设有观测值
X
n
1
\mathop{X}\limits_{n1}
n1X的非线性函数
Z
=
f
(
X
)
(1)
Z=f(X) \tag{1}
Z=f(X)(1)
或写成 Z = f ( X 1 , X 2 , ⋯ , X n ) Z=f\left(X_{1}, X_{2}, \cdots, X_{n}\right) Z=f(X1,X2,⋯,Xn)
已知
X
X
X的协方差矩阵
D
x
x
D_{xx}
Dxx,欲求Z的协方差
D
z
z
D_{zz}
Dzz
假定观测值
X
X
X有近似值
X
0
n
1
\mathop{X^0}\limits_{n1}
n1X0
X
0
=
[
X
1
0
X
2
0
⋯
X
n
0
]
T
\begin{aligned} &X^{0}=\left[\begin{array}{llll} X_{1}^{0} & X_{2}^{0} & \cdots & X_{n}^{0} \end{array}\right]^{T}\\ \end{aligned}
X0=[X10X20⋯Xn0]T
可将函数式按泰勒级数在点
X
1
0
,
X
2
0
,
⋯
X
n
0
X_{1}^{0},X_{2}^{0}, \cdots X_{n}^{0}
X10,X20,⋯Xn0处展开为
Z
=
f
(
X
1
0
X
2
0
⋯
X
n
0
)
+
(
∂
f
∂
X
1
)
0
(
X
1
−
X
1
0
)
+
(
∂
f
∂
X
2
)
(
X
2
−
X
2
0
)
+
⋯
+
(
∂
f
∂
X
n
)
(
X
n
−
X
n
0
)
+
(
二
次
以
上
项
)
.
(2)
\begin{aligned} &Z=f\left(X_{1}^{0} \quad X_{2}^{0} \quad \cdots \quad X_{n}^{0}\right)+\left(\frac{\partial f}{\partial X_{1}}\right)_{0}\left(X_{1}-X_{1}^{0}\right)\\ &+\left(\frac{\partial f}{\partial X_{2}}\right)\left(X_{2}-X_{2}^{0}\right)+\cdots+\left(\frac{\partial f}{\partial X_{n}}\right)\left(X_{n}-X_{n}^{0}\right)+(二次以上项). \end{aligned}\tag{2}
Z=f(X10X20⋯Xn0)+(∂X1∂f)0(X1−X10)+(∂X2∂f)(X2−X20)+⋯+(∂Xn∂f)(Xn−Xn0)+(二次以上项).(2)
式中，
(
∂
f
∂
X
i
)
0
(\frac{\partial f}{\partial X_{i}})_0
(∂Xi∂f)0是函数对各个变量所取的偏导数，并以近似值
X
0
X^0
X0带入所算得的数值，它们都是常数。当
X
0
X^0
X0与X非常接近时，上式中二次以上各项很微小，故可以略去。因此，可将上式写为
Z
=
(
∂
f
∂
X
1
)
0
X
1
+
(
∂
f
∂
X
2
)
0
X
2
+
⋯
+
(
∂
f
∂
X
2
)
0
X
0
+
f
(
X
1
0
,
X
2
0
,
⋯
,
X
n
)
−
∑
i
=
1
n
(
∂
f
∂
X
i
)
0
X
i
(3)
Z=\left(\frac{\partial f}{\partial X_{1}}\right)_{0} X_{1}+\left(\frac{\partial f}{\partial X_{2}}\right)_{0} X_{2}+\cdots+\left(\frac{\partial f}{\partial X_{2}}\right) _{0} X_{0}+f\left(X_{1}^{0}, X_{2}^{0}, \cdots, X_{n}\right)-\sum_{i=1}^{n}\left(\frac{\partial f}{\partial X_{i}}\right) _{0} X_i\tag{3}
Z=(∂X1∂f)0X1+(∂X2∂f)0X2+⋯+(∂X2∂f)0X0+f(X10,X20,⋯,Xn)−i=1∑n(∂Xi∂f)0Xi(3)
令
K
=
[
k
1
k
2
⋯
k
2
]
=
[
(
∂
f
∂
X
1
)
0
(
∂
f
∂
X
1
)
0
⋯
(
∂
f
∂
X
n
)
0
]
K=\left[\begin{array}{llll} k_{1} & k_{2} & \cdots & k_{2} \end{array}\right]=\left[\begin{array}{lll} (\frac{\partial f}{\partial X_{1}})_{0}&(\frac{\partial f}{\partial X_{1}})_{0}& \cdots(\frac{\partial f}{\partial X_{n}})_{0}] \end{array}\right.
K=[k1k2⋯k2]=[(∂X1∂f)0(∂X1∂f)0⋯(∂Xn∂f)0]
k 0 = f ( X 1 0 X 2 0 ⋯ X n 0 ) − ∑ i = 1 n ( ∂ f ∂ X i ) 0 X i (4) k_{0}=f\left(X_{1}^{0} \quad X_{2}^{0} \quad \cdots \quad X_{n}^{0}\right)-\sum_{i=1}^{n}\left(\frac{\partial f}{\partial X_{i}}\right)_{0} X_{i}\tag{4} k0=f(X10X20⋯Xn0)−i=1∑n(∂Xi∂f)0Xi(4)
则
Z
=
k
1
X
1
+
k
2
X
2
+
⋯
+
k
n
X
n
+
k
0
=
K
X
+
k
0
Z=k_{1} X_{1}+k_{2} X_{2}+\cdots+k_{\mathrm{n}} X_{\mathrm{n}}+k_{0}=K X+k_{0}
Z=k1X1+k2X2+⋯+knXn+k0=KX+k0
这样，就将非线性函数化为线性函数，因此可以按照线性函数的方法求取方差
D
z
z
=
K
D
x
x
K
T
(5)
D_{z z}=K D_{x x} K^{T}\tag{5}
Dzz=KDxxKT(5)
如果令
d
X
i
=
X
i
−
X
i
0
(
i
=
1
,
2
,
⋯
,
n
)
d
X
=
(
d
X
1
d
X
2
⋯
d
X
n
)
T
d
Z
=
Z
−
Z
0
=
Z
−
f
(
X
1
0
,
X
2
0
,
⋯
,
X
n
0
)
\begin{aligned} &\mathrm{d} X_{i}=X_{i}-X_{i}^{0} \quad(i=1,2, \cdots, n)\\ &\mathrm{d} X=\left(\begin{array}{llll} \mathrm{d} X_{1} & \mathrm{d} X_{2} & \cdots & \mathrm{d} X_{\mathrm{n}} \end{array}\right)^{\mathrm{T}}\\ &\mathrm{d} Z=Z-Z^{0}=Z-f\left(X_{1}^{0}, X_{2}^{0}, \cdots, X_{n}^{0}\right) \end{aligned}
dXi=Xi−Xi0(i=1,2,⋯,n)dX=(dX1dX2⋯dXn)TdZ=Z−Z0=Z−f(X10,X20,⋯,Xn0)
则可以写为
d
Z
=
(
∂
f
∂
X
1
)
0
d
X
1
+
(
∂
f
∂
X
2
)
0
d
X
2
+
⋯
+
(
∂
f
∂
X
n
)
0
d
X
n
=
K
d
X
\mathrm{d} Z=\left(\frac{\partial f}{\partial X_{1}}\right)_{0} \mathrm{d} X_{1}+\left(\frac{\partial f}{\partial X_{2}}\right)_{0} \mathrm{d} X_{2}+\cdots+\left(\frac{\partial f}{\partial X_{n}}\right)_{0} \mathrm{d} X_{n}=K \mathrm{d} X
dZ=(∂X1∂f)0dX1+(∂X2∂f)0dX2+⋯+(∂Xn∂f)0dXn=KdX
易知上式是非线性函数式的全微分。因为根据式（5）应用协方差传播率式求
D
z
z
D_{zz}
Dzz时，只要求知道式中的系数矩阵
K
K
K。所以，为了求非线性函数的方差，只要对它先求全微分并将非线性函数化为线性函数形式，再按协方差传播律就可求得该函数的方差。
如果有
t
t
t个非线性函数
Z
1
=
f
1
(
X
1
,
X
2
,
⋯
,
X
n
)
Z
2
=
f
2
(
X
1
,
X
2
,
⋯
,
X
n
)
⋯
⋯
Z
t
=
f
t
(
X
1
,
X
2
,
⋯
,
X
n
)
}
\left.\begin{array}{l} Z_{1}=f_{1}\left(X_{1}, X_{2}, \cdots, X_{n}\right) \\ Z_{2}=f_{2}\left(X_{1}, X_{2}, \cdots, X_{n}\right) \\ \cdots\cdots\\ Z_{t}=f_{t}\left(X_{1}, X_{2}, \cdots, X_{n}\right) \\ \end{array}\right\}
Z1=f1(X1,X2,⋯,Xn)Z2=f2(X1,X2,⋯,Xn)⋯⋯Zt=ft(X1,X2,⋯,Xn)⎭⎪⎪⎬⎪⎪⎫
将
t
t
t个函数求全微分得
d
Z
1
=
(
∂
f
1
∂
X
1
)
0
d
X
1
+
(
∂
f
1
∂
X
2
)
0
d
X
2
+
⋯
+
(
∂
f
1
∂
X
n
)
0
d
X
n
d
Z
2
=
(
∂
f
2
∂
X
1
)
0
d
X
1
+
(
∂
f
2
∂
X
2
)
0
d
X
2
+
⋯
+
(
∂
f
2
∂
X
n
)
0
d
X
n
⋯
⋯
d
Z
t
=
(
∂
f
t
∂
X
1
)
0
d
X
1
+
(
∂
f
t
∂
X
2
)
0
d
X
2
+
⋯
+
(
∂
f
t
∂
X
n
)
0
d
X
n
}
\left.\begin{array}{l} \mathrm{d} Z_{1}=\left(\frac{\partial f_{1}}{\partial X_{1}}\right)_{0} \mathrm{d} X_{1}+\left(\frac{\partial f_{1}}{\partial X_{2}}\right)_{0} \mathrm{d} X_{2}+\cdots+\left(\frac{\partial f_{1}}{\partial X_{n}}\right)_{0} \mathrm{d} X_{n} \\ \mathrm{d} Z_{2}=\left(\frac{\partial f_{2}}{\partial X_{1}}\right)_{0} \mathrm{d} X_{1}+\left(\frac{\partial f_{2}}{\partial X_{2}}\right)_{0} \mathrm{d} X_{2}+\cdots+\left(\frac{\partial f_{2}}{\partial X_{n}}\right)_{0} \mathrm{d} X_{n}\\ \cdots\cdots\\ d Z_{t}=\left(\frac{\partial f_{t}}{\partial X_{1}}\right)_{0} d X_{1}+\left(\frac{\partial f_{t}}{\partial X_{2}}\right)_{0} d X_{2}+\cdots+\left(\frac{\partial f_{t}}{\partial X_{n}}\right)_{0} d X_{n} \\ \end{array}\right\}
dZ1=(∂X1∂f1)0dX1+(∂X2∂f1)0dX2+⋯+(∂Xn∂f1)0dXndZ2=(∂X1∂f2)0dX1+(∂X2∂f2)0dX2+⋯+(∂Xn∂f2)0dXn⋯⋯dZt=(∂X1∂ft)0dX1+(∂X2∂ft)0dX2+⋯+(∂Xn∂ft)0dXn⎭⎪⎪⎪⎪⎪⎬⎪⎪⎪⎪⎪⎫
若记
Z
t
1
=
[
Z
1
Z
2
⋮
Z
i
]
,
d
Z
t
1
=
[
d
Z
1
d
Z
2
⋮
d
Z
i
]
\mathop{Z}\limits_{t1}=\left[\begin{array}{c} Z_{1} \\ Z_{2} \\ \vdots \\ Z_{i} \end{array}\right], \quad \mathop{dZ}\limits_{t1}=\left[\begin{array}{c} \mathrm{d} Z_{1} \\ \mathrm{d} Z_{2} \\ \vdots \\ \mathrm{d} Z_{i} \end{array}\right]
t1Z=⎣⎢⎢⎢⎡Z1Z2⋮Zi⎦⎥⎥⎥⎤,t1dZ=⎣⎢⎢⎢⎡dZ1dZ2⋮dZi⎦⎥⎥⎥⎤
K t n = [ ( ∂ f 1 ∂ X 1 ) 0 ( ∂ f 1 ∂ X 2 ) 0 ⋯ ( ∂ f 1 ∂ X n ) 0 ( ∂ f 2 ∂ X 1 ) 0 ( ∂ f 2 ∂ X 2 ) 0 ⋯ ( ∂ f 2 ∂ X n ) 0 ⋮ ⋮ ⋮ ( ∂ f 1 ∂ X 1 ) 0 ( ∂ f 1 ∂ X 2 ) 0 ⋯ ( ∂ f t ∂ X n ) 0 ] \mathop{K}\limits_{tn}=\left[\begin{array}{cc} \left(\frac{\partial f_{1}}{\partial X_{1}}\right)_{0} & \left(\frac{\partial f_{1}}{\partial X_{2}}\right)_{0} & \cdots & \left(\frac{\partial f_{1}}{\partial X_{n}}\right)_{0} \\ \left(\frac{\partial f_{2}}{\partial X_{1}}\right)_{0} & \left(\frac{\partial f_{2}}{\partial X_{2}}\right)_{0} & \cdots & \left(\frac{\partial f_{2}}{\partial X_{n}}\right)_{0} \\ \vdots & \vdots & \vdots & \\ \left(\frac{\partial f_{1}}{\partial X_{1}}\right)_{0} & \left(\frac{\partial f_{1}}{\partial X_{2}}\right)_{0} & \cdots & \left(\frac{\partial f_{t}}{\partial X_{n}}\right)_{0} \end{array}\right] tnK=⎣⎢⎢⎢⎢⎢⎢⎡(∂X1∂f1)0(∂X1∂f2)0⋮(∂X1∂f1)0(∂X2∂f1)0(∂X2∂f2)0⋮(∂X2∂f1)0⋯⋯⋮⋯(∂Xn∂f1)0(∂Xn∂f2)0(∂Xn∂ft)0⎦⎥⎥⎥⎥⎥⎥⎤
则有
d
Z
=
K
d
X
dZ=KdX
dZ=KdX
亦可求得
Z
t
1
\mathop{Z}\limits_{t1}
t1Z的协方差矩阵。
D
z
z
=
K
D
x
x
K
T
D_{z z}=K D_{x x} K^{T}
Dzz=KDxxKT
注：观测值 X X X的近似值 X 0 n 1 \mathop{X^0}\limits_{n1} n1X0通常是其期望，或者是当前的工作点。
参看文献
误差测量与平差基础 武汉大学出版社
4676
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
