# Projector-Gap Trust Region for Shared Updates

- ID: 3157
- Canonical URL: https://synthcore.org/idea/3157/projector-gap-trust-region-for-shared-updates
- API JSON: https://synthcore.org/api/idea/3157.json
- API Markdown: https://synthcore.org/api/idea/3157.md
- Verification status: unverified
- Source: [arXiv:2609.03921](https://arxiv.org/abs/2609.03921)
- Category: regularization
- Solves: stability, generalization, speedup
- ML areas: optimizer, rnn, ssm, training-dynamics
- Math tags: optimization, control-theory, dynamical-systems, linear-algebra, spectral-theory
- Ratings: usefulness 7/10; difficulty 6/10; novelty 9/10

## Idea description

Use the behavior-subspace gap as a trust-region constraint when applying a shared update to multiple recurrent modules or experts. A proposed common update is accepted only when post-update behavior subspaces remain close to their leader and their graph subspaces remain sufficiently transverse, preventing one shared optimizer step from destabilizing dynamically different members.

## Mathematical statement

Let P_i(theta_i) be the projector onto the finite-horizon behavior subspace of module i at parameters theta_i, and let Delta theta_k be a shared cluster update. Constrain the update by ||P_i(theta_i+Delta theta_k)-P_lk(theta_lk+Delta theta_k)||_2 <= epsilon for every follower i. Since exact post-update projectors are expensive, estimate the directional change using DP_i(theta_i)[Delta theta_k], obtained by finite differences or a Jacobian-vector approximation. Scale the proposed update by alpha_k so the predicted maximum gap remains below epsilon. This converts the paper's geometric compatibility condition into an adaptive controller for training dynamics.

## Key formulas

- $$\operatorname{gap}^{+}_{i,k}=\left\|P_i(\theta_i+\Delta\theta_k)-P_{\ell_k}(\theta_{\ell_k}+\Delta\theta_k)\right\|_2\leq\varepsilon,$$
- $$\widehat{P}_i^{+}=P_i(\theta_i)+D P_i(\theta_i)[\Delta\theta_k],\qquad D P_i[\Delta\theta]\approx\frac{P_i(\theta_i+h\Delta\theta)-P_i(\theta_i)}{h},$$
- $$\alpha_k=\min\left(1,\frac{\varepsilon}{\max_{i\in\mathcal{C}_k}\|\widehat{P}_i^{+}(1)-\widehat{P}_{\ell_k}^{+}(1)\|_2+\delta}\right),\qquad \Delta\theta_k\leftarrow\alpha_k\Delta\theta_k,$$
- $$\operatorname{gap}^{\mathcal{G}}_L(i,j)=\|P_i-P_j\|_2=\max\{\|P_j^{\perp}P_i\|_2,\|P_i^{\perp}P_j\|_2\},\qquad P_i^{\perp}=I-P_i.$

## Implementation notes

Place the mechanism after gradient aggregation and before the optimizer update for each cluster of GRU, SSM, or MoE parameters. All cluster members receive a proposed common Adam or SGD update, but a scalar trust coefficient alpha_k scales that update before it is committed. Maintain calibration trajectories and a rank-r orthonormal basis U_i for every module; form P_i=U_iU_i^T every E training steps. Calculate the proposed update Delta theta_k, predict the updated basis using a finite-difference Jacobian-vector estimate or by applying the update to shadow copies, and compute the maximum follower-leader projector gap. Set alpha_k=min(1,epsilon/(max_gap+delta)); update theta_i with alpha_k Delta theta_k. If the measured post-update gap exceeds epsilon, halve alpha and retry. The paper supplies the projector-gap metric and its graph-subspace interpretation; behavior projectors, their parameter sensitivity, epsilon, and delta are implementation estimates. First experiment: train a four-layer stack of heterogeneous GRUs on adding and copy-memory tasks, sharing one update per two modules. Compare ordinary shared updates with the gap trust region at equal wall-clock cost while sweeping learning rate and epsilon. Record loss spikes, gradient norms, hidden-state norms, actual gaps, and clipped-update frequency. The quantitative prediction is that loss-spike probability rises sharply when the proposed gap exceeds epsilon, while adaptive alpha keeps measured gaps below epsilon and increases the largest stable learning rate. Test reproducibility across seeds and horizons L.

## Disclaimer

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