Statistical Methods
Check Geometry Before Radial Basis Interpolation
Define the task
Radial basis interpolation can estimate values at new multidimensional coordinates, but coordinate geometry and scale matter. State what each coordinate represents and whether Euclidean separation is meaningful. Changing one axis from meters to millimeters without adjusting the model changes distances. Record the kernel, smoothing, polynomial degree and any shape parameter. Use independent validation points when possible; fitting the training coordinates alone does not establish useful prediction between or beyond them.
Carry out the check
With zero smoothing and a well-posed system, interpolation reproduces observed values. Positive smoothing relaxes exact fitting. Polynomial terms impose minimum-degree and rank requirements that depend on the kernel; inspect these requirements before choosing settings. Duplicate locations, poorly distributed coordinates or an unsuitable shape parameter can make the system unstable. Global fitting requires memory that grows quadratically with the number of points; a neighbor-based calculation changes the local fitting problem and should be documented. Avoid extrapolation claims without a justified scientific model.
Worked hypothetical example
Consider a fictional planar response f(x,y)=1+2x+3y observed at (0,0), (1,0) and (0,1), giving values 1,3,4. A compatible interpolator with a linear polynomial term reproduces the plane and gives 3.5 at (0.5,0.5) in exact arithmetic. If all training points instead lie on one straight line, the two-dimensional linear polynomial design lacks full rank. This example checks polynomial reproduction and geometry, not the accuracy of a real measured surface.
Method checked 16 November 2026 against SciPy documentation: RBFInterpolator.