[Haskell-cafe] ANNOUNCE: approximate-equality 1.0 -- Newtype
wrappers for approximate equality
Neil Brown
nccb2 at kent.ac.uk
Tue Aug 3 06:27:06 EDT 2010
On 03/08/10 05:32, Gregory Crosswhite wrote:
> I am pleased to announce the release of the package "approximate-equality", which provides newtype wrappers that allow one to effectively override the equality operator of a value so that it is/approximate/ rather than/exact/. The wrappers use type annotations to specify the tolerance; the 'Digits' type constructor has been provided for specifying the tolerance using type-level natural numbers. Instances for all of the classes in the numerical hierarchy have been provided for the wrappers, so the wrapped values can mostly be used in the same way as the original values. (In fact, most of the time one doesn't even have to wrap the starting values, since expressions such as (1+sqrt 2/3) are automatically wrapped thanks to the 'fromIntegral' method of the 'Num' typeclass.)
>
>
> The motivation behind this package is that there are times when working with floating point numbers that one would like the equality operator to check for approximate equality rather than exact equality. For example, in one of my own projects I have a walker that spits out a list of coordinate values that I keep track of, and since floating point arithmetic is inexact it will often arrive at the same point from two different paths and thus obtain slightly different values for the coordinates of that point. I could have just written a function to do approximate matching of the point and be done with it, but then I can't leverage pre-built data structures such as Data.Set and Data.Map which use (==) for determining whether a key is a member.
>
> This package is compatible with Haskell 2010, as it only uses the EmptyDataDecls extension, and the only package dependency is type-level-natural-numbers which itself is Haskell 2010.
>
> Any feedback from the community is, of course, very welcome.
>
>
I like the look of this. Eq and Ord instances that use epsilon values
look like they will be handy. I have a design question/suggestion. You
have:
class AbsoluteTolerance absolute_tolerance where
absoluteToleranceOf :: Fractional value =>
AbsolutelyApproximateValue absolute_tolerance value -> value
Why do you need this class (and the other two)? It looks like you can
just define:
absoluteToleranceOf :: (NaturalNumber n, Fractional value) =>
AbsolutelyApproximateValue (Digits n) value -> value
absoluteToleranceOf = toleranceFromDigits . getAbsoluteTolerance
So you can get the same function without needing to add a type-class.
Or is it that you envisage other tolerance specifications besides
Digit? If so, I wonder if this flexibility complicates your API
unnecessarily. If you removed those type-classes and just stuck with
Digits, the size of your API documentation would halve, and make the
docs a lot more readable.
Thanks,
Neil.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.haskell.org/pipermail/haskell-cafe/attachments/20100803/1cf1c761/attachment.html
More information about the Haskell-Cafe
mailing list