Is the condition that asymptotic bias and asymptotic variance goes to zero for infinite samples sufficient to guarantee the pointwise consistency of an estimator based on plug-in kernel density estimator (KDE)?


I am interested in the estimation of a log of ratio between two continuous dentisy functions $f=\log \frac{p_1(x)}{p_2(x)}$ with kernel density estimation (KDE).

Assume that we have two sets of samples $\{x_i^{(1)} \}_{i=1}^{N_1}$ and $\{x_j^{(2)} \}_{j=1}^{N_2}$ from two continuous $D$-dimensional density functions $p_1(x)$ and $p_2(x)$, respectively.

Then, a logarithm of a ratio of densities $\log\frac{p_1(x)}{p_2(x)}=\log p_1(x) - \log p_2 (x)$ at a test point x is estimated with KDEs as plug-in density estimates. The KDEs are defined as $$ \widehat{p}_1(x)=\frac{1}{N_1}\sum_{i=1}^{N_1}K_h(x, x_i^{(1)}),\quad \widehat{p}_2(x)=\frac{1}{N_2}\sum_{j=1}^{N_2}K_h(x, x_j^{(2)}), $$ where $K_h(\cdot, \cdot)$ is a kernel function with a bandwidth $h$. Let's just assume a Gaussian kernel. The resulting estimator is $\widehat{f}(x) =\log \frac{\widehat{p}_1(x)}{\widehat{p}_2(x)}$. I can assume some conditions for $p_1(x)$, $p_2(x)$, and $K_h(\cdot, \cdot)$ so that the logarithm is well defined.

I have obtained the bias, $\mathbb{E}[\widehat{f}]-f$, and the variance, $\mathbb{E}[(\widehat{f}(x)-\mathbb{E}[\widehat{f}(x)])^2]$, of the estimator $\widehat{f}(x)$, both of which goes to zero with the conditions $N\to\infty$, $h\to0$, $Nh^D\to\infty$.

I have found articles which say vanishing bias and variance are enough for pointwise consistency in KDE case. 1) Section 2.6 in Hansen's lecture note 2) Theorem 1 of Wied and Weissbach, 2012.

Therefore, currently, I suspect that the condition of bias and variance going to zero is sufficient for pointwise consistency of my estimator $\widehat{f}$. Is this correct? or if not, what kind of argument is necessary to prove that the estimator is pointwise consistent?


Your Answer

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

Browse other questions tagged or ask your own question.