[Caml-list] min, max and nan

[Caml-list] min, max and nan

[Yaron Minsky]

One of the oddities associated with NaNs in ocaml is what happens when
you take the min of something and nan.  Witness:

# min nan 3.;;
- : float = 3.
# min 3. nan;;
- : float = nan

Surprisingly, the result depends on the ordering.  Why it does so is
clear once you look at the implementation and ponder for a moment the
IEEE requirements on comparisons involving nan.  When using ocaml
3.07, I tried and failed to come up with a polymorphic min that
behaved reasonably in this case, that is, that returned nan when
either argument is nan.  With 3.08, this is now doable.  Here's how it
works:

# let contains_nan x = x <> x;;
val contains_nan : 'a -> bool = <fun>
# let nmin x y = 
  if contains_nan x then x 
  else if contains_nan y then y 
  else min x y;;
      val nmin : 'a -> 'a -> 'a = <fun>
# nmin 3. nan;;
- : float = nan
# nmin nan 3.;;
- : float = nan

Still, we don't quite escape from the oddities of nan.  We still get
order dependence in the case of larger data structures that include
nan's:

# nmin (1,nan) (2,nan);;
- : int * float = (1, nan)
# nmin (2,nan) (1,nan);;
- : int * float = (2, nan)

But in my mind, the current state of affairs is a big improvement.

Yaron

-------------------
To unsubscribe, mail caml-list-request@inria.fr Archives: http://caml.inria.fr
Bug reports: http://caml.inria.fr/bin/caml-bugs FAQ: http://caml.inria.fr/FAQ/
Beginner's list: http://groups.yahoo.com/group/ocaml_beginners

Re: [Caml-list] min, max and nan

[Markus Mottl]

On Wed, 14 Jul 2004, Yaron Minsky wrote:
> When using ocaml 3.07, I tried and failed to come up with a polymorphic
> min that behaved reasonably in this case, that is, that returned
> nan when either argument is nan.  With 3.08, this is now doable.
> Here's how it works:

Another possibility is to use "classify_float":

  let nmin x y =
    if classify_float x = FP_nan || classify_float y = FP_nan then nan
    else min x y

This is actually even more efficient.

> Still, we don't quite escape from the oddities of nan.  We still get
> order dependence in the case of larger data structures that include
> nan's:

I think this could be fixed by making use of the Obj-module, though this
would look a little bit messy.

Regards,
Markus

-- 
Markus Mottl          http://www.oefai.at/~markus          markus@oefai.at

-------------------
To unsubscribe, mail caml-list-request@inria.fr Archives: http://caml.inria.fr
Bug reports: http://caml.inria.fr/bin/caml-bugs FAQ: http://caml.inria.fr/FAQ/
Beginner's list: http://groups.yahoo.com/group/ocaml_beginners