spectral-born-machines

name	spectral-born-machines
description	Quantum generative models using group Fourier analysis for discrete data with classical trainability at scale (arXiv: 2607.06675)

Spectral Born Machines

Overview

Spectral Born machines are a class of quantum generative models that generalize IQP Born machines through group Fourier analysis. They exploit quantum Fourier transform to create inductive bias for integer-structured data while remaining classically hard to sample.

Key Innovation: View Born machines through group Fourier analysis lens, enabling spectral bias for structured discrete data.

Core Methodology

1. Theoretical Foundation

• Group Fourier Analysis: Generalize IQP Born machines using group representation theory
• Quantum Fourier Transform: Creates natural inductive bias for integer-structured data
• Classical Hardness: Maintains quantum advantage in sampling complexity

2. Training Approach

• Maximum Mean Discrepancy (MMD): Loss function based on graph spectral analysis
• Classical Trainability: Efficient training on classical hardware at scale
• Software Implementation: Available in PennyLane's new tcdq module

3. Key Results

• Parameter Efficiency: Spectral bias leads to significantly reduced parameter counts vs unstructured approaches
• Scalability: Successfully trained 190-qubit model with 1M+ parameters
• Overfitting Resistance: Highly over-parameterized models may be immune to overfitting in data-scarce regimes
• Biological Application: Learned distribution of 93-nucleotide ribosomal RNA

Technical Details

Architecture Components

1. IQP Circuit Structure: Instantaneous Quantum Polynomial-time circuits
2. Spectral Bias: Fourier-based inductive bias for discrete structures
3. Graph Spectral Analysis: MMD computation using spectral properties

Training Pipeline

1. Define target discrete distribution
2. Construct spectral Born machine ansatz
3. Compute MMD loss via graph spectral analysis
4. Optimize parameters using classical optimizer
5. Sample from trained quantum model

Use Cases

• Discrete Data Generation: Integer-structured data (sequences, graphs, combinatorial objects)
• Biological Sequences: RNA/DNA sequence modeling
• Combinatorial Optimization: Sampling from complex discrete distributions
• Quantum Advantage Demonstration: Classical hardness with quantum sampling

Implementation Notes

• Framework: PennyLane with tcdq module
• Scalability: Tested up to 190 qubits, 1M+ parameters
• Hardware: Classical simulation for training, quantum hardware for sampling
• Data Requirements: Works well in strongly data-scarce regimes

Activation Keywords

spectral Born machine, quantum generative model, group Fourier analysis, IQP Born machine, quantum Fourier transform, discrete data generation, MMD loss, graph spectral analysis, PennyLane tcdq, quantum sampling advantage

References

• arXiv: 2607.06675 (2026)
• Authors: Austin Huang, William Maxwell, Vasilis Belis, Evan Peters, Jason Pye, Soran Jahangiri, Joseph Bowles
• Software: PennyLane tcdq module