# Multiplicative Manifold Unscented Recurrent Cell

- ID: 675
- Canonical URL: https://synthcore.org/idea/675/multiplicative-manifold-unscented-recurrent-cell
- API JSON: https://synthcore.org/api/idea/675.json
- API Markdown: https://synthcore.org/api/idea/675.md
- Verification status: beats_baseline
- Source: [arXiv:2607.10072](https://arxiv.org/abs/2607.10072)
- Category: architecture
- Solves: stability, accuracy
- ML areas: rnn, ssm, world-model
- Math tags: geometry, dynamical-systems, probability, statistics, numerical-analysis
- Ratings: usefulness 7/10; difficulty 6/10; novelty 7/10

## Idea description

Replace the Euclidean hidden-state update of a recurrent or state-space neural network with a mixed manifold state containing a rotation component and Euclidean features. Propagate uncertainty with sigma points in tangent error coordinates, retract rotational perturbations through the exponential map, and compute the training loss from the predicted covariance. This avoids invalid rotations and captures second-order curvature effects that a first-order EKF-style recurrent cell misses at large time steps.

## Mathematical statement

Let the latent state be x = [q, z], where q is a unit quaternion on S3 and z is a d-dimensional Euclidean feature vector. Represent local uncertainty by e = [delta_theta, delta_z] in R^(3+d), with covariance P. A rotational perturbation is retracted as q = delta_q(delta_theta) multiplied by q_hat, where delta_q is the quaternion exponential of the rotation vector. For tangent dimension n = 3 + d, generate symmetric sigma points from P, retract each point to the manifold, propagate every point through the learned transition F_phi, and compute the output covariance from weighted tangent residuals. Q is a positive semidefinite process-noise covariance. The unscented construction is the mechanism that retains second-order variation under nonlinear propagation, while the multiplicative retraction preserves the rotation constraint exactly.

## Key formulas

- q_i = delta_q(delta_theta_i) * q_hat,     z_i = z_hat + delta_z_i
- delta_q(delta_theta) = [cos(||delta_theta||/2), sinc(||delta_theta||/2) delta_theta/2],     sinc(u) = sin(u)/u
- e_0 = 0,     e_i = column_i(sqrt((n + lambda) P)),     e_(i+n) = -e_i,     lambda = alpha^2 (n + kappa) - n
- P_plus = Q + sum_i W_i_c e_i_plus e_i_plus^T,     W_0_c = lambda/(n+lambda) + 1 - alpha^2 + beta,     W_i_c = 1/(2(n+lambda))

## Implementation notes

1. Integration point: use this as a replacement for the hidden update in a GRU-like recurrent model, a learned Kalman filter, or a latent state-space model. Reserve three latent coordinates for a quaternion and keep the remaining d coordinates Euclidean. Let a small MLP produce a deterministic transition and process covariance from the current latent state and input: (x_bar, log_Q) = F_phi(x_t, u_t). 2. Pseudocode: maintain q_hat, z_hat, and P. Compute L = Cholesky((n + lambda)P), create the zero sigma point and the positive and negative columns of L, and split every tangent point into delta_theta and delta_z. Retract the rotational part using delta_q(delta_theta) multiplied by q_hat, add delta_z to z_hat, and run the same neural transition on every sigma point. Estimate the quaternion mean by selecting an initial sigma-point quaternion and iteratively averaging weighted logarithmic rotation errors relative to that mean. Compute Euclidean means directly, express every propagated quaternion as a tangent residual relative to the mean quaternion, form P_plus with the displayed covariance equation, and add a softplus-parameterized Q. Normalize quaternions after neural propagation. Train with next-step negative log likelihood, supervised state loss, or both. 3. Computed quantities are sigma points, Cholesky factors, quaternion retractions, tangent logarithms, and covariance updates. The parameters alpha, beta, and kappa are chosen hyperparameters; Q can initially be a learned diagonal matrix and later be calibrated from residuals. If backpropagation through Cholesky is unstable, stop gradients through P for the first experiment. 4. First cheap experiment: create synthetic rigid-body sequences with known quaternion rotations plus a 32-dimensional Euclidean latent signal. Compare an ordinary GRU, an additive UKF-GRU, and this multiplicative UKF-GRU at transition intervals Delta_t in {1, 2, 4, 8, 16}. Measure quaternion geodesic error, covariance NEES = e_t^T P_t^(-1) e_t, the fraction of invalid quaternions, and the fraction of divergent runs. The quantitative prediction is that the multiplicative model preserves unit-quaternion validity exactly and maintains NEES near the tangent dimension n at larger Delta_t. The additive or first-order baseline should show a sharp consistency failure when the one-sigma rotational spread sqrt(trace(P_theta_theta)) becomes approximately 0.3 radians or larger; the measured divergence boundary should shift with this curvature scale.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built quaternion exponential/logarithm operations, multiplicative sigma-point retraction, iterative quaternion averaging, tangent covariance estimation, and a first-order control. The manifold method preserved unit quaternion norms to 2.22e-16 and reduced propagated-mean error from 0.04178 to 0.00180 on average for sigma <= 0.4. At sigma=0.8 it degraded, so the benefit is real but limited to moderate uncertainty.

### Mechanism check

- Verdict: Built quaternion exponential/logarithm operations, multiplicative sigma-point retraction, iterative quaternion averaging, tangent covariance estimation, and a first-order control. The manifold method preserved unit quaternion norms to 2.22e-16 and reduced propagated-mean error from 0.04178 to 0.00180 on average for sigma <= 0.4. At sigma=0.8 it degraded, so the benefit is real but limited to moderate uncertainty.
- Confidence: 8/10
- Limitations: No trainable GRU/MLP, sequence forecasting, learned process noise, NEES evaluation, GPU experiment, or runtime/FLOP comparison was tested. The experiment uses a synthetic quadratic tangent transition and a fixed UT configuration.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0317
- Paired wins: 6/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/296/bench_report.json)
- [bench_stage2.py](https://synthcore.org/code/296/bench_stage2.py)
- [manifold_ukf_experiment.py](https://synthcore.org/code/296/manifold_ukf_experiment.py)
- [report.md](https://synthcore.org/code/296/report.md)
- [report_bench_2026-09-02T161517.md](https://synthcore.org/code/296/report_bench_2026-09-02T161517.md)
- [results.json](https://synthcore.org/code/296/results.json)
- [Download all files as ZIP](https://synthcore.org/download/296)

## Disclaimer

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