Avoid NaN in transformed dists density and cdf

[venpopov]

This PR improves the stability of density calculations for transformed distributions (#97). Previously density values outside of the support region for transformed distributions were NaN, because of the numeric derivative procedure. The approach here is:

• Add a new argument to support(), interval, which is a vector of two values , "open" or "closed", for the lower and upper bound respectively
• support() for non-transformed functions automatically determines whether the interval is open or closed, by evaluating the density at the boundaries (0 -> open interval, 1 -> closed interval)
• for transformed distributions, the support is determined by applying the transformation function to the support of the parent
• warnings about NaN values during numeric derivatives are suppressed
• all values below or above the support (strict or non-strict, depending on the interval type) are replcaed with 0s
• similar for cdfs, but with 1 for values above upper bound

I added extensive tests to ensure this works properly.

[mitchelloharawild]

Thanks, looks great. Would be nice to use field() where appropriate.
The assumption about the support being the same will work for now, but will need appropriately vectorising later for improved performance via vectorisation (#25)

Could you please also add yourself as ctb for the package if you would like?

[venpopov]

Thanks, looks great. Would be nice to use field() where appropriate. The assumption about the support being the same will work for now, but will need appropriately vectorising later for improved performance via vectorisation (#25)

Could you please also add yourself as ctb for the package if you would like?

thanks for the tip about field(). I am completely new to the vctrs data structure so did not yet know about this :)

Will make the suggested changes now

[mitchelloharawild]

Yes, you are right. Following the structure of lim but with logical would be great, thanks!

…

On Tue, Apr 2, 2024, 1:16 AM Ven Popov ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In R/support.R <#98 (comment)> : > #' -new_support_region <- function(x, limits = NULL) { - vctrs::new_rcrd(list(x = x, lim = limits), class = "support_region") +new_support_region <- function(x, limits = NULL, interval = list(c('closed','closed'))) { we can do a logical vector, but it still needs to be logical(2L) because the ends of the interval can be different (open-open, open-closed, closed-open, closed-closed). — Reply to this email directly, view it on GitHub <#98 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AD3BJFZKF5G3UBSXR3K7CQ3Y3FT4JAVCNFSM6AAAAABFQOV3O2VHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMYTSNZRGM2TAMJSGI> . You are receiving this because your review was requested.Message ID: ***@***.***>

[venpopov]

I've implemented your requested changes, could you please take a look?

[mitchelloharawild]

Thanks! A few small suggestions.

[mitchelloharawild]

Thanks, update the docs and it looks good to merge.

[venpopov]

All done!

[mitchelloharawild]

Great, thanks!