SunBURST

Seeded Universe Navigation — Bayesian Unification via Radial Shooting Techniques

A GPU-accelerated Bayesian evidence calculator achieving machine-precision accuracy through 1024 dimensions with sub-linear scaling.

Features

• Extreme scalability: O(D^0.67) scaling vs O(exp(D)) for traditional methods
• GPU acceleration: 1000× speedup over dynesty/PolyChord at matched dimensions
• High dimensions: Works reliably from 2D to 1024D+
• Pure Python: No compilation required, NumPy/CuPy compatible
• Automatic mode detection: Handles multimodal posteriors automatically

Installation

pip install sunburst-bayes

For GPU acceleration (optional but recommended):

pip install cupy-cuda11x  # For CUDA 11.x
# or
pip install cupy-cuda12x  # For CUDA 12.x

Did it work for you?

If you used SunBURST — even just to try it — I'd genuinely like to know. Drop a note in Discussions or open an issue. Bug reports, use cases, and "it worked on my problem" are all welcome.

Quick Start

from sunburst import compute_evidence, get_array_module

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

# Define parameter bounds
dim = 64
bounds = [(-10, 10)] * dim

# Compute evidence
result = compute_evidence(log_likelihood, bounds)

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

Expected output (RTX 3080, 64D Gaussian):

log Z = -91.8939 ± 0.0001
Peaks found: 1
Time: 0.71s

The true value is log Z = -91.8939 for a 64D unit Gaussian on [-10, 10]^64. SunBURST recovers it to machine precision in under a second.

When Not to Use SunBURST

SunBURST works best on posteriors that are approximately Gaussian near their peaks. It is not the right tool when:

• Heavy-tailed posteriors (e.g. Student-t with low ν). The Laplace approximation underestimates probability mass in the tails. Errors can exceed 100%.
• Highly curved or banana-shaped posteriors where the Hessian at the peak misrepresents the global geometry.
• Ring/shell distributions (e.g. donut posteriors) where probability mass concentrates far from any peak.
• You need posterior samples, not evidence. SunBURST computes log Z; it does not produce MCMC-like samples. Use dynesty or PolyChord if you need the posterior itself.

For these cases, traditional nested samplers remain more reliable. See Table 4 in the paper for detailed failure-mode benchmarks.

Interactive GUI

An interactive Streamlit demo is available:

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

Performance Benchmarks

Tested on RTX 3080 Laptop GPU with n_oscillations=1:

Dimension	SunBURST	dynesty	UltraNest	Speedup
2D	0.39s	0.61s	0.87s	1.6–2.2×
8D	0.42s	37s	54s	88–129×
64D	0.71s	TIMEOUT	TIMEOUT	>1200×
256D	2.72s	—	—	∞
1024D	14.0s	—	—	∞

TIMEOUT = >600s (10 minutes)

SunBURST completes in seconds where traditional methods timeout.

Built-in Test

Verify your installation:

import sunburst
result = sunburst.test(dim=64)  # Runs Gaussian benchmark

Or from command line:

sunburst --test gaussian --dim 64

Configuration Options

result = compute_evidence(
    log_likelihood,
    bounds,
    n_oscillations=1,     # 1=fast, 3=conservative mode detection
    fast=True,            # Fast Hessian estimation
    return_peaks=True,    # Include peak locations in result
    verbose=False,        # Suppress progress output
    seed=42,              # Reproducibility
)

Result Object

result.log_evidence        # float: Estimated log Z
result.log_evidence_std    # float: Uncertainty estimate
result.n_peaks             # int: Number of modes found
result.peaks               # ndarray: (n_peaks, dim) peak locations
result.hessians            # list: Hessian matrices at peaks
result.log_evidence_per_peak  # ndarray: Evidence contribution per peak
result.wall_time           # float: Total computation time
result.module_times        # dict: Per-stage timing breakdown
result.n_likelihood_calls  # int: Total likelihood evaluations
result.config              # dict: Configuration used

GPU Utilities

from sunburst import gpu_available, gpu_info, get_array_module

if gpu_available():
    print(gpu_info())
    xp = get_array_module()  # Returns cupy if available, else numpy

Architecture

SunBURST uses a 4-stage pipeline, named after Guang Ping Yang Style Tai Chi forms:

1. CarryTiger (抱虎歸山): Mode detection via ray casting
2. GreenDragon (青龍出水): Peak refinement with L-BFGS
3. BendTheBow (彎弓射虎): Evidence calculation via Laplace approximation
4. GraspBirdsTail (攬雀尾): Optional dimensional reduction

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}
}

License

MIT License — see LICENSE for details.

Contributing

Contributions welcome! Please see our contributing guidelines.

Acknowledgments

Module names honor Master Donald Rubbo and the Guang Ping Yang Style Tai Chi (廣平楊式太極拳) tradition.