# Spectrahedral Obedience Layer

- ID: 1819
- Canonical URL: https://synthcore.org/idea/1819/spectrahedral-obedience-layer
- API JSON: https://synthcore.org/api/idea/1819.json
- API Markdown: https://synthcore.org/api/idea/1819.md
- Verification status: unverified
- Source: [arXiv:2608.04973](https://arxiv.org/abs/2608.04973)
- Category: architecture
- Solves: stability, accuracy, generalization
- ML areas: rl, moe-routing, architecture, loss
- Math tags: convex-analysis, linear-algebra, optimization, representation-theory
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Insert a differentiable equilibrium layer between a neural payoff/state encoder and the final action recommendations. The layer parameterizes a joint recommendation object and enforces all unilateral-deviation inequalities as positive-semidefinite constraints, preventing the network from producing recommendations that agents have a strict incentive to disobey. A quantum-inspired density-matrix parameterization can model correlated recommendations using PSD matrices rather than factorized action probabilities.

## Mathematical statement

Definition 3.3 defines obedience by requiring, for every player i, private type t_i, recommended action a_i, deviation b_i, and other-player action a_{-i}, the expected deviation gain to be nonnegative: sum_{omega,t_{-i},a_{-i}} psi(omega) pi(t_i,t_{-i}|omega) sigma(a_i,a_{-i}|t,omega)[u_i(a_i,a_{-i},omega)-u_i(b_i,a_{-i},omega)] >= 0. Here omega is the payoff state, psi(omega) its prior, t=(t_i,t_{-i}) the signal profile, pi the signal distribution, sigma the mediator's conditional recommendation distribution, and u_i the payoff. The quantum analogue replaces classical information objects with density operators rho_omega, where rho_omega is positive semidefinite and has trace one, and expresses obedience through Loewner domination X >=_L 0, meaning v^*Xv >= 0 for every complex vector v. For an implementable neural adaptation, X_{i,t_i,a_i,b_i} is the operator-valued deviation-surplus matrix obtained by weighting recommendation blocks by u_i(a_i,a_{-i},omega)-u_i(b_i,a_{-i},omega). Enforce X_{i,t_i,a_i,b_i} >=_L 0 for every constraint. The feasible set is a spectrahedron, an affine slice of the PSD cone, so it can be handled by an SDP projection or a log-det barrier.

## Key formulas

- $$\sum_{\omega,t_{-i},a_{-i}}\psi(\omega)\pi(t_i,t_{-i}\mid\omega)\sigma(a_i,a_{-i}\mid t,\omega)\big[u_i(a_i,a_{-i},\omega)-u_i(b_i,a_{-i},\omega)\big]\ge 0.$$
- $$X_{i,t_i,a_i,b_i}=\sum_{\omega,a_{-i}}\operatorname{Tr}_{-i}\!\left(R_{a_i,a_{-i},\omega}\right)\big[u_i(a_i,a_{-i},\omega)-u_i(b_i,a_{-i},\omega)\big]\succeq 0,$$
- $$R_{a,\omega}\succeq 0,\qquad \sum_{a}R_{a,\omega}=\rho_{\omega},\qquad \rho_{\omega}\succeq0,\quad\operatorname{Tr}(\rho_{\omega})=1.$$
- $$\mathcal{L}(R)=\mathcal{L}_{\mathrm{task}}(R)+\lambda\sum_{i,t_i,a_i,b_i}-\log\det\!\left(X_{i,t_i,a_i,b_i}+\varepsilon I\right),\qquad \lambda>0.$$

## Implementation notes

(1) Integration point: use this as the final recommendation head in a two-player multi-agent RL policy or as a router in a small mixture-of-experts model. Let the state/payoff encoder output unconstrained matrices Z_{a,omega}; convert them to PSD recommendation blocks with R_{a,omega}=L_{a,omega}L_{a,omega}^T, where L_{a,omega} is a learned lower-triangular factor. Normalize over actions for each payoff state using rho_omega=sum_a R_{a,omega}, followed by trace normalization. (2) Pseudocode: encode observations and payoff-state features; produce Cholesky factors L[a,omega]; form R[a,omega]=L[a,omega]@L[a,omega].T; compute rho[omega]=sum_a R[a,omega] and divide by Tr(rho[omega]); for every player i, type t_i, recommendation a_i, and alternative b_i, assemble X[i,t_i,a_i,b_i] by summing recommendation blocks weighted by the payoff differences in the first displayed formula; compute each minimum eigenvalue; optimize task loss plus the log-det barrier, or perform a small batched SDP projection after each forward pass. (3) The paper-derived quantities are PSD feasibility, Loewner order, trace normalization, and deviation-surplus construction. Estimate psi, pi, and unknown payoff tables from minibatch frequencies or a learned payoff model. Use epsilon=1e-4 and initialize lambda=1e-2, annealing toward 1e-4. (4) First experiment: train two agents on a two-state, two-action matrix game with randomly sampled payoff tables. Compare an unconstrained correlated-policy head, a scalar BCE linear-program projection, and this PSD head. Measure return, maximum positive deviation gain, fraction of recommendations obeyed, and wall-clock overhead. Success means deviation gain stays below 1e-3 while matching unconstrained return, or improved robustness on held-out payoff states. A secondary test should compare correlated PSD recommendations against factorized policies at equal parameter count.

## Disclaimer

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