Structural conditions on trust-region radius updates. Part II: second-order convergence and asymptotic radius behaviour
Abstract
Continuing Part I (arXiv:2606.30202), we ask not which structural conditions on {Δk} suffice for global convergence but what the radius does; the answers separate the mechanisms by whether the radius is tied to a criticality measure. We identify three asymptotic regimes: the criticality-anchored rules give ∑k Δk2/Mk < +∞, hence liminfk Δk = 0, and Δk → 0 once the model Hessians are bounded; the step-driven rules inherit the summability of the step series, hence ∑k Δk < +∞; the fixed-factor rules keep Δk bounded away from zero. Near a critical point with positive definite Hessian, convergence of a subsequence forces convergence of the full sequence. We then extend the framework to second-order criticality through τk = max{|gk|, −λk-}, obtaining convergence to second-order critical points with the worst-case complexity 𝒪(max{εg−2εH−1, εH−3}), which reduces to 𝒪(ε−3) for a common tolerance and is sharp for quadratic models. Finally we establish eventual inactivity of the constraint and superlinear convergence, isolating the two properties of the radius rule through which the mechanism enters; the rate itself is mechanism-independent. The criticality-anchored and capped gradient-scaled rules reach inactivity only under a side condition, ζ> κ̅ or μ̅ > κ̅, with κ̅ a constant of the solution that cannot be checked in advance. These conditions are necessary, not artefacts: on f(x) = sin(x)/x, κ̅ is unbounded over the minimisers, so no parameter value meets the condition at all of them. Among the rules tied to a criticality measure, only those with an unbounded multiplier escape such a constant.
Related
- Structural conditions on trust-region radius updates. Part I: first-order convergence and complexity
- High-Probability Convergence of SGD via Batched Updates
- Online adaptive non-intrusive model reduction via manifold interpolation and subspace updates: application to FSI convergence acceleration
- Asymptotic behaviour of the v-number of homogeneous ideals
- Boundary Blowup Solutions for the Finsler p-Laplacian: Wellposedness and Asymptotic Behaviour
BibTeX
@misc{rieussec2026structural,
title = {{Structural conditions on trust-region radius updates. Part II: second-order convergence and asymptotic radius behaviour}},
author = {Jérémy Rieussec and Fabian Bastin},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.26137},
eprint = {2610.26137},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26137},
note = {CC BY-SA 4.0},
}
RIS
TY - UNPB
TI - Structural conditions on trust-region radius updates. Part II: second-order convergence and asymptotic radius behaviour
AU - Jérémy Rieussec
AU - Fabian Bastin
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2610.26137
UR - https://arxiv.org/abs/2610.26137
AB - Continuing Part I (arXiv:2606.30202), we ask not which structural conditions on {Δ_k} suffice for global convergence but what the radius does; the answers separate the mechanisms by whether the radius is tied to a criticality measure. We identify three asymptotic regimes: the criticality-anchored rules give ∑_k Δ_k^2/M_k < +∞, hence liminf_k Δ_k = 0, and Δ_k → 0 once the model Hessians are bounded; the step-driven rules inherit the summability of the step series, hence ∑_k Δ_k < +∞; the fixed-factor rules keep Δ_k bounded away from zero. Near a critical point with positive definite Hessian, convergence of a subsequence forces convergence of the full sequence. We then extend the framework to second-order criticality through τ_k = max{|g_k|, −λ_k^-}, obtaining convergence to second-order critical points with the worst-case complexity 𝒪(max{ε_g^−2ε_H^−1, ε_H^−3}), which reduces to 𝒪(ε^−3) for a common tolerance and is sharp for quadratic models. Finally we establish eventual inactivity of the constraint and superlinear convergence, isolating the two properties of the radius rule through which the mechanism enters; the rate itself is mechanism-independent. The criticality-anchored and capped gradient-scaled rules reach inactivity only under a side condition, ζ> κ̅ or μ̅ > κ̅, with κ̅ a constant of the solution that cannot be checked in advance. These conditions are necessary, not artefacts: on f(x) = sin(x)/x, κ̅ is unbounded over the minimisers, so no parameter value meets the condition at all of them. Among the rules tied to a criticality measure, only those with an unbounded multiplier escape such a constant.
ER -