Skip to content

Latest commit

 

History

History
85 lines (47 loc) · 7.11 KB

GMRES.md

File metadata and controls

85 lines (47 loc) · 7.11 KB

Previously, we introduce several iterative algorims to solve the linear equation Ax = b, where A is n x n symmetric matrix. What if A is non-symmetric matrix? To address the issue, we apply the generalized minimal residual method (GMRES), which will be introduced in the following.

The basic idea of GMRES is to construct the approximate solution x_k, where image. Note that image is the k-dimension Krylov subspace, image is the initial guess, and image is the initial residual.

Change Target Function

Recall that our goal is to find the solution that minimizes the residual, shown as follow:

image

Let image, where image is k x k unitary matrix and can be computed by using Arnoldi iteration.

Therefore, we can rewrite the function to be minimized as:

image,

where (a) equation holds since

image,

(b) equation holds according to the Arnoldi equation image,

(c) equation holds since image is the first column of (k+1) x (k+1) identity matrix,

and (d) equation holds since image is orthonormal.

Algorithm

In summary, the algorithm is shown as follow

At k-th iteration:

  1. Compute image using Arnoldi iteration;
  2. Find image that minimizes image;
  3. Form the solution image;

Solve Least Square Problem

In the following, let us take a close look at the detailed implemention of how to find image that minimizes image.

To solve the least square problem

image,

we adopt QR decomposition, shwon as follow

image,

where image is (k+1) x (k+1) orthogonal matrix and image is (k+1) x k upper triangular matrix.

The difficulty is that we expect to update the decomposition of image cheaply at each step of Arnoldi iteration. That is,

image,

where image. This implies that we add a new column and a new row at each step.

To proceed, we apply Given rotations.

Let image represent the rotation matrix, shown as

image.

At next step, we premultiply image to the new column image and get the rotationed column image. Append the rotationed column and image to previous image, we have a almost triangular matrix

image.

In order to get a triangular matrix image, premultiply by image: image

where image,

image,

and

image.

Therefore, let us look at the QR decomposition image again. We find that image is the acumulated product of the rotation matrices and is unitary.

Finally, we can rewrite the target function as:

image,

where image.

The image that minimizes the target function is image.