Re: Table length reduction

Chris Rokusek wrote:
>
> D.C.,
>
> The more I think about this, the more I believe an even better solution is
> that the error function be computed for BOTH a linear and cubic spline and
> then a given point can only be removed if it simultaneously below a limit
> (or within a percentile) for both. This would prevent the cubic spline from
> removing elbows and keep linear from removing points in the curve "where
> there is about to be a change."

Sounds good, but since this is done on a sort basis there needs to be
a scalar metric for prioritizing culls. Looks like you may have to
learn Icon yet, Chris!

> Even with this contraint, it seems like there still be a lot of truly
> worthless points removed which is the whole point of the excersise. What do
> you think?

Don't see why not. In general, though, a linear interpolation is just
a very special case of cubic. Maybe the linear will accomplish what
you want.

> Another observation: Since the algorithm computes all the error values as a
> first pass and then begins culling starting with the lowest error first, it
> seems as though the error values near (2 neighbors) the first removed point
> are now invalid (cause they depended on the removed point) and this should
> be taken into consideration somehow (either by not allowing the neighbors to
> be culled or recomputing the error for the neighbors).

Already done. The algorithm culls the least-error point, then marks its four
neighbors as keepers until the next pass, when the errors are all recalculated.

> Is it getting hairy enough yet?

I'm just rusty enough on the math that I don't know if there's a real chance
of later cullings increasing the error for points culled earlier to unacceptable
degree. That's one of the reasons I put this thing out -- to have people push
real datasets through it to see if the results are OK.

> > -----Original Message-----
> > From: Weston Beal [mailto:weston_beal@mentorg.com]
> > Sent: Monday, May 08, 2000 2:02 PM
> > To: 'IBIS Mailing list'
> > Subject: RE: Table length reduction
> >
> >
> > D.C.,
> >
> > I think I understand what you're doing, but wouldn't it be easier
> > and maybe
> > more applicable to do linear interpolation? If the program (simulator or
> > translator) that reads the resulting IBIS file does cubic spline then you
> > are correct in doing the same in your filter. If the IBIS file is going
> > into a program that does piece-wise linear interpolation then the filter
> > should do the same, I think.
> >
> > On a related subject, what do ya'll think about a smoothing filter (least
> > square?) to reduce measurement noise in IBIS tables? Is it useful or
> > dangerous?
> >
> > Regards,
> > Weston

-- 
D. C. Sessions
dc.sessions@vlsi.com