服从正态分布的正弦函数期望
服从正态分布的正弦函数、余弦函数期望。
如果X
服从均值为
μ
\mu
μ,方差为
σ
2
\sigma^2
σ2的正态分布,计算sin(X)
与cos(X)
的数学期望。
利用特征函数(Characteristic Function)Wiki-Characteristic Function,我们知道
X
∼
N
(
μ
,
σ
2
)
X\sim N(\mu, \sigma^2)
X∼N(μ,σ2)的特征函数为:
φ
X
(
t
)
=
E
(
e
i
t
X
)
=
e
x
p
(
i
μ
t
−
σ
2
t
2
2
)
=
e
x
p
(
−
σ
2
t
2
/
2
)
e
x
p
(
i
μ
t
)
\varphi_{X(t)}=E(e^{itX})=exp\left({i\mu t - \dfrac{\sigma^2t^2}{2}}\right)=exp(-\sigma^2t^2/2)exp(i\mu t)
φX(t)=E(eitX)=exp(iμt−2σ2t2)=exp(−σ2t2/2)exp(iμt)
根据欧拉公式:
e
i
x
=
cos
(
x
)
+
i
sin
(
x
)
e^{ix} = \cos(x)+i\sin(x)
eix=cos(x)+isin(x)
E ( e i t X ) = e x p ( i μ t − σ 2 t 2 2 ) = e x p ( − σ 2 t 2 / 2 ) e x p ( i μ t ) = e x p ( − σ 2 t 2 / 2 ) [ cos ( μ t ) + i sin ( μ t ) ] = e x p ( − σ 2 t 2 / 2 ) cos ( μ t ) + e x p ( − σ 2 t 2 / 2 ) sin ( μ t ) i = R e a l ( E ( e i t X ) ) + i I m ( E ( e i t X ) ) \begin{aligned} E(e^{itX})&=exp\left({i\mu t - \dfrac{\sigma^2t^2}{2}}\right) \\ &=exp(-\sigma^2t^2/2)exp(i\mu t) \\ &=exp(-\sigma^2t^2/2)\left[\cos(\mu t)+i\sin(\mu t)\right]\\ &=exp(-\sigma^2t^2/2)\cos(\mu t) +exp(-\sigma^2t^2/2)\sin(\mu t)i \\ &=Real(E(e^{itX}))+iIm(E(e^{itX})) \end{aligned} E(eitX)=exp(iμt−2σ2t2)=exp(−σ2t2/2)exp(iμt)=exp(−σ2t2/2)[cos(μt)+isin(μt)]=exp(−σ2t2/2)cos(μt)+exp(−σ2t2/2)sin(μt)i=Real(E(eitX))+iIm(E(eitX))
上述数学期望变为:
E
(
e
i
t
X
)
=
E
(
cos
(
t
X
)
+
i
sin
(
t
X
)
)
=
E
(
c
o
s
(
t
X
)
)
+
i
E
(
s
i
n
(
t
X
)
)
=
R
e
a
l
(
E
(
e
i
t
X
)
)
+
i
I
m
(
E
(
e
i
t
X
)
)
\begin{aligned} E(e^{itX})&=E\left(\cos(tX)+i\sin(tX)\right)\\ &= E(cos(tX)) + iE(sin(tX))\\ &=Real(E(e^{itX}))+iIm(E(e^{itX})) \end{aligned}
E(eitX)=E(cos(tX)+isin(tX))=E(cos(tX))+iE(sin(tX))=Real(E(eitX))+iIm(E(eitX))
对比上述的实数部分和虚数部分,得到:
E
(
cos
(
t
X
)
)
=
e
x
p
(
−
σ
2
t
2
/
2
)
cos
(
μ
t
)
E
(
sin
(
t
X
)
)
=
e
x
p
(
−
σ
2
t
2
/
2
)
s
i
n
(
μ
t
)
E(\cos(tX))=exp(-\sigma^2t^2/2)\cos(\mu t)\\ E(\sin(tX))=exp(-\sigma^2t^2/2)sin(\mu t)
E(cos(tX))=exp(−σ2t2/2)cos(μt)E(sin(tX))=exp(−σ2t2/2)sin(μt)
最后,当t=1
的时候:
E
(
cos
(
X
)
)
=
e
x
p
(
−
σ
2
/
2
)
cos
(
μ
)
E
(
sin
(
X
)
)
=
e
x
p
(
−
σ
2
/
2
)
s
i
n
(
μ
)
E(\cos(X))=exp(-\sigma^2/2)\cos(\mu)\\ E(\sin(X))=exp(-\sigma^2/2)sin(\mu)
E(cos(X))=exp(−σ2/2)cos(μ)E(sin(X))=exp(−σ2/2)sin(μ)
参考资料:Mean and variance of Y=cos(bX) when X has a Gaussian distribution
应用:MipNeRF(ArxivLink)采用的IPE位置编码中,对服从高斯分布正弦函数的数学期望计算:
# Code Source:
# https://github.com/liuyuan-pal/NeRO/blob/3b4d421a646097e7d59557c5ea24f4281ab38ef1/network/field.py#L369-L378
def expected_sin(mean, var):
"""Compute the mean of sin(x), x ~ N(mean, var)."""
return torch.exp(-0.5 * var) * torch.sin(mean) # large var -> small value.
def IPE(mean,var,min_deg,max_deg):
scales = 2**torch.arange(min_deg, max_deg)
shape = mean.shape[:-1] + (-1,)
scaled_mean = torch.reshape(mean[..., None, :] * scales[:, None], shape)
scaled_var = torch.reshape(var[..., None, :] * scales[:, None]**2, shape)
return expected_sin(torch.concat([scaled_mean, scaled_mean + 0.5 * np.pi], dim=-1), torch.concat([scaled_var] * 2, dim=-1))