OctoInformBeta

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

Lemma 3.2 [Exact scalar pairing]: For an integrable hh of compact support, roots carry the hat coefficients and errors cost at most twice the defects, which telescope

Dyadic hat pieces split into constant-sign roots and zero-mass errors.

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 hh be an integrable real density on the plane, such as a scalar coordinate of the vector measure ∇gdx\nabla g\,dx of a real C1C^1 function gg of compact support, that vanishes outside the closed disk |x|≤R|x|\le R, and let V=∫|h|dxV=\int|h|\,dx be its total mass. Localize hh on dyadic grids of bilinear hats, at scales 2kr2^k r doubling from a finest scale r>0r>0, and record at each vertex the mass of hh under its hat. As the scale doubles, positive and negative mass under adjacent fine hats cancel in the coarse coefficient, and the sum of the absolute coefficients can fall below VV. The question is where that loss sits among the vertices and the levels, and whether each hat piece of hh can be traded for a density of one sign carrying its coefficient, the difference going into errors of mass zero whose L1L^1 cost is paid by the loss at their vertex.

Dyadic Hats and Defects

The tent function is the piecewise linear bump Λ\Lambda on ℝ\mathbb R of height one at the origin, Λ(t)=max⁡(1−|t|,0).\Lambda(t)=\max(1-|t|,\,0).(1)✦ At a scale t>0t>0, the product hat of an integer index p=(p1,p2)p=(p_1,p_2) is the real function φpt\varphi^t_p on ℂ\mathbb C given by φpt(x)=Λ(Re⁡xt−p1)Λ(Im⁡xt−p2),\varphi^t_p(x)=\Lambda\Bigl(\frac{\operatorname{Re}x}{t}-p_1\Bigr)\, \Lambda\Bigl(\frac{\operatorname{Im}x}{t}-p_2\Bigr),(2)✦ the piecewise bilinear bump of height one at the point tptp, supported on the closed square of side 2t2t centered there. Level k≥0k\ge0 has the dyadic scale 2kr2^k r, and for an integer index qq, called a vertex, φk,q=φq2kr\varphi_{k,q}=\varphi^{2^k r}_q is the hat of qq at that level. At every level, for a natural number NN, the vertices range over the square box QN={q∈ℤ2:|q1|≤N,|q2|≤N}Q_N=\{q\in\mathbb Z^2 : |q_1|\le N,\ |q_2|\le N\}(3)✦ of (2N+1)2(2N+1)^2 integer pairs.

The interpolation coefficient of a coarse index pp and a fine index qq is λpq=Λ(q12−p1)Λ(q22−p2),\lambda_{pq}=\Lambda\bigl(\tfrac{q_1}{2}-p_1\bigr)\,\Lambda\bigl(\tfrac{q_2}{2}-p_2\bigr),(4)✦ which is 11 when q=2pq=2p, 12\tfrac12 or 14\tfrac14 when qq differs from 2p2p by one in exactly one or in both coordinates, and 00 otherwise. The refinement coefficients λpq(k)\lambda^{(k)}_{pq} of depth kk, for p,q∈QNp,q\in Q_N, are the entries of the kkth power of the matrix (λpq)p,q∈QN(\lambda_{pq})_{p,q\in Q_N}: λpq(0)=δpq,λpq(d+1)=∑a∈QNλpaλaq(d).\lambda^{(0)}_{pq}=\delta_{pq},\qquad \lambda^{(d+1)}_{pq}=\sum_{a\in Q_N}\lambda_{pa}\,\lambda^{(d)}_{aq}.(5)✦

The coefficient of hh at the vertex qq of level kk is its mass under the hat, ck,q=∫φk,qhdx,c_{k,q}=\int\varphi_{k,q}\,h\,dx,(6) and the defect dk,qd_{k,q} is half the absolute mass the coefficient loses against what it merges: at level 00 against |h||h| under the hat, and at level k+1k+1 against the absolute coefficients of level kk weighted by interpolation, d0,q=12(∫φ0,q|h|dx−|c0,q|),dk+1,q=12(∑s∈QNλqs|ck,s|−|ck+1,q|).d_{0,q}=\tfrac12\Bigl(\int\varphi_{0,q}\,|h|\,dx-|c_{0,q}|\Bigr),\qquad d_{k+1,q}=\tfrac12\Bigl(\sum_{s\in Q_N}\lambda_{qs}\,|c_{k,s}|-|c_{k+1,q}|\Bigr).(7)

Roots and Errors

A cancellation scheme over a measure μ\mu on a space XX is given by a finite index set II, real weights wk(q,s)w_k(q,s) for each level k≥0k\ge0 and q,s∈Iq,s\in I, and real functions fqf_q on XX, the pieces, for q∈Iq\in I. The refined pieces Fk,qF_{k,q} of a cancellation scheme are the real functions on XX obtained by merging the pieces of the scheme level by level, F0,q=fq,Fk+1,q=∑s∈Iwk(q,s)Fk,s.F_{0,q}=f_q,\qquad F_{k+1,q}=\sum_{s\in I}w_k(q,s)\,F_{k,s}.(8)✦ The root g♯g^{\sharp} of an integrable real function gg with respect to μ\mu is the real function on XX of constant sign given by g♯=(∫Xgdμ/∫X|g|dμ)|g|g^{\sharp}=\bigl(\int_X g\,d\mu\big/\int_X|g|\,d\mu\bigr)\,|g| if ∫X|g|dμ>0\int_X|g|\,d\mu>0, and g♯=0g^{\sharp}=0 otherwise. The roots ρk,q\rho_{k,q} of a cancellation scheme are the real functions on XX with ρ0,q\rho_{0,q} the root of the piece fqf_q and ρk+1,q\rho_{k+1,q} the root of the weighted sum of the roots of level kk, and the errors βk,q\beta_{k,q} of the scheme are the real functions on XX given by the difference between the function whose root is taken and that root: ρ0,q=fq♯,ρk+1,q=(∑s∈Iwk(q,s)ρk,s)♯,\rho_{0,q}=f_q^{\sharp},\qquad \rho_{k+1,q}=\Bigl(\sum_{s\in I}w_k(q,s)\,\rho_{k,s}\Bigr)^{\sharp},(9)✦ β0,q=fq−ρ0,q,βk+1,q=∑s∈Iwk(q,s)ρk,s−ρk+1,q,\beta_{0,q}=f_q-\rho_{0,q},\qquad \beta_{k+1,q}=\sum_{s\in I}w_k(q,s)\,\rho_{k,s}-\rho_{k+1,q},(10)✦

The root and error densities ρk,q\rho_{k,q} and βk,q\beta_{k,q} of hh on the box QNQ_N are the roots and errors of the scheme over Lebesgue measure on ℂ\mathbb C with index set QNQ_N, the weights wk(q,s)=λqsw_k(q,s)=\lambda_{qs} at every level, and the pieces fq=φ0,qhf_q=\varphi_{0,q}h of hh under the hats of the finest level.

The Claim

Theorem 1. Let r>0r>0 and R≥0R\ge0, let NN be a natural number with R/r+1≤NR/r+1\le N, and let hh be an integrable real density on ℂ\mathbb C that vanishes outside the closed disk |x|≤R|x|\le R, with total mass VV, coefficients ck,qc_{k,q} and defects dk,qd_{k,q} as in (6) and (7), and root and error densities ρk,q\rho_{k,q} and βk,q\beta_{k,q} on the box QNQ_N. Then for every level k≥0k\ge0 and every vertex q∈QNq\in Q_N, φk,q(x)h(x)=ρk,q(x)+∑j=0k∑s∈QNλqs(k−j)βj,s(x)for every x∈ℂ,\varphi_{k,q}(x)\,h(x)=\rho_{k,q}(x) +\sum_{j=0}^{k}\sum_{s\in Q_N}\lambda^{(k-j)}_{qs}\,\beta_{j,s}(x) \qquad\text{for every }x\in\mathbb C,(11)✦ the root and the error at qq satisfy ∫ρk,qdx=ck,q,\int\rho_{k,q}\,dx=c_{k,q},(12)✦ ∫|ρk,q|dx=|ck,q|,\int|\rho_{k,q}|\,dx=|c_{k,q}|,(13)✦ ∫βk,qdx=0,\int\beta_{k,q}\,dx=0,(14)✦ ∫|βk,q|dx≤2dk,q,\int|\beta_{k,q}|\,dx\le 2\,d_{k,q},(15)✦ and ρk,q(x)=βk,q(x)=0for every x∈ℂ with φk,q(x)=0.\rho_{k,q}(x)=\beta_{k,q}(x)=0\qquad \text{for every }x\in\mathbb C\text{ with }\varphi_{k,q}(x)=0.(16)✦ For every n≥0n\ge0, ∑k=0n∑q∈QNdk,q=12(V−∑q∈QN|cn,q|).\sum_{k=0}^{n}\sum_{q\in Q_N}d_{k,q}=\frac12\Bigl(V-\sum_{q\in Q_N}|c_{n,q}|\Bigr).(17)✦ Moreover, in every cancellation scheme over a measure μ\mu on a space XX whose pieces are integrable, at every level kk and index qq with ∫XFk,qdμ≥0\int_X F_{k,q}\,d\mu\ge0, ρk,q(x)≥0for every x∈X.\rho_{k,q}(x)\ge0\qquad\text{for every }x\in X.(18)✦

In the paper, the result reads: for every c∈Cc\in C, the set of scalar coordinates cn,±c_{n,\pm}, there are absolutely continuous measures ρsc\rho^c_s, μs,+c\mu^c_{s,+} and μs,−c\mu^c_{s,-} at all vertices s∈Vls\in V_l, 1≤l≤m1\le l\le m, supported in supp⁡φs\operatorname{supp}\varphi_s, with ρsc\rho^c_s of constant sign, μs,±c≥0\mu^c_{s,\pm}\ge0, ρsc(ℝ2)=cs\rho^c_s(\mathbb R^2)=c_s, |ρsc|(ℝ2)=|cs||\rho^c_s|(\mathbb R^2)=|c_s| and μs,±c(ℝ2)=dsc\mu^c_{s,\pm}(\mathbb R^2)=d^c_s, the defects, such that βsc=μs,+c−μs,−c\beta^c_s=\mu^c_{s,+}-\mu^c_{s,-} satisfies βsc(ℝ2)=0\beta^c_s(\mathbb R^2)=0, |βsc|(ℝ2)≤2dsc|\beta^c_s|(\mathbb R^2)\le 2d^c_s and supp⁡βsc⊆supp⁡φs\operatorname{supp}\beta^c_s\subseteq\operatorname{supp}\varphi_s, and φqνc=ρqc+∑l=km∑s∈Vlλqsβsc\varphi_q\nu^c=\rho^c_q+\sum_{l=k}^{m}\sum_{s\in V_l}\lambda_{qs}\beta^c_s, with finitely many nonzero terms, for every q∈Vkq\in V_k, 1≤k≤m1\le k\le m, where VlV_l stands for the set of vertices of level ll, mm for the number of levels, φs\varphi_s and φq\varphi_q for the hats of the vertices ss and qq at their levels, νc\nu^c for hdxh\,dx, λqs\lambda_{qs} for the refinement coefficient of qq and ss, and the levels 1≤l≤m1\le l\le m are counted from the coarsest.

What Is Checked

The claim is stated over the tent function Λ\Lambda and the product hats φk,q\varphi_{k,q} at the dyadic scales 2kr2^k r, the box QNQ_N, the interpolation coefficients λpq\lambda_{pq} and the refinement coefficients λpq(k)\lambda^{(k)}_{pq}, the coefficients ck,qc_{k,q} and defects dk,qd_{k,q} of hh, and the refined pieces, roots and errors of cancellation schemes, among them the scheme with the hat pieces φ0,qh\varphi_{0,q}h and the interpolation weights that gives the root and error densities of hh.