Caractérisation des limites fondamentales de l'erreur quadratique moyenne pour l'estimation de signaux comportant des points de rupture

Abstract : This thesis deals with the study of estimators' performance in signal processing. The focus is the analysis of the lower bounds on the Mean Square Error (MSE) for abrupt change-point estimation. Such tools will help to characterize performance of maximum likelihood estimator in the frequentist context but also maximum a posteriori and conditional mean estimators in the Bayesian context. The main difficulty comes from the fact that, when dealing with sampled signals, the parameters of interest (i.e., the change points) lie on a discrete space. Consequently, the classical large sample theory results (e.g., asymptotic normality of the maximum likelihood estimator) or the Cramér-Rao bound do not apply. Some results concerning the asymptotic distribution of the maximum likelihood only are available in the mathematics literature but are currently of limited interest for practical signal processing problems. When the MSE of estimators is chosen as performance criterion, an important amount of work has been provided concerning lower bounds on the MSE in the last years. Then, several studies have proposed new inequalities leading to tighter lower bounds in comparison with the Cramér-Rao bound. These new lower bounds have less regularity conditions and are able to handle estimators’ MSE behavior in both asymptotic and non-asymptotic areas. The goal of this thesis is to complete previous results on lower bounds in the asymptotic area (i.e. when the number of samples and/or the signal-to-noise ratio is high) for change-point estimation but, also, to provide an analysis in the non-asymptotic region. The tools used here will be the lower bounds of the Weiss-Weinstein family which are already known in signal processing to outperform the Cramér-Rao bound for applications such as spectral analysis or array processing. A closed-form expression of this family is provided for a single and multiple change points and some extensions are given when the parameters of the distributions on each segment are unknown. An analysis in terms of robustness with respect to the prior influence on our models is also provided. Finally, we apply our results to specific problems such as: Gaussian data, Poisson data and exponentially distributed data.
Complete list of metadatas

Cited literature [139 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01957951
Contributor : Abes Star <>
Submitted on : Monday, December 17, 2018 - 4:03:06 PM
Last modification on : Thursday, July 11, 2019 - 3:26:09 AM
Long-term archiving on : Monday, March 18, 2019 - 4:02:15 PM

File

72783_BACHARACH_2018_archivage...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01957951, version 1

Citation

Lucien Bacharach. Caractérisation des limites fondamentales de l'erreur quadratique moyenne pour l'estimation de signaux comportant des points de rupture. Traitement du signal et de l'image [eess.SP]. Université Paris-Saclay, 2018. Français. ⟨NNT : 2018SACLS322⟩. ⟨tel-01957951⟩

Share

Metrics

Record views

170

Files downloads

642