100 lines
3.9 KiB
Markdown
100 lines
3.9 KiB
Markdown
# MOTIVATION
|
||
|
||
论文中的辨识是认为系统的阶数已知的辨识，而且 MCMC 的抽样的变量是特征多项式矩阵，我想使用特征值来抽样，并同时考虑系统阶数未知的情况。
|
||
|
||
采用RJMMC的方法
|
||
|
||
## 细致平稳条件
|
||
|
||
$$
|
||
\alpha \left( x,y \right) =\min \left\{ 1,R \right\}
|
||
\\
|
||
\pi \left( x \right) p\left( x,y \right) =\pi \left( y \right) p\left( y,x \right)
|
||
$$
|
||
|
||
## 跨维度游动
|
||
|
||
### k <-> k+1
|
||
|
||
1. Birth move
|
||
$$
|
||
\log R=\log \frac{P\left( x_{k+1} \right) P\left( y|x_{k+1} \right)}{P\left( x_k \right) P\left( y|x_k \right)}+\log \frac{P\left( dead \right)}{P\left( birth \right)}+\log \frac{\small{\frac{1}{k+1}}}{q\left( u \right)}+\log \left| J_1 \right|
|
||
$$
|
||
2. Death move
|
||
$$
|
||
\log R=\log \frac{P\left( x_k \right) P\left( y|x_k \right)}{P\left( x_{k+1} \right) P\left( y|x_{k+1} \right)}+\log \frac{P\left( birth \right)}{P\left( dead \right)}+\log \frac{\small{q\left( u \right)}}{\frac{1}{k+1}}-\log \left| J_1 \right|
|
||
$$
|
||
|
||
### k <-> k+2
|
||
|
||
1. Birth move
|
||
$$
|
||
\log R=\log \frac{P\left( x_{k+2} \right) P\left( y|x_{k+2} \right)}{P\left( x_k \right) P\left( y|x_k \right)}+\log \frac{P\left( dead \right)}{P\left( birth \right)}+\log \frac{\small{\frac{1}{m+1}}}{q\left( \theta \right) q\left( r \right)}+\log \left| J_2 \right|
|
||
$$
|
||
2. Death move
|
||
$$
|
||
\log R=\log \frac{P\left( x_k \right) P\left( y|x_k \right)}{P\left( x_{k+2} \right) P\left( y|x_{k+2} \right)}+\log \frac{P\left( birth \right)}{P\left( dead \right)}+\log \frac{\small{q\left( \theta \right) q\left( r \right)}}{\frac{1}{m+1}}-\log \left| J_2 \right|
|
||
$$
|
||
|
||
$其中 m$ 是当前系统中复共轭极点对的数量
|
||
$J_1$ 和 $J_2$ 是雅可比矩阵行列式
|
||
$\left| J_1 \right|=\left| \prod_{\boldsymbol{i}=1}^{\boldsymbol{k}}{\left( \lambda _i-u_i \right)} \right|$
|
||
$\left| J_2 \right|=\left| \prod_{\boldsymbol{i}=1}^{\boldsymbol{k}}{\left( \lambda _{k+1}-\lambda _i \right) \left( \lambda _{k+2}-\lambda _i \right)} \right|\left| \lambda _{k+1}-\lambda _{k+1} \right|$
|
||
|
||
## 同维度游动
|
||
|
||
### 根的类型转换
|
||
|
||
先确定映射关系，令
|
||
|
||
$$
|
||
a=\frac{r_1+r_2}{2}\text{、}b=\frac{r_1-r_2}{2}
|
||
\\
|
||
r_1=a+b\text{、}r_2=a-b
|
||
$$
|
||
|
||
|
||
再确定雅可比矩阵行列式
|
||
$$
|
||
J_{C\rightarrow R}=\left| \det \left( \begin{matrix}
|
||
\frac{\partial a}{\partial r_1}& \frac{\partial a}{\partial r_2}\\
|
||
\frac{\partial b}{\partial r_1}& \frac{\partial b}{\partial r_2}\\
|
||
\end{matrix} \right) \right|=\frac{1}{2}
|
||
\\
|
||
J_{R\rightarrow C}={J_{C\rightarrow R}}^{-1}=2
|
||
$$
|
||
|
||
1. 实根 <-> 复共轭根对
|
||
在$n_r$个实根中任选两个实根$r_1$和$r_2$，转换为复共轭根对$a\pm jb$，其中$a=\frac{r_1+r_2}{2}$，$b=\frac{r_1-r_2}{2}$。
|
||
|
||
$$
|
||
q\left( x|x' \right) =p_m\times \frac{1}{C\left( n_r,2 \right)}
|
||
$$
|
||
|
||
此时的接受率写作:$Merge\text{：}\alpha \left( x',x \right) =\min \left\{ 1,\frac{\pi \left( x \right)}{\pi \left( x' \right)}\cdot \frac{1}{q\left( x|x' \right)}\cdot \frac{1}{2} \right\} $
|
||
|
||
2. 复共轭根对 <-> 实根
|
||
在$n_c$个复共轭根对中任选一个复共轭根对$a\pm jb$，转换为两个实根$r_1=a+b$和$r_2=a-b$。
|
||
|
||
$$
|
||
q\left( x'|x \right) =p_c\times \frac{1}{n_c}
|
||
$$
|
||
|
||
此时的接受率写作:$Split\text{：}\alpha \left( x,x' \right) =\min \left\{ 1,\frac{\pi \left( x' \right)}{\pi \left( x \right)}\cdot \frac{1}{q\left( x'|x \right)}\cdot 2 \right\} $
|
||
|
||
### 同类型根的调整
|
||
|
||
1. 实根调整
|
||
选择一个实根$r$，通过添加噪声$\epsilon \sim N\left( 0,\sigma ^2 \right)$调整该实根的位置为$r' = r + \epsilon$。
|
||
|
||
$$
|
||
\alpha \left( x,x' \right) =\min \left\{ 1,\frac{\pi \left( x' \right)}{\pi \left( x \right)} \right\}
|
||
$$
|
||
|
||
2. 复共轭根对调整
|
||
选择一个复共轭根对$a\pm jb$，通过添加噪声$\epsilon _a \sim N\left( 0,\sigma _a^2 \right)$和$\epsilon _b \sim N\left( 0,\sigma _b^2 \right)$调整该复共轭根对的位置为$a' = a + \epsilon _a$和$b' = b + \epsilon _b$。
|
||
|
||
$$
|
||
\alpha \left( x,x' \right) =\min \left\{ 1,\frac{\pi \left( x' \right)}{\pi \left( x \right)} \right\}
|
||
$$
|