Jensen不等式
引言
?概率不等式是概率論和數(shù)理統(tǒng)計(jì)的理論研究中的重要工具,對(duì)于概率極限理論和統(tǒng)計(jì)大樣本理論,幾乎所有重要結(jié)果的論證是借助于概率不等式的巧妙應(yīng)用,Jensen\mathrm{Jensen}Jensen不等式和證明,并應(yīng)用其帶來(lái)解決一些相關(guān)問(wèn)題。
Jensen\mathrm{Jensen}Jensen不等式不同形式
?Jensen\mathrm{Jensen}Jensen不等式的形式有很多種,標(biāo)準(zhǔn)形式的如下:
Jensen\mathrm{Jensen}Jensen不等式: 如果f(x)f(x)f(x)為連續(xù)實(shí)值凸函數(shù),且x1≤x2≤?≤xnx_1\le x_2\le \cdots \le x_nx1?≤x2?≤?≤xn?,∑i=1nλi=1\sum\limits_{i=1}^n\lambda_i=1i=1∑n?λi?=1,λi≥0\lambda_i \ge0λi?≥0,i=1,2?,ni=1,2\cdots,ni=1,2?,n,則有∑i=1nλif(xi)≥f(∑i=1nλixi)\sum\limits_{i=1}^n\lambda_i f(x_i)\ge f(\sum\limits_{i=1}^n\lambda_i x_i)i=1∑n?λi?f(xi?)≥f(i=1∑n?λi?xi?)如果f(x)f(x)f(x)為連續(xù)實(shí)值凹函數(shù),則有∑i=1nλif(xi)≤f(∑i=1nλixi)\sum\limits_{i=1}^n\lambda_i f(x_i)\le f(\sum\limits_{i=1}^n\lambda_i x_i)i=1∑n?λi?f(xi?)≤f(i=1∑n?λi?xi?)
?在概率論中Jensen\mathrm{Jensen}Jensen不等式有:離散型,連續(xù)型,條件期望型和中位數(shù)型等形式
Jensen\mathrm{Jensen}Jensen不等式1: 設(shè)f(x)f(x)f(x)是區(qū)間[a,b][a,b][a,b]上的凸函數(shù),XXX是取值于[a,b][a,b][a,b]上子集AAA的離散型隨機(jī)變量,則有如下兩個(gè)結(jié)論成立
(1)E(f(X))≥f(E(X))\mathbb{E}(f(X))\ge f(\mathbb{E}(X))E(f(X))≥f(E(X));
(2)如果f(X)f(X)f(X)是嚴(yán)格凸的,則不等式中等號(hào)當(dāng)且僅當(dāng)P(X=E(X))=1P(X=\mathbb{E}(X))=1P(X=E(X))=1時(shí)成立。
證明:
(1)對(duì)XXX取值的個(gè)數(shù)進(jìn)行數(shù)學(xué)歸納法證明,首先對(duì)于兩點(diǎn)分布:X~{p(x1),p(x2)}X \sim \{p(x_1),p(x_2)\}X~{p(x1?),p(x2?)}簡(jiǎn)記p1=p(x1)p_1=p(x_1)p1?=p(x1?),p2=p(x2)p_2=p(x_2)p2?=p(x2?)。注意到p1=1?p2p_1=1-p_2p1?=1?p2?,則有E(f(X))=p1f(x1)+p2f(x2)≥f(p1x1+p2x2)=f(E(X))\mathbb{E}(f(X))=p_1f(x_1)+p_2f(x_2)\ge f(p_1x_1+p_2x_2)=f(\mathbb{E}(X))E(f(X))=p1?f(x1?)+p2?f(x2?)≥f(p1?x1?+p2?x2?)=f(E(X))假設(shè)XXX的值域AAA中元素個(gè)數(shù)為n?1(n≥2)n-1(n \ge 2)n?1(n≥2),A={x1,x2,?,xn?1}A=\{x_1,x_2,\cdots,x_{n-1}\}A={x1?,x2?,?,xn?1?}時(shí),結(jié)論(1)式成立,則對(duì)AAA中元素個(gè)數(shù)為n(n≥2)n(n\ge 2)n(n≥2),A=(x1,x2,?,xn)A=(x_1,x_2,\cdots,x_n)A=(x1?,x2?,?,xn?)時(shí),簡(jiǎn)記pi=p(xi)p_i=p(x_i)pi?=p(xi?),pi′=pi1?pn,i=1,2,?,np_i^{\prime}=\frac{p_i}{1-p_n},i=1,2,\cdots,npi′?=1?pn?pi??,i=1,2,?,n,則有{p1′,p2′,?,pn?1′}\{p_1^{\prime},p_2^{\prime},\cdots,p^{\prime}_{n-1}\}{p1′?,p2′?,?,pn?1′?}是一個(gè)概率分布,從而有E(f(X))=p1f(x1)+p2f(x2)+?+pnf(xn)=(1?pn)∑i=1n?1pi′f(xi)+pnf(xn)≥(1?pn)f(∑i=1n?1pi′xi)+pnf(xn)≥f(∑i=1npixi)=f(E(X))\begin{aligned}\mathbb{E}(f(X))&=p_1f(x_1)+p_2f(x_2)+\cdots+p_nf(x_n)\\&=(1-p_n)\sum\limits_{i=1}^{n-1}p^{\prime}_i f(x_i)+p_n f(x_n)\\&\ge(1-p_n)f(\sum\limits_{i=1}^{n-1}p_i^{\prime}x_i)+p_nf(x_n)\\&\ge f(\sum\limits_{i=1}^np_ix_i)=f(\mathbb{E}(X))\end{aligned}E(f(X))?=p1?f(x1?)+p2?f(x2?)+?+pn?f(xn?)=(1?pn?)i=1∑n?1?pi′?f(xi?)+pn?f(xn?)≥(1?pn?)f(i=1∑n?1?pi′?xi?)+pn?f(xn?)≥f(i=1∑n?pi?xi?)=f(E(X))?
(2)若f(x)f(x)f(x)是嚴(yán)格凸的,則總有E(f(x))≥f(E(X))\mathbb{E}(f(x))\ge f(\mathbb{E}(X))E(f(x))≥f(E(X))成立,除非當(dāng)且僅當(dāng)P(X=E(X))=1P(X=\mathbb{E}(X))=1P(X=E(X))=1時(shí),E(f(X))=f(E(X))\mathbb{E}(f(X))=f(\mathbb{E}(X))E(f(X))=f(E(X))成立。
Jensen\mathrm{Jensen}Jensen不等式2: 設(shè)XXX是mmm維隨機(jī)向量,f(x)f(x)f(x)為定義在Rm\mathbb{R}^{m}Rm上的凸函數(shù)(m=1,2,?)(m=1,2,\cdots)(m=1,2,?),其中E(X)<∞\mathbb{E}(X)<\inftyE(X)<∞,則有
(1)E(f(X))≥f(E(X))\mathbb{E}(f(X))\ge f(\mathbb{E}(X))E(f(X))≥f(E(X));
(2)如果f(X)f(X)f(X)是嚴(yán)格凸的,則不等式中等號(hào)當(dāng)且僅當(dāng)P(X=E(X))=1P(X=\mathbb{E}(X))=1P(X=E(X))=1時(shí)成立。
證明:
(1)由于y=f(x)y=f(x)y=f(x)是Rm+1\mathbb{R}^{m+1}Rm+1中的一個(gè)凸曲面,而點(diǎn)(E(X),f(E(X)))(\mathbb{E}(X),f(\mathbb{E}(X)))(E(X),f(E(X)))在次曲面上。存在一個(gè)過(guò)此點(diǎn)的平面,使得上述曲面全在此平面上的上方。若以y=f(E(X))+c′(x?E(X))y=f(\mathbb{E}(X))+c^{\prime}(x-\mathbb{E}(X))y=f(E(X))+c′(x?E(X))記此平面的方程,則有f(x)≥f(E(X))+c′(x?E(X))f(x)\ge f(\mathbb{E}(X))+c^{\prime}(x-\mathbb{E}(X))f(x)≥f(E(X))+c′(x?E(X))因而則有E(f(X))≥f(E(X))+c′E(X?E(X))=f(E(X))\mathbb{E}(f(X))\ge f(\mathbb{E}(X))+c^{\prime}\mathbb{E}(X-\mathbb{E}(X))=f(\mathbb{E}(X ))E(f(X))≥f(E(X))+c′E(X?E(X))=f(E(X))
(2)若f(x)f(x)f(x)是嚴(yán)格凸的,則除非x=E(X)x=\mathbb{E}(X)x=E(X),總有f(x)>f(E(X))f(x)>f(\mathbb{E}(X))f(x)>f(E(X)),總有f(x)>f(E(X))+c′(x?E(X))f(x)>f(\mathbb{E}(X))+c^{\prime}(x-\mathbb{E}(X))f(x)>f(E(X))+c′(x?E(X))成立,因而當(dāng)且僅當(dāng)P(X=E(X))=1P(X=\mathbb{E}(X))=1P(X=E(X))=1時(shí)E(f(X))=f(E(X))\mathbb{E}(f(X))=f(\mathbb{E}(X))E(f(X))=f(E(X))成立。
Jensen\mathrm{Jensen}Jensen不等式3: 設(shè)f(x)f(x)f(x)是連續(xù)凸函數(shù),XXX為關(guān)于ggg為σ\sigmaσ可積的隨機(jī)變量,則f(X)f(X)f(X)關(guān)于ggg的條件期望存在,且有f(E[X∣g])≥E(f(X)∣g)f(\mathbb{E}[X|g])\ge \mathbb{E}(f(X)|g)f(E[X∣g])≥E(f(X)∣g)幾乎必然成立。
證明: 令f′(x)f^{\prime}(x)f′(x)為f(x)f(x)f(x)的右導(dǎo)數(shù),則對(duì)任意實(shí)數(shù)xxx與yyy有f′(x)(y?x)≥f(y)?f(x)f^{\prime}(x)(y-x)\ge f(y)-f(x)f′(x)(y?x)≥f(y)?f(x)以E[X∣g]\mathbb{E}[X|g]E[X∣g]及XXX代替上式中的xxx與yyy得到f′(E[X∣g])(X?E[X∣g])+f(E[X∣g])≤f(X)f^{\prime}(\mathbb{E}[X|g])(X-\mathbb{E}[X|g])+f(\mathbb{E}[X|g])\le f(X)f′(E[X∣g])(X?E[X∣g])+f(E[X∣g])≤f(X)記上式左邊的隨機(jī)變量為YYY,則YYY關(guān)于ggg的條件期望存在,且E[Y∣g]=f(E[X∣g])\mathbb{E}[Y|g]=f(\mathbb{E}[X|g])E[Y∣g]=f(E[X∣g])將不等式兩邊同時(shí)取條件期望則有f(E[X∣g])≤E[f(X)∣g]f(\mathbb{E}[X|g])\le \mathbb{E}[f(X)|g]f(E[X∣g])≤E[f(X)∣g]幾乎必然成立。
總結(jié)
- 上一篇: Python飞机大战项目终篇(一步一步实
- 下一篇: python宿舍管理系统_python实