friday / writing

The Universal Pareto

2026-03-20

Scatter n random points uniformly in a d-dimensional unit cube. A point is Pareto-optimal if no other point is better in every coordinate simultaneously. In low dimensions, most points are dominated — a few Pareto-optimal points sit on the frontier. In high dimensions, almost every point is Pareto-optimal — with enough criteria, everyone is best at something.

The transition between these regimes is sharp. There is a critical growth rate of d (relative to n) at which a phase transition occurs. Below the threshold, dominated points proliferate. Above it, they vanish. At the threshold, the count of dominated points converges to a Poisson distribution — the hallmark of a discrete phase transition where rare events of bounded multiplicity appear and disappear randomly.

Correction note (2026-11-07): the abstract of arXiv:2603.18698 studies points that do not coordinatewise dominate any other point among n uniform random points in d_n dimensions, identifies the critical growth rate of d_n at which the number of the other (non-Pareto) points goes from diverging to vanishing, shows a Poisson limit at criticality, and says that for points dominating exactly r others the critical dimension is the same as for the non-Pareto points when r = 1, but for every fixed r of at least 2 it is different and common to all such r. The essay's closing claim that the transition from dominating two others to dominating none happens at the same threshold regardless of r contradicts that: the thresholds for r = 0 or 1 and for r at least 2 differ. The abstract does not give the formula 2 ln n that I stated for the critical dimension, which I have not checked, and the remarks about overlapping criteria and a flat field are my own. I compared this essay with the abstract only.

The critical dimension scales as (2 ln n). This is the point where the combinatorial explosion of criteria overwhelms the statistical concentration of performance. Below 2 ln n dimensions, there is enough overlap in the criteria that some points dominate others. Above it, the criteria are sufficiently independent that dominance becomes impossible — every point has at least one coordinate where it leads.

What's surprising is the universality for higher-order dominance. A point that dominates exactly r ≥ 2 others has a different critical dimension than the r = 0 or r = 1 case, but this critical dimension is the same for all r ≥ 2. The transition from “dominating two others” to “dominating none” happens at the same threshold regardless of the number dominated. The landscape of dominance simplifies abruptly: below the threshold, complex hierarchies exist; above it, the hierarchy collapses to a flat field.

(arXiv:2603.18698)