PhaseCoder: Multi-Domain Phase Encoding

• PhaseCoder is a framework that leverages phase coding to encode, transmit, and decode information in neuroscience, communication, spatial audio, and radar systems.
• It integrates methodologies such as neural mass modeling, genetic algorithms, and group theory to optimize phase separation and decoding accuracy across applications.
• Empirical results demonstrate high classification rates (over 95% in neural systems), notable SNR gains in communication, and improved signal processing in spatial audio and radar.

PhaseCoder is a term associated with several distinct research lines, all linked by their explicit use of phase—or phase-coded structures—for encoding, retrieving, or manipulating information across domains including neuroscience, digital communications, spatial audio, and radar. In contemporary literature, “PhaseCoder” has denoted neural dynamical classifiers using oscillatory phase encoding, pilotless communication protocols exploiting phase symmetry in modulation, microphone-geometry–invariant spatial audio transformers, and optimized phase-code sequence discovery for radar. This entry systematically reviews the principal instances and theoretical underpinnings of PhaseCoder, emphasizing advances in architecture, decoding, optimization, and application.

1. PhaseCoder in Neural Dynamical Systems

PhaseCoder designates a neural dynamical device that leverages phase coding in neural mass models to map discrete inputs onto the phase of limit-cycle oscillations, thereby enabling class-encoding and robust decoding via power measurements. The canonical construction employs a network of Jansen–Rit cortical columns—each with excitatory interneurons, inhibitory interneurons, and pyramidal cells—whose dynamics are governed by coupled second-order ODEs with postsynaptic kernels and sigmoid nonlinearities. Inputs are encoded as short pulses delivered to subsets of “input” columns, propagating through a multilayer feedforward network topology that culminates in a single output column (Pei, 7 Mar 2025).

Parameters (local gains, synaptic time-constants, and intercolumn feedforward weights) are globally optimized by a real-coded genetic algorithm. The fitness function is defined to maximize between-class phase separation of the output pyramidal cell voltage oscillations while penalizing unbounded responses.

For $K$ input classes, the optimized network generates distinct output waveforms,

$$V_3^{(k)}(t) \approx R \sin\bigl(\omega t + \phi_k\bigr),$$

with phase offsets $\phi_k$ that encode class identity. The phase $\phi_k$ is extracted by Hilbert transform, and inter-class separations $\Delta\phi_{ij}$ on the order of $30^\circ$–$90^\circ$ are observed with high statistical reliability ($p < 0.05$).

Decoding is implemented by delivering phase-locked, brief excitatory pulses at candidate phases $\{\phi_k\}$ and measuring the induced power gain at the fundamental frequency $\omega$. The class is inferred as $$\hat{k} = \arg\max_{k} \hat{P}(\omega; \phi_{\text{stim}}=\phi_k)$$, yielding decoding accuracy in excess of 95%. The device thus realizes a dynamical “oscillatory phase code” classifier with direct links to the broader phase-of-firing code principle in neural information theory (Cattani et al., 2015).

2. PhaseCoder in Pilotless Digital Communication

In physical-layer digital communications, PhaseCoder (or phase-equivariant coded modulation) capitalizes on the rotational symmetry of Gray-labeled QAM constellations and the code automorphism group of polar codes to enable pilotless, jointly phase-ambiguous decoding (Geiselhart et al., 2023). A blind fine phase estimator (e.g., Viterbi–Viterbi) removes the continuous phase offset $\phi_f$ but leaves a residual ambiguity among the set $\{0, \frac{\pi}{2}, \pi, \frac{3\pi}{2}\}$.

Let $c$ be the polar-coded bits and $x = \mathrm{map}(c)$ the mapped QAM symbols. Discrete phase rotations induce coordinated bit-permutations and bit-flips corresponding to code automorphisms. If $T$ is the block-cyclic permutation, and $b$ is the codeword effecting the required flips, the group action is $c \mapsto c T^m \oplus (b \oplus \dots)$ for $m \in \mathbb{Z}_4$.

To resolve the $\mathbb{Z}_4$ ambiguity, two frozen bit channels are reserved in the polar code. During decoding, the SC decoder recovers these bits, supplying the phase index $m$, which is then corrected via a companion code automorphism. This scheme jointly achieves phase disambiguation and source decoding with no reliance on pilot symbols, resulting in SNR gains up to 0.8 dB (QPSK) and 2 dB (16-QAM) over pilot-based benchmarks (Geiselhart et al., 2023).

3. PhaseCoder Architectures in Spatial Audio

In spatial audio and multimodal LLM pipelines, PhaseCoder designates a transformer-only spatial audio encoder that is agnostic to microphone geometry (Dementyev et al., 28 Jan 2026). The input is multichannel raw audio with linked 3D microphone coordinates. The architecture utilizes:

• Patch-level STFT feature extraction (magnitude and phase, 258-dim per channel/frame)
• Linear projection to D=256 embeddings
• Positional embeddings: sequential (1D sinusoidal), frame-level, and microphone-coordinate based (spherical encoding with parameterized trigonometric fusion)
• ViT-style encoder: 5 transformer blocks, 4-head self-attention

Embedding the microphone location vector directly allows for arbitrary array topologies, enabling generalization across diverse spatial configurations. Training incorporates synthetic augmentation (RIRs, random arrays, SNR variation). Tasks include joint azimuth/elevation/distance classification, multitask spatial reasoning, and diarized transcription.

For LOCATA (8-mic) benchmark, PhaseCoder achieves azimuth MAE 7.44° with accuracy@10° of 86.96%, exceeding the previous state-of-the-art GI-DOAEnet (82.48%). Integrated with Gemma 3n, PhaseCoder’s “spatial audio tokens” enable LLM-based complex spatial reasoning and targeted transcription, with gains of up to +28% in yes/no localization reasoning and >15 pp WER reduction on targeted transcription.

4. Genetic-Algorithm-Based Phase Code Discovery for Radar

In pulse-compression radar (PCR), PhaseCoder/GASeq refers to a genetic-algorithm framework for discovering binary phase codes that maximize signal-to-clutter ratio (SCR) when paired with the optimal mismatched filter (2207.14631). The fitness function is $f(s) = s^\top R(s)^{-1} s$, with $R(s)$ the clutter covariance structured by all aperiodic autocorrelation lags except zero.

A population-based GA (typically $N=59$ or 100; $P=10^4$; $K=200$ generations) employs elite selection, tournament selection, one-point crossover, and targeted mutation, with explicit measures to prevent early convergence. At $N=59$, GASeq finds a phase code with SCR=50.84 (surpassing deep RL and GAN approaches by significant margins), and successfully scales to previously intractable code lengths ($N=100$).

The process is dominated by $O(N^3)$ matrix operations for $R(s)^{-1}$ per candidate; however, parallelism renders evaluations tractable for moderate $N$. This approach yields codes with lower sidelobes and stronger clutter rejection than previously known classes (e.g., Legendre or GAN-designed codes).

5. Theoretical and Practical Significance

All instances of PhaseCoder—across neuroscience, information theory, array signal processing, and radar—exploit the information-carrying capacity of phase, either directly (oscillatory encoding, phase of firing, phase retrieval) or structurally (rotational symmetry in codes, phase embedding in transformers). The methodologies integrate dynamical systems modeling, group representation theory, genetic and evolutionary algorithms, and deep learning architectures. Robust phase coding enables pilotless transmission, geometry-invariance, and efficient classification, reducing resource demands while maintaining or improving decoding performance.

These systems demonstrate the efficacy of phase-based information representation both for biological modeling (neural mass phase encoding), cognitive/sensory device design (LLM spatial audio integration), advanced radar waveform optimization, and communication reliability in hostile or high-mobility scenarios.

6. Summary Table: Principal PhaseCoder Domains

Domain	PhaseCoder Construction	Key Results/Benchmarks
Dynamical neuroscience	Jansen–Rit network + GA optimization	Phase code decoding $\geq$95% acc., phase shift 30°–90°
Communication theory	Polar code automorphism, equivariance	Pilotless QPSK/16-QAM, +0.8–2 dB vs. pilots
Spatial audio/LLMs	Geometry-agnostic transformer encoder	SOTA on LOCATA: 7.44° MAE, 86.96% @ 10°, LLM integration
Radar coding	GA code discovery (GASeq)	SCR=50.84 (N=59), outperforms GAN/DRL/random

A plausible implication is that shared phase-coding abstractions across these domains could inspire new generic frameworks for resource-efficient, phase-centric signal processing and learning in both artificial and biological systems.