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)



On Wed, 11 Feb 2009 21:47:20 -0100, Michel Hack <hack@xxxxxxxxxxxxxx> wrote:

> |     If xx=[l,u] is compact, m and r shall have the values defined by
> |         set round up
> |         r = 0.5*(u-l);m=l+r.

That formula is due to Prof. Oishi from Waseda university, Tokyo.  See
   http://www.ti3.tu-harburg.de/paper/rump/OiRu02.pdf


I am sorry, I misread the text above. Oishi's method, as it is in
the above paper, is

   setround up
   m = (l+u)/2
   r = m-l

It follows that m-r <= l   AND   u <= m+r.



So we what do make of John Pryce's example?

    Input:  [0.9993, 1.001]    (format: 4-digit Decimal FP)
    r = 0.5 * (1.001 - 0.9993)  = 0.0008500   (exactly)
    m = 0.9993 + 0.0008500      = 1.0001500   (exactly)
                                = 1.001       (rounded up)
    Result:  1.001+-0.0008500  (decimal precision 4 digits)

    Hovever, 1.001 +- 0.00085  does not cover the original lower bound,
    as the exact range is 1.00015 to 1.00185 -- it is slightly shifted.


With the correct formula it looks as follows:

   Input:  [0.9993, 1.001]    (format: 4-digit Decimal FP)
   m = 0.5 * (0.9993 + 1.001)  = 0.5 * 2.0003   (exactly)
                               -> 0.5 * 2.001   (rounded up)
                               = 1.0005         (exactly)
                               -> 1.001         (rounded up)
   r = 1.001 - 0.9993          = 0.0017         (exactly)
                               = 0.0017         (rounded up)
   Result:  1.001+-0.0017  (decimal precision 4 digits)

A correct inclusion, though overestimated.
Sometimes the following formula, though a little more expensive, yields
a better result. In the example above there seems no difference.

   setround up
   m = l + (u-l)/2
   r = m-l




Now, the cited paper (after a very quick glance, I have to admit)
does not actually claim that:    m-r <= l   AND   u <= m+r

It claims that a certain quantity X contained in [l, u] is also
contained in [m-r, m+r], and that the MidRad form is easier to
compute and yet as good (or nearly as good) as the InfSup form.

That is an entirely different matter.



No, the quantity X is an arbitrary quantity within [l,u], thus
X in [m-r,m+r] implies  m-r <= X  AND   X <= m+r ,  and therefore
m-r <= l   AND   u <= m+r .

Best wishes

Siegfried M. Rump


--
=====================================================
Prof. Dr. Siegfried M. Rump
Institute for Reliable Computing
Hamburg University of Technology
Schwarzenbergstr. 95
21071 Hamburg
Germany
phone  +49 40 42878 3027
fax    +49 40 42878 2489
http://www.ti3.tu-harburg.de

and

Visiting Professor at Waseda University
Faculty of Science and Engineering
Shinjuku Lambdax Bldg. 902
2-4-12 Okubo, Shinjuku-ku
Tokyo 169-0072
Japan
phone/fax in Japan  +81 3 5286 3414