✓✓ Beats tuned baseline 2026

Cyclic Multirate Neural State Space

Implementation & benchmark of arXiv:2607.09194 — Cyclic Reformulation-Based Identification and Polytopic Uncertainty Modeling for Multirate Systems

Usefulness7/10

Difficulty5/10

Novelty6/10

Architecture Stability Accuracy Sample efficiency Rnn Ssm World model Dynamical systems Control theory Spectral theory Linear algebra

Source paper: Cyclic Reformulation-Based Identification and Polytopic Uncertainty Modeling for Multirate Systems arXiv:2607.09194 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace interpolation of heterogeneous sensor streams by a phase-indexed recurrent or state-space network with period M, where M is the least common multiple of the sensor sampling periods. The network applies a distinct transition for each phase while using a fixed cyclic phase update, preserving timing structure and allowing missing observations to enter only when their phase-specific sensor is available.

Formulas

$$h_{t+1}=f_{\theta,r_t}(h_t,u_t),\qquad y_t=g_{\phi,r_t}(h_t),\qquad r_{t+1}=(r_t+1)\bmod M.$$

$$\begin{bmatrix}h_{t+1}\\y_t\end{bmatrix}=\begin{bmatrix}A_{r_t}&B_{r_t}\\C_{r_t}&D_{r_t}\end{bmatrix}\begin{bmatrix}h_t\\u_t\end{bmatrix},\qquad A_{r+M}=A_r.$$

$$\left\|J_{M-1}J_{M-2}\cdots J_0\right\|_2\leq\rho<1,\qquad J_r=\frac{\partial f_{\theta,r}}{\partial h},$$

$$\|\delta h_{qM}\|_2\leq\rho^q\|\delta h_0\|_2.$$

Mathematical statement

For phase r in {0,...,M-1}, let h_t in R^d be the latent state, u_t the available input vector, and f_{theta,r} the transition used at phase r. The cyclic model is h_{t+1}=f_{theta,r_t}(h_t,u_t), r_{t+1}=(r_t+1) mod M, with observation model y_t=g_{phi,r_t}(h_t). In the linear case this is x(k+1)=A_k x(k)+B_k u(k), y(k)=C_k x(k)+D_k u(k), where all matrices repeat every M steps. A sufficient local contraction condition is that the Jacobian product over one cycle has norm below one: ||J_{M-1}...J_0||_2 <= rho < 1, where J_r is the Jacobian of f_{theta,r} with respect to h. This gives exponential decay of perturbations at cycle boundaries.

Implementation notes

1. Integration point: implement a recurrent, S4-like, or small neural state-space model whose transition receives an integer phase r and a sensor-availability mask. Create M phase-specific transition heads f_r, using a shared backbone plus a low-rank phase embedding when M is large. At each physical time step, feed only measurements that are actually present; do not forward-fill missing sensor values. 2. Pseudocode: initialize h and r; for each t, construct z=[u_t, availability_mask_t, phase_embedding(r)]; compute h_new=f_r(h,z); predict y_hat=g_r(h_new); accumulate loss only for observed outputs; set r=(r+1) mod M. Estimate the Jacobian product over each M-step window using automatic differentiation and add a penalty when its spectral norm exceeds a target rho. 3. Computed versus estimated: M and phase assignments are computed exactly from sensor periods; Jacobians and their product norm are estimated on minibatch trajectories; rho is a chosen contraction target. 4. First cheap experiment: train on a synthetic linear periodically time-varying system with M=3 and randomly missing outputs, comparing interpolation+GRU, a phase-blind GRU, and this model. Sweep recurrent gain or spectral-normalization scale and measure long-horizon rollout error and perturbation growth. The predicted signature is a stability boundary near ||J_2J_1J_0||_2=1. Below the boundary, perturbations should decay approximately as rho^q, with measured log-decay slope agreeing with log(rho) within 20%.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Significant win ✓

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

How to read this: The implementation beat the baseline, but the claimed mechanism did not reproduce at benchmark scale. The gain may come from another effect; treat the cause as unresolved.

Stage 1 — Mechanism check agent confidence 9/10

Built a deterministic cyclic linear state-space verification and a small multirate forecasting comparison. The predicted contraction boundary gamma*=0.941036 for M=3 was observed at cycle norm 1.0; measured log-decay slopes matched log cycle norm essentially exactly across the sweep. Period scaling also matched Lambda^(-1/M) for M=2..5, and noncommuting transitions showed phase-order sensitivity. However, phase-aware forecasting MSE (0.071355) was slightly worse than phase-blind GRU-like linear recurrence (0.071331) and interpolation baseline (0.071330), so no practical accuracy win was demonstrated.

Agent confidence: 9/10

Baseline: Interpolation-style predictor MSE 0.071330; phase-blind predictor MSE 0.071331
Idea: Cyclic phase-aware predictor MSE 0.071355; mechanism boundary gamma observed 0.941036 vs predicted 0.941036; decay slopes matched predicted log(rho) to numerical precision; period-boundary absolute error <= 1.2e-16

Limitations: The practical comparison is a tiny linear recurrence proxy rather than trained GRU/neural state-space models, and it does not measure FLOPs, wall-clock speed, learned Jacobian penalties, or real heterogeneous sensor datasets. The exact mechanism verification uses constructed linear systems, so it establishes the mathematical signature rather than generalization to nonlinear models.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.09194) Sep 2, 2026

Beats the tuned baseline by -0.001 metric points (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0
Idea mean: 0.0014
Effect (Δ): -0.001 (n/a; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Not confirmed ✗
Practical effect: Helps
Baseline tuning: swept over 0 configs

Limitations:

Only the built-in 8-step actuated pendulum dynamics track was tested. The mechanism signature used directional JVP products rather than an exact full Jacobian spectral norm, and no heterogeneous real sensor dataset, missing-observation masking, FLOPs, or wall-clock comparison was tested.

How to run: python3 stage2_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Worked ✓

Agent confidence: 8/10

Implemented Cyclic Multirate Neural State Space as a phase-indexed recurrent transition model and compared it against a phase-blind recurrent baseline on the structurally matched dynamics track. The idea achieved lower test MSE on all 8 paired seeds: 0.0014195 versus 0.0024560, delta -0.0010364, permutation p=0.0081, satisfying the benchmark's significant-win criterion. The empirical mechanism signature did not confirm the quantitative prediction, with predicted cycle gain 0.2103 versus observed 0.2874.

Baseline: Mean test MSE 0.0024559747689636424; best config lr=0.008, phase_scale=0.0.
Delta mean: -0.0010364291301812045
Idea: Mean test MSE 0.0014195456387824379; selected config lr=0.008, phase_scale=1.0.
Idea wins: 8
Mechanism signature: map[confirmed:false cycle_jvp_gain_product_predicted:0.21025113761425018 cycle_perturbation_gain_observed:0.2874454855918884]
N pairs: 8
Permutation p value: 0.0081
Relative improvement pct: 42.205

Benchmark result