Source-linked AI summary
JEPA-kMMVAE: CSI Prediction for Time-Varying Wireless Channels
Efstathios Chatziloizos, Konstantinos Vandikas, Aneta Vulgarakis Feljan, Zheng Chen, Nikolaos Pappas
TL;DR
Time-varying CSI can become outdated before use, and prediction from noisy history must distinguish channel evolution from observation noise. JEPA-kMMVAE combines temporally structured latent prediction with direct or residual CSI decoding. Residual prediction has the lowest mean NMSE at moderate and high input SNRs, while sequence baselines are more accurate under stronger noise.
Problem
The paper addresses how to predict future CSI from noisy historical observations while exploiting temporal dependence and distinguishing channel evolution from measurement noise.
Source: §1 · p. 1, §2.1 · p. 2, §2.2 · p. 2Method
JEPA-kMMVAE combines a pretrained k-MMVAE representation, temporal latent prediction, a future-CSI decoder, and clean-target latent supervision, with direct and residual variants.
Source: §1 · p. 1, §3.3 · p. 3Results
Residual JEPA-kMMVAE has the lowest mean NMSE at 15–25 dB, beating the strongest baseline by 1.70 dB at 20 dB and 1.94 dB at 25 dB, while Transformer and GRU have the lowest means at 5 and 10 dB, respectively.
Source: §4.2 · p. 4Takeaways & Limitations
The direct–residual crossover highlights the importance of observation reliability when the latest snapshot is used as a prediction reference.
Source: §4.2 · p. 4, §5 · p. 4The narrowband formulation applies per OFDM subcarrier and does not account for inter-subcarrier correlation.
Source: §2.1 · p. 2
Abstract
from arXiv · showhide
This paper considers channel state information (CSI) prediction using neural networks for time-varying multi-antenna channels from noisy historical observations. The main challenge is to exploit temporal correlation while distinguishing channel evolution from observation noise. We propose JEPA-kMMVAE, which combines a joint-embedding predictive architecture with a k-memory Markov variational autoencoder to predict future CSI through temporally structured latent representations. The framework supports direct prediction and residual corrections relative to the latest latent state and noisy CSI observation. Simulation results show that the residual configuration outperforms the evaluated baselines at moderate and high input CSI signal-to-noise ratios, while sequence baselines remain more accurate under stronger noise.
1. INTRODUCTION
Time-varying CSI can become outdated before use, making prediction from history relevant to timeliness while requiring channel evolution to be distinguished from observation noise. JEPA-kMMVAE extends temporally structured variational representations to causal next-CSI prediction from noisy history.
1. INTRODUCTION
- The central challenge is exploiting temporal dependence while distinguishing channel evolution from noise in the available observations.CSI may become outdated between pilot-based acquisition and use in multi-antenna systems.Source: §1 · p. 1
- Prior work used k-MMVAE for time-varying channel estimation, adapted it to context-aware CSI compression [5], and explored Transformer prediction and masked CSI pretraining.JEPA-kMMVAE extends temporally structured variational representations to one-step prediction without an observation at the target time.Source: §1 · p. 1
- JEPA latent supervision predicts target representations from context representations, while residual learning motivates correcting an available observation for noise and channel evolution.The paper combines these ideas with the CSI-domain objective.Source: §1 · p. 1
- JEPA-kMMVAE predicts the next noise-free CSI snapshot from noisy history using a pretrained k-MMVAE representation, temporal latent prediction, and direct or residual CSI decoding.A clean target branch provides latent supervision during training.Source: §1 · p. 1
2. SYSTEM MODEL AND PREDICTION PROBLEM
The paper formulates causal CSI prediction as estimating the next clean channel snapshot from a finite history of noisy observations. It models time-varying multi-antenna channels and evaluates predictions under squared error and NMSE.
2. SYSTEM MODEL AND PREDICTION PROBLEM
- A finite history of noisy CSI measurements is used to predict the next clean channel snapshot one sampling interval ahead.The future target is available only for training and evaluation, not as prediction input.Source: §2 · p. 1, Figure 1 · p. 2
2.1. Time-Varying Channel Observations
- The model considers a base station with R antennas and a single-antenna user; the narrowband formulation applies per OFDM subcarrier and omits inter-subcarrier correlation.Successive channel snapshots are correlated through propagation-environment and terminal-position evolution.Source: §2.1 · p. 2
- CSI observations differ from the true channel because of estimation errors and measurement noise, modeled as additive noise with a common variance within each window.For neural processing, real and imaginary channel components are stacked, preserving squared channel errors; input SNR is defined using the latest snapshot.Source: §2.1 · p. 2
2.2. Causal CSI Prediction
- At time t, the predictor uses only context Y_t to estimate x_{t+1}; the squared-error risk averages over channel evolution and observation noise.The unrestricted finite-second-moment minimizer is the conditional mean E[x_{t+1} | Y_t], and evaluation reports 10 log10(NMSE) in dB.Source: §2.2 · p. 2
3. PROPOSED PREDICTION FRAMEWORK
JEPA-kMMVAE combines a pretrained k-MMVAE representation, a trainable latent predictor, and a future-CSI decoder; its residual configuration uses noisy history for prediction and clean targets for training supervision.
3.1. Temporally Structured Channel Representation
- The k-MMVAE uses first-order Markov latent dynamics and finite-memory inference; k sets explicit preceding observations, while the previous latent carries recursively obtained information.Its backbone is pretrained to reconstruct the clean terminal snapshot from noisy context, then frozen; posterior means are propagated recursively for prediction.Source: §3.1 · pp. 2–3
3. PROPOSED PREDICTION FRAMEWORK
- The residual framework predicts from noisy history, while a clean-target branch supervises latent prediction during training and the CSI output is supervised by the clean future snapshot.The clean target branch is not part of the prediction path.Source: §3 · p. 2, Figure 2 · p. 3
3.2. Latent Prediction and Residual CSI Decoding
- The residual decoder combines a temporal correction to the latest inferred latent with the last k context latents to correct the latest noisy CSI observation.The correction addresses both observation noise and channel evolution; a zero correction recovers persistence, and both residual references are available at inference.Source: §3.2 · p. 3
3.3. Clean Target Embeddings and Training Objective
- A frozen copy of the context encoder forms clean target embeddings from shifted future CSI, and stop-gradient prevents updates to that target branch.The CSI loss supervises prediction accuracy, while the JEPA loss aligns predicted and clean target representations; both losses train the latent predictor and decoder.Source: §3.3 · p. 3
4. SIMULATION RESULTS
Residual JEPA-kMMVAE is strongest at 15–25 dB, while Transformer and GRU lead at 5 and 10 dB; the direct–residual ranking also changes with input SNR.
4.1. Experimental Setup
- QuaDRiGa 3GPP 38.901 UMa simulations use 2.1 GHz channels, 32 antennas, 100 snapshots spaced 0.5 ms, and speeds centered at 1, 10, and 30 m/s.Source: §4.1 · p. 3
4.2. Results and Discussion
- 1.70 dB at 20 dB and 1.94 dB at 25 dB are residual JEPA-kMMVAE’s advantages over the strongest baseline; Transformer and GRU lead at 5 and 10 dB.The advantage is described as modest at 15 dB.Source: Table 1 · p. 4, §4.2 · p. 4
4.1. Experimental Setup
- Evaluation uses 8 context observations, k = 3 and d = 64 for JEPA-kMMVAE, and all 92 rolling windows from each test trajectory.Training, validation, and test sets contain 8,000, 800, and 800 trajectories; learned predictors train on SNRs sampled uniformly from 0–25 dB.Source: §4.1 · p. 4
4.2. Results and Discussion
- Direct prediction is 1.18 dB better at 5 dB, but residual prediction leads by 4.08 dB at 25 dB.The ordering reverses by 10 dB.Source: Figure 3 · p. 4, §4.2 · p. 4
- At weak observation noise, the latest snapshot is a useful reference; with stronger noise, the decoder must also remove corruption in that reference.Source: §4.2 · p. 4
5. CONCLUSION
The conclusion presents JEPA-kMMVAE for causal CSI prediction using temporally structured latent representations. Its residual configuration leads at moderate and high observation SNRs, while learned baselines remain more accurate under stronger noise.
5. CONCLUSION
- JEPA-kMMVAE extends a pretrained Markov variational encoder with latent forecasting and direct or residual CSI decoding for causal prediction.Source: §5 · p. 4
- The evaluated residual configuration achieves the lowest mean NMSE at moderate and high observation SNRs, while learned baselines remain more accurate under stronger noise.Source: §5 · p. 4
- Future work will investigate noise-aware combinations of direct and residual prediction and extend the framework to multi-step channel forecasting.Source: §5 · p. 4