Admit a budget-limited trust-region step using its model cost
Compare a returned step with a feasible interior candidate using the original rank-deficient subproblem objective.
A subproblem before convergence status
Compare the returned step’s model cost with an independently feasible interior candidate before calling a finite-budget failure ordinary nonconvergence. A calculation that eventually converges under defaults may still be unusable under a limited budget. Being on the trust-region boundary alone proves neither correctness nor failure. The Truthound inputs are externally recomputed model cost and the cost of a feasible reference candidate. [1, 2, 3]
Default and limited budgets are different
SciPy #26416 reports unnecessary boundary steps and a negative LM parameter in the trf exact solver for rank-deficient subproblems. Its public examples converge with the default budget. The report explicitly uses model cost as evidence and notes that a boundary minimizer can be legitimate; this article does not reject every boundary step. [1]
Verify a feasible reference candidate
Retain J, f, the trust radius, and the returned step, then independently recompute the original q(p). Hold a path that spends evaluations on a worse candidate while comparing solver and budget alternatives. Verify the reference candidate’s norm as well as its cost under the same subproblem. [1, 3]
An independent basis for cost 0
For illustrative J=diag(1,0), f=(-0.001,0), and radius=1, p=(0.001,0) is feasible with q=0. A returned p=(1,0) has q=0.4990005. An observed cost 0 passes the <=0 comparison; 0.4990005 fails. The reference comes from substitution into the original objective and a feasible candidate. [1, 3]
Cost comparison is not optimization
ColumnPairValidator in Truthound Python SDK 3.3.2 compares externally computed returned_model_cost and independently_feasible_cost using <=. Operators supply J, f, radius, and feasibility checks. Core neither solves least_squares nor the trust-region problem; passing does not establish global nonlinear optimality or convergence. [1, 2, 3]
Compare alternatives with their budgets
Compare a verified alternative trust solver using the same objective, evaluation budget, constraints, and rank conditions. Recheck residuals rather than merely increasing a limit until execution succeeds. Existing independent cost and feasibility checks reduce the benefit of a duplicate audit. [1, 3]
Application conditions
Externally verify the costs and feasibility of actual and independent reference candidates under the same J, f, and radius.
Alternatives
Compare a verified trust solver under matching constraints, rank conditions, and evaluation budgets.
Limitations
Passing the cost comparison does not guarantee optimality or convergence of the full nonlinear problem.
Evidence
- Event checked
- Subject checked
- Modified
