矩阵特征值和特征向量¶
\[ \begin{aligned} A\vec{v} &= \lambda \vec{v} \\ |A - \lambda I| &= 0 \end{aligned} \]
求解行列式的代价很大!
幂法¶
假设 \(A\) 的特征值满足 \(\lambda_1 > \lambda_2 > \cdots > \lambda_n\),找 \(x_0 \neq 0\),
用特征向量表示 \(x_0\)
\[ x_0 = \sum_{i=1}^{n} \alpha_i u_i \]
\(u_i\) 是 \(\lambda_i\) 对应的特征向量。
那么
\[ A x_0 = \sum_{i=1}^{n} \alpha_i \underline{A u_i} = \sum_{i=1}^{n} \alpha_i \lambda_i u_i \]
经过多次迭代,
\[ \begin{aligned} x_{k+1} = A x_k &= A^{k+1} x_0 \\ &= A^{k+1} \sum_{i=1}^{n} \alpha_i u_i \\ &= \sum_{i=1}^{n} \alpha_i \lambda_i^{k+1} u_i \\ &= \lambda_1^{k+1} \left(\alpha_1 u_1 + \sum_{i=2}^{n} \alpha_i \left(\frac{\lambda_i}{\lambda_1}\right)^{k+1} u_i\right) \\ &\overset{k \to \infty}{\approx} \lambda_1^{k+1} \alpha_1 u_1 \end{aligned} \]
故得到
\[ \lambda_1 \approx \frac{x_i^{(k+1)}}{x_i^{(k)}} \]