Convex Programming and Decomposition of Discrete Frequency Spectra.
TEXAS UNIV AUSTIN CENTER FOR CYBERNETIC STUDIES
Pagination or Media Count:
The decomposition of a given empirical frequency distribution into a finite mixture of distributions from an admissible class is posed as a convex programming problem. The convexity is attained with any of the usual statistical principles, e.g., minimum variance, maximum likelihood, minimum absolute deviations, by discretizing on the parameters e.g., variance that would yield non-convexity. Computational properties are explored in the context of decomposition of a blood-cell volume distribution and employ minimum absolute deviations. A new class of linear programming problems is thereby uncovered with excessive iteration counts on standard large-scale commerical computer codes. Computational results thus far indicate stability in selection of components of decomposition. Author
- Statistics and Probability
- Operations Research