当前位置:   article > 正文

切诺夫界(Chernoff Bound)形式及其证明

chernoff bound

马尔可夫不等式

设随机变量X的取值为非负数,马尔可夫不等式形式为:
P ( X ≥ ξ ) ≤ E ( X ) ξ P(X \ge \xi) \le \frac{E(X)}{\xi} P(Xξ)ξE(X)

p r o o f . proof. proof.

设非负随机变量 X X X的概率密度函数为 f ( x ) f(x) f(x)
E ( X ) = ∫ 0 ∞ x f ( x ) d x = ∫ 0 ξ x f ( x ) d x + ∫ ξ ∞ x f ( x ) d x ≥ ∫ ξ ∞ x f ( x ) d x ≥ ∫ ξ ∞ ξ f ( x ) d x = ξ P ( X ≥ ξ )

E(X)=0xf(x)dx=0ξxf(x)dx+ξxf(x)dxξxf(x)dxξξf(x)dx=ξP(Xξ)
E(X)=0xf(x)dx=0ξxf(x)dx+ξxf(x)dxξxf(x)dxξξf(x)dx=ξP(Xξ)
所以 E ( X ) ≥ ξ P ( X ≥ ξ ) E(X) \ge \xi P(X \ge \xi) E(X)ξP(Xξ)

经变换得到结论:
P ( X ≥ ξ ) ≤ E ( X ) ξ P(X \ge \xi) \le \frac{E(X)}{\xi} P(Xξ)ξE(X)

切诺夫界

X = ∑ i = 1 n X i X=\sum_{i=1}^{n}X_i X=i=1nXi,其中 X 1 , … , X n X_1, \dots,X_n X1,,Xn之间相互独立,并且 P ( X i = 1 ) = p i ,   P ( X i = 0 ) = 1 − p i P(X_i=1)=p_i,\ P(X_i=0)=1-p_i P(Xi=1)=pi, P(Xi=0)=1pi

μ = E ( X ) = ∑ i = 1 n p i \mu = E(X)=\sum_{i=1}^{n}p_i μ=E(X)=i=1npi

  1. Upper Tail:
    ∀ δ > 0 ,   P ( X ≥ ( 1 + δ ) μ ) ≤ e − δ 2 2 + δ μ \forall \delta \gt 0,\ P(X \ge (1+\delta)\mu) \le e^{-\frac{\delta^2}{2+\delta}\mu} δ>0, P(X(1+δ)μ)e2+δδ2μ

  2. Lower Tail:
    ∀ 0 < δ < 1 ,   P ( X ≤ ( 1 − δ ) μ ) ≤ e − δ 2 2 μ \forall 0 \lt \delta \lt 1,\ P(X \le (1-\delta)\mu) \le e^{-\frac{\delta^2}{2}\mu} ∀0<δ<1, P(X(1δ)μ)e2δ2μ

p r o o f . proof. proof.

由马尔可夫不等式可知,设 X X X为一非负随机变量,对于 ∀ s > 0 \forall s \gt 0 s>0
P ( X ≥ a ) = P ( e s X ≥ e s a ) ≤ E ( e s X ) e s a P(X \ge a)=P(e^{sX} \ge e^{sa}) \le \frac{E(e^{sX})}{e^{sa}} P(Xa)=P(esXesa)esaE(esX)
类似可知
P ( X ≤ a ) = P ( e − s X ≥ e − s a ) ≤ E ( e − s X ) e − s a P(X \le a)=P(e^{-sX} \ge e^{-sa}) \le \frac{E(e^{-sX})}{e^{-sa}} P(Xa)=P(esXesa)esaE(esX)
现令 M X ( s ) = E ( e s X ) M_X(s)=E(e^{sX}) MX(s)=E(esX)

l e m m a   1 : lemma\ 1: lemma 1:

设随机变量 X = ∑ i = 1 n X i X=\sum_{i=1}^{n}X_i X=i=1nXi

M X ( s ) = E ( e s X ) = E ( e s ∑ i = 1 n X i ) = E ( ∏ i = 1 n e s X i ) = ∏ i = 1 n E ( e s X i ) = ∏ i = 1 n M X i ( s ) M_X(s)=E(e^{sX})=E(e^{s\sum_{i=1}^{n}X_i})=E(\prod_{i=1}^{n}e^{sX_i})=\prod_{i=1}^{n}E(e^{sX_i})=\prod_{i=1}^{n}M_{X_i}(s) MX(s)=E(esX)=E(esi=1nXi)=E(i=1nesXi)=i=1nE(esXi)=i=1nMXi(s)

即:
M X ( s ) = ∏ i = 1 n M X i ( s ) M_X(s)=\prod_{i=1}^{n}M_{X_i}(s) MX(s)=i=1nMXi(s)
l e m m a   2 lemma\ 2 lemma 2

设随机变量 X ∼ B ( p , 1 ) X \sim B(p,1) XB(p,1)

M X ( s ) = p e s + ( 1 − p ) = 1 + p ( e s − 1 ) M_X(s)=pe^{s}+(1-p)=1+p(e^{s}-1) MX(s)=pes+(1p)=1+p(es1)

因为 e x ≥ 1 + x e^x \ge 1+x ex1+x,得到结论
M X ( s ) ≤ e p ( e s − 1 ) M_X(s) \le e^{p(e^{s}-1)} MX(s)ep(es1)
上尾证明:
P ( X ≤ ( 1 + δ ) μ ) ≤ E ( e s X ) e s ( 1 + δ ) μ = M X ( s ) e s ( 1 + δ ) μ = ∏ i = 1 n M X i ( s ) e s ( 1 + δ ) μ ≤ ∏ i = 1 n e p i ( e s − 1 ) e s ( 1 + δ ) μ = e ( e s − 1 ) ∑ i = 1 n p i e s ( 1 + δ ) μ = e ( e s − 1 ) μ e s ( 1 + δ ) μ = [ e ( e s − 1 ) e s ( 1 + δ ) ] μ

P(X(1+δ)μ)E(esX)es(1+δ)μ=MX(s)es(1+δ)μ=i=1nMXi(s)es(1+δ)μi=1nepi(es1)es(1+δ)μ=e(es1)i=1npies(1+δ)μ=e(es1)μes(1+δ)μ=[e(es1)es(1+δ)]μ
P(X(1+δ)μ)es(1+δ)μE(esX)=es(1+δ)μMX(s)=es(1+δ)μi=1nMXi(s)es(1+δ)μi=1nepi(es1)=es(1+δ)μe(es1)i=1npi=es(1+δ)μe(es1)μ=[es(1+δ)e(es1)]μ
s = l n ( 1 + δ ) s=ln(1+\delta) s=ln(1+δ)

可得:
P ( X ≤ ( 1 + δ ) μ ) ≤ ( e δ ( 1 + δ ) ( 1 + δ ) ) μ P(X \le (1+\delta)\mu) \le \left(\frac{e^\delta}{(1+\delta)^{(1+\delta)}}\right)^\mu P(X(1+δ)μ)((1+δ)(1+δ)eδ)μ
又有: l n [ ( e δ ( 1 + δ ) ( 1 + δ ) ) μ ] = μ ( δ − ( 1 + δ ) l n ( 1 + δ ) ) ln\left[\left(\frac{e^\delta}{(1+\delta)^{(1+\delta)}}\right)^\mu\right]=\mu(\delta-(1+\delta)ln(1+\delta)) ln[((1+δ)(1+δ)eδ)μ]=μ(δ(1+δ)ln(1+δ)),且 ∀ x > 0 ,   l n ( 1 + x ) ≥ x 1 + x 2 \forall x \gt 0,\ ln(1+x) \ge \frac{x}{1+\frac{x}{2}} x>0, ln(1+x)1+2xx

可得:
μ ( δ − ( 1 + δ ) l n ( 1 + δ ) ) ≤ − δ 2 2 + δ μ \mu(\delta-(1+\delta)ln(1+\delta)) \le -\frac{\delta^2}{2+\delta}\mu μ(δ(1+δ)ln(1+δ))2+δδ2μ
即:
( e δ ( 1 + δ ) ( 1 + δ ) ) μ ≤ e − δ 2 2 + δ μ \left(\frac{e^\delta}{(1+\delta)^{(1+\delta)}}\right)^\mu \le e^{-\frac{\delta^2}{2+\delta}\mu} ((1+δ)(1+δ)eδ)μe2+δδ2μ
所以上尾得证:
P ( X ≤ ( 1 + δ ) μ ) ≤ e − δ 2 2 + δ μ P(X \le (1+\delta)\mu) \le e^{-\frac{\delta^2}{2+\delta}\mu} P(X(1+δ)μ)e2+δδ2μ

下尾类似:

s = l n ( 1 − δ ) s=ln(1-\delta) s=ln(1δ)

且使用不等式: ∀ 0 < δ < 1 ,   l n ( 1 + δ ) ≥ − δ + δ 2 2 \forall 0 \lt \delta \lt 1,\ ln(1+\delta) \ge -\delta + \frac{\delta^2}{2} ∀0<δ<1, ln(1+δ)δ+2δ2

非伯努利切诺夫界

在上述切诺夫界中,假设了 X i ∼ B ( p i , 1 ) X_i \sim B(p_i,1) XiB(pi,1),对于不服从伯努利分布的随机变量,同样有

X 1 , … , X n X_1, \dots, X_n X1,,Xn相互独立,其中 a ≤ X i ≤ b a \le X_i \le b aXib,令 X = ∑ i = 1 n X i ,   μ = E ( X ) ,   ∀ δ > 0 X=\sum_{i=1}^{n}X_i,\ \mu=E(X),\ \forall \delta \gt 0 X=i=1nXi, μ=E(X), δ>0有:

上尾:
P ( X ≥ ( 1 + δ ) μ ) ≤ e − 2 δ 2 μ 2 n ( b − a ) 2 P(X \ge (1+\delta)\mu) \le e^{-\frac{2\delta^2\mu^2}{n(b-a)^2}} P(X(1+δ)μ)en(ba)22δ2μ2
下尾:
P ( X ≤ ( 1 − δ ) μ ) ≤ e − δ 2 μ 2 n ( b − a ) 2 P(X \le (1-\delta)\mu) \le e^{-\frac{\delta^2\mu^2}{n(b-a)^2}} P(X(1δ)μ)en(ba)2δ2μ2

声明:本文内容由网友自发贡献,不代表【wpsshop博客】立场,版权归原作者所有,本站不承担相应法律责任。如您发现有侵权的内容,请联系我们。转载请注明出处:https://www.wpsshop.cn/w/知新_RL/article/detail/81506
推荐阅读
相关标签
  

闽ICP备14008679号