ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 1

Cramér-Rao Bound for Localization with A

Priori Knowledge on Biased Range
Measurements
Tao Wang
arXiv:1104.4056v1 [cs.IT] 20 Apr 2011

Abstract

This paper derives a general expression for the Cramér-Rao bound (CRB) of wireless localization
algorithms using range measurements subject to bias corruption. Specifically, the a priori knowledge
about which range measurements are biased, and the probability density functions (PDF) of the biases
are assumed to be available. For each range measurement, the error due to estimating the time-of-arrival
of the detected signal is modeled as a Gaussian distributed random variable with zero mean and known
variance. In general, the derived CRB expression can be evaluated numerically. An approximate CRB
expression is also derived when the bias PDF is very informative. Using these CRB expressions, we
study the impact of the bias distribution on the mean square error (MSE) bound corresponding to the
CRB. The analysis is corroborated by numerical experiments.

Index Terms

Cramér-Rao bound, localization, biased range measurements, wireless networks.

I. I NTRODUCTION

Wireless localization systems have been attracting intensive research interest in both academia and
industry lately. For indoor or dense urban environments, Global Positioning Systems do not function well
for geolocation purposes. Instead, a few beacons at fixed and known positions are exploited for geolocation
using measurements related to signal power decay, bearing, difference of range, and range between a

T. Wang is with Communications and Remote Sensing Laboratory, Institute of Information and Communication Tech-
nologies, Electronics and Applied Mathematics, Université Catholique de Louvain, 1348 Louvain-la-Neuve, Belgium. E-mail:
tao.wang@uclouvain.be.

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 2

target and the beacons [1]. Thanks to the superior multipath resolution and penetration capability of ultra-
wideband (UWB), a range measurement can be obtained by UWB pulsed signals with high accuracy [2].
Therefore, range based localization has become a promising option for short-range wireless networks.
Much work on ranging algorithms based on time of arrival (TOA) measurements has been reported
[2], [3]. Specifically, they are usually based on detecting the first arriving signal from a transmitter
and estimating its time of arrival (TOA) at a receiver. Note that the clocks of the transmitter and the
receiver need to synchronized, and the transmission time at the transmitter should be sent to the receiver
for producing the range measurements. In this paper, we consider the geolocation algorithms using
range measurements generated between the target and multiple beacons in the aforementioned way. It
is important to note that the considered algorithms are suitable for cooperative (blue-force) geolocation
applications, in that they require the cooperation of the target and the beacons in order to synchronize
their clocks for generating range measurements.
In general, a range measurement error can be modeled as the sum of two terms. The first one is due to
the error of estimating the TOA of the detected signal, while the second one corresponds to the difference
between the path length traveled by the detected signal and the transmitter-receiver distance. For instance,
the second term is nonzero when the detected signal does not come from the line of sight propagation
path. The first term is usually modeled as a zero mean random variable with a Gaussian distribution in
the literature [4]–[6]. As a consequence, the second term is equal to the bias of the range measurement.
It is well known that the Cramér-Rao bound (CRB) sets a fundamental lower limit to the covariance
of any unbiased estimator for a vector parameter (please refer to pages 63 to 72 in [7]). Specifically,
the mean square error (MSE) of any unbiased location estimator is lower bounded by the corresponding
CRB’s trace, referred to as the MSE bound hereafter. The CRB and the MSE bound are used extensively
to evaluate particular localization algorithms, as well as guide the geolocation system design fulfilling
certain accuracy requirements [4], [8], [9]. When some of the range measurements are biased, the CRB
has been derived in [10] by regarding each bias as a deterministic nuisance parameter, and it was shown
in [5] that using those bias-corrupted measurements does not improve the CRB.
Recently, much work has been reported on modeling the bias as a random variable with a particular
probability density function (PDF). Specifically, this PDF is obtained empirically by extensive experiments
[11], [12], or theoretically by scattering models [13]. In case the PDF of the bias is known a priori, a
Cramér-Rao-like bound has been derived in [14] for the joint estimation of the deterministic target location
and the random bias. This bound is referred to as the hybrid CRB in [15], [16], since the parameters
to be estimated consist of both deterministic and random parameters. It was shown in [5] that using the

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 3

bias-corrupted measurements improves the MSE bound associated with the hybrid CRB.
Specifically, the hybrid CRB was derived from the joint PDF of the distance measurements and the
random bias. Compared with the hybrid CRB, the CRB derived from the marginal PDF of range
measurements is much tighter and can be asymptotically attained by a maximum likelihood (ML)
estimator [7], [16]. When the bias PDF is approximated by a piecewise constant function, an expression
that can be numerically evaluated was derived for the CRB in [6]. Numerical results showed that for
uniformly distributed bias, the presence of the bias degrades the MSE bound and using the bias-corrupted
measurements improves the MSE bound.
Compared with the above existing work, this paper contains the following contributions:
• We derive a general expression that can be numerically evaluated for the CRB. When the prior bias
PDF is very informative, an approximate CRB expression is derived.
• Based on these expressions, the impact of the bias distribution on the MSE bound corresponding to
the trace of the CRB is studied analytically.
The rest of this paper is organized as follows. Section II describes a typical 2D localization system
and related models for range measurements. In Section III, a general CRB expression is derived for this
system. After that, an approximate CRB expression is derived in Section IV when the prior bias PDF is
very informative. Based on these expressions, the impact of the bias distribution on the MSE bound is
studied analytically in Section V. We will show some numerical results in Section VI, and complete this
paper by some conclusions in Section VII.
Notations: Upper (lower) boldface letters denote matrices (column vectors), and [·]T represents the
transpose operator. ∇x (f (x)) stands for a column vector which is the gradient of the function f (x) with
respect to x, and ∇2x (f (x)) represents a matrix which is the Hessian of the function f (x) with respect
to x.

II. SYSTEM SETUP AND RANGING MODELS

We consider a typical 2D geolocation system equipped with M beacons (M ≥ 3). In the following
sections, although only the CRB for this 2D system is derived, the CRB for a 3D system can be derived in
the same way and will be given as well. The coordinate of beacon m is pm = [xm , ym ]T (m = 1, · · · , M ),
and a target is located at u = [xu , yu ]T .
Suppose independent range measurements have been produced between a target and beacons, and the
one between beacon m and the target is denoted by rm . Without loss of generality, we assume the first L
range measurements are biased, while the other measurements are all unbiased. The measurement error

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 4

related to rm is modeled as:


 ε +b m = 1, · · · , L
m m
em = rm − dm = (1)
 εm m = L + 1, · · · , M,
where dm represents the distance between the target and beacon m, and εm is due to the TOA estimation
error of the detected signal, and bm corresponds to the difference of the path length traveled by the
detected signal and the distance from beacon m to the target. We assume bm has the a priori known
PDF p(bm ) with its mean and variance denoted by bm and κ2m , respectively. In addition, we assume bm
is independent of εm , and εm is Gaussian distributed with zero mean. Besides, we denote its variance
2 and assume it is known a priori. The motivation behind this assumption is twofold. One is that
as σm
it simplifies the mathematical derivation. The other is that the impact of the bias distribution on the
localization CRB can be studied when the TOA estimation has a guaranteed accuracy at a prescribed
level.
We stack all range measurements into the column vector r = [r1 , · · · , rM ]T . The log-likelihood function
of u is denoted as Λ = ln(p(r; u)) where p(r; u) is the PDF of r. Since all range measurements are
independent of each other, Λ = M
P
m=1 Λm , where Λm = ln(p(rm ; u)) is the log-likelihood function of

rm given u, and p(rm ; u) is the PDF of rm .

For an unbiased range measurement rm (m = L + 1, · · · , M ), p(rm ; u) can be formulated as:
(rm − dm )2
 
1
p(rm ; u) = p exp − 2
(2)
2
2πσm 2σm
For a biased range measurement rm (m = 1, · · · , L), we define Γm = ln(p(rm |bm ; u)) as the joint
log-likelihood function of rm given u and bm , where p(rm |bm ; u) is the PDF of rm conditional on bm
and u. p(rm |bm ; u) and p(rm ; u) can be expressed respectively as:
(rm − dm − bm )2
 
1
p(rm |bm ; u) = p exp − 2
(3)
2πσm2 2σm
Z +∞
p(rm ; u) = p(rm |bm ; u)p(bm )dbm (4)
−∞

III. D ERIVATIONS OF A GENERAL CRB EXPRESSION

The CRB is the inverse of the Fisher information matrix (FIM) F, which is expressed by:

F = Er;u (∇u Λ)(∇u Λ)T

 
(5)

where Er;u [·] denotes the expectation operator with respect to p(r; u). Since all distance measurements
are independent of each other, F = M
P
m=1 Fm where Fm is the FIM related to rm [7]:

Fm = Erm ;u (∇u Λm )(∇u Λm )T

 
(6)

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 5

where Erm ;u [·] denotes the expectation operator with respect to p(rm ; u).
When m = L + 1, · · · , M , we can show that:
(rm − dm )2
 
Fm = Erm ;u 4
φm φTm
σm
−2
= σm φm φTm , (7)

u−pm
where φm = dm represents the orientation of the target relative to beacon m.
When m = 1, · · · , L, we can first prove that:
Z +∞
1
∇u Λm = [∇u p(rm |bm ; u)] p(bm )dbm
p(rm ; u) −∞
Z +∞
p(rm |bm ; u)p(bm )
= [∇u Γm ]dbm
−∞ p(rm ; u)
= Ebm |rm ;u [∇u Γm ]
−2
= σm am φ m (8)

where am = Ebm |rm ;u [rm − dm − bm ] and Ebm |rm ;u [·] denotes the expectation operator with respect
to the posterior PDF p(bm |rm ; u). This means that am actually stands for the posterior mean of εm
conditional on rm . Inserting (8) into (6), we can compute Fm by:

Fm = λm φm φTm (9)

−4 · E
 2
where λm = σm rm ;u am .

Summing F1 , · · · , FM and taking the inverse, we can express the CRB as:
" L M
#−1
X X
−2
C= λm φm φTm + σm φm φTm (10)
m=1 m=L+1

The CRB for a 3-D localization system can be derived in the same way as given above. Most
interestingly, we can show after simple mathematical arrangement that the CRB for the 3-D localization
system can still be evaluated by (10), when u and pm are substituted with the 3-D coordinates of the
target and beacon m, respectively.
It is important to note that (10) is a general expression for the CRB, when each biased measurement
rm is Gaussian distributed conditional on bm . However, we have proven that ∇u Λm = Ebm |rm ;u [∇u Γm ]
holds in general for any particular form of p(rm |bm ; u). This means that a closed-form CRB expression
can be derived in a similar way when p(rm |bm ; u) takes other specific form.
To illustrate the generality of (10), we show that the CRB can be derived by (10) when bm is determinis-
tic. Note that this CRB will be used when illustrating the effectiveness of the approximate CRB derived in

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 6

Section IV. In this case, p(bm |rm ; u) = δ(bm − bm ), am = rm − dm − bm , p(rm ; u) = p(rm |bm = bm ; u),
−2 . This means that the contribution of r to the CRB is the same as if this range
and therefore λm = σm m

measurement were not bias corrupted. This is true since when bm is deterministic and known a priori,
its value can be subtracted from rm , which means that rm is in effect not subject to the bias corruption.
In general, an analytical expression for λm does not exist, in such a case λm can be evaluated
numerically, with C following from equation (10). This provides a way to study the effects of the
geometric configuration, the TOA estimation quality, as well as the bias distribution on the CRB. To this
end, it can easily be shown that
Z +∞
−4
λm = σm a2m p(rm ; u)drm (11)
−∞

and
+∞
p(rm |bm ; u)p(bm )
Z
am = (rm − dm − bm ) dbm , (12)
−∞ p(rm ; u)
respectively. For every value of rm , p(rm ; u) and am can be first evaluated by computing numerically
the integrals in (4) and (12) with respect to bm 1 , respectively. Then, λm can be evaluated by computing
numerically the integral with respect to rm in (11).
Note that the CRB expression in (10), as those presented in [4]–[6], is derived under the assumption
that we have the a priori knowledge about which range measurements are biased. In almost all practical
scenarios, this a priori knowledge may not be available [18]–[20]. In this case, the CRB in (10) is an
optimistic lower bound, since it assumes the implicit knowledge which is not really available for the
range measurements. We will illustrate the optimism of (10) when the a priori knowledge about which
range measurements are biased is not available by numerical experiments in Section VI. Relaxing the
assumption of knowing which measurements are biased, and generating an information reduction factor
that scales the CRB when the probability of each measurement getting biased is known a priori, in a
similar way as shown in [18]–[20], would be a valuable piece of future work.
We should also note that the derived CRB is for cooperative geolocation applications, since the
synchronization between the target and the beacons is required as mentioned in the introduction part. For
geolocating evasive targets, range measurements would only be available with active sensing, in which
case azimuth measurements, etc., would be available as well. It is another interesting piece of future work
to derive the CRB for localization using azimuth measurements, etc., in addition to range measurements.

1
The detailed procedure of numerical integration can be found in pages 198 − 200 in [17].

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 7

IV. D ERIVATION OF AN APPROXIMATE CRB EXPRESSION WHEN κm ≪ σm

We now derive an approximate CRB expression when bm has a small variance compared to σm (i.e.
κm ≪ σm ). First of all, we express Fm (m = 1, · · · , L) alternatively as [7]:

Fm = −Erm ;u ∇2u Λm
 
(13)

Next, we approximate p(rm |bm ; u) by the first-order Taylor-series expansion as follows:

′
p(rm |bm ; u) ≈ fm + fm · (bm − bm ) (14)

′ are expressed respectively as:

where fm and fm

fm = p(rm |bm = bm ; u) (15)

′ ∂p(rm |bm ; u)
fm = (16)
∂bm bm =bm

Based on (4) and (14), we can show that p(rm ; u) ≈ fm . In addition, we make the following
approximations:

′ · (b − b )
p(rm |bm ; u) fm + fm m m
≈
p(rm ; u) fm
∂Γm
=1 + · (bm − bm )
∂bm bm =bm
rm − dm − bm
=1 + 2
(bm − bm ) (17)
σm

Inserting (17) into (12), we can approximate am as follows:

κ2m
 
am ≈ 1 − 2 (rm − dm − bm ) (18)
σm
Based on (8) and (18), ∇2u Λm can be approximated by:

−2
∇2u Λm = σm (∇u am )φTm + am (∇2u dm )
 

κ2m
   
−2 T 2
≈ σm − 1 − 2 φm φm + am (∇u dm ) (19)
σm
Using (18), (19), and (13), Fm can be approximated by:

Fm ≈ λ′m φm φTm (20)

where

λ′m = σm
−2
1 − (κm /σm )2 .


(21)

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 8

From (9) and (20), we can see that λm ≈ λ′m . In addition, CRB can be approximated by:
" L M
#−1
X X
C≈ λ′m φm φTm + σm−2
φm φTm (22)
m=1 m=L+1

Note that a better approximation of the CRB can be derived with a higher order Taylor-series expansion
in the same way as given above, when the statistical moments of bm higher than the 2nd order are known
a priori. Here, we only use the first order Taylor-series expansion to derive an approximation when the
2nd order moment of bm (i.e. κm ) is known. The derived approximate CRB expression features a simple
mathematical structure but reveals the influence of the bias distribution on the CRB clearly.
We will illustrate the effectiveness of (22) by numerical experiments using measured bias distributions in
Section V. In fact, we can analytically illustrate the effectiveness of (22) when bm is Gaussian distributed.
Note that although in practice the Gaussian distribution is unlikely to be a good approximation of the
true bias distribution, it lends the CRB to be easily simplified analytically so that the effectiveness of
(22) can be examined. In this case, rm can be equivalently modeled as rm = dm + bm + ε′m where ε′m is
2 + κ2 . Therefore, λ = (σ 2 + κ2 )−1 according
Gaussian distributed with zero mean and variance σm m m m m

to the earlier analysis for the case when the bias is deterministic in Section III. Since κm ≪ σm , we can
see that C can indeed be approximately computed with (22), due to the fact that
−1
κ2m

−2
λm = σ m 1 + 2
σm
κ2
 
−2
≈ σm 1− m 2
σm
≈ λ′m . (23)

V. M SE BOUND ANALYSIS

It is important to note that the MSE bound, namely the trace of C and denoted by Tr(C), reduces
as λm increases. This is because M
P
m=1 Fm increases in a positive definite sense, therefore both C and

Tr(C) decrease, which means that the MSE bound improves, as λm increases.
When rm (m = 1, · · · , L) are discarded, we can compute C by setting λm to 0. Since λm is no smaller
than 0, using a bias-corrupted measurement results in the same or better MSE bound than discarding it.

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 9

An important property of λm is that λm ≤ σm −2 . This can be justified by:

h
2 i
−4
λm = σ m · Erm ;u Ebm |rm ;u [rm − dm − bm ]
−4
· Erm ;u Ebm |rm ;u (rm − dm − bm )2
  
≤ σm
−4
· Ebm Erm |bm ;u (rm − dm − bm )2
  
= σm
−2
= σm (24)

where Ebm [·] and Erm |bm ;u [·] stand for the expectation operator with respect to p(bm ) and p(rm |bm ; u),
respectively. Specifically, the second inequality is according to the Jensen’s inequality, and the equality
holds if bm is deterministic (please refer to pages 77 − 78 in [21]). This implies the presence of the
random bias bm degrades the MSE bound.
It is interesting to consider two special cases:

1) The first case is when bm has a very large variance compared to σm (i.e., κm ≫ σm ) so that for
any fixed rm , p(bm ) is approximately constant within the nonzero support of p(rm |bm ; u). This
case corresponds to the scenario where the prior bias PDF is not informative, and we can find that
λm ≈ 0. If bm is Gaussian distributed, this is the case when κ2m is sufficiently large, and λm is
indeed close to zero. This means that discarding a bias-corrupted measurement degrades the MSE
bound slightly when the prior bias PDF is not informative.
2) The second case is when bm has a very small variance compared to σm (i.e., κm ≪ σm ) so that for
any fixed rm , p(rm |bm ; u) varies slowly with respect to bm within the nonzero support of p(bm ).
This case corresponds to the scenario where the prior bias PDF is very informative. Based on the
analysis in Section IV, we can find that λm ≈ λ′m = σm
−2 1 − (κ /σ )2 ≈ σ −2 . This means that


m m m

the presence of the random bias bm results in a slight degradation to the MSE bound, compared to
the case in which there are no measurement biases.

VI. N UMERICAL EXPERIMENTS

For illustration purposes, we consider a wireless localization system with four beacons located respec-
tively at p1 = [0, 10]T , p2 = [0, −10]T , p3 = [−10, 0]T , p4 = [10, 0]T , and a target at u = [−1, −5]T , as
shown in Figure 1. We assume r1 is biased while r2 , r3 , and r4 are unbiased, and σm = 1 (m = 1, · · · , 4).
p(b1 ) is assumed to have the same shape as the measured bias distribution reported in [11]. Specifically,

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 10

10 p
1

y−axis (m)
0 p p
3 4

−2

−4
u
−6

−8

−10 p
2

−10 −5 0 5 10
x−axis (m)

Fig. 1. The localization system with 4 beacons considered in the numerical experiments. Note that only r1 , namely the range
measurement between u and p1 , is biased.

TABLE I
T HE VALUE OF Pi , i = 0, · · · , 8.

P0 P1 P2 P3 P4 P5 P6 P7 P8
0.12 0.03 0.31 0.12 0.24 0.12 0.03 0 0.03

p(b1 ) is expressed as

 0 if b1 ∈ (−∞, Ω0 ] ∪ (Ω9 , ∞)
p(b1 ) = (25)
 Pi if b1 ∈ (Ωi , Ωi+1 ] with i = 0, · · · , 8,
∆

where Ωi = 0.1 + i∆, and Pi is given in Table I for i = 0, · · · , 8. When ∆ = 0.1, p(b1 ) is shown in
Figure 2. It can be readily derived that b1 = 0.1 + 3.49∆ and κ1 = 1.83∆, which means that κ1 is a
linear function of ∆. When ∆ increases from 0.1 to 1, we have computed κ1 , and the results are shown
in Figure 3. Note that all the above values related to a coordinate or length have the units of meters.
To examine the effectiveness of the approximate CRB expression (22), we have evaluated the CRB
from (10) numerically, and the approximate CRB from (22) when ∆ increases with κ1 /σm < 1 satisfied.
The MSE bounds corresponding to those CRBs are shown with respect to κ1 /σm in Figure 4.a. We can
see that when κ1 /σm ≤ 0.5, the MSE bound corresponding to the approximate CRB is very close to the
one computed from (10). This illustrates the effectiveness of (22) when κ1 ≪ σm .
To examine the analysis in Section V, we have also evaluated the CRB from (10) numerically when ∆
increases with κ1 /σm > 1 satisfied, and the corresponding MSE bound is shown with respect to κ1 /σm

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 11

3.5

2.5

The bias PDF p(b )

1
2

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
bias (m)

Fig. 2. The PDF of b1 , which is the bias of the range measurement between u and p1 , when ∆ = 0.1 m.

1.8

1.6

1.4

1.2
κ (m)

1
1

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
∆ (m)

Fig. 3. κ1 , which is the standard deviation of the random bias b1 , when ∆ increases from 0.1 to 1 m.

in Figure 4.b. We have also computed the CRBs when r1 is discarded and when r1 is not bias corrupted,
respectively, and shown the corresponding MSEs in both Figure 4.a and Figure 4.b. It is clearly shown in
Figure 4.a that as κ1 reduces, the MSE bound approaches the one corresponding to the case when r1 is
not bias corrupted. On the other hand, as κ1 increases, the MSE bound approaches the one corresponding
to the case when r1 is discarded, as shown in Figure 4.b. These observations corroborate the analysis in
Section V.
Note that the derived CRB is an optimistic bound in scenarios when we do not have the a priori

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 12

1.4

1.35 MSE bound for CRB computed from (10)

MSE bound for approximate CRB computed from (22)
1.3 MSE bound for CRB from (10) when r is discarded
1
MSE bound for CRB from (10) when r is unbiased
1

MSE bound (m2)

1.25

1.2

1.15

1.1

1.05

1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
κ1/σm

(a) when κ1 /σm < 1

1.4

1.35

1.3
MSE bound (m2)

1.25

1.2
MSE bound for CRB computed from (10)
1.15 MSE bound for CRB computed from (10) when r is discarded
1
MSE bound for CRB computed from (10) when r is unbiased
1
1.1

1.05

1
0 5 10 15 20 25 30 35
κ /σ
1 m

(b) when κ1 /σm > 1

Fig. 4. The MSE bounds corresponding to the CRBs computed when κ1 /σm < 1 and κ1 /σm > 1, respectively.

knowledge about which range measurements are biased. To illustrate this remark explicitly, we compare
the MSE bound corresponding to the CRB computed from (10) with the MSE of two ML location
estimators as κ1 /σm increases, since the ML estimator is widely used and able to achieve the corre-
sponding CRB asymptotically. For the first ML estimator, the above mentioned a priori knowledge is not
available, and a binary parameter sm is introduced to indicate that rm (m = 1, · · · , 4) is bias corrupted
if sm = 0. Specifically, the first ML estimator produces a joint estimate of u and {sm , m = 1, · · · , 4}
as the maximizer of the log-likelihood function
4
X
L(u, s1 , s2 , s3 , s4 ) = Lm (u, sm ), (26)
m=1

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 13

2.2
MSE bound for CRB computed from (10)

MSE bound or MSE of the ML estimator (m )

2
MSE of the ML estimator with a priori knowledge
2
MSE of the ML estimator without a priori knowledge

1.8

1.6

1.4

1.2

1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
κ1/σm

Fig. 5. The MSEs of the ML location estimators and the MSE bound corresponding to the CRB computed from (10). Note
that the MSE bound is the same as the corresponding one depicted in Fig. 4.

where
(rm − dm )2

1
Lm (u, sm ) = ln( p ) + ln sm exp{− 2
}
2πσm2 2σm
Z +∞
(rm − dm − bm )2

+(1 − sm ) p(bm ) exp{− 2
}db m , (27)
−∞ 2σm
and p(bm ) = p(b1 ), m = 2, 3, 4. For the second ML estimator, the a priori knowledge is available, i.e.,
this estimator knows a priori that only r1 is biased, and produces an estimate of u as the maximizer of
L(u, s1 , s2 , s3 , s4 ) when s1 = 0 and sm = 1, m = 2, 3, 4. For every ∆, the MSE of each estimator is
computed by averaging the square errors of the location estimates for 1000 random realizations of range
measurements, and the results are shown in Figure 5 with respect to κ1 /σm . We can see that the MSE
of the second ML estimator is slightly above the MSE bound computed from (10), while the MSE of the
first ML estimator is much greater than that MSE bound. These observations indicate that the derived
CRB is indeed an optimistic lower bound for localization algorithms without the a priori knowledge
about which range measurements are biased.

VII. C ONCLUSION

We have derived a general expression for the CRB of wireless localization algorithms using range
measurements subject to bias corruption. Specifically, the knowledge about which range measurements
are biased, and the probability density functions (PDF) of the biases are assumed to be known a priori.
For each range measurement, the error due to estimating the TOA of the detected signal is modeled

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 14

as a Gaussian distributed random variable with zero mean and known variance. We have also derived
an approximate CRB expression when the bias PDF is very informative. Using these CRB expressions,
we have studied the impact of the bias distribution on the MSE bound corresponding to the CRB. The
analysis has been corroborated by numerical experiments.

ACKNOWLEDGEMENT

The author is very grateful to Prof. Claude Jauffret for coordinating the review, as well as the anonymous
reviewers for their valuable comments and suggestions to improve the quality of this paper.

R EFERENCES

[1] N. Bulusu, J. Heidemann, and D. Estrin, “GPS-less low-cost outdoor localization for very small devices,” IEEE Personal
Communications, vol. 7, no. 5, pp. 28–34, oct. 2000.
[2] J. Y. Lee and R. A. Scholtz, “Ranging in a dense multipath environment using an UWB radio link,” IEEE Journal on
Selected Areas in Communications, vol. 20, no. 9, pp. 1677–1683, Dec. 2002.
[3] C. Falsi, D. Dardari, L. Mucchi et al., “Time of arrival estimation for UWB localizers in realistic environments,” EURASIP
J. Applied Signal Proc., vol. 2006, pp. 1–13.
[4] F. Gustafsson and F. Gunnarsson, “Mobile positioning using wireless networks,” IEEE Signal Processing Magazine, vol. 22,
no. 4, pp. 41–53, Jul. 2005.
[5] Y. Qi, H. Kobayashi, and H. Suda, “Analysis of wireless geolocation in a non-line-of-sight environment,” IEEE Trans.
Wireless Communications, vol. 5, no. 3, pp. 672–681, Mar. 2006.
[6] D. B. Jourdan, D. Dardari, and M. Z. Win, “Position error bound for UWB localization in dense cluttered environments,”
IEEE Trans. Aerospace and Electronic Systems, vol. 44, no. 2, pp. 613–628, Apr. 2008.
[7] H. L. V. Trees, Detection, estimation, and modulation theory, Part I. John Wiley & Sons, 1968.
[8] S. Gezici, Z. Tian, G. B. Giannakis et al., “Localization via ultra-wideband radios,” IEEE Signal Processing Magazine,
vol. 22, no. 4, pp. 70–84, Jul. 2005.
[9] T. Wang, G. Leus, and L. Huang, “Ranging energy optimization for robust sensor positioning based on semidefinite
programming,” IEEE Transactions on Signal Processing, vol. 57, no. 12, pp. 4777–4787, Dec. 2009.
[10] Y. Qi and H. Kobayashi, “Cramér-Rao lower bound for geolocation in non-line-of-sight environment,” in Proc. ICASSP,
May 2002, pp. 2473–2476.
[11] D. B. Jourdan, J. J. Deyst, M. Z. Win et al., “Monte-carlo localization in dense multipath environments using UWB
ranging,” in Proc. IEEE Conf. UWB, Sep. 2005, pp. 314–319.
[12] B. Alavi and K. Pahlavan, “Modeling of the TOA-based distance measurement error using UWB indoor radio measure-
ments,” IEEE Communications Letters, vol. 10, no. 4, pp. 275–277, Apr. 2006.
[13] S. Al-Jazzar, J. J. Caffery, and H. You, “Scattering-model-based methods for TOA location in NLOS environments,” IEEE
Trans. Vehicular Technology, vol. 56, no. 2, pp. 583–593, Mar. 2007.
[14] Y. Qi and H. Kobayashi, “On geolocation accuracy with prior information in non-line-of-sight environment,” in Proc. VTC,
Sep. 2002, pp. 285–288.

May 28, 2018 DRAFT

ACCEPTED BY IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS ON 22/DEC/2010 15

[15] F. Gini and R. Reggiannini, “On the use of Cramér-Rao-Like bounds in the presence of random nuisance parameters,”
IEEE Trans. Communcations, vol. 48, no. 12, pp. 2120–2126, Dec. 2000.
[16] I. Reuven and H. Messer, “A Barankin-Type lower bound on the estimation error of a hybrid parameter vector,” IEEE
Trans. Information Theory, vol. 43, no. 3, pp. 1084–1093, May 1997.
[17] J. Kiusalaas, Numerical Methods in Engineering with Python. Cambridge University Press, 2005.
[18] M. Hernandez, A. Marrs, N. Gordon, S. Maskell, and C. Reed, “Cramér-Rao bounds for non-linear filtering with
measurement origin uncertainty,” in Proceedings of the 5th International Conference on Information Fusion, vol. 1, 2002,
pp. 18–25.
[19] X. Zhang, P. Willett, and Y. Bar-Shalom, “Dynamic Cramér-Rao bound for target tracking in clutter,” IEEE Transactions
on Aerospace and Electronic Systems, vol. 41, no. 4, pp. 1154–1167, Oct. 2005.
[20] M. Hernandez, A. Farina, and B. Ristic, “PCRLB for tracking in cluttered environments: measurement sequence
conditioning approach,” IEEE Transactions on Aerospace and Electronic Systems, vol. 42, no. 2, pp. 680–704, Apr. 2006.
[21] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge University Press, 2004.

Tao Wang received respectively B.E. and DoE degree in electronic engineering from Zhejiang University,
China, in 2001 and 2006, as well as civil electrical engineering degree (summa cum laude) from Université
Catholique de Louvain, Belgium. He had multiple research appointments in Motorola Electronics Ltd.
Suzhou Branch, China, since 2000 to 2001, the Institute for Infocomm Research, Singapore, since 2004
to 2005, Delft University of Technology and Holst Center in the Netherlands since 2008 to 2009. He is
now a researcher in Université Catholique de Louvain, Belgium. He has been an associate editor in chief
for Signal Processing: An International Journal (SPIJ) since Oct. 2010. His current research interests are in the optimization
of wireless localization systems with energy awareness, as well as resource allocation algorithms in wireless communication
systems.

May 28, 2018 DRAFT