Skip to content
Crypto.world

The Ledger

Preprint: ML-DSA-44 Per-Record Design Costs 8.7× ECDSA Even With Proposed Precompile

An arXiv preprint compares post-quantum signatures for authenticating IoT data on Ethereum, including designs that sign each record and designs that sign a Merkle root for a batch. It reports that batching changes the per-record costs, while data availability and signing latency remain relevant to deployment choices.

Preprint: ML-DSA-44 Per-Record Design Costs 8.7× ECDSA Even With Proposed Precompile

The 8.7× result has a specific scope

An arXiv preprint dated Nov. 5, 2026, reports that signing every record with ML-DSA-44 costs 8.7 times as much as ECDSA, even with a proposed ML-DSA precompile. That figure concerns the preprint’s per-record design, not every post-quantum signature or authentication setup. The preprint attributes the cost to Ethereum’s calldata floor: the data cost, rather than signature verification, sets the price in that case.

The study measures costs under current Ethereum execution rules, including the EIP-7623 calldata floor and EIP-7825 transaction gas cap, and under calldata repricing scheduled for the Glamsterdam upgrade. It combines device-side measurements of ML-DSA, Falcon, SLH-DSA, LMS and XMSS with verifier costs and complete transactions on an Osaka-rule EVM. Its proposed ML-DSA and Falcon precompiles are drafts and are not scheduled for an upgrade.

Root signing changes the comparison

Instead of signing every record, a device can sign a Merkle root representing a batch. For a root covering 100 records, the preprint reports about 3,350 gas per record for a standard LMS verifier, compared with 245 gas per record for ECDSA in the same design. It presents the LMS/HSS verifier as standard-compliant for SHA-256 and SHA-256/192. These figures describe a root-based design, where one device signature authenticates a batch rather than each record being signed separately.

LMS and XMSS are stateful, hash-based signature schemes. Each one-time key must be used only once, so a signer has to maintain a reliable counter. The preprint explains that a tree of height h yields 2h signatures, while HSS stacks trees to extend that capacity. It also describes the operational consequences of keeping that counter: with daily roots, the device persists it once per epoch, rather than once for every record. The choice of signing scheme therefore involves more than transaction gas.

Precompile estimates depend on the design

For a Merkle root covering 100 records, the preprint reports an ML-DSA-44 cost of 1,031 gas per record with proposed precompiles. A hybrid ECDSA and ML-DSA signature adds 2.5% to the cost of the ML-DSA signature in the reported comparison. These estimates depend on proposed precompiles and should not be read as costs for signing each record individually.

The preprint also reports that the scheduled Glamsterdam repricing raises floor-bound designs by up to 1.9 times, while the standard LMS root design is unaffected. Its comparisons account for signature sizes, verification costs, where keys are kept and what data is posted. For example, the preprint lists ML-DSA-44’s public key at 1,312 bytes and its signature at 2,420 bytes; Falcon-512’s compressed signature is 656 bytes.

Batching has application tradeoffs

The preprint considers several ways to publish data associated with signed roots. One design posts the leaves in transaction calldata; another posts only the root and related values, leaving the leaves off-chain. A further variant puts leaves in a blob. These choices affect what observers can access: the preprint says calldata and event logs are part of chain history and may be pruned, while blob contents are kept by consensus clients for about 18 days. Contract storage is permanent, but the root-only design does not publish the leaves.

Signing roots spreads the signature cost across multiple records, but requires waiting to collect records into a batch. The preprint’s assessment is that post-quantum device authentication can be deployed on Ethereum without protocol changes if the batching latency and the selected variant’s data-availability properties suit the application. Its results compare designs; they do not identify one signature scheme as the best fit for every IoT deployment.

Sources

  1. Original source · Preprint arXiv arXiv:2611.08108

Related stories

Explore

More articles