The paper tackles a fundamental question in the mathematics of sampling: how "well-distributed" are random samples? Discrepancy is a way of measuring how evenly a set of points covers a space, with low discrepancy meaning the points are spread out in a structured, uniform way. Purely random samples are known to have relatively high discrepancy, meaning they tend to cluster and leave gaps. This paper asks whether random samples can be made nearly as good as the best possible point sets, called low-discrepancy sets, with only minor adjustments.
The main result is that yes, they can. Given any probability distribution in any number of dimensions, if you draw a random sample of n points, you only need to move a small fraction of those points on average to obtain a point set whose discrepancy is nearly as small as theoretically possible. The target discrepancy achieved is roughly a factor of a power of the logarithm of n divided by n, which is essentially optimal up to those logarithmic corrections. Crucially, this works for arbitrary probability distributions, not just the uniform distribution on a simple geometric shape.
The practical implication is that random sampling is not fundamentally at odds with good coverage properties. It is, in a precise sense, close to being well-distributed, even if it does not look that way on the surface. This bridges a conceptual gap between probabilistic sampling methods, which are easy to generate, and quasi-random or low-discrepancy sequences, which are carefully engineered for even coverage. The result suggests that one could potentially take a random sample and correct it with minimal effort to get the accuracy benefits usually associated with quasi-random methods.