SpinGQE: A Generative Quantum Eigensolver for Spin Hamiltonians

1. Problem Statement

The search for the ground state of quantum systems is a fundamental challenge in quantum computing, with applications in quantum chemistry, condensed matter physics, and combinatorial optimization. The current leading approach, the Variational Quantum Eigensolver (VQE), faces significant limitations:

• Barren Plateaus: Optimization landscapes often become flat, making gradient-based training ineffective.
• Ansatz Limitations: Fixed circuit structures often lack the expressivity to capture complex ground states.
• Domain Dependency: VQE relies heavily on physical intuition and domain-specific symmetries (e.g., electronic structure), which do not generalize well to arbitrary Hamiltonians.
• Noise Sensitivity: Deep circuits required for complex systems are highly susceptible to noise on NISQ (Noisy Intermediate-Scale Quantum) devices.

The authors propose SpinGQE, an extension of the Generative Quantum Eigensolver (GQE) framework, specifically adapted for spin Hamiltonians. The goal is to reframe circuit design as a generative modeling task to bypass the limitations of traditional variational methods.

2. Methodology

SpinGQE treats the design of a quantum circuit as a sequence generation problem, utilizing a Transformer-based decoder to learn the distribution of quantum circuits that yield low-energy states.

A. Generative Framework

• Architecture: A Transformer model samples sequences of operators (gates) from a fixed Operator Pool.
• Training Loop (Online):
  1. The model generates a batch of operator sequences (circuits).
  2. Each sequence is executed on a quantum simulator or device to evaluate the expectation value of the target Hamiltonian $H$.
  3. Prefix Evaluation: Unlike standard methods that only evaluate the final circuit, SpinGQE evaluates the energy of every prefix subsequence of gates ($U_{1:t}$). This creates a trajectory of energies.
  4. Logit Matching: The model produces cumulative logits for each timestep. The training objective is to align these logits with the measured energies of the corresponding prefixes.

B. Loss Function

The authors employ a Weighted Mean-Squared Error (MSE) loss to prioritize low-energy regions:
$$\text{MSE}_{\text{weighted}} = \frac{1}{N} \sum_{i=1}^{N} w(E_N^{(i)}) \sum_{t=1}^{T^{(i)}} \left( l_t(x^{(i)}) - E_t^{(i)} \right)^2$$

• Weighting Function: $w(E) = \frac{1}{1 + e^{\beta E}}$.
• Mechanism: This function assigns higher weights to circuits with lower energies ($E$), controlled by the parameter $\beta$. This guides the model to focus on the "valleys" of the energy landscape, discouraging premature convergence to local minima.

C. Operator Pool Design

For spin systems (specifically the Heisenberg model), the operator pool is constructed to reflect the native Pauli structure:

• Gates: Pauli rotations $U(P, \theta) = e^{-i \theta P / 2}$.
• Operators: $P \in \{Z, XX, YY, ZZ\}$.
• Angles: Discrete set $\{\pm \pi/2^k \mid k = 1, \dots, 5\}$.
• Topology: Single-qubit rotations on all qubits; two-qubit rotations restricted to nearest-neighbor pairs (respecting the 1D chain structure).

D. Post-Processing Optimization

To overcome the constraints of the discrete operator pool and fixed angles, a two-stage hybrid optimization follows the generative training:

1. Angle Refinement: Continuous optimization (using L-BFGS-B or COBYLA) adjusts the rotation angles of the generated circuit.
2. Qubit Reassignment (Wire Swap): A greedy search iterates through gate positions, testing different qubit pairings (including non-nearest neighbors) to introduce long-range entanglement, re-optimizing angles after each swap.

3. Key Contributions

1. Domain Adaptation: Successfully extends the GQE framework from fermionic systems to spin Hamiltonians, demonstrating versatility across quantum domains.
2. Generative Circuit Design: Shifts the optimization burden from continuous parameter tuning (VQE) to discrete circuit structure learning via a Transformer, effectively navigating complex energy landscapes without relying on problem-specific symmetries.
3. Prefix-Based Supervision: Introduces a training strategy that evaluates energy at every gate subsequence, allowing the model to learn progressive energy descent patterns and assign credit to effective partial circuits.
4. Hybrid Refinement: Demonstrates that a lightweight generative model combined with classical post-processing can achieve near-exact ground states, outperforming training on larger, more complex operator pools.

4. Results

The method was validated on a 4-qubit one-dimensional Heisenberg model with an external magnetic field ($H = J \sum (X_n X_{n+1} + Y_n Y_{n+1} + Z_n Z_{n+1}) + h \sum Z_n$).

• Antiferromagnetic Regime ($h \le J$, specifically $h=J=10$):

  • Challenge: Highly frustrated, entangled ground state with a rugged energy landscape.
  • Performance: The base model (37M parameters) converged to $E \approx -60.78J$ (Exact: $-64.641J$).
  • Post-Processing: Angle refinement and wire swaps reduced the energy to $-64.64J$, effectively reaching the exact ground state.
  • Insight: The model learned specific gate-angle preferences (e.g., $YY$ gates on specific qubit pairs) and the post-processing successfully introduced necessary long-range entanglement (e.g., between qubits 0 and 3) that the nearest-neighbor pool could not generate.

• Field-Dominated Regime ($h > J$, specifically $h=10, J=1$):

  • Challenge: Simpler landscape approaching a product state.
  • Performance: The model converged directly to $-37.0J$ (Exact: $-37.0J$) without needing extensive post-processing.

• Hyperparameter Optimization:

  • Model Size: Smaller models (37M parameters) outperformed larger ones (113M), which failed to converge due to data scarcity (barren plateaus in the training data space).
  • Sequence Length: 12 gates was optimal.
  • Weighting ($\beta$) and Batch Size ($M$): Moderate $\beta$ (0.3) and small batch sizes ($M=10$) yielded the best balance between exploration and exploitation.

• Symmetry Protection: The algorithm successfully captured the physical phenomenon where the ground state energy remains constant for $h \le J$ due to the conservation of total magnetization ($S_z^{total}$), demonstrating physical consistency.

5. Significance and Future Outlook

• Scalability: SpinGQE offers a scalable alternative to VQE by shifting the computational load to classical generative modeling (Transformers), reducing the reliance on deep quantum circuits during the search phase.
• Generalizability: The approach does not require domain-specific ansatz construction, making it applicable to arbitrary Hamiltonians where physical intuition is insufficient.
• Efficiency: The method achieves high precision with relatively few quantum evaluations per epoch (sample efficiency), crucial for NISQ devices.
• Future Directions: The authors suggest that transfer learning could allow models trained on small spin chains to generalize to larger, more complex systems, mitigating the need for massive training datasets as system size grows.

In conclusion, SpinGQE represents a paradigm shift in ground state search, leveraging the power of generative AI to navigate quantum energy landscapes more effectively than traditional variational methods, particularly in complex, frustrated spin systems.