On Thu, 6 Jan 2011, Matthijs H.D. van der Wiel wrote:
> On Wed, Jan 5, 2011 at 7:07 PM, Peter W. Draper <[log in to unmask]> wrote:
>> On Mon, 3 Jan 2011, Matthijs H.D. van der Wiel wrote:
>>
>>> Additional question about the "deblend" script in starjava/bin/splat/
>>> : is it possible to get more decimals on the parameter errors? For
>>> example, I currently get this in the "report.log" output file:
>>>
>>> Post-fit: GaussianFitter[flux = 7.295523848504501, (+/- 0.0), scale =
>>> 1.1952008680668813, (+/- 0.0), centre = -6.87900360112492, (+/- 0.0),
>>> sigma = 2.435149604227493(+/- 0.0), FWHM = 5.734339100684084(+/- 0.0),
>>> ]
>>>
>>> All I can infer from this is that the errors on each parameter are
>>> smaller than 0.05 in the corresponding unit.
>>
>> Hi Matthijs,
>>
>> this probably indicates that your input spectrum does not have any variances
>> and in that case the attempt to guess errors using a reduced chi-square of
>> one isn't working (too good a fit, correlated data might do that).
>
> That's true, my spectrum does not have a variance array associated to it.
>
>> Otherwise the precision used to write out the errors is the same as for
>> the model values.
>
> So that means that a value of "1.1952008680668813, (+/- 0.0)" really
> indicates an error identical to zero? I guess if the reduced chi
> squares bottoms down to unity, there's not much else the fitter can
> give us... But how can the fit be that perfect if there is noise in my
> observed spectrum?
Hi Matthijs,
sorry, seems I confused what you get from the fitting of a composite model
with what you get from the single line fitting. In fact you will not
currently get any error estimates for the parameters from a composite fit
under any circumstances.
Since the change is relatively trivial I've changed SPLAT so that it now
also reports errors for the composite fits as well.
Tim, I've submitted these are they accepted for the new release?
Cheers,
Peter.
|