Parametric-Oscillator Ising Machines

• Parametric-Oscillator Ising Machines are systems where binary spin values are encoded in the two stable phase states (0/π) of driven oscillators.
• They exploit degenerate parametric oscillation in devices like DOPOs and Josephson oscillators to naturally map oscillator phases to an Ising Hamiltonian.
• A Kuramoto-style phase reduction reveals an energy descent dynamic that underpins the devices' ability to relax into optimal spin configurations.

Parametric-Oscillator Ising Machines (POIMs) are Ising machines in which binary spin variables are encoded in the phase-bistable states of parametrically driven oscillators, and optimization is performed by letting coupled oscillator networks relax toward low-energy collective phase configurations. In the formulation developed for networks of parametric oscillators such as degenerate optical parametric oscillators (DOPOs), Josephson parametric oscillators, and parametric frequency dividers, the central ingredients are pumping near $2\omega$, phase bistability at $0/\pi$, and coupling-induced synchronization or anti-synchronization that embeds an Ising objective in the network dynamics (Khan et al., 28 Oct 2025).

1. Physical basis and spin encoding

The defining physical mechanism of a POIM is degenerate parametric oscillation. When a nonlinear resonator is driven near twice its natural frequency, the resulting subharmonic oscillation acquires two stable phase states separated by $\pi$. Those two states encode Ising spins. In the canonical phase mapping used across multiple implementations, $s_i=+1$ corresponds to phase $0$, $s_i=-1$ corresponds to phase $\pi$, and an equivalent compact notation is $s_i=\cos\theta_i$ when $\theta_i\in\{0,\pi\}$ (Khan et al., 28 Oct 2025).

This phase-bistable encoding appears in several hardware classes. In Kerr-nonlinear parametric oscillators (KPOs), a two-photon drive and Kerr nonlinearity generate a double-well structure in phase space, and the two coherent states at the wells’ minima represent binary spin values (Kanao et al., 2020). In Josephson parametric oscillators (JPOs), a microwave resonator terminated by a SQUID is flux-pumped at $2\omega_0$, and above threshold the phase locks to two stable values separated by $0/\pi$0 (Razmkhah et al., 2023). In parametric frequency dividers (PFDs), pumping a nonlinear resonant circuit near twice its natural frequency generates a subharmonic oscillation at half the pump frequency and yields the same $0/\pi$1 phase bistability (Casilli et al., 2023). In micromechanical parametrons, period-doubled motion under stiffness modulation near twice the eigenfrequency similarly produces two stable phase states that encode $0/\pi$2 (Han et al., 2023).

Within this family, the DOPO-based coherent Ising machine established the basic optical paradigm: each spin is represented by the binary phase state of an above-threshold DOPO pulse, while mutual injection implements the couplings of the target Ising instance (Marandi et al., 2014). A general oscillator-based formulation had already shown that phase-bistable encoding can be expressed as a second-harmonic locking term in a Kuramoto-type phase model, with the binary states enforced by sub-harmonic injection locking or equivalent parametric forcing (Wang et al., 2017).

2. Canonical phase dynamics and the Kuramoto reduction

A central unifying result is that near-threshold, single-mode parametric oscillators can be reduced to a conjugate Stuart–Landau model. For $0/\pi$3 weakly coupled oscillators pumped near $0/\pi$4, the slow-flow complex amplitudes $0/\pi$5 obey

$0/\pi$6

where $0/\pi$7 are normal couplings and $0/\pi$8 are conjugate, phase-sensitive couplings (Khan et al., 28 Oct 2025).

With $0/\pi$9, real-imaginary separation yields coupled amplitude and phase equations. Under above-threshold, quasi-steady amplitudes and weak coupling, the phase-only reduction becomes

$\pi$0

This is a Kuramoto-style phase model, but not the standard Kuramoto model. It combines the familiar phase-difference term $\pi$1 with an intrinsic sum-phase term $\pi$2 generated by conjugate coupling (Khan et al., 28 Oct 2025).

That formulation clarifies a key distinction between POIMs and traditional oscillator-based Ising machines. In the traditional phase model,

$\pi$3

the $\pi$4 term is supplied by an externally injected second harmonic (Wang et al., 2017). In POIMs, the parametric pump naturally generates the conjugate term $\pi$5, and in phase form this directly produces the on-site bistability term $\pi$6. The explicit second-harmonic drive required in standard oscillator-based Ising machines is therefore unnecessary in the parametric case (Khan et al., 28 Oct 2025).

For real symmetric couplers and fixed pump phase, the complex dynamics admit a global Lyapunov function $\pi$7 with gradient flow $\pi$8, and the phase-reduced dynamics likewise minimize a phase energy $\pi$9. This energy-descent structure is what makes the Kuramoto reduction more than a formal analogy: it gives a canonical phase description of POIMs that preserves the optimization interpretation (Khan et al., 28 Oct 2025).

3. Mapping to the Ising Hamiltonian and operating regimes

The Ising mapping follows directly from phase bistability. When $s_i=+1$0, one sets

$s_i=+1$1

For the reduced phase dynamics, the energy can be written as

$s_i=+1$2

and at binary phases this becomes

$s_i=+1$3

Thus the phase-reduced POIM minimizes the standard Ising energy up to an additive constant, with the ferromagnetic or antiferromagnetic sense determined by the sign of $s_i=+1$4 (Khan et al., 28 Oct 2025).

The same mapping appears in KPO networks. In the rotating frame at half the pump frequency, the KPO Hamiltonian is

$s_i=+1$5

and above the pitchfork bifurcation the two coherent amplitudes $s_i=+1$6 encode the physical spin $s_i=+1$7 (Kanao et al., 2020). In time-multiplexed DOPO networks, the above-threshold $s_i=+1$8 phases of optical pulses likewise encode $s_i=+1$9, and mutual injections realize the Ising couplings (Marandi et al., 2014).

A persistent issue is the relation between threshold dynamics and combinatorial optimization. One line of work showed that near threshold, mode competition selects the eigenvector $0$0 of the coupling matrix associated with the eigenvalue of maximal real part, and that for random frustrated Ising instances this continuous spectral objective is generically misaligned with the discrete Ising ground state. In that regime, networks of parametric oscillators are therefore “intrinsically not Ising solvers” for random spin glasses; the correct solution is recovered only in an above-threshold nonlinear regime where local saturation and bistability dominate (Strinati et al., 2020). A related survey emphasized that in coherent Ising machines the minimum-eigenvalue mode of the Ising coupling matrix appears near threshold and can impede relaxation to true ground states, which is why pump schedules, nonlinear saturation, and error-correction feedback become central design elements (Yamamoto et al., 2020).

Experimentally, gradual pumping can materially change behavior. A 16-bit time-division-multiplexed femtosecond DOPO coherent Ising machine reported more than $0$1 success rates for one-dimensional Ising ring and nondeterministic polynomial-time (NP) hard instances, and the accompanying analysis attributed part