OctoInformBeta

OpenAI preprint, September 2026 · §4. Directions that change slowly in scale and position · §4

Lemma 4.5 [Intrinsic marks and expected switches]: For integrable hh supported in the centered disk of radius (N−1)r(N-1)r, marked mass and expected label switches are at most constants times ∫|h|\int|h|

Mass bounds for intrinsic marks and expected switches.

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 h:ℂ→ℂh:\mathbb C\to\mathbb C be an integrable vector density on the plane, in the application the gradient ∇g\nabla g of an input function gg, and V=∫ℂ|h|V=\int_{\mathbb C}|h| its total mass. Tested against the bilinear hats of a dyadic grid, hh has a vector mass at each scale and each vertex, and a nonzero vector mass has a direction on the unit circle. Weighted along paths of vertices by transition weights, the vector masses at the finest scale form a forest of finite trees, one rooted at each vertex of a box. The item asks how often the directions turn along this forest and what the turning costs against VV: the mass at the nodes where the direction has moved by more than a threshold δ>0\delta>0, the energy of the local disagreement of directions, and the number of times a label taken from a finite palette of directions has to switch as particles run down the trees.

Trees, Marks and Particles

A finite branching tree tt with labels in a finite set ι\iota and values in a normed space XX is either a leaf carrying a vector of XX or a branch with one child tit_i, again such a tree, for each label i∈ιi\in\iota.✦ A node of tt is named by the finite sequence pp of labels that leads to it from the root, one child per label; |p||p| is the length of pp, p|jp|_j its prefix of length jj, and ipip the sequence pp preceded by ii. A valid path of tt is a sequence of labels that, followed from the root, ends exactly at a leaf.✦ An active path of tt is a sequence of labels ending at a leaf that carries a nonzero vector.✦

The total vector v(t)v(t) of tt is the sum of the vectors at its leaves✦, and the terminal mass ∥t∥\|t\| of tt is the sum of their norms.✦ The direction of a vector v∈Xv\in X with fallback e∈Xe\in X is dir⁡ev=v/∥v∥\operatorname{dir}_ev=v/\|v\| when v≠0v\neq0 and ee when v=0v=0.✦ Fix a unit vector e∈Xe\in X. At a node ξ\xi, vξv_\xi is the total vector of the subtree hanging from ξ\xi, Mξ=∥vξ∥M_\xi=\|v_\xi\| is its mass and ωξ=dir⁡evξ\omega_\xi=\operatorname{dir}_ev_\xi its direction.

Fix also a real number δ\delta and a reference direction n∈Xn\in X. The marks of tt from nn are found from the root down: a node ξ\xi is a mark when ∥ωξ−nξ∥>δ,\|\omega_\xi-n_\xi\|>\delta,(1) where nξn_\xi is the direction ωζ\omega_\zeta of the nearest mark ζ\zeta among the strict ancestors of ξ\xi, or nn if there is none. The mark indicator of tt from nn is the function on finite sequences of labels equal to 11 at a sequence ending at a mark and 00 elsewhere✦, and the first crossing cost of tt from nn is the sum of MξM_\xi over the marks of tt from nn.✦ The intrinsic marks of tt are the root of tt, named by the empty sequence ∅\emptyset, and, in the subtree of each child, the marks from the root direction ω∅\omega_\emptyset. The marked cost ℳδ(t)\mathcal M_\delta(t) of tt is the sum of MξM_\xi over its intrinsic marks✦, and the root mark indicator χδt(p)\chi^t_\delta(p) is 11 when pp ends at an intrinsic mark and 00 otherwise.✦

The node charge of tt under a real weight Γ\Gamma on finite sequences of labels is the sum ∑ξMξΓ(ξ)\sum_\xi M_\xi\,\Gamma(\xi) over all nodes ξ\xi of tt.✦ The path charge ΣbΓ(p)\Sigma_b\Gamma(p) of a sequence pp of labels from a birth level b≤|p|b\le|p| under Γ\Gamma is ΣbΓ(p)=∑j=b|p|Γ(p|j).\Sigma_b\Gamma(p)=\sum_{j=b}^{|p|}\Gamma(p|_j).(2)✦

The particles of tt run down its valid paths, each from a birth level. The survival fraction σ(t)\sigma(t) of a branch tt is the proportion of its children, one for each label, that survive. Let GG be a real function on finite sequences of labels and FF a real function of a sequence and a birth level; write G(i⋅)G(i\,\cdot) for the function p↦G(ip)p\mapsto G(ip), F(i⋅,1)F(i\,\cdot,1) for p↦F(ip,1)p\mapsto F(ip,1) and F(i⋅,⋅+1)F(i\,\cdot,\cdot+1) for (p,b)↦F(ip,b+1)(p,b)\mapsto F(ip,b+1). The root birth evaluation Πt∘(G)\Pi^\circ_t(G) of GG is ∥v∥G(∅)\|v\|\,G(\emptyset) when tt is a leaf carrying vv, and Πt∘(G)=σ(t)∑i∈ιΠti∘(G(i⋅))\Pi^\circ_t(G)=\sigma(t)\sum_{i\in\iota}\Pi^\circ_{t_i}\bigl(G(i\,\cdot)\bigr)(3)✦ when tt is a branch. The late birth evaluation Πt+(F)\Pi^+_t(F) of FF is 00 when tt is a leaf, and Πt+(F)=(1−σ(t))∑i∈ιΠti∘(F(i⋅,1))+∑i∈ιΠti+(F(i⋅,⋅+1))\Pi^+_t(F)=\bigl(1-\sigma(t)\bigr)\sum_{i\in\iota}\Pi^\circ_{t_i}\bigl(F(i\,\cdot,1)\bigr) +\sum_{i\in\iota}\Pi^+_{t_i}\bigl(F(i\,\cdot,\cdot+1)\bigr)(4)✦ when tt is a branch. The particle evaluation of FF is Πt(F)=Πt∘(F(⋅,0))+Πt+(F)\Pi_t(F)=\Pi^\circ_t\bigl(F(\cdot,0)\bigr)+\Pi^+_t(F).✦

The Dyadic Vector Forest

The tent Λ\Lambda is the function on ℝ\mathbb R given by Λ(x)=max⁡(1−|x|,0)\Lambda(x)=\max(1-|x|,0).✦ For a scale rr and an index p=(p1,p2)∈ℤ2p=(p_1,p_2)\in\mathbb Z^2, the bilinear hat φr,p\varphi_{r,p} is the function on ℂ\mathbb C given by φr,p(x)=Λ(Re⁡x/r−p1)Λ(Im⁡x/r−p2).\varphi_{r,p}(x)=\Lambda(\operatorname{Re}x/r-p_1)\,\Lambda(\operatorname{Im}x/r-p_2).(5)✦ The grid point of an index p∈ℤ2p\in\mathbb Z^2 at scale rr is the complex number rp=rp1+irp2rp=rp_1+\mathrm i\,rp_2.✦ The vector mass hr,ph_{r,p} and the scalar mass ar,pa_{r,p} of hh at pp are hr,p=∫ℂφr,p(y)h(y)dy,ar,p=|hr,p|.h_{r,p}=\int_{\mathbb C}\varphi_{r,p}(y)\,h(y)\,dy,\qquad a_{r,p}=|h_{r,p}|.(6) The dyadic scale of level jj is 2jr2^jr, level 00 being the finest scale rr.✦ The box grid GNG_N is the square index box of half-width NN, the set of p∈ℤ2p\in\mathbb Z^2 with |p1|≤N|p_1|\le N and |p2|≤N|p_2|\le N.

At a scale ss, the mass direction at a vertex q∈ℤ2q\in\mathbb Z^2 is θs(q)=dir⁡1hs,q\theta_s(q)=\operatorname{dir}_1h_{s,q}, the direction of the vector mass with fallback 11.✦ The window of radius dd around qq is the square Wd(q)W_d(q) of the (2d+1)2(2d+1)^2 vertices zz with |z1−q1|≤d|z_1-q_1|\le d and |z2−q2|≤d|z_2-q_2|\le d✦, and a maximizing vertex of a real function ff on ℤ2\mathbb Z^2 in Wd(q)W_d(q) is a vertex of the window at which ff attains its maximum over the window, chosen once and for all.✦ The chosen vertex zs*z^*_s of zz maximizes the scalar mass as,⋅a_{s,\cdot} in W300(z)W_{300}(z)✦, and the chosen direction at zz is Θs(z)=θs(zs*)\Theta_s(z)=\theta_s(z^*_s).✦ The worst vertex ws(q)w_s(q) of qq maximizes z↦|Θs(z)−θs(q)|2z\mapsto|\Theta_s(z)-\theta_s(q)|^2 in W200(q)W_{200}(q)✦, and the energy of qq at scale ss is Es(q)=|Θs(ws(q))−θs(q)|2=maxz∈W200(q)|Θs(z)−θs(q)|2.E_s(q)=|\Theta_s(w_s(q))-\theta_s(q)|^2=\max_{z\in W_{200}(q)}|\Theta_s(z)-\theta_s(q)|^2.(7)✦

Fix the finest scale rr and a depth kk. The dyadic vector tree TqT_q rooted at q∈GNq\in G_N is the finite branching tree with labels in GNG_N and values in ℂ\mathbb C whose nodes of depth less than kk are branches and whose leaf at a sequence (z1,…,zk)(z_1,\dots,z_k) carries β(z0,z1)β(z1,z2)⋯β(zk−1,zk)hr,zk,z0=q,\beta(z_0,z_1)\,\beta(z_1,z_2)\cdots\beta(z_{k-1},z_k)\,h_{r,z_k},\qquad z_0=q,(8)✦ where β(z,z′)\beta(z,z') is the transition weight from zz to z′z'; the trees TqT_q, q∈GNq\in G_N, form the dyadic vector forest. The last vertex ℓ(q,p)\ell(q,p) of a sequence pp in GNG_N is its last term, or qq when pp is empty✦, and the path energy of pp in TqT_q is Pq(p)=E2k−|p|r(ℓ(q,p)).P_q(p)=E_{2^{k-|p|}r}\bigl(\ell(q,p)\bigr).(9)✦

Labels and Switching

A finite circle palette is a finite set of points of the unit circle in ℂ\mathbb C such that every point of the circle lies at distance less than 1/20001/2000 from one of them.✦ The fixed palette 𝒫\mathcal P is one such palette chosen once and for all✦, and its points are the labels.✦

A law ν\nu on a finite set AA is a probability distribution on AA, a function ν≥0\nu\ge0 on AA with ∑a∈Aν(a)=1\sum_{a\in A}\nu(a)=1.✦ The expectation of a real function ff on AA under a law ν\nu on AA is 𝔼νf=∑a∈Aν(a)f(a)\mathbb E_\nu f=\sum_{a\in A}\nu(a)f(a).✦

For a tolerance τ\tau, the switch probability ramp ψτ\psi_\tau is the function of a distance dd given by ψτ(d)=min⁡{1,((d−τ)+τ)2}.\psi_\tau(d)=\min\Bigl\{1,\Bigl(\frac{(d-\tau)_+}{\tau}\Bigr)^2\Bigr\}.(10) At a scale ss and with τ=1/100\tau=1/100, the switch chance of a label nn at a vertex z∈ℤ2z\in\mathbb Z^2 is πs(n,z)=ψτ(|n−Θs(z)|)\pi_s(n,z)=\psi_\tau(|n-\Theta_s(z)|)✦, and the switching rate of nn at a point x∈ℂx\in\mathbb C is the tent interpolation of the switch chance, Rs(x,n)=∑p∈ℤ2φs,p(x)πs(n,p).R_s(x,n)=\sum_{p\in\mathbb Z^2}\varphi_{s,p}(x)\,\pi_s(n,p).(11)✦

Along a sequence of radii (ρj)j≥0(\rho_j)_{j\ge0}, the label marginal gives for each level jj and point x∈ℂx\in\mathbb C a law νj(x)\nu_j(x) on the labels: ν0(x)\nu_0(x) is the initial label law, a law on the labels attached to the scale ρ0\rho_0 and the point xx, and νj+1(x)=∑nνj(x)(n)Kρj+1,x(n,⋅)\nu_{j+1}(x)=\sum_n\nu_j(x)(n)\,K_{\rho_{j+1},x}(n,\cdot) is the mixture, with the weights of νj(x)\nu_j(x), of the scale transition rows Kρj+1,x(n,⋅)K_{\rho_{j+1},x}(n,\cdot), the laws on the labels that the scale ρj+1\rho_{j+1} and the point xx attach to each label nn.✦ The switching cost, the expected chance that the label switches on passing to the next level, is κ0(x)=1,κj+1(x)=𝔼νj(x)Rρj+1(x,⋅).\kappa_0(x)=1,\qquad\kappa_{j+1}(x)=\mathbb E_{\nu_j(x)}R_{\rho_{j+1}}(x,\cdot).(12)✦

For the forest, take ρj=2k−jr\rho_j=2^{k-j}r. The path switch total of a sequence pp from a birth level bb at a point xx is S(x;p,b)=∑i=b|p|κi(x).S(x;p,b)=\sum_{i=b}^{|p|}\kappa_i(x).(13)✦ The particle switch count Sq(p,b)S_q(p,b) of TqT_q is the path switch total at the grid point of the last vertex of pp, Sq(p,b)=S(rℓ(q,p);p,b)S_q(p,b)=S\bigl(r\,\ell(q,p);p,b\bigr).

The Claim

Theorem 1. Let h:ℂ→ℂh:\mathbb C\to\mathbb C be integrable with V=∫ℂ|h|V=\int_{\mathbb C}|h|, let N,k∈ℕN,k\in\mathbb N and r∈ℝr\in\mathbb R, and let TqT_q, q∈GNq\in G_N, be the dyadic vector trees of depth kk, with directions taken with fallback 11. For every δ>0\delta>0, ∑q∈GNℳδ(Tq)≤(1+2δ2)V.\sum_{q\in G_N}\mathcal M_\delta(T_q)\le\Bigl(1+\frac{2}{\delta^2}\Bigr)V.(14)✦ Suppose moreover that r>0r>0 and that hh vanishes outside the closed disk of radius R≥0R\ge0 about the origin, where R/r+1≤NR/r+1\le N. Then the mass-weighted path energies of the forest equal the mass-weighted sum of the energies of the vertices of GNG_N over the scales 2jr2^jr, 0≤j≤k0\le j\le k, ∑q∈GN∑ξ∈TqMξPq(ξ)=∑j=0k∑q∈GNa2jr,qE2jr(q);\sum_{q\in G_N}\sum_{\xi\in T_q}M_\xi\,P_q(\xi) =\sum_{j=0}^{k}\sum_{q\in G_N}a_{2^jr,q}\,E_{2^jr}(q);(15)✦ the path switch totals satisfy, for all q∈GNq\in G_N, all active paths pp of TqT_q, all birth levels b≤|p|b\le|p| and all points x∈ℂx\in\mathbb C with φr,ℓ(q,p)(x)≠0\varphi_{r,\ell(q,p)}(x)\neq0, S(x;p,b)≤2+2Σbχ1/800Tq(p)+2000000ΣbPq(p);S(x;p,b)\le2+2\,\Sigma_b\chi^{T_q}_{1/800}(p)+2000000\,\Sigma_bP_q(p);(16)✦ and the particle switch counts of the forest satisfy ∑q∈GNΠTq(Sq)≤(2560004+2000000⋅576⋅40972)V.\sum_{q\in G_N}\Pi_{T_q}(S_q)\le\bigl(2560004+2000000\cdot576\cdot4097^2\bigr)V.(17)✦ Finally, let tt be a finite branching tree with labels in a finite set ι\iota and values in a normed space XX, let e∈Xe\in X be a unit vector and δ>0\delta>0, with directions taken with fallback ee. Then Πt((p,b)↦Σbχδt(p))≤(1+2δ2)∥t∥,\Pi_t\bigl((p,b)\mapsto\Sigma_b\chi^t_\delta(p)\bigr)\le\Bigl(1+\frac{2}{\delta^2}\Bigr)\|t\|,(18)✦ and if a real number cc, a real weight Γ\Gamma on finite sequences of labels and a real function FF of a sequence and a birth level satisfy F(p,b)≤1+2Σbχδt(p)+cΣbΓ(p)F(p,b)\le1+2\,\Sigma_b\chi^t_\delta(p)+c\,\Sigma_b\Gamma(p) for every valid path pp of tt and every b≤|p|b\le|p|, then Πt(F)≤(3+4δ2)∥t∥+c∑ξ∈tMξΓ(ξ).\Pi_t(F)\le\Bigl(3+\frac{4}{\delta^2}\Bigr)\|t\|+c\sum_{\xi\in t}M_\xi\,\Gamma(\xi).(19)✦

In the paper, the result reads: ∑ξ intrinsically markedMξ≤CV\sum_{\xi\text{ intrinsically marked}}M_\xi\le CV, ∫∑s=BmEs(qs)dμ≤CV\int\sum_{s=B}^{m}E_s(q_s)\,d\mu\le CV and ∫∑s=BmRs(y)dμ≤CV\int\sum_{s=B}^{m}R_s(y)\,d\mu\le CV, where μ\mu is the measure on the particles, BB a particle’s birth level, qsq_s its position at level ss, yy its terminal position, mm the number of levels and CC a constant.

What Is Checked

The claim is stated over finite branching trees of vectors in a normed space, with their intrinsic marks, node and path charges and particle evaluations, and over the dyadic vector forest of an integrable density hh on the box grid GNG_N, with the energies of its mass and chosen directions and the switching costs of the label law on the fixed palette; the energy identity and the two switch bounds of the forest assume in addition that r>0r>0 and that hh vanishes outside the closed disk of radius RR about the origin, with R/r+1≤NR/r+1\le N. Beside the claim, the library shows that a finite circle palette exists✦ and that every real function on ℤ2\mathbb Z^2 attains its maximum on every window Wd(q)W_d(q).✦