RE: Treatment of uncertainties from the TFractionFitter

Hi Ashish,

Multiply the templates by 1E6 (hopefully) rendering the "MC statistics" uncertainty negligible.

BTW: Are you really sure you want to use TFractionFitter? Wasn't there a paper arguing that even for the specific cases it's recommended for the mathematical basis isn't as good as it seemed to be when the technique was invented?

 cheers,
 Amnon

-----Original Message-----
From: owner-roottalk_at_root.cern.ch on behalf of Ashish Kumar Sent: Fri 02-Apr-10 12:45 PM
To: Tim Head
Cc: roottalk_at_root.cern.ch
Subject: Re: [ROOT] Treatment of uncertainties from the TFractionFitter  

Hello Tim,

   Thanks a lot for giving a thought to this problem and your suggestion! We are working in that line to estimate the uncertainties.

Ashish

Tim Head wrote:
> Hello Ashish,
>
> On 1 April 2010 11:12, Ashish Kumar <ashishk_at_fnal.gov> wrote:
>
>> Hello,
>> I am using TFractionFitter to fit the data distribution with the
>> templates from Monte Carlo simulation to extract their respective fractions.
>> The Fitter gives us the fractions with the errors. The problem is how to
>> disentangle the error into statistics and systematics parts. The error must
>> have some part as statistical, but, clearly it has systematics too since it
>> varies the shape of the templates while fitting. Can you please help me with
>> this?
>>
>
> As I understand it TFractionFitter will change the central value of
> each bin if that change will lead to a better fit. By how much it
> moves it depends on the statistical uncertainty on that bin of the
> template. This way our lack of knowledge about the central value of a
> bin is taken into account. If all templates were made with infinite
> statistics then TFractionFitter would not move the central value of
> each bin, I believe.
>
> IMHO the only "systematic" uncertainty which is then already part of
> the uncertainty on the fraction returned by TFractionFitter is the one
> often referred to as "MC statistics".
>
> It should be easy to check this with a little toy MC. Generate two
> pairs of templates, once using moderate statistics and once with
> "huge" statistics. Then generate lots of pseudo datasets with a known
> fraction of template A and template B. Now if you fit your pseudo data
> with the pair of template with moderate statistics TFractionFitter
> will move around the central value of some bins. If you use the
> templates with "huge" statistics it should stop doing it(or at least
> at a much smaller rate).
>
> Maybe an expert on this topic could comment?
>
> Let me know what you find out,
> Tim
>
> ps. Take these ideas with a pinch of salt, I am not an expert!
>
>
Received on Sun Apr 04 2010 - 22:47:37 CEST