Optimal supports for linear predictive models

The problem of finding the optimal set of causal pixels (support) for use in linear predictive models is addressed. After presenting counterexamples to popular intuitions about supports, a general result relating the distortion incurred with a small support to optimal coefficients of a larger support is derived. A geometrical interpretation is provided. Two algorithms that optimally increase/decrease support sizes by one at each step are presented. Experimental results illustrate the significant gains realized by the algorithms compared with commonly used supports.