sel.principleAxes() is not correct about mdanalysis

[GoogleCodeExporter]

Indeed, the values are exactly transposed (up to spurious numerical differences 
and the sign of one eigenvector (which is irrelevant)):

  >>> pa = numpy.array([[-0.67825811, -0.20538766, -0.70553656] ,[-0.51792553, -0.54748404,  0.657278  ], [-0.5212668,   0.81121954,  0.26495998]])
  >>> sv = numpy.array([[ 0.67825812  0.5179255   0.52126682] [-0.20538768 -0.54748404  0.81121951] [-0.70553654  0.657278    0.26495996]])
  >>> sv[0] *= - 1    # flip sign of first ev 
  >>> pa - sv.T       # difference
  array([[  9.99999994e-09,  -2.99999999e-08,   2.00000000e-08],
         [  2.00000000e-08,   0.00000000e+00,   2.99999999e-08],
         [ -2.00000000e-08,   0.00000000e+00,   2.00000000e-08]])  

The first eigenvector from principleAxes is pa[:,0]; this implementation 
followed the numpy philosophy of having the coordinates in individual columns 
so that one can pass them separately around.

Is this too confusing? 

Right now I am changing the docs to make this clear at least.

Many thanks for checking!

Original comment by orbeckst on 29 Jun 2010 at 9:55

• Changed state: Started

[GoogleCodeExporter]

I agree that this is confusing.

In r335 I added principalAxes() (correct spelling!) which does what you would 
expect:

  e1,e2,e3 = selection.principalAxes()

The old function principleAxes() with the old behaviour is kept but deprecated.

Original comment by orbeckst on 29 Jun 2010 at 10:20

• Changed state: Fixed

[GoogleCodeExporter]

I've compared the behaviour of the new principalAxes() and my results using svd 
using the python code in my original post and they are identical (to within a 
factor of -1)

Original comment by [email protected] on 29 Jun 2010 at 10:37

[GoogleCodeExporter]

That's good – many thanks for checking.

Original comment by orbeckst on 29 Jun 2010 at 10:54

• Changed state: Verified