Unverified 2026

Fractal surface-to-volume coordinate router

Implementation & benchmark of arXiv:2608.21246 — Space-filling surfaces: sharp Hölder continuous parameterizations from squares to cubes

Usefulness5/10

Difficulty6/10

Novelty6/10

Architecture Memory Scalability Accuracy Cnn Graph nn Embedding Memory Geometry Approximation theory Topology Combinatorics

Source paper: Space-filling surfaces: sharp Hölder continuous parameterizations from squares to cubes arXiv:2608.21246 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Build a fixed multiscale router that maps 2D coordinates to 3D voxel coordinates using the paper's X-shaped self-similar refinement. Use the router to run a 3D feature field or volumetric token mixer over a 2D-organized tensor, while retaining a mathematically controlled locality bound instead of an arbitrary flattening permutation. The first target is a 3D neural field or small voxel classifier where the router replaces either a dense 3D feature table or a naive raster-order token layout.

Formulas

$$M=2^m,\qquad \alpha=\frac{m}{m+1};\qquad m=2:\ M=4,\ \alpha=\frac23.$$

$$\mathcal A:=\{(p,q)\in\{0,\ldots,M-1\}^2:p=q\text{ or }p+q=M-1\}.$$

$$\phi_{p,q}(x)=\frac{x}{M}+t_{p,q},\qquad \phi_{p,q}(Q)=Q_{p,q},\qquad \operatorname{diam}(\phi_{a_k}\circ\cdots\circ\phi_{a_1}(Q))=M^{-k}\operatorname{diam}(Q).$$

$$\|f(u)-f(v)\|_\infty\le C\|u-v\|_\infty^{m/(m+1)}\quad\text{for }u,v\in[0,1]^m.$$

Mathematical statement

For input dimension $m$, the paper fixes $M=2^m$ and the sharp exponent $\alpha=m/(m+1)$. In the square-to-cube case, $m=2$, $M=4$, and $\alpha=2/3$. Let $Q=[-1/2,1/2]^2$, and partition it into $M^2$ subsquares $Q_{p,q}$ of side length $1/M$. The retained branch indices are exactly $\mathcal A=\{(p,q)\in\{0,\ldots,M-1\}^2:p=q\text{ or }p+q=M-1\}$, the X-shaped set of diagonal subsquares. For each $(p,q)\in\mathcal A$, $\phi_{p,q}:Q\to Q_{p,q}$ is the affine contraction consisting of dilation by $1/M$ and translation to $Q_{p,q}$. At depth $k$, recursively apply $\phi_{p_k,q_k}\circ\cdots\circ\phi_{p_1,q_1}$, so a parameter cell has diameter at most $M^{-k}$. The associated surface-to-volume parameterization has the scale law $\|f(u)-f(v)\|_\infty\le C\|u-v\|_\infty^{\alpha}$, with $f:[0,1]^m\to[0,1]^{m+1}$ and $\alpha=m/(m+1)$; for $m=2$, $\alpha=2/3$. The implementation uses the retained branch set as the exact finite-resolution router and estimates the constant $C$ empirically from sampled neighboring pairs. A branch-to-voxel bijection must be supplied at each scale; use a fixed deterministic bijection that assigns the retained branches to child cubes and alternates orientation so adjacent parent branches receive adjacent child addresses.

Implementation notes

Integrate the router before feature lookup or token mixing in a small volumetric model. Start with the square-to-cube case. Represent a 2D address $u=(u_1,u_2)\in[0,1]^2$ at depth $k$ by base-$M$ digits. At each level choose a branch $(p,q)$ from $\mathcal A$, apply $x\leftarrow x/M+t_{p,q}$, and use a fixed table $b_k(p,q)$ to select one of four child-cube orientations. Alternate reflected orientations according to the parent branch so neighboring retained squares map to neighboring 3D child cubes. The integer voxel address is the concatenation of child-cube digits. Pseudocode: for level in 1..k: digit=(floor(M*u1),floor(M*u2)); branch=project_to_A(digit); x=(x+t[branch])/M; child=b[level,parent_orientation,branch]; voxel=voxel*2+child; u=M*u-floor(M*u); orientation=update_orientation(orientation,branch). Gather 2D feature cells at addresses $u$, scatter them to routed 3D voxel addresses, and resolve duplicates by mean pooling or a learned gated sum. Add $L_{loc}=\mathbb E_{u,v}[\max(0,\|f(u)-f(v)\|_\infty-C_{emp}\|u-v\|_\infty^{2/3})]$, with $C_{emp}$ set to the 95th-percentile sampled ratio. The paper supplies the retained branches, contraction scale, and target exponent; the branch-to-cube bijection, orientation rule, and collision handling are engineering choices and must be ablated. First test a 2D-routed 3D CNN or neural radiance field on ModelNet10 or a 32^3 ShapeNet subset against dense 3D convolution, raster flattening, Morton order, and Hilbert-like ordering at equal feature count and FLOPs. Measure accuracy or PSNR, peak activation memory, throughput, duplicate-address rate, and empirical Hölder ratio. Success means at least 25% lower activation memory or improved throughput with no more than a 1 percentage point accuracy loss, while locality beats raster order and approaches Morton or Hilbert baselines.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.