1: \begin{abstract}

2: 	\noindent\small 

3: 	 Based on an idea in~\cite{APSS2016} we propose a new iterative

4:    multiplicative filtering algorithm for label assignment matrices

5:    which can be used for the supervised partitioning of data.

6:    Starting with a row-normalized matrix containing the averaged distances

7:    between prior features and the observed ones the method assigns in a very

8:    efficient way labels to the data. 

9:    We interpret the algorithm as a gradient ascent method with respect

10:    to a certain function on the product manifold of positive numbers followed

11:    by a reprojection 	onto a subset of the probability simplex consisting of

12:    vectors whose components are bounded away from zero by a small constant.

13:    While such boundedness away from zero is necessary to avoid an arithmetic

14:    underflow, our convergence results imply that they are also necessary for

15:    theoretical reasons.

16:    Numerical examples show that the proposed simple and fast algorithm leads to 

17:    very good results. 

18:    In particular we apply the method for the partitioning 

19:    of manifold-valued images.	

20: \end{abstract}

21: