夜空中最亮的星
1- Dirichlet 积分
设(I(a)=frac1piint_0^{+infty}frac{sin{at}}{t}dt),则有:
[I(a)=
egin{cases}
frac12& ext{a>0}\
0&a=0\
-frac12&a<0
end{cases}
]
为了证明 (Dirichlet 积分),我们先证明(int_0^{+infty}frac{sin{x}}{x}dx=fracpi2)
[egin{align}
设 frac1x=&int_0^{+infty}e^{-xs}ds\
int_0^Tfrac{sin{x}}{x}dx=&int_0^{T}(sin{x}{int_0^{+infty}e^{-xs}ds)}dx\
=&int_0^{+infty}({int_0^{T}sin{x} e^{-xs}dx)}ds\
=&int_0^{+infty}[frac{1}{1+s^2}-frac{scdotsin T+Tcdotcos{T}}{s^2+T^2}e^{-s}]ds\
=&fracpi2-int_0^{+infty}frac{scdotsin T+Tcdotcos{T}}{s^2+T^2}e^{-s}ds\
(ecause&lim_{s oinfty}frac{scdotsin T+Tcdotcos{T}}{s^2+T^2}=0)\
=&fracpi2\
则I(a)=&frac1piint_0^{+infty}frac{sin{at}}{t}dt\
=&acdotfrac1acdotfrac1piint_0^{+infty}frac{sin{at}}{at}d(at)\
=&frac1picdotfracpi2\
=½
end{align}
]
此积分的重要意义是可以将符号转化为积分表达式!
3-特征函数
设(X)是随机变量,(f(x))为概率密度函数,则特征函数为:
[varphi_X(t)=int_{-infty}^infty e^{itx}f(x)dx
]
由:(e^{itx}=sum_{n=0}^inftyfrac{(itx)^n}{n!}),则有:
[egin{align}
varphi_X(t)=&int_{-infty}^infty e^{itx}f(x)dx\
=&int_{-infty}^infty (sum_{n=0}^inftyfrac{(itx)^n}{n!})f(x)dx\
=&sum_{n=0}^inftyint_{-infty}^infty(frac{(itx)^n}{n!})f(x)dx\
=&sum_{n=0}^inftyfrac{(it)^n}{n!}int_{-infty}^infty x^nf(x)dx\
=&sum_{n=0}^inftyfrac{(it)^n}{n!}E(x^n)\
varphi_X(t)=&e^{it}cdot E(x^n)
end{align}
]
特征函数的反演公式:
[f(x)=frac1{2pi}int_{-infty}^{infty}e^{-itx}varphi(t)dt
]
4-特征函数性质
设({f(x)=P(xleq X) | E(X)=mu,Var(X)=sigma^2}),其特征函数的本质为概率密度函数的傅里叶变换:
[varphi_X(t)=int_{-infty}^infty e^{itx}f(x)dx
]
而特征函数的 k 阶导数与 k 阶矩之间有密切联系:
[varphi^{(k)}_X(t)=[int_{-infty}^infty e^{itx}f(x)dx]^{(k)}=int_{-infty}^infty (e^{itx})^{(k)}f(x)dx
=int_{-infty}^infty (ix)^{k}e^{itx}f(x)dx
]
于是,令(t=0):
[egin{align}
varphi_X(0)&=int_{-infty}^infty f(x)dx=1\
varphi^{'}_X(0)&=int_{-infty}^infty (ix)e^0f(x) dx\
&=iint_{-infty}^infty xf(x) dx\
&=icdot E(x)\
varphi^{''}_X(0)&=int_{-infty}^infty (ix)^2e^0f(x) dx\
&=-int_{-infty}^infty x^2f(x) dx\
&=-(int_{-infty}^infty (x-mu)^2f(x) dx+int_{-infty}^infty(2mu)xf(x) dx-mu^2int_{-infty}^infty f(x) dx)\
&=-sigma^2-mu^2
end{align}
]
特征函数的乘法性质:
[varphi_{X+Y}(t)=varphi_{X}(t)varphi_{Y}(t)
]
若(X hicksim N(0,1)),则其概率密度函数为:
[f(x)=frac{1}{sqrt{2pi}}e^{-frac{x^2}{2}}
]
其对应的特征函数为:
[egin{align}
varphi_X(t)=&int_{-infty}^infty e^{itx}(frac{1}{sqrt{2pi}}e^{-frac{x^2}{2}})dx=e^{-frac{t^2}{2}}
end{align}
]
关于对含参量反常积分可微性
[egin{align}
varphi_X(t)=&int_{-infty}^infty e^{itx}(frac{1}{sqrt{2pi}}e^{-frac{x^2}{2}})dx\
varphi'_X(t)=
=&frac{1}{sqrt{2pi}}int_{-infty}^infty ixe^{-frac{x^2}{2}}e^{itx}dx\
=&frac{i}{sqrt{2pi}}int_{-infty}^infty e^{itx}d(-e^{-frac{x^2}{2}})\
=&frac{i}{sqrt{2pi}}e^{(itx-frac{x^2}{2})}|_{-infty}^{+infty}
-frac{i}{sqrt{2pi}}int_{-infty}^infty -e^{-frac{x^2}{2}}d(e^{itx})\
=&frac{i}{sqrt{2pi}}e^{(itx-frac{x^2}{2})}|_{-infty}^{+infty}
-frac{i}{sqrt{2pi}}int_{-infty}^infty ite^{itx}(-e^{-frac{x^2}{2}})dx\
=&-frac{t}{sqrt{2pi}}int_{-infty}^infty e^{itx}(e^{-frac{x^2}{2}})dx\
=&-tvarphi(t)
end{align}
]
于是我们又获得一个重要的微分方程:
[frac{dvarphi(t)}{dt}=-tcdotvarphi(t)
]
变形得到:
[egin{align}
frac{dvarphi(t)}{varphi(t)}&=-tdt\
intfrac{dvarphi(t)}{varphi(t)}&=-int tdt\
ln{varphi(t)}&=-frac{t^2}{2}+C\
varphi(t)&=C_0cdot e^{-frac{t^2}{2}}\
(ecausevarphi(0)=1,带入公式&得到C_0=e^0=1,于是有)\
varphi(t)&=e^{-frac{t^2}{2}}\
end{align}
]
也就是说当概率密度函数为(f(x)=frac{1}{sqrt{2pi}}e^{-frac{x^2}{2}})时,特征函数为:(varphi(t)=e^{-frac{t^2}{2}})。
5-中心极限定理
设(X_1,X_2,dots,X_n)为n个独立同分布随机变量,(X_i hicksim N(mu,sigma^2)),不妨设(Y_i=X_i-mu,)那么(Y_i hicksim N(0,sigma^2)),设(Y_i)的特征函数为(varphi(t)),
设随机变量(eta=frac{Y_1+Y_2+dots+Y_n}{sigmasqrt{n}}),由(varphi_{X+Y}(t)=varphi_{X}(t)varphi_{Y}(t)),则(eta)的特征函数为:
[[varphi(frac{t}{sqrt{n}sigma})][varphi(frac{t}{sqrt{n}sigma})]dots[varphi(frac{t}{sqrt{n}sigma})]=[varphi(frac{t}{sqrt{n}sigma})]^n
]
将(varphi(frac{t}{sqrt{n}sigma}))在0点(Taylor)展开,
[egin{align}
varphi(frac{t}{sqrt{n}sigma})
&=varphi(0)+varphi'(0)(frac{t}{sqrt{n}sigma})+frac{varphi''(0)}{2!}(frac{t}{sqrt{n}sigma})^2+o((frac{t}{sqrt{n}sigma})^2)\
&=1+frac{imu t}{sqrt{n}sigma}-frac{(sigma^2+mu)t^2}{2nsigma^2}+o((frac{t}{sqrt{n}sigma})^2)\
&=1-frac{t^2}{2n}+o((frac{t}{sqrt{n}sigma})^2)\
herefore[varphi(frac{t}{sqrt{n}sigma})]^n&=[1-frac{t^2}{2n}+o((frac{t}{sqrt{n}sigma})^2)]^n\
&=[1-frac{t^2}{2n}+o((frac{t}{sqrt{n}sigma})^2)]^{(-frac{2n}{t^2})(-frac{t^2}{2})}\
lim_{n oinfty}[varphi(frac{t}{sqrt{n}sigma})]^n&=e^{-frac{t^2}{2}}
end{align}
]
因此(eta hicksim N(0,1))服从标准正态分布。