Kuramoto Oscillatory Phase Encoding (KoPE)

• Kuramoto Oscillatory Phase Encoding (KoPE) is a phase-based computational framework that encodes information in oscillator phases and exploits synchronization for computation.
• It employs diverse coding schemes—including binary, stochastic, and manifold-valued formulations—to implement Ising machines, associative memory, and phase-augmented neural architectures.
• KoPE integrates dynamical phase equations with gradient transport and learning paradigms, enabling robust optimization across hardware implementations and transformer-based systems.

Kuramoto Oscillatory Phase Encoding (KoPE) denotes a class of phase-based computational schemes in which information is encoded in the phases of coupled oscillators, computation proceeds through synchronization or phase-locked equilibria, and decoding uses phase differences, trigonometric readouts, or collective order parameters. Across the literature, KoPE appears in several distinct but compatible forms: binary phase encodings for Ising machines, phase-only learning in coupled Kuramoto networks, discrete-state stochastic phase codes, manifold-valued oscillator representations, and phase-augmented transformer architectures (Khan et al., 28 Oct 2025, Ahmadi, 11 Apr 2026, Jörg, 2017, Jacimovic, 2024, Xiao et al., 9 Apr 2026).

1. Representational principle and coding schemes

At its core, KoPE treats phase as the primary computational state. In the classical scalar setting, oscillator $i$ carries a phase $\theta_i \in [0,2\pi)$, and the collective state is summarized by the complex order parameter

$$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$

with $r$ measuring coherence and $\psi$ the mean phase. In Ising-oriented variants, the encoding is explicitly binary: parametric oscillators above threshold exhibit phase bistability at $0$ and $\pi$, and spins are mapped by $s_i = \cos \phi_i \in \{+1,-1\}$, equivalently $s_i=\operatorname{sign}[\cos\phi_i]$ with $\phi_i\in\{0,\pi\}$. The vector of phases then encodes a candidate Ising configuration (Khan et al., 28 Oct 2025).

KoPE is not restricted to binary alphabets. In the stochastic discrete-phase formulation, the phase interval $\theta_i \in [0,2\pi)$0 is discretized into $\theta_i \in [0,2\pi)$1 equally spaced states of size $\theta_i \in [0,2\pi)$2, each oscillator carries an integer phase state $\theta_i \in [0,2\pi)$3, and its physical phase is $\theta_i \in [0,2\pi)$4. The encoded symbol is the residue class $\theta_i \in [0,2\pi)$5, so KoPE becomes an explicitly $\theta_i \in [0,2\pi)$6-ary phase code implemented by a continuous-time Markov chain (Jörg, 2017).

The same representational idea extends beyond $\theta_i \in [0,2\pi)$7. In geometry-informed formulations, oscillators live on spheres, homogeneous spaces, or Lie groups: $\theta_i \in [0,2\pi)$8, $\theta_i \in [0,2\pi)$9, $$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$0, or $$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$1. Here, “phase” generalizes to orientation, principal angle, or group element, and decoding uses consensus directions, invariant traces, hyperbolic or Bergman barycenters, and symmetry-respecting probability families such as von Mises, wrapped Cauchy, von Mises–Fisher, spherical Cauchy, and Bergman–spherical Cauchy (Jacimovic, 2024).

In transformer-based KoPE, each token carries an additional evolving phase state. The implementation uses a concatenation of $$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$2 two-dimensional phase vectors $$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$3, each pair representing $$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$4. These phases are updated layer-wise and used to rotate query, key, value, and output subspaces, so that phase becomes a depth-evolving relational variable rather than a static positional tag (Xiao et al., 9 Apr 2026).

2. Dynamical formulations and energy structure

The canonical dynamical backbone of KoPE is the Kuramoto phase equation

$$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$5

or its mean-field specialization. Several KoPE variants augment this form while retaining a phase-only interpretation. For parametric-oscillator Ising machines, starting from a conjugate Stuart–Landau network

$$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$6

one obtains, after amplitude–phase decomposition and weak-coupling reduction,

$$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$7

This reduction adds two structures absent from the purely normal Kuramoto network: an intrinsic on-site even-harmonic term $$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$8, generated by parametric gain, and a conjugate-coupling channel $$r e^{i\psi} = \frac{1}{N}\sum_{j=1}^N e^{i\theta_j},$$9, generated by the $r$0 pathway. For real symmetric couplers and pump alignment, the dynamics are gradient with Lyapunov function

$r$1

and at binary phases this reduces, up to a constant, to the Ising energy (Khan et al., 28 Oct 2025).

A different route to exact phase dynamics is provided by Koopman phase–amplitude coordinates. For a stable limit-cycle oscillator $r$2, global coordinates $r$3 and $r$4 can be defined so that the uncoupled system obeys $r$5 and $r$6. Coupling along the dual vector $r$7 enforces

$r$8

while amplitude channels

$r$9

remain arbitrary. In this construction, phase dynamics are globally Kuramoto and exactly decoupled from amplitudes throughout the basin of the limit cycle, not merely near the orbit (Yawata et al., 5 Jan 2026).

The discrete-state stochastic formalism replaces ODE phase flow by Poisson phase hopping. The joint probability obeys a master equation in which intrinsic oscillation and coupling are encoded by ladder operators, and the continuous-phase limit recovers the stochastic Kuramoto SDE

$\psi$0

This makes explicit that coupling contributes both synchronizing drift and phase diffusion (Jörg, 2017).

3. Learning, inference, and gradient transport through phase

KoPE is also a learning paradigm. In equilibrium-propagation formulations, the computational state is the synchronized equilibrium $\psi$1 of a Kuramoto network in a rotating frame, inputs are encoded in natural frequencies of designated input oscillators, and outputs are read from designated output phases. For the loss

$\psi$2

the central identity states

$\psi$3

where $\psi$4 is the output-nudging strength and $\psi$5 are mean-centered frequencies. The same physical phase displacement that represents the inference state therefore carries the exact gradient with respect to frequencies. In the reported sparse layered architecture with $\psi$6 input, $\psi$7 hidden, and $\psi$8 output oscillators, frequency learning on converged seeds outperformed coupling-weight learning among converged seeds at matched parameter counts, $\psi$9 vs. $0$0, $0$1, while random initialization produced an approximately $0$2 convergence failure rate that was eliminated by topology-aware spectral seeding, improving success from $0$3 to $0$4 on the primary task and to $0$5 on three further settings (Ahmadi, 11 Apr 2026).

Readout in these learning systems is phase-based rather than amplitude-based. In the primary classification task, the predicted class is

$0$6

and weakly nudged phase differences directly supply the update

$0$7

Theoretical analysis relies on the Jacobian of the equilibrium equations being the negative of a coupling-weighted graph Laplacian, so gradient propagation becomes diffusion through the inverse reduced Laplacian (Ahmadi, 11 Apr 2026).

Transformer KoPE implements learning differently but preserves the same principle that phase is a computational carrier. With token activations $0$8 and phase state $0$9, the coupling matrix is attention-like,

$\pi$0

and the phase state is updated by a vector Kuramoto Euler step followed by pairwise normalization. Attention logits and value aggregation are modulated by complex rotations $\pi$1 or $\pi$2 rotation blocks, so synchronization and attention concentration are co-evolving (Xiao et al., 9 Apr 2026).

4. Ising machines, associative memory, and hardware realizations

In parametric-oscillator Ising machines, KoPE gives a phase-only interpretation of binary optimization. The target Ising Hamiltonian

$\pi$3

is realized by symmetric linear couplings at binary phases because both $\pi$4 and $\pi$5 reduce to $\pi$6 when $\pi$7. A central result is that explicit $\pi$8 driving is unnecessary in parametric oscillators: the $\pi$9 pump term already produces the on-site $s_i = \cos \phi_i \in \{+1,-1\}$0 bistability responsible for the $s_i = \cos \phi_i \in \{+1,-1\}$1 alphabet. The same analysis shows that amplitude heterogeneity rescales couplings according to

$s_i = \cos \phi_i \in \{+1,-1\}$2

so the machine minimizes a weighted Ising instance rather than the target one unless amplitudes are regulated. A convenient heterogeneity measure is $s_i = \cos \phi_i \in \{+1,-1\}$3, and solution quality typically degrades monotonically as AH increases. The implementation discussion explicitly includes DOPO networks, Josephson parametric oscillators, and electromechanical parametric dividers, as well as optical, electronic, and mechanical measurement and feedback pathways (Khan et al., 28 Oct 2025).

Higher-order KoPE extends associative memory beyond pairwise Kuramoto coupling. The generalized network

$s_i = \cos \phi_i \in \{+1,-1\}$4

stores memories as binary phase patterns $s_i = \cos \phi_i \in \{+1,-1\}$5, equivalent to $s_i = \cos \phi_i \in \{+1,-1\}$6, with Hebbian matrix and tensor embeddings. Mean-field analysis yields the free energy

$s_i = \cos \phi_i \in \{+1,-1\}$7

a tricritical point at $s_i = \cos \phi_i \in \{+1,-1\}$8 and $s_i = \cos \phi_i \in \{+1,-1\}$9, and a quartic-dominated bistable regime with Kramers escape time

$s_i=\operatorname{sign}[\cos\phi_i]$0

For finite load, the critical capacity obeys

$s_i=\operatorname{sign}[\cos\phi_i]$1

and the reported large-scale simulations show superlinear memory-capacity scaling with system size. Candidate physical platforms explicitly include optical parametric oscillators, four-wave mixing, nonlinear crystals with spatial light modulators, exciton–polariton condensate lattices, superconducting Kerr-parametric-oscillator arrays, and electronic or spintronic oscillators with engineered higher harmonics (Nagerl et al., 29 Jul 2025).

5. Geometric and neural-network generalizations

KoPE on manifolds recasts phase encoding as geometry-informed learning. On $s_i=\operatorname{sign}[\cos\phi_i]$2, the real higher-dimensional Kuramoto model uses

$s_i=\operatorname{sign}[\cos\phi_i]$3

while the complex model on $s_i=\operatorname{sign}[\cos\phi_i]$4 and the non-Abelian models on $s_i=\operatorname{sign}[\cos\phi_i]$5 and $s_i=\operatorname{sign}[\cos\phi_i]$6 replace scalar phase differences by projected vector fields or commutator-like couplings. For identical globally coupled systems, trajectories factor through groups of Möbius, hyperbolic, or Bergman isometries, and the continuum limit yields invariant statistical families suited to geometric deep learning, including von Mises, wrapped Cauchy, Kato–Jones, hyperbolic von Mises, von Mises–Fisher, spherical Cauchy, and Bergman–spherical Cauchy distributions (Jacimovic, 2024).

In vision transformers, KoPE augments each token with an evolving phase state and updates it layer-wise via

$s_i=\operatorname{sign}[\cos\phi_i]$7

while attention uses phase-rotated queries, keys, values, and outputs. The phase mechanism is explicitly synchronized by token-content couplings and initialized with multi-frequency $s_i=\operatorname{sign}[\cos\phi_i]$8-dimensional rotary positional embeddings. The reported empirical results state that KoPE improves training, parameter, and data efficiency; reaches similar accuracy with approximately $s_i=\operatorname{sign}[\cos\phi_i]$9 fewer training samples; adds approximately $\phi_i\in\{0,\pi\}$0–$\phi_i\in\{0,\pi\}$1 parameters and approximately $\phi_i\in\{0,\pi\}$2 additional FLOPs; improves ADE20K semantic segmentation from $\phi_i\in\{0,\pi\}$3 to $\phi_i\in\{0,\pi\}$4 mIoU for ViT-B and from $\phi_i\in\{0,\pi\}$5 to $\phi_i\in\{0,\pi\}$6 for ViT-L; improves COCO panoptic segmentation from $\phi_i\in\{0,\pi\}$7 to $\phi_i\in\{0,\pi\}$8 PQ and from $\phi_i\in\{0,\pi\}$9 to $\theta_i \in [0,2\pi)$00 PQ$\theta_i \in [0,2\pi)$01 for ViT-B; and achieves $\theta_i \in [0,2\pi)$02 on ARC-AGI-1 and $\theta_i \in [0,2\pi)$03 on ARC-AGI-2, with maxima $\theta_i \in [0,2\pi)$04 and $\theta_i \in [0,2\pi)$05, respectively. The accompanying analysis states that lowering softmax temperature does not replicate the gains, while attention-weighted synchronization

$\theta_i \in [0,2\pi)$06

rises across training and layers (Xiao et al., 9 Apr 2026).

A recurrent theme across these geometric and neural formulations is that KoPE need not be confined to oscillator hardware. The same principle is realized whenever phase or phase-like state variables are updated by synchronization dynamics and then used as the computational substrate for inference, binding, or equivariant representation learning (Jacimovic, 2024, Xiao et al., 9 Apr 2026).

6. Noise, partial locking, delays, and operational limits

Noise analysis shows that KoPE is intrinsically a drift–diffusion problem. For noisy limit-cycle oscillators governed by the Itô SDE

$\theta_i \in [0,2\pi)$07

phase reduction yields a reduced phase SDE

$\theta_i \in [0,2\pi)$08

where the drift includes both explicit Itô corrections and an amplitude-covariance term,

$\theta_i \in [0,2\pi)$09

Accordingly, phase noise is not pure diffusion: amplitude fluctuations shift the mean frequency even when the expected amplitude remains unchanged. The paper also gives a coherence-time interpretation and derives a phase-error estimate for $\theta_i \in [0,2\pi)$10-ary phase encoding under Gaussian phase noise (Bonnin, 2019).

KoPE does not require full synchronization. For the finite-$\theta_i \in [0,2\pi)$11 all-to-all Kuramoto model

$\theta_i \in [0,2\pi)$12

partially phase-locked clusters are guaranteed when a subset spread criterion

$\theta_i \in [0,2\pi)$13

holds, where $\theta_i \in [0,2\pi)$14 is the minimal frequency spread over subsets of size $\theta_i \in [0,2\pi)$15 and

$\theta_i \in [0,2\pi)$16

The associated invariant and attracting balls quantify robustness and asymptotic intra-cluster spread, and in the thermodynamic limit the largest cluster size satisfies deterministic lower and upper bounds derived from $\theta_i \in [0,2\pi)$17 and $\theta_i \in [0,2\pi)$18 (Bronski et al., 2020).

Delays, topology, and plasticity further diversify KoPE dynamics. In delayed Hebbian Kuramoto lattices,

$\theta_i \in [0,2\pi)$19

the combination of delay and learning yields half-integer phase-ramp modes in one dimension and spiral-like frequency-entrained patterns in two dimensions. In the fast-learning one-dimensional case, the reported mode transitions with increasing delay are $\theta_i \in [0,2\pi)$20. On Y-shaped trees, simulations classify regimes by the FFT of the order parameter $\theta_i \in [0,2\pi)$21: $\theta_i \in [0,2\pi)$22 peaks indicates synchronization, $\theta_i \in [0,2\pi)$23–$\theta_i \in [0,2\pi)$24 peaks a wave state, and more than $\theta_i \in [0,2\pi)$25 peaks chaos, with increasing main-branch length or frequency heterogeneity pushing the system toward wave and chaotic behavior (Timms et al., 2013, Nouhi et al., 2023).

The main analytical limitations are structural. Exact Lyapunov descent for parametric KoPE relies on real symmetric couplers and pump alignment; phase lags, asymmetries, or delays break exact gradient flow and can introduce non-conservative currents. Exact equilibrium-propagation identities require symmetric coupling, stable equilibria with nonsingular reduced Jacobian, and weak nudging. Global decoupling of phase from amplitude requires Koopman eigenfunctions and their dual vector fields on the full basin of a stable limit cycle. In the discrete-state setting, coarse phase discretization increases stochasticity and can suppress synchrony relative to the continuous Kuramoto limit (Khan et al., 28 Oct 2025, Ahmadi, 11 Apr 2026, Yawata et al., 5 Jan 2026, Jörg, 2017).

Taken together, these results define KoPE as a broad, technically unified program: encode information in phases or phase-like states, shape the collective dynamics by Kuramoto-type coupling, and exploit synchronization structure for optimization, memory, or learning. The specific implementation may be a parametric Ising machine, a stochastic phase Markov chain, a manifold-valued dynamical system, or a phase-augmented deep network, but the common invariant is that computation is organized by collective phase dynamics rather than by amplitudes alone.