Tangent Excess and Persistent Difference-of-Convex Regularization for Nonconvex QCQPs

Componentwise convexification can impose a local convergence cost on nonconvex quadratically constrained quadratic programs (QCQPs). We derive a lower bound on the relative tangent excess forced by negatively curved components with positive Lagrangian weights. QCQP examples attain this bound. We characterize corrections confined to normal directions. Established sensitivity formulas convert this excess κ into the factor κ/(1 + κ), reproduced by controlled trajectories. LazyQCQP reuses difference-of-convex corrections, refines components and preserves signed quadratic-equality curvature. A specified filter algorithm has stationary accumulation points under explicit assumptions. At a regular Karush–Kuhn–Tucker (KKT) limit, a sufficiently small curvature threshold makes additional diagonal regularization unnecessary. On 4,078 QCQPs, six configurations are compared with Ipopt using an independent original-KKT certificate. With MUMPS for both solvers, the baseline LazyQCQP configuration (R3) certifies 98.9% of 1,037 cases with quadratic inequalities and no quadratic equality, versus 97.1% for monotone Ipopt with certificate-matched stopping targets. On 797 mixed cases with both quadratic constraint types, the rates are 82.8% versus 53.1%. MA27 reverses the first ordering (96.0% versus 97.2%) but preserves the second (83.2% versus 55.0%). For instances certified by both solvers, with both process times above one second, the median ratio of R3 process time to Ipopt process time with MUMPS is 0.34 in the first class and 0.48 in the mixed class. A controlled radio-beamforming ablation shows that removing quadratically implied bounds lowers attained objectives but can reduce certificate counts. On 12 instances, reuse preserves trajectories and reduces median paired process times by approximately 15% with either linear solver.

Authors

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11732465/v1
Primary Topic
Advanced Optimization Algorithms Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Tangent Excess and Persistent Difference-of-Convex Regularization for Nonconvex QCQPs

François Bertrand Akoa
Research Square
Advanced Optimization Algorithms Research
preprint

Tangent Excess and Persistent Difference-of-Convex Regularization for Nonconvex QCQPs

François Bertrand Akoa
preprint en

Abstract

Componentwise convexification can impose a local convergence cost on nonconvex quadratically constrained quadratic programs (QCQPs). We derive a lower bound on the relative tangent excess forced by negatively curved components with positive Lagrangian weights. QCQP examples attain this bound. We characterize corrections confined to normal directions. Established sensitivity formulas convert this excess κ into the factor κ/(1 + κ), reproduced by controlled trajectories. LazyQCQP reuses difference-of-convex corrections, refines components and preserves signed quadratic-equality curvature. A specified filter algorithm has stationary accumulation points under explicit assumptions. At a regular Karush–Kuhn–Tucker (KKT) limit, a sufficiently small curvature threshold makes additional diagonal regularization unnecessary. On 4,078 QCQPs, six configurations are compared with Ipopt using an independent original-KKT certificate. With MUMPS for both solvers, the baseline LazyQCQP configuration (R3) certifies 98.9% of 1,037 cases with quadratic inequalities and no quadratic equality, versus 97.1% for monotone Ipopt with certificate-matched stopping targets. On 797 mixed cases with both quadratic constraint types, the rates are 82.8% versus 53.1%. MA27 reverses the first ordering (96.0% versus 97.2%) but preserves the second (83.2% versus 55.0%). For instances certified by both solvers, with both process times above one second, the median ratio of R3 process time to Ipopt process time with MUMPS is 0.34 in the first class and 0.48 in the mixed class. A controlled radio-beamforming ablation shows that removing quadratically implied bounds lowers attained objectives but can reduce certificate counts. On 12 instances, reuse preserves trajectories and reduces median paired process times by approximately 15% with either linear solver.

Research Square
Advanced Optimization Algorithms Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon