La validation d’un modèle d’astronomie via l’analyse de données expérimentales conduit souvent à résoudre un problème mal posé, numériquement et statistiquement. Ceci motive un interêt croissant pour les méthodes d’estimation robuste, parmi lesquelles la classe des M-estimateurs joue un rôle fondamental et optimal. Introduits par Huber, ils sont solutions d’un problème d’optimisation sans expression analytique des solutions, ni même toujours unicité. Nous montrons comment des méthodes d’analyse non lisse permettent de traiter élégamment le problème d’optimisation associé tant du point de vue primal que dual, et comment cette approche a été implémentée en routine pour l’analyse de la rotation terrestre.
In recent years, robustness is one problem that has been given much attention in statistical literature. Under the simultaneous occurence of outliers and/or collinearity, various alternatives to Least Squares are available. Among them, the Huber-M estimators are currently attracting attention when the errors have a contaminated Gaussian distribution. Since they cannot be expressed analytically, finding efficient algorithms to produce them in the case of large data sets is a field of active research. We present new highly parallel algorithms based on the Spingarn Partial Inverse-proximal approach. Practical implementation is described with application to Earth Rotation series.
@article{RFM_2002__6__27_0, author = {Bougeard, Mireille L.}, title = {M\'ethodes d{\textquoteright}analyse non lisse pour l{\textquoteright}estimation robuste}, journal = {Femmes & math}, pages = {27--40}, publisher = {Association femmes et math\'ematiques}, volume = {6}, year = {2002}, language = {fr}, url = {http://www.numdam.org/item/RFM_2002__6__27_0/} }
Bougeard, Mireille L. Méthodes d’analyse non lisse pour l’estimation robuste. Femmes & math, Tome 6 (2002), pp. 27-40. http://www.numdam.org/item/RFM_2002__6__27_0/
[1] Variational convergence for functions and operators, Pitman Publ., 1984 | MR | Zbl
[2] Lien entre referentiel dynamique et ICRS-Hipparcos, Journees Systemes de Reference spatio-temporels 1996 (1996)
[3] Robust Estimation with application to Hipparcos Minor Planets Data, Volume ESA-SP402 (july 1997) (1997), pp. 165-168
[4] Parallel Proximal Decomposition Algorithms for Robust Estimation, Annals of Op. Research, Volume 90 (1999), pp. 247-270 | DOI | MR | Zbl
[5] Computational Aspects of primal-dual algorithms for estimation with constraints, Approximation Optimization and math Economics (Lassonde, ed.), Springer-Verlag (2001), pp. 59-69 | DOI | MR | Zbl
[6] Connection between some statistical estimation criteria, lower c2 functions and Moreau-Yosida approximates, Bulletin International Statistical Institut 47th Session, Volume 1, INSEE Paris press (1989), pp. 159-160
[7] Morse theory for favorable classes of nondifferentiable functions in statistical optimization, Transactions 11th Prague Conference (1990) on Information theory, Academia Publ House of the Czechoslovak Academy of Sciences (1992), pp. 291-301 | Zbl
[8] An improved algorithm for discrete l1 linear approximation, SIAM J. Num. Anal., Volume 10 (1977), pp. 839-848 | DOI | MR | Zbl
[9] Operateurs maximaux monotones, North Holland, 1973
[10] Applications algorithmiques de l’analyse proximale, Masters thesis , University Lyon 1, France, mars (1995) (memoire D.E.A.)
[11] Management Science, 1 (1955), pp. 138-151 | DOI | MR | Zbl
[12] The mathematical structure of Huber’s estimator, SIAM J. of Scientific and Statistical Computing, Volume 6 (1985), pp. 209-219 | DOI | MR | Zbl
[13] J. of Statistical Computation and simulation, 15 (1977), pp. 207-238 | DOI | Zbl
[14] Robust Statistics, Wiley, New York, 1986 | MR | Zbl
[15] Sur quelques points du Système du monde, Mémoire Académie des Sciences Paris, 1793
[16] Duality results and proximal solutions of the Huber- estimator problem, J. Optimization Theory and Applications, Volume 30 (1994), pp. 203-221 | MR | Zbl
[17] Fonctions convexes duales, points proximaux, C. R. Acad. Sci. Paris, Volume A255 (1962), pp. 2897-2899 | Zbl
[18] Convex Analysis, Princeton University Press, 1970 | DOI | MR | Zbl
[19] Conjugate Duality and Optimization, SIAM, 1989 | MR | Zbl
[20] Bounded influence inference, Statistical Data analysis based on the norm, Elsevier Sci. (1987)
[21] Appl. Math. Optimization, 10 (1983), pp. 247-265 | DOI | MR | Zbl
[22] Méthodes de résolution de problèmes mal posés, 1972 (Mir Moscou, 1972) | MR
[23] Méthodes non paramétriques, Dunod éd., 1982