Tagging the Zeros of Zero-Noise Extrapolation: Protected Ones, Protected Zeros and a Blind Spot
Year: 2026
Type: preprint
Abstract: Zero-noise extrapolation (ZNE) amplifies the noise of a quantum circuit by a factor lambda and extrapolates themeasured expectation values back to lambda = 0. The value at zero is never measured; it is made by a fit model. Wetag what ZNE does to zeros and ones whose true values are known, using an exact density-matrix simulation of asix-qubit Trotterised Ising chain with depolarising noise, amplitude damping and coherent over-rotation, amplified byunitary folding (lambda = 1, 3, 5). Two runs were sealed, with hashes posted publicly before execution. Theprotected one: the parity P = prod X commutes with every gate, so its expectation decays as a pure exponentialwhose rate follows from the noise parameters; this held to 4e-15, exponential extrapolation returns 1, and polynomialfits stop at 0.53-0.86 (Richardson). The protected zero: <Z_i> = 0 is untouched by Pauli noise and moved only byamplitude damping; Richardson restores it to within 18% of the raw shift, while an exponential fit leaves at least 79%.The blind spot: folding does not amplify coherent over-rotation, so ZNE passes it through; in 229 terms of 60 freshrandom configurations the median passed fraction was 0.98. Our failure: the first run sealed that per-termexponential extrapolation beats one fit of the total energy, and it came out F because our exponential fit was definedonly for positive values and a term was negative. Repaired, and re-sealed on fresh configurations, the claim heldwhere it is defined (94% of configurations), a data-only curvature rule chose the better fit 96% of the time, and theshot budget of ZNE was predicted within 2%. Of 11 sealed claims, 9 came out R, 1 U and 1 F. The work is asimulation; it makes no claim about any device. (1) Text: CC-BY-4.0 | Code: GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later) (2) Data and code. Both seal documents with their SHA-256 files, frozen results, development and post-hoc logs, andall simulation, judging and figure scripts (AGPL-3.0-or-later) are inBridge_Paper_Tagging_the_Zeros_of_Zero_Noise_Extrapolation_scripts_v1.0.0.zip. The seal and result hashesare posted at github.com/watabe-masanori/cds-polytope, issue #2. No external data are used. (3) AI use. Designs, code, analyses and text were produced by Kurado (Claude Opus 5.5, Anthropic) working with M. Watabe, who chose every direction and posted the hashes. Draft comments by ChatGPT and Gemini, relayed by M. Watabe, prompted Appendix A, the standard-term column of Table 3 and Section 5; they are AI commentary, not peerreview.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7232204722
