1: \begin{abstract}

2: We describe the Median $K$-flats (MKF) algorithm, a simple online

3: method for hybrid linear modeling, i.e., for approximating data by a

4: mixture of flats. This algorithm simultaneously partitions the data

5: into clusters while finding their corresponding best approximating

6: $\ell_1$ $d$-flats, so that the cumulative $\ell_1$ error is

7: minimized. The current implementation restricts $d$-flats to be

8: $d$-dimensional linear subspaces. It requires a negligible amount of

9: storage, and its complexity, when modeling data consisting of $N$

10: points in $\reals^D$ with $K$ $d$-dimensional linear subspaces, is

11: of order $O(n_s \cdot K \cdot d \cdot D+n_s \cdot d^2 \cdot D)$,

12: where $n_s$ is the number of iterations required for convergence

13: (empirically on the order of $10^4$). Since it is an online

14: algorithm, data can be supplied to it incrementally and it can

15: incrementally produce the corresponding output. The performance of

16: the algorithm is carefully evaluated using synthetic and real data.

17: \vspace{.1in}

18: \end{abstract}

19: