OctoInformBeta

OpenAI preprint, September 2026 · §5. A hierarchy of branching lattice points

Proposition 5.1: A block hierarchy controls residuals and Gaussian weights with finite coordinate alphabets

Lattice patterns with binary suffix branching.

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 triangular lattice construction distinguishes raw lattice coordinates from physical coordinates, related by a linear automorphism MM. A vertical pattern PyP_y at a point y∈ℝDy\in\mathbb R^D is a finite set of raw lattice points ℓ\ell whose physical residuals y−Mℓy-M\ell lie in prescribed coordinate boxes. Its Gaussian weights use the standard density in (1): γ1(u)=e−u2/22π,γD(x)=∏j=1Dγ1(xj)=(2π)−D/2exp⁡(−∥x∥222).\gamma_1(u)=\frac{e^{-u^2/2}}{\sqrt{2\pi}},\qquad \gamma_D(x)=\prod_{j=1}^D\gamma_1(x_j) =(2\pi)^{-D/2}\exp\!\left(-\frac{\|x\|_2^2}{2}\right).(1) The branching condition concerns suffixes of the raw coordinates. The one-dimensional standard Gaussian density γ1:ℝ→ℝ\gamma_1:\mathbb R\to\mathbb R is the function in (1); the density γD\gamma_D is its product over the DD coordinates.

Blocks and Their Triangular Maps

A folded Gaussian block BB consists of a dimension bb, a prime pp, a positive height hh, a usable set U⊆𝔽pbU\subseteq\mathbb F_p^b, and a nonzero direction d∈𝔽pbd\in\mathbb F_p^b.✦ Its block group is the additive group 𝔽pb\mathbb F_p^b, and its real vector space is ℝb\mathbb R^b. Denote its block lattice, block automorphism and block determinant by LBL_B, ABA_B and ΔB\Delta_B.

For columns w1,…,wm∈𝔽pbw_1,\ldots,w_m\in\mathbb F_p^b, the column shear SB,w:ℝm→ℝbS_{B,w}:\mathbb R^m\to\mathbb R^b is the linear map in (2), where val\operatorname{val} denotes the representative in {0,…,p−1}\{0,\ldots,p-1\}.✦ Let rBr_B be the least nonnegative integer with p≤2rBp\le2^{r_B}; the terminal dimension is tB=brBt_B=br_B.✦ Writing eje_j for the standard coordinate vectors of 𝔽pb\mathbb F_p^b, the binary terminal columns are 2rej2^re_j for 1≤j≤b1\le j\le b and 0≤r<rB0\le r<r_B, and the terminal binary shear TBT_B uses these columns: (SB,wv)j=hp∑k=1mval⁡(wkj)vk,TB=SB,(2rej)j,r:ℝtB→ℝb.(S_{B,w}v)_j=\frac hp\sum_{k=1}^m\operatorname{val}(w_k{}_j)v_k, \qquad T_B=S_{B,(2^re_j)_{j,r}}:\mathbb R^{t_B}\longrightarrow\mathbb R^b.(2)

Coordinate concatenation places u∈ℝmu\in\mathbb R^m before v∈ℝnv\in\mathbb R^n in ℝm+n\mathbb R^{m+n}.✦ The direct product of additive subgroups Λ⊆ℝm\Lambda\subseteq\mathbb R^m and Γ⊆ℝn\Gamma\subseteq\mathbb R^n, denoted by Λ×Γ⊆ℝm+n\Lambda\times\Gamma\subseteq\mathbb R^{m+n}, is the image of their Cartesian product under coordinate concatenation.✦ For linear automorphisms e,fe,f and a linear map S:ℝn→ℝmS:\mathbb R^n\to\mathbb R^m, the block triangular map is 𝒯e,f,S\mathcal T_{e,f,S} in (3): 𝒯e,f,S(u,v)=(eu+Sv,fv).\mathcal T_{e,f,S}(u,v)=(eu+Sv,fv).(3)✦

A branching block chain c=(B1,…,Bs)c=(B_1,\ldots,B_s) is a single block or an extension of a chain ending at AA by a block BB, with one column in 𝔽pAbA\mathbb F_{p_A}^{b_A} for each coordinate of BB.✦ Its initial block is B1B_1.✦ Write ν(b)\nu(b) for the prescribed ceiling next size. A branching block chain c=(B1,…,Bs)c=(B_1,\ldots,B_s) has ceiling block sizes when every extension satisfies bi+1=ν(bi)b_{i+1}=\nu(b_i).✦ A blockwise property of a chain is a property holding for every constituent block.✦

The chain vector space is ℝn\mathbb R^n, where n=∑ibin=\sum_i b_i.✦ The last-block inclusion ιc:ℝbs→ℝn\iota_c:\mathbb R^{b_s}\to\mathbb R^n puts a vector in the last block and zero elsewhere; last-block subtraction is y−ιcvy-\iota_c v. The chain product lattice LcL_c, chain determinant DvD_v, and extended product lattice LL are given in (4). Here B=BsB=B_s, t=tBt=t_B, and the integer lattice ℤt\mathbb Z^t is the integer span of the standard coordinate basis.✦ For a function ff on blocks, the sum over blocks is Sc(f)=∑i=1sf(Bi)S_c(f)=\sum_{i=1}^s f(B_i).✦ The squared chain radius is Rc2R_c^2, and the extended chain vector space is ℝD\mathbb R^D: Lc=LB1×⋯×LBs,Dv=∏i=1sΔBi,L=Lc×ℤt,D=n+t,Rc2=Sc(B↦bBhB2),ℛ2=Rc2+t.\begin{gathered} L_c=L_{B_1}\times\cdots\times L_{B_s},\qquad D_v=\prod_{i=1}^s\Delta_{B_i},\qquad L=L_c\times\mathbb Z^t,\\ D=n+t,\qquad R_c^2=S_c\bigl(B\mapsto b_Bh_B^2\bigr),\qquad \mathcal R^2=R_c^2+t. \end{gathered}(4) The chain dimension nn is the sum of the dimensions of its constituent blocks. The extended chain dimension DD is the chain dimension plus the terminal dimension tt.

The chain triangular automorphism McM_c starts with ABA_B on a single block and uses the last-block shear at each extension. The extended triangular automorphism MM adds the terminal binary shear, as in (5): M(B)=AB,Mc+B(u,v)=(Mcu+ιcSA,wv,ABv),M(u,v)=(Mcu+ιcTBv,v),(u,v)∈ℝn×ℝt.\begin{gathered} M_{(B)}=A_B,\qquad M_{c+B}(u,v)=\bigl(M_cu+\iota_cS_{A,w}v,A_Bv\bigr),\\ M(u,v)=\bigl(M_cu+\iota_cT_Bv,v\bigr), \qquad (u,v)\in\mathbb R^n\times\mathbb R^t. \end{gathered}(5)

Residual Boxes and Coordinate Alphabets

For vectors yy and ℓ\ell in the chain vector space, the chain residual condition ℛc(y,ℓ)\mathscr R_c(y,\ell) is the recursive box condition in (6); at an extension write y=(y′,yB)y=(y',y_B) and ℓ=(ℓ′,ℓB)\ell=(\ell',\ell_B).✦ For vectors yy and ℓ\ell in the extended chain vector space, the extended residual condition ℰc(y,ℓ)\mathscr E_c(y,\ell) is the box condition in (6), with the terminal coordinates denoted by the subscript TT: ℛ(B)(y,ℓ)⇔∥y−hBℓ∥∞≤hB,ℛc+B(y,ℓ)⇔ℛc(y′−ιcSA,wℓB,ℓ′)and∥yB−hBℓB∥∞≤hB,ℰc(y,ℓ)⇔ℛc(y′−ιcTBℓT,ℓ′)and∥yT−ℓT∥∞≤1.\begin{aligned} \mathscr R_{(B)}(y,\ell) &\ \Longleftrightarrow\ \|y-h_B\ell\|_\infty\le h_B,\\ \mathscr R_{c+B}(y,\ell) &\ \Longleftrightarrow\ \mathscr R_c(y'-\iota_cS_{A,w}\ell_B,\ell') \ \text{and}\ \|y_B-h_B\ell_B\|_\infty\le h_B,\\ \mathscr E_c(y,\ell) &\ \Longleftrightarrow\ \mathscr R_c(y'-\iota_cT_B\ell_T,\ell') \ \text{and}\ \|y_T-\ell_T\|_\infty\le1. \end{aligned}(6) The coordinate heights are hih_i on block BiB_i; their extension ηj\eta_j equals one on the terminal coordinates.✦ The chain coordinate denominators are pip_i on block BiB_i, and their extension qjq_j equals one on the terminal coordinates.

For nonnegative integers QQ and pp, the rational alphabet A(Q,p)A(Q,p) is the finite image of the integer interval [−Qp,Qp][-Qp,Qp] under k↦k/pk\mapsto k/p.✦ The alphabet radius of the extended chain is QcQ_c, and its coordinate alphabets are 𝒜j\mathcal A_j: A(Q,p)={k/p:k∈ℤ,−Qp≤k≤Qp},Qc=1+(D+2)D,𝒜j=A(Qc,qj)=qj−1ℤ∩[−Qc,Qc].\begin{gathered} A(Q,p)=\{k/p:k\in\mathbb Z,\ -Qp\le k\le Qp\},\qquad Q_c=1+(D+2)^D,\\ \mathcal A_j=A(Q_c,q_j)=q_j^{-1}\mathbb Z\cap[-Q_c,Q_c]. \end{gathered}(7)

A finite set P⊂ℝDP\subset\mathbb R^D has binary suffix fibers, denoted by 𝖲𝖡(P)\mathsf{SB}(P), when for each coordinate jj and each fixed suffix (zj+1,…,zD)(z_{j+1},\ldots,z_D) there is an a∈ℝa\in\mathbb R satisfying (8): {xj:x∈P,xk=zk for every k>j}⊆{a,a+1}.\{x_j:x\in P,\ x_k=z_k\text{ for every }k>j\}\subseteq\{a,a+1\}.(8)✦

The Claim

Theorem 1. There are an integer threshold b0≥2b_0\ge2 and absolute constants C,k,κ>0C,k,\kappa>0, independent of b1b_1, such that for each integer b1b_1 above b0b_0 a chain cc with ceiling block sizes and initial block dimension b1b_1 can be chosen to satisfy (9): D≤C(b1+1),ℛ2≤C(b1log⁡(b1+1)+1),for every y∈ℝD there is a finite ⌀≠Py⊆L with𝖲𝖡(Py),ℰc(y,ℓ)(ℓ∈Py),DvγD(y−Mℓ)≤exp⁡(−kb10.70)(ℓ∈Py),κ≤∑ℓ∈PyDvγD(y−Mℓ),there is ty∈ℝt with ℓT=ty(ℓ∈Py).\begin{gathered} D\le C(b_1+1),\qquad \mathcal R^2\le C\bigl(b_1\log(b_1+1)+1\bigr),\\ \text{for every }y\in\mathbb R^D\text{ there is a finite } \varnothing\ne P_y\subseteq L\text{ with}\\ \mathsf{SB}(P_y),\qquad \mathscr E_c(y,\ell)\quad(\ell\in P_y),\\ D_v\gamma_D(y-M\ell)\le \exp(-k b_1^{0.70})\quad(\ell\in P_y),\qquad \kappa\le\sum_{\ell\in P_y}D_v\gamma_D(y-M\ell),\\ \text{there is }t_y\in\mathbb R^t\text{ with } \ell_T=t_y\quad(\ell\in P_y). \end{gathered}(9)✦ The constants and chains can be chosen with all these properties and the uniform alphabet estimate (10): log⁡(|𝒜j|+1)≤log⁡8+Dlog⁡(D+2)+b12(1≤j≤D).\log(|\mathcal A_j|+1) \le\log8+D\log(D+2)+b_1^2\qquad(1\le j\le D).(10)✦

For any chain cc and vectors y,ℓ∈ℝDy,\ell\in\mathbb R^D satisfying ℰc(y,ℓ)\mathscr E_c(y,\ell), the physical residual satisfies (11): |(y−Mℓ)j|≤ηj(1≤j≤D).|(y-M\ell)_j|\le\eta_j\qquad(1\le j\le D).(11)✦ For z∈[0,1)Dz\in[0,1)^D and ℓ∈L\ell\in L satisfying ℰc(Mz,ℓ)\mathscr E_c(Mz,\ell), the coordinate membership assertion is (12): ℓj∈𝒜j(1≤j≤D).\ell_j\in\mathcal A_j\qquad(1\le j\le D).(12)✦ For U≥0U\ge0 and a chain whose every block satisfies log⁡pi≤U\log p_i\le U, the logarithmic alphabet bound is (13): log⁡(|𝒜j|+1)≤log⁡8+Dlog⁡(D+2)+U(1≤j≤D).\log(|\mathcal A_j|+1)\le\log8+D\log(D+2)+U \qquad(1\le j\le D).(13)✦ For finite sets P′,P⊂ℝDP',P\subset\mathbb R^D with P′⊆PP'\subseteq P and 𝖲𝖡(P)\mathsf{SB}(P), deletion preserves the branching condition as stated in (14): 𝖲𝖡(P′).\mathsf{SB}(P').(14)✦ A finite set P⊂ℝDP\subset\mathbb R^D with binary suffix fibers satisfies (15): |P|≤2D.|P|\le2^D.(15)✦

In the paper, the result reads: there are absolute b0,C,k,κ>0b_0,C,k,\kappa>0 such that, for every integer b1≥b0b_1\ge b_0, the preceding construction has D≤C(b1+1)D\le C(b_1+1) and ℛ2=∑ibihi2≤C(b1log⁡(b1+1)+1)\mathcal R^2=\sum_i b_ih_i^2\le C(b_1\log(b_1+1)+1), with the terminal block included in this sum, and for every y∈ℝDy\in\mathbb R^D there is a finite nonempty Py⊆LP_y\subseteq L with ∥(y−Mℓ)i∥∞≤hi\|(y-M\ell)_i\|_\infty\le h_i, total normalized Gaussian weight ∑ℓ∈PyDvγD(y−Mℓ)≥κ\sum_{\ell\in P_y}D_v\gamma_D(y-M\ell)\ge\kappa and each weight at most exp⁡(−kb10.70)\exp(-k b_1^{0.70}), whose terminal block is constant and whose values in block ii, after fixing later blocks, lie in a translate λi+{0,1}bi\lambda_i+\{0,1\}^{b_i} with λi∈LBi\lambda_i\in L_{B_i} depending on that suffix, so that |Py|≤2D|P_y|\le2^D, while fixing any scalar suffix leaves at most two next values separated by one and deletion preserves this rule, and for y∈M[0,1)Dy\in M[0,1)^D the ordinary coordinates may be restricted to pi−1ℤ∩[−Qc,Qc]p_i^{-1}\mathbb Z\cap[-Q_c,Q_c] and terminal coordinates to ℤ∩[−Qc,Qc]\mathbb Z\cap[-Q_c,Q_c], with Qc=1+(D+2)DQ_c=1+(D+2)^D and logarithmic alphabet sizes O(Dlog⁡(D+2)+bilog⁡bi)O(D\log(D+2)+b_i\log b_i), omitting the second term in the terminal block.

What Is Checked

The statements range over folded Gaussian blocks and their finite chains, with the associated product lattices, triangular maps and residual conditions, and over finite sets of real coordinate vectors. Coordinate concatenation is a real linear equivalence; the block triangular map in (3) has the inverse in (16): 𝒯e,f,S−1(u,v)=(e−1(u−Sf−1v),f−1v).\mathcal T_{e,f,S}^{-1}(u,v) =\bigl(e^{-1}(u-Sf^{-1}v),f^{-1}v\bigr).(16)