Improved Methods for Determining Quantum Error Correcting Code Performance and Fault Tolerance

Source: arXiv:2607.27153 · Published 2026-07-29 · By Michael Mullan, Matthew Weippert, Winton Brown

TL;DR

This paper addresses a key challenge in quantum error correction: accurately estimating logical error rates at extremely low physical error probabilities required for utility-scale quantum computing. Monte Carlo methods become infeasible at these low rates due to enormous sampling demands, and traditional extrapolations based on code distance and minimum weight uncorrectable error patterns (w_min) can fail subtly if the actual w_min is lower than expected due to implementation details or decoder imperfections. The authors identify that typical logical failure patterns contain a small malignant core of errors surrounded by many easily correctable 'fluff' errors. Leveraging this structural insight, they introduce two main contributions: a pruning algorithm to remove fluff errors and isolate the minimal malignant core in failure patterns generated at higher error rates, and a novel subregion Metropolis-Hastings Markov chain Monte Carlo (MCMC) approach that interpolates between single-site MCMC and full Monte Carlo by resampling error subregions. Their subregion MCMC significantly improves mixing and convergence speed in sampling rare logical failures in full circuit-level simulations. Experiments with surface codes and concatenated Bacon-Shor codes demonstrate that their pruning effectively reveals minimum weight failures to validate fault tolerance assumptions, and subregion MCMC converges faster than prior methods, enabling more confident low logical error estimates.

Key findings

• Monte Carlo simulation to reach 10% confidence at logical error rate 10^-10 requires approximately 10^12 runs, making direct simulation infeasible for large codes.
• Extrapolation of logical error rates using P(p)=αp^w_min often fails due to unknown or implementation-dependent w_min, causing underestimates of failure rates.
• Typical failure patterns consist of a malignant core plus many easily correctable 'fluff' errors; fluff removal can reduce error patterns dramatically.
• Pruning algorithm reduces failure patterns to minimum size by iterative removal and retesting, needing 1000-10000 rounds and small removal subsets (<5 errors).
• At p=10^-3 for a distance-11 surface code (26,694 circuit locations), expected errors ~27, but minimum failure weight w_min=6 means fewer than 6 errors failure patterns are extremely rare (<10^-6 probability).
• Subregion MCMC introduces parameters region rate (p_r) and flip rate (p_f) to resample fractions of error patterns each step, interpolating between single-location MCMC and full Monte Carlo.
• Setting p_r=1/w_min and p_f=p_j (physical error rate) results in always accepted proposals that resample the core malignant error, dramatically improving Markov chain decorrelation times.
• Empirically, subregion MCMC mixes and converges much faster than the Bravyi-Vargo single-site Metropolis method, confirmed by coverage of multiple distinct logical failure cores in chains.

Threat model

The adversary is modeled by an independent, stochastic physical noise process applying Pauli errors with rate p_j at each circuit location, including gates and idle times. The adversary cannot manipulate the decoder or syndrome extraction circuits maliciously but noise models capture realistic error propagation and spreading. The defense is focused on accurately estimating logical error rates and minimal failure weight patterns under such noise rather than with an active adversary.

Methodology — deep read

Threat model & assumptions: The adversary is characterized by a noise model applying independent errors at circuit locations with rate p_j; errors are Pauli errors with assigned probabilities. The goal is to estimate logical failure rates in error correcting codes under realistic noise and imperfect decoding, particularly in low physical error regimes where direct Monte Carlo is infeasible. The decoder and syndrome extraction circuit implementation affect the minimal uncorrectable error weight w_min, which is unknown a priori and can cause extrapolation errors. The methodology assumes access to a full circuit noise model and decoding procedure but not explicit analytical access to w_min.

Data: The authors simulate full syndrome extraction circuits following the surface code and concatenated Bacon-Shor codes on a high-performance Quantum Virtual Machine (QVM). For each code, around d rounds of syndrome measurement are performed, with errors inserted at each circuit location under circuit depolarizing noise with equal error distribution among Pauli types. Datasets consist of Monte Carlo generated error patterns labeled by whether they cause logical failure after decoding. Sample sizes vary, with multiple Monte Carlo runs performed at several physical error rates (especially higher p) to generate failure patterns. Data splits and details on held-out sets are not described explicitly.

Architecture/algorithm: The pruning algorithm iteratively removes random subsets of errors from failure patterns and retests whether the remaining pattern still causes logical failure, identifying the minimal malignant core by removing fluff errors. The subregion MCMC algorithm generalizes the Bravyi-Vargo approach by resampling a random subregion of circuit locations at each step, parameterized by a region rate p_r that controls the expected fraction of resampled locations, and a flip rate p_f used in resampling error probabilities. The Metropolis-Hastings acceptance ratio formula is adapted for this proposal distribution. The special case p_r=1/w_min and p_f=p_j yields always accepted proposals that resample the malignant core efficiently.

Training regime: Not applicable for a classical sense of training; MCMC chains are run for many steps to achieve approximate convergence. The number of MCMC steps, number of chains, and burn-in procedures are heuristically chosen. Convergence diagnostics like the Gelman-Rubin metric R_hat are used by running multiple independent chains. The pruning algorithm is run for 1000-10000 rounds per error pattern removing small subsets.

Evaluation protocol: Logical error rates are estimated via the splitting method using multiple p_j levels, evaluating the ratio of logical failure probabilities at adjacent error rates from MCMC samples. The pruning algorithm is validated by comparing w_min discovered to expected theoretical distances and decoder performance. Comparison is made with the prior Bravyi-Vargo (BV) single-site MCMC algorithm via mixing speed and coverage of the logical failure space. No adversarial or distribution shift evaluation reported. Standard error bars are shown in figures for Monte Carlo estimates.

Reproducibility: The authors implement experiments in their internal QVM software; no public code release or frozen weights are mentioned. Data on exact simulation budgets and seeds is limited. Nonetheless, algorithms are described with sufficient detail for reproduction given similar simulators.

Concrete example: For a d=11 surface code with an intentionally faulty decoder (missing a hook), Monte Carlo simulations at high error p allow fitting Eq. (1) to estimate logical error rates and w_min. Pruning reduces failure patterns from ~27 errors to a minimal core of 6 which reveals the decoding fault. Subregion MCMC with p_r=1/6 rapidly explores failure modes with that core, improving convergence versus standard BV MCMC.

Technical innovations

• Introduction of a pruning algorithm to iteratively remove easily correctable errors ('fluff') from logical failure patterns to isolate a minimal malignant core.
• Development of subregion Metropolis-Hastings MCMC that resamples random subsets of the error pattern at each step with tunable parameters controlling proposal size and error resampling probabilities.
• Theoretical derivation showing acceptance ratio independence from region rate p_r and that setting flip rate p_f equal to physical error rate p_j yields acceptance probability one, improving sampling efficiency.
• Application and validation of pruning plus subregion MCMC on full circuit noise models for both surface and concatenated Bacon-Shor codes, demonstrating markedly faster mixing and better low error regime estimates.

Datasets

• Surface code simulations (distance 5,7,9,11,13) — sizes from ~2,000 to ~44,000 circuit locations — generated with internal QVM
• Two-level concatenated Bacon-Shor codes (3x3,5x5,7x7) — sizes 27,439 to 798,161 locations — generated with internal QVM

Baselines vs proposed

• BV MCMC single-site proposals: higher decorrelation time and slower convergence versus subregion MCMC with p_r=1/w_min, p_f=p_j
• Monte Carlo simulation: infeasible for logical error rates <10^-10 due to required ~10^12 runs; subregion MCMC enables estimation in this regime
• Pruning reduces failure error pattern sizes from tens or hundreds of errors (e.g., ~27 errors in d=11 surface code) down to minimum core sizes matching theoretical w_min (e.g., 6), greatly improving interpretability

Limitations

• No formal proof of MCMC convergence or mixing times for subregion method; empirical validation only
• Experiments limited to surface code and concatenated Bacon-Shor; generalization to other codes or noise models not demonstrated
• No adversarial attacks or decoder modifications evaluated to test robustness of w_min discovery
• Pruning algorithm requires running the decoder repeatedly on subsets, potentially expensive for very large codes
• Statistical error estimation for MCMC nearly relies on heuristic R_hat and lacks detailed variance analysis of correlated samples
• Code and data for reproducibility are not publicly released

Open questions / follow-ons

• How does subregion MCMC performance and pruning generalize to different quantum error correcting codes, noise models, or fault-tolerant schemes beyond surface and Bacon-Shor codes?
• Can more sophisticated MCMC proposals inspired by Hamiltonian or slice sampling further reduce correlation times in logical failure sampling?
• How can pruning and subregion MCMC be integrated into decoder design workflows to detect systematic decoder or implementation faults earlier?
• What are rigorous statistical error bounds for MCMC estimates given strong temporal correlation and rare event sampling in quantum error correction?

Why it matters for bot defense

For bot-defense and CAPTCHA system engineers, this paper demonstrates advanced statistical sampling methods to estimate extremely low failure probabilities where direct simulation is infeasible, similar to estimating rare attack success rates or adversarial bypass events under complex defenses. The pruning method to isolate minimal error cores can inspire techniques to identify minimal adversarial perturbations behind complex CAPTCHA bypasses. The subregion MCMC approach offers a mechanism to improve sampling efficiency over traditional Markov chain methods by targeting key regions of failure space, a principle applicable to modeling bot or adversarial interactions focused on critical vulnerabilities. Although quantum codes differ from classical CAPTCHA challenges, understanding how to extrapolate performance beyond direct measurements and design efficient rare-event samplers is valuable for bot-defense practitioners confronting subtle, low-frequency adversarial behaviors.

Cite

@article{arxiv2607_27153,
  title={ Improved Methods for Determining Quantum Error Correcting Code Performance and Fault Tolerance },
  author={ Michael Mullan and Matthew Weippert and Winton Brown },
  journal={arXiv preprint arXiv:2607.27153},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.27153}
}

Read the full paper

• arXiv:2607.27153 (abstract)
• arXiv:2607.27153 (PDF)