Statistical inference across time scales

Abstract

This thesis studies the problem of statistical inference across time scales for a stochastic process. More particularly we study how the choice of the sampling parameter affects statistical procedures. We narrow down to the inference of jump processes from the discrete observation of one trajectory over [0,T]. As the length of the observation interval T tends to infinity, the sampling rate either goes to 0 (microscopic scale) or to some positive constant (intermediate scale) or grows to infinity (macroscopic scale). We set in a case where there are infinitely many observations. First we specialise in a toy model: a compound Poisson process of unknown intensity with symmetric Bernoulli jumps. Chapter 2 highlights the concept of statistical estimation in the three regimes defined above and the phenomena at stake. We study the properties of the statistical experiments in each regime, we show that the Local Asymptotic Normality property holds in every regimes (microscopic, intermediate and macroscopic). We also provide the formula of the associated Fisher information in each regime. Then we study how a statistical procedure which is optimal (of minimal variance) at a given scale is affected when we use it on data coming from another scale. We focus on the empirical quadratic variation estimator, it is an optimal procedure at macroscopic scales. We apply it on data coming from intermediate and microscopic regimes. Although the estimator remains efficient at microscopic scales, it shows a substantial loss of information when used on data coming from an intermediate regime. That loss can be explicitly related to the sampling rate. We provide an unified procedure, efficient in all regimes. Chapters 3 and 4 focus on microscopic regimes, when the sampling rate decreases to 0. The nonparametric estimation of the jump density of a renewal reward process is studied. We propose an adaptive wavelet threshold density estimator. It achieves minimax rates of convergence for sampling rates that vanish polynomially with T, namely in T^{-alpha} for alpha>0. The estimation procedure is based on the inversion of the compounding operator in the same spirit as Buchmann and Grubel (2003), which specialiase in the study of discrete compound laws. The inverse operator is explicit in the case of a compound Poisson process (see Chapter 3), but has no closed form expression for renewal reward processes (see Chapter 4). In that latter case the inverse operator is approached with a fixed point technique. Finally Chapter 5 studies at which rate identifiability is lost in macroscopic regimes. Indeed when a jump process is observed at an arbitrarily large sampling rate, limit approximations, like Gaussian approximations, become valid and the specificities of the jumps may be lost, as long as the structure of the process is more complex than the one introduced in Chapter 2. First we study a toy model depending on a 2-dimensional parameter. We distinguish two different regimes: fast (macroscopic) regimes where all information on the parameter is lost and slow regimes where the parameter remains identifiable but where optimal estimators converge with slower rates than the expected usual parametric ones. From this toy model lower bounds are derived, they ensure that consistent estimation of Levy processes or renewal reward processes is not possible when the sampling rate grows faster than the square root of T. Finally we identify regimes where an experiment consisting in increments of a compound Poisson process is asymptotically equivalent to an experiment consisting in Gaussian random variables. We also give regimes where there is no consistent estimator for compound Poisson processes depending on too many parameters