Reduced-Order Lagrangian Particle Model

• Reduced-Order Lagrangian Particle Models are mathematical frameworks that simplify high-dimensional systems by using lower-dimensional particle, latent, or collective coordinates.
• They employ techniques such as gyrokinetic reductions, coarse-grained Langevin closures, and collective-coordinate methods to preserve transport properties and underlying variational structures.
• These models enable efficient simulation of complex phenomena in turbulence, soft robotics, and wave dynamics while balancing accuracy and computational efficiency.

Reduced-Order Lagrangian Particle Model denotes a class of reduced descriptions in which the state of a high-dimensional Lagrangian system is replaced by a lower-dimensional dynamical law written in particle coordinates, collective coordinates, latent coordinates, sparse material points, or explicitly memory-augmented observables. In current arXiv literature, closely related constructions appear in gyrokinetic reduced-particle dynamics, coarse-grained Langevin and Mori–Zwanzig closures, collective-coordinate reductions of dispersive waves, structure-preserving latent Lagrangian networks, and sparse-graph or basis-function simulators for Lagrangian continua (Tronko et al., 2017, Ma et al., 2018, Anderson et al., 2022, Friedl et al., 2024, Viswanath et al., 2024, Wit et al., 21 Jul 2025, Do et al., 8 Jun 2026).

1. Scope and state representations

The phrase covers several non-equivalent but structurally related notions of reduction. In the most literal particle sense, the reduced model keeps a lower-dimensional particle state and its induced characteristic equations. In gyrokinetics, the reduced-particle model eliminates fast gyromotion and evolves a gyrocenter state in coordinates such as $(\mathbf X,p_z,\mu)$, with $\mu$ an adiabatic invariant and $(\mathbf X,p_z)$ determined by a reduced Hamiltonian (Tronko et al., 2017). In coarse-grained Langevin dynamics, one splits a full particle state into resolved and unresolved components,

$$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$

and then eliminates $(\xi,\eta)$ to obtain effective dynamics for coarse positions $q$ and coarse velocities $p$ (Ma et al., 2018).

A second interpretation retains not full particle positions and velocities, but a reduced set of observables from which trajectories are reconstructed. In turbulent tracer modeling, the resolved variables are the particle acceleration $\mathbf a(t)$ and the local velocity-gradient tensor $\nabla \mathbf u(t)$, while velocity and position are recovered by

$$\mathbf v_{n+1}=\mathbf v_n+\mathbf a_n\Delta t,\qquad \mathbf x_{n+1}=\mathbf x_n+\mathbf v_n\Delta t,$$

with tracer kinematics

$\mu$0

The reduced state is therefore particle-centered but not identical to the full tracer trajectory state (Wit et al., 21 Jul 2025). In nonlinear Schrödinger reductions, the “particle-like” state is a finite set of collective coordinates such as

$\mu$1

or, for translating packets, $\mu$2, where amplitude, width, chirp, phase, and center position evolve by finite-dimensional ODEs (Anderson et al., 2022).

A third interpretation treats the Lagrangian state as a field over reference material coordinates rather than as an unordered cloud. OnlyDense defines

$\mu$3

with $\mu$4, and approximates snapshots by a linear subspace

$\mu$5

Lag-pDMD and LagCAE-pDMD similarly use augmented Lagrangian snapshots such as

$\mu$6

so that the reduced state contains both transported coordinates and attached field values (Do et al., 8 Jun 2026, Zhang et al., 9 Mar 2026). This suggests that reduced-order Lagrangian particle modeling is best understood as a family of reductions centered on moving-material descriptions rather than a single canonical state choice.

2. Variational and geometric foundations

A central line of work defines reduced models by preserving the Lagrangian structure itself. For a mechanical system with configuration manifold $\mu$7, Lagrangian

$\mu$8

and generalized forces $\mu$9, the full equations are

$(\mathbf X,p_z)$0

A reduced configuration manifold $(\mathbf X,p_z)$1 with embedding $(\mathbf X,p_z)$2 induces the reduced Lagrangian by pullback,

$(\mathbf X,p_z)$3

and the reduced mass metric

$(\mathbf X,p_z)$4

This Riemannian interpretation is explicit in learned reduced-order Lagrangian dynamics and latent control frameworks, where the latent space is treated as an embedded manifold and the mass matrix as the kinetic-energy metric (Friedl et al., 2024, Friedl et al., 9 Feb 2026).

The same principle appears in nonlinear model reduction for structural dynamics. Instead of hyper-reducing the equations term by term, one first approximates the “Lagrangian ingredients”—metric or mass matrix, potential-energy function, dissipation function, and external force—and then derives the ROM by the forced Euler–Lagrange equation. In that setting, the reduced configuration ansatz

$(\mathbf X,p_z)$5

yields a reduced model of the form

$(\mathbf X,p_z)$6

with symmetry and positive definiteness preserved by construction (Carlberg et al., 2014). The same paper emphasizes that this differs fundamentally from equation-based hyper-reduction, which can destroy symmetry, positive definiteness, and mechanical interpretability.

Nonintrusive operator-inference variants preserve the same second-order structure directly from data. Lagrangian Operator Inference and related structure-preserving ROMs learn reduced mass, damping, stiffness, and input operators constrained so that the reduced system remains a valid mechanical model. For soft-robot and wave-equation examples, the learned ROMs take forms such as

$(\mathbf X,p_z)$7

or, in conservative nonlinear-wave settings,

$(\mathbf X,p_z)$8

with symmetry and positive-definiteness constraints used to maintain Lagrangian structure (Sharma et al., 2022, Sharma et al., 2024). This structural viewpoint is one of the clearest distinguishing marks of the field.

3. Memory, stochasticity, and non-Markovian closure

A major theme in reduced-order Lagrangian particle modeling is that unresolved degrees of freedom reappear as memory and noise. The most explicit formulation is the Mori–Zwanzig generalized Langevin equation,

$(\mathbf X,p_z)$9

and its discrete counterpart

$$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$0

In turbulent tracer dynamics, the reduced model truncates the orthogonal dynamics $$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$1, keeps explicit finite memory, and evolves

$$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$2

where the memory kernels are learned as MLPs and organized with Takens-style delay embedding and logarithmically spaced lags $$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$3 (Wit et al., 21 Jul 2025). The paper reports that training only on short-horizon pointwise MSE with $$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$4, $$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$5, $$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$6, and $$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$7 suffices to recover both short-time trajectory accuracy and long-time heavy-tailed statistics at test time.

The same structural phenomenon appears in coarse-grained Langevin dynamics. Eliminating unresolved coordinates from a full Langevin system yields a generalized Langevin equation

$$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$8

with memory kernel

$$x(t)=\Phi q(t)+\Psi \xi(t),\qquad v(t)=\Phi p(t)+\Psi \eta(t),$$9

and colored noise satisfying the second fluctuation-dissipation theorem,

$(\xi,\eta)$0

Projection onto Krylov subspaces produces a finite-dimensional Markovian embedding for the hidden variables, and for order less than six the reduced model obtained from subspace projection automatically satisfies the fluctuation-dissipation theorem (Ma et al., 2018). In this literature, memory is not an auxiliary modeling convenience; it is the canonical reduced representation of unresolved particle physics.

4. Transport-aligned and collective-coordinate reductions

One historically important form of reduced-order Lagrangian particle model is the collective-coordinate reduction of coherent structures. For the nonlinear Schrödinger equation, a Gaussian packet ansatz

$(\xi,\eta)$1

with

$(\xi,\eta)$2

leads, in the co-moving frame, to the reduced equations

$(\xi,\eta)$3

together with the corresponding phase equation (Anderson et al., 2022). The same paper shows that the reduced Lagrangian method and Reduced-Order Nonlinear Solutions (RONS) are the imaginary and real parts of a single complex-valued master equation, and proves that they coincide for ansatz classes such as exponentials of complex polynomials in $(\xi,\eta)$4. It also provides an important critique: in the stationary-frame NLS with a translating Gaussian ansatz, the reduced Lagrangian contains no dependence on $(\xi,\eta)$5 or $(\xi,\eta)$6, so it yields no ODE for packet translation, whereas RONS correctly gives $(\xi,\eta)$7. For modified NLS, where no known Lagrangian exists, the reduced Lagrangian approach is inapplicable while RONS remains usable.

Transport-dominated ROMs motivate a different, but related, Lagrangian reduction: one augments the reduced state with moving coordinates. Lagrangian DMD replaces fixed-grid snapshots by observables

$(\xi,\eta)$8

so that particle or grid positions are reduced jointly with transported values (Lu et al., 2019). A more recent parametric formulation writes the Lagrangian snapshot as

$(\xi,\eta)$9

or, in higher dimensions, as coordinate channels plus solution channels, and then applies either linear reduced DMD or autoencoder compression in that augmented space (Zhang et al., 9 Mar 2026). The theoretical motivation is explicit: in simple transport problems the Kolmogorov $q$0-width decays slowly in the Eulerian frame but can collapse to zero after only a few dimensions in the Lagrangian frame, for example

$q$1

for a one-dimensional transport-with-jump example. This suggests that the essential reduced variable for transport is often not the field alone, but the pair consisting of the transported coordinates and the carried quantity.

5. Learned latent, sparse-graph, and basis-function particle simulators

Recent work has generalized reduced-order Lagrangian particle modeling into explicitly learned simulators. GIOROM evolves dynamics only on a sparse reduced-order set of particles, typically about $q$2–$q$3 points, and reconstructs dense states at arbitrary query points using a neural-field ROM. Its reduced state is the sparse graph of material points with velocity history, and its learned operator

$q$4

predicts reduced accelerations from recent reduced velocities. The framework is reported to operate at roughly $q$5 of the full-order degrees of freedom in some experiments, while reconstructing point clouds on the order of $q$6 points from sparse graphs with $q$7 points (Viswanath et al., 2024). Its physical interpretation is hybrid: reduced dynamics are evolved on sparse particles, while dense reconstruction uses a learned low-dimensional basis expansion.

A complementary line learns nonlinear latent manifolds and reduced Lagrangian mechanics on them. RO-LNN learns an embedding $q$8, reduction map $q$9, latent mass matrix, and latent potential for systems including a 192-DoF rope and a 600-DoF cloth whose generalized coordinates are Cartesian positions of mass centers. In one reported benchmark, a 16-DoF pendulum rollout over $p$0 steps takes $p$1 s for the symbolic full-order model and $p$2 s for the ODE-trained RO-LNN, roughly a $p$3 speedup (Friedl et al., 2024). A later latent-control framework extends this viewpoint to tracking and stabilization, including a 15-DoF augmented pendulum and a 351-coordinate plush puppet, and analyzes the latent closed-loop system as a mechanically structured disturbance-driven ROM (Friedl et al., 9 Feb 2026).

OnlyDense takes a different route by treating the state as a function over reference configuration and learning a linear subspace of neural basis functions. Snapshots are approximated by

$p$4

coefficients are obtained by projection or least squares,

$p$5

and latent dynamics are then learned in coefficient space (Do et al., 8 Jun 2026). On large-scale SPH data, including a projectile dataset with approximately $p$6 million particles, the paper reports $p$7 with as few as $p$8 basis functions. A related turbulence study constructs a hierarchy of reduced Lagrangian models ranging from Neural ODE acceleration laws to increasingly structured weakly compressible SPH closures at coarse resolutions $p$9, explicitly testing how much generalizability is gained by enforcing SPH symmetries and learnable kernels (Woodward et al., 2021). Together, these works show that “reduced-order Lagrangian particle model” now includes sparse graph simulators, latent geometric mechanics, and continuous basis-function ROMs, not only classical collective-coordinate or projection methods.

6. Applications, performance, and limitations

The application range is unusually broad. In turbulence, a Mori–Zwanzig tracer ROM is trained on a three-dimensional homogeneous isotropic turbulence DNS in a periodic cube of size $\mathbf a(t)$0, discretized on a $\mathbf a(t)$1 pseudospectral grid at $\mathbf a(t)$2, using up to $\mathbf a(t)$3M Lagrangian tracer trajectories from the TURB-Lagr database. It reproduces heavy-tailed acceleration PDFs, acceleration autocorrelation, $\mathbf a(t)$4–$\mathbf a(t)$5-type velocity-gradient diagnostics, joint statistics of $\mathbf a(t)$6 and $\mathbf a(t)$7, and recovers realistic turbulence after an out-of-distribution initialization from fluid at rest with only small perturbations; the KL divergence falls close to zero after a transient of order $\mathbf a(t)$8 (Wit et al., 21 Jul 2025). In gyrokinetics, the reduced-particle model is the central object from which both particle characteristics and polarization/magnetization terms in field equations are derived, and the literature distinguishes a second-order full-FLR Hamiltonian model from a second-order long-wavelength approximation suitable for particle-in-cell implementation (Tronko et al., 2017). In soft robotics, Lagrangian Operator Inference is demonstrated on a model with $\mathbf a(t)$9 degrees of freedom, where preserving the underlying Lagrangian structure is reported to improve predictive accuracy and robustness to unseen inputs (Sharma et al., 2024).

The literature also records several sharp limitations and corrections to naive intuition. Reduced Lagrangian collective-coordinate models are not universally reliable: for stationary-frame NLS they can fail to produce the packet-center equation, and for modified NLS they may be inapplicable altogether (Anderson et al., 2022). Transport-aligned reductions are powerful but not unconstrained: Lagrangian DMD is demonstrated only on one-dimensional shock-free examples, and the broader Lagrangian prediction literature states that strong deformations and shocks will require robust remapping and regularization strategies (Lu et al., 2019, Zhang et al., 9 Mar 2026). Learned latent Lagrangian models for deformables still lack explicit treatment of collisions, contact impulses, topology change, strong frictional dissipation, or particle permutation symmetries in their basic dense-autoencoder formulations (Friedl et al., 2024). OnlyDense, despite its million-particle demonstrations, is explicitly tied to a single reference configuration (Do et al., 8 Jun 2026). The tracer Mori–Zwanzig ROM is trained and tested only for stationary homogeneous isotropic turbulence at $\nabla \mathbf u(t)$0, and does not demonstrate transfer across Reynolds numbers, forcing conditions, anisotropic or wall-bounded flows, or inertial particles (Wit et al., 21 Jul 2025).

Taken together, these results support a precise but nontrivial conclusion. A reduced-order Lagrangian particle model is not merely a smaller simulator. It is a reduced dynamical law that attempts to preserve the moving-material character of the original system, whether by retaining particle coordinates directly, by encoding unresolved physics as memory and colored noise, by evolving collective coordinates of coherent structures, or by learning a latent manifold or sparse particle graph with mechanically or geometrically constrained dynamics. The literature suggests that reduction is most successful when the chosen state variables remain aligned with transport, symmetries, and variational structure, and least reliable when those structural features are replaced by purely kinematic compression.