【Numberical Optimization】5 Conjugate Gradient Methods (zen学习笔记)

5.1  The linear conjugate gradient methods

The conjugate gradient method is an itterative method for solving a linear system of equations:

                                                                                            Ax=b

which can be stated equivalently as the following minimization problem:

                                                                            min\,\phi (x)=\frac{1}{2}x^{T}Ax-b^{T}x

第k步的残差:r_{k}=Ax_{k}-b

5.1.1  Conjugate direction methods

1.阐述了共轭方向算法能在有限步内收敛到最优点(步数主要与特征值分布有关)

The conjugate vectors are given.

  • \left \{ p_{0} ,p_{1},...,p_{l}\right \} is said to be conjugate with respect to the symmetric positive definite martix A if p_{i}^{T}Ap_{j}=0,\quad for\, all\,i\neq j
  • \left \{ p_{0} ,p_{1},...,p_{l}\right \} is also linearly independent.
  • 有限步内能收敛到最优解,证明及算法如下:

 

Defining \hat x as 

                                                                                      \hat x=S^{-1}x

where S is defined by S=[p_{0}\,p_{1}...p_{n-1}]

then \hat \phi (\hat x)=\phi (S \hat x)=\frac{1}{2}\hat x^{T}(S^{T}AS)\hat x-(S^{T}b)^{T}\hat x

2.最优点在集合中      {\color{Blue} \left \{ x|x=x_{0} +span\left \{ p_{0},p_{1},...,p_{k-1} \right \}\right \}}即可由x_{0},\left \{ p_{0},p_{1},...,p_{k-1} \right \}线性表出

5.1.2  Basic properties of the conjugate gradient method

          A PRACTICAL FORM OF THE CONJUGATE GRADIENT METHOD

前面介绍了共轭梯度算法的一些性质,接下来的算法提出共轭方向\left \{ p_{0},p_{1},...,p_{k-1} \right \}的构造

共轭梯度初步算法及共轭梯度算法

改进之处在于:

  • ​​​​​​​\alpha _{k}=-\frac{r_{k}^{T}p_{k}}{p_{k}^{T}Ap_{k}}=-\frac{r_{k}^{T}(-r_{k}+\beta_{k}p_{k-1})}{p_{k}^{T}Ap_{k}}\rightarrow \frac{r_{k}^{T}r_{k}}{p_{k}^{T}Ap_{k}}
  • \beta _{k}=\frac{r_{k+1}^{T}Ap_{k}}{p_{k}^{T}Ap_{k}}=\frac{r_{k+1}^{T}(\frac{r_{k+1}-r_{k}}{\alpha _{k}})}{p_{k}^{T}(\frac{r_{k+1}-r_{k}}{\alpha _{k}})}\rightarrow \frac{r_{k+1}^{T}r_{k+1}}{r_{k}^{T}r_{k}}

5.1.3  RATE OF CONVERGENCE

解释收敛步数与特征值分布的关系、验证 CG算法在某种意义上是最优的

猜你喜欢

转载自blog.csdn.net/weixin_38716567/article/details/83005117