Alphabeta Math
LemmaStatement: AI-adaptedProof: AI-generatedPipeline-generatedprecheck pass
How statement and proof provenance work

The first chip identifies the source of the statement or construction; the second identifies the source of its local proof or verification.

  • Literature-sourced: the exact statement appears in a cited source; only wording and notation differ.
  • AI-adapted: a semantically identical restatement of literature-sourced material, modulo indexing, notation, and boundary cases adopted by the library.
  • AI-generated: a genuinely novel statement formulated by AI, with no source for the claim itself.

These labels describe origin, not correctness: citations and verification chips remain separate evidence.

Smoothed margin monotonicity along homogeneous Clarke flow

Statement

Let N,p≥1 and let q1,…,qN:Rp→R be locally Lipschitz functions that are positively homogeneous of common degree L>0 and admit the Clarke chain rule along absolutely continuous arcs. Set L(θ)=∑i=1Ne−qi(θ). Let θ:[0,∞)→Rp be a global Clarke gradient-flow trajectory for L, and suppose L(θ(t0))<1 for some t0≥0. For t≥t0, the loss is nonincreasing, the smoothed normalized margin γ~(θ(t))=−log⁡L(θ(t))/∥θ(t)∥2L is nondecreasing, L(θ(t))→0, and ∥θ(t)∥2→∞. Moreover, 0≤γˉ(θ(t))−γ~(θ(t))≤(log⁡N)/∥θ(t)∥2L, so the normalized margin and smoothed normalized margin have the same finite positive limit.

Facts & Assumptions

Given: The finite homogeneous margins, their local Lipschitz and Clarke chain-rule properties, the global Clarke flow, and a time t0 with loss below one.

[F1]

Clarke subdifferential and gradient-flow trajectory: a Clarke gradient-flow trajectory is locally absolutely continuous, and the subdifferential is the closed convex hull of limiting ordinary gradients, is nonempty compact convex, and has support function the Clarke generalized directional derivative.

[F2]

Homogeneous classifier margins and exponential training risk: qi(cθ)=cLqi(θ) and L=∑ie−qi.

[F3]

Euler identity for Clarke subgradients of homogeneous margins: every hi∈∂∘qi(θ) satisfies ⟨θ,hi⟩=Lqi(θ).

[F4]

Clarke chain rule for finite exponential empirical risk: for sample margins in a homogeneous-classifier model, along a Clarke flow, L′=−∥θ˙∥22 almost everywhere, and almost everywhere θ˙=∑ie−qihi with hi∈∂∘qi(θ).

[F5]

For a nonempty subset of Rn with n≥1, compactness, closedness and boundedness, pseudocompactness, and attainment of extrema by every continuous real-valued function are equivalent: continuous real-valued functions attain finite extrema on nonempty closed bounded subsets of finite-dimensional Euclidean space.

Proof

technique · use Euler's identity to differentiate the radius and exponential risk, then compare the weighted average margin with the log inverse loss
1.1F1F4F5givenconstructalgebra

We establish the identities of [F4] for the arbitrary margins allowed here. At any fixed parameter x, put ai=e−qi(x) and fi=e−qi. Since qi=−log⁡fi, the two functions have exactly the same differentiability points, and ∇fi=−e−qi∇qi there. Taking gradient limits and closed convex hulls in [F1] gives ∂∘fi(x)=−ai∂∘qi(x). Set C=∑i∂∘fi(x), a nonempty compact convex set. For any direction w, the finite-sum limsup inequality and [F1] give L∘(x;w)≤∑ifi∘(x;w)=max⁡c∈C⟨c,w⟩. Thus ∂∘L(x)⊆C: otherwise, for g∈∂∘L(x)∖C, let c0 minimize ∥g−c∥2 on C, using [F5]. Convexity gives ⟨g−c0,c−c0⟩≤0 for every c∈C, whereas ⟨g,g−c0⟩>⟨c0,g−c0⟩, contradicting the directional bound. Consequently the flow inclusion supplies θ˙=∑ie−qihi with hi∈∂∘qi(θ) almost everywhere. The assumed arc chain rules give L′=−⟨∑ie−qihi,θ˙⟩=−∥θ˙∥22 almost everywhere. All these compositions are locally absolutely continuous, since the margins are locally Lipschitz. No graph realization is needed.

2.1F1F2step 1.1givenalgebra

Put ϕ(t)=L(θ(t)), r(t)=∥θ(t)∥2, and z(t)=−log⁡ϕ(t). By [F1], θ is locally absolutely continuous; local Lipschitz continuity of the margins makes these compositions locally absolutely continuous on each compact time interval where r>0. Since ϕ(t0)<1 and [step 1.1] gives ϕ′=−∥θ˙∥22≤0 almost everywhere, ϕ(t)<1 for all t≥t0. Thus every qi(θ(t))>0. Homogeneity gives qi(0)=0, so θ(t)≠0 and r(t)>0.

3.1F3step 1.1step 2.1algebra

At almost every t≥t0, set ai=e−qi(θ(t)), pi=ai/ϕ, and μ=∑ipiqi(θ(t)). By [step 1.1], choose hi∈∂∘qi(θ(t)) with v:=θ˙=∑iaihi. Euler's identity [F3] yields ⟨θ,v⟩=Lϕμ, hence (log⁡r)′=Lϕμ/r2>0. Also qi=z−log⁡pi, so μ=z−∑ipilog⁡pi≥z>0 because pi>0, ∑ipi=1, and −log⁡pi≥0.

4.1F3step 1.1step 3.1algebra

The chain rule in [step 1.1] gives z′=∥v∥22/ϕ. Differentiating z/rL and applying Cauchy–Schwarz, γ~′=r−L(∥v∥22/ϕ−L2zϕμ/r2)≥L2ϕμ(μ−z)/rL+2≥0, because ∥v∥22≥⟨θ,v⟩2/r2=L2ϕ2μ2/r2. Thus the smoothed margin is nondecreasing.

5.1F2F5step 2.1step 4.1algebra

The function u↦qmin⁡(u):=min⁡iqi(u) is continuous on the unit sphere, so [F5] gives a finite maximum C. Homogeneity and ϕ≥e−qmin⁡ imply 0<γ~(t0)≤z/rL≤γˉ≤C. Since γ~ is nondecreasing by step 4.1, it has a finite positive limit. Writing A=∑ie−(qi−qmin⁡)∈[1,N] gives z=qmin⁡−log⁡A and 0≤γˉ−z/rL=log⁡A/rL≤(log⁡N)/rL.

6.1F2F5step 3.1step 5.1assume-contradischarge-contradiction∎

The nonincreasing loss has a limit ϕ∞≥0. If ϕ∞>0, then z≥z(t0)>0 and (r2)′=2Lϕμ≥2Lϕ∞z(t0)>0 almost everywhere by step 3.1, so r→∞ while z remains bounded. This forces z/rL→0, contradicting the positive lower bound in step 5.1. Hence ϕ→0. To show r→∞, suppose otherwise. Then some R<∞ and times tk→∞ satisfy r(tk)≤R. The closed ball of radius R is compact by [F5], and each continuous qi is bounded above there. Since N≥1, this gives a common positive lower bound for ϕ(tk)=∑ie−qi(θ(tk)), contradicting ϕ(t)→0. Thus r→∞. The error bound in step 5.1 now tends to zero, so the normalized and smoothed margins converge to the same positive limit.

Depends on

Used by

Dependency tree · two levels

13 results within two dependency steps of this one, each drawn at its shortest distance from it. An arrow runs from a result to what uses it, so the chart reads left to right and ends at this result, which carries a heavier outline. Every node is a link to that result. Click elsewhere on the chart to enlarge it.

Sources