Log-SNRAS

A Computationally Efficient Variance-Stabilized Metric for Vetting Heteroscedastic Light Curves

Overview

Log-SNRAS (Logarithmic Signal-to-Noise Ratio with Adjusted Statistics) is a computationally efficient, non-parametric post-detection vetting filter designed to automate the classification of exoplanetary transit candidates from space-based missions (Kepler, K2, TESS).

Standard SNR metrics and search algorithms (e.g., BLS, TPS) evaluate signal significance assuming stationary, homoscedastic noise. In real time-series photometry, instrumental anomalies, stellar variability, and non-stationary artifacts frequently induce false-positive detections. Log-SNRAS penalizes heteroscedastic candidates by evaluating the relative dispersion contrast ($\psi$) between in-transit and out-of-transit flux, scaling detection significance through a physically-motivated logarithmic penalty:

$$\text{Log-SNRAS} = \frac{\delta}{\hat{\sigma}_{\text{out}}} \times \frac{\sqrt{N_{\text{in}}}}{1 + \ln(1 + \psi_{\text{corr}})}$$

---

Key Breakthrough: Shape-Corrected Dispersion ($\sigma_{\text{in, res}}$)

In earlier single-host formulations, deep planetary transits were susceptible to false variance penalties because deterministic geometric transit curvature (ingress, flat bottom, egress) inflated the raw in-transit standard deviation $\sigma_{\text{in}}$.

Log-SNRAS resolves this by decoupling geometric transit curvature from stochastic photometric noise: $$\sigma_{\text{in, res}} = \sqrt{\frac{1}{N_{\text{in}} - 1} \sum_{i \in \text{in}} (f_i - \hat{f}_i)^2}$$ where $\hat{f}$ is a non-parametric transit profile fit in orbital phase space.

This correction completely eliminates false vetoes on deep planetary transits:

• TOI-201 ($R_p = 1.01 R_J$, depth $\sim 5{,}200$ ppm): Penalty drops from $\mathcal{P}{\text{raw}} = 1.640$ (was Tier 3 Veto) to $\mathcal{P}{\text{corr}} = 0.498$ (Restored to Clean Planetary Transit).
• Kepler-8 b (depth $\sim 10{,}000$ ppm): Penalty drops from $0.951$ to $0.004$ (Tier 1 Clean).
• Kepler-11 d (depth $\sim 10{,}000$ ppm): Penalty drops from $1.552$ to $0.031$ (Tier 1 Clean).
• Deep Transits from TESS & Kepler (tested on $\delta$ up to $56%$): Reliably classified as Tier 1 Clean.

---

Curated Multi-Host Benchmark ($N=17$) and LOHO Validation

The method is validated across a balanced, literature-verified multi-host benchmark:

• 10 Confirmed Exoplanets across 10 Distinct Hosts: Pi Mensae c, TOI-201 b, Kepler-10 b, Kepler-448 b, TrES-2 b, WASP-126 b, TOI-540 b, L 98-59 b, Kepler-8 b, and Kepler-11 d.
• 7 Literature-Proven Non-Planetary Artifacts: Boyajian's Star (Quarters 8 & 16 anomalous dips), KOI-1611 eccentric eclipsing binary, Kepler eccentric EB, contact binaries (W UMa KIC 11253226), and ellipsoidal variables.

Empirical Classification Performance:

Method	AUC	95% Bootstrap CI	Mean LOHO AUC
Shape-Corrected Penalty ($-\mathcal{P}_{\text{corr}}$)	0.757	[0.486, 0.971]	0.757 $\pm$ 0.036
Log-SNRAS Composite	0.700	[0.414, 0.943]	0.700 $\pm$ 0.035
Traditional SNR (T-SNR)	0.629	[0.321, 0.914]	0.629 $\pm$ 0.032
Robust SNR (MAD)	0.614	[0.314, 0.886]	0.614 $\pm$ 0.033
BLS SNR Proxy	0.614	[0.314, 0.886]	0.614 $\pm$ 0.033
Pont SNR (2006)	0.600	[0.271, 0.886]	0.600 $\pm$ 0.034
Inverse Depth ($1/\delta$)	0.471	[0.143, 0.857]	0.471 $\pm$ 0.045

Leave-One-Host-Out (LOHO) Cross-Validation:

• Excluding Pi Mensae yields AUC = 0.762.
• Excluding TOI-201 yields AUC = 0.778.
• Mean LOHO AUC = $0.757 \pm 0.036$, confirming that performance is host-invariant.

---

Repository Structure

Log-SNRAS/
├── src/
│   ├── log_snras/                 # Modular Python package
│   │   ├── __init__.py
│   │   ├── core.py               # Log-SNRAS math, shape correction, tiers
│   │   ├── masking.py            # Uniform ephemeris-based window definition
│   │   └── io.py                 # FITS named-column reader (PDCSAP_FLUX)
│   └── calculate_log_snras.m     # Standalone vectorized MATLAB function
├── data/
│   ├── curated_benchmark_catalog.csv    # 17 multi-host benchmark targets
│   ├── multi_host_evaluation_results.csv # Empirical metric evaluation
│   ├── multi_host_auc_performance.csv   # Statistical AUC comparison
│   └── evaluation_dataset_v2.csv        # Archival dataset (superseded)
├── scripts/
│   ├── evaluate_multi_host_benchmark.py # Automated end-to-end evaluation
│   ├── generate_publication_figures.py  # 100% real archival FITS figures
│   ├── test_nrebig.py                   # Empirical validation on TESS targets
│   └── test_nrebig_kepler.py            # Empirical validation on Kepler targets
└── tests/
    └── test_core.py                     # Unit tests for shape correction and tiers

---

Quickstart (Python)

# 1. Install dependencies
pip install numpy scipy astropy pandas matplotlib scikit-learn

# 2. Run automated multi-host evaluation
python scripts/evaluate_multi_host_benchmark.py

Python API Example:

from log_snras.core import compute_log_snras, compute_shape_corrected_dispersion

# time, flux, and in_mask (True = In-Transit)
# Compute shape-corrected in-transit dispersion
sig_in_res, fitted_profile = compute_shape_corrected_dispersion(time, flux, in_mask)

# Compute Log-SNRAS and vetting tier
result = compute_log_snras(flux, in_mask, out_mask, snr_trad=45.0, sig_in_res=sig_in_res)
print(f"Log-SNRAS Score: {result.log_snras:.2f}")
print(f"Vetting Tier: {result.tier} (Penalty = {result.penalty:.4f})")

---

Quickstart (MATLAB)

% Add src to path
addpath('src');

% Calculate Log-SNRAS
[score, penalty, psi, tier] = calculate_log_snras(flux, in_mask, out_mask, snr_trad);
fprintf('Tier: %s | Log-SNRAS: %.2f\n', tier, score);

---

Reproducibility Statement

Every table and figure in the manuscript is generated from 100% real NASA archival Kepler and TESS light curves using explicit named columns (PDCSAP_FLUX), uniform orbital ephemeris masking, and deterministic seeds. Run:

python scripts/generate_publication_figures.py

to regenerate all publication-grade figures (600 DPI) into figures/.

---

Citation

@article{jabbar2026logsnras,
  author  = {Jabbar, Ahmed Sattar},
  title   = {Log-SNRAS: A Computationally Efficient Variance-Stabilized Metric for Vetting Heteroscedastic Light Curves},
  journal = {Astronomy and Computing},
  year    = {2026}
}

License

MIT License. See LICENSE for details.