接着上一节的不等式(一)-Markov与Chebyshev不等式
本节将学习Hoeffding不等式。

Hoeffding不等式

作用与Chebyshev不等式类似,但区间更紧致(增加了独立性约束)

Hoeffding不等式
\(Y_1,...Y_n\)相互独立,且\(E(Y_i) = 0\),且\(a_i \leq Y_i \leq b_i\),令\(\epsilon > 0\),则对任意\(t>0\)

\[P(\sum_{i=1}^{n}Y_i \geq \epsilon) \leq e^{-t\epsilon} \prod_{i=1}^{n}e^{t^2(b_i-a_i)^2/8} \]

证明过程如下:

\[\begin{eqnarray} P \Bigg( \sum_{i=1}^{n}Y_i \geq \epsilon \Bigg) &=& P \Bigg( t \sum_{i=1}^{n}Y_i \geq t \epsilon \Bigg ) \\ &=& P \Bigg(e^{t \sum \limits_{i=1}^{n}Y_i} \geq e^{t \epsilon} \Bigg) \\ & \leq & e^{-t \epsilon} E\Bigg(e^{t \sum \limits_{i=1}^{n}Y_i} \Bigg) \qquad (Markov 不等式) \\ &=& e^{-t \epsilon} \prod_{i=1}^{n} E \Bigg( e^{t Y_i} \Bigg) \qquad (Y_1,...Y_n相互独立,期望的性质) \end{eqnarray}\]

对于凸函数\(f(x) = e^x\),对于任意\(\alpha \in [0, 1]\),和\(x \in [a, b]\)都满足

\[f(x) \leq \alpha f(a) + (1 - \alpha)f(b) \]

如图所示:

因为\(a_i \leq Y_i \leq b_i\),即\(Y_i \in [a_i, b_i]\),令\(\alpha = \frac{b_i-Y_i }{b_i - a_i} \in [0, 1]\),则

$$e^{tY_i} \leq \frac{(b_i - Y_i)}{(b_i - a_i)} e^{ta_i} + \frac{(Y_i - a_i)}{(b_i - a_i)} e^{tb_i} $$

又因为\(E(Y_i) = 0\),对两边同时取期望,

\[E(e^{tY_i}) \leq \frac{b_i - E(Y_i)}{(b_i - a_i)} e^{ta_i} + \frac{E(Y_i) - a_i}{(b_i - a_i)} e^{tb_i} = \frac{b_i}{(b_i - a_i)} e^{ta_i} - \frac{a_i}{(b_i - a_i)} e^{tb_i} \]

\(u = t(b_i - a_i)\),和\(\gamma = -a_i/b_i - a_i\),可以将上式右边写作$$e^{ -\gamma u + log(1 - \gamma + \gamma e^u)}$$

记为\(e^{g(u)}\),我们对\(g(u)\)作泰勒展开

\[g(u) = \frac{g(0)}{0!}(u-0)^0 + \frac{g^{'}(0)}{1!}(u-0)^1 + \frac{g^{''}(0)}{2!}(u-0)^2 + o(u^2) \]

易得\(g(0) = g^{'}(0) = 0\),而

\[g^{''}(u) = \frac{\gamma(1-\gamma)e^u}{(1- \gamma + \gamma e^u)^2} \]

得到\(g^{''}(0) = \gamma(1-\gamma) \leq 1/4\),带入泰勒展开式,得到

\[g(u) = \frac{g^{''}(0)}{2!}(u-0)^2 + o(u^2) \leq \frac{1/4}{2!}(u-0)^2 = \frac{u^2}{8} = \frac{1}{8}t^2(b_i - a_i)^{2} \]

因此,

\[E(e^{tY_i}) \leq e^{g(u)} \leq e^{t^2(b_i - a_i)^{2}/8} \]

不等式得证。

Hoeffding不等式
\(Y_1,...Y_n\)相互独立,且\(E(Y_i) = 0\),且\(a_i \leq Y_i \leq b_i\),令\(\epsilon > 0\),则对任意\(t>0\)

\[P(\sum_{i=1}^{n}Y_i \geq \epsilon) \leq e^{-t\epsilon} \prod_{i=1}^{n}e^{t^2(b_i-a_i)^2/8} \]

\(X_1,...X_n \backsim Bernoulli(p)\),则对任意\(\epsilon > 0\),有

\[P(|\overline{X} - p| > \epsilon) \leq 2e^{-2n\epsilon^2} \qquad (2) \]

其中 \(\overline{X_n} = \frac{1}{n}\sum_{i=1}^{n}X_i\)

对于(2)式证明如下:
\(Y_i = \frac{1}{n}(X_i - p)\),有\(E(Y_i) = 0\),且有

\[\sum_{i=1}^{n} Y_i = \frac{1}{n} \sum_{i=1}^{n} (X_i - p) = \frac{1}{n}\sum_{i=1}^{n} X_i - p = \overline{X_n} - p \]

又有\(a \leq Y_i \leq b\),因为\(X_1,...X_n \backsim Bernoulli(p)\),所以\(X_i\)的取值只能是\(0\)\(1\),因此\(a = - p/n\)\(b = (1-p)/n\),那么\((b-a)^2 = 1/n^2\)

\[P(\overline{X} - p > \epsilon) = p(\sum_{i=1}^{n} Y_i > \epsilon) \leq e^{-t\epsilon}e^{t^2/(8n)} \qquad (t>0) \]

\(t = 4n\epsilon\),我们得到$$P(\overline{X} - p > \epsilon) \leq e{-2n\epsilon2}$$
同理可得$$P(\overline{X} - p < -\epsilon) \leq e{-2n\epsilon2}$$
因此

\[P(|\overline{X} - p| > \epsilon) \leq 2e^{-2n\epsilon^2} \]

不同于其他的不等式是在收敛的情况下等式成立,Hoeffding不等式对于任意n都成立。

Hoeffding不等式应用

Reference

  1. 《All of Statistics: A Concise Course in Statistical Inference》by Wasserman, Larry
  2. 不等式 by 中科院 卿来云老师课件