1: \begin{abstract}

2:   We propose iterative proportional scaling (IPS) via decomposable

3:   submodels for maximizing likelihood function of a hierarchical model

4:   for contingency tables.  In ordinary IPS the proportional scaling is

5:   performed by cycling through the members of the generating class of

6:   a hierarchical model.  We propose to adjust more marginals at each

7:   step.  This is accomplished by expressing the generating class as a

8:   union of decomposable submodels and cycling through the decomposable

9:   models.  We prove convergence of our proposed procedure, if the

10:   amount of scaling is adjusted properly at each step.  We also

11:   analyze the proposed algorithms around the maximum likelihood

12:   estimate (MLE) in detail. Faster

13:   convergence of our proposed procedure is illustrated by numerical

14:   examples.

15: \end{abstract}

16: