給定隨機變數 $X$,我們可以定義其對應的特性函數 (characteristic function) 如下: \[
\phi_X(t):= E [e^{itX}]
\]
Comment:
1. 特性函數可視為 Fourier Transform。
2. 特性函數只與 $X$ 的 distribution 有關。
3. 特性函數滿足下列關係:$\phi(0)=1 $ 且
\[\left| {{\phi _X}(t)} \right| = \left| {E[{e^{itX}}]} \right| \le E\left[ {\left| {{e^{itX}}} \right|} \right] \le 1\]
4. 特性函數為 uniformly continuous on $\mathbb{R}$。亦即
對任意 $t \in \mathbb{R}$,我們有
\[\begin{array}{l}
\left| {\phi \left( {t + h} \right) - \phi \left( t \right)} \right| = \left| {E{e^{i\left( {t + h} \right)X}} - E{e^{i\left( t \right)X}}} \right|\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}&{}&{}
\end{array} = \left| {E\left[ {{e^{i\left( t \right)X}}\left( {{e^{i\left( h \right)X}} - 1} \right)} \right]} \right|\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}&{}&{}
\end{array} \le E\left[ {\left| {{e^{i\left( t \right)X}}\left( {{e^{i\left( h \right)X}} - 1} \right)} \right|} \right] \to 0 \;\; \text{as $\; h \downarrow 0$}
\end{array}\]上述極限成立因為 Dominated Convergence Theorem。
5. 若對任意 $t \in \mathbb{R}$ n-th moment 皆存在 則
\[{\phi _X}\left( t \right) = E{e^{itX}} = \sum\limits_{k = 0}^\infty {\frac{{E\left[ {{{\left( {itX} \right)}^k}} \right]}}{{k!}}} \]
且若 $E|X|^n <\infty$ 我們亦可對其 Taylor Expansion around $t=0$ 亦即
\[{\phi _X}\left( t \right) = \sum\limits_{k = 0}^n {\frac{{E\left[ {{{\left( {itX} \right)}^k}} \right]}}{{k!}}} + o\left( {{t^n}} \right)\]
Example: 若給定 隨機變數 $X,Y$ 且 $X,Y$ 彼此互為獨立,現在定義 $Z:=X+Y$,試求其 特性函數 $\phi_{Z}(t) =?$
由特性函數定義:
\[\begin{array}{l}
{\phi _Z}(t) = E\left[ {{e^{itZ}}} \right] = E\left[ {{e^{it\left( {X + Y} \right)}}} \right]\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}
\end{array} = E\left[ {{e^{it\left( X \right)}}{e^{it\left( Y \right)}}} \right]
\end{array}\]由於 $X$ 與 $Y$ 互為獨立,故
\[{\phi _Z}(t) = E\left[ {{e^{it\left( X \right)}}{e^{it\left( Y \right)}}} \right] = E\left[ {{e^{it\left( X \right)}}} \right]E\left[ {{e^{it\left( Y \right)}}} \right] = {\phi _X}(t){\phi _Y}(t)\]
Example: 現在考慮 隨機變數 sequence $\{X_i\}$ 為 i.i.d. ,現在定義 $S_n:=X_1+X_2+...+X_n$,試求其 特性函數 $\phi_{S_n}(t) =?$
Solution
\[\begin{array}{l}
{\phi _{{S_n}}}(t) = E\left[ {{e^{it{S_n}}}} \right] = E\left[ {{e^{it\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right]\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}
\end{array} = E\left[ {{e^{it\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right]
\end{array}\]由於 $\{X_i\}$ 為 i.i.d.,故其具有共同的 distribution,又 特性函數完全由 distribution 決定,故可知 $\phi_{X_1} = \phi_{X_2} = ... \phi_{X_n} := \phi$
\[{\phi _{{S_n}}}(t) = E\left[ {{e^{it\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right] = \prod\limits_{i = 1}^n {\underbrace {E\left[ {{e^{it\left( {{X_i}} \right)}}} \right]}_{ = \phi (t)}} = {\left( {\phi (t)} \right)^n}\]
Example: 同上題,考慮 隨機變數 sequence $\{X_i\}$ 為 i.i.d. ,並定義 $S_n:=X_1+X_2+...+X_n$,試求其 特性函數 $\phi_{S_n/n}(t) =?$
Solution
\[\begin{array}{l}
{\phi _{{S_n}/n}}(t) = E\left[ {{e^{it\frac{{{S_n}}}{n}}}} \right] = E\left[ {{e^{it\left( {\frac{{{X_1} + {X_2} + ... + {X_n}}}{n}} \right)}}} \right]\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}
\end{array} = E\left[ {{e^{i\frac{t}{n}\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right]
\end{array}\]由於 $\{X_i\}$ 為 i.i.d.,故其具有共同的 distribution,又 特性函數完全由 distribution 決定,故可知 $\phi_{X_1} = \phi_{X_2} = ... \phi_{X_n} := \phi$:
\[{\phi _{{S_n}/n}}(t) = E\left[ {{e^{i\frac{t}{n}\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right] = \prod\limits_{i = 1}^n {\underbrace {E\left[ {{e^{i\frac{t}{n}\left( {{X_i}} \right)}}} \right]}_{ = \phi (\frac{t}{n})}} = {\left( {\phi (\frac{t}{n})} \right)^n}\]
FACT:
若 $E|X|^2 < \infty$,則 $\varphi(t) = 1 + it EX - t^2 \frac{E[(itX)^2]}{2}+o(t^2)$
Proof:
事實上,由 先前的 comment 5 可知
\[{\phi _X}\left( t \right) = 1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}} + o\left( {{t^2}} \right) + H.O.T.\]但我們僅有假設 $E|X|^2 < \infty$ 故高階項不保證有界。
但所幸我們仍有以下不等式
\[E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \le E\left[ {\min \left\{ {{{\left| {tX} \right|}^3},2{{\left| {tX} \right|}^2}} \right\}} \right]\]且我們可進一步確認其具有上界為
\[E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \le E\left[ {\min \left\{ {{{\left| {tX} \right|}^3},2{{\left| {tX} \right|}^2}} \right\}} \right] \le E\left[ {{{\left| {tX} \right|}^2}} \right]\]故由 Dominated Convergence Theorem 可知
\[\begin{array}{l}
E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \le E\left[ {\min \left\{ {{{\left| {tX} \right|}^3},2{{\left| {tX} \right|}^2}} \right\}} \right] \le E\left[ {{{\left| {tX} \right|}^2}} \right]\\
\Rightarrow E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \to 0
\end{array}\]
\phi_X(t):= E [e^{itX}]
\]
Comment:
1. 特性函數可視為 Fourier Transform。
2. 特性函數只與 $X$ 的 distribution 有關。
3. 特性函數滿足下列關係:$\phi(0)=1 $ 且
\[\left| {{\phi _X}(t)} \right| = \left| {E[{e^{itX}}]} \right| \le E\left[ {\left| {{e^{itX}}} \right|} \right] \le 1\]
4. 特性函數為 uniformly continuous on $\mathbb{R}$。亦即
對任意 $t \in \mathbb{R}$,我們有
\[\begin{array}{l}
\left| {\phi \left( {t + h} \right) - \phi \left( t \right)} \right| = \left| {E{e^{i\left( {t + h} \right)X}} - E{e^{i\left( t \right)X}}} \right|\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}&{}&{}
\end{array} = \left| {E\left[ {{e^{i\left( t \right)X}}\left( {{e^{i\left( h \right)X}} - 1} \right)} \right]} \right|\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}&{}&{}
\end{array} \le E\left[ {\left| {{e^{i\left( t \right)X}}\left( {{e^{i\left( h \right)X}} - 1} \right)} \right|} \right] \to 0 \;\; \text{as $\; h \downarrow 0$}
\end{array}\]上述極限成立因為 Dominated Convergence Theorem。
5. 若對任意 $t \in \mathbb{R}$ n-th moment 皆存在 則
\[{\phi _X}\left( t \right) = E{e^{itX}} = \sum\limits_{k = 0}^\infty {\frac{{E\left[ {{{\left( {itX} \right)}^k}} \right]}}{{k!}}} \]
且若 $E|X|^n <\infty$ 我們亦可對其 Taylor Expansion around $t=0$ 亦即
\[{\phi _X}\left( t \right) = \sum\limits_{k = 0}^n {\frac{{E\left[ {{{\left( {itX} \right)}^k}} \right]}}{{k!}}} + o\left( {{t^n}} \right)\]
Example: 若給定 隨機變數 $X,Y$ 且 $X,Y$ 彼此互為獨立,現在定義 $Z:=X+Y$,試求其 特性函數 $\phi_{Z}(t) =?$
由特性函數定義:
\[\begin{array}{l}
{\phi _Z}(t) = E\left[ {{e^{itZ}}} \right] = E\left[ {{e^{it\left( {X + Y} \right)}}} \right]\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}
\end{array} = E\left[ {{e^{it\left( X \right)}}{e^{it\left( Y \right)}}} \right]
\end{array}\]由於 $X$ 與 $Y$ 互為獨立,故
\[{\phi _Z}(t) = E\left[ {{e^{it\left( X \right)}}{e^{it\left( Y \right)}}} \right] = E\left[ {{e^{it\left( X \right)}}} \right]E\left[ {{e^{it\left( Y \right)}}} \right] = {\phi _X}(t){\phi _Y}(t)\]
Example: 現在考慮 隨機變數 sequence $\{X_i\}$ 為 i.i.d. ,現在定義 $S_n:=X_1+X_2+...+X_n$,試求其 特性函數 $\phi_{S_n}(t) =?$
Solution
\[\begin{array}{l}
{\phi _{{S_n}}}(t) = E\left[ {{e^{it{S_n}}}} \right] = E\left[ {{e^{it\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right]\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}
\end{array} = E\left[ {{e^{it\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right]
\end{array}\]由於 $\{X_i\}$ 為 i.i.d.,故其具有共同的 distribution,又 特性函數完全由 distribution 決定,故可知 $\phi_{X_1} = \phi_{X_2} = ... \phi_{X_n} := \phi$
\[{\phi _{{S_n}}}(t) = E\left[ {{e^{it\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right] = \prod\limits_{i = 1}^n {\underbrace {E\left[ {{e^{it\left( {{X_i}} \right)}}} \right]}_{ = \phi (t)}} = {\left( {\phi (t)} \right)^n}\]
Example: 同上題,考慮 隨機變數 sequence $\{X_i\}$ 為 i.i.d. ,並定義 $S_n:=X_1+X_2+...+X_n$,試求其 特性函數 $\phi_{S_n/n}(t) =?$
Solution
\[\begin{array}{l}
{\phi _{{S_n}/n}}(t) = E\left[ {{e^{it\frac{{{S_n}}}{n}}}} \right] = E\left[ {{e^{it\left( {\frac{{{X_1} + {X_2} + ... + {X_n}}}{n}} \right)}}} \right]\\
\begin{array}{*{20}{c}}
{}&{}&{}&{}
\end{array} = E\left[ {{e^{i\frac{t}{n}\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right]
\end{array}\]由於 $\{X_i\}$ 為 i.i.d.,故其具有共同的 distribution,又 特性函數完全由 distribution 決定,故可知 $\phi_{X_1} = \phi_{X_2} = ... \phi_{X_n} := \phi$:
\[{\phi _{{S_n}/n}}(t) = E\left[ {{e^{i\frac{t}{n}\left( {{X_1} + {X_2} + ... + {X_n}} \right)}}} \right] = \prod\limits_{i = 1}^n {\underbrace {E\left[ {{e^{i\frac{t}{n}\left( {{X_i}} \right)}}} \right]}_{ = \phi (\frac{t}{n})}} = {\left( {\phi (\frac{t}{n})} \right)^n}\]
FACT:
若 $E|X|^2 < \infty$,則 $\varphi(t) = 1 + it EX - t^2 \frac{E[(itX)^2]}{2}+o(t^2)$
Proof:
事實上,由 先前的 comment 5 可知
\[{\phi _X}\left( t \right) = 1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}} + o\left( {{t^2}} \right) + H.O.T.\]但我們僅有假設 $E|X|^2 < \infty$ 故高階項不保證有界。
但所幸我們仍有以下不等式
\[E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \le E\left[ {\min \left\{ {{{\left| {tX} \right|}^3},2{{\left| {tX} \right|}^2}} \right\}} \right]\]且我們可進一步確認其具有上界為
\[E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \le E\left[ {\min \left\{ {{{\left| {tX} \right|}^3},2{{\left| {tX} \right|}^2}} \right\}} \right] \le E\left[ {{{\left| {tX} \right|}^2}} \right]\]故由 Dominated Convergence Theorem 可知
\[\begin{array}{l}
E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \le E\left[ {\min \left\{ {{{\left| {tX} \right|}^3},2{{\left| {tX} \right|}^2}} \right\}} \right] \le E\left[ {{{\left| {tX} \right|}^2}} \right]\\
\Rightarrow E\left| {{e^{itX}} - \left( {1 + itE\left[ X \right] - \frac{{{t^2}E\left[ {{X^2}} \right]}}{{2!}}} \right)} \right| \to 0
\end{array}\]
留言
張貼留言