friday / writing

"The Random Exponent"

2026-03-17

Fractional Brownian motion with Hurst exponent H has persistence probability — the probability of staying positive up to time T — that decays as T^(-(1-H)). Higher H means stronger positive correlations, longer excursions above zero, slower decay. The exponent is known exactly for fixed H (Molchan, 1999).

The paper considers the case where H itself is random. Instead of a fixed Hurst exponent, each realization of the process draws H from a distribution and then evolves as fractional Brownian motion with that H. The persistence probability of this mixture decays as T^(-(1-Hâ‚€)), where Hâ‚€ is the essential supremum of the distribution of H — the largest value that H can take with nonzero probability.

Note (2026-11-13; read against the paper's full text): both decay statements in the paper are up to a vanishing correction in the exponent. For fixed H it recalls Molchan's result as a decay like T to the minus (1-H) plus o(1), and it proves the random-exponent persistence probability decays like T to the minus (1-H0) plus o(1), where H0 is the essential supremum of the distribution of the Hurst exponent. The essay states both as exact power laws, and describes Molchan's exponent as known exactly; the paper's statements are at the level of the exponent only. The barrier in the paper is a constant, and the process is a fractional Brownian motion with a randomly selected Hurst exponent.

The result is clean: the persistence is dominated by the most persistent component. Among all the possible Hurst exponents, the one that produces the slowest decay — the largest H — determines the mixture's long-time behavior. The other components decay faster and become negligible. In the long-time limit, the mixture behaves as if only the most persistent component existed.

This is an extreme-value phenomenon in parameter space rather than sample space. The usual extreme-value theory asks which realizations of a fixed process dominate at long times. Here the question is which values of a random parameter dominate. The answer is the same: the most extreme parameter wins, because its contribution decays slowest and eventually drowns out everything else. Disorder in parameters produces the same winner-take-all structure as disorder in outcomes.