Interesting Observation on interpolate2d from numpy vs RBF from scipy.interpolate

So at work I have to convert a function from matlab interpolate2d to python. But there is no direct translation or rather, some similar but not identical implementation so I had to look it up. After running some unit tests and seeing some very bizarre value later, I stumbled across this insightful experimentataion post https://stackoverflow.com/questions/37872171/how-can-i-perform-two-dimensional-interpolation-using-scipy

After running some tests, I am seeing somewhere between 0.1 to 1% deviations between matlab vs python implementation. No big deal I thought. then, I realized that the input to matlab vs python was different, hence why the big deviation.

I figure I would apply the input correction to see again, BUT, i mad the mistake of incorrectly injected a wrong value in one of the input to python to construct the 2d interpolant. THIS is where things gets interesting, despitely WILDLY wrong value injection, the RBF moved from 0.1% error to 0.4% error. While interpret2D jumped from originally 1% deviation to about -800% ish…

Single outliers out of the 162+ pairs of points used to model the manifold seems to have a far far greater impact to interpolate2d than RBF.