OnlineLearning.separating_weight_sum_pos
theoremverifiedimported Lean
Positivity of the separator's total weight. Let $X, Y \subseteq \bbr$, let $f : \bbr \to \bbr \to \bbr$ be a payoff function, and let $u$ be a finite sample of columns drawn from $Y$. Suppose $X$ is nonempty, fix a vector $c \in \bbr^u$, and let $L : \bbr^u \to \bbr$ be a continuous linear functional that strictly separates $c$ from the finite upper image $U(X,f,u)$, in the sense that $L(z) < L(c)$ for every $z \in U(X,f,u)$. Writing $e_y \in \bbr^u$ for the standard basis vector supported at coordinate $y$, one then has \[ 0 \;<\; \sum_{y \in u} \bigl(-L(e_y)\bigr). \]
Statement
set_option relaxedAutoImplicit false in set_option autoImplicit false in open Real Finset BigOperators in lemma OnlineLearning.separating_weight_sum_pos {X Y : Set ℝ} {f : ℝ → ℝ → ℝ} {u : Finset Y} {cvec : u → ℝ} (hX : X.Nonempty) (L : (u → ℝ) →L[ℝ] ℝ) (hsep : ∀ z ∈ finiteUpperImage X f u, L z < L cvec) : 0 < ∑ y : u, -L (Pi.single y (1 : ℝ))
- Source
- TCSlib, `OnlineLearning.separating_weight_sum_pos`
- Lean proof
- By Karim Abdel Sadek, Mark Bedaywi, from github.com/Shilun-Allan-Li/tcslib at
c8a02591b1e5(Copyright (c) 2026 Karim Abdel Sadek and Mark Bedaywi. All rights reserved, Apache-2.0). Compiled here against leanprover/lean4:v4.33.1, Mathlib v4.33.1; written for leanprover/lean4:v4.25.0.OnlineLearning.separating_weight_sum_pos← OnlineLearning.separating_weight_sum_pos (TCSlib/LearningTheory/Minimax/ConvexMinimaxSeparation.lean:234): verbatim
- Verified
- 07 Oct 2026
- Axioms
Classical.choiceQuot.soundpropext- Built on 4
- OnlineLearning.continuousLinearMap_pi_apply_eq_sum_singleCoordinate expansion of a continuous linear functional. Let $\iota$ be a finite index set and let $L \colon \bbr^{\iota} \to \bbr$ be a…
- OnlineLearning.separating_coordinate_nonposSeparating functional is coordinatewise nonpositive. Let $X, Y \subseteq \bbr$, let $f : \bbr \to \bbr \to \bbr$ be a payoff, let $u$ be a…
- OnlineLearning.finiteUpperImageFor sets $X, Y \subseteq \bbr$, a payoff $f : \bbr \to \bbr \to \bbr$, and a finite sample $u$ of columns from $Y$, the finite upper image…
- OnlineLearning.separating_functional_ne_zeroA nonzero separating functional over the upper image. Let $X, Y \subseteq \bbr$, let $f : \bbr \to \bbr \to \bbr$ be a payoff, and let $u$…
- Used by 1
- OnlineLearning.finite_sublevel_intersections_by_separationNonempty finite intersection of minimax sublevel sets. Let $X, Y \subseteq \bbr$ and $f : \bbr \times \bbr \to \bbr$ satisfy the…
Lean source view module on GitHub
Choose a tactic — a marked line — to see the goals Lean had before and after it, as in Lean’s infoview. Step with ← → or j k.
set_option relaxedAutoImplicit false inset_option autoImplicit false inopen Real Finset BigOperators in/-- For a functional `L` separating the upper image of a finite column sample from a point, the negated coordinate coefficients `−L(e_y)` have strictly positive total mass `Σ_y −L(e_y) > 0`; dividing by this mass turns them into a probability vector. **Proof sketch.** Step 1: every `−L(e_y)` is nonnegative by `separating_coordinate_nonpos`. Step 2: some coefficient is nonzero, since otherwise the coordinate expansion `continuousLinearMap_pi_apply_eq_sum_single` would make `L = 0`, contradicting `separating_functional_ne_zero`. Step 3: that coefficient is then strictly negative, so the sum of nonnegative terms with one positive term is positive. -/lemma OnlineLearning.separating_weight_sum_pos {X Y : Set ℝ} {f : ℝ → ℝ → ℝ} {u : Finset Y} {cvec : u → ℝ} (hX : X.Nonempty) (L : (u → ℝ) →L[ℝ] ℝ) (hsep : ∀ z ∈ finiteUpperImage X f u, L z < L cvec) : 0 < ∑ y : u, -L (Pi.single y (1 : ℝ)) := by classical -- Step 1: all negated coefficients are nonnegative. have hnonpos := separating_coordinate_nonpos (X := X) (Y := Y) (f := f) hX L hsep have hnonneg : ∀ y : u, 0 ≤ -L (Pi.single y (1 : ℝ)) := by intro y linarith [hnonpos y] -- Step 2: some coefficient is nonzero, else `L = 0`. have hne : ∃ y : u, L (Pi.single y (1 : ℝ)) ≠ 0 := by by_contra hnone have hall : ∀ y : u, L (Pi.single y (1 : ℝ)) = 0 := by intro y exact not_not.mp (by simpa using (show ¬ L (Pi.single y (1 : ℝ)) ≠ 0 from fun hy => hnone ⟨y, hy⟩)) have hLzero : L = 0 := by ext z rw [continuousLinearMap_pi_apply_eq_sum_single L z] simp [hall] exact separating_functional_ne_zero (X := X) (Y := Y) (f := f) hX L hsep hLzero -- Step 3: that coefficient is strictly negative; the sum is positive. rcases hne with ⟨y, hy⟩ have hpos_y : 0 < -L (Pi.single y (1 : ℝ)) := by have hle := hnonpos y have hlt : L (Pi.single y (1 : ℝ)) < 0 := lt_of_le_of_ne hle hy linarith exact Finset.sum_pos' (fun i _ => hnonneg i) ⟨y, Finset.mem_univ y, hpos_y⟩