Efficient SNARK for Floating-Point Computation
Abstract
Floating-point arithmetic under the IEEE-754 standard is the foundation of modern numerical computing, yet as computation increasingly migrates to untrusted environments—such as outsourced HPC—verifying that floating-point operations are executed faithfully becomes critical. Succinct non-interactive arguments of knowledge (SNARKs) can provide such guarantees, but incorporating IEEE-754 semantics into SNARKs remains notoriously difficult due to the mismatch between prime-field arithmetic and bit-level floating-point logic. Existing approaches are limited in the operations they support. We present a sound, efficient, and expressive SNARK framework for IEEE-754 FP32 arithmetic. Our key idea is a unified generalized normal representation that collapses normal, subnormal, and zero cases into a single algebraic form, eliminating the costly case-by-case branching of prior work. Building on this, we design efficient constraints for addition, multiplication, division, and square root via structured-lookup tables, reducing constraints by 1.9×-2.9× over the state-of-the-art. For transcendental functions (exp, sin, cos), we propose an exponent-aware decomposition to reduce the prover overhead. Under the same proof backend, our framework improves prover efficiency by 1.9×-4.0× for basic operations compared with previous state-of-the-art ZKLP(S&P'25). Our exactness claim is limited to executions with finite FP32 inputs and finite FP32 outputs; ±∞ and NaN are outside the current relation.