Author: Lucas Luhur Last updated: 2026-07-24
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) | |
| network / links | connectivity (regular graph) | $C$ | 8 |
| channel noise | $\rho_{\text{net}}$ | 0 (noise-free, $\lambda\to\infty$; $T_e = d$) | |
| mean eccentricity ($C{=}8$, $N{=}1000$) | $E[\text{ecc}]$ | 5.0 hops | |
| baseline link delay (flat) | $d$ | 0.3 s: the homogeneous link; for the uniform law only (geo replaces it with continent pings + $\delta_{proc}$) | |
| mixing (AC path) | mix hops (layers) | $k$ | 3 |
| per-hop mix delay (mean) | $1/\lambda_M$ | swept axis (attribution) — $\text{Exp}(\lambda_M)$; operating point 1 s; $k$-hop total $\text{Gamma}(k,\lambda_M)$ | |
| sender hold (mean) | $1/\lambda_S$ | 0.25 s — $\text{Exp}(\lambda_S)$ (held) | |
| delay ratio | $\rho = \lambda_S/\lambda_M$ | 4 at the operating point (PoPETs 2024/0136 target $\rho \ge 4$) | |
| receiver delay | — | none ($\mathcal{R}$ does not delay $\Rightarrow$ $k$ stages) | |
| cover / sender set | attribution cover | count |
19 $\Rightarrow \lvert S\rvert = 20$ (fixed) |
| stake cover | $p_s$ | swept $\{0.02 \dots 0.9\}$ | |
| horizon | epoch length | $T$ | 388,800 slots (attribution + stake full-epoch) |
The model (recap): Every genuine proposal is routed down one shared path before the broadcast:
$$ \mathcal{S} \to M_1 \to \cdots \to M_k \to \mathcal{R} \longrightarrow \text{broadcast to } N \text{ nodes} $$
| AC-path | uniform | geo |
|---|---|---|
| $\mathcal{S}\to M_1$ ($d_i^S$) | $U(0,6.22)$, $\sigma_d \approx 1.75$ s | continent ping, $\sigma_d \approx 0.028$ s |
| $M_1\to\cdots\to M_k$ ($D_M$) | scalar $(k{-}1)d$ | $\sum$ inter-mix continent pings |
| $M_k\to\mathcal{R}$ ($d_r^R$) | $U(0,6.22)$ | continent ping |
| $\mathcal{R}\to\text{all}$ ($D_{br}$) | flat $d\cdot E[\text{ecc}] = 1.5$ s | weighted Dijkstra $\approx 1.25$ s |
| random hold $Z$ | $\text{Exp}(\lambda_S) + \text{Gamma}(k,\lambda_M)$ (geography-independent) | (same) |
uniform:
geo: Every AC-path link is a continent one-way ping $L(c(u),c(v))$ from the ping-calibrated continent mixture, with one continent per node, quenched per realisation:
Source: Real Internet Control Message Protocol (ICMP) round-trip times from wondernetwork.com/pings. 30 pings per ordered city pair, averaged, in ms. 12 cities, 4 per continent, chosen to spread within each continent
Continents and $\pi$: $\pi$ is the distribution of nodes across continents. the probability that a given node, and the first mix $c^*$, sits on each continent.
| continent | example cities | $\pi_c$ (renormalised) |
|---|---|---|
| Europe | Frankfurt, Madrid, … | 0.364 |
| North America | New York, Los Angeles, … | 0.364 |
| Asia | Tokyo, Hong Kong, Mumbai, … | 0.273 |
The continent-pair Round-Trip Time (RTT) matrix (symmetric): each entry a mean over 6 (intra) or 16 (inter) city pairs. The one-way link used in the model: $\text{RTT} \times 0.5$
| RTT (ms) | Europe | N. America | Asia |
|---|---|---|---|
| Europe | 30.3 | 113.2 | 182.3 |
| N. America | 113.2 | 38.8 | 187.6 |
| Asia | 182.3 | 187.6 | 89.2 |
A walkthrough: one message $\mathcal{S}\to M_1\to M_2\to M_3\to R$. Take a batch where the continents drew (quenched for the run): sender on Europe, mixes $M_1$ N. America ($= c^*$), $M_2$ Asia, $M_3$ Europe, receiver on Asia. One-way links are $\text{RTT}/2$ in seconds:
| leg | continents | value |
|---|---|---|
| $(d_i^S)$ $\mathcal{S}\to M_1$ | EU $\to$ NA | $113.2/2/1000 = 0.057$ s |
| $(D_M)$ $M_1\to M_2\to M_3$ | NA$\to$Asia$\to$EU | $(187.6 + 182.3)/2/1000 = 0.185$ s |
| $(d_r^R)$ $M_3\to R$ | EU $\to$ Asia | $182.3/2/1000 = 0.091$ s |
| $\mu(i,r) = d_i^S + D_M + d_r^R$ | $0.333$ s (deterministic) |
Then the random hold $Z$ (mean $0.25 + 3\times 1 = 3.25$ s at $1/\lambda_M = 1$).
The one part that is not an AC-path link is the final block broadcast: once $R$ holds the block it gossips it to all $N$ consensus nodes, and $D_{br}$, the mean time to reach the farthest node, is the denominator of latency $\ell$.
$E[\text{ecc}]$: the broadcast distance in hops**.** A source's eccentricity is the largest shortest-path distance from it to any other node, the hop count to the farthest node, i.e. how long the gossip takes to reach everyone; $E[\text{ecc}]$ averages it over sources.
$\Pr(\text{ecc}>\ell) = 1 - \big(1 - \Pr(L>\ell)\big)^{N-1}, \qquad E[\text{ecc}] = \sum_{\ell\ge 0}\Pr(\text{ecc}>\ell)$
$D_{br}$ : the broadcast latency. $D_{br}(s) = \max_{j\ne s}\text{dist}(s,j)$ is the source's weighted eccentricity, the same distance to the farthest node, but summing edge latencies instead of counting hops; $E[D_{br}]$ averages over broadcasters (‣).
The full $E[D_{br}]$: