Estimation of Parameters of Boundary Value Problems for Linear Ordinary Differential Equations with Uncertain Data ()
1. Introduction
The theory of finding estimates of solutions to stochastic differential equations has been intensively developing since the classical works of Kalman and Bucy [1] [2] . This theory found numerous applications in the treatment of the results of experiments in physics, biology, medicine, and many other areas of science and technology. Such successful and broad applications are explained by the fact that Kalman-Bucy methods provide differential equations for optimal mean square estimates which can be solved numerically in the real-time mode. At the same time, it should be noted that Krasovskii and Kurzhanskii proposed in [3] [4] an alternative approach to estimating the solutions of differential equations where perturbations and inaccuracies of additional data about solution were not known and the only thing given was that they belong to a certain domain.
Let us formulate a general approach to the problem of minimax estimation. If a state of a system is described by a linear ordinary differential equation

and a function
is observed in a time interval
, where
, 
and
are known matrices, the minimax estimation problem consists in the most accurate determination of a function
at the “worst” realization of unknown quantities
taken from a certain set. N.N. Krasovskii was the first who stated this problem in [3] . Under different constraints imposed on function
and for known function
he proposed various methods of estimating inner products
, where
in the class of operations linear with respect to observations that minimize the maximal error. Later these estimates were called minimax a priori or guaranteed estimates (see [3] [4] ).
This theory was further developed in the works of Chernous’ko, Pshenichnyi, Kuntzevich, Nakonechnyi, Kirichenko, Podilipenko, and their disciples; one may refer e.g. to [4] -[10] and the bibliography therein.
We note that the duality principle elaborated in [3] [4] , and [5] proved its efficiency for the determination of minimax estimates [5] . According to this principle, finding minimax a priori estimates can be reduced to a certain problem of optimal control of a system; this approach enabled one to obtain, under certain restrictions, recurrent equations, namely, the minimax Kalman-Bucy filter (see [5] ).
The essential results within the frames of
-theory are obtained in [11] .
In this paper, we study estimation of solutions of boundary value problems (BVPs) for ordinary differential equations at fixed points of interval from additional data about their solutions. Such settings may be considered as inverse problems when additional data are given with errors. We assume that these errors are random with unknown correlation functions. Similar problems arise in data processing of observations of the objects or processes described by BVPs for ordinary differential equations with unknown perturbations of the right-hand sides or boundary conditions. We solve the estimation problems using guaranteed linear estimates that minimize maximal mean square estimation errors. It is shown that optimal guaranteed estimates are expressed via solutions to special BVPs for ordinary differential equations.
2. Preliminaries and Auxiliary Results
Assume that it is given a vector-function
with the components belonging to space
and vectors
and
. Consider the following BVP: find a vector-function
that satisfies a system of linear first-order ordinary differential equations
(0.1)
almost everywhere on an interval
and the boundary conditions
(0.2)
at the points 0 and
. Here
is an
matrix with the entries
continuous on
;



and

are
and
matrices of rank
and
, respectively; upper index
denotes transposition of a matrix or a vector and the upper bar throughout the whole text of the paper that e.g. index 
takes all values from 1 to
;
is a space of functions absolutely continuous on
for which the derivative that exists almost everywhere on
belongs to space
and

The problem of finding a function
that satisfies on
the equation
(0.3)
and the boundary conditions
(0.4)
will be called the homogeneous BVP corresponding to BVP (0.1), (0.2).
The solution
to homogeneous BVP (0.3), (0.4) is called the trivial solution.
BVP (0.1), (0.2) can be written in a scalar form:
(0.5)
(0.6)

Let
(0.7)
be a fundamental system of solutions to (0.3) (for the definition, see e.g. [12] p. 179). Then the solutions to (0.3), (0.4) have the form

where, by virtue of (0.4), constants
must be such that
(0.8)
Thus, if the matrix
(0.9)
has rank
, the homogeneous BVP has only the trivial solution. The inverse statement is also valid: if the homogeneous BVP has only the trivial solution then the rank of matrix (0.9) equals
(see, for example [13] ).
Assume in what follows that homogeneous BVP (0.3), (0.4) corresponding to BVP (0.1), (0.2) has only the trivial solution or what is the same that matrix (0.9) has rank
As is known [13] , under this assumption, initial BVP (0.1), (0.2) is uniquely solvable at any right-hand sides
, and 
Formulate the notion of a BVP conjugate to (0.1), (0.2). To this end, introduce the following designations:
is the
unit matrix;
is the
null matrix; 

is a square nondegenerate
submatrix of the matrix
; 

is an
submatrix of
obtained as a result of deleting in
all columns of matrix
(so that
);
is an
matrix such that its
th column equals
th column of matrix
(its size is
),
and
th column equals lth column of matrix
is an
matrix such that its
th column equals k-th column of matrix
and
th column equals lth column of matrix
is an
matrix such that its
th column equals kth column of matrix
and
th column equals lth column of matrix

Introduce more similar notations: 

is a square nondegenerate
submatrix of the matrix
is a
submatrix of the matrix
obtained as a result of deleting in
all columns of matrix
(so that
);
is an
matrix such that its
th column equals kth column of matrix
(the size of the latter is
),
and
th column equals lth column of matrix
is an
matrix such that its
th column equals kth column of matrix
and
th column equals lth column of matrix
is an
matrix such that its
th column equals kth column of matrix
and
th column equals lth column of matrix

By

we will denote the inner product of vectors
and
in the Euclidean space 
Then, if
we have
(0.10)
where the differential operator

will be called formally conjugate to operator 
Let us show that the term
in (0.10) can be represented as
(0.11)
Note first that

where

Then
and

where
and
are vectors composed of components of vector
with the numbers equal to the numbers of components of vectors
and
, respectively. Taking into account that

we have

Analogously

These two equalities yield representation (0.11); using the latter and (0.10), we obtain
(0.12)
In order to write the sum of the first four terms on the right-hand side of (0.12) in a scalar form, introduce the following notations:
(0.13)
(0.14)
(0.15)
(0.16)
(0.17)
(0.18)
Then the Equality (0.12) can be written as
(0.19)
Now we can introduce the notion of the adjoint BVP.
Definition 2.1 The homogeneous BVP
(0.20)
(0.21)
is called adjoint to homogeneous BVP (0.3), (0.4).
Definition 2.2 The inhomogeneous BVP
(0.22)
(0.23)
where
is called adjoint to inhomogeneous BVP (0.1), (0.2).
The results contained in [13] [14] imply that the following statement is valid (for more detailed explanations see [9] , pp. 9-11).
Theorem 2.1 If homogeneous BVP (0.3), (0.4) has only the trivial solution, then the corresponding adjoint BVP (0.20), (0.21) also has only the trivial solution and inhomogeneous BVP (0.22), (0.23) has one and only one solution
at any
.
3. Statement of the Minimax Estimation Problem and Its Reduction to an Optimal Control Problem
Let a vector-function
(0.24)
with the values from the space
be observed on an interval
; here
is an
matrix with the entries that are continuous functions on
and
is an unknown random vector process whose realizations enter observations (0.24).
Denote by
the set of random vector processes
with zero expectation
and second moments
integrable on
such that their correlation matrices
satisfy the inequality
(0.25)
where
is a positive definite matrix of dimension
the entries of
and
are continuous on
is a given positive number,
denotes the trace of the matrix
.
Set
(0.26)
where
are given vectors;
is a given vector-function;
, and
are positive definite matrices of dimensions
, and
, respectively, the entries of
and
are continuous on
,
is a given positive number.
Assume that the right-hand sides
and
of Equation (0.1) and boundary conditions (0.2) are not known exactly and it is known only that the element
belongs to a set
and, additionally,
Further we also will assume, without loss of generality, that in (0.25) and (0.26) 
Let a vector-function
be a solution to BVP (0.1), (0.2).
We will look for an estimation of the inner product
(0.27)
in the class of estimates linear with respect to observations that have the form
(0.28)
where
1 and
is a vector belonging to
,
,
, and
is certain constant. Then 
Definition 3.1 An estimate

for which vector-function
and constant
are determined from the condition

where

is a solution to BVP (0.1), (0.2) at
and


will be called a minimax mean square estimate of inner product
The quantity

will be called an error of the minimax estimation.
We see that the minimax mean square estimate of inner product
is an estimate at which the maximum mean square estimation error calculated for the worst realization of perturbations attains its minimum.
We will show that solution to the minimax estimation problem is reduced to the solution of a certain optimal control problem.
For every fixed
introduce vector-functions
and
as a solution to the following BVP:
(0.29)
Lemma 3.1 Determination of the minimax mean square estimate of inner product
is equivalent to the problem of optimal control of the system described by BVP (0.29) with the cost function
(0.30)
Proof. Show first that BVP (0.29) is uniquely solvable under the condition that functions
and 
belong, respectively, to the spaces
and 
Since homogeneous BVP (0.3), (0.4) has only the trivial solution, the BVP
(0.31)
has, in line with Theorem 2.1, the unique solution for any right-hand side, in particular, at
(0.32)
Denote this solution by
and its restrictions on intervals
, and
by
and
, respectively. Note that function
is absolutely continuous on
(see [15] ).
Let us show that the problem
(0.33)
has one and only one solution at any vector 
Denote by
the coordinates of vector-function
Let
be the fundamental system of solutions of the equation system
on
Then we can represent functions
in the form

where
are constants. Taking into account the boundary conditions at the points
and transmission conditions at
in (0.33), we see that the solution to BVP (0.33) is equivalent to the solution of the following linear equation system with
unknowns

(0.34)
(0.35)
(0.36)
(0.37)
(0.38)
where



denote the coordinates of vector
and
and
denote the entries of matrices
and
, respectively.
Show that system (0.34)-(0.38) is uniquely solvable at any vector
To this end, note that homogeneous system (0.34)-(0.38) (at
) has only the trivial solution.
Indeed, setting
in Equations (0.35) and (0.36), taking into account (0.37) and the fact that

, and
because
is the fundamental system of solutions of the equation system
on
we obtain

Coefficients
satisfy Equations (0.34) and (0.38); therefore vector-function
with the components
is a solution to conjugate BVP (0.20), (0.21) which has only the trivial solution
on
by Theorem 2.1. This implies
, so the homogeneous linear equation system (0.34)- (0.38) (at
) has only the trivial solution. Consequently, system (0.34)-(0.38) and therefore BVP (0.33) which is equivalent to this system are uniquely solvable at any vector
Then vector-functions
form the unique solution to BVP (0.29).
Show next that the determination of the minimax estimate of inner product
is equivalent to the problem of optimal control of the system described by BVP (0.29) with the cost function (0.30).
Using the second and third equations in (0.29) and the fact that
is a solution to BVP (0.1), (0.2) at
and
we easily obtain the relationships


Taking into account the equalities



and that


(we refer to the reasoning on p. 4) we obtain

Taking into notice that


and therefore


we use the last equality to obtain
(0.39)
where

Recalling that
is a vector process with zero expectation, we use condition (0.25) and the known relationship
that couples the dispersion
of random variable
with its expectation
to obtain




which yields
(0.40)
In order to calculate the supremum on the right-hand side of (0.40) we apply the generalized CauchyBunyakovsky inequality [16] and write this inequality in the form convenient for further analysis: for any
, the generalized Cauchy-Bunyakovsky inequality holds

in which the equality is attained at

Setting in the generalized Cauchy-Bunyakovsky inequality


and denoting

we obtain, in line with (2.7), the inequality

where the equality is attained at

Thus,
(0.41)
at 
Calculate the second term on the right-hand side of (0.40). Applying the generalized Cauchy-Bunyakovsky inequality, we have
(0.42)
Here
can be placed under the integral sign according to the Fubini theorem because we assume that
is a random process of the integrable second moment. Transform the last factor on the right-hand side of (0.42):

Taking into account that (0.25) holds, we see that (0.42) yields
(0.43)
It is not difficult to check that here, the equality sign is attained at the element

where
is a random variable such that
and
We conclude that statement of the lemma follows now from (0.40), (0.41), and (0.43).
4. Representations for Minimax Mean Square Estimates of Functionals from Solutions to Two-Point Boundary Value Problems and Estimation Errors
In this section we prove the theorem concerning general form of minimax mean square estimates. Solving optimal control problem (0.29), (0.30), we arrive at the following result.
Theorem 4.1 The minimax mean square estimate of expression
has the form

where
(0.44)

(0.45)
and vector-functions
and
,
are determined from the solution to the problem
(0.46)
Here
and
The minimax estimation error
(0.47)
Problem (0.46) is uniquely solvable.
Proof. We will solve optimal control problem (0.29), (0.30). Represent solutions
of problem (0.29) as
where
and
denote the solutions to this problem at
and
respectively. Then function (0.30) can be represented in the form

where



Since solution
of BVP (0.31) is continuous2 with respect to right-hand side
defined by
(0.32), the function
is a linear bounded operator mapping the space
to

Thus,
is a continuous quadratic form corresponding to a symmetric continuous bilinear form

is a linear continuous functional defined on
and
is a constant independent of
. We have

Using Theorem 1.1 from [17] , we conclude that there is one and only one element
such that

Therefore

Taking into consideration the latter equality, (0.30), and designations on p. 11, we obtain
(0.48)
Introduce functions
and
as a unique solution to the following problem3:
(0.49)
Now transform the sum of the last for terms on the right-hand side of (0.48) taking into notice that
and
We have
(0.50)
From Equalities (0.48)-(0.50) it follows that

so that
(0.51)
Functions
and
are absolutely continuous on segments

, and
, respectively, as solutions to BVP (0.49); therefore, functions
and
that perform optimal control are continuous on
and
Replacing in (0.29) functions
and
by
and
defined by formulas (0.51) and denoting
we arrive at problem
(0.46) and equalities (0.44).
Taking into consideration the way this problem was formulated we can state that its unique solvability follows from the fact that functional (0.30) has one minimum point
.
Now let us prove representation (0.47). Substituting into formula
expressions (0.44) for 
and
we have
(0.52)
Next, we can apply the reasoning similar to that on p. 4 and use (0.46) to obtain

which yields
(0.53)
In a similar manner, using the equality

we obtain
(0.54)
Relationships (0.52)-(0.54) yield

which is to be proved.
It is easy to see that function
defined by (0.45) and the function
(0.55)
satisfy the following uniquely solvable BVP
(0.56)
where



and
is the characteristic function of interval
.
Now Theorem 4.1 can be restated as follows:
Theorem 4.1' The minimax estimate of expression
has the form

where
(0.57)

and vector-functions
and
are determined from the solution to problem 0.56.
Obtain now another representation for the minimax mean square estimate of quantity
which is independent of
and
. To this end, introduce vector-functions
and
as solutions to the problem
(0.58)
at realizations
that belong with probability 1 to space 
Note that unique solvability of problem (0.58) at every realization can be proved similarly to the case of (0.46). Namely, one can show that solutions to the problem of optimal control of the system







with the cost function

can be reduced to the solution of problem (0.58) where the optimal control
is expressed in terms of the solution to this problem as
; the unique solvability of the problem follows from the existence of the unique minimum point
of functional
.
Considering system (0.58) at realizations
it is easy to see that its solution is continuous with respect to the right-hand side. This property enables us to conclude, using the general theory of linear continuous transformations of random processes, that the functions
considered as random fields have finite second moments.
Theorem 4.2 The following representation is valid

Proof. By virtue of (0.44) and (0.58),
(0.59)
Next,
(0.60)
Similarly,
(0.61)
From (0.59), (0.60), and (0.61) it follows
(0.62)
However,
(0.63)
(0.64)
Subtracting from (0.63) equality (0.64), we obtain

or
(0.65)
Since


we can use the latter equalities, (0.65), and the fact that
is a symmetric matrix to obtain
(0.66)
Performing a similar analysis, one can prove that
(0.67)
From (0.66), (0.67), and (0.62) and the expression for
it follows

The theorem is proved.
As is easily seen from (0.58), the functions
and
defined by

and

satisfy the following uniquely solvable BVP:
(0.68)
at realizations
that belong with probability 1 to space
.
Thus, Theorem 4.2 can be restated in the following form.
Theorem 4.2' The minimax mean square estimate of expression
has the form

where vector-function
is determined from the solution to problem (0.68).
Remark. Function
can be taken as a good, in certain sense, estimate of solution
of initial BVP (0.1), (0.2) on 
As an example, consider the case when a vector-function
is observed on an interval
, where a vector-function
with values in
is a solution to the BVP
(0.69)
and operator
is defined by the relation

where
is a positive definite
-matrix whose entries are continuous functions on 
Note that this problem has the unique classical solution if
is continuous on
and the unique generalized solution if 
Assume that, as well as in the previous case,
is an
matrix with the entries that are continuous functions on
and
is a random vector process with zero expectation
and unknown
correlation matrix
. Assume also that domains
and
are given in the form (0.25) and (0.26) where matrices
, and
entering (0.26) have dimensions
, and 
Write Equation (0.69) as a first-order system by setting
and introducing a vector-function


with
components, a vector
with
components, a
-matrix

matrices
and
. Then system (0.69) can be written as
(0.70)
(0.71)
Applying Theorems 1 and 2 and performing necessary transformations in the resulting equations that are similar to (0.46) and (0.58) (in terms of the designations introduced above) we prove the following Theorem 4.3 The minimax mean square estimate of expression
has the form

where


and vector-functions
,
and
are determined from the solutions to the problems






and







respectively.
5. Minimax Mean Square Estimates of Solutions Subject to Incomplete Restrictions on Unknown Parameters
Assume again that observations have form (0.24) and unknown parameters
,
, and
belong to the domain
(0.72)
where
is given in (0.26). The correlation function of process
belongs to domain (0.25). Introduce the set
(0.73)
here
where
is the solution to BVP (0.29).
Lemma 5.1

where
(0.74)
This lemma can be proved using formula (0.39).
Lemma 5.2
is a convex closed set in the space
.
Proof. The convexity of set
is obvious. Let us prove that this set is closed.
Note that functions
and
can be represented as
(0.75)
where
and
are known matrix functions with the elements from
and
and
are vectors. Expression (0.75) can be obtained if we introduce a vector
such that
Then 

where
is a solution to the equation

and
is the unit matrix. Next,

Since BVP (0.29) is uniquely solvable, there exists one and only one vector
satisfying the system of algebraic equations

Solving this system we determine
in the form

where
is a known matrix function continuous on
and
is a known vector. Taking into account this equality, we obtain expressions (0.75). From these relationships, it follows that if a sequence
converges in
to a function
then


which proves that
is a closed set.
Assume now that
is nonempty. Then the following statement is valid.
Theorem 5.1 There exists the unique minimax mean square estimate of expression
which can be represented in the form (0.44) at
where vector-functions
and
solve the equations
(0.76)
Proof. Similarly to Theorem 4.1 one can show that for
the following equality holds

where

and
are solutions to Equations (0.29) at
and
is a strictly convex lower semicontinuous functional on a closed convex set
and
Therefore there exists one and only one vector
such that
This vector can be determined from the relationship

where


and
are Lagrange multipliers.
Further analysis is similar to the proof of Theorem 4.1. Let vector-functions
and
,
be solutions to the system
(0.77)
Theorem 5.2 Assume that for any vector
set
is nonempty. Then system (0.77) is uniquely solvable and the equality

holds Proof. Introduce functions
,
as a solution to the BVP







where
Define a set

Since
is nonempty, the same is valid for
for any vector
. Similarly to the case of
one can show that
is a convex closed set. Denote by
the functional of the form

One can show, following Theorem 5.1, that on set
there is one and only one point of minimum of functional
namely,

where functions
and
are determined from system (0.77). The proof of the equality

is similar to that in Theorem 4.2.
6. Conclusions
For a system described by a one-dimensional two-point BVP with decoupling boundary conditions at the endpoints of the interval and quadratic restrictions imposed on the unknown deterministic data and the second moments of observation noise, we have obtained guaranteed mean square estimates of inner product
where
is the unknown solution of the BVP at a point
and
. Guaranteed estimates are obtained using the duality of the problems of estimation and optimal control. We have shown that guaranteed mean square estimates and estimation errors are expressed via solutions to special optimal control problems for conjugate BVPs. The solutions to these optimal control problems enable one to find explicit expressions for estimates and estimation errors both for distributed and point observations.
The obtained results are applied to minimax estimation of solutions of two-point BVPs for linear ordinary second-order differential equations.
Methods and results of the paper may be used for estimation under uncertainties of the states of the systems described by more general linear BVPs for different classes of functional--differential equations; in particular, for systems of differential equations with impulse perturbations, differential equations with multipoint conditions, and in several other cases.
7. Results of Numerical Experiments
Let realizations of the random variables
(0.78)
be observed. Here
are independent random variables for which
;
is a solution of the BVP
(0.79)
(0.80)
and

are eigenfunctions of the operator
, where
.
We assume that function
is not known exactly and is chosen arbitrarily from the set
, where
is a certain constant.
Applying the technique similar to the estimation methods developed in Section 4, it is possible to obtain expressions for the minimax mean square estimates in the case when observations have the form (0.78). In particular, in this case the function
which approximates the solution
of BVP (0.79)-(0.80) on the interval
is determined from the system




and can be represented in the form

The exact solution
and its estimate
(bold curves) calculated on the basis of the modelled observations are presented in Figure 1 and Figure 2. Calculations are performed at
,
, and
for 
using the parameters
and
(Figure 1) and
and
(Figure 2).
As can be seen from these figures, parameter
plays the crucial role as far as the estimation quality is concerned. In fact, this parameter directly influences the signal-to-noise ratio.
The calculations were performed using Wolfram Mathematica.
NOTES
*Corresponding author.

1If
then the minimax estimation problem can be solved in a similar manner but somewhat simpler.

2This continuous dependence follows from the representation of function
in terms of Green's matrix
of BVP (0.31) (see , p. 115):


3The unique solvability of problem (0.49) can be proved similarly to problem (0.29).