Unverified 2026

Lorentzian SU(2) recurrent flow

Implementation & benchmark of arXiv:2607.11592 — The Lorentzian Problem on the Group $SU(2)$

Usefulness6/10

Difficulty5/10

Novelty6/10

Dynamics Stability Accuracy Memory Rnn Ssm Training dynamics Initialization Lie groups Differential geometry Dynamical systems Geometry Linear algebra

Source paper: The Lorentzian Problem on the Group $SU(2)$ arXiv:2607.11592 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Use the paper's explicitly solved SU(2)-based extremal flow as a structured recurrent transition instead of learning an unconstrained dense recurrent matrix. The transition has only two scalar parameters, a radius/frequency r and phase phi, while its rotating coefficient pattern continuously mixes four real state coordinates and can be integrated with a norm-preserving Cayley transform.

Formulas

$$\operatorname{SU}(2)=\{q\in\mathbb{C}^{2\times 2}\mid qq^{*}=q^{*}q=\operatorname{Id},\ \det q=1\},\qquad \mathfrak{su}(2)=\{X\in\mathbb{C}^{2\times 2}\mid X+X^{*}=0,\ \operatorname{tr}X=0\}.$$

$$\begin{cases}\dot{x}_{1}=\frac{r_{0}}{2}\left(y_{1}\sin(-2r_{0}t+\varphi_{0})+x_{2}\cos(-2r_{0}t+\varphi_{0})+y_{2}\right),\\ \dot{y}_{1}=\frac{r_{0}}{2}\left(-x_{1}\sin(-2r_{0}t+\varphi_{0})-x_{2}+y_{2}\cos(-2r_{0}t+\varphi_{0})\right),\\ \dot{x}_{2}=\frac{r_{0}}{2}\left(-x_{1}\cos(-2r_{0}t+\varphi_{0})+y_{1}-y_{2}\sin(-2r_{0}t+\varphi_{0})\right),\\ \dot{y}_{2}=\frac{r_{0}}{2}\left(-x_{1}-y_{1}\cos(-2r_{0}t+\varphi_{0})+x_{2}\sin(-2r_{0}t+\varphi_{0})\right).\end{cases}$$

$$X_{k+1}=\left(I-\frac{h}{2}A\left((k+\tfrac12)h;r,\varphi\right)\right)^{-1}\left(I+\frac{h}{2}A\left((k+\tfrac12)h;r,\varphi\right)\right)X_k.$$

$$\|X_{k+1}\|_{2}=\|X_k\|_{2}\quad\text{when }A(t;r,\varphi)^T=-A(t;r,\varphi).$$

Mathematical statement

The underlying group is SU(2) = {q in C^{2x2}: qq^* = q^*q = I, det(q)=1}, with Lie algebra su(2) = {X: X+X^*=0, tr(X)=0}; this supplies skew-Hermitian, norm-preserving infinitesimal generators. The extracted abnormal-extremal equations define a four-dimensional real state X=(x_1,y_1,x_2,y_2)^T, with radius r_0 and phase phi_0, through the time-periodic linear system shown below. Here r_0 >= 0 is the trajectory frequency, phi_0 is an initial phase, t is continuous time, and A(t;r_0,phi_0) is the real coefficient matrix read from the four equations. The coefficients are sinusoidal with angular frequency 2r_0 and are arranged according to the SU(2) geometry. We use the Cayley discretization X_{k+1}=(I-hA_k/2)^{-1}(I+hA_k/2)X_k, where h is the integration step and A_k=A((k+1/2)h;r,phi). For a skew-symmetric A_k, the Cayley map is orthogonal, so it preserves ||X_k||_2 without eigenvalue clipping.

Implementation notes

First integrate the construction into the hidden-to-hidden transition of a vanilla RNN or a linear state-space model. Partition the hidden state into blocks of four, X=(x_1,y_1,x_2,y_2)^T. For each block, parameterize r=softplus(r_raw) and phi=pi*tanh(phi_raw), and use an input projection W_in u_t as an additive drive. With step size h=1 initially, compute tau=k+1/2, form the 4x4 matrix A(tau;r,phi) from the four displayed differential equations, compute C=solve(I-hA/2, I+hA/2), and update X_{t+1}=C X_t + W_in u_t. Concatenate blocks and apply an output projection; optionally add a residual input pathway or a gated additive drive, but do not unconstrain C. The paper-derived quantities are the sinusoidal coefficient pattern and Cayley flow. Learned quantities are r, phi, W_in, the output projection, and optional input/residual gates. Log hidden norms, gradient norms, and the singular values of C over long sequences. Begin with adding problems, copy/retrieval tasks, and sequential MNIST using a small model, comparing against tanh RNNs and standard orthogonal/unitary RNNs at matched hidden size and FLOPs. Test lengths 100, 500, and 2000. Success requires lower hidden-norm drift, fewer exploding-gradient events, and better long-sequence validation accuracy or loss. Ablate Cayley versus Euler integration and fixed versus learned r and phi; reject the approach if it has no advantage over an equally sized orthogonal RNN.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.