Machine learning of quantum data using optimal similarity measurements

1. Problem Statement

Quantum Machine Learning (QML) aims to leverage quantum states to process data with potential advantages over classical methods. A fundamental operation in QML is evaluating the similarity (or overlap) between two quantum states, defined as $|\langle \psi_A | \psi_B \rangle|^2$. This metric serves as the kernel function in quantum classifiers and the cost function in training quantum models.

However, a critical bottleneck exists:

• Classical Intractability: Determining the overlap of unknown quantum states typically requires full tomographic reconstruction of individual states, which scales exponentially with the system dimension (number of modes or photons).
• Distributed Limitations: In quantum networks where data is transmitted as quantum states (e.g., from sensors or simulators), parties often cannot share copies of the states. Distributed strategies relying on local measurements and classical communication also suffer from exponential sample complexity for high-dimensional states.
• Hardware Overhead: Existing methods often require deep quantum circuits or entanglement-assisted protocols that are difficult to scale experimentally.

The paper addresses the need for a sample-optimal, hardware-efficient protocol to estimate quantum state overlap directly, without characterizing individual states, to enable scalable QML.

2. Methodology

The authors propose and experimentally validate a protocol based on bosonic quantum interference and photon-number parity readout.

Theoretical Framework

• Joint Measurement Strategy: Instead of measuring states individually (distributed) or using global measurements on $2N$ copies, the protocol interferes a single copy of state $\hat{\rho}^{(A)}$ and a single copy of $\hat{\rho}^{(B)}$ on a set of balanced beam splitters (BS).
• Parity Readout: The output modes of one register are measured using Photon-Number-Resolving Detectors (PNRDs) to determine the total photon-number parity ($\hat{\Pi} = (-1)^{\sum \hat{n}_i}$).
• Mathematical Derivation: The trace overlap is derived as:
  $$\text{Tr}[\hat{\rho}^{(A)}\hat{\rho}^{(B)}] = \text{Tr}^{(B)} \left[ \text{Tr}^{(A)} \left( \hat{U}_{BS}^\dagger \hat{\rho}^{(A)} \hat{\rho}^{(B)} \hat{U}_{BS} \right) \hat{\Pi} \right]$$
  This relationship holds for arbitrary mixed states, reducing to the squared inner product for pure states.
• Sample Complexity: The authors prove (Theorem 1) that the number of samples $N$ required to estimate the overlap within an additive error $\epsilon$ scales as $O(\epsilon^{-2})$. Crucially, this scaling is independent of the Hilbert space dimension, mode number, or photon number.
• Optimality: Using Helstrom's bound, they prove (Theorem 2) that this $O(\epsilon^{-2})$ scaling is information-theoretically optimal (up to a constant factor), even when compared to complex global measurements. In contrast, distributed strategies for high-dimensional states generally incur exponential costs.

Experimental Implementation

• Platform: The experiment was conducted on Prakash-1, a fully programmable integrated photonic quantum processor based on a silicon nitride (Si$_3$N$_4$) chip.
• State Preparation:
  • Source: Spontaneous Parametric Down-Conversion (SPDC) generates indistinguishable photon pairs at 1550 nm.
  • Encoding: Photons are encoded into qudit states (spatial superpositions over 4 modes) with tunable relative phases.
  • Circuit: A $10 \times 10$ mesh of thermo-optic phase shifters implements universal linear optics. The circuit routes the two qudits into a specific interference pattern (balanced BSs) and directs outputs to 8 Superconducting Nanowire Single-Photon Detectors (SNSPDs).
• Measurement: Since the detectors are not photon-number-resolving, the experiment uses coincidence detection between specific output modes. For single-photon inputs, an odd-parity event in one register corresponds to a coincidence between the two registers, allowing the overlap to be estimated as $1 - 2P_{\text{odd}}$.

3. Key Contributions

1. Sample-Optimal Protocol: Demonstrated a measurement scheme that achieves the theoretical lower bound for overlap estimation ($O(\epsilon^{-2})$) regardless of system dimension.
2. Hardware Efficiency: The protocol requires only linear optics (beam splitters) and parity measurements, avoiding the need for deep unitary circuits or ancilla qubits.
3. Scalability: The sample complexity does not grow with the number of modes or photons, making it suitable for high-dimensional quantum data (both Discrete-Variable and Continuous-Variable states).
4. Experimental Validation: Successfully implemented the scheme on a noisy, intermediate-scale photonic processor (Prakash-1) with thermal crosstalk mitigation.

4. Results

The authors demonstrated the utility of the overlap measurement in two QML tasks:

• Quantum Data Classification (Supervised Learning):

  • Task: Binary classification of quantum datasets using a Support Vector Machine (SVM) with a quantum kernel.
  • Datasets: Three types of 3D parameter-encoded qudit states: linearly separable, concentric spheres, and partially overlapping clusters.
  • Performance: Achieved test accuracies of 100% (separable), 98.47% (spherical), and 91.65% (overlapping). The quantum kernel successfully separated data that was not linearly separable in the original parameter space.

• Online Learning (Unsupervised/Generative):

  • Task: Training a parametrized quantum model to approximate an unknown target state using Simultaneous Perturbation Stochastic Approximation (SPSA).
  • Setup: The target state was a noisy, randomly initialized qudit. The model updated its phases iteratively to minimize infidelity ($1 - |\langle \psi_{target} | \psi_{model} \rangle|^2$).
  • Performance: The algorithm converged reliably within 500 iterations, achieving a median final infidelity of $1.7 \times 10^{-2}$ (98.3% fidelity). The method proved robust against thermal crosstalk noise and shot noise, even with low sample counts ($N=100$) per iteration.

5. Significance and Future Outlook

• Scalable QML Pathway: This work establishes joint overlap measurement as a practical, scalable building block for quantum machine learning, particularly for future quantum networks where data is natively quantum.
• Bypassing Characterization: It proves that relational information (similarity) can be extracted efficiently without the costly step of full state tomography.
• Hardware Native: The approach is native to photonic platforms, which are ideal for quantum networks due to low decoherence and telecom compatibility.
• Future Directions: The authors suggest extending the interference to multi-state interactions to access higher-order nonlinear functionals (analogous to multi-body attention in transformers) and applying the method to non-Gaussian, highly entangled states that are classically intractable.

In summary, the paper provides a rigorous theoretical proof and a robust experimental demonstration that bosonic interference offers an optimal, dimension-independent method for measuring quantum similarity, solving a critical bottleneck for scalable quantum machine learning.