Prerequisites & Notation

Before You Begin

Modern wireless receivers increasingly benefit from machine learning: deep networks that learn channel estimators, detectors, and resource allocators from data. OTFS's DD-domain structure — sparse, geometric, low-dimensional — is ideally suited to ML. This chapter develops the ML toolkit for OTFS: deep learning receivers that exploit DD sparsity, learned pilot patterns that outperform hand-designed ones, and model-based deep unfolding that fuses the physics of OTFS with NN expressivity.

OTFS modulation and detection(Review OTFS Ch. 6, 8)
Self-check: Can you explain the DD input-output relation and standard MP detection?
Channel estimation (embedded pilots)(Review OTFS Ch. 7)
Self-check: Do you recall how pilot-based estimation extracts path parameters?
Fractional Doppler and off-grid effects(Review OTFS Ch. 10)
Self-check: Are you familiar with the fractional Doppler penalty?
ML fundamentals (deep NN, backprop, loss functions)(Review ITA Ch. 28; Sci-Python Ch. 26-30)
Self-check: Do you know how to train a feedforward or conv NN?

Notation for This Chapter

ML-specific symbols introduced here.

Symbol	Meaning	Introduced
$\hat{\mathbf{h}}_{\theta}$	Neural network channel estimator with parameters $\theta$	s01
$\hat{\mathbf{x}}_{\theta}$	NN detector output	s01
$\mathcal{L}(\theta)$	Training loss (MSE or cross-entropy)	s01
$\mathbf{p}_{\theta}$	Learned pilot pattern, parameters $\theta$	s02
$T_{\text{iter}}$	Number of unfolded iterations	s03
$\eta$	Learning rate (Adam, RMSprop)	s04

← Ch 20 Deep Learning for OTFS Receivers