Resource-Efficient Variational Quantum Classifier

1. Problem Statement

Variational Quantum Classifiers (VQCs) are promising hybrid algorithms for Noisy Intermediate-Scale Quantum (NISQ) devices. However, they face a critical bottleneck during the prediction (inference) stage:

• Probabilistic Nature: Quantum measurements are inherently probabilistic. To obtain a reliable classification, standard VQCs require repeated circuit executions (shots), typically $T=1024$ or more, to estimate the expectation value or majority vote.
• Computational Overhead: This repeated execution creates significant computational costs, limiting the practicality of VQCs for real-world applications where many predictions are needed.
• Noise Sensitivity: Existing "single-shot" approaches often rely on strict assumptions (e.g., classes being orthogonal in Hilbert space) that are difficult to meet in practice, leading to poor robustness against noise.

2. Methodology

The authors propose a Unambiguous Quantum Classifier (Model M3) that combines a specific measurement strategy with classical post-processing to reduce resource requirements while maintaining or improving accuracy.

A. Three Comparative Models

The study evaluates three distinct approaches using a binary classification task:

1. Model M1 (Single-Qubit Baseline): Measures only the first qubit. The class is determined by the sign of the eigenvalue ($m_0 \in \{+1, -1\}$) after aggregating $T$ shots.
2. Model M2 (Hamming Distance Baseline): Measures all $k$ qubits. The class is determined by the sign of the sum of all eigenvalues ($\sum m_j$) after aggregating $T$ shots.
3. Model M3 (Unambiguous Classifier - Proposed):
   • Measurement: Measures all $k$ qubits (similar to M2).
   • Filtering (Post-Processing): Introduces an acceptance threshold parameter $t$. A measurement outcome is considered valid only if the absolute sum of eigenvalues satisfies $|\sum m_j| \geq 2t - k$.
   • Logic: Outcomes with small sums (ambiguous cases where the signal is weak) are discarded. Only "decisive" outcomes are used to calculate the final class probability. If no valid shots are obtained for a data point, a random label is assigned.

B. Circuit Architecture

• Data Encoding: Uses a ZZ-entangling feature map ($U_{FM}$) with $l_{FM}$ layers to map classical data into the Hilbert space.
• Ansatz: Employs a RealAmplitudes ansatz ($U_A$) with $l_A$ layers, featuring reverse linear entangling and data re-uploading (injecting input data at multiple layers) to enhance expressivity.
• Optimization: Trained using the SPSA (Simultaneous Perturbation Stochastic Approximation) optimizer, chosen for its robustness in noisy environments and low cost per step (2 evaluations regardless of parameter dimension).

C. Experimental Setup

• Dataset: Breast Cancer classification dataset (binary classification).
• Preprocessing: Standard scaling, PCA for dimensionality reduction, and Min-Max scaling.
• Simulations: Performed on PennyLane using lightning.qubit (noiseless) and default.mixed (noisy).
• Noise Model: Realistic noise channels were applied after every operation: Depolarizing (2%), Amplitude Damping (5%), and Phase Damping (3%).

3. Key Contributions

1. Resource Efficiency: The M3 model drastically reduces the number of circuit executions required per prediction by filtering out ambiguous measurements, rather than averaging over all shots.
2. Enhanced Robustness: The filtering mechanism acts as a natural error mitigation strategy. By discarding low-confidence (noisy) outcomes, the model becomes less susceptible to the stochastic errors inherent in NISQ devices.
3. Theoretical Justification: The authors provide a theoretical analysis using Chernoff bounds and Stirling's approximation, demonstrating that M3 achieves a lower error bound than M1/M2 provided the ansatz can rotate the state such that the probability of decisive correct outcomes ($q_0$) is high and decisive incorrect outcomes ($q_1$) are low.
4. Practical Validation: Demonstrated that the approach works effectively with standard optimizers (SPSA) and data re-uploading techniques, making it viable for near-term hardware.

4. Experimental Results

The models were tested on 3-qubit and 5-qubit configurations under both noiseless and noisy conditions.

Performance Metrics (3-Qubit, Noiseless):

• Accuracy:
  • M3 (Proposed): 90.00%
  • M2 (Baseline): 83.92%
  • M1 (Baseline): 75.15%
  • Improvement: M3 outperformed the best baseline (M2) by 6.9 percentage points.
• Resource Usage:
  • M1/M2 required 1024 shots per prediction.
  • M3 required an average of 128 shots per prediction, but only ~22 were "valid" (decisive) shots used for the final decision.
  • Efficiency: M3 achieved an 8x reduction in total circuit executions per prediction compared to baselines.

Performance Metrics (Noisy Conditions):

• Accuracy:
  • M3: 65.72% (dropped from 90% due to noise, but still the highest).
  • M1/M2: ~62.5%.
  • Improvement: M3 maintained a ~3.1 percentage point advantage over baselines in noisy settings.
• Robustness: The reduction in execution cost (8x) was preserved even under noise.

5-Qubit Results:

• M3 achieved 76.84% (noiseless) and 62.31% (noisy), consistently outperforming M1 and M2, though the gap narrowed slightly compared to the 3-qubit case.

5. Significance and Conclusion

This paper presents a viable pathway for making Variational Quantum Classifiers practical for near-term quantum hardware.

• Trade-off Optimization: It successfully balances the trade-off between classification accuracy and quantum resource consumption (circuit depth/execution count).
• Noise Resilience: The "unambiguous" filtering strategy provides a form of intrinsic error mitigation, making the classifier more robust to the decoherence and gate errors typical of NISQ devices.
• Scalability: The method does not require complex error correction codes or changes to the underlying ansatz, making it easily implementable on current quantum simulators and hardware.

The authors conclude that appropriate post-processing strategies are essential for the future of QML, allowing for high-performance classification with significantly reduced quantum overhead. Future work will focus on adaptive thresholds and larger datasets.