# Multiplicative-Noise Riccati Preconditioner

- ID: 1788
- Canonical URL: https://synthcore.org/idea/1788/multiplicative-noise-riccati-preconditioner
- API JSON: https://synthcore.org/api/idea/1788.json
- API Markdown: https://synthcore.org/api/idea/1788.md
- Verification status: unverified
- Source: [arXiv:2608.04386](https://arxiv.org/abs/2608.04386)
- Category: optimization
- Solves: stability, speedup, accuracy
- ML areas: optimizer, training-dynamics
- Math tags: stochastic-processes, optimization, dynamical-systems
- Ratings: usefulness 6/10; difficulty 6/10; novelty 6/10

## Idea description

Replace a standard diagonal optimizer preconditioner with a small Riccati-derived feedback controller for a block of neural parameters. The controller explicitly accounts for update-dependent stochasticity, potentially preventing unstable steps in noisy or strongly coupled training dynamics while permitting larger effective learning rates.

## Mathematical statement

The paper defines the effective control curvature and feedback terms as H_t = R_t + sum_h (D_t^h)^T P_t D_t^h, Theta_t = B_t^T P_t + sum_h (D_t^h)^T P_t C_t^h + S_t^T, and vartheta_t = B_t^T p_t + sum_h (D_t^h)^T P_t sigma_t^h + rho_t. Here P_t is the quadratic value Hessian, p_t is the linear value-gradient term, B_t and D_t^h describe how the control changes drift and diffusion, A_t and C_t^h describe uncontrolled drift and diffusion, R_t is the direct control-cost matrix, S_t is a state-control cross-cost matrix, and sigma_t^h and rho_t represent affine diffusion and affine cost terms. The feedback law is u_t = -H_t^{-1}(Theta_t x_t + vartheta_t). The quadratic Riccati equation is dot(P_t) + P_t A_t + A_t^T P_t + Q_t + sum_h (C_t^h)^T P_t C_t^h - Theta_t^T H_t^{-1} Theta_t = 0 with terminal condition P_T = G. The transferable property is the positive semidefinite correction (D_t^h)^T P_t D_t^h, which makes the optimizer more conservative exactly in directions where the update amplifies stochastic variation.

## Key formulas

- $$H_t=\mathcal{R}_t+\sum_{h=1}^{d_W}(\mathcal{D}_t^h)^{\top}P_t\mathcal{D}_t^h,\qquad \Theta_t=\mathcal{B}_t^{\top}P_t+\sum_{h=1}^{d_W}(\mathcal{D}_t^h)^{\top}P_t\mathcal{C}_t^h+\mathcal{S}_t^{\top}.$$
- $$\vartheta_t=\mathcal{B}_t^{\top}p_t+\sum_{h=1}^{d_W}(\mathcal{D}_t^h)^{\top}P_t\mathfrak{\sigma}_t^h+\mathfrak{\rho}_t.$$
- $$\dot{P}_t+P_t\mathcal{A}_t+\mathcal{A}_t^{\top}P_t+Q_t+\sum_{h=1}^{d_W}(\mathcal{C}_t^h)^{\top}P_t\mathcal{C}_t^h-\Theta_t^{\top}H_t^{-1}\Theta_t=0,\qquad P_T=G.$$
- $$u_t=-H_t^{-1}(\Theta_t x_t+\vartheta_t),\qquad \Delta\theta_t=\eta u_t.$$

## Implementation notes

Integrate the method into the optimizer update for one or more parameter blocks, not into the forward pass. Every K training steps, represent a block by a low-dimensional state x_t in R^m, for example x_t = [normalized parameter coordinates, momentum coordinates], with m between 16 and 64. Model local training dynamics as dx = (A x + B u)dt + sum_h (C^h x + D^h u)dW^h, where u is the proposed parameter update. Compute the minibatch gradient g_t and use it as the unfiltered control proposal. Estimate A and B by ridge regression of x_{t+1}-x_t on [x_t,g_t]. Estimate C^h and D^h from the covariance of regression residuals, using a rank-r eigendecomposition with r between 2 and 8. Choose Q = I, R = lambda I, S = 0, and solve the backward Riccati equation over a short horizon of 10 to 50 optimizer intervals. At the current step form H = R + sum_h D^{hT} P D^h and Theta = B^T P + sum_h D^{hT} P C^h, then apply u = -H^{-1} Theta x, optionally adding a small gradient forcing term -alpha H^{-1} g_t. Stabilize the solve with H <- H + epsilon I and clip the feedback gain norm. The Riccati correction and feedback are computed from the paper's equations; regression, covariance estimation, horizon choice, and clipping are empirical engineering choices. First test on a two-layer MLP trained on CIFAR-10 against AdamW and Shampoo at equal update FLOPs. Use blocks of size 32, refresh the model every 100 steps, and compare loss after a fixed number of examples, maximum stable learning rate, gradient-norm spikes, and final validation accuracy. A positive result is faster loss descent or a larger stable learning rate without degradation in final accuracy.

## Disclaimer

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