SunBURST

Seeded Universe Navigation — Bayesian Unification via Radial Shooting Techniques

A deterministic, GPU-accelerated algorithm for Bayesian model evidence calculation, achieving machine-precision accuracy through 1024 dimensions.

GitHub Repository arXiv Paper

Overview

SunBURST is a novel approach to computing Bayesian model evidence (marginal likelihood) that softens the "curse of dimensionality" plaguing traditional methods like nested sampling.

Key achievements:

• Machine-precision accuracy on Gaussian posteriors through 1024 dimensions
• Sub-linear scaling O(d^0.67) vs. traditional O(d²) methods
• Deterministic results with no MCMC convergence concerns
• Automatic multimodal detection and integration
• Full posterior characterization via Laplace approximation

The Pipeline

SunBURST operates in three stages, named after movements in Guang Ping Yang Style Tai Chi (廣平楊式太極拳) in honor of Master Donald Rubbo:

🐯 CarryTiger (抱虎歸山) — Mode detection via GPU ray casting from prior boundary

🐉 GreenDragon (青龍出水) — Peak refinement using GPU-accelerated L-BFGS optimization

🏹 BendTheBow (彎弓射虎) — Evidence calculation via Laplace approximation with radial verification

Quick Start

pip install sunburst-bayes

from sunburst import compute_evidence, get_array_module

# Define your log-likelihood function (GPU-native)
def log_likelihood(x):
    xp = get_array_module(x)  # CuPy if GPU, NumPy if CPU
    return -0.5 * xp.sum(x**2, axis=1)

# Set parameter bounds
bounds = [(-10, 10)] * 100  # 100-dimensional problem

# Compute evidence
result = compute_evidence(log_likelihood, bounds)

print(f"log Z = {result.log_evidence:.4f}")
print(f"Modes found: {result.n_peaks}")
print(f"Time: {result.wall_time:.2f}s")

Interactive GUI

An interactive Streamlit demo is available in the repository:

git clone https://github.com/beastraban/sunburst.git
cd sunburst/sunburst_super_gui
pip install streamlit
streamlit run app.py

Performance

Benchmarks on NVIDIA RTX 3090 demonstrate sub-linear scaling with dimension, achieving 1024D in 6.8 seconds while competitors timeout:

Applications

SunBURST is designed for high-dimensional Bayesian inference problems including:

• Cosmological parameter estimation — CMB analysis, dark energy constraints
• Gravitational wave astronomy — Source characterization, model selection
• Machine learning — Bayesian neural network evidence, model comparison
• Drug discovery — High-dimensional molecular parameter spaces
• Financial modeling — Multi-factor model selection

Citation

If you use SunBURST in your research, please cite:

@article{wolfson2026sunburst,
    title={SunBURST: Deterministic GPU-Accelerated Bayesian Evidence 
           via Mode-Centric Laplace Integration},
    author={Wolfson, Ira},
    journal={arXiv preprint arXiv:2601.19957},
    year={2026}
}

Contact

For questions, collaborations, or bug reports, please open an issue on GitHub or get in touch.