△ Mechanism confirmed, baseline not beaten 2026

Bilinear Input-Conditioned Koopman Cell

Implementation & benchmark of arXiv:2607.01819 — Koopman operator theory: fundamentals, control, and applications

Usefulness6/10

Difficulty5/10

Novelty5/10

Architecture Stability Accuracy Sample efficiency World model Ssm Rnn Rl Dynamical systems Linear algebra Control theory Tensor decomposition

Source paper: Koopman operator theory: fundamentals, control, and applications arXiv:2607.01819 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace an unconstrained input-conditioned recurrent transition with a bilinear latent update, so controls modulate a fixed linear latent dynamics matrix through low-rank state-input interactions. The resulting cell preserves the computational simplicity of linear propagation while representing multiplicative effects of actions that an additive control term cannot capture efficiently.

Formulas

$$\begin{bmatrix}x_{1}\\x_{2}\\x_{1}^{2}\\1\end{bmatrix}^{+}=\begin{bmatrix}a&0&0&0\\0&c&d&h\\0&0&a^{2}&0\\0&0&0&1\end{bmatrix}\begin{bmatrix}x_{1}\\x_{2}\\x_{1}^{2}\\1\end{bmatrix}+\begin{bmatrix}0&0&0&0\\e&0&0&f\\0&0&0&0\\0&0&0&0\end{bmatrix}\begin{bmatrix}x_{1}\\x_{2}\\x_{1}^{2}\\1\end{bmatrix}u.$$

$$z_{t+1}=Az_t+\sum_{j=1}^{p}u_{t,j}B_jz_t+Du_t+b,$$

$$B_j=U_jV_j^{\top},\qquad z_{t+1}=Az_t+\sum_{j=1}^{p}u_{t,j}U_j(V_j^{\top}z_t)+Du_t+b.$$

Mathematical statement

The paper derives finite-dimensional linear and bilinear Koopman forms for controlled systems. In the displayed bilinear case, the lifted state $z=[x_1,x_2,x_1^2,1]^{\top}$ evolves as $z^{+}=Az+Bzu$, where $A=\begin{bmatrix}a&0&0&0\\0&c&d&h\\0&0&a^2&0\\0&0&0&1\end{bmatrix}$ and $B=\begin{bmatrix}0&0&0&0\\e&0&0&f\\0&0&0&0\\0&0&0&0\end{bmatrix}$. The symbols a, c, d, h, e, and f are scalar system coefficients and u is the control input. Generalize this structure to a learned latent vector $z_t\in\mathbb{R}^m$ and action $u_t\in\mathbb{R}^p$: $z_{t+1}=Az_t+\sum_{j=1}^{p}u_{t,j}B_jz_t+Du_t+b$. The bilinear term is the transferable asset: it represents action-dependent changes in latent dynamics with parameter cost $O(pm^2)$, or $O(prm)$ with rank-r factors $B_j=U_jV_j^{\top}$.

Implementation notes

Use this as the transition cell in a model-predictive world model, controlled SSM, or action-conditioned recurrent predictor. Encode observation x_t into z_t in R^m. Store a base matrix A, action matrix D, bias b, and either dense B_j matrices or rank-r factors U_j in R^{m x r}, V_j in R^{m x r}. At every step compute q_j=V_j^T z_t, multiply each q_j by the scalar action u_{t,j}, map through U_j, sum across j, and update z_next=A z_t+sum_j u_{t,j} U_j q_j+D u_t+b. Decode z_next to predict the next observation. Train with multi-step teacher-forced and free-running rollout losses; optionally constrain A's spectral radius by parameterizing A=rho*tanh(S)/max(1,sigma_max(S)) with rho<1, while leaving the bilinear term expressive. Pseudocode: z=encoder(x); for t in horizon: interaction=0; for j in range(p): interaction += u[j]*U[j]@(V[j].T@z); z=A@z+interaction+D@u+b; prediction=decoder(z); accumulate loss. The paper supplies the bilinear structure; rank r, spectral constraint, and losses are empirical choices. Begin with m=32, r=4, and 2-4 actions on a simulated pendulum, CartPole, or MuJoCo task. Compare against an equal-parameter MLP transition and an additive linear model. Track rollout MSE, control reward under MPC or model-predictive planning, parameter count, and wall-clock cost. Success means materially lower long-horizon prediction error or higher planned control reward than additive dynamics, with no instability and lower cost than the MLP baseline.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Inconclusive

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

Stage 1 — Mechanism check agent confidence 9/10

Built a deterministic bilinear Koopman-style controlled predictor and compared it with an additive linear transition on a system containing an action-state interaction. The paper matrix identity held to 3.6e-15 maximum absolute error; the bilinear model recovered the planted coefficients and achieved essentially zero held-out 24-step rollout error, versus additive RMSE 0.0564. This clearly demonstrates the claimed multiplicative-effect advantage on the toy system, though it does not establish performance on noisy high-dimensional world models.

Agent confidence: 9/10

Baseline: Additive linear: one-step MSE 0.0054615; 24-step held-out rollout MSE 0.00318596, RMSE 0.05644; 3 parameters.
Idea: Bilinear feature model: one-step MSE 4.04e-32; 24-step held-out rollout MSE 3.48e-32, RMSE 1.86e-16; 4 parameters. Matrix identity maximum error 3.55e-15.

Limitations: Only a scalar, noiseless synthetic plant was tested, with offline least-squares fitting and one action dimension. No neural encoder/decoder, rank-constrained latent implementation, stability constraint, training-cost comparison, MPC reward, noise robustness, or CartPole/pendulum benchmark was evaluated.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.01819) Sep 4, 2026

Ahead of the baseline (−24.6%) but not statistically significant (p=0.0991, wins 6 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: custom_bilinear_vs_additive
Paired seeds: 8
Baseline mean: 0.002
Idea mean: 0.0015
Effect (Δ): -0.0005 (−24.6%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.0991 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track was tested. Dense bilinear factors, spectral constraints, multi-step free-running loss, MPC reward, and additional control environments were not evaluated.

How to run: python3 bench_experiment.py

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

Stage 2 — Benchmark Sep 4, 2026 latest Failed ✗

Agent confidence: 9/10

Built and tested a rank-4 bilinear input-conditioned recurrent cell on the structurally matched actuated-pendulum dynamics track. It reduced mean MSE from 0.0020132 to 0.0015174 with 6/8 paired wins and confirmed an action-dependent trained-model signature, but the paired permutation p-value was 0.0991, so there is no significant win.

Baseline: mean test MSE 0.0020132122, std 0.0006065339, lr=0.01, epochs=18, 8 seeds
Idea: mean test MSE 0.0015174268, std 0.0001158377, rank=4, lr=0.01, epochs=18, 8 seeds
Paired delta: -0.0004957854
Permutation p: 0.0991

Benchmark result