\[ \begin{aligned} \mathbf{x}^\mathrm{f}(t_{i+1}) &= \mathbf{M}(t_{i+1},\,t_i)\mathbf{x}^\mathrm{a}\\ \mathbf{P}^\mathrm{f}(t_{i+1}) &= \mathbf{M}(t_{i+1},\,t_i)\mathbf{P}^\mathrm{a}(t_{i})\mathbf{M}^\mathrm{T}(t_{i+1},\,t_i) + \mathbf{Q}(t_i)\\ \mathbf{K}_i &= \mathbf{P}^\mathrm{f}(t_i)\mathbf{H}_i^\mathrm{T}[\mathbf{H}_i\mathbf{P}^\mathrm{f}(t_i)\mathbf{H}_i^\mathrm{T} + \mathbf{R}_i]^{-1}\\ \mathbf{x}^\mathrm{a}(t_i) &= \mathbf{x}^\mathrm{f}(t_i) + \mathbf{K}_i[\mathbf{y}_i^\mathrm{o} - \mathbf{H}_i\mathbf{x}^\mathrm{f}(t_i)]\\ \mathbf{P}^\mathrm{a}(t_i) &= [\mathbf{I} - \mathbf{K}_i\mathbf{H}_i]\mathbf{P}^\mathrm{f}(t_i) \end{aligned} \]
EKF: extended Kalman filter
状態の予報と解析にそれぞれ非線型の\(M\)と\(H\)を用いる。
\[ \begin{aligned} \mathbf{x}^\mathrm{f}(t_{i+1}) &= \color{red}{M_{t_{i+1},\,t_i}(\mathbf{x}^\mathrm{a})}\\ \mathbf{P}^\mathrm{f}(t_{i+1}) &= \mathbf{M}(t_{i+1},\,t_i)\mathbf{P}^\mathrm{a}(t_{i})\mathbf{M}^\mathrm{T}(t_{i+1},\,t_i) + \mathbf{Q}(t_i)\\ \mathbf{K}_i &= \mathbf{P}^\mathrm{f}(t_i)\mathbf{H}_i^\mathrm{T}[\mathbf{H}_i\mathbf{P}^\mathrm{f}(t_i)\mathbf{H}_i^\mathrm{T} + \mathbf{R}_i]^{-1}\\ \mathbf{x}^\mathrm{a}(t_i) &= \mathbf{x}^\mathrm{f}(t_i) + \mathbf{K}_i[\mathbf{y}_i^\mathrm{o} - \color{red}{H_i(\mathbf{x}^\mathrm{f}(t_i))}]\\ \mathbf{P}^\mathrm{a}(t_i) &= [\mathbf{I} - \mathbf{K}_i\mathbf{H}_i]\mathbf{P}^\mathrm{f}(t_i) \end{aligned} \]
\[ \mathbf{P}^\mathrm{a} = [\mathbf{I} - \mathbf{K}\mathbf{H}]\mathbf{P}^\mathrm{f}[\mathbf{I} - \mathbf{K}\mathbf{H}]^\mathrm{T} + \mathbf{KRK}^\mathrm{T} \]
\[ \mathbf{MPM}^\mathrm{T} = \mathbf{M}(\mathbf{MP})^\mathrm{T} \]
\[ \begin{aligned} \frac{\mathrm{d}x}{\mathrm{d}t} &= x(a_1 + a_2x + a_3y) \\ \frac{\mathrm{d}y}{\mathrm{d}t} &= y(a_4 + a_5y + a_6x) \end{aligned} \]
Bouttier (1995)
データ同化夏の学校2025