# Hyperbolic Ring-Closure Regularizer

- ID: 910
- Canonical URL: https://synthcore.org/idea/910/hyperbolic-ring-closure-regularizer
- API JSON: https://synthcore.org/api/idea/910.json
- API Markdown: https://synthcore.org/api/idea/910.md
- Verification status: unverified
- Source: [arXiv:2607.14348](https://arxiv.org/abs/2607.14348)
- Category: regularization
- Solves: stability, accuracy, generalization
- ML areas: regularization, cnn, graph-nn, training
- Math tags: geometry, differential-geometry, numerical-analysis
- Ratings: usefulness 5/10; difficulty 5/10; novelty 8/10

## Idea description

Regularize a scalar feature field on a 2D grid by interpreting each feature value as the uniformizing variable of a hyperbolic ring and penalizing violations of local orthogonal-ring angle closure. Unlike a raw Laplacian penalty, this constrains the representation through positive hyperbolic radii and geometrically meaningful edge compatibility.

## Mathematical statement

The paper assigns each vertex v_i a hyperbolic ring with outer radius R_i, inner radius r_i, and shared ratio q=cosh(r_i)/cosh(R_i) less than or equal to 1, using the hyperbolic Pythagoras identity cosh(R_j)cosh(r_i)=cosh(r_j)cosh(R_i) on every incident edge. For an edge joining rings (R,r) and (R_k,r_k), the two right-triangle angles are phi_k=arctan(tanh(R_k)/sinh(r)) and psi_k=arctan(tanh(r_k)/sinh(R)), so the opening angle is theta_k=phi_k+psi_k. A valid orthogonal ring pattern satisfies the vertex closure condition sum_{k adjacent to v} theta_k=2 pi. We map a neural scalar feature u_i to a positive ring size using R_i=softplus(a u_i+b)+R_min and r_i=arcosh(q cosh(R_i)), with a,b,R_min fixed or learned and q in (0,1). The loss penalizes both edge-wise violations of the shared-q relation and squared angle-closure error.

## Key formulas

- $$\cosh R_j\cosh r_i=\cosh r_j\cosh R_i,\qquad q=\frac{\cosh r_i}{\cosh R_i}\leq 1.$$
- $$\varphi_k=\arctan\frac{\tanh R_k}{\sinh r},\qquad \psi_k=\arctan\frac{\tanh r_k}{\sinh R},\qquad \theta_k=\varphi_k+\psi_k.$$
- $$\mathcal{L}_{\mathrm{ring}}=\frac{1}{|E|}\sum_{(i,j)\in E}\left(\frac{\cosh r_i}{\cosh R_i}-\frac{\cosh r_j}{\cosh R_j}\right)^2+\frac{1}{|V|}\sum_{i\in V}\left(\sum_{j\sim i}\theta_{ij}-2\pi\right)^2.$$

## Implementation notes

(1) Integrate this into a CNN, vision transformer patch grid, or graph neural network immediately after a scalar feature head produces u in R^{H times W}. Use one scalar channel or a small learned projection of the feature tensor; do not apply the loss independently to every high-dimensional channel initially. Convert each feature to R_i=softplus(a u_i+b)+R_min, with R_min=10^{-3}. Choose q=sigmoid(q_0)(1-10^{-4}), and compute r_i=arcosh(q cosh(R_i)); clamp the arcosh argument to at least 1+10^{-6}. (2) Pseudocode: R=softplus(a*u+b)+R_min; q=sigmoid(q_0)*(1-eps); r=acosh(q*cosh(R)); for every directed grid edge i to j, compute phi[i,j]=atan(tanh(R[j])/sinh(r[i])), psi[i,j]=atan(tanh(r[j])/sinh(R[i])), and theta[i,j]=phi[i,j]+psi[i,j]; accumulate closure[i]=sum over j adjacent to i of theta[i,j]; compute Lring=mean over edges of (cosh(r[i])/cosh(R[i])-cosh(r[j])/cosh(R[j])) squared plus mean over vertices of (closure[i]-2*pi) squared. Add lambda_ring times Lring to the task loss. (3) The geometric formulas are computed exactly; q, a, and b are learned or fixed, while the closure target 2*pi and grid adjacency are fixed. Log the minimum r_i, maximum R_i, closure error, and gradient norm to detect numerical saturation. (4) First experiment: train a small CIFAR-10 CNN with a scalar feature head, comparing cross-entropy alone, a standard total-variation penalty, a standard Laplacian penalty, and this ring loss at equal parameter count and training steps. Test lambda_ring in {10^{-4},10^{-3},10^{-2},10^{-1}}. Success means lower validation loss or improved corruption robustness at the same training budget, with closure error decreasing during training and no increase in gradient explosions.

## Disclaimer

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