Prerequisites
Linear algebra, nullspaces, least squares, and quadratic optimization.
A null direction changes parameters without changing predictions. A penalty can select one estimate but does not add observational information. Separate identification by data from selection by a regularization rule.
Worked example
If only x1 + x2 = 8 is observed, (3,5) and (5,3) fit equally well. The minimum-norm exact fit is (4,4), selected by symmetry and the norm rather than by separate measurements.
Model and Assumptions
An invented noiseless model observes y = x1 + x2 = 6. For lambda > 0, minimize J = (x1 + x2 - 6)^2 + lambda(x1^2 + x2^2). No nonnegativity constraint is imposed.
Further inquiry
Derive the minimizer for arbitrary positive lambda. Compare its prediction bias with the unpenalized solution family. Design a second measurement and prove the combined matrix has full rank.
Review criteria
- Show x1 = x2 = 6/(2 + lambda).
- Explain why stabilization does not resolve data-only identifiability.
- Check independence of the proposed measurement.