CRAMER-RAO BOUND AND OPTIMUM SENSOR ARRAY

FOR SOURCE LOCALIZATION FROM TIME DIFFERENCES OF ARRIVAL

B. Yang and J. Scheuing

Chair of System Theory and Signal Processing

University of Stuttgart, Germany

ABSTRACT thus
1
This paper presents a theoretical analysis of the Cramer-Rao τi j =
di j (p) + ni j
v
lower bound for source localization from time differences of v is the wave propagation speed and ni j is the TDOA mea-
arrival. We derive properties of the Cramer-Rao bound and surement error.
design optimum sensor arrays which minimize the bound. For M sensors, there are a total number of M(M − 1)/2
possible sensor pairs and TDOA measurements. Let
1. INTRODUCTION
I0 = {(i, j)|1 ≤ j < i ≤ M} (1)
The technique of source localization by using time differ- denote the set of all sensor pairs. I is a subset of I0 and
ences of arrival (TDOA) plays an important role in many ap- contains these N (N ≤ M(M−1) 2 ) sensor pairs whose TDOA
plications like navigation, localization and tracking of acous- measurements are used in source localization. By introduc-
tic sources, and location services of mobile communication. ing the N × 1 vectors
In the literature, the main research effort was focused on the ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
development of various methods for the source position es- ⎢⎢⎢ τi j ⎥⎥⎥ ⎢⎢⎢ di j ⎥⎥⎥ ⎢⎢⎢ ni j ⎥⎥⎥
⎢
τ = ⎢⎣ .. ⎥⎦ ⎥ ⎢
, d = ⎢⎣ .. ⎥⎦ ⎥ , n = ⎢⎣ .. ⎥⎥⎦
⎢ ,
timation. A recent review of some methods can be found in . (i, j)∈I . (i, j)∈I . (i, j)∈I
[1]. In comparison, relatively less is known about the the-
oretical performance of these methods. There exist many the signal model for the TDOA measurements becomes
numerical comparisons between different methods, but the
1
impact of the sensor array geometry on the localization ac- τ= d(p) + n (2)
curacy is still not well understood yet. v
This paper aims at a better understanding of the theoret- The problem of source localization is to estimate the source
ical performance of source localization. We discuss proper- position vector p given q , τ, and v.
i
ties of the Cramer-Rao bound for TDOA based localization.
We derive necessary and sufficient conditions for the op-
3. CRAMER-RAO BOUND
timum sensor array geometry which minimizes the bound.
Examples of optimum arrays for both two-dimensional and The Cramer-Rao bound (CRB) is a lower bound for the co-
three-dimensional localization are given. variance matrix of unbiased estimators. It is often used as a
The following notations are used in the paper. Matrices benchmark against which the efficiency of unbiased estima-
and column vectors are represented by boldface and under- tors is tested. It is given by J−1 where
lined characters. The superscript T denotes transpose.
·
 
is the Euclidean vector norm. tr(·) is the matrix trace. J = E [∇ ln f (τ; p)][∇ ln f (τ; p)]T
is the Fisher information matrix. f (τ; p) is the probability
2. PROBLEM FORMULATION density function (PDF) of τ. ∇ is the gradient operator with
respect to p. E(·) denotes the expectation on τ.
We consider the source localization problem in D (D =
Assume that the measurement error vector n in (2) is
2, 3). The sensor array consists of M sensors at the locations
Gaussian with zero mean and the full rank covariance ma-
q ∈ D (i = 1, . . . , M). The source position is p ∈ D . The
i trix C which is independent of p. The PDF of τ is
distance between the source and sensor i is di (p) =
p −
q
(i = 1, . . . , M). The difference in the distances of sensor 1 1 1 1
i
i and j from the source is given by di j (p) = di (p) − d j (p). f (τ; p) = √ exp − (τ − d)T C−1 (τ − d)
(2π)N/2 det(C) 2 v v
The noisy TDOA measurement between sensor i and j is (3)

0-7803-8874-7/05/$20.00 ©2005 IEEE IV - 961 ICASSP 2005

The CRB is well known from [2] be the truncated Taylor series of d(p) around the solution
p found at the k-th iteration. The quadratic order and all
J−1 = v2 (GC−1 GT )−1 (4) k
higher-order terms of the Taylor series are ignored. Maxi-
with mizing the likelihood f (τ; p) in (3) while using the approx-
imation (7) results in an improved estimate
G = ∇ d T = [g . . .](i, j)∈I ,
ij
g = g −g , (5) p = p + [G(p )C−1 GT (p )]−1 ·
ij i j k+1 k k k
p−q G(p )C−1 [τv − d(p )] (8)
i k k
g = ∇di (p) = (i = 1, . . . , M)
i
p − q
The key issue of this method is to find a good initial guess
i

Clearly, g is a unit-length vector with
g
= 1. It points p to avoid local minima and divergence.
0
i i
from sensor i to the source. G is a D × N matrix. It depends Prop. 3: The more TDOA measurements we use, the smaller
on the source and sensor positions and on the set I of sensor the CRB is. Let Ii (i = 1, 2) be two sets of TDOA measure-
pairs which are used for source localization. ments and J−1 the corresponding CRBs. If I1 ⊂ I2 , then
Ii
J−1
I1
≥ J −1
I2
, i.e. J −1
I1
− J−1
I2
is non-negative definite.
4. PROPERTIES OF THE CRAMER-RAO BOUND Proof: The assumption I1 ⊂ I2 implies that all columns of
GI1 = [g . . .](i, j)∈I1 are also contained in GI2 = [g . . .](i, j)∈I2 .
In this section, we discuss some properties of the CRB. ij ij
This means GI1 C−1 GTI1 ≤ GI2 C−1 GTI2 and J−1 I1
≥ J−1
I2
. This
Prop. 1: One necessary condition for the existence of the property holds even for correlated TDOA measurements.
CRB is M ≥ D + 1. In the literature, many source localization methods rely
Proof: For the existence of the D × D inverse Fisher infor- on the spherical model instead of the hyperbolic model. They
mation matrix in (4), the D × N matrix G must have rank D. are spherical intersection [5], spherical interpolation [6] and
Due to the definition (5), any vectors g can be written as a further improvements of them [2, 1]. These methods are
ij
linear combination of the M − 1 vectors g (i = 2, . . . , M): non-iterative and computationally efficient. However, they
i1
g = g − g . Therefore, the maximum number of linearly suffer from the drawback that maximum M − 1 sensor pairs
ij i1 j1 I = {(2, 1), . . . , (M, 1)} (with sensor 1 being the reference
independent columns in G is M − 1. This means
sensor) can be used. This limits the accuracy of the source
M − 1 ≥ rank(G) = D localization. According to Prop. 3, the CRB of these meth-
ods is larger than the CRB when we use all M(M − 1)/2
sensor pairs.
Prop. 2: Under the Gaussian error model (3), no unbiased
Therefore, the following two-step procedure is suitable
estimators attain the CRB.
to improve the localization accuracy:
Proof: A necessary and efficient condition for an unbiased
efficient estimator is the existence of some functions J(·) • Use a non-iterative method like any of the above spher-
and h(·) such that [3] ical methods to find a reliable initial guess p for the
0
∇ ln f (τ; p) = J(p)[h(τ) − p] (6) source position.

Starting from (3), the gradient vector is calculated to • Use one or a few Gauss-Newton-iterations in (8), based
on all sensor pairs I0 , to further improve the localiza-
1 1 tion accuracy.
∇ ln f (τ; p) = G(p)C−1 τ − d(p)
v v
Since the condition (6) can never be satisfied, there are no Prop. 4: The CRB does not depend on the range
p − q

i
efficient estimators for the source position vector p. Never- between the source and sensors. It only depends on the di-
theless, the maximum-likelihood (ML) estimator rection of the vectors p − q .
i
Proof: This follows immediately from (5).
p̂ = arg max f (τ; p) Note that this is a property of the CRB. The covariance
ML p
matrices of source position estimators will depend on both
is able to attain the CRB asymptotically (N → ∞). range and direction of p − q .
i
One possibility to find the ML estimator is to use a Gauss-
Newton iteration [4]. Let
5. OPTIMUM ARRAY GEOMETRY
T
d(p) ≈ d(p ) + ∇ d(p) p=p (p − p )
k k
k The geometry of the source localization problem is deter-
T
= d(p ) + G (p )(p − p ) (7) mined by the sensor and source positions q and p. Starting
k k k i

IV - 962
from the CRB in (4), we now study how to optimize the ge- This means
ometry to minimize the CRB for a given number of sensors. D2
tr[(GGT )−1 ] ≥ (11)
This problem has been addressed in [7, 8] for the narrow- tr(GGT )
band far-field direction-of-arrival (DOA) estimation. To our
knowledge, there are no similar works for the TDOA based The equality holds if and only if all eigenvalues of GGT are
source localization which is usually a broadband near-field identical λi = λ (i = 1, . . . , D). This is equivalent to
estimation problem.
GGT = λI (12)
Since the CRB is a square matrix, we consider the mini-
mization of the trace of the CRB tr(J−1 ). It is a lower bound
Now we study the term tr(GGT ). Due to g = g − g ,
for the sum of variances of unbiased estimators for all el- ij i j
M(M−1)
ements of the sensor position vector p. For simplicity, we the relationship between the D × 2 matrix G in (9) and
consider white measurement noise with C = σ2 I in this pa- the D × M matrix g in C2 is
per. I is an identity matrix and σ2 is the variance of the
TDOA measurement noise. According to Prop. 3, we have G = gT
to use the set I0 of all M(M − 1)/2 sensor pairs in order to M(M−1)
One example for the M× 2 transform matrix T for
achieve a CRB as small as possible. This means
M = 4 is
G = [g ... g g ... g ... g ] (9) ⎡ ⎤
21 M1 32 M2 M,M−1 ⎢⎢⎢ −1 −1 −1 0 0 0 ⎥⎥
⎥
⎢⎢⎢ 1 0 0 −1 −1 0 ⎥⎥⎥⎥
The optimization problem becomes to find M unit-length T = ⎢⎢⎢⎢ ⎥
vectors g ∈ D (i = 1, . . . , M) in such a way that ⎢⎢⎣ 0 1 0 1 0 −1 ⎥⎥⎥⎥
i
⎦
0 0 1 0 1 1
tr(J−1 ) = (vσ)2 tr[(GGT )−1 ]
In general, each row of T has the squared norm M − 1 and
is minimized. each pair of rows of T has the inner product −1. Corre-
spondingly,
Theorem 1: If C = σ2 I and I = I0 , ⎡ ⎤
2 ⎢⎢⎢ M − 1 −1 ... −1 ⎥⎥⎥
D ⎢⎢⎢ −1 M−1 ... −1 ⎥⎥⎥
tr(J−1 ) ≥ (vσ)2 ⎢ ⎥⎥⎥
TTT = ⎢⎢⎢⎢⎢
(10)
M2 .. .. .. .. ⎥⎥⎥ = MI − 1 1T
⎢⎢⎢ . . . . ⎥⎥⎥
The equality holds if and only if ⎣ ⎦
−1 −1 ... M−1
M
C1) g = 0 and with 1 = [1, . . . , 1]T . This leads to
i=1 i

C2) the D × M matrix g = [g . . . g ] satisfies ggT = GGT = gTTT gT = MggT − (g1)(g1)T (13)
1 M
(M/D)I, i.e. g has orthogonal row vectors with equal
row norm. and

Proof: The proof consists of two parts. We first find a tr(GGT ) = Mtr(ggT ) − tr[(g1)(g1)T ]
lower bound for tr[(GGT )−1 ] in terms of tr(GGT ) because M
tr(GGT ) is simpler to calculate than tr[(GGT )−1 ]. Then we = M tr(g gT ) −
g1
2
maximize tr(GGT ). i=1
i i

Let λi > 0 (i = 1, . . . , D) be the eigenvalues of the D× D = M −
g1
2 ≤ M 2

2
(14)
symmetric and positive definite matrix GGT . The eigenval-
ues of (GGT )−1 are 1/λi . It is well known The equality holds if and only if the condition C1 is satis-
D D fied:
M
tr(GGT ) = λi , tr[(GGT )−1 ] = 1/λi
g1 = g =0 (15)
i=1 i=1 i
i=1
According to the Cauchy-Schwarz inequality, Combining (11) and (14), we obtain
⎡ √ ⎤T ⎡ √ ⎤ 
⎢⎢⎢ λ1 ⎥⎥⎥ ⎢⎢⎢ 1/ λ1 ⎥⎥⎥   D D D2 D2
⎢⎢⎢ . ⎥⎥⎥ ⎢⎢⎢ . ⎥⎥⎥ tr[(GGT )−1 ] ≥ ≥
⎢
D = ⎢⎢ . ⎥⎥ ⎢⎢ . ⎥ ⎢ . ⎥
⎥⎥⎥ ≤ λi · 1/λi tr(GGT ) M 2
⎢⎣ √ ⎥⎦ ⎢⎣ √. ⎦ i=1 i=1
λD 1/ λD
 Combining (12) with (13) and (15), we obtain ggT = cI.
= tr(GGT )tr[(GGT )−1 ] Since tr(ggT ) = Dc = M, the condition C2 yields.

IV - 963
5.1. Optimum 2D Array tetrahed. octahed. cube icosahed. dodecahed.

We have derived the necessary and sufficient conditions C1

and C2 to achieve the minimum CRB. But how does the
optimum array look like?
Let g = [cos αi , sin αi ]T for D = 2. The condition C1 is
i M=v=4 6 8 12 20
M
clearly equivalent to e jαi
= 0. The condition C2 implies e=6 12 12 30 30
i=1 f =4 8 6 20 12
M M M
cos2 αi = sin2 αi = M
and sin αi cos αi = 0. It is
i=1 i=1
2
i=1 Table 1. Platonic solids and their number of vertices v,
easy to verify that a “uniform angular array” (UAA) defined edges e, and faces f
by
2π
αi = α0 + (i − 1) (i = 1, . . . , M) (16) 7. REFERENCES
M
satisfies both C1 and C2. UAA, however, is sufficient but [1] Y. Huang, J. Benesty, and G. W. Elko, “Source localiza-
not necessary as shown by the next theorem. tion,” in Audio Signal Processing for Next-Generation
Multimedia Communication Systems, Y. Huang and
Theorem 2: A sensor array consists of K subarrays charac-
J. Benesty, Eds., chapter 9. Kluwer Academic Publish-
terized by the D × Mk matrices gk (k = 1, . . . , K). If each
ers, 2004.
subarray is optimum in the sense of C1 and C2, then the
K
overall array of M = Mk sensors with g = [g1 . . . gK ] is [2] Y. T. Chan and K. C. Ho, “A simple and efficient es-
k=1 timator for hyperbolic location,” IEEE Trans. Signal
optimum as well. Processing, vol. 42, pp. 1905–1915, 1994.
Proof: From gk 1 = 0 and gk gTk = const · I for all k, we con-
K K [3] J. M. Mendel, Lessons in Estimation Theory for Signal
clude immediately g1 = gk 1 = 0 and ggT = gk gTk = Processing, Communications, and Control, Prentice-
k=1 k=1
const · I. Hall, Englewood Cliffs, 2 edition, 1995.
Interestingly, we also know optimum arrays other than
[4] W. H. Foy, “Position-location solutions by taylor-series
superposition of UAAs.
estimation,” IEEE Trans. Aerospace and Electron. Sys-
tems, vol. 12, pp. 187–194, 1976.
5.2. Optimum 3D Array
[5] H. C. Schau and A. Z. Robinson, “Passive source
In the three-dimensional case (D = 3), we again look for localization employing intersecting spherical surfaces
“uniform angular arrays” whose vectors g are “equally” from time-of-arrival differences,” IEEE Trans. Acous-
i
distributed on a unit spherical surface. There are exactly five tic, Speech, and Signal Processing, vol. 35, pp. 1223–
solutions to this symmetry problem, the so called Platonic 1225, 1987.
solids: tetrahedron, octahedron, cube, icosahedron, and do-
decahedron [9]. Table 1 shows these solids and summarizes [6] J. O. Smith and J. S. Abel, “Closed-form least-squares
their number of vertices, edges, and faces. Each of the vec- source location estimation from range-difference mea-
tors g points from the center of solids to one of the vertices. surements,” IEEE Trans. Acoustic, Speech, and Signal
i Processing, vol. 35, pp. 1661–1669, 1987.
The number of vertices v is identical to the number of sen-
sors M. The vectors g are known from the literature. It is [7] C. Chambers, T. C. Tozer, et al., “Temporal and spa-
i
straightforward to show that all Platonic solids satisfy the tial sampling influence on the estimates of superim-
optimum array conditions C1 and C2. posed narrowband signals: When less can mean more,”
Theorem 2 also applies to the three-dimensional case. IEEE Trans. Signal Processing, vol. 44, pp. 3085–3096,
This means, any superposition of centered Platonic solids is 1996.
again an optimum sensor array. A superposition of a tetra-
hedron and an octahedron returns, for example, an optimum [8] A. B. Gershman and J. F. Böhme, “A note on most fa-
array with M = 10 sensors. vorable array geometries for DOA estimation and array
interpolation,” IEEE Signal Processing Letters, vol. 4,
pp. 232–235, 1997.
6. ACKNOWLEDGMENT
[9] H. Cundy and A. Rollett, Mathematical Models, Strad-
The authors would like to thank Mr. Zeile for valuable dis- broke, England: Tarquin Pub., 3 edition, 1989.
cussions.

IV - 964