Definition
Signal separation extracts individual sources from mixtures:
y = As \quad \xrightarrow{\text{separation}} \quad \hat{s}
In SCU terms: Separation untangles mixed χ-mode sources—recovering individual χ-mode signals from superposed observations.
The Cocktail Party Problem
Multiple χ-mode sources mix at sensors:
y_1 = a_{11}s_1 + a_{12}s_2 + ...
y_2 = a_{21}s_1 + a_{22}s_2 + ...
How to recover individual sources $s_i$?
Separation Approaches
| Method | χ-Mode Assumption |
|---|---|
| Spatial filtering | Different χ-mode directions |
| ICA | Statistically independent χ-modes |
| NMF | Non-negative χ-mode components |
| Sparsity | Distinct χ-mode frequency content |
Independent Component Analysis (ICA)
Sources are statistically independent:
P(s_1, s_2) = P(s_1) \cdot P(s_2)
Find unmixing matrix W to maximize independence.
Spatial Methods
Sensor arrays exploit χ-mode directionality:
\text{Beamforming: } y = \sum_i w_i x_i
Steer toward desired χ-mode source.
Requirements for Separation
| Requirement | Why Needed |
|---|---|
| Multiple sensors | Spatial χ-mode diversity |
| Statistical independence | Different χ-mode statistics |
| Non-Gaussianity | Distinguish from noise |
| Sparsity | Non-overlapping χ-mode support |
Challenges
| Challenge | Problem |
|---|---|
| Underdetermined | More sources than sensors |
| Non-stationary | Mixing changes over time |
| Convolutive | χ-modes delayed and filtered |
| Permutation | Source order ambiguous |
Applications
| Field | Separation Task |
|---|---|
| Audio | Separate voices, instruments |
| Brain imaging | Isolate neural χ-modes |
| Communications | Multi-user separation |
| Astrophysics | Foreground removal |
The Key Insight
Separation untangles χ-mode mixtures.
Recovering individual sources:
- Multiple χ-modes arrive superposed
- Statistical properties differ
- Algorithms exploit these differences
- Individual sources recovered
When we separate signals, we're using the statistical structure of χ-mode sources to untangle their superposition—recovering what came from where.