theorem
proved
additive_composition_is_minimal
show as:
view math explainer →
open explainer
Generate a durable explainer page for this declaration.
open lean source
IndisputableMonolith.Foundation.HierarchyForcing on GitHub at line 108.
browse module
All declarations in this module, on Recognition.
explainer page
Derivations using this theorem
depends on
used by
formal source
105/-- **Theorem (Phase 3)**: Among recurrence coefficients (a, b) with
106a ≥ 1 and b ≥ 1, the pair (1, 1) uniquely minimizes max(a, b).
107No axiom needed — this is pure arithmetic. -/
108theorem additive_composition_is_minimal (a b : ℕ) (ha : 1 ≤ a) (hb : 1 ≤ b) :
109 max a b = 1 → a = 1 ∧ b = 1 := by
110 intro h
111 constructor
112 · exact Nat.le_antisymm (by omega) ha
113 · exact Nat.le_antisymm (by omega) hb
114
115/-- The pair (1,1) achieves max = 1. -/
116theorem min_max_achieved : max 1 1 = 1 := by simp
117
118/-- Any other pair has max ≥ 2. -/
119theorem other_pairs_larger (a b : ℕ) (ha : 1 ≤ a) (hb : 1 ≤ b)
120 (h : ¬(a = 1 ∧ b = 1)) : 2 ≤ max a b := by omega
121
122/-- Construct the uniform scale ladder from forced data. -/
123noncomputable def hierarchy_forced
124 (M : NontrivialMultilevelComposition)
125 (no_free_scale : ∀ j k,
126 M.levels (j + 1) / M.levels j = M.levels (k + 1) / M.levels k)
127 (ratio_gt_one : 1 < M.levels 1 / M.levels 0) :
128 UniformScaleLadder :=
129 { levels := M.levels
130 levels_pos := M.levels_pos
131 ratio := M.levels 1 / M.levels 0
132 ratio_gt_one := ratio_gt_one
133 uniform_scaling := fun k => by
134 have hk := M.levels_pos k
135 have h0 := M.levels_pos 0
136 have hratio := no_free_scale k 0
137 rw [div_eq_div_iff (ne_of_gt hk) (ne_of_gt h0)] at hratio
138 field_simp; linarith }
papers checked against this theorem (showing 2 of 2)
-
20% high-entropy tokens suffice for effective RL on LLM reasoning
"utilizing only 20% of the tokens while maintaining performance comparable to full-gradient updates... training exclusively on the 80% lowest-entropy tokens leads to a marked decline"
-
Activation differences steer LLMs without retraining
"ActAdd is lightweight: it does not require any machine optimization and works with a single pair of data points"