Unverified 2026

Invariant Möbius latent mixer

Implementation & benchmark of arXiv:2607.03354 — Dynamics of integer zeroes of homogeneous quadratic equations over $\mathbb{R}^3$

Usefulness5/10

Difficulty5/10

Novelty7/10

Architecture Stability Sample efficiency Accuracy Vae Diffusion Embedding Training dynamics Dynamical systems Geometry Algebra Probability

Source paper: Dynamics of integer zeroes of homogeneous quadratic equations over $\mathbb{R}^3$ arXiv:2607.03354 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Insert a piecewise Möbius transformation as a deterministic latent mixing layer, using the paper's exact branch structure rather than a generic unconstrained MLP. The transformation repeatedly moves points between branches while preserving a known reference density, creating a cheap chaotic mixer with analytically computable Jacobian factors. Use a truncated, normalized version in practice so that the sigma-finite invariant measure becomes a valid finite training distribution.

Formulas

$$T_{2}(t)=\begin{cases}\frac{t}{1-3t},&0<t<\frac14,\\[2pt]\frac{1-3t}{t},&\frac14<t<\frac13,\\[2pt]\frac{3t-1}{1-2t},&\frac13<t<\frac25,\\[2pt]\frac{1-2t}{3t-1},&\frac25<t<\frac12,\\[2pt]\frac{2t-1}{t},&\frac12<t<1.\end{cases}$$

$$d\mu(t)=\frac{1}{t(1-t)}\,dt,\qquad \rho(y)=\sum_{x\in T^{-1}(y)}\frac{\rho(x)}{|T'(x)|}.$$

$$T(t)=\frac{at+b}{ct+d}\quad\Longrightarrow\quad T'(t)=\frac{ad-bc}{(ct+d)^2},$$

$$\rho_{\varepsilon}(t)=\frac{1}{2\log((1-\varepsilon)/\varepsilon)}\frac{1}{t(1-t)},\qquad t\in[\varepsilon,1-\varepsilon].$$

Mathematical statement

The paper gives the explicit five-branch map $T_{2}:(0,1)\to(0,1)$, with $T_{2}(t)=t/(1-3t)$ for $0<t<1/4$, $T_{2}(t)=(1-3t)/t$ for $1/4<t<1/3$, $T_{2}(t)=(3t-1)/(1-2t)$ for $1/3<t<2/5$, $T_{2}(t)=(1-2t)/(3t-1)$ for $2/5<t<1/2$, and $T_{2}(t)=(2t-1)/t$ for $1/2<t<1$. Here $t$ is a scalar latent coordinate and each branch is a fractional-linear map $T(t)=(at+b)/(ct+d)$. The stated invariant measure is $d\mu(t)=dt/[t(1-t)]$, meaning that the density $\rho(t)=1/[t(1-t)]$ satisfies the Frobenius-Perron invariance equation $\rho(y)=\sum_{x\in T^{-1}(y)}\rho(x)/|T'(x)|$ wherever all inverse branches exist. For a Möbius branch, $T'(t)=(ad-bc)/(ct+d)^2$, so both the transformed coordinate and the log-Jacobian are available exactly. Because $\mu$ is sigma-finite rather than a probability measure, define a finite implementation domain $I_{\varepsilon}=[\varepsilon,1-\varepsilon]$ and use the normalized density $\rho_{\varepsilon}(t)=\rho(t)/Z_{\varepsilon}$ with $Z_{\varepsilon}=\int_{\varepsilon}^{1-\varepsilon}dt/[t(1-t)]=2\log((1-\varepsilon)/\varepsilon)$.

Implementation notes

1. Integration point: apply the mixer independently to one or more normalized latent coordinates immediately after a VAE encoder, diffusion-noise projection, or flow coupling layer. Keep the other coordinates unchanged initially so the effect is isolated. Choose $\varepsilon=10^{-3}$ and map each unconstrained latent scalar $z$ to $t=\varepsilon+(1-2\varepsilon)\sigma(z)$, where $\sigma$ is the logistic function. 2. Pseudocode: receive $t$; select branch $k$ from the five intervals in the formula; compute $y=(a_k t+b_k)/(c_k t+d_k)$; if numerical roundoff leaves the truncated interval, clamp to $[\varepsilon,1-\varepsilon]$; return $y$. For an invertible flow, store the branch index and add $\log|dy/dt|+\log|dt/dz|$ to the log-density, using $dy/dt=(a_kd_k-b_kc_k)/(c_kt+d_k)^2$. Apply the map for $m=2$ to $8$ iterations, optionally alternating with a learned permutation across latent dimensions. 3. Mathematics versus estimation: the branch formulas, derivatives, and reference density are computed exactly; estimate mixing quality through empirical autocorrelation, effective sample size, held-out likelihood, and gradient-norm variance. Truncation slightly violates exact invariance at the boundaries, so record boundary hits and compare against larger $\varepsilon$ values. 4. First experiment: train 2D RealNVP models on two-moons and eight-Gaussians data, comparing identity/permutation, learned affine coupling, and the fixed Möbius mixer at equal parameter count. Measure negative log-likelihood, sample coverage, lag-1 autocorrelation after repeated latent mixing, and wall-clock cost. Success means lower autocorrelation and equal-or-better likelihood with no additional trainable parameters; target at least 20% faster decorrelation per scalar operation.

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.