OctoInformBeta

OpenAI preprint, September 2026 · §7. The range 2 < p < ∞

§7.1-§7.2 (setting and proof of Theorem 1.1): the ramp and its error: Pareto mixtures admit controlled retractions and scalar moment identities for p>2p>2

Dyadic ramps yield bounded gradient jumps.

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

The ramp construction concerns normalized differences of real functions on a finite vertex set, measured by a weighted LpL_p energy on edges. Let WW and EE be finite vertex and edge sets, with endpoint maps s,t:E→Ws,t:E\to W. Assume n=|W|≥2n=|W|\geq2, every vertex is incident to an edge, and every pair of distinct vertices is joined by an edge in either orientation. Give each edge a scale δe>0\delta_e>0 and a weight we>0w_e>0, with ∑ewe=1\sum_e w_e=1; write σ({e})=we\sigma(\{e\})=w_e for the edge probability measure.

The incidence matrix M:ℝW→ℝEM:\mathbb R^W\to\mathbb R^E and the normalized gradient A:ℝW→ℝEA:\mathbb R^W\to\mathbb R^E are the matrices in (1). The normalized-gradient space V=im⁡AV=\operatorname{im}A is a subspace of ℝE\mathbb R^E. Mea=𝟏{s(e)=a}−𝟏{t(e)=a},A=diag⁡(δe−1)M,(Ax)e=xs(e)−xt(e)δe.M_{ea}=\mathbf1_{\{s(e)=a\}}-\mathbf1_{\{t(e)=a\}}, \qquad A=\operatorname{diag}(\delta_e^{-1})M, \qquad (Ax)_e=\frac{x_{s(e)}-x_{t(e)}}{\delta_e}.(1) A retraction P:ℝE→ℝEP:\mathbb R^E\to\mathbb R^E onto VV is a matrix that fixes VV and factors through AA; its row ℓ1\ell_1 norm is denoted by ρe(P)\rho_e(P), as in (2). ℛ(A)={P:PA=A,P=AB for some B:ℝE→ℝW},ρe(P)=∑f∈E|Pef|.\mathcal R(A)=\{P:PA=A,\ P=AB\text{ for some }B:\mathbb R^E\to\mathbb R^W\}, \qquad \rho_e(P)=\sum_{f\in E}|P_{ef}|.(2)

For scalar thresholds aa and bb, the tail truncation Ta:ℝ→ℝT_a:\mathbb R\to\mathbb R is the map that retains a scalar rr when |r|>a|r|>a and otherwise vanishes; the band truncation Ba,b:ℝ→ℝB_{a,b}:\mathbb R\to\mathbb R retains it when a<|r|≤ba<|r|\leq b and otherwise vanishes.✦ The dyadic ramp Rℓ:ℝ→ℝR_\ell:\mathbb R\to\mathbb R, for an integer ℓ≥0\ell\geq0, is the sum of band truncations in (3).✦ Ta(r)=r𝟏{a<|r|},Ba,b(r)=r𝟏{a<|r|≤b},Rℓ(r)=∑k=0ℓ−1B2k,2k+ℓ(r).T_a(r)=r\mathbf1_{\{a<|r|\}},\qquad B_{a,b}(r)=r\mathbf1_{\{a<|r|\leq b\}},\qquad R_\ell(r)=\sum_{k=0}^{\ell-1}B_{2^k,2^{k+\ell}}(r).(3) The vector tail truncation TaT_a on ℝE\mathbb R^E is the map that applies the scalar tail truncation to each coordinate. For p>0p>0, the power intensity Λp\Lambda_p is the measure on the positive half-line with density pr−p−1pr^{-p-1}✦, and the Pareto law πp\pi_p is its restriction to (1,∞)(1,\infty): dΛp(r)=pr−p−1𝟏{r>0}dr,πp=Λp|(1,∞).d\Lambda_p(r)=pr^{-p-1}\mathbf1_{\{r>0\}}\,dr, \qquad \pi_p=\Lambda_p\big|_{(1,\infty)}.(4) Let II be a nonempty finite set, put N=|I|N=|I|, and choose vectors gi∈Vg_i\in V with ∑i∈I|gi(e)|p=1\sum_{i\in I}|g_i(e)|^p=1 for every edge ee. The uniform mixture Av⁡i∈Iηi\operatorname{Av}_{i\in I}\eta_i of measures ηi\eta_i on a common measurable space is the measure given by their arithmetic average; the Pareto vector mixture μ\mu is the measure on ℝE\mathbb R^E in (5), formed from gg and the law in (4).✦ The symbol #\# denotes pushforward. Av⁡i∈Iηi=1N∑i∈Iηi,μ=1N∑i∈I(r↦rgi)#πp.\operatorname{Av}_{i\in I}\eta_i=\frac1N\sum_{i\in I}\eta_i, \qquad \mu=\frac1N\sum_{i\in I}(r\mapsto rg_i)_\#\pi_p.(5) The coordinatewise ramp uℓ:ℝE→ℝEu_\ell:\mathbb R^E\to\mathbb R^E is the map that applies the scalar map in (3) on each edge, and bℓb_\ell is the scalar pp-energy under πp\pi_p. For p>2p>2, the ramp variance constant apa_p and the graph parameter HH are the numbers in (6). uℓ(v)e=Rℓ(ve),bℓ=∫|Rℓ(r)|pdπp(r),ap=p/(p−2)(1−2(2−p)/2)2,H=4log⁡n.u_\ell(v)_e=R_\ell(v_e),\qquad b_\ell=\int |R_\ell(r)|^p\,d\pi_p(r),\qquad a_p=\frac{p/(p-2)}{\bigl(1-2^{(2-p)/2}\bigr)^2},\qquad H=4\log n.(6)

The Claim

Theorem 1. For p,q>0p,q>0, a nonempty finite family gi∈ℝEg_i\in\mathbb R^E normalized by ∑i|gi(e)|p=1\sum_i|g_i(e)|^p=1 on every coordinate, and ℓ≥0\ell\geq0, the mixture in (5) satisfies the moment identity (7): ∫|Rℓ(ve)|qdμ(v)=1N∫|Rℓ(r)|qdπp(r)(e∈E).\int |R_\ell(v_e)|^q\,d\mu(v) =\frac1N\int |R_\ell(r)|^q\,d\pi_p(r) \qquad(e\in E).(7)✦ For every integer m≥1m\geq1, the even-length scalar ramp has the lower bound (8): plog⁡2mp+1≤∫|R2m(r)|pdπp(r).p\log2\,m^{p+1} \leq\int |R_{2m}(r)|^p\,d\pi_p(r).(8)✦

Suppose p>2p>2. There is Cp>0C_p>0 such that the following projection and retraction estimates hold for every integer ℓ≥0\ell\geq0. For the projection estimate, let η\eta be a probability measure on ℝE\mathbb R^E, let QQ be a real matrix on ℝE\mathbb R^E, and give EE nonnegative weights αe\alpha_e of sum one. Assume K≥1K\geq1, B≥0B\geq0, ρe(Q)≤K\rho_e(Q)\leq K for every ee, and Qv=vQv=v for η\eta-almost every vv. Assume also that T2k(ve)2T_{2^k}(v_e)^2 is integrable for every ee and 0≤k<2ℓ0\leq k<2\ell, with the second-moment bounds in (9): ∫T2k(ve)2dη(v)≤B2k(2−p),∑eαe∫(Q(T2k(v)))e2dη(v)≤4B2k(2−p)(0≤k<2ℓ).\begin{aligned} \int T_{2^k}(v_e)^2\,d\eta(v)&\leq B\,2^{k(2-p)},\\ \sum_e\alpha_e\int\bigl(Q(T_{2^k}(v))\bigr)_e^2\,d\eta(v) &\leq4B\,2^{k(2-p)} \qquad(0\leq k<2\ell). \end{aligned}(9) Then the projection error obeys (10): ∑eαe∫|(Quℓ(v))e−uℓ(v)e|pdη(v)≤CpKp−2ℓB.\sum_e\alpha_e\int \bigl|(Qu_\ell(v))_e-u_\ell(v)_e\bigr|^p\,d\eta(v) \leq C_pK^{p-2}\ell B.(10)✦

For the graph in (1) and the normalized family gi∈Vg_i\in V, with the positive weights and pairwise adjacency specified above, there is a retraction P∈ℛ(A)P\in\mathcal R(A) satisfying (11), with μ\mu and the constants from (5) and (6): ρe(P)≤H(e∈E),∫∑ewe|(Puℓ(v))e−uℓ(v)e|pdμ(v)≤CpHp−2ℓN,∑ewe∫(Puℓ(v))e2dμ(v)≤4apN,∫(Puℓ(v))e2dμ(v)≤H2apN(e∈E).\begin{aligned} \rho_e(P)&\leq H &&(e\in E),\\ \int\!\sum_e w_e\bigl|(Pu_\ell(v))_e-u_\ell(v)_e\bigr|^p\,d\mu(v) &\leq\frac{C_pH^{p-2}\ell}{N},\\ \sum_e w_e\int(Pu_\ell(v))_e^2\,d\mu(v)&\leq\frac{4a_p}{N},\\ \int(Pu_\ell(v))_e^2\,d\mu(v)&\leq\frac{H^2a_p}{N} &&(e\in E). \end{aligned}(11)✦

For each ε>0\varepsilon>0 there is Cp,ε>0C_{p,\varepsilon}>0, independent of the graph, its weights and the family gg, such that a probability measure ν\nu on ℝE\mathbb R^E satisfies the bounded gradient jump estimates (12): (−id⁡)#ν=ν,ν(V)=1,ν{z:|ze|≤H22ℓ for every e}=1,∑ewe|N∫|ze|pdν(z)−bℓ|≤εbℓ+Cp,εHp−2ℓ,∑eweN∫ze2dν(z)≤4ap,N∫ze2dν(z)≤H2ap(e∈E).\begin{aligned} (-\operatorname{id})_\#\nu&=\nu,\qquad \nu(V)=1,\\ \nu\{z:|z_e|\leq H2^{2\ell}\text{ for every }e\}&=1,\\ \sum_e w_e\left|N\int|z_e|^p\,d\nu(z)-b_\ell\right| &\leq\varepsilon b_\ell+C_{p,\varepsilon}H^{p-2}\ell,\\ \sum_e w_eN\int z_e^2\,d\nu(z)&\leq4a_p,\\ N\int z_e^2\,d\nu(z)&\leq H^2a_p\qquad(e\in E). \end{aligned}(12)✦

In the paper, the result reads: with Hpap=4log⁡(2n)H_{\mathrm{pap}}=4\log(2n) and scale parameter LL, for a vector vv of law νpap\nu_{\mathrm{pap}} and an integer ℓ≥1\ell\geq1, the coordinate truncations v≤a=(ve𝟏{|ve|≤a})ev_{\leq a}=(v_e\mathbf1_{\{|v_e|\leq a\}})_e give u=∑k=0ℓ−1(v≤2k+ℓ−v≤2k)u=\sum_{k=0}^{\ell-1}(v_{\leq2^{k+\ell}}-v_{\leq2^k}) and Z=PuZ=Pu, every edge has pp-energy b=plog⁡2(ℓp+2∑j=1ℓ−1jp)b=p\log2\bigl(\ell^p+2\sum_{j=1}^{\ell-1}j^p\bigr) of order ℓp+1\ell^{p+1}, the weighted second moment of ZZ is at most CpC_p and each coordinate second moment is at most CpHpap2C_pH_{\mathrm{pap}}^2, while |Ze|≤Hpap22ℓ|Z_e|\leq H_{\mathrm{pap}}2^{2\ell} and ∥Z−u∥Lp(σ⊗νpap)p≤CpHpapp−2(ℓ+1)\|Z-u\|_{L_p(\sigma\otimes\nu_{\mathrm{pap}})}^p\leq C_pH_{\mathrm{pap}}^{p-2}(\ell+1), giving relative error at most CpHpap1−2/p/ℓC_pH_{\mathrm{pap}}^{1-2/p}/\ell and the choice ℓ=⌈LHpap1−2/p⌉\ell=\lceil LH_{\mathrm{pap}}^{1-2/p}\rceil.

What Is Checked

The moment and scalar energy assertions concern the truncations and Pareto measures in (3) and (5); the projection assertion concerns a probability law fixed almost surely by a matrix with the tail bounds (9), and the retraction and jump assertions concern finite normalized-gradient spaces with positive edge scales and normalized positive weights, with every vertex incident to an edge and every pair of distinct vertices adjacent.