I was wondering whether there is an equivalent to Kernel Density Estimation to estimate nonparametrically the logarithm of a density. Or if there is any nonparametric method for that. (Taking the logarithm of the kernel density estimate seems suboptimal.)
Could you point towards references dealing with this?
EDIT. With suboptimal I mean that $\log \hat f_h(x)$ has some extra bias with respect to the bias of $\hat f_h(x)$, although it is still asymptotically unbiased. By a Taylor expansion, $$ \log \hat f_h(x) = \log(f(x) + \hat f_h(x) - f(x)) \approx \log f(x) + \frac{\hat f_h(x) - f(x)}{f(x)} - \frac{1}{2}\frac{(\hat f_h(x) - f(x))^2}{f(x)^2}. $$ Then, taking the expectation and pluging-in the bias and MSE expressions for $\hat f_h(x)$: \begin{eqnarray*} \mathbb{E}[\log \hat f_h(x)] &\approx& \log f(x) + \frac{\mu_2(K)f''(x)}{2f(x)} h^2 - \frac{1}{2f(x)^2} \Big(\frac{1}{4}\mu_2(K)^2f''(x)^2h^4 + \frac{R(K)f(x)}{nh}\Big)\\ & = &\log f(x) + O(h^2 + (nh)^{-1}) \end{eqnarray*} which is larger than the bias of $\hat f_h(x)$, $\mathbb{E}[\hat f_h(x)] = f(x) +O(h^2)$.