Exploiting many-body localization for scalable variational quantum simulation

1. Problem Statement

Variational Quantum Algorithms (VQAs), such as the Variational Quantum Eigensolver (VQE), are a leading approach for achieving quantum advantage on near-term hardware. However, their scalability is severely hindered by barren plateaus: regions in the optimization landscape where the gradients of the cost function vanish exponentially with the system size ($n$).

• Root Cause: Barren plateaus typically arise when the parameterized quantum circuit (ansatz) becomes too expressive, effectively forming a unitary 2-design (approximating a Haar-random distribution). In this regime, the entanglement entropy follows a volume law, causing cost gradients to concentrate around zero.
• Current Limitations: Existing mitigation strategies (e.g., layer-wise training, specific ansatz structures, or "close-to-identity" initialization) often struggle to balance expressibility with trainability or fail to scale effectively for generic Hamiltonians.

2. Methodology

The authors propose a novel initialization strategy based on Many-Body Localization (MBL) to prevent the circuit from entering the barren plateau regime during the critical early stages of optimization.

A. Floquet-Initialized Hardware-Efficient Ansatz

• Circuit Architecture: The authors utilize a universal, hardware-efficient layered ansatz consisting of $D$ layers. Each layer contains single-qubit ($R_x, R_z$) and two-qubit ($R_{xx}, R_{yy}, R_{zz}$) Pauli rotations.
• Floquet Initialization: Instead of randomizing parameters independently, the circuit is initialized with layer replication. All layers share identical parameters drawn from a specific distribution:
  • Steady Parameters: Sampled uniformly from $[-\pi, \pi)$ to introduce strong quenched disorder.
  • Kick Parameters: Sampled uniformly from $[-W, W]$, where $W$ is the kick strength.
• Physical Analogy: This setup simulates a periodically driven (Floquet) disordered quantum system.
  • Low $W$ (MBL Phase): The system remains localized, preserving memory of the initial state and exhibiting area-law entanglement.
  • High $W$ (Thermal Phase): The system thermalizes, leading to volume-law entanglement and randomization (unitary 2-design behavior).

B. Optimization Protocol

1. Trial State: A low-complexity classical trial state $|\psi_t\rangle$ (e.g., from a Matrix Product State or Hartree-Fock) is prepared.
2. Encoding: A shallow channel $E_t$ maps $|\psi_t\rangle$ to a computational basis state $|\phi_c\rangle$.
3. Variational Evolution: The Floquet-initialized circuit $\hat{U}(\vartheta)$ acts on $|\phi_c\rangle$.
4. Decoding: The inverse channel $E_t^{-1}$ maps the output back to the logical frame.
5. Training: Parameters are updated via gradient descent. Crucially, while the circuit is initialized in the MBL phase, parameters are allowed to evolve freely during optimization, but the trajectory is hypothesized to remain within a "trainable" region far from the unitary 2-design limit.

3. Key Contributions & Theoretical Framework

A. Characterization of the MBL–Thermalization Transition

The authors define a distinct phase transition in the variational circuit controlled by the kick strength $W$. They introduce three diagnostics to identify the transition point $W^*$:

1. Inverse Participation Ratio (IPR): Proves that in the MBL phase, the output state remains localized to the trial state (high IPR), whereas in the thermal phase, it delocalizes (low IPR). They establish a link between IPR and unitary $t$-designs (Theorem 1).
2. Entanglement Entropy: Demonstrates a transition from area-law scaling (MBL) to volume-law scaling (thermal).
3. Low-Weight Stabilizer Rényi Entropy ($M_{t,k}$): A novel metric introduced to quantify the "magic" (non-stabilizerness) of the state.
   • Theorem 2: Shows that $M_{t,k}$ serves as a witness against higher-order unitary designs ($2t$-designs).
   • In the MBL phase, $M_{t,k}$ remains $O(1)$, indicating the state is far from Haar-random. In the thermal phase, it approaches the Haar benchmark.

B. Theoretical Guarantees on Trainability

• Observation 1: Initializing within the MBL phase ($W < W^*$) ensures that the von Neumann entropy of local regions remains bounded below the Page value, mathematically precluding the occurrence of barren plateaus (which require high entropy).
• Theorem 3: Provides a lower bound on the gradient norm along a linear path between MBL and thermal parameters. It proves that if the initial state is in the MBL phase, the gradient magnitude remains inverse-polynomial in system size, ensuring trainability.

4. Results

A. Numerical Simulations

• Phase Transition: Simulations on 1D ring and circulant graph topologies confirmed a sharp transition at a critical kick strength $W^*$ (e.g., $W^* \approx 0.9$ for 1D rings). Below $W^*$, gradients remained stable as $n$ increased; above $W^*$, gradients decayed exponentially.
• Optimization Performance: The authors tested the Aubry-André model across four phases (Anderson localized, extended, MBL, and ergodic).
  • MBL Initialization: Consistently achieved approximation ratios close to 1.0 (exact ground state) for system sizes up to $n=18$.
  • Random/Thermal Initialization: Failed rapidly as $n$ increased, often getting stuck in local minima or failing to converge due to barren plateaus.
  • Trajectory Analysis: The optimization trajectory starting from MBL initialization stayed far from the Haar benchmark (low $M_{2,2}$) throughout the process, avoiding the barren plateau region.

B. Experimental Verification

• Hardware: Experiments were conducted on the IBM Brisbane processor (127 qubits).
• Setup: A kicked Heisenberg chain (ferromagnetic) was simulated on up to 31 qubits.
• Findings:
  • The experimental data showed a clear phase transition in gradient magnitude as $W$ increased.
  • A critical point $W^* \approx 0.5$ was observed (shifted from the theoretical $0.8$ due to hardware noise).
  • Below this threshold, gradients were preserved, and optimization successfully converged to lower energy states compared to thermal or random initializations.

5. Significance and Impact

1. Scalability Solution: This work provides a concrete, experimentally validated strategy to overcome the barren plateau problem, a major bottleneck for scaling VQAs to useful problem sizes.
2. Physics-Informed Algorithm Design: It bridges condensed matter physics (MBL) and quantum computing, demonstrating that physical phases of matter (localization vs. thermalization) can be directly mapped to algorithmic trainability.
3. Hardware Efficiency: The method works on current noisy intermediate-scale quantum (NISQ) devices without requiring error correction, leveraging the natural resilience of MBL systems to decoherence.
4. Generalizability: While demonstrated on specific models, the framework suggests that "warm-starting" variational circuits within a localized phase is a viable general strategy for preparing ground states of complex Hamiltonians, potentially enabling the solution of problems previously considered classically intractable due to optimization failures.

In summary, the paper demonstrates that by initializing variational quantum circuits within the Many-Body Localized phase, one can maintain robust gradients and avoid barren plateaus, enabling scalable and efficient quantum simulation on current hardware.