Error Flow Maps from Mirror Circuits

A mirror circuit checks a quantum computer by running a program forward and then running its exact inverse. On perfect hardware the qubits return to all zeros, so any nonzero output bit means something went wrong. No classical simulation of the intermediate state is needed. This is today's main scalable answer to the quantum verification gap, the problem that we cannot check a device we cannot simulate. Its weakness is the one video researchers hit with optical flow, the per-pixel motion field used to align frames. Video pipelines handle flow failures by computing motion forward and backward and masking the pixels where the round trip does not close. Mirror benchmarks, by contrast, collapse their round trip into a single number, the probability of returning to zero. That number cannot say which gate, qubit, or time step failed. It can also be fooled when an error in the forward half is undone by the matching error in the inverse half.

This note proposes to treat the full pattern of flipped output bits from mirror circuits as data for an inverse problem. For Clifford circuits, the gate set that classical computers can simulate efficiently, each candidate error location produces a computable signature of flipped bits. That makes it possible to infer a space-time map of error rates across the circuit from ordinary mirror runs, without inserting extra test sub-circuits. Borrowing the forward-backward check from optical flow, we run two or more physically different but logically equivalent inverses. Where their inferred maps disagree, the cause is coherent error or inverse-side fault rather than a simple random fault, and we mask those regions as unreliable instead of reporting a false clean bill of health.

Core Claim

Mirror benchmarking already has the information needed for fine-grained fault localization; current practice discards it by summarizing outputs as Hamming-distance polarization. Let the forward circuit be undressed, so structured algorithms keep the coherent error accumulation that Measuring the capabilities of quantum computers showed randomized circuits hide. Let the inverse come in several structurally distinct variants. Then a regularized linear inversion recovers per-gate, per-layer error rates. A cross-variant consistency mask then flags exactly the circuit regions where echo cancellation or non-Pauli noise makes those rates untrustworthy.

The closest prior work is Quantum circuit debugging and sensitivity analysis via local inversions. That paper localizes faults by physically inverting sub-slices or causal cones, which changes the circuit for every block tested. We instead localize from whole-circuit mirror data by classical inference. We also add a diagnostic for when that inference fails, which neither the local-inversion work nor the global mirror protocols provide.

Mechanism

In the small-error regime, a Pauli error of type $P$ at space-time location $g$ in a Clifford mirror circuit propagates deterministically to a final flip pattern $\sigma(g,P)$. That pattern can be computed in polynomial time by stabilizer propagation, using the kind of efficient Clifford arithmetic in Fast algorithms for classical specifications of stabiliser states and Clifford gates. The observed distribution over nonzero flip patterns $s$ is then approximately linear in the unknown error rates $p_{g,P}$:

$$q(s) \approx \sum_{g,P} p_{g,P}\,\mathbf{1}[\sigma(g,P)=s].$$

Many locations share a signature, so the system is underdetermined. This is the quantum analogue of the aperture problem in optical flow, where local image data only constrains motion along the brightness gradient. We resolve the ambiguity the way flow estimation does, with a smoothness prior. Error rates of the same physical gate type on the same qubit pair are tied across layers, and rates of neighboring couplers are softly coupled. A different inverse variant changes the propagation map $\sigma$ and so breaks some degeneracies directly, which is the second reason to run several inverses.

The variants share the logical identity $U^\dagger U = I$ but differ physically in three ways: native-gate decomposition of $U^\dagger$, ordering of commuting two-qubit gates, and a single Pauli frame change at the mirror plane. Under purely stochastic Pauli noise, all variants must yield the same forward-half map. A coherent over-rotation that cancels against its mirror image in one variant will not cancel in another. Persistent disagreement on a region therefore plays the role of the occlusion mask in Artistic style transfer for videos. There, forward-backward flow disagreement marks pixels where warping cannot be trusted; here it marks gates where the stochastic error map cannot be trusted. The same idea answers the contamination problem raised by An Unsupervised Temporal Consistency (TC) Loss To Improve the Performance of Semantic Segmentation Networks. That paper found that a faulty flow estimate penalizes correct outputs. In mirror circuits, the analogue is that the noisy inverse corrupts the verdict on the forward circuit, and variant comparison separates the two.

Research Program

We will build three components. The first is a signature compiler that takes a Clifford or Clifford-approximated structured circuit and outputs the sparse matrix of $\sigma(g,P)$ for each inverse variant. The second is a regularized non-negative solver with gate-type tying and coupler smoothness. The third is the consistency-mask rule, with thresholds calibrated by bootstrap on shot noise. Counting statistics set the expected precision: each location's rate is estimated from flip events, so error should scale roughly as $1/\sqrt{N_{\text{shots}}\, p_g}$ before degeneracy penalties.

We will analyze identifiability as a function of circuit structure. Periodic structured circuits are the hard case, because repeated layers create repeated signatures. We will also quantify how many inverse variants are needed before the map becomes identifiable.

The scope is limited to Clifford circuits and to structured algorithm circuits whose non-Clifford rotations are rounded to Clifford ones. The rounded circuits are the ones exercised by Scalable Full-Stack Benchmarks for Quantum Computers for Hamiltonian simulation. Universal circuits like those in Demonstrating Scalable Randomized Benchmarking of Universal Gate Sets are out of scope. So are circuits with mid-circuit measurement.

Baselines

The naive reference is global mirror polarization as computed in Scalable randomized benchmarking of quantum computers using mirror circuits and in mirror circuit fidelity estimation from Scalable Full-Stack Benchmarks for Quantum Computers. It gives no localization, so for localization tasks its score is chance. For echo detection we will report its return probability on circuits with injected coherent faults. Those papers report capability regions and Quantum Volume overestimation, not localization accuracy. The one quantitative anchor from the file is the finding in Measuring the capabilities of quantum computers that structured circuits fail up to 10 times faster than randomized ones. The file reports no per-layer polarization decay value.

A stronger naive baseline uses per-qubit marginal flip rates from the same mirror data, attributing error to the qubits that flip most. This tests whether our signature inversion adds anything beyond reading off which output bits are bad.

The strongest existing competitor is local inversion from Quantum circuit debugging and sensitivity analysis via local inversions. It is purpose-built for in-situ fault localization and measures each block directly instead of inferring it. We will run it on the same circuits and the same device in the same calibration window, with a matched total shot budget split across its sub-circuits. I found no reported localization accuracy for local inversions on Lacuna. To count as beating it, our method must match or exceed its top-k fault recall at equal shots, or match it at a clearly smaller shot and circuit-count budget.

For echo detection the competitor is Pauli-dressed mirror benchmarking, run with the same number of distinct circuits as our variant set.

Two ablations isolate the key ingredients. The first uses a single inverse instead of several, which tests whether variant diversity matters for identifiability and coherent-error detection. The second drops the smoothness prior, which tests whether the aperture-style regularization is what makes the inversion well posed.

Why not just…?

Why not just use local inversions? They measure locality directly and are the real threat. Their cost is one modified circuit per tested block. Each modification also changes the circuit being tested, so the faults they find are faults of a different circuit. If at matched shots local inversion localizes as well as our inference, our contribution shrinks to running unmodified circuits plus the consistency mask. The head-to-head shot-budget curve in our experiments decides this. We will also check whether the two methods agree on hardware, which validates both.

Why not just Pauli-dress everything to kill echo cancellation? Dressing does convert coherent errors to stochastic ones. But Measuring the capabilities of quantum computers shows that randomization shields exactly the systematic accumulation that makes structured programs fail. Our single mirror-plane frame change is a minimal intervention that leaves the forward structure intact. We will test whether full dressing detects our injected coherent faults at the same rate as variant disagreement. We expect that it detects the presence of error but underestimates its magnitude in the structured circuit.

Why not just characterize each gate separately and simulate? Per-gate characterization misses context-dependent crosstalk, which Scalable randomized benchmarking of quantum computers using mirror circuits detected only under simultaneous operation. For non-Clifford targets it also needs classical simulation that the verification gap rules out. Our comparison of predicted against observed flip-pattern distributions on held-out circuits will measure how much context-dependent error isolated characterization misses.

Why not just use classical shadows? As Efficient sampling for Pauli-measurement-based shadow tomography in direct fidelity estimation discusses, hardware-friendly Pauli shadows become expensive for highly entangled targets. Estimating fidelity also requires a classically known target state, which mirroring avoids. Shadows estimate state properties, not where in time an error entered, so they do not compete on localization.

Proposed Experiments

The first stage is stabilizer simulation with known ground truth. We will use mirror randomized benchmarking circuits up to the $15\times15$ lattice scale used in Scalable randomized benchmarking of quantum computers using mirror circuits. We will add structured periodic circuits in the style of Measuring the capabilities of quantum computers and Clifford-rounded Hamiltonian simulation circuits. Into these we inject stochastic Pauli faults at random space-time locations, correlated crosstalk on chosen couplers, and coherent over-rotations designed to echo-cancel under the literal inverse. The metrics are top-k recall of faulty locations, rank correlation between inferred and true per-gate rates, and the true-positive and false-positive rate of the consistency mask on coherent faults. We report all metrics as functions of shot budget and number of inverse variants.

The decisive check is a single echo-cancelling over-rotation buried in a structured circuit. Plain mirroring should return near-ideal polarization. Our method must flag the region through variant disagreement and place it within one layer of the true location.

The second stage runs on cloud superconducting hardware of the class used by Scalable Full-Stack Benchmarks for Quantum Computers, at widths up to about 14 qubits. We compare our inferred maps against local-inversion measurements taken in the same calibration window. We also test prediction: the map inferred from one set of mirror circuits must predict flip-pattern distributions on held-out structured circuits better than the global-polarization and marginal-flip baselines, as measured by total variation distance.

The proposal fails if one of three things happens. Variant diversity may add no identifiability over a single inverse. The mask may not separate coherent from stochastic faults in simulation. Local inversion may dominate at every matched shot budget. Any of these would be a clean, publishable negative result about how much information end-to-end mirror data actually carries.