Nonparametric Learning Non-Gaussian Quantum States of Continuous Variable Systems

1. Problem Statement

Continuous-variable (CV) quantum systems, such as optical fields (qumodes), are fundamental to quantum computing and sensing. However, reconstructing their quantum states is challenging due to the infinite-dimensional nature of the Hilbert space.

• Limitations of Traditional Methods: Standard approaches often rely on the Wigner function or density matrices. Reconstructing these from homodyne tomography data typically involves inverting the Radon transform, which is ill-posed and sensitive to sampling noise. This often leads to unphysical states (e.g., negative probabilities) or loss of non-classical features.
• Limitations of Parametric Methods: Many modern techniques use parametric models (e.g., Gaussian Mixture Models or Maximum Likelihood Estimation - MLE). These require prior assumptions about the state's structure (e.g., the number of modes or Gaussianity). They fail or perform poorly when applied to complex, non-Gaussian, or multimodal states (like cat states or GKP states) if the assumed model does not match the true state.
• The Gap: There is a lack of robust, non-parametric estimation techniques that can reconstruct arbitrary CV states (including non-Gaussian ones) from noisy data without prior knowledge, while providing rigorous convergence guarantees for trace characteristics (purity, overlap, trace distance).

2. Methodology: Kernel Quantum State Estimation (KQSE)

The authors propose Kernel Quantum State Estimation (KQSE), a non-parametric framework based on the symplectic tomographic representation of quantum mechanics.

Core Concepts

1. Symplectic Tomograms: Instead of working directly with the Wigner function (which can be negative), the method uses the symplectic tomogram $W(x|\mu, \nu)$, which is a genuine, non-negative probability density function (PDF) of the quadrature operator $X_{\mu,\nu} = \mu q + \nu p$.
2. Kernel Density Estimation (KDE): The tomogram is estimated non-parametrically using KDE. This avoids binning artifacts and model bias, allowing the method to adapt to complex, multimodal distributions.
3. Characteristic Function (CF) Estimation (KCFE): The method leverages the Fourier transform of the tomogram, known as the Characteristic Function (CF), denoted $\phi(t; \mu, \nu)$.
   • The CF uniquely determines the quantum state.
   • Crucially, the CF allows for the analytic calculation of trace quantities (purity, overlap, trace distance) and the reconstruction of the density matrix kernel $\rho(y, y')$ via inverse Fourier transforms.
4. Noise Filtering: The framework incorporates a deconvolution step to handle experimental noise (mode mismatch, detector inefficiency). By modeling the noise as an additive Gaussian component, the method filters noise in the CF domain by dividing by the known noise CF, ensuring stability.

The Algorithm

1. Data Collection: Measure rotated quadratures $X_{\mu,\nu}$ for various settings $(\mu, \nu)$.
2. CF Estimation: Estimate the CF of the signal from noisy data using a kernel-based estimator (KCFE) that includes a noise-correction factor.
3. State Reconstruction:
   • Density Matrix: Reconstruct the kernel $\rho(y, y')$ by integrating the estimated CF over the phase space parameters.
   • Trace Quantities: Calculate purity ($\text{tr}(\rho^2)$), overlap ($\text{tr}(\rho_1 \rho_2)$), and trace distance directly from the CF integrals.
4. Tail Correction: To handle the infinite integration domain, the method truncates the integral at a finite bound and uses parametric knowledge of the asymptotic decay (Gaussian tails) to correct the tail contribution, ensuring high precision.

3. Key Contributions & Theoretical Results

• Non-Parametric Universality: The method makes no assumptions about the state's shape (Gaussian or otherwise). It is proven to work for arbitrary quantum states, including highly non-Gaussian and multimodal ones.
• Convergence Rates:
  • The KQSE estimator achieves a near-optimal convergence rate of $\tilde{O}(T^{-1})$ for the squared error of the density matrix and trace quantities, where $T$ is the total number of measurements.
  • This matches the optimal rate of parametric MLE but without the model bias.
  • Theorem 3 establishes that to achieve precision $\epsilon$ in trace distance, the required sample size scales as $O(\epsilon^{-5/2})$, which is a universal bound for arbitrary states.
• Trace Distance Lower Bound: The authors prove that the trace distance between two quantum states is lower-bounded by the total variation distance between their corresponding tomograms. This provides a rigorous link between classical statistical distance and quantum distinguishability.
• Robustness to Noise: The built-in CF filtering mechanism allows the method to recover the true state characteristics even from noisy data, outperforming standard MLE which often fails under model mismatch.

4. Results

The paper validates the method through extensive simulations and experimental data analysis.

Simulation Studies (Coherent Cat States)

• Multimodality: The authors tested the method on Coherent Cat States (CCS), which have complex, multimodal tomograms.
• Comparison with MLE:
  • MLE: Parametric MLE (using Gaussian Mixture Models) failed to reconstruct the state accurately when the number of mixture components was misspecified (e.g., using 2 components for a 3-peak state). Even with a correct number of components, MLE showed higher errors.
  • KQSE: The non-parametric KDE/KCFE approach accurately reconstructed the multimodal tomogram and the density matrix kernel without any prior knowledge of the number of modes.
• Noise Handling: In noisy scenarios, the KQSE with noise correction recovered the state with errors nearly identical to the noiseless case, whereas uncorrected methods suffered significant degradation.

Experimental Validation (Schrödinger Kitten State)

• Dataset: The method was applied to real homodyne-tomographic data from a quantum optical experiment generating a "Schrödinger kitten" state (a superposition of vacuum and single-photon states).
• Performance:
  • The KQSE reconstruction yielded the lowest error ($L_\infty \approx 1.01 \times 10^{-2}$) compared to MLE with 1 or 2 Gaussian components.
  • MLE with a single component failed completely, and even the 2-component MLE showed visible deviations from the theoretical kernel.
  • The KQSE successfully captured the asymmetry and non-Gaussian features of the experimental data that parametric models missed.

5. Significance

• Scalability for Quantum Technologies: As quantum technologies move toward error correction and universal computation using non-Gaussian states (e.g., GKP codes, cat codes), robust state characterization is critical. KQSE provides a reliable tool for benchmarking these states without requiring complex, error-prone parametric modeling.
• Data-Driven Flexibility: By removing the need for prior assumptions about the state's structure, KQSE is ideal for "black-box" characterization of unknown quantum systems or systems where the theoretical model is incomplete.
• Theoretical Rigor: The paper provides a mathematically rigorous foundation for non-parametric quantum state estimation, establishing convergence rates and error bounds that were previously lacking for general CV states.
• Practical Robustness: The ability to filter noise directly in the characteristic function domain makes the method highly suitable for real-world experiments where detector inefficiencies and losses are unavoidable.

In summary, this work bridges the gap between classical non-parametric statistics and quantum state tomography, offering a robust, optimal, and assumption-free method for learning the complex quantum states essential for next-generation quantum information processing.