
Hi
I'm comparing my MathNet eigendecomposition with the one from Matlab. The eigen values are identical but the eigenvectors do not correlate..and it doesn't look like a scaling issue to me.
CC = new DenseMatrix(3, 3);
CC[0, 0] = new Complex32( 2.8123f, 0.3302f);
CC[0, 1] = new Complex32( 4.9281f, 0.1065f);
CC[0, 2] = new Complex32( 2.9926f, 0.1885f);
CC[1, 0] = new Complex32(0.7638f, 0.4657f);
CC[1, 1] = new Complex32( 1.2616f, 0.4053f);
CC[1, 2] = new Complex32(1.3960f, 0.2950f);
CC[2, 0] = new Complex32( 3.3723f, 0.1272f);
CC[2, 1] = new Complex32( 6.3473f, 0.1351f);
CC[2, 2] = new Complex32( 5.7357f, 0.3368f);
eigVR = new UserEvd(CC);
Yields:
[MathNet.Numerics.LinearAlgebra.Complex32.DenseMatrix] {
(0.2683224, 0.3892781), (0.4026686, 0.7535514), (0.8103176, 0.3028587),
(0.3325276, 0.04118589), (0.05912521, 0.0987411), (1.135981, 0.2766953),
(0.5707567, 0.5817487), (0.3471378, 0.3766167), (1.872847, 0.1478064)}
MatrixD {
(5.573158, 2.325578),(0, 0),(0, 0)
(0, 0),(0.7181261, 0.09491814),(0, 0)
(0, 0),(0, 0),(3.518316, 1.348196)}
Whereas Matlab produces:
[V,D]=eig(C) % using the same matrix values as above
>> [V,D]=eig(C) V = 0.4658 + 0.0811i 0.8520 0.3500  0.1002i
0.2623 + 0.2085i 0.0591  0.0984i 0.4675  0.1537i
0.8150 0.4944 + 0.1283i 0.7907
D = 5.5731 + 2.3257i 0 0
0 0.7181 + 0.0949i 0
0 0 3.5183  1.3483i
Did I do something wrong?
Thx



>Did I do something wrong?
No, there is a bug in our Complex/Complex EVD classes.
>eigVR = new UserEvd(CC);
In general, you should use:
eigVr = CC.Evd();
or
eigVr = new DenseEvd(CC);
instead when working with dense matrices. CC.Evd() will use the DenseEvd class which might be a little quicker (at least for larger matrices).



OK, thanks!
How do I get to know when this has been fixed?



If you follow this issue http://mathnetnumerics.codeplex.com/discussions/263117, you'll get notified when it has been fixed.


Jul 20, 2011 at 7:36 PM
Edited Jul 20, 2011 at 7:37 PM

Hi Cuda,
Maybe your eigen solver does compute the correct eigen vector. I've realized that eigen vectors are not unique up to some constant k.
I tried to divide Matlabs principal eigen vector with Math.Net's principal eigen vector:
>> k = EigV_Matlab./EigV_MathNet
k =
0.70036 + 0.71383i
0.7004 + 0.71377i
0.70035 + 0.71383i
Cheers



Hi guys,
I have stumbled into similar problem described in this discussion about eigenvectors mismatch between mathnet and matlab.
Has there been any update on this?
regards.

