scientific article; zbMATH DE number 424625

From MaRDI portal
Publication:3136462





The Japanese original of this book appeared in 1976. Its main theme is the canonical representation of Gaussian processes in terms of their innovations, with the classical case of stationary Gaussian processes featuring only as a special case. Any stationary, zero mean, purely nondeterministic Gaussian process \(\dots,X_{-1}, X_ 0, X_ 1, \dots\) can be represented in the form \(X_ n=\sum^ n_{j=-\infty} c_{n-j} \xi_ j\), in such a way that the \(\xi_ n\)'s are independent standard Gaussian and generate the same ``histories as the \(X_ n\)'s. Such a representation is termed a canonical representation and the \(\xi_ n\)'s are the innovations of the process \(\{X_ n\}\). In the continuous parameter case the canonical representation of a stationary Gaussian process \(\{X(t)\}\) is of the form \(X(t)=\int^ t_{-\infty} G(t- u)dB(u)\), where \(\{B(t)\}\) is a Brownian motion (Karhunen representation). The situation is more intricate in the non-stationary case, where the general form of the canonical representation is \[ \widetilde X(t)=\sum^ N_{i=1}\int^ t_ 0F_ i(t,u)dB_ i(u)+\sum_{\{t_ j \leq t\}} \sum^{L_ j}_{l=1} b^ l_ j(t) B^ l_ j. \] Here the \(t_ j\) \((1 \leq j \leq J\) or \(1 \leq j<\infty)\) are points on the parameter line, \(N\) and \(L_ j\) can be finite or \(\infty\) and the main conditions on the other entities are that the \(B_ i(t)\)'s are independent Gaussian processes with independent increments, the \(B_ j^ l\)'s are standard Gaussian variables, mutually independent and independent of the processes \(B_ i(t)\), and the process \(\{\widetilde X(t)\}\) has the same distribution as \(\{X(t)\}\). The book develops systematically the theory of canonical representations, answering questions of existence and uniqueness, investigating the properties of such representations and exploring their connections with other aspects of the theory of Gaussian processes. Each part of the exposition progresses from the simple to the more sophisticated, thus avoiding the perils of plunging the reader straight into the deep waters of the more advanced theory. No specialist knowledge is required on the part of the reader, other than a good command of the general theory of probability. The first chapter deals with foundational matters, such as random variables and their characteristics, conditional expectations, martingales etc. Systems of Gaussian random variables are introduced in the second chapter. This chapter gives the construction of a canonical representation in the simple and illuminating case of a Gaussian sequence \(X_ 1,X_ 2,\dots\), where the innovation sequence \(\xi_ 1,\xi_ 2,\dots\) is easily obtained by means of the Gram-Schmidt orthogonalisation procedure. This case illustrates some other features which are important in the general theory: necessary and sufficient conditions on the representing kernel or on the covariance matrix of a Gaussian sequence are given, for instance, which ensure that the latter is Markov. This chapter also introduces Brownian motion and points out that the continuous parameter analogue of the canonical representation obtained for sequences may not be adequate in general. In Chapter III the authors discuss stationary Gaussian processes with discrete or continuous parameter space. The heart of the chapter is devoted to the spectral representation and its role in the construction of the canonical representation. The construction of the canonical representation of a nonstationary, continuous parameter Gaussian process is presented in the fourth chapter. The representation is then linked to the reproducing kernel Hilbert space of the process, which admits a direct sum decomposition corresponding to the canonical representation. A number of examples is given and the chapter closes with an application to prediction theory. The fifth chapter is devoted to higher order Markov (called multiple Markov) Gaussian processes. In classical approaches the higher order Markov property in the continuous parameter case is expressed in terms of the process and a number of its derivatives. The authors discuss these approaches at length, showing the precise relationship between differential operators, their Green's functions and canonical representations. Paul Lévy's \(M(t)\)-process leads to the Markov property of order \(\infty\), which in fact boils down to Lévy's idea of conditional independence of future and past, given the \(\sigma\)-field of events occurring ``at or near time \(t\). The last section of this chapter discusses the concept of \(T\)-positivity (reflection positivity), inspired from quantum field theory. In Chapter VI the authors present a well-known dichotomy result for Gaussian processes: two Gaussian measures are either mutually absolutely continuous (equivalent) or singular. Conditions for equivalence are given, with the discussion focusing on canonical representations of Gaussian processes equivalent to Wiener processes or to other Gaussian processes. Here Girsanov's theorem plays an important role. In an appendix the reader is introduced to topics such as multiple Wiener integrals and Wiener's homogeneous chaos, stochastic integrals, Itô's formula and Girsanov's theorem. The book ends with some historical notes.




Cited in
(71)








This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3136462)