# Centro-affine spherical smoothness regularizer

- ID: 1183
- Canonical URL: https://synthcore.org/idea/1183/centro-affine-spherical-smoothness-regularizer
- API JSON: https://synthcore.org/api/idea/1183.json
- API Markdown: https://synthcore.org/api/idea/1183.md
- Verification status: unverified
- Source: [arXiv:2607.20223](https://arxiv.org/abs/2607.20223)
- Category: regularization
- Solves: stability, generalization
- ML areas: embedding, regularization, attention
- Math tags: geometry, differential-geometry, convex-analysis, spectral-theory
- Ratings: usefulness 4/10; difficulty 6/10; novelty 7/10

## Idea description

Add a centro-affine Dirichlet penalty to a neural module whose inputs or outputs lie on a sphere, such as normalized embeddings or attention directions. The penalty measures intrinsic variation under an unconditional convex-body metric while projecting out the constant and coordinate-affine modes excluded by the theorem.

## Mathematical statement

Let K be a smooth unconditional convex body in R to the n, h its support function on the unit sphere, bar-g its round-sphere metric, and bar-nabla its round-sphere connection. Define tau = bar-nabla-squared h plus h bar-g, and define the centro-affine metric g = tau divided by h. The associated measure is dV_K = h times det(bar-g inverse times tau) dx = h divided by Gauss curvature times dx. For a smooth function F satisfying integral F dV_K = 0 and integral F times x_i divided by h dV_K = 0 for every coordinate i, the theorem gives integral F squared dV_K <= 1 divided by n times integral g inverse applied to the centro-affine gradient of F twice dV_K. The neural adaptation uses this intrinsic gradient energy as a regularizer and weighted least-squares projection to approximate the orthogonality conditions.

## Key formulas

- X = bar-nabla h + h x; tau = bar-nabla-squared h + h bar-g; g = tau divided by h
- dV_K = h det(bar-g inverse tau) dx = h divided by Gauss curvature dx
- integral F dV_K = 0; integral F times x_i divided by h dV_K = 0 for i = 1,...,n
- integral F squared dV_K <= (1 divided by n) integral g-inverse(centro-affine-gradient F, centro-affine-gradient F) dV_K

## Implementation notes

(1) Integrate the regularizer into a spherical embedding head or attention-direction module. Given hidden vectors z in R to the n, normalize them as x = z divided by its Euclidean norm plus epsilon, and let the head produce a scalar or channel-wise function F(x). Start with an unconditional ellipsoid support function h(x) = square root of the sum of a_i squared x_i squared, with every a_i positive. This preserves coordinate-reflection symmetry and is easy to evaluate. Compute tau from the spherical Hessian of h, then g = tau divided by h and the quadrature weight w = h times determinant of the inverse round metric times tau. (2) For each minibatch, form B whose columns are the constant function and the n functions x_i divided by h. Compute c = (B-transpose W B plus epsilon I) inverse times B-transpose W F, where W is diagonal with entries w. Set the residual R = F minus Bc. Compute its spherical or autodifferentiated gradient, contract it with the inverse centro-affine metric, and add lambda times the weighted mean of this energy divided by n to the task loss. (3) The paper supplies the metric construction, measure, removed modes, and one-over-n scaling; automatic differentiation and minibatch quadrature estimate derivatives and integrals. (4) First test on CIFAR-10 using a small MLP or ViT projection head producing normalized 128-dimensional embeddings. Compare weight decay, an isotropic spherical Jacobian penalty, and this penalty at matched loss-weight sweeps. Measure test accuracy, gradient-norm variance, embedding collapse, and loss oscillation. Success is higher accuracy or lower gradient variance at equal training loss without increased collapse.

## Disclaimer

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