Coin Brief ENDE

Lean Ethereum study: folding loses to leanVM recursion for post-quantum signature aggregation

A post on ethresear.ch on 8 November reports measurements and negative results for one of the open questions in Lean Ethereum: how to aggregate post-quantum signatures. Lean Ethereum plans to replace BLS signatures with leanXMSS signatures aggregated by leanVM proofs, and the draft EIP-8292 sets a gossip target of about 256 KiB per aggregate. An Ethereum Foundation design study had contrasted two routes, running a full verifier inside a zkVM or folding proofs and proving a decider at the end, without numbers. The author, arianaraghi, ran leanVM and the WARP research code on an 8 GB Apple M2, treated as a low-end aggregator.

The baseline is concrete. Aggregating N signatures in a leaf costs about 18.5k plus 805.7 cycles per signature; 1,500 signatures reproduce the project README exactly at 1,227,023 cycles and a 265.1 KiB proof. Verifying one 250-signature child inside a merge costs 186.8k cycles, more than half of it in that child's WHIR opening. One README figure for two-to-one recursion did not reproduce.

Three accumulation designs lost. Folding leanXMSS verification directly would need at least 7.1 million R1CS constraints per signature, roughly 450 times what leanVM pays per signature. When the author folded leanVM's entire step relation instead, the WARP prover needed some 260 ns for each witness element, while leanVM commits a word in about 55 ns. Carrying the children's level-0 openings forward from merge to merge costs around 57k cycles each time and can defer no more than about 46k, so it never pays. A batching design that shares one WHIR opening across children saved 17.6 percent per child but was still 2.4 percent slower at the slot deadline than the tuned baseline. The post labels each figure as verified by measurement or inferred.

The useful positive result is a configuration rule. A merge node that crosses from 2^25 to 2^26 committed words takes 1.4 to 1.8 times longer for about 3 percent more data. Splitting 1,500 signatures into five leaves of 300 instead of six of 250 keeps the node below that step, cutting the deadline from 3.11 to 2.24 seconds and CPU per slot from 7.99 to 6.47 seconds. All results apply inside one aggregator: an accumulator would have to carry its leaf's witness, about 1 GiB, instead of a 265 KiB proof, so proofs passed over gossip remain recursive.

Lean Ethereum study: folding loses to leanVM recursion for post-quantum signature aggregation
Lean Ethereum study: folding loses to leanVM recursion for post-quantum signature aggregation — Coin Brief

Why it matters

Negative results with exact cycle counts save other teams months. They also show that for this workload the cheapest gain came from sizing the batches, not from a new proof system.

Written by Victoria Shinder.