SG++

Here we demonstrate how to refine a grid.
As a refinement indicator, we take the surpluses of the grid points directly. We start with a regular sparse grid of level 3 with linear basis functions and refine five times. In each refinement step, we refine the grid point with the highest absolute surplus.
The following example interpolates the (nonsymmetric) function f : [0,1]x[0,1] > R, f(x,y) := 16 (x1) * x * (y1) * y
The number of grid points is printed in each iteration. After refinement, the surplusses have to be set for all new grid points, i.e., the alphaVector has to be extended.
For instructions on how to run the example, please see Installation and Usage.
function to interpolate. This is a twodimensional parabola.  nonsymmetric(!)
create a twodimensional piecewise bilinear grid
create regular sparse grid, level 3
create coefficient vector with size corresponding to the grid size. Initially, all the values are set to zero.
Obtain function values and refine adaptively 5 times
Each time, we have to hierarchize the grid again, because in the previous interation, new grid points have been added.
Refine a single grid point each time. The SurplusRefinementFunctor chooses the grid point with the highest absolute surplus. Refining the point means, that all children of this point (if not already present) are added to the grid. Also all missing parents are added (recursively).
Extend alpha vector (new entries uninitialized). Note that right now, the surplus vector has the correct size again, but the values of the new points are set to zero. The correct surplus values will be inserted in the next iteration of the refinement loop.
This results in the following output:
dimensionality: 2 number of initial grid points: 17 length of alpha vector: 17 refinement step 1, new grid size: 21 refinement step 2, new grid size: 24 refinement step 3, new grid size: 27 refinement step 4, new grid size: 29 refinement step 5, new grid size: 33
There are clearly more efficient approaches than to set the function values for all grid points and to hierarchize the whole grid each time. But this works even where no efficient alternatives are available and suffices for demonstration purposes.
This use of the SurplusRefinementFunctor takes as arguments the coefficient vector (it doesn't have to be the coefficient vector, it could be something modified!) and the number of grid points to refine (if available). It bases its refinement decision on the absolute values of the vector's entries, choosing the largest ones. Other refinement functors are available or can be implemented.