Accepted Manuscript

Visibility graph analysis of wall turbulence time-series

Giovanni Iacobello, Stefania Scarsoglio, Luca Ridolfi

PII: S0375-9601(17)31023-X
DOI: https://doi.org/10.1016/j.physleta.2017.10.027
Reference: PLA 24790

To appear in: Physics Letters A

Received date: 19 June 2017

Revised date: 15 September 2017
Accepted date: 15 October 2017

Please cite this article in press as: G. Iacobello et al., Visibility graph analysis of wall turbulence time-series, Phys. Lett. A (2017),
https://doi.org/10.1016/j.physleta.2017.10.027

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing
this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is
published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
Highlights
• We apply the visibility algorithm to analyze turbulent channel flow time-series.
• Temporal structures of the series are inferred by the network global metrics.
• Peaks and irregularities of time-series are highlighted by the visibility networks.
• Different flow dynamics along the wall-normal direction are captured by the metrics.
• This method represents a promising tool for inhomogeneous turbulent flows analyses.
 1 Visibility graph analysis of wall turbulence time-series

2 Giovanni Iacobelloa,∗ , Stefania Scarsoglioa , Luca Ridolfib

3
a Departmentof Mechanical and Aerospace Engineering, Politecnico di Torino, 10129
4 Torino, Italy
5
b Department of Environmental, Land and Infrastructure Engineering, Politecnico di

6 Torino, 10129 Torino, Italy

7 Abstract
8 The spatio-temporal features of the velocity field of a fully-developed turbu-
9 lent channel flow are investigated through the natural visibility graph (NVG)
10 method, which is able to fully map the intrinsic structure of the time-
11 series into complex networks. Time-series of the three velocity components,
12 (u, v, w), are analyzed at fixed grid-points of the whole three-dimensional do-
13 main. Each time-series was mapped into a network by means of the NVG
14 algorithm, so that each network corresponds to a grid-point of the simulation.
15 The degree centrality, the transitivity and the here proposed mean link-length
16 were evaluated as indicators of the global visibility, inter-visibility, and mean
17 temporal distance among nodes, respectively. The metrics were averaged along
18 the directions of homogeneity (x, z) of the flow, so they only depend on the wall-
19 normal coordinate, y + . The visibility-based networks, inheriting the flow field
20 features, unveil key temporal properties of the turbulent time-series and their
21 changes moving along y + . Although intrinsically simple to be implemented, the
22 visibility graph-based approach offers a promising and effective support to the
23 classical methods for accurate time-series analyses of inhomogeneous turbulent
24 flows.
25 Key words: Turbulent channel flows, Complex networks, Time-series analysis,
26 Visibility graph, Direct numerical simulations.

27 1. Introduction
28 One of the most challenging research topic in classical physics is represented
29 by turbulent flows. Their great importance is evident through a number of nat-
30 ural phenomena (e.g., rivers, atmospheric and oceanic streams), industrial and
31 civil applications (e.g., flow through pumps, heat exchangers, wake flows of ve-
32 hicles and aircraft, wind-building interactions) in which turbulence is involved.
33 The study of wall-bounded turbulent flows, in particular, is a very active re-
34 search field, due to the great attention paid to the fluid-structure interaction.
35 Although deeply studied from a phenomenological and theoretical point of view,
36 the turbulence dynamics, due to their complexity, are still not fully understood
37 [1, 2]. Nowadays, several numerical simulations and experiments are performed,

∗ Corresponding author
Preprint submitted to Elsevier October 17, 2017
Email addresses: [email protected] (Giovanni Iacobello),
[email protected] (Stefania Scarsoglio), [email protected] (Luca
Ridolfi)
38 providing a massive amount of spatio-temporal data that needs to be properly
39 examined. Different approaches, mainly relying on statistical techniques, are
40 then typically used to explore and analyze turbulent flows.
41 Among all the proposed techniques, time-series analysis is a broadly adopted
42 approach to study the temporal evolution of dynamical systems, specifically
43 those with high intrinsic complexity. Different methods, such as Fourier and
44 wavelet transforms [3, 4], as well as nonlinear approaches [5–7], have been de-
45 veloped so far to extract information from time-series. However, since each
46 method unavoidably loses some information about the temporal structure of
47 the series analyzed, new approaches are continuously required to fill this lack.
48 In the last decades, complex networks, by combining elements from graph the-
49 ory and statistical physics [8–10], have turned out to be powerful tools to study
50 complex systems, specifically by mapping time-series to extract non-trivial in-
51 formation [11]. Recently, several improvements were gained in this field and
52 numerous advances were proposed based on different approaches [12], such as
53 correlation [13, 14], visibility [15, 16], phase-space reconstruction [17, 18], recur-
54 rence quantification [19–21], and transition probabilities [22, 23] algorithms.
55 Beside the well-established applications to Internet, World Wide Web, economy
56 and social dynamics [24, 25], growing attention has been given nowadays to the
57 application of complex networks to fluid flows and different flow regimes have
58 been explored, such as two-phase flows [18, 20, 26], geophysical flows [27, 28],
59 turbulent jets [23, 29], reacting flows [30], as well as fully developed turbulent
60 flows [31, 32] and isotropic turbulence [33, 34].
61 In this work, the natural visibility algorithm is exploited to investigate the
62 spatio-temporal characterization of a fully-developed turbulent channel flow,
63 solved through a direct numerical simulation (DNS) and available from the
64 Johns Hopkins Turbulence Database (JHTDB) [35, 36]. Time-series of the three
65 velocity components were analyzed at fixed spatial positions, and a single net-
66 work was built at each point. In so doing, an ensemble of networks was obtained,
67 where each network corresponds to a time-series. This novel approach allows
68 us to capture some important aspects of the temporal structure of the
69 signal and how these features change along the wall-normal direction.
70 In fact, we can systematically extract information about the occur-
71 rence and temporal collocation of extreme events (i.e., peaks) and
72 irregularities, which are fundamental features to characterize turbu-
73 lent flows. The statistical tools classically adopted in turbulence, such
74 as correlation function, higher-order statistics, structure functions,
75 energy spectrum and probability density functions, all fail in pre-
76 serving and discerning the temporal structure of a time-series (e.g.,
77 two different temporal signals can have the same probability density
78 function or energy spectrum). The visibility approach here presented
79 is instead able to fully inherit and point out the temporal structure of
80 the turbulent series: the different temporal dislocation of events such
81 peaks and fluctuations will lead, case by case, to a different network
82 topology.
83 A systematic approach to highlight temporal features of the time-series through

2
84 the most significant network metrics is thus proposed and discussed. Particular
85 care is given not only to relate the network topology to the temporal
86 structure of the series, but also paying attention to the physical in-
87 terpretation of the results. New insights into how the network topology is
88 affected by important temporal features of the mapped signal are thus provided.
89 Specific combinations of the trend of the network metrics are able to shed light
90 into the time-series structure. Furthermore, a qualitative correspondence be-
91 tween the network metrics and the flow dynamics is presented, underlying the
92 ability of the method to identify different flow regions.

93 2. Methods

94 2.1. Database description

95 The data here used were extracted from a DNS of a fully developed turbu-
96 lent channel flow [37], available from the JHTDB [35, 36]. The simulation is
97 performed at Reτ = 1000, where Reτ = huτ /ν is the friction velocity Reynolds
98 number, h = 1 is the half-channel height, ν = 5·10−5 Ub h is the viscosity, Ub = 1
99 is the bulk channel velocity, and uτ = 5·10−2 is the friction velocity (all physical
100 parameters are dimensionless). Periodic boundary conditions in the streamwise
101 (x) and spanwise (z) directions are adopted, while the no-slip condition is im-
102 posed at the top and bottom walls, y/h = ±1 (y is the direction normal to the
103 wall). Once the statistically stationary conditions were reached, the simulation
104 was carried on for approximately one flow-through time, t ∈ [0, 26]h/Ub , with a
105 storage temporal step δt = 0.0065. Thus Nt = 4000 temporal frames are avail-
106 able. Velocity (u, v, w) and pressure (p) fields were computed over the physical
107 domain, (Lx × Ly × Lz ) = (8πh × 2h × 3πh), and stored with a grid resolution
108 (Nx × Ny × Nz ) = (2048 × 512 × 1536). Other simulation parameters and flow
109 statistics are given elsewhere [37].
110 In this study, a subset of the domain was taken into account, exploiting the
111 geometrical features of the flow field along the three directions (x, y, z). In the
112 wall-normal direction, y, due to the geometric symmetry, only grid-points from
113 the bottom wall to the half-channel height, −1 ≤ y/h ≤ 0, were considered. As
114 a result, the values of the dimensionless distance from the wall, y + , defined as
115 [38] y + = (y/h + 1)Reτ , ranges in the interval [0, 103 ]. Along the y-direction the
116 distance between consecutive grid-points was selected to increase gradually from
117 the wall towards the center of the channel. Indeed, close to the wall the flow
118 is strongly inhomogeneous, and a finer spatial resolution is necessary to better
119 capture the features of the flow field. Differently, along the x and z directions,
120 a coarse uniform storage was adopted. In fact, along these two directions the
121 flow is statistically homogeneous and fewer uniformly spaced grid-points are
122 sufficient to guarantee the statistical stationarity of the results.
123 The selected sub-domain size is (SX × SY × SZ ) = (64 × 70 × 12), where the first
124 grid-point Y = 0 corresponds to the wall coordinates y/h = −1 and y + = 0.
125 Details of the sub-domain structure are reported in Appendix A.

3
126 2.2. Mapping time-series into networks: the visibility algorithm
127 In the time-series analysis, complex networks represent a recent and promis-
128 ing tool to highlight and characterize important structural properties[7, 9]. In
129 the present work, the natural visibility algorithm proposed by Lacasa and co-
130 authors [15] was adopted. According to this method, two values (ti , s(ti )) and
131 (tj , s(tj )) of a univariate time-series s(tn ), n = {1, 2, ..., N }, have visibility, and
132 consequently are two connected nodes of the associated network, if the following
133 condition
t j − tk
s(tk ) < s(tj ) + (s(ti ) − s(tj )) , (1)
t j − ti
134 is fulfilled for any ti < tk < tj (or equivalently i < k < j). From a geometrical
135 point of view, two nodes are linked if there is a straight line connecting them
136 without intersecting any intermediate data. The natural visibility criterion is,
137 therefore, a convexity criterion. A geometrically simpler version of the NVG can
138 be obtained considering only horizontal lines among data, defining the so called
139 horizontal visibility graph [16]. In this case, the horizontal visibility satisfies an
140 ordering criterion.
141 The visibility algorithm is simple to implement and has been applied in many
142 different fields (e.g., [39–43]), including fluid flows [29–32, 44, 45]. However,
143 the visibility approach has some drawbacks, related to the fact that it is invariant
144 under affine transformations [15] (i.e., rescaling and translation of both horizon-
145 tal and vertical axes), so this could lead to a lost of information in mapping the
146 time-series. Moreover, if time-series with a considerable number of observations
147 (indicatively Nt > 104 ) are analyzed, then the condition (1) requires to be veri-
148 fied many times. In these situations, as in this study, an optimized approach is
149 crucial to sharply decrease the computational costs (e.g., see [46]).

150 2.3. Complex network metrics

151 A summary of the network metrics investigated in the present work is here
152 reported [9, 10, 24]. A network is defined as a graph G(N, E) = (V, E), where
153 V = {1, 2, ..., N } is a set of N labeled nodes (or vertices) and E = {1, 2, ..., E} is
154 a set of E links (or edges), with non-trivial topological features. The adjacency
155 matrix, Aij , defined as

0, if {i, j} ∈ E,
Aij = (2)
1, if {i, j} ∈ E,
156 determines the existence of a link between a pair {i, j} of nodes. Beside being
157 unweighted (A is a binary matrix), in this study we only consider undirected
158 networks (Aij = Aji ) with no self-loops (Aii = 0).
159 In general, two kinds of metrics can be defined: metrics associated to nodes,
160 namely local metrics, and metrics related to the entire network, here referred
161 as global metrics. In the following, the • notation is adopted to indicate that
162 global metrics were obtained by averaging (over all nodes in the network) the
163 corresponding local metrics.

4
164 The degree centrality of a node is defined as

N
ki = Aij , (3)
j=1

165 and gives the number of topological neighbors of node i, that is the number of
166 nodes linked to it (the set of neighbors is called the neighborhood, Γi , of node
167 i). In particular, if all N nodes in a network are linked each other, the network
168 is said fully-connected and the degrees are all constant and equal to N − 1. The
169 average degree centrality of a network is then

1 
N
k = ki . (4)
N i=1
170 As a centrality measure, the degree, ki , is an indicator of the most important
171 vertices in a network. The fraction of nodes in the network that have degree
172 k is the degree distribution, pk , and it also represents the probability that a
173 randomly chosen node has degree k. In many real networks, pk is heavy-tailed,
174 because of an intrinsic noise due to the finiteness of the time-series [9]. In these
175 cases it may be useful to evaluate the cumulative degree distribution [10]:
∞

Pk = pk  , (5)
k =k

176 which is the probability to find a degree greater than or equal to k. The statis-
177 tical fluctuations present in the tails of the pk distribution are smoothed if Pk
178 is used [9].
179 The transitivity, T r, is a global clustering metric and is defined as [24]
3NΔ
Tr = , (6)
N3
180 where NΔ and N3 are the number of triangles and the number of connected
181 triples in the network, respectively. A triangle is a set of three nodes linked
182 between them. A connected triple, instead, is a set of three nodes where two of
183 them must be directly linked to the third node. The transitivity, 0 ≤ T r ≤ 1,
184 is therefore a measure of the presence of triangles in the network. Another
185 commonly used clustering metric is the clustering coefficient, Ci [9].
186 Both the transitivity and the clustering coefficient are measures of
187 the presence of triangles in the network, but Ci tends to weight the
188 contributions of low-ki vertices more heavily that T r [10], being its
189 denominator proportional to ki2 , and making its interpretation less
190 clear and general. For this reason, in the following we only focus on
191 the transitivity to capture the inter-node relations among nodes.
192 Finally, we propose a new local metric, based on the temporal length between
193 two mutually visible vertices [47], defined as the mean link-length:
1 
d1n (i) = |tj − ti |, (7)
ki
j∈Γi

5
194 where Γi and ki are the neighborhood and the degree centrality of node i,
195 respectively. From a time-series point of view, each node represents a temporal
196 event and, then, the physical distance between two nodes i and j can be defined
197 as |tj − ti |. It follows that d1n increases when a node is linked to nodes far
198 in time from it. Averaging over all nodes in the network, a global measure is
199 obtained:
1 
N
d1n  = d1n (i). (8)
N i=1

200 2.4. Building the networks

201 In this study, the velocity field (u, v, w) was focused on, being one of the
202 most basic and significant field to analyze a turbulent flow. Exploiting
203 the visibility-invariance under affine transformations, in the following
204 each time-series is normalized as u∗ (ti ) = (u(ti ) − μ)/σ, where μ and σ
205 are the local mean and standard deviation values of u(ti ), respectively.
206 By following the classical decomposition adopted for the statistical
207 description of turbulence [38], the resulting signal, u∗ , has zero mean
208 value and standard deviation equal to 1. At fixed point, (u∗ , v ∗ , w∗ )
209 represent the net turbulent fluctuations of the velocity field. The
210 adopted decomposition allows one to separate the complete signal,
211 u(t), into a mean term constant in time, μ, and a fluctuating tempo-
212 ral part, u∗ (t). In so doing, we can focus on the temporal variations
213 only, by comparing normalized signals having the same mean and
214 standard deviation values. Turbulent fluctuations are the basis of the
215 statistical description
 of turbulence. For example, root-mean-square
216 velocity, urms = u∗ (t)2 (the overbar represents the temporal aver-
217 age), is usually defined to quantify the turbulence strength. A high
218 urms indicates an elevate level turbulence. Thus, (u∗ , v ∗ , w∗ ) hold the
219 primary indication of the turbulence intensity of a velocity field.
220 For each grid-point in space, all the Nt = 4000 time frames were then exploited
221 to build the networks, being the velocity series dependent only on the time
222 (i.e., the series are univariate at fixed coordinates). Therefore, Sx × Sy × Sz =
223 53760 networks were constructed for each velocity component. According to
224 the visibility algorithm — since each of the resulting networks is connected (i.e.
225 every node has at least one neighbor) [15] — each temporal instant corresponds
226 to a node. Consequently, all the 53760 networks have the same number of nodes
227 N = Nt = 4000. The number of links of each network, E, instead, can
228 be obtained from the average degree values by applying the general
229 relation, E = kN/2. A sensitivity analysis on the number of nodes N is
230 reported in Appendix B, for which different temporal discretizations (namely
231 2 δt and 4 δt, where δt = 0.0065 is the temporal discretization leading to 4000
232 time frames) are considered, resulting in different cardinality of the networks.
233 We recall that in a fully developed turbulent channel flow the velocity and
234 pressure fields are statistically homogeneous along the streamwise, x, and span-
235 wise, z, directions. The wall normal coordinate, y + , is then the only direction

6
236 where spatial inhomogeneities develop. Since network measures inherit the prop-
237 erties of the mapped time-series [15], also the global metrics (i.e. averaged over
238 the nodes of each network) were assumed as statistically homogeneous in the
239 x-z directions (more details are reported in Appendix C, where few
240 representative plots of the global metrics calculated at fixed x and z
241 locations are reported). Consequently, the global metrics were firstly calcu-
242 lated for each single network. Subsequently, such global metrics were averaged
243 over the Sx × Sz grid-points as:

 )= 1 
F(Y F(X, Y, Z), (9)
Sx Sz
X Z

244 where where F represents the specific metric considered, namely F = {k; T r; d1n }.
245 In so doing, we obtained three averaged quantities:  
k, Tr, and d 1n , where the
246

notation (•) indicates the average over grid-points in the directions of spatial
247 homogeneity of the flow. This operation makes the above averaged metrics
248 dependent only on the wall normal coordinate y + and their plots statistically
249 meaningful.

250 3. Relating time-series structure and network metrics

251 It is known that the general structure of time-series is preserved in the topol-
252 ogy of the associated natural visibility graphs, as shown by Lacasa et al. [15]
253 and as emerges from successive works [29–32, 39, 41]. Specifically, periodic time-
254 series are converted into regular networks, i.e. graphs where nodes have constant
255 degrees related to the periods of the series. Fractal series, instead, convert into
256 networks with power-law degree distributions [48]. In particular, fractional
257 Gaussian noise with Hurst exponent equal to 0.5 (i.e., uncorrelated
258 random series) are mapped by the NVG method into networks with
259 power-law degree distribution with exponent, γ, equal to 4 [48].
260 Here, particular attention is paid relating the network topology and
261 the temporal structure of the series to a physical interpretation of the
262 network metrics, with respect to the flow dynamics. In fact, although
263 the overall features of the time-series are inherited by the corresponding visibility
264 graphs, it is not straightforward how topological network metrics are affected
265 by different temporal behaviors of the series. In order to explore this gap, we
266 qualitatively relate the global metrics behavior to the temporal structure of the
267 corresponding time-series.
268 In general, if two different time-series are compared, they can differ in several
269 ways. In this analysis, we focused on the presence of peaks and irregularities.
270 A point of a time-series, s(ti ), is said a peak if it is a local (or global) maximum
271 of s(ti ), with order of magnitude comparable with the maximum excursion of
272 the series, Δ = smax − smin . Peaks generally have higher probabilities to con-
273 nect to other points in the series, because obstacles to the visibility are avoided
274 from higher positions. However, in turn, the long-term visibility of points in

7
275 the surroundings of peaks is obstructed by the peaks, thus creating local bar-
276 riers to the visibility of lower points of the series. Irregularities are temporal
277 variations with order of magnitude much smaller than Δ, and defined as local
278 barriers decreasing the visibility of the surrounding points. Peaks and irregu-
279 larities are focused mainly for two reasons. First, the occurrence and
280 temporal collocation of extreme events (i.e., peaks) and irregularities
281 represent some of the fundamental features to characterize turbu-
282 lent flows. Second, the NVG is a suitable method to evidence this
283 kind of flow properties and translate them into the network topology.
284 In particular, among all the topological parameters investigated, the
285 transitivity, T r, the global mean link-length, d1n , and the average
286 degree, k, turned out to be the metrics that better capture the tem-
287 poral structure of the time-series in terms of peaks and irregularities,
288 inheriting important features of the turbulent flow dynamics.
289 In order to schematize how the occurrence of peaks and irregularities affects
290 the temporal structure and in turn the network topology, we consider four exem-
291 plifying time series, as reported in Fig. 1. The starting series (panel a) is a sine
292 function. With respect to panel (a), in panels (b)-(d) a uniform random noise
293 is added to account for irregularity, while in panels (c) and (d) the periodicity
294 is halved. The graphical representation of the networks corresponding to each
295 time-series is reported on the right panels of Fig. 1.

296 3.1. Transitivity analysis

297 Let us first focus on the transitivity, T r. We recall that, since each pair
298 (j, l) ∈ Γi (Γi is the neighborhood of node i) always forms a connected triple
299 with node i, the total number of triples in the network depends on the size of all
300 the neighborhoods Γ. On the other hand, triangles are formed only if the nodes
301 (j, l) ∈ Γi are also linked, that is if Aij = Ail = Ajl = 1. In general, short-term
302 connections are the most probable ones because time-series are not expected to
303 sharply change in time (except for random series), so that nodes which are close
304 in time are more likely to form triangles. If two neighbors (j, l) ∈ Γi are far in
305 time, instead, there are many nodes in between j and l so that there is a high
306 probability to find a node that obstructs the inter-visibility of j and l. As a
307 result, the total number of triangles and triples, and therefore the transitivity,
308 strongly depend on the inter-visibility of nodes inside each neighborhood. The
309 transitivity can be then interpreted as a measure to characterize the typical
310 convexity properties on some intermediate time-scale (i.e., the neighborhood
311 temporal lengths) [41].
312 To better describe the effects of peaks and irregularities on the transitivity,
313 let us consider the time-series, s(ti ), and the corresponding networks, G, re-
314 ported in Fig. 1. Time-series s(ti )(a),(c) are clearly more regular than the series
315 s(ti )(b),(d) , while s(ti )(c) and s(ti )(d) have three peaks instead of two. Therefore,
316 while the networks G(a) and G(c) are well organized in clusters (one cluster for
317 G(a) , two for G(c) ), G(b) and G(d) appear more complex (right panels of Fig. 1).
318 This happens because in G(b),(d) there are many nodes with low visibility due
319 to the presence of irregularities.

8
320 More in detail, any point in the ranges ti = (1 − 50) and ti = (51 − 100) of
321 s(ti )(c) has basically the same inter-visibility of corresponding points in s(ti )(a)
322 (i.e., nodes at the same relative altitude), because there are no substantial local
323 changes of regularity between s(ti )(a) and s(ti )(c) . As a consequence, the value
324 of transitivity of G(a) and G(c) are expected to be scarcely affected by different
325 occurrence of peaks, as evident from the Fig. 2 where T r(c) and T r(a) are
326 actually almost the same. The presence of more (or less) peaks in a time-series
327 then does not significantly modify the inter-visibility (i.e., the transitivity) of
328 nodes. Therefore, also T r of G(b) and G(d) are almost equal (see Fig. 2), being
329 the irregularities of time-series s(ti )(b) and s(ti )(d) very similar.
330 On the other hand, the time-series s(ti )(b),(d) clearly display irregularities
331 if compared with time-series s(ti )(a),(c) . The inter-visibility among neighbors
332 of a generic node is obstructed because of the irregularities in the time-series.
333 Let us consider an arbitrary node, for example i = 34, highlighted as a green-
334 colored dot in panels (a) and (b) of Fig. 1. While in G(a) the neighborhood
335 Γ34,(a) (highlighted in yellow in Fig. 1) includes either short-term, medium-term,
336 and long-term links, in G(b) the neighborhood Γ34,(b) includes only short-term
337 and long-term connections. Therefore, the number of triangles (relative to the
338 number of triples) in which is involved a generic node (e.g., i = 34) is generally
339 lower in irregular networks than in regular ones. As a result, the values of
340 T r(b) and T r(d) are much lower than T r(a) and T r(c) , as observed in Fig. 2.
341 Summarizing, the transitivity is much more affected by local variations due to
342 the presence of irregularities rather than the presence of local peaks in the series.
343 In terms of flow dynamics, the transitivity is related to the presence
344 of local fluctuations between consecutive peaks. Recalling that the
345 all signals are normalized with respect to their mean and standard
346 deviation values, the transitivity is thus a net measure of the intrinsic
347 fluctuation level of the time-series.

348 3.2. Mean link-length analysis

349 The second metric considered is the global mean link-length, d1n . If peaks
350 often occur in a series (as in Fig. 1, panels (c) and (d)), points far from each
351 other are not visible because far connections are hampered by peaks, and d1n  is
352 consequently strongly reduced. The visibility of a generic node in the networks
353 G(c) and G(d) is limited by the peak at i = 50 (green-colored dot in Fig. 2), which
354 in turn divides the networks into two main clusters (see bottom right panels of
355 Fig. 2). The value of d1n (c) and d1n (d) are indeed much lower than d1n (a)
356 and d1n (b) , respectively, as can be seen in Fig. 2. On the other hand, d1n 
357 is not essentially affected by the irregularities of a series. Indeed, irregularities
358 mostly prevent medium-term connections than short and long-term links but,
359 averaging over all nodes in the network, a value of the order of medium-term
360 links is generally obtained for d1n . In fact, in Fig. 2 the value of d1n (b) is
361 approximately equal to d1n (a) , while d1n (d) is almost the same of d1n (c) ,
362 indicating that there are no relevant changes in the global mean link-length due
363 to irregularities. To conclude, the global mean link-length is strongly influenced
364 by the occurrence of peaks, being slightly affected by the irregularities. The

9
365 mean link-length measures how isolate and sporadic extreme events
366 (i.e., peaks) are, with low d1n  values when the recurrence of peaks
367 is high. Differently to high-order statistics (such as, for example,
368 kurtosis), d1n  is able to fully capture the temporal dislocation of
369 such extreme events along the time-series.

370 3.3. Combining T r and d1n 

371 As a consequence of the previous observations, the visibility algorithm turns
372 out to be able to capture two main features of the temporal structure of a series:
373 the recurrence of peaks and the presence of irregularities. It is worth noting
374 that the two global (i.e., those associated to networks) measures analyzed so
375 far inherit the local structural features of the mapped time-series. Therefore, a
376 comparative temporal characterization of the time-series can be carried out by
377 combining the behaviors of the global metrics.
378 If T r and d1n  are focused on, a time-series can differ from another through
379 a combination of the metrics behaviors, namely T r and d1n  can increase,
380 decrease, or remain almost constant. Excluding the combination in which both
381 T r and d1n  are almost constant (i.e., the two compared time-series share the
382 same temporal features), four different cases can occur and they are explained
383 in table 1. Therefore, given the metric trends, it is possible to infer from table
384 1 how time-series differ in terms of peaks and irregularities.
385 Being the degree centrality, k, a direct measure of the visibility of nodes, it
386 can contemporarily account for both the recurrence of peaks and the presence
387 of irregularities. In other words, k combines the features of both T r and d1n 
388 in a single global metric. Therefore, due to its intrinsic definition, the degree
389 variation in general cannot be univocally related to a specific temporal fea-
390 ture (either peaks or irregularities occurrence). For this reason, although being
391 conceptually one of the easiest measure to interpret, here the degree centrality
392 will be mainly discussed as a posteriori validation of the transitivity and global
393 mean-length behaviors.

394 4. Results

395 The procedure described in the previous section is adopted to analyze the
396 velocity time-series of the turbulent channel flow, starting from the streamwise
397 component, u∗ , and then considering the other velocity components, v ∗ and w∗ .

398 4.1. Streamwise velocity component, u∗


399 In Fig. 3 the metrics  
k, Tr, d 1n are plotted as a function of the wall-normal
400 coordinate, y + . Substantial variations of these metrics occur moving along the
401 wall-normal direction, exhibiting clear and regular trends. The three metrics
402 have overall similar behaviors, rising from the wall up to a maximum value, then
403 decreasing to y +  100 − 200 and, finally, barely changing towards the center
404 of the channel. The maximum values of  
k, Tr, and d 1n are not exactly at the
405 same value of y , but they are quite close in the range y +  4 − 7. The global
+

10
406 network-metrics {k; T r; d1n }, computed for each of the Sx × Sz grid-point,
407 have regular trends similar to the averaged ones shown in Fig. 3, which are thus
408 representative of the global metrics measured along the wall-normal direction
409 and in different (x, z) coordinates (see also Fig. C9 in Appendix C).
410 To infer the temporal structure of time-series along the wall-normal coor-
411 dinate, we start from time-series close to the wall and then proceed towards
412 the center of the channel. In particular, we focus on three representative y +
413 stations, i.e. y + = 0.017, 15.4, 996.3. The three time-series, u∗ (t), at the se-
414 lected y + stations are shown in Fig. 4(a), while a graphical representation of
415 the corresponding networks is displayed in Fig. 4(b), revealing the presence of
416 different topological features at different distances from the wall.
417 Starting from the time-series at y + = 0.017, this series appears globally quite
418 smooth, with relatively slow variations in time, resulting in few pronounced
419 peaks. Moving from y + = 0.017 to y + = 15.4, the Fig. 3 shows that the
420 transitivity, Tr, consistently increases, while the average mean link-length, d1n ,
421 and the average degree,  k, noticeably decrease. This combination of metrics
422 corresponds to the Case D in table 1 (here T S(1) and T S(2) correspond to time-
423 series at y + = 15.4 and y + = 0.017, respectively). A normalized time-series
424 extracted at y + = 15.4 is then expected to be (on average) more regular than a
425 series extracted at y + = 0.017 (indicated by the growth of Tr), and with a more
426 frequent occurrence of peaks (indicated by the drop of d  
1n ). The reduction of k
427 suggests that the increasing occurrence of peaks affects the global visibility more
428 than the reduction in the irregularities. Looking at the time-series extracted at
429 y + = 15.4 of Fig. 4(a), it is indeed with more peaks than the series extracted
430 at y + = 0.017. This aspect is also evident in a more clustered topology of the
431 network built on the time-series at y + = 15.4 (see Fig. 4(b)). The regularities
432 appear globally similar but, as zoomed in the inset of Fig. 4(a), the two time-
433 series are locally different. In particular, the time-series at y + = 0.017 appears
434 more irregular, as indicated by the transitivity.
435 From y + = 15.4 to y + = 996.3 (i.e., close to the center of the channel,
436 h), all the average metrics substantially decrease. This combination of metrics
437 corresponds to the Case C in table 1. Accordingly, we expect that a time-series
438 extracted at the center of the channel is (on average) less regular than a series
439 at y + = 15.4 and with a more frequent recurrence of peaks. This behavior
440 can be clearly seen in Fig. 4 where the time-series at the center of the channel
441 is more fluctuating than the time-series at y + = 15.4, and the corresponding
442 network appears more clustered and disordered. It is interesting to note that
443 from y + ≈ 102 to the center of the channel, the three metrics barely change.
444 In summary, the temporal features of the series are actually as predicted
445 by combining the network metrics. Through the behavior of the met-
446 rics along the y + direction, Fig. 3 yields first important results on
447 the presence, dislocation and structure of extreme events and irreg-
448 ularities of the time-series. This kind of information can enrich the
449 comprehension of the flow dynamics. It is important to remark that the
450 behavior of a single metric is not a sufficient information, but a combination

11
451 of the two metrics, Tr and d 1n , instead, determines how two time-series differ
452 in terms of recurrence of peaks and/or irregularities. Moreover, we do not
453 refer to the specific value assumed by the metric, but the analysis is
454 comparative as it focuses on the trend each metric assumes as a func-
455 tion of the distance from the wall. Specifically, comparing time-series
456 at the wall and at the center of the channel, peaks are expected to
457 be remarkably closer, while irregularities do not substantially change
458

(d 
1n decreases while T r slightly increases). In fact, as shown in Fig.
459 4(a), in the center of the channel peaks occur more frequently but the
460 irregularity between them remains basically unvaried. However, this
461 trend is not monotonic along y + , since the time-series locally (around
462 y + = 15.4) change their regularity. In terms of the network topology,
463 as displayed in Fig. 4(b), close to the wall the network is composed by
464 different subnetworks, corresponding to the peaks of the series, which
465 are widely connected with each other and internally. Going towards
466 y + = 15.4, the simultaneous decrease of d  
1n and increase of T r mainly
467 break down long connections among the subnetworks. The drop of
468

d 1n plays a major role here, acting to split long-term links. The sub-
469 sequent decrease of both d  
1n and T r (from y
+
= 15.4 to y + = 996.3)
470 breaks principally intra-network connections. At this stage, the pre-
471 vailing effect is locally induced by the increase of irregularity, which
472 leads to a ramification of each subnetwork.
473 A comment on the degree centrality can be eventually carried out. A high
474 value of k indicates a globally convex time-series, while low values indicate a
475 strong fragmentation of the visibility network [41]. As a result, considering
476 the behavior of  k in Fig. 3, at high values of y + the time-series are globally
477 more fragmented than the time-series close to the wall, confirming what found
478 observing the trends of Tr and d 1n .
479 Finally, the average cumulative degree distributions, P k , are illustrated in
480 Fig. 5 (semi-log plot). As for the metrics, a degree distribution was computed for
481 each network and all the distributions were then averaged over the homogeneous
482 directions. The Pk of a visibility network can be thought of as a measure
483 of the (linear and nonlinear) temporal dependences existing in the time-series
484 [16]. However, differently from the horizontal visibility algorithm, the
485 behavior of the degree distribution also depends on the probability
486 density function (pdf ) of the mapped time-series when the natural
487 visibility algorithm is applied [49]. As evident from the Fig. 5, the tail of
488 the distributions reveals decreasing exponential trends, i.e. the higher degree
489 values (i.e., the hubs) are generally very infrequent. In particular, the exponent
490
k increases (in modulus) from the wall towards the center
of the fitting, γ, of the P
491 of the channel, y = h. This is consistent with the temporal integral scale
492 measurements [50], which decrease from the wall to the center of the
493 channel.
494 In order to isolate (from the pdf contribute) the net impact of (lin-
495 ear and nonlinear) dependences in the turbulent time-series, we built

12
496 four series by shuffling four velocity time-series (at arbitrary (x, z)
497 locations) at the same wall-normal distances considered, i.e., y + =
498 (0.017, 15.4, 106.2, 996.3). As shown in Fig. 5 (and highlighted in the in-
499 set), the slopes of Pk from the shuffled series are substantially steeper
500 than the turbulent time-series. This demonstrates the key role of the
501 (linear and nonlinear) correlation aspects of the turbulent series.
502 Up to now, we pointed out the ability of the visibility-based networks to shed
503 light on the temporal structure of the corresponding mapped time-series. Now
504 we try to relate the network metrics with the flow dynamics, that are responsible
505 for the time-series behavior. Looking at the Fig. 3, three regions are particu-
506 larly interesting (i.e. y +  7, 7  y +  150, and y +  150) where the average
507 metrics mostly change their trend. It should be noted that the values of y +
508 delimiting such regions are very close to the limit values, y + = 5 and y/h = 0.1,
509 of the viscous sub-layer and inner layer, respectively [38]. In particular, for
510 Reτ = 1000, the inner layer limit is about y + = 100. The region for y + < 5
511 is characterized by slow moving fluid and the flow dynamics are dominated by
512 the viscous shear stresses. The normalized time-series u∗ (ti ) here can be as-
513 sumed to roughly share a similar temporal structure (although their mean and
514 standard deviation values clearly change along y + ). The corresponding metrics
515 (see Fig. 3) highlight this behavior resulting in barely increasing trends. As
516 previously observed, around y +  4 − 7 (which is the upper bound of
517 the viscous sub-layer) the three metrics reach their maximum values.
518 Here we expect, in terms of time-series shape, a minimum number
519 of peaks along with the minimum irregularities. Recalling that all
520 signals are normalized with the local mean and standard deviation,
521 a possible interpretation is the following. Around y +  4 − 7, we are
522 approaching the buffer layer (5 < y + < 30), an intermediate region
523 where viscous shear stress starts decreasing while turbulence activity
524 begins to grow. However, at the very beginning (y +  4−7), turbulent
525 processes are very low, thus resulting in a minimum of irregularities,
526 which act over a signal that is still affected by slow temporal varia-
527 tions (i.e., low number of peaks). The combination of these dynamics
528 reasonably explains the maxima reached by all the metrics around
529 the region y +  4 − 7. For y + > 5 the flow dynamics are more affected by the
530 Reynolds shear stresses, and the flow shows a tendency to organize into coher-
531 ent turbulent patterns [38]. The structure of the time-series is then affected by
532 turbulent processes (e.g., ejections and sweeps [38]), leading to rapid temporal
533 variations. This behavior could be recognized in the drop of the average met-
534 rics (Fig. 3). As y + further increases (y + > 100), the turbulent patterns are
535 less affected by the wall and they can develop in larger structures. However,
536 the coexistence of multiple scales and the more complicated flow structure [51]
537 seems not to translate into a clear trend for the network metrics.

538 4.2. Transversal and spanwise velocity components, (v ∗ , w∗ )

539

In Fig. 6, the three metrics Tr, d 
1n and k are displayed for all the velocity
∗ ∗ ∗
540 components, (u , v , w ) to facilitate the comparison. In general, the mean link-

13
541 length and the average degree measured on the time-series of v ∗ and w∗ show
542 trends similar to those of u∗ , while different trends are obtained considering the
543 transitivity.
544 More in particular, for d +
1n the trends over y for the three velocity compo-
545 nents are similar, but values for the streamwise velocity, u∗ , are overall higher
546 than those displayed by v ∗ and w∗ . The relative difference decreases towards
547 the center of the channel. The scenario for the degree centrality,  k, is analo-
548 gous. Differences for the  k values of the three components are marked close to
549 the wall, while k values tend to coincide approaching the channel center. This
550 behavior can be explained by considering that close to the wall the presence
551 of the wall itself strongly influences and differently characterizes the flow dy-
552 namics in the three directions of the velocity, and consequently the networks
553 based on the corresponding time-series are affected. On the contrary, the wall
554 effects decrease moving far from the wall (y + > 100), thus differences among the
555 metrics built on u∗ , v ∗ and w∗ , reduce. As for the transitivity, Tr, the metric
556 difference among velocity components is even more accentuated. In fact, in the
557 region y + < 100, not only values are different but also metrics display differ-
558 ent trends. In particular, the wall-normal velocity component, v ∗ , is strongly
559 affected by the presence of the wall (recall that close to the wall the motion
560 corresponds to flow in planes parallel to the wall [38]) and this in turn involves
561 the transitivity. For example, spikes with large negative values can be found
562 in the time-series of v ∗ as a consequence of strong events that appear only in
563 the very near-wall region, revealed by high kurtosis levels [52]. Since these deep
564 peaks are negative and relatively short, the degree and the mean link-length of
565 the corresponding networks are barely affected, while the transitivity is strongly
566 reduced. Towards the channel center, similarly to the mean link-length d 1n , the
567 transitivity differences for the three velocity components tend to reduce.
568 In the end, the cumulative degree distributions, P k , of the networks built on
∗ ∗ ∗
569 the three velocity components, (u , v , w ), and averaged over the grid-points
570 in the homogeneous directions are displayed in Fig. 7. At fixed positions from
571 the wall ((a): y + = 0.0017, (b): y + = 15.4, (c): y + = 996.3), the slope of the
572 three components is pretty similar, confirming that a steeper decay is present
573 when moving far from the wall (from y + = 0.0017 to y + = 996.3).

574 5. Conclusions

575 In this work, the application of the natural visibility graph to time-series of a
576 fully-developed turbulent channel flow was studied. Our attention was focused
577 on the streamwise velocity component, u, although the other velocity compo-
578 nents were also explored. Velocity time-series were adopted to build the
579 corresponding networks as the velocity field is one of the most in-
580 tuitive quantity to characterize a fluid flow. However, the visibility
581 graph method can be applied to other quantities of turbulence in-
582 terest, such as the Reynolds shear stress, the kinetic energy, or the
583 vorticity field. Firstly, we provided some novel insights into how the net-

14
584 work metrics are affected by the different temporal structure of the mapped
585 time-series. The average transitivity, Tr, the here introduced mean link-length,
586

d 
1n , and the average degree, k, were chosen as the most representative metrics.
587 Their trends turned out to effectively highlight the temporal features, in terms
588 of peaks and irregularities, of the mapped time-series along the wall-normal co-
589 ordinate, y + . Furthermore, the cumulative degree distributions are found to
590 show a decreasing exponential tendency, but with fitting exponent values at
591 least one order of magnitude greater than uncorrelated random series. Differ-
592 ent metrics variations were also quite well associated to the flow dynamics, as
593 responsible of the time-series behavior.
594 Despite several statistical techniques are available to study nonlinear time-
595 series, specifically regarding turbulence, most of them are invariant under dif-
596 ferent temporal structures of the time-series. The visibility-network analysis,
597 instead, reveled to be a powerful and synthetic tool to handle big-data and to
598 explore specific temporal features of the mapped series, without losing infor-
599 mation about their temporal structures and also capturing the underlying flow
600 dynamics. In fact, each network is built holding the temporal dislo-
601 cation of important temporal features, such as extreme events and
602 irregularities. To extract and handle this information is crucial for
603 a deeper understanding of the flow dynamics, since the most com-
604 mon statistical tools adopted in turbulence, from spectral analysis to
605 higher-order moments, are not able to retain the temporal colloca-
606 tion of such phenomena. Our network-based approach demonstrates
607 that visibility graph method is able to give much information about
608 temporal structure of turbulent time-series and it will deserve future
609 efforts, such as community and neighborhood detection, to better
610 explore the network topology and its physical meaning. Future works
611 can also involve simulations with different Reynolds numbers, and weighted or
612 directed networks may be considered. Furthermore, finer spatial and temporal
613 simulation resolutions may be considered.
614 Based on present findings, the proposed procedure may thus provide a
615 promising support to the classical methods for accurate time-series analyses
616 of inhomogeneous turbulent flows. In particular, given a time-series and the
617 behaviors of the network metrics as a function of the distance from the wall, it
618 is possible to qualitatively infer the behavior of the time-series at another wall-
619 normal distance. This method can be then particularly useful as a predictive
620 and supportive tool when experimental measurements are difficult.

621 Acknowledgments

622 This work was supported by the scholarship Ernesto e Ben Omega Petrazz-
623 ini, awarded by the Accademia delle Scienze di Torino, Turin (Italy). A special
624 thank goes to the Accademia delle Scienze and the Petrazzini family. The au-
625 thors would also like to thank J. G. M. (Hans) Kuerten for the fruitful discussion
626 of the results.

15
627 Appendix A

628 The selected grid-points of the sub-domain (Sx , Sy , Sz ) ⊂ (Nx × Ny × Nz )

629 are reported below (according to the labeling of the online database) in the
630 form (a : d : b), where a and b are respectively the first and the last index of
631 a uniformly spaced interval, and d is a grid step size (e.g., (1 : 2 : 9) takes the
632 indices {1; 3; 5; 7; 9}):
633 • wall-normal direction,
⎧
⎪
⎪ (0 : 1 : 21)
⎪
⎪
⎪
⎪ (23 : 2 : 39)
⎨
(42 : 3 : 54) and (58 : 3 : 79)
Y =
⎪
⎪ (84 : 5 : 169)
⎪
⎪
⎪
⎪ (179 : 10 : 239)
⎩
255, i.e. y + = 996.3;

634 • streamwise direction, X = (0 : 32 : 2016);

635 • spanwise direction, Z = (110 : 128 : 1518).

636 Appendix B

637 A sensitivity analysis on the number of nodes N is here reported, by varying

638 the temporal discretization and consequently the cardinality of the correspond-
639 ing network. We recall that δt = 0.0065 leads to a number of nodes, N = 4000.
640 Two other time steps, namely 2 δt, and 4 δt, have been considered, resulting in
641 networks with N = 2000 and N = 1000, respectively. In Fig. B8, the met-
642 rics as function of y + are displayed for the three temporal samplings, c δt, with
643 c = 1, 2, 4. Mean link-length and degree centrality are reported as scaled with
644 c (namely c d  
1n and c k), to facilitate the comparison between samplings. The
645 transitivity, Tr, is not scaled with c as by definition varies between 0 and 1. It
646 can be observed that, apart from the specific values reached by the transitivity,
647 the metrics behavior along the wall-normal direction y + is not sensitive to the
648 choice of the temporal discretization (i.e., the number of nodes).

649 Appendix C

650 In this section we report few representative plots of the global

651 metrics calculated on networks at different single streamwise, x, and
652 spanwise, z, locations. In Fig. C9 we illustrate the behavior of the
653 transitivity, the mean link-length and the degree centrality as a func-
654 tion of y + before the averaging operation, performed according to the
655 Eq. (9), and compare them with the averaged behavior (as shown
656 in Fig. 3). Specifically, in Fig. C9, we plotted 48 curves for each
657 metric, obtained from 48 uniformly spaced grid-points in the (x, z)

16
658 directions and covering the whole domain. As can be seen, the aver-
659 aged behaviors (reported in black in Fig. C9) are representative of
660 the behavior of the global metrics for different streamwise and span-
661 wise grid-points. The distributions of the mean link-length and the
662 degree centrality appear less noisy (especially at the center of the
663 channel) than the plots of the transitivity because the latter is glob-
664 ally evaluated for each network (see Eq. (6)), while k and d1n  are
665 defined as averages over nodes (see Eq. (4) and Eq. (8)). Therefore,
666 we can conclude that the statistics homogeneity of the flow is inher-
667 ited by the visibility networks, making the average behavior along y +
668 statistically meaningful.

669 References

670 [1] Z. Warhaft, Proc. Natl. Acad. Sci. 99 (2002) 2481–2486.

671 [2] A. J. Smits, B. J. McKeon, I. Marusic, Ann. Rev. Fluid Mech. 43 (2011)
672 353–375.

673 [3] G. E. Box, et al., Time series analysis: forecasting and control, John Wiley
674 & Sons, 2015.
675 [4] D. B. Percival, A. T. Walden, Wavelet methods for time series analysis,
676 volume 4, Cambridge University Press, 2006.
677 [5] S. Strogatz, et al., Comput. Phys. 8 (1994) 532–532.

678 [6] H. Kantz, T. Schreiber, Nonlinear time series analysis, volume 7, Cam-
679 bridge University Press, 2004.
680 [7] A. S. Campanharo, et al., PLoS ONE 6 (2011) e23378.
681 [8] D. J. Watts, S. H. Strogatz, Nature 393 (1998) 440–442.

682 [9] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, D.-U. Hwang, Phys. Rep.
683 424 (2006) 175–308.
684 [10] M. E. Newman, SIAM Rev. 45 (2003) 167–256.

685 [11] Z.-K. Gao, M. Small, J. Kurths, Europhysics Letters 116 (2016) 50001.
686 [12] A. Nuñez, B. Luque, J. Gomez, L. Lacasa, Visibility algorithms: A short
687 review, Citeseer, 2012.
688 [13] J. Zhang, M. Small, Phys. Rev. Lett. 96 (2006) 238701.

689 [14] Y. Yang, H. Yang, Physica A 387 (2008) 1381–1386.

690 [15] L. Lacasa, B. Luque, F. Ballesteros, J. Luque, J. Nuno, Proc. Natl. Acad.
691 Sci. 105 (2008) 4972–4975.

17
692 [16] B. Luque, L. Lacasa, F. Ballesteros, J. Luque, Phys. Rev. E 80 (2009)
693 046103.
694 [17] X. Xu, J. Zhang, M. Small, Proc. Natl. Acad. Sci. 105 (2008) 19601–19605.
695 [18] Z. Gao, N. Jin, Chaos 19 (2009) 033137.

696 [19] R. V. Donner, et al., New J. Phys. 12 (2010) 033025.

697 [20] Z.-K. Gao, et al., Europhysics Letters 103 (2013) 50004.
698 [21] Z.-K. Gao, W. Dang, Y. Yang, Q. Cai, Chaos 27 (2017) 035809.

699 [22] G. Nicolis, A. G. Cantu, C. Nicolis, International Journal of Bifurcation

700 and Chaos 15 (2005) 3467–3480.
701 [23] A. Shirazi, et al., J. Stat. Mech. Theory Exp. 2009 (2009) P07046.
702 [24] L. d. F. Costa, F. A. Rodrigues, G. Travieso, P. R. Villas Boas, Adv. Phys.
703 56 (2007) 167–242.

704 [25] S. Havlin, et al., European Physical Journal Special Topics 214 (2012) 273–
705 293.
706 [26] Z.-K. Gao, S.-S. Zhang, W.-D. Dang, S. Li, Q. Cai, International Journal
707 of Bifurcation and Chaos 27 (2017) 1750059.

708 [27] M. Lindner, R. V. Donner, Chaos 27 (2017) 035806.

709 [28] L. Tupikina, N. Molkenthin, C. López, E. Hernández-Garcı́a, N. Marwan,
710 J. Kurths, PLoS ONE 11 (2016) e0153703.
711 [29] A. Charakopoulos, T. Karakasidis, P. Papanicolaou, A. Liakopoulos, Chaos
712 24 (2014) 024408.
713 [30] M. Murugesan, R. Sujith, J. Fluid Mech. 772 (2015) 225–245.
714 [31] C. Liu, W.-X. Zhou, W.-K. Yuan, Physica A 389 (2010) 2675–2681.
715 [32] P. Manshour, M. R. R. Tabar, J. Peinke, J. Stat. Mech. Theory Exp. 2015
716 (2015) P08031.
717 [33] K. Taira, A. G. Nair, S. L. Brunton, J. Fluid Mech. 795 (2016) R2.
718 [34] S. Scarsoglio, G. Iacobello, L. Ridolfi, International Journal of Bifurcation
719 and Chaos 26 (2016) 1650223.

720 [35] Y. Li, et al., J. Turbul. (2008) N31.

721 [36] E. Perlman, R. Burns, Y. Li, C. Meneveau, in: Proc. of the 2007
722 ACM/IEEE Conf. on Supercomputing, ACM, p. 23.
723 [37] J. Graham, et al., Journal of Turbulence 17 (2016) 181–215.

18
724 [38] S. B. Pope, Turbulent flows, Cambridge University Press, 2000.
725 [39] M. Stephen, C. Gu, H. Yang, PLoS ONE 10 (2015) e0143015.
726 [40] S. Supriya, S. Siuly, H. Wang, J. Cao, Y. Zhang, IEEE Access 4 (2016)
727 6554–6566.

728 [41] R. V. Donner, J. F. Donges, Acta Geophysica 60 (2012) 589–623.

729 [42] S. Mutua, C. Gu, H. Yang, Chaos 26 (2016) 053107.
730 [43] Z.-K. Gao, Q. Cai, Y.-X. Yang, N. Dong, S.-S. Zhang, International Journal
731 of Neural Systems 27 (2017) 1750005.

732 [44] J. Singh, R. Belur Vishwanath, S. Chaudhuri, R. Sujith, Chaos 27 (2017)

733 043107.
734 [45] A. Braga, L. Alves, L. Costa, A. Ribeiro, M. de Jesus, A. Tateishi,
735 H. Ribeiro, Physica A: Statistical Mechanics and its Applications 444
736 (2016) 1003–1011.

737 [46] X. Lan, H. Mo, S. Chen, Q. Liu, Y. Deng, Chaos 25 (2015) 083105.
738 [47] I. Bezsudnov, A. Snarskii, Physica A 414 (2014) 53–60.
739 [48] L. Lacasa, B. Luque, J. Luque, J. C. Nuno, EPL (Europhysics Letters) 86
740 (2009) 30001.

741 [49] P. Manshour, Chaos: An Interdisciplinary Journal of Nonlinear Science 25

742 (2015) 103105.
743 [50] M. Quadrio, P. Luchini, Physics of Fluids 15 (2003) 2219.
744 [51] J. Jiménez, Phys. Fluids (1994-present) 25 (2013) 101302.

745 [52] C. Xu, Z. Zhang, J. M. den Toonder, F. Nieuwstadt, Phys. Fluids 8 (1996)
746 1938–1944.

19
747 Figure Legends

748 Fig. 1 (Left) Examples of sine time-series with different temporal features. In
749 panels (a) and (b), the green-colored dot indicates the point s(ti = 34), while
750 yellow points highlight its neighborhood, Γ34 . In panels (c) and (d), the green-
751 colored dot evidences the point s(ti = 50). (Right) Networks corresponding to
752 the time-series on the left.
753 Fig. 2 Bar plot of the transitivity (blue) and global mean link-length (yellow)
754 values for the four time-series of Fig. 1.
755 Fig. 3 Global metrics averaged over the Sx × Sz networks as function of y + ,
756 reported
 in a log-linear plot. The values of the metrics at the wall (y + = 0) are
757
 
k, Tr, d 1n = (1.9995, 0, 1), resulting from constant time-series and thus not
758 shown here. Three representative values of y + are also highlighted.
759 Fig. 4 (a) Normalized time-series, u∗ (ti ), at the grid-points X = 1601, Z = 750
760 and y + = 0.017, 15.4, 996.3. The choice of the coordinate in the homogeneous
761 directions, x and z, is arbitrary. In the inset the time-series at y + = 0.017 (blue)
762 and y + = 15.4 (red) are highlighted and compared in the range ti ∈ [300, 800].
763 (b) Graphical representation of the networks extracted from the time-series of
764 panel (a).
765 Fig. 5 Cumulative degree distributions, P k , averaged over the grid-points in the
766 homogeneous x and z directions, reported in a linear-log plot. The distribu-
767 tions close to the vertical axis (highlighted in the inset) correspond
768 to networks built on shuffled time-series at wall-normal distances:
769 (×), y + = 0.017; (◦), y + = 15.4; (), y + = 106.2; (), y + = 996.3. The
770 fittings were performed as Pk ∼ exp(γk), with a trust-region method of opti-
771 mization. The resulting values of γ are (−1.01, −1.08, −2.85, −2.86) · 10−2
772 from y + = 0.017 to y + = 996.3, respectively; the slope for the shuffled
773 series is about γ ≈ 2 · 10−1 . The coefficient of determination, R2 , of the
774 fittings is always above 0.99.
775 Fig. 6 Average metrics Tr, d  
1n , and k evaluated from time-series extracted
776 from the velocity field, (u∗ , v ∗ , w∗ ).
777 Fig. 7 Cumulative degree distributions of the networks built on the time-series
778 of three velocity components, (u∗ , v ∗ , w∗ ), and averaged over the grid-points in
779 the homogeneous directions. (a): y + = 0.0017, (b): y + = 15.4, (c): y + = 996.3.
780 Fig. B1 Averaged metric behaviors, Tr, cd  
1n , ck, as function of y
+
for three
∗
781 different time sampling of the streamwise velocity time-series, u . The curves
782 are obtained with cδt, where the sampling is c = 1 (blue), which corresponds to
783 the case in Fig. 3, c = 2 (red), c = 4 (green).
784 Fig. C1 Metrics behaviors of networks built on the streamwise ve-
785 locity component, u∗ , and extracted at 48 different uniformly spaced
786 (x, z) locations. The black plots correspond to the averaged behavior,
787 as shown in Fig. 3.

20
Table 1: Scheme of the ways two time-series, T S(1) and T S(2) , can differ and corresponding
behaviors of the global network-metrics, T r and d1n .

Cases Temporal structure Metric behaviors TS1 TS2

features
Case A Peaks occur more frequently in T r(2) ≈ T r(1)
T S(2) than in T S(1) d1n (2) < d1n (1)

Case B T S(2) is more irregular than T r(2) < T r(1)

T S(1) d1n (2) ≈ d1n (1)

Case C Peaks occur more frequently in T r(2) < T r(1)

T S(2) than in T S(1) , and T S(2) d1n (2) < d1n (1)
is more irregular than T S(1)

Case D Peaks occur less frequently in T r(2) < T r(1)

T S(2) than in T S(1) , and T S(2) d1n (2) > d1n (1)
is more irregular than T S(1)

21
 (a)

1 10 20 30 40 50 60 70 80 90 100

(b)

1 10 20 30 40 50 60 70 80 90 100

(c)

1 10 20 30 40 50 60 70 80 90 100

(d)

1 10 20 30 40 50 60 70 80 90 100

Fig. 1:

22
 1 50
Tr
<d >
1n
0.8 40

0.6 30

0.4 20

0.2 10

0 0
(a) (b) (c) (d)

Fig. 2:

Averaged Metrics

Degree
Transitivity
200 0.5
0.017 Mean Link Distance

15.4
150 0.45

996.3
100 0.4

50 0.35

-2 -1 0 1 2 3
10 10 10 10 10 10

Fig. 3:

23
(a)
2
0
-2

2
0
-2

2
0
-2

0 400 800 1200 1600 2000 2400 2800 3200 3600 4000

Fig. 4:

100
10 0

-2
10 10 -2

10 -4
0 20 40 60 80
-4
10

y +=0.017
-6 y +=15.4
10
y +=106.2
y +=996.3
0 200 400 600 800 1000 1200 1400 1600 1800 2000

Fig. 5:

24
(a) (b) (c)
0.5 200
u 200 u u
v v v
0.45 w w 150 w
150

0.4 100 100

0.35 50 50

0.3 0 0
10-2 10-1 100 101 102 103 10-2 10-1 100 101 102 103 10-2 10-1 100 101 102 103

Fig. 6:

(a) (b) (c)

10
0
10
0 100 u
u u
v v v
-2 w -2 w w
10 10 10-2

-4 -4
10 10
10-4

10-6 10
-6

10-6
0 1000 2000 0 500 1000 1500 0 200 400 600

Fig. 7:

(a) (b) (c)

0.5 200
200

0.45 150
150

0.4 100 100

0.35 50 50
-2 -1 0 1 2 3 -2 -1 0 1 2 3
10-2 10-1 100 101 102 103 10 10 10 10 10 10 10 10 10 10 10 10

Fig. B8:

(a) (b) (c)

0.7 300 250

0.6 200
200
0.5 150

0.4 100
100
0.3 50

0.2 0 0
10-2 10-1 100 101 102 103 10-2 10-1 100 101 102 103 10-2 10-1 100 101 102 103

Fig. C9:

25