aiqbench analysis
This paper gives formal, computable bounds on how much a fixed ridge regression model's performance can differ from retraining on the full feature set, when the data contains uncertain or interval-valued entries. The results are tested on four real-world tasks, with certification reported at a width of 10^-5 and some speedups over a centered-moments baseline. For ML practitioners, the useful part is the verifiable conditions and the separation of certification dimension from feature dimension. The evidence is still limited to specific benchmarks and a rank-three setting, so generalization to other data regimes remains to be shown.
Certifying Retraining Regret through Residual Structure
arXiv:2610.41410v1 Announce Type: new Abstract: We certify a fixed ridge predictor's regret against full-feature retraining over every continuous completion of an interval-valued table. Our main theorem gives verifiable original-row residual conditions for a uniform error bound independent of nuisance-feature scale and rank, with an asymptotically sharp leading constant and quantitative leakage stability. Full-rank constructions separate certification dimension from feature dimension, including examples requiring directions beyond the candidate. Observed-row anchors make the conditions computable, while exact row geometry and global search return valid upper bounds and feasible original-table witnesses. A finite-scale contraction and common-domain moment fallback implement the structural result. Four real-world tasks retain all model features and certify twenty rank-three instances to width $10^{-5}$. Their primary median widths improve over centered moments. Joint residual refinement additionally certifies all twenty new response-independent masks to the same precision within 76.21 seconds. For a concrete fixed-point predictor contract, pure energy certifies thirteen of fifteen stopping decisions in ten seconds, versus five for moments. Refinement certifies all fifteen within 91.68 seconds at the same objective threshold.
Abstract:We certify a fixed ridge predictor's regret against full-feature retraining over every continuous completion of an interval-valued table. Our main theorem gives verifiable original-row residual conditions for a uniform error bound independent of nuisance-feature scale and rank, with an asymptotically sharp leading constant and quantitative leakage stability. Full-rank constructions separate certification dimension from feature dimension, including examples requiring directions beyond the candidate. Observed-row anchors make the conditions computable, while exact row geometry and global search return valid upper bounds and feasible original-table witnesses. A finite-scale contraction and common-domain moment fallback implement the structural result. Four real-world tasks retain all model features and certify twenty rank-three instances to width $10^{-5}$. Their primary median widths improve over centered moments. Joint residual refinement additionally certifies all twenty new response-independent masks to the same precision within 76.21 seconds. For a concrete fixed-point predictor contract, pure energy certifies thirteen of fifteen stopping decisions in ten seconds, versus five for moments. Refinement certifies all fifteen within 91.68 seconds at the same objective threshold.
46 pages, 5 figures. Code and reproduction materials: this https URL
Machine Learning (cs.LG)
arXiv:2610.41410 [cs.LG]
(or arXiv:2610.41410v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.41410 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
code/FROZEN_VALIDATION_V2.json
code/PROTOCOL_ADAPTIVE_V33.json
code/aggregate_split.py
code/anchor_energy.py
code/arrowhead_certificate.py
code/audit_adaptive_general_v33.py
code/audit_arrow_tree_v29.py
code/audit_arrowhead_exact.py
code/audit_compressed_arrow_v27.py
code/audit_compressed_exact.py
code/audit_contraction_real.py
code/audit_curvature_controls.py
code/audit_energy_decisions.py
code/audit_first_stages.py
code/audit_general_compact_v32.py
code/audit_general_refinement.py
code/audit_localized_refinement.py
code/audit_matched_controls.py
code/audit_panel_alignment.py
code/audit_profile_outer_v31.py
code/audit_real_world.py
code/audit_refinement_cover.py
code/audit_row_edge_refinement.py
code/audit_scip_objective.py
code/audit_sdp_exact.py
code/audit_structure.py
code/audit_tail_diagnostics.py
code/audit_zorro_matched.py
code/audit_zorro_refined.py
code/auto_energy.py
code/auto_energy_audit.py
code/auto_energy_protocol.md
code/auto_energy_run.py
code/automatic_basis.py
code/calibration_analyze.py
code/calibration_protocol.md
code/calibration_run.py
code/calibration_sensitivity.py
code/centered_signal_energy.py
code/compressed_edge_arrow_v27.py
code/compressed_offline_v29.json
code/compressed_polynomial.py
code/confirmation_protocol.md
code/contraction_energy.py
code/contraction_settings.json
code/curvature_controls.py
code/datasets.py
code/decision_settings.json
code/discounted_planar_oracle.py
code/download_contraction_data.py
code/download_extra.py
code/download_public.py
code/dual_certificate.py
code/edge_polynomial.py
code/epsilon_audit.py
code/epsilon_energy.py
code/epsilon_protocol.md
code/epsilon_run.py
code/experiment_settings.json
code/extend_validation_baselines.py
code/fast_interval.py
code/fixed_point_contract.py
code/fresh_arrow_v29.py
code/greedy_audit.py
code/greedy_basis.py
code/greedy_pool.json
code/greedy_protocol.md
code/greedy_run.py
code/halfspace_moment.py
code/independent_mask_settings.json
code/independent_moment_audit.py
code/independent_planar_rows.py
code/interval_arithmetic.py
code/intrinsic_dimension.py
code/leakage_audit.py
code/leakage_energy.py
code/leakage_protocol.md
code/leakage_run.py
code/local_residual.py
code/localized_moment.py
code/metric_audit.py
code/metric_basis.py
code/metric_protocol.md
code/metric_run.py
code/moment_reference.py
code/moment_tail.py
code/multirank_audit.py
code/multirank_run.py
code/nuisance_scale.py
code/nuisance_scale_protocol.md
code/observed_anchor.py
code/pilot_arrow_tree_v28.py
code/planar_audit.py
code/planar_density_run.py
code/planar_oracle.py
code/planar_protocol.md
code/planar_refine.py
code/planar_run.py
code/plot_final_scope.py
code/plot_paper_figures.py
code/polynomial_majorant.py
code/prepare_public_data.py
code/profile_offline_v32.json
code/profile_outer_v31.py
code/profile_query_v31.py
code/project_paths.py
code/protocol.md
code/public_energy_rank3_run.py
code/public_energy_run.py
code/public_problems.py
code/quadratic_majorant.py
code/rank_certificate.py
code/real_world_design.py
code/refinement_pipeline.py
code/rep_pool_pools.json
code/rep_pool_protocol.md
code/rep_pool_run.py
code/residual_energy.py
code/robust_reference.py
code/row_edge_exact.py
code/row_edge_input.py
code/row_edge_offline_checks.json
code/row_edge_oracle.py
code/row_edge_screen.py
code/run_adaptive_general_v33.py
code/run_additional_components.py
code/run_component_base.py
code/run_confirmation.py
code/run_contraction_real.py
code/run_dense_compressed.py
code/run_dense_refinement.py
code/run_energy_decisions.py
code/run_frozen_refinement.py
code/run_general_compact_v32.py
code/run_general_refinement.py
code/run_independent_masks.py
code/run_localized_refinement.py
code/run_matched_controls.py
code/run_multirank_exact.py
code/run_precision_controls.py
code/run_real_world.py
code/run_row_edge_refinement.py
code/run_saved_refinement.py
code/run_scip.py
code/run_scip_matched.py
code/run_structure.py
code/run_tail_diagnostics.py
code/run_zorro_matched.py
code/run_zorro_refined.py
code/sdp_residual.py
code/uniform_curvature_v28.py
code/verify_compressed_v29.py
code/verify_contraction_offline.py
code/verify_offline.py
code/verify_profile_v32.py
code/verify_refinement_offline.py
code/verify_row_edge_offline.py
code/zorro_refined.py
code/zorro_regret.py
requirements-general-linux.txt
requirements-refinement-windows.txt
requirements.txt
results/acquisition/calibration_audit.json
results/acquisition/calibration_plan.json
results/acquisition/calibration_summary.csv
results/acquisition/public_energy_rank3_results.json
results/contraction_real/adaptive_independent_audit.json
results/contraction_real/aggregate.json
results/contraction_real/independent_audit.json
results/contraction_real/pure_energy_summary.csv
results/contraction_real/summary.csv
results/dimension_and_decisions/calibration_endpoints.json
results/dimension_and_decisions/calibration_sensitivity.json
results/dimension_and_decisions/intrinsic_dimension_audit.json
results/energy_decisions/FROZEN_DESIGN.json
results/energy_decisions/aggregate.json
results/energy_decisions/conditions.json
results/energy_decisions/fresh_fixed_point_contract.json
results/energy_decisions/independent_audit.json
results/energy_decisions/primary_fixed_point_contract.json
results/energy_decisions/sensitivity.json
results/energy_decisions/summary.csv
results/energy_mechanism/auto_energy_audit_results.json
results/energy_mechanism/auto_energy_results.json
results/energy_mechanism/curvature_audit.json
results/energy_mechanism/curvature_controls.json
results/energy_mechanism/nuisance_scale_results.json
results/general_adaptive_v33/all_conditions.json
results/general_adaptive_v33/conditions.csv
results/general_adaptive_v33/environment.json
results/general_adaptive_v33/protocol.json
results/general_adaptive_v33/summary.json
results/general_coarsening/additional_components.json
results/general_coarsening/confirmation_summary.csv
results/independent_masks/aggregate.json
results/independent_masks/independent_audit.json
results/independent_masks/summary.csv
results/matched_controls/aggregate.json
results/matched_controls/independent_audit.json
results/matched_controls/regret_levels.json
results/matched_controls/summary.csv
results/matched_controls/tail_components.json
results/matched_external/anytime_curves.json
results/matched_external/midpoint_increments.json
results/matched_external/scip_all_masks.csv
results/matched_external/scip_all_masks.json
results/matched_external/scip_strict_results.json
results/matched_external/scip_summary.csv
results/matched_external/scip_target_results.json
results/matched_external/zorro_independent_audit.json
results/matched_external/zorro_refined_independent_audit.json
results/matched_external/zorro_refined_summary.json
results/matched_external/zorro_summary.json
results/metric_and_leakage/leakage_audit_results.json
results/metric_and_leakage/leakage_results.json
results/metric_and_leakage/metric_audit.json
results/metric_and_leakage/metric_comparison.json
results/multiple_directions/multirank_audit_results.json
results/multiple_directions/multirank_results.json
results/precision_and_scip/SCIP_FOLLOWUP_REPORT.json
results/precision_and_scip/precision_row_summary.csv
results/precision_and_scip/scip_followup_summary.csv
results/precision_and_scip/scip_objective_audit.json
results/real_world/confirmation_summary.csv
results/real_world/independent_audit_results.json
results/real_world/public_energy_aggregate.json
results/refinement_v92/FROZEN_GENERAL_V3.json
results/refinement_v92/FROZEN_VALIDATION_V2.json
results/refinement_v92/audit_summary.json
results/refinement_v92/controls_baseline_extended.json
results/refinement_v92/decisions_baseline_extended.json
results/refinement_v92/dense_edge_summary.json
results/refinement_v92/extended_decision_comparison.json
results/refinement_v92/frozen_controls.json
results/refinement_v92/frozen_decisions.json
results/refinement_v92/frozen_independent.json
results/refinement_v92/frozen_validation.json
results/refinement_v92/general_superconductivity.json
results/refinement_v92/multirank_exact.json
results/refinement_v92/summary.csv
results/refinement_v92/validation_baseline_extended.json
results/residual_controls/corrected_direction_summary.json
results/residual_controls/epsilon_independent_audit.json
results/residual_controls/epsilon_results.json
results/row_geometry/planar_audit.json
results/row_geometry/planar_density_results.json
results/row_geometry/planar_refinement.json
results/row_geometry/planar_results.json
results/row_geometry/structure_robust_results.json
results/structure/independent_structure_audit.json
results/structure/structure_results.json
results/structure/structure_summary.csv
results/tail_diagnostics/diagnostics.csv
results/tail_diagnostics/independent_gram_audit.json
results/tail_diagnostics/summary.json
(251 additional files not shown) You must enabled JavaScript to view entire file list.
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 541 words
Related coverage
From Mathematical to Executable Certificates for Machine Unlearning
arXiv · Oct 5, 2026
Certified Safety Curation: Distribution-Free Guarantees for Safe Offline Reinforcement Learning
arXiv · Sep 14, 2026
Erased, but Not Gone: Output Forgetting Is Not True Forgetting
arXiv · Oct 12, 2026
Loss is Not Behavior: A Unified Output-Space Analysis of Gradient-Based Machine Unlearning
arXiv · Oct 30, 2026