The Kalman Like Particle Filter : Optimal Estimation With Quantized Innovations/Measurements

0.21 MB·English

by  Ravi Teja Sukhavasi

The Kalman Like Particle Filter : Optimal Estimation With Quantized
Innovations/Measurements
Ravi Teja Sukhavasi and Babak Hassibi
Abstract—We study the problem of optimal estimation quantization levels. In both cases, it is assumed that the
using quantized innovations, with application to distributed conditional state density is approximately Gaussian leading
estimation over sensor networks. We show that the state
to a linear filter and a very simple characterization of its
probability density conditioned on the quantized innovations
errorperformance.UndertheGaussianassumption,theerror
can be expressed as the sum of a Gaussian random vector
and a certain truncated Gaussian vector. This structure bears covariance matrix associated with the state estimation error
close resemblance to the full information Kalman filter and so satisfiesamodifiedRiccatirecursionofthetypethatappears
allows us to effectively combine the Kalman structure with a in [11]. The only difference between this modified Riccati
particlefiltertorecursivelycomputethestateestimate.Wecall
and the traditional one is a scaling factor λ multiplying
the resuting filter the Kalman like particle filter (KLPF) and
9 the nonlinear term of the recursion. For the SOI Kalman
observe that it delivers close to optimal performance using far
0 fewer particles than that of a particle filter directly applied to filter (SOI-KF), λ is 2 while [10] presents a formula for
0 π
the original problem. We also note that the conditional state λ in the case of multiple quantization levels. Henceforth,
2
density follows a, so called, generalized closed skew-normal these filters will be referred to as SOI-KF and MLQ-KF,

p (GCSN) distribution.
and their associated Riccati recursions as SOI-Riccati and
e Index Terms—Distributed state estimation, Sign of Innova-
MLQ-Riccati respectively.
S tion,ClosedSkewNormalDistribution,ParticleFilter,Wireless
sensor network, Kalman Filter. For linear time invariant dynamical systems, if the Gaus-
5
sian assumption were realistic, convergence of the modified

] I. INTRODUCTION Riccati must mean the convergence of the corresponding
T linear filters. Using results presented in [11] one can find
Recent advances in very large-scale integration and mi-
I linear time invariant systems for which the MLQ-Riccati
. croelectromechanical system technology have led to the
s andSOI-Riccaticonverge.[12]providesexamplesforwhich
c availabilityofcheap,lowqualityandlowpowerconsumption
the actual error performance of SOI-KF and MLQ-KF do
[ sensors in the market. This has generated a great deal of
not converge to their respective Riccatis, which warrants a
interest in wireless sensor networks (WSNs) due to their
1
closer examination of the assumption of Gaussianity. This
v potential applications in several diverse fields [1]. Sensor
is one of the questions addressed in this paper. We derive
6 network constraints such as limited bandwidth and power
9 have inspired a considerable amount of research in de- a novel stochastic characterization of the conditional state
9 density (see theorem 3.1). Using this, it is straighforwardto
veloping energy efficient algorithms for network coverage
0 see that the conditionalstate density is notGaussian, as one
and decentralized detection and estimation using quantized
.
9 sensor observations [2]–[4]. wouldexpect,giventhenon-linearnatureofquantization.In
0 fact, it belongs to a class of distributions which we refer to
The problem of estimation with quantized measurements
9 as Generalized Closed Skew Normal (GCSN) distributions.
is almost as old as the Kalman filter itself. An early survey
0
A careful literature review reveals that a related observation
: on the subject can be found in [5]. However, most of the
v has been made in [13]. In particular, with some effort, [13]
earlier techniques centered around using numerical integra-
Xi tion methods to approximatethe optimalstate estimate. The can be used to derive theorem 3.1. Specializing this result
to state space models, we develop a novel particle filtering
r adventofparticlefiltering[6]–[8]createdawholenewsetof
a approach to optimally estimate the state using quantized
toolstohandlenon-linearestimationproblems.Forexample,
measurements/innovations. In the rest of the paper, we use
[4] proposes a particle filtering solution for optimal filtering
the words ‘measurements’ and ‘innovations’ interchangealy
usingquantizedsensormeasurements.But,quantizingsensor
since the analysiswillprovethatthe generalstructureof the
measurementscanleadtolargequantizationnoiseswhenthe
filter does not depend on whether sensor measurements or
observedvaluesarelargewhichthenleadstopoorestimation
innovations are quantized. The proposed filter requires far
accuracy.In[9],thislimitationisovercomebydevelopingan
fewer particles than that of a particle filter applied directly
elegant distributed estimation approach based on quantizing
to the original problem [12], as will be shown through
the innovation to one bit (the so called sign of innovation
various simulations. We also develop a precise formulation
or SOI). In [10], this is generalized to handle multiple
of the conditional state density and observe that it follows
RaviTejaSukhavasiiswiththedepartmentofElectricalEngineering,Cal- what we call a generalized closed skew-normal distribution,
iforniaInstitute ofTechnology, Pasadena, [email protected] which is very similar to those studied in [14]–[19]. Some
Babak Hassibi is a faculty with the department of Electri-
useful propertiesof this distributionare also providedin the
cal Engineering, California Institute of Technology, Pasadena, USA
[email protected] Appendix. The next section introduces the problem setup.
x(n+1)=A(n)x(n)+w(n) where x(n) ∈ Rd is the state, y(n) ∈ R is the observation,
and w(n) ∈ Rd and v(n) ∈ R are uncorrelated Gaussian
. . . . . white noises with zero means and covariances W(n) and
R(n) , σ2(n), respectively. The initial state, x(0), of the
v(0) x(0) x(1) v(1) v(n) x(n) v
+ + + system, is also a zero mean Gaussian with covariance P(0)
andisuncorrelatedwithbothw(n)andv(n).Asin[9],[10],
S(0) S(1) S(n)
we consider the sensor network configuration in which the
q(0) q(1) q(n) fusion center has sufficient power to broadcast its predicted
Ey(n)/qn−1 outputand the correspondingerrorcovarianceto its sensors.
FUSION CENTER Sensors are assumed to have limited power and hence their
transmission of information should be limited. Here, we
Fig.1. WSNwithafusioncenter:Thesensorsactasdatagathering
devices. assume that the energy required for receiving messages is
much less than that for transmitting.
Fig 1 outlines the overall filtering paradigm. Once a
II. PROBLEMSTATEMENTANDPRELIMINARIES scheduling algorithm is in place, at each time instant, a
The broader problem that one would like to solve can sensor S(n) makes a measurement y(n) and computes the
be cast as causal estimation of a random process {x(n)} innovation e˜(n) = y(n)−yˆ(n|n−1), where yˆ(n|n−1)=
usingaquantizedversionofa measurementprocess{y(n)}, hxˆ(n|n − 1) together with the variance of the innovation
where x(n) and y(n) can be vectors. Without any further ke˜(n)k2 are received by the sensor from the fusion center,
structure,thisisadifficultproblemtoanalyze.When{x(n)} withxˆ(n|n−1)beingtheonesteppredictorofthestate.[9],
and {y(n)} are jointly Gaussian, we will provide a novel [10]proposemethodstoquantizeǫ(n)andusethequantized
stochastic characterization of the probaility density of x(n) innovationstoupdatethestateestimate.Thesefilterstakethe
causally conditionedon the quantizedmeasurementprocess. following shape
Weuseittoproposeanovelfilteringtechniquefortheabove
P(n)(H(n))T
problem which reduces to an elegant particle filter when xˆ(n/n)=xˆ(n/n−1)+L(q(n))
H(n)P(n)(H(n))T +σ2
{x(n)} has state space structure. v
xˆ(n+1/n)=A(n)xˆ(n/n)
A. Notation
P(n)(H(n))TH(n)P(n)
Thefollowingnotationwillbeusedintherestofthepaper. P(n/n)=P(n)−λ (2a)
H(n)P(n)(H(n))T +σ2
1) If {u(n)}∞ is a discrete time random process, v
n=−∞ P(n+1),P(n+1/n)=A(n)P(n/n)(A(n))T +W(n)
u(i : j) denotes the collection of random vari-
(2b)
ables (u(i),...,u(j)). U denotes the random vector
n
[u(0),...,u(n)]T and u denotes a realization of U . whereq(n) denotesthe quantizedinnovationwhile L(q(n))
n n
2) ForarandomvectorY,kYk2 ,E(Y−EY)(Y−EY)T and the value of λ dependon the quantizationscheme used.
3) ForrandomvectorsX,Y,hX,Yi,E(X−EX)(Y − Eqs (2a) and (2b) constitute the modified Riccati recursion
EY)T with parameter λ. For SIO-KF, λ = 2 and for MLQ-KF,
π
4) For random variables (X ,...,X ), L(X ,...,X ) [10] provides a formula for λ and L(q(n)). The above
1 n 1 n
denotes their linear span. filter is derivedbasedon the assumptionthatthe conditional
5) For a vector x, x(i) denotes its i-th component. In the distributionf(x(n)/q )isGaussian,whichwewillprove
n−1
context of particle filtering, xi would denote the i-th is generally false. [12] provides examples where the error
particle. performanceof the filters in [9],[10]do nottrack the modi-
6) N (µ,Σ)denotesak-dimnormalrandomvariablewith fiedriccatirecursionsthattheywerepredictedto.Instead,the
k
mean µ and covariance Σ. N (a,b,µ,Σ) denotes a k- optimal filter, which was approximated by a particle filter,
k
dim normal truncated to lie in (a,b), where a, b are has been observed to have an error covariance matrix that
k-dim vectors and the truncation is component-wise. obeys the modified Riccatis. In order to approximate the
7) Φ(x) = P(X ≤ x), where X ∼ N(0,1), optimal filter, [12] employs a very basic particle filtering
Φ(a,b,µ,σ) = P(X ∈ (a,b)) when X ∼ N(µ,σ2) algorithm which is outlined below for easy reference.
and in general, Φ(I ,µ,σ)=P(X ∈I ).
∗ ∗
8) Thenotionofoptimalitytobeusedthroughoutthepaper Alg 1. Particle Filter
is mean squared error optimality.
1) Set n=0. For i=1,··· ,N, initialize the particles,
B. Problem Formulation xi(0|−1)∼f(x(0)) and set xˆ(0|−1)=0
Suppose{x(n)}and{y(n)}havethefollowingstatespace 2) At time n, set q(n)=Q(y(n)−H(n)xˆ(n|n−1)),
structure. where Q(.) is a quantization rule.
3) Suppose the quantized value q(n) implies that
x(n+1)=A(n)x(n)+w(n) (1a)
y(n)−H(n)xˆ(n|n−1)∈I(n), then
y(n)=H(n)x(n)+v(n) (1b)
v(n)+H(n)(x(n)−xˆ(n/n−1))∈I(n). The
importance weights are now calculated as follows Q(.)beageneralquantizationrulewhichgeneratesthequan-
tized measurement process {q(n)}, where q(n) = Q(y ).
wi(n)=Φ(I(n),H(n)(x(n)−xˆ(n/n−1)),σ (n)) n
v Note that we allow the quantization rule to depend on all
4) Measurement update is given by the past measurements y(i), i ≤ n. This includes, as a
special case, the method of quantizing the innovations first
N
xˆ(n/n)= wi(n)xi(n/n−1) proposedin[9].We willshowthattheprobabilitydensityof
Xi=1 x(n)conditionedonthequantizedmeasurementsQn admits
a characterization very similar to Eq (3). We state the result
where wi(n) are the normalized weights, i.e.,
in the following theorem.
wj(n)= wj(n)
PN wi(n)
5) Resample Ni=p1articles with replacement accoding to, Theorem 3.1: Theprobabilitydensityofx(n)conditioned
on the quantized measurement Q can be expressed as
Prob xi(n/n)=xj(n/n−1) =wj(n) n
(cid:0) (cid:1) x(n)/(Q =q )∼N (0,R )
6) For i=1,··· ,N, predict new particles according to, n n k ∆
+R (n)(R (n))−1(Y /(Q =q ))
xy y n n n
xi(n+1/n)∼f x(n+1)|xi(n|n) , i.e., (5)
xi(n+1/n)=A(n)(cid:0)xi(n/n)+N (0,(cid:1)W(n))
k
Proof: Theproofisfairlystraightforward.Thetheorem
where W(n) is the covariance of the process noise at will be proved by showing that the moment generating
time n. fuction of x(n)/(Q =q ) can be seen as the product
n n
7) Set xˆ(n+1|n)=A(n)xˆ(n|n). Also, set n=n+1 of two moment generating functions corresponding to the
and iterate from step 2. two random variables in Eq (5). For brevity, we will write
x(n)/(Q =q ) as x(n)/q .
n n n
The particles in Alg 1 describe the conditional state
f(x(n)/q )= f(x(n),y /q )dy
density f(x(n)/q ) and simulations suggest that one needs n Z n n n
n
upwards of a few hundred particles to get satisfactory error
Noting that f(x(n)/y ,q )=f(x(n)/y ), we can write
performance for most systems. In the following sections, n n n
we disprove the premise behind MLQ-KF and SOI-KF and EetTx(n)/q = etTx(n)f(x(n)/y )f(y /q )dx(n)dy
develop a novel particle filtering technique (KLPF) which n Z n n n n
schoonwvetrhgaetsthtoeKthLePoFpntiemedaslffialrteferwaesrypmaprttioctliecsaltlhya.nStihmeuplaarttiiocnles (=∗)e21tTR∆tZ etTRxy(n)(Ry(n))−1ynf(yn/qn)dyn
filterofAlg1.Thedifferencepartlyliesinusingparticlesto
describeaprobabilitydensityfuntionwithmuchlesssupport | ,mfgofRxy(n){(zRy(n))−1yn/qn }
than the conditional state density.
=⇒ M (t)=M (t)M ((R (n))−1R (n)t)
III. FULL INFORMATION VSQUANTIZED INNOVATIONS x(n)/qn Z yn/qn y yx (6)
Suppose{x(n)}and{y(n)}arejointlyGaussian,thenitis
where Z ∼N (0,R ). In getting (∗), we used the fact that
well known that the probability density of x(n) conditioned k ∆
x(n)/y ∼N (R (n)(R (n))−1y ,R ).Foranyrandom
on Y is a Gaussian with the following parameters n k xy y n ∆
n variable Y, it is easy to see that M (ATt)=M (t). The
Y AY
x(n)/(Y =y )∼N (0,R )+R (n)(R (n))−1y , result is now obvious from Eq (6).
n n k ∆ xy y n
(3)
Comparing Eqs (3) and (5), the only difference is the mea-
R (n),kx(n)k2, R (n),kY k2, R (n),hx(n),Y i
x y n xy n surement vector y being replaced by the random variable
n
and R∆ ,Rx(n)−Rxy(n)(Ry(n))−1Ryx(n) Yq(n),Yn/qn.ItiseasytoseethatYq(n)isamultivariate
gaussian random variable truncated to lie in the region
When {x(n)} has an underlying state space structure and
defined by q . It is worth noting that the covariance of
{y(n)} is a linear measurement of {x(n)} corrupted by n
x(n)/q , kx(n)/q k2, is given by
additive white Gaussian noise, as defined in Eq (1), it is n n
well known that the following Riccati recursion propogates R +R (n)(R (n))−1kY /q k2(R (n))−1R (n)
∆ xy y n n y yx
the error covariance P(n)=R
∆
Asthequantizationschemebecomesfiner(inanappropriate
P(n),P(n/n−1)=A(n)P(n−1/n−1)A(n)T +Q(n) manner), Y (n) clearly converges to Y and x(n)/q ap-
q n n
P(n)H(n)(H(n))TP(n) proaches a Gaussian as is well known. Using theorem 3.1,
P(n/n)=P(n)− (4)
H(n)P(n)(H(n))T +σ2 it is easy to see that x(n)/qn is not Gaussian in general,
v
contrary to the assumption made in [9], [10]. It belongs
P(0)=kx(0)k2
to a class of distributions, which we call the Generalized
We would like to address the problem of optimalestimation Skew Normal Distributions (GCSN), the details of which
using a quantized version of the observation process. Let are provided in the Appendix.
We will use theorem 3.1 to propose a novel particle Φ s (n),s (n),Ez(n)/z , R
1 2 n−1 ∆n
filtering scheme. We begin by noting that (cid:16) p (cid:17)
Also the conditional distribution of Z /Z is given by
n 1:n−1
Ex(n)/q =R (n)(R (n))−1EY (n) (7)
n xy y q
f (z(n)/z )=N z(n);s (n),s (n),Ez(n)/z ,R
n n−1 1 2 n−1 ∆n
The filter is implemented for the quantization scheme pro- (cid:16) (cid:17)
posed in [9], [10]. At time n, the sensor which is sched- Ez(n)/zn−1 =hz(n),Zn−1i(Rz(n−1))−1zn−1
uled to make the measurement receives a prediction of its R ,kz(n)k2−hz(n),Z i(R (n−1))−1hZ ,z(n)i
∆n n−1 z n−1
measurement yˆ(n) , yˆ(n/n−1) and the error covariance
ky(n)−yˆ(n/n−1)k2. The sensor, then quantizes e˜(n) , Note that fn(z(n)/zn−1) is a one dimensional truncated
ke˜(n)k normal.
where e˜(n) , y(n) − yˆ(n/n − 1), using the following
Proof: Theproofisstraightforwardandisnotpresented
quantizer and transmits the result to the fusion center.
here due to space limitations.
q if x>r The following Lemma describes a standard technique to
K K−1
Q(x)=q if r <x≤r , i≥2 generate scalar truncated normal random variables.
 i i−1 i Lemma 4.2: Let U ∼U(0,1), then
q if x≤r
1 1
 Y =Φ−1((Φ(b)−Φ(a))U +Φ(a))
where (r ,...,r ) are the quantization levels. Using the
1 K−1
convention, r0 = −∞ and rK = ∞, we can re-write is distributed as N(a,b,0,1)
the quantization levels as (r0,...,rK). The output of the Proof: If F(y) denotes the cumulative distribution
quantizerattimenwillbedenotedbyq(n)andthelowerand function of Y, then it is well known that F(Y) ∼ U(0,1).
upper limits of the interval implied by q(n) will be denoted We also have
by r (n) and r (n) respectively. We assume that the fusion
1 2
0 if y <a
center has access to the exact value of q(n) at every instant.
Notethatyˆ(n/n−1)is anestimate ofthemeasurement,not F(y)=Φ(y)−Φ(a) if a≤y <b
Φ(b)−Φ(a)
necessarily the optimal. Hence y(n)−yˆ(n/n−1) is not an 1 if y ≥b
innovationinthetruesenseoftheword,unlesstheestimator 
employed at the fusion center is optimal. Nevetheless, we So, if U ∼ U(0,1), then F−1(U) ∼ N(a,b,0,1)
will refer to q(n) as the quantized innovation. With this and hence the technique. Noting that N(a,b,µ,σ2) ≡
setup, we will develop a filtering technique which needs N(a−σµ,b−σµ,0,1), we can generate scalar truncated normal
random variables with arbitrary mean and variance.
far fewer particles to achieve optimal performance than the
simple particle filter applied to the original problem. We will now propose a particle filtering technique to
recursively compute the state estimate.
IV. THE FILTER
We shall begin with the observation that Y (n) is a
q Alg 2. Truncated Normal Particle Filter
multivariate normal distribution truncated as follows
1) At n=0, generate
Yq(n)=Yn/(cid:16)e˜(j)∈(r1(j),r2(j))∀j ≤n(cid:17) {yi(0)}Ni=1 ∼N(s1(0),s2(0),0,Ry(0)) using the
technique outlined in Lemma 4.2.
=Y / y(j)∈(r (j)+yˆ(j),r (j)+yˆ(j))∀j ≤n
n 1 2 2) At time n, for each particle {yi }, compute the
(cid:16) (cid:17) n−1
weight as
Let s (n) , r (j) + yˆ(j) and s (n) , r (j) + yˆ(j).
1 1 2 2
Then, clearly, Yq(n) is a multivariate normal with its j-th wi(n)=Φ s1(n),s2(n),Ey(n)/yni−1, R∆n (8)
component truncated to lie in the interval (s1(j),s2(j)) ∀ (cid:16) p (cid:17)
j ≤ n, i.e., Y (n) ∼ N (s ,s ,0,R (n)). From Eq 3) Generate
q n+1 1,n 2,n y
(7),itisclearthattheoptimalstateestimatecanbecomputed
{yi(n)}N ∼N s (n),s (n),Ey(n)/yi ,R
by firstcomputingthe meanof the truncatednormal.Before i=1 1 2 n−1 ∆n
(cid:0) (cid:1)
proposing the filter, we need a couple of results for and define yi = yni−1 .
(a) relating the distributions of Y (n) and Y (n−1) and n hyi(n)i
q q 4) Measurement update
(b) generating scalar truncated normal random variables.
N wi(n)yi
Lemma 4.1: Let (Z1,Z2,··· ,Zn)∼fn(zn), where xˆ(n|n)=R (n)(R (n))−1 i=1 n (9)
xy y P N wi(n)
f (z )=N (z ;s ,s ,0,R (n)), i=1
n n n n 1,n 2,n z P
5) Resample the N particles {yi}N with replacement
then the marginal of the first n−1 components is given by n i=1
accoding to Prob yi =wi(n) where the normalized
n
(Z ,··· ,Z )∼f (z ) weights are given(cid:0)by(cid:1)
1 n−1 n n−1
≡fn−1(zn−1) wi(n)= wj(n)
∝N (z ;s ,s ,0,R (n−1))× N wi(n)
n−1 n−1 1,n−1 2,n−1 z i=1
z }| {
P
6) Set xˆ(n+1|n)=A(n)xˆ(n|n). Also, set n=n+1 Now consider Eq (9), we can write it alternately as
and iterate from step 2.
N wi(n)yi
xˆ(n|n)=R (n)(R (n))−1 i=1 n
xy y P N wi(n)
i=1
Inthefilterproposedabove,itisimportanttonotethatthe N P N
dimension of the particle yi is n and hence increases with = wi(n)Rxy(n)(Ry(n))−1yni = wi(n)xi(n/n)
n X X
time.Intheabsenceofanystructureontherandomprocesses i=1 i=1
(12)
{x(n)} and {y(n)}, the computations involved in running
the above filter will quickly become infeasible. Alg 2 is of From Eqs (11) and (12), it is easy to see that all the
purely theoretical interest but the technique presented above information in yi is captured in xi(n/n). Hence, we only
n
is greatly simplified when x(n) has state space structure. In needtoworkwith{xi(n/n)}N atanytimen.We cannow
i=1
the following subsection, we will show how to overcome desribe the new filter as follows.
the problem of increasing particle dimension when {x(n)}
Alg 3. Kalman Like Particle Filter (KLPF)
is a Gauss Markov process and {y(n)} is a linear mea-
1) At n=0, generate
surement of {x(n)} corrupted by additive white Gaussian
{yi(0)}N ∼N(s (0),s (0),0,R (0)) using the
noise. The filter takes an elegant formulation in which the i=1 1 2 y
technique outlined in Lemma 4.2. Compute {xi(0)}
particle dimension remains fixed and is equal to the state
using Eq (10a) and xi(1/0)=A(0)xi(0/0)
dimension. This approach requires far fewer particles than
2) Generate
Alg 1 and the filtering technique is quite general, in that,
it can handle arbitrary number of quantization levels and {yi(n)}N ∼N s (n),s (n),H(n)xi(n),R
i=1 1 2 ∆n
arbitrary quantization intervals. We will also observe that (cid:0) (cid:1)
Use Eq (11b) to compute R . Then form {xi(n/n)}
thefilterrequiresfewerandfewerparticlesasthenumberof ∆n
using Eq (10).
quantization levels increases.
3) At time n, for each particle {xi(n)}, compute the
weight {wi(n)} as
A. Exploiting State Space Structure - The Kalman Like wi(n)=Φ s (n),s (n),H(n)xi(n), R
1 2 ∆n
Particle Filter (cid:16) p (cid:17)
4) (Measurement update)
Suppose{x(n)}and{y(n)}havethestatespacestructure Normalize the weights to get wi(n)= wj(n) and
defined in Eq (1). Then we know that Rxy(n)(Ry(n))−1yni compute the measurement updated statePeNis=ti1mwai(tne)using
is the optimal estimate of x(n) given the measurement Eq (12), i.e., xˆ(n/n)= N wi(n)xi(n/n)
vectoryi and hence can be computedby runninga Kalman i=1
n 5) Resample the N particlePs {xi(n/n)}N with
filter using yi, the specific realization of the measurement i=1
n replacement accoding to Prob xi(n/n) =wi(n) and
vector Y . Similarly Ey(n)/yi is the optimal estimate
n n−1 compute xi(n+1)=A(n+1(cid:0))xi(n/n)(cid:1).
of y(n) given Y = yi and R is the resulting
n−1 n−1 ∆n 6) Set xˆ(n+1|n)=A(n)xˆ(n|n). Also, set n=n+1
errorcovariance.Allthesequantitiesemergenaturallyin the
and iterate from step 2.
following Kalman filtering steps.
P(0)H(0)T
xi(0/0)= yi(0) (10a) Remark 1: It is easy to see that the above filter can be
H(0)P(0)(H(0))T +σ2(0)
v extended to the case when the quantizer employed at the
P(k+1),P(k+1/k)=AP(k/k)AT +W(k) (10b) sensor is of the more general form Q(x)=q if x∈I for
j j
P(k)H(k)(H(k))TP(k) 1≤j ≤K.
P(k+1/k+1)=P(k)−
H(k)P(k)(H(k))T +σ2(k)
v
(10c) From the aboveimplementation,in terms of complexity,the
xi(k+1),xi(k+1/k)=Axi(k/k) (10d) KLPF is clearly equivalent to running N parallel Kalman
filters. Hence, the complexity of the KLPF scales linearly
xi(k+1/k+1)=xi(k+1)+ (10e)
in N, the number of particles. Also, it converges to the
P(k)(H(k))T optimalfilter as N →∞. But formost systems, simulations
(yi(k+1)−H(k+1)xi(k+1))
H(k)P(k)(H(k))T +σ2(k) suggestthattheKLPFdeliversclosetooptimalperformance
v
(10f) for N ≤50. The particles in the KLPF describe the random
variable R (n)(R (n))−1Y (n). The support of its distri-
Now, R (n)(R (n))−1yi, R and Ey(n)/yi can be xy y q
xy y n ∆n n−1 bution clearly decreases with increasing quantization levels.
calculated as follows
As a result KLPF needs fewer particles as the quantization
R (n)(R (n))−1yi =xi(n/n) (11a) becomes finer, a property that Alg 1 does not share. This
xy y n
will be demonstrated through examples in Section V.
R =H(n)P(n)(H(n))T +σ2(n) (11b)
∆n v The scenario considered thus far involves one measure-
Ey(n)/yi =H(n)xi(n) (11c)
n−1 ment per time instant. But Alg 3 can be easily extended to
TABLEI
250
SUMMARYOFNUMERICALRESULTS Monte Carlo of SOI − KF
SOI−Riccati
Example 1 Example 2 ˆx(n)k2125000 PKaLrPtiFc,le N FKLilPteFr ,= N 5P0F0 = 2500
SOI-KF SnOoI 2-bit SyOesI 2-bit kx(n)-100
MLQ-KF - no - yes 50
Alg 1 2500 10000 500 750 00 10 20 30 40 tim50en 60 70 80 90 100
KLPF 500 90 25 3
(a) SOI-KF, Alg 1 and KLPF
350
MLQ−Riccati for 2 bits
300 Monte Carlo Simulation of MLQ−KF
hgaivnednletimmeulitnipslteanmt.eHaesurereimt iesntasssfurmomeddtihfafetrtehnetmseenassourrsematenat n)||2250 KPLaPrtiFc,le N FKLilPteFr ,= N 9P0F = 10000
ˆx(200
noise processes are uncorrelated across different sensors. n)-150
The proof is simple and is not detailed here due to space ||x(100
constraints.
50
V. SIMULATIONS 00 10 20 30 40 tim50en 60 70 80 90 100
(b) MLQ-KF, Alg 1 and KLPF
InAlg1,theparticlesdescribethefullprobabilitydensity
of the state conditioned on quantized measurements. While
in Alg 3, part of the information about the conditional state Fig. 2. In (a), SOI-KF clearly diverges while Alg 1 and KLPF
converge to the optimal filter. From (b), MLQ-KF with 4 levels of
density is captured neatly by the Kalman filter. So, the
quantization also diverges while KLPF converges to the optimal
particlesdescribe a truncatedGaussian which has muchless
filter with just 90 particles
support than the full conditional distribution. We give two
examples in this section to demonstrate the effectiveness of
KLPF. The following table summarizes the highlights from the optimal filter whereas MLQ-KF is itself clearly not the
the two examples optimal filter. In Examples 1 and 2, by optimality, we meant
In Table I, a ‘yes’ indicates that the filter works and is that the mean squared errors of Alg 1 and KLPF tracked
close to optimal, a ‘no’ indicates that its estimation error the corresponding modified Riccati recursions very closely.
diverges and a ‘-’ means that the quantization method does In fact, simulations showed that the mean squared errors of
not apply to the filter. ‘SOI’ stands for ‘sign of innovation’ Alg 1 and KLPF saturate at this value irrespective of the
and ‘2-bit’ stands for a quantization rule with quantization number of particles used.
intervals given by (−∞,−1.2437), (−1.2437,−0.3823),
(−0.3823,0.3823), (0.3823,1.2437) and (1.2437,∞). If B. Example 2
the innovation falls in the interval (−0.3823,0.3823), no
A simple tracking system can be characterized by the
measurement update is done, so that 2 bits will suffice to τ4 τ3
represent the output of the above quantizer. The numbersin followingparameters,A=h10τ1i,h=[10],W =hτ43 τ22i,
front of Alg 1 and KLPF denote the number of particles σ2 = 0.81 and P(0) = 0.01I and the sampling2period
v 3
required to approximate the optimal filter. Clearly, KLPF τ =0.1.
requires far fewer particles than Alg 1. Also evident from In Example 2, note that KLPF works with much fewer
Table I is the fact that KLPF needs dramatically fewer particles than in Example 1. One can attribute this to the
particles as the quantization becomes finer. much higher value of the optimal mean squared error in
Example1thaninExample2,ascanbeseenfromtheplots.
A. Example 1
We re-use this example from [12]. Consider a linear VI. CONCLUSIONS
time invariant system of the form (1) with the following
We propose a Kalman like particle filter (KLPF) that,
0.95 1 0
parameters: A = 0 0.9 10 , h = [102], W = 2I3, in the examples studied, required moderately small number
h 0 0 0.95i
R , σ2 = 2.5 and P(0) = 0.01I , where I denotes an of particles and therefore can obtain close to optimal per-
v 3 m
m×m identity matrix. Note that A is a stable matrix. As formance with a computational complexity comaparable to
can be seen from the plots, SOI-KF and MLQ-KF diverge the conventional Kalman filter. An important open issue is
but KLPF delivers optimal performance with much fewer to determine the number of particles necessary to closely
particles than Alg 1. With the addition of just 1 bit, the approximate the optimal filter. As earlier observed in [12],
requirednumberofparticlesdropsfrom500to 90.Butwith the error covariance matrix of the optimal filter seems to
Alg 1, the numberof particles goesup from 2500 to 10000. follow the modified Riccati recursion introduced in [11].
Remark 2: Extensive simulations suggest that the mod- Determining whether this is the case, and why, remains an
ified Riccati of Eq (2) describes the error performance of interesting open question.
0.8 compression,” Information Theory, IEEE Transactions on, vol. 16,
no.2,pp.152–161,1970.
0.7
[14] A Azzalini and A Dalla Valle, “The multivariate skew-normal
n)||20.6 distribution,” Biometrika, vol.83,no.4,pp.715–726, 1996.
ˆx(0.5 [15] Reinaldo B. Arellano-Valle and Marc G. Genton, “On fundamental
-
n)0.4 KLPF, NKLPF = 25 skew distributions,” J. Multivar. Anal., vol. 96, no. 1, pp. 93–116,
||x(0.3 2S OleIv −e lK−RFiccati 2005.
Particle Filter, NPF = 500 [16] LiHeMarcG.GentonandXiangweiLiu, “Momentsofskew-normal
0.2 randomvectors andtheirquadratic forms,” Statistics andProbability
0.10 10 20 30 40 tim50en 60 70 80 90 100 Letters, vol.51,no.4,pp.319–325,2001.
[17] Marc G. Genton, “Discussion of the skew normal,” Scandinavian
(a) SOI-KF, Alg 1 and KLPF JournalofStatistics, vol.32,no.2,pp.189–198, 2005.
[18] MarcG.GentonPhilippeNaveauandXilinShen, “Askewedkalman
0.6 filter,” J.Multivar. Anal.,vol.94,no.2,pp.382–400,2005.
0.55 [19] MarcG.Genton, Skew-Elliptical DistributionsandtheirApplications
0.5 -AJourney beyongNormality, ChapmanandHall/CRC, 2004.
n)-ˆx(n)||200..034.554 5P alertvicelle − F Riltieccr,a NtiPF = 750 APPENDIX
||x(0.02.53 K5 LlePvFe,l N−K KLPFF = 3 A. The Generalized Closed Skew-Normal Distribution
0.2
0.15 Definition 1: For x ∈ Rn, we define the generalized closed
0.10 10 20 30 40 tim50en 60 70 80 90 100 skew-normal distribution,GCSNk,n(x;µ,Σ,D,s1,s2,∆),asfol-
lows
(b) MLQ-KF, Alg 1 and KLPF
GCSN (x;µ,Σ,D,s ,s ,∆), N (x;µ,Σ)L (.)
k,n 1 2 k k,n
Φ (s ,s ;D(x−µ),∆)
Fig. 3. Both in (a) and (b), all the filters are close to optimal. L (.)= n 1 2 (13)
KLPFachievesoptimalperformancewithremarkablyfewparticles k,n Φn(s1,s2;0,∆+DΣDT)
and hence has a complexity of the same order as that of SOI-KF N (x;µ,Σ) is a k-dim gaussian r.v and Φ (y ,y ;ν,∆) =
k n 1 2
and MLQ-KF. P(y(1) ≤ X ≤ y(1),...,y(n) ≤ X ≤ y(n)) where
1 1 2 1 n 2
(X ,...,X )∼N (x;ν,∆).
1 n n
Thesizesofthematricesinvolvedfollowaccordingly.Thisisavery
REFERENCES simplegeneralizationoftheclosedskew-normal(CSN)distribution
defined in [19]. Infact, it reduces to the CSN if y = −∞
[1] I.F.Akyildiz,WeilianSu,Y.Sankarasubramaniam,andE.Cayirci, “A 1
in all its components. Naturally, it inherits most of its closure
survey on sensor networks,” Communications Magazine, IEEE, vol.
properties. Skew elliptical distributions generated a lot of interest
40,no.8,pp.102–114,2002.
because they provide a much needed tool to handle skewness
[2] A.Krasnopeev, Jin-Jun Xiao, andZhi-Quan Luo, “Minimum energy
decentralized estimation in sensor network with correlated sensor in statistical modeling and have a good number of properties in
noise,” inICASSP,2005,vol.3,pp.iii/673–iii/676. common with the standard normal distribution, such as closure
[3] Zhi-QuanLuoandJin-JunXiao, “Universal decentralized estimation undermarginlizationandconditioning.Wewillbrieflydiscusssome
inabandwidthconstrainedsensornetwork,” inICASSP,2005,vol.4, of the useful properties.
pp.iv/829–iv/832. Lemma 1.1: If X ∼ N (µ,Σ), Z ∼ N (0,∆) be
k n
[4] G Rickard Karlsson and Fredrik Gustafsson, “Particle filtering for independent, then X/(D(X−µ)−Z ∈(s ,s )) ∼
quantized sensor information,” in 13th European Signal Processing 1 2
GCSN (µ,Σ,D,s ,s ,∆)
Conference, EUSIPCO.2005,EURASIP,Turkey. k,n 1 2
In the following proposition, we provide a stochastic character-
[5] R. Curry, Estimation and Control with Quantized Measurements,
M.I.TPress,1970. ization of the GCSN which can be identified as a special case of
[6] N.J. Gordon, D.J. Salmond, and A.F.M. Smith, “Novel approach to theorem 3.1 for state space systems.
nonlinear/non-gaussian bayesian state estimation,” Radar andSignal Proposition 1.2: Suppose
Processing, IEEProceedings-F, vol.140,no.2,pp.107–113,1993.
[7] D.CrisanandA.Doucet,“Asurveyofconvergenceresultsonparticle V ∼N 0,Σ−ΣDT(∆+DΣDT)−1DΣ and
k
filtering methods for practitioners,” IEEE Transactions on Signal “ ”
Processing,vol.50,no.3,pp.736–746, 2002. U ∼N s ,s ,0,∆+DΣDT
n 1 2
[8] M.S.Arulampalam,S.Maskell,N.Gordon,andT.Clapp, “Atutorial “ ”
onparticlefiltersforonlinenonlinear/non-gaussianbayesiantracking,”
are independent, where N (a,b,0,R) denotes a p-variate normal
IEEETransactionsonSignalProcessing,vol.50,no.2,pp.174–188, p
distributiontruncatedcomponent-wise toliein(a,b),whereaand
2002.
b are p-dim vectors specifying the boundaries of a cube in Rp,
[9] A. Ribeiro, G. B. Giannakis, and S. I. Roumeliotis, “Soi-kf: Dis-
torfibiuntneodvaktailomnsa,n”fiIElteErEinTgrwanitshaclotiwon-csoosntcSoigmnmalunPircoacteiosnsisngu,sivnogl.th5e4,singon. i.e., Z∼Np(a,b,0,R) =⇒ fZ(z) = PN(Zp(z∈;0(a,,Rb)))IZ∈(a,b).
12,pp.4782–4795,2006. Then, Y = V + ΣDT ∆+DΣDT −1U is distributed as
[10] KeyouYou,LihuaXie,ShuliSun,andWendongXiao,“Multiple-level GCSN(0,Σ,D,ν,∆). ` ´
quantized innovation kalman filter,” in17thInternational Federation The proposition will be proved using lemmas 1.3 and 1.4 which
ofAutomaticControl. 2008,Seoul,Korea.
are stated below.
[11] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M.I. Jordan,
Lemma 1.3: If X ∼ GCSN (µ,Σ,D,s ,s ,∆), then the
andS.S.Sastry, “Kalmanfilteringwithintermittent observations,” in k,n 1 2
moment generating function of X is given by
Proceedings. 42ndIEEEConference onDecision andControl, 2003,
2003,vol.1,pp.701–708Vol.1.
Φ (s ,s ;DΣt,∆+DΣDT) 1
[12] Ravi Teja Sukhavasi and Babak Hassibi, “Particle filtering for M (t)= n 1 2 exp(µt+ tTΣt)
quantized innovations,” ICASSP,2009. X Φn(s1,s2;0,∆+DΣDT) 2
[13] W.V Velde R. Curry and J. Potter, “Nonlinear estimation with Now, we will compute the moment generating function of a
quantized measurements - pcm predictive quantization, and data truncated multivariate normal distribution.
Lemma 1.4: LetX ∼N (a,b,0,R),thenthemoment generat- given by the following recursions
p
ing function of X is given by
Σ(n/n)=Σ(n), D(n/n)= D(n/n−1) ,s = s1,n−1 ,
M (t)= Φn(s1,s2;DΣt,R)exp tTRt (14) » H – 1,n »s1(n)–
ProWofe:w[PilrloonfoXwofcPormoppolesΦtietniot(hnse11,p.s2r2o];o0f,oRf)the prop„osit2ion.« s2,n =»ss22,(nn−)1–, ∆(n/n)=»∆(0n) σ0v2– (16)
Let S = ΣDT ∆+DΣDT −1, then Y = V + SU and where xˆ(n) , xˆ(n/n−1) = E(x(n)/qn−1). In the rest of
MY(t) = MV(t)M` SU(t). Now´, MSU(t) = MU(STt). Plugging the appendix, we will use the notation R(n) = ∆(n/n) +
intheformulaforthemomentgeneratingfunctionderivedinlemma D(n/n)Σ(n)D(n/n)T (=∗)∆(n)+D(n)Σ(n)D(n)T ((∗)follows
1.4, we see that MY(t) is the moment generating function of a from Eq (15)).
GCSN (µ,Σ,D,s ,s ,∆) using lemma 1.3.
k,n 1 2
Proposition 1.5: LetX ∼GCSN (µ,Σ,D,s ,s ,∆),W ∼ C. Interpreting the Parameters
k,n 1 2
Nk(µW,Q) be independent and define Y =AX+W, then Y ∼ Theorem 1.8 outlines how the conditional state density evolves
GCSNk,n(µY,ΣY,DY,s1,s2,∆Y), where with time. But, there is little intuition in the recursions proposed
Σ =AΣAT +Q, D =DΣATΣ−1, and there.Inthefollowinglines,wewillexplaintheconnectionofeach
Y Y Y of the parameters to the state space model, so that one can afford
∆Y =∆+DΣDT −DYΣYDYT, µY =µ+µW abetterinsight intotheoverall filteringtechnique firstproposed in
This proof of the proposition follows directly from the following [9].WewillstartbynotingthatΣ(n)isequaltothestatecovariance
two lemmas. The first states that the GCSN is closed under full at time n.
ranklineartransformationswhilethesecondstatesitsclosureunder Lemma 1.9: Σ(n)=kx(n)k2
addition of gaussian noise. The next Lemma states that R(n) = kY k2, i.e., R(n) is the
n
Lemma 1.6: If Y ∼ GCSN (µ,Σ,D,s ,s ,∆) and A covariance of the measurement vector [y(0),...,y(n)]T
k,n 1 2
is a full column rank k × p (k ≥ p) matrix, then Lemma 1.10: R(n)=kYnk2 =Ry(n)
AY ∼ GCSNk,n Aµ,AΣAT,D(ATA)−1AT,s1,s2,∆ . If A The following Lemma gives an expression for P(qn)
is a full row ran`k p × k (p ≥ k) matrix, then A´Y ∼ Lemma 1.11: The probability of quantized innovations is given
GCSN (µ ,Σ ,D ,ν,∆ ), where by
k,n A A A A
µA =Aµ, ΣA =AΣAT, DA =DΣATΣ−A1 P(qn)=Φn+1(s1,n,s2,n;0,Ry(n))
In the following Lemma, we look at the cross covariance of x(n)
∆ =∆+DΣDT −D Σ DT
A A A A and Yn
The following lemma states that the GCSN is closed under Lemma 1.12: hY ,x(n)i=D(n/n)TΣ(n)
n
addition of gaussian noise. Now, applying Proposition 1.2 on theconditional statedensity, we
Lemma 1.7: If X ∼ GCSNk,n(µ,Σ,D,s1,s2,∆) and can write x(n)/Qn=qn stochastically as
W ∼ N (µ ,Σ ) are independent, then X + W ∼
GCSNk,nk(µXW+W,WΣX+W,DX+W,s1,s2,∆X+W), x(n)/qn=Z+Σ(n)D(n/n)T(R(n))−1Yq(n), where
Dwhere µ=X+DWΣ(Σ=+Σµ )+−1µaWnd, ∆ΣX+W ==∆ +ΣDΣ+DΣTW−, Z ∼Nk(0,Σ(n)−Σ(n)D(n/n)(R(n))−1D(n/n)TΣ(n))
DX+WΣ DT . W X+W Yq(n)∼Nn+1(s1,n,s2,n;0,Ry(n)) (17)
X+W X+W X+W
Using Lemmas 1.9, 1.10 and 1.12, one can see that
B. The Conditional State Density Σ(n)−Σ(n)D(n/n)(R(n))−1D(n/n)TΣ(n)=
We will now show that the probability density of the state kx(n)k2−hx(n),Y i kY k2 −1hY ,x(n)i and (18)
n n n
acoGndCitSioNn.edTiomnethueppdaastteqaunadntiMzeedasinunreomvaetniotnuspfda(txe(nre)c/uqrnsi)ofnosllowwilsl Σ(n)D(n/n)T(R(n))−`1 =hx´(n),Yni kYnk2 −1 (19)
be provided for the parameters of the conditional density. All the ` ´
parametershaveverysimpleinterpretationsintermsofthestatistics |,R{xyz(n)}|(Ry({nz))−1}
of the state space model. For ease of exposition, we will provide It is now easy to see that Eqs (18) and (19) are consistent with
all the recursions for a time-invariant state space model. But the theorem 3.1 and hence we have
derivations can be trivially extended to the time varying case.
Theorem 1.8 (Timeand Measurement Updates): The x(n)/qn=Nk(0,Pkal(n/n))+Rxy(n)(R(n))−1Yq(n) (20)
probability density function of x(n)/qm is given by wherePkal(n/n)istheerrorcovarianceoftheKalmanfilterattime
GCSNk,m+1(x(n);0,Σ(n/m),D(n/m),s1,n/m,s2,n/m, step n. The methodology outlined in the appendix can be seen as
∆(n/m)), for m=n−1, n and ∀ n ≥ 0. The recursions relating analternatewayofarrivingatAlg3.Howeveritiscircumlocutious
theparametersofthedistributionsofx(n−1)/Qn−1,x(n)/Qn−1 and not very intuitive. Moreover, the stochastic characterization in
and x(n)/Qn are given by Eq (20) cannot be extended to the more general case of theorem
3.1 using the theory of GCSN. However, it serves as conclusive
Σ(n),Σ(n/n−1)=AΣ(n−1/n−1)AT +W,
evidencethattheconditionalstatedensityisnotGaussian,contrary
D(n),D(n/n−1)=D(n−1/n−1)Σ(n−1/n−1)ATΣ(n)−1 to what is assumed in [9], [10].
s1,n−1 ,s1,n/n−1 =s1,n−1/n−1,
s2,n−1 ,s2,n/n−1 =s2,n−1/n−1,
∆(n),∆(n/n−1)=∆(n−1/n−1)+
D(n−1/n−1)Σ(n−1/n−1)D(n−1/n−1)−D Σ DT (15)
n n n
Eq (15) constitutes the time update. The measurement update is