Say we have samples $x_i$ from an unknown distribution $F(x)$. We want to know the number of samples $n$ such that we can say the distribution is "well sampled". Generally we need some sort of metric that is proportionally to "well-sampled"ness, and then we also need to be able to set some threshold that is "sufficient".

One idea I had for a metric is to take the average nearest-neighbor-distance ($d \equiv \langle \, \min(x_i - x_j)_j \, \rangle _i$ for $i ≠ j$) and divide by the standard deviation, $d/\sigma$. This ratio is always less than unity, but how to determine a threshold value is unclear.

If we were to construct an inferred estimate of the distribution, $f(x)$, e.g. using a KDE, then we could calculate a goodness-of-fit statistic (e.g. ks-test) between the CDF of $f(x)$ and that of the $x_i$ themselves. This approach has a more obvious way of deciding on a threshold value, but it requires numerous decisions to be made about how to construct the estimated $f(x)$ (e.g., for a KDE, a kernel and a bandwidth).

Are there any standard approaches or heuristics in the literature for this type of problem?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.