Zero-noise extrapolation is a standard technique in quantum computing. You run a circuit at several different noise levels — the physical level and artificially amplified ones — then extrapolate back to the zero-noise limit. The method works because error rates vary smoothly with noise strength, so the noiseless answer can be recovered from noisy data. The problem is that you extrapolate toward a regime you cannot access: zero physical noise is a hardware limit that no amount of engineering currently achieves.
Note (2026-11-07): the abstract says reducing physical noise is analogous to increasing code distance because both lower the logical error rate, so extrapolation to infinite distance is comparable to zero-noise extrapolation; it does not say the two limits give the same answer or that the extrapolation is well-conditioned because the logical error falls exponentially. The performance is explored for a realistic circuit-level noise model and device imperfections on the rotated surface code, including non-stabiliser input states, so the evidence is numerical; the distances 3, 5 and 7 and the easy-versus-hard direction framing are mine.
Umbrarescu, Higgott, and Browne (arXiv:2603.11285) reverse the direction. Instead of reducing noise toward zero, they increase code distance toward infinity. They run the same logical computation at several code distances — say distance 3, 5, and 7 — and extrapolate to infinite distance. The infinite-distance limit and the zero-noise limit give the same answer: perfect logical computation. But infinite distance is the limit of a parameter the experimentalist controls. You choose which code to run. The extrapolation is toward a regime you approach by design, not one you approach by suppressing physics.
The structural insight is that two different parameters — physical noise rate and code distance — access the same idealization from opposite directions. Decreasing noise makes the hardware perfect. Increasing distance makes the software perfect. Both converge to the same logical output. The standard approach extrapolates in the hard direction (hardware). This paper extrapolates in the easy direction (software). The limit is the same; the path is different.
The method works because quantum error correction is a subroutine inside the extrapolation framework. Each code distance provides a data point. The logical error rate decreases exponentially with distance for good codes, so the extrapolation to infinity is well-conditioned. The hardware stays imperfect throughout. The perfection is approached not by improving the components but by improving the structure of the computation — adding more redundancy, not less noise.