A Critical Comparison of Lagrangian Methods for Coherent Structure Detection

Alireza Hadjighasem; Mohammad Farazmand; Daniel Blazevski; Gary Froyland; George Haller

Source-linked AI summary

A Critical Comparison of Lagrangian Methods for Coherent Structure Detection

Alireza Hadjighasem, Mohammad Farazmand, Daniel Blazevski, Gary Froyland, George Haller

arXiv:1704.05716v1Apr 19, 2017physics.flu-dynmath.DSnlin.CD

TL;DR

The paper addresses the difficulty of choosing among diverse Lagrangian coherence-detection methods and the limited comparison between heuristic and mathematically justified approaches. It surveys and tests twelve methods and diagnostic scalar fields on three temporally aperiodic benchmark flows, finding divergent predictions and coherence failures exposed by passive advection. The authors distill method strengths and weaknesses and recommend minimal self-consistency requirements.

Problem
The paper addresses the difficulty of choosing among many Lagrangian structure-detection methods and the limited contrast between heuristic tools and mathematically justified methods.
Method
The paper surveys and tests twelve Lagrangian coherence-detection approaches, including mathematical methods and diagnostic scalar fields, on three benchmark flows.
Results
The methods often produce different coherent-structure predictions, while passive advection exposes false positives and negatives, including for mathematically justified methods.
Takeaways & Limitations
Coherence detections should be checked by material advection against the exact coherence principle underlying each method.
Takeaways & Limitations
The recommendations are especially consequential for now-casting, short-term forecasting, flow control, and real-time decisions in sensitive situations when a method fails any stated requirement.

Abstract

from arXiv · show

We review and test twelve different approaches to the detection of finite-time coherent material structures in two-dimensional, temporally aperiodic flows. We consider both mathematical methods and diagnostic scalar fields, comparing their performance on three benchmark examples: the quasiperiodically forced Bickley jet, a two-dimensional turbulence simulation, and an observational wind velocity field from Jupiter's atmosphere. A close inspection of the results reveals that the various methods often produce very different predictions for coherent structures, once they are evaluated beyond heuristic visual assessment. As we find by passive advection of the coherent set candidates, false positives and negatives can be produced even by some of the mathematically justified methods due to the ineffectiveness of their underlying coherence principles in certain flow configurations. We summarize the inferred strengths and weaknesses of each method, and make general recommendations for minimal self-consistency requirements that any Lagrangian coherence detection technique should satisfy.

I. INTRODUCTION

The paper systematically compares twelve Lagrangian coherent-structure detection methods across three increasingly difficult two-dimensional unsteady flows. It contrasts diagnostic scalar fields with mathematically defined approaches and evaluates them against objectivity, passive-scalar baselines, and material trajectory behavior.

The study addresses the difficulty of choosing among numerous heuristic and mathematical Lagrangian detection methods by comparing them on challenging benchmark flows.
Twelve methods are classified as diagnostic scalar-field approaches or analytic methods based on mathematically formulated coherence principles.
The benchmarks progress from an analytically known quasiperiodic Bickley jet to discretely sampled two-dimensional turbulence and an observational Jupiter wind field.
The comparison covers elliptic, hyperbolic, and parabolic material structures rather than focusing on one structure type.
Diagnostic fields generally lack a strict coherent-structure definition, so their features must at least outperform randomly chosen passively advected scalar fields.
Objectivity requires identifying the same coherent trajectories under Euclidean observer changes, especially in unsteady and rotating geophysical flows.

A. Finite-time Lyapunov exponent (FTLE)

FTLE and related deformation-based diagnostics use trajectory separation or deformation information to identify hyperbolic and parabolic structures, but their interpretations are flow- and parameter-dependent. The section also describes mesochronic analysis as limited by return-trajectory assumptions and frame dependence.

FTLE ridges are proposed as markers of strongest repelling or attracting hyperbolic LCSs, and the FTLE field is objective.
An FTLE ridge may represent high shear rather than a repelling material surface, so visual ridge identification is not definitive.
FTLE trenches can approximate parabolic jet cores but can also fail to coincide with the jet.
FTLE is not designed to detect elliptic vortex boundaries because low values near vortices do not generally create sharp boundaries.
FSLE emphasizes separation beyond a selected scale and avoids prescribing an end time, but its ridges have no general correspondence with FTLE ridges.
Mesochronic classification is difficult to interpret in general unsteady flows because it relies on trajectories returning exactly to their starting points and is not objective.

D. Trajectory length

Trajectory length and trajectory-complexity diagnostics are computationally simple ways to partition flows by apparent dynamical differences. However, trajectory length lacks an established mathematical connection to coherence and is frame-dependent, while complexity has an objective histogram-based construction.

Trajectory-length boundaries are identified from abrupt spatial variations in the arclength of trajectories.
The trajectory-length function is among the quickest diagnostics to compute and can use float data without a velocity field or neighboring trajectories.
Trajectory length has no established mathematical connection to material coherence and has documented counterexamples.
Trajectory length is neither objective nor Galilean invariant because a co-moving observer can assign zero arclength to a selected trajectory.
Trajectory complexity is quantified through the ergodicity defect, an L2 deviation of trajectory-point histograms from a constant histogram.
Large gradients in the complexity field are expected to mark boundaries between qualitatively different flow regions, and the construction is objective under rotations and translations.

F. Shape coherence

Shape coherence seeks closed material curves whose geometry remains nearly congruent under advection, using an angle between forward- and backward-time strain directions. The comparison tests this heuristic in unsteady flows, while transfer-operator methods formulate coherence through probabilistic mass retention.

F. Shape coherence: Shape-coherent boundaries are sought as closed material lines whose curvature distributions remain close to those of their advected images.
F. Shape coherence: The shape-coherence diagnostic minimizes the angle between dominant eigenvectors of forward- and backward-time Cauchy–Green tensors.
F. Shape coherence: Small splitting angles along closed boundaries are supported by steady linear-flow examples but remain a heuristic assertion for general unsteady flows.
F. Shape coherence: Reliable closed-level-curve detection is numerically challenging, so the study identifies small-angle regions and advects their seeded initial conditions for comparison.
Probabilistic transfer operator method: The transfer-operator and dynamic-Laplacian approaches are objective by construction, and the latter arises as a zero-diffusion limit of the former.
Probabilistic transfer operator method: Transfer-operator methods identify finite-time coherent sets as regions that maximize agreement or minimize dispersion of transported mass.
Probabilistic transfer operator method: The probabilistic method extracts initial and final coherent sets from level sets of singular vectors of a normalized transfer operator.

2. Dynamic Laplace operator method

The dynamic Laplace approach seeks disconnecting surfaces whose advected boundaries remain short relative to the volumes they separate, thereby minimizing boundary filamentation. It is a zero-diffusion-limit method whose numerical results are similar to probabilistic transfer operators and whose constructions are objective.

2. Dynamic Laplace operator method: The method seeks surfaces whose advected images remain short relative to the volumes of the disconnected regions.This minimizes filamentation of the boundary under nonlinear evolution.
2. Dynamic Laplace operator method: The dynamic Laplace operator is used to solve the variational problem over smooth disconnecting surfaces.
2. Dynamic Laplace operator method: Its eigenvectors reveal geometric information from which multiple coherent sets can be extracted, analogously to transfer-operator methods.
2. Dynamic Laplace operator method: As a zero-diffusion limit, the dynamic Laplace and probabilistic transfer-operator methods produce very similar numerical results and are objective by construction.
2. Dynamic Laplace operator method: Hierarchical transfer operators identify multiple coherent pairs through iterative refinement restricted to previously identified coherent sets.This repeated-bisection strategy differs from extracting multiple sets through multiple singular vectors.

C. Fuzzy cluster analysis of trajectories

Trajectory-clustering methods identify coherent regions by grouping particles with similar dynamical histories. Fuzzy C-means permits overlapping cluster membership, whereas spectral clustering approximates normalized graph cuts through Laplacian eigenvectors.

C. Fuzzy cluster analysis of trajectories: Fuzzy C-means clusters trajectories represented by concatenated particle positions over discrete times, using a dynamic distance between trajectory pairs.
C. Fuzzy cluster analysis of trajectories: Membership values quantify how strongly each trajectory point belongs to each cluster, with the fuzziness parameter controlling overlap between clusters.
C. Fuzzy cluster analysis of trajectories: At m = 1, fuzzy memberships converge to 0 or 1, producing non-overlapping clusters similar to K-means.
C. Fuzzy cluster analysis of trajectories: After optimization, each trajectory is assigned to the cluster with its maximum membership, while trajectories below the 0.9 threshold may be treated as non-coherent.
C. Fuzzy cluster analysis of trajectories: Spectral clustering builds a particle similarity graph in which coherent structures have strong within-cluster similarity and weak similarity to complementary particles.
C. Fuzzy cluster analysis of trajectories: The normalized-cut problem is NP-complete, but approximate solutions use generalized graph-Laplacian eigenvectors near zero to extract coherent structures.

E. Stretching-based coherence: Geodesic theory of LCS

Geodesic LCS theory defines coherent material curves through global variational principles based on finite-time stretching and shear. It distinguishes hyperbolic, parabolic, and elliptic structures, with elliptic LCSs forming non-filamenting vortex boundaries.

E. Stretching-based coherence: Geodesic theory of LCS: Geodesic LCS theory uses global variational principles for material surfaces that organize coherent, time-evolving tracer patterns.
E. Stretching-based coherence: Geodesic theory of LCS: Hyperbolic LCSs repel or attract neighboring material elements at locally highest finite-time rates and generalize stable and unstable manifolds.
E. Stretching-based coherence: Geodesic theory of LCS: Parabolic LCSs minimize Lagrangian shear and are constructed from alternating shrink–stretch line segments connecting tensorline singularities.They are described as more robust under perturbations than hyperbolic LCSs.
E. Stretching-based coherence: Geodesic theory of LCS: Elliptic LCSs are closed null-geodesics associated with stationary material-line-averaged strain and form nested families parameterized by λ.
E. Stretching-based coherence: Geodesic theory of LCS: Any subset of an elliptic LCS limit cycle stretches exactly by λ over the interval [t0, t1], so elliptic LCSs exhibit no filamentation when advected.
E. Stretching-based coherence: Geodesic theory of LCS: The outermost member of a nested elliptic limit-cycle family serves as a Lagrangian vortex boundary.

F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD)

LAVD identifies rotationally coherent vortex boundaries from objective, trajectory-averaged vorticity deviations. The approach addresses the non-additivity of classic polar rotation by using dynamically consistent rotation concepts, while defining boundaries as outermost nested tubular level surfaces.

F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD): Rotationally coherent LCSs are tubular material surfaces whose elements undergo identical mean material rotation over a finite interval.
F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD): Classic polar rotation is not additive over subintervals, so PRA does not match experimentally observed mean material rotation of finite tracers.
F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD): Dynamic polar decomposition resolves this inconsistency by obtaining dynamic rotation and stretch tensors as solutions of linear differential equations.
F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD): The dynamically consistent intrinsic angle provides an objective extension of PRA in both two- and three-dimensional flows.
F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD): The LAVD is the trajectory-averaged normed deviation of vorticity from its spatial mean and equals twice the intrinsic rotation angle.
F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD): LAVD vortex boundaries are outermost nested tubular level surfaces satisfying minimum-length and convexity-deficiency thresholds.
F. Rotational coherence from the Lagrangian-Averaged Vorticity Deviation (LAVD): LAVD boundaries may develop tangential filamentation, but such filaments rotate at the vortex body's average rate without global transverse breakaway.

V. METHOD COMPARISONS ON THREE EXAMPLES

The paper compares twelve detection methods across three increasingly difficult flow examples, using matched computational effort and passive advection to assess predicted material coherence.

Bickley jet: The Bickley jet is analytically defined, quasiperiodic, recurrent, and challenging because its non-near-integrable dynamics do not guarantee persistence of invariant structures.Its recurrent but nonperiodic time dependence also prevents using the classic Poincaré map to visualize coherence.
Two-dimensional turbulence: The turbulence example uses a finite-time, fully aperiodic and non-recurrent velocity data set containing coherent regions that move and merge.This makes it representative of several practical difficulties in identifying coherence from flow data.
Jupiter wind field: The Jupiter example contains one vortical structure, but its short data record relative to the Great Red Spot’s rotation period adds a severe time-scale limitation.The velocity field is reconstructed from enhanced video footage of Jupiter’s atmosphere.
Comparison design: The comparison equalizes the total number of advected trajectories across methods and examples while also reporting each method’s required parameters.Table I measures particle requirements for constructing a Lagrangian field, whereas Table II compares user-parameter requirements.
Comparison design: Parameter choices were selected to favor each method while remaining robust to small variations.The study therefore evaluates methods under experienced, intentionally favorable parameter settings rather than arbitrary defaults.
Evaluation criterion: Passive advection establishes coherence in a region, with missed structures counted as false negatives and predicted but incoherent structures as false positives.Geometric differences between candidate boundaries are not treated as failures when sustained material coherence is confirmed.

A. Quasi-periodically perturbed Bickley jet

In the quasiperiodic Bickley jet, methods differ substantially in which vortices and jet structures they identify and how well candidate boundaries remain coherent under advection. The comparison shows that no single method captures every relevant structure: spectral clustering detects all vortices, while transfer operator and LAVD methods provide sharper but more selective results.

Flow model: The Bickley jet combines a steady background flow with a time-dependent perturbation represented by three Rossby waves.The characteristic scales are U = 62.66 ms−1 and L = 1770 km, with the analysis spanning t ∈[0, 11] day.
Diagnostic comparison: Most diagnostic scalar fields indicate six vortices, but mesochronic, shape coherence, transfer operator, and geodesic methods miss some or all of them.The paper therefore examines these methods as exceptions to the broader vortex indication provided by diagnostic fields.
Diagnostic comparison: Shape coherence captures none of the coherent Bickley-jet vortices because its spiral candidate regions cannot contain closed vortex boundaries.This failure occurs even with a weakened version of the underlying criterion.
Analytic methods: The transfer operator sharply identifies the jet core but misses the vortices at its dominant level, while hierarchical partitioning captures three vortices at hierarchy level n = 5.Higher singular vectors reveal additional vortices, although extracting their boundaries requires thresholding choices and selecting how many vectors to inspect.
Analytic methods: Spectral clustering consistently detects all vortices and improves their size estimates, but its dominant eigenvectors do not reveal the meandering jet because shear separates jet particles.The seventh generalized eigenvector does reveal the jet, but the spectrum provides no a priori indication that it should be examined.
Structure-specific performance: Geodesic LCS identifies three of six vortical regions, whereas LAVD captures all vortices accurately but cannot detect the meandering jet or hyperbolic and parabolic LCSs.Only the FTLE, geodesic, and transfer operator methods provide sharp jet-core boundaries among the methods discussed for jet identification.

B. Two-dimensional turbulence

In two-dimensional turbulence, advection tests distinguish physically coherent vortices from coincidental or method-specific detections. Several methods identify vortex structures, but transfer-operator approaches and shape coherence exhibit important limitations in this flow.

Method comparison: Most methods indicate several vortex-type structures, whereas shape coherence and transfer-operator methods do not directly recover the turbulence vortices.Transfer-operator coherent sets remain bounded under advection but are unrelated to the vortices generally regarded as coherent structures.
Transfer operators: Hierarchical transfer-operator results include vortex-like structures among many additional patches and show no overall convergence across box-covering resolutions.The fixed termination threshold is affected by resolution-dependent reference measures and initial numerical diffusion.
Shape coherence: The shape-coherence candidate regions form spirals, so no closed vortex boundaries satisfy its coherence requirement in this temporally aperiodic flow.The underlying principle is argued to suit flows with the same behavior forward and backward in time, unlike this example.
Clustering methods: Spectral clustering and fuzzy clustering recover coherent vortices, while fuzzy clustering gives the sharpest, least conservative assessment for the reported larger vortices.Fuzzy-clustering boundaries show only tangential filamentation for three clearly coherent vortices.
Vortex detection: Geodesic vortex boundaries remain coherent under advection, while the geodesic method is too conservative to detect smaller vortices.The geodesic method identifies only three of six vortical regions as coherent, whereas other methods identify six arguably coherent material vortices.
Vortex detection: LAVD vortex boundaries preserve rotational coherence and display only tangential filamentation at the final time.The method’s derivation guarantees this tangential-filamentation behavior for the reported boundaries.
Diagnostic fields: Trajectory-length contours near vortex cores remain coherent, but they substantially underestimate the upper vortex’s size.The mesochronic vortex criterion finds no vortex-type structures in the selected regions because saddle-type critical points are present.

C. Wind field from Jupiter’s atmosphere

In Jupiter’s observational wind field, several boundary-based methods identify the Great Red Spot and geodesic LCSs identify jet cores, while other methods provide weaker or less definitive results. Advection tests expose stretching in some fuzzy-clustering candidates and support close agreement among three vortex-boundary methods.

Data and setup: The Jupiter comparison uses velocity vectors reconstructed from Cassini video footage covering 24 Jovian days.The velocity field was obtained with Advection Corrected Correlation Image Velocimetry.
Great Red Spot: Several methods signal a localized vortex-type coherent structure corresponding to Jupiter’s Great Red Spot, but transfer operator, shape coherence, fuzzy clustering, and mesochronic methods are exceptions.The fifth transfer-operator singular vector indicates the GRS, but the singular-value spectrum does not distinguish it a priori.
Great Red Spot: Spectral clustering, geodesic LCS detection, and LAVD give very close Great Red Spot boundaries, indicating a fairly well-defined material vortex.The reported GRS core has negligible material filamentation.
Jet identification: Most diagnostic methods suggest two jets near the GRS, but geodesic LCS, transfer-operator, and hierarchical transfer-operator methods provide clearer jet cores or boundaries.The geodesic signals coincide with visually observed zonal jet cores in Jupiter’s atmosphere.
Advection assessment: Fuzzy-clustering candidates can stretch substantially under advection because some are transverse to Jupiter’s shear-jet transport barriers.Together with the Bickley-jet example, this indicates difficulty identifying both vortical and jet-type structures.
Assessment: The assessment summarizes inferred strengths and weaknesses for all twelve compared Lagrangian methods.The paper’s broader comparison distinguishes diagnostic methods from analytic methods and evaluates their outputs across benchmark flows.
FTLE assessment: FTLE is described as simple and objective, with ridges capturing hyperbolic LCSs under additional mathematical conditions and trenches approximating jet cores.Its listed weaknesses include unreliable elliptic-LCS detection and false positives where high stretching coincides with high shear.
FSLE assessment: FSLE is described as simple, objective, and scale-focused without requiring a prescribed integration time or flow-map differentiation.Its weaknesses include limited correspondence to hyperbolic LCSs, unreliable elliptic and parabolic detection, mixed time-scale structures, and jump discontinuities.

3. Mesochronic analysis

The reviewed methods differ in computational demands, objectivity, mathematical grounding, and which coherent structures they can detect. Their limitations include missed structures, ambiguous boundaries, parameter sensitivity, and outputs that can become unphysical or difficult to interpret.

Trajectory length method: The trajectory length method is simple and often reveals features consistent with other methods, but it is nonobjective and lacks a clear mathematical meaning for non-periodic trajectories.It also lacks reliable detection of hyperbolic and parabolic LCSs, while its elliptic and hyperbolic classifications can conflict with classic stability notions.
Trajectory complexity method: The trajectory complexity method is simple, objective, physically intuitive, and topologically consistent for vortical features without differentiating the flow map.However, it provides no clear structure boundaries and has no clear mathematical connection to coherence.
Shape coherence method: The shape coherence method is objective and intuitive for steady and time-periodic flows, but assumes identical stretching history forward and backward in time.That assumption causes it to miss coherent structures in time-dependent data and leaves no clear recipe for extracting closed boundaries.
Transfer-operator method: The transfer-operator method has an appealing mathematical foundation, rigorous material-coherence estimates, broad dimensional applicability, and sharp boundaries when a region splits into two coherent sets.Its first nontrivial singular vector always produces a two-set partition, while additional structures require an a priori unclear number of further singular vectors and substantial computational cost.
Hierarchical transfer-operator method: The hierarchical transfer-operator method is objective, incrementally refinable, and not restricted to two coherent sets, but its hierarchy lacks overall convergence.The number of subsets can increase endlessly in homogeneous regions, producing unphysical output beyond some granularity.

9. Fuzzy clustering of trajectories

Fuzzy clustering offers an objective, simple, low-cost route to trajectory-based coherent-set detection without flow-map differentiation. Its outputs can find open low-dispersion regions, but shape fidelity, coherence under advection, parameter robustness, and detected structure count remain important constraints.

Strengths and weaknesses: Fuzzy clustering has a simple, appealing theoretical foundation, is objective, and requires no differentiation of the flow map with respect to initial conditions.The number of coherent structures is an input parameter, and detected structures may have convoluted shapes unlike known coherent boundaries.
Strengths and weaknesses: Most detected fuzzy-clustering structures have convoluted shapes, and some lose coherence further through stretching under advection.The method also cannot detect hyperbolic LCSs and has difficulty detecting elliptic and parabolic LCSs simultaneously.
Strengths and weaknesses: The method requires robustness checks across extraction parameters, while the number of coherent structures to locate must be supplied as an input.These constraints make parameter selection part of interpreting the resulting coherent-set candidates.
Strengths and weaknesses: Fuzzy clustering is objective, simple, and requires few user inputs; it consistently finds open, low-dispersion regions beyond elliptic LCSs.Unlike methods whose number of structures is prescribed, its number of coherent structures is reported as an output.
Strengths and weaknesses: The spectral method requires a well-defined spectral gap and can be computationally expensive for many trajectories, while its gap depends on the sparsification radius.It also cannot detect hyperbolic and parabolic LCSs and may produce structures whose robustness varies with extraction time.
Strengths and weaknesses: Variational methods can automate objective detection of hyperbolic, elliptic, and parabolic LCSs, with exact principles guaranteeing no filamentation for elliptic LCSs under advection.They are computationally involved and may miss elliptic LCSs with non-uniformly stretching boundaries; one implementation does not extend to three dimensions.
Strengths and weaknesses: LAVD-based detection is automated, simple, objective, inexpensive, and mathematically related to material rotation, but it cannot detect hyperbolic or parabolic LCSs.It relies on velocity-field derivatives and assumes a sufficiently large computational domain for the spatial mean vorticity to be representative.
General assessment: Random scalar fields perform favorably compared with heuristic Lagrangian diagnostics, but their observed features are consequences of material coherence rather than its root cause.Therefore, visual features in a diagnostic field do not validate the intuitive arguments used to construct that diagnostic.