Boost logo

Ublas :

From: Leopoldo Peralta (lperalta_at_[hidden])
Date: 2007-08-24 12:13:19

This is the point. The issue of error propagation due to repeated operations
on floats could conduct to inaccurate answers in matrix factorization (i.e
Gauss elimination, LU Equation Linear solving, Matrix inversion, Cholevsky,
QR etc.)

If the numbers you are dealing with can be represented as rational types,
the result will be exact, provided no integer overflow occurs at any
calculation phase. That requires that rational numerator and denominator are
between the values std::numeric_limits<__int>::max/min)().


Leopoldo Peralta

-----Mensaje original-----
De: ublas-bounces_at_[hidden] [mailto:ublas-bounces_at_[hidden]] En
nombre de Giuseppe Bilotta
Enviado el: Friday, August 24, 2007 6:47 AM
Para: ublas_at_[hidden]
Asunto: Re: [ublas] [patch] improve type deduction

On Friday 24 August 2007 13:22, Karl Meerbergen wrote:

> What is the advantage of boost::rational numbers compared to floating
> numbers?

No loss of precision for arithmetic operations, and consequently guarantee
of exactness of all operations.

Giuseppe "Oblomov" Bilotta
ublas mailing list