Author: Lucas Luhur Last updated: 2026-08-01

Summary

  1. Log-normal link latency and its parameters
  2. Mixnet on the Bayesian inference attack and its design choices

Parameters (current):

group parameter symbol value
consensus nodes $N$ 1000
active-slots coefficient $f$ $1/30$
Pareto shape (stake inequality) $k_{\text{stake}}$ $4/3$ ($G = 0.60$, Cardano-calibrated)
epoch length $T$ 388,800
network / links connectivity (random regular graph) $C$ 8
channel noise $\rho_{\text{net}}$ 0 (noise-free)
mean eccentricity $E[\text{ecc}]$ 5.0 hops
link law (every link, gossip graph included) $d$ $7.2\,\text{ms} + \text{LogNormal}$
broadcast latency (Dijkstra weighted eccentricity) $E[D_{br}]$ $\approx$ 0.251 s
mixing mix hops / layers $k$ 3
per-hop mix delay (mean) $1/\lambda_M$ swept, $0.0001 \to 1.0$ s
sender hold (mean) $1/\lambda_S$ coupled: $(1/\lambda_M)/\rho$
delay ratio $\rho = \lambda_S/\lambda_M$ 4 (target $\rho \geq 4$)
receiver delay — none ($\mathcal{R}$ does not delay $\Rightarrow$ $k$ stages)
cover / sender set cover count (fixed) count 19
statistics seed / repetitions 0 / 50 per cell

1. Single path mix model (lognormal)

1.1 The log-normal link law

AC-path lognormal single path mix model
$\mathcal{S}\to M_1$ ($d_i^S$) 0.0072 + lognormal
$M_1\to\cdots\to M_k$ ($D_M$) same law per link, summed
$M_k\to\mathcal{R}$ ($d_r^R$) same law
$\mathcal{R}\to\text{all}$ ($D_{br}$) Dijkstra, same law

The lognormal distribution of link latencies is used as it is:

  1. Mechanistic: it is the law of multiplicatively combined random effects ( https://academic.oup.com/bioscience/article-abstract/51/5/341/243981)
  2. Admissible: continuous and strictly positive
  3. Controllable: two parameters map one-to-one onto (mean, variance), so $\sigma_d$ , hence $\eta$, can be tuned.

The empirical paper: It measures real internet delay and fits gamma, lognormal and Weibull to it. Log-normal fits best, though none of the three passes the KS test, so it's a ranking between the candidates rather than a confirmation (https://etd.lib.metu.edu.tr/upload/1223155/index.pdf).

The information-theoretic paper: Maximum entropy. We assert three things about a link: it is positive, it has a typical size, and it has a spread. The log-normal is the least-committal law consistent with that (https://ideas.repec.org/a/eee/econom/v150y2009i2p219-230.html)

The model is $d = L + X$, $X \sim \text{LogNormal}(\mu, \sigma)$. All three numbers come from WonderNetwork ping data of 12 cities, across 3 continents. 66 city pairs.

value what it is
floor $L$ 7.2 ms the smallest measured one-way city-pair latency
mean 62.9 ms the ping population’s mean
sd 33.3 ms its standard deviation: the structural spread across node pairs

Moment-matched: The two log-normal parameters are chosen so the model reproduces the data's mean and standard deviation. The floor is a constant, so it shifts the mean and leaves the variance untouched: