[Ifeffit] running ifeffit under 64-bit windows7

[Matt Newville]

Hi Sameh,

On Sat, Mar 10, 2012 at 8:17 AM, Sameh Ibrahim Ahmed
<sameh2977 at gmail.com> wrote:
>
> Hi,
>
> I have used ARTEMIS to fit the EXAFS of a simple Cu foil with the two
> diffrent machines, a 32 bit and 64 bit ones, both running widows7, 32 ans 64
> bit respectively.
> The results obtained are slightly different, I have appended the message
> with these results. the very low value of R-factor produced with the 64bit
> system is difficult to interpret. the questions are;
> 1- how to account for these differences?
> 2- if I to publish something, which measure of the quality should I present?
> and how can I interpret it?
>
> regards
> Sameh
>
> ============================================================
> 32bit windows 7
> Independent points          =      12.172851562
> Number of variables         =       6.000000000
> Chi-square                  =      53.001967099
> Reduced Chi-square          =       8.586301900
> R-factor                    =       0.000163791 !!!
> Measurement uncertainty (k) =       0.000110566
> Measurement uncertainty (R) =       0.057163282
> Number of data sets         =       1.000000000
>
> Guess parameters +/- uncertainties  (initial guess):
>   amp             =     0.9907500   +/-      0.0312520    (1.0000)
>   enot            =     5.6814210   +/-      0.3815760    (0.0000)
>   delr1           =     0.0016460   +/-      0.0033990    (0.0000)
>   ss1             =     0.0099820   +/-      0.0004090    (0.0030)
>   w1 3rd cumulant =     0.0001710   +/-      0.0000340    (0.0000)
>   p1 4th cumulant =     0.0000240   +/-      0.0000070    (0.0000)
>
> ============================================================
> 64bit windows 7
> Independent points          =      12.172851562
> Number of variables         =       6.000000000
> Chi-square                  =      54.036163164
> Reduced Chi-square          =       8.753841335
> R-factor                    =   0.309664502E-06 !!!
> Measurement uncertainty (k) =       0.000097377
> Measurement uncertainty (R) =       0.050344538
> Number of data sets         =       1.000000000
>
> Guess parameters +/- uncertainties  (initial guess):
>   amp             =     1.0029100   +/-      0.0296670    (1.0000)
>   enot            =     5.7522760   +/-      0.4614510    (0.0000)
>   delr1           =     0.0024980   +/-      0.0040190    (0.0000)
>   ss1             =     0.0101430   +/-      0.0003790    (0.0030)
>   w1 3rd cumulant =     0.0001780   +/-      0.0000410    (0.0000)
>   p1 4th cumulant =     0.0000260   +/-      0.0000070    (0.0000)
> ============================================================
>

Except for R-factor, these differences are pretty small -- all
parameters are well within the estimated error bars.  I'm not sure why
R-factor is different.  I would say that there is essentially no
difference in what to report.... the R factors are both small enough
to mean "very good fit", and any difference between them would  really
only important when comparing two different fits -- in that case, just
be consistent.

But that's not to say that it's not worth trying to understand the
difference.... but that might take a bit of investigative work.
One thing I noticed in the projects you sent (only to me -- please use
the mailing list!!)  is that these fits use different versions of
Athena and ifeffit:

  32bit Win7: Artemis 0.8.012, ifeffit 1.2.11
  64bin Win7: Artemis 0.8.014, ifeffit 1.2.11c

Off hand, I don't know that either of these is actually significant.

--Matt