Gauge-Invariant Coherent Gate Error from Raw Readout Events

Quantum gates usually go wrong in two ways. Random noise scrambles the state. Coherent error over-rotates or mis-phases it, so the gate does a slightly wrong rotation. A clean way to measure how far an implemented gate $V$ is from its target $U$ is the "spectral convex hull distance". Take the eigenvalues of $U^\dagger V$, which are points on the unit circle in the complex plane. Draw the smallest convex polygon containing them, and measure its distance from the origin. For unitary gates, Operational distance and fidelity for quantum channels proves this number equals the worst-case "minimax" fidelity between the two operations, and it sandwiches the diamond-norm error that fault-tolerance thresholds depend on. Experiments rarely estimate this quantity directly, and they throw away information before they start. The detector produces a continuous signal, such as photon arrival times in fluorescence readout or voltages in dispersive readout. Each shot is thresholded to a single bit, and only then is the gate reconstructed from bit counts.

This note proposes to estimate the spectral convex hull distance directly from raw readout events. The tool is an unbinned extended maximum likelihood fit, the particle-physics method that scores every recorded event under a continuous probability model and also accounts for the random total number of events. The detector-response model is fitted in place, from the same data, alongside a physically valid gate model. The target is the gauge-invariant eigenphases of the gate rather than its full matrix. The core claim is testable: for moderate readout quality, this estimator reaches a given precision on the coherent-error distance with substantially fewer shots than thresholded tomography. It also avoids the readout-calibration bias that a fixed, pre-calibrated soft-readout model incurs, and the gain should vanish as readout becomes perfect.

Core Claim

Two existing results line up. Easy better quantum process tomography shows that a gate's eigenvalues are gauge-invariant, meaning they are unaffected by how errors are split between preparation, measurement and the gate. Its cheap "tomography on nothing" correction, $\hat G = \hat{\mathcal E}^{-1/2}\hat G_0\hat{\mathcal E}^{-1/2}$, recovers those eigenvalues with near-zero error as preparation-and-measurement noise grows. Standard tomography's eigenvalue error instead grows linearly. The method uses linear inversion on discrete counts, however. It can return non-physical maps, and the authors explicitly call for embedding the correction in a constrained maximum-likelihood fit.

Separately, the minimax-fidelity result tells us exactly which function of the eigenphases matters operationally. Global phase drops out. For a unitary, the phases of the superoperator eigenvalues are pairwise differences $\phi_i-\phi_j$ of the eigenphases of $U^\dagger V$, so they determine the arc spread directly. Let $\Delta$ be the length of the shortest arc of the unit circle containing all eigenphases. Then

$$ f(U,V)=\operatorname{dist}\big(0,\operatorname{conv}\,\operatorname{spec}(U^\dagger V)\big)=\cos(\Delta/2)\ \text{ for } \Delta\le\pi,\qquad f=0 \text{ otherwise}. $$

For a noisy gate we use the same formula, applied to the phases of the gauge-invariant superoperator eigenvalues of the fitted physical map. This is a definition we adopt for the coherent part of the error. It reduces exactly to the Belavkin–D'Ariano–Raginsky quantity in the unitary limit.

The new element is the likelihood that feeds this estimate. Consider fluorescence readout of a trapped-ion or neutral-atom qubit. Shot $s$ records a random number of photons $N_s$ with arrival times $t_{s1},\dots,t_{sN_s}$. The bright state emits at rate $S$ and can be optically pumped dark during the window. The dark state can be pumped bright. Background arrives at rate $B$. Each shot therefore already has the extended-likelihood form: a Poisson factor for $N_s$ times a product of arrival-time densities, mixed over the qubit's post-gate state probabilities $p_k(\theta)$:

$$ \mathcal L(\theta,\eta)=\prod_{s}\sum_{k} p_{k}(\theta)\,\frac{e^{-\mu_k(\eta)}\mu_k(\eta)^{N_s}}{N_s!}\prod_{i=1}^{N_s} P_k(t_{si}\mid\eta). $$

Here $\theta$ parametrizes the physical gate map plus preparation and measurement maps. $\eta$ collects the detector-response parameters: rates, pumping times and dead time. $\mu_k$ is the expected photon count in readout branch $k$. Thresholding $N_s$ keeps only one bit of this. Arrival-time structure is especially valuable when a bright atom goes dark mid-window, because that shot looks dark to a count threshold but has a telltale early burst.

Prior Work

The method draws on three lines of work.

The first is unbinned and extended likelihood fitting in rare-event physics. The CLEO τ→μγ search (Update of the search for the neutrinoless decay τ→μγ) fits a two-dimensional kinematic plane event by event and sets an upper limit near a physical boundary. BASiL (Signal estimation in on/off measurements including event-by-event variables) weights events by continuous single-event likelihoods and reports a 15–20% precision gain over hard-cut counting at 10% signal-to-noise. The PandaX-II calibration (Determination of responses of liquid xenon…) fits detector response in situ with an unbinned likelihood. We reuse three things from these papers: the event-level likelihood structure, in-situ response fitting, and boundary-aware interval construction. None of them touches quantum channels.

The second is single-photon timing likelihoods with dead time. Free-running vs Synchronous: Single-Photon Lidar jointly fits signal flux, background flux and a continuous parameter from photon arrival times. It shows that arrival-time likelihoods remove the pile-up bias of histogrammed timing. This motivates our readout model. We replace depth with qubit-state probabilities that come from a quantum map.

The third is quantum tomography itself. The maximum-likelihood state tomography methods, such as Soft-Bayes and Rigorous maximum-likelihood estimation for quantum states, maximize multinomial likelihoods of discrete outcome counts and reconstruct full matrices. We instead model continuous events and target a scalar spectral functional.

Two other results constrain the design rather than supplying it. Gate set tomography with randomized compiling (Benchmarking quantum logic operations relative to thresholds for fault tolerance) makes worst-case and average error agree, reporting a maximum worst-case error of 0.0197(3) on the AQT@LBNL processor with roughly 4% preparation-and-measurement error. It does so by twirling, which deliberately erases the coherent phases our quantity measures. The unitary-learning bounds in Learning Quantum States and Unitaries of Bounded Gate Complexity show that worst-case learning is exponentially expensive in general. We therefore confine the claim to few-qubit gates, where the spectrum has at most 16 points, and make no claim of beating those lower bounds.

Research Program

We will build the estimator for one- and two-qubit gates in four steps.

First, we parametrize the gate as a Choi matrix kept completely positive by a Cholesky-type factorization. Preparation and measurement are modelled as separate physical maps, following the Better-QPT structure but fitted jointly rather than inverted. Experiments include the empty "tomography on nothing" sequences and repeated-gate sequences $G^k$ for several depths $k$. Including $G^k$ amplifies small eigenphases, and since eigenvalues of $G^k$ are $\lambda^k$ the parametrization stays the same.

Second, we write the per-shot extended likelihood above for fluorescence readout. A Gaussian-mixture variant covers dispersive voltage readout, with an extra mixture branch for leakage events whose yield enters through the extended Poisson term.

Third, we maximize over $\theta$ and $\eta$ jointly. We then read off the superoperator eigenphases, compute $f$, and build profile-likelihood intervals for $f$. Because $f$ is non-smooth where the shortest-arc structure changes and is pinned at the boundary when the gate is nearly ideal, we will calibrate coverage by toy Monte Carlo in the Feldman–Cousins spirit used for rare-decay upper limits.

Fourth, we compute the Fisher information of raw-event readout versus thresholded readout for each response model. This gives a predicted shot-savings factor before any fit is run, so the empirical gains can be checked against theory rather than just reported.

Baselines

The naive reference point is standard linear-inversion process tomography on thresholded bits, with $f$ computed from the eigenphases of the reconstructed map. Better QPT reports that its eigenvalue error grows linearly with preparation-and-measurement noise, so it should fail visibly.

The strongest existing method is Better QPT (Easy better quantum process tomography) on thresholded bits. We run it in two forms: as published with linear inversion, and upgraded to a CPTP-constrained binned maximum-likelihood fit with the same gauge correction. The upgraded form is exactly the extension its authors request, and it is the competitor that most directly isolates the value of using raw events. The paper reports eigenvalue error "near zero" on its single-qubit $X_{\pi/2}$ simulation only graphically. No numerical eigenvalue-error result was found on Lacuna for this benchmark, so we will re-run it on our simulator rather than quote a number.

The key ablation is our unbinned likelihood with the detector-response parameters $\eta$ frozen at values from a separate calibration run. This "soft readout with a fixed model" separates the information gain from raw events from the robustness gain of fitting the response in place.

Gate set tomography serves as a cost-and-accuracy reference rather than a competitor. The only number on file is its 0.0197(3) maximum worst-case error on hardware we do not have.

All methods receive identical shot records, which are thresholded for the binned methods. They use the same sequence set, including the $G^k$ depths, and the same optimizer budget. Any threshold or bin choice is tuned on held-out simulated data. Beating the upgraded binned fit means two things. Its root-mean-square error on $f$ must be lower at equal shots, by a margin matching the Fisher-information prediction. Its intervals must also hold nominal coverage where the binned fit's coverage or bias degrades.

Why not just…?

Why not just threshold the readout and run the constrained Better-QPT fit? This is the main competitor, and for high-fidelity readout it may be nearly as good. The Fisher-information analysis predicts the gap. It should be small when bright and dark photon-count distributions barely overlap, and large when pumping during the window or background makes them overlap. The experiments sweep readout overlap specifically to locate where the proposal pays off and where it does not.

Why not just use soft-decision readout with a pre-calibrated detector model? This is a real threat, because it captures most of the information gain at little cost. Our claim against it rests on drift. If the response rates or pumping times shift between calibration and data-taking, a frozen model biases the state probabilities and hence the eigenphases. The frozen-$\eta$ ablation under injected drift settles whether fitting the response in place matters. If it does not, we will report that the simpler method suffices.

Why not just use randomized compiling plus randomized benchmarking? The fault-tolerance benchmarking paper shows this cleanly bounds worst-case error, but it does so by converting coherent error into stochastic error. That destroys the phases that define $f$, so it answers a different question: it reports how bad the twirled gate is, not how a calibrator should correct the untwirled one.

Why not just estimate the rotation angle by repeating the gate many times? Gate repetition is complementary rather than an alternative, and we include it in the sequence design. Its benefit, amplifying phase, and the readout-information benefit should multiply. We test this by comparing gains at each depth $k$ separately.

Proposed Experiments

All experiments use simulation, built on the Better-QPT single-qubit $X_{\pi/2}$ benchmark with depolarizing and coherent preparation-and-measurement noise, extended to a two-qubit entangling gate with injected coherent error. Readout is modelled physically by fluorescence with pumping, background and dead time, and by dispersive two- or three-cluster voltage readout. Preparation-and-measurement error rates are swept through about 4%, the level reported on the AQT device.

The metrics are bias and root-mean-square error on $f$ and on individual eigenphases, interval coverage near $f=1$, shots needed to reach a target error, and wall-clock fit time.

Three decisive checks follow from the claim. The measured shot savings must track the Fisher-information prediction across readout-overlap settings and collapse toward one as readout becomes perfect; if it does not, the claimed mechanism is wrong. Under injected response drift, the in-place fit must stay unbiased while the frozen-response ablation does not. Near the boundary and at arc-structure transitions in the two-qubit case, profile intervals must keep nominal coverage, while naive plug-in intervals from linear inversion are expected to undercover.

If hardware raw readout records become available, the same pipeline can be run on them. The headline result does not depend on them.