OctoInformBeta

OpenAI preprint, September 2026 · §3. Dyadic vector masses and common prices · §3

Proposition 3.3 [Common discounted prices]: On each cell, the price plus the bad price bounds in L1L^1 a majorant of the unfavorable cone parts of ∇ut\nabla u_t at neighborhood radii

Discounted prices of dyadic cells total at most CVCV..

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 gg be a real C1C^1 function on the plane ℂ\mathbb C whose gradient h=∇gh=\nabla g, a complex-valued function, is integrable and vanishes outside the closed disk |y|≤R|y|\le R, and let V=∫|h|V=\int|h|. For a real function GG on ℂ\mathbb C and a radius t>0t>0, the disk average of GG at xx is AtG(x)=1πt2∫B(x,t)G(y)dy,A_tG(x)=\frac{1}{\pi t^2}\int_{B(x,t)}G(y)\,dy,(1)✦ where B(x,t)B(x,t) is the open disk of radius tt about xx, and ut=Atgu_t=A_tg.

The question is how to charge the gradients ∇ut(x)\nabla u_t(x), at the radii that a maximal function of these averages selects, to the mass of hh on a dyadic grid, cell by cell, so that the charges of all cells at all scales add up to at most a constant times VV. Each cell receives a common discounted price, built from the cancellation of hh against piecewise bilinear hats from one scale to the next.

Masses and Cones

For a natural number NN, the box grid ΛN\Lambda_N is the set of the (2N+1)2(2N+1)^2 integer pairs q=(q1,q2)q=(q_1,q_2) with |q1|≤N|q_1|\le N and |q2|≤N|q_2|\le N.✦ The tent function is ψ(ξ)=max⁡(1−|ξ|,0)\psi(\xi)=\max(1-|\xi|,0) for real ξ\xi✦, and the hat of an index q∈ℤ2q\in\mathbb Z^2 at a scale δ>0\delta>0 is the function φδ,q(x)=ψ(Re⁡x/δ−q1)ψ(Im⁡x/δ−q2)\varphi_{\delta,q}(x)=\psi\bigl(\operatorname{Re}x/\delta-q_1\bigr)\, \psi\bigl(\operatorname{Im}x/\delta-q_2\bigr)(2)✦ on ℂ\mathbb C, the piecewise bilinear bump of height 11 at the grid point δq=δq1+iδq2\delta q=\delta q_1+i\delta q_2, supported on the closed square of side 2δ2\delta centered there. Levels are counted from a finest scale r>0r>0: level kk has scale 2kr2^kr, and its vertices are the indices q∈ΛNq\in\Lambda_N. The vector mass of a vector density h:ℂ→ℂh:\mathbb C\to\mathbb C at level kk and vertex qq is the complex number mk(q)=∫φ2kr,q(y)h(y)dym_k(q)=\int\varphi_{2^kr,q}(y)\,h(y)\,dy✦; its scalar mass is |mk(q)||m_k(q)|✦, and its direction θk(q)\theta_k(q) is the unit complex number mk(q)/|mk(q)|m_k(q)/|m_k(q)| when mk(q)≠0m_k(q)\ne0 and 11 when mk(q)=0m_k(q)=0.

A palette is a finite set of complex numbers of modulus 11, its labels, such that every point of the unit circle lies at distance less than 1/20001/2000 from some label✦; 𝒩\mathcal N is a palette fixed once and for all, with |𝒩||\mathcal N| labels.✦ For a label nn and a sign σ=±1\sigma=\pm1, the cone axis an(σ)=32n+σ2ina_n(\sigma)=\frac{\sqrt3}{2}\,n+\frac{\sigma}{2}\,in is the unit vector at angle σ⋅30∘\sigma\cdot30^\circ from nn✦, and the cone coordinate ℓn,σ\ell_{n,\sigma} is the real-linear functional on ℂ\mathbb C given by ℓn,σ(z)=13⟨n,z⟩+σ⟨in,z⟩,\ell_{n,\sigma}(z)=\tfrac{1}{\sqrt3}\langle n,z\rangle+\sigma\langle in,z\rangle,(3)✦ where ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle is the real inner product of ℂ=ℝ2\mathbb C=\mathbb R^2. The good sector of a label nn is the set of z∈ℂz\in\mathbb C with ℓn,1(z)≥0\ell_{n,1}(z)\ge0 and ℓn,−1(z)≥0\ell_{n,-1}(z)\ge0.✦ For a real number ss write s+=max⁡(s,0)s_+=\max(s,0); the unfavorable cone part of z∈ℂz\in\mathbb C for nn is the number Un(z)=(−⟨an(1),z⟩)++(−⟨an(−1),z⟩)+.U_n(z)=\bigl(-\langle a_n(1),z\rangle\bigr)_++\bigl(-\langle a_n(-1),z\rangle\bigr)_+ .(4) At level kk, a vertex qq is bad for nn when θk(q)\theta_k(q) lies outside the good sector of nn.✦

Defects and Prices

The interpolation coefficient of a coarse index pp and a fine index qq is βpq=ψ(q1/2−p1)ψ(q2/2−p2)\beta_{pq}=\psi(q_1/2-p_1)\,\psi(q_2/2-p_2).✦ The refinement coefficients of depth dd are the entries λpq(d)\lambda^{(d)}_{pq}, p,q∈ΛNp,q\in\Lambda_N, of the ddth power of the matrix (βpq)(\beta_{pq}) restricted to ΛN\Lambda_N: λpq(0)\lambda^{(0)}_{pq} is 11 if p=qp=q and 00 otherwise, and λpq(d+1)=∑w∈ΛNβpwλwq(d)\lambda^{(d+1)}_{pq}=\sum_{w\in\Lambda_N}\beta_{pw}\lambda^{(d)}_{wq}.✦

The density defects of a real density ff on ℂ\mathbb C are the numbers dj[f](q)d_j[f](q), for a level jj and a vertex q∈ΛNq\in\Lambda_N, given by d0[f](q)=12(∫φr,q|f|−|∫φr,qf|),dj+1[f](q)=12(∑s∈ΛNβqs|∫φ2jr,sf|−|∫φ2j+1r,qf|).\begin{aligned} d_0[f](q)&=\tfrac12\Bigl(\int\varphi_{r,q}\,|f|-\Bigl|\int\varphi_{r,q}\,f\Bigr|\Bigr),\\ d_{j+1}[f](q)&=\tfrac12\Bigl(\sum_{s\in\Lambda_N}\beta_{qs}\Bigl|\int\varphi_{2^jr,s}\,f\Bigr| -\Bigl|\int\varphi_{2^{j+1}r,q}\,f\Bigr|\Bigr). \end{aligned}(5)✦ The discount factor is ρ=2−α\rho=2^{-\alpha} with α=1/100\alpha=1/100.✦ The scalar price of a real density ff on ℂ\mathbb C at level kk and vertex q∈ΛNq\in\Lambda_N is the number Pk[f](q)P_k[f](q) that gathers the defects of the levels j≤kj\le k through the refinement coefficients, discounted by ρ\rho per level: Pk[f](q)=∑j=0kρk−j∑s∈ΛNλqs(k−j)dj[f](s).P_k[f](q)=\sum_{j=0}^{k}\rho^{\,k-j}\sum_{s\in\Lambda_N}\lambda^{(k-j)}_{qs}\,d_j[f](s).(6)✦ The common root price of a vector density hh at level kk and vertex q∈ΛNq\in\Lambda_N is the number Dk(q)D_k(q) that adds the scalar prices of the cone coordinates of hh over the palette and both signs: Dk(q)=∑n∈𝒩(Pk[ℓn,1∘h](q)+Pk[ℓn,−1∘h](q)).D_k(q)=\sum_{n\in\mathcal N}\Bigl(P_k[\ell_{n,1}\circ h](q)+P_k[\ell_{n,-1}\circ h](q)\Bigr).(7)✦

For a finite set S⊂ΛNS\subset\Lambda_N, the cell price at level kk is Dk(S)=∑q∈SDk(q)D_k(S)=\sum_{q\in S}D_k(q)✦, and the root mass Mk(S)M_k(S) is the sum of the scalar masses |mk(q)||m_k(q)| over q∈Sq\in S✦; for a label nn, the bad price Bkn(S)B^n_k(S) is the same sum taken over the vertices of SS that are bad for nn.✦ The roots of an index p∈ℤ2p\in\mathbb Z^2 are the vertices in the square window of half-width 2020 about pp, the set R(p)={q∈ΛN:|q−p|∞≤20}R(p)=\{q\in\Lambda_N:|q-p|_\infty\le20\}.✦

Cells and Radii

At a scale δ>0\delta>0, the physical cell with corner x0∈ℂx_0\in\mathbb C is the half-open square x0+[0,δ)2x_0+[0,\delta)^2, the set of points yy that the rescaling y↦(y−x0)/δy\mapsto(y-x_0)/\delta sends into the unit cell [0,1)2[0,1)^2.✦ The cell of index pp at level kk is Qk(p)=2krp+[0,2kr)2Q_k(p)=2^krp+[0,2^kr)^2, with corner at the grid point 2krp2^krp, and Gk,p(Y)=g(2krp+2krY)G_{k,p}(Y)=g(2^krp+2^kr\,Y) is gg read through the rescaling of that cell.

For a real function GG on ℂ\mathbb C, a point xx and a scale δ>0\delta>0, a neighborhood radius of GG at xx at scale δ\delta is a radius t∈[δ,2δ]t\in[\delta,2\delta] such that At′G(x)≤AtG(x)A_{t'}G(x)\le A_tG(x) for every t′>0t'>0 with |t′−t|≤δ/4|t'-t|\le\delta/4. The set of neighborhood radii of GG at xx at scale δ\delta is denoted IδG(x)I^G_\delta(x). An endpoint radius of GG at xx at scale δ\delta is a radius t∈[δ,2δ]t\in[\delta,2\delta] that is one of the endpoints δ\delta, 2δ2\delta or satisfies At′G(x)≤AtG(x)A_{t'}G(x)\le A_tG(x) for every t′∈[δ,2δ]t'\in[\delta,2\delta]. The set of endpoint radii of GG at xx at scale δ\delta is denoted JδG(x)J^G_\delta(x).

The Claim

Theorem 1. Let r>0r>0 and R≥0R\ge0, and let NN be a natural number with R/r+1≤NR/r+1\le N.

(a) If h:ℂ→ℂh:\mathbb C\to\mathbb C is integrable and vanishes outside the closed disk |y|≤R|y|\le R, then for every m≥0m\ge0 and all finite sets P0,…,PmP_0,\dots,P_m of cell indices in ℤ2\mathbb Z^2, ∑k=0m∑p∈PkDk(R(p))≤1681⋅2|𝒩|1−ρ∫|h|.\sum_{k=0}^{m}\sum_{p\in P_k}D_k\bigl(R(p)\bigr)\le\frac{1681\cdot2\,|\mathcal N|}{1-\rho} \int|h|.(8)✦

There is an absolute constant C>0C>0 with the following property. Let gg be a real C1C^1 function on ℂ\mathbb C whose gradient h=∇gh=\nabla g is integrable and vanishes outside the closed disk |y|≤R|y|\le R, let kk be a level, p∈ℤ2p\in\mathbb Z^2 a cell index and n∈𝒩n\in\mathcal N a label, and write δ=2kr\delta=2^kr, G=Gk,pG=G_{k,p}, D=Dk(R(p))D=D_k(R(p)), M=Mk(R(p))M=M_k(R(p)) and B=Bkn(R(p))B=B^n_k(R(p)).

(b) If ∇g\nabla g is bounded, there is a measurable F:ℂ→[0,∞]F:\mathbb C\to[0,\infty] with ∫Qk(p)F≤C(B+D)\int_{Q_k(p)}F\le C(B+D) such that Un(∇ut(y))≤F(y)for all y∈Qk(p) and t∈Iδg(y).U_n\bigl(\nabla u_t(y)\bigr)\le F(y)\qquad\text{for all }y\in Q_k(p)\text{ and } t\in I^g_\delta(y).(9)✦

(c) If ∇g\nabla g is bounded, there is a measurable F:ℂ→[0,∞]F:\mathbb C\to[0,\infty] with ∫Qk(p)F≤C(M+D)\int_{Q_k(p)}F\le C(M+D) such that, for all y∈Qk(p)y\in Q_k(p), |∇ut(y)|≤F(y) for t∈Jδg(y),|ut(y)−ut′(y)|δ≤F(y) for 0<t,t′≤4δ.|\nabla u_t(y)|\le F(y)\ \text{ for }t\in J^g_\delta(y),\qquad \frac{|u_t(y)-u_{t'}(y)|}{\delta}\le F(y)\ \text{ for }0<t,t'\le4\delta.(10)✦

(d) For each sign σ=±1\sigma=\pm1 and each set Q⊂ℂQ\subset\mathbb C of measure at most 11, there is a measurable F:ℂ→[0,∞]F:\mathbb C\to[0,\infty] with ∫QF≤Cδ−1(B+D)\int_QF\le C\delta^{-1}(B+D) such that (−⟨an(σ),∫B(X,t)∇G⟩)+≤F(X)for all |X|≤2 and t∈I1G(X).\Bigl(-\Bigl\langle a_n(\sigma),\int_{B(X,t)}\nabla G\Bigr\rangle\Bigr)_+\le F(X)\qquad \text{for all }|X|\le2\text{ and }t\in I^G_1(X).(11)✦

(e) For each set Q⊂ℂQ\subset\mathbb C of measure at most 11, there is a measurable F:ℂ→[0,∞]F:\mathbb C\to[0,\infty] with ∫QF≤Cδ−1(M+D)\int_QF\le C\delta^{-1}(M+D) such that |∫B(X,t)∇G|≤F(X)for all |X|≤2 and t∈J1G(X).\Bigl|\int_{B(X,t)}\nabla G\Bigr|\le F(X)\qquad\text{for all }|X|\le2\text{ and } t\in J^G_1(X).(12)✦

(f) If ∇g\nabla g is bounded, then for each set Q⊂ℂQ\subset\mathbb C of measure at most 11 there is a measurable F:ℂ→[0,∞]F:\mathbb C\to[0,\infty] with ∫QF≤Cδ−1(M+D)\int_QF\le C\delta^{-1}(M+D) such that |AtG(X)−At′G(X)|≤F(X)for all |X|≤2 and 0<t,t′≤4.\bigl|A_tG(X)-A_{t'}G(X)\bigr|\le F(X)\qquad\text{for all }|X|\le2\text{ and } 0<t,t'\le4.(13)✦

In the paper, the result reads: for p∈Vkp\in V_k, 1≤k≤m1\le k\le m, the cell QpQ_p has the price DQp=∑c∈𝒞∑q∈R(Qp)∑l=km2−α(l−k)∑s∈VlλqsdscD_{Q_p}=\sum_{c\in\mathcal C}\sum_{q\in R(Q_p)}\sum_{l=k}^{m}2^{-\alpha(l-k)} \sum_{s\in V_l}\lambda_{qs}\,d^c_s with α=1/100\alpha=1/100 and R(Qp)={q∈Vk:|q−p|∞≤20rk}R(Q_p)=\{q\in V_k:|q-p|_\infty\le20r_k\}, these prices satisfy ∑k∑pDQp≤4122(1−2−α)∑c∈𝒞∥c(∇g)∥1≤CV\sum_k\sum_pD_{Q_p}\le\frac{41^2}{2(1-2^{-\alpha})}\sum_{c\in\mathcal C}\|c(\nabla g)\|_1 \le CV, and for every such cell QQ and every n∈𝒩n\in\mathcal N, ∫Qsup⁡t∈Ik(x)∑±(−e±(n)⋅∇ut(x))+dx≤C(DQ+∑q∈R(Q),(q,n) badaq)\int_Q\sup_{t\in I_k(x)}\sum_\pm\bigl(-e_\pm(n)\cdot\nabla u_t(x)\bigr)_+\,dx\le C\bigl(D_Q+\sum_{q\in R(Q),\,(q,n)\text{ bad}}a_q\bigr) and ∫Q[sup⁡t∈Jk(x)|∇ut(x)|+rk−1vk(x)]dx≤C(DQ+∑q∈R(Q)aq)\int_Q\bigl[\sup_{t\in J_k(x)}|\nabla u_t(x)|+r_k^{-1}v_k(x)\bigr]\,dx\le C\bigl(D_Q+\sum_{q\in R(Q)}a_q\bigr), all constants absolute and independent of mm, where its levels k=1,…,mk=1,\dots,m run from the coarsest with rk=2−kr_k=2^{-k} and vertex sets VkV_k, 𝒞\mathcal C is the set of cone coordinates ℓn,±1\ell_{n,\pm1}, λqs\lambda_{qs} and dscd^c_s are refinement coefficients and the density defects of c(∇g)c(\nabla g), e±(n)=an(±1)e_\pm(n)=a_n(\pm1), aqa_q is the scalar mass at qq, Ik(x)I_k(x) and Jk(x)J_k(x) hold the neighborhood and endpoint radii, and vk(x)v_k(x) stands for the differences |ut(x)−ut′(x)||u_t(x)-u_{t'}(x)|.

What Is Checked

The claim is stated over the hats φδ,q\varphi_{\delta,q} on the box grid ΛN\Lambda_N and the vector masses mk(q)m_k(q), the fixed palette 𝒩\mathcal N with its cone axes and cone coordinates, the density defects and refinement coefficients, the scalar, root and cell prices discounted by ρ\rho, the root masses and bad prices over the roots R(p)R(p), the disk averages AtA_t with their neighborhood and endpoint radii, and the physical cells Qk(p)Q_k(p). The library also proves that a palette exists.✦