(d) = ω2 L tr Q−1 (ω) Q−1 −1 T −1
p (ω ) − Q p (ω )1 L −1 1 L −1 Q p (ω ) dω , (2)
2π 1+ l=1 S (ω)/ Q l (ω)
−π

where tr is the trace operator, 1i is the i × 1 vector with all elements 1, S (ω) and Q l (ω) represent the power spectra of s(n)
and ql (n), respectively, while Q(ω) = diag( Q 1 (ω), Q 2 (ω), . . . , Q L (ω)) and Q p (ω) = diag( Q 2 (ω), Q 3 (ω), . . . , Q L (ω)) where
diag(a1 , a2 , . . . , an ) is a diagonal matrix whose diagonal entries are a1 , a2 , . . . , an . It is clear from (2) that the optimum
TDOA estimates are correlated. Assuming that s(n) and {ql (n)} are uncorrelated white Gaussian processes with variances σs2
and σq2 , respectively, we have S (ω) = σs2 , Q l (ω) = σq2 , l = 1, 2, . . . , L, Q(ω) = σq2 I L and Q p (ω) = σq2 I L −1 where Ii represents
the i × i identity matrix. Under the white signal and noise assumption, (2) can be simplified to [11]:

π 2 N Λ2 ( LIL −1 − 1L −1 1TL−1 )
F(d) = , (3)
3(1 + L Λ)

where Λ = σs2 /σq2 is the signal-to-noise ratio (SNR). The covariance matrix of the optimum estimate for d when N → ∞ is
equal to the inverse of F(d), which has the form of [11]:

3(1 + L Λ)  
F−1 (d) = 2 N Λ2
I L −1 + 1 L −1 1 TL−1 . (4)
π
Let the position of the source and sensors be x = [x, y ] T and xl = [xl , yl ] T , l = 1, 2, . . . , L, respectively. The FIM for x using
the optimum TDOA estimates, denoted by F(x), is [5]:

GF(d)G T
F(x) = , (5)
c2
where

G = [g21 , g32 , . . . , g L1 ],

gl1 = gl − g1 ,
 x−xl 
√
g x,l (x−xl )2 +( y − yl )2 cos(θl )
gl = = y − yl =
g y ,l √ sin(θl )
(x−xl )2 +( y − yl )2

and c and θl denote the known signal propagation speed and the incline angle from the source to the lth sensor, respectively.
With the use of (3), (5) becomes

L ⎡  L  L L L ⎤
2
⎢ ( L − 1) cos (θl ) − cos (θ i ) cos (θ j ) L cos (θl ) sin (θ l ) − cos (θl ) sin (θ l ) ⎥
π 2 N Λ2 ⎢ ⎢ l=1 i = j l=1 l=1 l=1 ⎥
⎥.
F(x) = 2 ⎢ ⎥ (6)
3c (1 + L Λ) ⎣  L  L  L  L L
⎦
2
L cos(θl ) sin(θl ) − cos(θl ) sin(θl ) ( L − 1) sin (θl ) − sin(θi ) sin(θ j )
l=1 l=1 l=1 l=1 i = j

The optimum sensor placement strategy is obtained by minimizing the trace of the CRLB for positioning, that is, tr(F−1 (x)).
In Appendix A, we have proved that tr(F−1 (x)) can be expressed as f (θ ) where θ = [θ1 , θ2 , . . . , θ L ] T , and f (θ ) is of the
form:
L
L ( L − 1) − 2 i> j cos(θi − θ j )
f (θ ) = L L L . (7)
2
L ( L − 2) i> j sin (θl − θ j ) + 2L l=1 i> j sin(θl − θ j ) sin(θi − θl )
652 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659

Let the numerator and denominator of f (θ ) be u (θ ) and v (θ ), respectively. As f (θ ) > 0 because tr(F−1 (x)) should be
positive and u (θ ) is obviously larger than 0, we have v (θ ) > 0. Differentiating (7) with respect to θl , l = 1, 2, . . . , L, and
setting the resultant expressions to zero yield:
 
∂ u (θ )  ∂ v ( θ ) 
= f (θ ) , l = 1, 2, . . . , L , (8)
∂θl θl =θ̂l ∂θl θl =θ̂l

where the components of [θ̂1 , θ̂2 , . . . , θ̂ L ] T are the optimum sensor directions with respect to the source. We have shown in
Appendix B that (8) can be simplified to:
 
∂ u (θ )  ∂ v ( θ ) 
= 0 and = 0, l = 1, 2, . . . , L , (9)
∂θl θl =θ̂l ∂θl θl =θ̂l
which are only satisfied by the following conditions (see Appendix B):


L 
L 
L 
L
cos(θ̂l ) = 0, sin(θ̂l ) = 0, cos(2θ̂l ) = 0 and sin(2θ̂l ) = 0. (10)
l=1 l=1 l=1 l=1

An obvious solution for (10) is

2π
θ̂l = (l − 1) + φ, l = 1, 2, . . . , L , (11)
L
where φ ∈ (0, 2π ) which corresponds to an UAA. As an illustration, let L = 3 and the first incline angle is fixed at θ̂1 = 0.
After simple trigonometric manipulations, (10) is reduced to
3
1 + cos(2θ̂2 ) + cos(2θ̂3 ) = 0 ⇒ sin2 (θ̂2 ) + sin2 (θ̂3 ) = , (12)
4
0 + sin(2θ̂2 ) + sin(2θ̂3 ) = 0 ⇒ cos(θ̂2 ) sin(θ̂2 ) + cos(θ̂3 ) sin(θ̂3 ) = 0, (13)
1 + cos(θ̂2 ) + cos(θ̂3 ) = 0, (14)

0 + sin(θ̂2 ) + sin(θ̂3 ) = 0 ⇒ sin(θ̂2 ) = − sin(θ̂3 ). (15)

Solving (12) to (15) gives θ̂2 = 2π /3 and θ̂3 = 4π /3, which agree with (11). Nevertheless, a generalized solution is
2π
θ̂ m,i = (lm − 1) + φm , i = 1, 2, . . . , L m , m = 1, 2, . . . , M , (16)
Lm
where the optimal direction vector, θ̂ m,i = [θ̂1,1 , θ̂1,2 , . . . , θ̂1, L 1 , θ̂2,1 , . . . , θ̂2, L 2 , . . . , θ̂ M , L M ] T with L 1 + L 2 + · · · + L M = L,
φm ∈ [0, 2π ) and lm is the index of Lm . It is noteworthy that these findings agree with [7], although the latter assumes
uncorrelated TDOA information.
With the use of (10), F(x) with optimum array geometry is easily shown to be:

π 2 N L 2 Λ2
F(x) = I2 . (17)
6c 2 (1 + LSNR)
According to Appendix C, the minimum achievable trace of the CRLB for TDOA-based positioning is then:
  12c 2 (1 + L Λ) 12c 2 12c 2
tr F−1 (x) = = + , (18)
π 2 N L2Λ π 2 N L 2 Λ2 π 2 N LΛ
where we see that the CRLB decreases linearly with N. Moreover, the bound is inversely proportional to L 2 Λ2 for Λ 1
while its value decreases linearly with L and Λ for Λ 1. It is worthy to point out that the result of (18) is essentially
identical to [7] up to a scaling factor.

3. Numerical examples

Simulation results are presented in this section to validate our analytical findings. The source position is fixed at x =
[0, 0] T . First the superiority of the UAA over other typical geometries, namely, nonuniform angular array (NAA), corners,
L-shape and uniform linear array (ULA), in terms of mean square position error (MSPE) performance, is demonstrated
in Fig. 1. We consider a rectangular area of dimension 20 m × 10 m with four sensors whose coordinates for different
placement strategies are tabulated in Table 1. As optimum TDOA measurements are assumed, all the MSPEs are computed
using tr(F−1 (x)) with c = 360 m s−1 . It is observed in Fig. 1 that the UAA strategy has 2 dB to 7 dB improvement over other
standard placement schemes for different SNR conditions.
Second, we study the optimum angular separation by considering an angular array with L = 3 where the first sensor
position is fixed at x1 = [5 cos(0), 5 sin(0)] T . To avoid duplication, θ3 > θ2 is assigned. Fig. 2 shows the MSPE versus different
θ2 and θ3 at Λ = 50 dB. We see that the MSPE is minimized when θ2 = 2π /3 and θ3 = 4π /3, which conform to our
analytical calculations in (12)–(15).
 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659 653

Fig. 1. Mean square position errors of different sensor geometries versus SNR when L = 4.

Table 1
Senors positions for different placement strategies.

Geometry x1 (m) x2 (m) x3 (m) x4 (m)

UAA [5 cos(0), 5 sin(0)]T [5 cos(π /2), 5 sin(π /2)]T [5 cos(π ), 5 sin(π )]T [5 cos(3π /2), 5 sin(3π /2)]T
NAA [5 cos(π /3), 5 sin(π /3)]T [5 cos(2π /3), 5 sin(2π /3)]T [5 cos(π ), 5 sin(π )]T [5 cos(4π /3), 5 sin(4π /3)]T
Corners [−10, 5]T [10, 5]T [10, −5]T [−10, −5]T
L-shape [−10, 5] T
[0, −5] T
[10, −5] T
[−10, −5]T
ULA [−10, −5]T [10, −5]T [−2.5, −5]T [2.5, −5]T

Fig. 2. Mean square position error versus angular separation when L = 3.

654 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659

4. Conclusion

Via optimum time-difference-of-arrival (TDOA) estimation in an sensor array where each sensor receives a white source
signal in the presence of white noise, we have found that uniform angular arrays and its superpositions correspond to an
optimum sensor placement strategy in the two-dimensional scenarios. The minimum achievable Cramér–Rao lower bound
is also produced. Our future works include extension of our development to arbitrary TDOA covariance matrices and/or
three-dimensional positioning.

Acknowledgment

The work described in this paper was supported by a grant from CityU (Project No. 7002132).

Appendix A

In this appendix, we show that tr(F−1 (x)) can be expressed as f (θ ). Denoting the (i , j ) entry of 3c 2 (1 + L Λ)F−1 (x)/
(π 2
N Λ2 ) by f i j , F−1 (x) is

3c 2 (1 + L Λ) 1 f 22 − f 12
F −1 (x) = × . (A.1)
π 2 N Λ2 f 11 f 22 − f 12 f 21 − f 21 f 11

The trace of F −1 (x) is then

  3c 2 (1 + L Λ) f 11 + f 22
tr F−1 (x) = × . (A.2)
π 2 N Λ2 f 11 f 22 − f 12 f 21

With the use of (6), we have:


L 
L 
L 
L
f 11 + f 22 = ( L − 1) cos2 (θl ) − cos(θi ) cos(θ j ) + ( L − 1) sin2 (θl ) − sin(θi ) sin(θ j )
l=1 i = j l=1 i = j


L 
L
= L ( L − 1) − cos(θi − θ j ) = L ( L − 1) − 2 cos(θi − θ j ), (A.3)
i = j i> j


L 
L 
L 
L
f 11 f 22 = ( L − 1) cos2 (θl ) − cos(θi ) cos(θ j ) × ( L − 1) sin2 (θl ) − sin(θi ) sin(θ j )
l=1 i = j l=1 i = j
 2  L 2

L 
L 
L 
2 2
= L cos (θl ) − cos(θl ) × L sin (θl ) − sin(θl )
l=1 l=1 l=1 l=1
 2  2

L 
L 
L 
L
2 2 2
=L cos (θl ) sin (θl ) + cos(θl ) sin(θl )
l=1 l=1 l=1 l=1
 2  L 2

L 
L 
L 
−L cos2 (θl ) sin(θl ) −L sin2 (θl ) cos(θl ) (A.4)
l=1 l=1 l=1 l=1

and
2

L 
L 
L
2
f 12 f 21 = f 12 = L cos(θl ) sin(θl ) − cos(θl ) sin(θl )
l=1 l=1 l=1
 L 2  L 2  2
  
L
2
=L cos(θl ) sin(θl ) + cos(θl ) sin(θl )
l=1 l=1 l=1


L 
L 
L
− 2L cos(θl ) sin(θl ) cos(θl ) sin(θl ). (A.5)
l=1 l=1 l=1

From (A.4) and (A.5), we obtain:

 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659 655

 2

L 
L 
L
2 2 2 2 2
f 11 f 22 − f 12 =L cos (θl ) sin (θl ) − L cos(θl ) sin(θl )
l=1 l=1 l=1
 2  2

L 
L 
L L
−L sin2 (θl ) cos(θl ) −L cos2 (θl ) sin(θl )
l=1 l=1 l=1 l=1


L 
L 
L
+ 2L cos(θl ) sin(θl ) cos(θl ) sin(θl ). (A.6)
l=1 l=1 l=1

The five terms in (A.6) are computed as


L 
L 
L 
L
L2 cos2 (θl ) sin2 (θl ) = L 2 cos2 (θl ) sin2 (θl ) + L 2 cos2 (θi ) sin2 (θ j ) + sin2 (θi ) cos2 (θ j ), (A.7)
l=1 l=1 l=1 i> j
 2

L 
L 
L
L 2
cos(θl ) sin(θl ) = L2 cos2 (θl ) sin2 (θl ) + 2L 2 cos(θi ) sin(θi ) cos(θ j ) sin(θ j ), (A.8)
l=1 l=1 i> j
 2

L 
L 
L 
L 
L 
L
2
L sin (θl ) cos(θl ) =L cos2 (θl ) sin2 (θl ) + 2L sin2 (θl ) cos(θi ) cos(θ j ), (A.9)
l=1 l=1 l=1 l=1 l=1 i> j
 2

L 
L 
L 
L 
L 
L
L cos2 (θl ) sin(θl ) =L cos2 (θl ) sin2 (θl ) + 2L cos2 (θl ) sin(θi ) sin(θ j ) (A.10)
l=1 l=1 l=1 l=1 l=1 i> j

and

L 
L 
L
2L cos(θl ) sin(θl ) cos(θl ) sin(θl )
l=1 l=1 l=1
 

L 
L 
L
= 2L cos(θl ) sin(θl ) cos(θl ) sin(θl ) + cos(θi ) sin(θ j ) + sin(θi ) cos(θ j )
l=1 l=1 i> j
 L 2
 
L 
L
= 2L cos(θl ) sin(θl ) + 2L cos(θl ) sin(θl ) cos(θi ) sin(θ j ) + sin(θi ) cos(θ j ). (A.11)
l=1 l=1 i> j

Subtracting (A.8) from (A.7) yields

 L 2

L 
L 
2 2 2 2
L cos (θl ) sin (θl ) − L cos(θl ) sin(θl )
l=1 l=1 l=1


L 
L
= L2 cos2 (θi ) sin2 (θ j ) + L 2 sin2 (θi ) cos2 (θ j ) − 2L 2 cos(θi ) sin(θi ) cos(θ j ) sin(θ j )
i> j i> j


L
 2 
L
= L2 cos(θi ) sin(θ j ) − L 2 sin(θi ) cos(θ j ) = L2 sin2 (θi − θ j ). (A.12)
i> j i> j

Substituting (A.9)–(A.12) into (A.6) and with the use of (A.12), we get
 2

L 
L 
L 
L
2 2 2 2 2
f 11 f 22 − f 12 =L sin (θi − θ j ) − 2L cos (θl ) sin (θl ) − cos(θl ) sin(θl )
i> j l=1 l=1 l=1


L 
L
+ 2L cos(θl ) sin(θl ) cos(θi ) sin(θ j ) + sin(θi ) cos(θ j )
l=1 i> j


L 
L 
L 
L
− 2L cos2 (θl ) sin(θi ) sin(θ j ) − 2L sin2 (θl ) cos(θi ) cos(θ j )
l=1 i> j l=1 i> j

 
L 
L 
L
 
= L 2 − 2L sin2 (θi − θ j ) + 2L cos(θl ) sin(θl ) cos(θi ) sin(θ j ) + cos(θ j ) sin(θi )
i> j l=1 i > j
656 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659


L 
L 
L 
L
− 2L sin2 (θl ) cos(θi ) cos(θ j ) − 2L cos2 (θl ) sin(θi ) sin(θ j )
l=1 i > j l=1 i > j

  
L 
L 
L
= L 2 − 2L sin2 (θi − θ j ) + 2L sin(θl − θ j ) sin(θi − θl ). (A.13)
i> j l=1 i > j

Putting (A.3) and (A.13) into (A.2) yields (7).

Appendix B

In Appendix B, we first prove that (8) can be simplified to (9) and then derive (10). For an arbitrary index of θ , say,
s = 1, 2, . . . , L, the derivatives of u (θ ) and v (θ ) with respect to θs are

∂(u (θ )) ∂( iL> j cos(θi − θ j ))
= −2
∂θs ∂θs

L 
L 
L
=2 sin(θs − θl ) = 2 sin(θs ) cos(θl ) − 2 cos(θs ) sin(θl ) (B.1)
l=1 l=1 l=1

and
  
∂( v (θ )) ∂( iL> j sin2 (θi − θ j )) ∂( lL=1 iL> j sin(θl − θ j ) sin(θi − θl ))
= L ( L − 2) + 2L (B.2)
∂θs ∂θs ∂θs
with

∂( iL> j sin2 (θi − θ j )) 
L 
L
= 2 sin(θs − θ j ) cos(θs − θ j ) − 2 sin(θi − θs ) cos(θi − θs )
∂θs
s> j i >s


L
= 2 sin(θs − θl ) cos(θs − θl )
l=1


L
  
L 
L
= sin 2(θs − θl ) = sin(2θs ) cos(2θl ) − cos(2θs ) sin(2θl ) (B.3)
l=1 l=1 l=1

and
 
∂( lL=1 iL> j sin(θl − θ j ) sin(θi − θl ))
∂θs
 

L
∂ 
L
∂ 
L
= sin(θl − θ j ) sin(θi − θl ) + sin(θs − θ j ) sin(θi − θs )
∂θs ∂θs
l=1 i> j i> j
s=l


L 
L 
L
= cos(θs − θl ) sin(θl − θi ) − sin(2θs − θi − θ j )
l=1 l=1 i> j
s=l


L 
L 
L 
L
= cos(θs − θl ) sin(θl − θi ) − sin(2θs ) cos(θi + θ j ) − cos(2θs ) sin(θi + θ j ). (B.4)
l=1 l=1 i> j i> j
s=l

The three terms in (B.4) can be written as:


L 
L
cos(θs − θl ) sin(θl − θi )
l=1 l=1
s=l


L 
L 
L 
L
= cos(θs − θl ) sin(θl ) cos(θi ) − cos(θs − θl ) cos(θl ) sin(θi )
l=1 i =1 l=1 i =1
s=l s=l


L 
L 
L 
L
= cos(θs − θl ) sin(θl ) cos(θi ) − cos(θs − θl ) cos(θl ) sin(θi ), (B.5)
l=1 i =1 l=1 i =1
s=l s=l
 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659 657


L 
L 
L 
L
cos(θi + θ j ) = cos(θi + θ j ) − cos(2θl )
i> j i =1 j =1 l=1


L 
L 
L 
L 
L
= cos(θi ) cos(θ j ) + sin(θi ) sin(θ j ) − cos(2θl )
i =1 j =1 i =1 j =1 l=1
 2  2

L L 
L
= cos(θl ) + sin(θl ) − cos(2θl ) (B.6)
l=1 l=1 l=1

and


L 
L 
L 
L
sin(θi + θ j ) = sin(θi + θ j ) − sin(2θl )
i> j i =1 j =1 l=1


L 
L 
L 
L 
L
= sin(θi ) cos(θ j ) + cos(θi ) sin(θ j ) − sin(2θl )
i =1 j =1 i =1 j =1 l=1


L 
L 
L
=2 cos(θl ) sin(θl ) − sin(2θl ). (B.7)
l=1 l=1 l=1

Using (B.3)–(B.7), (B.2) becomes

∂( v ( θ))  L  L
= L ( L − 1) sin(2θs ) cos(2θl ) − cos(2θs ) sin(2θl )
∂θs
l=1 l=1


L 
L 
L 
L
+ 2L cos(θs − θk ) sin(θk ) cos(θl ) − cos(θs − θk ) cos(θk ) sin(θl )
k=1 l=1 k=1 l=1
s=k s=k
 2  2

L 
L 
L
− 2L sin(2θs ) cos(θl ) + sin(θl ) − cos(2θl )
l=1 l=1 l=1


L 
L 
L
− 2L cos(2θs ) 2 cos(θl ) sin(θ j ) − sin(2θl ) . (B.8)
l=1 j =1 l=1

From (7), when θs = θ̂s , s = 1, 2, . . . , L, all the derivatives of u (θ ) are identical and based on (B.1), we let


L 
L
2 sin(θ̂s ) cos(θ̂l ) − 2 cos(θ̂s ) sin(θ̂l ) = C , (B.9)
l=1 l=1

where C is a constant to be determined. Summing (B.9) for s = 1, 2, . . . , L, we easily obtain LC = 0 which gives C = 0. That
is, ∂ u (θ )/∂θl |θ =θ̂ = 0. With the use of f (θ ) > 0 and (8), we then get ∂ v (θ )/∂θl |θ =θ̂ = 0. As a result simplification of (8) to
l l l l
(9) is shown.
To prove the uniqueness of the solutions given in (10), we employ propositional calculus. Based on (B.9), we construct
the following statement:


L 
L
2 sin(θ̂m ) cos(θ̂l ) − 2 cos(θ̂m ) sin(θ̂l ) = 0,
l=1 l=1


L 
L
2 sin(θ̂n ) cos(θ̂l ) − 2 cos(θ̂n ) sin(θ̂l ) = 0, m = n, m, n = 1, 2, . . . , L . (B.10)
l=1 l=1
 L
We will prove that (B.10) contradicts with the opposite conditions of (10), namely, = 0 and lL=1 sin(θ̂l ) = 0.
l=1 cos(θ̂l )
L L
This means l=1 cos(θ̂l ) = 0 and l=1 sin(θ̂l ) = 0 are the only solutions for (B.10). The existence of θ̂m and θ̂n is given by
the following two statements, and the second statement can further divided into three sub-statements:

(1) ∃θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θn );

(2) ∀θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∨ cos(θ̂m ) = cos(θ̂n ),
(a) ∀θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ),
658 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659

(b) ∀θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ),

(c) ∀θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ),
where ∧ and ∨ are conjunction and disjunction operators [12], respectively.

Now the above statements are examined one by one. For ∃θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ), we first assume
the opposition of (10):


L 
L
cos(θ̂l ) = 0 ∨ sin(θ̂l ) = 0. (B.11)
l=1 l=1

The assumption of (B.11) corresponds to the following three possibilities:


L 
L
cos(θ̂l ) = 0 ∧ sin(θ̂l ) = 0 ⇒ sin(θ̂1 ) = sin(θ̂2 ) = · · · = sin(θ̂ L ) = 0  sin(θ̂m ) = sin(θ̂n ), (B.12)
l=1 l=1


L 
L
cos(θ̂l ) = 0 ∧ sin(θ̂l ) = 0 ⇒ cos(θ̂1 ) = cos(θ̂2 ) = · · · = cos(θ̂ L ) = 0  cos(θ̂m ) = cos(θ̂n ) (B.13)
l=1 l=1

or


L 
L
cos(θ̂l ) = 0 ∧ sin(θ̂l ) = 0 ⇒ sin(θ̂m ) cos(θ̂n ) = sin(θ̂n ) cos(θ̂m ) ⇒ sin(θ̂m − θ̂n ) = 0  L  3 (B.14)
l=1 l=1

where  is contradiction operator [12].

For ∀θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ), we have

1
L
sin(θ̂s ) = sin(θ̂l ), for s = 1, 2, . . . , L  sin(θ̂m ) = sin(θ̂n ). (B.15)
L
l=1

For ∀θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ), we have

1
L
cos(θ̂s ) = cos(θ̂l ), for s = 1, 2, . . . , L  cos(θ̂m ) = cos(θ̂n ). (B.16)
L
l=1

For ∀θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ), we have

F(x) = 02×2 ⇒ F−1 (x) is undefined  F−1 (x) is well defined. (B.17)

Therefore, there is no solution for (B.10) except


L 
L
cos(θ̂l ) = 0 ∧ sin(θ̂l ) = 0 (B.18)
l=1 l=1

with ∃θ̂m , θ̂n : sin(θ̂m ) = sin(θ̂n ) ∧ cos(θ̂m ) = cos(θ̂n ). Substituting (B.18) into (B.8) with θs = θ̂s , s = 1, 2, . . . , L, yields:


L 
L
( L + 1) sin(2θ̂s ) cos(2θ̂l ) − ( L − 3) cos(2θ̂s ) sin(2θ̂l ) = 0. (B.19)
l=1 l=1

By undergoing a similar procedure of (B.10)–(B.17), we obtain


L 
L
cos(2θ̂l ) = 0 ∧ sin(2θ̂l ) = 0 (B.20)
l=1 l=1

with ∃θ̂m , θ̂n : sin(2θ̂m ) = sin(2θ̂n ) ∧ cos(2θ̂m ) = cos(2θ̂n ). Combining (B.18) and (B.20) yields (10).

Appendix C

In this appendix, we produce the minimum value of the CRLB for TDOA-based positioning. Using (A.2) and f 12 = f 21 , we
obtain
f 11 + f 22 f 11 + f 22
 . (C.1)
f 11 f 22 − f 12 f 21 f 11 f 22
 K.W.K. Lui, H.C. So / Digital Signal Processing 19 (2009) 650–659 659

Applying the arithmetic mean–geometric mean inequality of f 11 and f 22 yields

f 11 + f 22  ( f 11 + f 22 )2
 f 11 f 22 ⇒  f 11 f 22  L 2 . (C.2)
2 4
From (C.1) and (C.2), we get ( f 11 + f 22 )2 /4, hence

f 11 + f 22 4
 . (C.3)
f 11 f 22 f 11 + f 22
On the other hand, (A.3) can be written as
 2  2

L 
L
2
f 11 + f 22 = L − 2 cos(θl ) −2 sin(θl )  L2. (C.4)
l=1 l=1

Using (C.3) and (C.4), we obtain

f 11 + f 22 4
 . (C.5)
f 11 f 22 − f 12 f 21 L2

Including the scaling term of π 2 NSNR2 /(3c 2 (1 + L Λ)) in (C.5) yields

  12c 2 (1 + L Λ)
tr F−1 (x)  (C.6)
π 2 N L 2 Λ2
which proves (18).

References

[1] G.C. Carter (Ed.), Special Issue on Time Delay Estimation, IEEE Trans. Acoust. Speech Signal Process. 29 (1981) 461–623.
[2] G.C. Carter, Coherence and Time Delay Estimation: An Applied Tutorial for Research, Development, Test, and Evaluation Engineers, IEEE Press, 1993.
[3] D.J. Torrieri, Statistical theory of passive location systems, IEEE Trans. Aerosp. Electron. Syst. 20 (1984) 183–197.
[4] M.A. Spirito, On the accuracy of cellular mobile station location estimation, IEEE Trans. Veh. Technol. 50 (2001) 674–685.
[5] Y.T. Chan, K.C. Ho, A simple and efficient estimator for hyperbolic location, IEEE Trans. Signal Process. 42 (1994) 1905–1915.
[6] A. Dogandzic, J. Riba, G. Seco, A.L. Swindlehurst (Eds.), Special section on positioning and navigation with applications to communications, IEEE Signal
Proc. Mag. 22 (4) (2005) 10–84.
[7] B. Yang, J. Scheuing, Cramer–Rao bound and optimum sensor array for source localization from time differences of arrival, in: Proc. ICASSP, vol. 4,
Philadelphia, USA, Mar. 2005, pp. 961–964.
[8] B. Yang, J. Scheuing, A theoretical analysis of 2D sensor arrays for TDOA based localization, in: Proc. ICASSP, vol. 4, Toulouse, France, May 2006, pp.
901–904.
[9] B. Yang, Different sensor placement strategies for TDOA based localization, in: Proc. ICASSP, vol. 2, Honolulu, Hawaii, USA, Apr. 2007, pp. 1093–1096.
[10] W.R. Hahn, S.A. Tretter, Optimum processing for delay-vector estimation in passive signal arrays, IEEE Trans. Inform. Theory 19 (1973) 608–614.
[11] H.C. So, Y.T. Chan, F.K.W. Chan, Closed-form formulae for time-difference-of-arrival estimation, IEEE Trans. Signal Process. 56 (6) (2008) 2614–2620.
[12] B. Davey, H.A. Prisetley, Introduction to Lattices and Order, Cambridge University Press, 2002.

Kenneth Wing Kin Lui received the B.Eng. (first class) degree in computer engineering from City University of Hong Kong in 2006. He is
currently working towards the Ph.D. degree in electronic engineering at the same university. His research interests include statistical signal
processing and their applications, with particular attention to frequency estimation, complex network and graph theory, optimization
techniques such as particle swarming and convex optimization, intelligent system and data mining, localization problems, and related
mathematics.

H.C. So was born in Hong Kong. He received the B.Eng. degree from City University of Hong Kong and the Ph.D. degree from The Chinese
University of Hong Kong, both in electronic engineering, in 1990 and 1995, respectively. From 1990 to 1991, he was an Electronic Engineer
at the Research & Development Division of Everex Systems Engineering Ltd., Hong Kong. During 1995–1996, he worked as a Postdoctoral
Fellow at The Chinese University of Hong Kong. From 1996 to 1999, he was a Research Assistant Professor at the Department of Elec-
tronic Engineering, City University of Hong Kong. Currently, he is an Associate Professor in the Department of Electronic Engineering at
City University of Hong Kong. His research interests include fast and adaptive filter, signal detection, parameter estimation and source
localization.