Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

For a positive signal strength the single-unit score has sharp threshold

Proved
GrokkingVector.unitScore_threshold

by raver1975 · Sep 17, 2026 · Mathlib c5ea003 (Lean v4.30.0)

aether-catalogmachinelearning

For a positive signal strength the single-unit score has sharp threshold 1/s.

theorem GrokkingVector.unitScore_threshold(s : ℝ) (hs : 0 < s) :
    (∀ t ≤ 1 / s, unitScore s t ≤ 0) ∧ (∀ t, 1 / s < t → 0 < unitScore s t) := by sorry

Formalization Note Transplanted verbatim from the Aether Catalog source MachineLearning/GrokkingDelayedTransition/VectorMargin.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.

Preamble
-- Thm stub generated from MachineLearning/GrokkingDelayedTransition/VectorMargin.lean
import Mathlib
import Definitions.Def_MachineLearning_GrokkingDelayedTransition_VectorMargin

/-!
# Delayed margin positivity for vector-valued two-layer ReLU networks

The catalog file `Catalog/MachineLearning/GrokkingPhaseTransition.lean` treats a
width-one *scalar* network.  This file carries out Future Direction 1 (finite
hidden width, matrix weights, a finite test set, and a classification margin)
and Direction 3 (train/test separation).

Main results.

* `exists_sharp_threshold`: any monotone continuous signal that starts negative
  and is eventually positive has a *sharp* threshold `τ`: it is `≤ 0` on
  `(-∞, τ]` and `> 0` on `(τ, ∞)`.  The threshold is unique
  (`sharp_threshold_unique`).
* `margin_sharp_threshold`: the same holds for the *minimum* over a finite test
  set of finitely many such signals — the delayed transition survives taking a
  worst-case margin over a test set.
* `netMargin_delayed_positivity`: for a genuinely vector-valued two-layer ReLU
  network (hidden width `m`, input dimension `d`, matrix weights, negative
  output bias, a finite two-class test set) the classification margin is
  nonpositive up to an explicit delay and strictly positive afterwards.
* `grokking_window_eq`: for a concrete dataset the set of times at which the
  training set is already perfectly classified while the test point is still
  misclassified is *exactly* the interval `(1/2, 2]` — a formal train/test
  separation window.
-/

open GrokkingVector

open Finset Set

/-! ### Sharp thresholds for monotone continuous signals -/



/-! ### Worst-case margin over a finite test set -/







/-! ### Vector-valued two-layer ReLU networks -/















/-! ### Delayed positivity of the classification margin -/



/-! ### An explicit train/test separation window -/








/-! ### A concrete instance: the hypotheses are satisfiable and the delay is exact

The general theorem `netMargin_delayed_positivity` is not vacuous: here is a
two-class test set on which all of its hypotheses hold, and for which the sharp
margin threshold can be computed exactly (it equals `1`).
-/

open Example

open GrokkingVector
















/-!
# Quantitative delay bounds, convex (tropical) structure, and robustness

Second research cycle.  `VectorMargin.lean` proves that the worst-case margin of
a vector-valued two-layer ReLU network has a *sharp* delay `τ`.  Here we

* locate `τ` quantitatively: `delay_lower_bound_of_threshold` and
  `delay_upper_bound_of_threshold` sandwich the delay between `|c| / S`, where
  `S = ∑ⱼ aⱼ gⱼ` is the total signal of the point, and
  `(|c|/a_{j₀} - b_{j₀})/g_{j₀}` coming from any single active unit — so the
  delay scales like *output-bias magnitude divided by signal strength*;
* connect the delay to the *tropical / convex-geometric* picture of the catalog
  file `Catalog/Tropical/NeuralCoding/GrokPhaseTransition.lean`: the ramped
  network output is a convex piecewise-linear (tropical) function of the ramp
  parameter (`netRamp_convexOn`), hence the set of times at which the network
  still fails is always an interval (`failure_set_convex`) — the delayed
  transition can never happen twice;
* prove robustness of the delayed transition itself
  (`perturbed_delayed_transition`): a uniformly `ε`-close trajectory keeps the
  transition, with the threshold moving by at most `ε/κ` where `κ` is the growth
  rate after the threshold;
* prove that the ratio between test delay and train delay is unbounded
  (`grokking_ratio_unbounded`): weakening the test signal makes the grokking
  window arbitrarily long compared with the time to fit the training set.
-/

open GrokkingVector

open Finset Set

/-! ### Convexity: the tropical structure of the ramped output -/






/-! ### Quantitative bounds on the delay -/





/-! ### Robustness of the delayed transition -/


/-! ### The grokking ratio can be made arbitrarily large -/
Formal statement
theorem GrokkingVector.unitScore_threshold(s : ℝ) (hs : 0 < s) :
    (∀ t ≤ 1 / s, unitScore s t ≤ 0) ∧ (∀ t, 1 / s < t → 0 < unitScore s t) := by sorry
Source
https://github.com/paulklemstine/Lean/blob/53c2925a02/Catalog/MachineLearning/GrokkingDelayedTransition/VectorMargin.lean#L706

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me