Correct Interpolation on a pre-calculated grid that is a vector function of a vector?

I have calculated vector data, such that the data can be written and represented as , where each component is a function of all three coordinates; i.e. . I need to interpolate to be able to represent the inputs of a different vector anywhere, and have it be a vector output. This different vector is . Written out fully, you can see that each component of the vector M is an array which depends on all three x, y, and z coordinates. As a result, I can't simply use Interp3 three times. This means that I need the G's as a grid, so that I can query at points (Gx_q, Gy_q, Gz_q) and find their (x,y,z) inputs that give G.

Interp3 repeatedly gives errors of "Input is not a Meshgrid."

My vector data arrays (, , and ) are sized 31x31x31. I'm so lost; I've tried so many tricks that have failed. Any tricks? Will scatteredInterpolant() work?

The situation is analogous to this user's question (https://www.mathworks.com/matlabcentral/answers/408712-interpolating-scattered-3-dimensional-data), but distinctly different. The answers provided in () also are not quite related – something there is wrong, as well.

Previously, I've poorly or incorrectly explained my problem, and so those questions (and, as a result, answers) were phrased incorrectly. This is fixed.

To avoid confusion, I'll present the case where the M vector depends only on one coordinate, each, such that , which lets us write a scalar G as , and just do the process three times. The code would look like:

load('data.mat')
x = 0:1000:30000;
y = 0:1000:30000;
z = 0:1000:30000;

Gx_q = 721.0085; %These are example query points
Gy_q = 15.1313;
Gz_q = 45.9302;

Gx = x - squeeze(Mx(1,1,:))'; %These should be 1D arrays; for simplicity, in this case I would only need one column of each M_i array.
Gy = y - squeeze(My(1,1,:))';
Gz = z - squeeze(Mz(1,1,:))';

Gx_out = interp1(Gx,x,Gx_q,'linear','extrap');
Gy_out = interp1(Gy,y,Gy_q,'linear','extrap');
Gz_out = interp1(Gz,z,Gz_q,'linear','extrap');

Answer 1

[Matt J]

so that I can query at points (Gx_q, Gy_q, Gz_q) and find their (x,y,z) inputs that give G.

But there may exist no such point or many of them. Suppose for example that Gx=1,Gy=1,Gz=1 everywhere in x,y,z. Then if Gx_q=5, Gx_q=6,Gz_q=7 there will exist no (x,y,z) agreeing with this query point for any reasonable interpolator. If, on the other hand Gx_q=1, Gx_q=1,Gz_q=1 then all (x,y,z) agree with it.

The best I think you can do is solve for one coordinate (x,y,z) at which the interpolation of G is closest to (Gx_q, Gy_q, Gz_q) according to some error metric. That can be approached as an inverse problem as in the following example:

[x,y,z]=ndgrid(0:1000:30000);

Gx=x-Mx;
Gy=y-My;
Gz=z-Mz;

%Find best grid point (x,y,z) - later refine that with fminsearch

Err = abs(Gx(:)-Gx_q)+abs(Gy(:)-Gy_q)+abs(Gz(:)-Gz_q);

[~,i]=min(Err);

xyzInitial = [x(i), y(i), z(i)];

%Refine:

Fx=griddedInterpolant(x,y,z,Gx);
Fy=griddedInterpolant(x,y,z,Gy);
Fz=griddedInterpolant(x,y,z,Gz);


xyz=fminsearch(@(xyz)lookupError(xyz,Fx,Fy,Fz),  xyzInitial);

x_out=xyz(1);
y_out=xyz(2);
z_out=xyz(3);



    function  Err=lookupError(xyz,Fx,Fy,Fz)

         gx=Fx(xyz);
         gy=Fy(xyz);
         gz=Fz(xyz);

        Err= abs(gx-Gx_q)+abs(gy-Gy_q)+abs(gz-Gz_q);


    end

[Blenndrman]

Conceptually, I've you're right on the money – thank you! There is a problem, however, when I try to run it: Matlab prints the error message

"Error using griddedInterpolant/subsref

Invalid arguments specified in evaluating the interpolant."

Do you think this has to do with strange behavior in the actual arrays Mx, My, and Mz? Or is the fact that I'm stuck running Matlab R2015b the problem? (I had to export the function lookupError to its own script, for example).

[Matt J]

Try it now.

[Blenndrman]

It seems to work. Holy moly. This was several days' struggle. Thank you so much!

[Blenndrman]

Last question: To minimize the error function, would you think it smarter to package this entire function into its function to be passed to fminsearch, or would it be better to simply make an ndgrid with a smaller stepsize?

Thanks!