It seam that C is wrong on this one and Go is right. The rationale is that a NaN must propagate through operations so that we can detect problems (avoid silent NaNs). See https://en.wikipedia.org/wiki/NaN
Thus any operation involving a NaN must return an NaN and this includes Max and Min. Le jeudi 22 avril 2021 à 12:43:23 UTC+2, jesper.lou...@gmail.com a écrit : > On Thu, Apr 22, 2021 at 11:54 AM 'Dan Kortschak' via golang-nuts < > golan...@googlegroups.com> wrote: > >> >> Does anyone know the reason for this? (Looking through other languages, >> the situation in general seems to be a mess). >> >> > I looked at OCaml, which provides both variants: > > Float.max : float -> float -> float > Float.max_num : float -> float -> float > > The `max` function includes NaNs in the computation, while `max_num` omits > them. Another important treatment the documentation mentions is `max > -0.0 +0.0`. > > I also looked at Matlab and R. They essentially provide ways to either > include the NaN or omit it from the computation, usually when you have a > vector of numbers where some of the numbers are unknown. I.e., they take > extra (optional) flags to their functions which tells you how to treat NaN > values. > > The "pure" solution is clearly the one where NaN's yield a NaN. It is > isomorphic to a Monoid based on optional values with the NaN forming the > unknown element, often called the option-monoid or the maybe-monoid. Or > said otherwise: for any type a, where a is a monoid, so is 'option a': > > forall a. Monoid a => Monoid (Option a) > > Since the max function itself is a monoid with the neutral element taken > as the minimal value, -inf, we can thus "lift" NaN values on top and still > have some mathematical consistency. > > But if you have an array of values, where some of them can be NaN, it is > probably smart to provide a function which filters the array of NaN values > and computes the remaining values. R provides the function `na.omit(..)` > for this. So do option types in languages such as OCaml or Haskell. Here, > we are not that concerned about what is sound from a formal point of view, > but more concerned with data cleanup before we start using the formality. > > Max of two values is a special case. The generalized version takes an > array/vector of values and provides the maximum in the structure. > > All of this leads me to conclude you probably want both variants in the > end. In some situations, it is better to include NaN values in the > computation, and in other situations, the right move is to omit/exclude > them. > > > -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/9950a06b-1222-4543-b979-763f37516e2an%40googlegroups.com.
