Boost C++ Libraries: Ticket #9672: PDF and CDF of a laplace distribution throwing domain_error

owner changed

John Maddock — Mon, 17 Feb 2014 09:56:32 GMT

owner changed from John Maddock to Paul A. Bristow

Originally we did not allow non-finite arguments anywhere, however, in the last release that should have been - mostly - fixed to allow infinite random variables when that made sense. Paul, I think you made those changes, were these distributions included?

Paul A. Bristow — Mon, 17 Feb 2014 14:44:20 GMT

I have reordered the checks in laplace.hpp to now allow + and - infinity.

Please can I email you a revised version for you to check that it meets your needs? Please email me at pbristow@…

anonymous — Mon, 17 Feb 2014 16:34:06 GMT

Paul, you can attach files direct to this ticket.

attachment set

Paul A. Bristow — Mon, 17 Feb 2014 16:47:58 GMT

attachment → laplace.hpp

Tue, 18 Feb 2014 06:19:29 GMT

Replying to johnmaddock:

Originally we did not allow non-finite arguments anywhere, however, in the last release that should have been - mostly - fixed to allow infinite random variables when that made sense. Paul, I think you made those changes, were these distributions included?

Dear John,

Thank you for your comment.

I actually would expect every distribution function to be defined on the whole real line. And since each distribution function is often for us a "terminal function"---eg. determining P-values or critical values of statistical tests, I would also expect its implementation to support infinities (due to overflow when calculating a test statistic) and NaN (from 0/0 say). However, I do expect the distribution parameters to take on valid finite values only---conceptually, we cannot use a "normal distribution with infinite mean"; here, I would expect an implementation to throw the domain_error like your team is doing, or---my preference---to return NaN.

But that is just my opinion as a statistician.

With best regards,
HS

Tue, 18 Feb 2014 06:26:31 GMT

Replying to pbristow:

I have reordered the checks in laplace.hpp to now allow + and - infinity.

Please can I email you a revised version for you to check that it meets your needs? Please email me at pbristow@…

Dear Paul,

Thank you very much for the fixes; it has cleared my test deck.

With best regards,
HS

Paul A. Bristow — Tue, 18 Feb 2014 10:24:37 GMT

Good :-)

Useful to have a professional statistician's view on when infinity should always be supported.

In a sense, we already provide a way of achieving this view policies. If you implement a policy to ignore domain and or overflow errors, then you will get a NaN. You can achieve this now but studying the examples of using policies.

When Boost.Math was first developed, the implementation of infinity and NaN was poorly and/or inconsistently supported, so our instinct was to duck the issues.

But then John devised the 'policies' control of what happens when overflow or domain error is encountered. (This allows us to mimic what happens when you have C view of error handling and error return codes).

Also other parts of Boost.Math have been extended to make it appear that infinity and NaN are fully supported. So we have started to permit infinity in some cases, but not all because of the work involved to changes things (and the many tests) in all the many functions. It has become a bit labyrinthine as a consequence. I'll put a fuller review on my TODO list.

status, milestone changed; keywords, resolution set

Paul A. Bristow — Tue, 18 Feb 2014 17:33:47 GMT

keywords laplace added
status new → closed
resolution → fixed
milestone To Be Determined → Boost 1.56.0

Fixed and pushed to develop.

SHA-1: bf40c602967e3f87a375b3dc9b684245e270c2c1

Random variate can be infinite.

Thu, 20 Feb 2014 07:53:37 GMT

Replying to pbristow:

Good :-)

Useful to have a professional statistician's view on when infinity should always be supported.

In a sense, we already provide a way of achieving this view policies. If you implement a policy to ignore domain and or overflow errors, then you will get a NaN. You can achieve this now but studying the examples of using policies.

When Boost.Math was first developed, the implementation of infinity and NaN was poorly and/or inconsistently supported, so our instinct was to duck the issues.

But then John devised the 'policies' control of what happens when overflow or domain error is encountered. (This allows us to mimic what happens when you have C view of error handling and error return codes).

Also other parts of Boost.Math have been extended to make it appear that infinity and NaN are fully supported. So we have started to permit infinity in some cases, but not all because of the work involved to changes things (and the many tests) in all the many functions. It has become a bit labyrinthine as a consequence. I'll put a fuller review on my TODO list.

Dear Paul,

Sorry for the late response---trying to get use to modular-boost and "git submodule".

Thanks for the reminder about policies---it is a beautiful pattern and I feel it is very well implemented in boost::math.

Yes, I do use my own policies. While exceptions and/or error numbers are great tools in programming, I (as a statistician) do not find them useful here for me with distributions: for example, if any of the distribution parameters are invalid, then any caller (programmer or actual user) should get a "non-answer" (implemented by silent NaN which is wonderful since it "propagates")---no need to "leak" an exception; and each "proper" distribution (meaning all its parameters are valid) should also returns NaN when the input value is NaN.

Thanks for including +-infinity as possible input values ... they do represent the sets of "large" (and "small") numbers ... and we do want to return 0 as the P-value when a test statistic is "infinite" (or "large" in mathematical sense).

Let me know if you would like an extra pair of eyes to look over your proposed changes: I am not very good in reading code but I think I am a better reader of documents. :)

With best regards,
HS

anonymous — Tue, 05 Dec 2017 18:40:08 GMT

Replying to Paul A. Bristow:

I have reordered the checks in laplace.hpp to now allow + and - infinity.

Please can I email you a revised version for you to check that it meets your needs? Please email me at pbristow@…

merciu — Tue, 05 Dec 2017 18:41:47 GMT

anonymous — Tue, 05 Dec 2017 19:06:02 GMT

anonymous — Wed, 09 May 2018 03:20:43 GMT

anonymous — Wed, 09 May 2018 03:22:03 GMT