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

Re: converting to a wider data type



Van, Baker and P1788

On 11 Nov 2008, at 12:43, R. Baker Kearfott wrote:
> Van et al,

Is it really that complicated?  This goes back to a basic difference
between us in point of view, reflected in both our view of floating
point conversions to intervals and interval-to-wider-interval
conversions.  ...  Looking at the underlying
mathematics, that means that, if [a,b] --> [aa,bb], what is
absolutely necessary is that aa <=a and bb <=b, when a, aa, b, and bb
are interpreted mathematically.  I think  IEEE 754-2008 directed rounding
between data types also specifies an "accuracy" of this conversion
in the sense that aa should be the largest number in the target
precision <= a and bb should be the smallest number in the target
precision >=a.  a and b are completely defined mathematical
numbers, and there is no need to imagine any bits that aren't
actually there in a and b themselves.  This statement is on the
"underlying model" level of discussion, but translates in a mechanical
way to bit patterns.
-----

Well stated, Baker! I agree 100%.

The great Jim Wilkinson used, and probably invented, the function
   fl(x) = "real number x rounded to the nearest floating point value y",
with "nearest" in the sense of a chosen rounding mode.

It's clear from how *he* used it that y is just another real number like x, not something (a half-open 
interval?) with "random bits". That's how everyone I know of has always done roundoff error analysis, let's 
stick with it.

Best wishes

John Pryce
j.d.pryce@xxxxxxxxxxxx




-----------------------------------------
Email sent from www.virginmedia.com/email
Virus-checked using McAfee(R) Software and scanned for spam