How to calculate expectation and variance of kernel density estimator?

[schniefen]

TL;DR Summary

What is the expectation and variance of $f_n(t)=\frac{1}{n}\sum_{i=1}^n \frac{1}{h}k\left(\frac{t-x_i}{h}\right)$?

This is a question from a mathematical statistics textbook, used at the first and most basic mathematical statistics course for undergraduate students. This exercise follows the chapter on nonparametric inference. An attempt at a solution is given. Any help is appreciated.

Exercise:

Suppose $x_1, ..., x_n$ are independent and identically distributed (i.i.d.) observations of a random variable $X$ with unknown distribution function $F$ and probability density function $f\in C^m$, for some $m>1$ fixed. Let
$$f_n(t)=\frac{1}{n}\sum_{i=1}^n \frac{1}{h}k\left(\frac{t-x_i}{h}\right)$$
be a kernel estimator of $f$, with $k\in C^{m+1}$ a given fixed function such that $k\geq 0$, $\int_{\mathbb{R}} k(u)\mathrm{d}u=1$, $\mathrm{supp} (k)=[-1,1]$ and bandwidth $h=h(u)$ (for the time being unspecified).
1. Show that $\mathbb{E}[f_n(t)]=\int_{\mathbb{R}} k(u) f(t-hu)\mathrm{d}u$.
2. Make a series expansion of $f$ around $t$ in terms of $hu$ in the expression for $\mathbb{E}[f_n(t)]$. Suppose that $k$ satisfies $\int_{\mathbb{R}} k(u)\mathrm{d}u=1$, $\int_{\mathbb{R}} k(u)u^l\mathrm{d}u=0$ for all $1<l<m$ and $\int_{\mathbb{R}} k(u)u^m\mathrm{d}u<\infty$. Determine the bias $\mathbb{E}[f_n(t)]-f(t)$ as a function of $h$.
3. Suppose that $\mathrm{Var}[k(X_1)]<\infty$ and determine $\mathrm{Var}[f_n(t)]$ as a function of $h$.
4. Determine the mean square error $\mathrm{mse}[f_n(t)]$ from 2 and 3 as a function of $h$.
5. For what value of $h$, as a function of $n$, is $\mathrm{mse}[f_n(t)]$ smallest?
6. For the value of $h=h(n)$ obtained from 5, how fast does $\mathrm{mse}[f_n(t)]$ converge to 0, when $n$ converges to $\infty$?

Attempt:

1. By linearity of the expectation, identical distribution of $x_1, ..., x_n$, the law of the unconscious statistician and the change of variables $u=(t−x)/h$,
\begin{align*}
\mathbb{E}[f_n(t)]&=\frac{1}{n}\sum_{i=1}^n \mathbb{E}\left[\frac{1}{h}k\left(\frac{t-x_i}{h}\right)\right]\\
&=\mathbb{E}\left[\frac{1}{h}k\left(\frac{t-x}{h}\right)\right]\\
&=\int_{\mathbb{R}}\frac{1}{h}k\left(\frac{t-x}{h}\right)f(x)\mathrm{d}x\\
&=\int_{\mathbb{R}}\frac{1}{h}k(u)f(t-hu)h\mathrm{d}u\\
&=\int_{\mathbb{R}}k(u)f(t-hu)\mathrm{d}u.
\end{align*}

2. From $f\in C^m$, it follows that $$f(t-hu)=\sum_{l=0}^m \frac{f^{(l)}(t)}{l!} (-hu)^l+o((hu)^m).$$
Then from 1 and linearity of integration,
\begin{align*}
\mathbb{E}[f_n(t)]&=\int_{\mathbb{R}}k(u)\left(\sum_{l=0}^m \frac{f^{(l)}(t)}{l!} (-hu)^l+o((hu)^m)\right)\mathrm{d}u\\
&=\sum_{l=0}^m\int_{\mathbb{R}}k(u)\frac{f^{(l)}(t)}{l!}\mathrm{d}u+\int_{\mathbb{R}}k(u)o((hu)^m)\mathrm{d}u.
\end{align*}
From the given conditions on $k$, the $l=0$ term reads
$$\int_{\mathbb{R}} k(u)f(t)\mathrm{d}u=f(t)\int_{\mathbb{R}} k(u) \mathrm{d}u=f(t)1=f(t).$$
The $1\leq l<m$ terms are
$$\int_{\mathbb{R}} k(u)\frac{f^{(l)}(t)}{l!} (-hu)^l\mathrm{d}u=\frac{f^{(l)}(t)(-h)^l}{l!}\int_{\mathbb{R}} k(u)u^l\mathrm{d}u=0.$$
Finally, the $l=m$ term is $$ \frac{f^{(m)}(t)(-h)^m}{m!}\int_{\mathbb{R}} k(u)u^m\mathrm{d}u<\infty.$$
The remainder term is simply,
$$\int_\mathbb{R} k(u) o((uh)^m)\mathrm{d}u = o(h^m)\int_\mathbb{R} k(u) u^m\mathrm{d}u = o(h^m).$$
Putting it all together:
$$\mathbb{E}[f_n(t)] = f(t) + \frac{f^{(m)}(t)(-h)^m}{m!} \int_{\mathbb{R}}k(u)u^m \mathrm{d}u + o(h^m),$$ and thus $$\mathbb{E}[f_n(t)]-f(t)=\frac{f^{(m)}(t)(-h)^m}{m!} \int_{\mathbb{R}}k(u)u^m \mathrm{d}u + o(h^m)=A(t)h^m+o(h^m),$$ where $A(t)=\frac{f^{(m)}(t)(-1)^m}{m!} \int_{\mathbb{R}}k(u)u^m \mathrm{d}u<\infty.$
Note that this solution assumes that $h\neq h(u)$ and that $\int_{\mathbb{R}} k(u)u^l\mathrm{d}u=0$ for all $1\leq l < m$. Is this reasonable?

3. This solution was partly suggested by someone else. By independence and the change of variables in 2,
\begin{align*}
\mathrm{Var}[f_n(t)]&=\mathrm{Var}\left[\frac{1}{n}\sum_{i=1}^n \frac{1}{h}k\left(\frac{t-x_i}{h}\right)\right]\\
&=\frac{1}{n^2h^2}\sum_{i=1}^n\mathrm{Var}\left[k\left(\frac{t-x_i}{h}\right)\right]\\
&=\frac{1}{nh^2}\mathrm{Var}\left[k\left(\frac{t-x}{h}\right)\right]\\
&=\frac{1}{nh^2}\left(\mathbb{E}\left[k^2\left(\frac{t-x}{h}\right)\right]-\left(\mathbb{E}\left[k\left(\frac{t-x}{h}\right)\right]\right)^2\right)\\
&=\frac{1}{nh^2}\left(h\int_{\mathbb{R}}k^2(u)f(t-hu)\mathrm{d}u-\left(h\int_{\mathbb{R}}k(u)f(t-hu)\mathrm{d}u\right)^2\right)\\
&=\frac{1}{nh^2}\left(h\int_{\mathbb{R}}k^2(u)\left(f(t)-(hu)f^{(1)}(t)+o(hu)\right)\mathrm{d}u +O(h^2) \right)\\
&=\frac{1}{nh^2}\left(h\cdot f(t)\int_{\mathbb{R}}k^2(u)\mathrm{d}u - h^2 f^{(1)}(t)\int_{\mathbb{R}}k^2(u)u\mathrm{d}u + o(h^2) +O(h^2) \right)\\
&=\frac{1}{nh^2}\left(h\cdot f(t)\int_{\mathbb{R}}k^2(u)\mathrm{d}u +O(h^2) \right)\\
&=\frac{f(t)}{nh}\int_{\mathbb{R}}k^2(u)\mathrm{d}u +o\left(\frac{1}{nh}\right),
\end{align*}
since $o(h^2) + O(h^2) = O(h^2)$ (2nd last equality) and $O(h^2) \cdot \frac{1}{nh^2} = O(h) \cdot\frac{1}{nh} = o(1)\cdot \frac{1}{nh} = o\left(\frac{1}{nh}\right)$ (last equality). What happens with $- h^2 f^{(1)}(t)\int_{\mathbb{R}}k^2(u)u\mathrm{d}u$ in the third to the second last equality?

 

[schniefen]

My assertion about $h\neq h(u)$ follows from...if $h=h(u)$ is in fact true, then $o((hu)^m)=u^mo(h^m)$ and thus $\int_\mathbb{R} k(u) o((uh)^m)\mathrm{d}u = \int_\mathbb{R} o(h^m)k(u) u^m\mathrm{d}u.$ However, is it possible to just pull the $o(h^m)$ out of the integral and use the condition that $\int_\mathbb{R} k(u) u^m\mathrm{d}u < \infty$?

 

[Office_Shredder]

I'm still lost on the first part. Is the $u$ in part 1 supposed to be the same $u$ that $h$ is a function of? You assumed it was constant when doing the change of variables.

 

[schniefen]

Thanks for the reply.

I'm still lost on the first part. Is the $u$ in part 1 supposed to be the same $u$ that $h$ is a function of?

As I look at the exercise again, I'm unsure if it isn't supposed to be $h=h(n)$ as specified in part 6. Would this make sense?

You assumed it was constant when doing the change of variables.

How come? The change of variables is $u=(t−x)/h$ and it implies $h\mathrm{d}u=\mathrm{d}x$.

 

[Office_Shredder]

How come? The change of variables is $u=(t−x)/h$ and it implies $h\mathrm{d}u=\mathrm{d}x$.

If h is a function of u (and hence of x also?) then this isn't true. This doesn't make a ton of sense as a set up, so I think your conjecture that the u is a typo seems right to me.

 

[schniefen]

If h is a function of u (and hence of x also?) then this isn't true. This doesn't make a ton of sense as a set up, so I think your conjecture that the u is a typo seems right to me.

Now, does $o((hu)^m)=u^mo(h^m)$ still hold if $h\neq h(u)$?

 

[schniefen]

Now, does $o((hu)^m)=u^mo(h^m)$ still hold if $h\neq h(u)$?

If one treats $u=u(n)$ (since $u=(t-x)/h(n)$), then one could pull out factors of $u$ if the variable inside $o$ suddenly changes from $hu$ to $n$. From the Taylor expansion, it is assumed it is $hu$, i.e. there is a function of $hu$ that goes to $0$ faster than $(hu)^m$ as $hu\to 0$. I guess one would have to find out what $hu\to 0$ would be equivalent to in terms of $n$. It is isn't stated in the exercise, but to obtain reasonable estimations, $h(n)\to 0$ as $n\to \infty$. Does this sound legitimate, @Office_Shredder? I have skimmed through different lecture notes about this and in many of them a similar calculation as above is carried out.

 

[Office_Shredder]

Now, does $o((hu)^m)=u^mo(h^m)$ still hold if $h\neq h(u)$?

Yes. In fact, you are going to run into more problems if h is a function of u, since now if h is small you have to consider what is happening to u. For example if $h=1/u$ the thing on the left is not a very impressive statement, and is not equal to the thing on the right.

 

[schniefen]

Yes. In fact, you are going to run into more problems if h is a function of u, since now if h is small you have to consider what is happening to u. For example if $h=1/u$ the thing on the left is not a very impressive statement, and is not equal to the thing on the right.

Why is $o((hu)^m)=o\left(\left(\frac{1}{u}u\right)^m\right)=o(1)$ not equal to $u^mo(h^m)=u^mo\left(\frac{1}{u^m}\right)=o(1)$?

 

[schniefen]

Summary:: What is the expectation and variance of $f_n(t)=\frac{1}{n}\sum_{i=1}^n \frac{1}{h}k\left(\frac{t-x_i}{h}\right)$?

This is a question from a mathematical statistics textbook, used at the first and most basic mathematical statistics course for undergraduate students. This exercise follows the chapter on nonparametric inference. An attempt at a solution is given. Any help is appreciated.

Exercise:

Suppose $x_1, ..., x_n$ are independent and identically distributed (i.i.d.) observations of a random variable $X$ with unknown distribution function $F$ and probability density function $f\in C^m$, for some $m>1$ fixed. Let
$$f_n(t)=\frac{1}{n}\sum_{i=1}^n \frac{1}{h}k\left(\frac{t-x_i}{h}\right)$$
be a kernel estimator of $f$, with $k\in C^{m+1}$ a given fixed function such that $k\geq 0$, $\int_{\mathbb{R}} k(u)\mathrm{d}u=1$, $\mathrm{supp} (k)=[-1,1]$ and bandwidth $h=\displaystyle\rlap{——}h(u)h(n)$ (for the time being unspecified).
1. Show that $\mathbb{E}[f_n(t)]=\int_{\mathbb{R}} k(u) f(t-hu)\mathrm{d}u$.
2. Make a series expansion of $f$ around $t$ in terms of $hu$ in the expression for $\mathbb{E}[f_n(t)]$. Suppose that $k$ satisfies $\int_{\mathbb{R}} k(u)\mathrm{d}u=1$, $\int_{\mathbb{R}} k(u)u^l\mathrm{d}u=0$ for all $1<l<m$ and $\int_{\mathbb{R}} k(u)u^m\mathrm{d}u<\infty$. Determine the bias $\mathbb{E}[f_n(t)]-f(t)$ as a function of $h$.
3. Suppose that $\mathrm{Var}[k(X_1)]<\infty$ and determine $\mathrm{Var}[f_n(t)]$ as a function of $h$.
4. Determine the mean square error $\mathrm{mse}[f_n(t)]$ from 2 and 3 as a function of $h$.
5. For what value of $h$, as a function of $n$, is $\mathrm{mse}[f_n(t)]$ smallest?
6. For the value of $h=h(n)$ obtained from 5, how fast does $\mathrm{mse}[f_n(t)]$ converge to 0, when $n$ converges to $\infty$?

Find attached a solution I have written.