![MathJax Logo](/templates/jsp/_style2/_tandf/pb2/images/math-jax.gif)
Abstract
The problem of optimal estimation of the linear functional depending on the unknown values of a stochastic sequence
with nth stationary increments from observations of the sequence
at points
, where
is a stationary sequence uncorrelated with
, is considered. Formulas for calculating the mean square error and the spectral characteristic of the optimal linear estimate of the functional are derived in the case where spectral densities of stochastic sequences are exactly known and admit the canonical factorizations. In the case of spectral uncertainty, where spectral densities are not known exactly, but sets of admissible spectral densities are specified, the minimax-robust method is applied. Formulas and relations that determine the least favourable spectral densities and the minimax-robust spectral characteristics are proposed for the given sets of admissible spectral densities. The filtering problem for a class of cointegrated sequences is investigated.
Public Interest Statement
The crucial assumption of application of traditional methods of finding solution to the filtering problem for random processes is that spectral densities of the processes are exactly known. However, in practical situations complete information on spectral densities is impossible and the established results cannot be directly applied to practical filtering problems. This is a reason to apply the minimax-robust method of filtering and derive the minimax estimates since they minimize the maximum value of the mean-square errors for all spectral densities from a given set of admissible densities simultaneously. In this article, we deal with the problem of optimal estimation of functionals depending on the unknown values of a random process with stationary increments based on observations of the process and a noise. In the case where spectral densities of the processes are not exactly known, relations for determining least favourable spectral densities and minimax-robust spectral characteristics are proposed.
1. Introduction
Basic results of the theory of wide sense stationary and related stochastic processes found their applications in analysis of models of economic and financial time series. The most simple examples are linear stationary models such as moving average (MA), autoregressive (AR) and autoregressive-moving average (ARMA) sequences, all of which refer to stationary sequences with rational spectral densities without unit AR-roots. Time series with trends and seasonal components are modelled by integrated ARMA (ARIMA) sequences which have unit roots in their autoregressive parts and are the examples of sequences with stationary increments. Such models are investigated during the last 30 years. The main points concerning model definition, parameter estimation, forecasting and further investigation of the models are discussed in the well-known book by Box, Jenkins, and Reinsel (Citation1994). While analysing financial data economists noticed that in some special cases linear combinations of integrated sequences become stationary. Granger (Citation1983) called this phenomenon cointegration. Cointegrated sequences found their application in applied and theoretical econometrics and financial time series analysis (see Engle & Granger, Citation1987).
The problem of estimation of unknown values of stochastic processes (extrapolation, interpolation and filtering problems) is an important part of the theory of stochastic processes. Effective methods of solution of the linear extrapolation, interpolation and filtering problems for stationary stochastic processes were developed by A.N. Kolmogorov, N. Wiener and A.M. Yaglom. See selected works by Kolmogorov (Citation1992), books by Wiener (Citation1966) and Yaglom (Citation1987a,Citation1987b). Further results one can find in the book by Rozanov (Citation1967).
Random processes with stationary nth increments are one of generalizations of the notion of stationary process that were introduced by Pinsker and Yaglom (Citation1954), Yaglom (Citation1955,Citation1957), Pinsker (Citation1955). They described the spectral representation of the stationary increment process and the canonical factorization of the spectral density, solved the extrapolation problem for such processes and discussed some examples. See books by Yaglom (Citation1987a,Citation1987b) for more relative results and references.
Traditional methods of finding solutions to extrapolation, interpolation and filtering problems for stationary and related stochastic processes are applied under the basic assumption that the spectral densities of the considered stochastic processes are exactly known. However, in most practical situations complete information on the spectral structure of the processes isn’t available. Investigators can apply the traditional methods considering the estimated spectral densities instead of the true ones. However, as it was shown by Vastola and Poor (Citation1983) with the help of some examples, this approach can result in significant increasing of the value of the error of estimate. Therefore, it is reasonable to derive estimates which are optimal for all densities from a certain class of spectral densities. The introduced estimates are called minimax-robust since they minimize the maximum of the mean-square errors for all spectral densities from a set of admissible spectral densities simultaneously. The minimax-robust method of estimation was proposed by Grenander (Citation1957) and later developed by Franke and Poor (Citation1984), Franke ( franke1985) for investigating the extrapolation and interpolation problems. For more details we refer to the survey paper by Kassam and Poor (Citation1985) who collected results in minimax (robust) methods of data processing till 1984. A wide range of results in minimax-robust extrapolation, interpolation and filtering of stochastic processes and sequences belongs to Moklyachuk (Citation1990,Citation2000,Citation2001,Citation2008a,Citation2015). Later Moklyachuk and Masyutka (Citation2006a,Citation2006b,Citation2007,Citation2008,Citation2011,Citation2012) developed the minimax technique of estimation for vector-valued stationary processes and sequences. Dubovets’ka, Masyutka, and Moklyachuk (Citation2012) investigated the problem of minimax-robust interpolation for another generalization of stationary processes – periodically correlated sequences. In the further papers Dubovetska and Moklyachuk (Citation2013a,Citation2013b,Citation2014a,Citation2014b) investigated the minimax-robust extrapolation, interpolation and filtering problems for periodically correlated processes and sequences. See the book by Golichenko and Moklyachuk (Citation2014) for more relative results and references. The minimax-robust extrapolation, interpolation and filtering problems for stochastic sequences and processes with nth stationary increments were solved by Luz and Moklyachuk (Citation2012,Citation2013a,Citation2013b,Citation2014a,Citation2014b,Citation2015a,Citation2015b,Citation2015c,Citation2016a,Citation2016b) Moklyachuk and Luz (Citation2013). The obtained results are applied to find solution of the extrapolation and filtering problems for cointegrated sequences (Luz and Moklyachuk, Citation2014b,Citation2015c). The problem of extrapolation of stochastic sequences with stationary increments from observations with non-stationary noise was investigated by Bell (Citation1984).
In the present article, we deal with the problem of optimal linear estimation of the functional which depends on the unknown values of a stochastic sequence
with nth stationary increments from observations of the sequence
at points
, where
is a stationary stochastic sequence uncorrelated with the sequence
. Solution to this problem based on the Hilbert space projection method is described in the paper by Luz and Moklyachuk (Citation2014b). The derived formulas for calculating the spectral characteristic and the mean square error of the optimal estimate
of the functional
are complicated for application and need to construct the inverse operator
which is also a complicated problem. On the other hand, most of spectral densities of stochastic sequences applied in time series analysis admit factorization. This is a reason to derive formulas for calculating the spectral characteristic and the mean square error of the optimal estimate of the functional which use coefficients of the canonical factorizations of spectral densities. In this article, it is shown that the derived formulas can be simplified under the condition that the spectral densities are such that the canonical factorizations of the functions hold true. The case of spectral certainty as well as the case of spectral uncertainty is considered. Formulas for calculating values of the mean-square errors and the spectral characteristics of the optimal linear estimate of the functional are derived under the condition of spectral certainty, where the spectral densities of the processes are exactly known. In the case of spectral uncertainty, where the spectral densities of the processes are not exactly known, but a class of admissible spectral densities is given, relations that determine the least favourable spectral densities and the minimax spectral characteristics are derived for some classes of spectral densities. The obtained results are applied to investigate the filtering problem for cointegrated sequences.
2. Stationary increment stochastic sequences. Spectral representation
In this section, we present a brief description of the properties of stochastic sequences with nth stationary increments. More detailed description of the properties of such sequences can be found in the books by Yaglom (Citation1987a,Citation1987b).
Definition 2.1
For a given stochastic sequence the sequence
(1)
(1)
where is a backward shift operator with step
, such that
, is called a stochastic nth increment sequence with step
.
For the stochastic nth increment sequence the following relations hold true:
(2)
(2)
(3)
(3)
where coefficients are determined by the representation
Definition 2.2
The stochastic nth increment sequence generated by the stochastic sequence
is wide sense stationary if the mathematical expectations
exist for all and do not depend on
. The function
is called a mean value of the nth increment sequence and the function
is called a structural function of the stationary nth increment sequence (or structural function of nth order of the stochastic sequence
).
The stochastic sequence which determines the stationary nth increment sequence
by formula (1) is called a sequence with stationary nth increments (or integrated sequence of order n).
Theorem 2.1
The mean value and the structural function
of the stochastic stationary nth increment sequence
can be represented in the following forms:
(4)
(4)
where c is a constant, is a left-continuous nondecreasing bounded function with
. The constant c and the structural function
are determined uniquely by the increment sequence
.
On the other hand, a function of the form (4) with a constant c and a function
of the form (5) with a function
satisfying the indicated conditions are the mean value and the structural function of some stationary nth increment sequence
, respectively.
Using representation (5) of the structural function of a stationary nth increment sequence and the Karhunen theorem (see Gikhman & Skorokhod, Citation2004; Karhunen, Citation1947), we obtain the following spectral representation of the stationary nth increment sequence
:
(5)
(5)
where is a random process with uncorrelated increments on
with respect to the spectral function
:
(6)
(6)
3. The filtering problem
Consider a stochastic sequence which generates the stationary nth increment sequence
with absolutely continuous spectral function
that has spectral density
. Let
be uncorrelated with the sequence
stationary stochastic sequence with absolutely continuous spectral function
which has spectral density
. Without loss of generality we will assume that the mean values of the increment sequence
and stationary sequence
equal to 0. Let us also assume that the step
.
Consider the problem of mean-square optimal linear estimation of the functional
which depends on the unknown values of the sequence from observations of the sequence
at points
.
We will suppose that coefficients a(k), , which determine the functional satisfy the inequalities
(7)
(7)
and the spectral densities and
satisfy the minimality condition
(8)
(8)
This condition (9) is sufficient in order that the mean-square error of the estimate of the functional is not equal to zero.
Note, that
is the spectral density of the stochastic sequence .
The functional can be represented in the form
, where
and
. In the case where conditions (8), the functional
has a finite second moment. To construct an estimate of the functional
it is sufficient to have an estimate of the functional
. Since the functional
depends on the values of the stochastic sequence
, which is observed, we have the following relations:
(9)
(9)
It follows from relation (10) that any estimate of the functional
can be represented in the form
(10)
(10)
where is the spectral characteristic of the estimate
.
Denote by the closed linear subspace of the Hilbert space
of random variables
that have zero first and finite second moment
,
, which is generated by values
,
. Denote by
the closed linear subspace of the Hilbert space
of square integrable on
functions with respect to the measure which has the density
, generated by functions
. It follows from the relation
that there exists a one to one correspondence between elements from the space
and elements
from the space
correspondingly.
Let , where by
we denote the least integer number among the numbers that are greater than or equal to x.
The mean square optimal estimate is a projection of the element
on the subspace
. This projection is described in the paper by Luz and Moklyachuk (Citation2014b). A solution of the filtering problem is described in the following theorem.
Theorem 3.1
Let be a stochastic sequence which defines stationary nth increment sequence
with absolutely continuous spectral function
which has spectral density
. Let
be uncorrelated with the sequence
stationary stochastic sequence with absolutely continuous spectral function
which has spectral density
. Let the coefficients
satisfy conditions (8). Let the spectral densities
and
of stochastic sequences
and
satisfy the minimality condition (9). The mean square optimal linear estimate
of the functional
based on observations of values
based on observations of the sequence
at points
can be calculated by formula (11). The spectral characteristic
and the mean square error
of the optimal estimate
are calculated by the formulas
(11)
(11)
and(12)
(12)
respectively, where ,
, coefficients
,
, are calculated by the formula
(13)
(13)
Here ,
,
are the linear operators in the space
determined with the help of matrices with the elements
,
,
for
,
This theorem gives us a possibility to find a solution to the filtering problem with the help of the Fourier coefficients of the functions
The derived formulas are complicated for application and need to construct the inverse operator which is also a complicated problem. On the other hand, most of spectral densities of stochastic sequences applied in time series analysis admit the factorization. This is a reason to derive formulas for calculating the spectral characteristic and the mean square error of the optimal estimate
of the functional
which use the canonical factorizations of the spectral densities. In particular, the proposed formulas (12) and (13) can be simplified under the condition that the spectral densities
and
are such that the following canonical factorizations of the functions hold true:
(14)
(14)
Denote by the linear operator in the space
which is determined by the matrix with elements
,
. The following lemmas from the paper by Luz and Moklyachuk (Citation2015b) give formulas for calculating operators
and
with the help of the coefficients of factorizations (15)–(16).
Lemma 3.1
Let the spectral densities and
be such that the canonical factorizations (15) – (16) hold true. Define linear operators
and
in the space
with the help of matrices with elements
and
for
,
and
for
. Then
(a) | the following factorization holds true | ||||
(b) | linear operator |
Lemma 3.2
Let canonical factorizations (15–16) hold true. Let the linear operators and
in the space
determined in the same way as in the lemma (3.1) and let the linear operator
in the space
determined by the matrix with elements
for
,
for
. Define also a linear operator
in the space
with the help of matrices with elements
,
, where the coefficients
,
, are determined in (17). Then
(a) | the linear operators | ||||
(b) | the inverse operator |
Lemma 3.3
Let the function admit the factorization (16) and let the linear operators
and
in the space
are determined by matrix with elements
and
,
. Then the operators
and
satisfy the relation
, where the linear operator
is determined in the lemma 3.1.
With the help of the introduced results we show that formulas (12) and (13) can be simplified in the case where the spectral densities and
are such that the canonical factorizations (15) – (16) hold true. Denote
. With the help of factorization (15) we have the next transformations:
where ,
, is the m-th elements of the vector
. Since
the following equality holds true(16)
(16)
With the help of factorization (17) and the relation
we have the next transformations(17)
(17)
Define coefficients in the following way:
,
for
,
for
, where coefficients
are determined by formula (14). Define the vectors
and
. Denote by
the linear operator which is determined by the matrix with elements
for
,
for
.
Making use of relations (18), (19) and the relation
we derive the following formula for calculating the spectral characteristic:(18)
(18)
In the last relation ,
, is the mth element of the vector
,
is the linear operator which is determined by the matrix with elements
,
,
,
is the linear operator which is determined by the matrix with elements
,
.
It is stated in Lemma 3.3 that the operator admits the representation
, where
is the linear operator which is determined by the matrix with elements
,
.
,
, is the m-th element of the vector
,
,
is the linear operator which is determined by the matrix with elements
,
,
,
is the linear operator which is determined by the matrix with elements
,
.
It follows from the Lemma 3.2 that the operator admits the representation
, where
is the linear operator which is determined by the matrix with elements
,
.
The mean square error is calculated by the formula(19)
(19)
These observations can be summarized in the form of the theorem.
Theorem 3.2
Let be a stochastic sequence which defines the stationary nth increment sequence
with an absolutely continuous spectral function
which has spectral density
. Let
be an uncorrelated with the sequence
stationary stochastic sequence with an absolutely continuous spectral function
which has spectral density
. Let the coefficients
satisfy condition (8), and let the spectral densities
and
of the sequences
and
admit canonical factorizations (15–16). The spectral characteristic
and the mean square error
of the optimal estimate
of the functional
based on observations of the sequence
at points
can be calculated by formulas (20) and (21).
Remark 3.1
Results described in theorem 3.2 can be used for finding the optimal estimate of the functional
based on observations of the sequence
at points
. For this purpose it is sufficient to take
for
in the formulas (11), (20), (21). In the case where
we have the smoothing problem. Solution of this problem is described in the following corollary.
corollary 3.1
The optimal estimate of the unknown value
based on observations of the sequence
at points
is calculated by formula
The spectral characteristic and the mean square error
of the optimal estimate
are calculated by the formulas
and
correspondingly, where ,
is an infinite dimension vector with elements
for
and
for
,
,
.
Remark 3.2
Since for all and
the condition
holds true, then there is a function such that
and
(see Gikhman & Skorokhod, Citation2004). In the case where factorization (15) holds true, the function
admits the factorization
(20)
(20)
The spectral density admits the canonical factorization
(21)
(21)
where the function has the radius of convergence
and does not have zeros in the region
.
Introduce the linear operators ,
and
in the space
with the help of the matrices with elements
,
and
for
,
,
and
for
. Denote
. The following relations hold true:
(22)
(22)
where ,
.
4. Filtering of cointegrated stochastic sequences
Consider two integrated stochastic sequences and
with absolutely continuous spectral functions
and
which have spectral densities
and
correspondingly.
Definition 4.1
Two integrated stochastic sequences and
are called cointegrated (of order 0) if there exists a constant
such that the sequence
is stationary.
The filtering problem for cointegrated stochastic sequences consists in finding the mean-square optimal linear estimate of the functional
which depends on the unknown values of the sequence from observations of the sequence
at points
. This problem can be solved by using results presented in the preceding section under the condition that sequences
and
are uncorrelated.
Suppose that the spectral densities and
are such that the following canonical factorizations hold true
(23)
(23)
Detemine the linear operators ,
and
with the help of the canonical factorizations (25–26) in the same way as operators
,
and
were defined. By using theorem 3.2, we derive that the spectral characteristic
of the optimal estimate
(24)
(24)
of the functional is calculated by the formula
(25)
(25)
where
the operator is determined in remark 3.2. The value of the mean-square error is calculated by the formula
(26)
(26)
Theorem 4.1
Let and
be two cointegrated stochastic sequences which have absolutely continuous spectral functions
and
with the spectral densities
and
, respectively. Let coefficients
satisfy conditions (8). If the spectral densities
and
admit canonical factorizations (25–26), and the sequences
and
are uncorrelated, then the spectral characteristic
and the mean-square error
of the optimal linear estimate
of the functional
of unknown elements
,
, from observations of the sequence
at points
is calculated by formulas (28) and (29).
Example 4.1
Consider two random sequences which are determined by the equations
where are two uncorrelated sequences of independent identically distributed random variables with
,
,
. Denote
and suppose that ,
. In this case the random sequences
and
are ARIMA(0, 1, 1) sequences with the spectral densities
The difference is a stationary sequence. That is why the integrated random sequences
and
are cointegrated with the parameter of cointegration
. Since the random sequences
and
are uncorrelated, then the sequences
and
are uncorrelated also.
Consider the problem of filtering of the functional from observations of the sequence
at points
. Making use of Theorem 4.1 we will have
. Since the first coordinate of the vector
is equal to 0, then
and
That is why the spectral characteristic of the optimal estimate
of the functional
is calculated by the formula
Denote by and
,
. The optimal estimate
of the functional
is calculated by the formula
The value of the mean-square error of the optimal estimate
of the functional
is calculated by the formula
5. Minimax-robust method of filtering
Formulas for calculation of values of the mean-square errors and spectral characteristics of the optimal linear estimates of the functional based on observations of the stochastic sequence
are derived under the condition that the spectral densities
and
of the stochastic sequences
and
are known. In the case where the spectral densities are not exactly known, but a set
of admissible spectral densities is given, the minimax (robust) approach to estimation of functionals which depend on the unknown values of stochastic sequence with stationary increments is reasonable. In other words, we are interested in finding an estimate that minimizes the maximum of mean-square errors for all spectral densities from a given class
of admissible spectral densities simultaneously.
Definition 5.1
For a given class of spectral densities the spectral densities
,
are called the least favourable densities in the class
for the optimal linear filtering of the functional
if the following relation holds true
Definition 5.2
For a given class of spectral densities the spectral characteristic
of the optimal linear estimate of the functional
is called minimax-robust if there are satisfied conditions
The following statements are consequences of the introduced definitions of least favourable spectral densities, minimax-robust spectral characteristic and 3.2.
Lemma 5.1
Spectral densities ,
which admit canonical factorizations (15) and (16) are least favourable in the class
for the optimal linear filtering of the functional
based on observations of the sequence
at points
if coefficients
of the canonical factorizations
(27)
(27)
determine a solution of the constrained optimization problem(28)
(28)
The minimax spectral characteristic is calculated by formula (20) if
.
Lemma 5.2
The spectral density which admits the canonical factorizations (15) – (16) with the known spectral density
is least favourable in the class
for the optimal linear filtering of the functional
based on observations of the sequence
at points
, if coefficients
of the canonical factorizations
(29)
(29)
determine a solution of the constrained optimization problem(30)
(30)
The minimax spectral characteristic is calculated by formula (20) if
.
Lemma 5.3
The spectral density which admit the canonical factorizations (15) with the known spectral density
is least favourable in the class
for the optimal linear filtering of the functional
based on observations of the sequence
at points
, if coefficients
of the canonical factorizations
(31)
(31)
determine a solution of the constrained optimization problem(32)
(32)
with the fixed coefficients . The minimax spectral characteristic
is calculated by formula (20) if
.
The minimax spectral characteristic and the pair
of least favourable spectral densities form a saddle point of the function
on the set
. The saddle point inequalities
hold true if and
, where
is a solution to the constrained optimization problem
(33)
(33)
This constrained optimization problem is equivalent to the unconstrained optimization problem(34)
(34)
where is the indicator function of the set
. A solution
to this unconstrained optimization problem is characterized by a condition
, which is the necessary and sufficient condition that the pair
belongs to the set of minimums of the convex functional
( see Moklyachuk, Citation2008b; Pshenichnyi, Citation1971; Rockafellar, Citation1997). Here the notion
determines a subdifferential of the functional
at the point
, which is a set of all linear bounded functionals
on
satisfying the inequality
In the case of investigation the cointegrated sequences we get the following optimization problem for determination of the least favourable spectral densities(35)
(35)
A solution to this unconstrained optimization problem is characterized by the condition
.
The form of the functionals and
allows us to find derivatives and differentials of these functionals in the space
. Hence, the complexity of the optimization problems (37) and (38) is characterized by the complexity of finding subdifferentials of the indicator functions
of the sets
.
6. Least favourable spectral densities in the class ![](//:0)
![](//:0)
Consider the problem of minimax-robust estimation of the functional based on observations of the sequence
at points of time
provided the spectral densities
and
admit canonical factorizations (15 and 16) and belong to the set of admissible spectral densities
, where
We use the Lagrange method of indefinite multiplies to find a solution to the constrained optimization problem (36), we get the following relations for determination the least favourable spectral densities ,
:
(36)
(36)
where the multiplies , matrices
,
, vector
are determined with the help of factorizations (16) and (22) of the functions
and
, relation (24) and condition
(37)
(37)
Making use the derived reasonings, we can formulate the following statements.
Proposition 6.1
The spectral densities and
which admit canonical factorizations (16) and (22) are least favourable in the class
for the optimal linear estimation of the functional
based on observations of the sequence
at points
, if they satisfy equations (39 and 40), relations (24), the problem (31) and conditions (41). The function
determined by formula (20), is minimax spectral characteristic of the optimal estimate of the functional
.
Proposition 6.2
Suppose that the spectral density is known and admits canonical factorization (16). The spectral density
from the class is least favourable for the optimal linear estimation of the functional
based on observations of the sequence
at points
, if the coefficient
, matrices
,
, vector
are determined from canonical factorizations (16), (22) of the functions
and
, relations (24), problem (33) and condition
. The function
, determined by formula (20), is minimax spectral characteristic of the optimal estimate of the functional
.
Proposition 6.3
Suppose that the spectral density is known and admits canonical factorization (16). The spectral density
from the class is least favourable for the optimal linear estimation of the functional
based on observations of the sequence
at points
, if the coefficient
, vector
are determined from canonical factorization (22) of the function
, relation (24), problem (35) and condition
. In this case, the matrices
,
are determined from the canonical factorization (16) of the given spectral density
. The function
, determined by formula (20), is minimax spectral characteristic of the optimal estimate of the functional
.
Consider the problem of minimax-robust estimation of the functional based on observations of the cointegrated sequence
at points of time
provided the spectral densities
and
admit canonical factorizations (25 and 26) and stochastic sequences
and
are uncorrelated. The least favourable spectral densities in the set of admissible spectral densities
, where
are determined by the condition . It follows from this condition that the least favourable spectral densities
,
are determined by the relations
(38)
(38)
where coefficients , vector
, matrices
,
are determined by factorization Equations (25) and (26) of the function
and
, relation (24) and conditions
(39)
(39)
Thus, we have the following statements.
Proposition 6.4
The spectral densities and
, that admit canonical factorizations (25) and (26), are least favourable in the class
for the optimal linear estimation of the functional
based on observations of the cointegrated with
sequence
at points
, if these densities satisfy equations (42 and 43) and are determined by relations (24), the problem (31) with
and conditions (44). The function
, determined by formula (28), is minimax spectral characteristic of the optimal estimate of the functional
.
7. Least favourable densities in the class ![](//:0)
![](//:0)
Consider the problem of minimax-robust estimation of the functional based on observations of the sequence
at points of time
provided the spectral densities
and
admit canonical factorizations (15 and 16) and belong to the set of admissible spectral densities
, where
The spectral densities ,
,
are known and fixed and the spectral densities
,
are bounded. It follows from the condition
that the least favourable spectral densities
,
satisfy the relations
(40)
(40)
where and
if
;
and
, if
;
and
if
. The coefficients
,
, matrices
,
, vector
are determined with the help of factorizations (16) and (22) of functions
and
, relations (24) and conditions (41).
The following theorems hold true.
Proposition 7.1
The spectral densities and
which admit the canonical factorizations (16) and (22) are least favourable in the class
for the optimal linear estimation of the functional
based on observations of the sequence
at points
, if they satisfy equations (45 and 46), relations (24), problem (31) and conditions (41). The function
determined by formula (20), is minimax spectral characteristic of the optimal estimate of the functional
.
Proposition 7.2
Suppose that the spectral density is known and admits canonical factorization (23). The spectral density
from the class is least favourable for the optimal linear estimation of the functional
based on observations of the sequence
at points
, if the coefficient
, matrices
,
, vector
are determined from the canonical factorizations (16), (22) of the functions
and
, relations (24), the problem (33) and condition
. The function
, determined by formula (20), is minimax spectral characteristic of the optimal estimate of the functional
.
Proposition 7.3
Suppose that the spectral density is known and admits the canonical factorization (16). The spectral density
from the class is least favourable for the optimal linear estimation of the functional
based on observations of the sequence
at points
, if the coefficient
, vector
are determined from the canonical factorization (22) of the function
, relation (24), problem (35) and condition
. In this case, the matrices
,
are determined from canonical factorization (16) of the given spectral density
. The function
, determined by formula (20), is minimax spectral characteristic of the optimal estimate of the functional
.
Consider the problem of minimax-robust estimation of the functional based on observations of the cointegrated sequence
at points of time
provided the spectral densities
and
admit canonical factorizations (25) – (26) and stochastic sequences
and
are uncorrelated. The least favourable spectral densities in the set of admissible spectral densities
, where
under the condition that the spectral densities and
admit the canonical factorizations (25 and 26). From the condition
, we get the following relations that determine the least favourable spectral densities
,
are determined by the relations
(41)
(41)
where and
if
;
and
if
;
and
if
. The coefficients
,
, matrices
,
, vector
are determined by the canonical factorizations (25) and (26) of functions
and
, relations (24) and condition (44).
Thus, we have the following statements.
Proposition 7.4
The spectral densities and
, that admit canonical factorizations (25 and 26), are least favourable in the class
for the optimal linear estimation of the functional
based on observations of the cointegrated with
sequence
at points
, if these densities satisfy equations (47 and 48) and are determined by relations (24), problem (31) with
and conditions 44). The function
, determined by formula (28), is minimax spectral characteristic of the optimal estimate of the functional
.
8. Conclusions
In this article, we propose a solution of the filtering problem for the functional which depends on unobserved values of a stochastic sequence
with stationary nth increments. Estimates are based on observations of the sequence
at points of time
, where
is a stationary sequence uncorrelated with
. We derive formulas for calculating the values of the mean-square errors and the spectral characteristics of the optimal linear estimate of the functional in the case where spectral densities
and
of the sequences
and
are exactly known. The obtained formulas are simpler than those obtained with the help of the Fourier coefficients of some functions determined by the spectral densities. In the case of spectral uncertainty, where spectral densities are not known exactly, but a set of admissible spectral densities is specified, the minimax-robust method is applied. Formulas that determine the least favourable spectral densities and the minimax (robust) spectral characteristics are derived for some special sets of admissible spectral densities. The obtained results are applied to find a solution of the filtering problem for a class of cointegrated sequences.
Acknowledgements
The authors would like to thank the referees for careful reading of the article and giving constructive suggestions.
Additional information
Funding
Notes on contributors
Maksym Luz
Maksym Luz is a PhD student, Department of Probability Theory, Statistics and Actuarial Mathematics, Taras Shevchenko National University of Kyiv. His research interests include estimation problems for random processes and sequences with stationary increments.
Mikhail Moklyachuk
Mikhail Moklyachuk is a professor, Department of Probability Theory, Statistics and Actuarial Mathematics, Taras Shevchenko National University of Kyiv. He received PhD degree in Physics and Mathematical Sciences from the Taras Shevchenko University of Kyiv in 1977. His research interests are statistical problems for stochastic processes and random fields. He is also a member of editorial boards of several international journals.
References
- Bell, W. (1984). Signal extraction for nonstationary time series. The Annals of Statistics, 12, 646–664.
- Box, G. E. P., Jenkins, G. M., & Reinsel, G. C. (1994). Time series analysis, Forecasting and control (3rd ed.). Englewood Cliffs, NJ: Prentice Hall.
- Dubovets’ka, I. I., Masyutka, O. Y., & Moklyachuk, M. P. (2012). Interpolation of periodically correlated stochastic sequences. Theory of Probability and Mathematical Statistics, 84, 43–56.
- Dubovets’ka, I. I., & Moklyachuk, M. P. (2013a). Filtration of linear functionals of periodically correlated sequences. Theory of Probability and Mathematical Statistics, 86, 51–64.
- Dubovets’ka, I. I., & Moklyachuk, M. P. (2013b). Minimax estimation problem for periodically correlated stochastic processes. Journal of Mathematics and System Science, 3, 26–30.
- Dubovets’ka, I. I., & Moklyachuk, M. P. (2014a). Extrapolation of periodically correlated processes from observations with noise. Theory of Probability and Mathematical Statistics, 88, 67–83.
- Dubovets’ka, I. I., & Moklyachuk, M. P. (2014b). On minimax estimation problems for periodically correlated stochastic processes. Contemporary Mathematics and Statistics, 2, 123–150.
- Engle, R. F., & Granger, C. W. J. (1987). Co-integration and error correction: Representation, estimation and testing. Econometrica, 55, 251–276.
- Franke, J. (1985). Minimax robust prediction of discrete time series. Zeitschrift far. Wahrscheinlichkeitstheorie und verwandte Gebiete, 68, 337–364.
- Franke, J., & Poor, H. V. (1984). Minimax-robust filtering and finite-length robust predictors (Robust and Nonlinear Time Series Analysis. Lecture Notes in Statistics, Vol. 26, 87–126. Heidelberg: Springer-Verlag.
- Gikhman, I. I., & Skorokhod, A. V. (2004). The theory of stochastic processes. I.. Berlin: Springer.
- Golichenko, I. I., & Moklyachuk, M. P. (2014). Estimates of functionals of periodically correlated processes. Kyiv: NVP “Interservis".
- Granger, C. W. J. (1983). Cointegrated variables and error correction models ( UCSD Discussion paper. 83–13a).
- Grenander, U. (1957). A prediction problem in game theory. Arkiv för Matematik, 3, 371–379.
- Ioffe, A. D., & Tihomirov, V. M. (1979). Theory of extremal problems (p. 460). Amsterdam: North-Holland.
- Karhunen, K. (1947). Uber lineare Methoden in der Wahrscheinlichkeitsrechnung. Annales Academiae Scientiarum Fennicae. Series A I. Mathematica, 37, 3–79.
- Kassam, S. A., & Poor, H. V. (1985). Robust techniques for signal processing: A survey. Proceedings of the IEEE, 73, 433–481.
- Kolmogorov, A. N. (1992). Selected works of A. N. Kolmogorov. In A. N. Shiryayev (Ed.), Probability theory and mathematical statistics (Vol. II). Dordrecht: Kluwer Academic.
- Luz, M. M., & Moklyachuk, M. P. (2012). Interpolation of functionals of stochastic sequences with stationary increments from observations with noise. Prykladna Statystyka. Aktuarna ta Finansova Matematyka, 2, 131–148.
- Luz, M. M., & Moklyachuk, M. P. (2013a). Interpolation of functionals of stochastic sequences with stationary increments. Theory of Probability and Mathematical Statistics, 87, 117–133.
- Luz, M. M., & Moklyachuk, M. P. (2013b). Minimax-robust filtering problem for stochastic sequence with stationary increments. Theory of Probability and Mathematical Statistics, 89, 117–131.
- Luz, M., & Moklyachuk, M. (2014a). Robust extrapolation problem for stochastic processes with stationary increments. Mathematics and Statistics, 2, 78–88.
- Luz, M., & Moklyachuk, M. (2014b). Minimax-robust filtering problem for stochastic sequences with stationary increments and cointegrated sequences. Statistics, Optimization & Information Computing, 2, 176–199.
- Luz, M., & Moklyachuk, M. (2015a). Minimax interpolation problem for random processes with stationary increments. Statistics, Optimization & Information Computing, 3, 30–41.
- Luz, M., & Moklyachuk, M. (2015b). Filtering problem for random processes with stationary increments. Contemporary Mathematics and Statistics, 3, 8–27.
- Luz, M., & Moklyachuk, M. (2015c). Minimax-robust prediction problem for stochastic sequences with stationary increments and cointegrated sequences. Statistics, Optimization & Information Computing, 3, 160–188.
- Luz, M., & Moklyachuk, M. (2016a). Filtering problem for functionals of stationary sequences. Statistics, Optimization & Information Computing, 4, 68–83.
- Luz, M., & Moklyachuk, M. (2016b). Minimax prediction of random processes with stationary increments from observations with stationary noise. Cogent Mathematics, 3, 1–17, 1133219.
- Moklyachuk, M. P. (1990). Minimax extrapolation and autoregressive-moving average processes. Theory of Probability and Mathematical Statistics, 41, 77–84.
- Moklyachuk, M. P. (2000). Robust procedures in time series analysis. Theory of Stochastic Processes, 6, 127–147.
- Moklyachuk, M. P. (2001). Game theory and convex optimization methods in robust estimation problems. Theory of Stochastic Processes, 7, 253–264.
- Moklyachuk, M. P. (2008a). Robust estimates for functionals of stochastic processes. Kyiv: Kyiv University.
- Moklyachuk, M. P. (2008b). Nonsmooth analysis and optimization (p. 400). Kyiv: Kyivskyi Universitet.
- Moklyachuk, M. P. (2015). Minimax-robust estimation problems for stationary stochastic sequences. Statistics, Optimization & Information Computing, 3, 348–419.
- Moklyachuk, M., & Luz, M. (2013). Robust extrapolation problem for stochastic sequences with stationary increments. Contemporary Mathematics and Statistics, 1, 123–150.
- Moklyachuk, M. P., & Masyutka, O Yu (2006a). Extrapolation of multidimensional stationary processes. Random Operators and Stochastic Equations, 14, 233–244.
- Moklyachuk, M. P., & Masyutka, O. Y. (2006b). Robust estimation problems for stochastic processes. Theory of Stochastic Processes, 12, 88–113.
- Moklyachuk, M. P., & Masyutka, O. Y. (2007). Robust filtering of stochastic processes. Theory of Stochastic Processes, 13, 166–181.
- Moklyachuk, M. P., & Masyutka, O. Y. (2008). Minimax prediction problem for multidimensional stationary stochastic sequences. Theory of Stochastic Processes, 14, 89–103.
- Moklyachuk, M. P., & Masyutka, O. Y. (2011). Minimax prediction problem for multidimensional stationary stochastic processes. Communications in Statistics - Theory and Methods, 40, 3700–3710.
- Moklyachuk, M. P., & Masyutka, O. Y. (2012). Minimax-robust estimation technique for stationary stochastic processes (p. 296). Lap Lambert Academic.
- Pinsker, M. S. (1955). The theory of curves with nth stationary increments in Hilbert spaces. Izvestiya Akademii Nauk SSSR. Seriya Matematicheskaya, 19, 319–344.
- Pinsker, M. S., & Yaglom, A. M. (1954). On linear extrapolation of random processes with nth stationary increments. Doklady Akademii Nauk SSSR, 94, 385–388.
- Pshenichnyi, B. N. (1971). Necessary conditions for an extremum. Pure and applied mathematics. 4 (Vol. XVIII, 230 p). New York, NY: Marcel Dekker.
- Rockafellar, R. T. (1997). Convex analysis (p. 451). Princeton, NJ: Princeton University Press.
- Rozanov, Y. A. (1967). Stationary stochastic processes. San Francisco, CA: Holden-Day.
- Vastola, K. S., & Poor, H. V. (1983). An analysis of the effects of spectral uncertainty on Wiener filtering. Automatica, 28, 289–293.
- Wiener, N. (1966). Extrapolation, interpolation, and smoothing of stationary time series. With engineering applications. Cambridge: MIT Press, Massachusetts Institute of Technology.
- Yaglom, A. M. (1955). Correlation theory of stationary and related random processes with stationary nth increments. Matematicheskii Sbornik, 37, 141–196.
- Yaglom, A. M. (1957). Some classes of random fields in n-dimensional space related with random stationary processes. Teoriya Veroyatnostej i Ee Primeneniya, 2, 292–338.
- Yaglom, A. M. (1987a). Correlation theory of stationary and related random functions. Basic results (Springer Series in Statistics, Vol. 1, p. 526). New York, NY: Springer-Verlag.
- Yaglom, A. M. (1987b). Correlation theory of stationary and related random functions. In Supplementary notes and references (Springer Series in Statistics, Vol. 2, p. 258). New York, NY: Springer-Verlag.