Distributed Sensor Network Localization from Local

Connectivity: Performance Analysis for the HOP-TERRAIN

Algorithm

Amin Karbasi Sewoong Oh

LTHC & LCAV Electrical Engineering Department
Schol of Compter and Communication Sciences Stanford Unviersity
EPFL, Switzereland Stanford, CA 94305, U.S.A
[email protected] [email protected]

ABSTRACT General Terms

This paper addresses the problem of determining the node Theory, Performance, Algorithms
locations in ad-hoc sensor networks when only connectiv-
ity information is available. In previous work, we showed Keywords
that the localization algorithm MDS-MAP proposed by Y.
Shang et al. is able to localize sensors up to a bounded er- distributed, localization, sensor network
ror decreasing at a rate inversely proportional to the radio
range r. The main limitation of MDS-MAP is the assump- 1. INTRODUCTION
tion that the available connectivity information is processed The problem we are interested in is determining the loca-
in a centralized way. tion of individual sensor nodes in a wireless ad-hoc sensor
In this work we investigate a practically important ques- network. This problem of localization plays an important
tion whether similar performance guarantees can be obtained role in wireless sensor networks when the positions of the
in a distributed setting. In particular, we analyze the per- nodes are not provided in advance. One way to acquire the
formance of the HOP-TERRAIN algorithm proposed by C. positions is by equipping all the sensors with a global posi-
Savarese et al. This algorithm can be seen as a distributed tioning system (GPS). This not only adds considerable cost
version of the MDS-MAP algorithm. More precisely, assume to the system, but more importantly, it does not work in in-
that the radio range r = o(1) and that the network consists door environments. As an alternative, we want an algorithm
of n sensors positioned randomly on a d-dimensional unit that can derive positions of sensors based on basic local in-
cube and d + 1 anchors in general positions. We show that formation such as proximity (which nodes are within com-
when only connectivity information is available, for every munication range of each others) or local distances (pairwise
unknown node i, the Euclidean distance between the esti- distances between neighboring sensors).
mate x̂i and the correct position xi is bounded by One consequence of the ad-hoc nature of the underly-
r0 ing networks is the lack of a central infrastructure. This
kxi − x̂i k ≤ + o(1), fact renders the use of common centralized positioning al-
r
gorithms [BY04, SRZF04]. We are interested in distributed
1
where r0 = Cd (log n/n) d for some constant Cd which only algorithms which can be employed on large-scale ad-hoc sen-
depends on d. Furthermore, we illustrate that a similar sor networks. Many distributed algorithms have been pro-
bound holds for the range-based model, when the approxi- posed where the success of them has mostly been measured
mate measurement for the distances is provided. experimentally [SLR02, NN03, SPS03]. To the best of our
knowledge, the theoretical guarantees associated with the
performance of these existing algorithms is rare with few
Categories and Subject Descriptors exceptions such as [AEG+ 06]. Such analytical bounds on
the performance of distributed localization algorithms can
C.2.1 [Computer Systems Organization]: Computer- provide answers to practical questions: for example, how
Communication Networks Network Architecture and Design large should the radio range be in order to get the error
[Wireless Communication] within a threshold? With this motivation in mind, our work
takes a step in this direction. In particular, our analysis fo-
cuses on providing a bound on the performance of a popular
robust positioning algorithm [SLR02] when applied to sen-
sor localization from only connectivity information, which
Permission to make digital or hard copies of all or part of this work for is a highly challenging problem. We prove that using Hop-
personal or classroom use is granted without fee provided that copies are
TERRAIN [SLR02], we are able to localize sensors up to
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to a bounded error in a connected network where most of the
republish, to post on servers or to redistribute to lists, requires prior specific pairwise distances are unknown and only local connectivity
permission and/or a fee. information is given.
SIGMETRICS’10, June 14–18, 2010, New York, New York, USA.
Copyright 2010 ACM 978-1-4503-0038-4/10/06 ...$10.00.
1.1 Related work proach of [OKM10] to analyze the performance of a truly dis-
tributed sensor localization algorithm. We show that Hop-
TERRAIN introduced in [SLR02] achieves a bounded error
Recently, a number of localization algorithms have been
when only local connectivity information is given.
proposed for sensor networks [SLR02, NN03, NSB03, RD07,
The organization of this paper is as follows. Section 2
SPS03, BY04, SRZF04, Sin08]. Based on the approach of
describes a distributed sensor localization algorithm known
processing the distance measurements, these algorithms can
as Hop-TERRAIN [SLR02] and states the main results on
be classified into two categories: centralized algorithms and
the performance. In Section 3, we provide detailed proof of
distributed algorithms. In centralized algorithms, all the
the main theorems.
distance measurements are sent to a single processor where
the estimated positions are computed. Two well-known cen-
tralized localization algorithms are multidimensional scaling 2. DISTRIBUTED LOCALIZATION ALGO-
(MDS) based approaches [SRZF04] and semidefinite pro- RITHM
gramming (SDP) based algorithms [BY04]. However, the
centralized algorithms typically have low energy efficiency In this section, we first define the basic mathematical
and low scalability due to dependency on a central proces- model for the sensors and the measurements. We then de-
sor and excessive communication overload. scribe the distributed localization algorithm and give a bound
on the error between the correct position of the sensors and
Perhaps a more practical and interesting case is when
the estimated positions returned by the algorithm.
there is no central infrastructure. [LR03] identifies a com-
mon three-phase structure of three popular distributed sen- 2.1 Model definition
sor localization algorithms, namely robust positioning [SLR02],
ad-hoc positioning [NN03] and N-hop multilateration [SPS03] Before discussing the distributed localization algorithm in
algorithm. Table 1 illustrates the structure of these algo- detail, we first define the mathematical model. First, we as-
rithms. In the first phase nodes share information to collec- sume that we have no fine control over the placement of the
tively determine the distances from each of the nodes to a sensors which we call the unknown nodes (e.g., the nodes
number of anchors. Anchors are special nodes with a priori are dropped from an airplane). Formally, n sensors, or un-
knowledge of their own position in some global coordinate known nodes, are distributed uniformly at random within
system. In the second phase nodes determine their position the d-dimensional hypercube [0, 1]d . Additionally, we as-
based on the estimated distances to the anchors provided by sume that there are m special sensors, which we call anchors,
the first phase and the known positions of the anchors. In with a priori knowledge of their own positions in some global
the last phase the initial estimated positions are iteratively coordinate. The anchors are assumed to be distributed uni-
refined. It is empirically demonstrated that these simple formly at random in [0, 1]d as well. However, it is reasonable
three-phase distributed sensor localization algorithms are to assume that we have some control over the positions of
robust and energy-efficient [LR03]. However, depending on these special sensors, or the anchors, and in the following
which method is used in each phase there are different trade- we show that we get similar performance bound with the
offs between localization accuracy, computation complexity minimum number of anchors (m = d + 1) if the positions of
and power requirements. In [NSB03], a distributed algo- the anchors are chosen properly.
rithm called the Gradient algorithm was proposed, which Let Va = {1, . . . , m} denote the set of m vertices corre-
is similar to ad-hoc positioning [NN03] but uses a different sponding to the anchors and Vu = {m+1, . . . , m+n} the set
method for estimating the average distance per hop. of n vertices corresponding to the unknown nodes. We con-
Another distributed approach introduced in [IFMW04] is sider the random geometric graph model G(n, r) = (V, E, P )
to pose the localization problem as an inference problem on where V = Vu ∪ Va , E ⊆ V × V is a set of undirected
a graphical model and solve it using Nonparametric Belief edges that connect pairs of sensors which are close to each
Propagation (NBP). It is naturally a distributed procedure other, and P : E → R+ is a non-negative real-valued func-
and produces both an estimate of sensor locations and a tion. We consider the function P as a mapping from a
representation of the location uncertainties. The estimated pair of connected nodes (i, j) to the approximate measure-
uncertainty may subsequently be used to determine the re- ment for the distance between i and j, which we call the
liability of each sensor’s location estimate. Despite recent proximity measurement. Let || · || be the Euclidean norm
developments in the algorithms for sensor localization lit- in Rd . Define a set of random positions of n + m sen-
tle is known about the theoretical analysis supporting their sors X = {x1 , . . . , xm , xm+1 , . . . , xm+n }, where xa ∈ Rd for
empirical simulation results. a ∈ {1, . . . , m} is the position of anchor a and and xi ∈ Rd
In an alternative line of work, the authors of [OKM10] for i ∈ {m + 1, . . . , m + n} is the position of unknown node i.
provided a bound on the performance of a centralized lo- We assume that only the anchors have a priori information
calization algorithm known as MDS-MAP. When n sensors about their own positions. Then, a common model for the
are randomly distributed in a unit d-dimensional hypercube, random geometric graph is the disc model where node i and
MDS-MAP is able to localize sensors up to a bounded er- j are connected if the Euclidean distance di,j = ||xi − xj || is
ror with only the pair-wise connectivity information, namely less than or equal to a positive radio range r. More formally,
whether two nodes are within a given radio range r or not. (i, j) ∈ E ⇔ di,j ≤ r .
This result crucially relies on the fact that there is a cen-
tral processor with access to the inter-sensor distance mea- To each edge (i, j) ∈ E, we associate the proximity mea-
surements. However, centralized algorithms suffer from the surement Pi,j between sensors i and j, which is a function
scalability problem and require higher computational com- of the distance di,j and random noise. In an ideal case when
plexity. Hence, a distributed algorithm with similar perfor- our measurements are exact, we have Pi,j = di,j . On the
mance bound is desirable. In this paper, we adapt the ap- other extreme, when we are given only network connectivity
 Table 1: Distributed localization algorithm classification
Phase Robust positioning Ad-hoc positioning N -hop multilateration
1. Distance DV-hop Euclidean Sum-dist
2. Position Lateration Lateration Min-max
3. Refinement Yes No Yes

Unknown sensors its own position. Moreover, communication is only possible

Anchors between adjacent neighboring nodes. The goal of distributed
sensor network localization is for each node to find its esti-
1 mated position that best fits the measured proximity with
small number of communications.
The main contribution of this paper is that, to the best of
our knowledge, we provide, for the first time, a performance
bound for a distributed sensor localization algorithm when
only the connectivity information is available. However, the
algorithm can be readily applied to the range-based model
without any modification. Further, given G(V, E, P ) from
from the range-based model we can easily produce G′ (V, E, P ′ )
′
where Pi,j = r for all (i, j) ∈ E. This implies that the per-
formance under the range-based model is also bounded by
the main result in Theorems 2.1 and 2.2.
r
2.2 Algorithm
0 1 Based on the robust positioning algorithm introduced in
[SLR02], the distributed sensor localization algorithm con-
sists of two steps :
Figure 1: An example of a random geometric graph
model in two dimensions where unknown sensors and Algorithm : Hop-TERRAIN [SLR02]
anchors are distributed randomly. A node is connected 1: Each node i computes the shortest paths
to all other nodes that are within distance r of itself. {dˆi,a : a ∈ Va } between itself and the anchors;
2: Each node i derives an estimated position x̂i
by triangulation with a least squares method.

information and no distance information, we have constant According to the three phase classification presented in Ta-
Pi,j ’s for all (i, j) ∈ E (see Figure 1). ble 1, this is closely related to the first two phases of the
The sensor localization algorithms can be classified into robust positioning algorithm. This algorithm uses a slightly
two different categories. For the connectivity-based model, different method for computing the shortest paths, which is
which is alternatively also known as the range-free model, compared in detail later in this section. Hence, through out
only the connectivity information is available. Formally, this paper, we refer to this algorithm as Hop-TERRAIN,
 which denotes the first two steps of robust positioning algo-
r if (i, j) ∈ E, rithm in [SLR02].
Pi,j =
∗ otherwise, Distributed shortest paths. The goal of the first step
is for each of the unknown nodes to estimate the distances
where a ∗ denotes that di,j > r. between itself and the anchors. This approximate distances
For the range-based model, which is also known as the will be used in the next triangulation step to derive an es-
range-aware model, the distance measurement is available timated position. The shortest path between an unknown
but may be corrupted by noise or have limited accuracy. node i and an anchor a in the graph G provides an estimate

[di,j + zi,j ]+ if (i, j) ∈ E, for the Euclidean distance di,a = ||xi − xa ||, and for a care-
Pi,j = fully chosen radio range r this shortest path estimation is
∗ otherwise,
close to the actual distance as will be shown in Lemma 3.1.
where zi,j models the measurement noise (in the noiseless Formally, the shortest path between an unknown node i
case zi,j = 0), possibly a function of the distance di,j , and and an anchor a in the graph G = (V, E, P ) is defined as a
[a]+ ≡ max{0, a}. Common examples are the additive Gaus- path between two nodes such that the sum of the proximity
sian noise model, where the zi,j ’s are i.i.d. Gaussian ran- measures of its constituent edges is minimized. We denote
dom variables, and the multiplicative noise model, where by dˆi,a the computed shortest path and this provides the
Pi,j = [(1 + Ni,j )di,j ]+ , for independent Gaussian random initial estimate for the distance between the node i and the
variables Ni,j ’s. anchor a. When only the connectivity information is avail-
In distributed sensor network localization not all the able and the corresponding graph G = (V, E, P ) is defined
information is available at each node. Given the graph as in the connectivity-based model, the shortest path dˆi,a is
G(n, r) = (V, E, P ) with associated proximity measurements equivalent to the minimum number of hops between a node
for each edges in E, we assume that each of the nodes is i and an anchor a multiplied by the radio range r.
aware of the proximity measurements between itself and its In order to find the minimum number of hops from an
adjacent neighbors and each of the anchors is also aware of unknown node i ∈ Vu to an anchors a ∈ Va in a distributed
way, we use a method similar to DV-hop [NN03]. Each Triangulation using least squares. In the second step,
unknown node maintains a table {xa , ha } which is initially each unknown node i ∈ Vu uses a set of estimated distances
empty, where xa ∈ Rd refers to the position of the anchor a {dˆi,a : a ∈ Va } together with the known positions of the
and ha to the number of hops from the unknown node to the anchors to perform a triangulation. the resulting estimated
anchor a. First, each of the anchors initiate a broadcast con- position is denoted by x̂i . For each node, the triangula-
taining its known location and a hop count of 1. All of the tion consists of solving a single instance of a least squares
one-hop neighbors surrounding the anchor, on receiving this problem (Ax = b) and this process is known as Lateration
broadcast, record the anchor’s position and a hop count of 1, [SRB01, LR03].
and then broadcast the anchor’s known position and a hop For an unknown node i, the position vector xi and the
count of 2. From then on, whenever a node receives a broad- anchor positions {xa : a ∈ {1, . . . , m}} satisfy the following
cast, it does one of the two things. If the broadcast refers to series of equations:
an anchor that is already in the record and the hop count
is larger than or equal to what is recorded, then the node ||x1 − xi ||2 = d2i,1 ,
does nothing. Otherwise, if the broadcast refers to an an- ..
chor that is new or has a hop count that is smaller, the node .
updates its table with this new information on its memory ||xm − xi ||2 = d2i,m .
and broadcast the new information after incrementing the
hop count by one. When every node has computed the hop
count to all the anchors, the number of hops is multiplied This set of equations can be linearized by subtracting each
by the radio range r to estimate the distances between the line from the next line.
node and the anchors. Note that to start triangulation, not
||x2 ||2 − ||x1 ||2 + 2(x1 − x2 )T xi
all the hop counts to all the anchors are necessary. A node
can start triangulation as soon as it has estimated distances = d2i,2 − d2i,1 ,
to d + 1 anchors. There is an obvious trade-off between ..
number of communications and performance. .
The above step of computing the minimum number of
hops is the same distributed algorithm as described in DV- ||xm ||2 − ||xm−1 ||2 + 2(xm−1 − xm )T xi
hop. However, the main difference is that instead of mul- = d2i,m − d2i,m−1 .
tiplying the number of hops by a fixed radio range r, in
DV-hop, the number of hops is multiplied by an average By reordering the terms, we get a series of linear equations
(i)
hop distance. The average hop distance is computed from for node i in the form A xi = b0 , for A ∈ R(m−1)×d and
m−1
the known pairwise distances between anchors and the num- b∈R defined as
ber of hops between the anchors. While numerical simula- 2
2(x1 − x2 )T
3
tions show that the average hop distance provides a better ..
A ≡ 4 5 ,
6 7
estimate, the difference between the computed average hop .
distance and the radio range r becomes negligible as n grows 2(xm−1 − xm )T
large.
||x1 ||2 − ||x2 ||2 + d2i,2 − d2i,1
2 3
We are interested in a scalable system of n unknown nodes
(i)
b0 ≡ 4 ..
for large value of n. As n grows large, it is reasonable to 5 .
6 7
.
assume that the average number of connected neighbors for ||xm−1 ||2 − ||xm ||2 + d2i,m − d2i,m−1
each node should stay constant. This happens, in our model,
if we chose the radio range r = C/n1/d . However, the num-
ber of hops is well defined only if the graph G is connected. Note that the matrix A does not depend on the particular
If G is not connected there might be a set of unknown nodes unknown node i and all the entries are known exactly to all
that are connected to too few anchors, resulting in under- the nodes after the distributed shortest paths step. However,
(i)
determined series of equations in the triangulation step. In the vector b0 is not available at node i, since di,a ’s are not
the unit square, assuming sensor positions are drawn uni- known. Hence we use an estimation b(i) , which is defined
(i)
formly at random as define in the previous section, the graph from b0 by replacing di,a by dˆi,a everywhere. Then, finding
is connected, with high probability, if πr 2 > (log n + cn )/n the optimal estimation x̂i of xi that minimizes the mean
for cn → ∞ [GK98]. A similar condition can be derived squared error is solved in a closed form using a standard
for generic d-dimensions as Cd r d > (log n + cn )/n, where least squares approach:
Cd ≤ π is a constant that depends on d. Hence, we focus
in the regime where the average number of connected neigh- x̂i = (AT A)−1 AT b(i) . (1)
bors is slowly increasing with n, namely, r = α(log n/n)1/d For bounded d = o(1), a single least squares has complex-
for some positive constant α such that the graph is connected ity O(m), and applying it n times results in the overall com-
with high probability. plexity of O(n m). No communication between the nodes is
As will be shown in Lemma 3.1, the key observation of necessary for this step.
the shortest paths step is that the estimation is guaranteed
to be arbitrarily close to the correct distance for properly 2.3 Main results
chosen radio range r = α(log n/n)1/d and large enough n.
Our main result establishes that Hop-TERRAIN [SLR02]
Moreover, this distributed shortest paths algorithm can be
achieves an arbitrarily small error for a radio range r =
done efficiently with total complexity of O(n m).
α(log n/n)1/d with a large enough constant α, when we
have only the connectivity information as in the case of the
connectivity-based model. The same bound holds immedi-
ately for the range-based model, when we have an approx-
imate measurements for the distances, and the same algo- Unknown sensors
rithm can be applied without any modification. The extra Deterministic anchors
information can be readily incorporated into the algorithm
to compute better estimates for the actual distances between
the unknown nodes and the anchors.
Define
«1
√
„
log n d
r0 ≡ 8 3 d3/2 , (2)
n

Theorem 2.1. Assume n sensors and m anchors are dis-

tributed uniformly at random in the [0, 1]d hypercube for a r
bounded dimension d = O(1). For a given radio range r > r0
and the number of anchors m = Ω(log n), the following is
true with high probability. For all unknown nodes i ∈ Vu , the
Euclidean distance between the estimate x̂i given by Hop-
TERRAIN and the correct position xi is bounded by
r0
||xi − x̂i || ≤ + O(r) , (3) Figure 2: Three anchors in fixed positions for a two-
r
dimensional sensor localization.
where r0 is defined in Eq. (2).

The proof is provided in Section 3. In the regime where

r = o(1), the above theorem implies that the error is in- The proof of this remark closely follows that of Theorem
versely proportional to the radio range r. As described in 2.1, and is provided in Section 3. However, there is nothing
the previous section, we are interested in the regime where particular about the position of the anchors in unit vectors.
r = α(log n/n)1/d for some constant α. Given a small posi- Any d+1 anchors in general position will give similar bound.
tive constant δ, this implies that Hop-TERRAIN is guaran- Only difference is that the constant term in the definition of
teed to produce estimated positions that satisfy ||xi − x̂i || ≤ r1 changes with the anchor positions.
δ for all i ∈ Vu with a large enough constant α and large When r = α(log n/n)1/d for some positive parameter α,
enough n. p
the error bound in (4) is ||xi − x̂i || ≤ C1 /α + C2 α log n/n
When the number of anchors is bounded and the posi- for some numerical constants C1 and C2 . The first term is
tions of the anchors are chosen randomly, it is possible that, inversely proportional to α and is independent of n, where
in the triangulation step, we get an ill-conditioned matrix as the second term is linearly dependent in α and vanishes
AT A, resulting in an large estimation error. This happens, as n grows large. This is illustrated in Figure 3, which shows
for instance, if three anchors fall close to a line. However, numerical simulations with n sensors randomly distributed
as mentioned in the previous section, it is reasonable to as- in the 2-dimensional unit square. We used 100 random sam-
sume that, for the anchors, the system designer has some ples for each point inP the figure to compute the root mean
control over where they are placed. In that case, the next squared error, {(1/n) n 2 1/2
i=1 ||xi − x̂i || } , averaged over the
remark shows that when the positions of anchors are prop- 100 samples. In the first figure, where we have only the con-
erly chosen, only d + 1 anchors suffice to get a similar bound nectivity information, we can see both contributions: the
on the performance. Note that this is the minimum num- linear term which depends on n and the 1/α term which
ber of anchors necessary for triangulation. For simplicity is less sensitive to n. In the second figure, where we have
we assume that one anchor is placed at the origin and d an- the exact distance measurements without noise, we can see
chors are placed at positions corresponding to d-dimensional that the linear term almost vanishes even for n = 5000,
unit vectors. Namely, the position of the d + 1 anchors are and overall error is much smaller. A network of n = 200
{[0, . . . , 0], [1, 0, . . . , 0], [0, 1, 0, . . . , 0], [0, . . . , 0, 1] }. (see fig- nodes randomly placed in the unit square is shown in Fig-
ure 2) ure 4. Three anchors in fixed positions are displayed by solid
blue circles. In this experiment the distance measurements
Theorem 2.2. Assume that n sensors are distributed uni-
are from the range based model with multiplicative noise,
formly at random in the [0, 1]d hypercube for a bounded di-
where Pi,j = [(1 + Ni,j )di,j ]+ for i.i.d. Gaussian Ni,j with
mension d = O(1). Also, assume that there are d + 1 an-
zero mean and variance 0.1. The noisy distance measure-
chors, one of which is placed at the origin, and the position
mentpis revealed only between the nodes within radio range
vectors of the d remaining anchors are the d-dimensional
r = 0.8 log n/n. Figure 5 shows the final solution of Hop-
unit vectors. For a given radio range r > r0 , the follow-
TERRAIN. The circles represent the correct positions and
ing is true with high probability. For all unknown nodes
the solid line represents the errors of the estimation from
i ∈ Vu , the Euclidean distance between the estimate x̂i given
the correct position. The average error in this example is
by Hop-TERRAIN and the correct position xi is bounded
about 0.075.
by
r1 2.4 Some technical remark and notations
||xi − x̂i || ≤ + O(r) , (4)
r In the following, whenever we write that a property A
√
where r0 is defined in Eq. (2) and r1 ≡ d r0 /2 3. holds with high probability (w.h.p.), we mean that P(A)
 0.14
n=5,000 1

0.13 n=10,000
0.9
Average Error

0.12 0.8

0.11 0.7

0.6
0.1
0.5
0.09
0.4
0.08
0.3
0.07
0.2

0.06 0.1
1 1.5 2 2.5 3 3.5 4 4.5
α 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.009
n=5,000
0.008 n=10,000

0.007
Figure 4: 200 nodes randomly placed in the unit square
Average Error

and
p 3 anchors in fixed positions. The radio range is r =
0.006 0.8 ∗ log n/n.
0.005
0.004
enough radio range r. Define
0.003
«1
√
„
0.002 (1 + β) log n d
r̃0 (β) ≡ 8 d . (5)
n
0.001
For simplicity we denote r̃0 (0) by r̃0 .
0
1 1.5 2 2.5 3 3.5 4 4.5 Lemma 3.1. (Bound on the shortest paths) Under the hy-
α pothesis of Theorem 2.1, w.h.p., the shortest paths between
Figure 3: Average distance between the correct position all pairs of nodes in the graph G(V, E, P ) are uniformly
{xi } and estimation p
{x̂i } using Hop-TERRAIN as a func- bounded by
tion of α, for r = α log n/n with n sensors in the unit r̃0
dˆ2i,j ≤ (1 + )d2i,j + O(r) ,
square under connectivity based model (above) and range r
based model (below). for r > r0 , where r̃0 is defined as in Eq. (5) and r0 as in
Eq. (2).
The proof of this lemma is given in Section 3.1. Since d2i,j ≤
d for all i and j, we have
approaches 1 as the number of sensors n goes to infinity.
Given a matrix A ∈ Rm×n , the spectral norm of A is denoted “ m−1 ´2 ”1/2
dk,2 − d2k,1 − dˆ2k,2 + dˆ2k,1
(i)
X` 2
by ||A||2 , and the Frobenius norm is denoted by ||A||F . For a ||b0 − b(i) || =
vector a ∈ Rn , ||a|| denotes the Euclidean norm. Finally, we k=1

use C to denote generic constants that do not depend on n. √ r̃0 √

≤ m−1 d + O( mr) , (6)
r
Next, to bound ||(AT A)−1 AT ||2 , we use the following lemma.
3. PROOF OF THE MAIN THEOREMS Lemma 3.2. Under the hypothesis of Theorem 2.1, the
In this section we provide the proofs of the theorems 2.1 following are true. Assuming deterministic anchor model,
and 2.2. Detailed proofs of the technical lemmas are pro- where m = d + 1 anchors are placed on the positions x1 =
vided in the following sections. [1, 0, . . . , 0], x2 = [0, 1, 0, . . . , 0], x3 = [0, . . . , 0, 1] and xm =
For an unknown node i, the estimation x̂i is given in [0, 0, . . . , 0].
Eq. (1). d
||(AT A)−1 AT ||2 ≤
.
2
(i)
||xi − x̂i || = ||(AT A)−1 AT b0 − (AT A)−1 AT b(i) || Assuming random anchor model, for m = Ω(log n) anchors
(i)
≤ ||(AT A)−1 AT ||2 ||b0 − b(i) || , chosen uniformly at random among n sensors,
r
3
(i)
First, to bound ||b0 − b(i) ||, we use the following key result ||(AT A)−1 AT ||2 ≤ ,
m−1
on the shortest paths. The main idea is that, for sensors
with high probability.
with uniformly random positions, shortest paths provide ar-
bitrarily close estimation to the correct distances for large and this finishes the proof of Theorems 2.1 and 2.2.
 function F : R+ → R+ as
1 
r if y ≤ r ,
0.9 F (y) =
(k + 1)r if y ∈ Lk for k ∈ {1, 2, . . .} ,
0.8 √
where L√k denotes the interval (r + (k − 1)(r − 2 d)δ, r +
0.7
k(r − 2 d)δ].
0.6 We will show, by induction, that for all pairs (i, j),

dˆi,j ≤ F (di,j ) .
0.5
(11)
0.4
First, in the case when di,j ≤ r, by definition of connectivity
0.3
based model, nodes i and j are connected by an edge in the
0.2 graph G, whence dˆi,j = r.
Next, assume that the bound in Eq. √ (11) is true for all
0.1
pairs (l, m) with dl,m ≤ r + (k − 1)(r − 2 d)δ. We consider
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
two nodes i and j at distance di,j ∈ Lk . Consider a line
segment li,j in a d-dimensional space with one end at xi and
the other at xj , where xi and xj denote the positions of
nodes i and j, respectively. √Let p be the point in the line
Figure 5: Location estimation using Hop-TERRAIN.
li,j which is at distance r − dδ from xi . Then, we focus
on the bin that contains p. By assumption that no bins are
empty, we know that we can always find at least one node
3.1 Proof of the bound on the shortest paths in the bin. Let k denote any one of those nodes in the bin.
Then we use following inequality which is true for all nodes
In this section, we prove a slightly stronger version of
(i, k, j).
Lemma 3.1. We will show that under the hypothesis of
Theorem 2.1, for any β ≥ 0 and all sensors i 6= j, there dˆi,j ≤ dˆi,k + dˆk,j .
exists a constant C such that, with probability larger than
Cn−β
1 − (1+β) , the shortest paths between all the pairs of Each of these two terms can be bounded using triangular
log n
nodes are uniformly bounded by inequality. To bound the first term, √ note that two nodes in
the same bin are at most distance dδ apart. √ Since p and xk
r̃0 (β) 2 are in the same bin and p is at distance r − dδ from node xi
dˆ2i,j ≤ (1 + )di,j + O(r) , (7)
r by construction, we know that di,k ≤ ||xi −p||+||p−xk || ≤ r,
for r > r0 , where r̃0 (β) is defined as in Eq. (5) and r0 as in whence dˆi,k = r. Analogously for the second√ term, dk,j ≤
Eq. (2). Especially, Lemma 3.1 follows if we set β = 0. ||xj − p|| + ||p − xk || ≤ r + (k − 1)(r − 2 d)δ, which implies
We start by applying bin-covering technique to the ran- that dˆk,j ≤ F (dk,j ) = kr. Hence, we proved that if Eq.√(11)
dom geometric points in [0, 1]d in a similar way as in [MP05]. is true for pairs (i, j) such that di,j ≤ r + (k − 1)(r − 2 d)δ,
We cover the space [0, 1]d with a set of non-overlapping hy- then dˆi,j ≤ (k + 1)r for pairs (i, j) such that di,j ∈ Lk . By
percubes whose edge lengths are δ. Thus there are total induction, this proves that the bound in Eq. (11) is true for
⌈1/δ⌉d bins, each of volume δ d . In formula, bin (i1 , . . . , id ) all pairs (i, j).
is the hypercube [(i1 − 1)δ, i1 δ) × · · · × [(id − 1)δ, id δ), for √ “ ”1/d
ik ∈ {1, . . . , ⌈1/δ⌉} and k ∈ {1, . . . , d}. When n nodes are Let µ = (r/2 d) n/((1 + β) log n) . Then, the func-
deployed in [0, 1]d uniformly at random, we say a bin is tion F (y) can be easily upper bounded by an affine function
empty if there is no node inside the bin. We want to en- Fa (y) = (1+1/(µ−1))y +r. Hence we have following bound
sure that, with high probability, there are no empty bins. on the shortest paths between any two nodes i and j.
For a given δ, define a parameter β ≡ (δ d n/ log n) − 1. 0 1
Since a bin is empty with probability (1 − δ d )n , we apply 1
union bound over all the ⌈1/δ⌉d bins to get, dˆi,j ≤ @1 + A di,j + r . (12)
B C
“ ”1/d
r n
l 1 md √
2 d (1+β) log n
−1
P(no bins are empty) ≥ 1 − (1 − δ d )n (8)
δ Figure 6 illustrates the comparison of the upper bounds
Cn “ (1 + β) log n ” n
F (di,j ) and Fa (di,j ), and the trivial lower bound dˆi,j ≥ di,j
≥1− 1− (9)
(1 + β) log n n in a p simulation with parameters d = 2, n = 6000 and
C n−β r = 64 log n/n. The simulation data shows the distri-
≥1− , (10) bution of shortest paths between all pairs of nodes with re-
(1 + β) log n
spect to the actual pairwise distances, which confirms that
where in (9) we used the fact that there exists a constant C shortest paths lie between the analytical upper and lower
such that ⌈1/δ⌉d ≤ C/δ d , and in (10) we used (1 − 1/n)n ≤ bounds. While the gap between the upper and p lower bound
e−1 , which is true for any positive n. is seemingly large, in the regime where r = α log n/n with
Assuming that, with high probability, no bins are empty, a constant α, the vertical gap r vanishes as n goes to infinity
we first show that the shortest paths is bounded by a func- and the slope of the affine upper bound can be made arbi-
tion F (di,j ) that only depends on the distance between the trarily small by increasing the radio range r or equivalently
two nodes. Let dˆi,j denote the shortest path between nodes taking large enough α. The bound on the squared shortest
i and j and di,j denote the Euclidean distance. Define a paths dˆ2i,j can be derived from the bound on the shortest
 2.5 3.2.1 Deterministic Model
simulation data
affine upper bound By putting the sensors in the mentioned positions the d×d
upper bound matrix A will be Toeplitz and have the following form.
2 lower bound 2 3
1 −1 0 · · · 0
6 0 1 −1 · · · 0 7
1.5 6 . . .. 7 .
6 7
A = 2 6 .. .. .. ..
dˆi,j 6 . . . 7
7
4 0 ··· 0 1 −1 5
1
0 ··· 0 0 1
We can easily find the inverse of matrix A.
0.5 2 3
1 1 1 ··· 1
6 0 1 1 ··· 1 7
0 16
A−1 = 6 ... .. .
7
.. ..
0 0.5 1 1.5 . .. 7.
6 7
26 . .
di,j 4 0 ··· 0 1 1
7
5
0 ··· 0 0 1
Figure 6: comparison of upper and lower bound of
shortest paths {dˆi,j } with respect to the correct distance Note that the maximum singular value of A−1 and the min-
{di,j } computed for n = 6000 sensors in 2-dimensional imum singular value of A are related as follows.
square [0, 1]2 under connectivity based model.
1
σmin (A) = . (18)
σmax (A−1 )

paths in Eq. (12) after some calculus. To find the maximum singular value of A−1 need to calculate
´T
the maximum eigenvalue of A−1 A−1 which has the form
`
n µ o2
dˆ2i,j ≤ di,j + r (13) 2
d d − 1 d −2 ··· 1
3
µ−1
6 d − 1 d − 1 d −2 ··· 1 7
µ2 µ 16
d2i,j + r 2 + 2 .. .. .. 7 .
´T 7
= di,j r (14) A−1 A−1 = 6
` 6 .. ..
(µ − 1)2 µ−1 46 . . . . . 7
7
“ 2µ − 1 ” 4 2 ··· 2 2 1 5
= 1+ d2i,j + O(r) (15) 1 ··· 1 1 1
(µ − 1)2
“ 4” 2 Using the Gershgorin circle theorem (see appendix B) we
≤ 1+ di,j + O(r) . (16)
µ can find an upper bound on the maximum eigenvalue of
´T
A−1 A−1 .
`
where in (15) we used the fact that (µ/(µ√ − 1))di,j = O(1),
which follows from the assumptions (r/4 d)d > (1+β) log n/n “ ´T ” d2
λmax A−1 A−1
`
and d = O(1). In (16), we used the inequality (2µ − 1)/(µ − ≤ , (19)
√ 4
1)2 ≤ 4/µ, which is true for µ ≥ 2 + 3. Substituting µ in
the formula, this finishes the proof of the desired bound in Hence, by combining (17) and (19) we get
Eq. (7). d
Note that although for the sake of simplicity, we focus ||(AT A)−1 A||2 ≤ . (20)
2
on [0, 1]d hypercube, our analysis easily generalizes to any
bounded convex set and homogeneous Poisson process model 3.2.2 Random Model
with density ρ = n. The homogeneous Poisson process Let the symmetric matrix B be defined as AT A. The
model is characterized by the probability that there are diagonal entries of B can be written as
exactly k nodes appearing in any region with volume A : m−1
k X
P(kA = k) = (ρA) k!
e−ρA . Here, kA is a random variable bi,i = 4 (xk,i − xk+1,i )2 , (21)
defined as the number of nodes in a region of volume A. k=1

for 1 ≤ i ≤ d and the off-diagonal entries as

3.2 Proof of Lemma 3.2
m−1
By using the singular value decomposition of a tall m−1× X
d matrix A, we know that it can be written as A = U ΣV T bi,j = 4 (xk,i − xk+1,i )(xk,j − xk+1,j ), (22)
k=1
where U is an orthogonal matrix, V is a unitary matrix and
Σ is a diagonal matrix. Then, for 1 ≤ i 6= j ≤ d where xk,i is the i-th element of vector xk .
T −1 −1 T In the following lemmas, we show that with high probability
(A A) A = UΣ V . as m increases the diagonal entries of B will all be of the
Hence, order of m, i.e., bi,i = Θ(m), and the off-diagonal entries
1
1 will be bounded from above by m 2 +ǫ , i.e., bi,j = o(m).
||(AT A)−1 A||2 = , (17)
σmin (A) Lemma 3.3. For any ǫ > 0 the diagonal entries of B are
where σmin (A) is the smallest singular value of A. This bounded as follows.
means that in order to upper bound ||(AT A)−1 A||2 we need 2ǫ
“ 1
”
P |bi,i − 2(m − 1)/3| > 4m 2 +ǫ ≤ 4e−m .
to lower bound the smallest singular value of A.
The idea is to use Hoeffding’s Inequality (see appendix A) for all 1 ≤ i ≤ d. As m grows, the RHS of (28) can be lower
for the sum of independent and bounded random variables. bounded by (m − 1)/3. By combining (27) and (28) we can
To this end, we need to divide the sum in (21) into sums of conclude that
even and odd terms as follows: „
(m − 1)
«
2ǫ
P λmin (B) ≥ ≥ 1 − 3d2 e−m . (29)
bi,i = bie + bio , 3
where As a result, from (17) and (29) we have
X
bie 2 !
= 4 (xk,i − xk+1,i ) , (23)
r
T 3 2ǫ
k∈even
−1
P ||(A A) A||2 ≤ ≥ 1 − 3d2 e−m , (30)
X m−1
bio = 4 (xk,i − xk+1,i ) . 2
(24)
k∈odd which shows that q
as m grows with high probability we have
This separation ensures that the random variables in sum- ||(AT A)−1 A||2 ≤ m−1
3
.
mations (23) and (24) are independent. Let the random
variable zki denote the term 4(xk,i − xk+1,i )2 in (23). Since 4. CONCLUSION
zki ∈ [0, 4] and all the terms in bie are independent of each In many applications of wireless sensor networks, it is cru-
other, we can use Hoeffding’s Inequality to upper bound the cial to determine the location of nodes. Distributed localiza-
probability of the deviation of bie from its expected value: tion of nodes is a key to enable most of these applications.
2ǫ For this matter, numerous algorithms have been recently
“ 1
”
P |bie − (m − 1)/3| > 2m 2 +ǫ ≤ 2e−m , (25)
proposed where the efficiency and success of them have been
for any fixed ǫ > 0. The same bound holds for bo . Namely, mostly demonstrated by simulations. In this paper, we in-
“ 1
” 2ǫ
vestigated the distributed sensor localization problem from a
P |bio − (m − 1)/3| > 2m 2 +ǫ ≤ 2e−m . (26) theoretical point of view and provided analytical bounds on
the performance of such an algorithm. More precisely, we
Hence, analyzed the HOP-TERRAIN algorithm and showed that
1 even when only the connectivity information was given, the
“ ”
P |bi,i − 2(m − 1)/3| > 4m 2 +ǫ
Euclidean distance between the estimate and the correct po-
(a)“ 1
” sition of every unknown node is bounded and decays at a
≤ P |be − (m − 1)/3| + |bo − (m − 1)/3| > 4m 2 +ǫ
rate inversely proportional to the radio range. In the case of
(b) 2ǫ noisy distance measurements, we observe the same behavior
≤ 4e−m , and a similar bound holds.
where in (a) we used triangular inequality and in (b) we
used the union bound. Acknowledgment
Lemma 3.4. For any ǫ > 0 the off-diagonal entries of B We thank Andrea Montanari and Rüdiger Urbanke for stim-
are bounded as follows. ulating discussions on the subject of this paper.
2ǫ
“ 1
”
P |bi,j | > 16m 2 +ǫ ≤ 4e−m .
5. REFERENCES
The proof follows the same lines as the proof of Lemma 3.3. [AEG+ 06] J. Aspnes, T. Eren, D. K. Goldenberg, A. S.
Using the Gershgorin circle theorem (see appendix B) we Morse, W. Whiteley, Y. R. Yang, B. D. O.
can find a lower bound on the minimum eigenvalue of B. Anderson, and P. N. Belhumeur. A theory of
network localization. IEEE Transactions on
λmin (B) ≥ min(bi,i − Ri ), (27)
i Mobile Computing, 5(12):1663–1678, December
where 2006.
X [BY04] P. Biswas and Y. Ye. Semidefinite programming
Ri = |bi,j |. for ad hoc wireless sensor network localization.
j6=i
In IPSN ’04: Proceedings of the 3rd
Now, let Bii denote the event that {bi,i < 2(m − 1)/3 − international symposium on Information
1
4m 2 +ǫ } and Bij (for i 6= j) denote the event that {bi,j > processing in sensor networks, pages 46–54,
1 New York, NY, USA, 2004. ACM.
16m 2 +ǫ }. Since the matrix B is symmetric, we only have
d(d + 1)/2 degrees of freedom. Lemma 3.3 and 3.4 provide [GK98] P. Gupta and P.R. Kumar. Critical power for
us with a bound on the probability of each event. Therefore, asymptotic connectivity. In Proceedings of the
by using the union bound we get 37th IEEE Conference on Decision and Control,
0 1 volume 1, pages 1106–1110 vol.1, 1998.
[ X [HJ85] R. Horn and C. Johnson. Matrix analysis.
P @ Bij A ≤ 1 − P(Bij ) Cambridge University Press, 1985.
i≤j i≤j [Hoe63] W. Hoeffding. Probability inequalities for sums
= 1 − 3d2 e−m .
2ǫ
of bounded random variables. Journal of the
2ǫ
American Statistical Association, 58:13âĂŞ30,
Therefore with probability at least 1 − 3d2 e−m we have 1963.
2(m − 1) 1 [IFMW04] A. T. Ihler, J. W. Fisher, R. L. Moses, and
bi,i − Ri ≥ − 16d · m 2 +ǫ , (28) A. S. Willsky. Nonparametric belief propagation
3
 for self-calibration in sensor networks. In IPSN [SPS03] A. Savvides, H. Park, and M. Srivastava. The
’04: Proceedings of the 3rd international n-hop multilateration primitive for node
symposium on Information processing in sensor localization problems. Mob. Netw. Appl.,
networks, pages 225–233, New York, NY, USA, 8(4):443–451, 2003.
2004. ACM. [SRB01] C. Savarese, J. Rabaey, and J. Beutel.
[LR03] K. Langendoen and N. Reijers. Distributed Locationing in distributed ad-hoc wireless
localization in wireless sensor networks: a sensor networks. In in ICASSP, pages
quantitative comparison. Comput. Netw., 2037–2040, 2001.
43(4):499–518, 2003. [SRZF04] Y. Shang, W. Ruml, Y. Zhang, and M. P. J.
[MP05] S. Muthukrishnan and G. Pandurangan. The Fromherz. Localization from connectivity in
bin-covering technique for thresholding random sensor networks. IEEE Trans. Parallel Distrib.
geometric graph properties. In SODA ’05: Syst., 15(11):961–974, 2004.
Proceedings of the sixteenth annual ACM-SIAM
symposium on Discrete algorithms, pages APPENDIX
989–998, Philadelphia, PA, USA, 2005.
[NN03] D. Niculescu and B. Nath. DV based A. HOEFFDING’S INEQUALITY
positioning in ad hoc networks. Journal of Hoeffding’s inequality [Hoe63] is a result in probability
Telecommunication Systems, 22:267–280, 2003. theory that gives an upper bound on the probability for
[NSB03] R. Nagpal, H. Shrobe, and J. Bachrach. the sum of random variables to deviate from its expected
Organizing a global coordinate system from value. Let z1 , z2 , . . . , zn be independent and bounded ran-
local information on an ad hoc sensor network. dom variables
P such that zk ∈ [ak , bk ] with probability one.
In IPSN ’03: Proceedings of the 2nd Let sn = n k=1 zk . Then for any δ > 0, we have
international conference on Information
− Pn 2δ2
processing in sensor networks, pages 333–348, P (|sn − E[sn ]| ≥ δ) ≤ 2e k=1
(bk −ak )2
.
2003.
[OKM10] S. Oh, A. Karbasi, and A. Montanari. Sensor
network localization from local connectivity: B. GERSHGORIN CIRCLE THEOREM
Performance analysis for the mds-map The Gershgorin circle theorem [HJ85] identifies a region
algorithm. In 2010 IEEE Information Theory in the complex plane that contains all the eigenvalues of a
Workshop (ITW 2010), January 6-8 2010. complex square matrix. For an n × n matrix A, define
http://infoscience.epfl.ch/record/140635. X
Ri = |ai,j |.
[RD07] M. Rudafshani and S. Datta. Localization in
j6=i
wireless sensor networks. In IPSN ’07:
Proceedings of the 6th international conference Then each eigenvalue of A is in at least one of the disks
on Information processing in sensor networks,
{z : |z − ai,i | ≤ Ri }.
pages 51–60, New York, NY, USA, 2007. ACM.
[Sin08] A. Singer. A remark on global positioning from
local distances. Proceedings of the National
Academy of Sciences, 105(28):9507–9511, 2008.
[SLR02] C. Savarese, K. Langendoen, and J. Rabaey.
Robust positioning algorithms for distributed
ad-hoc wireless sensor networks. In USENIX
Technical Annual Conference, pages 317–328,
Monterey, CA, June 2002.