[Haskell-beginners] How to improve the accuracy of floating point calculation?

Darren Grant therealkludgy at gmail.com
Wed Feb 6 02:50:07 CET 2013


It's a pathological situation that arises from the way enumFromThenTo is
defined for floating point values. Does anyone know why the specific
implementation isn't more accurate?

In simple cases you can sometimes convert an integer range to desired
floating point range to avoid cumulative precision errors.

Cheers,
d



On Tue, Feb 5, 2013 at 5:07 PM, Mateusz Kowalczyk
<fuuzetsu at fuuzetsu.co.uk>wrote:

> We just figured it out over here. The short version is that the 0.1 and
> 0.2 just happen to round pretty well (to what looks perfect to a human)
> while other figures don't. This can be easily represented by forcing a
> [Float] type for lower precision as opposed to the default double
> precision:
> *Main> [0.1, 0.2 .. 1] :: [Float]
>
> [0.1,0.2,0.3,0.40000004,0.50000006,0.6000001,0.7000001,0.80000013,0.90000015,1.0000002]
>
> Above we can easily see that with a float the errors compound so much
> that the numbers no longer round to a human-friendly representation.
>
> This can also be shown with doubles: the 0.1 was displayed as 0.1
> because it just happened to be fairly precise. Start compounding errors
> early enough and it no longer rounds so nicely:
>
> [-0.5,-0.4,-0.30000000000000004,-0.20000000000000007,-0.10000000000000009,-1.1102230246251565e-16,9.999999999999987e-2,0.19999999999999984,0.2999999999999998,0.3999999999999998,0.4999999999999998]
>
> You can see just how much precision we lost at mid-point by looking at
> what should be 0.
>
> Thanks
>
> On 06/02/13 00:47, Mateusz Kowalczyk wrote:
> > I thought that KC was asking why is 0.5 not displayed as 0.5 when it's
> > supported by the IEEE754. Even if that wasn't question, it is my
> > question now.
> >
> > IEEE754 is capable of representing 0.5 with perfect precision; in fact,
> > it would be `00111111000000000000000000000000'.
> >
> > My question is why [0.1, 0.2 .. 1.0] comes up with an imprecise
> > midpoint:
> >
> [0.1,0.2,0.30000000000000004,0.4000000000000001,0.5000000000000001,0.6000000000000001,0.7000000000000001,0.8,0.9,1.0]
> >
> > My initial guess is that the expansion is done by working out the
> > difference between the first two elements and using that to generate the
> > rest. This also makes me question why 0.1 is shown properly when it
> > can't be represented precisely in IEEE754. Quickly rolling own list
> > generation function:
> >
> > *Main> let f x y = x : f y (y + y - x)
> > *Main> take 10 $ f 0.1 0.2
> >
> [0.1,0.2,0.30000000000000004,0.4000000000000001,0.5000000000000001,0.6000000000000001,0.7000000000000001,0.8,0.9,1.0]
> >
> > reveals the same result as the short-hand which would imply that it's
> > exactly what it's doing (for the simple case). My question for why can
> > 0.1 be shown properly still stands in this case.
> >
> > On 06/02/13 00:12, Darren Grant wrote:
> >> I'm not sure how CReal implements its values, but IEEE754 also supports
> >> decimal formats preferred for accuracy in many applications. Take a
> look:
> >>
> >>    http://en.wikipedia.org/wiki/IEEE_floating_point
> >>
> >>
> >> Cheers,
> >> d
> >>
> >>
> >>
> >>
> >> On Tue, Feb 5, 2013 at 3:24 PM, Patrick Mylund Nielsen
> >> <haskell at patrickmylund.com <mailto:haskell at patrickmylund.com>> wrote:
> >>
> >>     http://floating-point-gui.de/
> >>     http://floating-point-gui.de/formats/fp/
> >>
> >>     http://en.wikipedia.org/wiki/Floating_point#Accuracy_problems
> >>
> >>
> >>     On Wed, Feb 6, 2013 at 12:08 AM, KC <kc1956 at gmail.com
> >>     <mailto:kc1956 at gmail.com>> wrote:
> >>
> >>         0.1 cannot be represented exactly in floating point.
> >>
> >>         0.5 can be represented exactly.  Why?
> >>
> >>
> >>         On Tue, Feb 5, 2013 at 2:41 PM, yi lu
> >>         <zhiwudazhanjiangshi at gmail.com
> >>         <mailto:zhiwudazhanjiangshi at gmail.com>> wrote:
> >>         > Hi,
> >>         >
> >>         > I found that in ghci, I input
> >>         > [0.1,0.2..2]
> >>         > and run, I get a result of
> >>         >
> >>         >
> >>
> [0.1,0.2,0.30000000000000004,0.4000000000000001,0.5000000000000001,0.6000000000000001,0.7000000000000001,0.8,0.9,1.0,1.1,1.2000000000000002,1.3000000000000003,1.4000000000000004,1.5000000000000004,1.6000000000000005,1.7000000000000006,1.8000000000000007,1.9000000000000008,2.000000000000001]
> >>         >
> >>         > But, as you know, it is not the exact answer.
> >>         >
> >>         > So, I wonder if there is something I can do to achieve a
> >>         better performance
> >>         > and get [0.1,0.2,0.3,0.4..] as the result.
> >>         >
> >>         > Thanks.
> >>         >
> >>         > _______________________________________________
> >>         > Beginners mailing list
> >>         > Beginners at haskell.org <mailto:Beginners at haskell.org>
> >>         > http://www.haskell.org/mailman/listinfo/beginners
> >>         >
> >>
> >>
> >>
> >>         --
> >>         --
> >>         Regards,
> >>         KC
> >>
> >>         _______________________________________________
> >>         Beginners mailing list
> >>         Beginners at haskell.org <mailto:Beginners at haskell.org>
> >>         http://www.haskell.org/mailman/listinfo/beginners
> >>
> >>
> >>
> >>     _______________________________________________
> >>     Beginners mailing list
> >>     Beginners at haskell.org <mailto:Beginners at haskell.org>
> >>     http://www.haskell.org/mailman/listinfo/beginners
> >>
> >>
> >>
> >>
> >> _______________________________________________
> >> Beginners mailing list
> >> Beginners at haskell.org
> >> http://www.haskell.org/mailman/listinfo/beginners
> >>
> >
> > _______________________________________________
> > Beginners mailing list
> > Beginners at haskell.org
> > http://www.haskell.org/mailman/listinfo/beginners
> >
>
> _______________________________________________
> Beginners mailing list
> Beginners at haskell.org
> http://www.haskell.org/mailman/listinfo/beginners
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/beginners/attachments/20130205/526c0a62/attachment-0001.htm>


More information about the Beginners mailing list