Why does non-Gaussian transient noise fool algorithms trained only on Gaussian white noise injections?

# Why Non-Gaussian Transient Noise Fools Gaussian-Trained Algorithms

The core problem is distribution mismatch: algorithms trained exclusively on Gaussian white noise learn to expect data where randomness is symmetric, unbounded, and lacks hidden structure. Non-Gaussian transients (called "glitches" in gravitational-wave detection) violate all three assumptions, and algorithms have no learned machinery to handle them.

How Gaussian training creates blind spots

When you train a machine-learning model on Gaussian noise alone, it learns statistical patterns specific to that distribution. The model's decision boundaries, thresholds, and feature representations are all calibrated for data where extreme outliers are rare and where noise is "white" (no preferred frequencies) [AG-2025.08-535].

Non-Gaussian transients—which can be brief, loud spikes or have internal structure like phase-coupled frequency components—look qualitatively different from anything in the training set. The model has no learned concept of what they are, so it either:

1. Misclassifies them as signal (false alarm): a glitch's envelope might resemble a gravitational-wave burst, fooling a detection algorithm.

2. Fails to remove them: denoising models trained on Gaussian noise can preserve glitch features because they were never taught to suppress them [AG-2024.03-180].

3. Biases parameter estimates: when a glitch overlaps a real signal in time and frequency, Bayesian inference methods trained on Gaussian likelihoods weight the glitch as part of the signal, distorting the inferred source properties [AG-2025.06-690].

The circular paradox and one solution

Here's the catch: you can't train on "real non-Gaussian data" without already knowing which parts are glitches—but you can't identify glitches without a trained detector. One elegant escape is to train on *statistical features* of Gaussian noise rather than the raw time series. The ABNORMAL system does this by learning to predict Gabor and wavelet statistics from pure Gaussian noise, then flagging data where those statistics deviate [AG-2025.08-535]. Real non-Gaussian transients will naturally break the learned statistical pattern, even though the model never saw them during training.

Why it matters

In gravitational-wave astronomy, a single glitch mimicking a merger signal can trigger false alarms and waste follow-up resources; more subtly, glitches overlapping real signals can shift inferred black-hole masses or spins by measurable amounts [AG-2025.06-690]. As detector networks expand (more telescopes = more glitches), this problem grows worse [AG-2024.03-419]. Robust methods that either don't assume Gaussian noise (hyperbolic likelihoods [AG-2026.02-519], normalizing flows [AG-2024.05-300]) or explicitly learn non-Gaussian structure (phase randomization, third-order statistics [AG-2024.01-486]) restore sensitivity where naive Gaussian algorithms fail.