Re: [boost] Bugs in boost random number distributions

From: Stephan Tolksdorf (andorxor_at_[hidden])
Date: 2006-08-08 16:38:48

>>> * a clumsy (and inexact) interface between the random generators and
>>>  the random distributions.
>> I wonder what you refer to here?
> 
> Here's one example.  In bernoulli_distribution we have
> 
>     if(_p == RealType(0))
>       return false;
>     else
>       return RealType(eng() - (eng.min)()) <=
>              _p * RealType((eng.max)()-(eng.min)());
> 
> where eng() returns integers between eng.min() and eng.max() incl.  The
> problems are:
> 
> * If the integers returned by eng() are signed (e.g.,
>   linear_congruential) then eng() - (eng.min)() and (eng.max)() -
>   (eng.min)() can overflow and the results are bogus.  
 > (...)

RealType(eng() - (eng.min)()) could just be replaced with
RealType(eng()) - RealType((eng.min)()), or alternatively template 
metaprogramming techniques be applied (result_type, min and max are 
compile time constants in the standard proposal).

> * The results depend on the granularity of the underlying engine.  Why
>   not return an exact result?  The answer lies in the overall structure
>   of the library where the bernoulli is expected to use the raw numbers
>   coming out of the generator.  If there were a
>   uniform_float01<RealType>(eng); 

See the generate_canonical template function in the standard proposal.

(...)
>>   template<class Engine>
>>   result_type operator()(Engine& eng)
>>
>> is sometimes awkward.  It doesn't allow compile-time inquiry of the
>> engine properties by the distribution class. 

min and max are static constants in the standard proposal, so there's no 
general problem with compile time inspection.

Stephan