Genetic optimization of ansatz expressibility for enhanced variational quantum algorithm performance

1. Problem Statement

Variational Quantum Algorithms (VQAs), such as the Variational Quantum Eigensolver (VQE), are the leading paradigm for near-term quantum computing on Noisy Intermediate-Scale Quantum (NISQ) devices. However, their performance is critically dependent on the design of the ansatz (the parameterized quantum circuit).

• The Trade-off: There is a fundamental tension between expressibility (the ability of the circuit to span the Hilbert space and represent complex target states) and trainability (the ability to optimize parameters without getting stuck in barren plateaus or local minima).
• Current Limitations:
  • Fixed/Heuristic Designs: Traditional ansatze like UCCSD (Unitary Coupled Cluster with Singles and Doubles) are chemically inspired but scale poorly ($O(n^3)$) and become too deep for NISQ devices.
  • Adaptive Methods: Algorithms like ADAPT-VQE build circuits iteratively but incur high classical computational overhead due to gradient calculations at every step.
  • Scalability: Existing Quantum Architecture Search (QAS) methods often suffer from high classical costs or lack generalizability across different problem Hamiltonians.

The core challenge is to design a problem-agnostic, scalable framework that generates ansatze with high expressibility while maintaining shallow depth and low parameter counts to ensure trainability.

2. Methodology

The authors propose a Genetic Algorithm (GA)-based framework to evolve quantum circuit architectures. Instead of designing a new circuit for every specific problem, they aim to evolve a single, highly expressive circuit structure that can be reused across different Hamiltonians.

A. Genetic Algorithm Framework

The GA treats quantum circuits as "individuals" in a population, evolving them over generations to maximize expressibility.

• Representation: An individual is a sequence of quantum gates arranged in layers. The "genome" consists of the ordered set of gates (e.g., $RX, RY, RZ, H, CNOT$) and their associated parameters.
• Initialization: A population of random circuits is generated up to a specified depth using a predefined gate set.
• Fitness Function (Expressibility Metric):
  • The fitness is defined by the circuit's expressibility.
  • The authors use the Jensen-Shannon Divergence (JSD) between the fidelity distribution of the circuit's generated states and the theoretical Haar-random distribution.
  • $E = D_{JSD}(\hat{P}_{PQC}(F) \parallel P_{Haar}(F))$.
  • A lower JSD value indicates higher expressibility (closer to a uniform distribution over the Hilbert space).
• Evolutionary Operators:
  • Selection: Top-performing circuits (lowest JSD) are selected as parents.
  • Crossover: Parent circuits are split at random layer indices (single-point or N-point crossover) and recombined to create offspring.
  • Mutation: Random gates are replaced with other gates from the set, or the order of gates is shuffled. For CNOT gates, control and target qubits may be swapped.

B. Optimization Strategy

• Depth-Awareness: The framework optimizes for a fixed target depth, ensuring the resulting circuits remain shallow.
• Gate Sets: The study evaluates nine different gate sets (A–I) with varying combinations of single-qubit rotations ($RX, RY, RZ$), Hadamard ($H$), identity ($I$), and CNOT gates (with or without nearest-neighbor constraints).
• Generalizability: Once a circuit architecture is evolved for a specific qubit count and depth, it is treated as a reusable template. Only the classical parameters ($\theta$) are re-optimized for new Hamiltonians, eliminating the need to re-run the GA for every new problem.

3. Key Contributions

1. Problem-Agnostic Ansatz Design: The framework generates a single circuit architecture that performs well across diverse Hamiltonians (molecular and spin systems) without requiring re-optimization of the circuit structure.
2. Balancing Expressibility and Trainability: By explicitly optimizing for expressibility within fixed depth constraints, the method avoids the "barren plateau" issues associated with overly deep circuits while ensuring sufficient state coverage.
3. Resource Efficiency: The evolved ansatze achieve competitive accuracy with significantly fewer layers and gates compared to standard methods like UCCSD and ADAPT-VQE.
4. Scalable Search: The GA approach effectively navigates the non-convex, high-dimensional search space of circuit architectures, outperforming greedy or fixed-design approaches.

4. Results and Benchmarks

The framework was tested on Molecular Hamiltonians ($H_2, LiH, BeH_2, H_2O$) and the Transverse Field Ising Model (TFIM).

• Convergence & Saturation:
  • Expressibility scores saturate around 10 generations, indicating the algorithm converges quickly.
  • Optimal circuit depth scales linearly with the number of qubits ($\approx 2n$). Beyond this depth, expressibility gains diminish.
• Gate Set Performance:
  • Gate Sets A, B, and H (featuring nearest-neighbor CNOT constraints) yielded the highest expressibility due to a higher density of entangling gates.
  • Gate Set E (containing $RY, RZ, CNOT, I$) offered the best trade-off: it provided sufficient expressibility with the fewest gates, leading to superior trainability and lower energy errors in VQE tasks.
• VQE Performance:
  • Accuracy: The GA-derived ansatze achieved ground-state energies comparable to UCCSD and ADAPT-VQE. For example, on $H_2$, the GA achieved chemical accuracy ($10^{-3}$ Hartree) with a depth of 12, whereas UCCSD required depth 20.
  • Depth Reduction: For larger molecules like $LiH$ and $H_2O$, the GA ansatz required depths of 48, compared to depths exceeding 1400 for UCCSD and 200–300 for ADAPT-VQE.
  • Noise Resilience: In noisy simulations (using IBM backend error models), the shallower GA circuits demonstrated better resilience and lower mean energy errors compared to deeper ADAPT circuits.
• Classical Overhead: While the GA search requires classical computation time (scaling with qubits and depth), this is a one-time cost. The resulting circuit is reusable, amortizing the overhead across multiple applications.

5. Significance

This work presents a viable strategy for automated, resource-efficient ansatz design for NISQ devices.

• Practical Utility: It solves the "design bottleneck" in VQAs by providing a method to generate circuits that are both expressive enough to solve complex problems and shallow enough to run on current hardware.
• Reusability: The ability to use a single evolved architecture for different chemical and physical systems simplifies the VQA workflow, making it more accessible for practical quantum chemistry and condensed matter simulations.
• Future Direction: The study validates evolutionary algorithms as a powerful tool for Quantum Architecture Search (QAS), suggesting that problem-agnostic, genetically optimized circuits can serve as a standard baseline for future quantum algorithm development.

In conclusion, the authors demonstrate that genetic optimization can produce highly expressive, low-depth, and reusable ansatze that outperform or match established methods while significantly reducing circuit complexity, thereby enhancing the practical applicability of variational quantum algorithms on near-term hardware.