# Geometrically Random Transport Network

- ID: 1544
- Canonical URL: https://synthcore.org/idea/1544/geometrically-random-transport-network
- API JSON: https://synthcore.org/api/idea/1544.json
- API Markdown: https://synthcore.org/api/idea/1544.md
- Verification status: unverified
- Source: [arXiv:2607.28579](https://arxiv.org/abs/2607.28579)
- Category: architecture
- Solves: stability, accuracy, scalability
- ML areas: ssm, rnn, transformer, initialization, training-dynamics
- Math tags: dynamical-systems, spectral-theory, random-matrix, graph-theory
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Construct a deep sequence model as a layered channel network with fixed random K-regular connections between neighboring depth layers, instead of dense or independently random weight matrices. Use norm-preserving edge normalization and a reversible residual update so that geometric randomness controls information transport while trainable nonlinear readouts provide task-specific computation. The architecture exposes a tunable crossover between quasi-one-dimensional ballistic or localized propagation and wide-layer diffusive propagation.

## Mathematical statement

The paper uses a Hamiltonian on layers indexed by x, with L_y channels per layer: H = J sum over x of the coupling between neighboring vectors through T^(x,x+1), plus its Hermitian conjugate. Here J is the uniform coupling, each layer vector has L_y components, and every binary matrix T^(x,x+1) has exactly K ones in every row and column. Each node therefore has K links to each adjacent layer, for total graph connectivity 2K. The transferable mechanism is that geometric randomness and the ratio L_y/K change transport: extensive L_y supports diffusive propagation and strong spectral level repulsion, while the quasi-one-dimensional limit supports localized and ballistic sectors. The neural analogue uses A_x = T^(x,x+1)/sqrt(K) in a reversible residual propagation operator. For a perturbation localized at layer x0, define p_x(t) as the normalized squared activation norm in layer x and M2(t) as its spatial second moment. The predicted signatures are M2 proportional to t squared for ballistic transport, M2 proportional to t for diffusion, and bounded M2 for localization.

## Key formulas

- H = J sum_(x=1)^(L_x-1) [Psi_(x+1)† T^(x,x+1) Psi_x + Hermitian conjugate], with T^(x,x+1) in {0,1}^(L_y by L_y), row sum K, and column sum K.
- A_x = T^(x,x+1)/sqrt(K), and h_x(t+1) = h_x(t) - i epsilon [A_(x-1)^T h_(x-1)(t) + A_x h_(x+1)(t)].
- p_x(t) = ||h_x(t)||^2 / sum_z ||h_z(t)||^2, bar_x(t) = sum_x x p_x(t), and M2(t) = sum_x (x - bar_x(t))^2 p_x(t).
- M2(t) is proportional to v^2 t^2 in the ballistic regime, to 2 D t in the diffusive regime, and approaches a constant in the localized regime.

## Implementation notes

1. Integration point: implement a recurrent or state-space backbone with L_x depth or time layers and L_y channels per layer. Sample one independent K-regular bipartite matrix T_x between every neighboring pair of layers, keep the binary masks fixed throughout training, and use A_x = T_x divided by sqrt(K). The simplest block is a reversible transport step followed by a learned pointwise real MLP and a task readout. For ordinary real-valued networks, represent the real and imaginary parts as two channel groups, or replace the imaginary update by a two-sided orthogonal real rotation. Initialize the residual strength epsilon to a small value and use a learned scalar gate initialized at zero. 2. Pseudocode: sample each T_x with K ones per row and column; construct sparse A_x; compute q_x = A_(x-1)^T h_(x-1) + A_x h_(x+1); update h_x <- h_x - i epsilon q_x; apply the pointwise gated nonlinearity; train the gates, nonlinear blocks, and readout. Optionally unfreeze edge weights later while preserving the K-regular mask. Estimate the largest Jacobian singular value on minibatches and reduce epsilon whenever the value exceeds one. 3. Computed from the paper: the layered Hamiltonian structure, uniform coupling, and K-regular geometric mask. Estimated empirically: the transport exponent alpha from a log-log fit to M2(t), the localization plateau, gradient propagation, and level-spacing statistics of the linearized block. 4. First cheap experiment: use L_x = 128, widths L_y in {4, 8, 16, 32, 64, 128}, and K in {1, 2, 4}. Compare the proposed mask against a dense residual SSM and an unstructured sparse mask on copying, delayed parity, and sequential MNIST. Before training, inject a one-channel impulse and measure M2 for 100 propagation steps. The quantitative prediction is that narrow fixed-K networks show an exponent alpha near 2 or a localization plateau, whereas increasing L_y at fixed K produces a crossover toward alpha near 1. Require the fitted exponent to differ by at least 0.3 between narrow and wide settings, and test whether this same crossover predicts the depth at which gradient norms cease to be localized or ballistic.

## Disclaimer

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