# Pair-Hydrodynamic Long-Memory State Space

- ID: 1420
- Canonical URL: https://synthcore.org/idea/1420/pair-hydrodynamic-long-memory-state-space
- API JSON: https://synthcore.org/api/idea/1420.json
- API Markdown: https://synthcore.org/api/idea/1420.md
- Verification status: unverified
- Source: [arXiv:2607.25938](https://arxiv.org/abs/2607.25938)
- Category: memory
- Solves: accuracy, stability, memory
- ML areas: ssm, rnn, world-model, training-dynamics
- Math tags: dynamical-systems, statistical-mechanics, pde, spectral-theory, linear-algebra
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Augment a stable diffusive state-space model with pair states formed from products of slow latent modes. Single modes represent ordinary long-wavelength diffusion, while pair modes represent the interacting hydrodynamic operators responsible for late-time tails in quartic observables. Use the pair states only for selected readout channels or a low-rank subset of mode pairs, preserving near-linear inference cost.

## Mathematical statement

Let z_k(t) be a latent Fourier mode with wave number k and diffusivity D greater than zero, obeying dot z_k equals minus D k squared z_k plus B_k u(t), where u(t) is the input and B_k is an input projection. The paper's mechanism motivates pair modes q_(k,l) equal to z_k z_l, whose free diffusive generator has eigenvalue minus D times (k squared plus l squared). An observable is y_a equal to the sum over k of C_(a,k) z_k plus the sum over k,l of P_(a,k,l) q_(k,l). For a translationally invariant quartic channel with pair momentum zero, use q_k equal to z_k z_(-k). Its memory kernel is K_2(t) equal to the integral of w(k) exp(-2 D k squared t) over k, where w(k) is a nonnegative overlap weight. If w(0) is nonzero, one-dimensional diffusion predicts K_2(t) asymptotic to w(0) times the square root of pi divided by 2 D t. Stability follows from D greater than zero because all retained modes have nonpositive decay rates.

## Key formulas

- $$\dot z_k(t)=-Dk^2z_k(t)+B_ku(t),\qquad k=\frac{2\pi m}{L}.$$
- $$q_{k,\ell}(t)=z_k(t)z_\ell(t),\qquad \dot q_{k,\ell}=-D(k^2+\ell^2)q_{k,\ell}+z_kB_\ell u+z_\ell B_ku.$$
- $$y_a(t)=\sum_k C_{a,k}z_k(t)+\sum_{k,\ell}P_{a,k\ell}q_{k,\ell}(t).$$
- $$K_2(t)=\int_{-\pi}^{\pi}w(k)e^{-2Dk^2t}\,dk\xrightarrow[t\to\infty]{w(0)>0}w(0)\sqrt{\frac{\pi}{2Dt}}.$$

## Implementation notes

First, insert the module between the input encoder and prediction head of a long-horizon forecasting model. Choose M low-frequency wave numbers k_m equal to 2 pi m divided by L and maintain complex latent states z_m; impose z_(-m) equal to the complex conjugate of z_m when real outputs are required. Add a low-rank pair bank q_m equal to z_m z_(-m), or a small set of q_(m,n) equal to z_m z_n for multichannel data. Pseudocode is: initialize z_m to zero; for each input u_t, update z_m to exp(-D k_m squared Delta t) times z_m plus B_m u_t; compute q_m equal to z_m z_(-m); concatenate z, q, and encoder features; predict y_hat_t equal to C z plus P q plus R h_t. Parameterize D as softplus(d) and use the exact exponential update, which guarantees decay for every positive time step. The paper-derived quantities are the diffusion eigenvalues, pair eigenvalues, and hydrodynamic pair construction. Estimate D, B, readout coefficients, and overlap weights w(k) from data. In a cheap experiment, train on synthetic one-dimensional diffusion data with the target y_t equal to the spatial integral of x(r,t) squared. Compare a diagonal SSM, a single-mode SSM, and the pair model at equal parameter count. Fit the late-time impulse response to A times t to the power minus alpha. The prediction is alpha approximately equal to one half in one dimension when w(0) is nonzero, with A proportional to w(0) D to the power minus one half. Removing pair modes should eliminate this algebraic tail or replace it with excessively rapid decay. Changing D by a factor of four should change the predicted tail amplitude by a factor of approximately two.

## Disclaimer

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