Filtering and smoothing of state vector for di use state space models

S. J. Koopman Department of Econometrics, Free University Amsterdam, de Boelelaan 1105, 1081 HV Amsterdam, Netherlands. J. Durbin Department of Statistics, London School of Economics and Political Science, Houghton Street, London WC2A 2AE, UK.
Abstract
This paper presents exact recursions for calculating the mean and mean square error matrix of the state vector given the observations for the multivariate linear Gaussian state space model in the case where the initial state vector is (partially) di use.
Keywords: Di use initialisation Kalman lter Smoothing State space.

1 Introduction
In this paper we develop e cient recursions for ltering and smoothing of the state vector for the multivariate linear Gaussian state space time series model given by

"t N(0 Ht) (1) t = 1 ::: n t t N(0 Qt ) where yt is the N 1 vector of observations and t is the m 1 unobserved state vector, in a situation where the initial state vector 1 is, at least partially, di use. The matrices Zt Tt Rt Ht and Qt are assumed to be known. The N 1 disturbance vector "t and the r 1 disturbance vector t are assumed to be serially independent and independent of each other at all time points. Similar results to those in this paper can also be derived for the more general model of de Jong (1991).
1

yt = Zt t + "t t+1 = Tt t + Rt

The initial state vector 1 is speci ed as

1 =a+A

+ R0 0

N(0 Q0)

(2)

where is a q 1 vector of unknown quantities, the m 1 vector a is known, the m q matrix A and the m (m ; q) matrix R0 are selection matrices, that is, they consist of columns of the identity matrix Im they are de ned so that when taken together, their columns constitute all the columns of Im so A R0 = 0 and = A 1. The matrix Q0 is assumed to be positive de nite and known. In most cases vector a will be treated as a zero vector unless some elements of the initial state vector are known constants. The vector is random and we assume that
0 0

N(0 Iq )

(3)

where > 0 is given. Therefore, the initial conditions for the state vector become E( 1) = a and Var( 1) = P where

P = P +P
1 0 0 1 1

(4)

and where we let ! 1 where appropriate. Here P = AA and P = R0 Q0R0 given the matrix structure of A, it follows that P is an m m diagonal matrix with q elements equal to one and the other elements equal to zero. Without loss of generality, when a diagonal element of P is non-zero we take the corresponding element of a to be zero. A vector with distribution N(0 Iq ) as ! 1 is said to be di use. The Kalman lter is a forward recursion for calculating at+1 = E( t+1jYt) with Yt = fy1 : : : ytg together with the mean square error matrix Pt+1 = Var( t+1 jYt) for t = 1 : : : n. The disturbance smoother is a backward recursion for calculating ^t = E ("t jYn) and ^t = " E ( t jYn), together with their corresponding mean square error matrices while the state smoother is a backward recursion for calculating ^ t = E ( t jYn) together with Vt = Var ( t jYn) for t = n : : : 1. Well-known Kalman lter and smoothing recursions are available for model (1) and (2) with given > 0. The objective in this paper is to derive limiting versions of these recursions when ! 1. A simple technique for obtaining an approximation to the exact limiting results is to replace in (4) by a large value such as 107 and use the standard recursions see Harvey and Phillips (1979). However, the numerical solution obtained in this way will contain numerical inaccuracies and is theoretically unsatisfactory. Two main approaches have appeared in the literature on the exact treatment of a di use initial vector for ltering and smoothing:
1

An exact solution as ! 1 was rst considered in detail by Ansley and Kohn (1985). They provided intricate equations for state ltering and smoothing. Ansley and Kohn (1990) presented a more simple treatment for mainly univariate models. The solution consists of modi cations to the Kalman lter which remain until the in uence of has disappeared from the updating equations. The transition to the usual Kalman lter at 2

some time point t = d is automatic there is no choice of doing a transition or not. The usual state smoothing recursions can be applied for t = n : : : d + 1. The modi cations for state smoothing are only required for the initial period t = d : : : 1. An alternative and simpler approach for di use likelihood evaluation and state ltering and smoothing was given by de Jong (1991). This solution requires augmenting the vector recursions of the Kalman lter and smoother to matrix recursions and adding an extra matrix recursion. Regression calculations applied to the auxiliary part of the Kalman lter and smoother provide the solution. A transition to the usual Kalman lter after some time point t = d is optional but it is normally advisable to collapse at the earliest time possible. However, as is the case for the Ansley and Kohn solution, modi cations for initial state smoothing are required when a collapse has taken place. Details of these modi cations are given by Chu-Chun-Lin and de Jong (1993). Other contributions on exact treatments of di use initialisation have appeared in the literature but these do not speci cally deal with the problem of both ltering and smoothing of the state vector. Bell and Hillmer (1991) consider likelihood evaluation of ARIMA models Snyder and Saligari (1996) present a strategy for initialising the Kalman lter using a square root formulation but their technique does not provide a solution for initial state smoothing (t = d : : : 1) Koopman (1997) presents a more transparent treatment of di use ltering and likelihood evaluation using the Ansley and Kohn approach in which the collapse to the Kalman lter is automatic. In this paper we provide a solution to di use state smoothing using the approach of Koopman (1997). The paper is organised as follows. Sections 2 and 3 provide exact initial recursions for di use state ltering and smoothing, respectively. Section 4 presents useful matrix formulations of the ltering and smoothing recursions. Proofs are given in Appendices.

2 State ltering
2.1 Non-di use state ltering
The Kalman lter for model (1) for given > 0 is

at+1 = Tt at + Ktvt
where

Pt+1 = Tt PtLt + Rt Qt Rt
0 0

t = 1 ::: n

(5)

vt = yt ; Ztat Kt = Tt PtZt Ft 1
0 ;

Ft = ZtPtZt + Ht Lt = Tt ; Kt Zt
0

(6)

with a1 = a and P1 = P . The proof can be found in Anderson and Moore (1979, Chapter 3) and Durbin and Koopman (2001, x4.2.1). 3

2.2 Di use state ltering

The mean square error matrix Pt is decomposed in a similar way to matrix P in (4), that is Pt = P t + P t + O( 1) t = 1 ::: n (7) where P t and P t do not depend on . It is shown by Ansley and Kohn (1985) and Koopman (1997) that the in uence of the term P t will disappear after a limited number of updates d of the exact initial Kalman lter. Therefore, the state ltering equations (5) apply without change for t = d + 1 : : : n when ! 1. Note that when all state elements follow stationary processes or are known, P = 0 and d = 0. Our approach relies on the expansion of the inverse of matrix Ft = F t + F t + O( 1 ) t = 1 ::: n which appears in the de nition of Kt in (6). The expression of Ft is analogous to the decomposition of (7). The solution that we shall give for both ltering and smoothing apply to all univariate series. For multivariate series we give explicit solutions for the special cases F t nonsingular and F t = 0. These two cases apply to nearly all time series occurring in practice. Although an explicit solution can be given as in Koopman (1997), we have found that for the rare cases where F t is a nonzero singular matrix, it is more e cient to convert the multivariate series to a univariate series in the way described in Koopman and Durbin (2000) and to apply the recursions of this paper to the resulting univariate series. In Appendix A we show that for t = 1 : : : d the exact initial state ltering equations when F t is nonsingular are given by at+1 = Tt at + K tvt (8) P t+1 = Tt P tL t P t+1 = Tt P tL t ; K tF tK t + Rt QtRt where vt = yt ; Zt at F t = Zt P tZt F t = ZtP tZt + Ht K t = Tt P tZt F 1t L t = Tt ; K tZt K t = (TtP t Zt ; K tF t)F 1t with the initialisations a1 = a, P 1 = P and P 1 = P . The recursions (8) are referred to as the exact initial Kalman lter following Koopman (1997). In the case F t = 0, we have at+1 = Tt at + K t vt (9) P t+1 = Tt P tTt P t+1 = Tt P tL t + RtQt Rt where vt = yt ; Ztat F t = ZtP t Zt + Ht (10) K t = TtP t Zt F t1 L t = Tt ; K tZt for t = 1 : : : d.
; 1 1 1 1 ; 1 1 1 1 1 1 0 1 1 1 0 0 0 1 1 1 0 0 1 1 0 ; 0 ; 1 1 1 1 1 1 1 1 1 1 0 1 1 0 0 0 0 ;

3 State smoothing
3.1 Non-di use state smoothing
For given , the conditional mean of the state t given all the observations Yn, that is ^ t = E ( t jYn), together with its mean square error matrix Vt can be evaluated recursively by wellknown smoothing recursions provided by Anderson and Moore (1979). A computationally more e cient set of recursions is developed by de Jong (1988), de Jong (1989) and Kohn and Ansley (1989). For model (1) and for given > 0 these are

rt 1 = Zt Ft 1vt + Lt rt Nt 1 = Zt Ft 1Zt + Lt Nt Lt (11) ^ t = at + Ptrt 1 Vt = Pt ; PtNt 1 Pt t = n n; 1 ::: 1 with rn = 0 and Nn = 0. Storage space is required for the quantities vt , Ft , Kt , at and Pt, which are calculated by the Kalman lter, for t = 1 : : : n.
0 ; 0 0 ; 0 ; ; ; ;

The state smoothing recursions (11) apply to the period t = n n ; 1 : : : d + 1 for which contribution to Pt of terms depending on vanishes as ! 1. We show in Appendix B that for the initial time period t = d d ; 1 : : : 1, the exact initial state smoothing equations when F t is nonsingular are given by
1

3.2 Di use state smoothing

^ t = at + P trt(0)1 + P trt(1)1 (12) Vt = P t ; P tNt(0)1 P t ; < P tNt(1)1 P t > ;P t Nt(2)1 P t where for any square matrix A, < A >= A + A , together with the backwards recursions
; 1 ; ; 1 ; 1 ; 1 0

rt(0)1 = L t rt(0) Nt(0)1 = L t Nt(0) L

0 ; 1 0 ; 1

rt(1)1 = Zt (F 1tvt ; K t rt(0) ) + L trt(1) Nt(1)1 = Zt F 1tZt + L tNt(1) L t ; < L tNt(0) K t Zt > (13) Nt(2)1 = Zt F# tZt + L tNt(2) L t; < L t Nt(1) K t Zt >
0 ; 0 0 ; 1 1 0 0 ; 0 0 ; ; 1 1 1 1 0 0 1 1 1

where

F# t = K t Nt(0) K t ; F 1tF t F
0 ; 1 1

(0) (1) and with initialisations rd = rd , rd = 0 and Nd(0) = Nd, Nd(1) = Nd(2) = 0. In the case F t = 0, we have

rt(1)1 = Nt(1)1 = t Nt(2)1 = Storage space is required for the quantities vt, F t , F t, K t, K calculated by the exact initial Kalman lter.
0 ; 0 ; ; 0 ; 0 ; ; ; 1

rt(0)1 = Zt F t1vt + L t rt(0) Nt(0)1 = Zt F t1Zt + L t Nt(0) L

Tt rt(1) Tt Nt(1) L t Tt Nt(2) Tt : t , at , P t and P

0 0 0

(14)
1

which are

4 Matrix formulations
A convenient representation of the exact initial state ltering equations for the case F gular is given by
1

t nonsin-

at+1 = Tt at + K tvt
1

Pt+1 = Tt Pt Lt + (Rt Qt Rt 0)
y y y0 0 y y 1 y y

t = 1 ::: d
#

(15)

with the initialisation a1 = a and P1 = P = (P P ). Elements of a corresponding to di use elements of 1 can be put equal to zero. The matrices Pt and Lt are de ned as

Pt = (P t P t)
y 1

Lt = L
y

"

t ;K t Zt : 0 L t
1

(16)

It is noteworthy that, apart from rede nitions of matrices, the initial recursions (15) have the same structure as the standard recursions (5). A similar representation to (15) for the case F t = 0 can be obtained by replacing K t in (15) by K t from (10) and the matrix Lt by
1 1 y

Lt = L t 0 : 0 Tt
De ne
y

"

Zt = Zt 0 0 Zt

"

vt =
y

vt 0
!

Ft =
y

"

0
;

F 1t ;F 1tF t F
; 1 ; 1

#
;

(0) (0) (1) rt = rt(1) Nt = Nt(1) Nt(2) : rt Nt Nt A convenient representation of the exact initial state smoothing equations for the case F nonsingular is given by
y y

"

rt 1 = Zt Ft vt + Lt rt ^ t = at + Pt rt 1
y y0 y y y0 ; y y ; y 0 0 y

Nt 1 = Zt Ft Zt + Lt Nt Lt Vt = P t ; Pt Nt 1 Pt t = d ::: 1
y y0 y y y0 y y ; y y y0 ; y y

(17)

with rd = (rd 0) and Nd = diag (Nd 0), where the partioned matrices Lt and Pt are as in (16). As with the lter (15), these recursions have the same structure as the standard recursions (11) apart from rede nition of matrices. These results are obtained after some minor manipulation of the equations in xx2.2 and 3.2. When F t = 0, the smoothing equations (17) apply with Ft and Lt replaced by
y y 1

1 Ft = F t 0 0 0
;

"

Lt = L t 0 0 Tt
6

"

respectively.

Acknowledgements
The rst author was working at the Center for Economic Research of Tilburg University, The Netherlands as a Research Fellow of the Royal Netherlands Academy of Arts and Sciences during the main part of this project. The nancial support of the Academy is gratefully acknowledged.

Appendix A Di use ltering

In this Appendix we derive the di use ltering results. The decomposition (7) leads to the similar decompositions for Ft = ZtPt Zt + Ht and Mt = Tt Pt Zt , that is
0 0

Ft = F
where

t + F t + O(

1)

Mt = M

t + M t + O(

1)

(18)

F M

t t

= ZtP tZt = Tt P tZt

1 0 1 ; 1

F t = Zt P tZt + Ht M t = Tt P tZt
0 0 1 ; ; ;

(19)

for t = 1 : : : d. The derivation of the exact initial Kalman lter for the case F t nonsingular is based on the expansion for Ft 1 = F t + F t + O( 1)] 1 as a power series in 1, that is

Ft 1 = Ft(0) +
;

1 F (1) +
t

2 F (2) + O ;
t

(20)
2 F (2) + : : : ):
t

for large . Since Ip = Ft Ft 1 we have

;

Ip = ( F

t+F t+ j

1F

at +

2F

b t + : : : )(Ft

(0) +

1 F (1) +
t

On equating coe cients of

for j = 0 ;1 ;2 : : : we obtain
1 1

F tFt(0) = 0 F tFt(0) + F tFt(1) = Ip Fa tFt(0) + F tFt(1) + F tFt(2) = 0 etc.

1 1

(21)

We need to solve equations (21) for Ft(0) , Ft(1) and Ft(2) further terms will not be required. For F t nonsingular, we have from (21),

Ft(0) = 0
; ;

Ft(1) = F

Ft(2) = ;F 1t F tF 1t:
; ; 1 1 ;

(22)

The matrix Kt = Mt Ft 1 depends on the inverse matrix Ft 1 so it also can be expressed as a power series in 1 . We have

Kt = M t + M t + O( 1)]( = K t + 1 K t + O( 2 )
; 1 ; ; 1

1 F (1) +
t

2 F (2) + : : : )
t

where

t t

= = = = =

M tFt(1) M tF 1t M t Ft(1) + M tFt(2) M t F 1t ; M tF 1tF tF (M t ; K tF t )F 1t:

1 ; 1 1 1 ; ; 1 1 1 ; 1 1

(23)

The updating equation (5) for at+1 can now be expressed as

at+1 = Ttat + Kt vt = Ttat + K t +

1

1K

t+O

; 2 ]vt
;

which becomes as ! 1,

at+1 = Tt at + K tvt :
1

(24)

The updating equation (5) for Pt+1 = P t+1 + P t+1 is

1

Pt+1 = Tt Pt Lt + Rt Qt Rt = Tt P t + P t + O (
0 0 1 1

1 )]

Tt ; K

t+

1K

t+O

; 2
;

Zt + Rt Qt Rt :
0 0

Consequently, the updates for P t+1 and P t+1 as ! 1 are given by

P P

= Tt P t(Tt ; K tZt) t+1 = Tt P t (Tt ; K t Zt ) ; Tt P t Zt K t + Rt Qt Rt :

t+1
0 1 1 0 0 0 0 1 1 ; ;

(25)

The matrix Pt+1 also depends on terms in 1 , 2, etc. but these terms will not be multiplied by or higher powers of within the Kalman lter recursions. Thus the updating equations for Pt+1 do not need to take account of these terms when ! 1. For the case F t = 0, we have P tZt = 0, M t = 0 and
1 1 1

Ft 1 = F t + O(
;

1 )] 1 = F 1 +
; ;

1F

a t + O(

2 ): 2 )]

It follows that

Kt = M t + 1Ma t + O( 2)] F t1 + = K t + 1 Ka t + O( 2)
; ; ; ; ;

1 Fa t + O(

where

K t = M t F t1 Ka t = M t Fa t + Ma t F t1:
; ;

(26)

We will show below that no explicit de nitions for Fa t , Ma t and Ka t are required. The ltering equations are given by

at+1 = Tt at + K t + 1Ka t + O( 2)]vt Pt+1 = Tt P t + P t + O( 1)]fTt ; K t +

; ; ; 1

1K

a t + O(

2 )]Z g
t

+ Rt Qt Rt :
0

When ! 1, we obtain

at+1 = Tt at + K tvt P t+1 = Tt P tTt P t+1 = Tt P t(Tt ; K tZt ) + Rt Qt Rt

0 1 1 0 0

(27)

since P tZt = 0.
0 1

Appendix B Di use smoothing

B1 Derivation of smoothed mean of state
To obtain the limiting recursions for the smoothing equation ^ t = at + Ptrt 1 given in (11) for t = d : : : 1, we consider the recursion rt 1 = Zt Ft 1vt + Lt rt . Since rt 1 depends linearly on Ft 1 and Kt which can both be expressed as power series in 1 we write
; 0 ; 0 ; ; ; ;

rt 1 = rt(0)1 +
; ; 1

1 r(1) + O( 2 ) t 1
; ;

t = d : : : 1:
;

(28)

Substituting the relevant expansions into the recursion for rt 1 in (11) we have for the case F t nonsingular,

rt(0)1 +
;

1 r(1) + : : : t 1
;

= Zt

1 F (1) +
t
1

2 F (2) + : : :
t
;

+ Tt ; K leading to recursions for rt(0) and rt(1) ,

t+

1K

t+O

; 2
;

vt

Zt

rt(0) +

1 r(1) + : : :
t

rt(0)1 = L trt(0) rt(1)1 = Zt F 1tvt ; K t rt(0) + L trt(1)

0 ; 1 0 ; 0 0 ; 1 1

(29)
1

(0) (1) for t = d : : : 1 with rd = rd and rd = 0. Note that L The smoothed state vector is
;

= Tt ; K tZt .

^ t = at + Ptrt 1 ; = at + P t rt(0)1 + 1 rt(1)1 + P t rt(0)1 + 1rt(1)1 + O 1 ; = at + P trt(0)1 + P trt(0)1 + P trt(1)1 + O 1 :

; ; ; ; ; 1 ; ; ; ; 1 ; 1 ;

(30)

It is obvious that for this expression to make sense we must have P trt(0)1 = 0 for all t. Thus we need to show that Var( tjYn) is nite as ! 1. Analogously to the arguments in de Jong (1991) we can express t as a linear function of 0 1 : : : t 1 . But Var( jYd) is nite by de nition of d so Vt = Var( t jYn) must be nite as ! 1 since d < n. Also, Qj = Var( j ) is nite so Var( j jYn) is nite for j = 0 : : : t ; 1. It follows that Var( t jYn) is nite for all t as ! 1 so from (30) P trt(0)1 = 0. Letting ! 1 we obtain the expression for ^ t in (12). For the case F t = 0, the recursion for rt becomes
1 ; ; 1 ; 1

rt(0)1 +
;

1 r(1) + : : : t 1
;

; = Zt F t1 + 1Fa t + : : : vt ; + Tt ; K t + 1Ka t O 2 Zt
0 ; ; ; ;

rt(0) +

1 r(1) + : : :
t

leading to recursions for rt(0) and rt(1) ,

; ;

rt(0)1 = Zt F t1vt + L t rt(0) rt(1)1 = Zt Fa t vt ; Ka t rt(0) + L t rt(1)

0 ; 0 0 0 0 ;

(31)
1

for t = d : : : 1 with matrix L t given by (10). Since rt(1)1 is premultiplied by matrix P (12) and since P tZt = 0, the recursion for rt(1) reduces to the one given in (14).
0 1

in

To obtain exact nite expressions for Vt and Nt 1 in (11) when F t is nonsingular and ! 1, for t = d : : : 1, we need to take three-term expansions instead of the two-term expressions previously employed:
; 1

B2 Derivation of smoothed variance of state

Nt = Nt(0) + 1Nt(1) + 2Nt(2) + O( 3): (32) Ignoring residual terms and on substituting in the expression Nt 1 = Zt Ft 1Zt + Lt Nt Lt , we obtain the recursion for Nt 1 as Nt(0)1 + 1 Nt(1)1 + 2Nt(2)1 + : : : = Zt 1Ft(1) + 2 Ft(2) + : : : Zt ; + Tt ; K t + 1K t + 2K# t + O 3 Zt Nt(0) + 1 Nt(1) + 2Nt(2) + : : : ; Tt ; K t + 1K t + 2K# t + O 3 Zt for the case F t is nonsingular. We will argue below that no explicit de nition for K# t is needed. This leads to the set of recursions Nt(0)1 = L tNt(0) L t Nt(1)1 = Zt F 1tZt + L tNt(1) L t; < Zt K tNt(0) L t > (33) Nt(2)1 = ;Zt F tF 1tF t Zt + L tNt(2) L t; < Zt K tNt(1) L t > (0) L > +Z K N (0) K Z ; < Zt K# tNt t t t t t t
; ; ; 0 ; 0 ; ; ; ; ; ; ; 0 ; ; ; ; ; 0 ; ; 1 ; ; ; 1 1 0 ; 1 0 1 ; 0 0 0 ; 1 1 1 0 1 0 0 ; 0 ; 1 0 1 1 1 0 0 0 1

10

with Nd(0) = Nd and Nd(1) = Nd(2) = 0. Note that < A >= A + A for any square matrix A. Substituting the power series in 1, 2 , etc. and the expression Pt = P t + P t into the relation Vt = Pt ; Pt Nt 1 Pt we obtain
0 ; ; 1 ;

Vt = P t + P t ; (P t + P t) Nt(0)1 + 1Nt(1)1 + 2Nt(2)1 + : : : (P t + P t) = ; 2 P tNt(0)1 P t + P t ; P tNt(0)1 P t ; P tNt(0)1 P t ; P tNt(1)1 P t +P t ; P t Nt(0)1 P t ; P tNt(1)1 P t ; P tNt(1)1 P t ; P tNt(2)1 P t ; +O 1 :
1 ; ; 1 ; ; ; 1 1 ; 1 1 1 ; ; 1 1 ; 1 ; ; 1 1 ; 1 ; 1 ;

(34)

It was argued in the previous section that Vt = Var( tjYn) is nite for t = 1 : : : n. Thus the two matrix terms associated with and 2 in (34) must be zero. Letting ! 1, the smoothed state variance matrix is given by

Vt = P t ; P tNt(0)1 P t; < P tNt(1)1 P

; ;

> ;P tNt(2)1 P t:
1 ; 1 0

(35)

Since we have argued that

P tNt(0)1 P
1 ; 1 1 ;

=0
1

where Nt(0)1 = L tNt(0) L

; 1

it follows that P tL t Nt(0) = 0, when F as if the recursion for Nt(2)1 in (33) is

0 ; 0 ; 1 1

is nonsingular. Therefore, we can proceed in e ect

0 0 0 1 1

Nt(2)1 = ;Zt F tF 1tF t Zt + L t Nt(2) L t ; < Zt K tNt(1) L

1

> +Zt K t Nt(0) K tZt :

0 0

It is convenient that K# t drops out from our calculations for Nt(2) . For the case F t = 0, the recursion for Nt becomes

Nt(0)1 + 1 Nt(1)1 + 2Nt(2)1 + : : : ; = Zt F t1 + 1Fa t + : : : Zt + Tt ; K t + 1Ka t + 2Kb t + O( Tt ; K t + 1 Ka t + 2Kb t + O(

; ; ; ; ; 0 ; ; ; ; ; ;

3)

Zt 3) Z t

Nt(0) +

1 N (1) +
t

2 N (2) + : : :
t

leading to the recursions

Nt(0)1 = Zt F t1Zt + L t Nt(0) L t Nt(1)1 = Zt Fa t Zt + L tNt(1) L t ; < Zt Ka t Nt(0) L t > Nt(2)1 = Zt Fb t Zt + L t Nt(2) L t ; < Zt Ka tNt(1) L t > ; < Zt Kb tNt(0) L t > +Zt Ka t Nt(0) Ka t Zt
0 ; 0 ; 0 0 0 0 ; 0 0 0 0 ; 0 0 0 0

(36)

11

for t = d : : : 1 with matrix L t given by (10). Similar to the arguments below (35), we have and P tNt(0)1 P t = P tL t Nt(0) L t P t = 0 since P tZt = 0. Further, matrix Nt(1)1 is pre- or post-multiplied by matrix P t in (12) and since P tZt = 0 and P tL t Nt(0) = 0, the recursion for Nt(1) reduces to the one given in (14). Similar reductions are obtained for the recursion for Nt(2) because it is pre- and post-multiplied by matrix P t in (12). We do not need to give explicit de nitions for the matrices Fa t , Fb t, Ka t and Kb t due to these reductions.
1 ; 1

P tNt(0)1 P

=0
0

References
Anderson, B. D. O. and J. B. Moore (1979). Optimal Filtering. Englewood Cli s: Prentice-Hall. Ansley, C. F. and R. Kohn (1985). Estimation, ltering and smoothing in state space models with incompletely speci ed initial conditions. Annals of Statistics 13, 1286{1316. Ansley, C. F. and R. Kohn (1990). Filtering and smoothing in state space models with partially di use initial conditions. J. Time Series Analysis 11, 275{93. Bell, W. and S. Hillmer (1991). Initializing the Kalman lter for nonstationary time series models. J. Time Series Analysis 12, 283{300. Chu-Chun-Lin, S. and P. de Jong (1993). A note on fast smoothing. Discussion paper, University of British Columbia. de Jong, P. (1988). A cross validation lter for time series models. Biometrika 75, 594{600. de Jong, P. (1989). Smoothing and interpolation with the state space model. J. American Statistical Association 84, 1085{8. de Jong, P. (1991). The di use Kalman lter. Annals of Statistics 19, 1073{83. Durbin, J. and S. J. Koopman (2001). Time Series Analysis by State Space Methods. Oxford: Oxford University Press. Harvey, A. C. and G. D. A. Phillips (1979). The estimation of regression models with autoregressivemoving average disturbances. Biometrika 66, 49{58. Kohn, R. and C. F. Ansley (1989). A fast algorithm for signal extraction, in uence and crossvalidation. Biometrika 76, 65{79. Koopman, S. J. (1997). Exact initial Kalman ltering and smoothing for non-stationary time series models. J. American Statistical Association 92, 1630{8. Koopman, S. J. and J. Durbin (2000). Fast ltering and smoothing for multivariate state space models. J. Time Series Analysis 21, 281{96. Snyder, R. D. and G. R. Saligari (1996). Initialization of the Kalman lter with partially di use initial conditions. J. Time Series Analysis 17, 409{24.

12