[NumPy-Tickets] [NumPy] #1885: einsum wrong result / crash

[NumPy Trac]

#1885: einsum wrong result / crash
--------------------------+-------------------------------------------------
 Reporter:  royf          |       Owner:  somebody
     Type:  defect        |      Status:  new     
 Priority:  normal        |   Milestone:  1.6.1   
Component:  numpy.core    |     Version:  1.6.0   
 Keywords:  einsum,wrong  |  
--------------------------+-------------------------------------------------
 numpy.einsum sometimes returns wrong results.
 Consider the following example:
 >>> p = numpy.array([.5, .5])
 >>> q = numpy.array([[0.26894142, 1-0.26894142], [1-0.26894142,
 0.26894142]])
 >>> r = numpy.identity(2)
 ...
 >>> p, q, r
 (array([ 0.5,  0.5]), array([[ 0.26894142,  0.73105858],
        [ 0.73105858,  0.26894142]]), array([[ 1.,  0.],
        [ 0.,  1.]]))
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 3.7305612660634556e+177
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 0.26894141999999999
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 3.7305612660634556e+177
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 0.26894141999999999
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 3.7305612660634556e+177
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 0.26894141999999999
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 0.26894141999999999
 >>> numpy.einsum('z,mz,zm->', p, q, r)
 0.13447070999999999

 In other runs this same example crashes the process.

 The correct answer is 0.26894142.
 Interestingly, one of the answers returned is half the correct one.
 This probably has to do with the specific value of p in the example.
 My guess is it's something like an off-by-one bug.

 I'd try to fix it myself,
 but the code is a scary mess :(

 Good luck!

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/1885>
NumPy <http://projects.scipy.org/numpy>
My example project