# Subdual cone energy layer

- ID: 2640
- Canonical URL: https://synthcore.org/idea/2640/subdual-cone-energy-layer
- API JSON: https://synthcore.org/api/idea/2640.json
- API Markdown: https://synthcore.org/api/idea/2640.md
- Verification status: unverified
- Source: [arXiv:2608.21005](https://arxiv.org/abs/2608.21005)
- Category: regularization
- Solves: accuracy, stability, generalization
- ML areas: embedding, loss, regularization
- Math tags: convex-analysis, geometry, linear-algebra
- Ratings: usefulness 6/10; difficulty 3/10; novelty 6/10

## Idea description

Add a cone-aware score to a latent representation by projecting each latent vector onto a closed convex cone K and using the norm of the projection as an order-sensitive energy. If K is subdual, any latent displacement in the cone order is guaranteed not to reduce this energy, providing a mathematically certified monotone feature rather than merely penalizing observed violations.

## Mathematical statement

Let H be a real Hilbert space with norm \|\cdot\|, let K \subseteq H be a closed convex cone, and let its dual be K^* = \{u \in H : \langle u,k\rangle \geq 0\     ext{for every } k\in K\}. The cone order is x\leq_K y if and only if y-x\in K. Let P_K(x)=\arg\min_{k\in K}\|x-k\| be the metric projection. The paper's characterization is K\subseteq K^* if and only if x\leq_K y implies \|P_Kx\|\leq\|P_Ky\|. Thus, for a subdual cone, the scalar function e_K(x)=\|P_Kx\| is order-monotone even when the stronger vector condition P_Kx\leq_K P_Ky is unavailable. For implementation, choose H=R^d and a fixed cone K. A particularly cheap choice is the nonnegative orthant K=R_+^d, where P_K(x)_i=max(x_i,0); a richer choice is the second-order cone K={ (t,v): t\geq\|v\|_2 }, whose projection is explicit. For matrix-valued embeddings, use the positive-semidefinite cone and P_K(X)=U\operatorname{diag}(\max(\lambda_i,0))U^T, where X=U\operatorname{diag}(\lambda_i)U^T. Subduality is guaranteed for these self-dual cones.

## Key formulas

- $$x\leq_K y\quad\Longleftrightarrow\quad y-x\in K.$$
- $$P_Kx=\arg\min_{k\in K}\|x-k\|.$$
- $$K\subseteq K^*\quad\Longleftrightarrow\quad x\leq_K y\ \Longrightarrow\ \|P_Kx\|\leq\|P_Ky\|.$$
- $$e_K(z)=\|P_Kz\|_2,\qquad \mathcal L_{\mathrm{ord}}=\frac{1}{|\mathcal P|}\sum_{(x,y)\in\mathcal P}\operatorname{softplus}\!\left(e_K(x)-e_K(y)+m\right),\quad y-x\in K.$$

## Implementation notes

Integrate the layer immediately after a d-dimensional encoder output h(x), before the prediction head. Compute both the projected latent vector p=P_K(h) and the scalar cone energy e=||p||_2; concatenate [h,p,e] to the downstream head, or use e as an auxiliary monotone score. For supervised ordered data, construct pairs (x,y) whose desired relation is y-x in K. Pseudocode: `h_x=encoder(x); h_y=encoder(y); p_x=project_K(h_x); p_y=project_K(h_y); e_x=sqrt(sum(p_x**2)+eps); e_y=sqrt(sum(p_y**2)+eps); L_ord=mean(softplus(e_x-e_y+m)); L=L_task+alpha*L_ord`. For the orthant, `project_K(z)=relu(z)`. For the second-order cone, write z=(t,v), r=||v||; return z if t>=r, zero if t<=-r, otherwise `0.5*(1+t/r)*(r,v)`. For PSD embeddings, eigendecompose each symmetric matrix and clamp negative eigenvalues, but begin with the orthant or SOC for cheap experiments. The theorem supplies the monotonicity guarantee when the pair relation is genuinely cone-positive; the network does not need to learn that guarantee. Empirically estimate violations on held-out ordered pairs using `max(0,e_x-e_y)` and compare against a plain latent-energy baseline. First test on a small MLP for synthetic order-preserving regression: sample x, k in K, set y=x+k, and predict a monotone target. Compare an ordinary MLP, a ReLU-constrained monotone head, and this auxiliary cone-energy model at matched parameter count. Success is lower held-out order-violation rate and equal-or-better task error, with no material increase in training time for orthant/SOC projections.

## Disclaimer

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