ERM
Finite-class ERM tail
Lean verified
finite-class ERM excess-risk tail under bounded loss
Lean declaration
Scope
finite class, bounded loss, iid sample, exact ERM
Boundary
finite class, bounded loss, iid sample, exact ERM
Verification dashboard
TheoremPath publishes a scoped set of Lean-verified theorems governed by a manifest. Statements in the manifest are the canonical claims; page prose is informal exposition. Verification is per-claim, not per-page.
Verified entries
56
Manifest declarations compiled in CI with zero recorded sorry/admit.
Axiom set
core only
propext, Classical.choice, Quot.sound
Lean toolchain
4.30.0-rc2
mathlib 25b7ac7d0c
Manifest version
2026-05-07
Most recent check: May 7, 2026.
These cards form the learning-theory front door: selected topic claims mapped to governed Lean declarations, with scope and proof boundaries shown beside the theorem. The complete audit log remains the manifest table below: 56 verified entries total, with 13 highlighted here.
ERM
Lean verified
finite-class ERM excess-risk tail under bounded loss
Lean declaration
Scope
finite class, bounded loss, iid sample, exact ERM
Boundary
finite class, bounded loss, iid sample, exact ERM
Symmetrization
Lean verified
E[genGap] <= 2 * E[empirical Rademacher complexity]
Lean declaration
Scope
finite index class, iid sample, bounded measurable losses
Boundary
finite index class, iid sample, bounded measurable losses
Rademacher
Lean verified
P(genGap >= 2 * E[Rad] + eps) has Azuma tail
Lean declaration
Scope
finite class, bounded loss, one-sided genGap, Azuma constant
Boundary
Lean formalization proves the one-sided genGap bound with Azuma constant (factor-of-4 looser than sharp McDiarmid). The governed page claim states the full two-sided uniformDeviation form with sharp McDiarmid constant and confidence-delta parametrization. The Lean result is a strict subset of the page claim.
Rademacher
Lean verified
empirical Rademacher complexity <= B * sqrt(2 log card / n)
Lean declaration
Scope
finite effective class, bounded loss, nontrivial cardinality case
Boundary
finite effective class, bounded loss, nontrivial cardinality case
VC
Lean verified
Rad <= B * sqrt(2d log(en/d) / n)
Lean declaration
Scope
finite class, user-supplied effective-growth assumption
Boundary
User supplies the effective-growth assumption for generic losses. For binary classifiers with 0-1 loss, the bridge is closed (BinaryVCBridge.lean).
VC
Lean verified
two-sided uniform deviation via VC growth and Azuma
Lean declaration
Scope
finite class, bounded loss, growth assumptions for loss and negated loss
Boundary
Azuma constant (8B² not 2B²). Binary VC bridge now closed for 0-1 loss. Generic losses still need user-supplied growth.
VC
Lean verified
0-1 loss effective class card = binary trace card
Lean declaration
Scope
binary classifiers, 0-1 loss, finite class
Boundary
Binary classifiers with 0-1 loss only. Does not cover real-valued loss or multi-class. No contraction.
Rademacher
Lean verified
Rad_S(phi o F) <= L * Rad_S(F)
Lean declaration
Scope
finite sample, finite class, scalar-valued functions
Boundary
Finite sample, finite nonempty hypothesis class, scalar outputs only; not full general Talagrand contraction.
Rademacher
Lean verified
Rad_S({x -> <w, x>}) <= R / n * sqrt(sum_i ||x_i||^2)
Lean declaration
Scope
finite-dimensional Euclidean features, finite indexed weights
Boundary
Finite sample, finite nonempty indexed weight family, finite-dimensional Euclidean feature space only.
PAC-Bayes
Lean verified
finite bounded-loss Catoni bad-event mass <= delta
Lean declaration
Scope
finite hypotheses, finite prior/posterior PMFs, [0,1] losses
Boundary
Finite data domain, finite nonempty hypothesis index, finite prior/posterior PMFs, scalar [0,1] losses, positive n/lambda/delta, and fixed lambda only. The fixed-budget and finite-grid square-root corollaries are separate finite claims; continuous posterior PAC-Bayes and infinite hypothesis classes remain outside this claim.
PAC-Bayes
Lean verified
R(rho) <= Rhat_S(rho) + sqrt(C / (2n)) outside a delta event
Lean declaration
Scope
finite hypotheses, finite PMFs, [0,1] losses, fixed budget C
Boundary
Finite data domain, finite nonempty hypothesis index, finite prior/posterior PMFs, scalar [0,1] losses, positive n/delta, and one fixed positive complexity budget C. The finite-grid theorem removes the single-budget restriction under an explicit grid-cover certificate; continuous posteriors and infinite hypothesis classes remain outside this claim.
PAC-Bayes
Lean verified
finite lambda-grid peeling removes the single-budget restriction
Lean declaration
Scope
finite hypotheses, finite PMFs, [0,1] losses, explicit grid cover
Boundary
Finite data domain, finite nonempty hypothesis index, finite prior/posterior PMFs, scalar [0,1] losses, positive n and grid confidence allocations, and an explicit finite grid-cover certificate. Does not prove continuous posterior PAC-Bayes, infinite hypothesis classes, or all-real-lambda optimization.
Stability
Lean verified
uniform stability beta => E[genGap(A(S), S)] <= beta
Lean declaration
Scope
finite iid product sample weights, scalar loss, one-sided expectation
Boundary
The Lean wrapper proves the finite one-sided expected bound E_S[R(A(S)) - Rhat_S(A(S))] <= beta. The page theorem also discusses an absolute-value expected form and a high-probability Bousquet-Elisseeff concentration bound, which remain outside this Lean entry.
One row per manifest entry, sorted by claim ID. The Lean source repository is currently private; the declaration column shows the file path inside the repo for traceability. A public Lean-only mirror is planned; until then, source paths are not clickable.
| Claim ID | Topic page | Lean declaration | Mode | Axioms | Last checked |
|---|---|---|---|---|---|
claim:algorithmic-stability::uniform-stability-generalization | algorithmic-stability | AlgorithmicStability.expectedFiniteGeneralizationGap_le_uniformStability_finiteProductverification/lean/TheoremPath/LearningTheory/AlgorithmicStability.lean | scoped bridge | propext, Classical.choice, Quot.sound | May 7, 2026 |
claim:asymptotic-statistics::continuous-mapping-theorem | asymptotic-statistics | AsymptoticStatistics.continuousMappingTheoremContinuousverification/lean/TheoremPath/Statistics/AsymptoticStatistics.lean | exact wrapper | core only | April 28, 2026 |
claim:asymptotic-statistics::slutsky-theorem | asymptotic-statistics | AsymptoticStatistics.slutskyTheoremPairverification/lean/TheoremPath/Statistics/AsymptoticStatistics.lean | exact wrapper | core only | May 2, 2026 |
claim:common-inequalities::cauchy-schwarz-inequality | common-inequalities | CommonInequalities.cauchySchwarzRealInnerverification/lean/TheoremPath/LinearAlgebra/CommonInequalities.lean | exact wrapper | core only | April 28, 2026 |
claim:common-inequalities::jensen-inequality | common-inequalities | CommonInequalities.jensenIntegralAverageConvexverification/lean/TheoremPath/LinearAlgebra/CommonInequalities.lean | exact wrapper | core only | April 30, 2026 |
claim:concentration-inequalities::chebyshev-inequality | concentration-inequalities | Concentration.chebyshevInequalityVarianceverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 28, 2026 |
claim:concentration-inequalities::hoeffding-one-sided-finite-sum | concentration-inequalities | Concentration.hoeffdingBoundedFiniteSumTailverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 28, 2026 |
claim:concentration-inequalities::markov-inequality | concentration-inequalities | Concentration.markovInequalityRealIntegrableverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 28, 2026 |
claim:concentration-inequalities::sub-gaussian-sum-tail-bound | concentration-inequalities | Concentration.subGaussianFiniteSumTailBoundverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 28, 2026 |
claim:empirical-risk-minimization::finite-class-hoeffding-approx-erm-excess-risk-tail | empirical-risk-minimization | ERMGeneralization.finiteClass_hoeffding_approxERM_excessRisk_tailverification/lean/TheoremPath/LearningTheory/ERMGeneralization.lean | scoped bridge | core only | May 3, 2026 |
claim:empirical-risk-minimization::finite-class-hoeffding-erm-excess-risk-tail | empirical-risk-minimization | ERMGeneralization.finiteClass_hoeffding_exactERM_excessRisk_tailverification/lean/TheoremPath/LearningTheory/ERMGeneralization.lean | scoped bridge | core only | May 3, 2026 |
claim:empirical-risk-minimization::vc-style-erm-excess-risk-tail | empirical-risk-minimization | VCSampleComplexity.vc_erm_excessRisk_tailverification/lean/TheoremPath/LearningTheory/VCSampleComplexity.lean | scoped bridge | core only | May 5, 2026 |
claim:expectation-variance-covariance-moments::law-of-total-variance | expectation-variance-covariance-moments | Moments.lawOfTotalVarianceverification/lean/TheoremPath/Probability/Moments.lean | exact wrapper | core only | April 30, 2026 |
claim:expectation-variance-covariance-moments::linearity-of-expectation | expectation-variance-covariance-moments | FiniteExpectation.linearityOfExpectationIntegrableverification/lean/TheoremPath/Probability/FiniteExpectation.lean | exact wrapper | core only | April 28, 2026 |
claim:expectation-variance-covariance-moments::variance-of-sum | expectation-variance-covariance-moments | Moments.varianceOfFiniteSumWithCovarianceverification/lean/TheoremPath/Probability/Moments.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-basic-identities | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureBasicIdentitiesverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-complement-rule | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureComplementRuleverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-continuity-from-above | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureContinuityFromAboveverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-continuity-from-below | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureContinuityFromBelowverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-countable-additivity | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureCountableAdditivityverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-countable-union-bound | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureCountableUnionBoundverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-finite-additivity | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureFiniteAdditivityverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:kolmogorov-probability-axioms::probability-measure-finite-union-bound | kolmogorov-probability-axioms | KolmogorovAxioms.probabilityMeasureFiniteUnionBoundverification/lean/TheoremPath/Probability/KolmogorovAxioms.lean | exact wrapper | core only | April 28, 2026 |
claim:law-of-large-numbers::strong-law-of-large-numbers | law-of-large-numbers | LawOfLargeNumbers.strongLawAverageTendstoAlmostSureverification/lean/TheoremPath/Probability/LawOfLargeNumbers.lean | exact wrapper | core only | April 30, 2026 |
claim:law-of-large-numbers::weak-law-of-large-numbers | law-of-large-numbers | LawOfLargeNumbers.weakLawAverageTendstoInMeasureverification/lean/TheoremPath/Probability/LawOfLargeNumbers.lean | exact wrapper | core only | April 30, 2026 |
claim:martingale-theory::azuma-hoeffding | martingale-theory | Concentration.azumaHoeffdingConditionalSubGaussianTailverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 30, 2026 |
claim:martingale-theory::doob-weak-maximal-inequality | martingale-theory | Martingale.doobWeakMaximalInequalityverification/lean/TheoremPath/Probability/Martingale.lean | exact wrapper | core only | April 30, 2026 |
claim:measure-theoretic-probability::borel-cantelli-first | measure-theoretic-probability | BorelCantelli.borelCantelliFirstLimsupMeasureZeroverification/lean/TheoremPath/Probability/BorelCantelli.lean | exact wrapper | core only | April 28, 2026 |
claim:measure-theoretic-probability::borel-cantelli-second | measure-theoretic-probability | BorelCantelli.borelCantelliSecondIndependentLimsupMeasureOneverification/lean/TheoremPath/Probability/BorelCantelli.lean | exact wrapper | core only | April 29, 2026 |
claim:measure-theoretic-probability::dominated-convergence-theorem | measure-theoretic-probability | MeasureConvergence.dominatedConvergenceIntegralverification/lean/TheoremPath/Probability/MeasureConvergence.lean | exact wrapper | core only | April 28, 2026 |
claim:measure-theoretic-probability::fatou-lemma | measure-theoretic-probability | MeasureConvergence.fatouLemmaLIntegralverification/lean/TheoremPath/Probability/MeasureConvergence.lean | exact wrapper | core only | April 28, 2026 |
claim:measure-theoretic-probability::monotone-convergence-theorem | measure-theoretic-probability | MeasureConvergence.monotoneConvergenceLIntegralverification/lean/TheoremPath/Probability/MeasureConvergence.lean | exact wrapper | core only | April 28, 2026 |
claim:pac-bayes-bounds::donsker-varadhan-finite-change-of-measure | pac-bayes-bounds | PACBayesKL.donsker_varadhanverification/lean/TheoremPath/LearningTheory/PACBayesKL.lean | scoped bridge | core only | May 6, 2026 |
claim:pac-bayes-bounds::finite-catoni-bounded-loss-pac-bayes-bound | pac-bayes-bounds | PACBayesBoundedLoss.finiteCatoni_badEventMass_le_deltaverification/lean/TheoremPath/LearningTheory/PACBayesBoundedLoss.lean | scoped bridge | core only | May 7, 2026 |
claim:pac-bayes-bounds::finite-grid-mcallester-optimized-pac-bayes-bound | pac-bayes-bounds | PACBayesBoundedLoss.finiteMcAllesterGridOptimized_badEventMass_le_deltaverification/lean/TheoremPath/LearningTheory/PACBayesBoundedLoss.lean | scoped bridge | core only | May 7, 2026 |
claim:pac-bayes-bounds::finite-mcallester-fixed-budget-pac-bayes-bound | pac-bayes-bounds | PACBayesBoundedLoss.finiteMcAllesterBoundedComplexity_badEventMass_le_deltaverification/lean/TheoremPath/LearningTheory/PACBayesBoundedLoss.lean | scoped bridge | core only | May 7, 2026 |
claim:pac-bayes-bounds::finite-pmf-kl-nonnegativity | pac-bayes-bounds | PACBayesKL.klDiv_nonnegverification/lean/TheoremPath/LearningTheory/PACBayesKL.lean | scoped bridge | core only | May 6, 2026 |
claim:pac-bayes-bounds::posterior-gap-finite-change-of-measure | pac-bayes-bounds | PACBayesKL.posterior_generalization_gap_change_of_measureverification/lean/TheoremPath/LearningTheory/PACBayesKL.lean | scoped bridge | core only | May 6, 2026 |
claim:pac-bayes-bounds::posterior-risk-finite-lambda-rearrangement | pac-bayes-bounds | PACBayesKL.posterior_risk_le_empiricalRisk_plus_of_log_moment_boundverification/lean/TheoremPath/LearningTheory/PACBayesKL.lean | scoped bridge | core only | May 6, 2026 |
claim:pac-bayes-bounds::prior-exp-moment-finite-markov-confidence-adapter | pac-bayes-bounds | PACBayesKL.priorExpMoment_tailMass_le_delta_of_expected_boundverification/lean/TheoremPath/LearningTheory/PACBayesKL.lean | scoped bridge | core only | May 6, 2026 |
claim:rademacher-complexity::finite-sample-contraction | rademacher-complexity | RademacherContraction.empiricalRademacherComplexity_contraction_lipschitzverification/lean/TheoremPath/LearningTheory/RademacherContraction.lean | scoped bridge | core only | May 6, 2026 |
claim:rademacher-complexity::high-probability-rademacher-gengap-bound | rademacher-complexity | HighProbRademacher.genGap_highProb_rademacherverification/lean/TheoremPath/LearningTheory/HighProbRademacher.lean | scoped bridge | core only | May 4, 2026 |
claim:rademacher-complexity::linear-predictor-rademacher-bound | rademacher-complexity | LinearPredictorRademacher.linearPredictor_rademacherverification/lean/TheoremPath/LearningTheory/LinearPredictorRademacher.lean | scoped bridge | core only | May 6, 2026 |
claim:rademacher-complexity::massart-finite-class-bound | rademacher-complexity | VCRademacher.empiricalRademacherComplexity_le_massart_effectiveverification/lean/TheoremPath/LearningTheory/VCRademacher.lean | scoped bridge | core only | May 5, 2026 |
claim:rademacher-complexity::one-lipschitz-empirical-rademacher-contraction | rademacher-complexity | RademacherContraction.contraction_empiricalverification/lean/TheoremPath/LearningTheory/RademacherContraction.lean | scoped bridge | core only | May 6, 2026 |
claim:rademacher-complexity::vc-style-rademacher-bound | rademacher-complexity | VCSampleComplexity.vcRademacher_pointwiseverification/lean/TheoremPath/LearningTheory/VCSampleComplexity.lean | scoped bridge | core only | May 5, 2026 |
claim:subgaussian-random-variables::hoeffding-lemma | subgaussian-random-variables | Concentration.hoeffdingLemmaBoundedCenteredSubgaussianMGFverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 29, 2026 |
claim:subgaussian-random-variables::subgaussian-linear-combination | subgaussian-random-variables | Concentration.subGaussianIndependentFiniteLinearCombinationMGFverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 29, 2026 |
claim:subgaussian-random-variables::subgaussian-mgf-characterization | subgaussian-random-variables | Concentration.subGaussianMGFImpliesRightTailBoundverification/lean/TheoremPath/Probability/Concentration.lean | exact wrapper | core only | April 29, 2026 |
claim:symmetrization-inequality::finite-sample-rademacher-symmetrization | symmetrization-inequality | RademacherSymmetrization.expected_genGap_le_two_expected_empiricalRademacherComplexityverification/lean/TheoremPath/LearningTheory/RademacherSymmetrization.lean | scoped bridge | core only | May 4, 2026 |
claim:symmetrization-inequality::one-sample-finite-class-symmetrization | symmetrization-inequality | Rademacher.oneSampleFiniteClassSymmetrizationverification/lean/TheoremPath/Statistics/Rademacher.lean | scoped bridge | core only | May 3, 2026 |
claim:uniform-convergence::epsilon-representative-sample | uniform-convergence | UniformConvergence.epsilonRepresentativeERMWorksverification/lean/TheoremPath/LearningTheory/UniformConvergence.lean | standalone | core only | April 28, 2026 |
claim:uniform-convergence::finite-class-uniform-convergence | uniform-convergence | UniformConvergence.finiteClassEpsilonRepresentativeHighProbabilityFromOneSidedTailsverification/lean/TheoremPath/LearningTheory/UniformConvergence.lean | scoped bridge | core only | April 30, 2026 |
claim:uniform-convergence::vc-style-uniform-deviation-bound | uniform-convergence | VCSampleComplexity.uniformDeviation_highProb_vcClassverification/lean/TheoremPath/LearningTheory/VCSampleComplexity.lean | scoped bridge | core only | May 5, 2026 |
claim:vc-dimension::binary-vc-zero-one-loss-bridge | vc-dimension | BinaryVCBridge.effectiveClass_zeroOneLoss_card_eq_binaryClassTraceverification/lean/TheoremPath/LearningTheory/BinaryVCBridge.lean | scoped bridge | core only | May 5, 2026 |
claim:vc-dimension::sauer-shelah-lemma | vc-dimension | VCDimension.sauerShelahFiniteSetFamilyverification/lean/TheoremPath/LearningTheory/VCDimension.lean | exact wrapper | core only | April 28, 2026 |
The probability layer behind the finite-class generalization theorem. Every arrow currently has a manifest entry behind it.
Diagram A. Finite-class ERM chain (probability layer). Solid arrows are Lean-verified. No dashed arrows yet at this layer.
The core learning-theory generalization layer. These six nodes are verified end to end (Stages 1, 2, 3 chained by transitivity, plus the high-probability bound via Azuma concentration). The contraction and linear-predictor entries are listed separately in the manifest table and assumption blocks.
Diagram B. Rademacher chain (learning-theory layer). Solid arrows are Lean-verified. Dashed arrows are planned or future.
Binary-classification VC-style ERM sample-complexity bound for 0-1 loss. Composes Sauer-Shelah polynomial bound, effective-class Massart, pointwise VC-Rademacher, high-probability genGap, two-sided uniform deviation, ERM excess-risk tail, and the binary-class VC bridge that closes the user-supplied growth assumption for 0-1 loss.
Diagram C. VC sample-complexity chain. All nodes and arrows are Lean-verified, including the binary-class VC bridge for 0-1 loss.
Scope, assumptions, and current boundaries for the finite-class ERM, Rademacher, VC, and finite chaining tracks. The four states are visually distinct and use no badges.
| Result | Status | Limitation |
|---|---|---|
| Exact ERM oracle inequality | verified | deterministic |
| Approximate ERM oracle inequality | verified | deterministic |
| Finite-class Hoeffding ERM bound | verified | finite class, bounded loss |
| Finite sign-vector Rademacher machinery | verified | combinatorial |
| Rademacher probability bridge | verified | expectation form only |
| Expected Rademacher symmetrization | verified | expectation form, finite index, bounded loss |
| High-probability Rademacher bound | verified | one-sided genGap, Azuma constant, finite class, bounded loss |
| Finite-sample scalar contraction | verified | finite sample, finite class, scalar-valued |
| Linear predictor Rademacher bound | verified | finite-dimensional Euclidean, finite indexed weights |
| Effective-class Massart bound | verified | finite class, bounded loss, effective class card > 1 |
| VC pointwise Rademacher bound | verified | user-supplied effective-growth assumption, finite class, bounded loss |
| VC high-probability genGap | verified | one-sided, Azuma constant, user-supplied growth, finite class |
| VC two-sided uniform deviation | verified | Azuma constant, user-supplied growth for ℓ and -ℓ, finite class |
| VC ERM excess-risk tail | verified | Azuma constant, user-supplied growth, finite class, bounded loss |
| Binary-class VC bridge (trace → effective class, 0-1 loss) | verified | binary classifiers only, 0-1 loss only, finite class |
| Experimental finite chaining foundation | verified | finite support layer only; not a headline governed claim |
| Generic PAC learnability | extension target | requires a separate PAC-learnability framework |
Each block records the informal statement, the literal Lean assumptions, and the verification boundary. The Lean assumptions are reproductions of the declaration signature, not paraphrase.
Statement (informal)
For a finite nonempty hypothesis class with iid bounded loss in [a, b], an exact ERM hypothesis, and a population-risk oracle, the excess-risk-tail probability is bounded by delta once the sample size exceeds (b - a)^2 log(2|H|/delta) / (2 epsilon^2).
Lean assumptions
[Fintype H] [Nonempty H] (hLossBounded : ∀ h x y, a ≤ ℓ h x y ∧ ℓ h x y ≤ b) (hOracle : IsERM oracle riskTrue) (hHat : IsERM hhat riskEmp) (hN : (b - a)^2 * Real.log (2 * |H| / delta) / (2 * epsilon^2) ≤ n)Boundary
Statement (informal)
Same setup as the exact-ERM theorem, but the empirical minimizer is replaced by a hypothesis whose empirical risk is within an explicit optimization slack of the ERM.
Lean assumptions
[Fintype H] [Nonempty H] (hLossBounded : ...) (hApproxERM : riskEmp hhat ≤ riskEmp (argmin riskEmp) + optError) (hN : ...)Boundary
Statement (informal)
On a finite nonempty index set, the empirical Rademacher complexity expressed as an integral against the uniform sign-vector PMF equals the discrete average over sign vectors of the supremum.
Lean assumptions
[Fintype ι] [Nonempty ι] (f : ι → SignVec n → ℝ) (hf : ∀ σ, BddAbove (Set.range (fun i => f i σ)))Boundary
Statement (informal)
For an iid sample S of size n drawn from a probability measure mu on Z, a finite nonempty index set iota of measurable losses ell_i bounded by B in absolute value, the expected supremum of (population risk minus empirical risk) is at most twice the expected empirical Rademacher complexity. Proved by transitivity through ghost-sample replacement (Stage 1) and Rademacher decoupling (Stage 2), with Fubini converting an iterated integral into a joint integral against mu^{⊗n} ⊗ mu^{⊗n}.
Lean assumptions
{ι : Type*} [Fintype ι] [Nonempty ι] {Z : Type*} [MeasurableSpace Z] {n : ℕ} (μ : Measure Z) [IsProbabilityMeasure μ] (ℓ : ι → Z → ℝ) {B : ℝ} (hB : 0 ≤ B) (hℓ_meas : ∀ i, Measurable (ℓ i)) (hℓ_bdd : ∀ i z, |ℓ i z| ≤ B) (hn : 0 < n)Boundary
Statement (informal)
For a finite nonempty hypothesis class with uniformly B-bounded measurable losses on an iid sample S ~ μⁿ: P(genGap(S) ≥ 2·E_S[empiricalRademacherComplexity] + ε) ≤ exp(-ε²n/(8B²)). Combines expected Rademacher symmetrization with the Azuma-style genGap tail bound via measure monotonicity.
Lean assumptions
{ι : Type*} [Fintype ι] [Nonempty ι] [Nonempty Z] [StandardBorelSpace Z] [IsProbabilityMeasure μ] {ℓ : ι → Z → ℝ} {B : ℝ} (hB : 0 < B) (hℓ_meas : ∀ i, Measurable (ℓ i)) (hℓ_bdd : ∀ i z, |ℓ i z| ≤ B) {n : ℕ} (hn : 0 < n) {ε : ℝ} (hε : 0 ≤ ε)Boundary
Statement (informal)
For a finite sample, finite nonempty hypothesis class, scalar-valued functions, positive Lipschitz constant L, and φ satisfying |φ s - φ t| ≤ L |s - t|: empiricalRademacherComplexity(φ ∘ ℓ, z) ≤ L · empiricalRademacherComplexity(ℓ, z).
Lean assumptions
{ι : Type*} [Fintype ι] [Nonempty ι] {Z : Type*} {n : ℕ} {φ : ℝ → ℝ} {L : ℝ} (hL : 0 < L) (hφ_lip : ∀ s t : ℝ, |φ s - φ t| ≤ L * |s - t|) (ℓ : ι → Z → ℝ) (z : Fin n → Z)Boundary
Statement (informal)
For finite weights w_i in EuclideanSpace ℝ (Fin d) with ‖w_i‖ ≤ R and finite sample z : Fin n → EuclideanSpace ℝ (Fin d), empiricalRademacherComplexity({x ↦ ⟪w_i, x⟫}, z) ≤ R · n⁻¹ · sqrt(Σ_k ‖z_k‖²). A bounded-input corollary proves the RB/sqrt(n) form.
Lean assumptions
{d n : ℕ} {ι : Type*} [Fintype ι] [Nonempty ι] (w : ι → EuclideanSpace ℝ (Fin d)) (z : Fin n → EuclideanSpace ℝ (Fin d)) (R : ℝ) (hR : 0 ≤ R) (hw : ∀ i, ‖w i‖ ≤ R) (hn : 0 < (n : ℝ))Boundary
Statement (informal)
For a hypothesis class with effective-class cardinality bounded by the growth function ∑_{k≤d} C(n,k) (for ℓ and -ℓ), B-bounded loss, and an iid sample S ~ μⁿ: P(risk(ERM) - risk(best) ≥ 4B·√(2d·log(en/d)/n) + 2ε) ≤ 2·exp(-ε²n/(8B²)). Chains Sauer-Shelah → Massart → VC Rademacher → Azuma → uniform deviation → ERM.
Lean assumptions
{ι : Type*} [Fintype ι] [Nonempty ι] [MeasurableSpace Z] [Nonempty Z] [StandardBorelSpace Z] {μ : Measure Z} [IsProbabilityMeasure μ] {ℓ : ι → Z → ℝ} {B : ℝ} (hB : 0 < B) (hℓ_meas : ∀ i, Measurable (ℓ i)) (hℓ_bdd : ∀ i z, |ℓ i z| ≤ B) {n d : ℕ} (hn : 0 < n) (hd : 0 < d) (hdn : d ≤ n) (hGrowth_uniform : ∀ z, (effectiveClass ℓ z).card ≤ ∑ k ∈ range (d+1), n.choose k) (hGrowth_neg : ∀ z, (effectiveClass (fun i w => -ℓ i w) z).card ≤ ...) (hhat : (Fin n → Z) → ι) (hERM : ∀ S, IsERM (empiricalRisk S ℓ) (hhat S)) (i_star : ι) {ε : ℝ} (hε : 0 ≤ ε)Boundary
These entries keep broader theorem ambitions separate from the narrower declarations already compiled by Lean. Boundaries are tracked as next-proof targets rather than inferred from adjacent results.
Sharp McDiarmid / two-sided Rademacher generalization bound
The one-sided high-probability Rademacher generalization bound is verified for finite classes with the Azuma constant. Current boundaries: the two-sided uniformDeviation form, the sharp McDiarmid constant exp(-ε²n/(2B²)), infinite hypothesis classes, and the full textbook confidence-delta parametrization are separate targets.
Dudley's entropy integral
The continuous entropy integral is not a governed manifest claim. The Lean tree has a verified finite chaining support layer, including finite dyadic entropy-budget and uniform entropy-cap bounds, but TheoremPath does not present it as a headline continuous Dudley theorem yet.
Full VC-to-PAC equivalence
The effective-growth VC-style sample-complexity theorem is verified (Sauer-Shelah → Massart → Rademacher → ERM). The binary-class VC bridge for 0-1 loss is closed (effectiveClass.card = binaryClassTrace.card). Current boundaries: generic real-valued-loss VC bridges, the full fundamental theorem of PAC learning, and infinite hypothesis classes remain future work.
Generic PAC learnability
TheoremPath verifies specific finite-class theorem chains today. A generic PAC-learnability framework is a larger formalization lane, not a consequence of a single checked declaration.
Page-level verification
Manifest verification certifies exact theorem statements. Topic prose, examples, diagrams, and citations stay governed by source review unless a page explicitly links to a matching Lean claim.
Non-core axioms
Every checked declaration uses only the mathlib core axiom set: propext, Classical.choice, Quot.sound. No third-party axioms.
Cross-references: evidence, methodology, and disclaimer.