We derive a regularized representation of the influence function using the spectral theory of positive semidefinite kernels.

We propose priori knowledge-free algorithms for distributionally robust average-reward estimation, with both finite-sample theoretical guarantees and numerical validation.