[daip] AIPS question: IMFIT vs JMFIT

[lisakov]

Dear Eric,

Thank you for your reply.
I was interested in principal feasibility to use circular Gaussians.
I don't think that playing with the data really worth your time if
circular Gaussians are not implemented.

Best regards,
Mikhail Lisakov.




On 09/16/2015 05:28 PM, Eric Greisen wrote:
> On 09/16/2015 03:18 AM, lisakov wrote:
>> Dear Sir,
>>
>> I have faced a need to perform an image plane fitting of a structure of
>> an AGN with several Gaussian components. Particularly it is a
>> multi-epoch stacked image of the 3C273 at 7 mm.
>>
>> I decided to use AIPS and found 2 tasks which satisfy my needs: IMFIT
>> and JMFIT.
>> I found then that these tasks perform differently. With the same
>> starting model (which is quite a good starting point, I believe) JMFIT
>> works well but IMFIT complains on strange value of some parameters (I
>> think, not the initial one, but the one which occurs during the evaluation)
>> IMFIT1: STRANGE VALUE FOR COMP= 1 PARAM= 4 VALUE= 5.111E+02
>> IMFIT1: RIDICULOUS VALUE FOR SOME PARAMETER AT NITER= 40
>> IMFIT1: CHECK INPUT MODEL. MODEL VALUE IS CRAZY
>>
>> and then "Purports to die of UNNATURAL causes".
>>
>>
>> Regarding this different behaviour I wonder, are there any caveats of
>> using JMFIT instead of the IMFIT?
>>
>> And yet another question. I am not totally satisfied with JMFIT because
>> I did not manage to make JMFIT use only pure circular Gaussians. I've
>> tried to let only major axes to vary, but this performs as it should -
>> it results in elliptical Gaussians. So my question: is there a
>> possibility to tell JMFIT to use pure circular Gaussian components?
>>
>>
>> AIPS version is 31DEC14.
>> Attached to the e-mail are
>> * image to fit
>> * commands to set initial parameters and run both IMFIT and JMFIT
> 
> I may play with your data later - but I know that the mathematician that 
> wrote both tasks did JMFIT second because he preferred that method.  So 
> there is not problem with taking the JMFIT results.  The two give very 
> similar results on relatively simple sources but with a more complex 
> situation I am not surprised that one fails to converge.
> I once followed the fitting in SAD (same math as JMFIT) and was amazed 
> at how strange some of the models were during the fit even when it 
> finally converged on something sensible.
> 
> Circular Gaussians would require significant changes in the code of the 
> tasks (doable but not with very much demand).
> 
> Eric Greisen
> 
> .
>