potential floating point error propagation in set shower scale

[Sung Hak Lim]

Dear MG5 team,

MG5 version: 2.6.0

I just noticed that NLO shower scale setter has a potential floating point error propagation.

In function emscainv in montecarlocounter.f, they return a rescaled random number in [0,1] by

(r - Sqrt[(r - r^2)])/(2 r - 1)

This expression is 0/0 form at r=0.5, therefore precision loss happen when we pick up r value near 0.5.

If this part give us crazy number, it seems it will be cured internally in other subroutine, such as taking scalemax instead in set_shower_scale in fks_singular.f. However, this will break smooth random number choice so I think it may generate unwanted flaws.

Isn't it better to use floting point arithmetic safe expression, like

Sqrt[r]/(Sqrt[r] + Sqrt[1 - r])

Best wishes,
Sung Hak

[Paolo Torrielli (paolo-torrielli)]

Dear Sung Hak,

it seems you opened a ticket, even it was already answered.
As I wrote, I think what you say is are fine. Just notice that
this did not produce instabilities because we explicitly set
emscainv =1/2 if r=1/2.
Anyway, your proposal is perfectly fine.
Cheers.
Paolo

> On 3 Nov 2017, at 09:15, Launchpad Bug Tracker <email address hidden> wrote:
>
> Olivier Mattelaer (olivier-mattelaer) has assigned this bug to you for MadGraph5_aMC@NLO:
>
> Dear MG5 team,
>
> MG5 version: 2.6.0
>
> I just noticed that NLO shower scale setter has a potential floating
> point error propagation.
>
> In function emscainv in montecarlocounter.f, they return a rescaled
> random number in [0,1] by
>
> (r - Sqrt[(r - r^2)])/(2 r - 1)
>
> This expression is 0/0 form at r=0.5, therefore precision loss happen
> when we pick up r value near 0.5.
>
> If this part give us crazy number, it seems it will be cured internally
> in other subroutine, such as taking scalemax instead in set_shower_scale
> in fks_singular.f. However, this will break smooth random number choice
> so I think it may generate unwanted flaws.
>
> Isn't it better to use floting point arithmetic safe expression, like
>
> Sqrt[r]/(Sqrt[r] + Sqrt[1 - r])
>
> Best wishes,
> Sung Hak
>
> ** Affects: mg5amcnlo
> Importance: Undecided
> Assignee: Paolo Torrielli (paolotorriell)
> Status: New
>
> --
> potential floating point error propagation in set shower scale
> https://bugs.launchpad.net/bugs/1729726
> You received this bug notification because you are a bug assignee.

[Sung Hak Lim (sunghak-lim)]

Dear Paolo,

Sorry, I didn't get your reply when I was writing this ticket.

I think setting emscainv=1/2 at r=1/2 doesn't help to solve this precision loss but it seems that impact is minimal so it seems fine.

Thanks a lot for answers!

Best wishes,
Sung Hak

[Olivier Mattelaer (olivier-mattelaer)]

Hi Paolo,

Can you close this bug report with the correct status?

Thanks,

Olivier

[Paolo Torrielli (paolotorriell)]

Hi all,
I finally implemented the suggestion, it will be available from 2.6.2 on.
Thanks again, I'll now close the thread.
Cheers.
Paolo