1
$\begingroup$

Given an n-dimensional sample $\left\lbrace X_{i}\right\rbrace$ of i.i.d. observations, let's consider a kernel density function $$\hat{f}_{h_{n},n}(x)=\dfrac{1}{nh_{n}}\cdot\sum_{i=1}^n K\left(\dfrac{X_{i}-x}{h_{n}}\right)$$ where:

  • $h_{n}$ is a smoothing parmeter
  • $K(u)=\mathbb{1}\left( u\in\left\lbrace -1/2,1/2\right\rbrace \right)$ is an indicator function

Question:

what's the optimal smoothing parameter $h_{n}$ in this case?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

The answer will depend on your criterion for "optimal" and on the true density $f$ that is to be estimated. A common criterion is the Mean Integrated Square Error $$MISE = E\left( \int \left(f- \hat{f}\right)^2 dx\right)$$ where $\hat{f}$ is the kernel density estimate of $f$. One approach for finding an optimal $h$ is by means of "AMISE", which is an asymptotic epxansion (for small $h$) of MISE, and then analytically finding its minimum through $AMISE'(h) = 0$. This yields $$h_{AMISE} = \left[\frac{\int K^2 dx}{n \int \left(f''\right)^2 dx \int x^2 K^2 dx}\right]^2$$ This is the startng points for most choices of $h$, e.g., by inserting a Gaussian for $f$, or by inserting the estimate $\hat{f}$ for $f$ and solving the resulting implicit equation for $h$ ("plug-in method").

For a gentle introdutcion into this topic, see

M. C. Jones, J. S. Marron, S. J. Sheather: "A Brief Survey of Bandwidth Selection for Density Estimation. Journal of the American Statistical Association 91,433, pp. 401-407 (1996)

or for a not so gentle, but more recent overview

SJ. Sheather: "Density Estimation." Statistical Science 19,4, pp. 588–597 (2004)

Improve this answer
$\endgroup$
2
  • $\begingroup$ thank you @cdalitz ! Could you explain to me how to replace $f$ with $\hat{f}$, in this case, please? Since the uniform kernel density is not differentiable we cannot compute the second-order derivative of $\hat{f}$. Is there a way to obtain a pseudo derivative? thank you $\endgroup$
    – Almostsurely
    Jul 3, 2020 at 23:21
  • $\begingroup$ @Almostsurely: You are right that$\hat{f}$ is twice differentientiable only if $K$ is twice differentiable (the derivative can be pulled under the sum in the kernel density estimator). Obviously, this fails for your kernel representing a uniform distribution. In this case, you can use "Silverman's rule of thumb" instead, as described in the above articles. $\endgroup$
    – cdalitz
    Jul 4, 2020 at 8:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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