You are on page 1of 7

2166 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO.

11, NOVEMBER 2007

Filtering of Stochastic Nonlinear Differential Systems via a particle filters [26], Gaussian sum approximations [19], the Unscented
Carleman Approximation Approach Kalman Filter (UKF) [21]. When measurements are considered in con-
tinuous time, the Extended Kalman Bucy Filter (EKBF) (see [13]) can
Alfredo Germani, Member, IEEE, Costanzo Manes, and be applied. More recently, a polynomial extension of the EKF (denoted
Pasquale Palumbo PEKF) has been proposed in [15], which is based on the application of
the optimal polynomial filter of [6], [7] to the Carleman approximation
of the nonlinear discrete-time system [22].
Abstract—This paper deals with the state estimation problem for This paper presents a procedure for the filter design for systems of the
stochastic nonlinear differential systems, driven by standard Wiener type (1) that generalizes the standard extended Kalman-Bucy approach,
processes, and presents a filter that is a generalization of the classical
Extended Kalman-Bucy filter (EKBF). While the EKBF is designed on
and avoids time discretization. The main step consists in computing a
the basis of a first order approximation of the system around the current Carleman-like approximation of a chosen degree  of the original non-
estimate, the proposed filter exploits a Carleman-like approximation of linear stochastic differential system (1), in the form of a bilinear system
a chosen degree 1. The approximation procedure, applied to both (linear drift and multiplicative noise) with respect to a suitably defined
the state and the measurement equations, allows to define an approximate extended state. In general, there may be good reasons for computing
representation of the system by means of a bilinear system, for which a
filtering algorithm is available from the literature. Numerical simulations bilinear approximations of nonlinear systems [3], and therefore bilin-
on an example show the improvement, in terms of sample error covariance, earization techniques have been used in the past in problems of systems
of the filter based on the first-order, second-order and third-order system approximation [28], [29], [31].
approximations ( = 1 2 3). Once the Carleman bilinear approximation of a system is computed,
Index Terms—Carleman approximation, extended Kalman-Bucy filter, the equations of the optimal linear filter for stochastic bilinear dif-
nonlinear filtering, Polynomial filtering. ferential systems presented in [9] can be applied without any further
approximation (note that the paper [9] presents the optimal polyno-
mial filter specifically worked out for bilinear systems, and therefore
I. INTRODUCTION the technique can not be directly applied to nonlinear systems of the
This note considers the filtering problem for nonlinear stochastic dif- form (1)). When  = 1 the proposed filtering algorithm reduces to
ferential systems described by the Itô equations the classical Extended Kalman-Bucy filter (EKBF), consisting of the
Kalman-Bucy filter equations applied to the linear approximation of
( ) = (x(t); u(t))dt + F dW 1 (t); x(0) = x0
dx t the differential system. Better performances of filters designed using
dy (t) = h(x(t); u(t))dt + GdW (t); y (0) = 0; a:s:; (1)
2
higher order system approximations are expected.
The paper is organized as follows: the next section presents the Car-
defined on a probability space (
; F ; P ), where x(t) 2 IRn is the leman approximation of stochastic nonlinear differential systems of the
state vector, u(t) 2 IRp is a known deterministic input, y (t) 2 IRq type (1); in Section Three the optimal linear filter for the Carleman bi-
is the measured output, W 1 (t) 2 IRs and W 2 (t) 2 IRq are in- linear approximation is derived; Section Four displays some numerical
dependent standard Wiener processes with respect to a family of in- results where the performances of the proposed algorithm are compared
creasing  -algebras fFt ; t  0g (i.e., the components of vectors with those of an EKBF.
W 1 (t) and W 2 (t) are a set of independent standard Wiener processes).
 : IRn 2 IRp 7! IRn and h : IRn 2 IRp 7! IRq are smooth non- II. CARLEMAN APPROXIMATION
linear maps. The initial state x0 is an F0 -measurable random variable, In order to compute the Carleman approximation of a chosen order
independent of both W 1 and W 2 . In order to avoid singular filtering  of the system (1), with  positive integer, it is necessary to assume
problems, see [5], the standard assumption of nonsingular output-noise that the random vector x0 has finite moments up to the degree 2
covariance is made here, i.e., rank(GGT ) = q . However, the approach
< +1; i = 1; . . . ; 2;
[i]
presented in [8] could be followed when the covariance of the measure- i = IE x0 (2)
ment noise is singular or even zero.
It is well known that the minimum variance state estimate requires where the square brackets at the exponent denote the Kronecker powers
the knowledge of the conditional probability density, whose computa- (see [7] for a quick survey on the Kronecker product and its main prop-
tion, in the general case, is a difficult infinite-dimensional problem [4], erties).
[23], [24], [32]. For this reason a great deal of work has been made Under standard analyticity hypotheses, both the state and the output
in the past to study implementable approximations of the optimal esti- equations can be written by using the Taylor polynomial approxima-
mator [11], [12], [16], [18]. tion around a suitably chosen state x  2 IRn . When using the Taylor
Another approach to state estimation consists in considering the time approximation for the filter design, the state x  can be chosen as a fixed
discretization of the original system and then to apply known filtering state that defines a nominal working point of the system or, as an alter-
procedures, like the Extended Kalman Filter (EKF), the most widely native, as the current state estimate (more details are given in Remark
used algorithm in nonlinear filtering problems (see, e.g., [13], [20]), 1, at the end of Section III). According to the Kronecker formalism, the
differential system in (1) becomes
1 s
Manuscript received March 13, 2007; revised July 27, 2007. Recommended
by Associate Editor M.-Q. Xiao. This work was supported by the Italian Min-
dx(t) = 8i (x; u(t))(x(t) 0 x)[i] dt + Fj dWj1 (t)
i=0 j =1
istry of University and Research (MIUR) under Grant PRIN 2005097521_003,
1 q
x; u(t))(x(t) 0 x
and the National Research Council (CNR).
dy (t) = Hi ( ) dt + Gj dWj2 (t) (3)
[i]
A. Germani and C. Manes are with the Dipartimento di Ingegneria Elettrica
e dell’Informazione, Università degli Studi dell’Aquila, 67040 L’Aquila, Italy i=0 j =1
(e-mail: germani@ing.univaq.it; manes@ing.univaq.it).
P. Palumbo is with the Istituto di Analisi dei Sistemi ed Informatica del CNR where Fj and Gj are the columns of F and G, respectively, and

8i (x; u) = i1! r[xi]


 ) = i1! r[xi]
h
“A. Ruberti,” 00185 Roma, Italy (e-mail: palumbo@iasi.rn.cnr.it).
Digital Object Identifier 10.1109/TAC.2007.908347 ; (
Hi x; u : (4)

0018-9286/$25.00 © 2007 IEEE


IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007 2167

The differential operator rx


applied to a function = (x) :
[i]
where
IRn 7! IRp is defined as follows
s

F0 = Fj[2] : (13)
r
=
[0]
x ; r [i+1]
x
=r
r
x
[i]
x ; i  1 (5) j =1

with rx = [@=@x1 1 1 1 @=@xn ] and rx


the Jacobian of the
The use of identities (9) and (10) gives the following expression for the
differentials d(x[k] (t)), for k  2:
vector function .
For the reader’s convenience, some useful properties of the Kro-
= U I
x[ 01] (t) (x(t); u(t))dt
necker product and of the differential operator (5) are reported below.
The following property is intensively used in this paper: for any ma- d x[k] (t) k
n n
k

trices A; B; C; D of suitable dimensions it is


+ 12 O I
x[ 02] (t) F0 dt
k
n n
k

+ U I
x[ 01] F dW 1 (t):
k k
(14)
(A 1 B )
(C 1 D ) = ( A
C ) 1 (B
D )
n n
(6)
Lemma 2: Consider system (1). For k  2, the differential of the
where A 1 B denotes the standard matrix product. Recall that the Kro- k -th order Kronecker power of the state x(t) can be written as
necker product is not commutative. Given a pair of integers (a; b), the
symbol Ca;b denotes a orthonormal commutation matrix in f0; 1ga1b2a1b
such that, given any two matrices A 2 IRr 2c and B 2 IRr 2c
1
d x[k] (t) = Akr (x; u(t))(x(t) 0 x)[r] dt
r =0
s

B
A = CrT ;r (A
B )Cc ;c : (7) + B jk x[k01] (t)dWj1 (t) (15)
j =1

The Kronecker power of a binomial, (a + b)[h] , allows the following


expansion: with

Akr (x; u) = Ak;a


r (x; u) + Ak;b
r (x) (16)
h

(a + b) [ ] =
h
Mjh a[j ]
b[h0j ] ; 8a; b 2 IR n
(8) where
j =0

with Mjh suitably defined matrix coefficients in IRn2n (see [7]).


r

Throughout the paper, the symbol In will denote the identity matrix
Ak;a
r (x; u) = Unk (8i (x; u)
In )
of order n.
i=0 _(r0k+1)
Lemma 1: For any x 2 IRn the following identities hold: 1
M 00 I
x 0 0
In k 1
r i n
[k r +i 1]
(17)

Ak;b
r (x) = 21 O (F
I )M 0 x 0 0
I
k
n 0 n
k
r
2 [k 2 r]
n ;
r
x =U
x
[h] h
n In
x 01] ; h  1
[h
(9) r k02 (18)

r
x =O
[2] [h] h
In
x[h02] ; h > 1 (10)
A k;b
r (x) = 0; r > k 0 2 (19)
x n
and

where Unh and Onh , for h > 1 are recursively computed as


B = U (F
I
jk
k
n j n ): (20)

Unh = In +C T
Unh01
In Proof: Consider the first term of (14) and replace (x; u) with
n ;n
the power expansion given in (3). The repeated use of (6) gives
Onh = Unh CnL ;n Unh01 CnL ;n
I n CnT ;n (11)

with initial value Un1 = In (the proof is reported in the Appendix). Unk In
x[k01] (x; u)
1
Unk In
x[k01]
In the derivation of the Carleman approximation of system (1) the Itô
formula for the computation of stochastic differentials written using the = 8 (x; u)(x 0 x)[ ]
i
i

i=0
Kronecker formalism is required. Using (70), proved in the Appendix 1
along with Lemma 1, the differential of the Kronecker power x[k] can
be written as
= Unk 8 (x; u)(x 0 x)[ ]
x[ 01]
i
i k

i=0
1
= Unk (8i (x; u)
In ) (x 0 x)[ ]
x[ 01] i k

d x[k] (t) = r
x[ ] (x(t); u(t))dt
x
k i=0
1
+ 12 r[2]
x[ ] F0 dt k = Unk (8i (x; u)
In )
x
i=0

+ r
x[ ] F dW 1 (t)
x
k
(12) 2 (x 0 x)
((x 0 x) + x) 0
[i] [k 1]
: (21)
2168 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007

By applying (8) to the (k 0 1)-th Kronecker power in (21) i.e., from definitions (18)

1
Unk In
x[k01] (x; u) 1 Ok
1 2 n In
x k 0 ( t)
[ 2]
F0 = r (
Ak;b x)(x 0 x
)[r] : (29)
r=0
= Unk (8i (x; u)
In )
i=0 Consider now the last term in (14). By repeated use of the property
k01
Mjk01 (x 0 x)[j]
x[k010j]
(6) it can be written as
2 (x 0 x) i
[ ]

j =0
1 k01 Unk In
x[k01] F dW 1
= x; u)
In
Unk (8i ( ) In
Mjk0 1
s
Unk In
x[k01]
i=0 j =0
= (Fj
1)dWj1
2 (x 0 x) [i+j ]

x 010j]
[k
j =1
1 k01 s
= x; u)
In
Unk (8i ( ) In
Mjk0 1 = Unk Fj
x k0
[ 1]
dWj1
i=0 j =0 j =1
s
2 In
x 010j] (x 0 x)[i+j] :
[k
(22) = Unk (Fj 1 1)
(In 1 x k0 )
[ 1]
dWj1
j =1
After the change of index r = i + j , the sums in (22) become s

1 i+k01
= Unk (Fj
In )x k 0 [ 1]
dWj1 ; (30)

Unk In
x [k 01] (x; u) = Unk (8i (x; u)
In )
j =1

i=0 r=i i.e., considering the definition (20)


2 In
Mrk00i 1
In
x k0r i0 (x 0 x) r :
[ + 1] [ ]
(23)
s
From the following equivalence between summations: Unk In
x[k01] F dW 1 = Bjk x k0 [ 1]
dWj1 : (31)
j =1
1 i+k01 1 r
( )i;k;r = ( )i;k;r ; (24)
The Lemma is proved by substitution of (25), (29), and (31) into (14).
i=0 r=i r=0 i=0_(r0k+1)
By neglecting in the summations in (3) and (15) the higher order
it follows that terms, greater than a chosen degree  , the differentials d(x[k] (t)); k =
1 1; 2; . . ., given by (1) for k = 1 and by (14) for k  2, and dy(t) are
Unk In
x[k01] (x; u) = r (x; u)(x 0 x
)[r] approximated as follows:
Ak;a (25)
r=0

where Ak;ar (x; u) is given in (17). Now consider the second term in d x[k] (t) ' x; u(t))(x(t) 0 x
Akr ( )[r] dt
(14). It can be written as r=0
s
+ Bjk x k0 (t) + Fjk
[ 1]
dWj1 (t);
1 Ok In
x[k02] F0 = 12 Onk F0
x[k02] j =1
2 n 

= 12 Onk (F0
In )x[k02] : (26) dy (t) '
i=0
x; u(t))(x(t) 0 x
Hi ( )[i] dt + GdW 2 (t) (32)

Since it is with Akr (


x; u) and Bjk , for k  2 defined by (16)–(20), and for k = 1
x[k02] = (x + (x 0 x))[k02]
k02
A1r (
x; u) = 8r (
x; u) 8r = 0 ; . . . ; 
= Mrk02  020r]
(x 0 x)[r]
x [k B j = 0 8j = 1 ; . . . ; s
1 (33)
r=0
k02 moreover
= Mrk02 x
[k020r]
In (x 0 x)[r] ; (27)
r=0

from (26) and (27) it follows Fj = Fj ; Fjk = 0; 8j = 1; . . . ; s; 8k > 1:


1 (34)

Consider now the following equivalence


k02
1 Ok In
x [k 02] (t) F0 =
1 Ok (F
I )M k02
2 n 2 n 0 n r
 
r=0
2 x[k020r]
In (x 0 x)[r] ; (28)
x; u)(x 0 x
Akr ( )[r] = Ajk (x; u)x j [ ]
(35)
r=0 j =0
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007 2169

with with X  (0) = (xT0 1 1 1 (x  )T )T ; k = 1; . . . ; ; Y  (0) = 0, and


[ ]
0

 A 1 1 1 A
1 1
A 1

Akj (x; u) = (01)r0j Akr (x; u)Mjr


x r0j
1 0

A = N =
[ ]
In (36) .. .. .. ..
. . . ; . (43)
r=j
A 111 A

1

 A 0
derived from the following identities 0 0 111 0 Fj 1
 Bj 0 1 1 1 0 Fj
B = F =
2 2
k
( )( 0 x)[r]
Ar x; u x

j .. .. .. ;

j .. (44)
r=0 0 . . . .
 r 0 0 Bj 0 Fj
= k
( )
Ar x; u Mj
r
x
[j ]

(0x) r0j [ ]
C  = [ C 1 1 1 C  ]; D  = C
1 0 (45)
r=0 j =0
 
= (01)r0j Akr (x; u)Mjr In
x r0j
[ ]
x
[j ]
: (37) III. THE FILTERING ALGORITHM
j =0 r=j
From the discussion in the previous section, it follows that the
Analogously, it comes that bilinear Carleman approximation (42) is an approximated generation
  model for the sequence (x(t); y (t)), produced by the exact model (1),
( )( 0 x)[i]dt =
Hi x; u x Cj (x; u)x j dt[ ]
(38) and therefore they can be used for the construction of an approximate
filter for system (1). From the approximation
i=0 j =0

with ( ) ' In On2(n 0n) X  (t)


x t (46)
 ~  (t) of the extended state X  (t) is known, then the
Cj (x; u) = (01)i0j Hi (x; u)Mji In
x i0j
[ ]
: (39)
if an estimate X
following estimate of x(t) is obtained
i=j
~(t) = In On2(n 0n) X~  (t):
x (47)
It follows that the approximation (32) of the differentials d(x[k] (t)) and
dy (t) can be rewritten, for k  1, as It is well known that the optimal estimate of X  (t) is provided by the
 conditional expectation w.r.t. all the Borel transformations of the mea-
d x
[k]
( t) ' Ajk (x; u(t)) x j (t) dt [ ] surements, whose computation in general can not be obtained through
j =0 algorithms of finite dimension. However, recalling that X  (t) is gen-
s erated by the bilinear system (42), the suboptimal filter presented in
+ Bjk x k0 (t) + Fjk
[ 1]
dWj t ;
1
() [9], specifically worked out for bilinear systems, can now be applied.
j =1 In particular, the best affine estimator is used here. The output of such
 filter is the projection of X (t) onto the space L(Yt ) of all the affine
dy t ( )' Cj (x; u(t)) x j (t)dt + GdW (t):
[ ] 2
(40) transformations of the random variables fY  ( ); t0    tg. Let us
j =0
denote X ^  (t) = 5[X  (t) j L(Yt )] such projection (formally, the pro-
The  -degree Carleman bilinear approximation of the stochastic dif- jection 5 on the subspace L(Yt ) is a random variable in L(Yt ) such
 2 IRn is a system with state
ferential system (1) around a given x that the difference X  (t) 0 5[X  (t)jL(Yt )] is orthogonal to L(Yt ),
X (t) 2 IR , with n = n + n + 1 1 1 + n , and output Y  (t) 2 IRq
 n 2 i.e., is uncorrelated with all random variables in L(Yt )). The estimate
that obey the equations of x(t) obtained from X ^  (t) using (47), is denoted x^ (t), i.e.
 ^ (t) =
x In On2(n 0n) X^  (t)

dXk t ( )= A ( ( )) ( )
k
j x; u t

Xj t dt
= In On2(n 0n) 5 [X  (t) j L (Yt )] : (48)
j =0
In the filter equations the mean and covariance of X  (t) are required.
s
+ (Bjk Xk01 (t) + Fjk ) dWj1 (t) Lemma 3: Let mX (t) = IE fX  (t)g and 9X (t) = Cov(X  (t))
denote the mean value and the covariance matrix of X  (t), respec-
j =1

dY

( t) = Cj (x; u(t))Xj (t)dt + GdW (t) 2
(41) tively. These obey the following equations:
j =0
_ (t) = A (x; u(t))mX (t) + N (x; u(t))
mX

where Xk (t) 2 IRn ; k = 1; . . . ;  , denotes the k -th block component 9_ X (t) = A (x; u(t))9X (t)
of the state X  (t). Comparing the (32) and (41), it is clear that Xk (t) + 9X AT (x; u(t)) + Q(mX (t); 9X (t)) (49)
is aimed to approximate x[k] (t); k = 1; . . . ;  , while Y  (t) should
approximate y (t). where
s
Bi 9X BT
The (41) can be put in a compact matrix form as follows
(
Q mX ; 9X ) = i
dX

(t) = A (x; u(t))X (t)dt + N (x; u(t))dt
   i=1
s
s
+ Bj X  (t) + Fj dWj1(t); + (Bi mX + Fi ) (Bi mX + Fi )T (50)
j =1 i=1
dY

(t) = C (x; u(t))X  (t)dt with initial values mX (0) = (1T 1 1 1 T )T ; 9X (0) = Cov(X0 ),
+ D (x; u(t))dt + GdW 2 (t); (42) where the vectors k are the moments defined in (2).
2170 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007

The proof of Lemma 3 can be found in [9].


^ ()
TABLE I
Theorem 1: The optimal linear estimate X  t of the process X  t() MEAN SQUARE ESTIMATION ERRORS FOR THE FILTER (51)–(53)
given by (42), obeys the equation

^ ( ) = A (x; u(t))X^  (t)dt


dX  t
+ N (x; u(t)) dt + P (t)CT (x; u(t))R01
1 (dY  (t) 0 (C (x; u(t))X^  (t) state that allows the derivation of filter equations with smaller size
+ D (x; u(t)))dt) (51) (see [15] for further details).

with R = GGT and P (t) is the error covariance matrix IV. SIMULATION RESULTS

P (t) = IE f(X  (t) 0 X^  (t))(X  (t) 0 X^  (t))T g (52) Numerical simulation results are here reported in order to show the
effectiveness of the proposed algorithm. Consider the following non-
linear system:
evolving according to the following equation
( ) = (0x1 (t) + x1 (t)x2 (t))dt + adW 1 (t);
dx1 t
P_ (t) = A (
x; u(t))P (t) + P (t)AT (
x; u(t)) ( ) = (02x2 (t) 0 2x1 (t)x2 (t))dt + bdW 1 (t)
dx2 t
+ Q(t) 0 P (t)CT (x; u(t))R01 C (x; u(t))P (t) (53) ( ) = (x1 (t) 0 x1 (t)x2 (k))dt + dW 2 (t)
dy t (56)

with P (0) = 9X (0), where Q(t) in (53) is defined as Q(t) = with a = b = 1; = 2. The initial state x0 is a Gaussian standard
Q(mX (t); 9X (t)), with Q(1; 1) defined by (50). random vector (zero mean and identity covariance). Both filters
Proof: The proof is a straightforward consequence of ([9], Thm. (51)–(53) and (54)–(55) have been implemented for  = 1; 2; 3;
4.4). The equations are somewhat different (and shorter), because the using MATLAB©. The bilinearization point x  = 0 has been
() ()
state noise W1 t and output noise W2 t have been assumed indepen- chosen for the filter (51)–(53). The computations needed for the
dent in this paper. filter derivation begin with the computation of the differential
Remark 1: The filter given in Theorem 1 provides the optimal affine () = [ () (
dx[2] t )( ) ( )( )
dx21 t d x1 x2 t d x1 x2 t dx22 t (note the re- ( )]
() ()
estimate of X  t as a function of the observations Y  t . From this, ()
dundancy in the components of dx[2] t ). The application of the Itô
^ ()
the estimate x t is computed using (48). However, the available mea- formula (68) gives
() ()
surement process is y t and not Y  t . Therefore, the differential
() ()
dY  t in (51) should be replaced with dy t . Note that in the filter dx21 t( ) = 2x1 (t)dx1 (t) + a2 dt
A N C D
(51)–(53) the matrices  ;  ;  and  depend on x and u t  () (
d x 1 x2 t)( ) = x2 (t)dx1 (t) + x1 (t)dx2 (t) + ab dt
8 ( ) ( )
through the terms i x; u and Hi x; u , defined in (4) as the co- dx22 t( ) = 2x2 (t)dx2 (t) + b2 dt
( )
efficients of the polynomial approximations of  x; u and h x; u ( ) (57)

around x. Following a standard EKBF approach, these coefficients can from which, substituting the differentials (56), we have
be re-computed at each time t as a function of the current estimate
^ ()  ^ ()
x t . Formally, this kind of filter is written replacing x with x t ( ) = 02x12 (t) + 2x21 (t)x2 (t) + a2 dt + 2x1 (t)adW 1 (t)
dx21 t
d(x1 x2 )(t) = 03x1 (t)x2 (t) + x1 (t)x22 (t) 0 2x21 (t)x2 (t) + ab dt
into the filter (51)–(53) as follows

^ ( ) = A (^x (t); u(t))X^  (t)dt + (ax2 (t) + bx1 (t))dW 1 (t)


dX  t
dx22 (t) = 04x22 (t) 0 4x1 (t)x22 (t) + b2 dt + 2x2 (t)bdW 1 (t)
+ N (^x (t); u(t))dt + P (t)CT (^x (t); u(t))R01
1 (dy(t) 0 (C (^x (t); u(t))X^  (t) + D (^x (t); u(t)))dt)
(58)

(54) Note that for the filter design with  =2


and x =0
the third order
_ ( ) = A (^x (t); u(t))P (t) + P (t)AT (^x (t); u(t)) + Q(t)
P t ()
terms x21 x2 and x1 x22 are neglected in the differential dx[2] t . Due to
0 P (t)CT (^x (t); u(t))R01 C (^x (t); u(t))P (t) (55) ()
lack of space, the differential dx[3] t is not reported.
The results displayed below are obtained simulating both system

Remark 2: For  = 1 the (54)–(55) coincide with those of the


(56) and filter (51)–(53), in the time interval [0, 100] according to the

EKBF. However, note that for  = 2 such equations do not coincide


Euler-Maruyama method [17] with integration step 1=01
: . The per-
formance of the filters are evaluated computing the mean of the squared
with those of the so called second-order EKBF (see [13]), in which estimation errors (MSE)
the Riccati equations are exactly the same of the plain EKBF, and a
= N 1+ 1
N
second order Taylor approximation is used for the state process, where
the second order state increments are substituted with the components x ; (xi (tk ) 0 x^i (tk ))2 (59)
k=0
of the error covariance matrix provided by the Riccati equation.
Remark 3: As discussed in Remark 1 of [15], the computational where tk = 1
k , with 1 = 01
: , and N = 1 000
. Let x ;0 de-
burden for real-time implementation can be reduced by eliminating note the sample mean square error of the trivial estimate x0i t  . ^() 0
the redundancies in the extended state vector X  . Recall that the
( + 1) The MSE x ; averaged over 5 000 simulation runs are reported in
n i0
Kronecker power x[i] has ni components, of which only Table I, where the improvement obtained by increasing the index  can
monomials are independent. It follows that X  has dimension
i
be recognized. Note that in this example the filter with  =1
(i.e., the

+
n n2 + ( + 1) ( + ) 1
1 1 1 n , but only i=1 n ii 0 (i.e., n n  0 )
Kalman-Bucy filter based on the linear approximation of the process
=0
around x ), provides worse performances than the trivial estimate
components are independent. This allows to define a reduced-extended ^() 0
x0i t  . It is interesting to note that in this example a considerable
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007 2171

V. CONCLUSION
The problem of state estimation for stochastic nonlinear differen-
tial systems using bilinearized system approximation has been investi-
gated in this paper. The filtering algorithm here proposed is based on
two steps: first the nonlinear system is approximated by using a Car-
leman-like bilinearization approach, taking into account all the powers
of the polynomial approximation of the drift and output functions, up
to a chosen degree  ; next, the equations of the optimal linear filter
for the approximating system are computed. This step is based on the
paper [9], concerning suboptimal state estimate of stochastic bilinear
systems. When the index  = 1, the proposed algorithm gives back the
classical Extended Kalman-Bucy Filter, i.e., the Kalman-Bucy filter ap-
plied to the linear approximation of the original differential system.

APPENDIX
The proofs here reported exploits the two following properties of the
rx operator:
Fig. 1. True and estimated states: the first component. rx
(f (x)
g(x)) = (rx
f )
g(x)+ Cp;q
T
((rx
g )
f (x))
(61)
7! IRp and g : IRn 7! IRq , and
for any differentiable f : IRn

8M 2 IRr2s ; 8N 2 C 1 (IRn; IRs2c );


rx
(M N (x)) = M (rx
N (x)): (62)
Proof of Lemma 1:
Proof: By finite induction. For h = 1 it is easy to compute rx

x = In , so that (9) is true for h = 1. Assume that (9) is true for some
h 0 1, with h > 1, i.e.

rx
x[h01] = Unh01 In
x[h02] : (63)
Then

rx
x[h] = rx x
x[h01]
[h01]
= In
x
T
+ Cn ;n rx
x[h01]
x
[h01]
= In
x
T
+ Cn ;n Unh01 In
x[h02]
x
[h01] h01
Fig. 2. True and estimated states: the second component. = In
x
T
+ Cn ;n Un
In In
x[h01]
= In + CnT ;n Unh01
In In
x[h01] : (64)
TABLE II
MEAN SQUARE ESTIMATION ERRORS FOR THE FILTER (54)–(55) This proves (9), taking into account the definition of Unh in (11). The
proof of (10) is obtained by expanding rx
x[h] , for h  2, as follows
[2]

r[2]
x
x
[h]

= rx
rx
x
[h]

[h01]
= r x
U n In
x
h
improvement is obtained when the second-order Carleman approxima-
[h01]
= U n r x
In
x
tion is used for the filter derivation ( = 2), while the third-order Car- h
leman approximation ( = 3) does not provide a significant improve-
ment. The CPU times for the execution of a single run on a laptop with
h L
= Un Cn ;n rx
x[h01]
In
Unh01 In
x[h01]
In
2 GHz clock are h L
= Un Cn ;n (65)
T=1 = 0:079 s; T=2 = 0:093 s; T=3 = 0:22 s: (60)
r[2]
x
x
[h]

Figs. 1 and 2 report the plots of the output of the last of the 5 000
h L
= Un Cn ;n Unh01 CnL ;n x[h01]
In
In

Unh01 CnL ;n
In x[h01]
In
simulation runs in the subinterval [0; 10]. Note that, due to the high h L
= Un Cn ;n
measurement noise, in some instants the estimates are quite far form
the true state. The mean square errors of the estimates produced by the h L
= Un Cn ;n Unh01 CnL ;n
In
filter (54)–(55), based on the system bilinearization around the current
estimate x 
^ (t), are reported in Table II. The performances are some- 2 CnT ;n In
x[h01] = Onh In
x[h01] : (66)
what better, especially for  = 1, but not dramatically different from
those produced by the filter (51)–(53).
2172 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007

In the derivation of the Carleman approximation of system (1) the Itô [10] C. K. Chui and G. Chen, Kalman Filtering with Real-Time Applica-
formula for the computation of stochastic differentials is needed (see tions. New York: Springer-Verlag, 1987.
[24]). Consider the stochastic process [11] G. B. Di Masi, M. Pratelli, and W. J. Runggaldier, “An approximation
for the nonlinear filtering problem, with error bound,” Stochastics, vol.
14, pp. 247–271, 1985.
p
[12] R. J. Elliot and R. Glowinski, “Approximations to solutions of the Zakai
filtering equation,” Stoch. Anal. Appl., vol. 7, no. 2, pp. 145–168, 1989.
dxt = f (xt )dt + g (xt )dWt = f (xt )dt + g:;k (xt )dWk;t
[13] A. Gelb, Applied Optimal Estimation. Cambridge, MA: MIT Press,
k=1 1984.
(67) [14] A. Germani, C. Manes, and P. Palumbo, “Filtering of differential
nonlinear systems via a Carleman approximation approach,” in 44th
where xt 2 IRn and Wt is a standard Wiener process in IRp . The IEEE Conf. on Decision and Control & European Control Conf.
(CDC-ECC05), Seville, Spain, 2005, pp. 5917–5922.
symbol g:;k (xt ) denotes the k -th column of matrix g (xt ). Consider a [15] A. Germani, C. Manes, and P. Palumbo, “Polynomial extended Kalman
transformation zt = r(t; x), where r(1; 1) is a scalar function, twice filter,” IEEE Trans. Automat. Contr., vol. 50, no. 12, pp. 2059–2064,
differentiable. The differential dzt , computed according to the Itô for- 2005.
mula, is as follows: [16] A. Germani and M. Piccioni, “Finite dimensional approximation for the
equations of nonlinear filtering derived in mild form,” Applied Math.
@r @r and Optim., vol. 16, pp. 51–72, 1987.
dzt = dt + dxt [17] D. J. Higham, “An algorithmic introduction to numerical simulation
@t @x
(t;x ) (t;x ) of stochastic differential equations,” SIAM Review, vol. 43, no. 3, pp.
1 @ r
2 525–546, 2001.
T
+ gi;: gj;: dt (68) [18] K. Ito, “Approximation of the Zakai equation for nonlinear filtering
2 @ xi @ xj theory,” SIAM J. Contr. Optim., vol. 34, no. 2, pp. 620–634, 1996.
i;j
(t;x )
[19] K. Ito and K. Xiong, “Gaussian filters for nonlinear filtering problems,”
IEEE Trans. Automat. Contr., vol. 45, pp. 910–927, 2000.
or, equivalently [20] A. H. Jazwinski, Stochastic Processes and Filtering Theory. New
2 York: Academic, 1970.
@r @r 1 @ r T
dzt = dt + dxt + tr (gg ) dt: [21] S. J. Julier and J. K. Uhlmann, “Unscented filtering and nonlinear esti-
@t @x 2 @x
2
(t;x ) (t;x ) (t;x ) mation,” Proc. of IEEE, vol. 92, pp. 401–422, 2004.
(69) [22] K. Kowalski and W. H. Steeb, Nonlinear Dynamical Systems and Car-
leman Linearization. Singapore: World Scientific, 1991.
[23] H. J. Kushner, “Nonlinear filtering: The exact dynamical equations sat-
Using the Kronecker formalism the Itô differential can be written as isfied by the conditional models,” IEEE Trans. Automat. Contr., vol. 12,
no. 3, pp. 262–267, 1967.
@r [24] R. S. Liptser and A. N. Shiryayev, Statistics of Random Processes I
dzt = + (rx
r ) f j(t;x ) and II. Berlin: Springer, 1977.
@t
(t;x ) [25] E. Pardoux, “Stochastic partial differential equations and filtering of
1
diffusion processes,” Stochastics, vol. 3, pp. 127–167, 1979.
[2] 2
+ rx
r ~
g dt [26] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tu-
2 (t;x ) torial on particle filters for online nonlinear/non-Gaussian bayesian
+ ( rx
r )gj(t;x ) dWt (70) tracking,” IEEE Trans. Sign. Proc., vol. 50, pp. 174–188, 2002.
[27] S. Sastry, Nonlinear Systems: Analysis, Stability and Control. New-
York: Springer-Verlag, 1999.
where [28] S. Svoronos, G. Stephanopoulos, and R. Aris, “Bilinear approximation
p
of general nonlinear dynamic systems with linear inputs,” Int. J. Contr.,
2 [2] vol. 31, no. 1, pp. 109–126, 1980.
~
g = g:;k (71) [29] J. T.-H. Lo, “Global bilinearization of systems with control appearing
k=1 linearly,” SIAM J. Control, vol. 13, no. 4, pp. 879–885, 1975.
[30] W. S. Wong and S. T. Yau, “The estimation algebra of nonlinear fil-
tering systems,” in Mathematical Control Theory, J. Baillieul and J. C.
Willems, Eds. : Springer Verlag, 1999, pp. 33–65.
REFERENCES [31] N. van de Wouw, H. Nijmeijer, and D. H. van Campen, “Statistical bi-
[1] B. D. O. Anderson and J. B. Moore, Optimal Filtering. Englewood linearization in stochastic nonlinear dynamics,” in Proc. of the 2nd Int.
Cliffs, NJ: Prentice-Hall, 1979. Conf. on Control of Oscillations and Chaos, 2000, vol. 3, pp. 394–399.
[2] Z. Bai, “Krylov subspace techniques for reduced-order modeling of [32] M. Zakai, “On the optimal filtering of diffusion processes,” Z.
large-scale dynamical systems,” Applied Numerical Mathematics, vol. Wahrschein. Verw. Geb., vol. 11, pp. 230–243, 1969.
43, pp. 9–44, 2002.
[3] C. Bruni, G. Di Pillo, and G. Koch, “Bilinear systems: An appealing
class of nearly linear systems in theory and applications,” IEEE Trans.
Automat. Contr., vol. 19, no. 4, pp. 334–348, 1974.
[4] R. Bucy and P. Joseph, Filtering for Stochastic Processes with Appli-
cations to Guidance. New York: Wiley, 1968.
[5] R. Bucy and E. Jonckheere, “Singular filtering problems,” Syst. Contr.
Lett., no. 13, pp. 339–344, 1989.
[6] F. Carravetta, A. Germani, and M. Raimondi, “Polynomial filtering for
linear discrete-time non-Gaussian systems,” SIAM J. Contr. Optim.,
vol. 34, no. 5, pp. 1666–1690, 1996.
[7] F. Carravetta, A. Germani, and M. Raimondi, “Polynomial filtering of
discrete-time stochastic linear systems with multiplicative state noise,”
IEEE Trans. Automat. Contr., vol. 42, pp. 1106–1126, 1997.
[8] F. Carravetta, A. Germani, and C. Manes, “Kalman-Bucy filtering for
singular output-noise covariance,” in 7th IEEE Med. Conf. on Control
and Aut. (MED’99), Haifa, Israel, 1999, pp. 1320–1324.
[9] F. Carravetta, A. Germani, and M. K. Shuakayev, “A new suboptimal
approach to the filtering problem for bilinear stochastic differential sys-
tems,” SIAM J. Contr. Optim., vol. 38, no. 4, pp. 1171–1203, 2000.

You might also like