矩阵取对数和 Fisher 变换之间关系
Published:
考虑一个 $n \times n$ 满秩(可逆)矩阵 $A$, 如何求解 $\log A$ ?
一种方法是进行级数展开:
\(\log(A)= log(I+B)=\sum _{k=1}^{\infty }{(-1)^{k+1}{\frac {B^{k}}{k}}}=B-{\frac {B^{2}}{2}}+{\frac {B^{3}}{3}}-{\frac {B^{4}}{4}}+\cdots\)
另外一种方法是积分法(注意:使用该方法的条件是 的特征值不为负值): $\log A=\int^1_0(A-I)[t(A-I)+I]^{-1}dt.$
例子:考虑如下矩阵: \(A = \begin{pmatrix} 1,&0\\ 1,&1\end{pmatrix},\) 求 $\log A$.
- 采用第一种方法 $A= I+B\ $
这里
\(B= \left(\begin{aligned} &0,0\\ &1,0\end{aligned}\right)\\\) 因此有:
\(\log A=\log(I+B) = \sum_{k=1}^{\infty }{(-1)^{k+1}{\frac {B^{k}}{k}}}=B-{\frac {B^{2}}{2}}+{\frac {B^{3}}{3}}-{\frac {B^{4}}{4}}+\cdots\)
注意到:
\(B^k= \begin{pmatrix} 0,&0\\ 1,&0\end{pmatrix} \begin{pmatrix} 0,&0\\ 1,&0\end{pmatrix} \cdot B^{k-2}=\begin{pmatrix} 0,&0\\ 0,&0\end{pmatrix},k=2,3,\cdots\\\)
所以
\(\log A=B= \begin{pmatrix} 0,&0\\ 1,&0\end{pmatrix}\)
- 采用第二种方法,首先验证矩阵 $A$ 的特征值为 1 不是负数,满足条件,
Fisher 变换 考虑两个随机变量之间的相关关系为 $\rho\in(0,1)$ ,则 Fisher 变换(也称Fisher Z-变换): \({\displaystyle z={1 \over 2}\ln \left({1+\rho\over 1-\rho}\right).}\)
对数相关矩阵和Fisher 变换之间的关系: 为解释二者之间的联系,我们考虑一个二维随机变量 $(X,Y)$ 形成的相关矩阵(记 $(X,Y)$ 之间的相关系数为 $\rho\in (0,1)$): \(\left( \begin{aligned} &1,\rho\\ &\rho,1 \end{aligned} \right)\)
则(这里采用积分的方法,因为特征根为 $1\pm \rho$ 不为负): \(\begin{aligned} \log \left( \begin{aligned} &1,\rho\\ &\rho,1 \end{aligned} \right) &=\int^1_0\left( \left( \begin{aligned} &1,\rho\\ &\rho,1 \end{aligned} \right)- \left( \begin{aligned} &1,0\\ &0,1 \end{aligned} \right)\right)\left[t\left( \left( \begin{aligned} &1,\rho\\ &\rho,1 \end{aligned} \right)- \left( \begin{aligned} &1,0\\ &0,1 \end{aligned} \right)\right)+ \left( \begin{aligned} &1,0\\ &0,1 \end{aligned} \right)\right]^{-1}dt\\ &=\int^1_0\left( \begin{aligned} &0,\rho\\ &\rho,0 \end{aligned} \right)\left( \begin{aligned} &1,t\rho\\ &t\rho,1 \end{aligned} \right)^{-1}dt\\ &=\int^1_0\left( \begin{aligned} &0,\rho\\ &\rho,0 \end{aligned} \right)\left( \begin{aligned} &\frac{1}{1-t^2\rho^2},\frac{-t\rho}{1-t^2\rho^2}\\ &\frac{1}{1-t^2\rho^2},\frac{1}{1-t^2\rho^2}\end{aligned} \right)dt\\ &=\int^1_0\left( \begin{aligned} &-\frac{t\rho^2}{1-t^2\rho^2},\frac{\rho}{1-t^2\rho^2}\\ &\frac{\rho}{1-t^2\rho^2},-\frac{1\rho^2}{1-t^2\rho^2}\end{aligned} \right)dt\\ &=\left( \begin{aligned} &\frac{1}{2}\ln(1-\rho^2),\frac{1}{2}\ln(\frac{1+\rho}{1-\rho})\\ &\frac{1}{2}\ln(\frac{1+\rho}{1-\rho}),\frac{1}{2}\ln(1-\rho^2)\end{aligned} \right).\\ \end{aligned}\)
至此,得到对数相关矩阵的结果。注意到副对角线正是 Fisher 变换。这个结论在 Arvhakov & Hansen 的 Econometrica 2021 的文章中也有提到。
参考文献:
- Archakov, I., & Hansen, P. R. (2021). A new parametrization of correlation matrices.Econometrica,89(4), 1699-1715.
- Higham, N. J. (2008).Functions of matrices: theory and computation. Society for Industrial and Applied Mathematics.