Choosing a quantum-noise model depends on how the experiment is designed

Two decisions that interact

Quantum-noise tomography estimates errors from measurements of a device. In the generator-based approach studied here, the model describes small rotations of individual qubits and interactions between neighboring qubits. A model family specifies which kinds of rotations and interactions are allowed.

Huang compares two families for a three-qubit chain. One includes individual-qubit rotations and dispersive interactions, represented by ZZ terms. The other includes individual-qubit rotations and exchange interactions, represented by XX and YY terms. They contain 11 and 13 parameters, respectively, and neither family contains the other.

The second decision concerns recovery gates: corrections applied at selected sites after an error syndrome is identified. The study asks both how accurately a design estimates parameters and how effectively it reveals that a candidate model is wrong.

The work is an arXiv preprint, with no journal publication reference supplied. Its main findings come from mathematical analysis and numerical calculations, not a hardware demonstration of the tomography protocol.

A measurement requirement for detecting mismatch

The analysis uses a 48-component “fingerprint”: a collection of quantities describing a physical reference channel and two channels associated with a three-qubit repetition code. Candidate models are fitted to these fingerprints using least-squares optimization with multiple starting points.

The residual is the mismatch left after fitting. Huang distinguishes its relative size from its absolute size in the measured fingerprint. The latter determines how difficult the mismatch is to detect against statistical fluctuations.

The paper derives the required number of measurements per fingerprint component as N* = λreq / (rb S)². Here, rb is the relative residual, S is the device’s signal scale, and λreq specifies the desired statistical detection performance. A “shot” is one repetition of a measurement. The calculation uses a chi-squared adequacy test, rather than judging a residual against a fixed percentage threshold.

For the reference device points with full six-gate recovery, rejecting the wrong family with 80% detection probability at a 5% false-positive level requires about 10.7 million shots per component for the dispersive device and 6.72 million for the exchange device.

At the study’s assumed budget of 100,000 shots per component, the corresponding detection probabilities are only 0.053 and 0.056. Under these modeled conditions, the test has little ability to distinguish the mismatch from measurement noise. A small reported residual therefore cannot settle the family choice without its measurement budget and statistical threshold.

Placement changes what can be learned

For three qubits, selecting correction sites separately for the bit-flip and phase-flip codes produces 64 recovery designs. Huang evaluates all of them. Every design is full rank in both families: the distinction is not whether the parameters are identifiable locally, but how sensitively they can be estimated.

Precision is calculated from Fisher information, which describes how strongly the fingerprint changes with the model parameters. At a fixed count of three recovery gates, changing placement alters the predicted root-mean-square uncertainty of the nine individual-qubit rotation angles by factors of 2.7 and 3.9 on the two platforms.

Rebuilding and refitting the fingerprints under each design also changes the family-detection requirement. Across designs, that requirement varies by a factor of 68 for the dispersive device, but only 1.7 for the exchange device. In the dispersive mismatch comparison, the design offering the most precise local-angle estimates requires about ten times as many shots for the family test as the cheapest-to-test design.

The paper attributes this difference to where the mismatch appears. For the exchange device, 98.1% of its squared magnitude lies in the reference block, measured identically by every design. For the dispersive device, only 48.0% lies there; the remainder reaches code blocks that recovery placement can change.

What the calculation does not establish

The analysis assumes coherent errors and ideal correction gates. Its statistical degrees of freedom use a linear-model approximation for a nonlinear fingerprint. Working-point scans are samples, not bounds.

An external check using calibration and measurement records from a real three-qubit register examines those assumptions, but does not implement the proposed protocol. It finds that readout bias and incoherent gate errors require attention before the calculated shot requirements become experimentally meaningful.

A five-qubit extension also finds design-dependent detection costs, but samples rather than exhaustively fits the design space; difficult optimization makes its reported floors lower bounds at the stated computational budget.

The practical recommendation is to report the model family, measurement budget, gate placement and estimation objective together. Neither a residual percentage nor a gate count alone describes what an experiment can establish.