griddedInterpolant bug

There is a bug with the griddedInterpolant function with nearest and linear methods. If the query set contains any out-of-range values, the entire evaluated set is NaN, not just the results corresponding to the out-of-range queries.

In contrast, the griddedInterpolant with cubic interpolation method returns NaN for only out-of-range queries and correctly evaluates the in-range queries.

This is an problem for me because I'd like to use the linear method. I'm currently filtering out the out-of-range values from the query before processing, but the best solution would be to get the linear griddedInterpolant to work as it should. How should I fix it?

This bug is demonstrated by the following example script:

[X,Y,Z] = ndgrid(1:10,1:10,1:10);
V = X.^.5 + Y.^.5 + Z.^.5;

nearestInterp = griddedInterpolant(X,Y,Z,V,'nearest');
linearInterp = griddedInterpolant(X,Y,Z,V,'linear');
cubicInterp = griddedInterpolant(X,Y,Z,V,'cubic');
splineInterp = griddedInterpolant(X,Y,Z,V,'spline');

[Xq,Yq,Zq] = ndgrid(0:.5:10,0:.5:10,0:.5:10);

Vq_nearest = nearestInterp(Xq,Yq,Zq);
Vq_linear = linearInterp(Xq,Yq,Zq);
Vq_cubic = cubicInterp(Xq,Yq,Zq);
Vq_spline = splineInterp(Xq,Yq,Zq);

Thanks, Nick

[Honglei Chen]

Hi Nick, which version of MATLAB are you using, I tried this in R2012a and couldn't reproduce it.

[Marc Lalancette]

I know this is an old thread but I also ran into this problem, using version 2011b. Is it possible to get this fixed without paying hundreds of dollars for a new version?

[Matt J]

The best solution, other than upgrading, is probably to create your own wrapper that separates out-of-bounds values from inbound values and only calls griddedInterpolant on the latter. You would do that only when the linear interpolation method is active, of course.

Answer 1

[Sean de Wolski]

This behavior was fixed in R2012a. If you are current on SMS, you can download R2012a or R2012b for free.

[Wen]

I'm seeing the same problem for spline in R2014a.

[Sean de Wolski]

Wen, can you provide a minimal working example?

[Wen]

Sorry, I realized the problem I found was a bit different then the topic question. It has to do with ignoring NaN values in V when making the interpolated surface. If nearest, linear, or cubic are used, I can get the values of the interpolated surface as long as it's not close to the x, y, z values at which V is NaN. However, the spline option makes everything NaN instead of only at those x, y, z values.

I added one line to the one Nick gave:

[X,Y,Z] = ndgrid(1:10,1:10,1:10);
V = X.^.5 + Y.^.5 + Z.^.5;
V(10, 10, 10) = NaN;
nearestInterp = griddedInterpolant(X,Y,Z,V,'nearest');
linearInterp = griddedInterpolant(X,Y,Z,V,'linear');
cubicInterp = griddedInterpolant(X,Y,Z,V,'cubic');
splineInterp = griddedInterpolant(X,Y,Z,V,'spline');
[Xq,Yq,Zq] = ndgrid(0:.5:10,0:.5:10,0:.5:10);
Vq_nearest = nearestInterp(Xq,Yq,Zq);
Vq_linear = linearInterp(Xq,Yq,Zq);
Vq_cubic = cubicInterp(Xq,Yq,Zq);
Vq_spline = splineInterp(Xq,Yq,Zq);

[Matt J]

That's not a bug, I don't think. With splines, the piecewise polynomial coefficients used to interpolate between the V(i) depend on all V(i) simultaneously. Any NaNs in the given data set can be expected to propagate everywhere.

[Wen]

Ah ok, makes sense. I see cubic has more NaNs in its interpolated values than linear, and it depends on 4 points (or is it 3?) rather than 2. Okay.