Polychronous Wave Computing (PWC)

• Polychronous Wave Computing is a computational model that uses propagating and interfering waves to execute parallel operations across multiple channels.
• It leverages engineered delay paths, frequency harmonics, and phase-coherent interference in platforms like time-modulated metasurfaces, spin-wave arrays, and photonic circuits.
• PWC offers high-throughput, low-latency analog computing for real-time signal processing and memory recall, despite challenges in device integration and stability.

Polychronous Wave Computing (PWC) encompasses a class of computational primitives and architectures in which information is carried, processed, and routed by propagating waves whose temporal, spatial, or spectral structure is harnessed directly for computation. The defining feature of PWC is that multiple tasks or logical operations are executed in parallel through the multiplexing of delay paths, frequency harmonics, phase-coherent interference, or spatiotemporal wavefront intersections. Unlike conventional digital systems, PWC natively encodes and manipulates information in the time, frequency, or phase domain—often bypassing analog-to-digital conversion and operating at substrate-limited latency and energy scales. PWC primitives have been realized in diverse platforms including metamaterial-based time-modulated metasurfaces, magnonic (spin-wave) arrays, programmable photonic circuits, phase-coherent oscillator networks, and resistive delay-line memories.

1. Physical Principles and Model Substrates

At its core, PWC leverages the propagation and interaction of waves—acoustic, electromagnetic, spin, or optical—subject to engineered modulation, boundary conditions, or interference. In time-modulated metasurfaces, explicit time-dependence is introduced in material or boundary parameters (e.g., ε(t), μ(t), or acoustic impedance Z(t)). This breaks time-invariance, permitting energy transfer between spectral components: an input at frequency ω₀ is converted to a comb of output frequencies ω₀ + nωₘ, with the amplitude of each harmonic governed by the nth-order Bessel function J_n(m), where m is the modulation index (Mousa et al., 2024).

In magnonic and oscillator arrays, wavefronts are launched by local excitation sources such as spin-torque nano-oscillators (STNOs) or optical modulators. These fronts propagate, and their engineered intersections in space-time activate logic, memory, or routing primitives, with detection and response implemented by thresholded transponder circuits (Macià et al., 2010). In spiking neural networks, PWC can be mapped onto networks with axonal delays and “polychronous groups”—sets of neurons firing in precise spatiotemporal chains due to matched delays (Marzi et al., 2018).

Phase-encoded architectures use programmable multiport interferometers, where input events are mapped to phases in a rotating reference frame and scored via coherent interference, with selection enforced by nonlinear winner-take-all dynamics (Berloff, 19 Jan 2026). Temporal memory modules based on one-transistor-one-resistor (1T1R) crossbars enable native storage and retrieval of wavefronts via arrival-time coding, relying on the intrinsic RC delays of memristive devices (Madhavan et al., 2020).

2. Mathematical Frameworks and the Theory of PWC

The theoretical foundation of PWC varies with the physical substrate. In time-modulated metasurfaces, the governing equation for a wave with time-varying boundary or permittivity is typically expressed as:

$$\partial^2p/\partial x^2 - (1/c^2)\partial^2p/\partial t^2 = 0$$

with boundary modulation $ξ(t) = A_m \cos(\omega_m t + φ_m)$, producing output harmonics described by a Floquet expansion:

$$p(x, t) = \sum_{n=-N}^N P_n\,e^{-i(\omega_0 + n\omega_m)t}$$

Amplitude conversion among harmonics follows from the Jacobi–Anger expansion and energy is partitioned as $|P_n|^2 \propto |J_n(m)|^2$. Each harmonic channel $n$ can be engineered to carry an independent spatial operator $H_n(k_y)$, enabling distinct mathematical operations per frequency channel (e.g., differentiation, integration, convolution) (Mousa et al., 2024).

Spin-wave implementations are governed by the Landau–Lifshitz–Gilbert equation augmented with spin-transfer torque, linearized as:

$$i\partial\tilde{m}/\partial t = (1 + i\alpha')\nabla^2\tilde{m} - (h-1 + i\alpha'(h-1))\tilde{m}$$

Superposition and controlled intersection of spin-wave packets encode both logic and memory operations (Macià et al., 2010).

In temporal neural codes, polychronous group firing patterns are characterized by high-dimensional binary or count codes, with pairwise distances and separation metrics scaling favorably for exponential capacity and robust linear separability. Let $x \in \mathbb{R}^N$ be a spike count vector and $c \in \{0,1\}^N$ its binarization, with Hamming and squared Euclidean distances controlling code separation and capacity (Marzi et al., 2018).

Phase-domain PWC operates by mapping timing to phase: each latency $t_j$ is encoded as $θ_j = Ω t_j \bmod 2π$, forming input phasors $u_j = e^{-iΩ t_j}$. Coherent linear scoring combines these via a programmable $N\times K$ matrix $J$:

$Ψ_k(x) = \sum_{j=1}^N J_{jk}u_j$

A subsequent nonlinear WTA array implements physical argmax selection, governed by saturable gain competition (Berloff, 19 Jan 2026).

3. Architectures, Unit Cells, and Composition

Metasurface-based PWC employs subwavelength unit cells (e.g., half-wavelength acoustic waveguides with side-branch Helmholtz resonators) with two core geometric degrees of freedom controlling transmission and phase. Time-invariance breaking is accomplished by harmonically driving cavity walls (e.g., using a deforming mesh in simulation), generating the required frequency-multiplexed harmonics for analog multitasking (Mousa et al., 2024).

Spin-wave PWC arranges STNOs and threshold transponders on thin films, leveraging geometric and temporal layout to realize cascaded computation via intersecting wavefronts. Logic gates and reverberating memory are constructed by controlling initial excitation timing, positions, and the subsequent detection/relaxation protocols (Macià et al., 2010).

Resistive temporal memories arrange memristive crossbars such that each memory cell’s conductance encodes a programmable RC delay. Capture uses an STDP-inspired write, translating pulse arrival times into conductance change. Recall applies a voltage step, eliciting a volley of output edges whose sequence recapitulates the input wavefront, enabling pipeline composition with temporal ALUs for sequential PWC steps (Madhavan et al., 2020).

Phase-domain PWC is realized using multiport interferometers (e.g., integrated photonic meshes) and gain-competition banks (e.g., semiconductor lasers or polariton condensates), requiring only a single stage of digitization at the output. Calibration techniques optimize phase margins using measurement-only feedback (Berloff, 19 Jan 2026).

4. Multiplexing Strategies and Parallel Computation

A central property of PWC is native parallelism via spectral, spatial, or phase multiplexing. In time-modulated metasurfaces, each generated harmonic ($ω_0 + nω_m$) constitutes an independent computational channel, capable of carrying out a distinct spatial or spectral operation $H^{(n)}(k_y)$. Algorithmic design proceeds by solving the inverse problem to find per-cell parameters that match the desired amplitude and phase for every operator across all channels, with energy partitioned according to $|J_n(m)|^2$ for each $n$ (Mousa et al., 2024).

Spin-wave PWC achieves polychronous logic by configuring multiple excitation and detection pathways whose interference encodes computation. Each wavefront intersection can be interpreted as the coincident arrival of logically-dependent input events, supporting AND, multi-input logic, and time-delay memory (Macià et al., 2010).

For phase-encoded spiking networks, PWC enables $K$-address parallel argmax routing by concurrently scoring all hypotheses in a single pass, with address selection determined by the argmax of WTA-enforced intensities. Timing-native lookup of spiking addresses (without per-spike digitization) offers throughput commensurate with input spike times and carrier coherence intervals (Berloff, 19 Jan 2026).

Temporal memories and race-logic ALUs process and propagate multi-channel wavefronts directly; by alternating memory and computation entirely in the time domain, complex, arbitrarily deep parallel-pipelined PWC graphs can be constructed (Madhavan et al., 2020).

5. Performance Metrics, Limitations, and Robustness

The efficiency and scalability of PWC are dictated by the underlying physics and circuit parameters. The number of independent harmonic channels $N$ in metasurface PWC is limited by the modulation index $m$; $|n| \leq 3\ldots5$ is practical for $m\sim1$, with energy per channel $η_n = |J_n(m)|^2$ and total energy conserved ($\sum_n η_n = 1$) (Mousa et al., 2024). Crosstalk and bandwidth are mitigated by ensuring spectral separation ($ω_m \gg$ operator bandwidth) and limiting frequency dispersion.

Spin-wave PWC is constrained by the group velocity of the medium, interference precision (packet width, dispersion), reflections, and integration density of transponders. Single STNO pulse energy is $\sim10$ fJ for 20 nm contacts, with operation in the sub-100 ps domain and GHz-range logic (Macià et al., 2010).

Temporal memory implementations are affected by memristor device variability, timing jitter induced by resistance spread, and leakage currents. With state-of-the-art device variation ($\sim1\%$), jitter is below 1 ns for dynamic ranges of tens of nanoseconds, and rank-order temporal codes maintain robustness to noise (Madhavan et al., 2020).

In phase-encoded architectures, operation is limited by phase-wrap (window $t_{\text{max}} < 2π/Ω$), carrier coherence time ($τ_{\text{coh}}\gtrsim t_{\text{max}}$), and a composite phase-noise budget ($$σ_{\text{eff}}^2 = (Ωσ_t)^2 + σ_θ^2 + σ_{\text{coh}}^2$$). Nonlinear WTA stages extend address-selection accuracy under noise compared to linear readout, and in situ hardware-in-the-loop calibration climbs from $55.9\%$ to $97.2\%$ accuracy under strong phase disorder (Berloff, 19 Jan 2026).

6. Applications, Scaling, and Outlook

PWC is suited for high-throughput, low-latency, parallel analog preprocessing—examples include real-time edge detection, convolutional filtering, ODE/PDE solving, structural health monitoring in extreme environments, and timing-domain address selection in sparse, LUT-style spiking networks (Mousa et al., 2024, Berloff, 19 Jan 2026).

Temporal memories with direct wavefront recall enable efficient chaining of sequence detection, pattern recognition, and dynamic programming entirely in the time domain, eliminating translation costs between analog/digital representations (Madhavan et al., 2020).

Spin-wave and photonic analogs offer pathways to ultra-compact, energy-efficient, high-speed logic, and memory, leveraging spatial and spectral parallelism far beyond conventional rate-coded or clocked-digital architectures (Macià et al., 2010, Berloff, 19 Jan 2026).

Scaling to large 2D metasurfaces or high-degree multiport interferometers enables spatial, spectral, and phase domain multiplexing, approaching “metacompute arrays” with theoretically massive parallel throughput (Mousa et al., 2024).

Challenges remain in device integration, phase stability, crosstalk, calibration, and error correction as system size grows. Ongoing work targets improved noise margins, dynamic reconfiguration, and efficient hybridization with standard CMOS control for practical deployment.

7. Comparative Table of PWC Modalities

Platform	Wave Modality / Channelization	Core Operation
Time-modulated metasurface	Frequency harmonics (ω₀ + nωₘ)	Parallel analog mathematical ops
Spin-wave STNO arrays	Spatiotemporal interference	Logic/memory by wavefront intersection
Phase-encoded photonics	Phase in rotating frame (Ω t_j)	Fast address selection (physical argmax)
Resistive crossbar memory	RC delays per row/column	Direct wavefront storage/recall

Each PWC modality exploits native physical processes for high-fidelity, parallel, and scalable computation, eschewing traditional clocked and rate-coded logic in favor of propagating, interfering, or multiplexed waves.