Gate Freezing Method for Gradient-Free Variational Quantum Algorithms in Circuit Optimization

1. Problem Statement

Variational Quantum Algorithms (VQAs) rely on Parameterized Quantum Circuits (PQCs) to solve complex problems in quantum chemistry, optimization, and machine learning. However, optimizing PQCs on Noisy Intermediate-Scale Quantum (NISQ) devices faces significant challenges:

• Resource Constraints: Optimization requires a large number of circuit evaluations (shots), which is costly on real hardware.
• Convergence Issues: As circuits deepen, optimization often stalls due to barren plateaus (vanishing gradients) or simply because the optimizer spends excessive resources tuning parameters that have already converged.
• Inefficiency: Standard gradient-free optimizers (like Rotosolve, Fraxis, and FQS) treat all gates equally in every iteration, even when many parameters have stabilized and require minimal adjustment.

The authors observe that after a few iterations, a large subset of PQC parameters changes very little between steps. The paper aims to leverage this observation to improve resource allocation without attempting to solve the barren plateau problem itself.

2. Methodology

The authors propose Gate Freezing Methods, inspired by layer freezing and fine-tuning techniques in deep learning. The core idea is to identify gates whose parameters have stabilized and "freeze" them (skip optimization) for a set number of iterations, thereby redirecting computational resources to gates that are still changing significantly.

A. Distance Metrics for Freezing

To determine if a gate should be frozen, the authors define a distance metric $D$ between the parameter values of the current iteration and the previous iteration. Two approaches are proposed:

1. Parameter-Based Distance:
   • Rotosolve: Uses the cyclic distance for rotation angles $\theta \in [-\pi, \pi)$.
   • Fraxis & FQS: Uses the geodesic distance on the unit sphere (for rotation axes) or the 4D unit sphere (for quaternions), accounting for the fact that $q$ and $-q$ represent the same physical rotation.
2. Matrix Norm Distance (Generalized):
   • Calculates the Frobenius norm distance between the unitary matrices of the gate before and after optimization: $D(U, V) = \sqrt{4 - 2|\text{Tr}(U^\dagger V)|}$ (normalized to $[0, 1]$).
   • This approach is hardware-agnostic and applicable to arbitrary $n$-qubit gates, not just single-qubit parameters.

B. Freezing Algorithms

Two specific algorithms are implemented:

1. Fixed Iteration Freezing: If the distance $D < T$ (a predefined threshold), the gate is frozen for a fixed number of iterations $\kappa$.
2. Incremental Gate Freezing: If a gate is frozen, its freezing duration $\kappa_d$ is incremented by 1 for the next time it is frozen. This heuristic increasingly penalizes gates that consistently show negligible change, dynamically allocating more optimization steps to "poorly optimized" gates.

3. Key Contributions

• Novel Optimization Strategy: Introduction of gate freezing specifically for sequential single-qubit gradient-free optimizers (Rotosolve, Fraxis, FQS).
• Dual Distance Metrics: Development of both parameter-based and generalized matrix-norm-based metrics to detect convergence, allowing flexibility for different ansatz structures.
• Dynamic Resource Allocation: The "Incremental Freezing" algorithm, which adaptively increases the freeze duration for stable gates, effectively shifting the optimization burden to active parameters.
• Comprehensive Benchmarking: Extensive testing across different Hamiltonians (Heisenberg, Fermi-Hubbard), circuit depths (shallow vs. deep), and noise conditions (statevector vs. shot noise).

4. Results

The methods were tested on the 1D Heisenberg model and the Fermi-Hubbard model using 5 to 9 qubits.

• Convergence Speed:
  • Rotosolve and Fraxis: Showed the most significant improvements. Gate freezing methods converged faster to the ground state energy and achieved lower relative errors compared to the base optimizers.
  • FQS: Showed smaller, problem-dependent improvements. In some cases (e.g., deep random circuits), the base FQS performed competitively, but freezing still offered marginal gains in specific thresholds.
• Threshold Sensitivity:
  • Optimal freezing thresholds ($T$) varied by optimizer (e.g., $T \approx 0.001$ for Rotosolve, $T \approx 0.025$ for Fraxis).
  • Incremental freezing generally outperformed fixed freezing, particularly for Rotosolve and Fraxis.
• Noise Resilience:
  • In simulations with shot noise (1024 and 4096 shots), the incremental freezing method provided clear benefits for Fraxis and Rotosolve in shallow circuits.
  • For deep circuits with shot noise, improvements were limited, consistent with the difficulty of optimizing deep PQCs under statistical noise (related to barren plateaus).
• Scalability:
  • As system size increased (6–9 qubits), the incremental freezing method consistently outperformed the base Rotosolve, maintaining a lower relative error to the ground state.
  • Analysis of gate freezing patterns revealed that gates in the last layers and middle qubits of the circuit tend to stabilize first and are frozen most frequently.

5. Significance and Conclusion

• Resource Efficiency: The primary contribution is the efficient reallocation of computational resources. By freezing stable gates, the method reduces the number of unnecessary circuit evaluations, which is critical for NISQ hardware where shot counts are expensive.
• Complementary to Barren Plateau Mitigation: The authors explicitly state that this method does not solve barren plateaus. Instead, it optimizes the process of existing gradient-free optimizers, making them more efficient within the regimes where they are already applicable.
• Practical Applicability: The method is compatible with current VQA frameworks (tested using PennyLane) and can be easily integrated into existing optimization loops.
• Future Directions: The paper suggests extending these methods to other VQAs (e.g., Variational Quantum Singular Value Decomposition) and exploring how gate freezing patterns could inform better ansatz designs or optimization sequences.

In summary, the paper demonstrates that Gate Freezing is a highly effective, low-overhead strategy to accelerate the convergence of gradient-free variational quantum algorithms, particularly for Rotosolve and Fraxis, by intelligently skipping the optimization of parameters that have already converged.