为什么 w T Σ w w^T\Sigma w wTΣw 代表类内散度?
在 LDA 的推导里,说 w T Σ w w^T\Sigma w wTΣw 代表了类内散度,但是为什么呢?
画图观察
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
S = np.array([[1,0],
[0,1]])
mu = np.array([0,0])
s3 = np.random.multivariate_normal(mu,S,500) # generate 1000 samples from the distribution
# plot the scatter
plt.figure(figsize=(10,10))
plt.scatter(s3[:,0],s3[:,1])
# set xlim and ylim
plt.xlim(-10,10)
plt.ylim(-10,10)
# add grid
plt.grid(True)
# display the scatter matrix in latex in title
plt.title(str(S[0,0])+' '+str(S[0,1])+'\n'+str(S[1,0])+' '+str(S[1,1]),fontsize=20)
结论
Σ
=
[
σ
11
σ
12
σ
21
σ
22
]
\Sigma =\begin{bmatrix} \sigma_{11} & \sigma_{12} \\ \sigma_{21} & \sigma_{22} \end{bmatrix}
Σ=[σ11σ21σ12σ22]
w
T
Σ
w
=
σ
11
w
1
2
+
σ
12
w
1
w
2
+
σ
21
w
1
w
2
+
σ
22
w
2
2
w^T\Sigma w = \sigma_{11}w_{1}^2+\sigma_{12}w_1w_2+\sigma_{21}w_1w_2+\sigma_{22}w_2^2
wTΣw=σ11w12+σ12w1w2+σ21w1w2+σ22w22
怎么样可以让散度减小呢?观察展开式,其实就是让协方差矩阵的这几个数都尽量小即可。