Wrong estimate for transit duration if long gaps are present

[hippke]

e.g., when combining two TESS sectors with a long gap in between: Transit duration estimate is wrong.

[hippke]

Suggested solution by @MNGuenther

ind_tr_phase = np.where( results['model_folded_model'] < 1. )[0]
lc_duration = results['period'] * (results['model_folded_phase'][ind_tr_phase[-1]] - results['model_folded_phase'][ind_tr_phase[0]])

[martindevora]

Hi. Is this planned to be solved any time soon? I'm finding the same issue. The transit duration value is different than the duration calculated from the model_folded_phase, which is the one looking correct.

Kind regards.

[mkunimoto]

While the above suggested fix correctly computes the transit duration from the best-fit transit model, I find that the best-fit transit model itself doesn't seem to be a good match when long gaps are present. Likewise, the estimated S/N (an important value for planet detection criteria) is much lower than one would expect.

As an example, here are the results for TIC-352682207 (TOI-4010), using a multi-sector TESS light curve (sectors 24, 25, 52, and 58). Using just sectors 24 and 25, TLS gives a good fit to the data and finds a duration of 2.1 hours (a little shorter than the expected 2.8 hours, but somewhat expected given this is 30-minute data) and S/N = 18.4:

And TLS shows a good fit when searching only sector 58, recovering a 2.8-hour duration (S/N = 12.9):

But when searching all four available sectors, the model is visibly worse. The model has only a 1.6-hour duration and the S/N is now 9.8:

[martindevora]

Hello @mkunimoto The problem comes when tls computes the intransit and out of transit points due to an estimated transit_duration_in_days using this method. When there are gaps in the curve, the inner calculate_stretch method returns wrong values that lead to smaller transit durations and hence, underestimated SNRs. In my case, I created a fork which I'm also publishing in pypi named foldedleastsquares where I re-compute the transit duration as done here. That way the SNR scales accordingly to the expected additional provided data. The API is the same as the one offered by tls with some small additions.

If my workaround was not found to be wrong, I'd be glad to request a merge of my fork to the official repository but I tried it some time ago and it hung on a pending pull request for several months; therefore I retired it.

Regards.

[hippke]

Hey both, sorry for the bug and sorry for not merging the pull-request. Not sure what went wrong back then. Happy to merge - how do we do this? Do you want to resubmit?

[martindevora]

My fork is also including a medium-sized code refactor to allow different transit templates to be applied and it grew in changes with time. Thus, I think that I could try to create a pull request from a simpler fork with only the changes mentioned in this issue or you could take a look at the changes I pointed out in my previous comment and cherry-pick them into the official project if you find them correct.

[hippke]

If possible I'd prefer just a bugfix for the issue mentioned here. I'd merge that and release a new version. So there is minimum risk for anything else to break.
The new features should be a subsequent change, and I'd want to do some tests on my own for them.

[mkunimoto]

Thanks @martindevora and @hippke - will the bugfix just fix the duration estimate? Or will it also address the transit model not fitting well for multi-year observations?