[Pharo-project] SIXX problem for ScaledDecimal
rpboland at gmail.com
Fri May 20 10:57:12 EDT 2011
> From: Chris Cunningham <cunningham.cb at gmail.com>
> Subject: Re: [Pharo-project] SIXX problem for ScaledDecimal
> To: Pharo-project at lists.gforge.inria.fr
> Message-ID: <BANLkTinP7E_oh2Fi2=axmra6JsoXVJYWkg at mail.gmail.com>
> Content-Type: text/plain; charset=ISO-8859-1
> For ScaledDecimal, SIXX should definite store it as the underlying
> fraction. Storing it in the printed fashion does change the value of
> the ScaledDecimal.
> (1/3) asScaledDecimal: 2 gives 0.33s2
> (33/100) asScaledDecimal: 2 gives 0.33s2
> 0.33s2 ~= ( (1/3) asScaledDecimal: 2 )
> ( (1/3) asScaledDecimal: 2 ) ~= ( (33/100) asScaledDecimal: 2 )
> In other words, the ScaledDecimal is about the precise internal number
> and the scale to display it - which is NOT the scale that it is stored
I don't agree. ScaledDecimal would be used where consistency is often
more important than accuracy, e.g. in Financial calculations where being off by
a few pennies doesn't matter but always getting the same answer matters a LOT.
Once a number is converted to a ScaledDecimal computations must always
produce the same result assuming enough digits. To ensure this asScaledDecimal
must produce a scaled decimal for use in computations and, as such, storing
a fraction internally is unacceptable.
>From a financial point of view 1/3 * 3 is not 1!
>From a programmers point of view asScaledDecimal changes the value
in much the same way asInteger does for a float.
Having said all of this allow me to contradict myself. It seems to me that
a scaled decimal should in general store more digits than are displayed.
For example money might be stored with a large (possibly arbitrary) number
of digits after the decimal point but only display two digits beyond the decimal
I don't know how ScaledDecimal is implemented but I would be inclined to
use arbitrary or large size integers and store the location of the decimal point
internally recomputing the position of the decimal point after each calculation.
This is an expensive way of doing computations but I expect users of
ScaledDecimal can live with that. When converting non decimal results to
ScaledDecimal then how many digits of precision to maintain must be specified
Dealing with ScaledDecimal is a complicated matter and I have probably failed
to appreciate many of the complications but whoever is implementing
(and Smalltalk needs ScaledDecimal) needs to appreciate those complications.
Anybody here a financial expert?
More information about the Pharo-dev