Marginalized jensen forking bound
OracleComp.EvalDist.marginalized_jensen_forking_bound
Project documentation
Marginalized Jensen / Cauchy-Schwarz step for the forking lemma. If a per-element bound acc x · (acc x / q − hinv) ≤ B x holds for every x (with acc x ≤ 1), and we marginalize over the output distribution of any mx : m X with [MonadLiftT m SPMF], then the marginalized expectation μ := ∑' x, Pr[= x | mx] · acc x satisfies the same forking-b...
Exact Lean statement
lemma marginalized_jensen_forking_bound
{X : Type} (mx : m X)
(acc B : X → ℝ≥0∞) (q hinv : ℝ≥0∞)
(hinv_ne_top : hinv ≠ ⊤)
(hacc_le : ∀ x, acc x ≤ 1)
(hper : ∀ x, acc x * (acc x / q - hinv) ≤ B x) :
(∑' x, Pr[= x | mx] * acc x) *
((∑' x, Pr[= x | mx] * acc x) / q - hinv) ≤
∑' x, Pr[= x | mx] * B xFormal artifact
Lean source
lemma marginalized_jensen_forking_bound {X : Type} (mx : m X) (acc B : X → ℝ≥0∞) (q hinv : ℝ≥0∞) (hinv_ne_top : hinv ≠ ⊤) (hacc_le : ∀ x, acc x ≤ 1) (hper : ∀ x, acc x * (acc x / q - hinv) ≤ B x) : (∑' x, Pr[= x | mx] * acc x) * ((∑' x, Pr[= x | mx] * acc x) / q - hinv) ≤ ∑' x, Pr[= x | mx] * B x := by classical set w : X → ℝ≥0∞ := fun x => Pr[= x | mx] set μ : ℝ≥0∞ := ∑' x, w x * acc x with hμ_def have hw_tsum_le_one : ∑' x, w x ≤ 1 := tsum_probOutput_le_one have hμ_le_one : μ ≤ 1 := by calc μ = ∑' x, w x * acc x := rfl _ ≤ ∑' x, w x * 1 := by gcongr with x; exact hacc_le x _ = ∑' x, w x := by simp _ ≤ 1 := hw_tsum_le_one have hμ_ne_top : μ ≠ ⊤ := ne_top_of_le_ne_top ENNReal.one_ne_top hμ_le_one have hμ_hinv_ne_top : ∑' x, w x * acc x * hinv ≠ ⊤ := by rw [ENNReal.tsum_mul_right]; exact ENNReal.mul_ne_top hμ_ne_top hinv_ne_top have hCS : μ ^ 2 ≤ ∑' x, w x * acc x ^ 2 := ENNReal.sq_tsum_le_tsum_sq w acc hw_tsum_le_one calc μ * (μ / q - hinv) = μ ^ 2 / q - μ * hinv := by rw [ENNReal.mul_sub (fun _ _ => hμ_ne_top), sq, mul_div_assoc] _ ≤ (∑' x, w x * acc x ^ 2) / q - μ * hinv := by gcongr _ = (∑' x, w x * acc x ^ 2 / q) - ∑' x, w x * acc x * hinv := by rw [hμ_def] simp_rw [div_eq_mul_inv, ENNReal.tsum_mul_right] _ ≤ ∑' x, (w x * acc x ^ 2 / q - w x * acc x * hinv) := tsum_sub_tsum_le_tsum_sub _ _ hμ_hinv_ne_top _ = ∑' x, w x * (acc x * (acc x / q - hinv)) := by refine tsum_congr fun x => ?_ have hwx_ne_top : w x ≠ ⊤ := ne_top_of_le_ne_top ENNReal.one_ne_top probOutput_le_one have hax_ne_top : acc x ≠ ⊤ := ne_top_of_le_ne_top ENNReal.one_ne_top (hacc_le x) rw [ENNReal.mul_sub (fun _ _ => hax_ne_top), sq, mul_div_assoc, ENNReal.mul_sub (fun _ _ => hwx_ne_top), mul_div_assoc, mul_assoc] _ ≤ ∑' x, w x * B x := by gcongr with x; exact hper x- Project
- VCVio
- License
- Apache-2.0
- Commit
- 2ceb2d825ee3
- Source
- VCVio/EvalDist/Inequalities.lean:66-106
Reuse this declaration
Bring the exact result into your workflow
The import identifies the source module. Your project still needs the pinned package dependency shown on this page.
What this badge means
This completion status comes from the project or community source. It has not yet been represented here as an independent rebuild and axiom audit.
Continue in this project
Related declarations
Expected Cost Nat eq sum tail probs of pathwise Cost At Most
AddWriterT.expectedCostNat_eq_sum_tail_probs_of_pathwiseCostAtMost
Plain-language statement
Finite tail-sum formula for natural-valued writer cost under a pathwise upper bound. If every execution path of oa incurs cost at most n, then the tail probabilities vanish above n, so the infinite tail sum truncates to Finset.range n.
Source project: VCVio
Person-level attribution pending.
IND CPA advantage to Real le sum step signed Advantage Real abs
AsymmEncAlg.IND_CPA_advantage_toReal_le_sum_step_signedAdvantageReal_abs
Plain-language statement
Planned generic one-time-to-many-time lift: bounded multi-query IND-CPA advantage is at most the sum of the extracted one-time signed advantages over the first q fresh LR queries.
Source project: VCVio
Person-level attribution pending.
IND CPA LR hybrid Game q eval Dist eq left of Makes At Most Queries
AsymmEncAlg.IND_CPA_LR_hybridGame_q_evalDist_eq_left_of_MakesAtMostQueries
Plain-language statement
If an adversary makes at most q fresh LR queries, then the leftUntil = q LR-hybrid is the all-left endpoint game.
Source project: VCVio
Person-level attribution pending.