quantum-boltzmann-machine-bilevel

name	quantum-boltzmann-machine-bilevel
description	Quantum Boltzmann Machine via Bilevel Optimization methodology. Extends QAOA circuit to bilevel optimization for fully connected QBMs, overcoming the fixed target Hamiltonian barrier. Use when building quantum generative models, training quantum Boltzmann machines, or extending QAOA for ML applications. arXiv:2605.07473

Quantum Boltzmann Machine via Bilevel Optimization

Description

Breaks QAOA's fixed target Hamiltonian barrier by enabling flexible Hamiltonian construction through bilevel optimization. Creates fully connected Quantum Boltzmann Machines that capture complex energy landscapes for optimization and machine learning.

Activation Keywords

• quantum boltzmann machine
• QBM bilevel optimization
• QAOA extension
• quantum generative model
• quantum energy-based model
• bilevel quantum optimization
• fully connected QBM
• quantum approximate optimization extension

Core Architecture

Bilevel Optimization Framework

Upper level:  Optimize QBM parameters (weights, biases)
  → Minimize KL divergence between model and data distributions
Lower level:  Optimize QAOA circuit parameters
  → Prepare thermal state of current Hamiltonian

Algorithm

def train_qbm_bilevel(data_samples, p_layers=3, n_qubits=None, max_iter=100):
    """Train fully connected QBM via bilevel optimization.
    
    Args:
        data_samples: Binary data vectors
        p_layers: Number of QAOA layers
        n_qubits: Number of qubits (defaults to data dimension)
    """
    n = n_qubits or len(data_samples[0])
    
    # Initialize Hamiltonian parameters
    # Fully connected: h_i (biases) + J_ij (couplings)
    biases = np.random.randn(n) * 0.1
    couplings = np.random.randn(n, n) * 0.1
    couplings = (couplings + couplings.T) / 2  # Symmetric
    
    for iteration in range(max_iter):
        # LOWER LEVEL: Optimize QAOA circuit for current H
        gamma, beta = optimize_qaoa_circuit(biases, couplings, p_layers)
        
        # UPPER LEVEL: Update Hamiltonian to match data
        model_expectation = compute_expectations(gamma, beta, biases, couplings)
        data_expectation = compute_data_expectations(data_samples)
        
        # Gradient update (contrastive divergence style)
        grad_biases = data_expectation['local_z'] - model_expectation['local_z']
        grad_couplings = data_expectation['ZZ'] - model_expectation['ZZ']
        
        biases += learning_rate * grad_biases
        couplings += learning_rate * grad_couplings
    
    return biases, couplings

Key Innovations

1. Flexible Hamiltonian: Unlike standard QAOA with fixed mixing/problem Hamiltonians, this approach learns the Hamiltonian itself
2. Full Connectivity: All-to-all couplings (J_ij for all i,j) vs QAOA's limited connectivity
3. Bilevel Structure: Inner loop prepares states, outer loop learns model
4. Generative Capability: Can sample from learned distribution

Comparison

Method	Connectivity	Hamiltonian	Training
Classical BM	Full	Fixed	CD/k-CD
Standard QBM	Limited	Fixed	Gradient-based
QAOA	Problem-specific	Fixed	Variational
QBM-Bilevel	Full	Learned	Bilevel

Use Cases

• Generative Modeling: Learn and sample from complex distributions
• Optimization: Find low-energy states of learned Hamiltonians
• Feature Learning: Extract hidden representations from data
• Quantum-Classical Hybrid: Bridge classical ML with quantum circuits

Pitfalls

• Bilevel optimization is computationally expensive
• Inner loop convergence affects outer loop quality
• Requires careful balancing of learning rates at both levels
• Near-term hardware limitations on qubit count and connectivity

Related Skills

• spectral-born-machines (2607.06675) — Quantum generative models using group Fourier analysis with QFT-based spectral inductive bias. Classically trainable via MMD with graph spectral kernels. Scalable to 190 qubits, 1M+ parameters. Over-parameterized models immune to overfitting in data-scarce regimes. Complementary to QBMs: SBMs exploit spectral structure for discrete data, QBMs use energy-based learning for general distributions.

References

• arXiv:2605.07473 - "Breaking QAOA's Fixed Target Hamiltonian Barrier"
• arXiv:2607.06675 - "Spectral Born machines: classically trainable quantum generative models for discrete data"