# Wave-Scale IMEX Latent Dynamics

- ID: 2371
- Canonical URL: https://synthcore.org/idea/2371/wave-scale-imex-latent-dynamics
- API JSON: https://synthcore.org/api/idea/2371.json
- API Markdown: https://synthcore.org/api/idea/2371.md
- Verification status: unverified
- Source: [arXiv:2608.15837](https://arxiv.org/abs/2608.15837)
- Category: dynamics
- Solves: stability, speedup, scalability
- ML areas: ssm, world-model, training-dynamics, graph-nn
- Math tags: dynamical-systems, numerical-analysis, linear-algebra, pde
- Ratings: usefulness 6/10; difficulty 6/10; novelty 5/10

## Idea description

Split a learned dynamical model into a slow nonlinear transport branch and a stiff fast-coupling branch, evaluating the former explicitly and solving only the latter with a small implicit iteration. This should permit larger rollout steps when latent fast modes have large Jacobian eigenvalues while retaining expressive nonlinear dynamics in the explicit branch.

## Mathematical statement

The paper decomposes the MHD evolution into an explicit advective subsystem and implicit magnetic and pressure subsystems. Its characteristic eigenvalues are lambda_{1,8}=u plus or minus c_f, lambda_{2,7}=u plus or minus c_a, lambda_{3,6}=u plus or minus c_s, lambda_4=u, and lambda_5=0, where u is fluid velocity, c_a=B_x/sqrt(rho) is the Alfven speed, and c_f,c_s are fast and slow magnetosonic speeds. The speeds obey c_{f,s}^2=1/2[(||B||^2/rho+a^2) plus or minus sqrt((||B||^2/rho+a^2)^2-4B_x^2a^2/rho)], with a^2=partial p/partial rho+(p/rho^2) partial p/partial e. Adapt this scale separation to a latent ODE dz/dt=f_slow(z;theta)+f_fast(z;phi). The IMEX step is z_{n+1}=z_n+dt f_slow(z_n;theta)+dt f_fast(z_{n+1};phi). For residual g(z)=z-z_n-dt f_slow(z_n)-dt f_fast(z), Newton uses [I-dt J_fast] delta z=-g, where J_fast=partial f_fast/partial z.

## Key formulas

- $$\lambda_{1,8}=u\pm c_f,\quad\lambda_{2,7}=u\pm c_a,\quad\lambda_{3,6}=u\pm c_s,\quad\lambda_4=u,\quad\lambda_5=0.$$
- $$c_{f,s}^{2}=\frac{1}{2}\left(\frac{\|\mathbf{B}\|^{2}}{\rho}+a^{2}\pm\sqrt{\left(\frac{\|\mathbf{B}\|^{2}}{\rho}+a^{2}\right)^{2}-\frac{4B_x^{2}a^{2}}{\rho}}\right).$$
- $$z_{n+1}=z_n+\Delta t\,f_{\mathrm{slow}}(z_n;\theta)+\Delta t\,f_{\mathrm{fast}}(z_{n+1};\phi).$$
- $$\left[I-\Delta t\,J_{\mathrm{fast}}(z_{n+1})\right]\Delta z=-g(z_{n+1}),\quad g(z)=z-z_n-\Delta t f_{\mathrm{slow}}(z_n)-\Delta t f_{\mathrm{fast}}(z).$$

## Implementation notes

Use this as a replacement for the ordinary explicit residual update in a latent ODE, recurrent world model, neural state-space model, or neural operator rollout. Split the latent dynamics into two subnetworks or two channel groups: f_slow is a standard MLP or GNN representing transport and ordinary interactions, while f_fast is a smaller residual MLP representing rapidly varying coupling. For each step, compute s=f_slow(z_n), initialize z_trial=z_n+dt*s, and perform two to five implicit iterations. The cheapest version is fixed point: z_trial <- z_n+dt*s+dt*f_fast(z_trial). A Newton version computes g=z_trial-z_n-dt*s-dt*f_fast(z_trial), forms J_fast with automatic differentiation, solves [I-dt*J_fast] delta=-g using conjugate gradients or GMRES, and updates z_trial <- z_trial+delta until ||g||/(||z_trial||+1e-8)<1e-4. Estimate which channels are stiff by five-step power iteration on the local Jacobian; assign the larger estimated spectral-radius component to f_fast. The paper contributes the characteristic-scale split, IMEX update, and implicit Jacobian system; the latent branch assignment and regularization are empirical choices. First test a small latent ODE on a damped stiff oscillator, Kuramoto system, or 2-D advection-reaction dataset. Compare against an explicit residual model with equal parameters and training FLOPs. Record rollout error, number of model evaluations, solver iterations, and stability at increasing dt. Success is stable prediction at two to five times larger dt, lower long-horizon error at equal evaluations, or faster loss descent when the explicit baseline diverges.

## Disclaimer

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