The Carnot Bound: Limits and Possibilities for Bandwidth-Efficient Consensus

Computer scientists Andrew Lewis-Pye and Patrick O'Grady have established a fundamental limit, dubbed 'The Carnot Bound,' for bandwidth-efficient consensus in distributed systems. Their paper proves that leader-based protocols with 2-round finality—common in many blockchain and database systems—are fundamentally limited to a data expansion rate of approximately 2.5. This means the leader must send 2.5 times more data than the actual payload, creating a hard bottleneck on throughput. The finding mathematically validates the performance ceiling seen in existing protocols like HotStuff and explains why simply improving network hardware has diminishing returns.

On the positive side, the researchers demonstrate that this bottleneck can be shattered by adding a third round of voting. They prove that 3-round protocols can theoretically achieve a data expansion rate arbitrarily close to 1, meaning nearly 100% of the leader's outgoing bandwidth carries useful payload data. The key innovation is using the extra round as a recovery mechanism: leaders can initially send aggressively encoded data fragments using erasure codes and, if reconstruction by followers fails, safely fall back to sending more redundant data in the next round without compromising system consistency.

To realize this theory, Lewis-Pye and O'Grady present two concrete protocols. 'Carnot 1' offers a clean design for systems with $n \geq 4f+1$ processors (tolerating $f$ Byzantine faults) and requires no extra message overhead. 'Carnot 2' achieves optimal resilience with $n \geq 3f+1$, at the cost of some additional fragment dissemination when under attack. In practice, under normal conditions with a correct leader, both protocols allow throughput to scale almost linearly with the underlying network bandwidth. Under worst-case adversarial conditions, they gracefully degrade to safe expansion rates of ~1.33x and ~1.5x, respectively, still outperforming the old 2.5x limit.