# Equilibrium-Seeking Predictive Optimizer

- ID: 2934
- Canonical URL: https://synthcore.org/idea/2934/equilibrium-seeking-predictive-optimizer
- API JSON: https://synthcore.org/api/idea/2934.json
- API Markdown: https://synthcore.org/api/idea/2934.md
- Verification status: unverified
- Source: [arXiv:2608.28180](https://arxiv.org/abs/2608.28180)
- Category: dynamics
- Solves: stability, accuracy, scalability
- ML areas: optimizer, training-dynamics, mlp
- Math tags: control-theory, optimization, dynamical-systems, linear-algebra
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Partition a neural network into heterogeneous parameter blocks or maintain several worker replicas, and model each block's optimizer state as a constrained linearized dynamical agent. At every synchronization interval, jointly optimize a finite sequence of parameter updates and a feasible common terminal parameter target, while enforcing consensus through distributed primal-dual iterations. Unlike ordinary gradient descent toward a fixed or implicit target, the target is selected together with the update trajectory and can absorb heterogeneous curvature, learning rates, and parameter constraints.

## Mathematical statement

For worker or parameter block i, let theta_i(k) in R^{n_i} be its parameter or optimizer state, v_i(k) in R^{m_i} its control update, A_i a local linearized state-transition matrix, and B_i a control matrix. The predictive model is theta_i(k+1)=A_i theta_i(k)+B_i v_i(k), with theta_i(k) in X_i and v_i(k) in U_i. The decision variable z_i is a dynamically feasible terminal equilibrium satisfying z_i=A_i z_i+B_i v_i^e, and neighboring equilibria are constrained toward consensus by graph incidence matrix E: (E tensor I)z=0. Use the finite-horizon quadratic surrogate J_i=sum over t from 0 to T-1 of [(theta_i(t)-z_i)^T Q_i(theta_i(t)-z_i)+(v_i(t)-D_i z_i)^T R_i(v_i(t)-D_i z_i)] plus (theta_i(T)-z_i)^T P_i(theta_i(T)-z_i), where Q_i, R_i, and P_i are positive semidefinite weighting matrices and D_i z_i is the equilibrium control. A distributed primal-dual method performs projected descent on local primal variables and dual ascent on consensus residuals. A practical sufficient local condition is alpha_i<2/L_i, where alpha_i is the primal step size and L_i is an estimate of the Lipschitz constant of the local gradient; for coupled consensus updates use alpha beta ||E||_2^2<1, where beta is the dual step size. The terminal matrix P_i should satisfy the discrete Lyapunov inequality A_cl,i^T P_i A_cl,i-P_i <= -Q_f,i, with A_cl,i=A_i+B_i K_i and Q_f,i positive definite, so the terminal cost decreases after applying the first planned control.

## Key formulas

- $$\theta_i(k+1)=A_i\theta_i(k)+B_i v_i(k),\qquad \theta_i(k)\in\mathcal{X}_i,\;v_i(k)\in\mathcal{U}_i.$$
- $$J_i=\sum_{t=0}^{T-1}\left[(\theta_i(t)-z_i)^\top Q_i(\theta_i(t)-z_i)+(v_i(t)-D_i z_i)^\top R_i(v_i(t)-D_i z_i)\right]+(\theta_i(T)-z_i)^\top P_i(\theta_i(T)-z_i).$$
- $$\theta_i^{p+1}=\Pi_{\mathcal{C}_i}\!\left(\theta_i^p-\alpha_i\left[\nabla J_i(\theta_i^p)+E_i^\top\lambda^p\right]\right),\qquad \lambda^{p+1}=\lambda^p+\beta(E\otimes I)z^{p+1},$$
- $$A_{cl,i}^{\top}P_iA_{cl,i}-P_i\preceq-Q_{f,i},\qquad \alpha\beta\|E\|_2^2<1,\qquad \alpha_i<\frac{2}{L_i}.$$

## Implementation notes

1. Integration point: use this as an outer optimizer around ordinary backpropagation. Split trainable parameters into blocks or replicas. At synchronization time k, collect each current parameter vector theta_i, gradient g_i, and a diagonal or low-rank local curvature estimate. Set A_i=I for a first implementation and B_i=-H_i^{-1}, so the control v_i represents a preconditioned gradient update; later estimate A_i from two successive optimizer states. Bound v_i by a per-block norm or coordinate box and bound theta_i using weight-norm constraints. 2. Pseudocode: initialize z_i=theta_i and dual lambda=0; for p=1,...,P, roll out theta_i[t+1]=A_i theta_i[t]+B_i v_i[t], evaluate the local quadratic surrogate plus a minibatch loss-gradient linearization, take projected primal descent on theta_i,z_i,v_i using alpha times (grad J_i+E_i^T lambda), then update lambda <- lambda+beta(E tensor I)z, and execute only v_i[0] on the real network before replanning. 3. Compute Q_i and R_i from parameter and update scaling and estimate L_i by two or three power iterations or backtracking; estimate ||E||_2 from the worker graph. P_i can initially be I, while the Lyapunov inequality is checked numerically for fitted A_i,B_i,K_i. 4. First experiment: train a small MLP on MNIST with four heterogeneous replicas, equalizing wall-clock time and gradient evaluations, and compare Adam, decentralized SGD, and the proposed one-step and five-step MPC variants. Ablate equilibrium optimization, dual consensus, and projection separately. 5. Pre-register the signature: on a quadratic loss with known Hessian H, divergence should begin near the predicted primal boundary alpha_c=2/lambda_max(H), and consensus residual ||E z_p|| should decay geometrically when alpha beta ||E||_2^2<1. Require measured alpha_c to be within 20% of prediction and a fitted exponential decay rate to be positive below the boundary. Reject the transfer if the method diverges substantially below the bound, fails to contract consensus on the quadratic test, or gives no improvement in terminal loss at equal update budget.

## Disclaimer

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