NOT FOR QUOTATION WITHOUT PERMISSION OF THE AUTHOR
CONTINUOUS TIMI3 AD- F'ILWZRING
k Yashin
December 1984 WP-84-105
Working h p e r s a r e interim reports on work of t h e international lnstitute for Applied Systems Analysis and have received only limited review. Views or opinions expressed herein do not necessarily r e p r e s e n t those of t h e Institute or of i t s National Member Organizations.
INTERNATIONAL
INSTITUTE
FOR APPLIED SYSTEMS ANALYSIS 2361 Laxenburg, AustriaThis paper deals with t h e problem of adaptive estimation in t h e continuous time Kalrnan filtration scheme. The necessary and sufficient conditions of t h e convergence of t h e p a r a m e t e r estimators are discussed. For systems which a r e characterized by constant but unknown parameters. t h e conditions of convergence can be checked before t h e observation start. The method of proof is based on the relations between singularity property of some probability measures a n d convergence of t h e Bayesian estimation algorithm.
CONTINUOUS ADAPTIVE
FIL7TRlNGk Yashin
1. INTRODUCTION
In m a n y cases, t h e coordinates of partially observed plants subjected t o random noise c a n now be successfully e s t i m a t e d (filtered) [1,2,3,4.5], by using good a @ o r i d a t a on t h e s t r u c t u r e a n d numerical values of t h e p a r a m e t e r s .
In r e a l life, however, some c h a r a c t e r i s t i c s of t h e plants m a y prove impossible t o determine before t h e experiment or observations begin. This leads t o c e r t a i n difficul- ties i n solving I!le filtering problem. U n c e r t a i n t y in t h e coefficients of t h e equation t h a t describe t h e plant subjected t o r a n d o m disturbances, m a y lead t o substantial performance dekerioration of coordinate estimating algorithms adjusted t o c e r t a i n fixed values of t h e p a r a m e t e r s [I].
Fortunately, t h e values of unknown p a r a m e t e r s c a n be updated reasonably often with t h e arrival of observed data. However, simultaneous estimation of both t h e p a r a m e t e r s and plant coordinates may prove t o be t h e nonlinear problem. Lack of
appropriate computer methods for t h e solution of nonlinear problems brings to life a host of heuristic estimation algorithms, which work in many actual situations but need more thorough analytical investigation [2.3.4].
The temptation to apply t h e available coordinate estimation algorithms (which are only effective if t h e parameter values are known) t o t h e solution of these non- linear problems leads t o adaptive filtering, whereby t h e filter equations use c u r r e n t parameter estimates which a r e obtained from processing of observations. The equa- tions or algorithms which lead t o parameter estimates a r e referred t o as adaptation, or adjustment, algorithms [5,6,7,8].
While in t h e case of available parameters the coordinate estimation algorithm is a kind of Kalman filter, with unknown parameters this a r r a n g e m e n t is referred to as adaptive Kalman filtering [9, lo].
The choice of t h e algorithm for t h e parameter estimation is somewhat arbitrary
[I].],
but one common property is frequently very important in all of these algorithms:t h e r e s u l t a n t p a r a m e t e r estimates should, in some sense, t e n d t o t h e i r t r u e values a s the number of observations grows. Such estimates a r e referred t o a s consistent. The plant with unknown p a r a m e t e r s is identifiable if t h e parameters' estimates are con- sistent.
It would be natural t o investigate t h e property of parameter estimate consistency through studying only t h e properties and characteristics of t h e initial plant dynamics equations, t h e properties of noises, and t h e specifics of t h e filtering algorithms and adaptation procedure. Such an a t t e m p t for systems described by discrete time equa- tions has been reported in [12.13,14].
The conditions sufficient for consistency of parameters' estimates, t h a t take on values from a c e r t a i n finite set, follow, in those papers, from t h e singularity of proba-
bilistic measures associated with various values of t h e parameter.
This paper provides t h e necessary and sufficient conditions for consistency of these parameter estimates in adaptive Kalman filtering for t h e case of t h e denumer- able s e t of the parameter values.
This proof relies on a close link between consistency a n d absolute continuity and singularity of a certain family of probabilistic measures. In t h e cases t o be discussed, singularity entails consistency of estimates; t h e conditions for consistency may be t h e conditions for t h e family of measures t o be singular. Recent research concerning absolute continuity a n d singularity of probabilistic measures associated with random processes [15] has made formulating t h e necessary and sufficient conditions for these properties possible. Representing some processes as solutions of stochastic dif- fererential equations permits formulating t h e singularity conditions in t e r m s of t h e characteristics of these equations. In adaptive Kalman filtering, t h e characteristics of t h e initial equations c a n be expressed as filter parameters. In this way, t h e condition for consistency of estimates can be tested in each specific case before t h e observa- tions are made. A study of consistency has been performed for Bayesian estimates in discrete time adaptive Kalman filtering in [16, 171.
2. ETATEXJZNT
OF
THE PROBLEMThe problem of p a r a m e t e r estimation in continuous time adaptive Kalman filter- ing can be investigated in t h e framework of the following formal description.
On probabilistic space
(R,H,P)
a random variable ~ ( w ) is specified which takes on values from a certain denumerable s e tjg,j,
i E N with apriori probabilitiesLet
fl
be a-algebra in Q t h a t is g e n e r a t e d by t h e values of t h e p a r a m e t e r @.
On t h e s a m e space t h e random process (6,c)
=
61,c1, t>
0 is specified. Denote asH:.(
a n dHIC
the a-algebrast h a t a r e g e n e r a t e d by values of t h e processes (zP,[) a n d
c
u p t o time t , t r 0 , respec- tively. Let~ t e ~ v ( =
H P V H:-(,t
2 0 a r e u-algebras g e n e r a t e d by t h e union of a-algebrasH P a n d H ~ Cwhile
H P ~ =
V H),t r
0 a r e u-algebras generated by t h e union of a - algebrasH@
and HI(.
t 2 0.
LetXdn( =
V tH:-(.
HPnd.(=
V HPQ+(=
H . HPB(=
vtHP-(.
1
Denote by I # , I@6.( a n d HP*( t h e respective nondecreasing right-continuous families of u-algebras.
Assume t h a t on probability s p a c e ( Q , H , P ) t h e process (19.6) c a n be r e p r e s e n t e d a s a s y s t e m of stochastic differential equations:
where ( 1 9 ~ ) ~ ~ is a sequence of k-dimensional vectors; ( [ t ) t , o is a s e q u e n c e of 1- dimensional vectors and W l a t , W Z a t a r e independent k
,
a n d k2-dimensional Wiener processes. respectively, independent of t h e initial values ofzPo,t0
a n d t h e random value of8.
The m a t r i x a ( @ , t ) is (k k ) , A ( B , t ) is (1 k ) and t h e m a t r i c e s b , ( ~ , t ) , b z ( @ , t ) . and B ( t ) a r e ( k k , ) , ( k k 2 ) , a n d ( 1 k z ) , respectively, a n d a r e t h e bounded functions of t i m e for a n y value of @.
The process 19 describes t h e t i m e variation of t h e unobservable coordinate of a c e r t a i n dynamic plant, while t h e process6
models t h e m e a s u r e m e n t of t h e coordinate 19 with random noise.Introduce on ( ~ , H P . ~ ' C ) a denumerable family of probabilistic measures
P, icN:
Assume t h a t matrix B ( t
) ~ * ( t )
is nonsingular for any t 1 0. LetPt . P
andpt
denotet h e restrictions of t h e m e a s u r e P to the u-algebras f f ! e d . C , f f C a n d
f f f ,
respectively, whileP;,
andR,
restriction of the measuresPi
toffB.'.C,
f f C andf f f .
Denote byz.j
the Radon-Nicodim derivative of t h e measure with respect t o t h e measure when i t exists.
Definition. HC-adapted p r o c e s s
B t , t
1 0 is c o n s i d e r e d t o b e a s t r o n g l y c o n s i s t e n t e s t i m a t e of t h e p a r a m e t e rB
i fIn this paper we will study t h e consistency conditions of t h e estimates
Bt
, tr
0 whereB t
=
E(BI
f f f ) .3.
RESULT
Assume t h a t the initial conditions go and
to
a r e such t h a twhere
E,
denotes mathematical expectation with respect t o t h e measureP.
From the fact t h a t the coefficients of t h e equations (1) are bounded a n d from the condition (2) i t follows t h a t with anyt <
= and i ENAssume also t h a t t h e joint distribution of t h e random variables do and
to
is Gaussian with respect t o each measureP',
EN. Hereafter b i t ) b 2 ( i , t ) . b , ( i . t ), b 2 ( i , t ) , a ( i , t ) and A ( i . t ) will replace b , ( B i , t ) , b z ( P i , t ) , a ( B i . t ) , A ( B , , ~ ) , respec-tively, for convenience of t h e notation.
The main r e s u l t of this paper is t h e following t h e o r e m .
Theorem 1 . The strong consistency property o f the estimate
pl,
t 2 0 takesplace if and onlyfor any i . j E N
where
and
einj
( t ) m e the solutions o f the following linear differential equatkns:and initial conditions
where
K ( t ) =
( b 2 ( i , t ) B e ( t )+
y,(t)A(i.t)*)(B(t )Be(t))-I
a n d yi
( t
) s a t i s f ythe
f o l l o w i n g e q u a t i o n sCorollary. A s s u m e t h a t t h e c o e f f i c i e n t s of
the
s t o c h a s t i c e q u a t i o n s ( 1 ) do n o t d e p e n d o n t i m e t , a n d a s t a t i o n a r y s o l u t i o n of e q u a t i o n (7) e z i s t s . T h e n f o r t h e e s t i - m a t eEn
t o b e c o n s i s t e n tit
is s u f f i c i e n t t h a t f o r a n y i , j EN ,
i= j
w h e r e
ai,,
is a s t a t i o n a r y s o l u t i o n of e q u a t i o n ( 1 0).The Corollary follows from equalities (4) and (5), and from t h e stationarity assumption.
Remark. h t h e c a s e of a f i n i t e n u m b e r o f possible v a l u e s of the p a r a m e t e r
B
a n d c o n s t a n t c o e f f i c i e n t s in e q u a t i o n( I ) ,
c o n d i t i o n (4) c a n be v e r i f i e d b e f o r e t h e o b s e r u a -tiom
a r e m a d e .4. PROOF
OF
THEOREX 1The proof will be preceded by several auxiliary lemmas.
kmma 1. I h e p r o c e s s ( 1 9 . f ) o n probabzlistic s p a c e s
( o M , H ~ . ~ , P )
c a n be r e p r e s e n t e d in t h e f o r n :dzPt = a ( i , t ) d t d f
+
b l ( i , t ) d W l , t + b z ( i , t ) d W ~ . ~ , $0 ( 1 3)where W , , , , W z , , are independent Wiener processes independent of d o , # o .
Proof. E r s t note t h a t for any i E
N
measuresP'
a r e absolutely continuous with respect t o m e a s u r e P a n d consequenty P,'<<
Pi for any t 2 0. Denote&'
.= - .
Let u sd Pt
consider t h e t r a n s f o r m a t i o n of t h e local c h a r a c t e r i s t i c s of t h e process $,[ with abso- lutely continuous transformation of t h e probabilistic m e a s u r e P i n t o t h e m e a s u r e
P,
i EN
on (Q,HB~*-F). From t h e definition of the m e a s u r e s P:, i EN,
t r 0 it follows t h a t for a n y t r 0&* =
l ( 8=
P i ) Pia n d consequently,
121, iEN,
t r 0 a r e independent on time. In compliance with t h e ana- log of t h e Girsanov t h e o r e m on transformation of t h e local c h a r a c t e r i s t i c s of t h e processes with absolutely c o n t i n u o u s transformation of t h e probabilistic m e a s u r e s , t h e local c h a r a c t e r i s t i c s of t h e process d , [ , in t h e case of a t i m e independent Radon- Nicodim derivative, r e m a i n u n c h a n g e d a n d consequently t h e process $,[ m a y be r e p r e s e n t e d by equation (1). Note now t h a t following t h e definitions of t h e m e a s u r e sPL,
i EN,
t h e n e x t equalities holdP
-a.s.while t h e processes W l , , , W Z . , r e t a i n t h e i r properties with r e s p e c t t o t h e m e a s u r e s
P,
i EN.
Thus on ( Q , H B . ~ . ~ ) t h e process ( 3 , ~ ) c a n be r e p r e s e n t e d by stochastic dif- ferential equation (13).Consider now t h e transformation of t h e local c h a r a c t e r i s t i c s of t h e process d , [ with t h e restriction of o-algebra H B . ~ . ~ t o
H d . 6 .
Since t h e coefficients a ( i . t ) d t ,A ( i , t ) d t , a r e Hb.4-adapted t h e innovation process ( d , ( ) which r e s u l t s from restricting
t h e o-algebras
H $ . ~ . ~ ~
t oH P . ~
, t 2 0 can be represented on( R , P . C , P ) ,
i E N in form ( 1 3 ).
I t is a well known f a c t t h a t with t h e above assumptions t h e problem of estimation of the coordinates of t h e process 19 from observations of t h e p a t h s of t h e process
#
on each space( R , P J , P )
c a n be solved in conditional Gaussian t e r m s . Denoting byt h e m e a n s q u a r e optimal e s t i m a t e of filtering for t h e process 29 and by 7 ,
( t
)= m;
((*,-
m m ;-
mf1 l
H f )t h e conditional variance of t h e e s t i m a t e i E
N,
t r 0 we have, for t h e s e variables, t h e well known equations [ l a ] :Denote
mg= E(d;
I
Hg.C), y f ( t )=
E((dt-
ma)(*;-
m!)*I HBJ)
The following assertion is t r u e .
]Lemma 2. m e process
#
c a n be r e p r e s e n t e d o n ( R , H ~ ~ . ( , P ) in the f o r md t ,
=
A ( @ , t ) m $ d t+
B ( t ) d W t ,to.
and W; is I?*€-adapted Wiener process,
The proof of this lemma can be done using the same arguments as in [ l a ] .
Lemma
3. On probabilisistic spaces( R , H ~ . [ , P ) ,
iE N
the p r o c e s s [ c a n be r e p r e s e n t e d asProof. Using the absolute continuity of the measure P:, i
E N ,
tr
0 in relation to d P f ,t r
0, i EN,
we the m e a s u r e Pt, 1 0 and time independent of the derivatives -d Pt
have, in compliance with t h e analog of the Girsanov theorem on transformation of local process characteristics with absolutely continuous transformation' of probabilis- tic measures, t h a t the characteristics of process
t,
which is represented by equation ( 1 6 ) , do not change with t h e replacement of measure P by measureP ,
i EN.
The fol- lowing equalities holdP
-a.s.whence follows presentation ( 1 7 ) .
The properties of process
t
a r e found to ensure equivalence between t h e proba- bilistic measuresFk
andFi,
i # j . Let us formulate a n d prove this assertion.Lemma
4. MeasuresFi
a n dFi
a r e e q u i v a l e n t o n m e a s u r a b l e space (R , H ~
).Proof. The boundedness of the coefficients A ( i , t ) , a ( i , t ) , b ( i , t ) , B ( t ) a n d non- singularity of the matrix B ( t ) B D ( t ) provide for any
t <
m t h e inequalitywhere
which is t r u e
P
and p - a . s . In accordance with the Liptzer-Shiryaev result[ l a ]
(chapter 7) it yields the equivalence of pti a n d
P)
on( o M , P J ) . I t
follows t h e n t h a t the restrictions of these measures to ( R,H( ) a r e equivalent, t h a t isFti - F/ .
Lemma 5.
7he processes
( m f ) t M , i EN are Guussian o n the pro babiListic spaces ( R , f i , ? ) ,
k E N ; k isnot necessarily equal to
i .Proof. I t is easy to see t h a t t h e processes ( m f ) t M , i E N a r e Gaussian on proba- bilistic spaces
( R , H ~ , P ) ,
i EN.
Let us prove t h a t the processes(m(i.t);,,
i E N a r e Gaussian on probabilistic spaces( R , H ( , ~ ) ,
k #i ,
k EN.
Because the simultaneous ini- tial distribution of t h e variables on any of the probabilistic spaces(R,P.(,P),
i EN
is Gaussian, t h e random variablesmi = E j ( g O 1 ~ 8 )
a r e linear func- tions of with any j E N and, consequently, simultaneous distributions of t h e vari- ablesm & , g O ,
are also Gaussian on probabilistic spaces( n , P . C . P )
where k E N, k # j . Substituting into equation (14) for t h e values ofmi.[t
from equation ( 1 ) we have+ ( b 2 0 ,t ) ~ ' ( t ) + y j ( t ) A '( j ,t ) ( B ( t ) B * ( t ))-'B(t)dW,.t
The fact t.hat t h e joint distribution of
(ma,290,[o)
is Gaussian and t h e formula for 4, a n dmi, j
E N,t
r 0 are linear, makes processes(mi)tao. j
EN
and ([, Gaussian on probabilistic spaces( Q , H ( , ~ ) ,
k EN ,
k # j which was required. Consequently, the variables( ~ ( i , t ) - ~ - ~ O , t ) m i ) ,
j EN
a r e Gaussian for any of t h e m e a s u r e sp.
k EN.
Lemma
6.Singularity conditions for measures F(i and F/ m a y be w r i t t e n as fol-
l o wProof. Using t h e main r e s u l t of paper [15] one can say t h a t measures and are singular if and only if
with F,-probability 1. Taking i n t o account t h a t processes rn/ are Gaussian on any pro- babilistic spaces ( R , H ~ . ? ), k E N and result [18], one can see t h a t this condition is equivalent to condition (18). Condition (19) is not very convenient for checking in a general case. For calculation of the mathematical expectation in (19) perform certain additional constructions.
Let @i.j(t) denote a block m a t r i x
and
q j ,
a block matrixUsing the well-known property of matrix multiplication [19] one can write
Direct verification shows t h a t t h e matrix
can be represented in t e r m s of m a t r i c e s G i . , ( t ) and q m j ( t ) as follows :
Consequently
Let u s find now t h e r e c u r r e n c e equation for t h e matrix G i j ( t ) . Take u p a block m a t r i x f i j ( t ) which has t h e form
Using r e c u r r e n c e equation (14) for mj a n d
mi
a n d equation ( 1 ) for 19; we have for block e l e m e n t s of t h e m a t r i x GiJ ( t )=
~j F, J ( t )Introducing the matrices
the formula for a i i ( t ) can be rearranged into a matrix form
Consequently, the conditions for singularity of t h e measures and
P',
i # j , i , j E N are equivalent to (4) of the theorem.1. Z.k Novoseltseva. "A priory Information in Optimal Problems," Automa-
tics
a n d Telemechanica ( i n Russinn) 39(6) (1968).2. P.I. Kitsul and kV. Lubkov, "On Design of Pseudogradient Algorithms For Adjust- m e n t of an Kalman Filter,"
Automatics
and T e l e m e c h m i c a 37(4) (1978).3. P. Eykhoff , a s t e r n Identification , Rzrameter and State Estimation, John Wiley, New York (1974).
4. J.J. Florentin, "Optimal Probing Adaptive Control of a Simple Bayesian System,"
Journal EZectronic Control(1 1 ) , p.571 (1962).
5. C.T. Leondes and J.O. Pearson, "Kalman Filtering of Systems with Parameter Uncertainties, a Survey.," h t e r n a t i o n a l Journal of Control 17(4) (1973).
6. K.J. Astrom and P. Eykhoff, "System Identification - a Survey," Automatica(7).
pp.123-162 (1971).
7. K.J. Astrom and B. Wittenmark , "Problems of Identification and Control," Journal of M a t h m a t i c a l Analysis and Applications 34, pp.90- 113 (197 1).
8. Y.Z. Tsypkin, Adaptation and Optimization, Nauka (1970).
9. C.T. Leonds and J.O. Pearson, "Kalman Filtering of Systems with Parameters Uncertainties.," h t e r n a t i o n a l J. Control 17(4) (1973).
10. P. Mehra, "On Line Identification of Linear Dynamic Systems with Applications t o Kalman Filtering,"
IEEE
iTSansaction Automatic Control AC-16 (1 ), pp.12-21 (1971).11. L. Ljung, "Convergence Analysis of Parametric Identification Methods.,"
IEEE
i'?rmaction Automatic Control AC-23(3) (1978).12. Y. Baram and N.R. Sandell , "An Information Theoretic Approach To ~ ~ n a m i c a l System Modeling And Identification,"
IEEE Transactions Automatic Control
AC- 23(1), pp.61-66 (1978).13. Y. Baram and N.Y. Sandell, "Consistent Estimation on Finite Parameter Sets with Application t o Linear Systems Identification.,"
IEEE lPransactions Automatic Con- tmL
AG23(3) (1978).14. Y. Baram. "A Sufficient Condition for Consistent Discrimination between Station- ary Gaussian Models.,"
IEEE Transactions o n Automutic Control
AC-23(5) (1978).15. Yu.M. Kabanov, R.S. Liptzer, and A.N. Shiryaev, "Absolute Continuity and Singu- larity of Locally Absolutely Continuous Probability ~ i s t r i b u t i o n s , "
Math. Sbornik
USSR ( i n Russia
) 35(5), pp.631-680 (1979).16.
N.M.
Kuznetsov and kl. Yashin, "On t h e Conditions of t h e Identifiability of Par- tially Observed Systems,"Docladi Akadefnii Nauk SSSR (in Russian)
259(4), pp. 790-793 (198 1).17. N.M. Kuznetsov and A.I. Yashin, "On Consistent Parameter Estimation In Adaptive Filtering,"
Problems Of Control And Information l k o r y
10(5), pp.317-327 (1981).18. R.S. Liptzer a n d AN. Shiryaev,
S u t i s t i c s of Random Processes,
Springer-Verlag, Berlin and New York (1978).19. F.R. Gantmacher,