Local-Shadow Predictive Control for Drifting Crosstalk on Qubit Lattices

Predictive controllers for quantum hardware plan a short pulse sequence, apply its first slice, measure, and replan. They work well on one or two qubits. They stop scaling because the "measure" step reconstructs the full quantum state, and that costs exponentially many measurements as qubits are added. On a real chip, though, the thing that goes wrong is mostly local and structured. Unwanted always-on couplings between neighboring qubits, called ZZ-crosstalk, drift over time and spread errors across the chip's wiring graph. We propose a controller that never reconstructs the global state. It measures only small patches of neighboring qubits using randomized "classical shadow" measurements. From those patches it estimates the crosstalk on each edge of the coupling graph, forecasts how the crosstalk will drift using a graph-based time-series model borrowed from traffic forecasting, and re-optimizes pulses patch by patch.

The claim is that feedback for nonlinear predictive quantum control can be made to scale with patch size rather than chip size. The number of measurement shots per control update should grow only logarithmically in the number of qubits. The graph structure of the chip then does double duty: it defines which patches to measure, and it defines the dependency structure of the drift forecaster. The question the proposal answers is whether this local, forecast-driven loop recovers the error that crosstalk adds on a surface-code lattice, at a measurement cost that full tomography cannot match beyond two qubits.

Where this sits

Three pieces of prior work motivate the gap. Model predictive control for robust quantum state preparation runs nonlinear model predictive control (NMPC), meaning replanning by solving a non-convex finite-horizon problem with sequential quadratic programming. The controller corrects detuning, leakage, and two-qubit crosstalk. Its feedback step, however, is full state tomography, and the authors name that step as the barrier to scaling. Robust Quantum Control via a Model Predictive Control Strategy shows that closed-loop replanning beats open-loop pulses by large margins on one qubit, but it relies on projective measurement of a single two-level system. Tractable Infinite-Horizon Stochastic Model Predictive Control for Quantum Filtering via Eigenstate Reduction reaches 8 qubits. It does so by exploiting commuting measurement and Hamiltonian structure, and it propagates a full density-matrix filter. We assume neither.

One paper supplies the target phenomenon. Alibaba Cloud Quantum Development Platform: Surface Code Simulations with Crosstalk simulates the 17-qubit surface code with ZZ-crosstalk using tensor networks. It finds that crosstalk raises the logical phase-flip probability to about 0.6%, a 60% increase over the crosstalk-free model, and pushes it past 1% as coupling strength approaches 0.05. It is a passive simulator with no control.

The estimation method comes from Predicting Features of Quantum Systems using Classical Shadows. A classical shadow is a randomized measurement record from which many local properties can be estimated at once, at a cost independent of the total state dimension. Two neural-shadow papers repair shadows' weakness on high-weight observables by fitting a global neural wavefunction: Neural-Shadow Quantum State Tomography and Bootstrapping Classical Shadows for Neural Quantum State Tomography. We deliberately do not do this. The controller only needs few-body quantities, where raw shadows are already efficient, and fitting a global network inside every control cycle would reintroduce the latency we are trying to remove.

The forecaster comes from traffic modeling. Spatio-Temporal Graph Convolutional Networks stacks graph convolutions between gated temporal convolutions to forecast quantities on sensor graphs. T-GCN embeds graph convolution inside a recurrent cell. We reuse the STGCN block. The graph's nodes are qubits, and its signals are per-edge crosstalk estimates over time.

Two lines of work constrain the design. Linear data-driven control built on Willems' lemma, such as Data-Driven Model Predictive Control with Stability and Robustness Guarantees, needs superposition of trajectories. Unitary dynamics does not provide that as a function of pulse parameters, so we keep a physics model rather than a trajectory library. Ellipsoidal tube methods such as Recursively Feasible Stochastic Model Predictive Control using Indirect Feedback assume measurement does not disturb the system. We therefore make no recursive-feasibility claim.

Core Claim

A drifting ZZ-crosstalk field on a qubit lattice can be tracked and suppressed by a receding-horizon controller whose feedback uses only $k$-qubit patches.

To estimate all patch observables to additive error $\epsilon$, the controller needs local Pauli shadows. Take $n$ qubits, patches of size $k$, and $M \approx n\,4^k$ patch observables. The shot count per update is then

$$N_{\text{shots}} = O\!\left(\frac{3^k \log(n\,4^k)}{\epsilon^2}\right),$$

where the $3^k$ factor is the local-Pauli shadow norm of a $k$-body observable and the logarithm comes from a union bound over the $M$ targets, following Huang and Kueng. Full tomography instead needs order $4^n$ measurement settings. The nonlinear pulse optimization also splits into overlapping patch subproblems, which are coordinated by exchanging boundary pulse plans with neighboring patches.

We commit to one honest restriction about measurement backaction, the disturbance that measuring a quantum state causes. Measuring destroys the state being controlled, so "feedback" here happens across repeated preparations. Each control cycle runs the current pulse schedule many times: some shots go to shadow estimation and the rest carry the actual computation. This is the same operating regime as recalibration, but with a model-based loop that is predictive and runs continuously. It is also the regime the NMPC and qTOMPC papers implicitly assume when they feed measured states back in simulation.

Research Program

The plant is a superconducting lattice Hamiltonian with single-qubit drives. Each edge carries a ZZ coupling $J_e(t)$ that drifts over time, with nominal strength in the 0.03–0.05 range used in the Alibaba study. Drift is generated in two regimes. In the correlated regime, edges share a common drift source, such as a smooth spatial field over the chip, plus independent noise. In the independent regime, each edge drifts on its own. This split matters because the forecaster can only help in the first regime.

Estimation happens in two steps. Local Pauli shadows collected after short probe sequences give reduced density matrices on each edge and its neighbors. From these, per-edge couplings $\hat J_e$ are fit by local Hamiltonian learning, meaning a least-squares fit of the two-qubit evolution to the shadow estimates.

Forecasting uses an STGCN over the coupling graph. It takes the history of $\hat J_e$ and outputs $\hat J_e(t+h)$ for the planning horizon $h$, together with a quantile-based uncertainty band.

The control step is patch-wise NMPC solved by sequential quadratic programming. It uses the physics model with forecast couplings, subject to amplitude and slew limits, and it averages the cost over the forecast uncertainty band. Pulses on the syndrome-extraction schedule of Surface-17 are the decision variables. Simulation of the full lattice uses the tensor-network simulator from the Alibaba platform (Alibaba Cloud QDP), so that measured logical error rates are exact rather than sampled. Scaling studies extend to rotated surface-code patches of 25 and 49 qubits.

Baselines

The naive reference point is no crosstalk-aware control at all: standard fixed pulses on Surface-17 under drifting crosstalk. The published anchor for this setting is the logical phase-flip probability of about 0.6% reported by Alibaba Cloud QDP. That figure is 60% above the crosstalk-free model, which implies a crosstalk-free floor of roughly 0.375%. The floor is our arithmetic from their two numbers, not a reported figure. We rerun this baseline under our drift processes. Beating it means closing a substantial fraction of the gap between the rerun baseline and our simulated crosstalk-free floor.

The strongest existing method is NMPC with full state tomography as feedback, from Model predictive control for robust quantum state preparation. On its two-coupled-qubit crosstalk benchmark, the paper reports reaching the target with settling time above 20 ns. We found no reported fidelity number on Lacuna for that benchmark. We run it with an identical physics model, solver, and horizon, and at every lattice size we give it the same total shot budget as our method. At two qubits it should match or beat us, because it has the full state. Our method beats it if, at 9 or more qubits under equal shot budgets, it reaches lower logical error. We also report the lattice size at which the tomography baseline cannot be completed within the budget.

The third baseline is periodic recalibration plus open-loop pulse optimization in the style of GRAPE (gradient-based pulse design). Couplings are re-estimated from the same local shadows at fixed intervals, and pulses are re-optimized offline. The paper Robust Quantum Control via a Model Predictive Control Strategy reports that GRAPE's accumulated tracking error was 8.39, versus 0.185 for closed-loop qTOMPC, under periodic uncertainty on one qubit. That result shows the gap is large in principle, but we must measure it on lattices.

Two ablations isolate the key ingredients. The first replaces the STGCN forecast with persistence, meaning the latest estimate is used as the forecast. This isolates the value of spatiotemporal forecasting. The second replaces the graph convolution with per-edge temporal models that ignore neighbors. This isolates the value of spatial coupling in the forecaster. All forecasters get the same data and the same hyperparameter search budget.

Why not just…?

Why not just recalibrate often and use persistence? This is the most serious threat. If drift is slow relative to the control cycle, or independent across edges, the last estimate is nearly as good as any forecast. We expect the forecaster to add nothing in the independent-drift regime, and we will report that. The forecaster earns its place only if it beats persistence in the correlated regime at equal shot cost. One way it could do so is by letting neighbors' estimates stand in for fewer shots on each edge. If it fails to beat persistence even there, the contribution reduces to local-shadow NMPC with persistence, which is still new relative to the tomography-based NMPC.

Why not just fit a neural quantum state from shadows, as NSQST does, and control from that? Neural-Shadow QST handles high-weight observables on 40-qubit GHZ states, but it is an offline fit. Crosstalk suppression needs only two- and three-body quantities, where raw shadows already have bounded variance. We include a timing comparison, the wall-clock time per fit versus the drift timescale, to test whether a neural fit could fit inside the loop at all.

Why not use linear data-driven control, avoiding the physics model? Willems-lemma methods such as Berberich et al. require superposition of trajectories. Qubit rotations compose nonlinearly in the pulse parameters. We include a linearized controller around nominal pulses as a secondary check, and we expect it to degrade as crosstalk grows toward 0.05.

Decisive Checks

Four results would settle the claim. First, the logical phase-flip probability on Surface-17 under drifting crosstalk should fall well below the rerun no-control baseline and approach the simulated crosstalk-free floor. Second, the shots per update needed to hold a fixed logical error should grow roughly logarithmically from 9 to 49 data qubits, consistent with the scaling equation, while the tomography baseline's cost explodes. Third, the forecaster should beat persistence in the correlated-drift regime and tie it in the independent regime. A win in the independent regime would suggest the forecaster is exploiting something other than spatial structure. Fourth, as a sanity check, the controller should reproduce the qualitative closed-loop advantage over GRAPE-style open loop on the single-qubit periodic-uncertainty setting of qTOMPC.