Adding skew_double_exponential family

[trinhdhk]

Adding experimental skew_double_exponential family, targeting #1836
Speed gains about 25%.
Caveat: currently I have noticed that sigma in Stan's implementation is twice larger than its counterpart in asym_laplace. This is a breaking change in terms of interpretation (and prior specs), so I'm still keeping asym_laplace.
I added one test but I think we need to add further test to ensure equivalence.
Please check if I added the test in the correct directory (currently in local test). And kindly let me know.
And sorry for CLRF (if there's still left-over).

[paul-buerkner]

I am fine with replacing asym_laplace Stan code with the native implementation in Stan. I would prefer not to add a new family for this. That causes too much confusion.

Do you know why one has sigma while the other has sigma / 2? Which one is the more "natural" parameterization?

[trinhdhk]

Good question. I believe it has to do with the fact that they have to guarantee its equivalence to double exponential distribution at tau = 0.5. See stan-dev/stan#2312 Trinh Dong

…

________________________________ From: Paul-Christian Bürkner ***@***.***> Sent: Friday, January 2, 2026 9:27:56 AM To: paul-buerkner/brms ***@***.***> Cc: Trinh Dong ***@***.***>; Author ***@***.***> Subject: Re: [paul-buerkner/brms] Adding skew_double_exponential family (PR #1845) [https://avatars.githubusercontent.com/u/12938496?s=20&v=4]paul-buerkner left a comment (paul-buerkner/brms#1845)<#1845 (comment)> I am fine with replacing asym_laplace Stan code with the native implementation in Stan. I would prefer not to add a new family for this. That causes too much confusion. Do you know why one has sigma while the other has sigma / 2? Which one is the more "natural" parameterization? — Reply to this email directly, view it on GitHub<#1845 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ALFCJX4NHI6W5LTWMBJOW6T4EY2ZZAVCNFSM6AAAAACPS2DGBWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTOMBUHA2DMNRSGQ>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[paul-buerkner]

Let's keep the current brms asym_laplace parameterization then but use Stan's built-in functions for fitting. Would you mind editing your PR accordingly? I am sorry that this means that we won't need most of the code you currently included in the PR.

[trinhdhk]

Let's keep the current brms asym_laplace parameterization then but use Stan's built-in functions for fitting. Would you mind editing your PR accordingly? I am sorry that this means that we won't need most of the code you currently included in the PR.

I'm happy to but it does mean a breaking change for past studies that have used asym laplace. For a mature software, this is not a good practice.

Most of my code now is just a wrapper for asym_laplace so an inplace replacement would reduce the time for testing equivalence etc. But like I said, I was thinking of the researchers (me).

[paul-buerkner]

What would it break exactly?

[trinhdhk]

What would it break exactly?

Ah I misread. Did you mean doubling sigma before feeding into Stan so that we keep the current formulation?

[paul-buerkner]

Correct.

[trinhdhk]

Clear!

…

________________________________ From: Paul-Christian Bürkner ***@***.***> Sent: Friday, January 2, 2026 10:07:56 AM To: paul-buerkner/brms ***@***.***> Cc: Trinh Dong ***@***.***>; Author ***@***.***> Subject: Re: [paul-buerkner/brms] Adding skew_double_exponential family (PR #1845) [https://avatars.githubusercontent.com/u/12938496?s=20&v=4]paul-buerkner left a comment (paul-buerkner/brms#1845)<#1845 (comment)> Correct. — Reply to this email directly, view it on GitHub<#1845 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ALFCJX3GBDWM2VNTYJTMCRL4EY7PZAVCNFSM6AAAAACPS2DGBWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTOMBUHE3DGMRQGM>. You are receiving this because you authored the thread.Message ID: ***@***.***>