no error in LO cross section when using bias

HI,

When you use bias, we do not estimate the "real" cross-section anymore but the integral of another function (the bias function).
We then use "re-weighting" to get the estimator of the original cross-section (that we print on screen).

We also have the statistical uncertainty for the function that we did integrate (that include the bias), but
we do not have an efficient/reliable way to estimate the statistical uncertainty associated to the original integral.
This is the reason why we set it to zero. What is clear is that the bias degrades the precision of the total cross-section so I would say that the shift of cross-section that you observe is not surprising.

Cheers,

Olivier

> On 20 Jul 2023, at 04:10, Benjamin Grinstein <email address hidden> wrote:
>
> New question #707335 on MadGraph5_aMC@NLO:
> https://answers.launchpad.net/mg5amcnlo/+question/707335
>
> I get slightly different cross sections when running with bias vs without. The process is p p > ta- vt~ (I don't think the specific process matters) and I get
> with bias: 3578 ± 0
> without bias: 3616 ± 6.9
> While the difference is small, it is much larger than the errors. I am particularly concerned about the zero uncertainty in the run with bias. For completeness the bias is with
> /ptl_bias = bias_module ! Bias type of bias, [None, ptj_bias, -custom_folder-]
> {'ptl_bias_target_ptl': 3000.0, 'ptl_bias_enhancement_power': 4.0} = bias_parameters
> where the ptl_bias module is just the ptj_bias module with is_a_l for is_a_j
>
> Incidentally the bias works: the histograms of pT of ta- (and that of MET) are very smooth at large pT (and MET) for the biased case, not so for the unbiased one (with a 10000 run).
>
> The question is then: what happened to the error in the biased run? (I understand that the difference is less than the correction that would come from NLO, but still....)
> Thanks
> Ben
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.