INTRODUCTION

Here I design filters and models of hyperbola superposition based on medians. Let r_i be a residual and $\Delta r_i$ be a residual perturbation caused by a change $\Delta \bold m$in the model $\bold m$.With $\bold m\leftarrow \bold m+\alpha\Delta \bold m$,choosing $\alpha =- \mbox{median}(r_i/\Delta r_r)$gives us residuals where as many components as possible of the residual are pushed towards zero. This procedure suggests a family of processes that should be immune to bursty noises in data and might quickly give good approximations for inversions of high dimensionality.