In leader-based consensus protocols, the leader must distribute each block to all processors. Without erasure coding, the leader sends full copies — a data expansion rate equal to the number of participants. Erasure coding reduces this by sending fragments that collectively reconstruct the block. The data expansion rate — total data sent divided by payload size — determines how close throughput approaches the network bandwidth limit.
Note (2026-11-12): checked against the paper's full text (Lewis-Pye and O'Grady). The numbers match: no protocol with two-round finality can go below a data expansion rate of about 2.5 (matched by E-Minimmit and Kudzu), and the three-round protocols approach 1 under favourable conditions and fall back to about 1.33 (Carnot 1, n at least 4f+1) and about 1.5 (Carnot 2, n at least 3f+1) under adversarial ones. Two details are left out. Carnot 1 solves Extractable state machine replication, a weaker task, while Carnot 2 solves full state machine replication at the optimal fault tolerance, which the paper presents as the trade-off between them. And the factor of 2.5 is the favourable-case gain; under adversarial conditions the gain over the 2.5 bound is a factor of about 1.7 to 1.9. I did not check the essay's remark that the extra round 'is not delay': the paper treats three-round finality as a protocol with a further voting round and compares latencies with other protocols, and the essay's sentence is its own gloss.
Protocols with two-round finality (one voting round) cannot achieve a data expansion rate below approximately 2.5 (arXiv:2603.11797). This bound is tight — existing protocols match it. The constraint is fundamental: with one voting round, the leader must ensure that any honest quorum can reconstruct the block, and the overlap requirements between quorums force redundancy.
Three-round finality breaks through. With two voting rounds, the data expansion rate approaches 1 — essentially no overhead. The mechanism: the second voting round provides a recovery path. The leader attempts an aggressive erasure code (low redundancy, high efficiency). If reconstruction fails — because Byzantine processors withheld their fragments — the second round detects the failure and falls back to a more conservative code. Consistency is never compromised because the protocol doesn't commit until reconstruction succeeds.
Two protocols realize this. Carnot 1 requires n >= 4f+1 processors and achieves clean design with no additional fragment dissemination. Carnot 2 operates at optimal resilience n >= 3f+1 but needs extra fragments when Byzantine processors interfere. Under favorable conditions, both approach expansion rate 1; under adversarial conditions, they revert to 1.33 and 1.5 respectively.
One additional voting round buys a factor-of-2.5 improvement in bandwidth efficiency. The extra round is not delay — it is a recovery mechanism that makes aggressive coding safe.