# From Large Deviations to Semidistances of Transport and Mixing: Coherence Analysis for Finite Lagrangian Data

- 720 Downloads
- 1 Citations

## Abstract

One way to analyze complicated non-autonomous flows is through trying to understand their transport behavior. In a quantitative, set-oriented approach to transport and mixing, finite time coherent sets play an important role. These are time-parametrized families of sets with unlikely transport to and from their surroundings under small or vanishing random perturbations of the dynamics. Here we propose, as a measure of transport and mixing for purely advective (i.e., deterministic) flows, (semi)distances that arise under vanishing perturbations in the sense of large deviations. Analogously, for given finite Lagrangian trajectory data we derive a discrete-time-and-space semidistance that comes from the “best” approximation of the randomly perturbed process conditioned on this limited information of the deterministic flow. It can be computed as shortest path in a graph with time-dependent weights. Furthermore, we argue that coherent sets are regions of maximal farness in terms of transport and mixing, and hence they occur as extremal regions on a spanning structure of the state space under this semidistance—in fact, under any distance measure arising from the physical notion of transport. Based on this notion, we develop a tool to analyze the state space (or the finite trajectory data at hand) and identify coherent regions. We validate our approach on idealized prototypical examples and well-studied standard cases.

## Keywords

Large deviation Coherent sets Lagrangian data Mixing Mixing distance Transport distance## Mathematics Subject Classification

Primary 37N10 60F10 Secondary 60J22 76F25 05C12 37M10## 1 Introduction

*Transport in Dynamical Systems*

Instrumental to understanding the essential behavior of complicated non-autonomous flows is to grasp how transport is happening in them. This leads on a qualitative level to objects that prohibit transport, commonly named *transport barriers*; often originating from the geometric picture for autonomous systems and that trajectories are unable to cross co-dimension 1 invariant manifolds (Haller 2000, 2001; Haller and Beron-Vera 2012, 2013). For periodically forced systems, invariant manifolds enclose regions called “lobes” that get transported across these periodically varying manifolds (MacKay et al. 1984; Rom-Kedar and Wiggins 1990).

On a quantitative level, one searches for surfaces of small flux (Balasuriya et al. 2014; Karrasch 2016; Froyland and Koltai 2017), so-called partial barriers (Wigner 1937; Meiss 1992). Instead of characterizing regions that do not mix with one another via enclosing them by boundaries of low flux, there are approaches that aim to describe these sets directly. Such *set-oriented* concepts are strongly interwoven with the theory of transfer operators (Perron–Frobenius and Koopman operators) and comprise *almost-invariant sets* (Dellnitz and Junge 1999), *ergodic partitions* (Mezić and Wiggins 1999) in autonomous, and *coherent sets* (Froyland et al. 2010; Froyland 2013) in the non-autonomous cases.

Distinctive attention has been given to coherent sets, which are a (possibly time-dependent) family of sets having little or no exchange with their surrounding in terms of transport and are robust to small diffusion over a finite time of consideration (Froyland 2013, 2015). Natural examples include moving vortices in atmospheric (Rypina et al. 2007; Froyland et al. 2010), oceanographic (Treguier et al. 2003; Dellnitz et al. 2009; Froyland et al. 2015), and plasma flows (Padberg et al. 2007). In such applications, one would like to be able to find coherent sets even in the cases when a dynamical model that can be evaluated arbitrarily often is not available, only a finite set of *Lagrangian trajectory data* (passive tracers moving with flow with positions sampled at discrete time instances). This problem has received lot of attention in recent years, and a diverse collection of tools has been developed to tackle it (Budišić and Mezić 2012; Allshouse and Thiffeault 2012; Ser-Giacomi et al. 2015; Froyland and Padberg-Gehle 2015; Allshouse and Peacock 2015; Williams et al. 2015; Hadjighasem et al. 2016; Banisch and Koltai 2017; Schlueter-Kuck and Dabiri 2017; Padberg-Gehle and Schneide 2017; Rypina and Pratt 2017; Fabregat et al. 2016; Froyland and Junge 2017).

While other current methods aim at collecting trajectories into coherent sets, in Banisch and Koltai (2017) it has been proposed to go one step further and analyze the connectivity structure of the state space under transport and mixing with “transport coordinates” and the “skeleton of transport”. Very similar observations have been made earlier in Budišić and Mezić (2012) in the infinite-time limit for periodically forced systems. While coherent sets (and transport barriers) aim at *partitioning* the state space, the skeleton is aiming at *“spanning”* the state space with respect to transport. In this respect, coherent sets can be associated with distinct “extremal regions” of the skeleton. Here we will only use this idea of extremality, more precisely that coherent sets are “maximally far” from one another, as measured by transport. To this end, we will need to measure “farness” of dynamical trajectories.

Several *dynamical distance measures* have been put forward already to measure the “distance” or “dissimilarity” of trajectories or initial states in dynamical systems (Mezić and Banaszuk 2004; Budišić and Mezić 2012; Froyland and Padberg-Gehle 2015; Hadjighasem et al. 2016; Fabregat et al. 2016; Karrasch and Keller 2016). The majority of them are shown to serve their purpose well in revealing coherent structures efficiently and reliably. However, they are either heuristic in the sense that they are not derived from the physical notion of transport and mixing, or no discretizations to finite scattered trajectory data have been developed.

The purpose of this paper is thus twofold. On the one hand, we develop a distance measure (a semidistance) between trajectories that is derived from the physical notion of transport and mixing subject to diffusion of vanishing strength, and we also derive a discretized distance measure for finite (also possibly sparse and incomplete) Lagrangian data that is consistent with its continuous counterpart in the limit of infinite data. On the other hand, we construct a tool to analyze with such distances the structure of the state space under transport, especially to find coherent sets. This tool makes use of the idea that coherent sets are some sort of extremal regions on a spanning structure with respect to transport, although in this work we will not investigate this “skeleton” in its entirety.

*Finite Time Coherent Sets*

*T*] for some \(T>0\). Throughout the paper, we will assume that \(v:[0,T]\times X\rightarrow \mathbb {R}^d\) is a continuous velocity field tangential at the boundary, such that the flow of (1), denoted by \(\phi _{s,t}[\cdot ]\), \(0\le s, t\le T\), is a diffeomorphism on appropriate subsets of

*X*. For \(t<s\) we flow backward in time: \(\phi _{s,t}= \phi _{t,s}^{-1}\).

*robust under noise*; without such a requirement, any non-intersecting characteristic of a singleton could be considered a coherent set. To this end, one typically perturbs the ODE (1) by a random noise (Denner et al. 2016; Froyland and Koltai 2017; Karrasch and Keller 2016), leading to the Itô stochastic differential equation (SDE)

^{1}

*finite time coherent pairs*(Froyland et al. 2010; Froyland 2013; Koltai et al. 2016), two sets \(A,B\subset X\) are coherent for times 0 and

*T*if most mass from set

*A*is likely to end up in set

*B*, and most mass ending up in set

*B*is likely to originate from set

*A*, that is,

*large deviations*as we explain below.

^{2}However, it turns out that the forward and backward conditions (3) are essentially equivalent in the large-deviation regime, and even worse, the large-deviation limits of (3) hardly give any quantitative information about how coherent two sets might be, as discussed in Appendix A. To conclude, the large deviations of conditions (3) do not yield sensible conditions for coherence.

*Large-Deviation-Based Semidistances*

In the current paper, we take a different approach. We study semidistances that quantify how unlikely it is for mass to flow from one point to another. These are semidistances in the sense that they satisfy all properties of a metric except for the triangle inequality. In the first part of this paper, in Sects. 2 and 3, we show how such semidistances can arise naturally from probabilistic arguments via large-deviation principles, as we explain below. In the second part, in Sects. 4 and 5, we discuss how (general) semidistances can be used to analyze coherent sets, and we apply the concepts of this paper to a number of examples.

*large deviations*of two probabilities. The first one is related to the probability that the endpoint \(\varvec{}{x}_T^{\scriptscriptstyle {(\varepsilon )}}\) of the random path is \(\phi _{0,T}[y]\), given that it starts in \(\varvec{}{x}_0^{\scriptscriptstyle {(\varepsilon )}}=x\), for any two initial positions \(x,y\in X\). As \(\varepsilon \rightarrow 0\), the process can no longer deviate from the deterministic flow of (1), and hence this probability will converge to 0 whenever \(x\ne y\). In fact, it converges exponentially fast (Freidlin and Wentzell 1998), i.e.,

*large-deviation principles*, and \(\mu _T(x\rightarrow y)\) is the

*large-deviation rate*. The less probable it is to reach one point from another, the larger the rate between them. The first semidistance is then obtained via symmetrization:

*cross*semidistance, since it arises from mass flowing from

*x*to \(\phi _{0,T}[y]\) and mass flowing from

*y*to \(\phi _{0,T}[x]\) simultaneously and independently, see Fig. 1.

*x*and

*y*, respectively, to meet at or before time

*T*(Fig. 2):

*meeting*semidistance is given by

By this procedure, we find two semidistances \(\mu _T^\mathrm {cross}\) and \(\mu _T^\mathrm {meet}\) that can be used as a measure of “farness” of points *x*, *y*, which will be low for points in the same coherent set, and high otherwise. Since both arise from large-deviation principles, they have a nice additional interpretation as a probabilistic cost or free energy that needs to be paid in order to deviate from the expected flows; such interpretation is common in statistical physics, see for example Onsager and Machlup (1953).

Nevertheless, we will see that in order to calculate these costs explicitly, the velocity field *v* needs to be known. As discussed above, this is in practice seldom the case; mostly one can only assume to have discrete-time snapshots of the positions of a limited number of floaters. With this in mind, we derive similar cost functions as above, that are based on such a finite data set *only*. This will be the content of Sect. 3. First, the *dynamics* is discretized in time and space by conditioning a usual time-stepping method for the SDE (2) on the event that the random continuous trajectories are to be found in the set of known floater positions at the \(K\in \mathbb {N}\) given time instances. As above, we then derive two large-deviation semidistances \(\nu _K^\mathrm {cross}(x,y)\) and \(\nu _K^\mathrm {meet}(x,y)\) that have a clear probabilistic interpretation, that can be used to characterize coherent sets, and that are based on the finite data set rather than on the explicit velocity field. In fact, we will show that these discrete-space–time semidistances are really specific discretizations of the continuous-space–time semidistances \(\mu _T^\mathrm {cross}\) and \(\mu _T^\mathrm {meet}\). As shown in Sect. 3.4, they can be computed as shortest path lengths in a time-dependent weighted graph. We give an algorithm to compute these shortest paths in Appendix B.

Let us stress that these semidistances are defined for deterministic dynamical systems. The random perturbation that is factored out by the large-deviation principle is merely acting as a catalyst to help quantify how strongly distinct trajectories mix—or, we should rather say how *poorly*, as the transport from one trajectory to another is inversely proportional to their semidistance.

*Coherence Analysis with Semidistances*

In Sect. 4, we describe how in general a semidistance on finite Lagrangian data can be used to analyze coherence. Key to our method is the notion of *cornerstone*: a point that is furthest away from all other points. Cornerstones are though of as “endpoints” of a spanning structure, and ideally each cornerstone is in some sense the center of a coherent set. As a next step, trajectories can be clustered around cornerstones to yield coherent sets. Of course, this approach is very close to the *k-means-* and *fuzzy c-means clustering* of trajectories with respect to dynamical distances in Hadjighasem et al. (2016); Froyland and Padberg-Gehle (2015), with the important difference that the centers are not chosen by the heuristics of these clustering approaches, but with regard to the properties of coherent sets in the light of transport and mixing.

To exemplify the usefulness of the theory put forth in this paper, in Sects. 4 and5 we test our approach on a number of standard test cases. Finally, Sect. 6 discusses possible combinations of this work with other concepts.

## 2 Large-Deviation Semidistances in Continuous Time and Space

In this section, we study large deviations of the forms (4) and (6). In large-deviation theory, it is often easier to first study large deviations in a larger space. In our setting, we first study the large deviations of paths in Sect. 2.1 before transforming to the large deviations of the endpoints in Sect. 2.2. We end with a discussion of the resulting semidistances \(\mu _T^\mathrm {cross},\mu _T^\mathrm {meet}\) in Sect. 2.3.

### 2.1 Large Deviations of Paths

We denote paths by \(w_{\scriptscriptstyle {(\cdot )}}\) to distinguish them from points *w*. Let \(\mathbb {P}\) be the Wiener measure, i.e., the probability that a Brownian path lies in a set \(U\subset C(0,T;\mathbb {R}^d)\) is \(\mathbb {P}[\varvec{}{w}_{(\cdot )}\in U]\). Recall that there does not exist a canonical probability measure on the space of paths, and so the Wiener measure cannot be identified with a meaningful density. This means that one always needs to consider sets rather than particular realizations of the Brownian path. Nevertheless, large-deviation rates are always local, in the sense that they depend on one realization only (the most likely one in the set *U* under consideration). This motivates writing \(\varvec{}{w}_{\scriptscriptstyle {(\cdot )}}\asymp f_{\scriptscriptstyle {(\cdot )}}\) if \(\varvec{}{w}_{\scriptscriptstyle {(\cdot )}}\) lies in an infinitesimal neighborhood *U* of the path \(f_{\scriptscriptstyle {(\cdot )}}\). We will make this more precise below.

The large deviations for the SDE (2) are a standard result by Freidlin and Wentzell (1998). This result can be derived via a combination of *Schilder’s theorem* and a *contraction principle* as we now explain.

^{3}

### 2.2 Large Deviations of Endpoints

*x*and ending in that given endpoint \(\phi _{0,T}[y]\), that is:

The sum (5) then defines the cross-semidistance \(\mu _T^\mathrm {cross}(x,y)\) and has a natural interpretation in terms of large deviations: As mentioned in the Introduction, it arises from two independent and simultaneous copies \(\varvec{}{x}^{\scriptscriptstyle {(\varepsilon )}}_t,\varvec{}{y}^{\scriptscriptstyle {(\varepsilon )}}_t\). By independence, the probability that \((\varvec{}{x}^{\scriptscriptstyle {(\varepsilon )}}_T,\varvec{}{y}^{\scriptscriptstyle {(\varepsilon )}}_T)\asymp (\phi _{0,T}[y],\phi _{0,T}[x])\) given \((\varvec{}{x}^{\scriptscriptstyle {(\varepsilon )}}_0,\varvec{}{y}^{\scriptscriptstyle {(\varepsilon )}}_0)=(x,y)\) is a product of one-way probabilities, yielding the sum of two one-way rates in the large deviations, see Fig. 1.

*x*and

*y*, respectively. We consider the probability that both trajectories end in a given point, say \(\phi _{0,T}[z]\) for some \(z\in \mathbb {R}^d\), see Fig. 2. Assuming independence of the two trajectories, we immediately get

*T*with zero cost; the time

*T*thus acts as a maximum time at which the paths should meet.

### 2.3 The Semidistances

*x*,

*y*,

- (i)
\(\mu _T^{\mathrm {cross}}(x,y)\ge 0\) and \(\mu _T^{\mathrm {meet}}(x,y)\ge 0\),

- (ii)
\(\mu _T^{\mathrm {cross}}(x,y)=0 \iff x=y\) and \(\mu _T^{\mathrm {meet}}(x,y)=0 \iff x=y\),

- (iii)
\(\mu _T^{\mathrm {cross}}(x,y)=\mu _T^{\mathrm {cross}}(y,x)\) and \(\mu _T^{\mathrm {meet}}(x,y)=\mu _T^{\mathrm {meet}}(y,x)\).

### Remark 2.1

*start*in a joint position and end in \(\phi _{0,T}[x]\) and \(\phi _{0,T}[y]\).

### 2.4 A Simple Example

Let us consider a very simple example, where the domain of interest is the interval [0, *L*], and there is no dynamics, i.e., \(v\equiv 0\). Its primary purpose is to form our intuition and expectations about how the semidistances work in more complicated settings. In particular, we shall see how the semidistances scale in time and system size.

*T*]. One can then easily see that \({\dot{x}}_t \equiv L/T\) is an optimal path in (9), thus giving

We will revisit this example in the next section and will realize that the behavior of the discrete semidistances deviates from the one observed here for continuous space and time.

## 3 Large-Deviation Semidistances in Discrete Time and Space

As mentioned in the Introduction, the cost functions \(\mu _T^\mathrm {cross}\) and \(\mu _T^\mathrm {meet}\) are difficult to calculate explicitly, and impossible if the velocity or flow field is not explicitly known. In this section, we take a more practical approach. We will assume that the only information at hand is the position at finite times of a finite number *I* of floaters.

*i*,

Our strategy is to study the probability that random particles described by the SDEs (12) deviate from the given floater trajectories (11), conditional to the fact that all our knowledge about the otherwise unknown flow field \(\phi _{s,t}\) comes from the time- and space-discrete set of trajectory data (11). We first approximate the SDEs (12) by discrete-time, continuous-space Markov processes in Sect. 3.1, as it is done in standard time-stepping methods for SDEs (Kloeden and Platen 2010). Next, in Sect. 3.2 we condition these discrete-time processes on the given floater positions. Then we calculate the large-deviation rate for trajectories in Sect. 3.3, and for endpoints in Sect. 3.4. Finally, we end the section with a discussion of the metric properties of the resulting large-deviation rates in Sect. 3.5.

### 3.1 Discrete-Time Approximation

*i*, and temporarily drop the superindex for brevity. Since the noise process is a standard Brownian motion, we know its density,

*splitting method*, since the deterministic evolution and purely noise parts of the SDE (2) are handled separately in the distinct steps; here it would be “noise-flow-noise”.

We would like to stress that our choice of discretization is made on the basis that we can use the available information on the flow given by (11). In the realm of one-step methods for SDEs we are bound to choices of the form (14), because there is no information on the drift other than the flow generated by it on prescribed time intervals \([k\tau , (k+1)\tau )\). Given the form of time-stepping (14), the optimal [in the sense of highest weak consistency order Kloeden and Platen (2010)] approximation of the SDE is obtained by choosing \(\alpha =1/2\). That is the so-called *Strang-splitting* (Strang 1968) and has weak order two, while for \(\alpha \ne 1/2\) we only get order one.^{4}

Having performed discretization in time, in the next section we derive a discrete-time, discrete-space, \(\alpha \)-dependent Markov chain that we will use to derive discrete semidistances.

### 3.2 Conditioning on Finite Data

*i*. For each \(k=0,\ldots ,K\), we introduce the set

*and*for each intermediate point \(\tilde{\varvec{}{x}}_k^+\in \mathcal {A}_k\). This automatically implies conditioning of the other intermediate points \(\tilde{\varvec{}{x}}_{k+1}^-\in \mathcal {A}_{k+1}\) due to (11). We choose to condition on the intermediate points for practical reasons; otherwise, we would not be able to perform the second step in (14), since the discrete trajectories are our only information about the flow, cf. Remark 3.1 below.

The conditioning on the finite data set results in replacing the discrete-time continuous-space process by a fully discrete-time discrete-space Markov chain that hops between the given trajectories. Therefore, the state of the new Markov chain can be represented by the labels \(j=1,\ldots ,I\); this is particularly useful since the deterministic flow (11) will change the positions but not the labels. Since the resulting process is still Markovian, we can fully characterize its behavior through its transition probabilities for one time-step \(k\rightarrow k+1\). We now calculate these transition probabilities, dealing with each step in (14) separately. See Fig. 3 for a sketch.

*one*time step (see Fig. 3):

- 1.
Start in \(x^{\scriptscriptstyle {(j)}}_k\), and perform a jump to some \(x^{\scriptscriptstyle {(\ell )}}_k\) with probability \(p^{\scriptscriptstyle {(k,+)}}_{j,\ell }\),

- 2.
Perform a deterministic jump from \(x^{\scriptscriptstyle {(\ell )}}_k\) to \(x^{\scriptscriptstyle {(\ell )}}_{k+1}\),

- 3.
Perform a jump from \(x^{\scriptscriptstyle {(\ell )}}_{k+1}\) to \(x^{\scriptscriptstyle {(m)}}_{k+1}\) with probability \(p^{\scriptscriptstyle {(k+1,-)}}_{\ell ,m}\).

### 3.3 Large Deviations of Discrete Trajectories

### Remark 3.1

### Remark 3.2

(Missing data and non-uniform time-sampling)

Note that the construction works exactly as described above even if information about trajectories is partially missing. The conditioning on the set \(\mathcal {A}_k\) works identically, but now these sets might have different cardinalities smaller or equal *I*. Observe that our only information about the deterministic flow for times in \([k\tau ,(k+1)\tau )\) comes from those trajectories that are available both in \(\mathcal {A}_k\) and \(\mathcal {A}_{k+1}\). If this intersection is empty, we need to skip that time slice completely. This is not a problem, since our choice of sampling time uniformly by the step size \(\tau \) was solely in order to ease presentation. As the reader has probably observed, the extension for varying time steps \(\tau _k\) is straightforward.

### 3.4 Large Deviations of Endpoints

*i*to

*j*in discrete time

*K*, and the meeting probability that two independent chains, starting from

*i*and

*j*, respectively, meet by discrete time

*K*or earlier. Since the paths (19) encode more information than the endpoints, we can now easily derive the large deviations of the one-way probability by a contraction principle. Indeed, for any two indices \(i,j=1,\ldots ,I\),

### 3.5 The Semidistances

- (i)
\(\nu _K^{\mathrm {cross}}(i,j)\ge 0\) and \(\nu _K^{\mathrm {meet}}(i,j)\ge 0\),

- (ii)
\(\nu _K^{\mathrm {cross}}(i,j)=0 \iff i=j\) and \(\nu _K^{\mathrm {meet}}(i,j)=0 \iff i=j\),

- (iii)
\(\nu _K^{\mathrm {cross}}(i,j)=\nu _K^{\mathrm {cross}}(j,i)\) and \(\nu _K^{\mathrm {meet}}(i,j)=\nu _K^{\mathrm {meet}}(j,i)\).

*known*positions of the floaters \(x_k^{\scriptscriptstyle {(\ell )}}\) only. Because of this: (1) these costs can be used in practice if the velocity field is unknown (Sects. 4 and 5); (2) these costs can even be applied to cases where there may not be an underlying velocity field, as for example in discrete-time dynamical system (Sect. 4.1).

### Remark 3.3

(Time reversal for discrete semidistances) Similarly to Remark 2.1, the one-way cost satisfies the time-reversal property \(\nu _K(i\rightarrow j) = {\overleftarrow{\nu }}\!\!_K(j\rightarrow i)\), provided \(\alpha =1/2\), where \(\overleftarrow{\nu }\!\!_K\) is the cost associated with the backward dynamics. Moreover, this time-reversal property also holds for the cross-semidistance, whereas for the meeting semidistance \(\nu _K^\mathrm {meet}(i,j)=\min _{\ell =1,\ldots ,I} \overleftarrow{\nu }\!\!_K(\ell \rightarrow i) + \overleftarrow{\nu }\!\!_K(\ell \rightarrow j)\). Apart from the superior consistency order discussed in Sect. 3.1, the invariance of semidistances under time reversal is another reason for choosing \(\alpha =1/2\).

### Remark 3.4

*K*. Naturally, this is simply the \(L^2\)-distance between two trajectories, as considered earlier in Froyland and Padberg-Gehle (2015). Although this construction is very easy to calculate and its square root is a genuine metric, it is less interpretable as a cost for transport and mixing.

### Remark 3.5

It should be noted that the semidistances \(\nu _K^\mathrm {meet},\nu _K^\mathrm {cross}\) scale quadratically in space; this becomes even more apparent in the example considered in Sect. 3.7. In the case of the \(L^2\)-distance (21), the cost becomes a genuine distance after taking the square root. However, if we take the square roots of \(\nu _K^\mathrm {meet}\) and \(\nu _K^\mathrm {cross}\), the triangle inequality still fails. We therefore stick to the quadratic scaling as this has the most direct interpretation as large-deviation costs.

### Remark 3.6

(Eulerian transport vs Lagrangian mixing) When speaking of transport in this paper, we mean “*transport* (of probability) *from a trajectory to another*”, to express how the dynamics is mixing up regions these two trajectories come in contact with. This can be seen as a Lagrangian perspective. We express with large-deviation rates the *unlikeliness* of transitions between trajectories, and these are then computed as shortest paths, cf. Sect. 3.4. Deceivingly similar mathematical constructions show up in Ser-Giacomi et al. (2015), where the authors consider “highly probable paths” of non-homogeneous Markov chains, which also leads to a time-dependent shortest path problem. Note, however, that this is orthogonal to our concept, as this is quantifying *likeliness*. A further important distinction is that their Markov chain is constructed in an Eulerian manner (opposed to our Lagrangian setting), meaning that it describes transport between fixed regions of state space; serving as a discretization of the flow field (Froyland et al. May 2007, 2010).

### 3.6 Discretization of the Continuous Semidistances

We now show that the one-way discrete-space–time cost \(\nu _K\) can also be obtained by discretizing the continuous-space–time cost \(\mu _T\). This means that discretization and derivation of the large-deviation principle are interchangeable operations (if done the right way). We will not be precise about the discretization error; of course one needs to assume that the number of floaters is sufficiently large.

### 3.7 The Simple Example Revisited

Let us now demonstrate how the results of this section apply to the example of Sect. 2.4.

*Discrete Time and Continuous Space*

*Discrete Time and Space*

*I*of equispaced trajectories of this system sampled at the same times as in the previous paragraph, the virtual random walker cannot make arbitrarily small jumps as in the continuous state case, thus

*I*grows. However, if \(K>I\), the smallest jumps are \(\Delta x_k = \frac{L}{I}\), thus

*I*/ 2 of them, such that we obtain \(\nu _K(0{\scriptstyle \rightarrow }L/2) = \nu _K(L/2{\scriptstyle \rightarrow }L) = \frac{I}{2}\cdot \frac{(L/I)^2}{\tau } = \frac{L^2}{4I\tau }\) (for even

*I*, and vanishing error for odd

*I*as

*I*grows).

The main lesson is that while in the continuous space case halving the Euclidean distance makes the semidistance scale by \(\frac{1}{4}\), if the spatial resolution of trajectories is coarse, the discrete semidistance scales only by \(\frac{1}{2}\). In general, if \(|x-y|=\delta \), then on a coarse resolution grid it takes about \(\frac{\delta }{\Delta x}\) jumps to travel between these two points, and we obtain \(\nu _K(x{\scriptstyle \rightarrow }y) \approx \frac{\delta }{\Delta x}\cdot \frac{(\Delta x)^2}{\tau } = \delta \cdot \frac{\Delta x}{\tau }\). Note that \(\Delta x\) and \(\tau \) are constant quantities, and thus the one-way discrete cost scales *linearly* in the Euclidean distance between the two points, as opposed to *quadratic* scaling in the continuous space case.

## 4 Coherence Analysis with Semidistances

Let us assume that we are given a set of discrete time and space trajectories \(\{x_k^{\scriptscriptstyle {(j)}}\}_{j=1,\ldots ,I,k=0,\ldots ,K}\), and a (semi)distance *d*. We now describe how such semidistances can be used to distinguish and analyze coherent sets from the finite data. We shall work with an unspecified semidistance *d*, but of course the semidistances that we have in mind are \(\nu _K^\mathrm {cross}\) and \(\nu _K^\mathrm {meet}\) that we derived in the previous section. Other—not large-deviation based—distance measures could be used just as well, as we discuss below. Nevertheless, the semidistances should not be completely arbitrary; we assume that they share the behavior of \(\nu _K^\mathrm {meet}\) and \(\nu _K^\mathrm {cross}\) that we discuss in Sects. 4.1 and 4.2.

To illustrate the ideas, we first analyze the behavior of two one-dimensional prototypical examples. These examples show the difference between two types of regions: “mixing” and “static” (also known as “regular”). In these one-dimensional and simple examples, one can easily determine the regions and whether they are mixing or static from the semidistances. One example has two invariant sets under the dynamics, which is (measure-theoretically and topologically) mixing on both of them. The other has two static regions, where the mutual physical distance of trajectories does not change under the dynamics, and these regions are separated by a third, mixing region.

After this we proceed with a more involved model: a two-dimensional periodically forced double-gyre flow, where the boundaries of the separate regions are no longer as clear-cut as in the one-dimensional example. Nevertheless, we will show that one can identify the separate regions via the tools that we present next.

To this end, we introduce the notion of cornerstones, representing possible coherent sets or mixing regions, and then discuss how to find them and when to stop searching for them. Finally, to obtain coherent sets, we assign the trajectories to cornerstones. The notion of fuzzy affiliations will be used to express the uncertainty whether a trajectory close to the boundary of a set belongs to it or not. Note that in the case of finite data such an uncertainty is always present.

### 4.1 Two Illustrative Model Cases

*Two Invariant, Mixing Subdomains*

^{5}

^{6}

*Two Static Regions Divided by a Mixing One*

From our analysis in Sect. 3.7, we would have expected to see quadratic growth of the one-way cost with respect to physical distance in the static regions, but we only observe linear growth. This is due to the finite number of considered trajectories, as also explained in the second paragraph of Sect. 3.7. All points of the mixing region have almost the same cost from any one point in the static regions, and approximately zero cost from one another. To obtain the cost between two points of different static regions, one has to consider the cost to go to the boundary of the static and mixing regions (linear cost in Euclidean distance), travel on a trajectory from there to the boundary of the other static region (at zero cost), and then go from there to the desired point. (Again, linear cost in Euclidean distance that needs to be covered.) Thus, the cost (and semidistance) between these two points is the sum of their one-way cost (and semidistance) to the mixing region, provided the time of consideration is sufficiently large for the mixing to take place.^{7} Since our fictive random walker uses trajectories of the mixing region to travel from one static region to the other, we will also call it *transition region* henceforth.

To conclude, from Fig. 8 we can easily identify three separate regions, and from the steepness of the slopes (linear/quadratic or flat), we can determine wether a region is static or mixing. As the next example shows, this distinction is usually not as clear as in these constructed examples, but the main ideas will be based on this observation.

### 4.2 The Periodically Forced Double Gyre

*X*. Equation (24) describes two counter-rotating gyres next to each other (the left one rotates clockwise), with the vertical boundary between the gyres oscillating periodically, see Fig. 9.

We choose a uniform \(50\times 25\) grid as initial conditions for the floaters at time \(t=0\); i.e., \(I=1250\). We sample the trajectories of these floaters at times \(t_k = k\tau \), \(k=0,1,\ldots ,K\), where \(K=100\) and \(\tau = 0.2\). That means, the length of trajectories in consideration is 20 times the period of the forcing.

Employing our large-deviation-based distance computations on this data set using Algorithm 1 and \(\alpha =1/2\), we get the one-way costs \(\nu _K(i{\scriptstyle \rightarrow }j)\), \(i,j=1,\ldots ,I\), from which we compute \(\nu _K^\mathrm {cross}(i,j)\) and \(\nu _K^\mathrm {meet}(i,j)\).

^{8}Since the semidistances grow linearly in the Euclidean distance inside the gyre, we see that we are in the low-resolution regime discussed in Sect. 3.7. Analogously to Fig. 8, we can again (vaguely) distinguish three regions: a steep (square-root-type) region, a flat region, and another steep (flipped square-root-type) region. As before, the flat region is typically strongly mixing, and the steep regions are static. We shall make this distinction more precise in the next sections.

### Remark 4.1

(Three-dimensional flows) Clearly, the scaling behavior shown in Figs. 8 and 10is dimension-dependent. For a three-dimensional static region, one would see a slope 1 / 3 on a log–log scale. The question remains, how does a typical coherent set behave there; can it be modeled by a static region? If an incompressible flow rotates uniformly in a plane, it necessarily has a constant shifting motion in the perpendicular axial direction, leading to cylindrical vortices. If the cylindrical rings of a vortex rotate at different angular frequencies, the flow speed along axial directions is non-uniform, and mixing-type behavior occurs in the vortex (Halász et al. 2007). We leave the analysis of such systems to future work and proceed with analyzing different aspects of prototypical two-dimensional flows here.

### 4.3 Cornerstones

*d*, we randomly choose a trajectory, represented by a label \(c_0\in \{1,\ldots ,I\}\), and compute the trajectory furthest from it, i.e, we set

*all*the other already identified “cornerstones” \(\{c_{q}\}_{q=1,\ldots ,Q}\), as in Rüdrich et al. (2017):

### 4.4 Number of Cornerstones

How to determine the number of cornerstones that should be used? Is there an optimal number, or is it up to our liking? In the case of the double gyre, as noted above, a fourth cornerstone would be part of one of the gyres, and assigning affiliations would thus split one gyre into two sets. If the gyres would consist of a continuum of periodic orbits, then we could proceed and split them this way into as many rings as we like. The same situation in an idealized framework appears in Sect. 4.1 for the static regions: since they are static, arbitrary subsets are perfectly coherent (even invariant in this case).

A good place to stop searching for further cornerstones would be when they would start to subdivide “maximal coherent” sets, as the gyres in the double-gyre example, or the static sets in the second, and the invariant sets in the first example of Sect. 4.1. To this end we make an idealized assumption: *Coherent sets appear as for the second example in Sect.* 4.1, *i.e., multiple static regions divided by one mixing region*. Here, “static” is meant in the sense that the mutual distances between points in the set barely change. Such an assumption was also utilized in (Hadjighasem et al. 2016).

Note that if there are \(C\ge 2\) coherent sets, the first *C* corner stones are going to be in them, one in each. This is due to the fact that to move from the center of one static region to another, the shortest path in (20) needs to move out of one set, travel in the transition region to the other set, and move to its center, hence maximizing the minimal distance to all other cornerstones. After finding all static regions, the next cornerstone is to be found in the transition region, if all static regions are approximately of the same size—which we assume here. The crucial observation is that this \((C+1)\)-st cornerstone is half as far from the other cornerstones,^{9} as they are from one another. In other words, \(d(c_i,c_{C+1}) + d(c_j,c_{C+1}) \approx d(c_i,c_j)\), \(i,j\le C\).

To summarize, our simple check when to stop searching for cornerstones is going to be, when the value of the objective function in (25) drops by at least a factor two compared with the previous value. Observe how nicely this works in the periodically forced double-gyre case: the rightmost points of the curves in Fig. 11 are the optimizers, and the corresponding value of the yellow curve is less than half of the values for the first two cornerstones. This indicates to stop with three cornerstones, as they will represent both the gyres and the transition region.

### 4.5 Clustering and Fuzzy Affiliations

*d*, we assign to each cornerstone \(c_1,c_2,c_3\) the trajectories that are closer to them than to the two other cornerstones, respectively. For the periodically forced double gyre and the meeting distance this is shown in Fig. 13.

Comparing this picture with the typical trajectories of the time-1 Poincaré map of the system [again, see Froyland and Padberg-Gehle 2014, Figure 1], it appears that the gyre regions in our figure are smaller. This is due to the nature of the transport distance at hand: the gyres are partly made up of so-called regular regions of the Poincaré map, meaning that typical trajectories move on periodic orbits that are approximately concentric circular lines. Transport between these trajectories is only possible through diffusion, and the price one has to pay for this transport in radial direction is reflected by the rate function (recall, this is what we model by the static regions in Sect. 4.1). The cost to get from the center of the gyre (the cornerstone \(c_1\) or \(c_2\)) to a regular trajectory in the same gyre is proportional to the “radial distance” between them (compare with the static part of the second example in Sect. 4.1). This behavior is not characteristic for the well-mixed transition region, because there the dynamics (eventually) brings any two trajectories close to each other. The effect is most prominent if the time frame of consideration grows infinitely large, and on our finite time horizon it appears as a flattening of the curve. This brings us back to why the blue and green regions in Fig. 13 are smaller than gyres in the Poincaré map. The answer is simply, because the outer periodic orbits are closer to the transition region than to the center of the gyre, hence also closer to the cornerstone \(c_3\) that is in the transition region, because the points in the transition region have very small distance from one another.

*fuzzy affiliations*\(q_{c_i}(\cdot )\), to obtain more refined information on coherence. For instance, let \(m>1\), and minimizing the affiliation-weighted penalty function

*fuzzy c-means algorithm*(Bezdek 1981), giving \(q_{c_i}(j)=1\) \(\Leftrightarrow \) \(d(c_i,j)=0\), i.e., affiliation is maximal if the distance is minimal. Further, the parameter

*m*controls the fuzziness of the clustering: large

*m*gives soft clusters, while

*m*approaching 1 gives more and more “crisp” clusters as the affiliations converge either to 0 or to 1 (Bezdek et al. 1987). The resulting affiliations (indicated at initial time) for \(m=2\) are shown in Fig. 14. For

*m*close to 1 we obtain affiliations very similar to the hard clusters in Fig. 13.

## 5 Numerical Results

In the previous section we already presented numerical results for the double-gyre system, where we used the results to motivate and develop the analysis tools. In this section we apply these tools to two other well-analyzed test cases: the perturbed Bickley Jet and the rotating (transitory) double gyre. They are different paradigmatic examples, as the Bickley Jet has a non-vortex coherent set (the jet core), and the transitory double gyre is not a periodically forced system, thus genuinely living on a finite time interval.

Let us also point out that the examples presented here and in the previous section are all one- or two-dimensional. Although we expect the analysis in higher dimensions to be at least qualitatively not very different from the two-dimensional case, dealing with the nevertheless arising subtle differences (see Remark 4.1) is beyond the scope of this conceptual work.

It turns out that the choice between the two semidistances \(\nu _K^\mathrm {cross}\) or \(\nu _K^\mathrm {meet}\) has only marginal influence on the results. In this section we shall mostly work with the cross-semidistance.

### 5.1 The Bickley Jet

The constants are chosen as in Rypina et al. (2007, Section 4) . In particular, we set \(k_n = 2n/r_e\) with \(r_e = 6.371\), \(U_0 = 5.414\), and \(L = 1.77\). The phase speeds \(c_n\) of the Rossby waves are \(c_1 = 0.1446U_0\), \(c_2 = 0.205U_0\), \(c_3 = 0.461U_0\), their amplitudes \(A_1 = 0.0075\), \(A_2 = 0.15\), and \(A_3 = 0.3\), as in Hadjighasem et al. (2016). The system is considered on a state space \(X = [0,\pi r_e]\times [-3,3]\) which is periodic in the horizontal \(x_1\) coordinate.

We choose a uniform \(60\times 18\) grid as initial conditions for the floaters at time \(t=0\); i.e., \(I=1080\). We sample the trajectories of these floaters at times \(t_k = k\tau \), \(k=0,1,\ldots ,K\), where \(K=80\) and \(\tau = 0.5\). In this time interval, typical trajectories cross the cylindrical state space horizontally 4–5 times, trajectories in the jet core (the wavy structure in Fig. 15) up to 9 times.

*one*mixing region, is not satisfied: the jet core is a coherent set itself, dividing two mixing regions (below and above it), each containing 3 further coherent sets (the gyres). Thus, \(c_1\) and \(c_2\) have maximal distance (\(\nu _K^\mathrm {cross}(c_1,c_2)=3.21\)), because the random walker needs to cross the jet core. Every further cornerstone \(c_3,\ldots ,c_6\) can be reached from either \(c_1\) or \(c_2\) through one of the mixing regions, and thus have a very similar cost. The deviation of these costs, \(2.14\text {--}2.45\), shows that we did not reach the state of full mixing on the chosen time interval.

Now, the seventh cornerstone lies in the jet core, which has to be crossed if traveling between cornerstones that are below and above it, respectively. The corresponding cost (1.42) is a bit larger than half of the previous cost, because \(c_7\) does not lie on the shortest path between cornerstones below and above the jet core. Intuitively, the “center line” of the jet core should be equally far from all cornerstones \(c_1,\ldots ,c_6\), if the time interval is large enough such that the regions around the gyres are truly mixing. Since it is not, there are points on the boundary of the jet core which are easier to reach from them, and thus easier to cross there. The cornerstone \(c_7\) represents the position where it is the hardest to cross. The eighth cornerstone has truly half the semidistance to the closest one than \(c_7\), and lies in one of the mixing regions.

^{10}The semidistances are shown in Fig. 16.

*m*is to 1, the more “crisp” the affiliation function is forced to be, and the mixing region is more easily reached from the thin jet core than from the gyres.

### 5.2 The Rotating Double Gyre

*X*, and repeating the analysis for their trajectories. We show the resulting fuzzy affiliations in Fig. 22.

## 6 Discussion and Outlook

### 6.1 The Dynamic Laplacian

Froyland (2015) has introduced the *dynamic Laplacian* as a transport-related tool to find coherent sets. Similarly to our approach, it makes use of a small random perturbation of size \(\varepsilon \), then \(\varepsilon \) is driven to zero.

Numerical methods so far discretize directly the dynamic Laplacian (Froyland and Junge 2015; Banisch and Koltai 2017; Froyland and Junge 2017). In light of our analysis, which can be used both ways (derive the large-deviation principle in continuous space, then discretize it to finite trajectories, cf. Sect. 3.6, or discretize the dynamics to finite trajectories, then derive the large-deviation principle on them, cf. Sect. 3.3), we ask whether there is a discrete dynamic Laplacian that can be derived from a discretization of the perturbed dynamics?

Mimicking the construction in Froyland (2015) and sketching the idea while skipping details, one should construct a discrete, \(\varepsilon \)-dependent transfer operator \(T_{\varepsilon }\in \mathbb {R}^{I\times I}\), that represents transition probabilities of a forward-backward process, then obtain a discrete dynamic Laplace operator \(L_{\text {dyn}} := \frac{d}{d\varepsilon }\big \vert _{\varepsilon =0}T_{\varepsilon }\). A discrete transfer operator \(T_\varepsilon \) that is a consistent approximation of the continuous dynamics can be obtained by a construction as in Sect. 3.2, by using the transition probabilities (17). Technical details aside, we see that the probabilities are linear combinations of terms of the form \(\text {e}^{-\Delta x / \varepsilon }\), where \(\Delta x\) here is a formal distance term that appears in the formulas. Differentiation with respect to \(\varepsilon \) immediately yields that all off-diagonal entries (basically, where \(\Delta x>0\)) of \(L_{\text {dyn}}\) are zero, in fact the matrix is the identity.

Thus, this approach of discretizing the dynamics first, and then factoring out the \(\varepsilon \)-small stochastic perturbation does not give a dynamically meaningful result. In analytic terms the very same problem occurred in a different attempt to introduce a discrete dynamic Laplacian from a discrete transfer operator, see (Banisch and Koltai 2017, Section IV) . In general, it would be desirable to understand when and how can the “first discretize, then factor out \(\varepsilon \)” methods work, such that they can complement the methods that directly discretize the (continuous) dynamic Laplace operator.

### 6.2 Other Distance Measures

The time-dependent shortest path problem used to compute our semidistances is computationally demanding in our current algorithmic realization, which theoretically limits the number of trajectories that can be handled. Moreover, they do not satisfy the triangle inequality, hence they are not a metric. Although numerical efficiency is not the main focus of this paper, and we demonstrated the usefulness of our semidistances in unraveling the underlying dynamical structure of the example systems, a more cheaply computable metric would enhance the utility and significance of the analysis methods presented here.

Ultimately, one would like to understand the intrinsic, possibly low-dimensional geometric organization of the state space with respect to transport and mixing, as pioneered in Banisch and Koltai (2017). Employing proper metrics would allow, e.g., the usage of low-dimensional embedding techniques, such as multidimensional scaling, to represent and better understand this geometric organization. One canonical candidate would be the metric structure related to the dynamic Laplacian, considered in Karrasch and Keller (2016). This will be subject of future studies.

To summarize, although other distance measures could be used to analyze complicated dynamic behavior, we showed that the semidistances we derived in this paper from the physical notion of transport and mixing in the vanishing diffusion setting are natural and effective diagnostic tools.

## Footnotes

- 1.
We denote random variables by boldface symbols.

- 2.
A different way of factoring out diffusion to obtain coherent sets for deterministic flows appeared in the set-oriented transfer operator-based characterization in Froyland (2015); Froyland and Kwok (2016), leading to the notion of the

*dynamic Laplacian*. See also our concluding remarks in Sect. 6.1. - 3.
More rigorously, (7) means \( \varepsilon \log \mathbb {P}\left[ \sqrt{\varepsilon }\varvec{}{w}_{(\cdot )} \in U\right] \xrightarrow [\varepsilon \rightarrow 0]{} -\inf _{w_{(\cdot )} \in U}\frac{1}{2}\int _0^T\!|\dot{w}_t|^2 \mathrm{d}t, \) where, for technical reasons, this convergence is realized by a liminf lower bound for open sets

*U*and limsup upper bound for closed sets*U*; see Dembo and Zeitouni (1987). - 4.Formally, this can be seen by denoting the generators of the noise process and advection by
*A*and*B*, respectively, and estimating the difference of the Markov propagators associated with (12) and (14) by performing formal Taylor expansions in \(\tau \) with the non-commuting operators*A*and*B*, to obtain$$\begin{aligned} e^{\tau (A+B)} - e^{(1-\alpha )\tau A} e^{\tau B} e^{\alpha \tau A} = \left\{ \begin{array}{ll} \mathcal {O}(\tau ^3), &{} \alpha = 1/2,\\ \mathcal {O}(\tau ^2), &{} \alpha \ne 1/2. \end{array} \right. \end{aligned}$$ - 5.
If a system \((X,\psi ,\mu )\) is mixing, then \((X\times X,\psi \times \psi ,\mu \times \mu )\) is ergodic (Walters 2000, Theorem 1.24) . Thus, for \(\mu \times \mu \)-almost every pair (

*x*,*y*), the trajectory \((\psi \times \psi )^t(x,y)\) will enter every set*A*of nonzero measure for some \(t\ge 0\). This shows (22) by taking \(\psi = \phi \vert _{(0,1/2)}\) or \(\psi = \phi \vert _{(1/2,1)}\), and \(A = \{(x,y)\,\vert \, |x-y|<\varepsilon \}\) for any fixed \(\varepsilon >0\). - 6.
Applying Footnote 5 to the case where \(\psi =\phi \vert _{(0,1/2)}\), we obtain that \((\phi \vert _{(0,1/2)} \times \phi \vert _{(0,1/2)} )^t(x,y)\) enters \(A = \{(x,y)\,\vert \, |x-1/2| + |y| < \varepsilon \}\) eventually. Noting that \(\phi \vert _{(1/2,1)} = \phi \vert _{(0,1/2)} (\cdot - \tfrac{1}{2}) + \tfrac{1}{2}\), the claim follows. Comparing the different slopes in Fig. 7, based on the reasoning in Footnote 5 and here we conjecture that the minimal distance of two trajectories decays faster in the case when they both start in the same invariant set, because the set \(\{(x,y)\,\vert \, |x-y|<\varepsilon \}\) is larger in measure than \(\{(x,y)\,\vert \, |x-1/2| + |y| < \varepsilon \}\).

- 7.
Note that for this argument, ergodicity of the dynamics in the “mixing” region would be sufficient, since one only needs to travel “from one static region to the other”. The crucial additional property we get from mixingness is that the mutual semidistances of points in this region go to zero.

- 8.
As on a regular grid there are \(\mathcal {O}(\delta ^2)\) points not further than Euclidean distance \(\delta \) from a reference point, the

*r*-th closest point to the reference point has distance \(\mathcal {O}(r^{1/2})\). - 9.
Here we assume that we are in the coarse spatial resolution case, where the semidistances scale linearly and not quadratically, cf. Sect. 3.7. Otherwise, the drop in the distance is more than a factor two (toward factor four).

- 10.
If we include additional cornerstones, results tend to deteriorate due to the low resolution and because the chosen time interval is not giving full mixing.

- 11.
It should be mentioned here that every time-dependent shortest path problem can be rephrased as a time-independent problem by considering each node at each time point as a distinct node of a large graph, and then it could be solved by standard methods. We do not take this approach here, as it might introduce memory requirement issues.

## Notes

### Acknowledgements

This work is supported by the Deutsche Forschungsgemeinschaft (DFG) through the Priority Programme SPP 1881 “Turbulent Superstructures”, and through the CRC 1114 “Scaling Cascades in Complex Systems”, projects A01 and C08.

## Supplementary material

## References

- Allshouse, M.R., Peacock, T.: Lagrangian based methods for coherent structure detection. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**25**(9), 97617 (2015)Google Scholar - Allshouse, M.R., Thiffeault, J.-L.: Detecting coherent structures using braids. Physica D: Nonlinear Phenom.
**241**(2), 95–105 (2012)CrossRefGoogle Scholar - Andres, S.: Diffusion processes with reflection. PhD thesis, TU Berlin (2009)Google Scholar
- Balasuriya, S., Froyland, G., Santitissadeekorn, N.: Absolute flux optimising curves of flows on a surface. J. Math. Anal. Appl.
**409**(1), 119–139 (2014)MathSciNetCrossRefzbMATHGoogle Scholar - Banisch, R., Koltai, P.: Understanding the geometry of transport: diffusion maps for Lagrangian trajectory data unravel coherent sets. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**27**(3), 035804 (2017)zbMATHGoogle Scholar - Bezdek, J.C.: Pattern Recognition with Fuzzy Objective Function Algorithms. Kluwer Academic Publishers, Dordrecht (1981)CrossRefzbMATHGoogle Scholar
- Bezdek, J.C., Hathaway, R.J., Sabin, M.J., Tucker, W.T.: Convergence theory for fuzzy c-means: counterexamples and repairs. IEEE Trans. Syst. Man Cybern.
**17**(5), 873–877 (1987)CrossRefzbMATHGoogle Scholar - Budišić, M., Mezić, I.: Geometry of the ergodic quotient reveals coherent structures in flows. Physica D: Nonlinear Phenom.
**241**(15), 1255–1269 (2012)MathSciNetCrossRefzbMATHGoogle Scholar - Dellnitz, M., Junge, O.: On the approximation of complicated dynamical behavior. SIAM J. Numer. Anal.
**36**, 491–515 (1999)MathSciNetCrossRefzbMATHGoogle Scholar - Dellnitz, M., Froyland, G., Horenkamp, C., Padberg-Gehle, K., Gupta, A.S.: Seasonal variability of the subpolar gyres in the Southern Ocean: a numerical investigation based on transfer operators. Nonlinear Proces. Geophys.
**16**, 655–664 (2009)CrossRefGoogle Scholar - Dembo, A., Zeitouni, O.: Large Deviations Techniques and Applications, vol. 38, 2nd edn. Springer, New York (1987)zbMATHGoogle Scholar
- Denner, A., Junge, O., Matthes, D.: Computing coherent sets using the Fokker–Planck equation. J. Comput. Dyn.
**3**(2), 163–177 (2016)MathSciNetCrossRefzbMATHGoogle Scholar - Dijkstra, E.W.: A note on two problems in connexion with graphs. Numer. Math.
**1**(1), 269–271 (1959)MathSciNetCrossRefzbMATHGoogle Scholar - Fabregat , A., Mezic, I., Poje, A. C.: Finite-time partitions for Lagrangian structure identification in Gulf Stream eddy transport. (2016). arXiv preprint arXiv:1606.07382
- Floyd, R.W.: Algorithm 97: shortest path. Commun. ACM
**5**(6), 345 (1962)CrossRefGoogle Scholar - Freidlin, M., Wentzell, A.: Random Perturbations of Dynamical Systems, 2nd edn. Springer, Berlin (1998)CrossRefzbMATHGoogle Scholar
- Froyland, G., Junge, O.: Robust FEM-based extraction of finite-time coherent sets using scattered, sparse, and incomplete trajectories. (2017). arXiv preprint arXiv:1705.03640
- Froyland, G., Kwok, E.: A dynamic Laplacian for identifying Lagrangian coherent structures on weighted Riemannian manifolds (2016). arXiv preprint arXiv:1610.01128
- Froyland, G., Padberg-Gehle, K.: Almost-invariant and finite-time coherent sets: directionality, duration, and diffusion. In: Ergodic Theory, Open Dynamics, and Coherent Structures, pp. 171–216. Springer (2014)Google Scholar
- Froyland, G.: An analytic framework for identifying finite-time coherent sets in time-dependent dynamical systems. Physica D: Nonlinear Phenom.
**250**, 1–19 (2013)MathSciNetCrossRefzbMATHGoogle Scholar - Froyland, G.: Dynamic isoperimetry and the geometry of Lagrangian coherent structures. Nonlinearity
**28**(10), 3587 (2015)MathSciNetCrossRefzbMATHGoogle Scholar - Froyland, G., Junge, O.: On fast computation of finite-time coherent sets using radial basis functions. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**25**(8), 087409 (2015)zbMATHGoogle Scholar - Froyland, G., Koltai, P.: Estimating long-term behavior of periodically driven flows without trajectory integration. Nonlinearity
**30**(5), 1948 (2017)MathSciNetCrossRefzbMATHGoogle Scholar - Froyland, G., Padberg-Gehle, K.: A rough-and-ready cluster-based approach for extracting finite-time coherent sets from sparse and incomplete trajectory data. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**25**(8), 087406 (2015)zbMATHGoogle Scholar - Froyland, G., Padberg, K., England, M.H., Treguier, A.M.: Detection of coherent oceanic structures via transfer operators. Phys. Rev. Lett.
**98**, 224503 (2007)CrossRefGoogle Scholar - Froyland, G., Santitissadeekorn, N., Monahan, A.: Transport in time-dependent dynamical systems: finite-time coherent sets. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**20**(4), 043116 (2010)zbMATHGoogle Scholar - Froyland, G., Horenkamp, C., Rossi, V., van Sebille, E.: Studying an Agulhas ring’s long-term pathway and decay with finite-time coherent sets. Chaos
**25**(8), 083119 (2015)MathSciNetCrossRefzbMATHGoogle Scholar - Hadjighasem, A., Karrasch, D., Teramoto, H., Haller, G.: Spectral-clustering approach to Lagrangian vortex detection. Phys. Rev. E
**93**(6), 063107 (2016)CrossRefGoogle Scholar - Halász, G., Gyüre, B., Jánosi, I.M., Szabó, K.G., Tél, T.: Vortex flow generated by a magnetic stirrer. Am. J. Phys.
**75**(12), 1092–1098 (2007)CrossRefGoogle Scholar - Haller, G., Beron-Vera, F.: Coherent Lagrangian vortices: the black holes of turbulence. J. Fluid Mech.
**731**, R4 (2013). https://doi.org/10.1017/jfm.2013.391 CrossRefzbMATHGoogle Scholar - Haller, G.: Finding finite-time invariant manifolds in two-dimensional velocity fields. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**10**(1), 99–108 (2000)zbMATHGoogle Scholar - Haller, G.: Distinguished material surfaces and coherent structures in three-dimensional fluid flows. Physica D
**149**(4), 248–277 (2001)MathSciNetCrossRefzbMATHGoogle Scholar - Haller, G., Beron-Vera, F.J.: Geodesic theory of transport barriers in two-dimensional flows. Physica D: Nonlinear Phenom.
**241**(20), 1680–1702 (2012)CrossRefzbMATHGoogle Scholar - Karrasch, D., Keller, J.: A geometric heat-flow theory of Lagrangian coherent structures (2016). arXiv Preprint arXiv:1608.05598
- Karrasch, D.: Lagrangian transport through surfaces in volume-preserving flows. SIAM J. Appl. Math.
**76**(3), 1178–1190 (2016)MathSciNetCrossRefzbMATHGoogle Scholar - Kloeden, P.E., Platen, E.: Numerical Solution of Stochastic Differential Equations, 3rd edn. Springer, Berlin (2010)zbMATHGoogle Scholar
- Koltai, P., Ciccotti, G., Schütte, C.: On metastability and Markov state models for non-stationary molecular dynamics. J. Chem. Phys.
**145**(17), 174103 (2016)CrossRefGoogle Scholar - MacKay, R., Meiss, J., Percival, I.: Transport in Hamiltonian systems. Physica D: Nonlinear Phenom.
**13**(1–2), 55–81 (1984)MathSciNetCrossRefzbMATHGoogle Scholar - Meiss, J.: Symplectic maps, variational principles, and transport. Rev. Mod. Phys.
**64**(3), 795 (1992)MathSciNetCrossRefzbMATHGoogle Scholar - Mezić, I., Banaszuk, A.: Comparison of systems with complex behavior. Physica D: Nonlinear Phenom.
**197**(1), 101–133 (2004)MathSciNetzbMATHGoogle Scholar - Mezić, I., Wiggins, S.: A method for visualization of invariant sets of dynamical systems based on the ergodic partition. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**9**(1), 213–218 (1999)MathSciNetzbMATHGoogle Scholar - Mosovsky, B.A., Meiss, J.D.: Transport in transitory dynamical systems. SIAM J. Appl. Dyn. Syst.
**10**(1), 35–65 (2011)MathSciNetCrossRefzbMATHGoogle Scholar - Øksendal, B.: Stochastic Differential Equations—An Introduction with Applications, 6th edn. Springer, Berlin (2003)zbMATHGoogle Scholar
- Onsager, L., Machlup, S.: Fluctuations and irreversible processes. Phys. Rev.
**91**(6), 1505–1512 (1953)MathSciNetCrossRefzbMATHGoogle Scholar - Padberg, K., Hauff, T., Jenko, F., Junge, O.: Lagrangian structures and transport in turbulent magnetized plasmas. New J. Phys.
**9**, 400 (2007)CrossRefGoogle Scholar - Padberg-Gehle, K., Schneide, C.: Network-based study of Lagrangian transport and mixing. Nonlinear Proces. Geophys. Discuss.
**1–14**, 2017 (2017)Google Scholar - Rom-Kedar, V., Wiggins, S.: Transport in two-dimensional maps. Arch. Ration. Mech. Anal.
**109**(3), 239–298 (1990)MathSciNetCrossRefzbMATHGoogle Scholar - Rüdrich, S., Sarich, M., Schütte, C.: Utilizing hitting times for finding metastable sets in non-reversible markov chains. To appear in J. Comput. Dyn. (2017) Preprint https://opus4.kobv.de/opus4-zib/frontdoor/index/index/docId/5120
- Rypina, I.I., Pratt, L.J.: Trajectory encounter volume as a diagnostic of mixing potential in fluid flows. Nonlinear Proces. Geophys.
**24**(2), 189 (2017)CrossRefGoogle Scholar - Rypina, I., Brown, M., Beron-Vera, F., Kocak, H., Olascoaga, M., Udovydchenkov, I.: On the Lagrangian dynamics of atmospheric zonal jets and the permeability of the stratospheric polar vortex. J. Atmos. Sci.
**64**(10), 3595–3610 (2007)CrossRefGoogle Scholar - Schlueter-Kuck, K.L., Dabiri, J.O.: Coherent structure colouring: identification of coherent structures from sparse data using graph theory. J. Fluid Mech.
**811**, 468–486 (2017)MathSciNetCrossRefzbMATHGoogle Scholar - Ser-Giacomi, E., Rossi, V., López, C., Hernández-García, E.: Flow networks: a characterization of geophysical fluid transport. Chaos: an Interdisciplinary. J. Nonlinear Sci.
**25**(3), 036404 (2015)Google Scholar - Ser-Giacomi, E., Vasile, R., Hernández-García, E., López, C.: Most probable paths in temporal weighted networks: an application to ocean transport. Phys. Rev. E
**92**(1), 012818 (2015)CrossRefGoogle Scholar - Strang, G.: On the construction and comparison of difference schemes. SIAM J. Numer. Anal.
**5**(3), 506–517 (1968)MathSciNetCrossRefzbMATHGoogle Scholar - Treguier, A.-M., Boebel, O., Barnier, B., Madec, G.: Agulhas eddy fluxes in a 1/6 degrees Atlantic model. Deep Sea Res. Part II
**50**(1), 251–280 (2003)CrossRefGoogle Scholar - Walters, P.: An Introduction to Ergodic Theory, vol. 79. Springer, Berlink (2000)zbMATHGoogle Scholar
- Wigner, E.P.: Calculation of the rate of elementary association reactions. J. Chem. Phys.
**5**, 720–725 (1937)CrossRefGoogle Scholar - Williams, M.O., Rypina, I.I., Rowley, C.W.: Identifying finite-time coherent sets from limited quantities of Lagrangian data. Chaos
**25**(8), 087408 (2015)MathSciNetCrossRefzbMATHGoogle Scholar

## Copyright information

**Open Access**This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.