Kurdyka-Łojasiewicz exponent via inf-projection (Q2162122)

The authors establish that under mild hypotheses the Kurdyka-Łojasiewicz exponent is preserved under inf-projections. In particular, it is \(1/2\), a value related to local linear convergence, for several important convex optimization problems including semidefinite-programming-representable functions and some functions that involve \(C^2\)-cone reducible structures, under conditions such as strict complementarity. The main results are then particularized on relevant optimization models like group-fused Lasso and overlapping group Lasso. In the nonconvex setting, the Kurdyka-Łojasiewicz exponent of many difference-of-convex (DC) functions turns out to be derivable from the one of their natural majorant functions, and in case of the Bregman envelope of a function it coincides with the one of the function itself. Last but not least, an estimation for the Kurdyka-Łojasiewicz exponent of the sum of the least squares function and the indicator function of the set of matrices of rank at most \(k\) are provided as well.