Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Finite Neural Networks, Gaussian Initialization, and the Limiting NTK

Definition
JGH_NTK_Model

by Minghui · Sep 26, 2026 · Mathlib c5ea003 (Lean v4.30.0)

machine-learningneural-tangent-kernelprobability

Mathematical model

Notation and probability model

Let d,q≥1d,q\ge1d,q≥1 be the input and output dimensions, h≥0h\ge0h≥0 the number of hidden layers, L=h+1L=h+1L=h+1, β>0\beta>0β>0, and σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R a Lipschitz activation with a nonnegative Lipschitz constant KKK. For widths n0=dn_0=dn0​=d, nL=qn_L=qnL​=q, and nℓ=wℓ−1+1n_\ell=w_{\ell-1}+1nℓ​=wℓ−1​+1 with wi∈Nw_i\in\mathbb Nwi​∈N, the probability space Ωw\Omega_wΩw​ is the finite real parameter space with every weight and bias coordinate independently N(0,1)\mathcal N(0,1)N(0,1). Its law is Pw\mathbb P_wPw​. The network has the recursion

zj(ℓ+1)(x)=1nℓ∑iWji(ℓ)ai(ℓ)(x)+βbj(ℓ),a(0)(x)=x,a(ℓ)(x)=σ(z(ℓ)(x)) (1≤ℓ≤h),z^{(\ell+1)}_j(x)=\frac{1}{\sqrt{n_\ell}}\sum_i W^{(\ell)}_{ji} a^{(\ell)}_i(x)+\beta b^{(\ell)}_j,\qquad a^{(0)}(x)=x,\quad a^{(\ell)}(x)=\sigma(z^{(\ell)}(x))\ (1\le\ell\le h),zj(ℓ+1)​(x)=nℓ​​1​i∑​Wji(ℓ)​ai(ℓ)​(x)+βbj(ℓ)​,a(0)(x)=x,a(ℓ)(x)=σ(z(ℓ)(x)) (1≤ℓ≤h),

with output fθ=z(L)f_\theta=z^{(L)}fθ​=z(L). The full kernel, including all weights and biases, is

Θkk′(L)(θ;x,y)=∑p∂θpfθ,k(x)∂θpfθ,k′(y).\Theta^{(L)}_{kk'}(\theta;x,y)=\sum_p \partial_{\theta_p}f_{\theta,k}(x)\partial_{\theta_p}f_{\theta,k'}(y).Θkk′(L)​(θ;x,y)=p∑​∂θp​​fθ,k​(x)∂θp​​fθ,k′​(y).

For a centered Gaussian pair (U,V)(U,V)(U,V) with covariance induced by Σ(ℓ)\Sigma^{(\ell)}Σ(ℓ) on (x,y)(x,y)(x,y), put

Σ(1)(x,y)=⟨x,y⟩/d+β2,Σ(ℓ+1)(x,y)=E[σ(U)σ(V)]+β2,\Sigma^{(1)}(x,y)=\langle x,y\rangle/d+\beta^2,\qquad \Sigma^{(\ell+1)}(x,y)=\mathbb E[\sigma(U)\sigma(V)]+\beta^2,Σ(1)(x,y)=⟨x,y⟩/d+β2,Σ(ℓ+1)(x,y)=E[σ(U)σ(V)]+β2, Σ˙(ℓ+1)(x,y)=E[σ′(U)σ′(V)],Θ∞(1)=Σ(1),Θ∞(ℓ+1)=Θ∞(ℓ)Σ˙(ℓ+1)+Σ(ℓ+1).\dot\Sigma^{(\ell+1)}(x,y)=\mathbb E[\sigma'(U)\sigma'(V)],\qquad \Theta_\infty^{(1)}=\Sigma^{(1)},\quad \Theta_\infty^{(\ell+1)}=\Theta_\infty^{(\ell)}\dot\Sigma^{(\ell+1)}+\Sigma^{(\ell+1)}.Σ˙(ℓ+1)(x,y)=E[σ′(U)σ′(V)],Θ∞(1)​=Σ(1),Θ∞(ℓ+1)​=Θ∞(ℓ)​Σ˙(ℓ+1)+Σ(ℓ+1).

All kernel products in the last expression are pointwise. Local index hhh in covarianceKernel and limitingNTK denotes paper depth h+1h+1h+1. The dataset X=(xi)i<NX=(x_i)_{i<N}X=(xi​)i<N​ is any fixed finite family; repetitions and N=0N=0N=0 are allowed. δkk′\delta_{kk'}δkk′​ is the Kronecker delta.

The limit takes n1n_1n1​ to infinity first and nhn_hnh​ last. More precisely, for any required error tolerance, the width condition is ∀eventuallywh−1⋯∀eventuallyw0\forall^{\mathrm{eventually}}w_{h-1}\cdots \forall^{\mathrm{eventually}}w_0∀eventuallywh−1​⋯∀eventuallyw0​; each inner threshold may depend on the fixed outer widths. For h=0h=0h=0 the filter is concentrated on the unique empty width vector, so the statements require the exact affine base case. This is not a simultaneous-width or whole-input-space uniform limit.

Formalization note: Gaussian measures are concrete Mathlib measures, including singular covariance. The covariance-validity milestone establishes their covariance interpretation; it is not a hypothesis of either convergence target. The activation assumption is only Lipschitz. Derivatives take Mathlib's zero value at points without derivatives, and proofs must justify the null exceptional set under positive Gaussian bias. Native convergence in distribution includes almost-everywhere measurability and weak convergence of probability laws. Primary source conventions: Jacot–Gabriel–Hongler, Section 2, PDF pp. 2–3; Section 4.1, PDF p. 5, Proposition 1, Theorem 1 and Remarks 2–3; Appendix A opening paragraphs, PDF p. 11, and Appendix A.1, PDF pp. 11–13. The relevant displays have no equation numbers.

Definition code
import Mathlib.Analysis.Calculus.FDeriv.Basic
import Mathlib.Analysis.Calculus.Deriv.Basic
import Mathlib.Probability.Distributions.Gaussian.Multivariate
import Mathlib.MeasureTheory.Function.ConvergenceInDistribution
import Mathlib.Data.Fin.Tuple.Basic

/-!
Finite networks and their initialization laws for Jacot--Gabriel--Hongler,
arXiv:1806.07572v4, Sections 2 and 4.1, PDF pp. 2--5; Appendix A, pp. 11--13.
The referenced architecture, kernel recurrences, and limit conventions are unnumbered displays.
-/

noncomputable section

open MeasureTheory ProbabilityTheory Filter
open scoped BigOperators Topology ENNReal

namespace JGH

/-- Layer `l`, destination neuron, and source neuron (`none` denotes its bias). -/
abbrev ParameterIndex (L : ℕ) (width : ℕ → ℕ) :=
  Σ l : Fin L, Fin (width (l.val + 1)) × Option (Fin (width l.val))

abbrev Parameters (L : ℕ) (width : ℕ → ℕ) := ParameterIndex L width → ℝ

/-- Independent standard normal law on every weight and bias coordinate. -/
def initialization (L : ℕ) (width : ℕ → ℕ) : Measure (Parameters L width) :=
  Measure.pi (fun _ ↦ gaussianReal 0 1)

instance (L : ℕ) (width : ℕ → ℕ) : IsProbabilityMeasure (initialization L width) := by
  unfold initialization
  infer_instance

/-- Preactivations, with the input at layer zero. Values beyond `L` are unused. -/
def preactivation (L : ℕ) (width : ℕ → ℕ) (σ : ℝ → ℝ) (β : ℝ)
    (θ : Parameters L width) (x : Fin (width 0) → ℝ) : (l : ℕ) → Fin (width l) → ℝ
  | 0 => x
  | l + 1 => fun j ↦
      if hl : l < L then
        (∑ i : Fin (width l), θ ⟨⟨l, hl⟩, j, some i⟩ *
          (if l = 0 then preactivation L width σ β θ x l i
            else σ (preactivation L width σ β θ x l i))) / Real.sqrt (width l) +
          β * θ ⟨⟨l, hl⟩, j, none⟩
      else 0

/-- The final affine preactivation, without applying the nonlinearity to the output. -/
def network (L : ℕ) (width : ℕ → ℕ) (σ : ℝ → ℝ) (β : ℝ)
    (θ : Parameters L width) (x : Fin (width 0) → ℝ) : Fin (width L) → ℝ :=
  preactivation L width σ β θ x L

/-- Parameter derivative in a coordinate direction; totalized by Mathlib off differentiability. -/
def parameterDerivative (L : ℕ) (width : ℕ → ℕ) (σ : ℝ → ℝ) (β : ℝ)
    (θ : Parameters L width) (x : Fin (width 0) → ℝ)
    (k : Fin (width L)) (p : ParameterIndex L width) : ℝ := by
  classical
  exact fderiv ℝ (fun η ↦ network L width σ β η x k) θ (Pi.single p 1)

/-- The full weight-and-bias neural tangent kernel, including cross-output entries. -/
def finiteNTK (L : ℕ) (width : ℕ → ℕ) (σ : ℝ → ℝ) (β : ℝ)
    (θ : Parameters L width) (x y : Fin (width 0) → ℝ)
    (k k' : Fin (width L)) : ℝ :=
  ∑ p, parameterDerivative L width σ β θ x k p *
    parameterDerivative L width σ β θ y k' p

abbrev Input (d : ℕ) := Fin d → ℝ
abbrev Kernel (d : ℕ) := Input d → Input d → ℝ

/-- Centered Gaussian pair, allowing singular covariance (in particular, repeated inputs). -/
def gaussianPair {d : ℕ} (K : Kernel d) (x y : Input d) :
    Measure (EuclideanSpace ℝ (Fin 2)) :=
  multivariateGaussian 0 (fun i j ↦ K (![x, y] i) (![x, y] j))

def gaussianProductExpectation {d : ℕ} (K : Kernel d) (φ : ℝ → ℝ)
    (x y : Input d) : ℝ :=
  ∫ z, φ (z 0) * φ (z 1) ∂gaussianPair K x y

/-- Index `h` means network depth `h + 1`, so `h = 0` is the affine base case. -/
def covarianceKernel (d : ℕ) (σ : ℝ → ℝ) (β : ℝ) : ℕ → Kernel d
  | 0 => fun x y ↦ (∑ i, x i * y i) / d + β ^ 2
  | h + 1 => fun x y ↦ gaussianProductExpectation (covarianceKernel d σ β h) σ x y + β ^ 2

/-- The deterministic NTK recurrence in Theorem 1, PDF p. 5. -/
def limitingNTK (d : ℕ) (σ : ℝ → ℝ) (β : ℝ) : ℕ → Kernel d
  | 0 => covarianceKernel d σ β 0
  | h + 1 => fun x y ↦ limitingNTK d σ β h x y *
      gaussianProductExpectation (covarianceKernel d σ β h) (deriv σ) x y +
        covarianceKernel d σ β (h + 1) x y

/-- Positive hidden width `w i + 1`; fixed input and output widths `d` and `q`. -/
def widths {h : ℕ} (d q : ℕ) (w : Fin h → ℕ) (l : ℕ) : ℕ :=
  if l = 0 then d else if hl : l - 1 < h then w ⟨l - 1, hl⟩ + 1 else q

/-- Earlier hidden widths tend to infinity first; the last width is outermost. -/
def sequentialWidths : (h : ℕ) → Filter (Fin h → ℕ)
  | 0 => pure Fin.elim0
  | h + 1 => (atTop : Filter ℕ).bind
      (fun last ↦ (sequentialWidths h).map (fun earlier ↦ Fin.snoc earlier last))

def inputForWidths {h : ℕ} (d q : ℕ) (w : Fin h → ℕ) (x : Input d) :
    Fin (widths d q w 0) → ℝ :=
  fun i ↦ x ⟨i.val, by simpa only [widths, if_pos rfl] using i.isLt⟩

def outputForWidths {h : ℕ} (d q : ℕ) (w : Fin h → ℕ) (k : Fin q) :
    Fin (widths d q w (h + 1)) :=
  ⟨k.val, by simpa only [widths, Nat.add_one_ne_zero, if_false, Nat.add_sub_cancel,
    Nat.lt_irrefl, dite_false] using k.isLt⟩

def initializedOutput {h N : ℕ} (d q : ℕ) (σ : ℝ → ℝ) (β : ℝ)
    (X : Fin N → Input d) (w : Fin h → ℕ)
    (θ : Parameters (h + 1) (widths d q w)) : EuclideanSpace ℝ (Fin N × Fin q) :=
  WithLp.toLp 2 (fun a ↦ network (h + 1) (widths d q w) σ β θ
    (inputForWidths d q w (X a.1)) (outputForWidths d q w a.2))

/-- Joint limit law on all dataset and output coordinates: independent output GPs. -/
def outputGaussian {N : ℕ} (d q : ℕ) (σ : ℝ → ℝ) (β : ℝ)
    (h : ℕ) (X : Fin N → Input d) : Measure (EuclideanSpace ℝ (Fin N × Fin q)) :=
  multivariateGaussian 0 (fun a b ↦
    if a.2 = b.2 then covarianceKernel d σ β h (X a.1) (X b.1) else 0)

instance {N : ℕ} (d q : ℕ) (σ : ℝ → ℝ) (β : ℝ) (h : ℕ) (X : Fin N → Input d) :
    IsProbabilityMeasure (outputGaussian d q σ β h X) := by
  unfold outputGaussian
  infer_instance

/-- Probability that some entry on the fixed finite dataset exceeds the tolerance. -/
def ntkBadProbability {h N : ℕ} (d q : ℕ) (σ : ℝ → ℝ) (β : ℝ)
    (X : Fin N → Input d) (ε : ℝ) (w : Fin h → ℕ) : ℝ≥0∞ :=
  initialization (h + 1) (widths d q w) {θ | ∃ i j : Fin N, ∃ k k' : Fin q,
    ε < |finiteNTK (h + 1) (widths d q w) σ β θ
      (inputForWidths d q w (X i)) (inputForWidths d q w (X j))
      (outputForWidths d q w k) (outputForWidths d q w k') -
        (if k = k' then limitingNTK d σ β h (X i) (X j) else 0)|}

end JGH
Source
Arthur Jacot, Franck Gabriel, Clément Hongler, Neural Tangent Kernel: Convergence and Generalization in Neural Networks, NeurIPS 2018, arXiv:1806.07572v4, https://arxiv.org/abs/1806.07572v4; Section 2, PDF p. 2 and PDF p. 3; Section 4.1, PDF p. 5, Proposition 1 and Theorem 1; Appendix A.1, PDF pp. 11–13. Architecture, covariance, and kernel displays are unnumbered.
Read-back

What the Lean code literally says, in plain math · gpt-6

ParameterIndex. For every natural number LLL and width function n:N→Nn:\mathbb N\to\mathbb Nn:N→N, this is the finite set IL,n={(ℓ,j,t):ℓ<L, j<nℓ+1, t∈{⊥}⊔{0,…,nℓ−1}}I_{L,n}=\{(\ell,j,t):\ell<L,\ j<n_{\ell+1},\ t\in\{\bot\}\sqcup\{0,\ldots,n_\ell-1\}\}IL,n​={(ℓ,j,t):ℓ<L, j<nℓ+1​, t∈{⊥}⊔{0,…,nℓ​−1}}. The distinguished index ⊥\bot⊥ represents a bias parameter; the other choices represent incoming weight parameters. Its cardinality is ∑ℓ=0L−1nℓ+1(nℓ+1)\sum_{\ell=0}^{L-1}n_{\ell+1}(n_\ell+1)∑ℓ=0L−1​nℓ+1​(nℓ​+1). No positivity condition is imposed. When L=0L=0L=0, it is empty. A layer with target width zero contributes no parameters; a layer with source width zero still contributes one bias parameter per target coordinate.

Parameters. For every L∈NL\in\mathbb NL∈N and width function n:N→Nn:\mathbb N\to\mathbb Nn:N→N, this is the space RIL,n\mathbb R^{I_{L,n}}RIL,n​ of all real assignments to the indices (ℓ,j,t)(\ell,j,t)(ℓ,j,t) with ℓ<L\ell<Lℓ<L, j<nℓ+1j<n_{\ell+1}j<nℓ+1​, and ttt either a bias symbol ⊥\bot⊥ or an incoming index i<nℓi<n_\elli<nℓ​. There are no restrictions or identifications among the assigned values. When there are no parameter indices, including L=0L=0L=0, this function space contains exactly one element: the empty assignment.

initialization. For every L∈NL\in\mathbb NL∈N and width function n:N→Nn:\mathbb N\to\mathbb Nn:N→N, this is the finite product probability measure ⨂p∈IL,nN(0,1)\bigotimes_{p\in I_{L,n}}\mathcal N(0,1)⨂p∈IL,n​​N(0,1) on the parameter space, where the indices comprise every weight (ℓ,j,i)(\ell,j,i)(ℓ,j,i) and bias (ℓ,j,⊥)(\ell,j,\bot)(ℓ,j,⊥) with ℓ<L\ell<Lℓ<L, j<nℓ+1j<n_{\ell+1}j<nℓ+1​, and i<nℓi<n_\elli<nℓ​. All these coordinates are independent standard real Gaussians, with mean zero and variance one. No width normalization or bias coefficient appears in this measure. For an empty parameter index set, it is the probability measure concentrated on the unique empty parameter assignment.

Probability instance for initialization. For every natural depth LLL and every natural-valued width function nnn, the product measure assigning independent N(0,1)\mathcal N(0,1)N(0,1) coordinates to all weights and biases for the LLL layers has total mass one. This also applies when the parameter index set is empty, since the resulting function space is a singleton. No positive-depth or positive-width assumption is required.

preactivation. Given L∈NL\in\mathbb NL∈N, widths n:N→Nn:\mathbb N\to\mathbb Nn:N→N, an arbitrary function σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, a real number β\betaβ, parameters θℓ,j,i\theta_{\ell,j,i}θℓ,j,i​ and θℓ,j,⊥\theta_{\ell,j,\bot}θℓ,j,⊥​ indexed by ℓ<L\ell<Lℓ<L, j<nℓ+1j<n_{\ell+1}j<nℓ+1​, and i<nℓi<n_\elli<nℓ​, and an input x∈Rn0x\in\mathbb R^{n_0}x∈Rn0​, this defines vectors aℓ∈Rnℓa_\ell\in\mathbb R^{n_\ell}aℓ​∈Rnℓ​ for every natural ℓ\ellℓ. It sets a0=xa_0=xa0​=x; for ℓ<L\ell<Lℓ<L, it sets aℓ+1,j=(∑i<nℓθℓ,j,iuℓ,i)/nℓ+βθℓ,j,⊥a_{\ell+1,j}=(\sum_{i<n_\ell}\theta_{\ell,j,i}u_{\ell,i})/\sqrt{n_\ell}+\beta\theta_{\ell,j,\bot}aℓ+1,j​=(∑i<nℓ​​θℓ,j,i​uℓ,i​)/nℓ​​+βθℓ,j,⊥​, where u0,i=xiu_{0,i}=x_iu0,i​=xi​ and uℓ,i=σ(aℓ,i)u_{\ell,i}=\sigma(a_{\ell,i})uℓ,i​=σ(aℓ,i​) for ℓ>0\ell>0ℓ>0; for ℓ≥L\ell\geq Lℓ≥L, it instead sets aℓ+1,j=0a_{\ell+1,j}=0aℓ+1,j​=0. Widths in square roots are interpreted as real numbers. Empty sums are zero and real division by zero is zero, so an active layer with source width zero has value βθℓ,j,⊥\beta\theta_{\ell,j,\bot}βθℓ,j,⊥​. A zero target width gives the empty vector. The input is not activated before the first affine map, and the recursive step does not activate its resulting vector. No regularity of σ\sigmaσ or restriction on β\betaβ is assumed. All positive layer indices beyond LLL are assigned zero vectors.

network. For natural LLL, widths n:N→Nn:\mathbb N\to\mathbb Nn:N→N, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, real β\betaβ, parameters θ\thetaθ comprising all layer weights and biases, and input x∈Rn0x\in\mathbb R^{n_0}x∈Rn0​, this returns Fθ(x)=aL∈RnLF_\theta(x)=a_L\in\mathbb R^{n_L}Fθ​(x)=aL​∈RnL​, where a0=xa_0=xa0​=x and aℓ+1,j=(∑i<nℓθℓ,j,iuℓ,i)/nℓ+βθℓ,j,⊥a_{\ell+1,j}=(\sum_{i<n_\ell}\theta_{\ell,j,i}u_{\ell,i})/\sqrt{n_\ell}+\beta\theta_{\ell,j,\bot}aℓ+1,j​=(∑i<nℓ​​θℓ,j,i​uℓ,i​)/nℓ​​+βθℓ,j,⊥​ for ℓ<L\ell<Lℓ<L, with u0,i=xiu_{0,i}=x_iu0,i​=xi​ and uℓ,i=σ(aℓ,i)u_{\ell,i}=\sigma(a_{\ell,i})uℓ,i​=σ(aℓ,i​) for ℓ>0\ell>0ℓ>0. Empty sums and division by zero are zero; a zero source width therefore leaves only the bias term, and a zero output width gives an empty output vector. For L=0L=0L=0, the output is exactly xxx, independently of the unique empty parameter assignment. No activation is applied after the last affine map. No continuity, measurability, or differentiability of σ\sigmaσ is assumed.

parameterDerivative. Given natural LLL, widths nnn, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, real β\betaβ, parameters θ∈RIL,n\theta\in\mathbb R^{I_{L,n}}θ∈RIL,n​, input x∈Rn0x\in\mathbb R^{n_0}x∈Rn0​, an output index k<nLk<n_Lk<nL​, and a parameter index p∈IL,np\in I_{L,n}p∈IL,n​, this is D(η↦Fη(x)k)(θ)[ep]D(\eta\mapsto F_\eta(x)_k)(\theta)[e_p]D(η↦Fη​(x)k​)(θ)[ep​]. Here IL,nI_{L,n}IL,n​ contains all weights (ℓ,j,i)(\ell,j,i)(ℓ,j,i) and biases (ℓ,j,⊥)(\ell,j,\bot)(ℓ,j,⊥) with ℓ<L\ell<Lℓ<L, j<nℓ+1j<n_{\ell+1}j<nℓ+1​, and i<nℓi<n_\elli<nℓ​; epe_pep​ equals one at ppp and zero elsewhere; and Fη(x)=aLF_\eta(x)=a_LFη​(x)=aL​ is determined by a0=xa_0=xa0​=x and the affine recursion aℓ+1,j=(∑i<nℓηℓ,j,iuℓ,i)/nℓ+βηℓ,j,⊥a_{\ell+1,j}=(\sum_{i<n_\ell}\eta_{\ell,j,i}u_{\ell,i})/\sqrt{n_\ell}+\beta\eta_{\ell,j,\bot}aℓ+1,j​=(∑i<nℓ​​ηℓ,j,i​uℓ,i​)/nℓ​​+βηℓ,j,⊥​, using u0,i=xiu_{0,i}=x_iu0,i​=xi​ and uℓ,i=σ(aℓ,i)u_{\ell,i}=\sigma(a_{\ell,i})uℓ,i​=σ(aℓ,i​) for ℓ>0\ell>0ℓ>0. The derivative is the real Fréchet derivative of the entire scalar output function on the parameter space. If that scalar function is not Fréchet differentiable at θ\thetaθ, its derivative is defined to be the zero linear map, so the declared value is zero. Merely having a derivative along coordinate ppp is not the condition tested. Empty sums and division by zero are zero. If no parameter index or output index exists, there are no corresponding admissible arguments ppp or kkk.

finiteNTK. For every natural LLL, width function nnn, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, real β\betaβ, parameters θ\thetaθ, inputs x,y∈Rn0x,y\in\mathbb R^{n_0}x,y∈Rn0​, and output indices k,k′<nLk,k'<n_Lk,k′<nL​, this is ∑p∈IL,nD(η↦Fη(x)k)(θ)[ep] D(η↦Fη(y)k′)(θ)[ep]\sum_{p\in I_{L,n}}D(\eta\mapsto F_\eta(x)_k)(\theta)[e_p]\,D(\eta\mapsto F_\eta(y)_{k'})(\theta)[e_p]∑p∈IL,n​​D(η↦Fη​(x)k​)(θ)[ep​]D(η↦Fη​(y)k′​)(θ)[ep​], with no further normalization. The index set contains every layer weight and bias, and epe_pep​ is the coordinate unit vector. For either input zzz, Fη(z)=aLF_\eta(z)=a_LFη​(z)=aL​ is given by a0=za_0=za0​=z and aℓ+1,j=(∑i<nℓηℓ,j,iuℓ,i)/nℓ+βηℓ,j,⊥a_{\ell+1,j}=(\sum_{i<n_\ell}\eta_{\ell,j,i}u_{\ell,i})/\sqrt{n_\ell}+\beta\eta_{\ell,j,\bot}aℓ+1,j​=(∑i<nℓ​​ηℓ,j,i​uℓ,i​)/nℓ​​+βηℓ,j,⊥​, where u0,i=ziu_{0,i}=z_iu0,i​=zi​ and uℓ,i=σ(aℓ,i)u_{\ell,i}=\sigma(a_{\ell,i})uℓ,i​=σ(aℓ,i​) for ℓ>0\ell>0ℓ>0. Each full scalar-output Fréchet derivative is set to zero if it does not exist. Consequently the displayed sum is zero if either scalar output function is not differentiable at θ\thetaθ. Empty sums and division by zero are zero. In particular, this kernel is zero at L=0L=0L=0, although the network then returns its input. If nL=0n_L=0nL​=0, there are no output indices k,k′k,k'k,k′.

Input. For each natural number ddd, this is the space Rd\mathbb R^dRd, represented by real-valued functions on the indices 0,…,d−10,\ldots,d-10,…,d−1, with no restrictions on their values. For d=0d=0d=0, it is a singleton containing the empty function, rather than an empty set of inputs.

Kernel. For each natural number ddd, this is the collection of all functions K:Rd×Rd→RK:\mathbb R^d\times\mathbb R^d\to\mathbb RK:Rd×Rd→R. No symmetry, positivity, positive semidefiniteness, continuity, or measurability is required. For d=0d=0d=0, there is one possible input pair, so a kernel is determined by one real value.

gaussianPair. Given any natural ddd, any function K:Rd×Rd→RK:\mathbb R^d\times\mathbb R^d\to\mathbb RK:Rd×Rd→R, and inputs x,y∈Rdx,y\in\mathbb R^dx,y∈Rd, form S=(K(x,x)K(x,y)K(y,x)K(y,y))S=\begin{pmatrix}K(x,x)&K(x,y)\\K(y,x)&K(y,y)\end{pmatrix}S=(K(x,x)K(y,x)​K(x,y)K(y,y)​). This defines a measure on the Euclidean plane as the law of S1/2GS^{1/2}GS1/2G, where GGG is a standard two-dimensional Gaussian and the mean is zero. If SSS is positive semidefinite, its positive semidefinite square root is used, and the resulting Gaussian has covariance SSS, including singular cases. If SSS is not positive semidefinite, the matrix-square-root operation returns zero and the measure is the point mass at (0,0)(0,0)(0,0). No positive-semidefiniteness or symmetry assumption on KKK is present. The definition also applies when d=0d=0d=0, with x=yx=yx=y the unique empty input.

gaussianProductExpectation. For any natural ddd, arbitrary functions K:Rd×Rd→RK:\mathbb R^d\times\mathbb R^d\to\mathbb RK:Rd×Rd→R and φ:R→R\varphi:\mathbb R\to\mathbb Rφ:R→R, and inputs x,y∈Rdx,y\in\mathbb R^dx,y∈Rd, this is the real Bochner integral ∫R2φ(z0)φ(z1) dγK,x,y(z)\int_{\mathbb R^2}\varphi(z_0)\varphi(z_1)\,d\gamma_{K,x,y}(z)∫R2​φ(z0​)φ(z1​)dγK,x,y​(z). The measure γK,x,y\gamma_{K,x,y}γK,x,y​ is centered Gaussian with covariance (K(x,x)K(x,y)K(y,x)K(y,y))\begin{pmatrix}K(x,x)&K(x,y)\\K(y,x)&K(y,y)\end{pmatrix}(K(x,x)K(y,x)​K(x,y)K(y,y)​) if that matrix is positive semidefinite, and is the point mass at zero otherwise. No measurability or integrability assumption on φ\varphiφ is included. If the integrand is not Bochner integrable, including failure of the requisite almost-everywhere measurability, the integral is defined to be zero. In the point-mass fallback case it equals φ(0)2\varphi(0)^2φ(0)2. Zero input dimension and singular positive-semidefinite covariance matrices are permitted.

covarianceKernel. For each natural ddd, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, and real β\betaβ, this defines real functions Σh\Sigma_hΣh​ on Rd×Rd\mathbb R^d\times\mathbb R^dRd×Rd, for all h∈Nh\in\mathbb Nh∈N, by Σ0(x,y)=(∑i<dxiyi)/d+β2\Sigma_0(x,y)=(\sum_{i<d}x_i y_i)/d+\beta^2Σ0​(x,y)=(∑i<d​xi​yi​)/d+β2 and Σh+1(x,y)=∫σ(z0)σ(z1) dγh,x,y(z)+β2\Sigma_{h+1}(x,y)=\int\sigma(z_0)\sigma(z_1)\,d\gamma_{h,x,y}(z)+\beta^2Σh+1​(x,y)=∫σ(z0​)σ(z1​)dγh,x,y​(z)+β2. The measure γh,x,y\gamma_{h,x,y}γh,x,y​ is centered Gaussian with covariance (Σh(x,x)Σh(x,y)Σh(y,x)Σh(y,y))\begin{pmatrix}\Sigma_h(x,x)&\Sigma_h(x,y)\\\Sigma_h(y,x)&\Sigma_h(y,y)\end{pmatrix}(Σh​(x,x)Σh​(y,x)​Σh​(x,y)Σh​(y,y)​) when that matrix is positive semidefinite, and the point mass at zero otherwise. Every integral is a real Bochner integral set to zero when its integrand is not integrable. Thus a nonintegrable product contributes zero, while a non-positive-semidefinite matrix contributes σ(0)2\sigma(0)^2σ(0)2 through the point-mass convention. No regularity assumptions on σ\sigmaσ occur. For d=0d=0d=0, the initial empty sum and its quotient by zero are zero, so Σ0=β2\Sigma_0=\beta^2Σ0​=β2 on the unique input pair. The definition contains no separate assertion that these functions are positive semidefinite or are actual covariances of random network outputs.

limitingNTK. For each natural ddd, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, and real β\betaβ, this defines functions Θh:Rd×Rd→R\Theta_h:\mathbb R^d\times\mathbb R^d\to\mathbb RΘh​:Rd×Rd→R for every h∈Nh\in\mathbb Nh∈N. Set Σ0(x,y)=(∑i<dxiyi)/d+β2\Sigma_0(x,y)=(\sum_{i<d}x_i y_i)/d+\beta^2Σ0​(x,y)=(∑i<d​xi​yi​)/d+β2 and recursively Σr+1(x,y)=∫σ(z0)σ(z1) dγr,x,y(z)+β2\Sigma_{r+1}(x,y)=\int\sigma(z_0)\sigma(z_1)\,d\gamma_{r,x,y}(z)+\beta^2Σr+1​(x,y)=∫σ(z0​)σ(z1​)dγr,x,y​(z)+β2, where γr,x,y\gamma_{r,x,y}γr,x,y​ is centered Gaussian with matrix entries Σr(x,x),Σr(x,y),Σr(y,x),Σr(y,y)\Sigma_r(x,x),\Sigma_r(x,y),\Sigma_r(y,x),\Sigma_r(y,y)Σr​(x,x),Σr​(x,y),Σr​(y,x),Σr​(y,y), or the point mass at zero when this matrix is not positive semidefinite. Then Θ0=Σ0\Theta_0=\Sigma_0Θ0​=Σ0​ and Θr+1(x,y)=Θr(x,y)∫σ˙(z0)σ˙(z1) dγr,x,y(z)+Σr+1(x,y)\Theta_{r+1}(x,y)=\Theta_r(x,y)\int\dot\sigma(z_0)\dot\sigma(z_1)\,d\gamma_{r,x,y}(z)+\Sigma_{r+1}(x,y)Θr+1​(x,y)=Θr​(x,y)∫σ˙(z0​)σ˙(z1​)dγr,x,y​(z)+Σr+1​(x,y). Here σ˙(t)\dot\sigma(t)σ˙(t) is the ordinary real derivative when it exists and zero otherwise. All integrals are real Bochner integrals and are zero for nonintegrable integrands. No regularity, integrability, or positive-dimension assumptions are imposed. At d=0d=0d=0, the normalized empty sum is zero, giving Θ0=Σ0=β2\Theta_0=\Sigma_0=\beta^2Θ0​=Σ0​=β2. This definition contains no convergence assertion.

widths. Given natural numbers h,d,qh,d,qh,d,q, a vector w∈Nhw\in\mathbb N^hw∈Nh, and a natural layer index ℓ\ellℓ, this defines n0=dn_0=dn0​=d, nℓ=wℓ−1+1n_\ell=w_{\ell-1}+1nℓ​=wℓ−1​+1 for 1≤ℓ≤h1\leq\ell\leq h1≤ℓ≤h, and nℓ=qn_\ell=qnℓ​=q for ℓ≥h+1\ell\geq h+1ℓ≥h+1. The subtraction in ℓ−1\ell-1ℓ−1 is truncated natural-number subtraction, but the first branch already handles ℓ=0\ell=0ℓ=0. All hidden widths are therefore positive, even if some wi=0w_i=0wi​=0. The input width ddd and output width qqq may be zero. When h=0h=0h=0, www is the unique empty vector, n0=dn_0=dn0​=d, and every positive layer index has width qqq.

sequentialWidths. For every natural hhh, this defines a filter Fh\mathcal F_hFh​ on Nh\mathbb N^hNh. For h=0h=0h=0, it is the point filter at the unique empty vector, so an eventual predicate must hold at that vector. Recursively, a predicate PPP is eventual for Fh+1\mathcal F_{h+1}Fh+1​ precisely when there exists M∈NM\in\mathbb NM∈N such that for every last coordinate b≥Mb\geq Mb≥M, the predicate a↦P(a0,…,ah−1,b)a\mapsto P(a_0,\ldots,a_{h-1},b)a↦P(a0​,…,ah−1​,b) is eventual for Fh\mathcal F_hFh​. For positive hhh, the complete quantifier order is ∃Mh−1 ∀wh−1≥Mh−1 ∃Mh−2 ∀wh−2≥Mh−2 ⋯ ∃M0 ∀w0≥M0, P(w0,…,wh−1)\exists M_{h-1}\ \forall w_{h-1}\geq M_{h-1}\ \exists M_{h-2}\ \forall w_{h-2}\geq M_{h-2}\ \cdots\ \exists M_0\ \forall w_0\geq M_0,\ P(w_0,\ldots,w_{h-1})∃Mh−1​ ∀wh−1​≥Mh−1​ ∃Mh−2​ ∀wh−2​≥Mh−2​ ⋯ ∃M0​ ∀w0​≥M0​, P(w0​,…,wh−1​). Each MjM_jMj​ may depend on the already chosen later coordinates wj+1,…,wh−1w_{j+1},\ldots,w_{h-1}wj+1​,…,wh−1​. In particular, for h=2h=2h=2 the order is exactly ∃M1 ∀w1≥M1 ∃M0 ∀w0≥M0, P(w0,w1)\exists M_1\ \forall w_1\geq M_1\ \exists M_0\ \forall w_0\geq M_0,\ P(w_0,w_1)∃M1​ ∀w1​≥M1​ ∃M0​ ∀w0​≥M0​, P(w0​,w1​); for h=1h=1h=1, it is the ordinary eventual condition at infinity on N\mathbb NN. The last coordinate has the outermost eventual quantifier, and coordinate zero has the innermost one. No single bound uniform in all coordinates is required.

inputForWidths. Given natural h,d,qh,d,qh,d,q, hidden-width parameters w∈Nhw\in\mathbb N^hw∈Nh, and x∈Rdx\in\mathbb R^dx∈Rd, define widths n0=dn_0=dn0​=d, nℓ=wℓ−1+1n_\ell=w_{\ell-1}+1nℓ​=wℓ−1​+1 for 1≤ℓ≤h1\leq\ell\leq h1≤ℓ≤h, and nℓ=qn_\ell=qnℓ​=q thereafter. This returns the element of Rn0\mathbb R^{n_0}Rn0​ with coordinate iii equal to xix_ixi​. It only changes the indexing type using n0=dn_0=dn0​=d; no coordinate value changes. At d=0d=0d=0, it sends the unique empty input vector to the unique empty vector.

outputForWidths. Given natural h,d,qh,d,qh,d,q, hidden-width parameters w∈Nhw\in\mathbb N^hw∈Nh, and k<qk<qk<q, define n0=dn_0=dn0​=d, nℓ=wℓ−1+1n_\ell=w_{\ell-1}+1nℓ​=wℓ−1​+1 for 1≤ℓ≤h1\leq\ell\leq h1≤ℓ≤h, and nℓ=qn_\ell=qnℓ​=q for ℓ≥h+1\ell\geq h+1ℓ≥h+1. This returns the same integer kkk, now indexed below nh+1=qn_{h+1}=qnh+1​=q. It neither permutes nor numerically changes the output index. If q=0q=0q=0, no admissible argument kkk exists.

initializedOutput. For all natural h,N,d,qh,N,d,qh,N,d,q, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, real β\betaβ, inputs Xi∈RdX_i\in\mathbb R^dXi​∈Rd for i<Ni<Ni<N, hidden-width parameters w∈Nhw\in\mathbb N^hw∈Nh, and parameters θ\thetaθ for h+1h+1h+1 affine layers with widths n0=dn_0=dn0​=d, nℓ=wℓ−1+1n_\ell=w_{\ell-1}+1nℓ​=wℓ−1​+1 for 1≤ℓ≤h1\leq\ell\leq h1≤ℓ≤h, and nh+1=qn_{h+1}=qnh+1​=q, this returns a vector YθY_\thetaYθ​ in the Euclidean space indexed by pairs (i,k)(i,k)(i,k), i<Ni<Ni<N, k<qk<qk<q. Its coordinate is (Yθ)i,k=ah+1,k(Xi)(Y_\theta)_{i,k}=a_{h+1,k}(X_i)(Yθ​)i,k​=ah+1,k​(Xi​), where a0(x)=xa_0(x)=xa0​(x)=x and aℓ+1,j(x)=(∑r<nℓθℓ,j,ruℓ,r(x))/nℓ+βθℓ,j,⊥a_{\ell+1,j}(x)=(\sum_{r<n_\ell}\theta_{\ell,j,r}u_{\ell,r}(x))/\sqrt{n_\ell}+\beta\theta_{\ell,j,\bot}aℓ+1,j​(x)=(∑r<nℓ​​θℓ,j,r​uℓ,r​(x))/nℓ​​+βθℓ,j,⊥​, using u0,r(x)=xru_{0,r}(x)=x_ru0,r​(x)=xr​ and uℓ,r(x)=σ(aℓ,r(x))u_{\ell,r}(x)=\sigma(a_{\ell,r}(x))uℓ,r​(x)=σ(aℓ,r​(x)) for ℓ>0\ell>0ℓ>0. Each weight and bias is a separate real parameter. Conversion to Euclidean-space representation preserves the coordinates. The function does not itself sample θ\thetaθ or assert a distribution, measurability, or differentiability. For h=0h=0h=0, it evaluates one affine layer with no activation; for d=0d=0d=0, the first sum and its quotient by zero vanish. If N=0N=0N=0 or q=0q=0q=0, the result is the unique empty-coordinate vector. Inputs may repeat and β\betaβ is unrestricted.

outputGaussian. For all natural N,d,q,hN,d,q,hN,d,q,h, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, real β\betaβ, and inputs Xi∈RdX_i\in\mathbb R^dXi​∈Rd for i<Ni<Ni<N, this is a centered Gaussian measure on the Euclidean coordinates (i,k)(i,k)(i,k), i<Ni<Ni<N, k<qk<qk<q, constructed from the matrix S(i,k),(j,k′)=Σh(Xi,Xj)S_{(i,k),(j,k')}=\Sigma_h(X_i,X_j)S(i,k),(j,k′)​=Σh​(Xi​,Xj​) when k=k′k=k'k=k′, and zero otherwise. The scalar functions satisfy Σ0(x,y)=(∑r<dxryr)/d+β2\Sigma_0(x,y)=(\sum_{r<d}x_r y_r)/d+\beta^2Σ0​(x,y)=(∑r<d​xr​yr​)/d+β2 and Σs+1(x,y)=∫σ(z0)σ(z1) dγs,x,y(z)+β2\Sigma_{s+1}(x,y)=\int\sigma(z_0)\sigma(z_1)\,d\gamma_{s,x,y}(z)+\beta^2Σs+1​(x,y)=∫σ(z0​)σ(z1​)dγs,x,y​(z)+β2, where γs,x,y\gamma_{s,x,y}γs,x,y​ is centered Gaussian with covariance (Σs(x,x)Σs(x,y)Σs(y,x)Σs(y,y))\begin{pmatrix}\Sigma_s(x,x)&\Sigma_s(x,y)\\\Sigma_s(y,x)&\Sigma_s(y,y)\end{pmatrix}(Σs​(x,x)Σs​(y,x)​Σs​(x,y)Σs​(y,y)​) if positive semidefinite, and the point mass at zero otherwise; nonintegrable products have integral zero. The output measure is the law of S1/2GS^{1/2}GS1/2G, for standard Gaussian GGG, when SSS is positive semidefinite. If SSS is not positive semidefinite, it is instead the point mass at zero. Thus identifying SSS as its covariance requires positive semidefiniteness. No regularity of σ\sigmaσ, distinctness of inputs, or positive dimensions are assumed. For d=0d=0d=0, the initial normalized sum is zero. For N=0N=0N=0 or q=0q=0q=0, the measure is concentrated on the unique empty-coordinate vector.

Probability instance for outputGaussian. For all natural N,d,q,hN,d,q,hN,d,q,h, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, real β\betaβ, and inputs Xi∈RdX_i\in\mathbb R^dXi​∈Rd, the measure on coordinates (i,k)(i,k)(i,k), i<Ni<Ni<N, k<qk<qk<q, constructed with zero mean and matrix S(i,k),(j,k′)=1k=k′Σh(Xi,Xj)S_{(i,k),(j,k')}=\mathbf1_{k=k'}\Sigma_h(X_i,X_j)S(i,k),(j,k′)​=1k=k′​Σh​(Xi​,Xj​), has total mass one. Here Σ0(x,y)=(∑r<dxryr)/d+β2\Sigma_0(x,y)=(\sum_{r<d}x_r y_r)/d+\beta^2Σ0​(x,y)=(∑r<d​xr​yr​)/d+β2, and Σs+1(x,y)\Sigma_{s+1}(x,y)Σs+1​(x,y) is β2\beta^2β2 plus the Bochner integral of σ(z0)σ(z1)\sigma(z_0)\sigma(z_1)σ(z0​)σ(z1​) under the centered pair Gaussian with entries Σs(x,x),Σs(x,y),Σs(y,x),Σs(y,y)\Sigma_s(x,x),\Sigma_s(x,y),\Sigma_s(y,x),\Sigma_s(y,y)Σs​(x,x),Σs​(x,y),Σs​(y,x),Σs​(y,y). Non-positive-semidefinite pair matrices give point masses at zero; nonintegrable products give zero integrals. The output measure likewise uses the square root of SSS applied to a standard Gaussian when positive semidefinite, and a point mass otherwise. Consequently this probability assertion requires no separate positive-semidefiniteness or regularity hypothesis. Empty coordinate sets still give total mass one on a singleton; zero input dimension uses the division-by-zero value zero.

ntkBadProbability. For all natural h,N,d,qh,N,d,qh,N,d,q, arbitrary σ:R→R\sigma:\mathbb R\to\mathbb Rσ:R→R, real β,ε\beta,\varepsilonβ,ε, inputs Xi∈RdX_i\in\mathbb R^dXi​∈Rd for i<Ni<Ni<N, and w∈Nhw\in\mathbb N^hw∈Nh, define widths n0=dn_0=dn0​=d, nℓ=wℓ−1+1n_\ell=w_{\ell-1}+1nℓ​=wℓ−1​+1 for 1≤ℓ≤h1\leq\ell\leq h1≤ℓ≤h, and nh+1=qn_{h+1}=qnh+1​=q. Give all weights θℓ,j,r\theta_{\ell,j,r}θℓ,j,r​ and biases θℓ,j,⊥\theta_{\ell,j,\bot}θℓ,j,⊥​, for 0≤ℓ≤h0\leq\ell\leq h0≤ℓ≤h, j<nℓ+1j<n_{\ell+1}j<nℓ+1​, and r<nℓr<n_\ellr<nℓ​, independent N(0,1)\mathcal N(0,1)N(0,1) distributions, and call their index set III. For parameters η\etaη, define Fη(x)=ah+1F_\eta(x)=a_{h+1}Fη​(x)=ah+1​ by a0=xa_0=xa0​=x and aℓ+1,j=(∑r<nℓηℓ,j,ruℓ,r)/nℓ+βηℓ,j,⊥a_{\ell+1,j}=(\sum_{r<n_\ell}\eta_{\ell,j,r}u_{\ell,r})/\sqrt{n_\ell}+\beta\eta_{\ell,j,\bot}aℓ+1,j​=(∑r<nℓ​​ηℓ,j,r​uℓ,r​)/nℓ​​+βηℓ,j,⊥​, where u0,r=xru_{0,r}=x_ru0,r​=xr​ and uℓ,r=σ(aℓ,r)u_{\ell,r}=\sigma(a_{\ell,r})uℓ,r​=σ(aℓ,r​) for ℓ>0\ell>0ℓ>0. Put Tθ(x,y)k,k′=∑p∈ID(η↦Fη(x)k)(θ)[ep]D(η↦Fη(y)k′)(θ)[ep]T_\theta(x,y)_{k,k'}=\sum_{p\in I}D(\eta\mapsto F_\eta(x)_k)(\theta)[e_p]D(\eta\mapsto F_\eta(y)_{k'})(\theta)[e_p]Tθ​(x,y)k,k′​=∑p∈I​D(η↦Fη​(x)k​)(θ)[ep​]D(η↦Fη​(y)k′​)(θ)[ep​], where epe_pep​ is the coordinate unit vector and a nonexistent full scalar-output Fréchet derivative is zero. Define the deterministic comparison by Σ0(x,y)=(∑r<dxryr)/d+β2\Sigma_0(x,y)=(\sum_{r<d}x_r y_r)/d+\beta^2Σ0​(x,y)=(∑r<d​xr​yr​)/d+β2, Θ0=Σ0\Theta_0=\Sigma_0Θ0​=Σ0​, Σs+1(x,y)=∫σ(z0)σ(z1) dγs,x,y(z)+β2\Sigma_{s+1}(x,y)=\int\sigma(z_0)\sigma(z_1)\,d\gamma_{s,x,y}(z)+\beta^2Σs+1​(x,y)=∫σ(z0​)σ(z1​)dγs,x,y​(z)+β2, and Θs+1(x,y)=Θs(x,y)∫σ˙(z0)σ˙(z1) dγs,x,y(z)+Σs+1(x,y)\Theta_{s+1}(x,y)=\Theta_s(x,y)\int\dot\sigma(z_0)\dot\sigma(z_1)\,d\gamma_{s,x,y}(z)+\Sigma_{s+1}(x,y)Θs+1​(x,y)=Θs​(x,y)∫σ˙(z0​)σ˙(z1​)dγs,x,y​(z)+Σs+1​(x,y). Here γs,x,y\gamma_{s,x,y}γs,x,y​ is the centered pair Gaussian with matrix (Σs(x,x)Σs(x,y)Σs(y,x)Σs(y,y))\begin{pmatrix}\Sigma_s(x,x)&\Sigma_s(x,y)\\\Sigma_s(y,x)&\Sigma_s(y,y)\end{pmatrix}(Σs​(x,x)Σs​(y,x)​Σs​(x,y)Σs​(y,y)​), replaced by a point mass at zero if this matrix is not positive semidefinite; σ˙\dot\sigmaσ˙ is the derivative set to zero where nonexistent; and nonintegrable Bochner integrals are zero. The declared value is the extended nonnegative real measure, under the independent Gaussian parameter law, of {θ:∃i,j<N ∃k,k′<q, ε<∣Tθ(Xi,Xj)k,k′−1k=k′Θh(Xi,Xj)∣}\{\theta:\exists i,j<N\ \exists k,k'<q,\ \varepsilon<|T_\theta(X_i,X_j)_{k,k'}-\mathbf1_{k=k'}\Theta_h(X_i,X_j)|\}{θ:∃i,j<N ∃k,k′<q, ε<∣Tθ​(Xi​,Xj​)k,k′​−1k=k′​Θh​(Xi​,Xj​)∣}. The inequality is strict, all input and output pairs are included, and no indices must be distinct. No measurable-event assumption, regularity assumption on σ\sigmaσ, or positivity assumption on ε\varepsilonε is supplied; for arbitrary sets this uses the measure’s canonical outer-measure extension. The value lies between zero and one. If N=0N=0N=0 or q=0q=0q=0, the existential event is empty and the value is zero. If N,q>0N,q>0N,q>0 and ε<0\varepsilon<0ε<0, the event is the whole parameter space and the value is one. Empty sums and division by zero are zero, and h=0h=0h=0 permits the unique empty width vector. No limiting assertion is included.

Human review
  • Endorsed by Shuze Chen · Sep 26, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Minghui · Sep 26, 2026

    Confirmed by the mission captain (proposal self-audit).

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