{
 "artifacts": null,
 "category": "dynamics",
 "description": "Replace a monolithic nonlinear latent transition in a neural world model or sequence predictor with two lifted latent channels: a global channel encoding scene-wide or sequence-wide structure and local channels encoding patches, segments, tokens, or objects. Propagate both channels with a block-structured linear operator and decode them jointly, so the encoder remains nonlinear but multi-step latent rollouts do not repeatedly apply a deep transition network.",
 "formulas_latex": [
  "$$\\dot{\\mathbf{x}}=\\mathbf{f}(\\mathbf{x},\\mathbf{u})$$",
  "$$\\mathbf{z}_{k+1}^{\\mathrm{Combined}}=\\begin{bmatrix}A^{\\mathrm{Global}}\u00260\\\\0\u0026A^{\\mathrm{Local}}\\end{bmatrix}\\begin{bmatrix}\\mathbf{z}_{k}^{\\mathrm{Global}}\\\\\\mathbf{z}_{k}^{\\mathrm{Local}}\\end{bmatrix}+\\begin{bmatrix}B^{\\mathrm{Global}}\\\\B^{\\mathrm{Local}}\\end{bmatrix}\\mathbf{u}_{k},\\qquad \\tilde{\\mathbf{x}}_{k}=\\begin{bmatrix}C^{\\mathrm{Global}}\u00260\\\\0\u0026C^{\\mathrm{Local}}\\end{bmatrix}\\mathbf{z}_{k}^{\\mathrm{Combined}}$$",
  "$$\\begin{aligned}z^g_{t+1}\u0026=A_gz^g_t+B_gu_t,\\\\z^l_{t+1}\u0026=A_lz^l_t+B_lu_t,\\\\\\hat{x}_{t+j}\u0026=D([z^g_{t+j};z^l_{t+j}]),\\\\z^g_t\u0026=\\phi_g(x_{t-r+1:t}),\\quad z^l_t=\\phi_l(x_{t-r+1:t}).\\end{aligned}$$",
  "$$\\mathcal{L}=\\sum_{j=1}^{H}\\left(\\|\\hat{x}_{t+j}-x_{t+j}\\|_2^2+\\lambda_g\\|P_g(\\hat{x}_{t+j}-x_{t+j})\\|_2^2+\\lambda_l\\|P_l(\\hat{x}_{t+j}-x_{t+j})\\|_2^2\\right)+\\beta\\sum_{j=0}^{H-1}\\left(\\|A_gz^g_{t+j}+B_gu_{t+j}-z^g_{t+j+1}\\|_2^2+\\|A_lz^l_{t+j}+B_lu_{t+j}-z^l_{t+j+1}\\|_2^2\\right)$$"
 ],
 "id": 3130,
 "implementation": "(1) Integration point: use this as the latent transition inside a small video-prediction, robot-world-model, or sequence-model backbone. Given an observation window $x_{t-r+1:t}$, implement two encoders: a global encoder $\\phi_g$ producing $z^g_t\\in\\mathbb{R}^{d_g}$ and a shared local encoder $\\phi_l$ producing $z^l_t\\in\\mathbb{R}^{N\\times d_l}$ for patches, objects, graph nodes, or body segments. Flatten local latents for the first MVP. Replace the usual GRU or MLP transition with the block update from the formula, and use a decoder $D([z^g;z^l])$ to reconstruct the next observation. (2) Pseudocode: encode $(z_g,z_l)=(\\phi_g(x_{t-r+1:t}),\\phi_l(x_{t-r+1:t}))$; for $j=0,\\ldots,H-1$, compute $z_g\\leftarrow A_gz_g+B_gu_{t+j}$ and $z_l\\leftarrow A_lz_l+B_lu_{t+j}$, then $\\hat{x}_{t+j+1}\\leftarrow D([z_g;z_l])$; minimize the displayed horizon loss, with $P_g$ selecting downsampled or global features and $P_l$ selecting local patches or segments. (3) Estimate $A_g,A_l,B_g,B_l$ either as trainable matrices or by least squares on detached encoder outputs. Initialize each $A$ near identity and optionally constrain its largest singular value to at most $1+\\epsilon$ using spectral normalization or projection. The zero cross-block structure comes from the paper; nonlinear coupling is learned by the encoders and decoder. $H$ is the rollout horizon, $r$ is the history length, $N$ is the number of local entities, and $\\lambda_g,\\lambda_l,\\beta$ are validation-tuned weights. (4) First cheap experiment: train on Moving-MNIST, bouncing balls, or a small robot-arm trajectory dataset using a two-layer GRU latent baseline with the same parameter count and rollout horizon. Compare one-step and 20-step prediction error, gradient norm growth, and rollout wall-clock time. Success is lower long-horizon MSE and less error explosion at equal FLOPs; for a control dataset, additionally test whether a simple shooting planner using the linear latent rollout achieves lower trajectory-tracking error without retraining under altered payload or disturbance conditions.",
 "math_summary": "The paper begins from the controlled nonlinear system $\\dot{\\mathbf{x}}=\\mathbf{f}(\\mathbf{x},\\mathbf{u})$, where $\\mathbf{x}\\in\\mathbb{R}^{n}$ is the state, $\\mathbf{u}\\in\\mathbb{R}^{m}$ is the control or conditioning input, and $\\mathbf{f}$ is the nonlinear vector field. Its combined lifted model has the form $\\mathbf{z}_{k+1}^{\\mathrm{Combined}}=\\begin{bmatrix}A^{\\mathrm{Global}}\u00260\\\\0\u0026A^{\\mathrm{Local}}\\end{bmatrix}\\begin{bmatrix}\\mathbf{z}_{k}^{\\mathrm{Global}}\\\\\\mathbf{z}_{k}^{\\mathrm{Local}}\\end{bmatrix}+\\begin{bmatrix}B^{\\mathrm{Global}}\\\\B^{\\mathrm{Local}}\\end{bmatrix}\\mathbf{u}_{k}$ and $\\tilde{\\mathbf{x}}_{k}=\\begin{bmatrix}C^{\\mathrm{Global}}\u00260\\\\0\u0026C^{\\mathrm{Local}}\\end{bmatrix}\\mathbf{z}_{k}^{\\mathrm{Combined}}$. Here $\\mathbf{z}^{\\mathrm{Global}}\\in\\mathbb{R}^{d_g}$ and $\\mathbf{z}^{\\mathrm{Local}}\\in\\mathbb{R}^{d_l}$ are lifted observables, $A^{\\mathrm{Global}}$ and $A^{\\mathrm{Local}}$ are learned linear transition matrices, $B^{\\mathrm{Global}}$ and $B^{\\mathrm{Local}}$ inject controls or conditioning variables, and $C^{\\mathrm{Global}}$, $C^{\\mathrm{Local}}$ decode the latent states. The zero off-diagonal blocks deliberately impose a modeling approximation in lifted space; physical coupling is retained through the joint nonlinear observable map and decoder. Adapt this as $z^g_t=\\phi_g(x_{\\le t})$, $z^l_t=\\phi_l(x_{\\le t})$, with $[z^g_{t+1};z^l_{t+1}]=\\operatorname{diag}(A_g,A_l)[z^g_t;z^l_t]+[B_g;B_l]u_t$ and $\\hat{x}_{t}=D([z^g_t;z^l_t])$. The useful property is linear, cheap, recursively structured rollout in lifted coordinates while preserving nonlinear cross-channel interactions in $\\phi_g$, $\\phi_l$, and $D$.",
 "math_tags": [
  "dynamical-systems",
  "linear-algebra",
  "control-theory",
  "optimization"
 ],
 "ml_areas": [
  "world-model",
  "ssm",
  "graph-nn"
 ],
 "paper": {
  "arxiv_id": "2609.03175",
  "arxiv_url": "https://arxiv.org/abs/2609.03175",
  "summary_what_math_gives_to_ml": "The paper's transferable asset is a structured Koopman lifting in which global and segment-local observables are propagated by a linear latent model, while the nonlinear encoder and decoder carry the coupling information. The block-diagonal lifted dynamics are not claimed to decouple the physical system; instead, they provide a cheaper and potentially better-conditioned representation for long-horizon prediction and optimization. This suggests a neural sequence or world-model architecture with separate global and local latent channels, constrained linear latent rollouts, and multi-scale reconstruction losses. The expected benefit is more stable long-horizon prediction and planning than an unconstrained recurrent latent dynamics model, especially when local changes must coexist with globally coherent structure.",
  "title": "Real-Time Shape Control of Multi-Segment Soft Robotic Arms Using Koopman Operators with Global and Local Observables",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 5,
  "usefulness": 7
 },
 "solves": [
  "stability",
  "accuracy",
  "scalability"
 ],
 "title": "Global-Local Koopman Latent Dynamics",
 "url": "https://synthcore.org/idea/3130/global-local-koopman-latent-dynamics",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)",
  "verification_axes": {
   "benchmark_mechanism": {
    "confirmed": null,
    "tested": false
   },
   "practical_benchmark": {
    "beats_baseline": null,
    "tested": false
   },
   "toy_mechanism_gate": {
    "confirmed": null,
    "tested": false
   }
  }
 }
}