theorem
proved
J_log_quadratic_approx
show as:
view math explainer →
open explainer
Generate a durable explainer page for this declaration.
open lean source
IndisputableMonolith.Foundation.DiscretenessForcing on GitHub at line 286.
browse module
All declarations in this module, on Recognition.
explainer page
Derivations using this theorem
depends on
formal source
283
284/-- **THEOREM (GROUNDED)**: Quadratic approximation of J_log.
285 For small perturbations, the cost is approximately quadratic in the log-ratio. -/
286theorem J_log_quadratic_approx (ε : ℝ) (hε : |ε| < 1) :
287 |J_log ε - ε^2 / 2| ≤ |ε|^4 / 20 := by
288 -- J_log ε = Jcost (exp ε) = Real.cosh ε - 1
289 have h_cosh : J_log ε = Real.cosh ε - 1 := by
290 simp [J_log, Real.cosh_eq, Real.exp_neg]
291 rw [h_cosh]
292
293
294 have h_abs : |ε|^4 = ε^4 := by
295 calc |ε|^4 = (|ε|^2)^2 := by ring
296 _ = (ε^2)^2 := by rw [sq_abs]
297 _ = ε^4 := by ring
298 rw [h_abs]
299 exact cosh_quadratic_bound ε hε
300
301
302
303
304
305
306
307
308
309/-! ## Instability in Continuous Spaces -/
310
311/-- A configuration is "stable" if it's a strict local minimum of J.
312 This means there's a neighborhood where it's the unique minimizer. -/
313def IsStable (x : ℝ) : Prop :=
314 ∃ ε > 0, ∀ y : ℝ, y ≠ x → |y - x| < ε → defect x < defect y
315
316/-- In a path-connected space with continuous J, there are no isolated stable points.
papers checked against this theorem (showing 2 of 2)
-
Survey maps paths to efficient LLM reasoning
"RL with Length Reward Design... length reward assigns higher scores to short, correct answers while penalizing lengthy or incorrect ones"
-
Mixup on example pairs improves neural network generalization
"mixup regularizes the neural network to favor simple linear behavior in-between training examples."