Theoretically, you can choose this distance arbitrarily. In practice, you need to be more careful. If the distance is too small, the support regions around each node will never intersect with each other. Thus, the equations effectively decouple from eachother and no meaningful answer can result from this. In the other extreme, choosing the distance too large means that a lot of nodes couple to each other, making the linear system dense and more costly to solve. Balancing the need to guarantee sufficient coupling while promoting sparsity is non-trivial.
Ideally, you want to choose the distance large enough to have the support around each node overlap with the neighbors' support regions (as we usually observe with the traditional galerkin finite element method). This really depends on how your points are distributed. If you have points that roughly evenly throughout the domain (e.g. poisson disk sampling), you may want to use the average nearest neighbor distance for this purpose. If you have the internodal distances stored in a matrix, it is very easy to calculate. For each node $x_i$, find the nearest neighbor $x_j$ whose distance $d^*_i=d_{ij}$ from $x_i$ is smallest compared to all the other nodes. If you have n nodes, then simply take the average of all the minimum distances $\frac{\sum_i d^*_i}{n}$. In practice, you may want use a distance slightly larger than this to overcome any slight biases in the point distribution.
However, you need to be cautious about using this approach for two reasons:
- If you don't already have the internodal distances, you'll need to compute them and that can be costly! ($O(n^2)$ operations).
- If your points are clustered into groups, then the average nearest neighbor distance can also lead to decoupled systems.
In general, you'll need to analyze the statistics of your spatial distribution to make the wisest choice of your support size. No method works well for all possible distributions.