Statistical Methods
Review Active Bounds in a Constrained Linear Fit
Define the task
A coefficient bound should represent a justified scientific or operational restriction, not a device for forcing a preferred result. Document units, lower and upper limits, and why each applies before fitting. A bounded linear least-squares problem minimizes half the squared residual norm subject to coefficient bounds. This differs from clipping predictions afterward or replacing negative coefficients in an unconstrained fit with zero.
Carry out the check
Check matrix and outcome alignment, rank and scaling. Inspect convergence and active bounds, then independently recompute predictions and residuals. The objective is convex, but numerical convergence still needs checking, and rank deficiency may permit nonunique coefficients. A coefficient touching its bound is a substantive finding about the constrained fit. Ordinary unconstrained standard errors do not automatically apply; decide how uncertainty will be assessed under the restrictions.
Worked hypothetical example
Hypothetical example: fit one coefficient β to design column [1,1] and outcomes [−1,3], with 0≤β≤0.5. The unconstrained optimum is the mean 1, outside the interval. The constrained optimum is 0.5, giving residuals [1.5,−2.5] and objective (2.25+6.25)/2=4.25. The derivative there is −1, so the objective would decrease by moving beyond the permitted upper bound. Record that active limit instead of presenting 0.5 as an unrestricted estimate.
Method checked 22 November 2026 against Primary method documentation: lsq_linear.