是否存在两个分布之间的Hellinger距离的无偏估计量?


20

在一个观察密度为的分布的分布的环境中,我想知道是否存在一个对密度为另一分布即 的Hellinger距离的无偏估计量(基于)。X1,,XnfXif0

H(f,f0)={1Xf(x)f0(x)dx}1/2.

4
因此f0是已知的且固定的。但是f是已知的还是参数家族的,或者您是在非参数框架中使用您从样本中获得的关于f的全部知识来进行此操作的?我认为尝试答案时会有所不同。
Michael R. Chernick 2012年

3
@MichaelChernick:假设您对所有知识f就是样本X1,,Xn
西安

2
我不认为它已经过计算(如果存在)。如果存在,那么AIC会有一个失散的兄弟。

4
如果假设ff0是离散的,则对这个问题进行攻击似乎是可行的。这导致一个明显的估计量(计算EDF和之间的Hellinger距离f0)。自举(理论上,不是通过模拟!)将为我们提供可能的偏差的处理方法,以及减少(甚至消除)偏差的方法。我希望通过平方距离而不是距离本身能够成功,因为它在数学上更容易处理。离散的假设f在应用中没有问题。无论如何,离散的空间f是一个密集子集。
ub

2
想到罗森布拉特的证明,没有 “善意”无偏估计量。我们可以克服这个问题并得到H f f 0)的无偏估计量吗?我不知道。fH(f,f0)
2012年

Answers:


5

对于任何合理的广义非参数分布类别中的f,均不存在H 2的无偏估计。HH2f

我们可以用以下简单的论证来证明这一点

比克尔和莱曼(1969)。凸族的无偏估计。《数学统计年鉴》,40(5)1523–1535。(项目欧几里得

固定一些分布FG,其密度分别为f 0fg。让ħ ˚F 分别表示ħ˚F ˚F 0,并让ħX来的一些估计ħ ˚F 基于Ñ IID样本X ˚FF0FGf0fgH(F)H(f,f0)H^(X)H(F)nXiF

假设ħ是无偏从以下形式的任何分布样品 中号α= α ˚F + 1 - α ģ 但是然后 Q α H^

Mα:=αF+(1α)G.
从而使Qα必须是在多项式α的度最多ñ
Q(α)=H(Mα)=x1xnH^(X)dMα(x1)dMα(xn)=x1xnH^(X)[αdF(x1)+(1α)dG(x1)][αdF(xn)+(1α)dG(xn)]=αnEXFn[H^(X)]++(1α)nEXGn[H^(X)],
Q(α)αn

现在,让我们专门研究一个合理的情况,并证明相应的不是多项式。Q

是一些分布这对恒定密度[ - 1 1 ]˚F 0X = c ^对所有| x | 1。(其行为之外的范围并不重要)设˚F是只在支持的一些分布[ - 1 0 ],和G ^一些分布只支持上[ 0 1 ]F0[1,1]f0(x)=c|x|1F[1,0]G[0,1]

现在 其中˚F=[R

Q(α)=H(mα,f0)=1Rmα(x)f0(x)dx=110cαf(x)dx01c(1α)g(x)dx=1αBF1αBG,
,对于BG同样。注意,对于任何具有密度的分布FGBF>0BG>0BF:=Rf(x)f0(x)dxBGBF>0BG>0FG

1αBF1αBGH^HMα

1αBF1αBGH2Mα

This excludes pretty much all reasonable nonparametric classes of distributions, except for those with densities bounded below (an assumption nonparametric analyses sometimes make). You could probably kill those classes too with a similar argument by just making the densities constant or something.


13

I don't know how to construct (if it exists) an unbiased estimator of the Hellinger distance. It seems possible to construct a consistent estimator. We have some fixed known density f0, and a random sample X1,,Xn from a density f>0. We want to estimate

H(f,f0)=1Xf(x)f0(x)dx=1Xf0(x)f(x)f(x)dx
=1E[f0(X)f(X)],
where Xf. By the SLLN, we know that
11ni=1nf0(Xi)f(Xi)H(f,f0),
almost surely, as n. Hence, a resonable way to estimate H(f,f0) would be to take some density estimator fn^ (such as a traditional kernel density estimator) of f, and compute
H^=11ni=1nf0(Xi)fn^(Xi).

3
@Zen: Good point! I consider this answer as the answer because it made me realise H sounds very much like a standard deviation, for which there exists no unbiased estimator. As for the variance of H^n2, no worries: E[(f0(X)/f(X))2]=1 implies that this estimator has a finite variance.
Xi'an

1
Thanks for the clarification about the variance of the estimator, Xi'an!
Zen

2
Some work on other consistent estimators: (a) arxiv.org/abs/1707.03083 and related work based on k-NN density estimators; (b) arxiv.org/abs/1402.2966 based on correcting kernel density estimates; (c) ieeexplore.ieee.org/document/5605355 based on a connection to classification. (Many of these are based on samples from both f and f0, because that's the work I knew about offhand, but I think there are variants for known f0.)
Dougal
By using our site, you acknowledge that you have read and understand our Cookie Policy and Privacy Policy.
Licensed under cc by-sa 3.0 with attribution required.