Pilotless Communication

A technique where pilot and data symbols are transmitted simultaneously by arithmetically adding them at the transmitter. The receiver uses a neural network to separate the known training sequence from the unknown data, enabling continuous channel tracking without sacrificing a dedicated time-frequency resource block.

• Eliminates pilot overhead entirely
• Enables tracking of fast-fading channels
• Requires careful power allocation between data and training components

The transmitter learns a constellation geometry and encoding scheme that is inherently robust to channel distortion without explicit pilot insertion. The receiver is jointly trained to perform blind detection by learning the statistical structure of the transmitted signal, treating the channel as a nuisance variable to be marginalized out.

• No explicit training symbols required
• Constellation points are optimized for blind separability
• Relies on high-dimensional learned manifolds

An end-to-end learned transceiver designed to operate without any channel state information (CSI) at the receiver. The neural network learns an encoding manifold that is invariant to common channel impairments like phase rotation and amplitude scaling, enabling reliable detection in scenarios where traditional coherent demodulation fails.

• Robust to phase noise and Doppler shift
• Ideal for high-mobility mmWave links
• Learns a channel-agnostic symbol mapping

A deep learning receiver that performs joint channel equalization and symbol detection directly from the received signal sequence without a separate channel estimation step. The network implicitly learns to invert channel effects using only the statistical properties of the transmitted signal, such as constant modulus or finite alphabet constraints.

• Replaces pilot-based channel estimation
• Operates on raw I/Q sample streams
• Adapts to channel variations via online learning

A learned extension of classical differential phase-shift keying where a neural network encodes information in the relative transition between consecutive symbols rather than absolute constellation points. The receiver decodes by comparing adjacent received symbols, making the system inherently immune to slow-varying channel phase without any pilot overhead.

• Inherent phase ambiguity resolution
• No channel estimation required
• Suitable for continuous-phase modulation schemes

The transmitter learns to map source bits to a high-dimensional embedding space where the geometric structure is preserved under channel distortions. The receiver learns a corresponding demapping function that is invariant to specific channel realizations, effectively factoring out the channel's impact through the learned representation geometry.

• Invariant representation learning
• Generalizes across unseen channel conditions
• Enables zero-shot adaptation to new environments

An end-to-end learned transceiver designed to operate without explicit channel state information (CSI). The neural network learns robust representations that are invariant to unknown channel phase and amplitude variations.

• Key Mechanism: Trains on a distribution of random channel realizations, forcing the latent space to encode information in a way that survives unknown multiplicative distortions
• Contrast: Unlike a standard channel autoencoder, this architecture does not receive CSI as a side input at the receiver
• Application: Enables robust communication over rapidly varying channels where pilot-based estimation would consume prohibitive overhead

A neural network receiver that jointly performs channel equalization and symbol detection directly from a received signal sequence without requiring a separate pilot-based channel estimation step.

• Implicit Learning: The network learns to invert channel effects by exploiting statistical structure in the transmitted signal, such as constant modulus or finite alphabet properties
• Architecture: Typically uses convolutional or recurrent layers to process temporal sequences of received symbols
• Advantage: Eliminates the spectral efficiency loss associated with dedicated pilot symbols while maintaining robustness to multipath fading

A classical precursor to pilotless communication where a low-power training sequence is arithmetically added to the data signal rather than occupying dedicated time or frequency slots.

• Spectral Efficiency: Recovers the throughput lost to pilot overhead by transmitting data and training simultaneously
• Power Allocation Trade-off: A fraction of transmit power is diverted to the known training sequence; the receiver uses this to estimate the channel and then subtracts the training contribution before decoding
• Neural Extension: Modern systems replace the linear training sequence with a learned, non-linear superimposed embedding optimized end-to-end with the receiver

A technique where the receiver neural network internalizes channel estimation as a latent computation rather than producing an explicit channel matrix.

• Mechanism: The network's hidden layers learn to condition symbol decisions on the received signal's structure without ever outputting an intermediate channel estimate
• Gradient Flow: During training, backpropagation forces the receiver to extract sufficient channel information to minimize bit error rate, effectively learning a blind estimation algorithm
• Benefit: Avoids the two-step estimation-then-detection pipeline, reducing computational latency and eliminating error propagation from imperfect explicit estimates

A deep learning framework that learns a compressed stochastic representation of an input signal that is maximally informative about a target task.

• Relevance to Pilotless Systems: Used to design encoders that strip away channel-dependent redundancies, transmitting only information that is robust to unknown channel transformations
• Objective: Maximizes mutual information between the compressed representation and the desired output while minimizing mutual information with the input, creating a bottleneck that discards nuisance variables like channel phase
• Training: Uses a variational approximation to the intractable mutual information, enabling gradient-based optimization of the rate-distortion trade-off

A classical non-coherent scheme where information is encoded in the phase difference between consecutive symbols rather than absolute phase, eliminating the need for a phase reference.

• Legacy Foundation: DPSK (Differential Phase Shift Keying) is the traditional baseline that pilotless neural systems aim to outperform
• Limitation: Suffers a 3 dB signal-to-noise ratio penalty compared to coherent detection with perfect CSI
• Neural Successor: Learned differential autoencoders use neural networks to discover non-linear differential encoding schemes that close the performance gap while maintaining reference-free operation