English

Algorithmic statistics: normal objects and universal models

Information Theory 2015-12-15 v1 math.IT

Abstract

Kolmogorov suggested to measure quality of a statistical hypothesis PP for a data xx by two parameters: Kolmogorov complexity C(P)C(P) of the hypothesis and the probability P(x)P(x) of xx with respect to PP. P. G\'acs, J. Tromp, P.M.B. Vit\'anyi discovered a small class of models that are universal in the following sense. Each hypothesis SijS_{ij} from that class is identified by two integer parameters i,ji,j and for every data xx and for each complexity level α\alpha there is a hypothesis SijS_{ij} with jil(x)j\le i\le l(x) of complexity at most α\alpha that has almost the best fit among all hypotheses of complexity at most α\alpha. The hypothesis SijS_{ij} is identified by ii and the leading iji-j bits of the binary representation of the number of strings of complexity at most ii. On the other hand, the initial data xx might be completely irrelevant to the the number of strings of complexity at most ii. Thus SijS_{ij} seems to have some information irrelevant to the data, which undermines Kolmogorov's approach: the best hypotheses should not have irrelevant information. To restrict the class of hypotheses for a data xx to those that have only relevant information, Vereshchagin introduced a notion of a strong model for xx: those are models for xx whose total conditional complexity conditional to xx is negligible. An object xx is called normal if for each complexity level α\alpha at least one its best fitting model of that complexity is strong. In this paper we show that there are "many types" of normal strings. Our second result states that there is a normal object xx such that all its best fitting models SijS_{ij} are not strong for xx. Our last result states that every best fit strong model for a normal object is again a normal object.

Cite

@article{arxiv.1512.04510,
  title  = {Algorithmic statistics: normal objects and universal models},
  author = {Alexey Milovanov},
  journal= {arXiv preprint arXiv:1512.04510},
  year   = {2015}
}

Comments

19 pages, 5 figures

R2 v1 2026-06-22T12:09:33.525Z