Fast Probabilistic Consensus (FPC)

• FPC is a fully distributed binary consensus protocol that uses iterative, randomized polling to achieve agreement even in adversarial, Byzantine-rich environments.
• The protocol leverages weighted voting based on node reputation via a Zipf-law distribution to mitigate adaptive adversary strategies and improve security.
• FPC employs rapid drift amplification and minimal peer-to-peer communication, making it scalable and effective for large networks such as the IOTA cryptocurrency.

Fast Probabilistic Consensus (FPC) is a fully distributed binary consensus protocol designed for efficiency and robustness in permissionless, adversarial environments with potentially large fractions of Byzantine nodes. FPC achieves consensus by iterative, decentralized polling with an external source of randomness that disrupts adversarial attempts to stall or split the global state, thereby enabling agreement with high probability under well-defined parameter regimes. The protocol has also been extended to incorporate weighted voting power according to node reputation, notably employing a Zipf-law distribution, with further enhancements targeting the detection and mitigation of adaptive ("berserk") adversary strategies. FPC has been adopted and extensively analyzed in the context of the IOTA cryptocurrency, but the results generalize to any scenario with authenticated, reputation-weighted nodes and reliable peer sampling (Müller et al., 2020, &&&1&&&, Popov et al., 2019).

1. Protocol Structure and Update Rules

FPC operates over a set of $N$ nodes, each maintaining a binary opinion $s_i(t) \in\{0,1\}$ at round $t$. The protocol iterates as follows:

• Peer Sampling: At each round $t$, each node $i$ selects a quorum $C_i(t)$ of $k$ peers, either uniformly at random (unweighted) or according to a distribution reflecting node reputation or "mana."
• Opinion Aggregation: Node $i$ computes the (possibly weighted) empirical mean $\eta_i(t+1)$ of the sampled peer opinions:

$$\eta_i(t+1) = \frac{1}{k_i(t)} \sum_{j\in C_i(t)} s_j(t)$$

In the weighted regime (reputation $m_j$ for peer $j$), the update is

$$\eta_i(t+1) = \frac{\sum_{j\in C_i(t)} g(m_j) \cdot s_j(t)}{\sum_{j\in C_i(t)} g(m_j)}$$

with $g(m) = 1$ for fairness (Müller et al., 2020).

• Random Thresholding: The protocol employs a round-indexed sequence of thresholds $\{U_t\}$:

$$U_t \sim \mathrm{Uniform}\left([ \beta, 1-\beta ]\right), \quad 0 \leq \beta \leq 1/2$$

The threshold for $t=0$ is a constant $\tau\in(0,1)$.

• Update and Finalization: Node $i$ updates its local opinion:

$$s_i(t+1) = \begin{cases} 1 & \text{if } \eta_i(t+1) > U_t \ 0 & \text{if } \eta_i(t+1) < U_t \ s_i(t) & \text{otherwise} \end{cases}$$

A node terminates and finalizes its opinion if it remains unchanged for $\ell$ consecutive rounds. The round cap is $\mathrm{maxIt}$.

The protocol avoids all-to-all communication—each node samples only $O(k)$ peers per round—and requires only a common, unpredictable random beacon for synchronization of $\{U_t\}$.

2. Byzantine Models and Security Parameters

FPC differentiates between several Byzantine adversary models, parameterized by the fraction $q$ of total nodes (or reputation) under adversarial control (Popov et al., 2021, Popov et al., 2019):

• Cautious adversary: Replies must be consistent across all queries in a round.
• Berserk adversary: Arbitrary, inconsistent replies to different queriers in a round.
• Semi-cautious adversary: Possible omission of replies, but all sent replies per round must be consistent.

Resilience thresholds are as follows:

• Cautious: Protocol achieves safety and liveness with high probability if $q < 1/2$ (strictly $q < \beta \lesssim 1/2$).
• Berserk: Secure and convergent for $q < 1/3$ in the unweighted case (since both $q < \beta$ and $q < 1-2\beta$ must hold).
• Semi-cautious: Secure for $q < (3-\sqrt{5})/2 \approx 0.38$.

The protocol is proven to terminate in $O(\log N)$ rounds with total $O(N\log^2 N)$ messages when $k = C\log N$ peers per round (Popov et al., 2019).

3. Role of Randomization and Drift Amplification

Crucial to FPC’s resilience is the use of global round randomization for the threshold $U_t$:

• Symmetry Breaking: The unpredictability of $U_t$ prevents Byzantine coalitions from stalling the system at the metastable midpoint ($\eta\approx 0.5$), which is known to block majority-dynamics protocols with fixed thresholds.
• Rapid Drift: When $U_t$ is drawn outside the adversarially influenced narrow band, a strong drift toward either $0$ or $1$ is induced among the honest nodes (Popov et al., 2021). Once a modest supermajority forms, successive rounds of independent randomization rapidly amplify this into unanimous agreement.

The probability of protocol completion in $m_0+\ell$ rounds is lower-bounded by expressions such as

$$P[\text{agreement}] \geq 1 - c_0 N \ell e^{-c_1 k} - e^{-c_2 m_0 \log k} \quad\text{(cautious case)}$$

for constants $c_0,c_1,c_2>0$ (Popov et al., 2021).

4. Weighted Voting and Protocol Enhancements

FPC naturally extends to settings in which voting power is nonuniform, modeling node reputation $m_i$ according to a Zipf distribution:

$$m_{(n)} = C n^{-s}, \quad C^{-1} = \sum_{n=1}^{N} n^{-s}$$

where $s\geq 0$ is the Zipf exponent. Peers are sampled and votes counted proportionally to $m_i$ (Müller et al., 2020).

Improvements proposed for the weighted FPC include:

• Fixed Threshold in Final Rounds: Replacing random $U_t$ with $\tau^*=0.5$ in the last $\ell_2$ rounds reduces the finalization error rate.
• Self-bias in Opinion Update: To avoid excessive flipping by small-reputation nodes, set

$\eta_i(t+1) := m_i s_i(t) + (1-m_i) \eta^*_i(t+1)$

with $\eta^*_i$ the peers’ mean.

• Distinct Peer Enforcement: If a peer is sampled more than once, only a single vote is counted per round; raw sample size is increased adaptively to assure near-constant effective quorum size.

Simulations show that with these tweaks, FPC is empirically robust to Byzantine reputational weight up to $q \approx 0.25-0.30$ for failure probability $10^{-3}$, a substantial improvement over the unweighted threshold of $q\approx 0.15$ (Müller et al., 2020).

5. Adversary Detection: Berserk Strategy and "v-list" Auditing

The protocol includes a defense against berserk adversaries:

• v-list requests: With small probability $p_B$ per poll, an honest node requests from the respondent its previous round's quorum and observed opinions ("v-list").
• Cross-audit and gossip: If a node is observed to return inconsistent v-lists (proving equivocation), honest nodes can cryptographically confirm and collectively exclude it.

The probability of catching a berserk node per round increases quadratically in $p_B$ and the adversary’s queried-mass; for typical network parameters detection probability per round exceeds $0.23$, ensuring rapid adversary isolation (Müller et al., 2020).

6. Empirical Performance and Topological Dependencies

Large-scale simulations (Müller et al., 2020) validate that FPC failure probability decays exponentially in both $N$ and $k$, with empirical results:

• $N$ grows from $500$ to $5000$: failure drops from $10^{-2}$ to $<10^{-4}$.
• $k$ increases (e.g., from $10$ to $40$ at Zipf $s=0$, $q=0.25$): failure plummets from $10^{-2}$ to $10^{-6}$.

Empirical studies of FPC in IOTA-style networks (Mamache et al., 2021) show that, absent high connectivity and small-world structure in the peer-to-peer overlay, FPC's convergence rate drops precipitously even under limited adversarial activity. For instance:

• Grid and torus topologies: negligible tolerance to even $10\%$ Byzantine participation.
• Watts–Strogatz small-world graphs with degree $K=10$–$20$: tolerates up to $10\%$ adversarial share with $\sim80\%$ agreement; performance collapses at $20\%$.

No analytic convergence bounds are given for these topologies, and the observed boundaries are determined empirically.

7. Implementation Guidelines and Applicability

To obtain optimal convergence and security, the following recommendations are provided (Müller et al., 2020):

• Quorum size: Chosen per desired failure target via $k \approx \frac{\log\epsilon^{-1}}{I(1/2)}$, with $I(\cdot)$ the binary KL divergence.
• Thresholding: Initial $\tau$ should match the initial majority; randomization width $\beta \in [0.2,0.4]$.
• Termination rounds: $8$–$12$ stable rounds for finalization.
• v-list audit: $p_B \in [0.05,0.1]$ achieves rapid berserk detection.

FPC's independence from IOTA-specific cryptography makes it suitable for any system with verifiable, nontransferable reputation and authenticated peer-sampling. Its communication complexity (per binary consensus) is $O(N \log^2 N)$, with protocol duration typically $O(\log N)$ rounds for practical parameter settings (Popov et al., 2019).

---

FPC fundamentally leverages randomized dynamics to defeat information symmetry exploited by Byzantine actors, requires only lightweight peer-to-peer operations, and achieves scalable, robust consensus under realistic network and adversarial models. Its weighted extension and adversary-detection mechanisms further enable operation in settings where voting power is highly skewed or actively targeted by sophisticated attacks (Müller et al., 2020, Popov et al., 2021, Popov et al., 2019).