Physics-Inspired Extrapolation for efficient error mitigation and hardware certification

1. Problem Statement

Quantum computers in the Noisy Intermediate-Scale Quantum (NISQ) era and the emerging Early Fault-Tolerant (EFT) era suffer from significant noise, preventing the execution of complex, industry-relevant algorithms. While Quantum Error Correction (QEC) is the ultimate solution, it requires prohibitive resource overheads (thousands of physical qubits per logical qubit). Consequently, Quantum Error Mitigation (QEM) is essential for improving accuracy on current hardware.

Existing QEM methods face critical limitations:

• Probabilistic Error Cancellation (PEC): Provides unbiased estimates but incurs an exponential sampling overhead, making it impractical for large circuits.
• Zero-Noise Extrapolation (ZNE): A popular heuristic method that amplifies noise and extrapolates to the zero-noise limit. However, standard ZNE relies on heuristic fitting functions (linear, polynomial, or exponential) without a rigorous theoretical foundation. This leads to:
  • Uncertainty in choosing the correct extrapolation function.
  • Lack of physical interpretation for fitting parameters.
  • High variance in estimates, particularly with exponential extrapolation.
  • Inability to quantify the distance between noisy and ideal circuits (hardware certification) without additional overhead.

2. Methodology: Physics-Inspired Extrapolation (PIE)

The authors propose Physics-Inspired Extrapolation (PIE), a linear circuit runtime protocol derived from the Error Mitigation by Restricted Evolution (EMRE) framework. Unlike heuristic ZNE, PIE is grounded in the operational interpretation of quantum channels.

Core Theoretical Framework:

• Generalized Quasi-Probabilistic Decomposition (GQPD): PIE leverages the concept that an ideal unitary operation $U$ can be expressed in terms of a noisy operation $E \circ U$ and a noise channel $M$:
  $$U = s_{opt} (E \circ U) - (s_{opt} - 1)M$$
  Here, $s_{opt}$ is the max-relative entropy between the ideal and noisy operations.
• Approximation: By restricting the evolution to the positive component of the decomposition, the ideal expectation value $\langle O \rangle_{ideal}$ is approximated as:
  $$\langle O \rangle_{ideal} \approx s \cdot \langle O \rangle_{noisy}$$
  where $s$ is the max-relative entropy.
• Noise Amplification via Folding: To create a dataset for extrapolation, the authors use circuit folding (inserting identity-equivalent operations $U^\dagger U$) to increase the effective noise level. If the circuit is folded $n$ times, the noise scaling factor becomes $\lambda = 2n + 1$.
• The Extrapolation Function: Based on the theoretical derivation, the relationship between the noisy expectation value and the noise scaling factor $\lambda$ is:
  $$\langle O \rangle_{err} \approx \frac{\langle O \rangle_{ideal}}{s^\lambda}$$
  Taking the logarithm yields a linear relationship:
  $$\log(\langle O \rangle_{err}) = \log(\langle O \rangle_{ideal}) - \lambda \log(s)$$
  This allows for a linear regression where:
  • The intercept ($\lambda=0$) gives the log of the ideal expectation value.
  • The slope corresponds to $-\log(s)$, where $s$ is the max-relative entropy.

Key Advantages over Standard ZNE:

1. Linear Runtime: The protocol scales linearly with the number of data points, avoiding the exponential overhead of PEC.
2. Theoretical Grounding: The fitting parameters have clear physical meanings (unlike heuristic ZNE).
3. Low Variance: The linear fit is numerically stable and produces lower variance estimates compared to non-linear exponential fits.
4. Hardware Certification: The slope of the fit directly quantifies the max-relative entropy, providing a metric for how far the hardware deviates from the ideal, effectively certifying hardware performance without extra experiments.

3. Key Contributions

• Derivation of PIE: A rigorous derivation of an extrapolation function based on max-relative entropy, moving beyond heuristic curve fitting.
• Dual Functionality: The method simultaneously performs error mitigation (estimating ideal values) and hardware certification (quantifying noise distance via the slope).
• Experimental Validation: Successful implementation on IBM Eagle (ibm_kyiv) and IBM Heron (ibm_torino) superconducting processors.
• Scalability Demonstration: Efficient simulation of 84-qubit quantum dynamics (transverse-field Ising model), a scale previously difficult to mitigate accurately.
• Comparison with Alternatives: Benchmarking against standard ZNE (linear, quadratic, exponential) and learning-based methods (Clifford Data Regression), showing superior stability and accuracy in low-to-moderate noise regimes.

4. Results

The authors validated PIE through numerical simulations and real hardware experiments:

• Ising Model Simulation (84 Qubits):

  • On the IBM Eagle processor, PIE was used to estimate the global magnetization of an 84-qubit Ising chain.
  • Accuracy: PIE results were consistent with noise-free benchmarks and outperformed standard ZNE methods in terms of statistical stability.
  • Variance: The variance of the PIE estimate was significantly lower than that of exponential ZNE, which often suffers from large confidence intervals due to non-linear fitting instability.
  • Depth Scaling: PIE remained the most accurate protocol for low-to-moderate circuit depths. While accuracy degrades at very high depths (due to the growing contribution of the neglected noise channel $M$), it still outperformed quadratic extrapolation.

• Quantum Chemistry (H2 and LiH):

  • PIE was applied to calculate ground-state energies for Hydrogen (4 qubits) and Lithium Hydride (12 qubits).
  • H2: PIE mitigated the energy to $-1.86 \pm 0.01$ Ha, closely matching the reference value of $-1.852$ Ha and the noise-free simulation of $-1.855 \pm 0.005$ Ha.
  • LiH: PIE improved the result from $-7.90$ Ha (unmitigated) to $-8.8 \pm 0.4$ Ha, approaching the reference of $-8.875$ Ha.

• Hardware Certification:

  • The slope of the PIE fit was used to compare different quantum processors. A shallower slope indicated a hardware device closer to the ideal (lower max-relative entropy), while steeper slopes indicated higher noise susceptibility. This was demonstrated on different IBM processors without additional computational cost.

• Comparison with Learning-Based Methods:

  • Compared to Clifford Data Regression (CDR), PIE offered comparable accuracy in low-noise regimes but avoided the heavy computational overhead required to train CDR models on classically simulable circuits.

5. Significance and Impact

• Bridging NISQ and EFT: PIE is particularly well-suited for the "Early Fault-Tolerant" era, where logical qubits exist but still suffer from residual noise. It offers a scalable strategy that does not require the massive overhead of full QEC or the exponential cost of PEC.
• Operational Interpretability: By linking the extrapolation slope to max-relative entropy, PIE transforms error mitigation from a "black box" data-fitting exercise into a tool for quantitative hardware diagnostics.
• Resource Efficiency: The linear scaling and low variance make PIE a practical candidate for near-term applications in quantum chemistry, materials science, and optimization, where high-fidelity results are needed from noisy hardware.
• Future Directions: The authors suggest extending PIE to handle variable noise across circuits and integrating it with randomized compiling to further suppress coherent errors.

In conclusion, Physics-Inspired Extrapolation (PIE) represents a significant step forward in quantum error mitigation by providing a theoretically grounded, efficient, and dual-purpose protocol that enhances both the accuracy of quantum computations and the certification of quantum hardware.