Lagrangian Descriptors in Dynamical Systems

• Lagrangian Descriptors are finite-time, trajectory-based scalar diagnostics that expose the geometric skeleton of dynamical systems by accumulating observables along orbits.
• They employ both continuous and discrete formulations to identify sharp transitions in phase-space, effectively detecting stable/unstable manifolds, separatrices, and chaotic regions.
• LDs have broad applications in fluid dynamics, chemical reactions, astrodynamics, and quantum systems, offering improved numerical diagnostics compared to traditional methods like FTLE and SALI.

Searching arXiv for recent and foundational papers on Lagrangian descriptors to ground the encyclopedia entry. arXiv Search Query: "Lagrangian descriptors review invariant manifolds Hamiltonian" Lagrangian descriptors (LDs) are finite-time, trajectory-based scalar diagnostics designed to expose the geometric skeleton of dynamical systems by accumulating a positive observable along orbits. Across the literature, LDs are used to reveal stable and unstable manifolds, separatrices, transport barriers, invariant tori, slow manifolds, chaotic layers, and bounded-motion regions by identifying sharp, non-smooth, or rapidly varying features in scalar fields over initial conditions (Lopesino et al., 2017). Their scope now spans autonomous and nonautonomous flows, discrete maps, stochastic differential equations, dissipative systems, chemical reaction dynamics, fluid transport, celestial mechanics, and recent semiclassical and quantum generalizations (Mancho et al., 2011).

1. Definitions and principal formulations

The original continuous-time formulation treats an LD as a finite-time accumulation of trajectory arc length. For a system

$\dot{\mathbf{x}}=\mathbf{v}(\mathbf{x},t),$

the basic symmetric descriptor is

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$

with forward and backward components $M^+$ and $M^-$ obtained by integrating on $[t_0,t_0+\tau]$ and $[t_0-\tau,t_0]$, respectively (Mancho et al., 2011). A closely related componentwise $p$-norm family, emphasized in the theoretical literature, is

$$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$

or equivalently, in autonomous notation, by integrating $\sum_i |f_i(\mathbf{x}(t),t)|^p$ along the orbit (Lopesino et al., 2017). The choice $0

Discrete-time analogues replace the time integral by orbitwise sums of step increments. For maps, one standard definition is

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$0

while componentwise forms

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$1

are used in higher-dimensional symplectic maps (Lopesino et al., 2017). This discrete viewpoint is central in the analysis of area-preserving maps and in later LD-based chaos diagnostics (Zimper et al., 2023).

The observable accumulated by an LD is not unique. Beyond arc length, the literature includes acceleration-based, curvature-based, and higher time-derivative variants, as well as formulations that integrate phase-space coordinates and momenta directly rather than velocities. In the LiCN$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$2LiNC application, for example, the descriptor is

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$3

with $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$4, and $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$5 denotes the phase-space coordinates themselves, not the velocity field (Revuelta et al., 2021). In stochastic settings, the corresponding stochastic LD is defined along sample paths by summing path increments,

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$6

so that randomness enters through the path rather than through an explicit diffusion term in the integrand (Balibrea-Iniesta et al., 2017).

2. Geometric mechanism and theoretical basis

The theoretical rationale for LDs is that invariant manifolds separate trajectories with qualitatively different forward or backward histories. When a positive quantity is accumulated along those trajectories, the resulting scalar field changes abruptly across the separating set. In the rigorous framework developed for $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$7, a “singular feature” is defined by nonexistence of a normal derivative of the LD across a surface, and such singular features are proven to coincide with stable and unstable manifolds in several two-dimensional hyperbolic settings (Lopesino et al., 2017).

For the linear autonomous saddle

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$8

the descriptor takes the explicit form

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$9

For $M^+$0, derivatives such as $M^+$1 become unbounded at $M^+$2, and likewise at $M^+$3. The stable and unstable manifolds are therefore singular sets of $M^+$4 for any finite $M^+$5 (Lopesino et al., 2017). Analogous results are established for nonlinear autonomous saddles via Moser normal form, and for linear and nonlinear nonautonomous hyperbolic systems via exact calculations or conjugacy arguments (Lopesino et al., 2017).

Forward and backward integrations isolate different invariant objects. In the standard hyperbolic picture, forward LDs emphasize stable manifolds and backward LDs emphasize unstable manifolds, while symmetric windows reveal both simultaneously (Lopesino et al., 2017). In time-dependent chemical dynamics this same principle is used operationally: minima of $M^+$6 identify $M^+$7, minima of $M^+$8 identify $M^+$9, and their intersection defines the anchor for a non-recrossing dividing surface (1705.00248).

A complementary 1-DoF viewpoint removes time altogether. For planar Hamiltonian systems with one degree of freedom, the geometric LD is defined as the Euclidean arc length of an energy level curve,

$M^-$0

In the pendulum, Duffing, and fish-tail separatrix problems, $M^-$1 attains a cusp-like maximum at the separatrix energy, and its derivative diverges as

$M^-$2

which isolates the geometric mechanism behind LD singularities without reference to a finite time window (Remi et al., 2021).

3. Variants across deterministic, stochastic, and dissipative dynamics

The map-based extension preserves the same structural idea. For two-dimensional area-preserving autonomous and nonautonomous maps, explicit formulae show that discrete LDs with $M^-$3 are non-differentiable across stable and unstable manifolds of hyperbolic saddles. In linear autonomous maps of the form

$M^-$4

the resulting discrete LD is proportional to $M^-$5, so directional derivatives become unbounded along the invariant axes. Rotated saddles and nonlinear maps exhibit the same phenomenon, with singular directions converging to the true invariant manifolds as the iteration horizon increases (Lopesino et al., 2017).

For stochastic differential equations, LDs are interpreted within the random dynamical systems framework. The stochastic extension

$M^-$6

uses sample-path increment sums $M^-$7, optionally ensemble averaged over Wiener realizations. In the noisy saddle, stochastic Duffing oscillator, and stochastic double gyre, the resulting fields reveal random fixed points, random stable and unstable manifolds, and transport barriers. The mechanism remains the same: for $M^-$8, non-differentiability appears when the dominant forward or backward increment terms vanish on the relevant random manifolds (Balibrea-Iniesta et al., 2017).

Dissipative systems require additional care because forward and backward dynamics are strongly asymmetric and backward integrations may blow up in finite time. In that setting LDs have been used to detect hyperbolic saddles with unequal expansion and contraction rates, limit cycles, slow manifolds, strange attractors, and transition ellipsoids in damped Hamiltonian systems. The reported practice includes asymmetric forward and backward times to balance stable and unstable signatures, and variable-time stopping when trajectories escape a bounded computational region (García-Garrido et al., 2021). In the Hopf normal form, LDs distinguish the case $M^-$9, where the origin is a stable focus, from $[t_0,t_0+\tau]$0, where a stable limit cycle of radius $[t_0,t_0+\tau]$1 appears as a ring in the LD field (García-Garrido et al., 2021). In slow–fast settings such as the van der Pol oscillator and the rotating-hoop model, LD ridges align with slow invariant manifolds (García-Garrido et al., 2021).

4. Computation, parameter choice, and numerical practice

The numerical workflow is usually minimal: choose a grid of initial conditions, integrate trajectories over a finite horizon, accumulate the chosen observable, and visualize the resulting scalar field. In continuous-time problems, the LD can be accumulated by augmenting the ODE with a scalar equation

$[t_0,t_0+\tau]$2

and integrating the extended system (Mancho et al., 2011). In discrete-time problems, the procedure is an orbit sum of step lengths or increment norms (Lopesino et al., 2017).

Parameter selection is problem dependent, but several quantitative prescriptions recur. For hyperbolic structures, subunit exponents are favored because they sharpen LD singularities. Reported choices include $[t_0,t_0+\tau]$3 in LiCN isomerization, $[t_0,t_0+\tau]$4 in the double pendulum and in several dissipative examples, and $[t_0,t_0+\tau]$5 in stochastic computations (Revuelta et al., 2021). Integration time is crucial. In the LiCN study, the practical criterion

$[t_0,t_0+\tau]$6

is used to resolve manifolds associated with a periodic orbit whose unstable Floquet exponent is $[t_0,t_0+\tau]$7. At $[t_0,t_0+\tau]$8, the reported values are $[t_0,t_0+\tau]$9 a.u. for the transition-state periodic orbit and $[t_0-\tau,t_0]$0 a.u. for the unstable saddle-node branch; production maps use $[t_0-\tau,t_0]$1 a.u. (Revuelta et al., 2021).

Nonautonomous and stochastic applications impose additional numerical requirements. The stochastic LD paper uses Euler–Maruyama integration, two-sided Brownian motion when backward integration is needed, and ensemble averaging when single-realization structures are noisy (Balibrea-Iniesta et al., 2017). In the double pendulum, forward LDs are computed with adaptive Dormand–Prince $[t_0-\tau,t_0]$2, tolerance $[t_0-\tau,t_0]$3, and $[t_0-\tau,t_0]$4 in dimensionless time units (López et al., 2024). Recent work on discrete maps replaces finite-difference estimates of LD second derivatives by differential algebra, obtaining machine-precision values of

$[t_0-\tau,t_0]$5

thereby improving threshold robustness in thin resonant webs and depleted meshes (Căliman et al., 2024).

5. Applications in fluids, chemistry, astrodynamics, and transition-state dynamics

In geophysical fluid dynamics, LDs were introduced as global descriptors for aperiodically time-dependent flows and applied to altimetric ocean data. In the Kuroshio region, the symmetric arc-length LD identified distinguished trajectories, hyperbolic manifolds, and non-hyperbolic eddy cores on the same scalar map. For $[t_0-\tau,t_0]$6 and $[t_0-\tau,t_0]$7 days, singular features aligned with independently computed stable and unstable manifolds, while the loss of smoothness inside an eddy core between $[t_0-\tau,t_0]$8 and $[t_0-\tau,t_0]$9 days was interpreted as a finite confinement-time indicator, with $p$0 days giving an upper bound on core trapping (Mendoza et al., 2010).

In chemical reaction dynamics, LDs have become a diagnostic for invariant manifolds associated with reaction barriers. For the LiCN$p$1LiNC isomerization, the rotationless molecule is modeled by a 2-DoF Hamiltonian with two minima at $p$2 and $p$3, energies $p$4 and $p$5, and a barrier top at $p$6. On a Poincaré section built along the minimum-energy path, singular features of the LD locate the invariant manifolds of the transition-state periodic orbit and of the unstable orbit born at the saddle-node bifurcation $p$7. The same framework supports a Morse-based equivalent 2-DoF potential and an adiabatic 1-DoF reduction that captures the principal separatrices and the emergence of a secondary dynamical barrier as radial excitation $p$8 increases (Revuelta et al., 2021).

In nonautonomous reaction dynamics, LDs are used constructively rather than just diagnostically. For a driven Gaussian barrier coupled to a bath mode, minima of forward and backward LDs on reactive-coordinate sections define $p$9 and $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$0; their intersection over all bath coordinates forms an anchor surface $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$1. The corresponding time-dependent dividing surface is empirically non-recrossing. In the reported 2-DoF model, the LD-based dividing surface yields zero recrossings for all 160,000 test trajectories, and for a thermal ensemble at $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$2 gives a rate $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$3, with Arrhenius parameters $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$4 and $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$5 (1705.00248).

Astrodynamics provides a different use case: finite-time boundedness rather than asymptotic invariant sets. In the perturbed planar bi-elliptic restricted four-body problem for the Didymos–Dimorphos system, forward LDs in the state $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$6 reveal phase-space structures that delimit bounded motion around Dimorphos over a finite horizon. Without solar radiation pressure, regions of bounded motion are visually identifiable; with SRP, the majority of those structures break down, leaving a large region of unstable motion with rare exceptions (Raffa et al., 2023).

6. LD-based chaos diagnostics and comparison with FTLE-, SALI-, and Poincaré-based methods

Although LDs were introduced as geometric visualization tools, a substantial literature now uses them as quantitative chaos indicators. One route computes finite differences of LD values for neighboring initial conditions. In low-dimensional conservative systems, the Difference of Neighboring orbits’ Lagrangian Descriptors and Ratio of Neighboring orbits’ Lagrangian Descriptors,

$$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$7

classify regular and chaotic motion with better than $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$8 agreement with SALI in both the Hénon–Heiles flow and the standard map, while using only short integration windows and forward LDs (Hillebrand et al., 2022). In the 4D symplectic standard map, the related indices $$M_p(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\sum_{i=1}^{n}\big|\dot{x}_i(t;\mathbf{x}_0)\big|^p\,dt,\qquad 0<p\le 1,$$9, $\sum_i |f_i(\mathbf{x}(t),t)|^p$0, and

$\sum_i |f_i(\mathbf{x}(t),t)|^p$1

achieve $\sum_i |f_i(\mathbf{x}(t),t)|^p$2 against SALI for a balanced choice $\sum_i |f_i(\mathbf{x}(t),t)|^p$3, $\sum_i |f_i(\mathbf{x}(t),t)|^p$4, $\sum_i |f_i(\mathbf{x}(t),t)|^p$5, with $\sum_i |f_i(\mathbf{x}(t),t)|^p$6 and $\sum_i |f_i(\mathbf{x}(t),t)|^p$7 generally outperforming $\sum_i |f_i(\mathbf{x}(t),t)|^p$8 (Zimper et al., 2023).

A simpler recent proposal uses only one neighboring orbit: $\sum_i |f_i(\mathbf{x}(t),t)|^p$9 The heuristic argument is that $0Jiménez-López et al., 20 Jun 2025).

Comparison with FTLEs is one of the oldest themes in the LD literature. In the periodically forced double gyre, forward and backward FTLE ridges approximate stable and unstable manifolds, but the paper reports that FTLE ridge locations can shift with $0KAM islands as smooth low-$0Getscher, 2021). In the ocean-flow study, LDs recover both hyperbolic and non-hyperbolic regions in a single field, while FTLEs require separate forward and backward computations and may exhibit “ghost” structures in transient settings (Mendoza et al., 2010). In deterministic-chaos applications, the second-derivative LD index $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$00 is presented as a non-variational alternative to FLI, MEGNO, and SALI, with about $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$01 agreement against SALI on 4D coupled standard maps (Daquin, 2023). The differential-algebra refinement of $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$02 goes further: finite-difference implementations can misclassify up to $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$03 of initial conditions in thin resonant webs, whereas differential algebra yields clean bimodal distributions and sensitivity comparable to MEGNO, FLI, and SALI across regimes (Căliman et al., 2024).

7. Conceptual issues, controversies, and emerging directions

A recurrent conceptual issue is objectivity. The theoretical framework explicitly argues that LDs are not objective under time-dependent Euclidean frame changes and that this is not necessarily a defect: in nonautonomous dynamics, different frames can carry genuinely different phase portraits, so a strictly objective scalar would be unable to distinguish them (Lopesino et al., 2017). This stands in contrast to diagnostics designed to be frame invariant. A related point is that some applied papers still describe LDs as heuristic and note that their precise mathematical connection to invariant manifolds remains debated outside the classes where rigorous results are available (Raffa et al., 2023). The present literature therefore supports two statements simultaneously: rigorous manifold-detection theorems exist for several hyperbolic settings, but general interpretive claims should still be tied to the dynamical context and to validation against independent structure calculations where possible (Lopesino et al., 2017).

Recent work extends LDs beyond classical deterministic mechanics. In the path-integral formulation, a quantum LD is defined as the fluctuation average of the classical LD around extremal trajectories. For the 1-DoF Hamiltonian saddle

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$04

stable and unstable manifolds $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$05 broaden into finite-width structures under quantum fluctuations. With a mode cutoff $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$06, time window $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$07, and Planck constant restored, the reported width estimate is

$$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$08

and Monte Carlo sampling agrees with this prediction within $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$09 across the explored range (Jiménez-López et al., 5 Apr 2026). A related Bohmian construction moves from a single wavefunction to a two-parameter preparation space of Gaussian wavepackets labeled by $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$10. For the inverted harmonic oscillator, the wavepacket-center dynamics are exactly classical, and the preparation-space stability matrix yields an $$M(\mathbf{x}_0,t_0,\tau)=\int_{t_0-\tau}^{t_0+\tau}\|\mathbf{v}(\mathbf{x}(t;\mathbf{x}_0),t)\|\,dt,$$11 bound on LD sensitivity, suggesting a geometric counterpart to the exponential growth familiar from semiclassical OTOC analyses (Wiggins, 21 Mar 2026).

These developments suggest that LDs have evolved from a finite-time flow-visualization tool into a broader geometric framework for transport, instability, and phase-space organization. The common thread remains unchanged: a scalar field built from trajectories can encode qualitative dynamical transitions as singularity patterns, whether the setting is a Kuroshio eddy, a chemical transition state, a stochastic transport barrier, a resonance web in a symplectic map, or a broadened quantum separatrix (Mancho et al., 2011).