| How to do surface fit? [message #10790] |
Tue, 27 January 1998 00:00  |
Gang Chen
Messages: 4
Registered: January 1998
|
Junior Member |
|
|
Hi,
I am trying to fit a surface (probably 2nd degree polynomial:
z=a+b*x+c*y+d*x*x+e*xy+f*y*y) on a two dimensional grids with IDL. But it
seems to me that SFIT and SVDFIT would not serve my purpose: SFIT only
generates the values on those grids without giving the coefficients, and
it uses different format of polynomial (z=sum of k_i,j * x^i * y^j); While
SVDFIT works only for curve fit. Does anybody have any idea on how to do
this by simply calling some existing functions instead of writing more
complicate code?
Many thanks,
Gang Chen == gang@cochise.biosci.arizona.edu
|
|
|
|
| Re: How to do surface fit? [message #10973 is a reply to message #10790] |
Thu, 29 January 1998 00:00  |
Evilio del Rio
Messages: 17
Registered: December 1997
|
Junior Member |
|
|
On Wed, 28 Jan 1998, Gang Chen wrote:
>> Use the KX keyword to SFIT, it will return an array of coefficients, in
>> your case:
>>
>> IDL> zfit = SFIT(z,2,KX=k)
>> IDL> help,k
>> COEFF FLOAT = Array[3, 3]
>>
>> They are calculated with the convention X= 0,...,NX and Y=0,...,NY
>> (so you must to make the needed conversion).
>
>
> HI Evilio,
>
> I really appreciate your help on this. It seems that it works the way I
> wanted. But I have another question: Since the default grids are set with
> X= 0,...,NX and Y=0,...,NY in the source code for SFIT, how can I do for a
> non-uniform grids (e.g., X=0, 1, 4, 5, 9, ..., and Y=3, 4, 7, 11, ..)?
> Do I have to resort to interpolation so that values at uniform grids can be
> distributed?
>
> Many thanks,
> Gang Chen == gang@cochise.biosci.arizona.edu
>
Humm... Well, in fact you could use multivarite linear fitting, this is
fine if you have low-degree polynimials. You should use the function
REGRESS
but you must prepare before the independent variables. Let's call u to the
REGRESS dependent variable (they call it Y in the doc) and v the set of
independient variables (X in the doc). The fit REGRESS does is:
u = const+a0*v0+a1*v1+...+an-1*vn-1
if you compare with your problem:
z = a+b*x+c*y+d*x*x+e*xy+f*y*y
you must set:
u <=> z
const <=> a
a0 <=> b, v0 <=> x
a1 <=> c, v1 <=> y
a2 <=> d, v2 <=> x*x
(etc...)
This is not strictly correct since the "independent" variables are not (x,
is strongly correlated with x^2, x*y, etc.) but it will work if your data
are not very abnormal. The correct procedure, however, should be an
improved SVDFIT so that it can accept multiple independent variables since
the values of "X" are only used to calculate the user-supplied function
FUNC (see IDL doc on SVDFIT and "Numerical Recipes in FORTRAN", ch. 15.4,
pag. 675).
Maybe we could suggest this change to RSI people.
Cheers,
____________________________________________________________ ________
Evilio Jose del Rio Silvan Institut de Ciencies del Mar
E-mail: edelrio@icm.csic.es URL: http://www.ieec.fcr.es/~evilio/
"Anywhere you choose,/ Anyway, you're gonna lose"- Mike Oldfield
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
