For log_pdf, I use wolfram alpha to derive the formula

Interesting - I would just have used the standard logarithm rules to cancel exps etc.

Do you have an opinion on which representation is more stable, numerically?

Intuitively, at least, I would think cancelling as many exps as possible is better.

PS: no need to close PR and reopen, if you update your branch on your fork, the PR automatically gets updated.

PS: no need to close PR and reopen, if you update your branch on your fork, the PR automatically gets updated.

Previously, I incorrectly named the branch, it should be logistic, but I named the branch as weibull, sorry about that 😅😅

Interesting - I would just have used the standard logarithm rules to cancel exps etc.

Thank you so much on your thoughts about this! Based on my manual calculation it should be

$(-\frac{x-\mu}{s}) - ln(s) - 2ln(1+e^{-\frac{x-\mu}{s}})$

Do you have an opinion on which representation is more stable, numerically?

I'm uncertain myself about which representation is numerically more stable 😅😅

Previously I leaned towards the result from Wolfram Alpha as it offered a relatively simpler equation, but I am also not really sure whether it is the best solution

After some careful consideration, I think using the standard logarithm rule would be more intuitive to use and to be read. I will change it accordingly 😅😅

Thank you so much on your thoughts about this! Based on my manual calculation it should be

Looks correct to me. I'd use numpy logaddexp (or the scipy distribution directly)

I'd use numpy logaddexp

Sure, I have adjusted the code to use the function

could you kindly add the distribution to the api reference in docs?

Sure, I have added the new class to the API reference

not blocking, but could you kindly look at the comment re logs and cancellation?

I am not sure I understand this, what does comment re logs and cancellation mean? Thank you!

Please let me know if there is any concern or feedback, thank you so much!

Interesting - I would just have used the standard logarithm rules to cancel exps etc.

The above is what I meant, yes

I'm assuming you're no longer working on this, so I'm wrapping this up now.

Hi @fkiraly, I very much apologize for the delay in completing the PR without notifying you further. Thank you so much for wrapping this PR up. Please kindly let me know if there's anything I can assist with or if there are any adjustments needed 🙏🙏

no worries - a message would have been helpful, but it is not unusual that PRs do not get completed, and we wrap those up that are close to the finish line.

Thanks for your contibutions with the scipy adapter, these are really nice!