0
$\begingroup$

I was given the following question as part of a homework assignment. Any help would be greatly appreciated!

The following image shows the steps of a preliminary version of the conjugate gradient algorithm and was taken from the book Numerical Optimization by Jorge Nocedal and Stephen J. Wright.

Preliminary version of the conjugate gradient algorithm taken from the book Numerical Optimization by Jorge Nocedal and Stephen J. Wright

We know from a previous theorem that $p_{k+1} \in \text{span}\{r_0, A r_0, \dots, A^{k+1} r_0\}$, $A \in \mathbb{R}^{n \times n}$ is symmetric positive definite and $r_k \in \mathbb{R}^{n \times 1}$, therefore $$ p_{k+1} = \xi_0 r_0 + \xi_1 A r_0 + \dots + \xi_{k+1} A^{k+1} r_0 $$ Show that $\xi_{k+1} = (-1)^{k+1}(\alpha_0 \times \alpha_1 \times \dots \times \alpha_k)$ using the preliminary version of the conjugate gradient method (Algorithm 5.1).

I have tried to solve this problem by Mathematical induction on $k$ since the base step is easy to prove, however, I haven't figured out a way to succesfully complete the induction step. I have tried using the following results:

  1. $r_{k+1} = r_k + \alpha_{k} A r_{k}$
  2. $r_k^T p_i = 0$ for $i = 0, 1, \dots, k-1$
  3. $p_k^T A p_i = 0$ for $i = 0, 1, \dots, k-1$
  4. $r_k^T r_i = 0$ for $i = 0, 1, \dots, k-1$

Note: the instructions never stated this, but I assumed that $\xi_0 = -1$ since it's essentially a definition in Algorithm 5.1.

$\endgroup$

1 Answer 1

1
$\begingroup$

I managed to prove it (or at least something like it) using induction. Here's what we'll need:

  1. $r_{k+1} = r_k + \alpha_k A p_k$. (deduced from Algorithm 5.1)
  2. If $p_k \in \text{span}\{r_0, Ar_0, \dots, A^k r_0\}$, then $Ap_k \in \text{span}\{Ar_0, A^2 r_0, \dots, A^{k+1} r_0\}$.
  3. $r_k \in \text{span}\{r_0, Ar_0, \dots, A^k r_0\}$.
  4. By definition, $p_0 = -r_0$.

Base Step

Let $k = 1$, then

$$p_k = p_1 = -r_1 + \beta_0 p_0 = -r_0 - \alpha_0 A p_0 + \beta_0 p_0$$

Since $p_0 = -r_0$, we get

$$p_1 = -r_0 + \alpha_0 A r_0 - \beta_0 r_0 = -(1 + \beta_0) r_0 + \alpha_0 Ar_0$$

Therefore, $p_k = p_1 \in \text{span}\{r_0, A r_0\} = \text{span}\{r_0, \dots, A^k r_0\}$.

Inductive step

Let $k$ be a natural number. Our induction hypothesis is as follows:

$$\xi_k = (-1)^{k+1} (\alpha_0 \times \alpha_1 \times \dots \times \alpha_{k-1})$$

We now show that this equation holds for $k+1$.

$$p_{k+1} = -r_{k+1} + \beta_k p_k = -r_k - \alpha_k Ap_k + \beta_k p_k$$

therefore,

$$-\alpha_k A p_k = r_k - r_{k+1}$$

Since $r_k \in \text{span}\{r_0, Ar_0, \dots, A^k r_0\}$, we have

$$-\alpha_k A p_k = r_k - r_{k+1} = \sum_{i=0}^k \xi_i A^i r_0 - \sum_{i=0}^{k +1} \xi_i A^i r_0 = -\xi_{k+1} A^{k+1} r_0$$

Multiply by $A^{-1}$ from the left

$$\alpha_k p_k = \xi_{k+1} A^k r_0 \Rightarrow \alpha_k p_k - \xi_{k+1} A^k r_0 = 0$$

Now, dividing by $\alpha_k$ and substituting $p_k$ for its expression as a linear combination, we have

$$p_k - \frac{\xi_{k+1}}{\alpha_k} A^k r_0 = \sum_{i = 0}^k \xi_i A^i r_0 - \frac{\xi_{k+1}}{\alpha_k} A^k r_0 = \sum_{i = 0}^{k-1} \xi_i A^i r_0 + (\xi_k - \frac{\xi_{k+1}}{\alpha_k}) A^k r_0 = 0$$

$\{r_0, A r_0, \dots, A^k r_0\}$ are linearly independent, therefore every coefficient in the last equation must be equal to zero.

$$\xi_k - \frac{\xi_{k+1}}{\alpha_k} = 0 \Longleftrightarrow \xi_{k+1} = \alpha_k \xi_k = (-1)^2 \alpha_k \xi_k$$

and, using our inductive hypothesis, we get

$$\xi_{k+1} = (-1)^2 \alpha_k \xi_k = (-1)^2 \alpha_k \Big((-1)^k (\alpha_0 \times \alpha_1 \times \dots \times \alpha_{k-1}) \Big) = (-1)^{k+2} (\alpha_0 \times \alpha_1 \times \dots \times \alpha_{k-1} \times \alpha_k)$$

Note: I strongly suspect that the exponent associated to $-1$ in the original question is incorrect and the general form of the coefficients is actually $\xi_{k+1} = (-1)^{k+2} (\alpha_0 \times \alpha_1 \times \dots \times \alpha_{k-1} \times \alpha_k)$.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .