# Projected Play-State Memory

- ID: 665
- Canonical URL: https://synthcore.org/idea/665/projected-play-state-memory
- API JSON: https://synthcore.org/api/idea/665.json
- API Markdown: https://synthcore.org/api/idea/665.md
- Verification status: unverified
- Source: [arXiv:2607.09974](https://arxiv.org/abs/2607.09974)
- Category: architecture
- Solves: stability, generalization, accuracy
- ML areas: ssm, rnn, world-model
- Math tags: convex-analysis, geometry, dynamical-systems
- Ratings: usefulness 6/10; difficulty 4/10; novelty 5/10

## Idea description

Turn a recurrent or state-space memory into a constrained hereditary state: the latent state remains in a learned convex domain, and only input motion that reaches the boundary changes the plastic component. This creates a nonexpansive, rate-independent memory that should suppress unstable state growth and make the representation depend on meaningful cumulative changes rather than arbitrary update frequency.

## Mathematical statement

Let \(C\subset F\) be a closed convex elastic domain, \(\pi(t)\in F\) the driving latent history, and \(P_C\) metric projection onto \(C\) in the stress-space norm \(\|\sigma\|_F^2=\mathbb{C}^{-1}\sigma\cdot\sigma\). The discrete vector play or return-map update is \(\sigma_k=P_C(\sigma_{k-1}+\Delta\pi_k)\), where \(\Delta\pi_k=\pi_k-\pi_{k-1}\). The correction \(r_k=(\sigma_{k-1}+\Delta\pi_k)-\sigma_k\) belongs to the normal cone \(N_C(\sigma_k)\), where \(N_C(x)=\{u:\langle u,y-x\rangle_F\le 0\ \forall y\in C\}\). For the absolutely continuous limit, \(\sigma(t)\in C\) and \(\dot\pi(t)-\dot\sigma(t)\in N_C(\sigma(t))\) almost everywhere. Metric projection is nonexpansive, \(\|P_C(x)-P_C(y)\|_F\le\|x-y\|_F\), so perturbations in the driving history cannot amplify in the constrained state.

## Key formulas

- $$\langle\sigma^{1},\sigma^{2}\rangle_{F}=\mathbb{C}^{-1}\sigma^{1}\cdot\sigma^{2},\qquad \|\sigma\|_{F}=\sqrt{\mathbb{C}^{-1}\sigma\cdot\sigma}.$$
- $$\sigma_k=P_C(\sigma_{k-1}+\Delta\pi_k),\qquad \Delta\pi_k=\pi_k-\pi_{k-1}.$$
- $$r_k=(\sigma_{k-1}+\Delta\pi_k)-\sigma_k\in N_C(\sigma_k),\qquad N_C(x)=\{u:\langle u,y-x\rangle_F\le 0\ \forall y\in C\}.$$
- $$\|P_C(x)-P_C(y)\|_F\le\|x-y\|_F,\qquad \sigma(t)\in C,\quad \dot\pi(t)-\dot\sigma(t)\in N_C(\sigma(t))\ \text{a.e.}$$

## Implementation notes

Insert the module as a recurrent memory update in an SSM, GRU-like block, or recurrent policy. Let the network produce an unconstrained drive increment Delta_pi_t=g_theta(x_t,h_{t-1}), and maintain a latent state sigma_t in a simple convex set C, initially a box |sigma_i|<=b_i or an l2 ball. Compute u_t=sigma_{t-1}+Delta_pi_t, then project with sigma_t=P_C(u_t). For a box, projection is elementwise clipping; for an l2 ball, use P_C(u)=u min(1,b/||u||_2). Feed sigma_t to the output head and optionally expose the correction r_t=u_t-sigma_t as a second channel, analogous to an elastic/plastic decomposition. Use a learned positive diagonal metric by scaling coordinates before projection, or begin with the Euclidean metric to avoid instability. The paper supplies the projection, normal-cone characterization, and nonexpansive guarantee; the set C, metric, drive network, and whether to expose r_t are engineering choices. Compare against the same backbone with an unconstrained additive state update and identical parameter count and truncated-BPTT length. Start with copy, adding-long-memory, and sequential-MNIST tasks using a two-layer GRU or SSM. Track hidden-state norms, gradient norms, input-perturbation amplification, validation accuracy, and performance when test sequences are much longer than training sequences. Success means bounded states and lower perturbation amplification without loss of short-sequence accuracy, together with improved length extrapolation.

## Disclaimer

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