Could HepLean's formal verification catch precision losses that arise from simplified numerical algorithms?

# HepLean and Numerical Precision in Algorithms

HepLean is designed to catch a different class of error than numerical precision loss — and the provided preprints suggest it may not be the right tool for that particular problem.

HepLean is a formal verification system that checks the *logical correctness* of mathematical proofs and definitions in high energy physics [AG-2024.05-806]. A striking example is its discovery of an error in a widely-cited 2006 paper on Higgs doublet model stability — an error that went unnoticed for two decades because the authors' mathematical reasoning contained a logical flaw [AG-2026.03-830]. Formal verification excels at catching these kinds of logical mistakes.

However, numerical precision loss in simplified algorithms is a *different beast*. It arises when you implement a mathematically correct formula in floating-point arithmetic, truncate series expansions, or use coarse approximations. These are not logical errors in the proof — they're approximation errors introduced by the computational method itself.

The archive does address precision in numerical contexts, but via other approaches: [AG-2024.06-1166] describes physics-driven algorithms that achieve numerical stability in scattering amplitude calculations by modifying propagators and vertices to avoid ill-conditioned regions. [AG-2026.01-1029] shows how to quantify and improve precision in machine-learned amplitude surrogates by identifying poorly-predicted phase space regions. But neither of these uses formal verification.

In principle, one could imagine formalizing numerical error bounds in Lean (proving that an algorithm's output lies within a certified error margin of the true value), but none of the provided preprints demonstrate this capability for HepLean.