Technical Summary: Virtual Error Cancellation for Logical Quantum Circuits

Problem Statement

Early fault-tolerant quantum processors will operate with moderate code distances, inevitably retaining non-negligible residual logical error rates. While increasing code distance suppresses these errors, it incurs steep hardware overheads. Logical quantum error mitigation offers a pathway to suppress these residual errors without additional physical qubits, but existing techniques face three primary constraints:

1. Noise Characterization: Many methods (e.g., Probabilistic Error Cancellation, PEC) require detailed prior noise tomography.
2. Circuit Modification: Techniques like active PEC necessitate inserting quasi-probability operations or modified circuits during execution.
3. Sampling Overhead: Mitigation often incurs substantial sampling costs. While syndrome-based postselection avoids circuit modifications, it introduces systematic bias by discarding data. Conversely, standard syndrome-aware PEC can be unbiased but often suffers from high sampling overheads (scaling with $\eta \approx 4$ for logical PEC or $\eta \approx 2$ for single-decoder approaches) and requires noise characterization.

The paper addresses the challenge of eliminating logical bias in universal quantum circuits without requiring prior noise characterization, circuit modifications, or the systematic bias inherent in postselection, while minimizing sampling overhead.

Methodology: Virtual Error Cancellation (VEC)

The authors propose Virtual Error Cancellation (VEC), a method that performs logical error mitigation entirely in classical post-processing using only the syndrome records generated during standard fault-tolerant execution.

Core Principles

1. Post-Processing Reweighting: Instead of discarding shots (postselection) or modifying circuits (PEC), VEC assigns a classical filter weight $w(s_i)$ to each shot based on its recorded syndrome $s_i$. The estimator is constructed as a weighted average: $\langle O \rangle_{\text{VEC}} = N^{-1} \sum w(s_i) m_i$.
2. Unbiasedness Conditions: To ensure the estimator is unbiased, the weights must satisfy two conditions derived from the requirement that the syndrome-averaged logical channel reduces to the identity:
   • Trace Preservation: $E_s[w(s)] = 1$.
   • Error Cancellation: $E_s[w(s)p_\alpha(s)] = 0$ for all logical error channels $\alpha$, where $p_\alpha(s)$ is the conditional error probability.
   • This necessitates negative weights for error-prone syndromes to cancel positive contributions on average.
3. Dual-Decoder Architecture: To avoid the intractable computation of exact posterior probabilities for every syndrome (which is $\#P$-hard), the authors introduce a coarse-grained protocol using two decoders:
   • Real-Time Decoder ($D_{rt}$): A fast, low-latency decoder (e.g., Minimum-Weight Perfect Matching) used during the experiment to determine the Pauli frame and feedforward.
   • Offline Decoder ($D_{off}$): A high-complexity, accurate decoder (e.g., Maximum-Likelihood or Tensor Network) run in post-processing.
   • Sector Partitioning: Syndromes are partitioned into sectors based on the agreement or disagreement between $D_{rt}$ and $D_{off}$. The consensus sector ($D_{rt} = D_{off}$) contains the majority of shots with suppressed errors. Disagreement sectors ($D_{rt} \neq D_{off}$) identify shots where the real-time decoder likely erred.
4. Noise Learning: Instead of separate calibration circuits, VEC learns the underlying fault distribution (Detector Error Model) directly from the stored syndrome records using linear inversion of syndrome fidelities. This allows the generation of mitigation weights entirely from experimental data.

Theoretical Framework

The paper establishes that for a "gapless" regime (where competing logical error chains have equal minimal weight, common in joint surface-code decoding), the optimal sampling overhead scales linearly with the unmitigated logical error rate ($C^* - 1 = O(p_L)$). This corresponds to an overhead factor $\eta \to 1$ in the large-distance limit, significantly outperforming standard logical PEC ($\eta \approx 4$) and single-decoder syndrome-aware PEC ($\eta \approx 2$).

Key Contributions

1. Virtual Error Cancellation Protocol: A complete framework for logical error mitigation that operates solely in classical post-processing, requiring no circuit modifications or prior noise characterization.
2. Dual-Decoder Implementation: A scalable scheme that pairs a real-time decoder with an offline reference decoder to define coarse-grained sectors. This approach preserves nearly all optimal statistical contrast while avoiding the computational cost of per-syndrome Bayesian inference.
3. End-to-End Noise Learning: A method to learn detector error models and logical error rates directly from syndrome histories, eliminating the need for separate calibration experiments.
4. Theoretical Bounds:
   • Proof that the optimal syndrome-resolved filter minimizes sampling overhead.
   • Demonstration that the dual-decoder coarse-grained protocol approaches this theoretical optimum.
   • Derivation of the sampling overhead scaling, showing a quadratic advantage over syndrome-aware PEC and a reduction to $\eta \approx 1$ in the favorable large-distance limit.
5. Residual Error Analysis: Theoretical derivation showing that residual bias due to finite-sample noise mischaracterization is suppressed by the agreement error suppression ratio $\alpha$, making the method more robust against calibration errors than physical PEC or postselection.

Results

Numerical simulations were performed on a distance-7 rotated surface code under code-capacity depolarizing noise:

• Error Suppression: The full-stack dual-decoder algorithm reduced the residual logical error rate by more than three orders of magnitude compared to the decoded logical error rate in an experimentally achievable setting.
• Comparison to Postselection: The method achieved more than two orders of magnitude better suppression compared to postselection based on the agreement of two decoders.
• Sampling Overhead: The excess sampling overhead ratio $\eta = (C^* - 1)/p_L$ approached $\approx 1.3$ at distance $d=7$, significantly lower than the $\eta=2$ ceiling of single-decoder syndrome-aware PEC and the $\eta=4$ of standard logical PEC.
• Noise Learning Accuracy: The learned physical and logical error rates showed high fidelity (median relative error $< 0.35\%$) across multiple noise realizations, validating the feasibility of learning from syndrome data alone.

Significance and Claims

The paper claims to establish syndrome records as a resource for correcting universal computational estimates after quantum execution. By moving mitigation entirely to post-processing, VEC decouples the error correction workflow from real-time hardware latency constraints.

The authors position VEC as a solution that bypasses the first two constraints of logical error mitigation (noise learning and circuit modification) completely. They argue that by leveraging the rich information in syndrome records and the differential power of dual decoders, the method achieves near-optimal sampling efficiency while eliminating the systematic bias of postselection. The results suggest that early fault-tolerant algorithms can achieve significantly higher fidelity without the steep hardware overhead of scaling code distances or the experimental complexity of active error cancellation circuits.

The work is presented as a foundational step toward scalable quantum computing, offering a practical pathway to suppress residual logical errors in the near term. The authors note that while the analysis focuses on Pauli noise models, the framework generalizes to circuit-level noise and can be integrated into existing quantum compilers to co-optimize code distance, decoder pairing, and sampling budgets.