Non-Coherent Autoencoder

The transmitter learns a geometric arrangement of symbols that is inherently robust to random phase rotations. Unlike fixed QAM schemes that fail under rotation, these learned constellations often exhibit circular or spherical symmetry, ensuring that a phase offset simply maps one valid symbol to another. This eliminates the need for pilot-based phase tracking, maximizing spectral efficiency in high-mobility scenarios where the channel changes faster than pilots can estimate it.

Rather than dedicating separate time-frequency resources to known pilot symbols, the autoencoder learns to superimpose training information onto the data-bearing signal. The receiver neural network is trained to simultaneously perform channel estimation and symbol detection from the raw received waveform. This pilotless communication paradigm recovers the throughput loss traditionally sacrificed for channel estimation overhead, approaching the capacity of coherent systems without requiring explicit CSI feedback.

The autoencoder learns a latent representation where information is encoded in the relative differences between consecutive symbols rather than their absolute positions. This generalizes classical differential phase-shift keying (DPSK) to high-dimensional learned manifolds. Key properties include:

• Invariance to slow-varying channel gain and phase
• Robustness to oscillator phase noise
• No need for absolute phase reference at the receiver

The transmitter jointly optimizes both the positions of constellation points (geometric shaping) and their transmission probabilities (probabilistic shaping) under a non-coherent objective. The loss function maximizes mutual information between transmitted bits and received symbols, marginalized over the unknown channel distribution. This produces energy-efficient signaling schemes where low-energy inner points are transmitted more frequently, approaching the Shannon capacity of the non-coherent channel.

During training, the autoencoder is exposed to a worst-case distribution of channel impairments rather than a fixed statistical model. An adversarial network or randomized augmentation injects severe phase noise, frequency offset, and amplitude fading. The transceiver learns representations that are invariant not just to expected channel conditions but to distributional shifts and extreme events, providing robust blind detection in contested or jammed electromagnetic environments.

The receiver sub-network functions as a blind equalization network, recovering transmitted symbols from a received sequence corrupted by unknown intersymbol interference and phase distortion. Unlike classical blind equalizers (e.g., Constant Modulus Algorithm) that rely on hand-crafted cost functions, the neural receiver learns an optimal equalization strategy directly from data. This is particularly effective for frequency-selective channels where traditional blind methods fail to converge.

In V2X and high-speed rail scenarios, the channel coherence time is extremely short due to rapid Doppler shifts. Pilot-based channel estimation becomes obsolete before it can be used.

• Learns a representation invariant to instantaneous phase rotation
• Eliminates the pilot overhead that consumes up to 30% of resources in LTE/NR
• Maintains link stability at Doppler frequencies exceeding 1 kHz
• Enables seamless handover between roadside units without CSI acquisition

For IoT sensor networks with billions of devices, per-device channel estimation is computationally intractable and energy-prohibitive. Non-coherent autoencoders enable grant-free, blind detection.

• Sensors transmit sporadically without scheduling requests
• Receiver jointly performs activity detection and data decoding
• Eliminates the need for orthogonal pilots, supporting massive overloading
• Reduces device power consumption by removing the channel estimation DSP block

Underwater channels exhibit extreme multipath delay spreads and rapid temporal variation due to surface waves and platform motion. Explicit channel estimation is often impossible.

• Learns to decode directly from the distorted, time-varying received waveform
• Robust to the doubly-selective fading characteristic of acoustic channels
• Avoids the prohibitive overhead of long training sequences in low-bandwidth links
• Enables adaptive modulation without feedback from the receiver

For bandwidth-constrained analog sensor data, such as images from a drone, non-coherent autoencoders perform deep joint source-channel coding without CSI.

• Directly maps raw pixel values to channel symbols
• Learns a robust embedding that survives unknown channel phase and amplitude scaling
• Outperforms separate source coding, channel coding, and modulation pipelines
• Graceful degradation: image quality scales smoothly with SNR without cliff effects

Non-coherent autoencoders can be trained to create a wiretap channel advantage. The legitimate receiver learns a decoding function that is invariant to the common channel, while an eavesdropper with a different channel cannot decode.

• Jointly learns a secure constellation and blind decoder
• Exploits the physical uniqueness of the legitimate channel as a key
• No key distribution or upper-layer cryptography required
• Security derives from the learned manifold geometry, not computational hardness

In diffusion-based molecular communication, the channel impulse response is stochastic and governed by Brownian motion. CSI is fundamentally unknowable at the symbol level.

• Learns to decode concentration pulses without knowing the diffusion coefficient
• Robust to the random inter-symbol interference caused by lingering molecules
• Enables practical nanonetworking for targeted drug delivery
• Adapts to varying flow conditions without recalibration