Stochastic equations in the problems of semimartingale parameter estimation (Q1780285)

Stochastic equations, both differential and nondifferential, play an important role in many problems of statistics of random processes, in particular, in problems of estimating the unknown parameters of semimartingales. Many problems of the estimation theory are reduced to the investigation of questions connected with the solvability (both strong and weak) of stochastic differential equations (SDEs) and with the asymptotic behavior of solutions of stochastic estimation equations. The formalization of a statistical problem and the desire to study it from the standpoint of generality which is implied by the very essence of a statistical problem make it necessary to consider new specific problems of stochastic analysis and to use new methods of investigating previously known problems. Such problems include the investigation of special types of SDEs and properties of their solutions, and also the asymptotic behavior of the roots of estimation equations in the case of a model disturbance in various formulations. Using the methods of martingale theory, one can make a general statistical model with filtration and an important particular case of it, in which models are associated with semimartingales, the objects of the research. This monograph is concerned with studies of this kind. Chapter 1, ``Structure of solutions of a one-dimensional SDE with unit diffusion coefficient'', deals with the structure of all solutions of the Carathéodory-type stochastic differential equation whose drift coefficient satisfies the well-known Carathéodory condition from the theory of ordinary differential equations. The statistical problem that leads to such SDEs is the innovation problem for the nonlinear filtering (estimation). The assumption is that the investigated processes are Itô processes. The aim is to present such a process without loosing information as a simpler and more convenient process for studying a diffusion-type process. In doing so, it should only be assumed that the drift coefficient of the desired Itô process is square integrable (with respect to the Lebesgue measure) with probability 1. It is desirable not to strengthen the principal assumption which is natural in many senses (in investigating the absolute continuity of measures of the corresponding processes, the structure of Itô functionals and diffusion-type processes and so on). However, the problem in such a general formulation has turned out to be rather difficult and still remains unsolved. Under various additional restrictions both on the structure of the processes considered and on the kind of dependence of the processes participating in the scheme, this problem is solved by many authors. The author speaks on this in more detail below when discussing the results of Chapter 2. Here, in Chapter 1, the drift process is assumed to be a random variable independent of a Wiener process participating in the scheme. Then the principal assumption holds trivially. The aim is to confirm the validity of the hypothesis about the existence of innovations (under the above-mentioned assumption only) in this particular case. However, it is turned out that the corresponding SDE (this problem is reduced to the proof of its strong solvability) has a singularity at the point \(t = 0\) and, therefore, the well-known results from the SDE theory are not applicable in this case. Therefore, the author develops a theory of such SDEs and studies the global and local properties of solutions, as well as the structure of the integral funnel of all solutions and so on. As a result, the author obtains a solution of the considered statistical problem. Chapter 2, ``Partially observable diffusion-type processes. Construction of an innovation process'', deals with the construction of an innovation process for the observed component of a partially observable diffusion-type process in the one- and multidimensional cases. Such problems arise in the estimation theory of partial likelihood schemes, the nonlinear filtering theory, and the theory of stochastic control by incomplete data. As mentioned above, the considered problem belongs to the range of previously existing problems which demand new methods for their investigation. In particular, such methods include the generalized Bayes formula, linear integral inequalities derived for the functional of ``filter'', the stochastic version of the Gronwall-Bellman lemma and so on. It is interesting to note that to solve even a one-dimensional problem, one should use the multidimensional version of the lemma mentioned above. The use of these methods made it possible to get rid of many assumptions like the assumption of smoothness on the coefficients of the scheme. In Chapter 3, ``Estimational stochastic equations and robust estimations in statistical models associated with semimartingales. Continuous alternatives'', the robust estimators for statistical models associated with semimartingales are constructed. The author considers models with shrinking contamination neighborhoods, where a sequence of alternative measures is contiguous to a sequence of basic measures. As the basic class of estimators, the author considers the class of generalized CLAN (consistent, linear, asymptotically normal) estimators. Note that here stochastic equations also play an important role, in particular, in the construction of these estimators. One of the construction methods consists in studying the solvability of stochastic estimation equations and the asymptotic behavior of their solutions for the model disturbance. Thus, the author studies the problem of the local limiting behavior of the roots of such equations by the appropriately generalized Dugue-Kramer-Le Breton method. The author also studies the global limiting behavior of the roots of these equations and obtains the desired CLAN estimators. Using the results obtained, the author constructs B-robust estimators with respect to the risk functional determined by the asymptotic mean-square error. In considering the general model of statistical experiments the author gives the definition of the notion of ``shrinking contamination neighborhoods'' and formulates the minimax optimization problem. The author also develops methods for finding optimal score functions, which, as it turns out, are the Huber functions. The author investigates the problem of robustness in two stages. In stage 1, he studies separately and in great detail an important particular case of the discrete time. The well-known special models of time series are discussed. In stage 2, he considers the general case associated with semimartingales. Finally, in Chapter 4, ``The Robbins-Monro-type stochastic differential equations. Convergence of solutions and strong consistency'', the author introduces and investigates the Robbins-Monro-type stochastic differential equation and, in particular, studies the question whether the solution of this equation is convergent with probability 1. Many generalized schemes of stochastic approximation and recursive estimation can be reduced to equations of the considered type. In this context, the author obtains a theorem which includes as particular cases many familiar results. The question how the results obtained here are related to the previously known results is studied when treating the special cases.