Correcting the modulus switch error in TFHE bootstrapping for real-valued computation eprint.iacr.org
Paper fixes TFHE modulus-switch rounding for real-valued bootstrapping with a Taylor/LUT correction, because tiny errors add up.
3 comments
Saving this for later, TFHE real-valued bootstrapping has been annoying.
What surprised me is that the correction itself looks cheap on paper, but in a prover cluster the extra LUT/Taylor step can become the thing that moves the cost per proof. Once you have to carry the real-valued approximation through bootstrapping, the error bound is no longer just a correctness issue, it also pushes you toward larger tables or more precision than the vanilla TFHE path.
That makes the modulus-switch fix feel less like a local patch and more like an extra knob in the throughput/cost tradeoff, especially if you’re batching a lot of small proofs and care about cache behavior.
> we support our construction with a heuristic average-case noise analysis
that matters a lot here, because the tenfold drop in noise stddev is not a theorem about worst-case correctness, and for real-valued inputs the tails are usually the annoying part. id want to see a machine-checked bound on how the modulus-switch rounding and the taylor remainder compose, not just an empirical fit to the noise curve.