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

Re: Motion9 ExactDot Product



Arnold Neumaier schrieb:
Ulrich Kulisch wrote:
Arnold Neumaier wrote:

> I had therefore asked for providing evidence for applications that really
> need the exact dot product, but this hasn't generated any response.
There was some response: I copy a few sentences of my response mail (dated October 20):

The IFIP Working Group letter requests an "exact" dot product as support for long interval arithmetic (for details see [6] or [9]). Long interval arihtmetic opens a new area of applications for interval arithmetic. I give a few examples:

The Krawczyk operator frequently is used to compute a verified enclosure for the solution of a system of linear equations. But for ill conditioned problems (take the Hilbert matrices of dimension larger than 11) the Krawczyk operator fails to find the solution.

In such cases a more powerful operator (called the Rump operator in [9]) solves the problem. In its extended form the Rump operator can be used to compute a verified solution even for extemely ill conditioned problems, see [11]. The Rump operator uses the "exact" dot product.

But all this can be done as efficiently with the optimally rounded dot product, since all that matters is that sufficiently many significant digits are produced.

This is the reason, I think, why Siegfried Rump got interested in
algorithms for the latter. In any case, I'd like to see his comment
on this issue.



Another very nice example is given in [7], an iteration with the logistic equation. Double precision floaint-point or interval arithmetic totally fail (no correct digit) after 30 iterations while long interval arithmetic after 2790 iterations still computes correct digits of a guaranteed enclosure. Long interval arithmetic allows computing highly accurate bounds for real arithmetic expressions.

Rump published work on how to get least significant bit accuracy for arbitrary arithmetic expressions
The original work of Siegfried Rump and Harald Böhm (done at Karlsruhe in 1982/83) use the exact dot product to prove this assertion. Of course there are applications where less accuracy suffices. If it does not suffice you can start an iteration. However, the exact dot product is the ultimate tool.
(and this includes 2790 iterations of the logistic equation)
I recommend to do what you claim here. The experience could help to understand the meaning and the simplicity of the exact dot product.

Ulrich Kulisch
Arnold Neumaier