How does source separation complexity scale when sources overlap in real observational data?

# Source Separation Complexity and Overlapping Signals

The complexity of separating overlapping sources in real data grows dramatically, but the nature of that growth depends on which method you use.

The Core Problem

When gravitational-wave signals overlap—in time, frequency, or both—traditional sequential methods (finding one source, subtracting it, then searching for the next) become unreliable. The reason is cascading error: misidentifying or partially subtracting the first signal corrupts all downstream parameter estimates [AG-2025.07-149]. As next-generation detectors reach unprecedented sensitivity, they will see so many overlapping signals that sequential analysis becomes impractical [AG-2024.09-171].

How Complexity Scales: The Methods

Traditional Bayesian inference (joint parameter estimation of all sources simultaneously) remains theoretically optimal but scales exponentially in computational cost with the number of overlapping signals. This is why blind source separation—methods that don't require knowing signal models in advance—have become essential [AG-2024.09-171].

Deep learning approaches offer a different scaling regime. An encoder-decoder architecture trained on LISA data can separate massive black-hole binaries, Galactic binaries, and instrumental glitches in a single pass, scaling more gracefully than sequential subtraction [AG-2025.03-625]. The UnMixFormer, an attention-based neural network, has demonstrated reliable separation of up to five overlapping compact binary coalescences simultaneously, maintaining near-constant computational cost across this range [AG-2024.12-534].

Iterative subtraction with neural density estimators offers a middle path: each iteration subtracts one signal using a neural network posterior sample, then re-infers remaining sources. The key insight is that with enough iterations and likelihood-based resampling to accelerate convergence, error does *not* accumulate catastrophically [AG-2025.07-149]. This method scales adaptively to source type and timing without retraining.

Matrix-based methods (Hankel low-rank approximation) exploit the mathematical structure that overlapping sinusoids form a low-rank matrix. Algorithms like ESPRIT and Cadzow iterations achieve near-optimal separation of closely spaced frequencies and even black-hole quasinormal modes, with mismatch scaling inversely with signal-to-noise ratio squared—a hallmark of statistical optimality [AG-2026.03-369].

Why This Matters

The diversity of methods matters because real LISA data will contain multiple *types* of overlapping sources (individual binaries plus confusion noise from unresolved populations) in mixed frequency bands [AG-2025.08-558]. No single technique dominates; instead, the field is converging on modular pipelines that apply different separators to different source classes, a pragmatic adaptation to the actual complexity observatories will face.