Re: Failed test in linsolve.m

[Julien Bect]

Le 02/04/2015 10:18, Marco Atzeri a écrit :
> > ***** test
> >   n = 4;
> >   A = triu (rand (n));
> >   x = rand (n, 1);
> >   b = A' * x;
> >   opts.UT = true;
> >   opts.TRANSA = true;
> >   assert (linsolve (A, b, opts), A' \ b);
> > !!!!! test failed
> > ASSERT errors for:  assert (linsolve (A, b, opts),A' \ b)
> >
> >    Location  |  Observed  |  Expected  |  Reason
> >      (1)        0.42063      0.42063      Abs err 5.5511e-17 exceeds 
> > tol 0
> >      (2)        0.68722      0.68722      Abs err 4.1078e-15 exceeds 
> > tol 0
> >      (3)        0.25533      0.25533      Abs err 2.1094e-15 exceeds 
> > tol 0
> >      (4)        0.52907      0.52907      Abs err 7.3275e-15 exceeds 
> > tol 0
> >
> > Has anybody seen this ? Perhaps should we introduce a tolerance for
> > these tests ?
>
>
> a zero tolerance is seldom a good idea.
> It should be several eps as minimum. 

I agree, but in this case it would require 34 * eps to pass this test... 
Is it acceptable ?

Actually, this test uses random data and therefore fails randomly if i 
hit "test linsolve" repeatedly in Octave.

Perhaps would it be more reasonable to specify a 4x4 matrix explicitely 
*and* set some tolerance (say, 50 * eps) for the comparison ?