Pcg method is imprecise

Question

Fedor ，2023-3-7

0
链接

Direct link to this question

https://ww2.mathworks.cn/matlabcentral/answers/1924365-pcg-method-is-imprecise

评论： Fedor ，2023-3-11

I must be wrong in my code, so feel free to note where I've mistaken. I'm testing pcg method's precision through following scheme: generate random positive definitive matrix, and let it be linear system with the desired solution as (1, 1, 1, 1...) - to simplify testing. We can solve that linear system with division or simply state that solution should be (1, 1, 1, 1...). Then we solve system again with pcg method and test that all xi are nearby 1. And it doesn't work - pcg's solution does not meet needed tolerance(precision). Maybe I understand the meaning of tolerance incorrectly?

Code:

tol = 1e-8;
iterations = 150;
matrix_1_size = 50;
matrix_1 = sprand(matrix_1_size,matrix_1_size,.5);
matrix_1 = matrix_1'*matrix_1;
B_1 = sum(matrix_1, 2);
expected_solution_1 = matrix_1 \ B_1;
pcg_solution_1 = pcg(matrix_1, B_1, tol, iterations);
diff = abs(pcg_solution_1 - expected_solution_1);
tolerance = zeros(matrix_1_size, 1) + tol;
assert(all(diff <= tolerance));

0 个评论
显示 -1更早的评论隐藏 -1更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Chunru 2023-3-8

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1924365-pcg-method-is-imprecise#answer_1187985

The pcg method use tol as a relative tolerance such that norm(residual) < tol*norm(rhs). So you need a different assert.

tol = 1e-8;
iterations = 150;
matrix_1_size = 50;
matrix_1 = sprand(matrix_1_size,matrix_1_size,.5);
matrix_1 = matrix_1'*matrix_1;
B_1 = sum(matrix_1, 2);
expected_solution_1 = matrix_1 \ B_1;
pcg_solution_1 = pcg(matrix_1, B_1, tol, iterations);
pcg converged at iteration 74 to a solution with relative residual 9.8e-09.
diff = abs(pcg_solution_1 - expected_solution_1);
norm(diff)
ans = 1.3296e-06
norm(pcg_solution_1-expected_solution_1)
ans = 1.3296e-06
norm(B_1)
ans = 772.9681
norm(pcg_solution_1-expected_solution_1)/norm(B_1) < tol
ans = logical
   1
assert(norm(pcg_solution_1-expected_solution_1)/norm(B_1) < tol)