Spacetime mitigation of logical errors

AI summary

Overview

Research area: quantum error mitigation and error detection on near-term superconducting quantum processors. Technical level: Advanced. Scope: This paper develops a spacetime Pauli-Lindblad framework for modeling post-selected logical noise, combines error detection with probabilistic error cancellation (PEC), and validates the protocol experimentally on the ibm_aachen processor.

What This Paper Is About

Error detection can suppress noise by rejecting runs with nontrivial syndromes, but the surviving runs still contain residual logical errors that bias results. PEC can correct modeled bias, but its sampling cost grows with circuit error strength. The paper combines the two so that post-selected logical errors are mitigated at lower sampling cost than PEC alone, and it provides a practical way to construct and normalize the required correction on real hardware.

Key Contributions

1. A spacetime Pauli-Lindblad representation of post-selected logical noise that combines a sparse physical noise model with the circuit’s syndrome information, using an auxiliary Pauli space with one tensor factor per circuit layer to track fault locations.
2. A perturbative construction of the post-selected logical noise generator in which individually undetected faults contribute exactly at first order, while detected faults with canceling syndromes generate higher-order terms coupling distinct spacetime locations.
3. A linearized post-selection protocol that resolves the mismatch between nonlinear post-selection and linear PEC by working with unnormalized post-selected maps, multiplying each post-selected expectation by its post-selection probability before averaging, and normalizing only after averaging.
4. An experimental validation and demonstration on ibm_aachen, including Clifford-circuit validation of the noise model through second order and first-order spacetime PEC for non-Clifford transverse-field Ising dynamics using 22 data and 27 check qubits.

Main Findings

• Clifford noise-model validation: On a random 3-regular graph state prepared on 20 data qubits with d = 10 brickwork repetitions, and a checked 24-qubit ring with d = 12 repetitions and four checks, the mean relative deviations between predicted and measured stabilizers were 3.4% without checks, 7.3% with four checks but no post-selection, and 3.5% with four checks and post-selection.
• Perturbative scaling confirmed: Under rate rescaling λ_ν → c λ_ν, the mean absolute truncation error across the 20 stabilizers scaled as c² at first order and c³ at second order. At the physical model strength c = 1, including the second-order generator term reduced this error from 0.04 to 0.01.
• Predicted sampling-cost reduction in the Clifford experiment: Without checks, Γ_PEC = 105.4. Post-selecting on four checks reduced γ_L² to 24.4, and with measured post-selection probability α = 0.2963, the combined overhead was Γ_ED+PEC = 82.3, approximately 22% lower.
• Non-Clifford Ising demonstration: ED+PEC recovered the ideal mean magnetization within plotted statistical uncertainty at all three Trotter depths d = 2, 4, 6. At d = 6, standalone PEC deviated strongly from the ideal value and had a much larger uncertainty.
• Site-resolved improvement: At d = 4, error detection reduced bias relative to the unmitigated result but left a residual discrepancy, while ED+PEC closely followed the ideal site-resolved magnetizations across the lattice.
• Sampling-overhead savings: Relative to standard PEC, ED+PEC reduced the total sampling overhead, including the cost of syndrome post-selection, by factors of 3.7×, 15.9×, and 63× at d = 2, 4, and 6, respectively. The corresponding overheads were 44.1 versus 12.1, 1941 versus 122, and 85545 versus 1359.
• Post-selection model consistency: The learned models reproduced the measured post-selection rates to about one percentage point.
• Residual-bias control: First-order inversion leaves higher-order contributions, but the second-order contribution to the bias was bounded and remained smaller than the statistical error for all considered observables.
• Nonlinearity demonstrated analytically: An amplitude-damping example shows that separately normalizing each randomized post-selected circuit produces a nonlinear map and residual biases of order O(p²) at small damping p, while rescaling by post-selection probabilities preserves the linear logical channel and permits an exact PEC inverse.

Methodology in Plain English

The circuit is split into n data qubits and k check qubits. The checks are designed so that ideal operation returns them to |0⟩, while faults can flip their measured syndrome. Runs with a nontrivial syndrome are discarded. The authors model physical noise as sparse single-qubit and nearest-neighbor two-qubit Pauli faults occurring after each circuit layer, learn the corresponding Pauli-Lindblad rates from calibration circuits, and then work out what the noise looks like after post-selection.

Because post-selection can make individually detectable faults survive when their syndromes cancel, the surviving logical noise is not simply the original noise with some terms removed. The authors represent faults at different circuit locations in an auxiliary space, take the logarithm of the post-selected noise channel to isolate connected contributions, and truncate the resulting expansion. Undetected faults enter at first order; pairs of detected faults with canceling syndromes enter at second order; higher orders involve larger syndrome-canceling combinations.

To apply PEC, they sample the inverse of this truncated logical noise channel, inserting Pauli corrections at the relevant spacetime locations and weighting measurement outcomes accordingly. The normalization issue is handled by keeping post-selected maps unnormalized during averaging and dividing by the overall post-selection probability only at the end. The protocol is tested first with Clifford circuits, where predictions can be evaluated efficiently, and then with Trotterized transverse-field Ising dynamics on a six-plaquette hexagonal lattice embedded in the heavy-hex topology of ibm_aachen. The Ising circuit uses n = 22 data qubits and k = 27 check qubits, with φ = -π/4, θ = 3π/8, d = 2, 4, 6 Trotter steps, and up to 648 CZ gates. The authors also apply a non-Markovian-error filter for leakage and TREX rescaling for state-preparation and measurement errors.

Why This Matters

The work shows that error detection and error mitigation can be combined in a way that is both experimentally practical and theoretically consistent, rather than treating post-selection as a separate preprocessing step. It provides a route to lower sampling overhead for near-term quantum computations and connects syndrome information directly to logical-noise modeling.

Real-world applications:

• Quantum simulation of spin systems: The demonstrated transverse-field Ising dynamics is a standard model for magnetic materials and interacting spin systems.
• Near-term hardware workflows: The protocol offers a way to improve accuracy on superconducting processors without full fault tolerance.
• Resource-aware quantum computing: Lower sampling overhead can reduce the number of shots and the cost of running quantum workloads.
• Foundation for error-corrected workflows: The spacetime noise model could inform mitigation of residual logical errors after active error correction.

Industry relevance: The experiments were performed on IBM’s ibm_aachen superconducting processor, and the results are the subject of a patent application (Application No. 19/365421, filed 22 Oct 2025). The approach targets practical accuracy and sampling-cost improvements for cloud-accessible quantum hardware.

Future Directions

• Midcircuit measurements: Extending the protocol from terminal checks to midcircuit post-selection, and optimizing how frequently such measurements are performed.
• Measurement and reset noise: Developing a consistent treatment of midcircuit-measurement and reset noise, building on recent noise-learning techniques.
• Post-error-correction mitigation: Extending spacetime PEC to residual logical errors after active error correction, using higher-order expansions organized by the logical action of faults after decoding.
• Non-Markovian logical dynamics: Capturing corrections to effective Markovian descriptions and correlations from physical noise with memory, including correlated correction cycles.

Target Audience

Quantum computing researchers and graduate students working on error mitigation, error detection, quantum error correction, and superconducting hardware. The paper is most useful to readers with background in Pauli noise models, probabilistic error cancellation, syndrome measurements, and quantum simulation; experimentalists and algorithm developers evaluating near-term accuracy-versus-overhead tradeoffs will also benefit.

Authors’ abstract

Error detection can suppress noise at a lower sampling cost than probabilistic error cancellation (PEC), but leaves residual logical errors. We combine the two approaches to mitigate these errors while lowering the cost of PEC. We introduce a spacetime Pauli-Lindblad representation and use it to construct the post-selected logical noise perturbatively from physical noise and syndrome information. Individually undetected faults contribute at first order, while detected faults with canceling syndromes generate higher-order terms coupling distinct spacetime locations. To reconcile the nonlinearity of post-selection with the linearity of PEC, we formulate the combined protocol in terms of unnormalized post-selected maps, with normalization performed only after averaging. On the ibm_aachen superconducting quantum processor, Clifford experiments validate the post-selected noise model and its perturbative scaling. We then demonstrate error detection with first-order PEC for transverse-field Ising dynamics using 22 data and 27 check qubits. The combined protocol recovers the mean magnetization within statistical uncertainty through six Trotter steps and reduces the inferred sampling overhead by up to a factor of 63 relative to PEC without post-selection.