[Haskell-cafe] Why not some subclass of Floating to model NaNs as some handleable bottom?

[Michal J Gajda]

The infamous `NaN /= NaN` makes only sense for `NaN` originating as a
result, since we cannot compare `NaN`s originating from different
computations.
But it breaks `Eq` instance laws as needed for property tests.
That is why comparison on `NaN` is better handled by `isFinite`,
`isANumber` predicates.
Note that beside `NaN` we also have other anomalous values, like
Underflow, Overflow, +inf and -inf.
These are all error values, and can hardly be treated in any other way.
And they all need to be handled for floating points.

Yes, comparing `NaN` with anything should give a rise to another error value.
That means that the only way out is making Either Error Float, and
then `(>=) :: Either Error Float -> Either Error Float -> Either Error
Bool`
So basically we need to lift all `Num` operations to the `Either Error` Monad.

That is probably best way to fix the problem: once error value
appears, we need to treat it consistently throughout entire
computation.
At the same time, we do not want a single error value to dominate
entire computation, so that is why we treat collections of
computations as computations that give a collection of good results
and a collection of errors separately.
If we take this into consideration, we notice that most interesting
computations occur on collections of values, and thus yield a
collection of results, not just a single output.

That is one takeaway from the referenced presentation on data
analytics in Haskell. (Similar presentation was also well received on
Data Science Europe. It should be on YouTube by now.)

Example of a 3D rotation is instructive: if NaN appears for any single
coordinate, we can get a useful results for all other coordinates, and
thus narrow impact of an error.
If the next step is projection on X-Y coordinates, then NaN or
Over/Under-flow within Z does not affect the result.

To my understanding, that is also the reason why IEEE mandated special
treatment of error values: most of the computations happen on large
matrices, vectors etc, and crashing for each single NaN would be a
true disaster.
It can be even ignored, when the NaN is computed for an energy
component within a single frame of long-running simulation, and the
error disappears within a single time step.
--
  Cheers
    Michał

On Wed, Aug 4, 2021 at 4:00 PM YueCompl <compl.yue at icloud.com> wrote:
>
> Thanks Michał,
>
> I feel less confused as I realized the non-halting possibility per bottoms, from your hint.
>
> I too think the signaling NaN is dreadful enough, so fortunately it's rarely seen nowadays.
>
> Actually what's on my mind was roughly something like "Maybe on steroids", I'm aware that NaN semantics breaks `Num` (or descendants) laws, as seen at https://gitlab.haskell.org/ghc/ghc/-/blob/master/libraries/base/GHC/Float.hs
>
> > Note that due to the presence of @NaN@, not all elements of 'Float' have an additive inverse.
>
> > Also note that due to the presence of -0, Float's 'Num' instance doesn't have an additive identity
>
> > Note that due to the presence of @NaN@, not all elements of 'Float' have an multiplicative inverse.
>
> So it should have been another family of `Num` classes, within which, various NaN related semantics can be legal, amongst which I'd think:
>
> * Silent propagation of NaN in arithmetics, like `Maybe` monad does, seems quite acceptable
> * Identity test, namely `NaN` /= `NaN` - this lacks theoretical ground or not?
> * Comparison, neither `NaN` > 1 nor `NaN` <= 1 - whether or not there's a theoretical framework for this to hold? Maybe `Boolean` type needs enhancement too to do it?
>
> No such family of `Num` classes exists to my aware by now, I just can't help wondering why.
>
> Cheers,
> Compl
>
> On 2021-08-04, at 02:38, Michał J Gajda <mjgajda at gmail.com> wrote:
>
> Dear Yue,
>
> Bottom has much weaker semantics than an exception: it means You may never get a result and thus will never handle it!
>
> Another reason is convenience: it is frequently the case that giving NaN in a row of numbers is much more informative than crashing a program with an exception and never printing the result anyway.
>
> Finally IEEE special values have clear propagation semantics: they are basically Maybe on steroids.
>
> The problem with this approach is indeed a silent handling.
>
> But in order to fix this, it is better to add preconditions to specific algorithms that do not allow IEEE special value on input (`isFinite` or `isNotNaN`) and then track the origin of the special value with the methods like those described here: https://skillsmatter.com/skillscasts/14905-agile-functional-data-pipeline-in-haskell-a-case-study-of-multicloud-api-binding
>
> Never throw an error without telling exactly why it happened and exactly where to fix it :-). Using bottom is last resort; exceptions likewise.
> --
>   Cheers
>     Michał
>
>


-- 
  Pozdrawiam
    Michał