✓✓ Beats tuned baseline 2026

Stable Rotating-Memory State Space

Implementation & benchmark of arXiv:2608.00344 — Memory with Onsager-Casimir symmetry: Rotating particle in a viscoelastic fluid

Usefulness8/10

Difficulty5/10

Novelty6/10

Architecture Stability Accuracy Rnn Ssm Training dynamics Dynamical systems Control theory Linear algebra Spectral theory Stochastic processes

Source paper: Memory with Onsager-Casimir symmetry: Rotating particle in a viscoelastic fluid arXiv:2608.00344 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Replace an unconstrained recurrent transition with a decaying symmetric memory operator plus a skew-symmetric rotational operator. The skew component creates phase-shifted cross-channel memory and can represent oscillatory or circulatory temporal dependencies without requiring eigenvalues with large positive real parts.

Formulas

$$\dot h(t)=\left[-D+R\right]h(t)+Bx(t),\qquad D=D^T\succeq0,\quad R=-R^T,$$

$$\frac{d}{dt}\|h(t)\|_2^2=-2h(t)^TDh(t)\le0,$$

$$\bar A=\exp\!\left(\Delta t[-D+R]\right),\qquad \text{stable memory iff }\rho(\bar A)<1,$$

$$\mathbb{C}(t,t')\text{ and }\boldsymbol{\chi}(t)\text{ obey }\frac{\partial}{\partial t}\mathbb{C}(t,t')=-k_BT[\boldsymbol{\chi}(t-t')-\boldsymbol{\chi}(t)],\quad t>t'.$$

Mathematical statement

Use a continuous-time hidden state $h(t)\in\mathbb{R}^d$ with transition matrix $A_c=-D+R$, where $D=D^T\succeq0$ is damping and $R=-R^T$ is a rotational coupling. Since $h^TDh\ge0$ and $h^TRh=0$, the energy derivative satisfies $\frac{d}{dt}\|h\|^2=-2h^TDh\le0$. Discretize with $\bar A=\exp(\Delta t A_c)$, where $\Delta t$ is the sampling interval. Stability requires spectral radius $\rho(\bar A)<1$; strict damping in every active mode gives asymptotic decay. Reversing the learned rotation means replacing $R$ by $-R$, the neural analogue of reversing the paper's rotation vector.

Implementation notes

Integrate this module as a drop-in replacement for the recurrent transition in a small RNN, gated RNN, or linear state-space layer. Parameterize $D=LL^T+\epsilon I$ using a diagonal or low-rank factor $L$, and parameterize $R=Q-Q^T$ from an unconstrained matrix $Q$. Form $A_c=-D+R$ and compute $\bar A=\exp(\Delta tA_c)$ for small hidden dimensions. For larger dimensions use the Cayley discretization $\bar A=(I-\frac{\Delta t}{2}A_c)^{-1}(I+\frac{\Delta t}{2}A_c)$ or a stable matrix-exponential approximation. The update is: compute $D$ and $R$; form $A_c$; obtain $\bar A$; update $h_{k+1}=\bar Ah_k+Bx_k$; produce $y_k=Ch_k$. The paper contributes the rotating non-reciprocal-memory mechanism and Onsager–Casimir reversal interpretation; damping scales and input/output projections are learned. The first cheap test should use copy, adding, delayed-MNIST, and permuted-MNIST sequence tasks, comparing a vanilla RNN, diagonal SSM, and this module at matched parameter count. Sweep rotational scale $\omega=\|R\|_2$ and damping. The prediction is a divergence boundary at $\rho(\bar A)=1$: hidden norms remain bounded below the boundary and grow exponentially above it. The measured boundary should match the computed spectral-radius threshold within 10%. At $R=0$, antisymmetric lag correlations should vanish; replacing $R$ by $-R$ should reverse their sign.

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 stable rotating-memory state-space module using A_c=-D+R and matrix-exponential discretization, with a delayed-copy comparison against a tanh RNN. The mechanism predictions were confirmed: spectral radius matched exp(-damping*dt) to about 2.3e-16 relative error independent of rotation, cross-channel impulse response was zero at omega=0 and reversed sign under R -> -R, and measured angular frequency and decay matched omega and damping to below 1e-14 absolute error. The rotating model achieved lower delayed-copy test MSE (0.3002 vs 0.3317), though this comparison was not parameter-matched, so the primary evidence is successful mathematical verification rather than a definitive accuracy win.

Agent confidence: 9/10

Baseline: Vanilla tanh RNN: test MSE 0.3317087293, 97 parameters
Idea: Stable rotating SSM: test MSE 0.3001681268, 37 parameters; spectral-radius relative error <=2.3e-16; rotation reversal sums exactly 0; frequency and decay absolute errors <=8.4e-15

Limitations: Only a small synthetic delayed-copy task was tested; copy, adding, delayed-MNIST, and permuted-MNIST were not tested. The learning comparison used different parameter counts and one seed, and no Cayley discretization, larger hidden dimensions, noisy inputs, or training-speed/FLOP comparison was evaluated.

How to run: python3 stable_rotating_memory.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.00344) Aug 31, 2026

Beats the tuned baseline by 75.5% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0005
Idea mean: 0.0001
Effect (Δ): -0.0004 (−75.5%; 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 6 configs

Limitations:

Only the built-in dynamics track was tested; sequence, vision, and tabular tracks were not evaluated. The implementation uses matrix_exp at hidden size 64 and did not benchmark Cayley discretization, wall-clock speed, FLOPs, or longer horizons.

How to run: python3 bench_rotating_dynamics.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: 9/10

Implemented the stable rotating-memory transition A_c=-D+R with positive diagonal damping, skew-symmetric rotation, and matrix-exponential discretization as a matched replacement for the dynamics-track RNN. On the full 8-seed paired pendulum benchmark, the idea achieved significantly lower MSE than the tuned vanilla RNN: 0.0001262 vs 0.0005149, paired delta -0.0003887, permutation p=0.0081. The trained-model stability signature was confirmed: observed spectral radius 0.988963 remained below one and below the damping-based bound 0.989220.

Baseline: Tuned vanilla tanh RNN, 4481 parameters, full 8-seed mean test MSE 0.0005148992 (std 0.0001690044), best lr=0.01, weight_decay=0.0001, epochs=12
Idea: Stable rotating-memory RNN, 4481 parameters, full 8-seed mean test MSE 0.0001261922 (std 0.0000828951), best lr=0.01, weight_decay=0.0001, epochs=12
Paired delta: -0.00038870695; idea wins 8/8 seeds
Permutation p: 0.0081

Benchmark result