# Buffered partition-of-unity gating

- ID: 1903
- Canonical URL: https://synthcore.org/idea/1903/buffered-partition-of-unity-gating
- API JSON: https://synthcore.org/api/idea/1903.json
- API Markdown: https://synthcore.org/api/idea/1903.md
- Verification status: unverified
- Source: [arXiv:2608.06679](https://arxiv.org/abs/2608.06679)
- Category: architecture
- Solves: stability, scalability, memory
- ML areas: cnn, attention, moe-routing, inference-speedup
- Math tags: geometry, combinatorics, linear-algebra
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Replace an unconstrained spatial gate or interpolation kernel by a compactly supported function whose translates under a lattice exactly sum to one. Impose zero products between translates under a second lattice, so active gates do not collide; thresholding a positive superlevel set then provides a nonzero separation margin and predictable sparse computation.

## Mathematical statement

Let T=A\\mathbb{Z}^{d} be a full-rank lattice with fundamental parallelepiped F=A[0,1)^{d}. The paper constructs a nonnegative continuous compactly supported function \\varphi such that its T-translates form a partition of unity, \\sum_{t\\in T}\\varphi(x+t)=1 for every x\\in\\mathbb{R}^{d}. It also imposes \\varphi(x)\\varphi(x-p)=0 for every nonzero p\\in P, where P is a second full-rank lattice. The first identity gives exact coverage and convex interpolation; the second gives pointwise disjointness of nontrivial P-translates. For a threshold \    au>0, K_{\    au}=\\{x:\\varphi(x)\\geq\    au\\} is compact, and continuity plus disjointness imply that for sufficiently small \\varepsilon>0, (K_{\    au}+B_{\\varepsilon})\\cap(K_{\    au}+p)=\\varnothing for every p\\in P\\setminus\\{0\\}. In the normalized commensurable case, the existence criterion is n\\geq m+d, where m and n are the finite row and column counts induced by T and P.

## Key formulas

- $$F=A[0,1)^{d},\qquad T=A\mathbb{Z}^{d}$$
- $$\sum_{t\in T}\varphi(x+t)=1\quad(x\in\mathbb{R}^{d})$$
- $$\varphi(x)\varphi(x-p)=0,\qquad p\in P\setminus\{0\}$$
- $$K_{\tau}=\{x\in\mathbb{R}^{d}:\varphi(x)\geq\tau\},\qquad (K_{\tau}+B_{\varepsilon})\cap(K_{\tau}+p)=\varnothing\quad(p\neq0)$$

## Implementation notes

Integrate this at the spatial gating or interpolation operation of a CNN, vision transformer, neural field, or local-attention layer. Let each lattice site t in T own a feature vector h_t, and compute y(x)=sum_t phi_theta(x-t)h_t. Parameterize phi_theta by nonnegative vertex values on a finite triangulation of one fundamental cell F, interpolate piecewise affinely inside each simplex, and periodically extend the result by T. Initialize with compact tent functions, then train the vertex values with the task loss plus sampled constraints. For sampled x_k in F, compute s_k=sum over nearby t of phi_theta(x_k+t), and for nonzero p in a finite neighborhood of P compute c_{k,p}=phi_theta(x_k)phi_theta(x_k-p). Optimize L_task+lambda_1 sum_k(s_k-1)^2+lambda_2 sum_{k,p}c_{k,p}; use softplus vertex parameters and normalize each sampled partition sum if necessary. At inference, retain only sites with phi_theta(x-t)>=tau. Estimate epsilon empirically as the minimum distance between K_tau and K_tau+p over nearby nonzero p. The exact mathematical objects are the partition identity and zero-product constraint; finite sampling and threshold selection are approximations. First test on a small CIFAR-100 CNN against a dense 3x3 convolution at equal parameter count. Measure active sites, FLOPs, boundary artifacts, and validation accuracy. Success is lower compute and memory with no accuracy loss, plus stable outputs under small coordinate perturbations near lattice-cell boundaries.

## Disclaimer

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