Stavros Macrakis wrote:
> It is interesting that you are using the same formulas regardless of the
> precision; I would have thought that the coefficients would have to be
> scaled for each precision to avoid overflow.
>
> In any case, I don't think you'll be able to get very far using the
> exact formulae involving rounding, since very few algebraic operations
> (simplifications) are possible. Just as in floating-point, you should
> probably work with an error term for each operation; of course with
> fixed-point, it is an absolute, not a relative, error.
You are right: I should not use the same formula, but compute the
subterms - like a compiler - and balance precision.
Using float error terms might be a vehicle to get a good error
estimation.
I'll try it.