# Triangle-weighted support regularizer

- ID: 1670
- Canonical URL: https://synthcore.org/idea/1670/triangle-weighted-support-regularizer
- API JSON: https://synthcore.org/api/idea/1670.json
- API Markdown: https://synthcore.org/api/idea/1670.md
- Verification status: unverified
- Source: [arXiv:2608.01393](https://arxiv.org/abs/2608.01393)
- Category: geometry
- Solves: accuracy, stability, generalization
- ML areas: embedding, loss, regularization
- Math tags: geometry, optimization, functional-analysis
- Ratings: usefulness 5/10; difficulty 4/10; novelty 6/10

## Idea description

Add an exact angle-aware support-function penalty to a network that predicts a convex region around points, trajectories, or latent embeddings. For each triangle orientation, the penalty checks the paper's weighted support inequality rather than sampling many boundary points, encouraging globally valid geometric coverage with only three directional support evaluations.

## Mathematical statement

The paper proves that for a triangle with base angles alpha and beta, satisfying 0 < alpha, 0 < beta, and alpha + beta < pi, coverage of all triangle starts is equivalent to the weighted support inequality sin(alpha) sin(beta) <= sin(beta) H1 + sin(alpha) H2 + sin(alpha + beta) H3. Here H1, H2, and H3 are nonnegative distances or support values of the curve in the three triangle-related directions. The theorem exact_support_function_characterization further states that, for a nonnegative support function h:R -> R, coverage at every phase t is equivalent to satisfying the same inequality for every t, with H1 = h(t + pi + alpha), H2 = h(t + pi - beta), and H3 = h(t). In the neural adaptation, h_theta(phi) is the predicted support extent of a convex set in direction u(phi) = (cos(phi), sin(phi)); violations are penalized with a squared hinge.

## Key formulas

- $$0<\alpha,\quad 0<\beta,\quad \alpha+\beta<\pi,\quad H_1,H_2,H_3\geq 0,$$
- $$\sin\alpha\,\sin\beta\leq \sin\beta\,H_1+\sin\alpha\,H_2+\sin(\alpha+\beta)\,H_3,$$
- $$\mathcal{L}_{\mathrm{tri}}(\theta)=\frac{1}{|T|}\sum_{t\in T}\left[\sin\alpha\sin\beta-\sin\beta\,h_\theta(t+\pi+\alpha)-\sin\alpha\,h_\theta(t+\pi-\beta)-\sin(\alpha+\beta)h_\theta(t)\right]_+^2,$$
- $$h_\theta(\varphi)=\max_{j=1,\ldots,m} z_{\theta,j}^{\top}u(\varphi),\qquad u(\varphi)=(\cos\varphi,\sin\varphi),$$

## Implementation notes

Integrate this as an auxiliary loss in a network that predicts a two-dimensional convex uncertainty set or geometric enclosure. The network receives an input x and outputs nonnegative vertex coordinates z_theta,1,...,z_theta,m in R^2, using either a softplus parameterization or unconstrained coordinates followed by a validity loss. For every minibatch example, define the predicted support function h_theta(phi) as the maximum dot product z_theta,j^T u(phi), where u(phi) = (cos(phi), sin(phi)). During training, replace the hard maximum with tau log sum_j exp(z_j^T u(phi) / tau), using tau = 0.02 or annealing tau from 0.1 to 0.01. Select triangle angles alpha and beta, for example alpha = beta = pi/4, and sample 16 uniformly spaced phases t per example. Compute H1 = h_theta(t + pi + alpha), H2 = h_theta(t + pi - beta), and H3 = h_theta(t). Compute v_t = sin(alpha)sin(beta) - sin(beta)H1 - sin(alpha)H2 - sin(alpha + beta)H3, then add lambda mean_t max(v_t, 0)^2 to the task loss. The theorem supplies the angle weights and exact inequality; phase count, temperature, and lambda are empirical choices. First test on synthetic randomly rotated line segments and trajectories, training an MLP to output an m-vertex convex enclosure. Compare a standard pointwise inclusion loss against the same model plus this regularizer. Evaluate directional violation over 360 test orientations, enclosure area, calibration or coverage, and training throughput. The expected signal is lower worst-case directional violation and improved coverage at equal predicted area, with less computation than dense boundary sampling.

## Disclaimer

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