Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: MidRad to/from InfSup (was: the "set paradigm" is harmful)



Shin'ichi Oishi schrieb:

what I proposed is

set round up
m=0.5*(l+u)
r=m-l

This gives a correct enclosure but has the blemish that
m need not be in [l,u].

E.g., in 3-digit decimal arithmetic when l=5.20, u=5.21,
we have l+u=10.5, m=5.25, r=0.05.

In my opinion, m in [l,u] is important to avoid unintended
surprises since the midpoint might be later used as a typical
point of the original interval.


----- Original Message ----- From: "Michel Hack" <hack@xxxxxxxxxxxxxx>
   Hovever, 1.001 +- 0.00011  does not cover the original lower bound,
   as the exact range is 1.00001 to 1.00111 -- it is slightly shifted.


I conclude that the proposed formula does not always provide
a valid MidRad enclosure of a given InfSup interval.

Now I'll open my proposed alternative to attack; my head is spinning:

   Set round nearest
   m = 0.5 * (l+u)
   Set round up
   r = max(u-m, m-l);

Sorry for this sloppiness in the Vienna Proposal.

I think the formula that should be in the standard is
    set round up
    m=l+0.5*(u-l);r=m-l

This ensures that m in [l,u] for any basis, and gives in the above example the enclosure m=5.21, r=0.01, which is 5 times narrower than
the Oishi enclosure.

In binary arithmetic, or when intermediate results are kept to higher
precision, the two formulas appear to be equivalent.


Arnold Neumaier