Documentation

MIPRE.Background.LIDT.MIPStarRE.LDT.SelfImprovement.Theorems.Thresholds.Final

Section 9 — final-stage numerical threshold absorptions #

This module contains the final-stage arithmetic comparisons which absorb the natural projective-output error terms into the literal selfImprovementError threshold. The estimates formalize the exponent-monotonicity bookkeeping in references/ldt-paper/self_improvement.tex, lines 803--810, and are used by the final-field construction theorems in the self-improvement theorem.

The helper-stage estimates, including the square-root bound on the global-variance error, are in Thresholds.Helper.

Final-stage threshold: 30 ζ̂ ≤ ζ #

For ε, δ, d/q ∈ [0, 1], the helper-stage error is dominated by selfImprovementError / 30. This is the exponent-monotonicity step in the paper's chained absorption (self_improvement.tex, lines 803--810): for unit-interval x and 0 ≤ z ≤ y, x^y ≤ x^z.

noncomputable def MIPStarRE.LDT.SelfImprovement.finalStagePowerSum (params : Parameters) (eps delta p : Error) :

The three-term power sum used in the final-stage self-improvement threshold bookkeeping.

For exponent p, this is the Lean counterpart of ε^p + δ^p + (d/q)^p. Keeping this expression named makes the nested absorptions in self_improvement.tex, lines 772--810, insensitive to future constant adjustments.

Equations
Instances For
    theorem MIPStarRE.LDT.SelfImprovement.finalStagePowerSum_nonneg (params : Parameters) (eps delta p : Error) (heps : 0 eps) (hdelta : 0 delta) :
    0 finalStagePowerSum params eps delta p

    The final-stage power sum is nonnegative when its base parameters are nonnegative.

    This is the positivity fact used when multiplying coefficient comparisons by ε^p + δ^p + (d/q)^p.

    theorem MIPStarRE.LDT.SelfImprovement.finalStagePowerSum_le_of_exponent_ge (params : Parameters) (eps delta : Error) (heps : 0 eps) (heps_le_one : eps 1) (hdelta : 0 delta) (hdelta_le_one : delta 1) (hdq_le_one : params.d / params.q 1) {pLarge pSmall : Error} (hpSmall_nonneg : 0 pSmall) (hpSmall_le_pLarge : pSmall pLarge) :
    finalStagePowerSum params eps delta pLarge finalStagePowerSum params eps delta pSmall

    Exponent monotonicity for the final-stage power sum on the unit interval.

    If ε, δ, and d/q all lie in [0, 1], then increasing the exponent decreases each term of the power sum. This is the formal arithmetic step behind the final threshold comparison in self_improvement.tex, lines 803--810.

    theorem MIPStarRE.LDT.SelfImprovement.sqrt_finalStagePowerSum_le (params : Parameters) (eps delta p : Error) (heps : 0 eps) (hdelta : 0 delta) :
    (finalStagePowerSum params eps delta p) finalStagePowerSum params eps delta (p / 2)

    The square root of a final-stage power sum is bounded by the power sum with half the exponent.

    The proof is the elementary estimate sqrt (x + y + z) ≤ sqrt x + sqrt y + sqrt z, applied to the three powers ε^p, δ^p, and (d/q)^p.

    theorem MIPStarRE.LDT.SelfImprovement.selfImprovementHelperError_eq_finalStagePowerSum (params : Parameters) [FieldModel params.q] (eps delta : Error) :
    selfImprovementHelperError params eps delta = 100 * params.m * finalStagePowerSum params eps delta (1 / 2)

    The helper-stage threshold as the final-stage power sum with exponent 1/2.

    This is definitionally equal to selfImprovementHelperError; the named lemma is used to avoid unfolding the error definition inside later absorptions.

    theorem MIPStarRE.LDT.SelfImprovement.selfImprovementError_eq_finalStagePowerSum (params : Parameters) [FieldModel params.q] (eps delta : Error) :
    selfImprovementError params eps delta = 3000 * params.m * finalStagePowerSum params eps delta (1 / 32)

    The final self-improvement threshold as the final-stage power sum with exponent 1/32.

    This records the literal error parameter appearing in the formal statement of the self-improvement theorem.

    The orthogonalization threshold written as an iterated square root of the helper threshold.

    The paper writes the corresponding quantity as 100 * \widehat{\zeta}^{1/4}. This lemma gives the equivalent square-root form used by the estimates which follow orthonormalization.

    The data-processing threshold written with an ordinary square root.

    This is the displayed expression used after projecting the helper output: 8\widehat{\zeta} + 8\sqrt{\widehat{\zeta}_{\mathrm{ortho}}}.

    The data-processing threshold contains the term 8 * selfImprovementHelperError.

    This lower bound is the scalar source of the alphabet-size obstruction in the older total-gap point-consistency transport: once the total-overlap estimate contributes sqrt (#F_q * selfImprovementDataProcessingError), this positive summand carries the cardinality of the alphabet into the final-stage error.

    The alphabet-size square-root term is bounded below by the corresponding term coming from 8 * selfImprovementHelperError.

    This records that, in the older total-gap final-fields transport, the cardinality factor cannot be removed by lower-bounding selfImprovementDataProcessingError through 8 * selfImprovementHelperError. The structural obstruction is the surrounding estimate carrying sqrt (#F_q * selfImprovementDataProcessingError) into a final threshold which has no corresponding alphabet-size term.

    theorem MIPStarRE.LDT.SelfImprovement.final_fields_projective_residual_error_le_selfImprovementError (params : Parameters) [FieldModel params.q] (eps delta : Error) (heps : 0 eps) (heps_le_one : eps 1) (hdelta : 0 delta) (hdelta_le_one : delta 1) (hd_le_q : params.d params.q) :

    Final projective-residual threshold absorption (self_improvement.tex, lines 803--810).

    The natural error emitted by the projective residual construction is ζ̂ + √ζ̂_dataprocess. Under the standard unit-interval hypotheses for ε, δ, and d/q, this theorem absorbs that natural error into the literal selfImprovementError threshold used by SelfImprovementConclusion. Lean records the third unit-interval hypothesis as d ≤ q, equivalently d/q ≤ 1 because q is positive.

    theorem MIPStarRE.LDT.SelfImprovement.final_fields_point_consistency_total_difference_error_le_selfImprovementError (params : Parameters) [FieldModel params.q] (eps delta : Error) (heps : 0 eps) (heps_le_one : eps 1) (hdelta : 0 delta) (hdelta_le_one : delta 1) (hd_le_q : params.d params.q) :

    Final point-consistency threshold absorption from a total-difference bound.

    The fallback submeasurement-total route contributes the natural error

    2 * ζ̂ + √ζ̂_dataprocess + 2 * √ζ̂_ortho.

    Under the standard unit-interval hypotheses for ε, δ, and d/q, this is still absorbed by the literal selfImprovementError threshold.

    theorem MIPStarRE.LDT.SelfImprovement.final_fields_completeness_error_le_selfImprovementError (params : Parameters) [FieldModel params.q] (eps delta : Error) (heps : 0 eps) (heps_le_one : eps 1) (hdelta : 0 delta) (hdelta_le_one : delta 1) (hd_le_q : params.d params.q) :
    2 * selfImprovementHelperError params eps delta + 2 * (selfImprovementOrthogonalizationError params eps delta) selfImprovementError params eps delta

    Final completeness threshold absorption (self_improvement.tex, lines 803--810).

    The natural error emitted by the projective completeness transport is 2ζ̂ + 2 sqrt ζ̂_ortho. Under the standard unit-interval hypotheses for ε, δ, and d/q, this theorem absorbs that natural error into the literal selfImprovementError threshold used by SelfImprovementFinalFields. Lean records the third unit-interval hypothesis as d ≤ q, equivalently d/q ≤ 1 because q is positive.

    theorem MIPStarRE.LDT.SelfImprovement.final_fields_self_closeness_error_le_selfImprovementError (params : Parameters) [FieldModel params.q] (eps delta : Error) (heps : 0 eps) (heps_le_one : eps 1) (hdelta : 0 delta) (hdelta_le_one : delta 1) (hd_le_q : params.d params.q) :

    Final self-closeness threshold absorption (self_improvement.tex, lines 803--810).

    The natural error emitted by the projective self-closeness transport is 3 * (ζ̂_ortho + 2ζ̂ + ζ̂_ortho). Under the standard unit-interval hypotheses for ε, δ, and d/q, this theorem absorbs that natural error into the literal selfImprovementError threshold used by SelfImprovementFinalFields. Lean records the third unit-interval hypothesis as d ≤ q, equivalently d/q ≤ 1 because q is positive.