An Exact and Efficient Sampler for Dynamic Discrete Distributions (2506.14062v1)

Abstract: Sampling from a dynamic discrete distribution involves sampling from a dynamic set of weighted elements, where elements can be added or removed at any stage of the sampling process. Although efficient for static sets, the Alias method becomes impractical in dynamic settings due to the need to reconstruct the sampler after each update, which incurs a computational cost proportional to the size of the distribution, making it unsuitable for applications requiring frequent insertions, deletions, or weight adjustments. To address this limitation, different approaches have been studied, such as the Forest of Trees method and the BUcket Sampling (BUS) method. However, all previous methods suffered from numerical issues which can bias the sampling process. In this paper, we describe EBUS (Exact BUcket Sampling), the first exact algorithm with $O(1)$ sampling and update cost. The sampler can be updated by base-$b$ numbers with bounded precision and exponent, and sample the distribution of its elements exactly and efficiently. We provide also a state of the art implementation of the method using IEEE 64-bit floating point numbers which we empirically show to be more efficient than several implementations of previous inexact methods.