Asymptotic convergence rate of the EM algorithm for Gaussian mixtures

被引:67
作者
Ma, JW [2 ]
Xu, L
Jordan, MI
机构
[1] Shantou Univ, Inst Math, Shantou 515063, Guangdong, Peoples R China
[2] Chinese Univ Hong Kong, Dept Comp Sci & Engn, Shatin, Hong Kong, Peoples R China
[3] Univ Calif Berkeley, Dept Comp Sci, Berkeley, CA 94720 USA
[4] Univ Calif Berkeley, Dept Stat, Berkeley, CA 94720 USA
关键词
D O I
10.1162/089976600300014764
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
It is well known that the convergence rate of the expectation-maximization (EM) algorithm can be faster than those of convention first-order iterative algorithms when the overlap in the given mixture is small. But this argument has not been mathematically proved yet. This article studies this problem asymptotically in the setting of gaussian mixtures under the theoretical framework of Xu and Jordan (1996). It has been proved that the asymptotic convergence rate of the EM algorithm for gaussian mixtures locally around the true solution Theta* is o(e(0.5-epsilon)(Theta*)), where epsilon > 0 is an arbitrarily small number, o(x) means that it is a higher-order infinitesimal as x --> 0, and e(Theta*) is a measure of the average overlap of gaussians in the mixture. In other words, the large sample local convergence rare for the EM algorithm tends to be asymptotically superlinear when e(Theta*) tends to zero.
引用
收藏
页码:2881 / 2907
页数:27
相关论文
共 18 条
  • [1] Delyon B, 1999, ANN STAT, V27, P94
  • [2] MAXIMUM LIKELIHOOD FROM INCOMPLETE DATA VIA EM ALGORITHM
    DEMPSTER, AP
    LAIRD, NM
    RUBIN, DB
    [J]. JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-METHODOLOGICAL, 1977, 39 (01): : 1 - 38
  • [3] GERALD SR, 1980, MATRIX DERIVATIVES
  • [4] Acceleration of the EM algorithm by using quasi-Newton methods
    Jamshidian, M
    Jennrich, RI
    [J]. JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-METHODOLOGICAL, 1997, 59 (03): : 569 - 587
  • [5] Convergence results for the EM approach to mixtures of experts architectures
    Jordan, MI
    Xu, L
    [J]. NEURAL NETWORKS, 1995, 8 (09) : 1409 - 1431
  • [6] HIERARCHICAL MIXTURES OF EXPERTS AND THE EM ALGORITHM
    JORDAN, MI
    JACOBS, RA
    [J]. NEURAL COMPUTATION, 1994, 6 (02) : 181 - 214
  • [7] LANGE K, 1995, J ROY STAT SOC B MET, V57, P425
  • [8] LANGE K, 1995, STAT SINICA, V5
  • [9] LIU CH, 1994, BIOMETRIKA, V81, P633
  • [10] The EM algorithm - An old folk-song sung to a fast new tune
    Meng, XL
    vanDyk, D
    [J]. JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-METHODOLOGICAL, 1997, 59 (03): : 511 - 540