OctoInformBeta

OpenAI preprint, September 2026 · Section 5. Pressure from weighted source layers

Section 5 caps, component weights and losses: Tent normalization and the associated nonnegative extended losses

Angular caps assign weights to source layer components.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Let E=ℝnE=\mathbb R^n be Euclidean space, let S={e∈E:|e|=1}S=\{e\in E:|e|=1\}, and let σ\sigma be its unnormalized surface measure. For positive powers p,qp,q and real weights A,BA,B, consider continuous, strictly positive functions u,vu,v on EE that are C2C^2 off the origin and satisfy the weighted elliptic system there. Its sources are −Δu=g,−Δv=f(x≠0),f(x)=|x|Bu(x)q,g(x)=|x|Av(x)p(x≠0),f(0)=g(0)=0.\begin{aligned} -\Delta u=g,\quad -\Delta v=f&\qquad(x\ne0),\\ f(x)=|x|^B u(x)^q,\quad g(x)=|x|^A v(x)^p&\qquad(x\ne0),\\ f(0)=g(0)&=0. \end{aligned}(1) Angular localization of the sources in (1) uses a radial cutoff weight and weights assigned to short portions of a source layer along a line.

The Cutoff and Line Geometry

For a real number rr, its positive part is r+=max⁡{r,0}r_+=\max\{r,0\}, and its positive-part power with integer exponent j≥0j\ge0 is r+jr_+^j.✦ The cutoff parameters m,k,Nm,k,N and radial profiles D,η0:ℝ→ℝD,\eta_0:\mathbb R\to\mathbb R are defined by m=n+3,k=n(m+1),N=k+n+2,D(r)=min⁡{1,(2−r)+},η0(r)=(1−(r−1)+3)+N.\begin{aligned} m&=n+3,& k&=n(m+1),& N&=k+n+2,\\ D(r)&=\min\{1,(2-r)_+\},& \eta_0(r)&=\bigl(1-(r-1)_+^3\bigr)_+^N. \end{aligned}(2) The boundary gap d:E→ℝd:E\to\mathbb R, compact cutoff η:E→ℝ\eta:E\to\mathbb R, and cutoff derivative weight w:E→ℝw:E\to\mathbb R associated with (2) are d(x)=D(|x|),η(x)=η0(|x|),w(x)=−Dη(x)[x].d(x)=D(|x|),\qquad \eta(x)=\eta_0(|x|),\qquad w(x)=-D\eta(x)[x].(3) The radial direction of a nonzero point x∈Ex\in E is the unit vector θx=x/|x|\theta_x=x/|x|.✦

Nonnegative extended quantities take values in [0,∞][0,\infty]; in an extended formula, r+r_+ denotes the finite extended value max⁡{r,0}\max\{r,0\}. For a∈[0,∞]a\in[0,\infty], write area^{\mathrm{re}} for its real value when a<∞a<\infty and for zero when a=∞a=\infty. The set infimum ℐI(h)\mathcal I_I(h) of a real function h:ℝ→ℝh:\mathbb R\to\mathbb R on I⊂ℝI\subset\mathbb R is the nonnegative extended number ℐI(h)=infs∈I(h(s))+.\mathcal I_I(h)=\inf_{s\in I}(h(s))_+.(4)✦

For a set XX, a subset U⊂X×ℝU\subset X\times\mathbb R, and a point z∈Xz\in X, an admissible rational endpoint rr is one whose entire segment from zero lies in the fiber of UU at zz: QU(z)={r∈ℚ:(z,s)∈U for every s∈[min⁡{0,r},max⁡{0,r}]}.Q_U(z)=\{r\in\mathbb Q:\ (z,s)\in U \text{ for every }s\in[\min\{0,r\},\max\{0,r\}]\}.(5)✦ For U⊂X×ℝU\subset X\times\mathbb R and z∈Xz\in X, the forward and backward exit distances aU+(z)a_U^+(z) and aU−(z)a_U^-(z), component length λU(z)\lambda_U(z), and component infimum ℐU(H;z)\mathcal I_U(H;z) of a real function H:X×ℝ→ℝH:X\times\mathbb R\to\mathbb R are the extended quantities aU+(z)=supr∈QU(z)r+,aU−(z)=sup−r∈QU(z)r+,λU(z)=aU+(z)+aU−(z),ℐU(H;z)=infr∈QU(z)(H(z,r))+.\begin{aligned} a_U^+(z)&=\sup_{r\in Q_U(z)}r_+,& a_U^-(z)&=\sup_{-r\in Q_U(z)}r_+,\\ \lambda_U(z)&=a_U^+(z)+a_U^-(z),& \mathcal I_U(H;z)&=\inf_{r\in Q_U(z)}(H(z,r))_+. \end{aligned}(6) For real functions H,RH,R on X×ℝX\times\mathbb R and δ∈ℝ\delta\in\mathbb R, the component weight 𝒞U(H,R;δ;z)\mathcal C_U(H,R;\delta;z) at z∈Xz\in X, defined from (6) for U⊂X×ℝU\subset X\times\mathbb R, is the nonnegative extended quantity 𝒞U(H,R;δ;z)={ℐU(H;z),(z,0)∈U,λU(z)<∞,λU(z)≤(2δ)+ℐU(R;z),0,otherwise.\mathcal C_U(H,R;\delta;z)= \begin{cases} \mathcal I_U(H;z),& (z,0)\in U,\quad \lambda_U(z)<\infty,\\ &\lambda_U(z)\le(2\delta)_+\mathcal I_U(R;z),\\ 0,&\text{otherwise}. \end{cases}(7)✦

The positive source subgraph 𝒮f⊂E×ℝ\mathcal S_f\subset E\times\mathbb R of a real function f:E→ℝf:E\to\mathbb R is the set in (8).✦ The layer measure νfμ\nu_f^\mu of ff for a measure μ\mu on EE is the restriction of product measure given by 𝒮f={(x,t):0<t<f(x)},νfμ=(μ⊗dt)|𝒮f.\mathcal S_f=\{(x,t):0<t<f(x)\},\qquad \nu_f^\mu=(\mu\otimes dt)|_{\mathcal S_f}.(8)✦

The Claim

Definition 1 (Caps, component weights, and losses). On a metric space TT with measure τ\tau, the tent cap profile Pθ,l:T→ℝP_{\theta,l}:T\to\mathbb R of center θ∈T\theta\in T and real width ll is the real function Pθ,l(e)=max⁡{1−dist⁡(e,θ)l,0}.P_{\theta,l}(e)=\max\left\{1-\frac{\operatorname{dist}(e,\theta)}{l},0\right\}.(9)✦ A quotient with zero denominator is assigned the value zero. The normalization factor Zτ(θ,l)Z_\tau(\theta,l) of the tent cap profile Pθ,lP_{\theta,l} on the measured metric space (T,τ)(T,\tau) is the real integral Zτ(θ,l)=∫TPθ,l(e)dτ(e).Z_\tau(\theta,l)=\int_T P_{\theta,l}(e)\,d\tau(e).(10)✦ The normalized cap weight Kτ(θ,l;⋅):T→ℝK_\tau(\theta,l;\cdot):T\to\mathbb R associated with the normalization factor in (10) is the real function Kτ(θ,l;e)=Pθ,l(e)Zτ(θ,l).K_\tau(\theta,l;e)=\frac{P_{\theta,l}(e)}{Z_\tau(\theta,l)}.(11)✦ The adaptive cap weight hκ:S×E→ℝh_\kappa:S\times E\to\mathbb R is indexed by κ∈ℝ\kappa\in\mathbb R; for 0<κ≤10<\kappa\le1 it satisfies the directional integral identity ∫Shκ(e,x)dσ(e)=w(x)(x∈E).\int_S h_\kappa(e,x)\,d\sigma(e)=w(x)\qquad(x\in E).(12)✦

For real functions h,ρh,\rho on ℝ\mathbb R, a real parameter δ\delta, and a set I⊂ℝI\subset\mathbb R of Lebesgue measure |I||I|, the interval component weight χI(h,ρ;δ)\chi_I(h,\rho;\delta) defined using (4) is the nonnegative extended number χI(h,ρ;δ)={ℐI(h),I≠⌀,|I|<∞,|I|≤(2δ)+ℐI(ρ),0,otherwise.\chi_I(h,\rho;\delta)= \begin{cases} \mathcal I_I(h),& I\ne\varnothing,\quad |I|<\infty,\quad |I|\le(2\delta)_+\mathcal I_I(\rho),\\ 0,&\text{otherwise}. \end{cases}(13)✦

The line layer UfU_f of a real source f:E→ℝf:E\to\mathbb R is the subset Uf={((e,x,t),s)∈(S×E×ℝ)×ℝ:0<t<f(x+se)}.U_f=\{((e,x,t),s)\in(S\times E\times\mathbb R)\times\mathbb R: 0<t<f(x+se)\}.(14)✦ Write Hκ((e,x,t),s)=hκ(e,x+se)H_\kappa((e,x,t),s)=h_\kappa(e,x+se) and R((e,x,t),s)=d(x+se)R((e,x,t),s)=d(x+se). The spatial line weight χf:S×E×ℝ→ℝ\chi_f:S\times E\times\mathbb R\to\mathbb R of a real source f:E→ℝf:E\to\mathbb R on the layer (14), with the component weight from (7), is the real function χf(e,x,t)=(𝒞Uf(Hκ,R;δ;(e,x,t)))re.\chi_f(e,x,t)= \bigl(\mathcal C_{U_f}(H_\kappa,R;\delta;(e,x,t))\bigr)^{\mathrm{re}}.(15)✦ For real functions u,fu,f on ℝn\mathbb R^n, the weighted layer loss Lf[u](e)L_f[u](e) in direction e∈Se\in S is the nonnegative extended integral Lf[u](e)=∫E×ℝ+((hκ(e,x)−χf(e,x,t))u(x))+dνfdx(x,t).L_f[u](e)=\int_{E\times\mathbb R}^{+} \bigl((h_\kappa(e,x)-\chi_f(e,x,t))u(x)\bigr)_+ \,d\nu_f^{dx}(x,t).(16)✦

In the paper, the result reads: under the reduced solution hypotheses, with normalized directional expectation 𝔼e\mathbb E_e, parameters 0<κ<10<\kappa<1 and 0<δ<1/40<\delta<1/4, and ℓ(x)=δd(x)m\ell(x)=\delta d(x)^m, choose a nonnegative smooth profile ψ≤1\psi\le1, equal to one near zero and supported in [0,1][0,1], set Zˆ(l)=𝔼eψ(|e−θ|/l)\widehat Z(l)=\mathbb E_e\psi(|e-\theta|/l) for any unit vector θ\theta, put hˆe(x)=w(x)ψ(|e−θx|/(κd(x)))/Zˆ(κd(x))\widehat h_e(x)=w(x)\psi(|e-\theta_x|/(\kappa d(x)))/ \widehat Z(\kappa d(x)) on 1<|x|<21<|x|<2 and zero elsewhere, so that 𝔼ehˆe=w\mathbb E_e\widehat h_e=w, assign to the component II of {f>t}\{f>t\} on x+ℝex+\mathbb Re containing xx the weight χˆI=inf⁡Ihˆe\widehat\chi_I=\inf_I\widehat h_e when II is bounded and |I|≤2δinf⁡Id|I|\le2\delta\inf_I d and zero otherwise, set χˆf(e,x,t)=χˆI\widehat\chi_f(e,x,t)=\widehat\chi_I and define χˆg\widehat\chi_g analogously, and define the losses Lˆf(e)=∫Eu(x)∫0f(x)(hˆe(x)−χˆf(e,x,t))dtdx\widehat L_f(e)=\int_E u(x)\int_0^{f(x)} (\widehat h_e(x)-\widehat\chi_f(e,x,t))\,dt\,dx and Lˆg(e)=∫Ev(y)∫0g(y)(hˆe(y)−χˆg(e,y,t′))dt′dy\widehat L_g(e)=\int_E v(y)\int_0^{g(y)} (\widehat h_e(y)-\widehat\chi_g(e,y,t'))\,dt'\,dy.

What Is Checked

The cap definitions (9) and (11) are stated on measured metric spaces, the component definitions (13) and (7) on arbitrary real functions and sets, and the spatial definitions (14), (15), and (16) on real functions on Euclidean space; the directional identity (12) assumes 0<κ≤10<\kappa\le1.