How does ICA's performance degrade when sources have correlated rather than independent noise?

# ICA Performance Under Correlated Noise

Your question asks about Independent Component Analysis (ICA)—a signal processing technique that separates mixed signals into independent sources—and how it breaks down when noise sources become correlated rather than independent.

The standard assumption in ICA is encoded in its name: the underlying *sources* are statistically independent. When this assumption holds, ICA can cleanly separate a mixture of signals. But when noise becomes correlated across channels, the mathematical foundation cracks.

Here's why it matters: if noise components are correlated, ICA will misattribute that shared noise structure to the signal sources themselves, degrading both source separation and parameter estimation [AG-2025.09-247]. In gravitational wave detection, for instance, correlated environmental noise (like seismic vibrations felt simultaneously by multiple detectors) can fool standard ICA into treating the correlation as part of the astrophysical signal.

The practical consequence varies by context. In gravitational wave detectors, neglecting correlated noise entirely leads to significant biases in reconstructing the actual signal parameters [AG-2025.01-134]. The archive shows that gravitational wave collaborations have moved beyond simple ICA: they now use variational inference methods and Bayesian modeling that explicitly account for off-diagonal correlations in the noise covariance matrix [AG-2024.09-331], and newer frameworks use *nonlinear* ICA variants that can handle non-linearly coupled correlated noise [AG-2025.09-247].

For space-based detectors like LISA, the problem is structural—different time-delay interferometry channels inherently have correlated noise—so standard ICA assumptions fail from the start [AG-2025.04-728].

The deeper lesson: ICA works well in the idealized regime but requires careful extension (Bayesian priors, explicit noise modeling, adaptive covariance reconstruction) to handle real-world correlated noise.