Re: [eigen] help: eigen cann't solve this simple ax correctly, but Armad

I was able to confirm the results of Eigen in Matlab using a simple x = A.lu().solve(b); (your matrix has a condition number of about 200, so we're good on the numerical issue front).

My best guess is that you have I/O problems. Did you confirm that the matrix you are using in Eigen is the matrix you intend to use? In particular: are you using the full precision of the file, and, most importantly, did you pay attention to the fact Eigen's matrices are column major (while reading your file sequentially is akin to having row major data) ?

Adrien

On Tue, Sep 8, 2020 at 8:32 AM ztdepyahoo <ztdepyahoo@xxxxxxxxx> wrote:

Thanks I　solve　it with
Vectorxd x = A.partialPivLu().solve(rhs);

I have a exact solution of "cos(6*Pi*x)*cos(6*Pi*y)", so i can obtain the absolute error. the order of error from Armadillo is "-14", while Eigen's error has an order of "1". I also tried other methods "fullPivLu()", "householderQr()", the situation doesn't change.

But if i compute the relative error with Eigen's " double relative_error = (A*x - b).norm() / b.norm()", the relative_error is very small. I am confused by this situations.

ztdepyahoo

ztdepyahoo@xxxxxxxxx

签名由网易邮箱大师定制

On 9/8/2020 07:18，Adrien Escande<adrien.escande@xxxxxxxxx> wrote：

Hi there,

could you be more precise: how did you try to solve your problem with Eigen? How do you know/check that the problem is not solved correctly ?

Best regards,
Adrien Escande

On Tue, Sep 8, 2020 at 12:55 AM ztdepyahoo <ztdepyahoo@xxxxxxxxx> wrote:

Dear sir:
The matrix is generated from the spectral method of a simple heat diffusion problems, the dimension is 841.
Attachment please find the a and rhs output from Armadillo.
I am wandering why eigen's dense linear sover cann't solve it correctly.
Regards

ztdepyahoo

ztdepyahoo@xxxxxxxxx

签名由网易邮箱大师定制