Source Localization:

Applications and Algorithms

Hing Cheung So

http://www.ee.cityu.edu.hk/~hcso

Department of Electronic Engineering

City University of Hong Kong

H. C. So Page 1
Outline

 Introduction
 Applications
 Positioning Principles and Measurement Models
 Positioning Algorithms
 Performance Analysis
 List of References

H. C. So Page 2
Introduction
What is Source Localization?
Determine the position of a source
 Position: , , latitude and longitude (e.g., HK
latitude and longitude are around 22°18’ N and 114°10’
E,), a point on a map, street number, building, etc.
 Source: target of interest, e.g., mobile phone, tablet PC,
person, car, ship, sensor node
Similar terminologies include wireless location, radiolocation
and position location
When the source is moving and we want to find the its
trajectory versus time, it may be referred to as target
tracking or position tracking

H. C. So Page 3
Elements in Positioning
Consider finding the absolute position of a source in terms
of , we need:
 Signals emitted from the source which contain position
information
 Sensors or receivers with known coordinates which collect
the signals
 An algorithm to compute the location using the received
signals. There are two categories:
 Directly uses the received measurements to obtain the
location, and it is referred to as direct approach [1]
 First extracts location-bearing information such as
time-of-arrival and energy estimates from the signals,
then base on them to determine , which may be
referred to as two-stage approach [2]
H. C. So Page 4
Applications

Emergency Assistance

A person with a mobile phone is in an emergency situation

but unable to describe his location, e.g.,

 During a hiking activity, a person gets lost from his team

and does not know where he is; He calls 911 and then
the police can determine his location

 A person sees a terrible accident and he calls the police;

However, he is too afraid and cannot tell his position
although he should know it

H. C. So Page 5
  U.S. Federal Communications Commission (FCC) has
mandated the Enhanced 911 (E911) rules where wireless
operators are required to provide position information of
wireless 911 callers:
http://transition.fcc.gov/pshs/services/911-
services/enhanced911/Welcome.html
 Phase I: Require wireless operators to report the
telephone number of a wireless 911 caller and the
location of the antenna or base station that received the
call, upon appropriate request by a local Public Safety
Answering Point (PSAP)
 Phase II: Require wireless carriers to provide more
precise location information, within 50 to 300 meters in
most cases

H. C. So Page 6
Personal Localization and Tracking

The position of a person is known time-by-time, e.g.,

 A parent can know where his child is; For example,
the parent is notified when his child has arrived home
or school; Or when his child gets lost in a crowded
area, his parent can look for him easily
http://www.positionlogic.com/industries-gps-tracking-
solutions/gps-tracking-solution-child-protection/
 A bodyguard can keep track of an important person to
ensure his/her safety, e.g., monitoring the position of
a president in a cocktail party
 The position of a mentally impaired person, e.g.,
elderly with Alzheimer's disease can be known by his
relative or caregiver regularly
H. C. So Page 7
Fleet Management and Asset Tracking
A user can track the location and status of a specific group
or individual, e.g.,
 Hong Kong Fire Services Department employs the Third
Generation Mobilizing System to manage fire engines and
ambulances
 Asset tracking represents the supervision of an asset,
which includes constant reporting of its position and
providing alert to the supervisor and client when the
asset leaves a specific area. Company examples include
FedEx and UPS
 A manager of a logistic company can know the locations
of the company’s vehicles and then use the position
information to increase the transport efficiency
H. C. So Page 8
 Hong Kong PCCW provides a service which marks the
location of client’s field workforce or packages on a web-
based digital map, enabling an easy way to manage
resources in terms of people, cargo and vehicles

http://www.pccw-hkt.com/en/Location-Based-Service/
H. C. So Page 9
Travel Services

Provide useful information for tourists, e.g.,

 A tourist can consult a location system where he/she can

go shopping or dining nearby, or to find out which is the
fastest or cost-effective way to go to his/her next
destination. For example, the position or requested route
will be shown on the digital map on a mobile phone
 Hong Kong Tourism Board has launched some phone and
tablet Apps for visitors with navigation services:
http://www.discoverhongkong.com/eng/plan-your-
trip/travel-kit/mobile-apps.jsp

H. C. So Page 10
Location-based Advertising and Marketing
The idea is to broadcast advertisement and marketing
information to users when they enter a geographical area:
 A cinema offers a discount of watching a movie to mobile
phone users nearby when the movie will be shown soon
 When a user walks by a store, a special offer
advertisement will ring on his mobile device
 Location-based broadcaster sends a mass text to
everyone with cell phones in the room at once
http://www.dailymail.co.uk/sciencetech/article-
2653449/The-shocking-car-safety-ad-hijacks-
cinemagoers-mobile-phones-exactly-distracting-text-
message-be.html

H. C. So Page 11
Location-based Billing
Location-based billing allows a wireless operator to charge
different rates to mobile subscribers based on where they
are

Automated Camera Steering

A representative application is video conferencing where we
need to automatically capture the face of an active talker
via determining his/her position. A microphone array is
employed as the receivers for collecting the speech signals

H. C. So Page 12
Distant Speech Acquisition
A distant speech corresponds to low signal-to-noise ratio
(SNR) conditions. If we are able to locate its source,
beamforming can then be applied to obtain the speech at
much higher SNR

Wireless Sensor Network

A wireless sensor network consists of many small,
inexpensive, low-power nodes which collect surrounding
data, perform small-scale computations and communicate
among their neighbors. It has great potential in numerous
remote monitoring and control applications such as:
 Smart home, e.g., when you enter a room, the light will
be automatically turned on
 Environmental monitoring, e.g., detection of fire source
H. C. So Page 13
Positioning Principles and Measurement Models

We consider the two-stage approach and assume that the

location-bearing information has been extracted from the
raw measurements in the first stage. They include:

 Time-of-Arrival (TOA)
 Time-Difference-of-Arrival (TDOA)
 Received Signal Strength (RSS)
 Direction-of-Arrival (DOA)

Given the TOA, TDOA, RSS or DOA estimates, our task is to

find source position

H. C. So Page 14
TOA
One-way signal propagation time between source and
receiver
In principle, the signals can be represented as

transmitted signal: s (t )
and
received signal: s (t − TOA)
From TOA, the distance between them can be determined:
Distance =
where is the speed of light

The target must lie on a circle centered at the receiver with

radius
H. C. So Page 15
Clock synchronization among source and all receivers is
needed
Synchronization in the source is not required if two-way
(round-trip) propagation is used
H. C. So Page 16
Let ( xi , yi ) be position of the ith receiver
Let ti be the TOA between target and ith sensor, we have:
di
ti = ⇒ d i = c ⋅ ti
c
where d i is the distance between them:
d i = ( x − xi )2 + ( y − yi )2

Practically, the TOA will be subject to error:

ri = d i + ni = ( x − xi )2 + ( y − yi )2 + ni
Suppose there are M receivers, we have M equations to
solve for the unknown ( x, y ) with the known ( xi , yi ) .

H. C. So Page 17
TDOA

In principle, signals collected at the mth and nth receivers

are
mth receiver: s (t − TOA m )
and
nth receiver: s (t − TOA n )

TDOA is the difference in TOAs between two receivers:

TDOAm,n = TOAm - TOAn

Time synchronization is needed among all sensors only

Each TDOA corresponds to one hyperbola and source

position is given by intersection of at least two hyperbolae
H. C. So Page 18
Let 1st sensor at ( x1, y1 ) be the reference and let ti,1 be the
TDOA of the source between the ith and 1st sensors, we
have:
d i,1
ti,1 = ⇒ d i,1 = c ⋅ ti,1, i ≠1
c
where d i,1 is the difference between d i and d1:

d i,1 = d i − d1 = ( x − xi )2 + ( y − yi )2 − ( x − x1 )2 + ( y − y1 )2

Practically, it is not possible to obtain noise-free TDOA:

ri,1 = d i,1 + ni,1 = ( x − xi )2 + ( y − yi )2 − ( x − x1 )2 + ( y − y1 )2 + ni,1

M sensors correspond to (M-1) TDOAs, so we have (M-1)

equations to solve for unknown ( x, y ) with known ( xi , yi )

H. C. So Page 19
RSS
Propagation path loss of the signal traveling from the source
to the receiver and their distance can be computed from it.
Let the power transmitted by the source be
The received signal power can be expressed as:

where is the propagation constant

Positioning concept is same as TOA but the distance derived
from RSS is not very accurate
Clock synchronization is not required
RSS is available at current wireless network, e.g., WiFi
(IEEE 802.11) and ZigBee (IEEE 802.15.4)
H. C. So Page 20
The received signal power at the ith sensor can be modeled
as:

where is the transmitted power, accounts for all factors

which affect the received power

However, the noise in RSS is log-normal distributed and the

observed RSS in dB is:

or

Suppose there are M sensors to measure the RSSs, we have

M equations to solve for ( x, y ) with the known ( xi , yi )
H. C. So Page 21
DOA
Arrival angle of signal from the source at the receiver and
each corresponds to a line-of-bearing (LOB)
Intersection of at least two LOBs gives source position
Antenna array is required at the receiver

H. C. So Page 22
Let the DOA of signal from the source at the ith sensor be θi ,
we get:
y − yi  y − yi 
tan( θi ) = ⇒ θi = tan −1  
x − xi  x − xi 
In the presence of disturbance ni , the noisy DOA
measurement is
−1 
y − yi 
ri = θi + ni = tan   + ni
 x − xi 
Suppose there are M receivers to measure the DOAs, we
have M equations to solve for ( x, y ) with the known ( xi , yi )

H. C. So Page 23
Positioning Algorithms
For simplicity, all disturbances { ni } are assumed to be zero-
mean and uncorrelated with each other and have the same
powers, i.e., independent and identically distributed (i.i.d.)
Optimum Solution [2]
The basic idea is to estimate the source position via
minimizing a nonlinear least squares (NLS) cost function
For TOA equations:

ri = d i + ni = ( x − xi )2 + ( y − yi )2 + ni , i = 1,2,, M

The NLS cost function is

M
(
J ( x, y ) = ∑ ri − ( x − xi ) + ( y − yi )
i =1
2 2
)
2

H. C. So Page 24
The position estimate is obtained as:
( xˆ, yˆ ) = arg min { J ( x, y )}
x, y

which means that when x = ˆ

x and y = yˆ , J ( x, y ) has the
minimum value.
However, finding the minimum is difficult because of
nonlinear J ( x, y) and thus global solution is not guaranteed
Newton’s method can be used with proper initialization:
−1
∂ J 2
∂ J 2
 ∂J 
xk   ∂x 2   ∂x 
ˆxk +1   ˆ ∂x∂y 
 y  =  y  −  ∂2J ∂2J 
 ∂J 
 ˆk +1   ˆk   
 ∂x∂y 2   ∂y 
 ∂y 
xk , y = ˆ
x =ˆ yk

H. C. So Page 25
For TDOA equations:

ri,1 = ( x − xi )2 + ( y − yi )2 − ( x − x1)2 + ( y − y1)2 + ni,1, i = 2,  , M

The NLS cost function is

M

i=2
(
J ( x, y) = ∑ ri − ( x − xi ) + ( y − yi ) + ( x − x1) + ( y − yi )
2 2 2 2
)
2

For RSS equations:

The NLS cost function is

H. C. So Page 26
For DOA equations:
y − yi 
−1 
ri = tan   + ni , i = 1,2  , M
 x − xi 

The NLS cost function is

2
M  −1  y − yi  
J ( x, y) = ∑ ri − tan  
i =1   x − xi 

When all disturbances { ni } are Gaussian distributed, NLS is

equivalent to maximum likelihood (ML) approach
NLS is optimum but it is computationally demanding if
stochastic approach (e.g., genetic algorithm, particle swarm
optimization) or grid search is used and initial estimates are
required for gradient-based techniques
H. C. So Page 27
Cramér-Rao Lower Bound (CRLB) [3]

CRLB is performance bound in terms of minimum achievable

variance provided by any unbiased estimators

Its derivation requires knowledge of the noise probability

density function (PDF) in closed-form

Let be the observations where is a known

function, is the vector of parameters of
interest, and is the noise vector. Denote its PDF by

The CRLB for can be obtained in two steps:

 Compute the Fisher information matrix
 CRLB for is the entry of ,

H. C. So Page 28
 has the form of:

In our case of two-dimensional positioning, is a 2x2

matrix, and we use as CRLB for location

H. C. So Page 29
Suboptimal but Simple Solutions
1. Linear Least Squares (LLS) [4]-[8]
Key idea is to convert nonlinear equations into linear via
introducing extra variable and then apply least squares (LS)
For TOA equations:

ri = d i + ni = ( x − xi )2 + ( y − yi )2 + ni , i = 1,2,, M

Squaring both sides, we have:

ri = ( x − xi )2 + ( y − yi )2 + ni + 2ni ( x − xi )2 + ( y − yi )2
2 2

Now the noise component is

mi = ni + 2ni ( x − xi )2 + ( y − yi )2
2

H. C. So Page 30
Expanding the equation and letting R = x 2 + y 2 yields

ri 2 = ( x − xi )2 + ( y − yi )2 + mi
⇒ ri 2 = x 2 − 2 xxi + xi 2 + y 2 − 2 yyi + yi 2 + mi
⇒ −2 xxi − 2 yyi + R = ri 2 − xi 2 − yi 2 − mi , R = x 2 + y 2
⇒ −2 xxi − 2 yyi + R ≈ ri 2 − xi 2 − yi 2

In matrix form:
Ax ≈ b
or
 − 2 x1 − 2 y1 1  x   r12 − x12 − y12 
      
  y ≈  
  
2 
− 2 xM − 2 yM 1  R  rM2 − xM
2
− yM
 

H. C. So Page 31
The LLS cost function is
J ( x) = ( Ax - b )T ( Ax - b )
= xT AT Ax - xT AT b - bT Ax - bT b
= xT AT Ax - 2xT AT b - bT b

The position estimate is obtained as:

x = arg min{J (x)}
ˆ
x

which can be easily obtained as

dJ (x)
= 2 AT Aˆ
x - 2AT b = 0
dx
⇒ AT Aˆ
x = AT b

⇒ˆ
x= A A ( T
) -1
AT b
H. C. So Page 32
The solution is simple because
 A and b are easy to construct
 Only simple matrix operations are involved

However, it is not optimal because

 The disturbances { mi } are different powers but LLS is

optimal only for zero mean i.i.d. noises
 At sufficiently large noise conditions, the means of { mi }
are not close to 0 and this also affects LLS accuracy
 The introduced variable R = x 2 + y 2 is a function of x and y
but this known information is not utilized in LLS

H. C. So Page 33
For TDOA equations:

ri,1 = ( x − xi )2 + ( y − yi )2 − ( x − x1)2 + ( y − y1)2 + ni,1, i = 2,  , M

⇒ ri,1 + ( x − x1)2 + ( y − y1)2 = ( x − xi )2 + ( y − yi )2 + ni,1

Let
mi = ni,1 + 2ni,1 ( x − xi )2 + ( y − yi )2
2

and
R1 = ( x − x1 )2 + ( y − y1 )2

Squaring both sides and rearranging the result:

( xi − x1 )( x − x1 ) + ( yi − y1 )( y − y1 ) + ri,1R1
[
= 0.5 ( xi − x1 )2 + ( yi − y1 )2 − ri,1 + 0.5mi
2
]
≈ 0.5[( x − x ) + ( yi − y1)2 − ri,1 ], i = 2,3,  , M
2 2
i 1

H. C. So Page 34
In matrix form, we have
Ax ≈ b
or
 x2 − x1 y2 − y1 r2,1   x − x1 
 
      y − y1  ≈
 
 xM − x1 y M − y1 rM ,1   R1 
 ( x2 − x1)2 + ( y2 − y1)2 − r2,12 
1 
  
2 2
( x − x1) + ( y M − y1) − rM ,1
2 2
 M 
The solution is then:

(
x= A A
ˆ
T
)
−1
AT b

H. C. So Page 35
For RSS equations:

Similar to TOA, we can get an estimate of :

such that .

In matrix form:

Ax ≈ b or

H. C. So Page 36
The solution is:

(
x= A A
ˆ
T
)
−1
AT b
For DOA equations:
y − yi 
−1 
ri = tan   + ni , i = 1,2  , M
 x − xi 
Recall
y − yi
tan(θi ) =
x − xi
Taking tangent operation on both sides:

−1 y − yi  y − yi y − yi
ri − ni = tan 
 
 ⇒ tan(ri − ni ) = ⇒ tan(ri ) ≈
 x − xi  x − xi x − xi

H. C. So Page 37
Conversion to linear form can be achieved via:
y − yi sin(ri ) y − yi
tan(ri ) ≈ ⇒ ≈
x − xi cos(ri ) x − xi
⇒ sin(ri )x − cos(ri ) y ≈ sin(ri )xi − cos(ri ) yi , i = 1,2,  , M
In matrix form:
Ax ≈ b
or
 sin(r1) − cos(r1)   sin(r1)x1 − cos(r1) y1 
   x  
  ≈ 
   y   
sin(rM ) − cos(rM ) sin(rM )xM − cos(rM ) y M 

As a result, the solution is also of the form:

(
x= A A
ˆ
T
)
−1
AT b
H. C. So Page 38
 Typical mean square error performance of LLS (=LSC)
H. C. So Page 39
2. Subspace Solution [9]-[12]

Recall TOA model:

ri = d i + ni = ( x − xi )2 + ( y − yi )2 + ni , i = 1,2,, M

Define a rank-2 matrix:

where

and

H. C. So Page 40
Represent D using eigenvalue decomposition:

where
Λ = diag(λ1 , λ 2 ,0,,0) , Λ s = diag(λ1 , λ 2 )
U = [ u1 , u 2 ,, u M ], U s = [ u 1 , u 2 ]
X is obtained up to a rotation:

X is then determined as:

,
where

H. C. So Page 41
In practice, we only have:

LS estimate of is:

Hence a signal subspace algorithm is resulted from:

or

and

is a set of linear equations which can be solved by LLS

H. C. So Page 42
Alternatively, using:

we have a noise subspace algorithm:

where

LS estimate is:

H. C. So Page 43
 Typical mean square error performance of subspace method
H. C. So Page 44
Solutions with Improved Performance
Design ideas are based on circumventing the drawbacks of
the LLS and subspace algorithms
Recall the drawbacks in LLS TOA-based method:
 The disturbances { mi } are not of same powers but when
we use LLS, equal weighting is assumed in each equation
Solution: Proper weighting for each equation, i.e., LLS is
modified to linear weighted least squares (LWLS)

At sufficiently small noise condition, we can rewrite the

linear equations by including the noise component as:

H. C. So Page 45
 , ,

we have which corresponds to linear unbiased

model and the best unbiased linear estimator (BLUE) is [5]

where

H. C. So Page 46
 Typical mean square error performance of BLUE
H. C. So Page 47
 At sufficiently large noise conditions, the means of { mi }
are not close to 0 and this affects LLS accuracy
Solution: Approximate NLS minimization using convex
optimization [13]-[16]

Two basic steps are involved:

 Transform the NLS or ML estimator to an equivalent
constrained optimization problem
 Relax the constrained optimization problem to a
convex optimization problem such that all constraints
are convex and linear

The performance of convex optimization approaches ML if

the constraints are tight

H. C. So Page 48
Typical mean square error performance of semidefinite relaxation
H. C. So Page 49
 The introduced variable R = x 2 + y 2 is a function of x and y
but this information has not been utilized in LLS
Solution: This information is utilized as a constraint in the
minimization, i.e., the position estimate is given by
x = arg min{J (x)}
ˆ subject to R = x2 + y2
x

Combining with WLS, the problem

Minimize J ( x ) = ( Ax − b)T C p-1 ( Ax − b) subject to R = x2 + y2

which is a constrained WLS (CWLS) problem and can be

solved by method of Lagrange multipliers [4]

H. C. So Page 50
 Typical mean square error performance of CWLS method
H. C. So Page 51
One drawback in subspace method:

 Noises in are not i.i.d. and LS is suboptimal

Solution: Incorporating an optimal weighting matrix in

[12]

Two basic steps are involved:

 Reorganize into the form of Ax ≈ b
 Incorporate the optimal weighting matrix in Ax ≈ b

H. C. So Page 52
Performance Analysis [17]
We start with analyzing the bias and mean square error
(MSE) in estimation of a scalar

Suppose its estimate is obtained by minimizing a

differentiable cost function constructed from , :

This implies

H. C. So Page 53
At small estimation error conditions, is close to . Applying
Tayler series expansion yields:

If is sufficiently smooth around , then

Hence

Similarly,

H. C. So Page 54
For estimation of a vector from minimizing , the
formulas are generalized as follows:

and

where is the gradient vector and is the

Hessian matrix:

H. C. So Page 55
As a result,

Similarly, the covariance matrix is:

The MSE of is given by entry of

H. C. So Page 56
Consider positioning of a source at by
sensors at known coordinates ,

Take TOA model as an illustration:

where and is white

Taking ML or NLS approach as an illustration, the cost

function is

H. C. So Page 57
To determine the bias and MSE, the steps include:

because

Similarly,

H. C. So Page 58
As a result,

With tedious calculation, we have

which is the inverse of the Fisher information matrix

That is, the estimator is optimum

H. C. So Page 59
List of References
[1] A. J. Weiss, “Direct position determination of narrowband radio
frequency transmitters,” IEEE Signal Processing Letters, vol.11, no.5,
pp.513-516, May 2004
[2] H. C. So, “Source localization: Algorithms and analysis,” Handbook of
Position Location: Theory, Practice and Advances, Chapter 2, S. A.
Zekavat and M. Buehrer, Eds., Wiley-IEEE Press, 2011
[3] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory, Prentice-Hall, NJ: Englewood Cliffs, 1993
[4] K. W. Cheung, H. C. So, W.-K. Ma and Y. T. Chan, “Least squares
algorithms for time-of-arrival based mobile location,” IEEE Transactions
on Signal Processing, vol.52, no.4, pp.1121-1128, April 2004
[5] F. K. W. Chan, H. C. So, J. Zheng and K. W. K. Lui, “On best linear
unbiased estimator approach for time-of-arrival based localization,” IET
- Signal Processing, vol.2, no.2, pp.156-162, June 2008
[6] K. W. Cheung, H. C. So, Y. T. Chan and W.-K. Ma, “A constrained least
squares approach to mobile positioning: algorithms and optimality,”
EURASIP Journal on Applied Signal Processing, vol.2006, Article ID
20858, pp.1-23, 2006
[7] H. C. So and L. Lin, “Linear least squares approach for accurate

H. C. So Page 60
 received signal strength based source localization,” IEEE Transactions
on Signal Processing, vol.59, no.8, pp.4035-4040, August 2011
[8] L. Lin, H. C. So and Y. T. Chan, “Accurate and simple source
localization using differential received signal strength,” Digital Signal
Processing, vol.23, no.3, pp.736–743, May 2013
[9] H. C. So and K. W. Chan, “A generalized subspace approach for mobile
positioning with time-of-arrival measurements,” IEEE Transactions on
Signal Processing, vol.55, no.10, pp.5103-5107, October 2007
[10] H.-W. Wei, Q. Wan, Z.-X. Chen and S.-F. Ye, “A novel weighted
multidimensional scaling analysis for time-of-arrival-based mobile
location,” IEEE Transactions on Signal Processing, vol.56, no.7,
pp.3018–3022, July 2008
[11] F. K. W. Chan, H. C. So and W.-K. Ma, “A novel subspace approach for
node localization in wireless sensor networks using range
measurements,” IEEE Transactions on Signal Processing, vol.57, no.1,
pp.260-269, January 2009
[12] F. K. W. Chan and H. C. So, “Efficient weighted multidimensional
scaling for wireless sensor network localization,” IEEE Transactions on
Signal Processing, vol.57, no.11, pp.4548-4553, Nov. 2009
[13] K. W. Cheung, W.-K. Ma and H. C. So, “Accurate approximation
algorithm for TOA-based maximum likelihood mobile location using
H. C. So Page 61
 semidefinite programming,” Proceedings of the International
Conference on Acoustics, Speech, and Signal Processing, vol.2, pp.145-
148, May 2004, Montreal, Canada
[14] K. W. K. Lui, F. K. W. Chan and H. C. So, “Semi-definite programming
approach for range-difference based source localization,” IEEE
Transactions on Signal Processing, vol.57, no.4, pp.1630-1633, April
2009
[15] K. W. K. Lui, F. K. W. Chan and H. C. So, “Accurate time delay
estimation based passive localization,” Signal Processing, vol.89,
pp.1835-1838, September 2009
[16] K. W. K. Lui, W.-K. Ma, H. C. So and F. K. W. Chan, “Semi-definite
programming algorithms for sensor network node localization with
uncertainties in anchor positions and/or propagation speed,” IEEE
Transactions on Signal Processing, vol.57, no.2, pp.752-763, February
2009
[17] H. C. So, Y. T. Chan, K. C. Ho and Y. Chen, “Simple formulas for bias
and mean square error computation,” IEEE Signal Processing Magazine,
vol.30, no.4, pp.162-165, July 2013

H. C. So Page 62