Let ($x_1, ..., x_n$) be i.i.d. samples drawn from some distribution $P$ with an unknown probability density function $f$. Its kernel density estimator is \begin{align} \hat{f}_h(x) = \frac{1}{n}\sum_{i=1}^n K_h (x - x_i) \quad = \frac{1}{nh} \sum_{i=1}^n K\Big(\frac{x-x_i}{h}\Big), \end{align} where $K$ a symmetric non-negative function that integrates to one.

All consistency proofs (1,2) require that $h(n) \to 0$ for $n \to \infty$ and I think remember that this is a necessary condition for strong and even weak consistency. Are there any analysis about what happens when the bandwidth $h$ is a constant? Maybe error some bounds on the convergence to $f$?

  • $\begingroup$ Isn't this a duplicate of the question you asked earlier: math.stackexchange.com/questions/914789/… $\endgroup$ – user76844 Sep 2 '14 at 13:21
  • $\begingroup$ They are similar. The question earlier was about consistency in general and the got correctly answered. Here I am explicitly interested in a constant bandwidth. $\endgroup$ – Manuel Schmidt Sep 2 '14 at 19:32

Its not a question of error bounds (a so-called second-order condition), but of first order convergence. If an estimator is not consistent, then it does not converge to the true value (i.e., bias), its not that it has a high degree of inter sample variability (i.e., sampling variance). See this paper, especially remark 6 on page 4.

As a simple example, what is the bound on the error of the following estimator:

Let $X_i \sim \mathcal{N}(\mu,1),\;\;\; \hat \mu = \frac{1}{n}\sum_{i=1}^N X_i + 50$


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.