No label defined (Q4214973)
From MaRDI portal
!
This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:
scientific article; zbMATH DE number 1217626
| Language | Label | Description | Also known as |
|---|---|---|---|
| default for all languages | No label defined |
||
| English | No label defined |
scientific article; zbMATH DE number 1217626 |
Statements
1996
0 references
EMS algorithm
0 references
EM algorithm
0 references
convergence
0 references
maximum likelihood estimation
0 references
parameter estimation
0 references
emission tomography
0 references
positive linear systems
0 references
fixed point iteration
0 references
stability
0 references
smoothing matrix
0 references
maximal eigenvector
0 references
0.8522918224334717
0 references
0.8128020167350769
0 references
0.80343097448349
0 references
0.7905977368354797
0 references
The EMS (estimate, maximize, smooth) algorithm, with the underlying (square) nonnegative linear system \(P^t\theta= b\), which is defined by \[ \theta^{n+1}= SF(\theta^n)\theta^n,\quad n= 0,1,2,\dots,\tag{1} \] where \(S\in\mathbb{R}^{m\times m}\) is a nonnegative smoothing matrix, \(F(\theta)\) is a diagonal matrix with entries \(F_j= \sum^m_{k= 1}\widetilde p_{jk}b_k/(P^t\theta)_k\) and \(\widetilde p_{jk}= p_{jk}/\sum_i p_{ji}\), has been proposed as a stabilized and faster converging replacement for the EM (estimate, maximize) algorithm \[ \theta^{n+ 1}= F(\theta^n) \theta^n,\quad n= 0,1,2,\dots,\tag{2} \] which has the undesirable features of very slow and noisy convergence, even when the underlying linear system has a unique positive solution. In this form, one route to the derivation of (2) is the considerable machinery of maximum likelihood estimation and the EM methodology applied to parameter estimation for Poisson data as exemplified by emission tomography (which is also the source of the notation). Here however, we will leave aside this background and consider the EM and EMS algorithms only in the context of algorithms for the solution of positive linear systems. In this paper, we discuss a new fixed point iteration for a well-posed nonnegative linear system, based on (1), which, when it converges, does so to the unique positive solution \(\theta_0\) and yet retains the improved convergence rate and stability properties of the EMS algorithm. This is achieved by adaptively choosing the smoothing matrix \(S\) as the iteration proceeds until, finally, it has the desired maximal eigenvector \(\theta_0\). For the entire collection see [Zbl 0897.00034].
0 references