In MCMC sampling methods, a transition kernel, as found in Metropolis(/Hastings) algorithm, is the comparison of the likelihood of the current position and the likelihood of the proposed position. However, in support vector machines and gaussian processes, the kernel is an operation defining the similarity of any two points.
In both cases, these definitions might be too simplistic. However, they very loosely are related in that they compare points with one another in support of some decision. Is this the most basic, valid definition of a kernel? Is there a better definition that encompasses any kernel in mathematics?