摘要 本文主要参考《统计信号处理基础》，一些性质后面标了页码。从联合正态分布的一些性质出发推导标量和向量形式的卡尔曼滤波，以及重点推导了均值不为0的各个公式。下面的一些性质出于方便考虑给起了名字，不一定是正式的学术命名。
公式与符号说明
离散系统
x
(
k
)
=
A
x
(
k
−
1
)
+
B
u
(
k
−
1
)
+
v
(
k
−
1
)
y
(
k
)
=
C
x
(
k
)
+
W
(
k
)
\begin{aligned} & x(k)=Ax(k-1)+Bu(k-1)+v(k-1) \\ & y(k)=Cx(k)+W(k) \end{aligned}
x(k)=Ax(k−1)+Bu(k−1)+v(k−1)y(k)=Cx(k)+W(k)
其中
V
,
W
V,W
V,W为零均值高斯白噪声的协方差矩阵。
卡尔曼滤波器递推公式如下
x
^
(
k
∣
k
−
1
)
=
A
x
^
(
k
−
1
)
+
B
u
(
k
−
1
)
x
^
(
k
)
=
x
^
(
k
∣
k
−
1
)
+
K
(
k
)
[
y
(
k
)
−
C
x
^
(
k
∣
k
−
1
)
]
P
(
k
∣
k
−
1
)
=
A
P
(
k
−
1
)
A
T
+
V
P
(
k
)
=
[
I
−
K
(
k
)
C
]
P
(
k
∣
k
−
1
)
K
(
k
)
=
P
(
k
∣
k
−
1
)
C
T
[
C
P
(
k
∣
k
−
1
)
C
T
+
W
]
−
1
\begin{aligned} & \hat{x}(k|k-1)=A\hat{x}(k-1)+Bu(k-1) \\ & \hat{x}(k)=\hat{x}(k|k-1)+K(k)[y(k)-C\hat{x}(k|k-1)] \\ & P(k|k-1)=AP(k-1)A^{\text{T}}+V \\ & P(k)=[I-K(k)C]P(k|k-1) \\ & K(k)=P(k|k-1)C^{\text{T}}[CP(k|k-1)C^{\text{T}}+W]^{-1} \\ \end{aligned}
x^(k∣k−1)=Ax^(k−1)+Bu(k−1)x^(k)=x^(k∣k−1)+K(k)[y(k)−Cx^(k∣k−1)]P(k∣k−1)=AP(k−1)AT+VP(k)=[I−K(k)C]P(k∣k−1)K(k)=P(k∣k−1)CT[CP(k∣k−1)CT+W]−1
- Y ( k ) = [ y ( 0 ) , y ( 1 ) , ⋯ , y ( k ) ] \mathbf{Y}(k)=[y(0),y(1),\cdots,y(k)] Y(k)=[y(0),y(1),⋯,y(k)] 表示前 k k k 个时刻的观测数据
- x ^ ( k ) = E [ x ( k ) ∣ Y ( k ) ] \hat{x}(k)=\text{E}[x(k)|Y(k)] x^(k)=E[x(k)∣Y(k)] 表示根据前 k k k 个时刻的观测数据对第 k k k时刻的实际状态 x ( k ) x(k) x(k) 的预测值
- x ^ ( k ∣ k − 1 ) = E [ x ( k ) ∣ Y ( k − 1 ) ] \hat{x}(k|k-1)=\text{E}[x(k)|Y(k-1)] x^(k∣k−1)=E[x(k)∣Y(k−1)] 表示根据前 k − 1 k-1 k−1 个时刻的观测数据对第 k k k 时刻的实际状态 x ( k ) x(k) x(k) 的预测值
- P ( k ∣ k − 1 ) = D [ x ( k ) − x ^ ( k ∣ k − 1 ) ] P(k|k-1)=\text{D}[x(k)-\hat{x}(k|k-1)] P(k∣k−1)=D[x(k)−x^(k∣k−1)] 称预测方差，或预测协方差矩阵
- P ( k ) = D [ x ( k ) − x ^ ( k ) ] P(k)=\text{D}[x(k)-\hat{x}(k)] P(k)=D[x(k)−x^(k)] 称估计方差，或估计协方差矩阵
- V = D v ( k ) V=Dv(k) V=Dv(k) 为状态噪声方差或协方差矩阵
- W = D w ( k ) W=Dw(k) W=Dw(k) 为观测噪声方差或协方差矩阵
几个性质及部分证明
贝叶斯最小均方误差估计量(Bmse)
观测到
x
x
x 后
θ
\theta
θ 的最小均方误差估计量
θ
^
\hat{\theta}
θ^ 为（[p255]）
θ
^
=
E
(
θ
∣
x
)
\hat{\theta}=\text{E}(\theta|x)
θ^=E(θ∣x)
证明的主要思路是求
θ
^
\hat{\theta}
θ^ 使得最小均方误差最小。证明：
Bmse
(
θ
^
)
=
E
[
(
θ
−
θ
^
)
2
]
=
∬
(
θ
−
θ
^
)
2
p
(
x
,
θ
)
d
x
d
θ
=
∫
[
∫
(
θ
−
θ
^
)
2
p
(
θ
∣
x
)
d
θ
]
p
(
x
)
d
x
J
=
∫
(
θ
−
θ
^
)
2
p
(
θ
∣
x
)
d
θ
∂
J
∂
θ
^
=
−
2
∫
θ
p
(
θ
∣
x
)
d
θ
+
2
θ
^
∫
p
(
θ
∣
x
)
d
θ
=
0
θ
^
=
2
∫
θ
p
(
θ
∣
x
)
d
θ
2
∫
p
(
θ
∣
x
)
d
θ
=
E
(
θ
∣
x
)
\begin{aligned} \text{Bmse}(\hat{\theta}) &= \text{E}[(\theta-\hat{\theta})^2] \\ &= \iint(\theta-\hat{\theta})^2p(x,\theta)\text{d}x\text{d}\theta \\ &=\int\left[\int(\theta-\hat{\theta})^2p(\theta|x)\text{d}\theta\right] p(x)\text{d}x \\ J &= \int(\theta-\hat{\theta})^2p(\theta|x)\text{d}\theta \\ \frac{\partial J}{\partial\hat{\theta}} &= -2\int\theta p(\theta|x)\text{d}\theta +2\hat{\theta}\int p(\theta|x)\text{d}\theta=0 \\ \hat{\theta} &= \frac{2\displaystyle\int\theta p(\theta|x)\text{d}\theta} {2\displaystyle\int p(\theta|x)\text{d}\theta}=\text{E}(\theta|x) \\ \end{aligned}
Bmse(θ^)J∂θ^∂Jθ^=E[(θ−θ^)2]=∬(θ−θ^)2p(x,θ)dxdθ=∫[∫(θ−θ^)2p(θ∣x)dθ]p(x)dx=∫(θ−θ^)2p(θ∣x)dθ=−2∫θp(θ∣x)dθ+2θ^∫p(θ∣x)dθ=0=2∫p(θ∣x)dθ2∫θp(θ∣x)dθ=E(θ∣x)
零均值应用定理(标量形式)
设
x
x
x和
y
y
y为联合正态分布的随机变量，则（[p264]）
E
(
y
∣
x
)
=
E
y
+
Cov
(
x
,
y
)
D
x
(
x
−
E
x
)
D
(
y
∣
x
)
=
D
y
−
Cov
2
(
x
,
y
)
D
x
\begin{aligned} & \text{E}(y|x)=\text{E}y+\frac{\text{Cov}(x,y)}{\text{D}x}(x-\text{E}x) \\ & \text{D}(y|x)=\text{D}y-\frac{\text{Cov}^2(x,y)}{\text{D}x} \\ \end{aligned}
E(y∣x)=Ey+DxCov(x,y)(x−Ex)D(y∣x)=Dy−DxCov2(x,y)
两式可另外写作
y
^
−
E
y
D
y
=
ρ
x
−
E
x
D
x
D
(
y
∣
x
)
=
D
y
(
1
−
ρ
2
)
\begin{aligned} & \frac{\hat{y}-\text{E}y}{\sqrt{\text{D}y}} =\rho\frac{x-\text{E}x}{\sqrt{\text{D}x}} \\ & \text{D}(y|x)=\text{D}y(1-\rho^2) \\ \end{aligned}
Dyy^−Ey=ρDxx−ExD(y∣x)=Dy(1−ρ2)
证明：
y
^
=
a
x
+
b
J
=
E
(
y
−
y
^
)
2
=
E
[
y
2
−
2
y
(
a
x
+
b
)
+
(
a
x
+
b
)
2
]
=
y
2
−
2
E
x
y
⋅
a
−
2
b
E
y
+
E
x
2
⋅
a
2
+
2
E
x
⋅
a
b
+
b
2
d
J
=
(
−
2
E
x
y
+
2
a
E
x
2
+
2
b
E
x
)
d
a
+
(
−
2
E
y
+
2
a
E
x
+
2
b
)
d
b
∂
J
∂
b
=
−
2
E
y
+
2
a
E
x
+
2
b
=
0
b
=
E
y
−
a
E
x
∂
J
∂
a
=
−
2
E
x
y
+
2
a
E
x
2
+
2
b
E
x
=
0
a
E
x
2
=
E
x
y
−
b
E
x
=
E
x
y
−
E
x
E
y
+
a
(
E
x
)
2
a
=
Cov
(
x
,
y
)
D
x
y
^
=
a
x
+
b
=
a
x
+
E
y
−
a
E
x
=
E
y
+
Cov
(
x
,
y
)
D
x
(
x
−
E
x
)
\begin{aligned} \hat{y} &= ax+b \\ J &= \text{E}(y-\hat{y})^2 \\ &= \text{E}[y^2-2y(ax+b)+(ax+b)^2] \\ &= y^2-2\text{E}xy\cdot a-2b\text{E}y +\text{E}x^2\cdot a^2+2\text{E}x\cdot ab+b^2 \\ \text{d}J &= (-2\text{E}xy+2a\text{E}x^2+2b\text{E}x)\text{d}a +(-2\text{E}y+2a\text{E}x+2b)\text{d}b \\ \frac{\partial J}{\partial b} &= -2\text{E}y+2a\text{E}x+2b = 0 \\ b &= \text{E}y-a\text{E}x \\ \frac{\partial J}{\partial a} &= -2\text{E}xy+2a\text{E}x^2+2b\text{E}x = 0 \\ a\text{E}x^2 &= \text{E}xy-b\text{E}x = \text{E}xy-\text{E}x\text{E}y+a(\text{E}x)^2 \\ a &= \frac{\text{Cov}(x,y)}{\text{D}x} \\ \hat{y} &= ax+b = ax+\text{E}y-a\text{E}x = \text{E}y+\frac{\text{Cov}(x,y)}{\text{D}x}(x-\text{E}x) \\ \end{aligned}
y^JdJ∂b∂Jb∂a∂JaEx2ay^=ax+b=E(y−y^)2=E[y2−2y(ax+b)+(ax+b)2]=y2−2Exy⋅a−2bEy+Ex2⋅a2+2Ex⋅ab+b2=(−2Exy+2aEx2+2bEx)da+(−2Ey+2aEx+2b)db=−2Ey+2aEx+2b=0=Ey−aEx=−2Exy+2aEx2+2bEx=0=Exy−bEx=Exy−ExEy+a(Ex)2=DxCov(x,y)=ax+b=ax+Ey−aEx=Ey+DxCov(x,y)(x−Ex)
该定理只对包括正态分布在内的满足线性关系的随机变量有效(具体什么地方满足线性暂时没搞清楚)。例如，对两个联合均匀分布
f
(
x
,
y
)
=
2
,
{
0
<
x
<
1
,
0
<
y
<
x
}
f
(
x
,
y
)
=
3
,
{
0
<
x
<
1
,
x
2
<
y
<
x
}
\begin{aligned} & f(x,y)=2,\{0<x<1,0<y<x\} \\ & f(x,y)=3,\{0<x<1,x^2<y<\sqrt{x}\} \end{aligned}
f(x,y)=2,{0<x<1,0<y<x}f(x,y)=3,{0<x<1,x2<y<x}
第一个成立，第二个由于非线性的存在而不成立，也就是说
y
y
y在数据
x
x
x下的线性贝叶斯估计量不是最佳估计量，线性贝叶斯估计量和最佳贝叶斯估计量分别为
y
^
=
E
y
+
Cov
(
x
,
y
)
D
x
(
x
−
E
x
)
=
133
x
+
9
153
y
^
=
E
(
y
∣
x
)
=
∫
∞
∞
y
f
(
y
∣
x
)
d
y
=
x
−
x
4
2
(
x
−
x
2
)
\begin{aligned} & \hat{y}=\text{E}y+\frac{\text{Cov}(x,y)}{\text{D}x}(x-\text{E}x) =\frac{133x+9}{153} \\ & \hat{y}=\text{E}(y|x)=\int_\infty^\infty yf(y|x)\text{d}y=\frac{x-x^4}{2(\sqrt x-x^2)} \\ \end{aligned}
y^=Ey+DxCov(x,y)(x−Ex)=153133x+9y^=E(y∣x)=∫∞∞yf(y∣x)dy=2(x−x2)x−x4
下面是最佳贝叶斯估计量（条件期望）的详细推导
f
(
x
)
=
∫
a
(
x
)
b
(
x
)
f
(
x
,
y
)
d
y
=
∫
x
2
x
3
d
y
=
3
(
x
−
x
2
)
,
{
x
∈
(
0
,
1
)
}
f
(
y
∣
x
)
=
f
(
x
,
y
)
f
(
x
)
=
3
,
{
x
∈
(
0
,
1
)
,
y
∈
(
x
2
,
x
)
}
3
(
x
−
x
2
)
,
{
x
∈
(
0
,
1
)
}
=
1
x
−
x
2
,
{
y
∈
(
x
2
,
x
)
}
E
(
y
∣
x
)
=
∫
a
(
x
)
b
(
x
)
y
f
(
y
∣
x
)
d
y
=
∫
x
2
x
y
x
−
x
2
d
y
=
[
y
2
2
(
x
−
x
2
)
]
x
2
x
=
x
−
x
4
2
(
x
−
x
2
)
\begin{aligned} & f(x)=\int_{a(x)}^{b(x)} f(x,y)\text{d}y =\int_{x^2}^{\sqrt x}3\text{d}y=3(\sqrt x-x^2),\quad\{x\in(0,1)\} \\ & f(y|x)=\frac{f(x,y)}{f(x)} =\frac{3,\quad\{x\in(0,1),y\in(x^2,\sqrt x)\}}{3(\sqrt x-x^2),\quad\{x\in(0,1)\}} =\frac{1}{\sqrt x-x^2},\quad\{y\in(x^2,\sqrt x)\} \\ & \text{E}(y|x)=\int_{a(x)}^{b(x)} yf(y|x)\text{d}y =\int_{x^2}^{\sqrt x}\frac{y}{\sqrt x-x^2}\text{d}y =\left[\frac{y^2}{2(\sqrt x-x^2)}\right]_{x^2}^{\sqrt x} =\frac{x-x^4}{2(\sqrt x-x^2)} \end{aligned}
f(x)=∫a(x)b(x)f(x,y)dy=∫x2x3dy=3(x−x2),{x∈(0,1)}f(y∣x)=f(x)f(x,y)=3(x−x2),{x∈(0,1)}3,{x∈(0,1),y∈(x2,x)}=x−x21,{y∈(x2,x)}E(y∣x)=∫a(x)b(x)yf(y∣x)dy=∫x2xx−x2ydy=[2(x−x2)y2]x2x=2(x−x2)x−x4
零均值应用定理(向量形式)
x
\boldsymbol{x}
x和
y
\boldsymbol{y}
y为联合正态分布的随机向量，
x
\boldsymbol{x}
x是m×1，
y
\boldsymbol{y}
y是n×1，分块协方差矩阵
C
=
[
C
x
x
C
x
y
C
y
x
C
y
y
]
\mathbf{C}=\left[\begin{matrix} \mathbf{C}_{xx} & \mathbf{C}_{xy} \\ \mathbf{C}_{yx} & \mathbf{C}_{yy} \end{matrix}\right]
C=[CxxCyxCxyCyy]
则
E
(
y
∣
x
)
=
E
(
y
)
+
C
y
x
C
x
x
−
1
(
x
−
E
(
x
)
)
\text{E}(\boldsymbol{y}|\boldsymbol{x})=\text{E}(\boldsymbol{y}) +\mathbf{C}_{yx}\mathbf{C}_{xx}^{-1}(\boldsymbol{x}-\text{E}(\boldsymbol{x}))
E(y∣x)=E(y)+CyxCxx−1(x−E(x))
其中
C
x
y
C_{xy}
Cxy表示
Cov
(
x
,
y
)
\text{Cov}(x,y)
Cov(x,y)。证明：
(其中省略的步骤见下文)
y
^
=
A
x
+
B
J
=
E
(
y
−
y
^
)
⊤
(
y
−
y
^
)
=
E
(
y
⊤
y
−
2
y
⊤
y
^
+
y
^
⊤
y
^
)
K
=
y
⊤
y
−
2
y
⊤
y
^
+
y
^
⊤
y
^
d
K
=
d
(
−
2
y
⊤
(
A
x
+
B
)
+
(
A
x
+
B
)
⊤
(
A
x
+
B
)
)
=
−
2
y
⊤
(
d
A
x
+
d
B
)
+
d
(
x
⊤
A
⊤
A
x
+
2
B
⊤
A
x
+
B
⊤
B
)
=
−
2
x
y
⊤
d
A
−
2
y
⊤
d
B
+
2
x
x
⊤
A
⊤
d
A
+
2
x
⊤
A
⊤
d
B
+
2
x
B
⊤
d
A
+
2
B
⊤
d
B
=
(
−
2
x
y
⊤
+
2
x
x
⊤
A
⊤
+
2
x
B
⊤
)
d
A
+
(
−
2
y
⊤
+
2
x
⊤
A
⊤
+
2
B
⊤
)
d
B
∂
J
∂
B
=
−
2
y
+
2
A
x
+
2
B
=
0
B
=
E
y
−
A
E
x
∂
J
∂
A
=
−
2
y
x
⊤
+
2
A
x
x
⊤
+
2
B
x
⊤
=
0
E
(
A
x
x
⊤
)
=
E
(
y
x
⊤
−
B
x
⊤
)
=
E
(
y
x
⊤
−
(
E
y
−
A
E
x
)
x
⊤
)
=
E
y
x
⊤
−
E
y
E
x
⊤
+
A
E
x
E
x
⊤
A
=
(
E
y
x
⊤
−
E
y
E
x
⊤
)
(
E
(
x
x
⊤
)
−
E
x
E
x
⊤
)
−
1
=
C
y
x
C
x
x
−
1
y
^
=
C
y
x
C
x
x
−
1
x
+
E
y
−
C
y
x
C
x
x
−
1
E
x
=
E
y
+
C
y
x
C
x
x
−
1
(
x
−
E
x
)
\begin{aligned} \hat{y} &= Ax+B \\ J &= \text{E}(y-\hat{y})^\top(y-\hat{y}) \\ &= \text{E}(y^\top y-2y^\top\hat{y}+\hat{y}^\top\hat{y}) \\ K &= y^\top y-2y^\top\hat{y}+\hat{y}^\top\hat{y} \\ \text{d}K &= \text{d}(-2y^\top(Ax+B)+(Ax+B)^\top(Ax+B)) \\ &= -2y^\top(\text{d}Ax+\text{d}B) +\text{d}(x^\top A^\top Ax+2B^\top Ax+B^\top B) \\ &= -2xy^\top\text{d}A-2y^\top\text{d}B+2xx^\top A^\top\text{d}A \\ &+ 2x^\top A^\top\text{d}B+2xB^\top\text{d}A+2B^\top\text{d}B \\ &= (-2xy^\top+2xx^\top A^\top+2xB^\top)\text{d}A +(-2y^\top+2x^\top A^\top+2B^\top)\text{d}B \\ \frac{\partial J}{\partial B} &= -2y+2Ax+2B =0\\ B &= \text{E}y-A\text{E}x \\ \frac{\partial J}{\partial A} &= -2yx^\top+2Axx^\top+2Bx^\top =0\\ \text{E}(Axx^\top) &= \text{E}(yx^\top-Bx^\top) \\ &= \text{E}(yx^\top-(\text{E}y-A\text{E}x)x^\top) \\ &= \text{E}yx^\top-\text{E}y\text{E}x^\top+A\text{E}x\text{E}x^\top \\ A &= (\text{E}yx^\top-\text{E}y\text{E}x^\top) (\text{E}(xx^\top)-\text{E}x\text{E}x^\top)^{-1} =C_{yx}C_{xx}^{-1} \\ \hat{y} &= C_{yx}C_{xx}^{-1}x+\text{E}y-C_{yx}C_{xx}^{-1}\text{E}x \\ &= \text{E}y+C_{yx}C_{xx}^{-1}(x-\text{E}x) \\ \end{aligned}
y^JKdK∂B∂JB∂A∂JE(Axx⊤)Ay^=Ax+B=E(y−y^)⊤(y−y^)=E(y⊤y−2y⊤y^+y^⊤y^)=y⊤y−2y⊤y^+y^⊤y^=d(−2y⊤(Ax+B)+(Ax+B)⊤(Ax+B))=−2y⊤(dAx+dB)+d(x⊤A⊤Ax+2B⊤Ax+B⊤B)=−2xy⊤dA−2y⊤dB+2xx⊤A⊤dA+2x⊤A⊤dB+2xB⊤dA+2B⊤dB=(−2xy⊤+2xx⊤A⊤+2xB⊤)dA+(−2y⊤+2x⊤A⊤+2B⊤)dB=−2y+2Ax+2B=0=Ey−AEx=−2yx⊤+2Axx⊤+2Bx⊤=0=E(yx⊤−Bx⊤)=E(yx⊤−(Ey−AEx)x⊤)=Eyx⊤−EyEx⊤+AExEx⊤=(Eyx⊤−EyEx⊤)(E(xx⊤)−ExEx⊤)−1=CyxCxx−1=CyxCxx−1x+Ey−CyxCxx−1Ex=Ey+CyxCxx−1(x−Ex)
其中矩阵求导的部分见 矩阵求导术(上)，其中的推导过程省略了迹的符号
tr
\text{tr}
tr，注意分辨。另外因为3个符号
d
\text{d}
d、
E
\text{E}
E、
tr
\text{tr}
tr均为线性算符，因此可交换计算顺序，推导中也省略了
E
\text{E}
E。
上面推导中用到的一些矩阵微分与求迹公式详细推导如下：
d
(
B
⊤
B
)
=
tr
(
d
B
⊤
B
+
B
⊤
d
B
)
=
tr
(
d
B
⊤
B
)
+
tr
(
B
⊤
d
B
)
=
tr
(
B
⊤
d
B
)
+
tr
(
B
⊤
d
B
)
=
tr
(
2
B
⊤
d
B
)
d
(
x
⊤
A
⊤
A
x
)
=
tr
(
x
⊤
d
(
A
⊤
A
)
x
)
=
tr
(
x
⊤
2
A
⊤
d
A
x
)
=
tr
(
2
x
x
⊤
A
⊤
d
A
)
d
(
B
⊤
A
x
)
=
tr
(
d
B
⊤
A
x
+
B
⊤
d
A
x
)
=
tr
(
x
⊤
A
⊤
d
B
+
x
B
⊤
d
A
)
\begin{aligned} \text{d}(B^\top B) &= \text{tr}(\text{d}B^\top B+B^\top\text{d}B) \\ &= \text{tr}(\text{d}B^\top B)+\text{tr}(B^\top\text{d}B) \\ &= \text{tr}(B^\top\text{d}B)+\text{tr}(B^\top\text{d}B) \\ &= \text{tr}(2B^\top\text{d}B) \\ \text{d}(x^\top A^\top Ax) &= \text{tr}(x^\top\text{d}(A^\top A)x) \\ &= \text{tr}(x^\top2A^\top\text{d}Ax) \\ &= \text{tr}(2xx^\top A^\top\text{d}A) \\ \text{d}(B^\top Ax) &= \text{tr}(\text{d}B^\top Ax+B^\top\text{d}Ax) \\ &= \text{tr}(x^\top A^\top\text{d}B+xB^\top\text{d}A) \\ \end{aligned}
d(B⊤B)d(x⊤A⊤Ax)d(B⊤Ax)=tr(dB⊤B+B⊤dB)=tr(dB⊤B)+tr(B⊤dB)=tr(B⊤dB)+tr(B⊤dB)=tr(2B⊤dB)=tr(x⊤d(A⊤A)x)=tr(x⊤2A⊤dAx)=tr(2xx⊤A⊤dA)=tr(dB⊤Ax+B⊤dAx)=tr(x⊤A⊤dB+xB⊤dA)
投影定理(正交原理)
当利用数据样本的线性组合来估计一个随机变量的时候，当估计值与真实值的误差和每一个数据样本正交时，该估计值是最佳估计量，即数据样本
x
x
x与最佳估计量
y
^
\hat{y}
y^满足
E
[
(
y
−
y
^
)
x
⊤
(
n
)
]
=
0
n
=
0
,
1
,
⋯
,
N
−
1
\text{E}[(y-\hat{y})x^\top(n)]=0\quad n=0,1,\cdots,N-1
E[(y−y^)x⊤(n)]=0n=0,1,⋯,N−1
零均值的随机变量满足内积空间中的性质。定义变量的长度
∣
∣
x
∣
∣
=
E
x
2
||x||=\sqrt{\text{E}x^2}
∣∣x∣∣=Ex2，变量
x
x
x和
y
y
y的内积
(
x
,
y
)
(x,y)
(x,y)定义为
E
(
x
y
)
\text{E}(xy)
E(xy)，两个变量的夹角定义为相关系数
ρ
\rho
ρ。当
E
(
x
y
)
=
0
\text{E}(xy)=0
E(xy)=0时称变量
x
x
x和
y
y
y正交。
均值不为零时，定义变量的长度
∣
∣
x
∣
∣
=
D
x
||x||=\sqrt{\text{D}x}
∣∣x∣∣=Dx，变量
x
x
x和
y
y
y的内积
(
x
,
y
)
(x,y)
(x,y)定义为
Cov
(
x
y
)
\text{Cov}(xy)
Cov(xy)，两个变量的夹角定义为相关系数
ρ
\rho
ρ。均值不为零的情况是我自己的猜测，很多资料都没有详细说明，但卡尔曼滤波的推导里全是均值非零的。
将
x
x
x和
y
y
y对应成数据的形式，即
x
(
0
)
x(0)
x(0)是
x
x
x，
x
(
1
)
x(1)
x(1)是
y
y
y，
x
^
(
1
∣
0
)
\hat{x}(1|0)
x^(1∣0)是
y
^
\hat{y}
y^，得到
E
[
(
x
(
1
)
−
x
^
(
1
∣
0
)
)
⊤
x
(
0
)
]
=
0
\text{E}[(x(1)-\hat{x}(1|0))^\top x(0)]=0
E[(x(1)−x^(1∣0))⊤x(0)]=0
其中
x
~
(
k
∣
k
−
1
)
=
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
\widetilde{x}(k|k-1) = x(k)-\hat{x}(k|k-1)
x
(k∣k−1)=x(k)−x^(k∣k−1)
称为新息(innovation)，与旧数据
x
(
0
)
x(0)
x(0)正交。
标量形式证明：
E
[
x
(
y
^
−
y
)
]
=
E
[
x
E
y
+
Cov
(
x
,
y
)
D
x
(
x
2
−
x
E
x
)
−
x
y
]
=
E
x
E
y
+
Cov
(
x
,
y
)
D
x
(
E
x
2
−
(
E
x
)
2
)
−
E
x
y
=
Cov
(
x
,
y
)
+
E
x
E
y
−
E
x
y
=
0
\begin{aligned} \text{E}[x(\hat{y}-y)] &= \text{E}[x\text{E}y +\frac{\text{Cov}(x,y)}{\text{D}x}(x^2-x\text{E}x)-xy] \\ &= \text{E}x\text{E}y +\frac{\text{Cov}(x,y)}{\text{D}x}(\text{E}x^2-(\text{E}x)^2)-\text{E}xy \\ &= \text{Cov}(x,y)+\text{E}x\text{E}y-\text{E}xy=0 \end{aligned}
E[x(y^−y)]=E[xEy+DxCov(x,y)(x2−xEx)−xy]=ExEy+DxCov(x,y)(Ex2−(Ex)2)−Exy=Cov(x,y)+ExEy−Exy=0
向量形式证明：
E
[
(
y
^
−
y
)
x
⊤
]
=
E
[
E
y
x
⊤
+
C
y
x
C
x
x
−
1
(
x
−
E
x
)
x
⊤
−
y
x
⊤
]
=
E
y
E
x
⊤
+
E
[
C
y
x
C
x
x
−
1
x
x
⊤
]
−
C
y
x
C
x
x
−
1
E
x
E
x
⊤
−
E
y
x
⊤
=
−
C
y
x
+
C
y
x
C
x
x
−
1
(
E
x
x
⊤
−
E
x
E
x
⊤
)
=
−
C
y
x
+
C
y
x
C
x
x
−
1
C
x
x
=
0
\begin{aligned} \text{E}[(\hat{y}-y)x^\top] &= \text{E}[\text{E}yx^\top+C_{yx}C_{xx}^{-1}(x-\text{E}x)x^\top-yx^\top] \\ &= \text{E}y\text{E}x^\top+\text{E}[C_{yx}C_{xx}^{-1}xx^\top] -C_{yx}C_{xx}^{-1}\text{E}x\text{E}x^\top-\text{E}yx^\top \\ &= -C_{yx}+C_{yx}C_{xx}^{-1}(\text{E}xx^\top-\text{E}x\text{E}x^\top) \\ &= -C_{yx}+C_{yx}C_{xx}^{-1}C_{xx} \\ &= 0 \end{aligned}
E[(y^−y)x⊤]=E[Eyx⊤+CyxCxx−1(x−Ex)x⊤−yx⊤]=EyEx⊤+E[CyxCxx−1xx⊤]−CyxCxx−1ExEx⊤−Eyx⊤=−Cyx+CyxCxx−1(Exx⊤−ExEx⊤)=−Cyx+CyxCxx−1Cxx=0
由于
E
(
y
−
y
^
)
=
0
\text{E}(y-\hat{y})=0
E(y−y^)=0，所以均值非零时正交条件也恰好成立。由图可得投影定理的另一个公式
E
[
(
y
−
y
^
)
y
^
]
=
0
\text{E}[(y-\hat{y})\hat{y}]=0
E[(y−y^)y^]=0
证明：
E
[
y
^
(
y
^
−
y
)
]
=
E
[
(
E
y
+
k
x
−
k
E
x
)
(
E
y
+
k
x
−
k
E
x
−
y
)
]
=
(
E
y
)
2
+
k
E
x
E
y
−
k
E
x
E
y
−
(
E
y
)
2
+
k
E
x
E
y
+
k
2
E
x
2
−
k
2
(
E
x
)
2
−
k
E
x
y
−
k
E
x
E
y
−
k
2
(
E
x
)
2
+
k
2
(
E
x
)
2
+
k
E
x
E
y
=
k
2
D
x
−
k
Cov
(
x
,
y
)
=
[
Cov
(
x
,
y
)
]
2
[
D
x
]
2
D
x
−
Cov
(
x
,
y
)
D
x
Cov
(
x
,
y
)
=
0
\begin{aligned} & \text{E}[\hat{y}(\hat{y}-y)] \\ =& \text{E}[(\text{E} y+k x-k \text{E} x)(\text{E} y+k x-k \text{E} x-y)] \\ =& (\text{E}y)^{2}+k\text{E}x\text{E}y-k\text{E}x\text{E}y-(\text{E}y)^{2}\\ &+ k\text{E}x\text{E}y+k^{2}Ex^2-k^{2}(\text{E}x)^{2}-k\text{E}xy \\ &- k\text{E}x\text{E}y-k^{2}(\text{E}x)^{2}+k^{2}(\text{E}x)^{2}+k\text{E}x\text{E}y \\ =& k^{2}\text{D}x-k\text{Cov}(x,y) \\ =& \frac{[\text{Cov}(x,y)]^2}{[\text{D}x]^2}\text{D}x-\frac{\text{Cov}(x,y)}{\text{D}x}\text{Cov}(x,y) \\ =&0 \end{aligned}
=====E[y^(y^−y)]E[(Ey+kx−kEx)(Ey+kx−kEx−y)](Ey)2+kExEy−kExEy−(Ey)2+kExEy+k2Ex2−k2(Ex)2−kExy−kExEy−k2(Ex)2+k2(Ex)2+kExEyk2Dx−kCov(x,y)[Dx]2[Cov(x,y)]2Dx−DxCov(x,y)Cov(x,y)0
期望可加性
E
[
y
1
+
y
2
∣
x
]
=
E
[
y
1
∣
x
]
+
E
[
y
2
∣
x
]
\text{E}[y_1+y_2|x]=\text{E}[y_1|x]+\text{E}[y_2|x]
E[y1+y2∣x]=E[y1∣x]+E[y2∣x]
独立条件可加性
若
x
1
x_1
x1和
x
2
x_2
x2独立，则
E
[
y
∣
x
1
,
x
2
]
=
E
[
y
∣
x
1
]
+
E
[
y
∣
x
2
]
−
E
y
\text{E}[y|x_1,x_2]=\text{E}[y|x_1]+\text{E}[y|x_2]-\text{E}y
E[y∣x1,x2]=E[y∣x1]+E[y∣x2]−Ey
证明：
令
x
=
[
x
1
⊤
,
x
2
⊤
]
⊤
x=[x_1^\top,x_2^\top]^\top
x=[x1⊤,x2⊤]⊤，则
C
x
x
−
1
=
[
C
x
1
x
1
C
x
1
x
2
C
x
2
x
1
C
x
2
x
2
]
−
1
=
[
C
x
1
x
1
−
1
O
O
C
x
2
x
2
−
1
]
C
y
x
=
[
C
y
x
1
C
y
x
2
]
E
(
y
∣
x
)
=
E
y
+
C
y
x
C
x
x
−
1
(
x
−
E
x
)
=
E
y
+
[
C
y
x
1
C
y
x
2
]
[
C
x
1
x
1
−
1
O
O
C
x
2
x
2
−
1
]
[
x
1
−
E
x
1
x
2
−
E
x
2
]
=
E
[
y
∣
x
1
]
+
E
[
y
∣
x
2
]
−
E
y
\begin{aligned} C_{xx}^{-1} &= \left[\begin{matrix} C_{x_1x_1} & C_{x_1x_2} \\ C_{x_2x_1} & C_{x_2x_2} \end{matrix}\right]^{-1} = \left[\begin{matrix} C_{x_1x_1}^{-1} & O \\ O & C_{x_2x_2}^{-1} \end{matrix}\right] \\ C_{yx} &= \left[\begin{matrix} C_{yx_1} & C_{yx_2} \end{matrix}\right] \\ \text{E}(y|x) &= \text{E}y+C_{yx}C_{xx}^{-1}(x-\text{E}x) \\ &= \text{E}y+\left[\begin{matrix} C_{yx_1} & C_{yx_2} \end{matrix}\right] \left[\begin{matrix} C_{x_1x_1}^{-1} & O \\ O & C_{x_2x_2}^{-1} \end{matrix}\right] \left[\begin{matrix} x_1-\text{E}x_1 \\ x_2-\text{E}x_2 \end{matrix}\right] \\ &= \text{E}[y|x_1]+\text{E}[y|x_2]-\text{E}y \end{aligned}
Cxx−1CyxE(y∣x)=[Cx1x1Cx2x1Cx1x2Cx2x2]−1=[Cx1x1−1OOCx2x2−1]=[Cyx1Cyx2]=Ey+CyxCxx−1(x−Ex)=Ey+[Cyx1Cyx2][Cx1x1−1OOCx2x2−1][x1−Ex1x2−Ex2]=E[y∣x1]+E[y∣x2]−Ey
非独立条件可加性(新息定理)
若
x
1
x_1
x1和
x
2
x_2
x2不独立，则根据投影定理取
x
2
x_2
x2与
x
1
x_1
x1独立的分量
x
~
2
\widetilde{x}_2
x
2，满足
E
[
y
∣
x
1
,
x
2
]
=
E
[
y
∣
x
1
,
x
~
2
]
=
E
[
y
∣
x
1
]
+
E
[
y
∣
x
~
2
]
−
E
y
\text{E}[y|x_1,x_2] =\text{E}[y|x_1,\widetilde{x}_2] =\text{E}[y|x_1]+\text{E}[y|\widetilde{x}_2]-\text{E}y
E[y∣x1,x2]=E[y∣x1,x
2]=E[y∣x1]+E[y∣x
2]−Ey
其中
x
~
2
=
x
2
−
x
^
2
=
x
2
−
E
(
x
2
∣
x
1
)
\widetilde{x}_2=x_2-\hat{x}_2=x_2-\text{E}(x_2|x_1)
x
2=x2−x^2=x2−E(x2∣x1)，由投影定理，
x
1
x_1
x1与
x
~
2
\widetilde{x}_2
x
2独立，
x
~
2
\widetilde{x}_2
x
2称为新息。
证明：
(下面的每个式子是先求部分后求整体，为便于理解可以从下往上看)
(标量情况下，
C
x
1
x
2
=
C
x
2
x
1
C_{x_1x_2}=C_{x_2x_1}
Cx1x2=Cx2x1)
Cov
(
y
,
x
^
2
)
=
E
[
y
(
E
x
2
+
Cov
(
x
2
,
x
1
)
D
x
1
(
x
1
−
E
x
1
)
)
]
−
E
y
E
x
^
2
=
E
y
E
x
2
+
Cov
(
x
2
,
x
1
)
D
x
1
(
E
x
1
y
−
E
x
1
E
y
)
−
E
y
E
x
2
=
C
x
1
x
2
C
y
x
1
D
x
1
Cov
(
x
2
,
x
^
2
)
=
E
x
2
x
^
2
−
E
x
2
E
x
^
2
=
E
[
x
2
(
E
x
2
+
Cov
(
x
2
,
x
1
)
D
x
1
(
x
1
−
E
x
1
)
)
]
−
(
E
x
2
)
2
=
Cov
(
x
2
,
x
1
)
D
x
1
(
E
x
2
x
1
−
E
x
2
E
x
1
)
=
C
x
1
x
2
2
D
x
1
D
x
^
2
=
D
(
E
x
2
+
Cov
(
x
2
,
x
1
)
D
x
1
(
x
1
−
E
x
1
)
)
=
C
x
1
x
2
2
(
D
x
1
)
2
D
x
1
=
C
x
1
x
2
2
D
x
1
D
x
~
2
=
D
x
2
+
D
x
^
2
−
2
Cov
(
x
2
,
x
^
2
)
=
D
x
2
−
C
x
1
x
2
2
D
x
1
E
[
y
∣
x
1
]
+
E
[
y
∣
x
~
2
]
−
E
y
=
E
y
+
C
x
1
y
D
x
1
(
x
1
−
E
x
1
)
+
Cov
(
y
,
x
~
2
)
D
x
~
2
(
x
~
2
−
E
x
~
2
)
=
OMIT
+
Cov
(
y
,
x
2
)
−
Cov
(
y
,
x
^
2
)
D
x
~
2
(
x
2
−
x
^
2
)
=
OMIT
+
C
y
x
2
−
C
x
1
x
2
C
y
x
1
D
x
1
D
x
2
−
C
x
1
x
2
2
D
x
1
(
x
2
−
E
2
−
Cov
(
x
2
,
x
1
)
D
x
1
(
x
1
−
E
1
)
)
=
OMIT
+
C
y
x
2
D
x
1
−
C
x
1
x
2
C
y
x
1
D
x
1
D
x
2
−
C
x
1
x
2
2
(
x
2
−
E
2
−
C
x
1
x
2
D
x
1
(
x
1
−
E
1
)
)
=
E
y
+
A
(
x
1
−
E
x
1
)
+
B
(
x
2
−
E
2
)
\begin{aligned} \text{Cov}(y,\hat{x}_2) &= \text{E}[y\left(\text{E}x_2 +\frac{\text{Cov}(x_2,x_1)}{\text{D}x_1}(x_1-\text{E}x_1)\right)] -\text{E}y\text{E}\hat{x}_2 \\ &= \text{E}y\text{E}x_2+\frac{\text{Cov}(x_2,x_1)}{\text{D}x_1} (\text{E}x_1y-\text{E}x_1\text{E}y)-\text{E}y\text{E}x_2 \\ &= \frac{C_{x_1x_2}C_{yx_1}}{\text{D}x_1} \\ \text{Cov}(x_2,\hat{x}_2) &= \text{E}x_2\hat{x}_2-\text{E}x_2\text{E}\hat{x}_2 \\ &= \text{E}[x_2(\text{E}x_2+\frac{\text{Cov}(x_2,x_1)}{\text{D}x_1} (x_1-\text{E}x_1))]-(\text{E}x_2)^2 \\ &= \frac{\text{Cov}(x_2,x_1)}{\text{D}x_1}(\text{E}x_2x_1-\text{E}x_2\text{E}x_1) \\ &= \frac{C_{x_1x_2}^2}{\text{D}x_1} \\ \text{D}\hat{x}_2 &= \text{D}(\text{E}x_2+\frac{\text{Cov}(x_2,x_1)}{\text{D}x_1} (x_1-\text{E}x_1)) \\ &= \frac{C_{x_1x_2}^2}{(\text{D}x_1)^2}\text{D}x_1 =\frac{C_{x_1x_2}^2}{\text{D}x_1} \\ \text{D}\widetilde{x}_2 &= \text{D}x_2+\text{D}\hat{x}_2 -2\text{Cov}(x_2,\hat{x}_2) \\ &= \text{D}x_2-\frac{C_{x_1x_2}^2}{\text{D}x_1} \\ \text{E}[y|x_1]+\text{E}[y|\widetilde{x}_2]-\text{E}y &= \text{E}y +\frac{C_{x_1y}}{\text{D}x_1}(x_1-\text{E}x_1) +\frac{\text{Cov}(y,\widetilde{x}_2)}{\text{D}\widetilde{x}_2} (\widetilde{x}_2-\text{E}\widetilde{x}_2) \\ &= \text{OMIT}+\frac{\text{Cov}(y,x_2)-\text{Cov}(y,\hat{x}_2)} {\text{D}\widetilde{x}_2}(x_2-\hat{x}_2) \\ &= \text{OMIT}+\frac{C_{yx_2}-\displaystyle\frac{C_{x_1x_2}C_{yx_1}}{\text{D}x_1}} {\text{D}x_2-\displaystyle\frac{C_{x_1x_2}^2}{\text{D}x_1}}(x_2-\text{E}_2 -\frac{\text{Cov}(x_2,x_1)}{\text{D}x_1}(x_1-\text{E}_1)) \\ &= \text{OMIT}+\frac{C_{yx_2}\text{D}x_1-C_{x_1x_2}C_{yx_1}} {\text{D}x_1\text{D}x_2-C_{x_1x_2}^2}(x_2-\text{E}_2 -\frac{C_{x_1x_2}}{\text{D}x_1}(x_1-\text{E}_1)) \\ &= \text{E}y+A(x_1-\text{E}x_1) +B(x_2-\text{E}_2) \\ \end{aligned}
Cov(y,x^2)Cov(x2,x^2)Dx^2Dx
2E[y∣x1]+E[y∣x
2]−Ey=E[y(Ex2+Dx1Cov(x2,x1)(x1−Ex1))]−EyEx^2=EyEx2+Dx1Cov(x2,x1)(Ex1y−Ex1Ey)−EyEx2=Dx1Cx1x2Cyx1=Ex2x^2−Ex2Ex^2=E[x2(Ex2+Dx1Cov(x2,x1)(x1−Ex1))]−(Ex2)2=Dx1Cov(x2,x1)(Ex2x1−Ex2Ex1)=Dx1Cx1x22=D(Ex2+Dx1Cov(x2,x1)(x1−Ex1))=(Dx1)2Cx1x22Dx1=Dx1Cx1x22=Dx2+Dx^2−2Cov(x2,x^2)=Dx2−Dx1Cx1x22=Ey+Dx1Cx1y(x1−Ex1)+Dx
2Cov(y,x
2)(x
2−Ex
2)=OMIT+Dx
2Cov(y,x2)−Cov(y,x^2)(x2−x^2)=OMIT+Dx2−Dx1Cx1x22Cyx2−Dx1Cx1x2Cyx1(x2−E2−Dx1Cov(x2,x1)(x1−E1))=OMIT+Dx1Dx2−Cx1x22Cyx2Dx1−Cx1x2Cyx1(x2−E2−Dx1Cx1x2(x1−E1))=Ey+A(x1−Ex1)+B(x2−E2)
其中
OMIT
\text{OMIT}
OMIT用于代替式
E
y
+
C
x
1
y
D
x
1
(
x
1
−
E
x
1
)
\text{E}y+\frac{C_{x_1y}}{\text{D}x_1}(x_1-\text{E}x_1)
Ey+Dx1Cx1y(x1−Ex1)
以及
B
=
C
y
x
2
D
x
1
−
C
x
1
x
2
C
y
x
1
D
x
1
D
x
2
−
C
x
1
x
2
2
A
=
C
x
1
y
D
x
1
−
B
C
x
1
x
2
D
x
1
=
C
x
1
y
D
x
1
(
D
x
1
D
x
2
−
C
x
1
x
2
2
)
−
C
x
1
x
2
D
x
1
(
C
y
x
2
D
x
1
−
C
x
1
x
2
C
y
x
1
)
D
x
1
D
x
2
−
C
x
1
x
2
2
=
C
x
1
y
D
x
2
−
C
x
1
x
2
C
y
x
2
D
x
1
D
x
2
−
C
x
1
x
2
2
\begin{aligned} B &= \frac{C_{yx_2}\text{D}x_1-C_{x_1x_2}C_{yx_1}} {\text{D}x_1\text{D}x_2-C_{x_1x_2}^2} \\ A &= \frac{C_{x_1y}}{\text{D}x_1}-B\frac{C_{x_1x_2}}{\text{D}x_1} \\ &= \frac{\displaystyle\frac{C_{x_1y}}{\text{D}x_1} (\text{D}x_1\text{D}x_2-C_{x_1x_2}^2) -\displaystyle\frac{C_{x_1x_2}}{\text{D}x_1} (C_{yx_2}\text{D}x_1-C_{x_1x_2}C_{yx_1})} {\text{D}x_1\text{D}x_2-C_{x_1x_2}^2} \\ &= \frac{C_{x_1y}\text{D}x_2-C_{x_1x_2}C_{yx_2}} {\text{D}x_1\text{D}x_2-C_{x_1x_2}^2} \\ \end{aligned}
BA=Dx1Dx2−Cx1x22Cyx2Dx1−Cx1x2Cyx1=Dx1Cx1y−BDx1Cx1x2=Dx1Dx2−Cx1x22Dx1Cx1y(Dx1Dx2−Cx1x22)−Dx1Cx1x2(Cyx2Dx1−Cx1x2Cyx1)=Dx1Dx2−Cx1x22Cx1yDx2−Cx1x2Cyx2
与另一个式子
E
(
y
∣
x
1
,
x
2
)
=
E
y
+
[
C
y
x
1
C
y
x
2
]
[
D
x
1
C
x
1
x
2
C
x
2
x
1
D
x
2
]
−
1
[
x
1
−
E
x
1
x
2
−
E
x
2
]
=
E
y
+
[
C
y
x
1
C
y
x
2
]
D
x
1
D
x
2
−
C
x
1
x
2
2
[
D
x
2
−
C
x
1
x
2
−
C
x
1
x
2
D
x
1
]
[
x
1
−
E
x
1
x
2
−
E
x
2
]
\begin{aligned} \text{E}(y|x_1,x_2) &= \text{E}y+\left[ \begin{matrix} C_{yx_1} & C_{yx_2} \end{matrix}\right] \left[\begin{matrix} \text{D}x_1 & C_{x_1x_2} \\ C_{x_2x_1} & \text{D}x_2 \end{matrix}\right]^{-1} \left[\begin{matrix} x_1-\text{E}x_1 \\ x_2-\text{E}x_2 \end{matrix}\right] \\ &= \text{E}y+\frac{\left[ \begin{matrix} C_{yx_1} & C_{yx_2} \end{matrix}\right]} {\text{D}x_1\text{D}x_2-C^2_{x_1x_2}} \left[\begin{matrix} \text{D}x_2 & -C_{x_1x_2} \\ -C_{x_1x_2} & \text{D}x_1 \end{matrix}\right] \left[\begin{matrix} x_1-\text{E}x_1 \\ x_2-\text{E}x_2 \end{matrix}\right] \\ \end{aligned}
E(y∣x1,x2)=Ey+[Cyx1Cyx2][Dx1Cx2x1Cx1x2Dx2]−1[x1−Ex1x2−Ex2]=Ey+Dx1Dx2−Cx1x22[Cyx1Cyx2][Dx2−Cx1x2−Cx1x2Dx1][x1−Ex1x2−Ex2]
中对应的
A
A
A和
B
B
B相等。
高斯白噪声中的直流电平
这个例子可以作为铺垫，有助于理解卡尔曼滤波各个公式的来源，比如
x
(
k
)
x(k)
x(k)和
x
(
k
−
1
)
x(k-1)
x(k−1)之间为什么还要有一个
x
^
(
k
∣
k
−
1
)
\hat{x}(k|k-1)
x^(k∣k−1)等。考虑模型
x
(
k
)
=
A
+
w
(
k
)
x(k)=A+w(k)
x(k)=A+w(k)
其中
A
A
A是待估计参数，
w
(
k
)
w(k)
w(k)是均值为0、方差为
σ
2
\sigma^2
σ2的高斯白噪声，
x
(
k
)
x(k)
x(k)是观测。这里需要注意的是，
A
A
A也是一个先验随机变量，也就是说在测量
A
A
A之前，预先猜测比方说
A
A
A应该在10左右，大概率不超过7~13的范围，因此假设
A
∼
N
(
10
,
1
)
A\sim N(10,1)
A∼N(10,1)，然后开始测量。
一开始可以得到
x
^
(
0
)
=
x
(
0
)
\hat{x}(0)=x(0)
x^(0)=x(0)，然后根据
x
(
0
)
x(0)
x(0)和
x
(
1
)
x(1)
x(1)预测
k
=
1
k=1
k=1时刻的值
E
[
x
(
1
)
∣
x
(
1
)
,
x
(
0
)
]
\text{E}[x(1)|x(1),x(0)]
E[x(1)∣x(1),x(0)]时，需要用到联合正态分布的条件可加性，但由于
x
(
1
)
x(1)
x(1)和
x
(
0
)
x(0)
x(0)不独立，需要使用投影定理计算出两个独立的变量
x
(
0
)
x(0)
x(0)与
x
~
(
1
∣
0
)
\widetilde{x}(1|0)
x
(1∣0)，进而计算
x
^
(
1
)
\hat{x}(1)
x^(1)，即
x
^
(
1
)
=
E
[
x
(
1
)
∣
x
(
1
)
,
x
(
0
)
]
=
E
[
x
(
1
)
∣
x
(
0
)
,
x
~
(
1
∣
0
)
]
=
E
[
x
(
1
)
∣
x
(
0
)
]
+
E
[
x
(
1
)
∣
x
~
(
1
∣
0
)
]
−
E
x
(
1
)
\begin{aligned} \hat{x}(1) &= \text{E}[x(1)|x(1),x(0)] \\ &= \text{E}[x(1)|x(0),\widetilde{x}(1|0)] \\ &= \text{E}[x(1)|x(0)]+\text{E}[x(1)|\widetilde{x}(1|0)]-\text{E}x(1) \end{aligned}
x^(1)=E[x(1)∣x(1),x(0)]=E[x(1)∣x(0),x
(1∣0)]=E[x(1)∣x(0)]+E[x(1)∣x
(1∣0)]−Ex(1)
其中
E
[
x
(
1
)
∣
x
(
0
)
]
=
x
^
(
1
∣
0
)
\text{E}[x(1)|x(0)]=\hat{x}(1|0)
E[x(1)∣x(0)]=x^(1∣0)，
x
^
(
1
∣
0
)
=
E
x
(
1
)
+
Cov
(
x
(
1
)
,
x
(
0
)
)
D
x
(
0
)
(
x
(
0
)
−
E
x
(
0
)
)
=
A
+
E
(
A
+
w
(
1
)
)
(
A
+
w
(
0
)
)
−
E
(
A
+
w
(
1
)
)
E
(
A
+
w
(
0
)
)
E
(
A
+
w
(
1
)
)
2
−
[
E
(
A
+
w
(
1
)
)
]
2
(
x
(
0
)
−
A
)
=
A
\begin{aligned} \hat{x}(1|0) &= \text{E}x(1)+\frac{\text{Cov}(x(1),x(0))} {\text{D}x(0)}(x(0)-\text{E}x(0)) \\ &= A+\frac{\text{E}(A+w(1))(A+w(0))-\text{E}(A+w(1))\text{E}(A+w(0))} {\text{E}(A+w(1))^2-[\text{E}(A+w(1))]^2}(x(0)-A) \\ &= A \end{aligned}
x^(1∣0)=Ex(1)+Dx(0)Cov(x(1),x(0))(x(0)−Ex(0))=A+E(A+w(1))2−[E(A+w(1))]2E(A+w(1))(A+w(0))−E(A+w(1))E(A+w(0))(x(0)−A)=A
此时式中就出现了
x
^
(
k
∣
k
−
1
)
\hat{x}(k|k-1)
x^(k∣k−1)，和另一个未知式
E
[
x
(
1
)
∣
x
~
(
1
∣
0
)
]
\text{E}[x(1)|\widetilde{x}(1|0)]
E[x(1)∣x
(1∣0)]。由零均值应用定理，
E
[
x
(
1
)
∣
x
~
(
1
∣
0
)
]
=
E
x
(
0
)
+
Cov
(
x
(
1
)
,
x
~
(
1
∣
0
)
)
D
x
~
(
1
∣
0
)
(
x
~
(
1
∣
0
)
−
E
x
~
(
1
∣
0
)
)
\text{E}[x(1)|\widetilde{x}(1|0)] = \text{E}x(0)+\frac{\text{Cov}(x(1),\widetilde{x}(1|0))} {\text{D}\widetilde{x}(1|0)}(\widetilde{x}(1|0)-\text{E}\widetilde{x}(1|0))
E[x(1)∣x
(1∣0)]=Ex(0)+Dx
(1∣0)Cov(x(1),x
(1∣0))(x
(1∣0)−Ex
(1∣0))
其中
x
~
(
1
∣
0
)
=
x
(
1
)
−
x
^
(
1
∣
0
)
E
x
~
(
1
∣
0
)
=
E
(
x
−
x
^
)
=
0
D
x
~
(
1
∣
0
)
=
E
(
x
(
1
)
−
x
^
(
1
∣
0
)
)
2
=
E
(
A
+
w
(
1
)
)
2
=
a
2
P
(
0
)
+
σ
2
=
P
(
1
∣
0
)
\begin{aligned} \widetilde{x}(1|0) &= x(1)-\hat{x}(1|0) \\ \text{E}\widetilde{x}(1|0) &= \text{E}(x-\hat{x}) = 0 \\ \text{D}\widetilde{x}(1|0) &= \text{E}(x(1)-\hat{x}(1|0))^2 \\ &= \text{E}(A+w(1))^2 \\ &= a^2P(0)+\sigma^2 \\ &= P(1|0) \end{aligned}
x
(1∣0)Ex
(1∣0)Dx
(1∣0)=x(1)−x^(1∣0)=E(x−x^)=0=E(x(1)−x^(1∣0))2=E(A+w(1))2=a2P(0)+σ2=P(1∣0)
卡尔曼滤波正式推导
标量形式
x
^
(
k
∣
k
−
1
)
=
E
[
x
(
k
)
∣
Y
(
k
−
1
)
]
=
E
[
A
x
(
k
−
1
)
+
B
u
(
k
−
1
)
+
v
(
k
−
1
)
∣
Y
(
k
−
1
)
]
=
A
E
[
x
(
k
−
1
)
∣
Y
(
k
−
1
)
]
+
B
E
[
u
(
k
−
1
)
∣
Y
(
k
−
1
)
]
+
E
[
v
(
k
−
1
)
∣
Y
(
k
−
1
)
]
=
A
x
^
(
k
−
1
)
+
B
u
(
k
−
1
)
x
~
(
k
∣
k
−
1
)
=
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
=
[
A
x
(
k
−
1
)
+
B
u
(
k
−
1
)
+
v
(
k
−
1
)
]
−
[
A
x
^
(
k
−
1
)
+
B
u
(
k
−
1
)
]
=
A
x
~
(
k
−
1
)
+
v
(
k
−
1
)
y
^
(
k
∣
k
−
1
)
=
E
[
y
(
k
)
∣
Y
(
k
−
1
)
]
=
E
[
C
x
(
k
)
+
w
(
k
−
1
)
∣
Y
(
k
−
1
)
]
=
C
x
^
(
k
∣
k
−
1
)
\begin{aligned} \hat{x}(k|k-1) &= \text{E}[x(k)|Y(k-1)] \\ &= \text{E}[Ax(k-1)+Bu(k-1)+v(k-1)|Y(k-1)] \\ &= A\text{E}[x(k-1)|Y(k-1)]+B\text{E}[u(k-1)|Y(k-1)]+\text{E}[v(k-1)|Y(k-1)] \\ &= A\hat{x}(k-1)+Bu(k-1) \\ \widetilde{x}(k|k-1) &= x(k)-\hat{x}(k|k-1) \\ &= [Ax(k-1)+Bu(k-1)+v(k-1)]-[A\hat{x}(k-1)+Bu(k-1)] \\ &= A\widetilde{x}(k-1)+v(k-1) \\ \hat{y}(k|k-1) &= \text{E}[y(k)|Y(k-1)] \\ &= \text{E}[Cx(k)+w(k-1)|Y(k-1)] \\ &= C\hat{x}(k|k-1) \\ \end{aligned}
x^(k∣k−1)x
(k∣k−1)y^(k∣k−1)=E[x(k)∣Y(k−1)]=E[Ax(k−1)+Bu(k−1)+v(k−1)∣Y(k−1)]=AE[x(k−1)∣Y(k−1)]+BE[u(k−1)∣Y(k−1)]+E[v(k−1)∣Y(k−1)]=Ax^(k−1)+Bu(k−1)=x(k)−x^(k∣k−1)=[Ax(k−1)+Bu(k−1)+v(k−1)]−[Ax^(k−1)+Bu(k−1)]=Ax
(k−1)+v(k−1)=E[y(k)∣Y(k−1)]=E[Cx(k)+w(k−1)∣Y(k−1)]=Cx^(k∣k−1)
由条件可加性，
x
^
(
k
)
=
E
[
x
(
k
)
∣
Y
(
k
)
]
=
E
[
x
(
k
)
∣
Y
(
k
−
1
)
,
y
(
k
)
]
=
E
[
x
(
k
)
∣
Y
(
k
−
1
)
,
y
~
(
k
∣
k
−
1
)
]
=
E
[
x
(
k
)
∣
Y
(
k
−
1
)
]
+
E
[
x
(
k
)
∣
y
~
(
k
∣
k
−
1
)
]
−
E
x
(
k
)
=
x
^
(
k
∣
k
−
1
)
+
E
[
x
(
k
)
∣
y
~
(
k
∣
k
−
1
)
]
−
E
x
(
k
)
\begin{aligned} \hat{x}(k) &= \text{E}[x(k)|Y(k)] \\ &= \text{E}[x(k)|Y(k-1),y(k)] \\ &= \text{E}[x(k)|Y(k-1),\widetilde{y}(k|k-1)] \\ &= \text{E}[x(k)|Y(k-1)]+\text{E}[x(k)|\widetilde{y}(k|k-1)]-\text{E}x(k) \\ &= \hat{x}(k|k-1)+\text{E}[x(k)|\widetilde{y}(k|k-1)]-\text{E}x(k) \end{aligned}
x^(k)=E[x(k)∣Y(k)]=E[x(k)∣Y(k−1),y(k)]=E[x(k)∣Y(k−1),y
(k∣k−1)]=E[x(k)∣Y(k−1)]+E[x(k)∣y
(k∣k−1)]−Ex(k)=x^(k∣k−1)+E[x(k)∣y
(k∣k−1)]−Ex(k)
由零均值应用定理，
E
[
x
(
k
)
∣
y
~
(
k
∣
k
−
1
)
]
=
E
x
(
k
)
+
Cov
(
x
(
k
)
,
y
~
(
k
∣
k
−
1
)
)
D
y
~
(
k
∣
k
−
1
)
(
y
~
(
k
∣
k
−
1
)
−
E
y
~
(
k
∣
k
−
1
)
)
\text{E}[x(k)|\widetilde{y}(k|k-1)] = \text{E}x(k)+\frac{\text{Cov}(x(k),\widetilde{y}(k|k-1))} {\text{D}\widetilde{y}(k|k-1)}(\widetilde{y}(k|k-1)-\text{E}\widetilde{y}(k|k-1))
E[x(k)∣y
(k∣k−1)]=Ex(k)+Dy
(k∣k−1)Cov(x(k),y
(k∣k−1))(y
(k∣k−1)−Ey
(k∣k−1))
其中
E
y
~
(
k
∣
k
−
1
)
=
E
(
y
−
y
^
)
=
0
P
(
k
∣
k
−
1
)
=
D
x
~
(
k
∣
k
−
1
)
=
E
(
a
x
~
(
k
−
1
)
+
v
(
k
−
1
)
)
2
=
a
2
P
(
k
−
1
)
+
V
\begin{aligned} \text{E}\widetilde{y}(k|k-1) &= \text{E}(y-\hat{y}) = 0 \\ P(k|k-1) &= \text{D}\widetilde{x}(k|k-1) \\ &= \text{E}(a\widetilde{x}(k-1)+v(k-1))^2 \\ &= a^2P(k-1)+V \end{aligned}
Ey
(k∣k−1)P(k∣k−1)=E(y−y^)=0=Dx
(k∣k−1)=E(ax
(k−1)+v(k−1))2=a2P(k−1)+V
令
K
(
k
)
=
Cov
(
x
(
k
)
,
y
~
(
k
∣
k
−
1
)
)
D
y
~
(
k
∣
k
−
1
)
K(k)=\frac{\text{Cov}(x(k),\widetilde{y}(k|k-1))} {\text{D}\widetilde{y}(k|k-1)}
K(k)=Dy
(k∣k−1)Cov(x(k),y
(k∣k−1))
则
E
[
x
(
k
)
∣
y
~
(
k
∣
k
−
1
)
]
=
E
x
(
k
)
+
K
(
k
)
(
y
(
k
)
−
y
^
(
k
∣
k
−
1
)
)
x
^
(
k
)
=
x
^
(
k
∣
k
−
1
)
+
E
[
x
(
k
)
∣
y
~
(
k
∣
k
−
1
)
]
−
E
x
(
k
)
=
x
^
(
k
∣
k
−
1
)
+
K
(
k
)
(
y
(
k
)
−
C
x
^
(
k
∣
k
−
1
)
)
\begin{aligned} \text{E}[x(k)|\widetilde{y}(k|k-1)] &= \text{E}x(k)+K(k)(y(k)-\hat{y}(k|k-1)) \\ \hat{x}(k)&= \hat{x}(k|k-1)+\text{E}[x(k)|\widetilde{y}(k|k-1)]-\text{E}x(k) \\ &= \hat{x}(k|k-1)+K(k)(y(k)-C\hat{x}(k|k-1)) \end{aligned}
E[x(k)∣y
(k∣k−1)]x^(k)=Ex(k)+K(k)(y(k)−y^(k∣k−1))=x^(k∣k−1)+E[x(k)∣y
(k∣k−1)]−Ex(k)=x^(k∣k−1)+K(k)(y(k)−Cx^(k∣k−1))
下面计算其中的
K
(
k
)
K(k)
K(k)，
D
y
~
(
k
∣
k
−
1
)
=
D
[
C
x
(
k
)
+
w
(
k
)
−
C
x
^
(
k
∣
k
−
1
)
]
=
C
2
D
[
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
]
+
D
w
(
k
)
=
C
2
P
(
k
∣
k
−
1
)
+
W
Cov
(
x
(
k
)
,
y
~
(
k
∣
k
−
1
)
)
=
Cov
(
x
(
k
)
,
C
x
(
k
)
+
w
(
k
)
−
C
x
^
(
k
∣
k
−
1
)
)
=
C
Cov
(
x
(
k
)
,
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
=
C
E
[
x
(
k
)
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
]
−
C
E
x
(
k
)
E
[
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
]
=
C
E
[
x
(
k
)
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
−
x
^
(
k
∣
k
−
1
)
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
]
=
C
E
[
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
2
]
=
C
P
(
k
∣
k
−
1
)
K
(
k
)
=
Cov
(
x
(
k
)
,
y
~
(
k
∣
k
−
1
)
)
D
y
~
(
k
∣
k
−
1
)
=
C
P
(
k
∣
k
−
1
)
C
2
P
(
k
∣
k
−
1
)
+
W
\begin{aligned} \text{D}\widetilde{y}(k|k-1) &= \text{D}[Cx(k)+w(k)-C\hat{x}(k|k-1)] \\ &= C^2\text{D}[x(k)-\hat{x}(k|k-1)]+\text{D}w(k) \\ &= C^2P(k|k-1)+W \\ \text{Cov}(x(k),\widetilde{y}(k|k-1)) &= \text{Cov}(x(k),Cx(k)+w(k)-C\hat{x}(k|k-1)) \\ &= C\text{Cov}(x(k),x(k)-\hat{x}(k|k-1)) \\ &= C\text{E}[x(k)(x(k)-\hat{x}(k|k-1))] -C\text{E}x(k)\text{E}[(x(k)-\hat{x}(k|k-1))] \\ &= C\text{E}[x(k)(x(k)-\hat{x}(k|k-1)) -\hat{x}(k|k-1)(x(k)-\hat{x}(k|k-1))] \\ &= C\text{E}[(x(k)-\hat{x}(k|k-1))^2] \\ &= CP(k|k-1) \\ K(k) &= \frac{\text{Cov}(x(k),\widetilde{y}(k|k-1))} {\text{D}\widetilde{y}(k|k-1)} \\ &= \frac{CP(k|k-1)}{C^2P(k|k-1)+W} \\ \end{aligned}
Dy
(k∣k−1)Cov(x(k),y
(k∣k−1))K(k)=D[Cx(k)+w(k)−Cx^(k∣k−1)]=C2D[x(k)−x^(k∣k−1)]+Dw(k)=C2P(k∣k−1)+W=Cov(x(k),Cx(k)+w(k)−Cx^(k∣k−1))=CCov(x(k),x(k)−x^(k∣k−1))=CE[x(k)(x(k)−x^(k∣k−1))]−CEx(k)E[(x(k)−x^(k∣k−1))]=CE[x(k)(x(k)−x^(k∣k−1))−x^(k∣k−1)(x(k)−x^(k∣k−1))]=CE[(x(k)−x^(k∣k−1))2]=CP(k∣k−1)=Dy
(k∣k−1)Cov(x(k),y
(k∣k−1))=C2P(k∣k−1)+WCP(k∣k−1)
其中显然
C
E
x
(
k
)
E
[
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
]
=
0
C\text{E}x(k)\text{E}[(x(k)-\hat{x}(k|k-1))]=0
CEx(k)E[(x(k)−x^(k∣k−1))]=0，而
x
^
(
k
∣
k
−
1
)
(
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
]
\hat{x}(k|k-1)((x(k)-\hat{x}(k|k-1))]
x^(k∣k−1)((x(k)−x^(k∣k−1))]是由投影定理的下面两个公式中的第二个
E
[
y
(
x
−
x
^
)
]
=
0
,
E
[
x
^
(
x
−
x
^
)
]
=
0
\text{E}[y(x-\hat{x})]=0,\quad\text{E}[\hat{x}(x-\hat{x})]=0
E[y(x−x^)]=0,E[x^(x−x^)]=0
下面出于递推需要而计算
P
(
k
)
P(k)
P(k)
P
(
k
)
=
D
[
x
(
k
)
−
x
^
(
k
)
]
=
D
[
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
−
K
(
k
)
(
y
(
k
)
−
C
x
^
(
k
∣
k
−
1
)
)
]
=
D
[
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
−
K
(
k
)
C
(
x
(
k
)
−
x
^
(
k
∣
k
−
1
)
)
−
K
(
k
)
w
]
=
D
[
(
1
−
K
(
k
)
C
)
x
~
−
K
(
k
)
w
]
=
(
1
−
K
(
k
)
C
)
2
P
(
k
∣
k
−
1
)
+
K
2
(
k
)
W
=
(
W
C
2
P
(
k
∣
k
−
1
)
+
W
)
2
P
(
k
∣
k
−
1
)
+
(
C
P
(
k
∣
k
−
1
)
C
2
P
(
k
∣
k
−
1
)
+
W
)
2
W
=
W
2
P
(
k
∣
k
−
1
)
+
C
2
P
2
(
k
∣
k
−
1
)
W
(
C
2
P
(
k
∣
k
−
1
)
+
W
)
2
=
W
P
(
k
∣
k
−
1
)
C
2
P
(
k
∣
k
−
1
)
+
W
=
(
1
−
K
(
k
)
C
)
P
(
k
∣
k
−
1
)
\begin{aligned} P(k) &= \text{D}[x(k)-\hat{x}(k)] \\ &= \text{D}[x(k)-\hat{x}(k|k-1)-K(k)(y(k)-C\hat{x}(k|k-1))] \\ &= \text{D}[x(k)-\hat{x}(k|k-1)-K(k)C(x(k)-\hat{x}(k|k-1))-K(k)w] \\ &= \text{D}[(1-K(k)C)\widetilde{x}-K(k)w] \\ &= (1-K(k)C)^2P(k|k-1)+K^2(k)W \\ &= \left(\frac{W}{C^2P(k|k-1)+W}\right)^2P(k|k-1) +\left(\frac{CP(k|k-1)}{C^2P(k|k-1)+W}\right)^2W \\ &= \frac{W^2P(k|k-1)+C^2P^2(k|k-1)W}{(C^2P(k|k-1)+W)^2} \\ &= \frac{WP(k|k-1)}{C^2P(k|k-1)+W} \\ &= (1-K(k)C)P(k|k-1) \\ \end{aligned}
P(k)=D[x(k)−x^(k)]=D[x(k)−x^(k∣k−1)−K(k)(y(k)−Cx^(k∣k−1))]=D[x(k)−x^(k∣k−1)−K(k)C(x(k)−x^(k∣k−1))−K(k)w]=D[(1−K(k)C)x
−K(k)w]=(1−K(k)C)2P(k∣k−1)+K2(k)W=(C2P(k∣k−1)+WW)2P(k∣k−1)+(C2P(k∣k−1)+WCP(k∣k−1))2W=(C2P(k∣k−1)+W)2W2P(k∣k−1)+C2P2(k∣k−1)W=C2P(k∣k−1)+WWP(k∣k−1)=(1−K(k)C)P(k∣k−1)
向量形式
x
^
(
k
∣
k
−
1
)
=
E
[
x
(
k
)
∣
Y
(
k
−
1
)
]
=
E
[
A
x
(
k
−
1
)
+
B
u
(
k
−
1
)
+
v
(
k
−
1
)
∣
Y
(
k
−
1
)
]
=
A
E
[
x
(
k
−
1
)
∣
Y
(
k
−
1
)
]
+
B
E
[
u
(
k
−
1
)
∣
Y
(
k
−
1
)
]
+
E
[
v
(
k
−
1
)
∣
Y
(
k
−
1
)
]
=
A
x
^
(
k
−
1
)
+
B
u
(
k
−
1
)
\begin{aligned} \hat{x}(k|k-1) &= \text{E}[x(k)|Y(k-1)] \\ &= \text{E}[Ax(k-1)+Bu(k-1)+v(k-1)|Y(k-1)] \\ &= A\text{E}[x(k-1)|Y(k-1)]+B\text{E}[u(k-1)|Y(k-1)]+\text{E}[v(k-1)|Y(k-1)] \\ &= A\hat{x}(k-1)+Bu(k-1) \\ \end{aligned}
x^(k∣k−1)=E[x(k)∣Y(k−1)]=E[Ax(k−1)+Bu(k−1)+v(k−1)∣Y(k−1)]=AE[x(k−1)∣Y(k−1)]+BE[u(k−1)∣Y(k−1)]+E[v(k−1)∣Y(k−1)]=Ax^(k−1)+Bu(k−1)
参考
- Steven M.Kay, 罗鹏飞. 统计信号处理基础[M]. 电子工业出版社, 2014.
- 孙增圻. 计算机控制理论与应用[M]. 清华大学出版社, 2008.
- 赵树杰, 赵建勋. 信号检测与估计理论[M]. 电子工业出版社, 2013.
- 卡尔曼滤波的推导过程详解
344
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
