The failure set of the hat network is not convex: the network fails at
ProvedGrokkingNextCycle.hatNet_failure_not_convexaether-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 sorrySource