Simulation-Based Quantum System Inference with Neural Posterior Estimation

AI summary

Overview

Research area: Quantum characterization and machine learning—specifically simulation-based inference (SBI) with neural posterior estimation (NPE) applied to inverse problems in quantum physics.

Technical level: Intermediate to Advanced. The paper assumes familiarity with Bayesian inference, quantum noise channels, and generative models, though the framework itself is presented as a reusable template.

Scope (one sentence): The paper introduces a unified, likelihood-free framework that trains conditional normalizing flows on data from polynomial-cost classical quantum simulators to produce reusable, amortized posteriors for noise learning, error mitigation, state tomography, and Hamiltonian learning.

What This Paper Is About

Quantum models predict measurement outcomes from parameters, but the reverse direction—recovering states, noise processes, or Hamiltonians from measurement data—requires evaluating a likelihood that is intractable because the underlying Hilbert space grows exponentially. The authors' goal is to bypass that likelihood entirely: use cheap classical simulators to generate synthetic parameter–observation pairs, train a neural density estimator on them once, and then map any new experimental measurement record to a full posterior distribution in a single forward pass.

Key Contributions

1. A unified likelihood-free framework. The authors formalize "simulation-based quantum system inference" as a single template specified by a tuple of parameters, prior, observation, and simulator, pairing polynomial-cost classical simulators (Pauli propagation, tensor networks) with normalizing flows or other neural density estimators. One trained model serves unlimited future experiments, converting per-experiment inference into a fixed up-front cost.

2. Demonstration across four distinct inverse problems. The same template is applied to sparse Pauli noise learning, posterior-driven quantum error mitigation, quantum state tomography, and Hamiltonian learning in Rydberg arrays, spanning systems of up to 81 qubits and inference over as many as 735 latent parameters.

3. Posterior uncertainty as a diagnostic of non-identifiability. Rather than only reporting point estimates, the authors show that broad posterior distributions isolate exactly which parameters cannot be resolved—for example, the 6(n−2) error coefficients affected by gauge freedom in an n-qubit brickwall circuit—complementing analytic treatments and extending to circuits where the algebra becomes cumbersome.

4. A digital-twin strategy for machine-learning error mitigation. Noise samples drawn from the inferred posterior parameterize a classical simulator (an efficient Clifford/stabilizer simulator) that generates paired noisy–ideal training data, replacing the hardware data that ML-QEM normally requires.

Main Findings

• Amortized Pauli noise learning scales to 50 qubits. For a 50-qubit brickwall circuit with a 735-parameter noise model, an amortized NPE model trained on 30,000 simulations recovered most Pauli error coefficients tightly around ground truth, using 2,000 posterior samples. The latent vector has dimension 15(n−1) and the observation dimension is 12n−9; the Pauli-propagation simulator runs at O(n²) cost, approximately 0.1 s at n = 50 on a single CPU thread.

• Gauge freedom, not model failure, explains broad posteriors. Pauli errors from adjacent gates couple into gauge-equivalent combinations, so 6(n−2) coefficients in the n-qubit brickwall circuit are unidentifiable even with tomographically complete inputs and measurements. In the gauge-free subspace, convergence metrics reach high accuracy, with thresholds of R²_gauge-free ≥ 0.99 and ≥ 0.999 used to define the required budget.

• Training-data budget grows approximately linearly with system size. Extracted from sequential NPE runs starting from 10,000 simulations and incrementing by 5,000 per round, the data requirement grows roughly linearly from n = 10 to n = 50, suggesting characterization beyond 50 qubits for the sparse Pauli noise model.

• Drifting noise is tracked without per-instance calibration. Across ten runs under prescribed sinusoidal drifts, the estimator accurately tracked p_XX with low posterior uncertainty, systematically overestimated p_YY while capturing its phase, and gave a broad, out-of-phase estimate for p_IX—consistent with its behavior on static instances.

• QEM with inferred noise matches ground-truth-parameter performance. For the 15-qubit probe state and observable ⟨Z₅Z₆Z₇Z₈⟩, ZNE (linear, quadratic, and exponential extrapolation from gains G ∈ {1.6, 2.2, 2.8}, 2,000 random circuit samples per gain, 256 shots per instance) and PEC (converged at 2,000 quasiprobability samples) using posterior mean or median produced mitigated values close to those from ground-truth noise, with nearly identical PEC sampling overhead. In several cases inferred-noise results were closer to ideal, which the authors attribute to finite-sampling fluctuations rather than a systematic advantage.

• Digital-twin ML-QEM outperforms ZNE and extrapolates in time. On Trotterized 12-qubit Ising chains (δt = 0.1, 14 Trotter steps up to t = 1.4, 500 random Clifford variants per step, 50 test coupling strengths), the SBI-DT denoiser achieved improvement factors of 5.3-fold, versus 2.1-fold for quadratic ZNE and 1.2-fold for linear ZNE, relative to the noisy baseline—over twofold better than quadratic ZNE—while requiring no additional quantum samples at deployment and generalizing beyond the training boundary at t > 1.4 where ZNE degraded.

• Reusable state tomography without per-state optimization. For 12-qubit pure states from a single-layer PQC ansatz simulated exactly with bond dimension χ = 2 and 10⁴ shots, median reconstruction infidelity fell by more than two orders of magnitude to approximately 2×10⁻³ at the largest budget of 9×10⁵ simulations, with best-case infidelity below 1.5×10⁻³ throughout; 86% of all posterior samples over 200 test states achieved fidelity above 0.99. A complementary 4-qubit dense Cholesky parameterization reached median fidelities above 0.98 for most tested states.

• Hamiltonian learning localizes atom displacements. In simulated 9×9 Rydberg atom arrays, displacements inferred from local observables after short-time evolution agreed closely with prescribed values (R² = 0.983), and site-resolved posterior uncertainties flagged locations warranting further calibration.

Methodology in Plain English

The researchers treat the classical simulator as the statistical model. They draw parameter values from a prior, push each through a fast approximate simulator, and record the resulting synthetic measurement outcomes. This dataset of (parameters, outcomes) pairs implicitly encodes the likelihood without ever computing it.

A conditional normalizing flow—here, neural spline flows combining rational-quadratic splines with autoregressive transformations—is then trained by maximum likelihood to map an observation back to a distribution over parameters. Because the flow is invertible with a tractable Jacobian, the posterior density can be evaluated exactly, and samples can be drawn for any new observation.

Two operating modes exist: a fully amortized model trained once and reused for all future observations, and sequential NPE, which uses the previous posterior as a new proposal distribution to reduce simulation cost when only one observation matters. For error mitigation, the inferred noise posterior is fed into standard protocols (PEC quasiprobabilities, ZNE amplification channels) or into a digital twin that generates synthetic training pairs for a regression network; for that regression task, the full posterior is replaced by a direct point-prediction network because the categorical Clifford prior is ill-suited to normalizing flows and only a point correction is needed.

Why This Matters

The work reframes quantum characterization as a fixed up-front cost rather than a per-experiment expense, which matters because hardware time is expensive and characterization tasks such as real-time noise tracking and cross-device benchmarking must be repeated indefinitely.

Real-world applications:

• Processor benchmarking and calibration: gate-by-gate Pauli error posteriors inform calibration and show which parameters can and cannot be resolved from the chosen measurement protocol.
• Error mitigation on near-term devices: inferred noise models directly parameterize PEC and ZNE, and digital-twin training removes the hardware-data bottleneck for ML-QEM.
• Device drift monitoring and feedback: rapid tracking of drifting parameters could turn noise characterization into actionable feedback for stabilizing performance.
• Quantum sensing and array calibration: site-resolved displacement posteriors in Rydberg arrays indicate which atoms need recalibration.

Industry relevance: The demonstrated scale is comparable to leading hardware accessible to external researchers, including Google's Willow, IBM's Heron, and QuEra's Aquila. For suitably noisy or structured problems, the authors note that state-of-the-art classical simulators can surpass contemporary error-mitigated quantum computers in both runtime and precision, making the training pipeline practical today.

Future Directions

• Non-Markovian noise learning. The current demonstrations assume stationary Markovian Pauli noise held fixed across repeated circuit layers; extending to non-Markovian noise is explicitly named as an anticipated direction.
• Validation on real hardware. All observations, both training and ground truth, come from classical simulators, so the gap between simulated and physical noise—and the value of fine-tuning on real data—remains untested.
• Pushing past 100 qubits. The approximately linear growth of the training-data budget from n = 10 to n = 50 suggests a route beyond 100 qubits, but this is an extrapolation from coarse training increments, not a demonstrated result.
• Integration with control and sensing. Quantum sensor calibration and feedback control are proposed as natural extensions, alongside using inferred parameters to initialize gradient-descent tomography for higher reconstruction accuracy.

Target Audience

Quantum information researchers working on characterization, benchmarking, and error mitigation; machine learning researchers interested in scientific applications of simulation-based inference and normalizing flows; and quantum computing practitioners at the device or systems level who need reusable, uncertainty-aware tools for diagnosing hardware. Readers without background in Bayesian inference or quantum noise channels will find the methods section demanding but the application sections accessible.

Authors’ abstract

Models of quantum systems faithfully map system parameters to observations, but the inverse problem of parameter inference from measurement data presents a fundamental challenge: computationally intractable likelihoods due to an exponentially large Hilbert space. Here, we introduce simulation-based quantum system inference, a unified, likelihood-free framework that learns parameter posteriors directly from classical simulation data. The central idea is to pair polynomial-cost classical simulators, such as Pauli propagation and tensor networks, with normalizing flows or other neural density estimators for accurate, reusable inference. A single model, trained once, maps any new measurement record to its posterior in one forward pass---turning per-experiment inference into a fixed, up-front cost. We numerically demonstrate the framework's versatility across Pauli noise learning, quantum error mitigation, quantum state tomography, and Hamiltonian learning, with examples involving 81-qubit shallow circuits and 735-parameter inference. In each case, the approach yields accurate estimates of identifiable parameters, while posterior uncertainty provides additional diagnostics of non-identifiability and indicates where further characterization is needed. Our framework reduces data-acquisition requirements in quantum experiments and accelerates parameter inference, providing a practical route to characterizing and improving large-scale quantum systems.