Coherent Ising Machines: Models & Architectures

• Coherent Ising Machines are optical oscillator networks that encode continuous optical states to represent binary spins for combinatorial Ising optimization.
• They employ diverse architectures, including time-multiplexed DOPOs, measurement-feedback, and all‐optical systems to dynamically explore low-energy configurations.
• Recent studies highlight practical advancements in error correction, sampling efficiency, and scalability through hybrid digital and physical implementations.

Searching arXiv for recent and foundational papers on coherent Ising machines, including measurement-feedback, all-optical, thermodynamic, and error-correction variants. Coherent Ising machines (CIMs) are optical oscillator networks devised to search for low-energy states of classical Ising Hamiltonians and thereby address combinatorial optimization tasks such as MAX-CUT, spin-glass instances, and related QUBO formulations. Across the literature, the physical state is typically not a hard binary spin during the evolution: each spin is represented by a continuous optical degree of freedom whose sign or phase yields $\sigma_i=\pm1$ at readout, while the couplings are realized by mutual optical injection, measurement-feedback through electronic coprocessors, or fully optical spatial transformations (Marandi et al., 2014, Strinati et al., 2021, Quinn et al., 2024, Aonishi et al., 2024). The field encompasses time-multiplexed degenerate optical parametric oscillator (DOPO) cavities, measurement-feedback fiber-loop systems, all-optical spatial architectures, polarization-based Kerr-resonator machines, and FPGA “cyber CIMs,” as well as a continuing debate over whether the computationally relevant mechanism is predominantly quantum-optical or effectively classical mean-field dynamics (King et al., 2018, Khosravi et al., 19 Jul 2025).

1. Problem formulation and spin representation

The optimization target in CIM research is a classical Ising cost function, but the sign convention varies across papers. Representative forms include

$H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$

$$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$

and

$$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$

with local fields or Zeeman terms $h_r$ included when required by the application (Marandi et al., 2014, Strinati et al., 2021, Gunathilaka et al., 2023). QUBO instances are commonly converted into Ising form, and several later CIM variants also support continuous or mixed formulations rather than only strict binary-spin objectives (Aonishi et al., 2024, Khosravi et al., 19 Jul 2025).

A defining feature of CIMs is that the physical degree of freedom is usually continuous. In measurement-feedback and mean-field formulations, one writes $s_i\in[-1,1]$ or $x_i\in\mathbb{R}$, and the discrete spin is recovered by $\tilde s_i=\mathrm{sign}(s_i)$ or $\sigma_i=\operatorname{sgn}(x_i)$ (King et al., 2018, Tiunov et al., 2019). In DOPO-based architectures, the two stable optical phases $0$ and $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$0 encode $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$1; in spatial OPO proposals the sign of the real field amplitude $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$2 plays the same role; and in a Kerr-resonator CIM the spin is encoded in a polarization state inferred from alternating hybrid-mode intensities rather than phase-sensitive homodyne output (Marandi et al., 2014, Strinati et al., 2021, Quinn et al., 2024).

This analog encoding is also the source of one recurrent difficulty: local-field or Zeeman terms do not naturally match pairwise interaction terms when pulse amplitudes are nonuniform. The mismatch between interaction and Zeeman scales in mean-field and measurement-feedback CIMs motivated several artificial-Zeeman constructions, including absolute mean amplitude, auxiliary-spin, and chaotic-amplitude-control methods (Gunathilaka et al., 2023).

2. Physical architectures

The foundational experimental CIM architecture is the time-division-multiplexed DOPO network in a single ring cavity. In the 4-spin demonstration, the cavity round-trip time satisfied $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$3, so multiple temporally separated pulses circulated simultaneously; delay lines implemented the mutual injections that encoded $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$4; and a one-bit-delay Michelson interferometer read out the relative phases. That system solved a 4-vertex NP-hard MAX-CUT instance with no computational error detected in 1000 runs (Marandi et al., 2014). A later 16-bit telecom-wavelength implementation used 16 time-multiplexed femtosecond DOPO pulses and reported success rates above $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$5 on one-dimensional ferromagnetic and antiferromagnetic rings and on 16-vertex cubic-graph instances, with gradual pumping and multimode pulse dynamics identified as key ingredients (Takata et al., 2016).

A distinct family comprises measurement-feedback CIMs. In these systems, each pulse is partially measured every round trip, the quadrature values are digitized, and an FPGA computes feedback fields via matrix-vector multiplication before reinjection. The Stanford and NTT systems analyzed in the literature are of this type, as is the 512-signal-pulse fiber-ring machine used to study a $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$6 ferromagnetic square-lattice Ising model (King et al., 2018, Takesue et al., 2023). The same architectural template underlies several quantum-optical models: discrete-time Kraus-operator descriptions, Gaussian-state models with measurement backaction, and exact positive-$H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$7 simulations of the feedback master equation (Yamamura et al., 2017, Ng et al., 2021, Kiesewetter et al., 2021).

Later work broadened the architectural space. One proposal replaced the optical-electronic feedback bottleneck with an all-optical spatial CIM built from a $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$8 parametric cavity and an intra-cavity spatial light modulator (SLM), using either a Fourier-plane circulant-coupling scheme or a real-space optical vector-matrix multiplication scheme (Strinati et al., 2021). A conceptually different experimental platform used spontaneous polarization symmetry breaking in a driven fiber Kerr resonator, with spins encoded in polarization and read out by intensity-only detection; a localized birefringent defect induced a topologically symmetry-protected regime intended to suppress bias (Quinn et al., 2024). Digital descendants include FPGA “cyber CIMs,” which implement physics-derived CIM equations directly in reconfigurable logic rather than in optical hardware (Aonishi et al., 2024).

3. Operating principles and dynamical models

The common physical picture is threshold-driven collective selection in a nonlinear gain-loss landscape. In the original OPO CIM formulation, the network has a phase-state-dependent photon decay rate corresponding to the Ising energy landscape; gradual increase of parametric gain lets the lowest-loss configuration reach threshold first, after which gain saturation suppresses higher-loss modes (Marandi et al., 2014). In the 16-bit femtosecond system, gradual pumping improved the probability of obtaining the ground state, while multimode dynamics enabled “multimode tunneling,” in which higher temporal or spectral modes transiently assist phase flips without complete pulse extinction (Takata et al., 2016).

At the equation level, CIM models span several levels of detail. For time-multiplexed OPOs, coupled c-number Langevin equations describe in-phase and quadrature components $H=-\sum_{ij}^{N}J_{ij}\sigma_i\sigma_j,$9 and $$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$0 with pump gain, nonlinear saturation, mutual coupling, and vacuum noise (Marandi et al., 2014). Measurement-feedback quantum models decompose each round trip into phase-sensitive amplification, output coupling, homodyne measurement, and coherent feedback injection, yielding discrete-time conditional state updates in the quadrature basis (Yamamura et al., 2017). Exact positive-$$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$1 phase-space simulations of the full feedback master equation, and discrete-time Gaussian-state models including pump depletion and measurement backaction, were later developed for feedback CIMs (Kiesewetter et al., 2021, Ng et al., 2021).

More abstract formulations make the continuous-state character explicit. SimCIM simplifies the pulse-amplitude dynamics to a real update

$$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$2

followed by hard saturation $$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$3, thereby treating the CIM as a noisy continuous-amplitude dynamical system rather than a native binary-spin machine (Tiunov et al., 2019). A later synthesis goes further and interprets both delay-line and measurement-feedback CIMs as approximate integrators of Langevin-type stochastic differential equations,

$$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$4

with pump/loss/nonlinearity in $$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$5, objective information in $$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$6, and optical or measurement noise in the diffusion term (Khosravi et al., 19 Jul 2025). This suggests a wider computational scope than Ising optimization alone.

4. Classical emulation, quantum character, and the central controversy

One influential line of argument holds that measurement-feedback CIMs, at least in their Stanford and NTT realizations, are not computationally quantum in any essential many-body sense. The architectural basis of this claim is that the pulses are repeatedly measured, interact only through a classical FPGA, and therefore do not sustain useful inter-pulse entanglement or coherent many-spin dynamics. In that view, the computational core is the repeated classical evaluation of mean fields $$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$7, and the optical subsystem acts as an analog dynamical wrapper around a classical mean-field feedback loop (King et al., 2018). A closely related classical emulator, noisy mean-field annealing (NMFA), was shown to track CIM success-probability and time-to-solution scaling closely on SK and MAX-CUT benchmarks while running roughly 20 times faster than the NTT machine and roughly 130 times faster than the Stanford machine for dense $$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$8 MAX-CUT, under the comparison conventions used there (King et al., 2018).

A second classical reinterpretation, SimCIM, replaced nonlinear optical saturation by hard clipping and treated the machine as a stochastic continuous-spin iteration with momentum. It outperformed both the physical CIM and NMFA on the paper’s benchmark set, and it handled arbitrary real-valued couplings that the cited experimental FPGA implementation restricted for throughput reasons (Tiunov et al., 2019). More recently, a broader survey argued that CIMs are inherently continuous-state machines that approximately implement Langevin dynamics, while also stressing that repeated ADC/DAC conversions and electronic feedback in hybrid architectures create a major bottleneck for speed and energy efficiency (Khosravi et al., 19 Jul 2025).

This classical reading does not eliminate the quantum-optical content of CIM dynamics; it localizes it. A discrete-time quantum model of a two-pulse measurement-feedback CIM found transient anti-squeezing, macroscopic coherence between opposite-sign states, and Wigner negativity in weak-measurement regimes, even though the architecture generates no inter-pulse entanglement because the couplings are mediated by local operations and classical communication (Yamamura et al., 2017). Exact positive-$$E_{\rm Ising}(\tau)= -\frac{1}{2}\sum_{ij} J_{ij}\sigma_i(\tau)\sigma_j(\tau),$$9 simulations of the feedback master equation likewise included internal quantum noise, measurement-feedback diffusion, and above-threshold DOPO dynamics (Kiesewetter et al., 2021). A discrete-time Gaussian-state model argued that quantum-noise-dominated, short-photon-lifetime operation can improve sampling of degenerate ground and low-energy states, and that low finesse can reduce the number of round trips needed relative to high-finesse continuous-time expectations (Ng et al., 2021). In the strong-gain-saturation, few-photon regime, standard Gaussian CIM theory becomes inaccurate, and a skew-Gaussian extension retaining $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$0 and $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$1 was proposed to better match quantum-master-equation success probabilities (Inui et al., 2024). The controversy is therefore not whether CIMs exhibit quantum-optical effects locally, but whether those effects are the decisive source of observed optimization performance.

5. Thermodynamic behavior, sampling, and control mechanisms

Beyond optimization, several papers study CIMs as finite-temperature or low-energy samplers. In a 512-pulse measurement-feedback DOPO network implementing a $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$2 ferromagnetic square-lattice Ising model, the sampled configurations were analyzed under a canonical-ensemble hypothesis. Effective inverse temperature $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$3 was estimated by maximum likelihood using the Wang-Landau density of states, and the resulting internal energy, entropy, and RMS magnetization showed phase-transition-like behavior near $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$4, close to Onsager’s critical value $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$5 and markedly closer to exact and finite-size numerical results than mean-field theory (Takesue et al., 2023). The same study found that noisy mean-field annealing tracked the mean-field curve rather than the experimental CIM data, and that the canonical interpretation degraded at sufficiently large $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$6, consistent with freeze-out or selected-mode effects (Takesue et al., 2023).

Sampling of degenerate ground states and low-lying excitations is another recurring theme. A discrete-time Gaussian quantum model of a measurement-feedback CIM showed that, for SK1 or binary-signed MAX-CUT instances, the number of round trips sufficient to sample all configurations up to the first-excited energy, including all degeneracies, scaled as $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$7, with a median sufficient sampling time of $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$8 round trips at $$H = - \dfrac{1}{2}\sum_{r=1}^N\sum_{r'=1}^NJ_{rr'}\sigma_r \sigma_{r'} - \sum_{r=1}^N h_r\sigma_r,$$9; at a 10 GHz repetition rate this corresponds to about $h_r$0 ms (Ng et al., 2021). A closed-loop error-correction CIM with adaptive pump, target intensity, and nodewise coupling-strength control displayed “migration” among local minima rather than one-shot convergence, improved success probability relative to an open-loop baseline, and approximately Boltzmann-like energy sampling with fitted $h_r$1 on an $h_r$2 instance possessing degenerate ground and excited states (Kako et al., 2020).

A major control theme is the suppression of amplitude heterogeneity and the deliberate destabilization of local minima. Optical error-correction circuits were proposed to realize chaotic amplitude control (CAC), chaotic feedback control (CFC), and separated feedback control (SFC) directly in optics, with programmable all-to-all Ising coupling and even directional $h_r$3 couplings that induce chaotic behavior (Reifenstein et al., 2021). For local fields, mean-field CIM studies found that CAC outperformed both absolute-mean-amplitude and auxiliary-spin methods for artificial Zeeman terms in both mean-field and more physically detailed models (Gunathilaka et al., 2023). Frustrated Ising models motivated a different extension: by adding ancillary modes and modified collective-loss couplings, frustrated problems can be embedded into frustration-free enlarged optical systems whose target states become dark states; the same ancillas also support error detection and excited-state search (Zhou et al., 2024). A plausible implication is that CIM control theory increasingly treats the machine not as a passive bifurcating network but as an actively shaped non-equilibrium search process.

6. Scalability, applications, and open limitations

Claims of scalability and versatility differ sharply across architectures. The all-optical spatial CIM with an intra-cavity SLM argued that optical state evolution and coupling can both occur in parallel, avoiding the optical-electronic bottleneck. In its circulant Fourier-plane scheme, the transverse SLM grid could in principle encode $h_r$4 OPOs, with $h_r$5 for commercial $h_r$6; in the fully programmable real-space matrix-multiplication scheme, arbitrary couplings are supported but only $h_r$7, giving $h_r$8 (Strinati et al., 2021). The same proposal suggested a short cavity with $h_r$9, round-trip time $s_i\in[-1,1]$0, and total computation time around $s_i\in[-1,1]$1 for $s_i\in[-1,1]$2 round trips, but it also stated that the size-independent-runtime claim is an architectural ideal, not a general complexity result, and that the numerical evidence compared only $s_i\in[-1,1]$3 and $s_i\in[-1,1]$4 (Strinati et al., 2021).

Several later platforms emphasized engineering robustness. A polarization-based Kerr-resonator CIM demonstrated continuous operation for over one hour with no resetting, no rejected trials, and no manual adjustments, used only standard telecommunications fiber components at 1550 nm, and implemented 1D antiferromagnetic spin chains up to 100 spins; its measured time-to-solution on the tested chain family was consistent with $s_i\in[-1,1]$5, although the benchmark topology was intentionally simple (Quinn et al., 2024). A femtosecond-pumped fiber-based DOPO CIM reported a 55% average success rate on the 100-vertex Möbius Ladder MAX-CUT benchmark, with 464 $s_i\in[-1,1]$6s per run, and maintained operation for over 8 hours on a 100-vertex random graph, finding the exact optimum in 50.58% of 18,000 runs and a cut at least 99% of optimum in 98.58% of runs (Wei et al., 8 Dec 2025). These works suggest that stability and control are now central performance metrics alongside raw success probability.

Digital descendants extend CIM dynamics into large-scale programmable hardware and domain applications. A cyber CIM implemented on a single FPGA supported open-loop CIM, closed-loop CIM, and Jacobi SOR, used FP32 couplings and FP32 Zeeman terms, reached $s_i\in[-1,1]$7 with full coupling on one FPGA, and was reported to run more than ten times faster than a GPU implementation while enabling applications such as CDMA multi-user detection and $s_i\in[-1,1]$8 compressed sensing (Aonishi et al., 2024). A later SNN-derived CIM with global mean-amplitude feedback introduced a stabilization term $s_i\in[-1,1]$9, reported up to a 27% improvement in Max-Cut success rate over a baseline SNN-CIM, and was applied to traffic assignment on both a $x_i\in\mathbb{R}$0 grid and a 481-spin Beijing road-network model (Jiang et al., 17 Sep 2025). These cyber and physics-inspired variants reinforce the view that “CIM” has become a family of dynamical algorithms and hardware realizations rather than a single fixed optical device.

The limitations are equally persistent. Hybrid measurement-feedback machines derive programmability from electronic processing, yet that same ADC/FPGA/DAC loop can dominate latency and power, obscuring the speed and energy advantages of optics (Khosravi et al., 19 Jul 2025). Fully optical systems avoid that bottleneck but often sacrifice coupling generality or remain at proof-of-principle scale (Strinati et al., 2021, Quinn et al., 2024). Benchmarking remains heterogeneous: some papers emphasize simple but exactly soluble graph families, others use SK or G-set instances, and several application papers rely on problem-specific mappings or approximations (Wei et al., 8 Dec 2025, Jiang et al., 17 Sep 2025). The cumulative record suggests that CIMs are best regarded as a broad class of open-dissipative, often continuous-state, optical or cyber-physical optimizers whose most distinctive scientific questions concern analog dynamics, control, and physical sampling, while their most pressing engineering questions concern coupling scalability, interface overhead, and reproducible system-level benchmarking.