OctoInformBeta

OpenAI preprint, September 2026 · §7. Likelihood cancellation and the forced binary subtree

Lemma 7.1: Coefficient energy and reference moments under typical metadata, with the scheduled level budget

Conditional cancellation bounds signed score moments.

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

A nested block has a binary orientation at every leaf. After the paired reports are observed, some of the hidden orientations remain fair bits. Averaging the leaf likelihood factors over these bits gives two conditional densities, one for each initial orientation. The library bounds the signed density through its coefficient energy and specializes the resulting estimate to the prescribed block schedule.

A finite probability law PP on a finite space XX is a nonnegative mass function with ∑x∈XP(x)=1\sum_{x\in X}P(x)=1.✦ For a finite probability law PP on a finite space XX and a real function ff on XX, the expectation of ff under PP is 𝔼Pf=∑xP(x)f(x)\mathbb E_P f=\sum_xP(x)f(x).✦ For finite probability laws PP on XX and QQ on YY, the product law P⊗QP\otimes Q is the finite probability law on the Cartesian product of their spaces with mass P(x)Q(y)P(x)Q(y)✦, and the coordinate product of finite laws PjP_j is the finite probability law with mass ∏jPj(xj)\prod_jP_j(x_j)✦; P⊗mP^{\otimes m} denotes the product of mm copies of PP.✦ All finite slots below are numbered from zero.

A paired mechanism on a side outcome space SS consists, at each trial jj, of finite laws Fj,Gj,HjF_j,G_j,H_j and a weight vj≥0v_j\geq0 satisfying vjHj(x)≤Fj(x),Gj(x)v_jH_j(x)\leq F_j(x),G_j(x).✦ A report zj=(dj,xj,yj)z_j=(d_j,x_j,y_j) contains an optional orientation flag dj∈{⊥,0,1}d_j\in\{\bot,0,1\} and two chronologically ordered side outcomes. The absent flag ⊥\bot records a jointly hidden pair. For a single trial with initial orientation τ∈{0,1}\tau\in\{0,1\}, the hidden mass, oriented pair law, flagged mass and report mass are respectively the functions in (1): hv(x,y)=v2H(x)H(y),O0=F⊗G,O1=G⊗F,fτ(d,x,y)={hv(x,y),d=⊥,Oτ(x,y)−hv(x,y),d=τ,0,d∈{0,1}\{τ},ρ(d,x,y)=12∑τ=01fτ(d,x,y).\begin{aligned} h_v(x,y)&=v^2H(x)H(y),& O_0&=F\otimes G,& O_1&=G\otimes F,\\ f_\tau(d,x,y)&= \begin{cases} h_v(x,y),&d=\bot,\\ O_\tau(x,y)-h_v(x,y),&d=\tau,\\ 0,&d\in\{0,1\}\setminus\{\tau\}, \end{cases} &\rho(d,x,y)&=\tfrac12\sum_{\tau=0}^1f_\tau(d,x,y). \end{aligned}(1) The joint report law ρM\rho_M of paired mechanisms MM is the coordinate product of their paired report laws.✦ The reported orientation posterior is a binary probability law that fixes a visible flag and assigns probability 1/21/2 to each orientation at an absent flag.✦ The joint orientation posterior Πz\Pi_z for a report array zz is the coordinate product of the reported orientation posteriors of its trials.

The Complete Window Tree

For child counts T(j)T(j) and local and global trial counts n(j),g(j)n(j),g(j), the complete window tree of depth p+1p+1 has T(p)T(p) children of depth pp. A window leaf index in a complete window tree is a sequence of child choices along a path to a leaf; at depth zero there is one leaf.✦ A window trial index names a root local trial, a root global trial, or a trial below one of the children; at depth zero there are no trials.✦ A window assignment on a complete window tree with values in a set AA is a function assigning a value in AA to every window trial index, including those on unoccupied paths and paths that do not fit.✦ The leaf and trial counts λp,mp\lambda_p,m_p, and the scheduled global counts, are specified by (2): λ0=1,λp+1=T(p)λp,m0=0,mp+1=n(p)+g(p)+T(p)mp,g(0)=0,g(j+1)=⌊T(j+1)/4⌋n(j).\begin{aligned} \lambda_0&=1,&\lambda_{p+1}&=T(p)\lambda_p,\\ m_0&=0,&m_{p+1}&=n(p)+g(p)+T(p)m_p,\\ g(0)&=0,&g(j+1)&=\lfloor T(j+1)/4\rfloor n(j). \end{aligned}(2) Nesting levels are numbered from the bottom. The assignment bijection reads values at their window trial indices.✦ Enumerating the finite trial set then identifies nested assignments with arrays of length mpm_p✦; the inverse places the flags of a report array into the nested tree.✦ Leaf arrays use the analogous finite enumeration of the window leaf indices.

The scheduled displacement Dj(u)D_j(u) at child uu is zero for j=0j=0. At level j+1j+1, global trials are grouped into blocks vv of n(j)n(j) trials, and the trial in within-block slot aa has global index vn(j)+avn(j)+a.✦ Let ξ\xi be a complete binary orientation assignment, with orientations treated as the integers zero and one. The net global displacement and the local center cjc_j are given by (3): Dj+1(u)=∑0≤v<⌊T(j+1)/4⌋v≤u,u+v<T(j+1)(∑a=0n(j)−1ξvn(j)+a−⌊n(j)/2⌋),cj=⌊T(j)/2⌋−1+⌊n(j)/2⌋.\begin{aligned} D_{j+1}(u)&= \sum_{\substack{0\leq v<\lfloor T(j+1)/4\rfloor\\v\leq u,\ u+v<T(j+1)}} \left(\sum_{a=0}^{n(j)-1}\xi_{vn(j)+a}-\lfloor n(j)/2\rfloor\right),\\ c_j&=\lfloor T(j)/2\rfloor-1+\lfloor n(j)/2\rfloor. \end{aligned}(3) Fix a nonnegative motif scale ss. The recursive leaf orientation starts at an initial orientation τ\tau. At a nonleaf it passes the integer template orientation at the node anchor to the chosen child, with that child’s scheduled displacement as incoming shift. At depth zero it returns the incoming orientation. With centers cjc_j, zero incoming shift at the root, and a flattened complete assignment ξ\xi, write θτ(ξ)\theta_\tau(\xi) for the flattened leaf orientation array.✦

For centers cjc_j, child displacements Dj(h)D_j(h) and sets AjA_j of child indices, central motif support is the following recursive condition on a complete binary assignment: at a node of depth j+1j+1 with incoming shift uu, every child hh at which the scale-ss motif indicator Is(h−cj+∑aξa+u)I_s(h-c_j+\sum_a\xi_a+u) is nonzero belongs to AjA_j, and every child satisfies the same condition with incoming shift Dj(h)D_j(h).✦ The depth-zero condition is vacuous. The central third is Aj={h<T(j):T(j)≤3h≤2T(j)}A_j=\{h<T(j):T(j)\leq3h\leq2T(j)\}.✦

The window posterior law for nested optional flags dd is the probability law on complete assignments that fixes reported orientations and makes all unreported orientations independent fair bits, with each global orientation drawn once.✦ It is the point mass on the empty assignment at depth zero and the product of the local, global and child posterior laws at positive depth.✦ For specified centers, displacements and sets of child indices, the noncentral posterior mass νd(u)\nu_d(u) is the probability under this law that central motif support fails with incoming shift uu.✦ For a report zz, write ν(z)\nu(z) for this mass using its nested flags, the local source centers, scheduled displacements, central thirds and zero root shift.

A paired report zz at depth b+1b+1 is typical metadata for common weight ω\omega and fitting threshold aa when it has at least n(j)ω2/2n(j)\omega^2/2 hidden local trials and g(j)ω2/2g(j)\omega^2/2 hidden global trials at every node of depth j+1j+1, and satisfies ν(z)<a\nu(z)<a.✦ Write 𝒯\mathcal T for this set of reports and 𝒰\mathcal U for the set of pairs (z,x)(z,x) with z∈𝒯z\in\mathcal T and an arbitrary complete leaf sample xx.✦

Coefficients and Budgets

For a binary leaf array θ\theta and a selection σ\sigma of leaf slots, the selection coefficient cθ(σ)c_\theta(\sigma) is the inclusion indicator that every selected slot has orientation one.✦ For a leaf law PP on a finite space XX, a real score Z:X→ℝZ:X\to\mathbb R and a leaf sample xx, the selection density fθ(Z,x)f_\theta(Z,x) multiplies 1+Z(xj)1+Z(x_j) over the slots of orientation one.✦ For a map θ\theta from complete assignments ξ\xi to binary leaf arrays, the averaged selection coefficient and selection density under a finite law RR on assignments are defined in (4): cθ(σ)=∏j∈σθj,fθ(Z,x)=∏j:θj=1(1+Z(xj)),c¯R(σ)=𝔼Rcθ(ξ)(σ),f¯R(Z,x)=𝔼Rfθ(ξ)(Z,x).\begin{aligned} c_\theta(\sigma)&=\prod_{j\in\sigma}\theta_j,& f_\theta(Z,x)&=\prod_{j:\theta_j=1}(1+Z(x_j)),\\ \overline c_R(\sigma)&=\mathbb E_R c_{\theta(\xi)}(\sigma),& \overline f_R(Z,x)&=\mathbb E_R f_{\theta(\xi)}(Z,x). \end{aligned}(4) The source oracle coefficients cτ,zc_{\tau,z} and densities gτ(z,x)g_\tau(z,x) are these averages with R=ΠzR=\Pi_z and θ=θτ\theta=\theta_\tau at depth b+1b+1. The signed source coefficient, weighted coefficient energy at a real scale ww, and source reference law are defined by (5); the coefficient sum uses counting measure, whereas reference moments use the probability law 𝖰\mathsf Q. Az(σ)=c0,z(σ)−c1,z(σ),ℰw(A)=∑σ⊆{0,…,λb+1−1}A(σ)2w2|σ|,𝖰=ρM⊗P⊗λb+1.\begin{aligned} A_z(\sigma)&=c_{0,z}(\sigma)-c_{1,z}(\sigma),\\ \mathcal E_w(A)&=\sum_{\sigma\subseteq\{0,\ldots,\lambda_{b+1}-1\}} A(\sigma)^2w^{2|\sigma|},\\ \mathsf Q&=\rho_M\otimes P^{\otimes\lambda_{b+1}}. \end{aligned}(5)

The hidden binomial cap hω(m)h_\omega(m) and hidden smoothing bound Hs,ω(m)H_{s,\omega}(m) are the functions in (6): hω(m)=4mω2+2,us,ω(m)=4(s+1)mω2+2,Hs,ω(m)=2s2(2us,ω(m))sus,ω(m).\begin{aligned} h_\omega(m)&=\sqrt{\frac4{m\omega^2+2}},& u_{s,\omega}(m)&=\sqrt{\frac{4(s+1)}{m\omega^2+2}},\\ H_{s,\omega}(m)&=2^{s^2}\bigl(2u_{s,\omega}(m)\bigr)^s u_{s,\omega}(m). \end{aligned}(6)

A finite ordered role shape of depth pp and degree bound rr is an ordered rooted tree with at most rr children per internal vertex and a role in {0,1,2}\{0,1,2\} on each child; there is one shape at depth zero.✦ The shape and role count S(r,p)S(r,p) is the number of finite ordered role shapes of depth pp and degree bound rr. For level costs CjC_j, the binary tree budget 𝔅(C,p)\mathfrak B(C,p) has 𝔅(C,0)=1\mathfrak B(C,0)=1 and 𝔅(C,p+1)=Cp𝔅(C,p)2\mathfrak B(C,p+1)=C_p\mathfrak B(C,p)^2.✦

Put ds=2sd_s=2^s and x+=max⁡{x,0}x_+=\max\{x,0\}. The interior label budget Ib,rI_{b,r}, source sparsity count 𝒩b\mathcal N_b, and forced contribution budget Jb,rJ_{b,r} are given by (7): Ib,r=(∏j=1bT(j))(r−2b+1)+,𝒩b=ds∏j=1bT(j),Cjtree=4ds6hω(g(j+1))hω(n(j)),Jb,r=S(r,b+1)Ib,rds2(r−2b+1)+(2ds3hω(n(b)))𝔅(Ctree,b).\begin{aligned} I_{b,r}&=\left(\prod_{j=1}^{b}T(j)\right)^{(r-2^{b+1})_+},& \mathcal N_b&=d_s\prod_{j=1}^{b}T(j),\\ C_j^{\mathrm{tree}}&=4d_s^6h_\omega(g(j+1))h_\omega(n(j)),\\ J_{b,r}&=S(r,b+1)I_{b,r}d_s^{2(r-2^{b+1})_+} \bigl(2d_s^3h_\omega(n(b))\bigr) \mathfrak B(C^{\mathrm{tree}},b). \end{aligned}(7) For cutoff RR, smoothing allowance η\eta, fitting threshold aa and score weight ww, the sharp degree budget Db,rD_{b,r} and sharp energy budget 𝔈(T,n,s,b,R;ω,η,a,w)\mathfrak E(T,n,s,b,R;\omega,\eta,a,w) are defined by (8): Db,r=4⋅3(b+1)r[(λb+1r)3(b+1)r(η2+a2)+𝟏{2b+1≤r}Jb,r],𝔈(T,n,s,b,R;ω,η,a,w)=∑r=0R−1w2r𝟏{r>0}Db,r+4(w2𝒩b)R1−w2𝒩b.\begin{aligned} D_{b,r}&=4\cdot3^{(b+1)r} \left[\binom{\lambda_{b+1}}{r}3^{(b+1)r}(\eta^2+a^2) +\mathbf1_{\{2^{b+1}\leq r\}}J_{b,r}\right],\\ \mathfrak E(T,n,s,b,R;\omega,\eta,a,w) &=\sum_{r=0}^{R-1}w^{2r}\mathbf1_{\{r>0\}}D_{b,r} +\frac{4(w^2\mathcal N_b)^R}{1-w^2\mathcal N_b}. \end{aligned}(8)

The Block Schedule

For maximum dimension BB and block count DD, the warmup count is 5⋅2B5\cdot2^B.✦ The dimension bi+1b_{i+1} belongs to zero-based block index ii, hence to block i+1i+1. Floors and ceilings with subscript ++ are truncated at zero. The block gain β\beta, total score exponent djd_j, error allowance ε\varepsilon, degree cutoff RR, work size WW and cancellation order ss are specified in (9): bi+1={2,i<5⋅2B,B,i≥5⋅2B,β(b)=b−14−120,dj=115+∑i=0j−1β(bi+1),ε=124B+100(D+1),R=210B+40,W=dD+R+B+10,s=⌈100Wε⌉+.\begin{aligned} b_{i+1}&=\begin{cases}2,&i<5\cdot2^B,\\B,&i\geq5\cdot2^B,\end{cases}& \beta(b)&=\frac{b-1}{4}-\frac1{20},\\ d_j&=\frac{11}{5}+\sum_{i=0}^{j-1}\beta(b_{i+1}),& \varepsilon&=\frac1{2^{4B+100}(D+1)},\\ R&=2^{10B+40},&W&=d_D+R+B+10,\\ s&=\left\lceil\frac{100W}{\varepsilon}\right\rceil_+. \end{aligned}(9) The source moment envelope MM, weight envelope VV and gap lower bound γ\gamma are the constants in (10): M=(2⋅108210B(B+4)+1)(D+1)2BD,V=144M2,γ=1128⋅22B.\begin{aligned} M&=\bigl(2\cdot10^8\,2^{10B}(B+4)+1\bigr)(D+1)2^{BD},\\ V&=144M^2,\qquad \gamma=\frac1{128\cdot2^{2B}}. \end{aligned}(10) For target scale NN and block dimension bb, the scale exponents AjA_j, window exponents κj\kappa_j, integer window sizes TjwinT_j^{\mathrm{win}} and extended block windows T(j)T(j) are defined in (11). Aj=(1+ε)2j2b−ε,κj=A(j−1)+−3(2j−1)ε,Tjwin=4⌊Nκj+1/4⌋+,T(j)={Tjwin,j<b,24,j≥b.\begin{aligned} A_j&=(1+\varepsilon)\frac{2^j}{2^b}-\varepsilon,& \kappa_j&=A_{(j-1)_+}-3(2^j-1)\varepsilon,\\ T_j^{\mathrm{win}}&=4\left\lfloor N^{\kappa_{j+1}}/4\right\rfloor_+,& T(j)&=\begin{cases}T_j^{\mathrm{win}},&j<b,\\24,&j\geq b.\end{cases} \end{aligned}(11)

A block length and overlap state is a quadruple of nonnegative integers giving length, overlap count, copies and trials.✦ With overlap multiplier CC, initial length ℓ\ell and initial overlap count hh, start from (ℓ,h,1,0)(\ell,h,1,0) and write ℓj,hj\ell_j,h_j for the length and overlap count at level jj. Let JjJ_j be the root trial count at target NN, length ℓj\ell_j, window TjwinT_j^{\mathrm{win}} and overlap count hjh_j, and let Loc⁡(P,hj)\operatorname{Loc}(P,h_j) be the local trial count for PP copies and overlap count hjh_j. The nested arithmetic states use copy counts PjP_j and trial counts tjt_j as in (12); these two counts describe entry into state j+1j+1. Pj={Jj(2hj+1)+Tjwin,j+1=b,⌊NAj+1/ℓj⌋+,j+1≠b,tj={Jj,j+1=b,Loc⁡(Pj,hj),j+1≠b,ℓj+1=ℓjPj,hj+1=⌈ChjPj⌉+.\begin{aligned} P_j&=\begin{cases} J_j(2h_j+1)+T_j^{\mathrm{win}},&j+1=b,\\ \lfloor N^{A_{j+1}}/\ell_j\rfloor_+,&j+1\ne b, \end{cases}& t_j&=\begin{cases}J_j,&j+1=b,\\ \operatorname{Loc}(P_j,h_j),&j+1\ne b,\end{cases}\\ \ell_{j+1}&=\ell_jP_j,&h_{j+1}&=\lceil Ch_jP_j\rceil_+. \end{aligned}(12) For count factor FF, the source arithmetic bounds at C,N,F,ε,s,b,ℓ,hC,N,F,\varepsilon,s,b,\ell,h are the conditions in (13) on these states: 12NAj≤ℓj≤NAj,ℓj≤hj≤C2jFℓj(0≤j≤b),12NAj+ε≤Pj≤2NAj+ε,Nε600C2bF≤tj≤2Nε(0≤j<b),24≤T(j),24⋅2s≤T(j),2T(j)≤Pj,4(tj+1)≤T(j)(0≤j<b).\begin{aligned} \tfrac12N^{A_j}\leq\ell_j&\leq N^{A_j},& \ell_j\leq h_j&\leq C^{2j}F\ell_j &&(0\leq j\leq b),\\ \tfrac12N^{A_j+\varepsilon}\leq P_j&\leq2N^{A_j+\varepsilon},& \frac{N^\varepsilon}{600C^{2b}F}\leq t_j&\leq2N^\varepsilon &&(0\leq j<b),\\ 24\leq T(j),\quad24\cdot2^s&\leq T(j),& 2T(j)\leq P_j,\quad4(t_j+1)&\leq T(j) &&(0\leq j<b). \end{aligned}(13)✦

For a real cap KK, the bad role factor ℛbad\mathcal R_{\mathrm{bad}} and forced role factor ℛforce\mathcal R_{\mathrm{force}}, together with the skeleton prefactor Σ\Sigma and smoothing prefactor Υ\Upsilon, are defined by (14): ℛbad(b,R)=∑r=0R−14⋅32(b+1)r,ℛforce(b,R)=∑r=0R−14⋅3(b+1)rS(r,b+1),Σ(s,b,K)=2(2s)3K𝔅(j↦4(2s)6K2,b),Υ(s,K)=2s2+s((s+1)K)s+1.\begin{aligned} \mathcal R_{\mathrm{bad}}(b,R)&=\sum_{r=0}^{R-1}4\cdot3^{2(b+1)r},\\ \mathcal R_{\mathrm{force}}(b,R)&=\sum_{r=0}^{R-1}4\cdot3^{(b+1)r}S(r,b+1),\\ \Sigma(s,b,K)&=2(2^s)^3K\, \mathfrak B\bigl(j\mapsto4(2^s)^6K^2,b\bigr),\\ \Upsilon(s,K)&=2^{s^2+s}\bigl((s+1)K\bigr)^{s+1}. \end{aligned}(14)

The Claim

Theorem 1. Let T(j)>0T(j)>0 for every jj, let s,b,Rs,b,R be nonnegative integers, and let ω,η\omega,\eta be real. Suppose a,w≥0a,w\geq0, w2𝒩b<1w^2\mathcal N_b<1, and Hs,ω(n(j))≤ηH_{s,\omega}(n(j))\leq\eta for 0≤j<b+10\leq j<b+1. For every typical report z∈𝒯z\in\mathcal T, the conditional coefficient energy satisfies (15): ℰw(Az)≤𝔈(T,n,s,b,R;ω,η,a,w).\mathcal E_w(A_z)\leq \mathfrak E(T,n,s,b,R;\omega,\eta,a,w).(15)✦

For the reference bound, take paired mechanisms MM on a finite side outcome space and a finite leaf law PP. Let ZZ be a real score with 𝔼PZ=0\mathbb E_PZ=0. Suppose kk is an integer with k≥1k\geq1, and write ri=2kr_i=2k for the even score exponent. Suppose L>0L>0, 𝔼P|Z|2k≤L2k\mathbb E_P|Z|^{2k}\leq L^{2k}, a≥0a\geq0, Hs,ω(n(j))≤ηH_{s,\omega}(n(j))\leq\eta for j<b+1j<b+1, and (6riL)2𝒩b<1(6r_iL)^2\mathcal N_b<1. The signed reference moment satisfies (16): 𝔼𝖰|𝟏𝒰(z,x)(g1(z,x)−g0(z,x))|2k≤𝔈(T,n,s,b,R;ω,η,a,6riL)k.\mathbb E_{\mathsf Q} \left|\mathbf1_{\mathcal U}(z,x) \bigl(g_1(z,x)-g_0(z,x)\bigr)\right|^{2k} \leq\mathfrak E(T,n,s,b,R;\omega,\eta,a,6r_iL)^k.(16)✦

For the block schedule, let B,D,i,ℓ,h,kB,D,i,\ell,h,k be nonnegative integers. Suppose B≥2B\geq2, i<Di<D, C,N,F≥1C,N,F\geq1 and ω>0\omega>0. Set b=bi+1b=b_{i+1} and take the windows and trials of (11) and (12). Assume the source arithmetic bounds (13), k≤Mk\leq M, and the budget conditions (17): K≥200C2bFω,8V(2s)2≤Nε,384ℛbad(b−1,R)≤Nε,192ℛforce(b−1,R)V2bΣ(s,b−1,K)≤Nε,1536≤Nε,Υ(s,K)≤Nε.\begin{aligned} K&\geq\frac{200C^{2b}F}{\omega},& 8V(2^s)^2&\leq N^\varepsilon,\\ 384\mathcal R_{\mathrm{bad}}(b-1,R)&\leq N^\varepsilon,& 192\mathcal R_{\mathrm{force}}(b-1,R)V^{2^b}\Sigma(s,b-1,K)&\leq N^\varepsilon,\\ 1536&\leq N^\varepsilon,&\Upsilon(s,K)&\leq N^\varepsilon. \end{aligned}(17) With L=N−A0di/2L=N^{-A_0d_i/2}, w=6(2k)Lw=6(2k)L, η=a=N−20W\eta=a=N^{-20W} and 𝒩b−1=2s∏j=1b−1T(j)\mathcal N_{b-1}=2^s\prod_{j=1}^{b-1}T(j), the level bounds are (18): w2𝒩b−1≤18N−(di+1)γ/2,w2𝒩b−1≤12,64𝔈(T,t,s,b−1,R;ω,η,a,w)≤N−di+1,Hs,ω(tj)≤N−20W(0≤j<b).\begin{aligned} w^2\mathcal N_{b-1}&\leq\frac18N^{-(d_i+1)\gamma/2},\qquad w^2\mathcal N_{b-1}\leq\tfrac12,\\ 64\mathfrak E(T,t,s,b-1,R;\omega,\eta,a,w)&\leq N^{-d_{i+1}},\\ H_{s,\omega}(t_j)&\leq N^{-20W}\quad(0\leq j<b). \end{aligned}(18)✦

In the paper, the result reads: for the block of Proposition 5.1 and every exact typical metadata value dd, if QQ is the product of the old first laws over all leaves and G+,G−G_+,G_- are the conditional densities obtained by averaging the type-dependent product of 1+Zold1+Z_{\mathrm{old}}, with ZoldZ_{\mathrm{old}} the old score, over all remaining hidden bits, including failed layouts, then ∥G+−G−∥Lri(Q)2≤N−d′−0.01\|G_+-G_-\|_{L^{r_i}(Q)}^2\leq N^{-d'-0.01} with the paper’s score exponent d′d', uniformly in dd and block level ii.

What Is Checked

The conditional bound (15) is stated over complete window assignments and typical reports. The reference bound (16) is over finite paired mechanisms and a centered leaf score with an even moment bound. The scheduled bound (18) assumes the specified arithmetic and prefactor inequalities. Under nonnegative common weight and common-submass domination, the paired report masses in (1) are nonnegative✦; for every real common weight their sum is one.✦