While spherical waves do not give perfect-reconstruction at half-wavelength overlap-add, one wavelength out, we see from the dB plots in Fig.19 that the error is quite small in audio terms. The ``ripple'' stays small down to well under wavelength before source proximity effects increase the ripple significantly. Note, incidentally, that we are ignoring the reactive ``mass'' component of the point-source field near its center, which is non-propagating (Morse and Ingard, 1968).
The ripple reduces as we get farther out from the array ( ) because the interpolation kernels expand, giving more overlap. (Increase the number of sources by the same factor to see this more clearly without additional array-truncation error.) Since we normally listen to arrays at some distance away, we see that a larger issue than sampling density is array extent; looking in the direction a wave is coming from, we should see plenty of samples surrounding the point that our ``look direction'' intersects along the array. The commonly used surrounding ring/sphere architectures (often set up for ambisonics or VBAP) are excellent in this respect: Every look direction has samples uniformly about/around it. Linear/planar arrays are at a disadvantage because they must be truncated or windowed, limiting the virtual stage view.
http://arxiv.org/abs/1911.07575
.