WorksPhysics and Astronomy
Emergence of chaos and fractality in the basin boundary of subcritical shear flow
Abstract
From a dynamical systems perspective of subcritical transition in shear flows, the basin boundary separating the laminar and turbulent attractors, along with the edge state that governs the long-term dynamics on that boundary, are of fundamental interest. As the Reynolds number is increased from small values, a multiplicity of simple exact coherent structures (ECS) appear in phase space, of which one often undertakes the role of the edge state. At higher values of the Reynolds number, however, the dynamics on the basin boundary and the edge state are sensitive to initial conditions, as shown by a wealth of numerical and experimental studies. The mechanism behind this transition remains unclear. To address this, we examine the subcritical regime of Taylor-Couette flow using a minimal computational box and reveal a generic mechanism whereby the edge state becomes chaotic. The first step in this process involves the formation of a heteroclinic tangle between a travelling-wave-type ECS and an independently engendered chaotic saddle. The interaction causes the basin boundary to incorporate the saddle, thus inheriting its fractal structure, while the edge state itself remains the simple ECS. The transition of the edge state to chaos requires a second step: the formation of a heteroclinic cycle involving the ECS and the chaotic saddle. In consequence, the observation of a simple non-chaotic edge state at given values of the parameters is no guarantee that the same situation will hold at nearby values.
Introduction
Comprehending the physical mechanisms behind subcritical transition is a major unresolved challenge in fluid dynamics that dates back to the celebrated pipe flow experiments of Reynolds [46]. A puzzling phenomenon already observed in Reynolds’ apparatus, and in many subsequent studies of shear flows (e.g., [8, 48, 28, 17, 39] 2007), is that the occurrence of turbulent transition depends on both the shape and amplitude of the initial disturbance. This amplitude dependence evinces the need for a mathematical tool that reaches beyond the limited realm of linear stability analysis. Moreover, the strong sensitivity to disturbance shape hinders the study of subcritical transition through brute-force approaches relying solely on statistical analysis of data generated by experiments or direct numerical simulation.
Over the past two decades, dynamical systems theory has become an increasingly useful tool for addressing the subcritical transition problem [32, 15, 23]. The usual strategy is to analyse the system in phase space to identify the key invariant objects that regulate the dynamics, namely the attractors, saddles and connecting manifolds. As shown in Figure 1, two attractors typically coexist in phase space in the subcritical regime of shear flows: the linearly stable laminar solution and a nontrivial attractor. The latter attractor may be turbulence or, in the early subcritical regime, some precursor to turbulence. The manifold that separates the basins of attraction of the attractors (illustrated by the red curve) is known as the basin boundary. By definition, initial conditions on one side of this boundary decay to laminar flow, while those on the other side evolve towards the nontrivial attractor. Accordingly, understanding subcritical transition necessarily depends on elucidating the properties of the basin boundary.
Figure 1. A schematic of the phase space illustrating a typical subcritical transition problem, in which two attractors coexist.
When the nontrivial attractor is chaotic, like turbulence, the basin boundary is sometimes referred to as the edge of chaos [52]. The basin boundary is often depicted as a curve in two-dimensional sketches, but it is actually a hypersurface that has one less dimension than the full phase space. Visualising this hypersurface for high-dimensional systems is challenging, but it has been accomplished with some success by Skufca et al. [52] for a low dimensional model. Their method consisted in systematically varying the initial condition and computing the duration of the transients before laminar decay (also known as lifetimes; see also [48]). If the trajectory did not relaminarise after a pre-determined but sufficiently long integration time, the initial condition was regarded as appertaining to the basin of attraction of the nontrivial pseudo-attractor (i.e. mildly unstable saddle). Their results showed that the basin boundary hypersurface may develop a large number of folds, thus introducing critical sensitivity to initial conditions.
Lifetime studies have been widely used in the investigation of subcritical shear flows, in particular to characterise the sensitivity to initial conditions and to reveal the apparent fractal structure of the phase space. For example, Kreilos et al. [34] demonstrated that similar representations as Skufca et al. [52] could be generated for plane Couette flow using appropriate phase map projections. Their findings suggest that the phase space may have a fractal structure, but its self-similar nature is yet to be confirmed. The parameter dependence of lifetimes was also explored by Eckhardt & Mersmann [14] and Moehlis et al. [43] for simple model problems. The promising results of these preliminary studies soon sparked numerical research on pipe flow, such as that by Schneider et al. [51]. Around the same time, experiments ([45, 29]) and numerical simulations ([17, 62]) showed that the reverse transition – the natural decay to laminar flow from the transitional chaotic state – occurs invariably in the early transitional regime. Interestingly, the decay process follows a memoryless (Poisson) probabilistic distribution, providing further evidence of the phase space complexity. For the latest developments of the transition problem in pipes, readers are referred to the review article by Avila et al. [2].
Generalising the notion of a basin boundary to a pseudo-attractor is convenient in shear flow turbulence, because at moderate Reynolds numbers it may decay after extremely long times. Indeed, for the cases studied in [17, 45, 29, 62], the laminar state is very likely the sole attractor. For symmetry restricted flows at sufficiently low values of the Reynolds number, however, the existence of a non-trivial attractor is sometimes relatively easy to ascertain and a basin boundary can be defined rigorously ([38, 33, 3, 63, 47]).
A widespread method for systematically analysing the basin boundary is the computation of edge states. A key observation in mathematically defining an edge state is that any trajectory starting exactly on the basin boundary will remain on it indefinitely (the basin boundary is an invariant set). Edge states are attractors of the dynamical system that arise when the dynamics are restricted to the basin boundary. In practice, an edge state emerges as the final state of trajectories when the initial condition is controlled to deliberately inhibit evolution toward either full-space attractor. The first implementation of such a control algorithm, known as the edge-tracking technique, in a subcritical shear flow problem was presented by Itano & Toh [30]. Edge tracking in small periodic domains, typically of the size of the minimal flow unit (i.e. the smallest periodic domain in which turbulence is sustained, see [31, 26]) often converges to simple solutions such as equilibria, travelling waves or limit cycles, depending on the shear flow under scrutiny. This highlights the importance of non-chaotic solutions, known as exact coherent structures (ECS), in understanding the subcritical transition problem ([44, 7, 16, 60]) and explains why they are observed repeatedly, if only briefly, in carefully controlled experiments ([27, 9]).
A necessary condition for an ECS to be an edge state is that it has a single unstable eigenvalue, the one that pushes trajectories away from the basin boundary. ECS with such stability properties typically arise from saddle-node bifurcations. When this occurs, the nodal (or upper branch) solution forms the backbone of the non-trivial attractor, while the saddle (or lower branch) solution has a single unstable eigenvalue and acts as the edge state; see Wang et al. [59] who studied plane Couette flow in the minimal box, for example. Although the lower branch solution tends to retain a single unstable eigenvalue up to increasingly high Reynolds numbers, a behaviour that is consistent with asymptotic analyses ([11]), there is no guarantee that the ECS will maintain its role as an edge state indefinitely. Indeed, given the sensitivity to initial conditions that is inherent in prototypical subcritical transition, it seems unlikely that a simple ECS should remain the edge state at sufficiently high Reynolds numbers. If the edge state were to become chaotic, which mechanisms could be held responsible? The existence of chaotic edge states has indeed been confirmed in shear flows (see, e.g., Schneider & Eckhardt [49]).
All of the complex scenarios of subcritical transition described above involve a chaotic saddle, i.e. an unstable chaotic set in phase space. Chaotic saddles can play different roles in phase space, and we classify them comprehensively into three mutually exclusive types: (i) those not on the basin boundary, (ii) those on the basin boundary but not an edge state, and (iii) those that are edge states, as illustrated in Figure 1. As already noted, when turbulence is only transient, a true basin boundary does not exist, and only type (i) is possible. Even in this case, edge tracking can converge to a saddle solution, which has only one unstable direction (Schneider & Eckhardt [50]). Strictly speaking, this is not an edge state, in the sense that it is not attached to any basin boundary. However, since a chaotic state that persists beyond feasible time horizons can be considered as an attractor for all practical purposes, the outcome of edge tracking effectively behaves as an edge state attached to the basin of attraction of that pseudo-attractor ([52, 15]). While such generalisations of edge states are now common in shear flow studies, in this paper we adopt the strict definition.
Type (i) chaotic saddles typically emerge from the collision of a chaotic attractor with its basin boundary in a global bifurcation known as boundary crisis ([24, 25]). The collision usually involves a saddle that was contained in the basin boundary and may have even acted as an edge state prior to the crisis. In plane Couette flow, a chaotic attractor arising from a period doubling cascade undergoes a boundary crisis that turns it into a type (i) chaotic saddle ([33, 3, 63]). The transient turbulence observed in pipe flow at low Reynolds numbers ([45]) is also type (i), and its origin would be related to a boundary crisis akin to that identified by Avila et al. [3]; Ritter et al. [47]. Thus far, the mechanisms by which chaotic saddles of type (ii) or (iii) appear has not been explored.
It is noteworthy that Lustro et al. [36] and Budanur et al. [6] recently applied the homoclinic tangency theory to plane Couette and pipe flow, respectively. To fully appreciate the motivation behind these studies, a profound understanding of discrete dynamical systems theory is required. Unlike in continuous dynamical systems, the unstable and stable manifolds of fixed points can intersect transversally in discrete systems. The presence of such a transversal homoclinic point entails the existence of infinitely many homoclinic points forming a homoclinic tangle. This, in turn, leads to chaos on a certain Cantor set, as described by the Smale-Birkhoff theorem ([4, 53]). The theory of discrete dynamical systems can be applied to shear flows by defining an appropriate Poincaré section in phase space. In the light of the numerical results of Lustro et al. [36], the tangency they identified likely relates to the emergence of a chaotic saddle of either type (ii) or (iii). However, their analysis did not include a detailed investigation of this aspect.
The aim of this paper is to demonstrate clearly and directly how fractal structures and memoryless processes can emerge at the basin boundary and produce chaotic edge states in subcritical shear flows. With this goal in mind, we select a flow configuration that favours the appearance of a chaotic set that sequentially evolves from type (i) to type (ii) and, finally, to type (iii). Specifically, we focus on the subcritical parameter region of Taylor-Couette flow investigated by Wang et al. [55]. At the root of the analysis is the stabilisation of an ECS (a drifting wave originally found by Deguchi et al. [12], when considered in the parallelogram-annular domain of Deguchi & Altmeyer [10]).
The paper is structured as follows. In § 2, we set out to review relevant known results of the system and present the full bifurcation diagram. The various types of temporal dynamics observed at selected values of the Reynolds number are presented in order to motivate the analysis that follows. Return map analysis is then employed in § 3 to explore how a type (i) chaotic saddle emerges and acquires a fractal structure in phase space. Then § 4 uncovers how the chaotic saddle merges with the basin boundary and becomes type (ii). Its eventual transition into a type (iii) saddle, i.e. a chaotic edge state, is addressed in § 5. Finally, we summarise the key findings and draw conclusions in § 6.
Problem statement, bifurcation diagram and typical time series
We consider the Taylor-Couette flow setup, taking the radial gap and as units for length and time, respectively. Here, and are the dimensional inner and outer cylinder radii, respectively, and is the kinematic viscosity. The cylinders rotate with angular velocities and . The flow obeys the incompressible Navier–Stokes equations, which we express in cylindrical coordinates . A zero axial mass flux condition is imposed to mimic the experimental conditions in a numerical setup of infinite axial extent. We fix the radius ratio and the outer cylinder Reynolds number following previous studies (Meseguer et al. 2009a,b; Deguchi et al. 2014; Wang et al. 2022, 2025a,b). The only parameter we vary throughout the study is the inner cylinder Reynolds number .
The total velocity field is decomposed into a base flow and a perturbation . The base state of the system is the circular Couette flow (CCF) solution, which has a well-known closed-form expression
with and .
The numerical methods used to solve the governing equations were thoroughly presented in Wang et al. [55], to which we refer the reader for details. The computational domain is of parallelogram-annular shape and periodic in the parallelogram directions . The spatial discretisation for the direct numerical simulations (DNS) employs Chebyshev–Fourier–Fourier spectral methods, and the spatial resolution of modes deployed by Wang et al. [56] has been confirmed as sufficient and kept. To compute and track unstable solutions in parameter space, we use a Poincaré–Newton–Krylov method, duly coupled with the method of slices [22, 5, 58] to cope with drift.
We characterise flow states by the normalised kinetic energy of the perturbation velocity and by the corresponding inner and outer cylinders normalised torque, and ,
where
is the volume-averaged kinetic energy of some velocity field . The volume of the computational domain is , and implies averaging in both parallelogram directions. With these definitions, and for CCF.
We start by summarising in § 2.1 what is known regarding the onset of chaos in counter-rotating Taylor-Couette flow. We then extend the bifurcation diagram of Wang et al. [56] to include new solution branches and investigate their role in the transition process and the morphing of phase space with varying . In § 2.2 we present a collection of time series at selected values of to illustrate the most relevant dynamical phenomena encountered along the complex bifurcation scenario. The key critical values of the Reynolds number for which the qualitative changes in the time series occur are briefly commented in § 2.3.
The onset of chaos as reported in previous studies
Figure 2a shows a schematic representation of the transition diagram in – parameter space, with the grey vertical line indicating an -sweep at constant for , which is the focus of our study. CCF is linearly unstable above the linear stability boundary (orange line), which is crossed at in our exploration. The solid black curve encapsulates a region of spatio-temporal intermittency, characterised by a flow consisting of a turbulent helical band coherently sustained over a laminar background flow, the spiral turbulence regime (Figure 2b). Wang et al. [55] found a finite-amplitude drifting rotating wave (DRW) that exists for . This solution, computed in the small parallelogram domain shown in Figure 2c, persists in the intermittency region and appears to play an important role in the formation and sustainment of spiral turbulence. The use of this specific parallelogram domain, with two of its sides aligned with the tilt of the turbulent spiral, has been instrumental in our analysis. As evident from the bifurcation diagram in Figure 2d, the upper branch of DRW (the one with higher torque) is stable at onset () in this particular domain. This stability is briefly preserved up to the Hopf bifurcation point H (), whence a family of stable relative periodic orbits (P, blue loops) is issued supercritically. To better convey the time periodic nature of P, we have included a third axis, , in the bifurcation diagram, such that cross-sections at constant correspond to phase map projections on the – plane. Since relative equilibria must satisfy , the DRW solution branch lies within the plane (grey plane). To simplify the ensuing analysis, we introduce a Poincaré section
Figure 2. Summary of results from Wang et al. [55]. (a) Transition diagram for counter-rotating Taylor–Couette flow with radius ratio . The bullets on the line indicate the parameter values used for flow field visualisations in panels (b) and (c). (b) Spiral turbulence at visualised by radial vorticity at . The dimensionless height of the domain is . (c) DRW at SN () visualised in the same format as panel (b), but with . The white parallelogram shows the computational domain. (d) Bifurcation diagram at . SN and H denote saddle-node and Hopf bifurcation points, respectively. The black curve represents DRW (solid for the stable portion, dashed when unstable). The P relative periodic orbit is represented by the blue closed curves at selected values of . Its representation on the Poincaré section is given by the intersection of the loops with the grey plane.
where is the set of all possible , representing velocity fields modulo spatial drifts. Relative periodic orbits are represented by discrete points on . In particular, the P solution branch is represented by the blue curve within the grey plane.
Figure 3 depicts a bifurcation diagram summarising all solutions involved in the transition process, as viewed on . Since on , we dispense with the subscripts in the torque (i.e. ). The P solution (blue) of Figure 2d undergoes a Feigenbaum cascade (green), as found by Wang et al. [56]. PD indicates the first period-doubling bifurcation as is increased, where P loses stability and a new periodic solution P of about twice the period of P emerges. The new stable solution crosses the Poincaré section twice every period, hence the subscript and its representation as two branches in the figure. At somewhat larger , P exhibits another period doubling bifurcation, PD, leading to P, which crosses at four distinct points. A period doubling cascade follows as is increased further, with periodic orbits of ever increasing period accumulating at ([56]), in accordance with Feigenbaum’s self-similar scenario ([18, 19, 20, 21]). From this point onwards long term dynamics become unpredictable, following the emergence of a chaotic attractor (cloud of green dots beyond in Figure 3). We shall denote the chaotic set originating from P as C. Strictly speaking, the cascade starts from P, but we adopt the subscript 2 because P plays a more significant role in the -range of interest. Wang et al. [56] and Wang et al. [57] demonstrated that the chaotic attractor is embedded in a nearly one-dimensional manifold in . This particularly low dimensionality is precisely the key enabling feature of our analysis. The unexplained features in the figure are findings specific to the present study and will be discussed later.
Figure 3. Bifurcation diagram constructed from points on . At , the orbit (blue line) undergoes a period-doubling bifurcation producing , which subsequently initiates a cascade leading to the chaotic set beyond the accumulation point (, the cascade, and are all shown in green). The orbit (dashed red line) emanates from a cyclic fold . Its upper branch (), exhibits another period-doubling cascade (magenta) generating the chaotic set . The black dots represent transient chaotic data initiated from , collected before the departure towards (see Figure 6b). The black dashed curve represents the lower branch of the drifting rotating wave (). The grey vertical lines at , and indicate global bifurcation points. The shaded region indicates the range of between the first heteroclinic bifurcation point () and the homoclinic bifurcation point () shown in Figure 5a.
In the following, we will examine the dynamics employing various types of diagrams, all of which are exemplified in Figure 4 at the accumulation point of the cascade . The green curve in Figure 4a shows the time series of the chaotic attractor as obtained with DNS. Points on are represented with green bullets. While time signals are valuable for their simplicity, projecting both the phase space trajectory and onto the three-dimensional space spanned by , as done in Figure 4b, reveals their relationship in a more clear way. Collecting the torque at consecutive crossings in an ordered sequence , and plotting each value as a function of the one preceding it, yields the green points in the return map of Figure 4c. The map appears quasi-one-dimensional, as revealed by the transients (black points) obtained from perturbations of the ‘period-3 orbit’ identified by Wang et al. [57] (red triangles). We refer to this periodic orbit as , and represent it with a red line in Figure 4a and Figure 4b, its three crossings shown as red triangles. While the details of will be discussed in the next section, the rationale behind its naming should already be clear from Figure 4b. We shall shortly see that has upper and lower branches, the one depicted in Figure 4 corresponding to the lower branch .
Figure 4. Dynamics at the accumulation point . (a) Time series of for the chaotic attractor obtained with DNS (green) and for P (red). Symbols indicate points on . The P solution here corresponds to P in Figure 3, which is unstable and has been computed with the Poincaré Newton-Krylov method. (b) Projection of the phase space onto the coordinates. The grey plane indicates . (c) Iteration map constructed from the torque sequence recorded on . The black dots represent transient dynamics obtained by slightly perturbing P.
Crisis bifurcations and the origin of the P3 solution family
Now let us briefly outline some of the new findings already apparent in the bifurcation diagram of Figure 3. The chaotic attractor generated by the Feigenbaum cascade loses stability at the critical Reynolds number . At this point, the chaotic attractor (green points) collides with the (red dashed curve) in a global bifurcation known as boundary crisis in dynamical systems theory ([24, 25]). The phenomenon is clearly observed in the bifurcation diagram thanks to the appropriate choice of .
To find the origin of P, we have tracked the branch to lower values of . The branch has a turning point at FC, where a fold bifurcation of limit cycles occurs. We refer to the solutions before and after this turning point as P (lower branch) and P (upper branch), respectively, the former being the solution responsible for the boundary crisis at . Remarkably, P is stable in the vicinity of FC and triggers its own period doubling cascade as is increased (see the inset of Figure 3). The resulting chaotic attractor (C, magenta points) eventually collides with P at , also in a boundary crisis. When exceeds this value, the chaotic state can no longer be sustained indefinitely and trajectories are ultimately attracted to P, which is still stable in that region of parameter space. A representation of the resulting transient chaos can be obtained by initialising DNS simulations with slight perturbations of P and collecting the initial transients (black dots in the bifurcation diagram, corresponding to multiple runs with varying initial perturbations of P; note that they do not form an invariant set and are therefore not a chaotic saddle as defined in dynamical systems theory). As is increased, the transient dynamics encompass an increasingly larger portion of the phase space until colliding with DRW (dashed black line in Figure 3) at .
The transient chaos does not disappear at this point, but its qualitative properties change, as will be clarified later in § 2.3.
At much larger , the P branch in Figure 3 undergoes another cyclic fold bifurcation, FC, at . Shortly after the turn, the P branch collides with DRW and disappears. This critical point is denoted as in the enlarged view of the bifurcation diagram shown in Figure 5a, and corresponds to the formation of a homoclinic orbit from DRW to itself. Figure 5b compares the time series along the unstable manifold of DRW at (dashed black line), i.e. the approximate homoclinic orbit, and the P solution (solid red) calculated at the closest attainable value of the Reynolds number. The excellent match between the time series provides strong evidence for the occurrence of a homoclinic bifurcation involving P and DRW. Note that Wang et al. [55] confirmed that DRW has only one real unstable eigenvalue and, therefore, its unstable manifold can be easily approximated using DNS.
Figure 5. Disappearance of P in a homoclinic bifurcation. (a) Enlarged view of the bifurcation diagram (Figure 3) near the homoclinic bifurcation point . (b) Time series at of the unstable manifold of DRW (dashed black) and the P solution (solid red) at the closest attainable .
Changes in the dynamics due to global bifurcations
In general, global bifurcations such as boundary crises induce a significant transformation of the phase space topology. Among the critical Reynolds numbers discussed so far, produces the most obvious alteration, as the dynamics on C cease to be permanent and all initial perturbations eventually decay to CCF, the sole remaining attractor. As a consequence, C can be straightforwardly classified as a type (i) chaotic saddle, similar to that found by Budanur et al. [6]. Relaminarision events of saddle C, particularly in the vicinity of , are often preceded by dynamics that are reminiscent of the P orbit, an observation that motivated its discovery in the first place (Wang et al. [57]). However, a connection between P and CCF is not immediately apparent from the bifurcation diagram of Figure 3, which shows no direct branching between the two solutions. The relation, of a dynamical nature, will become clear in section 5.
The bifurcation diagram of Figure 3 shows all stable states along with the most relevant unstable solution branches, i.e. DRW (dashed black), the unstable portion of P (dashed red) and the chaotic transients generated by the C chaotic saddle (cloud of black dots). In the parameter region of interest, from PD to , there are at least two attractors in the phase space. One is CCF, which exists and is stable all along, and the other is either P (up to PD) or the cascade that follows. In the interval between FC1 and , however, the attractor count rises to three. The third attractor is initially , then the cascade it spawns. In order to gain a qualitative understanding of the phase space changes, it is useful to monitor trajectories starting from the close neighbourhoods of and . From such trajectories, computed with DNS, the following observations can be readily made:
(a) For , trajectories starting from close to either lead to (grey circles in Figure 6a) or evolve towards the chaotic attractor (black crosses). On the other hand, starting from the vicinity of , the choice is between CCF (black line in Figure 7a) and (grey line). The edge tracking algorithm confirms that both and act as edge states at this stage, as will be argued later.
Figure 6. Time series of corresponding to DNS started from small random perturbations of at (a) , (b) and (c) . Only two time series of the many run are shown in each panel to illustrate the two qualitatively different behaviours observed at each of the values of . Panels (a) and (b) only depict the values of on , while (c) superposes also the continuous time series to capture the decay towards CCF. Perturbation amplitudes are of order relative to .
Figure 7. Unstable manifold of at (a) and (b) . The manifolds are closely approximated with DNS started from very slight perturbations of to either side in the direction of its only unstable eigenfunction. Perturbation amplitudes are of order relative to .
(b) For , has become a chaotic saddle and the attractor count is reduced to two. As illustrated by Figure 6b, trajectories starting from close to can only reach , but two different paths may be observed. The departure towards can either be uneventful (grey) or transiently visit prior to relaxing on (black). Since the transient dynamics of lead always to , the chaotic saddle is type (i). The clouds of black points in Figure 3 and Figure 5a depict transient visits to obtained from multiple DNS simulations started from perturbations of . The duration of these transients is highly sensitive to initial conditions; we shall see later that this is a consequence of the fractal structure of in phase space. As before, starting from close to may still lead to CCF or to , but now is the only solution acting as an edge state.
(c) For , CCF has become within reach of trajectories starting from the vicinity of . The dynamics can still end on , either directly (grey line and circles in Figure 6c) or after transiently approaching (not shown), or can alternatively decay to CCF after short visits to (black). We will see in section 3 that the chaotic saddle is now in the boundary separating the basins of attraction of CCF and , but edge tracking invariably converges on , which remains the sole edge state. Thus, is not an edge state and is therefore classified as type (ii).
(d) For sufficiently large , the route from to has been severed altogether. Figure 7b shows typical trajectories starting in the vicinity of at the accumulation point . Perturbing in one or the other direction of its one-dimensional unstable manifold always results in laminar decay, sometimes uneventfully (black line), sometimes after some transients (grey). Despite still having a single unstable eigenvalue, is no longer an edge state. A natural question arises as to when and how the qualitative change in the unstable manifold of , from Figure 7a to Figure 7b, occurs. We will show later that this happens at indicated in Figure 5a. For , the edge state becomes chaotic, which, according to our classification, corresponds to a type (iii) chaotic saddle.
The emergence of the chaotic saddle
As noted in the previous section, the first appearance of transient chaos occurs at . The main aim of this section is to examine the structure of the underlying chaotic saddle and to elucidate its emergence. The starting point is Figure 8a, which shows the representation on of exact solutions (see the legend) and both permanent (magenta dots) and transient (black dots) chaotic trajectories at . The magenta points represent the chaotic attractor at . These points have been obtained by evolving the black trajectory in Figure 6a for a long time and discarding the initial transients. Other key invariant objects are indicated with symbols in Figure 8a. In particular, (down-pointing red triangle) is located on one edge of the chaotic attractor, while (up-pointing red triangle) is at its
Figure 8. The transition of from attractor to saddle around . All data is recorded on . The magenta points represent the attractor obtained from DNS at , while the symbols shown in the legend correspond to Poincaré-Newton-Krylov solutions. The black dots show the transient dynamics obtained by perturbing at . (a) Projection of the points onto the – plane. (b) First return map of the torque sequence. (c) Sixth return map. The region enclosed by the dashed box is magnified in panel (d).
core. The black dots in Figure 8a denote transient chaos at . They have been obtained from many DNS runs initialised with small random perturbations of , one example of which is shown in Figure 6b. Roughly speaking, the chaotic saddle at exists in the region where the black and magenta points overlap (see inset of Figure 8a). Note that the black and grey trajectories of Figure 6a and Figure 6b correspond, respectively, to perturbations of toward and away from the chaotic set .
Figure 8a reveals that the discrete dynamics on evolve on a nearly one-dimensional manifold of negligible thickness. This motivates the construction of the return map shown in Figure 8b. The function relating consecutive torques on , defined by , is multi-valued, but a straightforward analysis reveals that the branch selection rules are univocal. In the following, we first dissect in §3.1 the branch selection rules of the return map using the period-doubling bifurcation starting from as a guide, before revisiting the global bifurcations that turn into a saddle in §3.2 and §3.3.
Structure of the return map and period-doubling originated from P2
Figure 9a shows a sketch of the map , with the two distinct branches labelled as A and B. The blue arrows indicate how acts on points of the map. Points corresponding to a periodic orbit form a closed loop as illustrated for (triangles). Meanwhile, maps the local minimum of branch A (black circle on top of the red plus sign) onto the point where branch B bifurcates from branch A (orange circle). In general, application of to the branching point (black circle) does not map back onto the minimum of branch A (red plus sign), as exemplified, at the same , in Figure 9b. The figure also shows the action of on the two branches as a whole. Branch B (dashed orange) maps onto a portion of branch A (solid black), while branch A maps onto the union of branch B and the portion of itself to the right of the branching point. The immediate implication of this, is that predicting precisely which branch the trajectory will be on after iterations is troublesome when is large.
Figure 9. Sketch of the iteration maps. Panels (a) and (b) are based on the data from Figure 8b. The branches A and B labeled in panel (a) are mapped under as illustrated by panel (b). Downward triangles represent , and the arrows indicate the action of . Panels (c) and (d) show the return maps when and are superstable, respectively.
An exception to this occurs when the minimum of the map (red plus sign) is a point of a periodic orbit, in which case the orbit is said to be superstable (the multiplier is zero). Figure 9c shows an example for a superstable . In this case, the map can be partitioned at the periodic points of , so that all the points on a partition are mapped by onto a partition or a combination of partitions. The map is said to admit a Markov partition. Figure 9d is a sketch of the map when is superstable, this being another case for which a Markov partition is possible. The cobweb of successive superstable orbits helps better understand the action of , as well as how the period-doubling bifurcations accumulate toward the limiting case shown in Figure 4c. The figures also confirm that the multivalued nature of does not lead to any peculiar behaviour in the period-doubling cascade, as the map consistently behaves like a quadratic function near the local minimum of branch A.
Return map analysis around
Let us now examine the flow dynamics around . As clear from Figure 6b, the and orbits play key roles in the dynamics. This observation motivates the depiction of the 6th return map in Figure 8c, making use of the fact that the least common multiple of 2 and 3 is 6 and, therefore, , and are all fixed points of .
Figure 8d is a close-up of the region enclosed by the dashed box in Figure 8c. Since is single-valued in this region, the usual cobweb analysis can be readily applied. The W-shaped magenta structure corresponds to the chaotic attractor that exists at .
The shape of the map is similar to the second iteration of the logistic map, which suggests that a local unimodal structure must have appeared in , as is the case indeed. Returning to the original map of Figure 8b, the unimodal structure originates from the lowermost cluster of magenta points. That is, the period doubling cascade generating can be traced back to the local minimum of the branch labeled B in Figure 9a.
The shaded square enclosing in Figure 8d has been defined using the furthermost point that maps onto . Cobweb analysis shows that, for , any trajectory started from within the square remains in the square. In other words, no trajectory can escape the shaded region. is embedded in the chaotic attractor, whereas small perturbations to , located at a corner of the square, can lead to either the chaotic attractor or , depending on whether they fall within the square or outside. This is consistent with observation (a) in § 2.3.
The critical point corresponds to the situation in which the local extrema of touch the shaded square. For , the W-shaped segment of the map is fully contained within the square. When the local extrema of the W-shaped map extend beyond the square region for , an escape pathway toward is formed, as can be seen from the cobweb analysis. The black dots in Figure 8 depict a collection of trajectories escaping from the square at . Notice, however, that these trajectories are not on the chaotic saddle itself but represent instead transient dynamics in its vicinity that eventually lead to , as will be clarified in § 3.3.
Fractality of the chaotic saddle
Figure 10 shows the same diagrams as Figure 8 but at the higher Reynolds number (the same value used in Figure 6c). Figure 10a and Figure 10b are qualitatively similar to Figure 8a and Figure 8b. However, as evident from Figure 10d, trajectories can escape the square region more easily (a larger proportion of points maps outside) at this parameter value than at depicted in Figure 8d. This makes the higher value of more convenient in analysing the properties and nature of the chaotic saddle.

Figure 10. The phase space at . Panels, styles and colours as for Figure 8.
Figure 11a provides an enlarged view of the square region in Figure 10d. The DNS data (black dots) has been interpolated with splines (black solid lines) for clarity and the resulting one-dimensional function named . The spline representation evinces that our DNS computations embody a textbook example of a one-dimensional chaotic saddle, as described in standard dynamical systems literature. Let us denote by the full range of torque values encompassed by the square region (see bottom of Figure 11a). Points in the set remain within the square for at least one iteration of . In other words, and correspond to the range and domain of , respectively. Naturally, consists of four intervals because consists of four disjoint segments. Initial points chosen in may escape the square under , since points in may escape under . To ensure that the trajectory stays inside the square for two iterations of the map, the initial value must be taken within . The set consists of subintervals. Similarly, to ensure that a point remains in the square under , the initial value must be chosen from . In the figure, each of the 16 apparent bands consists in actuality of a set of 4 sub-bands, resulting in a total of 64 subintervals (see the enlarged view in the red box). The chaotic saddle is the invariant set of all points that remain indefinitely within the square, and can therefore be formally expressed as , where . Each set consists of subintervals and provides an increasingly accurate approximation of the chaotic saddle as increases. As evident from its construction, the chaotic saddle is a Cantor set. To the best of the authors’ knowledge, this is the first direct demonstration of the fractal nature of a chaotic saddle in a shear flow.

Figure 11. Analysis of the chaotic saddle at . (a) Magnified view of the shaded square region in Figure 10d. The DNS data (black dots) has been interpolated with splines (function , black solid lines). The up- and down-pointing triangles denote and , respectively. The intervals below the map correspond to , where is the full range of . (b) Survival probability of the points (uniformly distributed in ) as a function of the number of iterates taken of . (c) Survival rate of points from one iterate to the next.
We can now estimate the escape rate away from the saddle by taking a large value of initial conditions in (e.g. uniformly distributed), and then counting the amount of points that remain in after the th iterative application of . The logarithm of drops linearly (Figure 11b), pointing at a nearly constant escape rate from one iteration to the next (Figure 11c), which is a hallmark of a Poisson distribution.
It should also be noted that, according to dynamical systems theory, a one-dimensional chaotic map such as that in Figure 11a exhibits topological transitivity [13]. In plain words, this means that a trajectory starting in an arbitrarily small neighbourhood of any point in the chaotic set will eventually reach an arbitrarily small neighbourhood of any other point in the set. Topological transitivity in our map implies that from the neighbourhood of , one can reach any neighbourhood of . For this reason, we were justified in choosing as the starting point for the investigation of the chaotic saddle (as done, e.g., to produce Figure 6).
The emergence of chaos in the basin boundary
We now turn our attention to the transition of the chaotic saddle from type (i) to type (ii) at the critical parameter . The basin boundary becomes chaotic at this point but, as we will see in § 4.1, remains the edge state.
To address the advent of chaos in the basin boundary and the later emergence of a chaotic edge state, we will exploit the theory of homoclinic and heteroclinic tangencies, which occur when manifolds in phase space come into contact. This theory has the advantage of greater generality than the simpler one-dimensional discrete map approximation, though the ensuing analysis tends to be more intricate. As an example, the period-doubling bifurcations leading to (see Figure 9) can also be interpreted as arising from homoclinic tangencies between the stable and unstable manifolds of . These tangencies ultimately result in chaos according to the Smale-Birkhoff theorem. In dynamical systems theory, the relationship between one-dimensional dynamics and manifold tangencies is often illustrated using the logistic and Hénon maps [54, 13]. The phenomena we observe here are analogous to those reported for these classical examples.
As we will see in §4.2, the dynamics in the neighbourhood of cannot be fully grasped with the one-dimensional discrete map approximation, hence the need for tangency theory.
Edge tracking at
We will now show that, although having been subsumed within the basin boundary, is still not an edge state immediately after . We will take advantage of its topological transitivity, as noted in § 3.3, to edge-track between trajectories originated in a vicinity of .
Figure 12a illustrates the result of edge tracking at , well above (see Figure 3 and Figure 5a), starting from a close neighbourhood of . The algorithm has been initiated by selecting two initial conditions in this neighbourhood, one decaying onto CCF (black) and the other departing towards (grey). Then, the standard process of alternating trajectory bisection and time-advance has been implemented to compute the dynamics on the basin boundary within prescribed tolerance. The two trajectories evolve together for as long as the edge tracking algorithm is kept on and, as expected, diverge thereafter. It is clear, however, that keeping the algorithm on for sufficiently long has the dynamics settle on a plateau ( in Figure 12a). The converged solution is , which remains the edge state at this value of despite the significant reconfiguration of the basin boundary.

Figure 12. Edge state analysis at . (a) Two converged bounding trajectories starting from the neighbourhood of and eventually leaving towards CCF (black) and (grey). The edge tracking algorithm is enforced up to to guarantee convergence onto the edge state, DRW (plateau), and then turned off. (b) Three dimensional projection of the trajectories on space while edge tracking is on (blue), up to ). Also shown are (red loop) and DRW (square). (c) Projection of the trajectories once edge tracking is turned off for (yellow lines). (green loop) and CCF (black circle) are also shown for reference.
Figure 12b shows the same edge tracking up to (blue line), projected onto the three-dimensional phase space . The two bounding trajectories start very close to (red closed loop) and approach (black square) before parting (yellow arrows) for . According to the definition in dynamical systems theory, the edge tracking trajectory is contained in the stable manifold of , , with the final approach approximating the least stable of its eigenmodes.
The yellow lines in Figure 12c correspond to the same phase map representation but for the diverging portion of the trajectories in Figure 12a, which result from turning off the edge tracking algorithm for . These trajectories start as close approximations of the one-dimensional unstable manifold of , .
Note that, despite having only one unstable eigenvalue, is not an edge state according to the definition given in § 1. As noted there, states possessing a single unstable eigenvalue can, in most cases, be computed through edge tracking by exploiting the different transient
behaviour of trajectories issued on either side of their unstable manifold as the separation criterion. The periodic orbit , however, provides an example that edge tracking offers no general guarantee of convergence for all solutions featuring one-dimensional unstable manifolds.
Heteroclinic tangency at
Let us revisit the return maps in Figure 10b-Figure 10d, plotted at . Recall that the black dots mark points on the unstable manifold of , i.e., . The global bifurcation at can be ascribed to a heteroclinic tangency between and . Indeed, the analysis of Figure 10d provides compelling evidence for a manifold tangency: the cob-web trajectories linking the local minima of the map to their image lands in close proximity of , as indicated by the arrow.
Figure 13 presents sketches of the phase space on near the critical parameter , constructed from the results obtained so far. The white and grey regions represent the basins of attraction of and CCF, respectively. They are separated by , with acting as the edge state. As noted earlier, possesses only one unstable eigenvalue, so forms a one-dimensional manifold in the phase space , whose two branches are directed toward CCF and at either side of (see Figure 12c). Although belongs in , its unstable manifold (orange line) has no representation on it, hence its depiction as dashed in Figure 13. It immediately departs from and does not intersect it again until after reaching the neighbourhood of on one side and converging on CCF on the other.

Figure 13. Sketch of the phase space near the critical Reynolds number . Each periodic orbit is represented by its intersection with . The unstable manifold , which does not lie in , is shown using dashed curves. (a) Before the tangency (). (b) After the tangency ().
Figure 13a and Figure 13b illustrate the phase space for and , respectively. In the former case, trajectories originating from a neighbourhood of always depart towards , implying that must lie entirely within the basin of attraction of . At the critical parameter, a heteroclinic tangency occurs between and , which is followed by a manifold tangle (see Figure 13b). Note that the occurrence of an intersection between the manifolds implies the existence of infinitely many intersections (heteroclinic points), which is why the manifolds must appear tangled [35, 61]. The effect of this tangency extends throughout the entire chaotic saddle due to its topological transitivity. It is evident that after the tangency, the basin boundary becomes extremely intricate, exhibiting fractal geometry. Transitions from this region therefore show strong sensitivity to initial conditions.
Comparison to plane Couette flow
The above results up to share certain features with the transition scenario in plane Couette flow reported by Lustro et al. [36]. This suggests some degree of genericity in the transition process of shear flows. For example, Figure 4 in Lustro et al. [36] also shows two tangencies occurring at the critical Reynolds numbers and . At the first tangency, the upper branch of a solution called gives rise to a chaotic attractor that makes contact with the lower branch, producing a chaotic saddle. This chaotic saddle is type (i) by our definition, and the tangency at is akin to our observation at . At the second tangency, the chaotic saddle seems to collide with the lower branch of another pre-existing solution, , which is the edge state known as the gentle periodic orbit (GPO). This scenario closely resembles the tangency we observe at . The presence of these shared global bifurcations suggests some degree of genericity in the transition process of shear flows.
We must, however, highlight a significant difference between the results by Lustro et al. [36] and ours: while the edge state was a periodic orbit (the GPO) there, it is a travelling wave () here. Recall that we specifically designed the Poincaré section to contain travelling waves such as . However, travelling waves are singular points on where the Poincaré map cannot be properly defined. This renders the theoretical analysis harder in our case, and for many other subcritical shear flows featuring steady or travelling-wave edge states (see § 1).
Another important difference is that Lustro et al. [36] reported both homoclinic and heteroclinic tangencies near , while we have only found a heteroclinic tangency at . As sketched in Figure 13, a homoclinic orbit from to itself is clearly impossible, since there is no pathway from the edge state to the chaotic saddle. This route emerges only at higher , as we will show in § 5. It is likely that, in Lustro et al. [36], the opening of the pathway precedes .
The emergence of a chaotic edge state
As evident from Figure 12a, remains an edge state at least up to , with its stable manifold separating the basins of attraction of and CCF. At somewhat larger , however, all trajectories starting near are eventually captured by CCF (see Figure 7b). In § 5.1 we shall show that this qualitative change in dynamics occurs at (see Figure 5a). The replacement edge state is then analysed in § 5.2 at just above this critical point.
Heteroclinic cycle at
Figures 14a and 14b track in its two directions, employing DNS at and , respectively. As approaches from below, the time series that eventually leads to stays for an increasingly long lapse shadowing (of which one period is shown with a red dashed curve for reference). The same shadowing occurs when decreasing towards from above for one of the trajectories, even if its final destination is now CCF. As a result, the 3rd, 6th, and 9th crossings in Figure 14a,Figure 14b are graphically indistinguishable from the corresponding values for . This is a sign that a (non-robust) heteroclinic connection from to occurs at the critical point, . The yellow curve in Figure 14c shows a three-dimensional phase map projection of the same time series that visits in Figure 14b, up to . This trajectory provides a good approximation of the heteroclinic orbit, i.e., the intersection of and , which should actually happen at .

Figure 14. The heteroclinic bifurcation at . Two trajectories started from a close vicinity of and leaving in the two different directions of are shown at (a) and (b) . The triangles indicate the 3rd, 6th and 9th Poincaré crossings, all three approaching (red dashed) at very close quarters. (c) Phase map representation at . The orange trajectories are , while the red orbit is . (d) Sketch of the phase space at exactly .
Recall that the robust heteroclinic connection from back to exists for (blue line in Figure 12b), since lies on the basin boundary. The emergence of the non-robust heteroclinic orbit from to at (yellow line in Figure 14c) generates a heteroclinic cycle. This heteroclinic cycle has been sketched in Figure 14d using a format similar to that of Figure 13.
We have analysed in Figure 15a the dependence of on by recording the 3rd, 6th and 9th -crossings (see the legend) of trajectories departing from a vicinity of and transiently approaching before heading towards either (up-pointing triangles) or CCF (down-pointing triangles). From this analysis, we estimate as the value of the parameter for which the sequence of -crossings asymptotically approaches (top red dashed line) and ceases to be accessible from . Incidentally, additional heteroclinic bifurcation points seem to occur in succession with increasing and, possibly, accumulate as is approached (see the region bounded by the blue rectangle and the magnification in Figure 15b; two such bifurcations are shown by the vertical grey lines). At these points, the connection from to is alternatively restored and broken. A detailed study of the apparent cascade of global bifurcations leading up to would require specifically designed numerical methods and is, therefore, beyond the scope of the paper.

Figure 15. Parameter dependence of . The 3rd, 6th and 9th crossings are indicated with triangles (see the legend and Figure 14ab). Up-pointing triangles indicate that the final state is , while down-pointing triangles mark trajectories ending in CCF. The red dashed curve is (the upper one is ). The black dashed curve is .
Edge tracking immediately above
Trajectories originating in a neighbourhood of at always relaminarise, so the solution is no longer an edge state. Since there are still two attractors, CCF and , the question arises as to which solution replaces in the role of the edge state.
Figure 16 shows the result of edge tracking at . Only the statistically converged edge state is shown, having discarded all preliminary transient evolution on the critical threshold. The long time series of Figure 16a reveals that the edge state is chaotic, the dynamics transiently slowing down at irregular intervals. A close inspection (see magnification in Figure 16b) reveals that the long lapses of fast dynamics orbit in close proximity of , while the intermittent slowdowns correspond to transient approaches to (horizontal grey dashed line). The phase map representation of Figure 16c further substantiates the claim that the chaotic edge state looks like (overlaid in dashed red) most of the time but occasionally visits (see the inset).

Figure 16. Edge state at . (a) Time series of the statistically converged edge state. (b) Magnification of the portion enclosed in the magenta rectangular region. The horizontal dashed line indicates . (c) Phase map representation of the edge state. is overlaid with a red dashed line. The green square region is magnified in the inset. is indicated with a gray square.
The peculiar dynamics in Figure 16 strongly suggest that the chaotic edge state must have originated from the heteroclinic cycle identified in §5.1. Figure 17 provides a sketch of how the phase space topology could have evolved for . In this figure, the CE (i.e. the dark blue loop) is the chaotic edge state originating from the heteroclinic cycle (see Figure 14d). The stable manifold now forms the basin boundary; although it is not shown in the figure, one may imagine it as approaching CE along a third coordinate axis.

Figure 17. Sketch of the phase space for slightly larger than . CE represents the chaotic edge state.
Figure 17 is drawn in a format similar to the previous sketches (Figure 13, Figure 14d). However, unlike earlier sketches, it does not merely show the discrete map on the Poincaré section. This is evinced by the fact that, while periodic orbits and are represented as points, CE is shown as a simple closed loop. This abstraction is a compromise made to illustrate the complications of an inherently high-dimensional phase space topology as we have in
the problem at hand. It should also be noted that the basin boundary is a hypersurface that contains the chaotic trajectory CE, and hence has a structure too complex to be visualised even in a three dimensional representation.
Summary and conclusions
Subcritical transition in minimal box shear flows can often be explained in the light of simple ECS arising at saddle-node bifurcations, with the saddle branch, together with its stable manifold, serving as the threshold for turbulent transition. However, as we show here, chaotic solutions emerged from global bifurcations may sometimes play a consequential part in the transition, even at relatively low Reynolds numbers.
Of the three possible types of chaotic saddles, those dwelling on the basin boundary, i.e., types (ii) and (iii), are instrumental in elucidating the turbulent transition of subcritical flows. However, only the mechanisms for the formation of type (i) saddles were known. By investigating the origin of type (ii) and type (iii) saddles, this paper contributes a theoretical framework that addresses the sensitive dependence on initial conditions observed in subcritical shear flow transition scenarios. The morphing of a chaotic saddle from one type to another happens through global bifurcations.
To summarise our findings, it is convenient to revisit the bifurcation diagram in Figure 3. The chaotic attractor , emerged from a period doubling cascade, becomes a saddle in a boundary crisis at . The saddle is type (i) because all trajectories starting from its neighbourhood eventually leak to the non-trivial attractor (), none relaminarising (see § 3). In other words, does not belong in the boundary that separates the basins of attraction of and the laminar circular Couette flow. The edge state is the simple solution , which has acted as such since its emergence in a saddle-node bifurcation together with the non-trivial attractor that later evolves into . The type (i) saddle is then subsumed into the basin boundary and becomes type (ii) in a heteroclinic bifurcation (another crisis) at involving the edge state, which at this point remains (see § 4). Finally, the occurrence of a second heteroclinic bifurcation at generates a heteroclinic loop between and a solution () at the boundary of the chaotic set that transforms the saddle into type (iii) (see § 5). From this point on, the chaotic saddle replaces in the role of edge state.
Our theoretical framework, supported by detailed numerical simulations of Taylor–Couette flow, explains some unaccounted for phenomena previously reported in the context of plane Couette and pipe flows. Notably, our findings in § 3 provide a precise rationale for the Poisson processes observed in various subcritical shear flows in the form of memoryless decay of turbulence (Eckhardt et al. [15]; Avila et al. [2]). Our results show how Poisson processes in type (i) saddles are closely related to the fractal structure in phase space of the chaotic set. To the best of our knowledge, this is the first study that accounts directly for the origin of the fractal structure of a chaotic saddle in a fluid mechanics system.
Complexity in the transition to sustained turbulence requires the presence of type (ii) or type (iii) chaotic saddles. We show that one way this may come about is through the transformation of a type (i) saddle. The underlying fractal structure is preserved during saddle type conversion. Assuming that this is a generic phase-space evolution scenario, our findings support the conjecture by Schneider et al. [51]; Kreilos & Eckhardt [33]; Kreilos et al. [34], that the basin boundary of subcritical shear flows has a fractal structure.
Also, the fact that the emergence of type (iii) chaotic saddles requires two heteroclinic bifurcations appears to possess a certain degree of universality. The importance of heteroclinic bifurcations to the onset of transient turbulence was emphasised by Lustro et al. [36] in the context of plane Couette flow. In our system, the final step producing the type (iii) saddle involves the replacement of the edge state via a global bifurcation. It is noteworthy that Lustro et al. [37] independently reported a similar edge switching phenomenon, although their case did not entail chaos. If the edge state is a simple ECS, transition control may be feasible (as suggested by Wang et al. 2007). However, once a type (iii) saddle appears, sensitivity to initial conditions leads to significant difficulties in devising effective control.
Finally, we note that global bifurcations that qualitatively change the nature of the edge state as the Reynolds number is increased seem to occur far more frequently than anticipated. We hypothesise that this is caused by the tangle of manifolds near the edge state. If global bifurcations that change the character of edge states or attractors are ubiquitous, this may explain why identifying a critical Reynolds number is notoriously difficult in subcritical shear flow transition. In long pipe flows, a critical Reynolds number has been identified statistically based on the competition of puff decay and splitting rates, turbulence remaining locally transient [1]. However, whether it is possible to determine a critical Reynolds number beyond which turbulence becomes sustained and the basin boundary can be rigorously defined remains an open question, even for short periodic pipes.
Acknowledgements. This research is supported by the Australian Research Council Discovery Project DP230102188 and the Ministerio de Ciencia, Innovación y Universidades, Agencia Estatal de Investigación, project PID2023-150029NB-I00 (MCIU/AEI/10.13039/501100011033/FEDER, UE). BW’s and RA’s research has been funded by the European Union’s Horizon 2020 research and innovation programme (Marie Skłodowska-Curie Grant Agreement No. 101034413). RA has also been funded by the Austrian Science Fund (FWF) 10.55776/ESP1481224. FM is a Serra-Húnter fellow.
Author ORCIDs. B. Wang, https://orcid.org/0000-0002-6229-0336; R. Ayats, https://orcid.org/0000-0001-6572-0621; K. Deguchi, https://orcid.org/0000-0002-3709-3242; A. Meseguer, https://orcid.org/0000-0002-2022-2001; F. Mellibovsky, https://orcid.org/0000-0003-0497-9052
Declaration of Interests. The authors report no conflicts of interest.
References
- [1]Avila, K., Moxey, D., de Lozar, A., Avila, M., Barkley, D. & Hof, B. 2011 The onset of turbulence in pipe flow. Science 333 (6039), 192–196.DOI
- [2]Avila, M., Barkley, D. & Hof, B. 2023 Transition to turbulence in pipe flow. Ann. Rev. Fluid Mech. 55, 575–602.
- [3]Avila, M., Mellibovsky, F., Roland, N. & Hof, B. 2013 Streamwise-localized solutions at the onset of turbulence in pipe flow. Phys. Rev. Lett. 110, 224502.DOI
- [4]Birkhoff, G. D. 1950 Nouvelles recherches sur les systèmes dynamiques. Collected Mathematical Papers 2, 530–662.
- [5]Budanur, N. B., Cvitanović, P., Davidchack, R. L. & Siminos, E. 2015 Reduction of SO(2) symmetry for spatially extended dynamical systems. Phys. Rev. Lett. 114 (8), 084102.DOI
- [6]Budanur, N. B., Dogra, A. S. & Hof, B. 2019 Geometry of transient chaos in streamwise-localized pipe flow turbulence. Phys. Rev. Fluids 4 (10), 102401.DOI
- [7]Clever, R. M. & Busse, F. H. 1992 Three-dimensional convection in a horizontal fluid layer subjected to a constant shear. J. Fluid Mech. 234, 511.DOI
- [8]Darbyshire, A.G. & Mullin, T. 1995 Transition to turbulence in constant-mass-flux pipe flow. J. Fluid Mech. 289, 83–114.DOI
- [9]De Lozar, A., Mellibovsky, F., Avila, M. & Hof, B. 2012 Edge state in pipe flow experiments. Phys. Rev. Lett. 108 (21), 214502.pubmed.ncbi.nlm.nih.gov/23003266
- [10]Deguchi, K. & Altmyer, S. 2013 Fully nonlinear mode competitions of nearly bicritical spiral or Taylor vortices in Taylor-Couette flow. Phys. Rev. E 87, 043017.pubmed.ncbi.nlm.nih.gov/23679523
- [11]Deguchi, K. & Hall, P. 2016 On the instability of vortex–wave interaction states. J. Fluid Mech. 802, 634–666.DOI
- [12]Deguchi, K., Meseguer, A. & Mellibovsky, F. 2014 Subcritical equilibria in Taylor-Couette flow. Phys. Rev. Lett. 112, 184502.pubmed.ncbi.nlm.nih.gov/24856700
- [13]Devaney, R. L. 2022 An Introduction to Chaotic Dynamical Systems, 3rd edn. Boca Raton, FL: CRC Press.
- [14]Eckhardt, B. & Mersmann, A. 1999 Transition to turbulence in a shear flow. Phys. Rev. E 60 (1), 509.DOI
- [15]Eckhardt, B., Schneider, T. M.; Hof, B. & Westerweel, J. 2007 Turbulence transition in pipe flow. Ann. Rev. Fluid Mech. 39, 447–468.pubmed.ncbi.nlm.nih.gov/19792506
- [16]Faisst, H. & Eckhardt, B. 2003 Travelling waves in pipe flow. Phys. Rev. Lett. 91, 224502.DOI
- [17]Faisst, H. & Eckhardt, B. 2004 Sensitive dependence on initial conditions in transition to turbulence in pipe flow. J. Fluid Mech. 504, 343–352.
- [18]Feigenbaum, M. J. 1978 Quantitative universality for a class of nonlinear transformations. Journal of Statistical Physics 19, 25–52.DOI
- [19]Feigenbaum, M. J. 1979 The onset spectrum of turbulence. Physics Letters A 74 (6), 375–378.DOI
- [20]Feigenbaum, M. J. 1980 The transition to aperiodic behavior in turbulent systems. Communications in mathematical physics 77, 65–86.
- [21]Feigenbaum, M. J. 1982 Some formalism and predictions of the period-doubling onset of chaos. In Nonlinear Problems: Present and Future (ed. Alan Bishop, David Campbell & Basil Nicolaenko), North-Holland Mathematics Studies, vol. 61, pp. 379–394. North-Holland.DOI
- [22]Froehlich, S. & Cvitanović, P. 2012 Reduction of continuous symmetries of chaotic flows by the method of slices. Commun. Nonlinear Sci. Numer. Simul. 17 (5), 2074–2084.DOI
- [23]Graham, M. D. & Floryan, D. 2021 Exact coherent states and the nonlinear dynamics of wall-bounded turbulent flows. Ann. Rev. Fluid Mech. 53, 227–253.DOI
- [24]Grebogi, C., Ott, E. & Yorke, J. A. 1983 Crises, sudden changes in chaotic attractors, and transient chaos. Physica D: Nonlinear Phenomena 7, 181–200.
- [25]Grebogi, C., Ott, E. & Yorke, J. A. 1986 Critical exponent of chaotic transients in nonlinear dynamical systems. Physical Review Letters 57, 1284–1287.pubmed.ncbi.nlm.nih.gov/10033406
- [26]Hamilton, J. M., Kim, J. & Waleffe, F. 1995 Regeneration mechanisms of near-wall turbulence. J. Fluid Mech. 287, 317–348.DOI
- [27]Hof, B., van Doorne, C. W. H, Westerweel, J., Nieuwstadt, F. T. M, Faisst, H., Eckhardt, B., Wedin, H., Kerswell, R.R. & Waleffe, F. 2004 Experimental observation of nonlinear traveling waves in turbulent pipe flow. Science 305, 1594–1598.pubmed.ncbi.nlm.nih.gov/15361619
- [28]Hof, B., Juel, A. & Mullin, T. 2003 Scaling of the turbulence transition threshold in a pipe. Phys. Rev. Lett. 91(24), 244502–4.DOI
- [29]Hof, B., Westerweel, J., Schneider, T. M. & Eckhardt, B. 2006 Finite lifetime of turbulence in shear flows. Nature 443(7107), 59–62.pubmed.ncbi.nlm.nih.gov/16957725
- [30]Itano, T. & Toh, S. 2001 The dynamics of bursting process in wall turbulence. Journal of the Physical Society of Japan 70 (3), 703–716.DOI
- [31]Jiménez, J. & Moin, P. 1991 The minimal flow unit in near-wall turbulence. J. Fluid Mech. 225.
- [32]Kerswell, R. R. 2005 Recent progress in understanding the transition to turbulence in a pipe. Nonlinearity 18, R17–R44.DOI
- [33]Kreilos, T. & Eckhardt, B. 2012 Periodic orbits near onset of chaos in plane Couette flow. CHAOS 22, 047505.DOI
- [34]Kreilos, T., Eckhardt, B. & Schneider, T. M. 2014 Increasing lifetimes and the growing saddles of shear flow turbulence. Phys. Rev. Lett. 112, 044503.DOI
- [35]Kuznetsov, Y. A. 2004 Elements of Applied Bifurcation Theory, 3rd Ed.. Springer-Verlag.
- [36]Lustro, JRT., Kawahara, G., van Veen, L., Shimizu, M. & Kokubu, H. 2019 The onset of transient turbulence in minimal plane Couette flow. J. Fluid Mech. 862, R2.DOI
- [37]Lustro, J. R. T., Shimizu, Y. & Kawahara, G. 2023 Homoclinic bifurcation and switching of edge state in plane Couette flow. Chaos: An Interdisciplinary Journal of Nonlinear Science 33 (6), 063112.DOI
- [38]Mellibovsky, F. & Eckhardt, B. 2011 Takens–Bogdanov bifurcation of travelling-wave solutions in pipe flow. J. Fluid Mech. 670, 96–129.DOI
- [39]Mellibovsky, F. & Meseguer, A. 2006 The role of streamwise perturbations in pipe flow transition. Phys. Fluids 18, 074104.DOI
- [43]Moehlis, J., Faisst, H. & Eckhardt, B. 2004 A low-dimensional model for turbulent shear flows. New Journal of Physics 6 (1), 56.
- [44]NAGATA, M. 1990 Three-dimensional finite-amplitude solutions in plane Couette flow: bifurcation from infinity. J. Fluid Mech. 217, 519–527.DOI
- [45]PEIXINHO, J. & MULLIN, T. 2006 Decay of turbulence in pipe flow. Phys. Rev. Lett. 96 (9), 094501.pubmed.ncbi.nlm.nih.gov/16606268
- [46]REYNOLDS, O. 1883 An experimental investigation of the circumstances which determine whether the motion of water shall be direct or sinuous, and of the law of resistance in parallel channels. Phil. Trans. R. Soc. A 174, 935–982.DOI
- [47]RITTER, P., MELLIBOVSKY, F. & AVILA, M. 2016 Emergence of spatio-temporal dynamics from exact coherent solutions in pipe flow. New J. Phys. 18, 083031.DOI
- [48]SCHMIEGEL, ARMIN & ECKHARDT, BRUNO 1997 Fractal stability border in plane couette flow. Phys. Rev. Lett. 79 (26), 5250.DOI
- [49]SCHNEIDER, T. M & ECKHARDT, B. 2006 Edge of chaos in pipe flow. Chaos: An Interdisciplinary Journal of Nonlinear Science 16 (4), 041103.pubmed.ncbi.nlm.nih.gov/17199375
- [50]SCHNEIDER, T. M & ECKHARDT, B. 2009 Edge states intermediate between laminar and turbulent dynamics in pipe flow. Phil. Trans. R. Soc. A 367 (1888), 577–587.DOI
- [51]SCHNEIDER, T. M., ECKHARDT, B. & YORKE, J. A. 2007 Turbulence transition and the edge of chaos in pipe flow. Phys. Rev. Lett. 99 (3), 034502.DOI
- [52]SKU FCA, J. D., YORKE, J. A. & ECKHARDT, B. 2006 Edge of chaos in a parallel shear flow. Phys. Rev. Lett. 96, 174101.pubmed.ncbi.nlm.nih.gov/16712300
- [53]SMALE, S. 1965 Diffeomorphisms with many periodic points. Differential and Combinatorial Topology pp. 63–80.
- [54]STROGATZ, STEVEN H. 2015 Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering, 2nd edn. Boca Raton, FL: CRC Press.
- [55]WANG, B., AYATS, R., DEGUCHI, K., MELLIBOVSKY, F. & MESEGUER, A. 2022 Self-sustainment of coherent structures in counter-rotating Taylor-Couette flow. J. Fluid Mech. 951, A21.DOI
- [56]WANG, B., AYATS, R., DEGUCHI, K., MESEGUER, A. & MELLIBOVSKY, F. 2025a Feigenbaum universality in subcritical Taylor-Couette flow. J. Fluid Mech. 1010, A36.arxiv.org/abs/2407.16097
- [57]WANG, B., AYATS, R., DEGUCHI, K., MESEGUER, A. & MELLIBOVSKY, F. 2025b Mathematically established chaos and forecast of statistics with recurrent patterns in Taylor-Couette flow. J. Fluid Mech. 1011, R2.arxiv.org/abs/2409.09234
- [58]WANG, B., MELLIBOVSKY, F., AYATS, R., DEGUCHI, K. & MESEGUER, A. 2023 Mean structure of the supercritical turbulent spiral in Taylor–Couette flow. Phil. Trans. R. Soc. A 381 (2246), 20220112.DOI
- [59]WANG, J., GIBSON, J. & WALEFFE, F. 2007 Lower branch coherent states in shear flows: transition and control. Phys. Rev. Lett. 98 (20), 204501.DOI
- [60]WEDIN, H. & KERSWELL, R. R. 2004 Exact coherent structures in pipe flow: travelling wave solutions. J. Fluid Mech. 508, 333–371.
- [61]WIGGINS, S. 2010 Introduction to Applied Nonlinear Dynamical Systems and Chaos, 2nd edn., Texts in Applied Mathematics, vol. 2. Springer New York.
- [62]WILLIS, A. P. & KERSWELL, R. R. 2007 Critical behavior in the relaminarization of localized turbulence in pipe flow. Phys. Rev. Lett. 98 (1), 014501.pubmed.ncbi.nlm.nih.gov/17358478
- [63]ZAMMERT, S. & ECKHARDT, B. 2015 Crisis bifurcations in plane Poiseuille flow. Phys. Rev. E 91 (4), 041003.pubmed.ncbi.nlm.nih.gov/25974431
