Why not flatten first?
Flattening preserves the numerical entries but can hide the meaning carried by separate modes.
The fluorescence-unmixing example makes the difference measurable:
- CP recovers the true component amounts at correlation 1.00.
- Flatten-then-SVD reaches only mean
|corr| ≈ 0.536. - The SVD representation can produce negative amount directions, even though a physical concentration cannot be negative.
Flattening is not necessarily wrong. The problem appears when we fuse two modes whose separate meaning was precisely the information we wanted to interpret.