Efficient Implementations of Small-Modulus Lattice KEM Components
ePrint 2026/2692 by Ying Liu, Yu Zhang and Xianhui Lu gives implementation techniques for small-modulus lattice KEMs at q=257 and q=769, integrated into BAT and END. It covers a zero-selector public-key encoding (1033 bytes for n=1024, overflow probability at most 2^-128), specialized rejection sampling and residual NTT multiplication. Encapsulation cost falls 39.93-55.25% for BAT and 41.36-48.84% for END. No code link is given.
What it is
ePrint 2026/2692 (preprint, received 30 October 2026) by Ying Liu, Yu Zhang and Xianhui Lu. It describes modulus-specific implementation work for the small-modulus lattice KEMs BAT (TCHES 2022) and END (ASIACRYPT 2026, to appear). The moduli are q=257 and q=769. The paper covers three components: public-key encoding, uniform public polynomial generation, and residual polynomial multiplication after an incomplete NTT.
How it works
Zero-selector encoding (q=257). Zero payload bytes mark the positions of ambiguous coefficients, and selector bits resolve their values. The paper analyzes selector capacity, overflow and key-generation retries. For n=1024, a 1033-byte configuration overflows with probability at most 2^-128, assuming independent uniform coefficients. Checked AVX2 decoding takes 213 cycles, against 2203 cycles for BAT's 1040-byte format.
q=769 encoding. The paper uses an existing compact radix format at the same encoded length as BAT, which lowers conversion cost.
Rejection sampling. At q=257 the sampler is specialized to simplify coefficient extraction. At q=769 it is specialized to cut random-bit demand. For (q,n)=(769,1024), expected candidate-bit demand falls 24.39% against direct 10-bit rejection. Four-polynomial generation uses 7.14-13.21% fewer AVX2 cycles.
Residual multiplication. Proven range bounds justify omitting strict input-centering corrections. Register scheduling reduces stack accesses.
Results in the full KEMs
The codecs and multiplication components were integrated into BAT and END and tested in C and AVX2.
| Encapsulation | Decapsulation | |
|---|---|---|
| BAT | 39.93-55.25% lower | 5.42-8.55% lower |
| END, C | 41.36-48.84% lower (range covers C and AVX2) | 23.99-45.89% lower |
| END, AVX2 | (see above) | 6.98-8.04% lower |
The authors say the END AVX2 gains come mainly from the public-key codec.
Caveats
- The 2^-128 overflow bound depends on coefficients being independent and uniform.
- Decapsulation gains are much smaller than encapsulation gains, especially in AVX2 (about 5-8%).
- The page gives no code link, no benchmark platform and no constant-time or side-channel discussion, so adoption would need reimplementation and separate review.
- The 1033-byte format differs from BAT's 1040-byte key format, so it is not wire-compatible with it.
Organisations named here
- ASIACRYPT
- TCHES
- zero-selector encoding
Technologies named here
- AVX2
- C
- BAT
- END
Techniques named here
- Compact radix codec
- Incomplete NTT residual polynomial multiplication
- Specialized rejection sampling
- Zero-selector encoding
Concepts named here
- End-to-end performance reductions
- Overflow probability
- Small-modulus lattice KEM
Related in other

Rerandomizable Garbling for Arithmetic Circuits

