两个复高斯分布的乘积（高维）

假设$p(\mathbf{x})\propto \mathcal{C}\mathcal{N}({\mathbf{\mu }}_1,{\mathbf{\Sigma }}_1)\mathcal{C}\mathcal{N}({\mathbf{\mu }}_2,{\mathbf{\Sigma }}_2)$，有
$$\begin{array}{rl}p(\mathbf{x})& \propto \exp \left\{-(\mathbf{x}-{\mathbf{\mu }}_1{)}^H{\mathbf{\Sigma }}_1^{-1}(\mathbf{x}-{\mathbf{\mu }}_1)-(\mathbf{x}-{\mathbf{\mu }}_2{)}^H{\mathbf{\Sigma }}_2^{-1}(\mathbf{x}-{\mathbf{\mu }}_2)\right\}\\ & \propto \exp \left\{-{\mathbf{x}}^H{\mathbf{\Sigma }}_1^{-1}\mathbf{x}+2\mathcal{R}\left\{{\mathbf{\mu }}_1^H{\mathbf{\Sigma }}_1^{-1}\mathbf{x}\right\}-{\mathbf{x}}^H{\mathbf{\Sigma }}_2^{-1}\mathbf{x}+2\mathcal{R}\left\{{\mathbf{\mu }}_2^H{\mathbf{\Sigma }}_2^{-1}\mathbf{x}\right\}\right\}\\ & =\exp \left\{-{\mathbf{x}}^H\left({\mathbf{\Sigma }}_1^{-1}+{\mathbf{\Sigma }}_2^{-1}\right)\mathbf{x}+2\mathcal{R}\left\{\left({\mathbf{\mu }}_1^H{\mathbf{\Sigma }}_1^{-1}+{\mathbf{\mu }}_2^H{\mathbf{\Sigma }}_2^{-1}\right)\mathbf{x}\right\}\right\}\\ & =\exp \left\{-\left[{\mathbf{x}}^H\left[\underset{={\Sigma }_x^{-1}}{\underbrace{{\Sigma }_1^{-1}+{\Sigma }_2^{-1}}}\right]\mathbf{x}-2\mathcal{R}\left\{{\left[\underset{={\mathbf{\mu }}_x}{\underbrace{{\Sigma }_x\left({\Sigma }_1^{-1}{\mathbf{\mu }}_1+{\Sigma }_2^{-1}{\mathbf{\mu }}_2\right)}}\right]}^H{\mathbf{\Sigma }}_x^{-1}\mathbf{x}\right\}\right]\right\}\\ & =\exp \left\{-\left[{\mathbf{x}}^H{\mathbf{\Sigma }}_x^{-1}\mathbf{x}+2\mathcal{R}{\mathbf{\mu }}_x^H{\mathbf{\Sigma }}_x^{-1}\mathbf{x}\right]\right\}\\ & \propto \exp \left\{-(\mathbf{x}-{\mathbf{\mu }}_x{)}^H{\mathbf{\Sigma }}^{-1}(\mathbf{x}-{\mathbf{\mu }}_x)\right\}\end{array}$$
因此$x\sim \mathcal{C}\mathcal{N}({\mathbf{\mu }}_x,{\mathbf{\Sigma }}_x)$，其中
$$\begin{array}{rl}{\mathbf{\mu }}_x& ={\Sigma }_x\left({\Sigma }_1^{-1}{\mathbf{\mu }}_1+{\Sigma }_2^{-1}{\mathbf{\mu }}_2\right)\\ {\mathbf{\Sigma }}_x& ={\left({\mathbf{\Sigma }}_1^{-1}+{\mathbf{\Sigma }}_2^{-1}\right)}^{-1}\end{array}$$
注意，当$\mathbf{x}$为标量$x$时，$x\sim \mathcal{C}\mathcal{N}({\mu }_x,{\sigma }_x)$
$$\begin{array}{rl}{\sigma }_x& ={\left(\frac{1}{{\sigma }_1^2}+\frac{1}{{\sigma }_2^2}\right)}^{-1}=\frac{{\sigma }_1^2{\sigma }_2^2}{{\sigma }_1^2+{\sigma }_2^2}\\ {\mu }_x& =\frac{{\sigma }_1^2{\sigma }_2^2}{{\sigma }_1^2+{\sigma }_2^2}\left(\frac{{\mu }_1}{{\sigma }_1^2}+\frac{{\mu }_2}{{\sigma }_2^2}\right)\end{array}$$
然而，在实际编写代码的过程中，考虑到数值的稳定性，我们一般按照如下顺序执行：
$$\begin{array}{rl}g& =\frac{{\sigma }_1^2}{{\sigma }_1^2+{\sigma }_2^2}\\ {\mu }_x& =g\cdot ({\mu }_2-{\mu }_1)+{\mu }_1\\ {\sigma }_x& =g\cdot {\sigma }_1^2\end{array}$$

另外，考虑$\text{常见的线性模型}$：
$$\mathbf{y}=\mathbf{A}\mathbf{x}+\mathbf{w}$$

其中$\mathbf{x}$的先验分布：$\mathbf{x}\sim \mathcal{C}\mathcal{N}(\mathbf{x};\mathbf{r},{\mathbf{\Sigma }}_1)$，似然分布$\mathbf{y}|\mathbf{A}\mathbf{x}\sim \mathcal{C}\mathcal{N}(\mathbf{y};\mathbf{A}\mathbf{x},{\mathbf{\Sigma }}_2)$，则关于$\mathbf{x}$的后验分布：
$$\begin{array}{rl}p(\mathbf{x}|\mathbf{y})& \propto \mathcal{C}\mathcal{N}(\mathbf{x};\mathbf{r},{\mathbf{\Sigma }}_1)\cdot \mathcal{C}\mathcal{N}(\mathbf{y};\mathbf{A}\mathbf{x},{\mathbf{\Sigma }}_2)\\ & \propto \exp \left\{-(\mathbf{x}-\mathbf{r}{)}^H{\mathbf{\Sigma }}_1^{-1}(\mathbf{x}-\mathbf{r})-(\mathbf{A}\mathbf{x}-\mathbf{y}{)}^H{\mathbf{\Sigma }}_2^{-1}(\mathbf{A}\mathbf{x}-\mathbf{y})\right\}\\ & \propto \exp \left\{{\mathbf{x}}^H{\mathbf{\Sigma }}_1^{-1}\mathbf{x}-2\mathcal{R}\left\{{\mathbf{r}}^H{\mathbf{\Sigma }}_1^{-1}\mathbf{x}\right\}-{\mathbf{x}}^H{\mathbf{A}}^H{\mathbf{\Sigma }}_2^{-1}\mathbf{A}\mathbf{x}-2\mathcal{R}\left\{{\mathbf{y}}^H{\mathbf{\Sigma }}_2^{-1}\mathbf{A}\mathbf{x}\right\}\right\}\\ & =\exp \left\{{\mathbf{x}}^H\left({\mathbf{\Sigma }}_1^{-1}+{\mathbf{A}}^H{\mathbf{\Sigma }}_2^{-1}\mathbf{A}\right)\mathbf{x}-2\mathcal{R}\left\{\left({\mathbf{r}}^H{\mathbf{\Sigma }}_1^{-1}+{\mathbf{y}}^H{\mathbf{\Sigma }}_2^{-1}\mathbf{A}\right)\mathbf{x}\right\}\right\}\\ & =\exp \left\{-{\mathbf{x}}^H\left[\underset{={\Sigma }_x^{-1}}{\underbrace{{\Sigma }_1^{-1}+{{\mathbf{A}}^H{\mathbf{\Sigma }}_2^{-1}\mathbf{A}}^{}}}\right]\mathbf{x}+2\mathcal{R}\left\{{\left[\underset{={\mathbf{\mu }}_x}{\underbrace{{\Sigma }_x\left({\Sigma }_1^{-1}\mathbf{r}+{\mathbf{A}}^H{\Sigma }_2^{-1}\mathbf{y}\right)}}\right]}^H{\mathbf{\Sigma }}_x^{-1}\mathbf{x}\right\}\right\}\end{array}$$

所以有
$$\begin{array}{rl}{\mathbf{\Sigma }}_x& ={\left({\Sigma }_1^{-1}+{\mathbf{A}}^H{\mathbf{\Sigma }}_2^{-1}{\mathbf{A}}^{}\right)}^{-1}\\ \mathbf{x}& ={\left({\Sigma }_1^{-1}+{\mathbf{A}}^H{\mathbf{\Sigma }}_2^{-1}{\mathbf{A}}^{}\right)}^{-1}\left({\Sigma }_1^{-1}\mathbf{r}+{\mathbf{A}}^H{\Sigma }_2^{-1}\mathbf{y}\right)\end{array}$$
事实上，$\mathbf{x}=\mathbb{E}[\mathbf{x}|\mathbf{y}]$，所以该均值也是MMSE估计的结果，因为其估计结果是线性的，所以也称为“LMMSE”。