Information criteria for modelling and identification

Proposes a method to approximate probability laws by histograms. These histograms have to approximate optimally in the sense of the maximum likelihood, and of a mean squares cost, the unknown law of a random process from a single N-sample. The determining of a histogram, that is to say the obtaining of the bins number defining the histogram and the distribution of these bins, is driven by three information criteria. The comparison between two histograms allows the detection of laws changes in real signals. Then, the authors extend the use of these criteria with the aim of extracting the useful information from statistical tables. The aim is to give, from several tables of contingency of characteristics of a population, the one or those which are the most representative of this population. The authors give the first results of an application in pattern recognition: the classification of handwritten digits