CRB based-design of linear antenna arrays for near-field

source localization
Houcem Gazzah, Jean-Pierre Delmas

To cite this version:

Houcem Gazzah, Jean-Pierre Delmas. CRB based-design of linear antenna arrays for near-field
source localization. IEEE Transactions on Antennas and Propagation, 2014, 62 (4), pp.1965 - 1974.
�10.1109/TAP.2014.2298882�. �hal-01255287v2�

HAL Id: hal-01255287

https://hal.science/hal-01255287v2
Submitted on 30 May 2016

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 1

CRB Based-Design of Linear Antenna Arrays

for Near-Field Source Localization
Houcem Gazzah and Jean Pierre Delmas

Abstract

This paper is devoted to the Cramer Rao bound (CRB) on the angle and range of a narrow-band
near-field source localized by means of an arbitrary linear array using the exact expression of the time
delay parameter. First, we prove that the conditional and unconditional CRBs are proportional for an
arbitrary parametrization of the steering vector. Then, a Taylor expansion of the CRB is conducted to
obtain accurate non-matrix closed-form expressions of the CRB on angle and range. In contrast to the
existing expressions, our expressions are simple, interpretable and more general because the sensors are
only constrained to be placed along some axis. Our analysis leads to the characterization of a class of
centro-symmetric linear arrays with better near-field angle and range estimation capabilities.

Index Terms

Cramer Rao bounds, linear antenna arrays, direction-of-arrival and range estimation, near-field source
localization.

I. I NTRODUCTION

Uniform linear arrays (ULA) are the most commonly used type of linear antenna arrays because they
are ambiguity-free and allow for fast algorithms, when direction of arrival (DOA) estimation of far-field
sources is sought. However, when the source is located in the antenna near-field, a change of the signal
model occurs as a new parameter is to be considered: the source-to-antenna distance. Fast algorithms
are no longer applicable, and, more seriously, the new (range) parameter will affect DOA estimation
accuracy, and, for some applications, is itself of interest and needs to be estimated. In this context, we

Houcem Gazzah is with the Department of Electrical and Computer Engineering, University of Sharjah, 27272, UAE. E-
mail: hgazzah@sharjah.ac.ae, Tel.: (971) 6.5050.917, Fax.: (971) 6.5050.872. Jean-Pierre Delmas is with Telecom SudParis,
Departement CITI, CNRS UMR 5157, Evry, France. E-mail: jean-pierre.delmas@it-sudparis.eu.

December 7, 2013 DRAFT

 2

prove that the ULA configuration is not the best choice. Alternative (other than uniform) placement of
the array sensors is shown to improve range estimation.
To motivate our design, we study the algorithm-independent CRB which constitutes the minimum
achievable variance on the estimated source parameters, here DOA and range of the near-field source.
Despite the huge literature about DOA estimation [1], research has been mostly dedicated to far-field
sources. In fact, when the source is in the array far-field, the (planar) waveform reaches two sensors
with a time difference that is proportional to the spacing between the two sensors. Hence, it is possible
to obtain simple and interpretable non-matrix expressions for the CRB (see e.g., [2]). In contrast, when
the source is in the antenna near-field, the time delay expression is more intricate and only approximate
non-matrix expressions of this CRB can be obtained. Inspired by subspace-based DOA algorithms, early
ones were based on an approximate propagation model based on second-order Taylor expansion of the
time delay parameter [3], [4]. Only lately has the exact time delay formula been used [5], but restricted
to the ULA.
We start by proving a fact about the so-called conditional and unconditional CRBs. Often, they have
been considered as independent (see e.g., the recent papers [15] and [5] which even concludes by
”extension of this work for stochastic sources is under consideration”). In this paper, we show that
they are, actually, proportional, an issue previously overlooked. Then, we investigate accurate non-matrix
expressions of the CRB on both DOA and range, using linear arrays of arbitrarily spaced sensors. They
are more general than those of [5] because we do not assume uniform linear arrays (neither punctured
nor sparse) where inter-sensor spacings1 are multiples of a minimum distance [7]. They are also more
compact than those from [4], [5] and [6] if applied to the special case of ULA, and so thanks to a
different coordinate system that implies a different definition of DOA and range.
The obtained CRB expressions allow for a rich interpretation of the array estimation capabilities
when the source is in the antenna near-field. For instance, they highlight the interest of a class of
centro-symmetric linear arrays made of pairs of sensors symmetrically located along the two sides of
the linear antenna array. Attractive features of such centro-symmetric linear arrays include lower DOA
and range CRBs and faster convergence to the lower far-field DOA CRB. Also, we show that within
centro-symmetric linear arrays, ULA is not the best choice. Centro-symmetric linear arrays can be found
that have identical DOA CRB as the ULA but significantly lower range CRB (by as much as 50%).

1
Constraints on the array positions, often adopted to limit array ambiguities [8], may, at the same time, affect estimation
performance [9].

December 7, 2013 DRAFT

 3

In fact, a geometric parameter is identified that controls the near-field estimation performance of the
centro-symmetric linear array.
The paper is organized as follows. Sec. II formulates the problem and specifies the data model. Sec. III
is dedicated to new expressions of the CRB. First, assuming an arbitrary parametrization of the steering
vector, we prove that the conditional and unconditional CRB are proportional. Then, we focus on the
DOA and range estimation of near-field sources. Using Taylor expansion, new expressions of the CRB are
derived, and numerically validated. The important class of centro-symmetric arrays is studied in details
in Sec. IV where better-than-ULA are obtained. This paper is concluded in Sec. V.

II. DATA M ODEL

θ h Source


R
r

x x x x x x- x
C1 C2 O Cp CP
xp -

Fig. 1. Source in the near-field of the arbitrary linear array.

As depicted in Fig. 1, we consider a linear antenna array made of P sensors C1 , · · · , CP . They

are located along a straight line at coordinates x1 , · · · , xP , respectively. Without loss of generality, we
assume the array centroid to be at the origin O of this axis. This choice will also allow for more compact
expressions of the CRB, compared to [4], [5]. A narrow-band signal s(t), with wavelength λ, is emitted
towards the antenna array by a source S located at a range r from the origin O and forming an angle θ
w.r.t. the normal of the array. The snapshot collected by sensor p at time index t is

yp (t) = gp exp (iτp ) s(t) + np (t), (1)

where s(t) and np (t) represent, respectively, the source signal collected at the origin and the ambient
additive noise collected by sensor p. Amplitude gp may depend on both r and θ, while τp is defined as
τp = 2π(SO − SCp )/λ, which can be rewritten as
r( √ )
τp = 2π 1 − βp
λ
def x2p
with βp = 1 − 2 xrp sin(θ) + r2 . Based on N snapshots {yp (t)}p=1,...,P ;t=t1 ,...,tN , estimates of both the
range r and the DOA θ are obtained using a variety of algorithms, among which a few are capable of
achieving the stochastic CRB [10].

December 7, 2013 DRAFT

 4

Estimation accuracy is evaluated in terms of the CRB, for which the usual statistical properties about
np (t) and s(t) are the following: (i) np (t) and s(t) are independent, (ii) {np (t)}p=1,...,P ;t=t1 ,...,tN are
independent, zero-mean circular Gaussian distributed with variance σn2 , (iii) {s(t)}t=t1 ,...,tN are assumed
to be either deterministic unknown parameters (the so-called conditional or deterministic model), or
independent zero-mean circular Gaussian distributed with variance σs2 (the so-called unconditional or
stochastic model).

III. E XPRESSIONS OF THE CRB

A. A general result about the CRB

We prove that the stochastic and deterministic CRBs are equal, up to a multiplicative constant, in the
specific case of a single source. Let’s, first, consider an arbitrary number K of sources (with K < P ) with
an arbitrary parametrization α = [α1 , ...., αL ]T of the steering vectors a(α) related to array geometry or
polarization, defined by

[a(α)]p=1,..,P = gp eiτp ,

where gp and τp denote the gain and the delay of the p-th sensor w.r.t. the origin O. gp includes in
particular possible power profiles and/or directional gains. General compact expressions of the CRB,
concentrated on the parameters of the K steering vectors alone, have been derived for these two models
of sources (see e.g., [13]) for one parameter per source. The expression of the stochastic CRB has been
extended for several parameters per source in [14, Appendix D], and following the proof given in [13],
the expression of the deterministic CRB can be also extended to several parameters per source. These
expressions are given respectively by

σn2 [ ( )]−1
CRBsto (α) = Re H ⊙ ((Rs AH R−1
y ARs ) ⊗ 1L )
T
(2)
2N
σn2 [ ( )]−1
CRBdet (α) = Re H ⊙ (RTs ⊗ 1L ) , (3)
2N
def def def def
where Ry = E[y(t)yH (t)] with y(t) = [y1 (t), ..., yP (t)]T , Rs = E[s(t)sH (t)] with s(t) =
def ∑ def
[s1 (t), ..., sK (t)]T in (2) and Rs = N1 N H
n=1 s(tn )s (tn )] in (3), A = [a1 , ..., aK ] with ak is the
def [ ]
steering vector of the k -th source parameterized by αk ∈ RL , H = DH I − A(AH A)−1 AH D and
defda1
D = [ dα 1
daK
, ..., dαK
], and where ⊗, ⊙ and 1L are the Kronecker product, the Hadamard product and the
L × L matrix of 1s, respectively.

December 7, 2013 DRAFT

 5

def 1 ∑N
Specialized to a single source for which Ry = σs2 a(α)aH (α) + σn2 I where σs2 = n=1 |s(tn ) |
2
N

for the deterministic model of the source, it is straightforward to prove the following result
( )
σn2
CRBsto (α) = 1 + CRBdet (α) = [F(α)]−1 (4)
∥a(α)∥2 σs2
where
[ ]
F(α) = cσ (α)Re ∥a(α)∥2 DH (α)D(α) − DH (α)a(α)aH (α)D(α)
[ ]
def def 2N σs4
with D(α) = ∂a(α)
∂α1 , ..., ∂a(α)
∂αL and where cσ (α) = σ2 (σ2 +∥a(α)∥ 2 σ 2 ) is independent of the source
n n s

and sensors positions for constant modulus steering vectors, only.

Thanks to (4), we only consider the stochastic source model for which the elements of the matrix F
(??) are given by
     
[F]i,j ∑
P ∑P ∑
P ∑
P
=  gp2   ′
gp,i ′
gp,j ′
+ τp,i ′
τp,j gp2  −  ′
gp,i gp   ′
gp,j gp 
cσ (α)
p=1 p=1 p=1 p=1
  
∑
P ∑
P
−  ′
τp,i gp2   ′
τp,j gp2  (6)
p=1 p=1
′ and τ ′ are the derivative of g and τ w.r.t. α , respectively.
where gp,i p,i p p i

This expression simplifies when the gain gp does not depend on the sensor (gp = g(α)) for which the
following expression holds:
  
[F]i,j ∑
P ∑
P ∑
P
= P ′
τp,i ′
τp,j − ′ 
τp,i ′ 
τp,j (7)
c′σ (α)
p=1 p=1 p=1

def 2N σs4 g 4 (α)

where now c′σ (α) = 2 (σ 2 +P g 2 (α)σ 2 ) .
σn Consequently the CRB (CRBDIR (α)) associated with directional
n s

sensors (with identical orientations and power profiles) is related to the CRB (CRBISO (α)) associated
with isotropic (i.e., omnidirectional with no power profile, gp = 1) sensors by the following relation:
2
2 σs
DIR 1 1 + P g (α) σn2
CRB (α) = 4 2 CRBISO (α). (8)
g (α) 1 + P σ2s
σn
Note that this theoretical relation does no apply in practice for directional sensors in the near filed
because...

B. Taylor expansion of the near-field matrix F

In the addressed problem, α1 = θ and α2 = r. To obtain interpretable closed-form expressions of

the CRB on θ and r, we consider a unitary modulus gain for each sensor,2 as often assumed in the

2
This condition, stronger than ∥a(α)∥2 = P can be extended to arbitrary common gain g(α) thanks to (8).

December 7, 2013 DRAFT

 6

2N σs4
CRB literature, for which c′σ (α) simplifies to cσ = 2 (σ 2 +P σ 2 ) ,
σn independent of the source and sensors
n s

positions.
The following Taylor expansion of the matrix F (7) is proved in Appendix A:
2c S3 3P S4 − S22
[F]1,2 = P + sin(θ) + o(ϵ4 ) (9)
r cos3 (θ) r3 r4
[ ]
c S2 S3 S4 P 4 sin2 (θ) − 1 − S22 sin2 (θ)
[F]1,1 = P 2 + 2P sin(θ) 3 + + o(ϵ4 ), (10)
r2 cos2 (θ) r r r4
c 1 P S4 − S22 P S5 − S2 S3
[F]2,2 = + sin(θ)
cos4 (θ) 4 r4 r5
[ ] [ ]
P S6 23 sin2 (θ) − 3 − 3S2 S4 5 sin2 (θ) − 1 − 8S32 sin2 (θ)
+ + o(ϵ6 ), (11)
8r6
def def def ∑
where ϵ = 1r maxp |xp |, limϵ→0 o(ϵ)/ϵ = 0, c = 4πλ2 cσ and Sk = Pp=1 xkp are array geometry dependent
2

constants, with, in particular, S1 = 0.

C. Taylor expansion of the near-field CRB

From the expression (9) of the matrix F, we see that DOA and range are decoupled to the second-order
in ϵ if and only if S3 = 0. Consequently, the DOA and range estimates given by any efficient algorithm
are uncorrelated to the second order in ϵ, if and only if S3 = 0. This special, yet important case, will be
studied in Sec. IV. For the moment, we give results about the general case of antenna arrays for which
S3 ̸= 0.
From the matrix F given by (9), (10) and (11), the following expressions of the CRB on the DOA
and range are proved in Appendix B.

c 1 1 + γ1 sin(θ)
r
CRB(θ) = 2 + o(ϵ), (12)
P S2 − P S3 2 cos2 (θ)
P S4 −S 2

CRB(r) 4c 1 + γ2 sin(θ)
r
= 2 + o(ϵ), (13)
r4 P S4 − S22 − P SS32 cos4 (θ)

3 −S2 S4 +P S4 −P S3 S5 3P S2 S3 S4 +S2 S3 −P S3 −2P S2 S5

def 2 2 2 def 3 3 2
where γ1 = 4P S3 (PSS2 S4 −S 2
)(P S2 S4 −S 3 −P S 2 ) and γ2 = 2 S2 (P S2 S4 −S 3 −P S 2 ) . Both depend on
2 2 3 2 3

S2 , S3 , S4 and S5 , but not on S6 .

D. The case of centro-symmetric arrays

Note that if S3 = 0 [resp., if S3 = S5 = 0], the expression (12) [resp., (13)] is still valid. However,
the term in 1/r in (12) of CRB(θ) [resp., in (13) of CRB(r)] vanishes. This scenario is far from being
marginal, as it notably includes the ULA. Specific results are developed to cover such arrays. We will

December 7, 2013 DRAFT

 7

discuss, in particular, the so-called centro-symmetric arrays, ones for which if a sensor is placed at some
position xp , then another one is placed at coordinate −xp .
First, we prove that for antenna arrays with S3 = 0 we have
[ ( ) ]
1 + 1 + 1 + P 4P S4
S4 −S22 sin2
(θ) S4 1
S2 r2
CRB(θ) = c + o(ϵ2 ). (14)
cos2 (θ)P S2
Second, for antenna arrays that satisfy both S3 = 0 and S5 = 0, we prove that
[ ]
CRB(r) 1 4 γ3 (θ)
=c 4 1+ + o(ϵ2 ), (15)
r4 cos (θ) P S4 − S22 2(P S4 − S22 )r2
def 18P 2 S42 +2S24 +3P S22 S4 −23P 2 S2 S6
where γ3 (θ) = P S2 sin2 (θ) + 3P S6 − 3S2 S4 . The proofs are summarized in
Appendix C.

E. Numerical validation

Let us, first, highlight similarities between the different obtained CRBs (12), (13), (14) and (15). For
def 1 def 1
this purpose we define C1 = c CRB(θ) and C2 = cr4 CRB(r), which also have the advantage of not
depending on the noise and signal power, nor on the signal wavelength. They are purely geometrical
functions of alone sensors and source positions.
For arbitrary linear arrays, (12) and (13) can be rewritten using the unique expression
(1) [ ]
Ti (2) sin(θ)
Ci = 1 + Ti + o(ϵ) (16)
cos2i (θ) r
(1) (2)
Expressions of constants Ti and Ti can be easily found and depend only on the array sensor positions.
For centro-symmetric arrays (more explicitly, for arrays satisfying S3 = 0 for C1 and S3 = S5 = 0 for
C2 ), unified expressions can be found as well. Indeed, (14) and (15) are rewritten as a unique expression
[ ]
Ti
(3) ( ) T (6)
(4) (5)
Ci = 1 + Ti sin2 (θ) + Ti i
+ o(ϵ2 ) (17)
cos2i (θ) r2
(3) (4)
where constants Ti and Ti can be easily verified to depend only on the array sensors positions. We
intend to validate every single coefficient in the Taylor expansions in (12), (13), (14) and (15).
cos2i (θ)Ci
(1) −1
def 2i def
ρ1,i = cos (1)
(θ)Ci T
First, for arbitrary linear arrays, for i = 1, 2, we define and ρ2,i = r i
(2) . All
Ti T i sin(θ)
converge to 1 when 1/ϵ = r/ maxp |xp | converges to infinity. This is verified in Fig. 2(a) that represents
a randomly chosen linear array (for which S2 = 0.3162 and S3 = 0.0904) and source.
def cos2i (θ)Ci def
Second, for centro-symmetric arrays, for i = 1, 2, we introduce ρ′1,i = (3) and ρ′2,i =
Ti
cos2i (θ)Ci
(3) −1
, which, also, converge to 1 when 1/ϵ = r/ maxp |xp | converges to infinity. This
T
r2 i
(T sin (θ)+T )Ti(6)
(4)
i
2 (5)
i

is confirmed by the numerical evaluations for the 6 sensors ULA, reported in Fig. 2(b).

December 7, 2013 DRAFT

 8

IV. A NALYSIS OF C ENTRO -S YMMETRIC A RRAYS

A. Centro-symmetric vs. non-centro-symmetric arrays

CRB expressions (12)-(13) as opposed to CRB expressions (14)-(15) suggest that there are two classes
of antenna arrays with different geometrical properties and estimation performance. In particular, we
are interested in the so-called centro-symmetric arrays because they have a better far-field estimation
performance. To highlight this fact, we connect our near-field DOA CRB (12) and (14) to the stochastic
far-field DOA CRB. The latter is given, for arbitrary linear arrays, by [7, rel. (5)]
( )
1 λ2 1 c
CRBFF (θ) = S σ 2 1 + σ2 = 2
. (18)
N 8π 2 cos2 (θ) 2 2s P 2s cos (θ)P S2
P σn σn

Normalized to the above, our near-field DOA CRBs (12) and (14) lead to, respectively,
[ ]
CRB(θ) 1 γ1 sin(θ)
= 2 1+ + o(ϵ) , (19)
CRBFF (θ) 1 − P S2PSS43−S 3 r
[ 2
( ) ]
CRB(θ) 2 4P S4 S4 1
= 1 + sin (θ) 1 + +1 + o(ϵ2 ). (20)
CRBFF (θ) P S 4 − S2
2 S2 r 2
From (20), we see that arrays for which S3 = 0 (e.g., for centro-symmetric arrays) do achieve CRBFF (θ)
when the source-to-array distance tends to infinity. At the same time, estimation of θ and r are decoupled
in the matrix F to the second-order in ϵ. In contrast, non-centro-symmetric arrays in (19), for which
S3 ̸= 0, verify limr→∞ CRB(θ) > CRBFF (θ) because P S4 − S22 > 0 (see Sec. IV-C). This strange
behavior is explained by the coupling between θ and r in the matrix F to the second-order in ϵ [see (9)].
More precisely, in the former case, the square of [F]1,2 tends to zero more rapidly than [F]2,2 when r
tends to ∞, in contrast to the latter case for which the square of [F]1,2 and the term [F]2,2 tend to zero
with the same speed. Consequently, from a practical point of view, as far as only the DOA parameter
is considered, the far-field model of propagation, although approximative may be preferable to the exact
near-field model for non centro-symmetric arrays with S3 ̸= 0.
If we take the range into consideration, the domain of validity of our approximations is larger for
centro-symmetric arrays than for arbitrary arrays, as a result of a convergence in 1/r2 compared to 1/r.
Furthermore, when comparing (14) (15) to (12) (13), we realize that, for centro-symmetric arrays, the
CRBs are symmetric w.r.t. positive/negative θ. However, for arbitrary arrays, they are not.
To illustrate the different behavior of centro-symmetric and non centro-symmetric arrays in the near-
field region, we test in Fig. 3 antenna arrays of P = 4 sensors forming either (i) a ULA with a constant
inter-sensors spacing d and for which S3 = S5 = 0, or (ii) a minimum hole and redundancy linear array
(MHRLA) with inter-spacings d, 3d, 2d [12] and for which S3 ̸= 0. Thanks to a larger aperture, the

December 7, 2013 DRAFT

 9

MHRLA exhibits a lower far-field CRB, for instance, CRBMHRLA

FF FF (θ) ≈ 0.22. However,
(θ)/CRBULA
due to the coupling of θ and r in the matrix F of the MHRLA, we have limr→∞ CRB(θ) > CRBFF (θ)
for this array. Furthermore, this figure confirms that the domain of validity of our approximations is much
larger for the centro-symmetric than for non centro-symmetric arrays.
Finally, notice that because we fix the time reference at the centroid of the array (and not at the left-end
as in [5]), we obtain simple and much easier to interpret closed-form CRB expressions that contrast with
the intricate expressions [5, (10-11)] that are valid for the only ULA. In particular, we note the monotone
behavior of CRB(θ) w.r.t. r and the symmetry of CRB(θ) and CRB(r) w.r.t. positive/negative θ with a
minimum for θ = 0. Also notice that, due to the change of time reference, our definition of the couple
(θ, r) is different from the one in [5] (it is, actually, significantly different if the source is in the very
near-field region).

B. Conditions of centro-symmetry

By centro-symmetric, we mean that the array is made of pairs of sensors placed at opposite coordinates,
i.e., if a sensor is placed at xp , then another one is placed at −xp . A sensor may be placed at the origin
and, then, P is odd. We determine that an array is centro-symmetric iff
Si = 0 for all i odd and less or equal to P
This is proved by induction (see details in Appendix D) thanks to the Newton-Girard formula [16, pp.
69-74] that allows one to calculate the different Si in an iterative manner.

C. Key geometric parameters for near-field performance

The rewriting of (12-15) allows us to point out two geometric parameters that shape the near-field
accuracy of antenna arrays. They are the unit-less

def S22 def S23

κ = and η =
P S4 P 2 S6
which remain unchanged if either a sensor is added/removed at/from the origin, or, more importantly, if
sensor coordinates are scaled by the same constant3 . Before showing where κ and η appear in the CRB
expressions and how they impact them, we highlight some of their intrinsic properties. First, we prove
in Appendix E that
η ≤ κ ≤ 1. (21)

3
This property is useful to conduct an affordable systematic search of linear arrays under the non-restrictive condition S2 = 1.

December 7, 2013 DRAFT

 10

Very interesting is the fact that there is (almost) a one-to-one correspondence between κ and η . In fact,
we prove in Appendix F that
def 1
For P = 4 : η = η1 (κ) = if S1 = S3 = 0 (22)
3
κ −2
def 4
For P = 5 : η = η2 (κ) = (3 ) if S1 = S3 = S5 = 0. (23)
5 κ − 52
For larger P , the two functions [especially η2 (κ)] provide good approximations of the exact η , as validated
by Fig. 4.

D. CRB in terms of κ and η

The two CRBs in (14) and (15) can now be rewritten as follows
[( ) ]
1 1 4 2 1
S2 + κ 1 + 1−κ sin (θ) + 1 P r2
CRB(θ) = c + o(ϵ2 ) (24)
cos2 (θ)P
 ( ) 
18 + 3κ + 2κ 2 − 23 sin2 (θ) + 3 κ2 −3κ
CRB(r) 1 4  1 S2 η η
 +o(ϵ2 ). (25)
= c 4 +
r4 cos (θ) S22 κ1 − 1 2P r2 (1 − κ)2

It becomes clear that while the array far-field (DOA estimation) performance is determined by S2 only
[which concurs with (18)], κ and η play a role in the (DOA and range estimation) near-field performance.
If S2 is fixed (which has no impact on κ and η ), and if we consider the most significant terms4 of,
respectively, (25) and (24), then it becomes clear that the array estimation performance is controlled by
κ through, respectively, (i) 1/(1/κ − 1) (an increasing function of κ) and (ii) function fθ (κ) defined as
[( ) ]
def 1 4 2
fθ (κ) = 1+ sin (θ) + 1
κ 1−κ
def
= f1 (κ) sin2 (θ) + f2 (κ)

The behavior of fθ (κ), illustrated in Fig. 5, suggests that, for DOA estimation, an antenna with κ loosely
close to 1/2 ensures limited degradation in all look directions. Values of κ close to, but lower than, 1/2
are preferred however, because they also lead to better estimation of the range parameter.

E. Comparison with ULA

The present analysis shows that, if the source is in the near-field of the linear antenna, then placing the
sensors at a regular spacing will not ensure the best performance. For instance, we prove in Appendix G

4 S2
The first term is the most significant term in (25), for all configurations where 2P r 2
< 0.01, which covers a large domain
of practical situations.

December 7, 2013 DRAFT

 11

that, in the case of ULA, κ tends to 5/9 (and η converges to 7/27) if the number of sensors increases
to infinity, which, by the way, leads to the following refinements of (24) and (25)
{ ( 2 2 )}
3P 2 d2 [ ] P d
CRB(θ) ≈ CRBFF (θ) 1 + 2
1 + 10 sin (θ) + o , (26)
20r2 r2
{ [ ] ( 2 2 )}
CRB(r) 720c 9P 2 d2 970 P d
≈ 1+ 1− 2
sin (θ) + o , (27)
r4 P 6 d4 cos4 (θ) 28r2 27 r2

where d denotes the spacing between two consecutive sensors.

From the discussion in Sec. IV-D, a (centro-symmetric) linear antenna with such a value of κ has near-
optimum performance for DOA estimation but not for range estimation. To better illustrate the impact of
κ on the estimation performance (of both DOA and range), we compare the 6-sensors ULA (with sensors
placed at ±0.1195, ±0.3586 and ±0.5976) against a non-ULA array of 6 sensors located at ±0.0674,
±0.2023 and ±0.6742. Both arrays exhibit the same S2 = 1 (and, hence, have identical far-field DOA
estimation CRBs). However, κ is equal to 0.5776 for the ULA and to 0.4 for the non-ULA. In Fig.
CRB(θ)|non−ULA CRB(r)|non−ULA
6, we report the ratios CRB(θ)|ULA and CRB(r)|ULA , calculated using the exact CRB expressions
and the approximate CRB expressions in (24) and (25). There, we can see that while we obtain similar
DOA performance, the non-ULA array has better range estimation capabilities. Hence, within the family
of centro-symmetric linear arrays characterized by a given size P and a given value of S2 , all verify
CRB(θ)|non−ULA
limr→∞ CRB(θ)|ULA = 1, from (24). However, from (25), the κ-dependent function

κULA − 1
1
def CRB(r)|non−ULA
RP (κ) = lim =
κ −1
r→∞ CRB(r)|ULA 1

can be seen as an indicator of improvement (over the ULA) whenever it is lower than one. For instance,
if P ≫ 1,
4 1
RP (κ) = .
5 1
κ −1

The above ratio is illustrated in Fig. 7 for the domain5 [0.3, 0.7] of κ outside which DOA near-field
performance degrades severely (as clear from Fig. 5(c)). It can be seen from Fig. 7 that the (far-field)
range CRB can be reduced by a much as 50% by antenna arrays with a κ moderately lower than that of
the ULA.

5
In fact, extreme values of κ (i.e. 0 and 1) are achieved by impractically co-localized sensors, either at the origin, or at the
same distance (and on both sides) from the origin.

December 7, 2013 DRAFT

 12

V. C ONCLUSION

Assuming no constraints other than sensors deployed along a straight line and using the exact expression
of the time delay parameter, accurate, simple and interpretable closed-form CRB expressions have been
obtained for both angle and range parameters of a near-field narrow-band source.

They show the exact geometric condition for the antenna array to have an attractive behavior in its
near-field: better precision and faster convergence to the lower far-field DOA CRB. Such a class of centro-
symmetric arrays includes, but is not restricted to, ULAs. Furthermore, it is proved that appropriately
designed centro-symmetric non-ULA can largely improve the range estimates without deteriorating the
DOA estimates under near-field conditions. Because they potentially have better estimation performance,
non-ULAs geometries may be adopted when array ambiguity can be tolerated or counter-measures are
made available [7]. Hence, our analysis gives a deeper insight into the array near-field performance and
shows that more flexibility is available for array design.

A PPENDIX

A. Taylor expansion of the matrix F: Proof of (9), (10) and (11)

def
The main steps of the proof are as follows. First, note that with ϵp = xp /r, we have

λ ∑ ′ ∑ √
P P
τp,1 = cos(θ)rϵp / βp ,
2π
p=1 p=1

λ ∑ ′ ∑ ∑
P P P
ϵp 1
τp,2 = P + sin(θ) √ − √ ,
2π βp βp
p=1 p=1 p=1

λ2 ∑ ∑
P P
′
ϵ2p
(τp,1 )2 = r2 cos2 (θ) ,
4π 2 βp
p=1 p=1

λ2 ∑ ∑ ∑
P P P
′
+ 1 − 2ϵp sin(θ)
ϵ2p sin2 (θ) ϵp sin(θ) − 1
(τp,2 )2 = P+ +2 √
4π 2 βp βp
p=1 p=1 p=1
( )
λ2 ∑ ′ ′ ∑
P P
1 ϵp sin(θ) − 1
τp,1 τp,2 = r cos(θ) ϵp √ + .
4π 2 βp βp
p=1 p=1

December 7, 2013 DRAFT

 13

√
All these sums appear to involve either 1/β or 1/ β , whose Taylor expansions are obtained subsequently:
1
= 1 + 2ϵp sin(θ) + (4 sin2 (θ) − 1)ϵ2p − 4 sin(θ) cos(2θ)ϵ3p + (1 − 12 sin2 (θ) + 16 sin4 (θ))ϵ4p
βp
+ sin(θ)(6 − 32 sin2 (θ) + 32 sin4 (θ))ϵ5p + [−1 + 24 sin2 (θ)1 − 80 sin4 (θ) + 64 sin6 (θ)]ϵ6p + o(ϵ6p ),
1 3 sin2 (θ) − 1 2 5 sin2 (θ) − 3 3 3 − 30 sin2 (θ) + 35 sin4 (θ) 4
√ = 1 + ϵp sin(θ) + ϵp + sin(θ) ϵp + ϵp
βp 2 2 8
15 − 70 sin2 (θ) + 63 sin4 (θ) 5 −5 + 105 sin2 (θ) − 315 sin4 (θ) + 231 sin6 (θ) 6
+ sin(θ) ϵp + ϵp + o(ϵ6p ).
8 16
The above expansions are used to obtain Taylor expansion of the different sums appearing in the right
hand side of (7). After tedious manipulations, (9), (10) and (11) are obtained in similar fashions.

B. Taylor expansion of the CRB for arbitrary arrays: Proof of (12) and (13)

First, note that by replacing r by maxp |xp |/ϵ in the [F]i,j terms (9-11), the matrix F can be written
in the following form:
 
b1,1 1,1 1,1 2 2
0 + b1 ϵ + b2 ϵ + o(ϵ ) b1,2 2 1,2 3 3
2 ϵ + b3 ϵ + o(ϵ )
F= ,
b1,2 2 1,2 3 3
2 ϵ + b3 ϵ + o(ϵ ) ϵ4 [b2,2 2,2 2,2 2 2
4 + b5 ϵ + b6 ϵ + o(ϵ )]

cos2 (θ)
where for e.g., b01,1 = c P S2 .
This allows one to obtain, after straightforward algebraic manipulations,
[ ( ) ]
1 b2,2 b1,1 2,2
b + b 1,1 2,2
b − 2b 1,2 1,2
b
CRB(θ) = [F−1 ]1,1 = 1+ϵ 2,2 −
5 0 5
1,1 2,2
1 4
1,1 2
2 3
+ o(ϵ) ,
1,1 (b21,1 )2 −
b0 − b2,2 b 4 b0 4 b (b 2 )
   
4

b52,2 b11,1 2b21,2 b31,2

2,2 + 1,1 −
1,1
1 b 1,1 2,2
CRB(r) = [F−1 ]2,2 = ( ) 1 + ϵ  1,11
− 4
b b0
1,1
b0 b4 
+ o(ϵ) .
2,2 (b1,1
2 )
2 b
ϵ b4 1 − b1,1 b2,2
4 b0 1 − b1,1 b2,2
2

0 4 0 4

By replacing the different terms bi,j

k by their respective values, and after simple but tedious manipulations,
we ultimately prove (12) and (13).

C. Taylor expansion of the CRB for centro-symmetric arrays: Proof of (14) and (15)

We use the same approach as in Sec. B. On one hand, we have, for S3 = 0,

 
1,1 1,1 2 2 1,2 3 3
b0 + b2 ϵ + o(ϵ ) b3 ϵ + o(ϵ )
F= .
1,2 3 3 4 2,2 2,2 2,2 2 2
b3 ϵ + o(ϵ ) ϵ [b4 + b5 ϵ + b6 ϵ + o(ϵ )]

which allows to conclude that

[ ( ) ]
−1 1 b1,1 (b1,2
3 )
2
CRB(θ) = [F ]1,1 = 1−ϵ 2 2
− + o(ϵ ) .2
b1,1
0 b1,1
0 b1,1
0 b4
2,2

December 7, 2013 DRAFT

 14

On the other hand, when both S3 = S5 = 0, we have

 
b1,1 + b1,1 2
ϵ + o(ϵ 2) b 1,2 3
ϵ + o(ϵ 3)
F= ,
0 2 3
1,2 3
b3 ϵ + o(ϵ ) 3 ϵ [b4 + b2,2
4 2,2
6 ϵ 2 + o(ϵ2 )]

which eventually leads to

[ ( ) ]
−1 1 b2,2 (b1,3
3 )
2
CRB(r) = [F ]2,2 = 1−ϵ 2 6
− 2
+ o(ϵ ) .
ϵ4 b2,2
4 b2,2
4 b1,1
0 b4
2,2

Replacing the different terms bi,j

k by their values, (14) and (15) are also proved after tedious manipulations.

D. Proof of the condition of centro-symmetry of Sec. IV-B

def
We consider real numbers x1 , · · · , xP and form the polynomial Q(x) = (x − x1 ) · · · (x −
def def
xP ) = xP − σ1 xP −1 + σ2 xP −2 + · · · + (−1)P σP . The coefficients σ1 , · · · , σP , given by σk =
∑ ∑P
1≤i1 <i2 <···<ik ≤P xi1 xi2 · · · xik , are known to be linked to S1 , · · · , SP defined as Sk =
k
p=1 xp by

means of the Newton-Girard formula [16, pp. 69-74]

∑
k−1
Sk = (−1)l−1 σl Sk−l + (−1)k−1 kσk , k ≥ 2 (28)
l=1
where by definition σP +1 = σP +2 = · · · = 0. Let 1, 3, · · · , 2I + 1 be all the odd integers ≤ P . Let’s
assume that S1 = S3 = · · · = S2I+1 = 0. We will prove that σ1 = σ3 = · · · = σ2I+1 = 0. We proceed
by induction to show that σ2i+1 = 0 for i = 0, 1, · · · , I . This is already verified for i = 0 because
σ1 = S1 . Let’s assume σ1 = σ3 = · · · = σ2i+1 = 0 for some i ≤ I , and let’s prove that σ2i+3 = 0.
∑
From (28), we have S2i+3 = 2i+2
l=1 (−1)
l−1 σ S
l 2i+3−l + (−1)
2i+2 (2i + 3)σ
2i+3 is necessarily zero. In

fact, 2i + 3 is odd, so that if l is odd, then 2i + 3 − l is even and vice versa, for l = 1, · · · , 2i + 2. Also,
l and 2i + 3 − l both are ≤ 2i + 1 and whenever one is odd, the corresponding σ and S coefficients are
zero. Hence, for l = 1, · · · , 2i + 2, we have necessarily σl S2i+3−l = 0; and so is σ2i+3 . Finally, Q(x)
is either xP + σ2 xP −2 + σ4 xP −4 + · · · + σP if P is even or xP + σ2 xP −2 + σ4 xP −4 + · · · + σP −1 x if
P is odd. In the first (resp. second) case, zeros of Q(x), i.e. x1 , · · · , xP , are of the form ±α1 , ±α2 , · · ·
(resp. 0, ±α1 , ±α2 , · · · ).

E. Proof of inequalities (21)

∑ ∑
For arbitrarily chosen a1 ≤ a2 ≤ · · · ≤ aP and b1 ≤ b2 ≤ · · · ≤ bP , we have Pi=1 Pj=1 (ai −
∑ (∑ ) (∑ )
aj )(bi − bj ) ≥ 0 which implies that P Pi=1 ai bi ≥ P
i=1 ai
P
i=1 bi , the so-called Tchebychev’s
def def def def
sum inequality. If we let ai = bi = x2i , we obtain P S4 ≥ S22 i.e. κ ≤ 1. If we let ai = x2i and bi = x4i ,
we obtain P S6 ≥ S2 S4 i.e. η ≤ κ.

December 7, 2013 DRAFT

 15

F. Proof of (22) and (23)

By virtue of (28), we have S2 = σ1 S1 − 2σ2 , S4 = σ1 S3 − σ2 S2 + σ3 S1 − 4σ4 and S6 = σ1 S5 −

σ2 S4 + σ3 S3 − σ4 S2 + σ5 S1 − 6σ6 . If S1 = S3 = S5 = 0 (if P = 4, only if S1 = S3 = 0, because
this is sufficient to have a centro-symmetric array and automatically implies S5 = 0), and after proper
S23
replacement, we obtain S6 = 3S2 S4
4 − 8 − 6σ6 which can be transformed into 3
κ = 4 1
P η + P
2 + 24P Sσ63
2

By definition, σ6 = 0 if P < 6 and the following becomes clearly justified

3 1
P = 4, S1 = S3 = 0 → = +2
κ η
3 41 5
P = 5, S1 = S3 = S5 = 0 → = + .
κ 5η 2

G. κ and η for large-sized ULAs

Consider a ULA centered at the origin made of P sensors spaced by d. The proof is given for odd
∑
P = 2Q + 1 (extension to even P can be conducted in a similar way). Using the identities of Qk=1 k
α

for α = 2, 4 and 6, we obtain

∑
Q
Q(Q + 1)(2Q + 1)
S2 = 2d2 k 2 = d2
3
k=1

∑
Q
3Q2 + 3Q − 1
S4 = 2d4 k 4 = d2 S2
5
k=1

∑
Q
3Q4 + 6Q3 − 3Q + 1
6
S6 = 2d k 6 = d4 S2 ,
7
k=1

which directly implies

S22 5 Q(Q + 1) S23 7 Q2 (Q + 1)2
κ= = and η= = .
P S4 3 3Q2 + 3Q − 1 P 2 S6 9 3Q4 + 6Q3 − 3Q + 1

R EFERENCES

[1] H. Krim and M. Viberg, “Two decades of array signal processing research,” IEEE Signal Process. Mag., vol. 13, no. 4, pp.
67-94, Jul. 1996.
[2] H. Gazzah and S. Marcos, “Cramer-Rao bounds for antenna array design,” IEEE Trans. Signal Process., vol. 54, no. 1, pp.
336-345, Jan. 2006.
[3] E. Grosicki, K. Abed-Meraim, and Y. Hua, ”A weighted linear prediction method for near field source localization,” IEEE
Trans. Signal Process., vol. 53, no. 10, pp. 3651-3660, Oct. 2005.

December 7, 2013 DRAFT

 16

[4] M. N. El Korso, R. Boyer, A. Renaux, and S. Marcos, ”Conditional and unconditional Cramer Rao bounds for near-field
Source localization,” IEEE Trans. Signal Process., vol. 58, no. 5, pp. 2901-2906, May 2010.
[5] Y. Begriche, M. Thameri, and K. Abed-Meraim, ”Exact Cramer Rao bound for near field source localization,” in Proc.
International Conference on Information Science, Signal Processing and their Applications, 2012, pp. 718-721.
[6] Y. Begriche, M. Thameri, and K. Abed-Meraim, ”Exact conditional and unconditional Cramer Rao bound for near field
localization,” in arXiv.
[7] C. ElKassis, J. Picheral, and C. Mokbel, “Advantages of nonuniform arrays using root-MUSIC”, Signal Processing, Elsevier,
vol. 90 (210), pp. 689-695.
[8] M. Gavish and A. J. Weiss, “Array geometry for ambiguity resolution in direction finding,” IEEE Trans. Antennas Propagat.,
vol. 39, pp. 143-146, Feb. 1991.
[9] A. J. Weiss, A. S. Willsky, and B. C. Levy, “Non uniform array processing via the polynomial approach,” IEEE Trans.
Aerosp. Electron. Syst., vol. 25, no. 1, pp. 48-55, 1989.
[10] H. Gazzah and J.-P. Delmas, “Spectral efficiency of beamforming-based parameter estimation in the single source case,”
in Proc. IEEE SSP, Nice, pp. 153-156, 2011.
[11] J.-P. Delmas and H. Gazzah, “CRB Analysis of near-field source localization using uniform circular arrays,” in Proc.
ICASSP, Vancouver pp. 3396-3400, 2013.
[12] Y. Meurisse and J.P. Delmas, ”Bounds for sparse planar and volume arrays,” IEEE Trans. Inform. Theory, vol. 47, no. 1,
pp. 464-468, January 2001.
[13] P. Stoica and A. Nehorai, ”Performances study of conditional and unconditional direction of arrival estimation,” IEEE
Trans. ASSP, vol. 38, no. 10, pp. 1783-1795, Oct. 1990.
[14] H. Abeida, J.-P. Delmas, ”Efficiency of subspace-based DOA estimators,” Signal Processing, vol. 87, no. 9, pp. 2075-2084,
Sept. 2007.
[15] D.T. Vu, A. Renaux, R. Boyer, and S. Marcos, ”A Cramer Rao bounds based analysis of 3D antenna array geometries
made from ULA branches,” Multidimensional Systems and Signal Processing, Springer, Oct. 2011.
[16] L. E. Dickson, “Elementary Theory of Equations,” New York, Wiley and Sons, 1914.

December 7, 2013 DRAFT

 17

1.5
ρ
1,1
ρ1,2
ρ
2,1
ρ2,2

0.5

0
5 10 15 20 25 30 35 40 45
r / maxP(|xp|)

(a)

1.5

0.5

ρ’1,1
ρ’1,2
ρ’2,1
ρ’2,2
0
5 10 15 20 25 30 35 40 45
r / maxP(|xp|)

(b)

Fig. 2. Validation of the DOA and range CRBs for an increasing source-to-array distance. The array is made of
6 sensors. In (a), they are placed at −0.1672, −0.0737, 0.0960, −0.1444, 0.4694 and −0.1801, forming a non
centro-symmetric array. In (b), they form a ULA. The source is placed, in (a), with θ = 47.13◦ ; and, in (b), with
θ = 146.64◦ .

December 7, 2013 DRAFT

 18

2.5
Exact
Approximate
ULA
MHRLA

Normalized DOA CRB

1.5

0.5
2 4 6 8 10 12
r / maxp(|xp|)

Fig. 3. Approximative and exact ratios CRB(θ)/CRBFF (θ) for 4-sensors ULA and MHRLA and a source at
θ = 60◦ . Approximate ratios are calculated using (19) for the (non-centro-symmetric) MHRLA, and (20) for the
(centro-symmetric) ULA.

1
η1(κ)

0.9 η2(κ)

0.8

0.7

0.6

0.5
η

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
κ

Fig. 4. Systematic search for centro-symmetrical arrays of P = 15 sensors verifying S2 = 1. κ and η are reported
as ’+’ dots; and so in comparison with η1 (κ) and η2 (κ).

December 7, 2013 DRAFT

 19

110

100

90

80

70

f (κ)
60

1 50

40

30

20

10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
κ

(a) f1 (κ)
20

18

16

14

12
f (κ)

10
2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
κ

(b) f2 (κ)

25
θ=π/2
θ=π/8
θ=π/16
θ=0
20

15
fθ(κ)

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
December 7, 2013 κ DRAFT

(c) fθ (κ)

Fig. 5. Near-field DOA estimation performance, expressed by fθ (κ) [and its constituent functions f1 (κ) and f2 (κ)], as function
 20

1.1

0.9
Normalized CRB
DOA CRB
Range CRB
0.8 Exact
Approximate

0.7

0.6

0.5

0.4
0.5 1 1.5 2 2.5
Range r

Fig. 6. DOA and range CRBs of the non-ULA (κ = 0.4) normalized to that of the equivalent ULA (κ = 0.5776).
Both arrays are made of P = 6 sensors and are such that S2 = 1.

2
P=4,5,...,10
P=∞
1.8

1.6

1.4

1.2
RP(κ)

0.8

0.6

0.4

0.2
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
κ

Fig. 7. Centro-symmetric non-ULA vs. ULA: Compared range estimation performance of far-field sources.

December 7, 2013 DRAFT