home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky sci.math:10680 sci.math.stat:1766
- Newsgroups: sci.math,sci.math.stat,sci.math.num_analysis
- Path: sparky!uunet!cs.utexas.edu!zaphod.mps.ohio-state.edu!pacific.mps.ohio-state.edu!linac!att!att!cbnewsi!shankx
- From: shankx@cbnewsi.cb.att.com (n.k.shankaranarayanan)
- Subject: Fitting a few coefficients to a 'peaky' function
- Organization: AT&T
- Date: Fri, 28 Aug 1992 18:56:22 GMT
- Message-ID: <1992Aug28.185622.15405@cbnewsi.cb.att.com>
- Lines: 48
-
- SUMMARY:
- This question is about extracting a few 'coefficients'
- from a complex (peaky) funtion, so that changes in the
- function can be tracked from sample to sample.
- Would appreciate general advice and suggestions,
- in the hope of uncovering some methods unknown to me.
- Please also recommend books/references.
-
- PROBLEM:
- I have a datsets - sets of points, representing a function
- y = f(x). The points are not equally-spaced along the
- x-axis. Some data sets can contain a very sharp peak
- around an x-value that varies from dataset to dataset.
- The datasets are probability distributions.
- The shape of the function changes slightly from
- dataset to dataset; the nature (height, width, location) of the
- peak also changes. There is also some noise.
-
- I need to extract a FEW coefficients
- to describe the significant parts of the function, to
- enable further investigation - sort of looking for
- a pattern dependency of some kind.
-
- APPROACH:
- I tried cubic splines - that gives many coefficients, and
- requires some personal inspection of each data-set.
-
- A Fourier analysis was better. The first few coefficients
- are very useful to track some low-frequency information from
- dataset to dataset.
-
- The problem of the peak remains. The Fourier (FFT) approach is
- not good at measuring the position and nature of the peak with
- a few parameters. I am now thinking in terms of a hybrid method -
- use the FFT for some low-frequency stuff and study the
- peak separately. Ideally, the algorithm should be robust and
- should and detect and locate a peak, and branch off accordingly.
-
- This is in the early stages of a project. So, it is
- acceptable to learn and apply a new technique altogether,
- and maybe change the dataset creation process.
-
- Any suggestions?
-
- Thank you for your time.
-
- -Shankar
-
-