AIPENOVA AZIZA SRAILOVNA

AIPENOVA AZIZA SRAILOVNA
NONPARAMETRIC ESTIMATION OF DERIVATIVES OF DENSITIES
USING BESOV NORMS
ABSTRACT
of the dissertation of Aipenova A.S. for the scientific degree of Doctor of
Philosophy (PhD) in the specialty 6D060100 - Mathematics
Relevance of the theme of dissertation.
Among the tools of mathematical statistics, nonparametric estimation
occupies a central place. It replaced the parametric estimates, which are much
inferior to it in accuracy and flexibility. The first results in this area have appeared
about 50 years ago in the works of Rosenblatt and Parsen. Currently, the basic
theory is well developed, but at the same there are a number of areas where
significant improvement is possible. Often the improvement can be achieved
through the application of the theory of functions, which are not well known
among statisticians. Considered dissertation belongs to the class of works in this
direction. At the same time can improve even acknowledged classical results.
The asymptotic properties of ̂
have been studied by, among others,
Bhattacharya (1967), Schuster (1969) and Silverman (1978). Schuster (1969)
showed that the uniform continuity of was necessary for uniform consistency,
under the condition ∑
. This condition is substantially weakened
in Silverman (1978). Singh (1977, 1979, 1987) show that it is possible to reduce
bias and improve the mean integrated squared error (MISE) of ̂
by
considering restrictions on the class of kernels used in its construction. Similar
efforts have been undertaken by Muller (1984), Coppejans and Sieg (2005), Bearse
and Rilstone (2009) and Henderson and Parmeter (2012).
In the theory of nonparametric estimation, there are many different methods.
Recently, Mynbaev and Martins-Filho (2010) proposed a class of nonparametric
kernel density estimators that attains bias reduction relative to ̂ by imposing
global higher order Lipschitz conditions on .
In this dissertation we extend the nonparametric class of kernel density
estimators proposed by Mynbaev and Martins-Filho (2010) to the estimation of order density derivatives. Contrary to some existing derivative estimators, the
estimators in our proposed class have a full asymptotic characterization, including
uniform consistency, asymptotic normality and exact convergence rates. In
addition, we discuss the criterion of the bandwidth choice that minimizes an
asymptotic approximation for the estimators' integrated squared error. Further, we
give an integral representation for the bias and exact orders of the bias and
variance of the proposed derivative estimator.
Also, we consider in the dissertation bias reduction using higher order kernels
which is popular in statistics, and provide construction of higher order kernels. In
addition, we extend the nonparametric estimation method to estimation of a
differential expression. It determines the relevance of the theme of the dissertation.
Research objective. Research objective is to extend the nonparametric class
of kernel density estimators proposed by Mynbaev and Martins-Filho (2010) to the
estimation of order
density derivatives and to construct new higher-order
kernels that reduce bias. We show the effectiveness of the proposed methods by
the Monte-Carlo method.
Subject of research. Subject of research is the nonparametric class of kernel
density estimators proposed by Mynbaev and Martins-Filho (2010) and
construction of higher order kernels by way of multiplication of densities by
polynomials.
Purpose of research: Purpose of research is to obtain exact asymptotic
characterization for density derivative estimators and show the effectiveness of the
proposed method by the Monte-Carlo method.
Scientific novelty of research: The dissertation presents the following
results:
The first result: the nonparametric class of density estimators proposed by
Mynbaev and Martins-Filho (2010) is extended to the estimation of -order
density derivatives.
The second result: the asymptotic properties of the proposed derivative
estimator are found, including the uniform consistency, asymptotic normality and
exact convergence rates.
The third result: the integral representation for the bias of the proposed
derivative estimator is found.
The fourth result: the criterion of the optimal bandwidth choice is proved.
The fifth result: the exact orders of the bias and variance of the proposed
derivative estimator are obtained.
The sixth result: the effectiveness of the proposed method is shown by the
Monte-Carlo method.
The seventh result: the method for improving bias in kernel density
estimation using higher order kernels is obtained and the effectiveness of the
method is shown by the Monte-Carlo method.
The eighth result: the nonparametric estimation method is extended to the
estimation of differential expressions.
The nineth result: the method to construct higher order kernels is obtained.
In addition, the expressions and the moments for the higher order kernels are
obtained.
Theoretical significance of research: Results of the dissertation make a
significant contribution to the theory nonparametric kernel density estimation.
Practical significance of the results: The work can be used in mathematical
statistics and in numerous applications in engineering, finance, medicine and
biology.
Structure and volume of the dissertation: The dissertation consists of
introduction, three chapters, conclusion and list of references. The work is
presented in 76 pages, including 6 figures, 10 tables and a list of references of the
47 items.