Professional Documents
Culture Documents
Extensions To The Theory of Widely Linear Complex Kalman Filtering
Extensions To The Theory of Widely Linear Complex Kalman Filtering
Extensions To The Theory of Widely Linear Complex Kalman Filtering
net/publication/51955025
CITATIONS READS
13 393
2 authors:
All content following this page was uploaded by Wenbing Dang on 01 April 2014.
Abstract—For a zero mean, proper, complex random vector estimators for random complex signals are developed in [4],
x, the Hermitian covariance ExxH is a complete second-order [6], [7]. Also the WLMMSE estimators for second order
characterization. However, if the vector x is improper, it is stationary process are given by [8]. WL Wiener filters are
correlated with its complex conjugate, meaning ExxT 6= 0. This
improper or complementary covariance must be accounted for developed in [3] and [4]. And a widely linear Kalman filter
in a complete second-order characterization. is studied in [9]. The problems of detection and equalization
The improper covariance has been exploited for widely linear in communication networks are also addressed in [10]–[16]
arXiv:1105.5432v1 [cs.SY] 27 May 2011
(WL) Wiener filters and WL minimum mean squared error using WL processing.
(MMSE) estimators, and the improvement in performance of It turns out that widely linear processing brings improve-
the WLMMSE estimator over the LMMSE estimator has been
quantified. In this paper we consider the design of the widely ment in performance over strictly linear processing [4], [17],
linear Kalman filter (WLKF). We analyze the WLKF, extended [18]. For instance, in [4] it has been shown that by assuming
WLKF, and unscented WLKF. The key idea of this paper is proper white measurement noises, the maximum performance
to modify the error covariance matrices and the construction advantage of WLMMSE estimation over LMMSE estimation
of effective sigma points in the WLKF in a systematic way is a factor of two.
that exploits the Hermitian and complementary covariance of
improper states and noises. In the past few decades the Kalman filter [19] has been
modified to apply to nonlinear problems, producing Extended
Index Terms—Widely linear transformation, complementary Kalman filters [20], Unscented Kalman filters [21], and parti-
covariance, Kalman filter, unscented Kalman filtering, sigma
points. cle filters [22]–[25]. The motivation of this paper is to make
use of widely linear processing to develop novel Kalman filters
and their nonlinear versions for improper complex states.
I. I NTRODUCTION We show that for improper complex states, complementary
Complex signals are ubiquitous in science and engineering, covariance matrices may be used to create widely linear KFs,
arising as they do as complex representations of two real as well as extended and unscented widely linear KFs. Widely
channels or of two dimensional fields. Consider a zero mean linear Kalman filters (WLKF) and extended WLKFs were
complex random vector x. The usual covariance matrix defined derived in [9], [26]. However, in these papers, the authors
as ExxH describes its Hermitian second order covariance. have studied WLKFs for dynamical systems with very special
But when x and its complex conjugate x∗ are correlated, the state and measurement matrices. In other words, the updates
complementary covariance matrix ExxT does not vanish, so of random states and measurement do not depend on the
it carries useful second order information about the complex conjugates of states and noises.
random vector x. We call a complex random vector proper In this paper we show that the WLKF can address more
as long as its complementary covariance matrix vanishes and general dynamical models that may have widely linear state
improper otherwise. Proper complex vectors have a statisti- and measurement equations. The duality between widely linear
cal description similar to real vectors, but improper random and dual real channel adaptive filtering [27] motivates us
vectors do not. A comprehensive second order analysis of to demonstrate that the WLKF can be used for LMMSE
improper random vectors and processes is considered in [1]– estimation in dual channels when those channels have ar-
[6]. bitrarily coupled state equations. In other words the WLKF
For any improper random vector x, for which x is correlated may be implemented into real equations, or by using complex
with its complex conjugate x∗ , intuition suggests that a good augmented notation, into complex equations. Moreover, by
estimator of x should depend on x∗ . This requires a method- studying a complex scalar random state with proper noises,
ology of widely linear processing instead of strictly linear we show that compared to the conventional KF, the WLKF
processing [1], [2]. The merit of widely linear processing has reduces MSE by a factor of two. This advantage may be more
been exploited in various estimation and filtering problems. significant for more general models.
The widely linear minimum mean squared error (WLMMSE) We propose the extended WLKF for the widely linear
approximation of nonlinear state and measurement equations,
This work is supported by the National Science Foundation under grants which is equivalent to the results in [9], [26]. For the EWLKF,
CCF-1018472 and CCF-0916314, and the Air Force Office of Scientific however, the estimated posterior means and covariances are
Research under contract FA9550-10-1-0241.
precise only to first order in a Taylor expansion. To obtain
W. Dang is with the Department of Electrical and Computer Engineering,
Colorado State University, Fort Collins, CO 80523, USA, Phone: (970) 817- accuracy to second order, we study the unscented WLKF. The
0249 (email: dwb87514@gmail.com). Unscented WLKF of [9] uses sigma points that represent the
L. L. Scharf is with the Department of Electrical and Computer Engi- mean and Hermitian covariance of states. However, the sigma
neering, and the Department of Statistics, Colorado State University, Fort
Collins, CO 80523, USA, Phone: (970) 491-2979, Fax: (970) 491-2249 (e- points should also represent the corresponding complementary
mail: Louis.Scharf@colostate.edu). covariance. In this paper we develop a systematic way to
2
construct modified sigma points with the property that they The matrix H = 12 Tm MTH n is called an augmented matrix
can preserve the complete first and second order statistical with the property that its southeast block is the conjugate of
information of complex random vectors. its northwest block, and its southwest block is the conjugate
The remainder of this paper is organized as follows. In sec- of the northeast block:
tion II we briefly review some rudiments of complex random
H1 H2
vectors, including widely linear transformations, augmented H= , (8)
H∗2 H∗1
covariances, and the definition of improper random signals.
In section III the widely linear minimum mean squared error where
estimator is introduced. In section IV we derive the widely 1
H1 = [M11 + M22 + j (M21 − M12 )] , (9)
linear Kalman filter (WLKF). We give two representations of 2
WLKF: the augmented WLKF and the Cartesian WLKF. In 1
H2 = [M11 − M22 + j (M21 + M12 )] . (10)
addition, we give an example demonstrating the performance 2
enhancement of the WLKF over the KF. In section V, we de- Therefore the augmented matrix H rules the widely linear
rive the extended WLKF. In section VI, we obtain the effective transformation
sigma points for any improper complex random vector. Then
y = H1 x + H2 x∗ ⇔ a + jb = H1 (u + jv) + H2 (u − jv)
we give an algorithm for constructing an unscented WLKF
(UWLKF). In a simulation for real phase demodulation, we = (H1 + H2 )u + j(H1 − H2 )v
show that an UWLKF outperforms a complex, unscented (11)
Kalman filters that assumes proper states and noises in making We see that if and only if H2 = 0, the widely linear transfor-
real estimates from complex baseband measurements. mation is a strictly linear transformation. This corresponds to
the special case M11 = M22 and M12 = −M21 in the linear
II. B RIEF R EVIEW OF C OMPLEX R ANDOM V ECTORS transformation of u, v into a, b.
Let Ω be the sample space of a random experiment that The augmented mean vector of the complex random vector
generates two channels of real signals u, v ∈ Rn defined on Ω. x is
From this we obtain the real composite random vector z ∈ R2n µx µu + jµv
µx = Ex = = , (12)
as zT = [uT , vT ]. We also have the complex random vector µ∗x µu − jµv
x ∈ Cn , obtained by composing u and v into its real and and the augmented covariance matrix of x is
imaginary parts: " #
x = u + jv. (1) H Rxx R e xx
Rxx = E(x − µx )(x − µx ) = e ∗ . (13)
Rxx R∗xx
The complex augmented random vector x corresponding to x
is defined as The matrix Rxx is the conventional Hermitian covariance
x matrix
x= ∗ . (2)
x Rxx = E(x − µx )(x − µx )H = RHxx , (14)
From here the complex augmented random vector will always
and the matrix R
e xx is the complementary covariance matrix
be underlined. It’s easy to check that the real composite vector
z and the complex augmented vector x are related as e xx = E(x − µx )(x − µx )T = R
R eT . (15)
xx
x = Tn z. (3) Definition 1: If the complementary covariance matrix R e xx
The real-to-complex transformation Tn is is zero, then x is called proper, otherwise x is improper.
The random vector x = u + jv is proper if and only if
I jI Ruu = Rvv and Ruv = −RTuv , where u and v are the real
Tn = , (4)
I −jI and imaginary parts of x respectively.
which is unitary within a factor of 2:
Tn TH H III. W IDELY L INEAR MMSE E STIMATION
n = Tn Tn = 2I. (5)
The general linear MMSE estimation problem is to estimate
In fact, it is equation (3) that governs the equivalence between a complex random vector x from a complex random measure-
dual channel filtering for z and complex filtering for x. Given a ment vector y. The objective is to minimize the mean square
real linear transformation M ∈ R2m×2n and a composite real error between the signal x and its estimator x̂. Suppose the
vector z ∈ R2n , then the most general linear transformation output of the linear estimator is x̂ = Wy. The orthogonality
of the real channels u and v into the real channels a, b is principle says the estimator error e = Wy − x is orthogonal
a M11 M12 u to the measurement y:
= , (6)
b M21 M22 v
E[(Wy − x)yH ] = 0. (16)
Call y = a + jb. Then the corresponding complex augmented
vector y is We shall assume that both the signal x and the measurement
y have zero means, since this doesn’t change the structure of
y a 1 u
y = ∗ = Tm = Tm MTH n Tn = Hx, the optimum linear estimator. This yields the normal equations
y b 2 v
(7) WRyy = Rxy , (17)
3
where Ryy = EyyH and Rxy = ExyH . We wish to extend IV. W IDELY-L INEAR K ALMAN F ILTER
this result to a widely linear MMSE estimator. From section
II we know that the widely linear MMSE estimator should be Consider the widely linear state and measurement equations
of the form
xt = A1 xt−1 + A2 x∗t−1 + B1 wt−1 + B2 wt−1
∗
, t = 1, 2, ...,
x̂ = W1 y + W2 y∗ . (18) (27)
This is equivalent to yt = C1 xt + C2 x∗t + nt , t = 0, 1, ..., (28)
x̂ = Wy, (19) where xt ∈ Cn and yt ∈ Cp , A1 , A2 ∈ Cn×n , B1 , B2 ∈
Cn×m , and C1 , C2 ∈ Cp×n . The initial state vector x0 is
where assumed to have zero mean. Our procedure is to augment these
y W1 W2 equations as
y= , and W = . (20)
y∗ W2∗ W1∗
xt = Axt−1 + Bwt−1 , t = 1, 2, ..., (29)
Since the mean square error may be written
1 yt = Cxt + nt , t = 0, 1, ..., (30)
Ekx̂ − xk2 = Ekx̂ − xk2 , (21)
2
it is enough to minimize the mean-squared error between where xT = [xT , xH ], yT = [yT , yH ], and
x̂ and x. Applying the orthogonality principle, we have the
corresponding augmented normal equations A1 A2 B1 B2 C1 C2
A= ,B = ,C = . (31)
A∗2 A∗1 B∗2 B∗1 C∗2 C∗1
WRyy = Rxy (22)
Suppose Ex0 = 0, and the augmented covariance of x0 is
Therefore the remaining problem is to specify the matrices H T
" #
W1 and W2 . Actually the key question is to determine the Ex x
0 0 Ex x
0 0 Π 0 Π
e 0
Ex0 xH0 = Ex∗ xH = = Π0 . (32)
inverse of Ryy . Using the matrix-inversion lemma, the matrix 0 0 Ex∗0 xT0 e ∗ Π∗
Π 0 0
R−1
yy may be written as
Assume that Eut = 0 and Ent = 0 for all t. Further assume
" #−1 " # the real and imaginary parts of x0 are uncorrelated with the
R R P−1 −P−1 −∗
yy Ryy Ryy real and imaginary parts of ut and nt , meaning Ex uH = 0
R−1 = e yy
e yy yy
e
yy = , 0 t
R∗yy R∗yy −Q−1 e ∗ −1
yy Ryy Ryy Q−1 and Ex0 nH
yy t = 0 for t ≥ 0. Using the representation advocated
(23) in [28], the joint augmented second-order characterization of
where the matrix Pyy = Ryy − R e yy R−∗ Re ∗ is the (x , u , n ) is given by
yy yy 0 t t
Schur complement of Ryy . It is the error covariance matrix
for linearly estimating y from y∗ . And Qyy = R∗yy −
x0 Π0 0 0 0
e ∗ R−1 Re yy = P∗ . Thus R−1 may be simplified to the E wn xH H
R yy yy yy yy 0 wHm nH
m 1 = 0 δnm Q δnm S 0 .
augmented matrix nn 0 δnm SH δnm R 0
" # (33)
H H
P −1
−P −1 e
R R
yy yy
−∗ For n ≥ m, Ew n x m = 0 and En n xm = 0. For n > m,
R−1
yy =
yy
−∗ e ∗ −1
yy
−∗
(24) Ew yH = 0 and En yH = 0. This is the same setup as that
−Pyy Ryy Ryy Pyy n m n m
of the usual Kalman filter, but with covariances augmented to
account for non-zero complementary covariance.
From the Woodbury identities we know that P−1 −∗
yy Ryy Ryy =
e
Assume the estimator of xt−1 from measurements YTt−1 =
R−1 −∗ −∗ e ∗ −1
yy Ryy Pyy , and Pyy Ryy Ryy = (Pyy Ryy Ryy )
e −1 e −∗ ∗
=
−1 (yT1 , . . . , yt−1 )T is known to be x̂t−1|t−1 . Here we require
R−∗ e ∗ −1
yy Ryy Pyy . Thus Ryy may also be written as x̂t−1|t−1 to be a complex augmented vector. Then the LMMSE
"
−1 −1 e −∗
# prediction of xt from YTt−1 is x̂t|t−1 = Ax̂t−1|t−1 , and
P −R R P
yy yy T
R−1
yy =
yy yy
. (25) the prediction of yt from Yt−1 is ŷt|t−1 = Cx̂t|t−1 . These
−R−∗ e ∗ −1
yy Ryy Pyy P−∗
yy augmented prediction updates are equivalent to widely linear
prediction updates x̂t|t−1 = A1 x̂t−1|t−1 + A2 x̂∗t−1|t−1 and
So the widely linear estimator is [4]
ŷt|t−1 = C1 x̂t|t−1 + C2 x̂∗t|t−1 .
x̂ = (Rxy −R e xy R−∗ Re ∗ )P−1 y+(R e xy −Rxy R−1 R e yy )P−∗ y∗ . Suppose we are given the error covariance matrix Pt−1|t−1
yy yy yy yy yy
(26) for êt−1|t−1 = x̂t−1|t−1 − xt−1 . Then the error covariance
The linear MMSE estimator is a special case of this widely matrix Pt|t−1 for êt|t−1 = x̂t|t−1 − xt is
linear MMSE estimator. In [4] it has been shown that when es- " #
timating an improper complex signal in additive white proper H H Pt|t−1 P e t|t−1
Pt|t−1 = APt−1|t−1 A + BQB = e ∗ ,
noise, the maximum performance advantage of WLMMSE Pt|t−1 Pt|t−1
estimation over LMMSE estimation is a factor of 2. (34)
4
where the NW block of Pt|t−1 is the Hermitian error covari- where the NW block of Pt|t is the Hermitian error covariance
ance and the NE block is the complementary error covariance: and the NE block is the complementary error covariance:
+ B1 QBT2 + B2 Qe ∗ BT B1 QB
e T + B2 Q∗ BT . Algorithm 1 Widely Linear Kalman Filter
2 1 1
(35) Initialize with x̂0|0 = 0, P0|0 = Π0 and P e 0|0 = Πe 0 as defined
in equation (32);
The error covariance matrix St|t−1 for the innovation n̂t|t−1 = At the t-th iteration, t = 1, 2, · · · ,
ŷt|t−1 − yt is 1. Update the predicted state as x̂t|t−1 = A1 x̂t−1|t−1 +
" # A2 x̂∗t−1|t−1 ;
S S 2. Update the predicted Hermitian and complementary error
= CPt|t−1 C + R = e t|t−1
H
e t|t−1
St|t−1 , (36)
S∗t|t−1 S∗t|t−1 covariance matrices Pt|t−1 , P e t|t−1 , St|t−1 , and Se t|t−1 by
using equations (35) and (37);
where the NW block of St|t−1 is the Hermitian error covari- 3. Apply equation (40) to derive the widely linear Kalman
ance and the NE block is the complementary error covariance: gains Kt and K e t , wherein the Schur complement is PS =
St|t−1 =C1 Pt|t−1 CH e∗ H H St|t−1 − St|t−1 S−∗
e e∗
t|t−1 St|t−1 ;
1 + C2 Pt|t−1 C1 + C1 Pt|t−1 C2
e
4. Use the predictions x̂t|t−1 and ŷt|t−1 = C1 x̂t|t−1 +
+ C2 P∗t|t−1 CH
2 + R, C2 x̂∗t|t−1 to generate the widely linear estimator of xt
S T
=C1 Pt|t−1 C + C2 Pe∗ T e t|t−1 CT
t|t−1 C2 + C1 P
e t|t−1
2 1
1 for constant scalar |α| = 1. This is irrespective of whether connection between the complex augmented and real compos-
êt|t−1 is improper. For instance, assume yt = Re(xt ) + nt ite representations is
with nt real. We know that yt is a noisy widely linear " #
Rxx R e xx
transformation of xt and ŷt|t−1 = Re(x̂t|t−1 ), meaning the Rxx = e ∗ = ExxH = E[(Tz)(Tz)H ] = TRzz TH ,
innovation n̂t|t−1 is real and hence maximally improper. One Rxx Rxx
can also readily see that no widely linear processing is needed (47)
because yt is only dependent on the real part of xt . and hence
Example 1 (dual channel KF using WLKF): This example Rxx = Ruu + Rvv + j(RTuv − Ruv ),
extends the duality between widely linear and dual channel e xx = Ruu − Rvv + j(RT + Ruv ).
R uv
adaptive filtering made by Mandic et. al. [27] to Kalman filters.
Start with two real channels worth of random states ut and By applying the above connection between real composite and
vt , with the same dimension. Denote zTt = [uTt vtT ] as the complex augmented vectors, the conventional Kalman filter
corresponding real composite state. Suppose the composite [19], [28] of the dual real channel model (43)-(44) can be
state and measurement equations are written in the following complex augmented form:
x̂0|0 = 0, P0|0 = Π0 , (48)
u
zt = t = Ezt−1 + Fω t−1
vt H
Pt|t−1 = APt−1|t−1 A + BQB , H
E11 E12 ut−1 F11 F12 µt−1
= + , t = 1, 2, ..., x̂t|t−1 = Ax̂t−1|t−1 ,
E21 E22 vt−1 F21 F22 σ t−1
(43) ŷt|t−1 = Bx̂t|t−1 ,
and St|t−1 = CPt|t−1 CH + R,
ψt =
ξt
= Gzt + η t =
G11 G12 ut ρ
+ t , Kt = CPt|t−1 CH S−1
t|t−1 ,
κt G21 G22 vt φt
x̂t|t = x̂t|t−1 + Kt (yt − ŷt|t−1 ).
t = 0, 1, ...,
(44) Pt|t = (I − Kt C)Pt|t−1 .
where ω Tt = [µTt σ Tt ] and η Tt = [ρTt φTt ] are the composite The matrices Pt|t−1 , Pt|t , St|t−1 , Q, and R are augmented
driving and measurement noises, and ψ Tt = [ξ Tt κTt ] is the covariance matrices of the augmented errors xt − x̂t|t−1 and
composite measurement. This dynamical model allows the xt − x̂t|t , innovation yt − ŷt|t−1 , and noises wt and nt
states and measurements on the respective real channels to respectively. The matrix Kt is the augmented Kalman gain.
be arbitrarily coupled. In the following we will show that a By checking Algorithm 1, we see that the equations in (48)
WLKF can be used as the KF for zt . For the real composite are identical with the WLKF for the complex augmented
vectors zt , ω t , η t , and ψ t , establish their complex augmented dynamical model in (45)-(46). This means that the KF for
representations as a dual real channel model can be always implemented as
a WLKF for the corresponding complex augmented model.
T
zTt = [uTt vtT ] −−→ xTt = [xTt xH
t ],
Conversely, the WLKF may be organized into a real KF
over dual channels. Therefore, the duality between dual real
T
ω Tt = [µTt σ Tt ] −−→ wTt = [wtT wtH ], channel KF and the single complex channel WLKF holds if
and only if the conditions of state and measurement matrices,
T
η Tt = [ρTt φTt ] −−→ nTt = [nTt nH
t ],
A = 12 TETH , B = 12 TFTH , and C = 21 TGTH are
satisfied. This is implicit in [1]–[4], [6], [7] and explicit in
T
ψ Tt = [ξ Tt κTt ] −−→ yTt = [ytT ytH ], [27]. The key reason for the duality observed here is that for
any augmented random vector x and its estimator error e, the
The resulting augmented state and measurement equations are orthogonality principle says ExeH = TEz(z − ẑ)T TH = 0.
This happens if and only if Ez(z − ẑ)T = 0.
xt = Axt−1 + Bwt−1 , t = 1, 2, ..., (45) Example 2 (Performance Comparison between WLKF and
KF): Here we are interested in the evolution of the aug-
yt = Cxt + nt , t = 0, 1, ..., (46)
mented covariance matrix Pt|t for estimator error êt|t . Given
with A = 12 TETH , B = 12 TFTH , and C = 21 TGTH . Pt−1|t−1 , the augmented covariance matrix Pt|t is
For a real composite random vector zT = [uT vT ] with Pt|t =(I − Kt Ct )Pt|t−1
zero mean, its covariance matrix Rzz is −1
=(I − Pt|t−1 CH
t St|t−1 Ct )Pt|t−1
EuuT EuvT
T Ruu Ruv h
−1
i
Rzz = Ezz =
EvuT EvvT
=
RTuv Rvv = I − Pt|t−1 CH H
t (Ct Pt|t−1 Ct + R) Ct Pt|t−1
= I − (At Pt−1|t−1 AH H H H
The composite covariance matrix Rzz is determined by three t + Bt QBt )Ct (Ct (At Pt−1|t−1 At
−1
+ Bt QBH H
Ct · (At Pt−1|t−1 AH H
real blocks: Ruu , Ruv , and Rvv . Recall the complex aug- t )Ct + R) t + Bt QBt )
mented representation for z is x = Tz. The second order (49)
6
Suppose we want to estimate a complex random scalar xt . The Substituting (53) into (52), we have the minimum ξtWL :
corresponding state equations are
1
min ξtWL =
xt = at−1 xt−1 + bt−1 wt−1 , t = 1, 2, ..., (50) qt (2P0|0 ) + qt (0)
2
yt = ct xt + nt , t = 0, 1, ... (51) The ratio of min ξtWL to ξtSL is:
where at , bt , ct ∈ C for all t ≥ 0. The augmented matrices are min ξtWL qt (2P0|0 ) + qt (0)
At = diag(at , a∗t ), Bt = diag(bt , b∗t ), and Ct = diag(ct , c∗t ). θt = SL
=
ξt 2qt (P0|0 )
We make the assumption that Q = N1 I, and R = N2 I, where
N1 , N2 > 0. From (49) we may find the recursion for the 2 It’s obvious that 21 ≤ θt ≤ 1. This is because qt is concave, and
by 2 augmented covariance matrix Pt|t as qt (2P0|0 ) + qt (0) ≤ 2qt (P0|0 ) for any P0|0 . Also qt (2P0|0 ) ≥
Pt|t = I − |ct |2 (|at |2 Pt−1|t−1 + |bt |2 N1 I)(|ct |2 (|at |2 Pt−1|t−1
qt (P0|0 ) for any P0|0 and qt (0) ≥ 0. Actually the condition for
achieving the lower bound is N1 N2 1. This coincides
+ |bt |2 N1 I) + N2 I)−1 · (|at |2 Pt−1|t−1 + |bt |2 N1 I),
with the MMSE analysis in [4]. Here we have assumed: (i) the
t = 1, 2, ... complex state equations (50) and (51) are strictly linear, (ii)
the state is scalar, and (iii) the noises wt and nt are white and
Suppose the initial augmented state covariance is
" # proper for all t. Only the prior state is improper. For a vector-
P0|0 Pe0|0 valued state coupled with widely linear update equations and
P0|0 = e∗ ∗ , improper colored noises, the advantage of the WLKF over the
P0|0 P0|0
SLKF may be much greater.
with eigenvalues {λ01 , λ02 }. We can show that given the eigen-
values {λt−1 t−1
1 , λ2 } of matrix Pt−1|t−1 , the eigenvalues of
Pt|t are V. E XTENDED W IDELY L INEAR K ALMAN F ILTER
λti = gt (λt−1i ), i = 1, 2, In this section it is assumed that the complex state sequence
where the function gt is given by {xt } evolves according to the nonlinear model
their Taylor expansion around the state estimate x̂t−1|t−1 : Therefore we can derive the extended WLKF along the lines
of Section III and summarize the filter as Algorithm 2 on the
∂ft−1
xt =ft−1 (x̂t−1|t−1 , 0) + (x̂t−1|t−1 , 0)(xt−1 − x̂t−1|t−1 ) top of next page. It can be observed from Alg. 2 that the
∂x nonlinear functions f and h are used to predict xt and yt as
∂ft−1 in the conventional EKF. The complex Jacobians are used to
+ (x̂t−1|t−1 , 0)wt−1
∂w modify the Hermitian and complementary covariance matrices
∂ft−1 and Kalman gain. This treatment of the EWLKF is equivalent
= (x̂t−1|t−1 , 0)xt−1 + ft−1 (x̂t−1|t−1 , 0)
∂x to the original treatment of Mandic, et al. [9], [26]. The only
∂ft−1 ∂ft−1
− (x̂t−1|t−1 , 0)x̂t−1|t−1 + (x̂t−1|t−1 , 0)wt−1 difference is that the authors of [9], [26] define updates for
∂x ∂w augmented covariances, which of course can be resolved for
=Ft,1 xt−1 + Ft,2 x∗t−1 + v̂t−1 + Lt,1 wt−1 + Lt,2 wt−1 ∗ updates of Hermitian and complementary covariances, whereas
=Ft xt−1 + v̂t−1 + Lt wt−1 , our updates for Hermitian and complementary covariances can
be assembled into updates for augmented covariances.
where
∂ft−1
Ft = Ft,1 Ft,2 = (x̂t−1|t−1 , 0) VI. U NSCENTED W IDELY L INEAR K ALMAN F ILTER
∂x
∂ft−1 ∂ft−1
The major defect of extended Kalman filters is that, since
= (x̂t−1|t−1 , 0) (x̂t−1|t−1 , 0) , the approximated model is obtained by linearizing state and
∂x ∂x∗
measurement equations, the posterior means and covariances
∂ft−1 are accurate only to the first order in a Taylor expansion.
Lt = Lt,1 Lt,2 = (x̂t−1|t−1 , 0)
∂w Unscented Kalman filters are motivated by the inaccuracy
∂ft−1 ∂ft−1 issues of the EKF. An UKF uses the unscented transformation
= (x̂t−1|t−1 , 0) (x̂t−1|t−1 , 0) ,
∂w ∂w ∗ (UT) to generate a fixed set of sigma points to represent the
distribution of a random variable [30], [31]. After propagating
and the term v̂t−1 = ft−1 (x̂t−1|t−1 , 0) − Ft x̂t−1|t−1 depends
sigma points through nonlinearities, the estimated posterior
only on the estimator x̂t−1|t−1 . Similarly we may approximate
mean and covariance are precise at least to second order in
yt as
a Taylor expansion. The common approach is to compose
∂ht complex random states and noises into a complex vector
yt =ht (x̂t|t−1 , 0) + (x̂t|t−1 , 0)(xt − x̂t|t−1 )
∂x sT = [xT wT nT ] and construct the sigma points of s. The
∂ht first and second order statistical information of s is
+ (x̂t|t−1 , 0)nt
∂n
∂ht µTs = [µTx µTw µTn ], (59)
= (x̂t|t−1 , 0)xt + ht (x̂t|t−1 , 0)
∂x
Rxx 0 0 R
e xx 0 0
∂ht ∂ht
− (x̂t|t−1 , 0)x̂t|t−1 + (x̂t|t−1 , 0)nt Rss = 0 Q 0, R e ss =
0 Q
e 0 .
∂x ∂n
0 0 R 0 0 R
=Ht,1 xt + Ht,2 x∗t + ŵt + Mt,1 nt + Mt,2 n∗t
e
+ Mt,1 RMH e∗ H e H ∗ H
t,1 + Mt,2 R Mt,1 + Mt,1 RMt,2 + Mt,2 R Mt,2 ,
S e∗
e t|t−1 = Ht,1 Pt|t−1 HT + Ht,2 P T T ∗ T
t|t−1 Ht,1 + Ht,1 Pt|t−1 Ht,2 + Ht,2 Pt|t−1 Ht,2
e
t,1
4. Use the predictions x̂t|t−1 and ŷt|t−1 = ht (x̂t|t−1 , 0) to generate the widely linear estimator of xt
e t yt − ŷt|t−1 ∗ ,
x̂t|t = x̂t|t−1 + Kt yt − ŷt|t−1 + K
5. Update the Hermitian and complementary error covariance matrices
Pt|t = I − Kt Ht,1 − K e t H∗ Pt|t−1 − Kt Ht,2 + K e t H∗ P e∗
t,2 t,1 t|t−1 ,
P e t H∗ P
e t|t = I − Kt Ht,1 − K e t H ∗ P∗
e t|t−1 − Kt Ht,2 + K
t,2 t,1 t|t−1 .
Denote the vector bk as the k-th column of matrix B for augmented vectors {X k } as
k = 1, 2, ..., 2N . Then the sigma points {Z k } of z are [30]–
Xk
[32] X k = TZ k =
X ∗k
Z 0 = µz , (62) µs , √
k = 0,
√ = µs + 2N + λTbk , k = 1, ..., 2N,
Z k = µz + 2N + λbk , k = 1, ..., 2N, √
√
µs − 2N + λTbk , k = 2N + 1, ..., 4N.
Z k = µz − 2N + λbk , k = 2N + 1, ..., 4N, (63)
We can show that all the X k compose the sigma points of the
corresponding to the following weights: augmented vector s, since X 0 = µs and
Rss = TRzz TH
(
λ/(2N + λ), k = 0,
mean weights: Wm (k) = 2n
1/[2(2N + λ)], k = 1, ..., 4N, X
= T( bk bH
k )T
H
Algorithm 3 Unscented Widely Linear Kalman Filter Hermitian covariance Rss , but also complementary covariance
Initialize with x̂0|0 = 0, P0|0 = Π0 and P e 0|0 = Πe 0 as defined Re ss . In [9] the authors proposed complex sigma points of s
in equation (32); which carry moments µs and Rss , but not R e ss .
At the t-th iteration, t = 1, 2, · · · , But what really concerns us is whether the modified sigma
Replace Pt−1|t−1 with Rxx , and P e t−1|t−1 with R e xx , and points will refine the propagation of mean and covariance
µTs,t = [x̂Tt−1|t−1 0 0] with µTs in (59). Construct effective through the nonlinearities f and h. Suppose we push each
sigma points {Xk,t T x
} = {[(Xk,t−1 )T (Xk,t−1
w
)T (Xk,t
n T
) ]} by modified sigma point X k through any non-linearity Y k =
using equations (59)-(63). g(X k ). Compute the following deterministic sample averages:
Prediction updates: 4N
X 4N
X
µ̂y = Wm (k)Y k = Wm (k)g(X k ),
X xk,t|t−1 = ft−1 (X xk,t−1 , X w
k,t−1 ) k=0 k=0
4N
X 4n
X
x̂t|t−1 = Wm (k)X xk,t|t−1 , Pyy = Wc (k)(Y k − µ̂y )(Y k − µ̂y )H ,
k=0 k=0
4N 4n
X
X
Pt|t−1 = Wc (k)(X xk,t|t−1 − x̂t|t−1 )(X xk,t|t−1 − x̂t|t−1 )H , P
e yy = Wc (k)(Y k − µ̂y )(Y k − µ̂y )T .
k=0 k=0
4N
X From the argument of [30], it can be verified that the sample
P
e t|t−1 = Wc (k)(X xk,t|t−1 − x̂t|t−1 )(X xk,t|t−1 − x̂t|t−1 )T , averages µ̂y , and Pyy , and Pe yy may be used to approximate
k=0 the mean E(f (s)), the Hermitian, and complementary covari-
Y k,t|t−1 = gt (X xk,t|t−1 , X nk,t ), ance matrices of f (s). The approximation is precise up to at
least second order in a Taylor expansion. The unscented widely
4N
X linear Kalman filter is described in Algorithm 3, in notation
ŷt|t−1 = Wm (k)Y k,t|t−1 , that follows [32].
k=0 Example 3 (Phase Demodulation Problem): Consider a
Measurement updates: scalar real random phase θt which is updated as
4N
X θt = aθt−1 + bwt−1 , t = 1, 2, ..., (64)
St|t−1 = Wc (k)(Y k,t|t−1 − ŷt|t−1 )(Y k,t|t−1 − ŷt|t−1 )H ,
k=0 where wt is a real driving noise. So, the phase is real. The
4N
measurement of θt is a noisy complex signal modulated by
X θt :
S
e t|t−1 = Wc (k)(Y k,t|t−1 − ŷt|t−1 )(Y k,t|t−1 − ŷt|t−1 )T ,
yt = eiθt + nt , t = 0, 1, ... (65)
k=0
4N
X where each nt is assumed to be a zero mean, scalar complex
Et|t−1 = Wc (k)(X xk,t|t−1 − x̂t|t−1 )(Y k,t|t−1 − ŷt|t−1 )H , Gaussian random variable [2], [4] with Hermitian variance
k=0 R and complementary variance R. e Suppose nt = ut + jvt ,
4N where ut and vt have variances Ruu and Rvv respectively and
covariance Ruv . The complex correlation coefficient between
X
E
e t|t−1 = Wc (k)(X xk,t|t−1 − x̂t|t−1 )(Y k,t|t−1 − ŷt|t−1 )T ,
k=0 nt and n∗t is
Fig. 1. Phase estimated by UWLKF at each iteration, SNR = 30dB, |ρ| = 0.5.
Fig. 3. Performance improvement r of UWLKF over UKF vs impropriety
of nt .