\qquad DC电平估计为例,估计值 A ^ \hat{A} A^为数据样本 x [ N ] x[N] x[N]均值,则LSE变为:
A ^ [ N ] = 1 N + 1 ∑ n = 0 N x [ n ] \hat{A}[N] = \frac{1}{N+1}\sum_{n = 0}^{N}x[n] A^[N]=N+11n=0∑Nx[n] \qquad 在加权LS问题中,当加权矩阵为 W \bold{W} W(通常为噪声矩阵)是对角矩阵时,其中 [ W ] i i = 1 / σ i 2 [\bold{W}]_{ii}=1/\sigma_{i}^{2} [W]ii=1/σi2,加权LSE表达式为:
A ^ [ N ] = ∑ n = 0 N x [ n ] σ n 2 ∑ n = 0 N 1 σ n 2 \hat{A}[N]=\frac{\sum_{n=0}^{N}\frac{x[n]}{\sigma_{n}^{2}}}{\sum_{n=0}^{N}\frac{1}{\sigma_{n}^{2}}} A^[N]=∑n=0Nσn21∑n=0Nσn2x[n] \qquad 分解为序贯形式为:
A ^ [ N ] = ∑ n = 0 N − 1 x [ n ] σ n 2 + x [ N ] σ N 2 ∑ n = 0 N 1 σ n 2 = ∑ n = 0 N − 1 1 σ n 2 A ^ [ N − 1 ] + x [ N ] σ N 2 ∑ n = 0 N 1 σ n 2 = A ^ [ N − 1 ] + 1 σ N 2 ∑ n = 0 N 1 σ n 2 ∗ ( x [ N ] − A ^ [ N − 1 ] ) \hat{A}[N] = \frac{\sum_{n=0}^{N-1}\frac{x[n]}{\sigma_{n}^{2}}+\frac{x[N]}{\sigma_{N}^{2}}}{\sum_{n=0}^{N}\frac{1}{\sigma_{n}^{2}}}= \frac{{\sum_{n=0}^{N-1}\frac{1}{\sigma_{n}^{2}}}\hat{A}[N-1] + \frac{x[N]}{\sigma_{N}^{2}}}{\sum_{n=0}^{N}\frac{1}{\sigma_{n}^{2}}} \\ =\hat{A}[N-1] + \frac{\frac{1}{\sigma_{N}^{2}}}{\sum_{n=0}^{N}\frac{1}{\sigma_{n}^{2}}} * (x[N] - \hat{A}[N-1]) A^[N]=∑n=0Nσn21∑n=0N−1σn2x[n]+σN2x[N]=∑n=0Nσn21∑n=0N−1σn21A^[N−1]+σN2x[N]=A^[N−1]+∑n=0Nσn21σN21∗(x[N]−A^[N−1]) \qquad 其中,LSE即为最优线性无偏估计(BLUE),因此:
v a r ( A ^ [ N − 1 ] ) = 1 ∑ n = 0 N − 1 1 σ n 2 var(\hat{A}[N-1]) = \frac{1}{\sum_{n=0}^{N-1}\frac{1}{\sigma_{n}^{2}}} var(A^[N−1])=∑n=0N−1σn211 \qquad 由此,增益 K [ N ] K[N] K[N]可以表示为:
K [ N ] = v a r ( A ^ [ N − 1 ] ) v a r ( A ^ [ N − 1 ] ) + σ N 2 K[N]=\frac{var(\hat A[N-1])}{var(\hat A[N-1]) + \sigma_{N}^{2}} K[N]=var(A^[N−1])+σN2var(A^[N−1]) \qquad 由于增益 K [ N ] K[N] K[N]取决于 v a r ( A ^ [ N − 1 ] ) var(\hat A[N-1]) var(A^[N−1]),则可以表示为:
v a r ( A ^ [ N ] ) = 1 ∑ n = 0 N 1 σ n 2 = 1 1 v a r ( A ^ [ N − 1 ] ) + 1 σ N 2 = v a r ( A ^ [ N − 1 ] ) σ N 2 v a r ( A ^ [ N − 1 ] ) + σ N 2 = ( 1 − v a r ( A ^ [ N − 1 ] ) v a r ( A ^ [ N − 1 ] ) + σ N 2 ) v a r ( A ^ [ N − 1 ] ) = ( 1 − K [ N ] ) v a r ( A ^ [ N − 1 ] ) var(\hat A[N]) = \frac{1}{\sum_{n=0}^{N}\frac{1}{\sigma_{n}^{2}}}=\frac{1}{\frac{1}{var(\hat A[N-1])} + \frac{1}{\sigma_{N}^{2}}}=\frac{var(\hat A[N-1])\sigma_{N}^{2}}{var(\hat A[N-1]) + \sigma_{N}^{2}} \\ =(1-\frac{var(\hat A[N-1])}{var(\hat A[N-1]) + \sigma_{N}^{2}})var(\hat A[N-1])=(1-K[N])var(\hat A[N-1]) var(A^[N])=∑n=0Nσn211=var(A^[N−1])1+σN211=var(A^[N−1])+σN2var(A^[N−1])σN2=(1−var(A^[N−1])+σN2var(A^[N−1]))var(A^[N−1])=(1−K[N])var(A^[N−1]) \qquad 由此, A ^ [ N ] , K [ N ] , v a r ( A ^ [ N ] ) \hat A[N],K[N],var(\hat A[N]) A^[N],K[N],var(A^[N])可以递推获得。
估计量更新为: A ^ [ N ] = A ^ [ N − 1 ] + K [ N ] ( x [ N ] − A ^ [ N − 1 ] ) \hat A[N] = \hat A[N-1] + K[N](x[N] - \hat A[N-1]) A^[N]=A^[N−1]+K[N](x[N]−A^[N−1])
方差更新为: v a r ( A ^ [ N ] ) = ( 1 − K [ N ] ) v a r ( A ^ [ N − 1 ] ) var(\hat A[N]) = (1-K[N])var(\hat A[N-1]) var(A^[N])=(1−K[N])var(A^[N−1]) \qquad 具有矢量参数的序贯LSE,使用 C C C表示零均值噪声的矩阵,即:
J = ( x − H θ ) T C − 1 ( x − H θ ) \bold J = (\bold x-\bold H\bold\theta)^{T}\bold C^{-1}(\bold x-\bold H\bold\theta) J=(x−Hθ)TC−1(x−Hθ)其中,包含与BLUE相同的条件即 θ ^ \hat\theta θ^无偏且线性。则最优的估计值为:
θ ^ = ( H T C − 1 H ) H − 1 C − 1 x C θ ^ = ( H T C − 1 H ) − 1 \hat\bold\theta=(\bold H^{T}\bold C^{-1}H)\bold H^{-1}\bold C^{-1}\bold x \\ \bold C_{\hat\theta}=(\bold H^{T}\bold C^{-1}\bold H)^{-1} θ^=(HTC−1H)H−1C−1xCθ^=(HTC−1H)−1其中, C \bold C C是对焦矩阵,即噪声不相关,因此 θ ^ \hat\bold\theta θ^可按照时间顺序获得,令:
C [ n ] = d i a g { σ 1 2 , σ 2 2 , . . . , σ n 2 } H [ n ] = [ H [ n − 1 ] h T [ n ] ] = [ n × p 1 × p ] x [ n ] = [ x [ 1 ] , x [ 2 ] , . . . , x [ n ] ] T \bold C[n] = diag\{\sigma_{1}^{2},\sigma_{2}^{2},...,\sigma_{n}^{2}\} \\ \bold H[n]=\begin{bmatrix} \bold H[n-1] \\ \bold h^{T}[n] \end{bmatrix}=\begin{bmatrix} n \times p \\ 1 \times p \end{bmatrix} \\ \bold x[n] = [x[1],x[2],...,x[n]]^{T} C[n]=diag{σ12,σ22,...,σn2}H[n]=[H[n−1]hT[n]]=[n×p1×p]x[n]=[x[1],x[2],...,x[n]]T \qquad 使用 θ ^ [ n ] \hat\theta[n] θ^[n]表示基于 x [ n ] \bold x[n] x[n]或者基于前 n + 1 n+1 n+1个数据样本的加权LSE,估计量为:
θ ^ [ n ] = ( H T [ n ] C − 1 [ n ] H [ n ] ) − 1 H [ n ] T C − 1 [ n ] x [ n ] \hat\bold\theta[n]=(\bold H^{T}[n]\bold C^{-1}[n]\bold H[n])^{-1}\bold H[n]^{T}\bold C^{-1}[n]\bold x[n] θ^[n]=(HT[n]C−1[n]H[n])−1H[n]TC−1[n]x[n] \qquad 协方差矩阵 ∑ [ n ] \bold{\sum}[n] ∑[n]为:
C θ ^ [ n ] = ∑ [ n ] = ( H T [ n ] C − 1 [ n ] H [ n ] ) − 1 \bold C_{\hat\theta[n]} = \bold{\sum}[n] = (\bold H^{T}[n]\bold C^{-1}[n]\bold H[n])^{-1} Cθ^[n]=∑[n]=(HT[n]C−1[n]H[n])−1 θ ^ [ n ] = ( [ H T [ n − 1 ] , h [ n ] ] [ C [ n − 1 ] 0 0 σ n 2 ] − 1 [ H [ n − 1 ] h T [ n ] ] ) − 1 ( [ H T [ n − 1 ] , h [ n ] ] [ C [ n − 1 ] 0 0 σ n 2 ] − 1 [ x [ n − 1 ] x [ n ] ] ) = ( H T [ n − 1 ] C − 1 [ n − 1 ] H [ n − 1 ] + 1 σ n 2 h [ n ] h T [ n ] ) − 1 ( H T [ n − 1 ] C − 1 [ n − 1 ] x [ n − 1 ] + 1 σ n 2 h [ n ] x [ n ] ) \hat\bold \theta[n]=([\bold H^{T}[n-1],\bold h[n]] \begin{bmatrix} \bold C[n-1] & 0 \\ 0 & \sigma^{2}_{n} \end{bmatrix}^{-1} \begin{bmatrix} \bold H[n-1] \\ \bold h^{T}[n] \end{bmatrix})^{-1} \\ ([\bold H^{T}[n-1],\bold h[n]] \begin{bmatrix} \bold C[n-1] & 0 \\ 0 & \sigma^{2}_{n} \end{bmatrix}^{-1} \begin{bmatrix} \bold x[n-1] \\ x[n] \end{bmatrix}) \\ =(\bold H^{T}[n-1]\bold C^{-1}[n-1]\bold H[n-1] + \frac{1}{\sigma^{2}_{n}}\bold h[n]\bold h^{T}[n])^{-1} \\ (\bold H^{T}[n-1]\bold C^{-1}[n-1]\bold x[n-1] + \frac{1}{\sigma^{2}_{n}}\bold h[n]\bold x[n]) θ^[n]=([HT[n−1],h[n]][C[n−1]00σn2]−1[H[n−1]hT[n]])−1([HT[n−1],h[n]][C[n−1]00σn2]−1[x[n−1]x[n]])=(HT[n−1]C−1[n−1]H[n−1]+σn21h[n]hT[n])−1(HT[n−1]C−1[n−1]x[n−1]+σn21h[n]x[n])令 ∑ [ n − 1 ] = ( H T [ n − 1 ] C − 1 [ n − 1 ] H [ n − 1 ] ) − 1 \sum[n-1] = (\bold H^{T}[n-1]\bold C^{-1}[n-1]\bold H[n-1])^{-1} ∑[n−1]=(HT[n−1]C−1[n−1]H[n−1])−1
则 θ ^ [ n ] = ( ∑ − 1 [ n − 1 ] + 1 σ n 2 h [ n ] h T [ n ] ) − 1 ( H T [ n − 1 ] C − 1 [ n − 1 ] x [ n − 1 ] + 1 σ n 2 h [ n ] x [ n ] ) \hat\theta[n] = (\bold{\sum}^{-1}[n-1] + \frac{1}{\sigma^{2}_{n}}\bold h[n]\bold h^{T}[n])^{-1} \\ (\bold H^{T}[n-1]\bold C^{-1}[n-1]\bold x[n-1] + \frac{1}{\sigma^{2}_{n}}\bold h[n]\bold x[n]) θ^[n]=(∑−1[n−1]+σn21h[n]hT[n])−1(HT[n−1]C−1[n−1]x[n−1]+σn21h[n]x[n])因为 ∑ [ n ] = ( ∑ − 1 [ n − 1 ] + 1 σ n 2 h [ n ] h T [ n ] ) − 1 \bold{\sum}[n] = (\bold{\sum}^{-1}[n-1] + \frac{1}{\sigma^{2}_{n}}\bold h[n]\bold h^{T}[n])^{-1} ∑[n]=(∑−1[n−1]+σn21h[n]hT[n])−1由Woodbury恒等式 ( A + u u T ) − 1 = A − 1 − A − 1 u u T A − 1 1 + u T A − 1 u (A + uu^{T})^{-1} = A^{-1} - \frac{A^{-1}uu^{T}A^{-1}}{1+u^{T}A^{-1}u} (A+uuT)−1=A−1−1+uTA−1uA−1uuTA−1得: ∑ [ n ] = ∑ [ n − 1 ] − ∑ [ n − 1 ] h [ n ] h [ n ] T ∑ [ n − 1 ] σ n 2 + h [ n ] T ∑ [ n − 1 ] h [ n ] = ( I − K [ n ] h T [ n ] ) ∑ [ n − 1 ] \bold{\sum}[n] = \bold{\sum}[n-1] - \frac{\bold{\sum}[n-1]\bold h[n]\bold h[n]^{T}\bold{\sum}[n-1]}{\sigma^{2}_{n}+\bold h[n]^{T}\bold{\sum}[n-1]\bold h[n]} \\ =(\bold I-\bold K[n]\bold h^{T}[n])\sum[n-1] ∑[n]=∑[n−1]−σn2+h[n]T∑[n−1]h[n]∑[n−1]h[n]h[n]T∑[n−1]=(I−K[n]hT[n])∑[n−1]其中, K [ n ] = ∑ [ n − 1 ] h [ n ] σ n 2 + h [ n ] T ∑ [ n − 1 ] h [ n ] \bold K[n] = \frac{\bold{\sum}[n-1]\bold h[n]}{\sigma^{2}_{n}+\bold h[n]^{T}\bold{\sum}[n-1]\bold h[n]} K[n]=σn2+h[n]T∑[n−1]h[n]∑[n−1]h[n]所以 θ ^ [ n ] = ( ( I − K [ n ] h T [ n ] ) ∑ [ n − 1 ] ) ( H T [ n − 1 ] C − 1 [ n − 1 ] x [ n − 1 ] + 1 σ n 2 h [ n ] x [ n ] ) \hat\theta[n] = ((\bold I-\bold K[n]\bold h^{T}[n])\sum[n-1]) \\(\bold H^{T}[n-1]\bold C^{-1}[n-1]\bold x[n-1] + \frac{1}{\sigma^{2}_{n}}\bold h[n]\bold x[n]) θ^[n]=((I−K[n]hT[n])∑[n−1])(HT[n−1]C−1[n−1]x[n−1]+σn21h[n]x[n])且 θ ^ [ n − 1 ] = ∑ [ n − 1 ] H T [ n − 1 ] C − 1 [ n − 1 ] x [ n − 1 ] \hat\bold\theta[n-1] = \sum[n-1]\bold H^{T}[n-1]\bold C^{-1}[n-1]\bold x[n-1] θ^[n−1]=∑[n−1]HT[n−1]C−1[n−1]x[n−1]化简的: θ ^ [ n ] = θ ^ [ n − 1 ] + K [ n ] ( x [ n ] − h T [ n ] θ ^ [ n − 1 ] ) \hat\bold\theta[n] = \hat\bold\theta[n-1] + \bold K[n](x[n] - \bold h^{T}[n]\hat\bold \theta[n-1]) θ^[n]=θ^[n−1]+K[n](x[n]−hT[n]θ^[n−1])以上得到方差更新方程和状态更新方程。