13 views

Uploaded by avivro1

save

- Continuum Mechanics
- Tut Problems 1
- CHE 5480
- alggeo_20130802e
- Hopf Bif Phenomena Bajaj
- zebra_striping_paper_21
- 091367
- Maneuvering & Control of Marine Vehicles
- EF Pseudocode
- The Extended Graetz Problem With Axial Conduction
- 24330220 Maneuvering and Control of Marine Vehicles
- A New Method for Determining Multiple Harmonic Source Locations in a Power Distribution System
- 15. Ijhss - Port Selection Using Analytic Hierarchy Process
- Class of Irregularity
- Suli E., Mayers D.-an Introduction to Numerical Analysis - Solutions
- Engineering Mathematics
- Wdowicki Paper 2
- Matlab Commands for Control System
- Social Security: wp98
- Scientia Magna, Vol 7, No. 1, 2011
- Substitution
- 10.1.1.144.4052 (1)
- IB Mathematical Methods III (Cambridge)
- HOMOCLINIC BIFURCATIONS
- A Generalized Framework for Distrubute Power Control in Wireless Networks
- A very quick review of differential equations.
- Kalmar-Nagy Subcritical Hopf Bifurcation in the Delay Equation Model for Machine Tool Vibrations
- Getting Started With MATLAB - David Hart
- Manuscript
- Qualitative Analysis of Linear Autonomous Systems With Constant Coefficients
- The Innovators: How a Group of Hackers, Geniuses, and Geeks Created the Digital Revolution
- Dispatches from Pluto: Lost and Found in the Mississippi Delta
- Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future
- The Unwinding: An Inner History of the New America
- Yes Please
- Sapiens: A Brief History of Humankind
- The Prize: The Epic Quest for Oil, Money & Power
- The Emperor of All Maladies: A Biography of Cancer
- A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story
- This Changes Everything: Capitalism vs. The Climate
- Grand Pursuit: The Story of Economic Genius
- Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America
- John Adams
- Smart People Should Build Things: How to Restore Our Culture of Achievement, Build a Path for Entrepreneurs, and Create New Jobs in America
- The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers
- Rise of ISIS: A Threat We Can't Ignore
- The World Is Flat 3.0: A Brief History of the Twenty-first Century
- The New Confessions of an Economic Hit Man
- Team of Rivals: The Political Genius of Abraham Lincoln
- Bad Feminist: Essays
- Steve Jobs
- Angela's Ashes: A Memoir
- How To Win Friends and Influence People
- The Incarnations: A Novel
- You Too Can Have a Body Like Mine: A Novel
- Extremely Loud and Incredibly Close: A Novel
- The Sympathizer: A Novel (Pulitzer Prize for Fiction)
- The Silver Linings Playbook: A Novel
- Leaving Berlin: A Novel
- The Light Between Oceans: A Novel
- The Flamethrowers: A Novel
- Brooklyn: A Novel
- The Rosie Project: A Novel
- The Blazing World: A Novel
- We Are Not Ourselves: A Novel
- The First Bad Man: A Novel
- Bel Canto
- The Master
- A Man Called Ove: A Novel
- The Love Affairs of Nathaniel P.: A Novel
- Life of Pi
- The Cider House Rules
- A Prayer for Owen Meany: A Novel
- The Perks of Being a Wallflower
- The Bonfire of the Vanities: A Novel
- Lovers at the Chameleon Club, Paris 1932: A Novel
- Beautiful Ruins: A Novel
- The Kitchen House: A Novel
- Interpreter of Maladies
- The Wallcreeper
- The Art of Racing in the Rain: A Novel
- Wolf Hall: A Novel
- A Tree Grows in Brooklyn

You are on page 1of 5

PROGRAMMING

Joseph S. Picard and Anthony J. Weiss

School of Electrical Engineering - Systems Department

Tel Aviv University, Tel Aviv 69978, Israel

phone: + 97236407460, fax: + 97236405027, email: {picard,ajw}@eng.tau.ac.il

ABSTRACT

Precise geolocation have attracted considerable interest in the

engineering literature. Almost all previous publications con-

sider small measurement errors. In this paper we discuss ge-

olocation in the presence of outliers, where several measure-

ments are severely corrupted while other measurements are

reasonably accurate. It is known that Maximum Likelihood

or Least Squares provide poor results under these conditions.

We demonstrate how using the ℓ

1

norm and linear program-

ming we can detect the outliers and use only the good mea-

surements for providing the ﬁnal location estimate. More-

over, we provide bounds on the number of outliers that can

be detected and eliminated.

1. INTRODUCTION

This work has been motivated by a frequent problem in a

large scale commercial geolocation system mainly used for

tracking stolen luxury cars, controlling vehicle ﬂeets and per-

sonal navigation [1]. The system consists of sensors dis-

tributed over a large geographical area, very much like rural

cellular base-stations. The stations intercept signals trans-

mitted by the mobile devices. The time synchronized receiv-

ing stations record the mobile device signal time of arrival

(TOA). The TOA measurements are transferred to a central

processing unit for obtaining an estimate of the mobile de-

vice location. Geolocation based on TOA has been known

for long time and is a standard non-linear estimation prob-

lem with known solutions [2]. However, the commercial

system mentioned above provides between 7 and 40 TOA

measurements for each mobile device where a considerable

number of these measurements (up to 20%) are unreliable

(outliers). The large errors may be the result of multipath, in-

terference, jamming, or even poor synchronization of some

stations. In order to provide a reliable location estimate it

is required to identify the outliers (or, equivalently, the set

of consistent measurements) and either correct the large er-

rors or eliminate the wrong measurements from the data se-

lected for location estimation. The identiﬁcation of the out-

liers is known to be NP hard and therefore the relatively sim-

ple problem of location estimation based on TOA requires

excessive computer resources in the presence of outliers.

For example, the identiﬁcation of k (say 10) outliers out of

m (say 40) TOA measurements requires the examination of

m

m−k

=

40

30

**= 847, 660, 528 subsets of measurements when
**

k is known. Since the number of outliers is not known in

advance it is required to examine even more subsets.

This research was supported by the Israel Science Foundation (grant

No. 218/08), the Institute for Future Technologies Research named for the

Medvedi, Shwartzman and Gensler Families and by the Center for Absorp-

tion in Science, Israel.

We propose to identify the outliers by using recent re-

sults in sparse representation of signals. Consider the lin-

ear set of equations s = Br where s is a vector that has a

representation r in the span of B, a.k.a. as dictionary. In

some applications the sparsest r is desired. The representa-

tion is sparse if the dictionary is over complete, i.e., B has

more columns than rows. In this case, s = Br is an under-

determined set of equations that cannot be solved uniquely

by classical methods. In recent publications, it has been

shown that if the sparsest representation is sufﬁciently sparse,

then not only it is unique, but it can also be easily obtained by

linear programming [3]-[5]. Conditions ensuring uniqueness

of the sparsest solution were established in [6]. The prob-

lem has been extended to a wider variety of bases in [7],[8]

including dictionaries built by concatenation of non-unitary

matrices [9],[10]. Since these approaches are all based on

ℓ

1

-norm minimization, the validity of sparse representation

by ℓ

1

-norm relaxation was also studied in [11]. The theory

of sparse signal representation has also been used for outliers

identiﬁcation in [12].

Recent works have exploited these advances by formulat-

ing geolocation as a sparse representation problem involving

the signals collected at several sensors [13]. This method is

known as Compressive Sensing [14], and sparsity is consid-

ered at the level of array signal processing.

Our approach is different since we address a problem oc-

curring after the extraction of location parameters from the

received signals. We assume that TOA, Angle of Arrival

(AOA), Time Difference of Arrival (TDOA) or Received Sig-

nal Strength (RSS) measurements are available, although a

subset is severely corrupted. In a linear formulation of the ge-

olocation problem, an over-determined set of equations must

be solved.

The solution that we are proposing here consists of sev-

eral ingredients. First, we write the quadratic equations of

the TOA measurements as linear equations. See [15]-[16]

for earlier applications of transforming non-linear equations

to linear equations in geolocation applications. Second, we

deﬁne an optimization problem in terms of ℓ

1

norm mini-

mization. Third, we solve the optimization problemusing ef-

ﬁcient linear programming methods. As a result we get a list

of outliers, the size of each error and location estimate. Us-

ing the location estimate and the list of good measurements

one can now apply a ﬁnal, non-linear, localization step using

digital terrain map (DTM) for the ultimate result.

The method described above can also be used for other

geolocation systems such as geolocation based on AOA or

geolocation based on RSS. The main goal of this paper is

to ﬁnd the limitations of the proposed outliers identiﬁcation

methods.

2. OVERVIEW

It is well known that geolocation based on AOA, TOA,

TDOA and RSS can be based on a solution of a linear set

of N equations (more details will be provided shortly),

y =Ax+e (1)

The vector x is the vector of unknown parameters which ob-

viously include the coordinates of the transmitter to be lo-

cated. The matrix A and the N×1 data vector y are func-

tions of the available measurements and the known parame-

ters such as the coordinates of the sensing devices. The en-

tries of the error vector e are assumed to be zero (or “small”)

except for a fewentries which are large and therefore referred

to as outliers. The distinction between “small” and “large”

errors is not obvious and therefore we assume, as a ﬁrst at-

tempt to solve the problem at hand, that most of the entries

in e are zero and only few are different from zero. Thus, the

error vector e is sparse. We search for the vector x that is

associated with the minimal number of outliers. Deﬁne [12],

(P

0

) : min

x

Ax−y

0

(2)

(P

1

) : min

x

Ax−y

1

(3)

where ·

0

stands for the ℓ

0

-norm (the number of vector ele-

ments that are not zero) and ·

1

stands for the ℓ

1

-norm (the

sum of absolute values of the vector elements). Formally, we

look for the solution of the minimization problem (P

0

). Un-

fortunately, this problem is NP hard, and therefore one may

consider solving (P

1

). The problem (P

1

) can be solved by

linear programming and according to our numerical experi-

ence works amazingly well.

The challenge is to establish a condition ensuring that the

solution of the tractable problem (P

1

) is also the solution of

the hard problem(P

0

). In this paper, we derive this condition

and show that it depends on the number of outliers. The idea

is to identify a matrix F =0 so that FA= 0. Then equation

(1) reduces to Fy =Fe. Deﬁne

(P

′

0

) : min

e

e

0

s.t. Fy =Fe (4)

(P

′

1

) : min

e

e

1

s.t. Fy =Fe (5)

In [12] it has been shown that the solution e

0

of the alter-

native problem (P

′

0

) yields the solution x

0

of (P

0

). Since

the ℓ

0

-norm problem (P

′

0

) is hard, it is replaced by the alter-

native problem (P

′

1

). Deﬁne the matrix U whose columns

are the normalized eigenvectors of AA

T

associated with the

non-zero eigenvalues. We propose F=I

N

−UU

T

where I

N

is the N ×N identity matrix. We look for a condition en-

suring that the solution of (P

′

1

) is identical to the solution

of (P

′

0

). Denote by e

0

the solution of (P

′

0

) and by S the

set of indices corresponding to non-zero entries of e

0

i.e.,

S = {i ∈ [1, ..., N]|e

0

(i) = 0}. Denote by .

2

the ℓ

2

-norm,

by s the number of elements in S and by δ(S) the s×1 vector

obtained from a vector δ by removing the entries δ

i

, i / ∈ S.

Lemma 1 If any vector δ =0 that satisﬁes Fδ =0 also sat-

isﬁes δ(S)

1

<

1

2

δ

1

then the solution of (P

′

0

) coincides

with the solution of (P

′

1

).

Lemma 2 Any δ = 0 that satisﬁes Fδ = 0 also satisﬁes

δ(S)

1

≤sδ

1

max

ℓ

U(ℓ, :)

2

2

¸

.

For the proofs of the lemmas, see [6], [8], [17] and [18]. By

combining the two lemmas we can get a condition that guar-

antees equivalence between (P

′

1

) and (P

′

0

). This condition is

sδ

1

max

ℓ

{U(ℓ, :)

2

2

} <

1

2

δ

1

which proves Theorem 1.

Theorem 1 If the number of non-zero entries, s, in the solu-

tion of (P

′

0

) satisﬁes

s <

1

2max

ℓ

U(ℓ, :)

2

2

∆

= B (6)

Then (P

′

1

) and (P

′

0

) have the same solution and the outliers

can be identiﬁed exactly.

The bound above depends on Uwhich is strongly related

to A. In the next section, we establish the structure of the

matrix A for AOA, TOA and TDOA data. Then, in sections

4, 5 and 6 we ﬁnd corresponding expressions for U. The

effects of small errors on outliers identiﬁcation are discussed

in section 8. Finally, numerical results are given in section 9.

3. REPRESENTATION OF GEOLOCATION BY

LINEAR EQUATIONS

In this section we cast the usually non-linear geolocation

equations as a set of linear equations.

3.1 Geolocation based on AOA

Assume that N stations collect AOA measurements of a sin-

gle target. Let p

i

= [x

i

, y

i

]

T

represent the i-th station coordi-

nates and p

t

= [x

t

, y

t

]

T

the target coordinates. Let φ

i

be the

AOA (measured anticlockwise w.r.t. the x-axis) at the i-th

station. Then, in the absence of noise, the i-th AOA mea-

surement satisﬁes (y

t

−y

i

)cosφ

i

= (x

t

−x

i

)sinφ

i

. Deﬁne

s [sinφ

1

, · · · , sinφ

N

]

T

c [cosφ

1

, · · · , cosφ

N

]

T

A [s,−c] x [x

t

, y

t

]

T

y [(x

1

sinφ

1

−y

1

cosφ

1

), · · · , (x

N

sinφ

N

−y

N

cosφ

N

)]

T

Then, y =Ax, and in the presence of errors y =Ax+e.

3.2 Geolocation based on TOA

Let m

i

be the measured TOA, c the propagation speed, and

t

0

the known transmit time. In the absence of errors, the

i-th TOA measurement is given by m

i

= c

−1

p

i

−p

t

2

+

t

0

. Substituting the target and sensor coordinates we get,

c

2

(m

i

−t

0

)

2

= p

i

2

+p

t

2

−2p

T

i

p

t

. Collecting the mea-

surements from N sensors, we get y =Ax with

y [p

1

2

, · · · , p

N

2

]

T

−c

2

m⊙m+t

2

0

1

N

−2t

0

m

A

2[p

1

, p

2

, · · · , p

N

]

T

, 1

N

x [p

T

t

, −p

t

2

]

T

where m⊙mis the element-wise product, and 1

N

is an N×1

vector of ones. In the presence of errors we get y =Ax+e.

3.3 Geolocation based on TDOA

If the transmit time, t

0

, is unknown the localization method

is called TDOA. In this case we have,

y [p

1

2

, · · · , p

N

2

]

T

−c

2

(m⊙m)

A

2[p

1

, p

2

, · · · , p

N

]

T

, 1

N

, cm

x [p

T

t

, (c

2

t

2

0

−p

t

2

), −2ct

0

]

T

Again, y =Ax, and in the presence of errors y =Ax+e.

4. NUMBER OF IDENTIFIABLE OUTLIERS IN

AOA GEOLOCATION

We nowestablish the bound on the number of outliers s given

in (6) for AOA geolocation. As shown in Section 3, we have

A = [s, −c]. The eigenvectors of A are c/c and s/s

provided that c

T

s = 0. This orthogonality condition can

always be satisﬁed by rotating the coordinate system [18].

Thus, U= [s/s, c/c]. According to theorem 1,

B =

c

2

s

2

2max

ℓ

c

2

sin

2

φ

ℓ

+s

2

cos

2

φ

ℓ

(7)

4.1 Application to Uniform Distribution of Sensors

Assume that the sensors are uniformly distributed within a

circular area whose radius is R and its center is at (0, 0). De-

note by (r

ℓ

, θ

ℓ

) the polar coordinates of the ℓ-th sensor. The

angles θ

ℓ

are uniformly distributed in [0, 2π] and the ranges

r

ℓ

are distributed in [0, R] according to the probability density

function f

r

(r) =

2r

R

2

. For large N, c

2

and s

2

tend to

N

2

.

Then B(N) =

N

4

+o(N).

4.2 Application to Circular Sensor Array

The sensors are uniformly distributed on a circle with radius

R. The coordinates of the ℓ-th sensor are x

ℓ

= Rcosθ

ℓ

, y

ℓ

=

Rsinθ

ℓ

where θ

ℓ

= 2πℓ/N. Assume that the transmitter is

located on the x-axis at coordinates (d, 0) where d > 0. We

show in [18] that for large N the bound tends to N/4 for

a transmitter within the circle of sensors, and to N/6 for a

remote transmitter. Furthermore, for any N, including small

N, if d ∈ [0, R] then B(N) ≥

N

4

−

1

2

.

4.3 Application to Concentric Circular Sensor Arrays

Consider K concentric circles. Assume that N

k

sensors are

located on the k-th circle and that N

k

= R

k

N

1

, where N

1

≫1

is the number of sensors on the smallest circle whose radius

is R

1

= 1. Further, assume that R

1

< R

2

< · · · < R

K

. Then,

for a transmitter located at coordinates (d, 0),

B

a

(N) =

N

4

d ≤R

1

N

1

4

1

d

2

∑

k−1

j=1

R

3

j

+∑

K

j=k

R

j

d ∈ [R

k−1

, R

k

]

N

2

2R

2

K

∑

K

j=1

R

j

∑

K

j=1

R

3

j

+1

−1

d ≫R

K

converges asymptotically to a bound stricter than B(N) [18].

5. NUMBER OF IDENTIFIABLE OUTLIERS IN

TOA GEOLOCATION

As shown in section 3, A = [2x

b

, 2y

b

, 1

N

] where x

b

, y

b

are

the x and y coordinates of all the sensors. The eigenvec-

tors of A are x

b

/x

b

, y

b

/y

b

and 1

N

/

√

N provided that

1

T

N

x

b

= 1

T

N

y

b

= x

T

b

y

b

= 0. These orthogonality assump-

tions can always be satisﬁed by a coordinate system rotation

and an addition of an appropriate artiﬁcial outlier measure-

ment [18], which slightly loosens the bound but in turn yields

closed form expressions. Then U =

x

b

x

b

,

y

b

y

b

,

1

N

√

N

. Ac-

cording to Theorem 1,

B =

x

b

2

y

b

2

2max

ℓ

¸

y

b

2

x

2

ℓ

+x

b

2

y

2

ℓ

¸

+2x

b

2

y

b

2

/N

(8)

5.1 Application to Uniform distribution of sensors

Assume that the sensors are uniformly distributed as in 4.1.

Then x

b

2

=y

b

2

=

NR

2

4

+o(N). Thus, B(N) =

N

10

+o(N)

without any limitation on the location of the transmitter.

5.2 Application to Uniform Circular Sensor Array

Consider a uniform circular array of N sensors as in 4.2.

Then x

b

2

= y

b

2

= NR

2

/2. From (8) we obtain the

bound on the number of outliers is B = N/6.

5.3 Application to Concentric Circular Sensor Arrays

Consider L uniform concentric circular sensor arrays cen-

tered on (0, 0). Assume that the ℓ-th circular array consists

of N

ℓ

≥ 3 sensors. The coordinates of the k-th sensor of the

ℓ-th array of radius r

ℓ

are

r

ℓ

cos(θ

ℓ,k

+α

ℓ

), r

ℓ

sin(θ

ℓ,k

+α

ℓ

)

where α

ℓ

∈ [0, 2π] and θ

ℓ,k

= 2πk/N

ℓ

for k = 1, ..., N

ℓ

. De-

ﬁne r

2

1

N

∑

L

ℓ=1

N

ℓ

r

2

ℓ

. Then we get from (8) the bound

B = N/(4max

¸

r

2

ℓ

/r

2

¸

+2) for any N.

5.4 Application to Uniform Linear Sensor Array

Consider a uniform linear array of sensors placed on the

x axis between −R/2 and R/2. The sensor separation is

R/(N −1). Observe that y

b

2

= 0 and x

b

2

= R

2 N

12

N+1

N−1

.

Since y

b

= 0 the eigenvectors of the matrix A are U =

[

x

b

x

b

,

1

N

√

N

]. Then, B = N/(6

N−1

N+1

+2) and B tends to

N

8

.

6. NUMBER OF IDENTIFIABLE OUTLIERS IN

TDOA GEOLOCATION

As shown in section 3, A= [2x

b

, 2y

b

, 1

N

, cm]. The eigen-

vectors of A are

x

b

x

b

,

y

b

y

b

,

1

N

√

N

and

m

m

provided that they

are orthogonal. These orthogonality conditions can always

be satisﬁed by coordinate system rotation and translation and

an addition of an appropriate artiﬁcial outlier measurement

[18]. Under these orthogonality conditions the bound is

B =

2max

ℓ

x

2

ℓ

x

b

2

+

y

2

ℓ

y

b

2

+

m

2

ℓ

m

2

¸

+

2

N

−1

(9)

6.1 Application to Uniform Distribution of Sensors

Consider a uniform distribution of sensors as in 5.1. Without

loss of generality, select t

0

=−

E{r

ℓ

}

c

. We show in [18] that

B(N) =

N

18

+o(N).

6.2 Application to Uniform Circular Sensor Array

Consider a uniformcircular array of sensors as in 4.2. Again,

assume that t

0

= −

E{r

ℓ

}

c

where r

ℓ

is the range between the

transmitter and the ℓ-th sensor. We show in [18] that for

N ≥3 the bound is B(N) =

2max

ℓ

m

2

ℓ

m

2

+

6

N

−1

and

B(N) =

N/6 d = 0

N/10 d = 0

+

or d ≫R

Np/(4p +4)+o(N) d = R where p = 2 −

16

π

2

Further, for ﬁnite N a good approximation for B is

B

a

(N) =

(N/10)

1 −

1

3

d

R

0 < d ≤R

(N/10)

1 −

R

5d−2R

d > R

7. HIDDEN ERRORS

We show nowthat some measurement errors do not affect the

consistency of the linear equations set. According to section

3, the AOA k-th linear equation is {y}

k

=x

k

sinφ

k

−y

k

cosφ

k

where φ

k

is the errorless AOA measurement. In the presence

of perturbations, denote by ε

k

the k-th measurement error and

by

ˆ

φ

k

=φ

k

+ε

k

the k-th actual measurement. We showin [18]

that e

k

= 2r

k

sin(ε

k

/2)cos(ε

k

/2 +φ

k

−θ

k

).The cosine term

is zero for ε

k

= 2(θ

k

−φ

k

) ±π. Then e

k

= 0 although ε

k

= 0.

Thus, there are cases of errors in the measurements that do

not appear as an error in the linear equations and therefore

do not affect the correct solution.

Similarly, according to section 3, and under the assump-

tion that t

0

= 0, the exact TOA measurements are m

k

=

1

c

p

k

−p

t

and the corresponding linear equation is {y}

k

=

p

k

2

−c

2

m

2

k

. If the k-th TOA measurement ˆ m

k

is impre-

cise, we have ˆ m

k

= m

k

+

ε

k

c

where ε

k

is the ranging error.

Then e

k

= −ε

k

(2p

k

−p

t

+ε

k

), see [18] for details. If

ε

k

= −2p

k

−p

t

then e

k

= 0, meaning that this measure-

ment error does not affect the linear equations and their solu-

tion. Note that this measurement error corresponds to ˆ m

k

<t

0

and therefore is likely to be removed from the TOA data set.

If t

0

is unknown, as in TDOA, the error cannot be detected

easily before emitter localization.

We have revealed the existence of hidden errors that af-

fect the measurements and therefore the nonlinear estima-

tion problem, but do not affect the linear set of equations.

Of course, hidden errors are not likely to occur since they

must satisfy the condition e

k

= 0 described above. However,

the probability of small values for e

k

may not be negligible.

Thus, we extend the concept of hidden outliers to any large

ε

k

for which e

k

is small.

Proposition 2 Robustness to Hidden Outliers - In the pres-

ence of small errors, solving nonlinear localization problems

by means of a linear set of equations ensures robustness to

hidden outliers.

8. STABILITY IN THE PRESENCE OF SMALL

ERRORS

Recall that our model in (1) assumed that the error vector e

consists of a few large entries called outliers and all other

entries are zero. This model assumes that small errors do not

exist. The original problem can be written as (P

′

1,0

) where

(P

′

1,ε

) : min

e

e

1

s.t. Fy−Fe

2

2

≤ε

In the presence of small errors we add a non-sparse vector

n, y = Ax+e +n. Assuming bounded small errors, pro-

jection using F yields Fy = Fe +Fn and Fn ≤ ε. Thus

we are interested in solving (P

′

1,ε

) with ε > 0. Denote by e

and e

n

the solutions of (P

′

1,ε

) in the absence and in the pres-

ence of small errors. By continuity of the penalty and con-

straint functions, we can justify the existence of a function

f (ε) so that lim

ε→0

f (ε) =0 and |e(k)−e

n

(k)| < f (ε) for any k.

For a non-outlier entry we have e(k) = 0 and |e

n

(k)| < f (ε).

Thus, the solutions e and e

n

have identical support provided

that all the entries of e

n

satisfying |e

n

(k)| < f (ε) are set to

0. Therefore, outliers identiﬁcation in the presence of sufﬁ-

ciently small errors can be achieved by linear programming.

0 5 10 15 20 25 30

0

1

2

3

4

5

6

7

8

9

10

Number of sensors N

B

o

u

n

d

o

n

t

h

e

n

u

m

b

e

r

o

f

o

u

t

l

i

e

r

s

B

e

B

B

a

Figure 1: The experimental bound, (B

e

), the theoretical

bound on the number of outliers, (B), and its approximation,

(B

a

), versus the number N of TOA sensors in the circular

array

9. NUMERICAL EXAMPLES

In previous sections, we discussed a sufﬁcient condition en-

suring the identiﬁcation of outliers within a set of AOA, TOA

or TDOA measurements by means of linear programming.

The condition limits the number of identiﬁable outliers by an

upper bound. In this section we examine the tightness of the

upper bound by numerical simulations.

9.1 Outliers identiﬁcation in TOA models

Recall that a range measurement error ε

k

generates a linear

equation error e

k

= −ε

k

(2r

k

+ε

k

) where r

k

is the distance

of the k-th sensor from the transmitter. We deﬁne an er-

ror as an outlier if |ε

k

| ≥ αr

k

with α = 0.1 Furthermore,

for a given topology one can generally establish an upper

bound on the largest distance between the transmitter and the

sensors. Range measurements with excessively large errors

can be removed before processing. Thus, one can deﬁne a

topology-dependent scalar D so that 0 ≤r

k

+ε

k

≤D for any

k. Thus, an outlier error e

k

resulting from ε

k

, satisﬁes

e

k

∈

−D

2

+r

2

k

, −α(2 +α)r

2

k

∪

αr

2

k

, (2 −α)r

2

k

(10)

In the simulations the outlier error have been uniformly dis-

tributed within the interval deﬁned in (10).

In this example, small errors have not been simulated.

Consider a uniform circular array of TOA sensors. Assume

the array has a radius of 1 distance unit [DU] and that the

emitter is located within this array. The maximum distance

between the transmitter and the sensors is 2 [DU]. Thus,

D = 2 [DU]. The bound B

a

= N/6 is independent of the

transmitter location. For each of the N

e

= 200 experiments

the transmitter location is selected at random within the cir-

cular area of the sensor array, and we identify the smallest

number of outliers for which identiﬁcation fails. The result

of the i-th experiment is denoted by k

i

. The experimental

bound is B

e

= min

1≤i≤N

e

k

i

. Denote by ˆ e the estimate of e by

solving (P

′

1

) using linear programming. We deﬁne failure

when some outliers are not identiﬁed or some good measure-

ments are classiﬁed as outliers. Since the experiments are

based on numerical solutions, we deﬁne perfect outliers iden-

tiﬁcation when max

1≤i≤N

| ˆ e

i

−e

i

| < 10

−6

. In Fig. 1, we plotted

the bounds B

e

, B and B

a

versus N. The bound B appears to

be rather tight.

0 0.05 0.1 0.15 0.2 0.25 0.3

0

0.1

0.2

0.3

0.4

0.5

Standard deviation σ of ranging measurement noise [DU]

R

M

S

E

o

f

l

o

c

a

t

i

o

n

e

s

t

i

m

a

t

e

s

[

D

U

]

0 outliers, LS

0 outliers, LP+LS

5 outliers, LS

5 outliers, LP+LS

10 outliers, LS

10 outliers, LP+LS

Figure 2: RMSE of LS location estimates versus standard

deviation of ranging small errors. Thin lines show results of

straightforward LS while thick lines show results with out-

liers removal step.

9.2 Localization in the Presence of Small Errors and

Outliers

In this section we examine the precision of least squares (LS)

location estimates with and without outliers identiﬁcation.

The measurements in the presence of noise (small errors) and

outliers are given by y =Ax+e+n. The LS estimate of x

is given by ˆ x

0

=A

†

y where A

†

is the pseudo-inverse of A.

A better way to estimate x may be to ﬁrst apply linear pro-

gramming in order to estimate the error vector

e+n. Then

apply LS to the modiﬁed data ˆ x

e

=A

†

(y−

e+n).

We performed 200 experiments with a uniform circular

array of radius 1 [DU] consisting of N = 30 TOA sensors.

The number of outliers s was successively set to 0, 5 and 10,

corresponding to 0%, 16% and 33% of the equations. The

outliers size correspond to (10). The small ranging errors

are realizations of zero-mean gaussian randomvariables with

standard deviation σ. Fig. 2 shows the RMSE of location

estimates versus σ. Note that σ = 0.3 [DU] is not negligi-

ble compared with the error-free measurements and therefore

the number of outliers is occasionally increased. According

to Fig. 2, in the presence of 5 or 10 outliers, it is always

signiﬁcantly advantageous to remove the outliers by linear

programming before LS estimation, regardless of the noise

variance. In the absence of outliers, a slight performance

degradation may be observed if preprocessing is used. Note

however that if no outlier is found the preprocessing can be

ignored.

10. CONCLUSION

We have invoked the theory of sparse signal representation

in the context of positioning. We have shown that in the ab-

sence of noise, AOA, TOA and TDOA data sets corrupted

by unknown outliers can be corrected using linear program-

ming, provided that the number of outliers does not exceed

a limit given in Theorem 1. In Sections 4, 5 and 6 the limit

given in (6) has been evaluated for several geometries. For

each case, the exact expression of the bound and simple ap-

proximations were presented. In followup work we focus on

outliers in the presence of small errors.

REFERENCES

[1] http://www.ituran.co.il

[2] D. J. Torrieri, “Statistical Theory of Passive Location

System,” IEEE Trans. on Aerospace and Electronic

Systems, Vol. AES-20, No. 2, pp. 183-198, Mar. 1984.

[3] I. Daubechies, “Time-frequency localization operators:

A geometric phase space approach”, IEEE Trans. In-

form. Theory, vol. 34, pp. 605-612, July 1988.

[4] S. Mallat, Z. Zhang, “Matching Pursuit in a Time-

Frequency Dictionary”, IEEE Trans. Signal Proc., vol.

41, pp. 3397-3415, Dec. 1993.

[5] A. Feuer, A. Nemirovsky, “On Sparse Representations

in Pairs of Bases”, IEEE Trans. Inf. Theory, vol. 49, no.

6, pp. 1579-1581, Jun. 2003.

[6] D. L. Donoho, X. Huo, “Uncertainty Principles and

Ideal Atomic Decomposition”, IEEE Trans. on Info.

Theory, vol. 47, no. 7, pp. 2845-2862, Nov. 2001.

[7] S. S. Chen, D. L. Donoho, M. A. Saunders, “Atomic

Decomposition by Basis Pursuit”, SIAM Journal on

Scientiﬁc Comp., vol. 40, no. 1, pp. 33-61, 1999.

[8] D. L. Donoho, M. Elad, “Optimally Sparse Representa-

tion in General (Nonorthogonal) Dictionaries via Min-

imization”, Proc. National Academy of Science, vol.

100, no. 5, pp. 2197-2202, Mar. 2003.

[9] J. J. Fuchs, “On Sparse Representations in Arbitrary

Redundant Bases”, IEEE Trans. on Information The-

ory, vol. 50, no. 6, pp. 1341-1344, June 2004.

[10] D. L. Donoho, “For Most Large Underdetermined Sys-

tems of Linear Equations the Minimal ℓ

1

-norm Solu-

tion is also the Sparsest Solution”, Comm. on pure and

applied math., vol. 59, no. 7, pp. 907-934, Wiley, 2006.

[11] Y. Tsaig, D. L. Donoho, “Breakdown of Equivalence

between the Minimal ℓ

1

-normSolution and the Sparsest

Solution”, Signal Processing, vol. 86, no. 3, pp. 533-

548, Mar. 2006.

[12] E. Candes, T. Tao, “Decoding by Linear Program-

ming”, IEEE Trans. on Information Theory, vol. 51, no.

12, pp. 4203-4215, Dec. 2005.

[13] D. Malioutov, M. Cetin, A. S. Willsky, “A Sparse Sig-

nal Reconstruction Perspective for Source Localization

With Sensor Arrays”, IEEE Trans. on Signal Proc., vol.

53, no. 8, pp. 3010-3022, Aug. 2005.

[14] V. Cevher, A. C. Gurbuz, J. H. McClellan, M. Chel-

lappa, “Compressive Wireless Arrays for Bearing Esti-

mation”, Proc. ICASSP 2008, pp. 2497-2500, 2008.

[15] B. Friedlander, “A Passive Localization Algorithm and

its Accuracy Analysis,” IEEE Journal of Ocean Engi-

neering, Vol. OE-12, No. 1, pp. 234-245, Jan. 1987.

[16] J. O. Smith, J. S. Abel, “The Spherical Interpola-

tion Method of Source Localization,” IEEE Journal of

Ocean Eng., Vol. OE-12, No. 1, pp. 246-252, Jan. 1987.

[17] E. J. Candes, J. Romberg, T. Tao, “Robust Uncertainty

Principles: Exact Signal Reconstruction From Highly

Incomplete Frequency Information”, IEEE Trans. on

Info. Theory, vol. 52, no. 2, pp. 489-509, Feb. 2006.

[18] J. S. Picard, A.J. Weiss, “Accurate Geolocation in the

Presence of Outliers using Linear Programming”, IEEE

Trans. on Signal Proc., submitted for publication, 2009.

with the solution of (P1 Lemma 2 Any δ = 0 that satisﬁes Fδ = 0 also satisﬁes δ (S) 1 ≤ s δ 1 max U(ℓ. / Lemma 1 If any vector δ = 0 that satisﬁes Fδ = 0 also sat′ isﬁes δ (S) 1 < 1 δ 1 then the solution of (P0 ) coincides 2 ′ ). · · · . Then. p2 . · · · .2 Geolocation based on TOA min e In [12] it has been shown that the solution e0 of the alter′ native problem (P0 ) yields the solution x0 of (P0 ). Deﬁne [12].t. − pt 2 ]T where m⊙m is the element-wise product. Then equation (1) reduces to Fy = Fe. pN ]T .. it is replaced by the alter′ native problem (P1 ). 3. The idea is to identify a matrix F = 0 so that FA = 0. cos φN ]T [s. and therefore one may consider solving (P1 ). yi ]T represent the i-th station coordinates and pt = [xt . yt ]T the target coordinates. this problem is NP hard. −2ct0 ]T Again.. y = Ax. . TOA and TDOA data. s. Finally. 2 the ℓ2 -norm.t. The matrix A and the N × 1 data vector y are functions of the available measurements and the known parameters such as the coordinates of the sensing devices. · · · .−c] x [xt . We look for a condition en′ suring that the solution of (P1 ) is identical to the solution ′ ). Denote by e the solution of (P′ ) and by S the of (P0 0 0 set of indices corresponding to non-zero entries of e0 i. y = Ax. t0 . y = Ax + e (1) For the proofs of the lemmas. · · · . ℓ The vector x is the vector of unknown parameters which obviously include the coordinates of the transmitter to be located. This condition is 2 1 s δ 1 max{ U(ℓ.2. i ∈ S. We search for the vector x that is associated with the minimal number of outliers. and t0 the known transmit time. the x-axis) at the i-th station. c2 (mi − t0 )2 = pi 2 + pt 2 − 2pT pt . Since ′ the ℓ0 -norm problem (P0 ) is hard. 3. REPRESENTATION OF GEOLOCATION BY LINEAR EQUATIONS In this section we cast the usually non-linear geolocation equations as a set of linear equations. 3. Fy = Fe s. we look for the solution of the minimization problem (P0 ). Collecting the meai surements from N sensors. The distinction between “small” and “large” errors is not obvious and therefore we assume. · · · . is unknown the localization method is called TDOA. :) 2 . The effects of small errors on outliers identiﬁcation are discussed in section 8. Deﬁne the matrix U whose columns are the normalized eigenvectors of AAT associated with the non-zero eigenvalues. In this case we have. 3. numerical results are given in section 9. (P0 ) : (P1 ) : min Ax − y x Theorem 1 If the number of non-zero entries. The entries of the error vector e are assumed to be zero (or “small”) except for a few entries which are large and therefore referred to as outliers.e. the error vector e is sparse. [8]. sin φN ]T c [cos φ1 . 5 and 6 we ﬁnd corresponding expressions for U. The challenge is to establish a condition ensuring that the solution of the tractable problem (P1 ) is also the solution of the hard problem (P0 ). Substituting the target and sensor coordinates we get. p2 . and in the presence of errors y = Ax + e. We propose F = IN − UUT where IN is the N × N identity matrix. The problem (P1 ) can be solved by linear programming and according to our numerical experience works amazingly well. we derive this condition and show that it depends on the number of outliers. (xN sin φN − yN cos φN )]T 0 1 (2) (3) min Ax − y x where · 0 stands for the ℓ0 -norm (the number of vector elements that are not zero) and · 1 stands for the ℓ1 -norm (the sum of absolute values of the vector elements). Deﬁne ′ (P0 ) : ′ (P1 ) : min e e e 0 1 s. we get y = Ax with y A 2 [ p1 2 . 1N . · · · . Thus. see [6]. in the solu′ tion of (P0 ) satisﬁes s< ′ (P1 ) ′ (P0 ) 1 2 max U(ℓ. yt ]T [(x1 sin φ1 − y1 cos φ1 ) . In this paper. [17] and [18]. Formally. cm 2 [ptT . Deﬁne s A y [sin φ1 . TDOA and RSS can be based on a solution of a linear set of N equations (more details will be provided shortly). pN ]T . Fy = Fe (4) (5) Then. pN 2 ]T − c2 (m ⊙ m) 2[p1 . y A x [ p1 2 . Let pi = [xi .r. (c2t0 − pt 2 ). as a ﬁrst attempt to solve the problem at hand. and 1N is an N ×1 vector of ones. In the presence of errors we get y = Ax + e.. pN 2 ]T − c2 m ⊙ m + t0 1N − 2t0m 2[p1 .t. in the absence of noise. we establish the structure of the matrix A for AOA. TOA. In the absence of errors.. by s the number of elements in S and by δ (S) the s × 1 vector obtained from a vector δ by removing the entries δi . Denote by .1 Geolocation based on AOA Assume that N stations collect AOA measurements of a single target. Unfortunately. 1N x [ptT . Then. that most of the entries in e are zero and only few are different from zero. · · · . The bound above depends on U which is strongly related to A. the i-th AOA measurement satisﬁes (yt − yi ) cos φi = (xt − xi ) sin φi . in sections 4. :) 2 } < 2 δ 1 which proves Theorem 1. In the next section. . c the propagation speed. N]|e0 (i) = 0}. and in the presence of errors y = Ax + e. OVERVIEW It is well known that geolocation based on AOA. 2 ℓ Let mi be the measured TOA. S = {i ∈ [1. the i-th TOA measurement is given by mi = c−1 pi − pt 2 + t0 .3 Geolocation based on TDOA If the transmit time. :) ℓ 2 2 =B ∆ (6) Then and have the same solution and the outliers can be identiﬁed exactly. Let φi be the AOA (measured anticlockwise w. By combining the two lemmas we can get a condition that guar′ ′ antees equivalence between (P1 ) and (P0 ).

U = [s/ s . assume that R1 < R2 < · · · < RK . 4 2 4. θℓ ) the polar coordinates of the ℓ-th sensor. yb . B = N/(4 max rℓ 5.2.. Furthermore. 5. The coordinates of the ℓ-th sensor are xℓ = R cos θℓ .3 Application to Concentric Circular Sensor Arrays Consider K concentric circles.2 Application to Uniform Circular Sensor Array Consider a uniform circular array of N sensors as in 4. These orthogonality conditions can always be satisﬁed by coordinate system rotation and translation and an addition of an appropriate artiﬁcial outlier measurement [18]. Observe that yb 2 = 0 and xb 2 = R2 12 N+1 . Thus. N d ≤ R1 4 N1 1 k−1 3 K d ∈ [Rk−1 . 2π ] and the ranges rℓ are distributed in [0. Then. Nℓ . Rk ] Ba (N) = 4 d 2 ∑ j=1 R j + ∑ j=k R j −1 ∑K R j N 2R2 j=1 + 1 d ≫ RK 2 K ∑K 3 j=1 R j 6.1 Application to Uniform Distribution of Sensors Consider a uniform distribution of sensors as in 5.k = 2π k/Nℓ for k = 1. 0). NUMBER OF IDENTIFIABLE OUTLIERS IN AOA GEOLOCATION We now establish the bound on the number of outliers s given in (6) for AOA geolocation. This orthogonality condition can always be satisﬁed by rotating the coordinate system [18]. 2yb . . which slightly loosens the bound but in turn yields 1 closed form expressions. The sensor separation is N R/(N − 1). 2yb . B= 2 max ℓ m2 ℓ 6 +N −1 and d=0 N/6 d = 0+ or d ≫ R B(N) = N/10 16 N p/(4p + 4)+ o(N) d = R where p = 2 − π 2 Further. A = [2xb . The angles θℓ are uniformly distributed in [0. −c]. The eigenvec√ tors of A are xb / xb . 0). for a transmitter located at coordinates (d. For large N. NUMBER OF IDENTIFIABLE OUTLIERS IN TOA GEOLOCATION As shown in section 3. Then xb 2 = yb 2 = NR2 /2. R] then B(N) ≥ N − 1 . yb are the x and y coordinates of all the sensors. B = N/(6 N−1 + 2) and B tends to N . As shown in Section 3. √N ]. Again.1. 4 4. and to N/6 for a remote transmitter. Assume that the ℓ-th circular array consists of Nℓ ≥ 3 sensors.2. 2π ] and θℓ. 0) where d > 0. According to theorem 1. Denote by (rℓ .. 2 Then B(N) = N + o(N). These orthogonality assumpb tions can always be satisﬁed by a coordinate system rotation and an addition of an appropriate artiﬁcial outlier measurement [18]. 5. 2 N Then xb 2 = yb 2 = NR + o(N).k + αℓ ). N−1 Since yb = 0 the eigenvectors of the matrix A are U = 1 [ xb . B= 2 max ℓ 5. √N . yℓ = R sin θℓ where θℓ = 2π ℓ/N. yb / yb and 1N / N provided that T T 1N xb = 1N yb = xT yb = 0. for any N.1 Application to Uniform Distribution of Sensors Assume that the sensors are uniformly distributed within a circular area whose radius is R and its center is at (0. B(N) = 10 + o(N) 4 without any limitation on the location of the transmitter.1. From (8) we obtain the bound on the number of outliers is B = N/6. R] according to the probability density 2r function fr (r) = R2 . Thus. We show in [18] that c N B(N) = 18 + o(N). The eigenvectors of A are c/ c and s/ s provided that cT s = 0. Assume that Nk sensors are located on the k-th circle and that Nk = Rk N1 . Then U = xb . Further. 6.1 Application to Uniform distribution of sensors Assume that the sensors are uniformly distributed as in 4. The eigen1 vectors of A are xb .3 Application to Concentric Circular Sensor Arrays Consider L uniform concentric circular sensor arrays centered on (0. c 2 and s 2 tend to N . where N1 ≫ 1 is the number of sensors on the smallest circle whose radius is R1 = 1. Under these orthogonality conditions the bound is B = 2 max ℓ x2 ℓ xb 2 + y2 ℓ yb 2 + m 2 m2 ℓ 2 +N −1 (9) 6. De1 L 2 ﬁne r2 N ∑ℓ=1 Nℓ rℓ . 1N ] where xb .4. for ﬁnite N a good approximation for B is Ba (N) = d (N/10) 1 − 1 R 3 R (N/10) 1 − 5d−2R xb 2 yb 2 yb 2 x2 + xb 2 y2 + 2 xb ℓ ℓ 2 yb 2 /N (8) 0<d ≤R d>R . cm].4 Application to Uniform Linear Sensor Array Consider a uniform linear array of sensors placed on the x axis between −R/2 and R/2. c/ c ]. NUMBER OF IDENTIFIABLE OUTLIERS IN TDOA GEOLOCATION As shown in section 3. √N and m provided that they xb yb m N are orthogonal.2 Application to Uniform Circular Sensor Array Consider a uniform circular array of sensors as in 4. We show in [18] that for large N the bound tends to N/4 for a transmitter within the circle of sensors. rℓ sin(θℓ. 0). select t0 = − E{rℓ } . we have A = [s. Acxb yb N cording to Theorem 1. if d ∈ [0. Without loss of generality.k + αℓ ) where αℓ ∈ [0. N+1 8 x N b c 2 s 2 c 2 sin2 φℓ + s 2 cos2 φ ℓ (7) 4. We show in [18] that for N ≥ 3 the bound is B(N) = 2 max ℓ m 2 converges asymptotically to a bound stricter than B(N) [18]. A = [2xb .. Then we get from (8) the bound 2 /r2 + 2) for any N. Then. E{r } assume that t0 = − c ℓ where rℓ is the range between the transmitter and the ℓ-th sensor. 1N . including small N.2 Application to Circular Sensor Array The sensors are uniformly distributed on a circle with radius R. yb . The coordinates of the k-th sensor of the ℓ-th array of radius rℓ are rℓ cos(θℓ. Assume that the transmitter is located on the x-axis at coordinates (d. 5.

we discussed a sufﬁcient condition ensuring the identiﬁcation of outliers within a set of AOA.1 Outliers identiﬁcation in TOA models Recall that a range measurement error εk generates a linear equation error ek = −εk (2rk + εk ) where rk is the distance of the k-th sensor from the transmitter. We deﬁne failure when some outliers are not identiﬁed or some good measurements are classiﬁed as outliers. 1. In Fig. Similarly. we deﬁne perfect outliers identiﬁcation when max |ei − ei | < 10−6 . outliers identiﬁcation in the presence of sufﬁciently small errors can be achieved by linear programming. Consider a uniform circular array of TOA sensors. The bound B appears to be rather tight. Then ek = 0 although εk = 0. versus the number N of TOA sensors in the circular array 9. According to section 3. If t0 is unknown. 8. B and Ba versus N. STABILITY IN THE PRESENCE OF SMALL ERRORS Recall that our model in (1) assumed that the error vector e consists of a few large entries called outliers and all other entries are zero. Assuming bounded small errors. Thus. We deﬁne an error as an outlier if |εk | ≥ α rk with α = 0.In the presence of small errors. and its approximation. Proposition 2 Robustness to Hidden Outliers . Since the experiments are based on numerical solutions. We have revealed the existence of hidden errors that affect the measurements and therefore the nonlinear estimation problem. For each of the Ne = 200 experiments the transmitter location is selected at random within the circular area of the sensor array. satisﬁes 2 2 2 2 ek ∈ −D2 + rk . an outlier error ek resulting from εk . Denote by e ′ ) in the absence and in the presand en the solutions of (P1. we have mk = mk + c where εk is the ranging error. the probability of small values for ek may not be negligible. The maximum distance between the transmitter and the sensors is 2 [DU]. Of course.0 ) where ′ (P1. one can deﬁne a topology-dependent scalar D so that 0 ≤ rk + εk ≤ D for any k. as in TDOA. ˆ Then ek = −εk (2 pk − pt + εk ). (B). −α (2 + α )rk ∪ α rk . there are cases of errors in the measurements that do not appear as an error in the linear equations and therefore do not affect the correct solution. we plotted ˆ 1≤i≤N 1≤i≤Ne For a non-outlier entry we have e(k) = 0 and |en (k)| < f (ε ). y = Ax + e + n. If the k-th TOA measurement mk is impreˆ k εk cise. the bounds Be . NUMERICAL EXAMPLES In previous sections. The original problem can be written as (P1. hidden errors are not likely to occur since they must satisfy the condition ek = 0 described above. the error cannot be detected easily before emitter localization. In this example. Note that this measurement error corresponds to mk < t0 ˆ and therefore is likely to be removed from the TOA data set. Thus. denote by εk the k-th measurement error and ˆ by φk = φk + εk the k-th actual measurement. The experimental bound is Be = min ki . (2 − α )rk (10) min e e 1 s. Thus ′ we are interested in solving (P1. projection using F yields Fy = Fe + Fn and Fn ≤ ε . the solutions e and en have identical support provided that all the entries of en satisfying |en (k)| < f (ε ) are set to 0. D = 2 [DU]. the AOA k-th linear equation is {y}k = xk sin φk − yk cos φk where φk is the errorless AOA measurement. solving nonlinear localization problems by means of a linear set of equations ensures robustness to hidden outliers. we can justify the existence of a function f (ε ) so that lim f (ε ) = 0 and |e(k) − en (k)| < f (ε ) for any k. Thus. TOA or TDOA measurements by means of linear programming. ε →0 In the simulations the outlier error have been uniformly distributed within the interval deﬁned in (10). In the presence of perturbations. small errors have not been simulated. Thus. The bound Ba = N/6 is independent of the transmitter location. Thus. and we identify the smallest number of outliers for which identiﬁcation fails. the exact TOA measurements are mk = 1 c pk − pt and the corresponding linear equation is {y}k = pk 2 − c2 m2 .t. . If εk = −2 pk − pt then ek = 0. The result of the i-th experiment is denoted by ki . However. but do not affect the linear set of equations. (Be ). We show in [18] that ek = 2rk sin (εk /2)cos (εk /2 + φk − θk ). and under the assumption that t0 = 0. (Ba ). This model assumes that small errors do not ′ exist. according to section 3. Denote by e the estimate of e by ˆ ′ solving (P1 ) using linear programming. see [18] for details.ε ) with ε > 0.1 Furthermore. Assume the array has a radius of 1 distance unit [DU] and that the emitter is located within this array. the theoretical bound on the number of outliers. Range measurements with excessively large errors can be removed before processing. HIDDEN ERRORS We show now that some measurement errors do not affect the consistency of the linear equations set. we extend the concept of hidden outliers to any large εk for which ek is small. Thus. Fy − Fe 2 2 ≤ε In the presence of small errors we add a non-sparse vector n. meaning that this measurement error does not affect the linear equations and their solution.7. 9.The cosine term is zero for εk = 2(θk − φk ) ± π . Therefore.ε ) : Bound on the number of outliers 10 9 8 7 6 5 4 3 2 1 0 0 5 10 15 20 Number of sensors N 25 30 Be B Ba Figure 1: The experimental bound. By continuity of the penalty and constraint functions. for a given topology one can generally establish an upper bound on the largest distance between the transmitter and the sensors. The condition limits the number of identiﬁable outliers by an upper bound. In this section we examine the tightness of the upper bound by numerical simulations.ε ence of small errors.

52. No. vol. J. 1984. 100. 6.ituran. Donoho. pp. Mallat. RMSE of location estimates [DU] 0. The number of outliers s was successively set to 0. 12.1 0. [9] J. “On Sparse Representations in Pairs of Bases”. Abel. In Sections 4. 1. A. M. pp.il [2] D. 183-198. IEEE Trans. Donoho. 2005. Thin lines show results of straightforward LS while thick lines show results with outliers removal step. 2. S. C. “Time-frequency localization operators: A geometric phase space approach”. Tsaig. Fig. LS 0 outliers. pp. on Information Theory. CONCLUSION We have invoked the theory of sparse signal representation in the context of positioning. on pure and applied math. L. “A Passive Localization Algorithm and its Accuracy Analysis. Mar. June 2004. IEEE Trans. on Signal Proc. Cetin. TOA and TDOA data sets corrupted by unknown outliers can be corrected using linear programming. IEEE Trans. Mar. “Decoding by Linear Programming”. Jan. 50. Theory.25 Standard deviation σ of ranging measurement noise [DU] 0. No. pp. AOA. it is always signiﬁcantly advantageous to remove the outliers by linear programming before LS estimation.0. on Info. no. OE-12. “The Spherical Interpolation Method of Source Localization. [6] D. “Robust Uncertainty Principles: Exact Signal Reconstruction From Highly Incomplete Frequency Information”. Proc. A. Signal Proc. on Information Theory. no.1 0 0 0. pp. Fuchs. L. IEEE Trans. 2. no. Weiss. pp. 2008. J. 9. Chen. Note that σ = 0. vol. Tao. The measurements in the presence of noise (small errors) and outliers are given by y = Ax + e + n. 5 and 6 the limit given in (6) has been evaluated for several geometries. vol. no. 53. “For Most Large Underdetermined Systems of Linear Equations the Minimal ℓ1 -norm Solution is also the Sparsest Solution”. 1579-1581. [13] D. [17] E. LP+LS REFERENCES [1] http://www. D. 86. T. Daubechies. 49. “Matching Pursuit in a TimeFrequency Dictionary”. [3] I. 1. Gurbuz. “Compressive Wireless Arrays for Bearing Estimation”. 10. vol. OE-12. 41. The outliers size correspond to (10). Saunders. [18] J. In followup work we focus on outliers in the presence of small errors. ˆ A better way to estimate x may be to ﬁrst apply linear programming in order to estimate the error vector e + n. A. “A Sparse Signal Reconstruction Perspective for Source Localization With Sensor Arrays”. No. In the absence of outliers. J. pp. IEEE Trans. Romberg. 5. no.” IEEE Journal of Ocean Engineering. 2003. Malioutov. Friedlander. McClellan.J. pp. J. 40. 3. IEEE Trans. 1341-1344. 2497-2500. 4203-4215. 234-245.. 2197-2202. Vol. pp. Chellappa. . 51. vol. Picard. Z.. 1999. M. no. 7. “Accurate Geolocation in the Presence of Outliers using Linear Programming”. 246-252. S. [11] Y. 2005.3 0. Jan. pp. “Optimally Sparse Representation in General (Nonorthogonal) Dictionaries via Minimization”. Theory. Then apply LS to the modiﬁed data xe = A† (y − e + n). Mar. no. 533548. vol.. “Uncertainty Principles and Ideal Atomic Decomposition”. pp. L. X. 2845-2862. on Signal Proc. Signal Processing. 2006.5 0 outliers. pp. [12] E. Smith. SIAM Journal on Scientiﬁc Comp. provided that the number of outliers does not exceed a limit given in Theorem 1. M. 907-934. L. Jun. AES-20. IEEE Trans. Vol. 2006. vol. no. The LS estimate of x is given by x0 = A† y where A† is the pseudo-inverse of A. 33-61.. J. vol.2 0.3 [DU] is not negligible compared with the error-free measurements and therefore the number of outliers is occasionally increased.2 Localization in the Presence of Small Errors and Outliers In this section we examine the precision of least squares (LS) location estimates with and without outliers identiﬁcation.05 0. 605-612. 16% and 33% of the equations. Zhang. [7] S. T. pp. IEEE Trans. A. For each case. Tao.4 0. 8. Candes. “Statistical Theory of Passive Location System. Aug. “Breakdown of Equivalence between the Minimal ℓ1 -norm Solution and the Sparsest Solution”. LS 5 outliers. 47. on Aerospace and Electronic Systems. 7. Wiley. L. vol. Torrieri. 2001. Vol. The small ranging errors are realizations of zero-mean gaussian random variables with standard deviation σ . Feb. [16] J. regardless of the noise variance.” IEEE Trans. Huo. 5 and 10. Proc. Nemirovsky. a slight performance degradation may be observed if preprocessing is used. Dec. Theory.co. no. National Academy of Science. in the presence of 5 or 10 outliers. LP+LS 10 outliers. vol. Donoho. Inform. Nov. Note however that if no outlier is found the preprocessing can be ignored. 3010-3022.2 0. Donoho. 34. 6. [5] A. 2006. 489-509. vol. 1993. [15] B. Feuer. Donoho. [4] S. Cevher. Dec. pp. Willsky. 2 shows the RMSE of location estimates versus σ . July 1988. “Atomic Decomposition by Basis Pursuit”. vol. O. ICASSP 2008. 2. A.. [8] D. 59. pp. Theory. 1987. pp. 2009. on Info. corresponding to 0%. IEEE Trans. Elad.15 0. We have shown that in the absence of noise. D.. [14] V.” IEEE Journal of Ocean Eng. “On Sparse Representations in Arbitrary Redundant Bases”.3 Figure 2: RMSE of LS location estimates versus standard deviation of ranging small errors. According to Fig. S. Candes. submitted for publication. 1987. M. the exact expression of the bound and simple approximations were presented. S. LS 10 outliers. no. [10] D. Inf. 2003. 1. 3397-3415. LP+LS 5 outliers. Comm. J. H. ˆ We performed 200 experiments with a uniform circular array of radius 1 [DU] consisting of N = 30 TOA sensors.

- Continuum MechanicsUploaded bySambit Das
- Tut Problems 1Uploaded byMoumita Dey
- CHE 5480Uploaded byGelbert Silot
- alggeo_20130802eUploaded byIdrissou Fmsb
- Hopf Bif Phenomena BajajUploaded byDávid Borbély
- zebra_striping_paper_21Uploaded byshagcum1500
- 091367Uploaded bykiasaritopol
- Maneuvering & Control of Marine VehiclesUploaded byAnil Kumar Dash
- EF PseudocodeUploaded bynuwan samera
- The Extended Graetz Problem With Axial ConductionUploaded byDeny Arief Rusamsi
- 24330220 Maneuvering and Control of Marine VehiclesUploaded byphuocthontu
- A New Method for Determining Multiple Harmonic Source Locations in a Power Distribution SystemUploaded bylordcx
- 15. Ijhss - Port Selection Using Analytic Hierarchy ProcessUploaded byiaset123
- Class of IrregularityUploaded bySowmya Hanumanthaiah
- Suli E., Mayers D.-an Introduction to Numerical Analysis - SolutionsUploaded byGrant Rutherford
- Engineering MathematicsUploaded byanoop_nifty
- Wdowicki Paper 2Uploaded byReynaldo Wilson Flores López
- Matlab Commands for Control SystemUploaded byPavan Kumar Nadimpalli
- Social Security: wp98Uploaded bySocial Security
- Scientia Magna, Vol 7, No. 1, 2011Uploaded byscience2010
- SubstitutionUploaded byEuwan Tyrone Prias
- 10.1.1.144.4052 (1)Uploaded byГригорий Колесников
- IB Mathematical Methods III (Cambridge)Uploaded byucaptd3
- HOMOCLINIC BIFURCATIONSUploaded byManias Dimitris
- A Generalized Framework for Distrubute Power Control in Wireless NetworksUploaded byVirali Shah
- A very quick review of differential equations.Uploaded byHao Lian
- Kalmar-Nagy Subcritical Hopf Bifurcation in the Delay Equation Model for Machine Tool VibrationsUploaded bykalmarnagy
- Getting Started With MATLAB - David HartUploaded byapi-3696482
- ManuscriptUploaded byxfliu
- Qualitative Analysis of Linear Autonomous Systems With Constant CoefficientsUploaded byChristian Sarmiento