U-statistics assume that the kernel remain fixed. I wonder if theorems in u-stat still hold true when the kernel is random. For instance, I estimate the kernel $h$ using data. The estimated kernel is denoted by $h_n$ with $h_n \xrightarrow{p} h.$ So $h_n$ changes with sample size.

My guess is it can work if $h$ belongs to P-Donsker class and $\mathbb{E}[(h_n-h)^2]\xrightarrow{p}0$. I am not sure if my guess is correct.


This is certainly going to depend on which theorem you're talking about, and how you're trying to insert the randomness in $h$ into the theorem.

But you might be able to use results from $U$-processes, which can be used to show uniform convergence over all possible choices of $h$ (and hence for whichever random one you pick). An excellent resource for an overview, though surely there have been some new results in the last twenty years, is Decoupling: From Dependence to Independence by Victor H. de la Peña and Evarist Giné.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.