Inertial Kuramoto Model Overview

• Inertial Kuramoto model is a second-order extension that adds inertia and damping to classical phase oscillators, enabling underdamped dynamics and bistability.
• It shows how inertia and network topology raise synchronization thresholds, induce hysteresis, and support both phase-locked and running states.
• The model applies to power grids, Josephson junctions, and more, offering insights into transient dynamics, clustering, and synchronization transitions.

The inertial Kuramoto model, also called the second-order Kuramoto model, is the canonical extension of the classical Kuramoto phase-oscillator system obtained by adding inertia and damping to the phase dynamics. In its globally coupled form it is written as

$$m\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\frac{\lambda}{N}\sum_{j=1}^N \sin(\theta_j-\theta_i),$$

while on a network it is commonly written as

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$

Here $\theta_i$ is the phase, $\dot{\theta}_i$ the instantaneous angular velocity, $\ddot{\theta}_i$ the angular acceleration, $\omega_i$ the intrinsic torque or natural frequency, $m$ the inertia, $\alpha$ the damping, and $A_{ij}$ the adjacency matrix. Relative to the first-order Kuramoto model, the inertial formulation enlarges the state space from phases alone to phase–velocity pairs and thereby changes the synchronization problem from overdamped relaxation to underdamped collective dynamics, with direct relevance to power-grid swing dynamics, Josephson junction arrays, and other systems in which transient frequency excursions are essential (Rodrigues et al., 2015).

1. Formulation and basic dynamical structure

The first-order Kuramoto model on a complete graph is

$$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$

with order parameter

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$0

so that the dynamics can be rewritten as

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$1

On a general network this becomes

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$2

The inertial model adds a second time scale and converts the phase dynamics into a damped second-order system. In the review literature it is presented as the principal second-order extension of the Kuramoto framework and as the natural model whenever phase-only dynamics is too overdamped to represent realistic transients (Rodrigues et al., 2015).

A one-node reduction,

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$3

is interpreted as a driven damped pendulum, a Josephson junction, or a one-machine infinite-bus power-grid model. This analogy is central because it explains why the inertial model supports both stable fixed points and stable rotating solutions. In the first-order model oscillators instantaneously relax toward the mean field; in the inertial model they can overshoot, oscillate, and sustain running states. The addition of inertia therefore produces bistability between phase-locked and running states, delayed approach to synchrony, richer cluster dynamics, and protocol dependence under adiabatic continuation in coupling (Rodrigues et al., 2015).

The statistical-mechanics formulation with noise makes the distinction sharper. In the generalized inertial noisy model,

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$4

the continuum limit is a Kramers equation in phase–velocity space. For $$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$5 the model is an equilibrium Brownian mean-field system, whereas for nonzero quenched frequency disorder detailed balance is violated and the stationary state is a nonequilibrium stationary state (Gupta et al., 2014).

2. Bifurcation structure, locking thresholds, and hysteresis

The defining phenomenology of the inertial Kuramoto model is the asymmetry between forward and backward synchronization. In the small-damping pendulum picture, the homoclinic bifurcation line is

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$6

and this local coexistence of a stable fixed point with a stable limit cycle underlies network-level hysteresis (Rodrigues et al., 2015). In the fully connected mean-field treatment of Tanaka, Lichtenberg, and Oishi, the dynamics is written as

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$7

and the total coherence splits into locked and drifting contributions,

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$8

Under increasing and decreasing continuation in $$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$9, distinct branches $\theta_i$0 and $\theta_i$1 arise, with corresponding critical couplings $\theta_i$2 and $\theta_i$3 (Rodrigues et al., 2015).

The key asymmetry is encoded in the locking boundaries. For decreasing coupling, the locking boundary is

$\theta_i$4

whereas for increasing coupling and small inertia,

$\theta_i$5

Thus the backward desynchronization threshold coincides with the classical first-order threshold, while the onset threshold is shifted upward by inertia. For unimodal symmetric $\theta_i$6, the standard Kuramoto threshold

$\theta_i$7

is replaced by

$\theta_i$8

which reduces to the classical value as $\theta_i$9. For Lorentzian $\dot{\theta}_i$0 of width $\dot{\theta}_i$1,

$\dot{\theta}_i$2

These formulas make explicit that inertia raises the critical coupling and makes synchronization harder (Rodrigues et al., 2015).

Finite-size studies refine this picture. For Gaussian frequencies and sufficiently large inertia, the transition is hysteretic and the coexistence region contains many clustered partially synchronized states. The minimal coupling required to observe the coherent state is weakly size dependent, whereas the maximal coupling sustaining incoherence increases with system size and, in the thermodynamic limit, grows proportionally to the mass. Large inertia also produces coherently drifting or “whirling” clusters, whose birth and merger cause step-like synchronization profiles and oscillations in the order parameter $\dot{\theta}_i$3 (Olmi et al., 2014). Within the same framework, a generalized self-consistency construction indexed by an arbitrary maximal locking frequency $\dot{\theta}_i$4 yields families of intermediate partially synchronized states rather than only the two extremal hysteresis branches (Olmi et al., 2014).

3. Finite-$\dot{\theta}_i$5 stability, relaxation, and the small-inertia regime

Rigorous finite-dimensional analysis has established that phase-locked states of the inertial Kuramoto model can be nonlinearly stable in robust norms. For globally coupled oscillators with finite inertia, a class of phase-locked states with diameter below $\dot{\theta}_i$6 is orbitally $\dot{\theta}_i$7-stable, and perturbations converge to the same locked profile up to a uniform phase shift. The asymptotic shift is explicit: $\dot{\theta}_i$8 This formula isolates a specifically inertial effect: the average initial frequency, multiplied by $\dot{\theta}_i$9, selects the orbit representative within the rotational equivalence class (Choi et al., 2011).

For identical oscillators with homogeneous inertia and damping, complete synchronization has been proved under a simple energy–coherence condition. If

$\ddot{\theta}_i$0

then relative frequencies vanish asymptotically and the solution converges to a phase-locked state of $\ddot{\theta}_i$1-type, namely either a one-cluster synchronized state or a bipolar state with two groups separated by $\ddot{\theta}_i$2 (Choi et al., 2017). The same work establishes uniform-in-time approximation of the finite system by the kinetic mean-field model and proves global existence and large-time asymptotics of measure-valued solutions (Choi et al., 2017).

Recent work has shifted the emphasis from near-locked initial conditions to generic initial phase configurations. In the all-to-all inertial model, one sufficient framework for asymptotic phase-locking can be summarized heuristically as

$\ddot{\theta}_i$3

Under such conditions, the relaxation dynamics decomposes into three stages: an initial layer, during which inertial memory and initial velocities dominate; a condensation stage, during which the order parameter becomes quasi-monotone and a majority cluster forms; and a persistence/relaxation stage, during which that majority cluster remains confined and the full configuration converges to a phase-locked traveling state. The proof combines Duhamel estimates, quasi-monotonicity of the order parameter, nonlinear Grönwall inequalities for majority-cluster diameter, and a Łojasiewicz-type argument for inertial analytic gradient flows (Cho et al., 2 Mar 2025).

The small-inertia regime admits an even sharper perturbative treatment. The inertial system can be written as a singular slow–fast perturbation of the first-order Kuramoto model,

$\ddot{\theta}_i$4

with slow manifold

$\ddot{\theta}_i$5

A quantitative Tikhonov theorem then yields explicit bounds on phase and velocity differences between the inertial and first-order flows, with velocity error of the form

$\ddot{\theta}_i$6

The same analysis proves asymptotic phase-locking in a low-inertia, high-coupling, low-dispersion regime and shows that exact reconstruction of phase velocities from phase positions breaks down beyond a sharp threshold: it is globally possible only if $\ddot{\theta}_i$7, whereas for $\ddot{\theta}_i$8 it is possible only up to a finite explicit time $\ddot{\theta}_i$9 (Cho et al., 15 Aug 2025).

4. Networks, topology, and graph-coupled inertial synchronization

On networks, inertia interacts with topology in ways that do not simply parallel the first-order theory. A prominent example is degree–frequency correlation. For uncorrelated networks with

$\omega_i$0

the inertial network equation becomes

$\omega_i$1

and in mean field each degree class maps to an effective damped driven pendulum. The resulting synchronization is not all-at-once explosive synchronization but “cluster explosive synchronization”: oscillators synchronize in degree classes, with low-degree nodes potentially locking simultaneously and higher-degree classes joining successively as $\omega_i$2 increases (Rodrigues et al., 2015). The same review emphasizes that a full inertial analytical theory for assortativity, modularity, weighted couplings, and other complex topologies remains unavailable (Rodrigues et al., 2015).

A low-dimensional degree-based reduction has been proposed for large uncorrelated networks. Writing

$\omega_i$3

one obtains degree-class order parameters $\omega_i$4 and an effective inertial closure in which the locking scale is modified from $\omega_i$5 to

$\omega_i$6

For Lorentzian natural frequencies, the reduced evolution is supplemented by a fitted correction term

$\omega_i$7

yielding a practical low-dimensional description of both stationary and transient synchronization in scale-free networks (Ji et al., 2014).

Local graph structure also enters through rigorous synchronization estimates. For connected symmetric weighted networks with inertia and frustration,

$\omega_i$8

one line of work proves that the phase diameter enters a region $\omega_i$9 after finite time and then the frequency diameter decays exponentially,

$m$0

under explicit small-inertia, small-frustration, strong-coupling assumptions (Zhu et al., 2024). A complementary treatment uses convex combinations of ordered oscillators and hypocoercive energy functionals to overcome the lack of a usable second-order gradient structure and the singularity of second derivatives of diameters, again obtaining exponentially fast complete frequency synchronization on symmetric connected weighted graphs (Zhu et al., 2024).

At the kinetic level, graph topology changes the instability of the mixing state. On convergent families of graphs with graphon $m$1, the inertial mean-field limit leads to a Vlasov equation whose linear stability is governed by the dispersion relation

$m$2

where $m$3 is an eigenvalue of the graphon operator. For unimodal even $m$4, loss of stability occurs through a pitchfork bifurcation. For bimodal $m$5, Penrose-diagram analysis identifies both pitchfork and Andronov–Hopf bifurcations, with graph eigenmodes determining whether the emergent patterns are homogeneous or twisted, stationary or traveling. In the inertial setting this includes stationary antiphase partially locked states at negative coupling and traveling twisted cluster states that do not arise in the same way in the non-inertial model (Chiba et al., 2022).

5. Stochastic, kinetic, and asynchronous formulations

The noisy inertial Kuramoto model is naturally formulated in phase–velocity space. In mean field,

$m$6

with Gaussian white noise

$m$7

The one-oscillator density $m$8 then satisfies a Kramers-type Fokker–Planck equation,

$m$9

For identical oscillators, the critical coupling from incoherence to coherence is independent of inertia, but for bimodal natural-frequency distributions inertia destabilizes incoherence and can turn a supercritical transition into a subcritical one. In the small-inertia limit, the velocity-averaged Smoluchowski equation yields a self-consistency expansion with a quadratic term in $\alpha$0; that term is absent in the standard first-order symmetric case and is identified as the mechanism behind hysteresis. The same review states that noise reduces hysteresis and raises the critical coupling (Rodrigues et al., 2015).

A broader nonequilibrium formulation uses three dimensionless control parameters $\alpha$1, $\alpha$2, and $\alpha$3, where $\alpha$4 is the width of the natural-frequency distribution. In this representation the inertial noisy model has an explicit incoherent stationary state,

$\alpha$5

and the synchronization problem becomes a Kramers equation for $\alpha$6. The central result is that, for $\alpha$7, $\alpha$8, and $\alpha$9, the model generically exhibits a nonequilibrium first-order synchronization transition surface bounded by spinodals $A_{ij}$0 and $A_{ij}$1, with hysteresis, metastability, and finite-$A_{ij}$2 switching between incoherent and synchronized states (Gupta et al., 2014).

The PDE theory of the inertial mean-field equation has also been developed directly as a degenerate Kolmogorov–Fokker–Planck problem. For the nonlinear kinetic equation

$A_{ij}$3

existence, positivity, normalization, uniqueness under additional support assumptions, and a priori exponential decay estimates in $A_{ij}$4 have been proved using the theory of degenerate Kolmogorov operators. The same work constructs a numerical operator consistent with the hypoelliptic structure and uses it to study phase coherence

$A_{ij}$5

and velocity coherence

$A_{ij}$6

finding that larger coupling increases phase coherence, larger inertia significantly increases asymptotic frequency coherence, and smaller noise increases both (Pecorella et al., 2024).

In the asynchronous regime, inertia has a further nontrivial effect. For a disordered inertial Kuramoto model with random couplings, an iterative mean-field closure computes self-consistent power spectra of the local network noise and of single oscillators. In that asynchronous state the method agrees well with direct network simulations and reveals that temporal correlations are shortest at an intermediate mass: the oscillator correlation time is minimized, the power spectra are flattest and most white-noise-like, and the Kolmogorov–Sinai entropy is maximal near the same mass. Thus inertia reorganizes not only synchronization thresholds but also the fluctuation statistics of incoherent states (Kati et al., 11 Mar 2025).

6. Generalizations, applications, and open problems

Several important generalizations retain the second-order structure while modifying the coupling term. In the forced inertial Kuramoto model,

$A_{ij}$7

the forcing acts as a pinning term. For symmetric bimodal frequency distributions, forcing competes with the intrinsic two-cluster structure, suppresses the standing-wave regime, and makes the backward transition discontinuous for moderate forcing, in contrast with the continuous behavior known for forced unimodal inertial populations. For the bi-delta distribution with $A_{ij}$8, the backward branch is explicit: $A_{ij}$9 The same work reports that sufficiently strong forcing can entrain all oscillators and largely remove hysteresis (Agnihotri et al., 24 Apr 2026).

Phase-lag variants produce two distinct lines of development. In the second-order Kuramoto–Sakaguchi model,

$$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$0

the running-state reduction contains an inertia-induced effective phase shift $$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$1, so inertia can be interpreted as generating an additional state-dependent phase lag. This viewpoint explains why small inertia can soften or remove discontinuous transitions induced by Sakaguchi frustration and why sufficiently large inertia and phase lag can prevent complete synchronization even at very large coupling, replacing it with clustered oscillating states. A threshold for this non-synchronizing forward process is

$$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$2

(Gao et al., 2020). A different modification introduces a binary node-dependent phase lag after the lag-free inertial system has reached steady state. There the phase lag does not change the local Melnikov threshold directly; instead it changes the global self-consistent quantities $$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$3 and $$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$4, shifts the left and right boundaries of the primary cluster asymmetrically, and enables the primary synchronized cluster to merge with secondary clusters. This enhancement mechanism appears only when inertia is large enough to support multiple synchronized clusters, with a reported threshold $$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$5 for their parameter set (Yi et al., 5 Jun 2026).

Higher-order and adaptive extensions further enrich the model class. On mean-field simplicial complexes with pairwise and triadic interactions,

$$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$6

the effective coupling is

$$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$7

and inertia and triadic interactions have largely separated roles: inertia primarily determines the forward synchronization threshold, while triadic coupling primarily determines the backward desynchronization threshold, producing prolonged hysteresis. For Lorentzian frequencies,

$$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$8

(Sabhahit et al., 2023). In the inertial Kuramoto model with Hebbian learning,

$$\dot{\theta}_i=\omega_i+\frac{\lambda}{N}\sum_{j=1}^N\sin(\theta_j-\theta_i),$$9

the $$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$00 problem reduces to a two-dimensional longitudinal subsystem and a three-dimensional transverse subsystem. The equilibrium condition in the transverse subsystem is

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$01

so equilibria exist iff $$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$02; the same reduced system is dissipative, with divergence $$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$03, and numerically exhibits periodic, heteroclinic, and equilibrium-dominated regimes (Ruangkriengsin et al., 2022).

Applications remain a primary motivation. For power grids, writing the machine phase as

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$04

the swing equation reduces to

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$05

which becomes the inertial Kuramoto form under homogeneous parameters and lossless lines (Rodrigues et al., 2015). Basin stability has therefore become an important nonlocal complement to linear stability, with

$$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$06

as the numerical fraction of random perturbations from which node $$\ddot{\theta}_i=-\alpha\dot{\theta}_i+\omega_i+\lambda\sum_{j=1}^N A_{ij}\sin(\theta_j-\theta_i).$$07 returns to synchrony. In networked inertial systems, low basin stability is associated not simply with low degree but with particular motifs such as dead ends, dead trees, and gateway structures (Rodrigues et al., 2015). Finite-size simulations of the Italian high-voltage power grid further show that whirling clusters can generate quasi-periodic order-parameter oscillations and non-monotone synchronization profiles (Olmi et al., 2014).

Across this literature, a consistent picture emerges. The inertial Kuramoto model is the natural framework whenever rotational energy, damping, and large transients matter; it raises synchronization thresholds, generates bistability and hysteresis, supports coexisting fixed-point and running solutions, and amplifies sensitivity to topology, continuation protocol, and perturbation size (Rodrigues et al., 2015). At the same time, several analytical problems remain open: a complete mean-field theory for asymmetric frequency distributions, a network-resolved theory comparable to Ott–Antonsen methods for assortativity and modularity, systematic treatment of delays in second-order systems, and an inertia-independent global synchronization theory valid beyond the small-inertia perturbative regime (Rodrigues et al., 2015, Cho et al., 15 Aug 2025).