Noise-Robust Estimation of Quantum Observables in Noisy Hardware

1. Problem Statement

Quantum computing on Noisy Intermediate-Scale Quantum (NISQ) devices is hindered by hardware noise, which corrupts expectation values and limits computational utility. While Quantum Error Mitigation (QEM) techniques exist, they face significant challenges:

• Noise-Aware Methods (e.g., Probabilistic Error Cancellation): Require precise noise characterization, which is difficult to learn and prone to drift, leading to residual bias.
• Noise-Agnostic Methods (e.g., Zero-Noise Extrapolation - ZNE):
  • Model Mismatch: They often assume simple noise decay models (e.g., single-exponential), whereas real non-Clifford circuits often follow complex multi-exponential decays, introducing bias.
  • Imperfect Amplification: Techniques like unitary folding do not scale noise uniformly because experimental noise channels often do not commute with circuit unitaries.
  • Sampling Overhead: Many methods incur exponential sampling costs as circuit size or error rates increase.

There is a critical need for a noise-agnostic framework that systematically suppresses estimation bias without requiring detailed noise modeling, while maintaining manageable sampling overhead.

2. Methodology: Noise-Robust Estimation (NRE)

The authors introduce Noise-Robust Estimation (NRE), a two-stage post-processing protocol that combines a target circuit with a structurally similar "noise-canceling" companion circuit.

Core Concept

NRE constructs an auxiliary estimator ($A$) that is less sensitive to noise than the raw observable. The framework relies on the discovery that the magnitude of the residual bias in the baseline estimator is correlated with the normalized dispersion of this auxiliary estimator across amplified noise realizations.

The Two-Stage Protocol

Stage 1: Baseline Estimation

• Circuit Construction:
  • Target Circuit ($t$): The circuit of interest.
  • Noise-Canceling Circuit ($ncc$): A circuit with the same gate structure and connectivity as the target, but with non-Clifford gates replaced by the "closest" Clifford gates (minimizing Frobenius norm). This allows the ideal noiseless expectation value $\langle O \rangle_{ncc}$ to be computed classically.
• Data Collection: Measurements are taken at $M$ different noise scale factors ($\lambda_i$), typically using global unitary folding.
• Auxiliary Quantity Construction:
  The auxiliary quantity $A$ is defined as:
  $$A_{ncc}(n, \lambda_i) = P_1(\lambda_i) + n P_2(\lambda_i)$$
  Where:
  • $P_1(\lambda_i) = \langle \tilde{O} \rangle_t(\lambda_i) \cdot \frac{\langle O \rangle_{ncc}}{\langle \tilde{O} \rangle_{ncc}(\lambda_i)}$ (Corrects for multiplicative noise).
  • $P_2(\lambda_i) = \log\left(\frac{\langle O \rangle_{ncc}}{\langle \tilde{O} \rangle_{ncc}(\lambda_i)}\right)$ (Captures multiplicative deviations in log-space).
  • $n$ is a tunable control parameter.
• Optimization: The control parameter $n$ is optimized ($n_{op}$) such that the Taylor expansion of $A$ around the smallest noise scale ($\lambda_1$) eliminates the leading-order bias terms. This yields the Baseline Estimator ($\langle O \rangle_{b-NRE}$).

Stage 2: Bias–Dispersion Correlation & Extrapolation

• Diagnostic Metric: The Normalized Dispersion ($D$) is calculated as the Mean Absolute Deviation (MAD) of the auxiliary estimator $A$ across noise scales, normalized by the MAD of the raw target observable.
  $$D = \frac{\text{MAD}[\{A(\lambda_i)\}]}{\text{MAD}[\{\langle \tilde{O} \rangle_t(\lambda_i)\}]}$$
• Theoretical Insight: Analytical derivation shows that the residual bias of the baseline estimator is governed by the same hierarchy of derivatives that controls the variation of $A$. Therefore, a smaller $D$ implies a smaller residual bias.
• Extrapolation:
  1. Bootstrapping: Generate an ensemble of baseline estimators and their corresponding $D$ values by resampling experimental counts.
  2. Regression: Perform a weighted regression of the baseline estimators against $D$.
  3. Final Estimate: Extrapolate to the zero-dispersion limit ($D \to 0$) to obtain the final NRE estimate ($\langle O \rangle_{NRE}$). This effectively filters out realizations with high bias.

3. Key Contributions

• Novel Framework: Introduction of NRE, a noise-agnostic method that does not require explicit noise modeling or training on near-Clifford circuits (unlike CDR/vnCDR).
• Diagnostic Metric: Discovery and validation of the correlation between the normalized dispersion of an auxiliary estimator and the magnitude of the residual bias. This provides a measurable, data-driven way to quantify and mitigate bias.
• Two-Stage Post-Processing: A structured approach that first constructs a robust baseline and then uses statistical extrapolation to the $D \to 0$ limit to suppress remaining bias.
• Efficiency: The method requires only one noise-canceling circuit per target circuit, avoiding the large training sets required by regression-based methods.

4. Experimental Results

The authors validated NRE on a 20-qubit IQM Garnet superconducting quantum processor using circuits with up to 480 entangling CZ gates.

• Benchmark 1: Transverse-Field Ising Model (TFIM)

  • Setup: 20 qubits, 240 CZ gates, varying noise scales.
  • Performance: NRE consistently achieved substantially lower bias compared to ZNE, CDR, vnCDR, and the Urbanek et al. method.
  • Robustness: Unlike ZNE, which is sensitive to the choice of noise scale factors and extrapolation models, NRE provided stable estimates across various noise configurations.
  • Extreme Noise: In highly noisy regimes (where target values were suppressed by ~80%), NRE remained the only method capable of recovering an accurate estimate (within ~10% of ideal).

• Benchmark 2: Quantum Chemistry (H4 Molecule)

  • Setup: 8 qubits, deep circuits (up to 264 depth), high-weight observables (Pauli weight up to 6).
  • Performance: Despite strong noise effects (target values suppressed by ~70-90%), NRE restored the ground-state energy close to the ideal value, achieving chemical precision (error < 1.59 mHa) in optimal settings.
  • Comparison: NRE outperformed ZNE and the Urbanek method, which failed to recover accurate energies in deep, noisy circuits.

• Sampling Overhead Analysis

  • Numerical simulations showed that while NRE has a higher sampling overhead than ZNE (approx. 7x), the second post-processing step (extrapolation to $D \to 0$) reduces the overhead by a factor of ~6.2 compared to the baseline estimator alone.
  • The method maintains a moderate overhead while providing significantly better bias suppression than existing techniques.

5. Significance and Impact

• Practicality for NISQ: NRE offers a practical, broadly applicable strategy for extracting reliable results from current noisy hardware without the need for complex noise tomography.
• Robustness to Imperfections: The method is inherently robust to imperfect noise amplification (a common failure point in ZNE) because the extrapolation to $D \to 0$ naturally compensates for scaling mismatches.
• Scalability: By requiring only a single companion circuit and avoiding exponential growth in training data, NRE is more scalable than data-driven regression methods like CDR.
• Future Directions: The paper outlines potential integration with error correction (logical-level NRE) and the use of machine learning to optimize the noise-canceling circuit construction and regression strategies.

In conclusion, NRE represents a significant advancement in error mitigation by leveraging a measurable statistical diagnostic (dispersion) to systematically suppress bias, establishing a new standard for reliable quantum computation on noisy hardware.