The document discusses measuring sample quality using kernels. It introduces the kernel Stein discrepancy (KSD) as a new quality measure for comparing samples approximating a target distribution. The KSD is based on Stein's method and uses reproducing kernels. It can detect when a sample sequence is converging to the target distribution or not. Computing the KSD reduces to pairwise evaluations of kernel functions and is feasible. The KSD converges to zero if and only if the sample sequence converges to the target distribution for certain choices of kernels like the inverse multiquadric kernel with parameter between -1 and 0.