⚡ electrical engineering

Structured Latent Dynamics in Wireless CSI via Homomorphic World Models

This paper presents a self-supervised JEPA-based framework that models wireless channel state information dynamics in a structured latent space using Lie algebra-derived homomorphic updates, achieving superior topology preservation and future forecasting on the DICHASUS dataset to enable scalable applications like localization and mobility-aware scheduling.

Original authors: Salmane Naoumi, Mehdi Bennis, Marwa Chafii

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Salmane Naoumi, Mehdi Bennis, Marwa Chafii

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to navigate a giant, invisible maze made of radio waves. Every time you move, the "shape" of the maze changes slightly because your position relative to walls, furniture, and other objects shifts the signal. This is the world of Wireless Communication, and the data describing these signals is called Channel State Information (CSI).

The problem? This data is messy, huge, and changes every millisecond. Trying to predict where a signal will go next is like trying to guess the path of a leaf blowing in a chaotic windstorm just by looking at a single snapshot.

This paper introduces a clever new way to solve this using a concept called a "World Model." Here is the breakdown in simple terms:

1. The Core Idea: Learning to "Dream" in Radio

Think of a World Model like a video game engine inside a computer's brain. Instead of just reacting to what it sees right now, the model builds a tiny, simplified map (a "latent space") of the world. It then uses this map to simulate the future.

The Analogy: Imagine you are playing a driving game. You don't need to see every single pebble on the road to know that if you turn the steering wheel left, the car will go left. You have an internal model of "physics."
In this paper: The AI learns an internal model of the "physics" of radio waves. It learns that if a user moves North at a certain speed, the radio signal will change in a specific, predictable way.

2. The Secret Sauce: "Homomorphic" Math (The Lego Block Approach)

Most AI models try to learn these changes by guessing and checking, which can be messy. This paper introduces a special mathematical trick called Homomorphic Dynamics based on Lie Algebra.

The Analogy: Imagine you have a set of Lego blocks.
- Standard AI: Tries to glue the blocks together in a random way to build a tower. Sometimes it falls over.
- This Paper's AI: Uses a special "Lego system" where every block has a specific shape that only fits with other blocks in a perfect, logical way. If you add a "move North" block, it snaps perfectly onto the current tower, extending it exactly where it should go.
Why it matters: This ensures that the AI's internal map stays geometrically consistent. If you move in a circle, the AI's internal map draws a circle, not a jagged mess. It respects the "rules of the universe" of radio waves.

3. How It Learns: The "Teacher-Student" Game

The model doesn't need a human teacher telling it, "You are wrong, the user is actually at point X." It learns self-supervised, like a student playing a game against a slightly older version of themselves.

The Student (Online Encoder): Looks at the current radio signal and guesses what the signal will look like in the future, based on how fast the user is moving.
The Teacher (Target Encoder): A "slow-moving" version of the student that has seen more data. It provides the "correct" answer for what the future signal actually looks like.
The Lesson: The student tries to match the teacher's prediction. Over time, the student gets so good at predicting the future that it builds a perfect, compact map of the entire wireless environment.

4. The Result: A "Channel Chart"

The end product is a Channel Chart.

The Analogy: Imagine a map of a city. Usually, you need GPS coordinates to know where you are. This AI creates a map where distance on the map equals distance in the real world, even though it never saw a GPS coordinate during training.
If two people are standing close together in a room, their "radio fingerprints" will be close together on this AI map. If they are far apart, the fingerprints are far apart.
The Magic: Because the model learned the rules of motion (the homomorphic part), it can predict where a user will be in the future just by knowing their speed and direction, even in a room it has never visited before.

Why Should You Care?

This isn't just about math; it's about making future wireless networks smarter and faster.

No More Lag: The network can "see" a user moving before they arrive and prepare the connection instantly.
Better Localization: It can pinpoint your location without needing expensive GPS hardware, just by listening to the radio waves.
Generalization: Because it learned the rules of movement rather than just memorizing specific rooms, it works in new buildings, new cities, and new environments without needing to be retrained.

In a nutshell: The authors built an AI that learns the "physics" of radio waves. Instead of just memorizing the past, it learned how to "dream" about the future, creating a perfect, geometric map of the wireless world that helps networks anticipate user movement and stay connected seamlessly.

1. Problem Statement

The paper addresses the challenge of learning predictive and structured representations of wireless channels from Channel State Information (CSI). While World Models (WMs) have proven effective in AI for simulating future states, applying them to wireless systems is difficult due to:

High-dimensional, noisy observations: CSI data is complex and voluminous.
Lack of temporal awareness: Existing Channel Charting (CC) methods (which map CSI to a low-dimensional space preserving spatial proximity) are typically static, relying on handcrafted metrics or contrastive learning without modeling temporal dynamics.
Unstructured dynamics: Prior predictive models often treat state transitions as generic neural network mappings, lacking geometric consistency and failing to generalize to unseen motion patterns or environments.

The core question is: Can a World Model's latent space serve as a meaningful, metrically faithful representation of the radio environment that supports both spatial charting and temporal prediction?

2. Methodology

The authors propose a self-supervised, predictive World Model based on the Joint Embedding Predictive Architecture (JEPA) framework, specifically tailored for wireless channels.

A. System Model & Data

Input: Wideband multi-antenna CSI tensors ( $H_t$ ) paired with user velocity ( $a_t$ ) derived from ground-truth positions.
Formulation: The problem is modeled as a Markov Decision Process (MDP) where the user's motion acts as the "action" driving the evolution of the channel state.
Preprocessing: Complex CSI is converted to real-valued tensors (magnitude and phase) and processed by a ResNet-style encoder.

B. Core Architecture: Homomorphic Latent Dynamics

The key innovation is replacing generic transition functions with homomorphic updates derived from Lie Algebra.

Encoder ( $\zeta$ ): Maps raw CSI to a latent vector $z_t$ .
Action Projection: A learnable mapping $\phi_\Theta$ projects the velocity control vector $a_t$ into a Lie algebra $\mathfrak{gl}(D)$ .
Group Element Generation: The Lie algebra element is exponentiated to form a Lie group element (a linear operator):
$\rho_\Theta(a_t) = \exp(\phi_\Theta(a_t))$
Structured Transition: Future latent states are predicted by composing these operators sequentially:
$\hat{z}_{t+h} = \left( \prod_{i=0}^{h-1} \rho_\Theta(a_{t+i}) \right) z_t$
This ensures that the latent dynamics are compositional (order of actions matters algebraically) and geometrically consistent (smooth transformations).

C. Training Objective (Self-Supervised)

The model is trained end-to-end without spatial labels using a Teacher-Student setup (JEPA style):

Teacher: An Exponential Moving Average (EMA) of the online encoder generates target embeddings ( $\bar{z}$ ).
Student: Predicts future embeddings ( $\hat{z}$ ) based on the current state and action sequence.
Loss Function ( $L_{total}$ ): A weighted sum of five components:
1. Teacher-Forcing Loss ( $L_{TF}$ ): Aligns predicted steps with targets.
2. Rollout Loss ( $L_{roll}$ ): Supervises the final horizon step to ensure long-term consistency.
3. VICReg Regularization ( $L_{var}, L_{cov}$ ): Prevents representation collapse by enforcing variance and decorrelation among latent dimensions.
4. Inverse Dynamics Loss ( $L_{IDM}$ ): Encourages the latent space to encode motion information by predicting the action from consecutive latent states.

3. Key Contributions

Unified Predictive World Model: A self-supervised framework that learns channel charting and temporal dynamics simultaneously from raw CSI, eliminating the need for pre-training or handcrafted dissimilarity metrics.
Homomorphic Latent Transitions: The first application of Lie algebra parameterizations to wireless channel modeling. This enforces a structured, compositional geometry in the latent space, allowing the model to generalize to arbitrary action sequences.
Metrically Faithful Channel Charting: Demonstrates that the learned encoder acts as a Forward Charting Function (FCF), mapping high-dimensional CSI to a latent space where Euclidean distances correlate with physical spatial displacements.
Generalization: The model successfully predicts future channel states and reconstructs trajectories in unseen environments without retraining.

4. Experimental Results

Dataset: Evaluated on the DICHASUS dataset (indoor wideband measurements with 4 base stations and 32 antennas).
Baselines: Compared against Multi-Layer Perceptron (MLP), Gated Recurrent Unit (GRU), and Feature-wise Linear Modulation (FiLM) predictors.
Metrics: Used standard Channel Charting metrics:
- Trustworthiness (TW) & Continuity (CT): Measure local structure preservation (Higher is better).
- Kruskal's Stress (KS) & Rajski's Distance (RD): Measure global topological alignment (Lower is better).
Performance:
- The proposed method achieved TW = 0.9948 and CT = 0.9764, outperforming all baselines.
- It significantly reduced global distortion (RD = 0.7473) compared to GRU (0.7818) and MLP (0.7828).
- Ablation Study: Removing the homomorphic structure (using generic MLPs) or regularization (VICReg/IDM) caused a sharp drop in performance, confirming the necessity of both geometric constraints and motion-aware supervision.
Visualization: PCA projections of the latent rollouts closely matched ground-truth physical trajectories, even for test paths never seen during training.

5. Significance and Future Impact

Scalable Foundation: This work provides a principled, scalable foundation for wireless representation learning, moving beyond static mapping to dynamic, predictive modeling.
Downstream Applications: The resulting "predictive channel charts" enable robust downstream tasks such as:
- Mobility-aware scheduling: Anticipating channel changes before they happen.
- Localization: Inferring user position from CSI without explicit GPS labels.
- Beam Tracking & Handover: Proactive resource management based on simulated future states.
Theoretical Advance: By bridging World Models and Wireless Physics via Lie groups, the paper demonstrates that enforcing algebraic structure on latent dynamics leads to superior generalization and geometric fidelity in complex physical environments.

The authors conclude that while the current work assumes a static environment with user-motion-driven dynamics, this framework is a critical precursor for developing autonomous, adaptive wireless systems capable of reasoning about complex, time-varying radio scenes.