Stat.Concentration.Rademacher.Contraction
The contraction principle: composing every function of a class with an L-Lipschitz map φ multiplies its empirical Rademacher complexity by at most L in the signed form (no absolute value inside the supremum), and by at m
Signed 3 core · 4 supporting This file proves the finite-index Ledoux--Talagrand contraction principle for signed empirical Rademacher averages. ★ rademacher_contraction
Signed Rademacher contraction
This file proves the finite-index Ledoux--Talagrand contraction principle for
signed empirical Rademacher averages. It defines LipschitzAt0, develops the
hybrid sign-swap argument, and exports rademacher_contraction together with
the signed scaling law. Absolute-value and arbitrary-index consequences are
provided by Contraction/Absolute.lean.
For a real-valued transformation and a real constant, the transformation is Lipschitz at zero with constant exactly when (1) it maps zero to zero and (2) for every two real numbers and , its increment has absolute value at most .
Definition (Lean source)
The sign atom is the two-element type containing the signs minus one and one. It provides the individual Rademacher signs used in the signed contraction argument.
If the proposed Lipschitz constant is nonnegative and the scalar transform is Lipschitz with that constant, then the empirical Rademacher complexity without absolute values of the transformed class is at most that constant times the original complexity.
Formal statement
Proof (Lean source)
4 supporting declarations (lemmas, instances)
-
finite_iSup_eq_valuelemma — A real-valued function on a nonempty finite population attains a largest value, and its supremum is that value.hypothesesα :Type*Nonempty αFinite αf :α → ℝconclusion∃ a : α,(⨆ x, f x) = f aconclusion 1x :f x ≤ f aProof (Lean source)
lemma finite_iSup_eq_value {α : Type*} [Nonempty α] [Finite α] (f : α → ℝ) : ∃ a : α, (⨆ x, f x) = f a ∧ ∀ x, f x ≤ f a := by classical rcases Finite.exists_max f with ⟨a, ha⟩ refine ⟨a, ?_, ha⟩ exact le_antisymm (ciSup_le ha) (le_ciSup (Finite.bddAbove_range f) a) -
sup_pair_lipschitz_scaledlemma — For a finite nonempty collection, the sum of the largest values obtained by adding and subtracting a nonnegative multiple of a transformation with a given Lipschitz constant is no greater than the corresponding sum using that linear bound.hypothesesι :Type*Nonempty ιFinite ιφ :ℝ → ℝL c :ℝhc :0 ≤ chφ :∀ x y, |φ x - φ y| ≤ L * |x - y|a :ι → ℝb :ι → ℝconclusion(⨆ i, a i + c * φ (b i)) + (⨆ i, a i - c * φ (b i))≤ (⨆ i, a i + c * (L * b i)) + (⨆ i, a i - c * (L * b i))Proof (Lean source)
lemma sup_pair_lipschitz_scaled {ι : Type*} [Nonempty ι] [Finite ι] (φ : ℝ → ℝ) {L c : ℝ} (hc : 0 ≤ c) (hφ : ∀ x y, |φ x - φ y| ≤ L * |x - y|) (a : ι → ℝ) (b : ι → ℝ) : (⨆ i, a i + c * φ (b i)) + (⨆ i, a i - c * φ (b i)) ≤ (⨆ i, a i + c * (L * b i)) + (⨆ i, a i - c * (L * b i)) := by classical rcases finite_iSup_eq_value (fun i : ι => a i + c * φ (b i)) with ⟨i₁, hi₁eq, _hi₁max⟩ rcases finite_iSup_eq_value (fun i : ι => a i - c * φ (b i)) with ⟨i₂, hi₂eq, _hi₂max⟩ rw [hi₁eq, hi₂eq] have hdiff : φ (b i₁) - φ (b i₂) ≤ L * |b i₁ - b i₂| := by exact (le_abs_self _).trans (hφ (b i₁) (b i₂)) have hcdiff : c * (φ (b i₁) - φ (b i₂)) ≤ c * (L * |b i₁ - b i₂|) := by exact mul_le_mul_of_nonneg_left hdiff hc by_cases hcase : b i₂ ≤ b i₁ · have habs : |b i₁ - b i₂| = b i₁ - b i₂ := abs_of_nonneg (sub_nonneg.mpr hcase) have hmain : a i₁ + c * φ (b i₁) + (a i₂ - c * φ (b i₂)) ≤ (a i₁ + c * (L * b i₁)) + (a i₂ - c * (L * b i₂)) := by rw [habs] at hcdiff linarith refine hmain.trans ?_ exact add_le_add (le_ciSup (Finite.bddAbove_range (fun i : ι => a i + c * (L * b i))) i₁) (le_ciSup (Finite.bddAbove_range (fun i : ι => a i - c * (L * b i))) i₂) · have hcase' : b i₁ ≤ b i₂ := le_of_not_ge hcase have habs : |b i₁ - b i₂| = b i₂ - b i₁ := by rw [abs_sub_comm] exact abs_of_nonneg (sub_nonneg.mpr hcase') have hmain : a i₁ + c * φ (b i₁) + (a i₂ - c * φ (b i₂)) ≤ (a i₂ + c * (L * b i₂)) + (a i₁ - c * (L * b i₁)) := by rw [habs] at hcdiff linarith refine hmain.trans ?_ have hp := le_ciSup (Finite.bddAbove_range (fun i : ι => a i + c * (L * b i))) i₂ have hm := le_ciSup (Finite.bddAbove_range (fun i : ι => a i - c * (L * b i))) i₁ linarith -
signAtom_coe_real_eq_neg_one_or_onelemma — Every sign atom, viewed as a real number, is either minus one or one.Proof (Lean source)
lemma signAtom_coe_real_eq_neg_one_or_one (s : SignAtom) : (s : ℝ) = -1 ∨ (s : ℝ) = 1 := by rcases s with ⟨z, hz⟩ simp only [Finset.mem_insert, Int.reduceNeg, Finset.mem_singleton, Int.cast_eq_one] at hz ⊢ rcases hz with rfl | rfl · left norm_num · right norm_num -
empiricalRademacherComplexity_without_abs_smul_classtheorem — Without-abs analogue of empiricalRademacherComplexity_smul_class: the signed scaling carries c, not |c|.hypothesesconclusionempiricalRademacherComplexity_without_abs n (fun i x => c * F i x) S= c * empiricalRademacherComplexity_without_abs n F SProof (Lean source)
theorem empiricalRademacherComplexity_without_abs_smul_class (F : ι → 𝒳 → ℝ) (c : ℝ) (hc : 0 ≤ c) (n : ℕ) (S : Fin n → 𝒳) : empiricalRademacherComplexity_without_abs n (fun i x => c * F i x) S = c * empiricalRademacherComplexity_without_abs n F S := by unfold empiricalRademacherComplexity_without_abs have hsum : (∑ σ : Signs n, ⨆ i, (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (c * F i (S k))) = c * ∑ σ : Signs n, ⨆ i, (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k) := by rw [Finset.mul_sum] refine Finset.sum_congr rfl fun σ _ => ?_ calc (⨆ i, (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (c * F i (S k))) = ⨆ i, c * ((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) := by refine iSup_congr fun i => ?_ calc (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (c * F i (S k)) = (n : ℝ)⁻¹ * (c * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) := by congr 1 rw [Finset.mul_sum] refine Finset.sum_congr rfl fun k _ => by ring _ = c * ((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) := by ring _ = c * ⨆ i, (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k) := by exact (Real.mul_iSup_of_nonneg hc (fun i => (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k))).symm rw [hsum] ring
Absolute 1 core · 8 supporting Starting from the signed finite-index contraction theorem, this file derives the standard factor-two absolute-value contraction bound. ★ rademacher_contraction_abs
Absolute Rademacher contraction
Starting from the signed finite-index contraction theorem, this file derives the standard factor-two absolute-value contraction bound. It also supplies scaling and subtraction rules for empirical Rademacher complexity and the arbitrary-index contraction theorem under boundedness of the relevant suprema.
Contraction principle, absolute-value form. If φ fixes 0 and is L-Lipschitz, i.e. φ 0 = 0 and |φ x - φ y| ≤ L * |x - y| for all x, y, then composing each function of the family F with φ multiplies the (absolute-value) empirical Rademacher complexity on the sample S by at most 2 * L.
Formal statement
Proof (Lean source)
8 supporting declarations (lemmas, instances)
-
lipschitzAt0_neglemma — Negating a function that fixes zero preserves the same Lipschitz constant.Proof (Lean source)
lemma lipschitzAt0_neg (φ : ℝ → ℝ) {L : ℝ} (hφ : LipschitzAt0 φ L) : LipschitzAt0 (fun x => -φ x) L := by refine ⟨by simp [hφ.1], ?_⟩ intro x y rw [show -φ x - -φ y = -(φ x - φ y) from by ring, abs_neg] exact hφ.2 x y -
iSup_abs_le_iSup_add_iSup_neg_of_exists_zerolemma — If a finite collection of real numbers includes zero, its largest absolute value is at most the sum of its largest value and the largest value after negation.hypothesesα :Type*Finite αx :α → ℝh0 :∃ a : α, x a = 0conclusion(⨆ a, |x a|) ≤ (⨆ a, x a) + (⨆ a, -x a)Proof (Lean source)
lemma iSup_abs_le_iSup_add_iSup_neg_of_exists_zero {α : Type*} [Finite α] (x : α → ℝ) (h0 : ∃ a : α, x a = 0) : (⨆ a, |x a|) ≤ (⨆ a, x a) + (⨆ a, -x a) := by classical letI := ofFinite α rcases h0 with ⟨a0, ha0⟩ letI : Nonempty α := ⟨a0⟩ have hsup_nonneg : 0 ≤ ⨆ a, x a := by rw [← ha0] exact le_ciSup (Finite.bddAbove_range x) a0 have hsup_neg_nonneg : 0 ≤ ⨆ a, -x a := by rw [← show -x a0 = 0 by rw [ha0]; simp] exact le_ciSup (Finite.bddAbove_range fun a => -x a) a0 refine ciSup_le ?_ intro a by_cases hx : 0 ≤ x a · rw [abs_of_nonneg hx] have hxle : x a ≤ ⨆ a, x a := le_ciSup (Finite.bddAbove_range x) a linarith · have hxle : -x a ≤ ⨆ a, -x a := le_ciSup (Finite.bddAbove_range fun a => -x a) a rw [abs_of_neg (lt_of_not_ge hx)] linarith -
empirical_abs_withZero_le_no_abs_plus_neglemma — If a function class contains the zero function, its empirical Rademacher complexity with absolute values is bounded by the sum of the corresponding unsigned complexities for the class and its negation.hypothesesι :sharedType u_1𝒳 :sharedType u_2Finite ιF :Option ι → 𝒳 → ℝhzero :∀ x, F none x = 0n :ℕS :Fin n → 𝒳conclusionempiricalRademacherComplexity n F S≤ empiricalRademacherComplexity_without_abs n F S+ empiricalRademacherComplexity_without_abs n (fun i x => -F i x) SProof (Lean source)
lemma empirical_abs_withZero_le_no_abs_plus_neg [Finite ι] (F : Option ι → 𝒳 → ℝ) (hzero : ∀ x, F none x = 0) (n : ℕ) (S : Fin n → 𝒳) : empiricalRademacherComplexity n F S ≤ empiricalRademacherComplexity_without_abs n F S + empiricalRademacherComplexity_without_abs n (fun i x => -F i x) S := by classical letI := ofFinite ι unfold empiricalRademacherComplexity empiricalRademacherComplexity_without_abs rw [← mul_add] refine mul_le_mul_of_nonneg_left ?_ (by positivity) rw [← Finset.sum_add_distrib] refine Finset.sum_le_sum ?_ intro σ _ have hzero_inner : (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F none (S k) = 0 := by simp [hzero] have hpoint := iSup_abs_le_iSup_add_iSup_neg_of_exists_zero (fun i : Option ι => (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) ⟨none, hzero_inner⟩ have hneg : (⨆ i : Option ι, -((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k))) = (⨆ i : Option ι, (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (-F i (S k))) := by refine iSup_congr fun i => ?_ calc -((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) = (n : ℝ)⁻¹ * (-(∑ k : Fin n, (σ k : ℝ) * F i (S k))) := by ring _ = (n : ℝ)⁻¹ * (∑ k : Fin n, -((σ k : ℝ) * F i (S k))) := by rw [Finset.sum_neg_distrib] _ = (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (-F i (S k)) := by congr 1 refine Finset.sum_congr rfl fun k _ => by ring simpa [hneg] using hpoint -
empiricalRademacherComplexity_smul_classtheorem — Scalar-multiplication law for empirical Rademacher complexity. Scaling each element of a function class by c scales the absolute-value empirical Rademacher complexity by |c|.hypothesesconclusionempiricalRademacherComplexity n (fun i x => c * F i x) S= |c| * empiricalRademacherComplexity n F SProof (Lean source)
theorem empiricalRademacherComplexity_smul_class (F : ι → 𝒳 → ℝ) (c : ℝ) (n : ℕ) (S : Fin n → 𝒳) : empiricalRademacherComplexity n (fun i x => c * F i x) S = |c| * empiricalRademacherComplexity n F S := by unfold empiricalRademacherComplexity have hsum : (∑ σ : Signs n, ⨆ i, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (c * F i (S k))|) = |c| * ∑ σ : Signs n, ⨆ i, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)| := by rw [Finset.mul_sum] refine Finset.sum_congr rfl fun σ _ => ?_ calc (⨆ i, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (c * F i (S k))|) = ⨆ i, |c| * |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)| := by refine iSup_congr fun i => ?_ have hlin : (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (c * F i (S k)) = c * ((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) := by calc (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (c * F i (S k)) = (n : ℝ)⁻¹ * (c * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) := by congr 1 rw [Finset.mul_sum] refine Finset.sum_congr rfl fun k _ => by ring _ = c * ((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) := by ring rw [hlin, abs_mul] _ = |c| * ⨆ i, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)| := by exact (Real.mul_iSup_of_nonneg (abs_nonneg c) (fun i => |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)|)).symm rw [hsum] ring -
absInner_le_of_boundlemma — A signed empirical average of a uniformly bounded function has absolute value no larger than the same uniform bound.hypothesesι :sharedType u_1𝒳 :sharedType u_2H :ι → 𝒳 → ℝM :ℝhM0 :0 ≤ MhH :∀ i x, |H i x| ≤ Mn :ℕS :Fin n → 𝒳σ :Signs ni :ιconclusion|(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * H i (S k)| ≤ MProof (Lean source)
lemma absInner_le_of_bound (H : ι → 𝒳 → ℝ) {M : ℝ} (hM0 : 0 ≤ M) (hH : ∀ i x, |H i x| ≤ M) (n : ℕ) (S : Fin n → 𝒳) (σ : Signs n) (i : ι) : |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * H i (S k)| ≤ M := by by_cases hn : n = 0 · subst hn; simpa using hM0 · have hn_pos : 0 < (n : ℝ) := by positivity rw [abs_mul, abs_of_nonneg (by positivity : (0:ℝ) ≤ (n:ℝ)⁻¹)] calc (n : ℝ)⁻¹ * |∑ k : Fin n, (σ k : ℝ) * H i (S k)| ≤ (n : ℝ)⁻¹ * ∑ k : Fin n, M := by refine mul_le_mul_of_nonneg_left ?_ (by positivity) refine (Finset.abs_sum_le_sum_abs _ _).trans (Finset.sum_le_sum ?_) intro k _ rw [abs_mul] rcases signAtom_coe_real_eq_neg_one_or_one (σ k) with h | h · rw [h]; simpa using hH i (S k) · rw [h]; simpa using hH i (S k) _ = M := by rw [Finset.sum_const, Finset.card_univ, Fintype.card_fin, nsmul_eq_mul] field_simp -
absInner_bddAbovelemma — Signed empirical averages of a uniformly bounded function class are bounded above by the common absolute bound.hypothesesι :sharedType u_1𝒳 :sharedType u_2H :ι → 𝒳 → ℝM :ℝhM0 :0 ≤ MhH :∀ i x, |H i x| ≤ Mn :ℕS :Fin n → 𝒳σ :Signs nProof (Lean source)
lemma absInner_bddAbove (H : ι → 𝒳 → ℝ) {M : ℝ} (hM0 : 0 ≤ M) (hH : ∀ i x, |H i x| ≤ M) (n : ℕ) (S : Fin n → 𝒳) (σ : Signs n) : BddAbove (range (fun i => |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * H i (S k)|)) := by refine ⟨M, ?_⟩ rintro _ ⟨i, rfl⟩ exact absInner_le_of_bound H hM0 hH n S σ i -
empiricalRademacherComplexity_sub_letheorem — Empirical Rademacher complexity is sub-additive over differences of classes. For classes F, G sharing the index ι, each with a uniform bound, the complexity of fun i x => F i x - G i x is at most the sum of the individual complexities.hypothesesι :sharedType u_1𝒳 :sharedType u_2Nonempty ιF G :ι → 𝒳 → ℝMF MG :ℝhMF0 :0 ≤ MFhMG0 :0 ≤ MGhF :∀ i x, |F i x| ≤ MFhG :∀ i x, |G i x| ≤ MGn :ℕS :Fin n → 𝒳conclusionempiricalRademacherComplexity n (fun i x => F i x - G i x) S≤ empiricalRademacherComplexity n F S + empiricalRademacherComplexity n G SProof (Lean source)
theorem empiricalRademacherComplexity_sub_le [Nonempty ι] (F G : ι → 𝒳 → ℝ) {MF MG : ℝ} (hMF0 : 0 ≤ MF) (hMG0 : 0 ≤ MG) (hF : ∀ i x, |F i x| ≤ MF) (hG : ∀ i x, |G i x| ≤ MG) (n : ℕ) (S : Fin n → 𝒳) : empiricalRademacherComplexity n (fun i x => F i x - G i x) S ≤ empiricalRademacherComplexity n F S + empiricalRademacherComplexity n G S := by classical unfold empiricalRademacherComplexity rw [← mul_add, ← Finset.sum_add_distrib] refine mul_le_mul_of_nonneg_left (Finset.sum_le_sum ?_) (by positivity) intro σ _ have hbddF := absInner_bddAbove F hMF0 hF n S σ have hbddG := absInner_bddAbove G hMG0 hG n S σ refine ciSup_le (fun i => ?_) have hsplit : (n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * (F i (S k) - G i (S k)) = ((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * F i (S k)) - ((n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * G i (S k)) := by rw [← mul_sub, ← Finset.sum_sub_distrib] refine congrArg _ (Finset.sum_congr rfl fun k _ => by ring) rw [hsplit] refine (abs_sub _ _).trans ?_ exact add_le_add (le_ciSup hbddF i) (le_ciSup hbddG i) -
empiricalRademacherComplexity_contraction_abs_of_bddAbovetheorem — Ledoux–Talagrand contraction over an arbitrary (possibly infinite) index. For an L-Lipschitz φ with φ 0 = 0 and a class F with a uniform bound M, R̂_n(φ ∘ F) ≤ 2L · R̂_n(F) for *any* nonempty index ι. The infinite-index case is reduced to the finite-index rademacher_contraction_abs by choosing, for each of the finitely many sign vectors, an ε-approximate maximizer; their finite collection is a finite subindex on which the Fintype contraction applies.hypothesesι :sharedType u_1𝒳 :sharedType u_2Nonempty ιφ :ℝ → ℝL :ℝhφ :LipschitzAt0 φ LF :ι → 𝒳 → ℝM :ℝhM0 :0 ≤ MhM :∀ i x, |F i x| ≤ Mn :ℕS :Fin n → 𝒳conclusionempiricalRademacherComplexity n (fun i x => φ (F i x)) S≤ 2 * L * empiricalRademacherComplexity n F SProof (Lean source)
theorem empiricalRademacherComplexity_contraction_abs_of_bddAbove [Nonempty ι] (φ : ℝ → ℝ) {L : ℝ} (hφ : LipschitzAt0 φ L) (F : ι → 𝒳 → ℝ) {M : ℝ} (hM0 : 0 ≤ M) (hM : ∀ i x, |F i x| ≤ M) (n : ℕ) (S : Fin n → 𝒳) : empiricalRademacherComplexity n (fun i x => φ (F i x)) S ≤ 2 * L * empiricalRademacherComplexity n F S := by classical have hL : 0 ≤ L := by have h := hφ.2 0 1 norm_num [hφ.1] at h exact (abs_nonneg _).trans h haveI hSigns : Nonempty (Signs n) := ⟨fun _ => ⟨1, by decide⟩⟩ -- uniform bound for the composed class `φ ∘ F` have hLM0 : 0 ≤ L * M := mul_nonneg hL hM0 have hφM : ∀ i x, |φ (F i x)| ≤ L * M := by intro i x have h := hφ.2 (F i x) 0 rw [hφ.1, sub_zero, sub_zero] at h exact h.trans (mul_le_mul_of_nonneg_left (hM i x) hL) have hcard : (0 : ℝ) < (card (Signs n) : ℝ) := by exact_mod_cast Fintype.card_pos (α := Signs n) refine le_of_forall_pos_le_add (fun ε hε => ?_) -- per-sign ε-approximate maximizers (explicit supremand form) have hex : ∀ σ : Signs n, ∃ i, (⨆ j, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F j (S k))|) - ε < |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F i (S k))| := by intro σ exact exists_lt_of_lt_ciSup (sub_lt_self _ hε) choose iσ hiσ using hex -- the finite subindex collecting all per-sign maximizers set T : Finset ι := Finset.image iσ Finset.univ with hT haveI : Nonempty {x // x ∈ T} := ⟨⟨iσ (Classical.arbitrary (Signs n)), Finset.mem_image.mpr ⟨_, Finset.mem_univ _, rfl⟩⟩⟩ -- Fintype contraction on the subindex `↥T` have hcontr := rademacher_contraction_abs (ι := {x // x ∈ T}) φ hφ (fun j => F j.val) n S -- (b) restricting the index only lowers the (with-abs) complexity have hb : empiricalRademacherComplexity n (fun j : {x // x ∈ T} => F j.val) S ≤ empiricalRademacherComplexity n F S := by unfold empiricalRademacherComplexity refine mul_le_mul_of_nonneg_left (Finset.sum_le_sum ?_) (by positivity) intro σ _ exact ciSup_le (fun j => le_ciSup (absInner_bddAbove F hM0 hM n S σ) j.val) -- (a) the ε-maximizers recover the full supremum up to ε have ha : empiricalRademacherComplexity n (fun i x => φ (F i x)) S ≤ empiricalRademacherComplexity n (fun (j : {x // x ∈ T}) x => φ (F j.val x)) S + ε := by unfold empiricalRademacherComplexity have hpt : ∀ σ : Signs n, (⨆ i, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F i (S k))|) ≤ (⨆ j : {x // x ∈ T}, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F j.val (S k))|) + ε := by intro σ have hmem : iσ σ ∈ T := Finset.mem_image.mpr ⟨σ, Finset.mem_univ _, rfl⟩ have hbddT : BddAbove (range (fun j : {x // x ∈ T} => |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F j.val (S k))|)) := Finite.bddAbove_range _ have hle := le_ciSup hbddT (⟨iσ σ, hmem⟩ : {x // x ∈ T}) have := hiσ σ linarith calc (card (Signs n) : ℝ)⁻¹ * ∑ σ : Signs n, ⨆ i, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F i (S k))| ≤ (card (Signs n) : ℝ)⁻¹ * ∑ σ : Signs n, ((⨆ j : {x // x ∈ T}, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F j.val (S k))|) + ε) := by refine mul_le_mul_of_nonneg_left (Finset.sum_le_sum fun σ _ => hpt σ) (by positivity) _ = (card (Signs n) : ℝ)⁻¹ * ∑ σ : Signs n, (⨆ j : {x // x ∈ T}, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F j.val (S k))|) + (card (Signs n) : ℝ)⁻¹ * (card (Signs n) : ℝ) * ε := by rw [Finset.sum_add_distrib, Finset.sum_const, Finset.card_univ, nsmul_eq_mul] ring _ = (card (Signs n) : ℝ)⁻¹ * ∑ σ : Signs n, (⨆ j : {x // x ∈ T}, |(n : ℝ)⁻¹ * ∑ k : Fin n, (σ k : ℝ) * φ (F j.val (S k))|) + ε := by rw [inv_mul_cancel₀ (ne_of_gt hcard)]; ring calc empiricalRademacherComplexity n (fun i x => φ (F i x)) S ≤ empiricalRademacherComplexity n (fun (j : {x // x ∈ T}) x => φ (F j.val x)) S + ε := ha _ ≤ 2 * L * empiricalRademacherComplexity n (fun j : {x // x ∈ T} => F j.val) S + ε := by linarith [hcontr] _ ≤ 2 * L * empiricalRademacherComplexity n F S + ε := by nlinarith [mul_le_mul_of_nonneg_left hb (by linarith : (0:ℝ) ≤ 2 * L)]
Coordinatewise 4 core · 0 supporting This module gives the Ledoux--Talagrand contraction principle when the scalar contraction may depend on the sample coordinate. ★ rademacher_contraction_coordinatewise★ rademacher_contraction_abs_coordinatewise
Coordinatewise Rademacher contraction
This module gives the Ledoux--Talagrand contraction principle when the scalar contraction may
depend on the sample coordinate. Thus the transformed signed sum contains
σ k * φ k (F i (S k)), with a common Lipschitz constant but a different map φ k at every
coordinate. The signed theorem has constant L; the absolute-value form has constant 2L.
Given a sample size, coordinatewise scalar maps, a function class, and a realized sample, the coordinatewise transformed signed Rademacher average averages the supremum of the signed transformed sums over all sign vectors.
Given a sample size, coordinatewise scalar maps, a function class, and a realized sample, the coordinatewise transformed absolute Rademacher average averages the supremum of the absolute signed transformed sums.
Coordinatewise Ledoux--Talagrand contraction, signed form. If the common constant is nonnegative and every coordinate map φ k is L-Lipschitz, then the coordinatewise transformed signed Rademacher average of the class F on S is at most L times its ordinary signed empirical Rademacher complexity.
Formal statement
Proof (Lean source)
Coordinatewise Ledoux--Talagrand contraction, absolute-value form. If every coordinate map fixes zero and is L-Lipschitz, then the coordinatewise transformed absolute Rademacher average of F on S is at most 2L times the ordinary absolute empirical Rademacher complexity.