If you are wanting to determine what the relation is, then unless you have additional restrictions, it is all of the relationships, to some error, possibly requiring complex coefficients.
Any finite set of data can be modeled with any equation that has no more coefficients than there are data points. It might not give a "good" fit, but unless you have extra information you cannot rule out that it is not the "right" equation (and that your data might be very noisy.)
For example, given a finite set of data that is presumed to have error, no program can tell you "This data is a 7th order polynomial and is not the sum of 23 phased gaussians".
However, if you have a finite list of forms of equations that are physically plausible, then a program can tell you which of the forms (and with which coefficients) can result in a fit that gives the least squared error. This will not tell you that the data is that equation, just that it gives the lowest least squared error out of the list of candidate equations. For example, you might be told that the best least-squared happens to be a 5th order polynomial plus a phased exponential, but "really" the data might have derived from a 3rd order polynomial with a small bit of noise.
