Kalman Filters
Dynamical Signal Models
一阶高斯-马尔可夫过程(first-order Gauss-Markov process):描述采样点之间(相邻)的相关性:
s
[
n
]
=
a
s
[
n
−
1
]
+
u
[
n
]
(1)
s[n] = as[n-1] + u[n] \tag{1}
s[n]=as[n−1]+u[n](1)
其中 u [ n ] u[n] u[n]是高斯白噪声(White Gaussian Noise, WGN),方差为 σ u 2 \sigma^2_u σu2, s [ − 1 ] ∼ N ( μ s , σ s 2 ) s[-1]\sim \mathcal{N}(\mu_s,\sigma^2_s) s[−1]∼N(μs,σs2), s [ − 1 ] s[-1] s[−1]与 u [ n ] u[n] u[n]相互独立( ∀ n ≥ 0 \forall n \geq 0 ∀n≥0)。该模型也常被称为dynamical model / state model
将
s
[
n
]
s[n]
s[n]表示为初始条件
s
[
−
1
]
s[-1]
s[−1]的函数形式:
s
[
0
]
=
a
s
[
−
1
]
+
u
[
0
]
s
[
1
]
=
a
s
[
0
]
+
u
[
1
]
=
a
2
s
[
−
1
]
+
a
u
[
0
]
+
u
[
1
]
etc.
\begin{aligned} s[0] &= as[-1] + u[0] \\ s[1] &= as[0] + u[1] \\ &= a^2 s[-1] + au[0] + u[1] \\ & \text{etc.} \end{aligned}
s[0]s[1]=as[−1]+u[0]=as[0]+u[1]=a2s[−1]+au[0]+u[1]etc.
一般地,我们有
s
[
n
]
=
a
n
+
1
s
[
−
1
]
+
∑
k
=
0
n
a
k
u
[
n
−
k
]
(2)
s[n] = a^{n+1} s[-1] + \sum_{k=0}^n a^k u[n-k] \tag{2}
s[n]=an+1s[−1]+k=0∑naku[n−k](2)
高斯随机过程(Gaussian random process):给定 k k k,对任意的采样点 { s [ n 1 ] , ⋯ , s [ n k ] } \{s[n_1],\cdots,s[n_k]\} {s[n1],⋯,s[nk]}, k k k维随机向量 s = [ s [ n 1 ] , ⋯ , s [ n k ] ] T \boldsymbol{s} = [s[n_1],\cdots,s[n_k]]^T s=[s[n1],⋯,s[nk]]T的分布为高维的高斯PDF,就认为 s [ n ] s[n] s[n]是一个高斯随机过程。
因为随机变量
s
[
−
1
]
s[-1]
s[−1]和
u
[
⋅
]
u[\cdot]
u[⋅]都是高斯随机变量并且相互独立,不难看出,
s
[
n
]
s[n]
s[n]是一个高斯随机过程。另外,
E
[
s
[
n
]
]
=
a
n
+
1
E
[
s
[
−
1
]
]
=
a
n
+
1
μ
s
\mathbb{E} [s[n]] = a^{n+1} \mathbb{E}[s[-1]]=a^{n+1} \mu_s
E[s[n]]=an+1E[s[−1]]=an+1μs
采样点
s
[
m
]
s[m]
s[m]与
s
[
n
]
s[n]
s[n]之间的协方差为:
c
s
[
m
,
n
]
=
E
[
(
s
[
m
]
−
E
[
s
[
m
]
]
)
(
s
[
n
]
−
E
[
s
[
n
]
]
)
]
=
E
[
(
a
m
+
1
(
s
[
−
1
]
−
μ
s
)
+
∑
k
=
0
m
a
k
u
[
m
−
k
]
)
⋅
(
a
n
+
1
(
s
[
−
1
]
−
μ
s
)
+
∑
l
=
0
n
a
l
u
[
n
−
l
]
)
]
=
a
m
+
n
+
2
σ
s
2
+
∑
k
=
0
m
∑
l
=
0
n
a
k
+
l
E
[
u
[
m
−
k
]
u
[
n
−
l
]
]
\begin{aligned} c_s[m,n] &= \mathbb{E} \left [ (s[m] - \mathbb{E} [s[m]]) (s[n] - \mathbb{E} [s[n]]) \right] \\ &= \mathbb{E} \left [ \left( a^{m+1}(s[-1] - \mu_s) + \sum_{k=0}^m a^k u[m-k] \right ) \cdot \left( a^{n+1}(s[-1] - \mu_s) + \sum_{l=0}^n a^l u[n-l] \right ) \right] \\ &= a^{m+n+2} \sigma_s^2 + \sum_{k=0}^m \sum_{l=0}^n a^{k+l} \mathbb{E} [u[m-k] u [n-l]] \end{aligned}
cs[m,n]=E[(s[m]−E[s[m]])(s[n]−E[s[n]])]=E[(am+1(s[−1]−μs)+k=0∑maku[m−k])⋅(an+1(s[−1]−μs)+l=0∑nalu[n−l])]=am+n+2σs2+k=0∑ml=0∑nak+lE[u[m−k]u[n−l]]
但是
E
[
u
[
m
−
k
]
u
[
n
−
l
]
]
=
σ
u
2
δ
[
k
−
(
l
+
m
−
n
)
]
\mathbb{E} [u[m-k] u [n-l]] = \sigma^2_u \delta [k - (l+m-n)]
E[u[m−k]u[n−l]]=σu2δ[k−(l+m−n)]
因此,当
m
≥
n
m \geq n
m≥n时
c
s
[
m
,
n
]
=
a
m
+
n
+
2
σ
s
2
+
σ
u
2
a
m
−
n
∑
l
=
0
n
a
2
l
\begin{aligned} c_s[m,n] &= a^{m+n+2} \sigma_s^2 + \sigma^2_u a^{m-n} \sum_{l=0}^n a^{2l} \end{aligned}
cs[m,n]=am+n+2σs2+σu2am−nl=0∑na2l
当
m
<
n
m < n
m<n时,
c
s
[
m
,
n
]
=
c
s
[
n
,
m
]
c_s[m,n] = c_s[n,m]
cs[m,n]=cs[n,m]。基于上述协方差,可以得到方差为
var
[
s
[
n
]
]
=
c
s
[
n
,
n
]
=
a
2
n
+
2
σ
s
2
+
σ
u
2
∑
l
=
0
n
a
2
l
\begin{aligned} \text{var}[s[n]] &= c_s[n,n] \\ &= a^{2n+2} \sigma^2_s + \sigma^2_u \sum_{l=0}^n a^{2l} \end{aligned}
var[s[n]]=cs[n,n]=a2n+2σs2+σu2l=0∑na2l
显然,因为
E
[
s
[
n
]
]
=
a
n
+
1
μ
s
\mathbb{E} [s[n]] = a^{n+1} \mu_s
E[s[n]]=an+1μs与
n
n
n相关,且协方差与
m
,
n
m,n
m,n相关,因此
s
[
n
]
s[n]
s[n]不是一个广义平稳过程(Wide-sense stationary, WSS)。然而,当
n
→
∞
n \rightarrow \infty
n→∞时
E
[
s
[
n
]
]
→
0
c
s
[
m
,
n
]
→
σ
u
2
a
m
−
n
1
−
a
2
\begin{aligned} \mathbb{E} [s[n]] & \rightarrow 0 \\ c_s[m,n] &\rightarrow \frac{\sigma_u^2 a^{m-n}}{1 - a^2} \end{aligned}
E[s[n]]cs[m,n]→0→1−a2σu2am−n
因为 ∣ a ∣ < 1 |a| < 1 ∣a∣<1(该条件对于整个过程的稳定是必要的,否则,均值和方差将会随着 n n n呈指数形式增长)
因为高斯-马尔可夫过程的特殊形式,其均值和方差也可以被迭代地表征(式(3,4)被称为均值与方差传播公式)
E
[
s
[
n
]
]
=
a
E
[
s
[
n
−
1
]
]
(3)
\mathbb{E} [s[n]] = a \mathbb{E} [s[n-1]] \tag{3}
E[s[n]]=aE[s[n−1]](3)
var [ s [ n ] ] = E [ ( s [ n ] − E [ s [ n ] ] 2 ) 2 ] = E [ ( a s [ n − 1 ] + u [ n ] − a E [ s [ n − 1 ] ] ) 2 ] = a 2 var [ s [ n − 1 ] ] + σ u 2 (4) \begin{aligned} \text{var}[s[n]] &= \mathbb{E} \left [ (s[n] - \mathbb{E}[s[n]]^2)^2 \right] \\ &= \mathbb{E} \left [ {\left ( as[n-1] + u[n] - a\mathbb{E}[s[n-1]] \right)}^2 \right] \\ &= a^2 \text{var}[s[n-1]] + \sigma^2_u \tag{{4}} \end{aligned} var[s[n]]=E[(s[n]−E[s[n]]2)2]=E[(as[n−1]+u[n]−aE[s[n−1]])2]=a2var[s[n−1]]+σu2(4)
其中我们使用了 E [ u [ n ] s [ n − 1 ] ] = 0 \mathbb{E}[u[n]s[n-1]] = 0 E[u[n]s[n−1]]=0,这是因为 s [ n − 1 ] s[n-1] s[n−1]只取决于 { s [ − 1 ] , u [ 0 ] , ⋯ , u [ n − 1 ] } \{s[-1], u[0], \cdots, u[n-1]\} {s[−1],u[0],⋯,u[n−1]},且这些随机变量独立于 u [ n ] u[n] u[n]。注意到,在式(4)中,第一项 a s [ n − 1 ] as[n-1] as[n−1]会造成方差减小,第二项的积累 σ u 2 \sigma^2_u σu2会造成方差增大,在达到稳态(steady state)后,或者 n → ∞ n\rightarrow\infty n→∞,两项的作用相互平衡,收敛为 σ u 2 / ( 1 − a 2 ) \sigma^2_u / (1-a^2) σu2/(1−a2).
考虑一个
p
p
p阶的高斯-马尔可夫过程:
s
[
n
]
=
−
∑
k
=
1
p
a
[
k
]
s
[
n
−
k
]
+
u
[
n
]
(5)
s[n] = - \sum_{k=1}^p a[k] s[n-k] + u[n] \tag{5}
s[n]=−k=1∑pa[k]s[n−k]+u[n](5)
因为
s
[
n
]
s[n]
s[n]取决于前
p
p
p个采样点,所以均值和方差传播式变得更加复杂。为了拓展之前的结论,我们指定
{
s
[
n
−
1
]
,
s
[
n
−
2
]
,
⋯
,
s
[
n
−
p
]
}
\{s[n-1],s[n-2],\cdots,s[n-p]\}
{s[n−1],s[n−2],⋯,s[n−p]}为采样时刻
n
n
n的系统状态(system state),我们定义状态向量:
s
[
n
−
1
]
=
[
s
[
n
−
p
]
s
[
n
−
p
+
1
]
⋮
s
[
n
−
1
]
]
(6)
\boldsymbol{s}\left[ n-1 \right] =\left[ \begin{array}{c} \begin{array}{c} s\left[ n-p \right]\\ s\left[ n-p+1 \right]\\ \end{array}\\ \vdots\\ s\left[ n-1 \right]\\ \end{array} \right] \tag{6}
s[n−1]=⎣⎢⎢⎢⎡s[n−p]s[n−p+1]⋮s[n−1]⎦⎥⎥⎥⎤(6)
我们可以把式(10)写为
[
s
[
n
−
p
+
1
]
s
[
n
−
p
+
2
]
⋮
s
[
n
−
1
]
s
[
n
]
]
=
[
0
1
0
⋯
0
0
0
1
⋯
0
0
0
0
⋯
0
⋮
⋮
⋮
⋱
⋮
−
a
[
p
]
−
a
[
p
−
1
]
−
a
[
p
−
2
]
⋯
−
a
[
1
]
]
⏟
A
[
s
[
n
−
p
]
s
[
n
−
p
+
1
]
⋮
s
[
n
−
2
]
s
[
n
−
1
]
]
+
[
0
0
⋮
1
]
⏟
B
u
[
n
]
\left[ \begin{array}{c} \begin{array}{c} s\left[ n-p+1 \right]\\ s\left[ n-p+2 \right]\\ \end{array}\\ \vdots\\ \begin{array}{c} s\left[ n-1 \right]\\ s\left[ n \right]\\ \end{array}\\ \end{array} \right] =\mathop {\underbrace{\left[ \begin{matrix}{} 0& 1& 0& \cdots& 0\\ 0& 0& 1& \cdots& 0\\ 0& 0& 0& \cdots& 0\\ \vdots& \vdots& \vdots& \ddots& \vdots\\ -a\left[ p \right]& -a\left[ p-1 \right]& -a\left[ p-2 \right]& \cdots& -a\left[ 1 \right]\\ \end{matrix} \right] }} \limits_{\boldsymbol{A}}\left[ \begin{array}{c} \begin{array}{c} s\left[ n-p \right]\\ s\left[ n-p+1 \right]\\ \end{array}\\ \vdots\\ \begin{array}{c} s\left[ n-2 \right]\\ s\left[ n-1 \right]\\ \end{array}\\ \end{array} \right] +\mathop {\underbrace{\left[ \begin{array}{c} 0\\ 0\\ \vdots\\ 1\\ \end{array} \right] }} \limits_{\boldsymbol{B}}u\left[ n \right]
⎣⎢⎢⎢⎢⎢⎡s[n−p+1]s[n−p+2]⋮s[n−1]s[n]⎦⎥⎥⎥⎥⎥⎤=A
⎣⎢⎢⎢⎢⎢⎡000⋮−a[p]100⋮−a[p−1]010⋮−a[p−2]⋯⋯⋯⋱⋯000⋮−a[1]⎦⎥⎥⎥⎥⎥⎤⎣⎢⎢⎢⎢⎢⎡s[n−p]s[n−p+1]⋮s[n−2]s[n−1]⎦⎥⎥⎥⎥⎥⎤+B
⎣⎢⎢⎢⎡00⋮1⎦⎥⎥⎥⎤u[n]
其中的前
(
p
−
1
)
(p-1)
(p−1)个方程为方阵,根据定义,将上式写为状态向量的形式:
s
[
n
]
=
A
s
[
n
−
1
]
+
B
u
[
n
]
(7)
\boldsymbol s[n] = \boldsymbol A \boldsymbol s[n-1] + \boldsymbol B \boldsymbol u[n] \tag{7}
s[n]=As[n−1]+Bu[n](7)
其中
A
\boldsymbol{ A}
A是一个
p
×
p
p \times p
p×p的非奇异矩阵(称为状态转移矩阵:state transition matrix),
B
\boldsymbol{ B}
B是一个
p
×
1
p \times 1
p×1的向量。式(7)的形式被称为向量高斯-马尔可夫模型(Vector Gauss-Markov Model)。更一般的模型可表示为,
s
[
n
]
=
A
s
[
n
−
1
]
+
B
u
[
n
]
(8)
\boldsymbol s[n] = \boldsymbol A \boldsymbol s[n-1] + \boldsymbol B \boldsymbol u[n] \tag{8}
s[n]=As[n−1]+Bu[n](8)
其中 A , B \boldsymbol{A,B} A,B都是固定的矩阵, A \boldsymbol{ A} A的维度为 p × p p \times p p×p, B \boldsymbol{B} B的维度为 p × r p \times r p×r。 s [ n ] \boldsymbol{ s}[n] s[n]是一个 p × 1 p \times 1 p×1的信号向量, u [ n ] \boldsymbol{ u}[n] u[n]是一个驱动噪声矢量(driving noise vector)。我们称式(8)为状态模型(state model),该模型的统计假设有:
-
输入的 u [ n ] \boldsymbol{u}[n] u[n]是高斯白噪声,q向量,即 u [ n ] \boldsymbol{ u}[n] u[n]是一个不相关的联合高斯分布的序列,且 E [ u [ n ] ] = 0 \mathbb{E}[\boldsymbol{u}[n]] = \boldsymbol{ 0} E[u[n]]=0,
E [ u [ m ] u T [ n ] ] = 0 , m ≠ n \mathbb{E}[\boldsymbol u[m] \boldsymbol u^T[n]] = \boldsymbol{ 0}, \ \ \ \ m\neq n E[u[m]uT[n]]=0, m=n u [ n ] \boldsymbol{u}[n] u[n]的协方差为:
E [ u [ n ] u T [ n ] ] = Q \mathbb{E}[\boldsymbol u[n] \boldsymbol u^T[n]] = \boldsymbol{Q} E[u[n]uT[n]]=Q 其中 Q \boldsymbol{ Q} Q 是一个 r × r r \times r r×r的正定矩阵。 -
初始状态 s [ − 1 ] \boldsymbol{ s}[-1] s[−1]是随机向量: s [ − 1 ] ∼ N ( μ s , C s ) \boldsymbol{s}[-1] \sim \mathcal{N}(\boldsymbol{\mu}_s, \boldsymbol{C}_s) s[−1]∼N(μs,Cs)独立于 u [ n ] , ∀ n ≥ 0 \boldsymbol{u}[n], \forall n \geq 0 u[n],∀n≥0
我们进一步推导向量高斯-马尔可夫模型的统计特征(标量模型的扩展),依据式(8),
s
[
0
]
=
A
s
[
−
1
]
+
B
u
[
0
]
s
[
1
]
=
A
s
[
0
]
+
B
u
[
1
]
=
A
2
s
[
−
1
]
+
A
B
u
[
0
]
+
B
u
[
1
]
etc.
\begin{aligned} \boldsymbol s [0] & = \boldsymbol A \boldsymbol s [-1] + \boldsymbol{ B}\boldsymbol u [0] \\ \boldsymbol s [1] & = \boldsymbol A \boldsymbol s [0] + \boldsymbol{ B}\boldsymbol u [1] \\ &= \boldsymbol A^2 \boldsymbol s [-1] + \boldsymbol{A B}\boldsymbol u [0] + \boldsymbol{ B}\boldsymbol u [1] \\ & \text{etc.} \end{aligned}
s[0]s[1]=As[−1]+Bu[0]=As[0]+Bu[1]=A2s[−1]+ABu[0]+Bu[1]etc.
一般地,我们可以推广得到
s
[
n
]
=
A
n
+
1
s
[
−
1
]
+
∑
k
=
0
n
A
k
B
u
[
n
−
k
]
\boldsymbol s[n] = \boldsymbol A^{n+1} \boldsymbol s[-1] + \sum_{k=0}^n \boldsymbol A^k \boldsymbol B \boldsymbol u[n-k]
s[n]=An+1s[−1]+k=0∑nAkBu[n−k]
其中
A
0
=
I
\boldsymbol{A}^0=\boldsymbol{I}
A0=I,可以看出,
s
[
n
]
\boldsymbol{ s}[n]
s[n]初始条件
s
[
−
1
]
\boldsymbol{s}[-1]
s[−1]和
u
[
⋅
]
\boldsymbol{u}[\cdot]
u[⋅]的线性组合,因此,
s
[
n
]
\boldsymbol{s}[n]
s[n]是一个高斯随机过程,那么就只需要决定其均值和方差。
E
[
s
[
n
]
]
=
A
n
+
1
E
[
s
[
−
1
]
]
=
A
n
+
1
μ
s
(9)
\mathbb{E}[\boldsymbol s[n]] = \boldsymbol A^{n+1} \mathbb{E}[\boldsymbol s[-1]] = \boldsymbol A^{n+1} \boldsymbol \mu_s \tag{9}
E[s[n]]=An+1E[s[−1]]=An+1μs(9)
其协方差:
C
s
[
m
,
n
]
=
E
[
(
s
[
m
]
−
E
[
s
[
m
]
]
)
(
s
[
n
]
−
E
[
s
[
n
]
]
)
T
]
=
E
[
(
A
m
+
1
(
s
[
−
1
]
−
μ
s
)
+
∑
k
=
0
m
A
k
B
u
[
m
−
k
]
)
⋅
(
A
n
+
1
(
s
[
−
1
]
−
μ
s
)
+
∑
l
=
0
n
A
l
B
u
[
n
−
l
]
)
T
]
=
A
m
+
1
C
s
A
n
+
1
T
+
∑
k
=
0
m
∑
l
=
0
n
A
k
B
E
[
u
[
m
−
k
]
u
T
[
n
−
l
]
]
B
T
A
l
T
\begin{aligned} \boldsymbol C_s[m,n] &= \mathbb{E} \left [ {\left( \boldsymbol s[m] - \mathbb{E}[\boldsymbol s[m]] \right)} {\left( \boldsymbol s[n] - \mathbb{E}[\boldsymbol s[n]] \right)}^T \right] \\ & = \mathbb{E} \left [ \left( \boldsymbol A^{m+1} (\boldsymbol s[-1] - \boldsymbol \mu_s) + \sum_{k=0}^m \boldsymbol A^k \boldsymbol B \boldsymbol u[m-k] \right) \cdot {\left( \boldsymbol A^{n+1} (\boldsymbol s[-1] - \boldsymbol \mu_s) + \sum_{l=0}^n \boldsymbol A^l \boldsymbol B \boldsymbol u[n-l] \right)}^T \right] \\ &= \boldsymbol A^{m+1} \boldsymbol C_s \boldsymbol A^{{n+1}^T} + \sum_{k=0}^m \sum_{l=0}^n \boldsymbol A^k \boldsymbol B \mathbb{E} \left [ \boldsymbol{ u}[m-k] \boldsymbol{u}^T[n-l]\right] \boldsymbol B^T \boldsymbol A^{l^T} \end{aligned}
Cs[m,n]=E[(s[m]−E[s[m]])(s[n]−E[s[n]])T]=E⎣⎡(Am+1(s[−1]−μs)+k=0∑mAkBu[m−k])⋅(An+1(s[−1]−μs)+l=0∑nAlBu[n−l])T⎦⎤=Am+1CsAn+1T+k=0∑ml=0∑nAkBE[u[m−k]uT[n−l]]BTAlT
注意到,
E
[
u
[
m
−
k
]
u
T
[
n
−
l
]
]
=
Q
δ
[
l
−
(
n
−
m
+
k
)
]
\mathbb{E} \left [ \boldsymbol{ u}[m-k] \boldsymbol{u}^T[n-l]\right] = \boldsymbol Q \delta [l-(n-m+k)]
E[u[m−k]uT[n−l]]=Qδ[l−(n−m+k)]
因此,当
m
≥
n
m \geq n
m≥n时,
C
s
[
m
,
n
]
=
A
m
+
1
C
s
A
n
+
1
T
+
∑
l
=
0
n
A
l
+
m
−
n
B
Q
B
T
A
l
T
(10)
\boldsymbol C_s[m,n] = \boldsymbol A^{m+1} \boldsymbol C_s \boldsymbol A^{{n+1}^T} + \sum_{l=0}^n \boldsymbol A^{l+m-n} \boldsymbol {BQB}^T \boldsymbol A^{l^T} \tag{{10}}
Cs[m,n]=Am+1CsAn+1T+l=0∑nAl+m−nBQBTAlT(10)
当
m
<
n
m<n
m<n时,
C
s
[
m
,
n
]
=
C
s
T
[
n
,
m
]
\boldsymbol C_s[m,n] = \boldsymbol C_s^T[n,m]
Cs[m,n]=CsT[n,m]
那么协方差矩阵可以表示为:
C
[
n
]
=
C
s
[
n
,
n
]
=
A
n
+
1
C
s
A
n
+
1
T
+
∑
k
=
0
n
A
k
B
Q
B
T
A
k
T
(11)
\begin{aligned} \boldsymbol C[n] &= \boldsymbol C_s[n,n] \\ &= \boldsymbol A^{n+1} \boldsymbol C_s \boldsymbol A^{{n+1}^T} + \sum_{k=0}^n \boldsymbol A^k \boldsymbol{BQB}^T \boldsymbol A^{k^T} \tag{11} \end{aligned}
C[n]=Cs[n,n]=An+1CsAn+1T+k=0∑nAkBQBTAkT(11)
期望和方差的传播方程可以写为:
E
[
s
[
n
]
]
=
A
E
[
s
[
n
−
1
]
]
(12)
\boldsymbol E[\boldsymbol s[n]] = \boldsymbol A \boldsymbol E[\boldsymbol s [n-1]] \tag{12}
E[s[n]]=AE[s[n−1]](12)
C [ n ] = A C [ n − 1 ] A T + B Q B T \boldsymbol C[n] = \boldsymbol A \boldsymbol C[n-1] \boldsymbol A^T + \boldsymbol {BQB}^T C[n]=AC[n−1]AT+BQBT
注意,只有当 A \boldsymbol{A} A的特征值幅度都小于1,才是一个稳定的过程(steady process)。
当
n
→
∞
n \rightarrow \infty
n→∞时,
E
[
s
[
n
]
]
=
A
n
+
1
μ
s
→
0
\mathbb{E} [\boldsymbol s [n]] = \boldsymbol A^{n+1} \boldsymbol \mu_s \rightarrow \boldsymbol 0
E[s[n]]=An+1μs→0
A n + 1 C s A n + 1 T → 0 \boldsymbol A^{n+1} \boldsymbol C_s \boldsymbol A^{{n+1}^T} \rightarrow 0 An+1CsAn+1T→0
因此,
C
[
n
]
→
C
=
∑
k
=
0
∞
A
k
B
Q
B
T
A
k
T
(13)
\boldsymbol C[n] \rightarrow \boldsymbol C = \sum_{k=0}^{\infty} \boldsymbol A^k \boldsymbol {BQB}^T \boldsymbol A^{k^T} \tag{13}
C[n]→C=k=0∑∞AkBQBTAkT(13)
另外,当
n
→
∞
n\rightarrow\infty
n→∞,
C
[
n
−
1
]
=
C
[
n
]
\boldsymbol{C}[n-1]=\boldsymbol{C}[n]
C[n−1]=C[n],那么稳态的协方差矩阵为方程(14)的解:
C
=
A
C
A
T
+
B
Q
B
T
(14)
\boldsymbol C = \boldsymbol {ACA}^T + \boldsymbol {BQB}^T \tag{14}
C=ACAT+BQBT(14)
该方程被称为Lyapunov equation.
将上述模型和定理总结如下:
定理-1:向量高斯马尔可夫模型(Vector Gauss-Markov Model):对一个
p
×
1
p \times 1
p×1的信号向量
s
[
n
]
\boldsymbol{ s}[n]
s[n],其高斯-马尔可夫模型为:
s
[
n
]
=
A
s
[
n
−
1
]
+
B
u
[
n
]
,
n
≥
0
(15)
\boldsymbol s[n] = \boldsymbol A \boldsymbol s[n-1] + \boldsymbol B \boldsymbol u [n], \ \ \ \ n \geq 0 \tag{15}
s[n]=As[n−1]+Bu[n], n≥0(15)
A
(
p
×
p
)
\boldsymbol{A} (p \times p)
A(p×p)和
B
(
p
×
r
)
\boldsymbol{ B} (p \times r)
B(p×r)已知,假设
A
\boldsymbol A
A的特征值幅度小于1,
u
[
n
]
(
r
×
1
)
\boldsymbol{u}[n] (r \times 1)
u[n](r×1)为高斯白噪声向量,
u
[
n
]
∼
N
(
0
,
Q
)
\boldsymbol{u}[n] \sim \mathcal{N}(\boldsymbol{0},\boldsymbol{Q})
u[n]∼N(0,Q)且
{
u
[
n
]
}
\{\boldsymbol{u}[n]\}
{u[n]}之间相互独立。初始条件
s
[
−
1
]
∼
N
(
μ
s
,
C
s
)
\boldsymbol{s}[-1] \sim \mathcal N(\boldsymbol{ \mu}_s,\boldsymbol C_s)
s[−1]∼N(μs,Cs),独立于
{
u
[
n
]
}
\{\boldsymbol{u}[n]\}
{u[n]},那么该信号过程是高斯的,且其均值为
E
[
s
[
n
]
]
=
A
n
+
1
μ
s
(16)
\mathbb{E} [\boldsymbol s [n]] = \boldsymbol A^{n+1} \boldsymbol \mu_s \tag{16}
E[s[n]]=An+1μs(16)
当
m
≥
n
m \geq n
m≥n时,协方差为
C
s
[
m
,
n
]
=
A
m
+
1
C
s
A
n
+
1
T
+
∑
l
=
0
n
A
l
+
m
−
n
B
Q
B
T
A
l
T
(17)
\boldsymbol C_s[m,n] = \boldsymbol A^{m+1} \boldsymbol C_s \boldsymbol A^{{n+1}^T} + \sum_{l=0}^n \boldsymbol A^{l+m-n} \boldsymbol {BQB}^T \boldsymbol A^{l^T} \tag{{17}}
Cs[m,n]=Am+1CsAn+1T+l=0∑nAl+m−nBQBTAlT(17)
当
m
<
n
m<n
m<n时,
C
s
[
m
,
n
]
=
C
s
T
[
n
,
m
]
\boldsymbol C_s[m,n] = \boldsymbol C_s^T[n,m]
Cs[m,n]=CsT[n,m]
那么协方差矩阵可以表示为:
C
[
n
]
=
C
s
[
n
,
n
]
=
A
n
+
1
C
s
A
n
+
1
T
+
∑
k
=
0
n
A
k
B
Q
B
T
A
k
T
(18)
\begin{aligned} \boldsymbol C[n] &= \boldsymbol C_s[n,n] \\ &= \boldsymbol A^{n+1} \boldsymbol C_s \boldsymbol A^{{n+1}^T} + \sum_{k=0}^n \boldsymbol A^k \boldsymbol{BQB}^T \boldsymbol A^{k^T} \tag{18} \end{aligned}
C[n]=Cs[n,n]=An+1CsAn+1T+k=0∑nAkBQBTAkT(18)
期望和方差的传播方程可以写为:
E
[
s
[
n
]
]
=
A
E
[
s
[
n
−
1
]
]
(19)
\boldsymbol E[\boldsymbol s[n]] = \boldsymbol A \boldsymbol E[\boldsymbol s [n-1]] \tag{19}
E[s[n]]=AE[s[n−1]](19)
C [ n ] = A C [ n − 1 ] A T + B Q B T (20) \boldsymbol C[n] = \boldsymbol A \boldsymbol C[n-1] \boldsymbol A^T + \boldsymbol {BQB}^T \tag{20} C[n]=AC[n−1]AT+BQBT(20)