Mahalanobis马氏距离，正态随机变量，协方差矩阵与相关矩阵

正交矩阵与正交变换

如果n阶矩阵A满足$A^{T}A=E$（即$A^{-1}=A^T$），那么称A为正交矩阵，简称正交阵。

对称矩阵的对角化

设$A$为$n$阶对称矩阵，则必有正交矩阵P，使$P^{-1}AP=P^{T}AP=\mathrm{\Lambda }$，其中$\mathrm{\Lambda }$ 是以$A$的$n$个特征值为对角元的对角矩阵。

为什么要求马氏距离?

在一个身高样本中，衡量其中的一个身高$l_i$，可以用与均值的距离$l_i-u_l$，即随机变量与其均值的偏离程度。如果将这个样本中所有变量标准化，随机变量与均值的距离为$\frac{l_i-u_l}{\sigma }$。比较其中两个身高间的距离，我们可以用$\frac{l_i-u_l}{{\sigma }_l}-\frac{l_j-u_l}{{\sigma }_l}$。（标准化后的均值为0）

同理，在一个体重样本中，衡量其中的一个体重$m_i$，可以用其标准化后的“距离”$\frac{m_i-{\mu }_m}{{\sigma }_m}$。

那么在一个(身高，体重) 二维样本中，一个单位是长度，一个单位是重量，该如何衡量随机变量$\left(l_i,m_i\right)$呢？是不是可以分别标准化以后，得到$\left(\frac{l_i-{\mu }_l}{{\sigma }_l},\frac{m_i-{\mu }_m}{{\sigma }_m}\right)$，再按照$\mathrm{\Delta }\mathrm{L}=\sqrt{x^2+y^2}$，求这个点到中心$\left(0,0\right)$的距离呢？ 而对于两个随机变量$\left(l_i,m_i\right)$与$\left(l_j,m_j\right)$的比较，是不是可以分别标准化以后，得到$\left(\frac{l_i-{\mu }_l}{{\sigma }_l},\frac{m_i-{\mu }_m}{{\sigma }_m}\right)$与$\left(\frac{l_j-{\mu }_l}{{\sigma }_l},\frac{m_j-{\mu }_m}{{\sigma }_m}\right)$，再按照$\mathrm{\Delta }\mathrm{L}=\sqrt{{\left(x_1-x_2\right)}^2+{\left(y_1-y_2\right)}^2}$，求这两个点之间的距离呢？

马氏距离定义（Baidu）

马氏距离(Mahalanobis distance)是由印度统计学家马哈拉诺比斯(P. C. Mahalanobis)提出的，表示点与一个分布之间的距离。它是一种有效的计算两个未知样本集的相似度的方法。与欧氏距离不同的是，它考虑到各种特性之间的联系（例如：一条关于身高的信息会带来一条关于体重的信息，因为两者是有关联的），并且是尺度无关的(scale-invariant)，即独立于测量尺度。对于一个均值为$\mu$，协方差矩阵为$C$的多变量向量，其马氏距离为$\sqrt{{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)}$。

正态随机变量与马氏距离

二维正态随机变量X1,X2的概率密度为
$$f\left(x_1,x_2\right)=\frac{1}{2\pi {\sigma }_1{\sigma }_2\sqrt{1-{\rho }^2}}exp\left\{\frac{-1}{2\left(1-{\rho }^2\right)}\left[\frac{{\left(x_1-{\mu }_1\right)}^2}{{\sigma }_1^2}-2\rho \frac{\left(x_1-{\mu }_1\right)\left(x_2-{\mu }_2\right)}{{\sigma }_1{\sigma }_2}+\frac{{\left(x_2-{\mu }_2\right)}^2}{{\sigma }_2^2}\right]\right\}$$
将上式中花括号内的式子写成矩阵形式，为此引入下面的列矩阵
$$X=\left(\genfrac{}{}{0ex}{}{x_1}{x_2}\right),\mu =\left(\genfrac{}{}{0ex}{}{{\mu }_1}{{\mu }_2}\right)。$$
$(X_1,X_2)$的协方差矩阵为
$$C=\left[\begin{array}{cc}{c}_{11}& c_{12}\\ c_{21}& c_{22}\end{array}\right]=\left[\begin{array}{cc}{\sigma }_1^2& \rho {\sigma }_1{\sigma }_2\\ \rho {\sigma }_1{\sigma }_2& {\sigma }_2^2\end{array}\right]$$
它的行列式 $detC={\sigma }_1{\sigma }_2\left(1-{\rho }^2\right)$，$C$的逆矩阵为
$$C^{-1}=\frac{1}{detC}\left[\begin{array}{cc}{\sigma }_2^2& -\rho {\sigma }_1{\sigma }_2\\ -\rho {\sigma }_1{\sigma }_2& {\sigma }_1^2\end{array}\right]$$
经过计算可知
${\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)=\frac{1}{detC}\left(\begin{array}{cc}{x}_1-{\mu }_1& x_2-{\mu }_2\end{array}\right)\left[\begin{array}{cc}{\sigma }_2^2& -\rho {\sigma }_1{\sigma }_2\\ -\rho {\sigma }_1{\sigma }_2& {\sigma }_1^2\end{array}\right]\left(\genfrac{}{}{0ex}{}{x_1-{\mu }_1}{x_2-{\mu }_2}\right)$
$$=\frac{1}{1-{\rho }^2}\left[\frac{{\left(x_1-{\mu }_1\right)}^2}{{\sigma }_1^2}-2\rho \frac{\left(x_1-{\mu }_1\right)\left(x_2-{\mu }_2\right)}{{\sigma }_1{\sigma }_2}+\frac{{\left(x_2-{\mu }_2\right)}^2}{{\sigma }_2^2}\right]$$
于是$(X_1,X_2)$的概率密度可写成
$$f\left(x_1,x_2\right)=\frac{1}{{2\pi }^{2/2}{detC}^{1/2}}exp\left\{-\frac{1}{2}{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)\right\}$$
上式容易推广到$n$维正态随机变量$\left(X_1,X_2,X_3,\cdots ,X_n\right)$的情况。
引入矩阵
$$X=\left[\begin{array}{c}{x}_1\\ x_2\\ ⋮\\ x_n\end{array}\right]和u=\left[\begin{array}{c}{\mu }_1\\ {\mu }_2\\ ⋮\\ {\mu }_n\end{array}\right]=\left[\begin{array}{c}E\left(X_1\right)\\ E\left(X_2\right)\\ ⋮\\ E\left(X_n\right)\end{array}\right]$$
$n$维正态随机变量$\left(X_1,X_2,X_3,\cdots ,X_n\right)$的概率密度定义为
$$f\left(x_1,x_2,\cdots ,x_n\right)=\frac{1}{{2\pi }^{n/2}{detC}^{1/2}}exp\left\{-\frac{1}{2}{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)\right\}$$
其中$C$是$\left(X_1,X_2,X_3,\cdots ,X_n\right)$的协方差矩阵。
比较马氏距离定义中的$\sqrt{{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)}$与$n$维正态随机变量的概率密度定义，可以看出，都包含了${\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)$。

公式解读

$X_1$和$X_2$相互独立$(\rho =0)$

二维正态随机变量$\left(X_1,X_2\right)$，如果$X_1$和$X_2$相互独立（即不相关），即$\rho =0$，则
$$C=\left[\begin{array}{cc}{c}_{11}& c_{12}\\ c_{21}& c_{22}\end{array}\right]=\left[\begin{array}{cc}{\sigma }_1^2& \rho {\sigma }_1{\sigma }_2\\ \rho {\sigma }_1{\sigma }_2& {\sigma }_2^2\end{array}\right]=\left[\begin{array}{cc}{\sigma }_1^2& \\ & {\sigma }_2^2\end{array}\right]$$
$$C^{-1}=\frac{1}{detC}\left[\begin{array}{cc}{\sigma }_2^2& -\rho {\sigma }_1{\sigma }_2\\ -\rho {\sigma }_1{\sigma }_2& {\sigma }_1^2\end{array}\right]=\left[\begin{array}{cc}1/{\sigma }_1^2& \\ & 1/{\sigma }_2^2\end{array}\right]$$
${\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)=\left(\begin{array}{cc}{x}_1-{\mu }_1& x_2-{\mu }_2\end{array}\right)\left[\begin{array}{cc}1/{\sigma }_1^2& \\ & 1/{\sigma }_1^2\end{array}\right]\left(\genfrac{}{}{0ex}{}{x_1-{\mu }_1}{x_2-{\mu }_2}\right)$
$$=\left(\begin{array}{cc}\frac{x_1-{\mu }_1}{{\sigma }_1}& \frac{x_2-{\mu }_2}{{\sigma }_2}\end{array}\right)\left(\genfrac{}{}{0ex}{}{\frac{x_1-{\mu }_1}{{\sigma }_1}}{\frac{x_2-{\mu }_2}{{\sigma }_2}}\right)$$
$$={\left(\frac{x_1-{\mu }_1}{{\sigma }_1}\right)}^2+{\left(\frac{x_2-{\mu }_2}{{\sigma }_2}\right)}^2$$
根据马氏距离定义，
$$\sqrt{{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)}=\sqrt{{\left(\frac{x_1-{\mu }_1}{{\sigma }_1}\right)}^2+{\left(\frac{x_2-{\mu }_2}{{\sigma }_2}\right)}^2}$$
显然，$\frac{x_1-{\mu }_1}{{\sigma }_1}、\frac{x_2-{\mu }_2}{{\sigma }_2}$都是标准化后的变量，马氏距离为点$\left(X_1,X_2\right)$标准化后到中心$\left(0,0\right)$的距离。
这样，是不是就符合我们前面衡量一个（身高，体重），猜想的方法呢？
事实是，现实中，身高与体重并不是毫不相关的，即相关系数$\rho \ne 0$。
那么，当$X_1$和$X_2$相关（并不独立）时，如何衡量一个（身高，体重）呢？如何理解$\sqrt{{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)}$呢?

$X_1$和$X_2$相关$(\rho \ne 0)$

二维正态随机变量（略）。
对多维正态随机变量，
$C=\left[\begin{array}{cc}\begin{array}{cc}{c}_{11}& c_{12}\\ c_{21}& c_{22}\end{array}& \begin{array}{cc}\dots & c_{1n}\\ \dots & c_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ c_{n1}& c_{n2}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & c_{nn}\end{array}\end{array}\right]=\left[\begin{array}{cc}\begin{array}{cc}{\sigma }_1^2& {\rho }_{12}{\sigma }_1{\sigma }_2\\ {\rho }_{12}{\sigma }_1{\sigma }_2& {\sigma }_2^2\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}{\sigma }_1{\sigma }_n\\ \dots & {\rho }_{2n}{\sigma }_2{\sigma }_n\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}{\sigma }_1{\sigma }_n& {\rho }_{2n}{\sigma }_2{\sigma }_n\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & {\sigma }_n^2\end{array}\end{array}\right]$
$$=\left[\begin{array}{cc}\begin{array}{cc}{\sigma }_1& \\ & {\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\sigma }_n\end{array}\end{array}\right]\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]\left[\begin{array}{cc}\begin{array}{cc}{\sigma }_1& \\ & {\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\sigma }_n\end{array}\end{array}\right]$$

根据逆矩阵运算规律
$C^{-1}={\left(\left[\begin{array}{cc}\begin{array}{cc}{\sigma }_1& \\ & {\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\sigma }_n\end{array}\end{array}\right]\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]\left[\begin{array}{cc}\begin{array}{cc}{\sigma }_1& \\ & {\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\sigma }_n\end{array}\end{array}\right]\right)}^{-1}$
$$={\left[\begin{array}{cc}\begin{array}{cc}{\sigma }_1& \\ & {\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\sigma }_n\end{array}\end{array}\right]}^{-1}{\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]}^{-1}{\left[\begin{array}{cc}\begin{array}{cc}{\sigma }_1& \\ & {\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\sigma }_n\end{array}\end{array}\right]}^{-1}$$
$$=\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]{\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]}^{-1}\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]$$
即协方差矩阵$C$可表示为方差与相关矩阵$R$的关系。

则马氏距离也可以表示为 相关矩阵与标准化后变量的关系，如下：
$\sqrt{{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)}$
$=\sqrt{\left(\begin{array}{cc}\begin{array}{cc}{x}_1-{\mu }_1& x_2-{\mu }_2\end{array}& \begin{array}{cc}\dots & x_n-{\mu }_n\end{array}\end{array}\right)\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]{\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]}^{-1}\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]\left(\begin{array}{c}\begin{array}{c}{x}_1-{\mu }_1\\ x_2-{\mu }_2\end{array}\\ \begin{array}{c}⋮\\ x_n-{\mu }_n\end{array}\end{array}\right)}$
$=\sqrt{\left(\begin{array}{cc}\begin{array}{cc}\frac{x_1-{\mu }_1}{{\sigma }_1}& \frac{x_2-{\mu }_2}{{\sigma }_2}\end{array}& \begin{array}{cc}\dots & \frac{x_n-{\mu }_n}{{\sigma }_n}\end{array}\end{array}\right){\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]}^{-1}\left(\begin{array}{c}\begin{array}{c}\frac{x_1-{\mu }_1}{{\sigma }_1}\\ \frac{x_2-{\mu }_2}{{\sigma }_2}\end{array}\\ \begin{array}{c}⋮\\ \frac{x_n-{\mu }_n}{{\sigma }_n}\end{array}\end{array}\right)}$
$=\sqrt{{\left(\frac{X-\mu }{\sigma }\right)}^T{\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]}^{-1}\left(\frac{X-\mu }{\sigma }\right)}$

即马氏距离中偏差$\left(X-\mu \right)$与协方差矩阵$C$的计算，变为了标准化矩阵$\left(\frac{X-\mu }{\sigma }\right)$与相关矩阵R的计算。
已知相关矩阵$R=\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]$为对称矩阵，可知有正交矩阵$P$，使$R=P_R^{-1}\left[\begin{array}{cc}\begin{array}{cc}{\lambda }_{1R}& \\ & {\lambda }_{2R}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\lambda }_{nR}\end{array}\end{array}\right]P_R，\left({\lambda }_{1R},{\lambda }_{2R},{\lambda }_{3R},\cdots ,{\lambda }_{nR}\right)$是$R$的$n$个特征值.

继续马氏距离公式的变换，
$\sqrt{{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)}$
$=\sqrt{{\left(\frac{X-\mu }{\sigma }\right)}^T{\left[\begin{array}{cc}\begin{array}{cc}1& {\rho }_{12}\\ {\rho }_{12}& 1\end{array}& \begin{array}{cc}\dots & {\rho }_{1n}\\ \dots & {\rho }_{2n}\end{array}\\ \begin{array}{cc}⋮& ⋮\\ {\rho }_{1n}& {\rho }_{2n}\end{array}& \begin{array}{cc}\ddots & ⋮\\ \dots & 1\end{array}\end{array}\right]}^{-1}\left(\frac{X-\mu }{\sigma }\right)}$
$=\sqrt{{\left(\frac{X-\mu }{\sigma }\right)}^T{\left(P_R^{-1}\left[\begin{array}{cc}\begin{array}{cc}{\lambda }_{1R}& \\ & {\lambda }_{2R}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\lambda }_{nR}\end{array}\end{array}\right]P_R\right)}^{-1}\left(\frac{X-\mu }{\sigma }\right)}$
$=\sqrt{{\left(\frac{X-\mu }{\sigma }\right)}^T{P}_R^{-1}{\left[\begin{array}{cc}\begin{array}{cc}{\lambda }_{1R}& \\ & {\lambda }_{2R}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\lambda }_{nR}\end{array}\end{array}\right]}^{-1}{P}_R\left(\frac{X-\mu }{\sigma }\right)}$
$=\sqrt{{\left(\frac{X-\mu }{\sigma }\right)}^T{P}_R^{-1}\left[\begin{array}{cc}\begin{array}{cc}1/{\lambda }_{1R}& \\ & 1/{\lambda }_{2R}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\lambda }_{nR}\end{array}\end{array}\right]P_R\left(\frac{X-\mu }{\sigma }\right)}$
$=\sqrt{{\left(\frac{X-\mu }{\sigma }\right)}^T{P}_R^T{\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]}^T\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left(\frac{X-\mu }{\sigma }\right)}$
$=\sqrt{{\left(\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left(\frac{X-\mu }{\sigma }\right)\right)}^T\left(\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left(\frac{X-\mu }{\sigma }\right)\right)}$
$=\sqrt{{\left(\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]\left(\begin{array}{c}\begin{array}{c}{x}_1-{\mu }_1\\ x_2-{\mu }_2\end{array}\\ \begin{array}{c}⋮\\ x_n-{\mu }_n\end{array}\end{array}\right)\right)}^T\left(\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]\left(\begin{array}{c}\begin{array}{c}{x}_1-{\mu }_1\\ x_2-{\mu }_2\end{array}\\ \begin{array}{c}⋮\\ x_n-{\mu }_n\end{array}\end{array}\right)\right)}$

观察$\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]\left(\begin{array}{c}\begin{array}{c}{x}_1-{\mu }_1\\ x_2-{\mu }_2\end{array}\\ \begin{array}{c}⋮\\ x_n-{\mu }_n\end{array}\end{array}\right)$，可以看到
$X\stackrel{标准化}{===\Rightarrow }\frac{X-\mu }{\sigma }\stackrel{正交变换(旋转)}{======\Rightarrow }{P}_R\left(\frac{X-\mu }{\sigma }\right)\stackrel{缩放变换}{====\Rightarrow }\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left(\frac{X-\mu }{\sigma }\right)$

$X$标准化后，先经过正交变换，再经过缩放变换，才得到变换后的点。马氏距离即是求该点到中心$\left(0,0\right)$的距离。

以上通过对公式的解读过程，分析了标准化后变量$\frac{X-\mu }{\sigma }$,相关矩阵与马氏距离的关系,

如果不需要求标准化的变量，可以直接用马氏距离公式求解。
已知协方差矩阵$C$为对称矩阵，可知有正交矩阵$P_C$，使$C=P_C^{-1}\left[\begin{array}{cc}\begin{array}{cc}{\lambda }_{1C}& \\ & {\lambda }_{2C}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & {\lambda }_{nC}\end{array}\end{array}\right]P_C$，$\left({\lambda }_{1C},{\lambda }_{2C},{\lambda }_{3C},\cdots ,{\lambda }_{nC}\right)$是$C$的$n$个特征值.
则马氏距离公式可变为，
$\sqrt{{\left(X-\mu \right)}^T{C}^{-1}\left(X-\mu \right)}$
$=\sqrt{{\left(\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1C}}& \\ & 1/\sqrt{{\lambda }_{2C}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nC}}\end{array}\end{array}\right]P_C\left(\begin{array}{c}\begin{array}{c}{x}_1-{\mu }_1\\ x_2-{\mu }_2\end{array}\\ \begin{array}{c}⋮\\ x_n-{\mu }_n\end{array}\end{array}\right)\right)}^T\left(\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1C}}& \\ & 1/\sqrt{{\lambda }_{2C}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nC}}\end{array}\end{array}\right]P_C\left(\begin{array}{c}\begin{array}{c}{x}_1-{\mu }_1\\ x_2-{\mu }_2\end{array}\\ \begin{array}{c}⋮\\ x_n-{\mu }_n\end{array}\end{array}\right)\right)}$

观察$\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1C}}& \\ & 1/\sqrt{{\lambda }_{2C}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nC}}\end{array}\end{array}\right]P_C\left(\begin{array}{c}\begin{array}{c}{x}_1-{\mu }_1\\ x_2-{\mu }_2\end{array}\\ \begin{array}{c}⋮\\ x_n-{\mu }_n\end{array}\end{array}\right)$，可以看到
$X\stackrel{与均值的偏差}{=====\Rightarrow }X-\mu \stackrel{正交变换(旋转)}{=====\Rightarrow }{P}_C\left(X-\mu \right)\stackrel{缩放变换}{===\Rightarrow }\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1C}}& \\ & 1/\sqrt{{\lambda }_{2C}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nC}}\end{array}\end{array}\right]P_C\left(X-\mu \right)$

用图像表示上述变换，

由$\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1R}}& \\ & 1/\sqrt{{\lambda }_{2R}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nR}}\end{array}\end{array}\right]P_R\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]=\left[\begin{array}{cc}\begin{array}{cc}1/\sqrt{{\lambda }_{1C}}& \\ & 1/\sqrt{{\lambda }_{2C}}\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/\sqrt{{\lambda }_{nC}}\end{array}\end{array}\right]P_C$可知，$P_C$对应的特征向量经过$\left[\begin{array}{cc}\begin{array}{cc}1/{\sigma }_1& \\ & 1/{\sigma }_2\end{array}& \\ & \begin{array}{cc}\ddots & \\ & 1/{\sigma }_n\end{array}\end{array}\right]$缩放变换后，得到了$P_R$对应特征向量的角度。在此不做分析。

从马氏距离公式的分析中可以看到，协方差矩阵的逆矩阵中已经包含了对选取变量$\left(X_1,X_2,X_3,\cdots ,X_n\right)$的标准化、旋转、缩放。故在计算马氏距离的过程中，无需求解旋转角度（即单位特征向量），缩放比例（特征值）。
在主成分分析和因子分析中，需要进一步求解协方差矩阵的逆矩阵中已经包含的旋转、缩放信息，即单位特征向量与特征值。主成分分析和因子分析在这里不做介绍。

参考资料:

1. 概率论与数理统计（浙大第四版）
2. 线性代数（同济第六版）