[Haskell-cafe] Uncertainty analysis library?

[Tom Nielsen]

so if you want to do it the quick, easy and approximately correct, why not
just use a monte-carlo monad?

say you have two values x and y, for which you know the means and
standard deviations mx, my, sdx, sdy. and you have a function f(x,y)
that expresses wheat you want to know. so then in the random-fu monad
you say (i can't member exactly how it works; the following is valid
in "probably", my unreleased probability/stats library at
https://github.com/glutamate/samfun)

xs = sampleN 1000 $ do
         x <- normal mx sdx
         y <- normal my sdy
         return $ f(x,y)

and then take the mean and standard deviation of xs? that's every bit
as correct as propagating the uncertainty of f, except for the finite
number of samples. (assuming your original uncertainties are gaussian)

Tom

On Sun, Mar 20, 2011 at 11:59 PM, Edward Amsden <eca7215 at cs.rit.edu> wrote:
> I'm actually a CS undergrad in a physics lab class. I have permission
> from my professor to use computer programs for analysis of lab data. I
> need to do calculations on data with uncertainty, but uncertainty
> analysis on many formulae in physics is rather tedious. I was hoping
> for something with instances for Num, Fractional, Floating, etc. that
> would allow me to combine two uncertain values and get a new value
> with uncertainty. I've been working on writing one myself and I don't
> find the concept hard, but it's a lot of effort that I don't want to
> duplicate if it's been done already.
>
> On Sun, Mar 20, 2011 at 5:46 PM, Tom Nielsen <tanielsen at gmail.com> wrote:
>> Interval arithmetic is of course not the same as uncertainty, although
>> computer scientists like to pretend that is the case. (and uncertainty
>> estimates do not have the be "rough".)
>>
>> In general the propagation of errors depends on whether the errors are
>> independent or not. The rules are given in Taylor: An introduction to
>> Error analysis (1997). Interval artihmetic corresponds to the worst
>> case of non-independent and non-random errors. In the case of
>> independent of random errors, you get:
>>
>> data Approximately a = a :+/-: a
>>
>> instance Num a => Num (Approximately a) where
>>  (m1 :+/-: err1) +  (m2 :+/-: err2) = (m1+m2) :+/-: (sqrt(err1^2+err2^2)
>>  (m1 :+/-: err1) -  (m2 :+/-: err2) = (m1-m2) :+/-: (sqrt(err1^2+err2^2)
>>  (m1 :+/-: err1) *  (m2 :+/-: err2) = (m1*m2) :+/-:
>> (sqrt((err1/m1)^2+(err2/m2)^2)
>>
>> the general rule is
>>
>> if y = f xs where xs :: [Approximately a], i.e f :: [Approximately a]
>> -> Approximately a
>>
>> the error term= sqrt $ sum $ map (^2) $ map (\(ym :+/-: yerr) ->
>> partial-derivative-of-yerr-with-respect-to-partial-ym * yerr) xs
>>
>> You can verify these things by running your calculation through soem
>> sort of randomness monad (monte-carlo or random-fu packages) Anyways,
>> I ended up not going down this route this because probabilistic data
>> analysis gives you the correct error estimate without propagating
>> error terms.
>>
>> Tom
>>
>> PS if you're a scientist and your accuracy estimate is on the same
>> order as your rounding error, your are doing pretty well :-) At least
>> in my field...
>>
>> On Sun, Mar 20, 2011 at 8:38 PM, Edward Kmett <ekmett at gmail.com> wrote:
>>> I have a package for interval arithmetic in hackage
>>> http://hackage.haskell.org/package/intervals-0.2.0
>>> However it does not currently properly adjust the floating point rounding
>>> mode so containment isn't perfect.
>>> However, we are actively working on fixing up the Haskell MPFR bindings,
>>> which will let us reliably set rounding modes, cleaning up the interval
>>> arithmetic library to be just a little bit more pedantic. Due to the way GHC
>>> interacts with GMP this is a disturbingly difficult process.
>>> I have an unreleased library for working with Taylor models that builds on
>>> top of that and my automatic differentiation library, but without working
>>> MPFR bindings, it isn't sufficiently accurate for me to comfortably release.
>>> -Edward
>>>
>>> On Sun, Mar 20, 2011 at 4:27 PM, Edward Amsden <eca7215 at cs.rit.edu> wrote:
>>>>
>>>> Hi cafe,
>>>>
>>>> I'm looking for a library that provides an instance of Num,
>>>> Fractional, Floating, etc, but carries uncertainty values through
>>>> calculations. A scan of hackage didn't turn anything up. Does anyone
>>>> know of a library like this?
>>>>
>>>> Thanks!
>>>>
>>>> --
>>>> Edward Amsden
>>>> Student
>>>> Computer Science
>>>> Rochester Institute of Technology
>>>> www.edwardamsden.com
>>>>
>>>> _______________________________________________
>>>> Haskell-Cafe mailing list
>>>> Haskell-Cafe at haskell.org
>>>> http://www.haskell.org/mailman/listinfo/haskell-cafe
>>>
>>>
>>> _______________________________________________
>>> Haskell-Cafe mailing list
>>> Haskell-Cafe at haskell.org
>>> http://www.haskell.org/mailman/listinfo/haskell-cafe
>>>
>>>
>>
>
>
>
> --
> Edward Amsden
> Student
> Computer Science
> Rochester Institute of Technology
> www.edwardamsden.com
>
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe at haskell.org
> http://www.haskell.org/mailman/listinfo/haskell-cafe
>