Why quantum computers can't see one of their own worst errors

Quantum computers are only as useful as their calibration. A new analysis argues the field's default cheap calibration is structurally blind to one class of coherent gate error, and shows that a logarithmically small set of extra measurement settings restores visibility at a cost that grows only with how poorly conditioned the problem is.

The paper, "Absent, Not Faint: Fisher-Information Limits and a Logarithmic Measurement-Design Cure for Passive Characterization of Coherent Qubit Noise," appeared on arXiv on 23 July 2026 (arXiv:2607.21663). It targets a routine step in calibrating a quantum processor: run a fixed-input circuit many times, read out the resulting histogram, and estimate the error parameters of the gates from those counts. This is the cheapest data a quantum device returns, and the cheapest way to populate an error model used to predict or correct future runs.

For one family of faults, the small systematic miscalibrations the authors call coherent over-rotations, that approach fails in a specific way. The error is not merely faint, meaning weakly present in the data. It is absent to first order. A histogram from the cheapest measurement looks the same whether the miscalibration is there or not, because the miscalibration can be exactly compensated by a second, random error. Two numbers cannot be separated from their sum, and the standard calibration cannot tell which side of the trade the device is on.

For commuting one- and two-qubit transverse over-rotations acting on a known input state, the measurement's Fisher information, a standard statistical bound on how much a given data set can learn about an unknown parameter, is singular along the fault's direction at zero angle. The Cramér–Rao lower bound, the best-achievable variance for any unbiased estimator, is therefore infinite. No finite-variance, locally unbiased estimator can recover that error from this data, no matter how many shots the experimenter takes. The fault sits in a direction the histogram literally does not see.

The obstruction lifts as the fault grows. At a generic nonzero angle the singularity partly resolves. Beyond four qubits it clears entirely, leaving a residual conditioning problem rather than outright blindness. The distinction matters in practice: a faint signal can be recovered by collecting more data, while an absent direction cannot, because the data carries no information about it at all. Throwing more shots at the same protocol is a way of making the same nothing more precisely.

The fix the authors propose is not a richer error model or a more expensive simulation. It is a richer measurement: a fixed, logarithmically small set of extra measurement settings that makes every such fault visible. Because the remaining obstruction is conditioning rather than coverage, the sampling cost is governed by how well-conditioned the problem is, with a complete-family closed form that is exponentially small in the qubit count. In other words, the price of seeing the error does not scale with the size of the device, only with how much the redesigned measurement still leaves the problem poorly conditioned.

The authors validate the argument in three ways. They prove the impossibility and the cure. They confirm both in exact numerical simulation, and show the conditioning analysis predicts recovery error across hundreds of randomized measurement designs. They also report a 3–5x bias gap on IBM Heron hardware, framed by the authors as a consistency check rather than an independent benchmark. The Heron figure should be read as the authors' own demonstration that their analysis matches a real device's behavior, not as a vendor performance result.

Two open problems remain. The proof and cure cover a specific class of faults, commuting one- and two-qubit over-rotations with known support on the input. Non-commuting faults and faults of unknown support are not covered, and would need separate treatment.

The practical upshot is a design principle. When a calibration protocol returns a histogram that looks indistinguishable from a different, simpler model, the usual response is more data. The paper identifies a class of quantum error where the correct response is a different measurement, and shows that response is both necessary and cheap.