Re: Smoothing allowance for raw data

In article <[EMAIL PROTECTED]>, [EMAIL PROTECTED] says...

>I don't want to smooth it - my client does :)  or at least they are looking
>at it as an option.
>
>The EPA does allow a small amount of smoothing, which can 'help' with the
>data systems peak detection routines.  I don't agree with smoothing, but if
>it's valid, legal and the client requests it, then so be it.  Smoothing
>already occurs to some extent in the detector electronics, either as a side
>effect of the electronic design, or intentionally through routines in the
>DSP.  But with hi-res GC/MS systems, there are a lot of sources of noise,
>and sometimes, a little smoothing will help the computers peak detection
>"eye" just as it would help an analysts "eye".  If done correctly, the
>smooth will greatly enhance the efficiency of peak detection, without
>invalidating the quantitation results.  The document that I vaguely remember
>specifies what level this smoothing is.
>

In my experience, there are 2 classes of smoothing algorithms:
(a) those that acknowledge the noise and do the maximum amount of smoothing 
possible - this class starts with maximum entropy or bayesian principles.
(b) those that do not acknowledge the noise, and so the user must use them 
with care.  They can easily oversmooth.  

Furethermore there are 2 types of algorithms: 
(1) the ones that will do even more damage when used a second time, 
(2) the ones that don't do anything the second time. You got the optimum 
amount of smoothing the first time, and a second pass can't improve it.

All the smoothing methods of class (a) are type (2).
In class (b), Fourier smooth is type (2).  

I recommend going with type (2).  Maximum entropy, or Fourier, or Wiener 
smoothing.

If you ever saw what S-G smoothing looks like in the Fourier domain, you 
would probably never use it again.  

Best wishes,
Lin DeNoyer