From: H. Olin Spivey <ospivey@bmb-fs1.biochem.okstate.edu>
  To  : HOLLADAYL@aol.com, rasmb@alpha.bbri.org
  Date: Tue, 16 May 2000 14:38:57 -0500

Re: who has a good fitting function with which to computeMw(r,app)?

Les,

   I can sympathize with your desire to avoid smoothing strategies.
However, you might still wish to try the smoothing routine by C. H.
Reinsch, Numerische Mathematik 10 (1967) 177-183.  It is not a sliding
algorithm.  Instead it smooths globally and to an extent consistent with
the smoothing factor S and standard deviation(s) Sig(i) you designate.
With S approx. = NPTS, your sum of squares of normalized residuals,
R(i)/Sig(i), will be consistent with the specified Sig(k).  I start with
this S, but allow for input of other values if I desire.  You definitely
don't won't a strict spline fit (S = 0), which would force the curve
through all the experimental points.  I have used this Reinsch routine for
years and specifically wrote a program for calculating the Mw vs.
absorbance from the XL-A data using this routine.

   The original Savitsky & Golay sliding algorithm requires equally spaced
points and will give very large artifacts upon deletion of intervening
points.  J. Steinien, et. al. corrected some of the problems, but not all.
We are not familiar with Dierckx's algorithm.

   I could send you a FORTRAN copy of the Reinsch algorithm as written by
my colleague in the Computer Science Dept. if you wish.

   Cordially,
   Olin
__________________________________________________________________


At 8:33 PM -0400 5/12/00, HOLLADAYL@aol.com wrote:
>HI all your gurus,  wizards, and elves...
>
>I'm searching for a function with which to fit A vs. r^2 with the aim of
>computing pointwise Mw estimates from absorbance scans.  Now I want to avoid
>a sliding fit of any kind in favor of a single function for the entire scan
>that replicates the essential features of the A vs. r^2 curve without
>introducing artifacts.
>
>Ages ago Jim Osborne used (I believe) a fitting function that was a quadratic
>in r^2 divided by a linear function of r^2.   I've tried Dierckx's
>interpolated cubic spline method, but find it is numerically problematical in
>that deleting a few points drastically alters the results.
>
>Now please do not tell me that this is a bad thing to do, and that it is not
>the best way to analyze sed equil data.  I know this already.  But a friend
>desperately needs a reliable way to do this, and I solicit any ideas, advice,
>or suggestions you might have as to fitting functions.
>
>I promise to put at rasmb any resulting software with a variety of fitting
>options.
>
>thanks all...
>
>Les Holladay


H. Olin Spivey                       Phone: (405) 744-6192
Dept. Biochem. & Molec. Biology      Fax:   (405) 744-7799
246 NRC                              Email: OSpivey@Biochem.Okstate.Edu
Oklahoma State University
Stillwater, OK 74078-3035