The paper studies how well a probability distribution spread across space can be approximated by a finite set of representative points, a problem called quantization. The authors develop a new way to measure the complexity of a distribution using a tree-based zooming process: start with a coarse grid of boxes covering the space, then keep subdividing any box that contains a meaningful chunk of the distribution's mass, stopping when boxes become too small to matter. The average depth at which this process halts turns out to encode the so-called geometric-mean quantization dimension, a number that describes how efficiently the distribution can be compressed. This connects an abstract geometric quantity to a concrete, computable algorithm.
A central finding is that this dimension is closely related to a family of statistical summaries called Renyi spectra, which track how mass is distributed across different scales. Specifically, the dimension exists as a well-defined single number precisely when a certain weighted sum over all boxes converges in a particular limiting sense. When this happens, the dimension can be recovered from that limit. The paper also shows that the geometric-mean dimension sits neatly between two other well-known concepts, entropy dimension and classical quantization dimension, unifying them within a single framework. Sharp inequalities and variance-based identities precisely quantify how far apart these two bounding dimensions can be.
The paper also examines what happens when the local behavior of the distribution, meaning how mass concentrates near individual points, follows a statistical pattern rather than being perfectly regular. When these local statistics obey a law of large numbers in a suitable sense, the upper and lower versions of the dimension coincide and can be read off as the arithmetic and harmonic means of the limiting distribution of local scaling exponents. The authors construct explicit examples, including mixtures of coin-flip processes and systems that switch between different scaling regimes, to show that various combinations of convergence and non-convergence are genuinely possible, confirming that their conditions are tight and not merely technical assumptions.