forum.alglib.net

Hello Sergey.

I use your library to aproximate probability distribution. I don't know the kind of distribution, thus prefered to use penalized cubic spline aproximation. I know, that the value of function is always greater than zero, but in case of luck of data or on the range boundaries the aproximated function sometimes become negative. How can i avoid this?

Thank you.

It is hard to make regression spline strictly positive. I think, one possible solution is to fit logarithm of your function, not the function itself. I.e. instead of S(x) (which is value of regression spline) you will have to calculate exp(S(x)) in order to get values of your function.