Description
PCA is the linear projection that maximises variance along orthogonal axes. In CSI sensing it serves three roles: (1) compact representation of redundant subcarriers, (2) removal of dominant static components so motion energy stands out, and (3) baseline against learned encoders. It is fast, reproducible, and a sensible starting point before reaching for autoencoders.
When it's used
- Subcarrier-redundancy reduction
- Motion-component extraction (subtracting top components leaves motion energy)
- Visualisation / clustering of CSI feature spaces
Limitations
- Linear; cannot capture interactions
- Top components dominated by hardware DC offsets if not preprocessed
- Outperformed by learned representations on rich CSI