Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

The failure set of the hat network is not convex: the network fails at

Proved
GrokkingNextCycle.hatNet_failure_not_convex

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

aether-catalogmachinelearning

The failure set of the hat network is not convex: the network fails at times 0 and 2 but succeeds at the intermediate time 1. Hence the nonnegativity assumption on the output weights in failure_set_convex is necessary.

theorem GrokkingNextCycle.hatNet_failure_not_convex: ¬ Convex ℝ {t : ℝ | hatNet t ≤ 0} := by sorry

Formalization Note Transplanted verbatim from the Aether Catalog source MachineLearning/GrokkingDelayedTransition/NextCycle.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/NextCycle.lean
import Mathlib
import Definitions.Def_MachineLearning_GrokkingDelayedTransition_GradientFlowThreshold
import Definitions.Def_MachineLearning_GrokkingDelayedTransition_NextCycle
import Definitions.Def_MachineLearning_GrokkingDelayedTransition_VectorMargin

/-!
# Closing the three next-cycle sub-conjectures of the grokking thread

`FUTURE_DIRECTIONS.md` of the previous cycle listed three concrete sub-conjectures
(N1)–(N3), and a sharpness half for Conjecture 4.  This file resolves all of
them, building directly on the definitions of
`GrokkingDelayedTransition/GradientFlowThreshold.lean` and
`GrokkingDelayedTransition/VectorMargin.lean`.

* **(N1) Filter-level divergence of the delay.**  `crossTime_tendsto_atTop`:
  the crossing time of the weight-decayed gradient flow tends to `+∞` as the
  weight decay increases to its critical value `λ_c = s/θ`.  This upgrades the
  purely existential `crossTime_diverges_at_criticality` to a genuine limit
  statement along the filter `𝓝[<] λ_c`.

* **(N2) An exact `1/m` width law.**  `zeroBias_sharp_threshold`: for a network
  with vanishing hidden biases the *sandwich* `delay_scaling_sandwich` collapses
  to an **equality**, the sharp delay being exactly `|c| / S` with
  `S = ∑ⱼ aⱼ gⱼ` the total signal.  Specialized to a symmetric width-`m`
  network (`symNet_sharp_threshold`) this gives `τ(m) = |c| / (m a g)`, hence
  the exact width law `m · τ(m) = |c|/(a g)` (`symNet_delay_width_law`) and
  `τ(m) → 0` (`symNet_delay_tendsto_zero`).

* **(N3) Grokking can happen twice.**  `hatNet_failure_set` computes the failure
  set of an explicit *sign-indefinite* width-three network to be exactly
  `(-∞, 1/2] ∪ [3/2, ∞)`, and `hatNet_failure_not_convex` shows it is not
  convex.  So the nonnegativity of the output weights in `netRamp_convexOn` /
  `failure_set_convex` cannot be dropped: with signed output weights a network
  can grok, un-grok and re-grok.

* **Sharpness of the robustness bound.**  `perturbed_threshold_shift_exact`:
  the displacement `ε/κ` of `perturbed_delayed_transition` is attained by the
  constant perturbation `-ε`, so that bound is sharp.
-/

open GrokkingNextCycle

open Filter Topology Set

/-! ### (N1) The delay diverges, as a limit along `𝓝[<] λ_c` -/


/-! ### (N2) An exact delay formula, and the `1/m` width law -/

-- open removed: section is not a namespace











/-! ### (N3) A sign-indefinite network that groks twice -/
Formal statement
theorem GrokkingNextCycle.hatNet_failure_not_convex: ¬ Convex ℝ {t : ℝ | hatNet t ≤ 0} := by sorry
Source
https://github.com/paulklemstine/Lean/blob/53c2925a02/Catalog/MachineLearning/GrokkingDelayedTransition/NextCycle.lean#L318

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