La tesi di dottorato prende spunto dallo studio di un nuovo approccio all’inferenza statistica, chiamato “prequential approach”, proposto da A. Philip Dawid. Nel caso di un insieme di osservazioni ottenute in modo sequenziale, in cui ogni nuova osservazione è preceduta da una previsione probabilistica circa il suo possibile valore, fatta sulla base delle osservazioni passate e di un qualche modello, questo approccio richiede che la valutazione empirica di tale modello dipenda solo dalla sequenza delle previsioni effettivamente fatte. Seguendo lo sviluppo di alcune idee emerse da questo approccio, si propone una definizione algoritmica di sequenza casuale basata completamente sul concetto primitivo di martingala. Per tali sequenze, per la cui definizione non si richiede l’introduzione di nessuna distribuzione di probabilità, e quindi senza fare uso degli assiomi della probabilità di Kolmogorov, si mostrano alcune leggi forti dei grandi numeri, la metà superiore della legge del logaritmo iterato, alcuni analoghi del teorema del limite centrale forte di Schatte, ed alcuni risultati che forniscono l’analogo della convergenza forte (sotto gli assiomi di Kolmogorov) della funzione di ripartizione empirica nel caso di alcuni processi stocastici elementari. In particolare, l’analogo del risultato di Schatte sembra essere la prima versione algoritmica del teorema del limite centrale forte (per qualsivoglia schema assiomatico di riferimento) presente nella letteratura.
Following an axiomatic introduction to the prequential (predictive sequential) principle to statistical inference proposed by A. P. Dawid, in which we consider some of the questions it raises, we examine a conjecture on the supposed prequential asymptotic behaviour of significance levels based on a particular class of test statistics. Then, after a presentation of some martingale probability frameworks recently proposed by V. G. Vovk, algorithmic constraints are introduced to give a definition of random sequences on the lines of Martin-Lof's classical approach. This definition, instead of being given, as in the classical algorithmic approach, with respect to a Kolmogorovian probability distribution P, is given only with respect to a sequence of measurable functions by using the principle of the excluded gambling strategy. The idea underlying this approach is that if we are to play an infinite sequence of fair games against an infinitely rich bookmaker, then, whatever computable strategy we choose, we shall never become richer and richer as the game goes on. These random sequences, apart from some basic properties, have been shown to satisfy: an analogue of Kolmogorov's strong law of large numbers; an analogue of the upper half of Kolmogorov's law of the iterated logarithm for binary martingales; and an analogue of Schatte's strong central limit theorem for the coin-tossing process. Besides, for these random sequences, we also investigated the distribution of the values of the corresponding infinite single realizations, in the case of two basic processes. These last results, together with the strong central limit theorem, would provide an instance in which `empirical' distribution functions are derived without the assumption of any Kolmogorovian probability distribution.
On some aspects of the prequential and algorithmic approaches to probability and statistical theory
MINOZZO, Marco
1996-01-01
Abstract
Following an axiomatic introduction to the prequential (predictive sequential) principle to statistical inference proposed by A. P. Dawid, in which we consider some of the questions it raises, we examine a conjecture on the supposed prequential asymptotic behaviour of significance levels based on a particular class of test statistics. Then, after a presentation of some martingale probability frameworks recently proposed by V. G. Vovk, algorithmic constraints are introduced to give a definition of random sequences on the lines of Martin-Lof's classical approach. This definition, instead of being given, as in the classical algorithmic approach, with respect to a Kolmogorovian probability distribution P, is given only with respect to a sequence of measurable functions by using the principle of the excluded gambling strategy. The idea underlying this approach is that if we are to play an infinite sequence of fair games against an infinitely rich bookmaker, then, whatever computable strategy we choose, we shall never become richer and richer as the game goes on. These random sequences, apart from some basic properties, have been shown to satisfy: an analogue of Kolmogorov's strong law of large numbers; an analogue of the upper half of Kolmogorov's law of the iterated logarithm for binary martingales; and an analogue of Schatte's strong central limit theorem for the coin-tossing process. Besides, for these random sequences, we also investigated the distribution of the values of the corresponding infinite single realizations, in the case of two basic processes. These last results, together with the strong central limit theorem, would provide an instance in which `empirical' distribution functions are derived without the assumption of any Kolmogorovian probability distribution.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.