betaFeedback

WorksPhysics and Astronomy

Effect of finite Reynolds number on self-similar crossing statistics and fractal measurements in turbulence

Abstract

Stochastic simulations are used to create synthetic one-dimensional telegraph approximation (TA) signals based on turbulent zero crossings, where the interval between crossings is governed by a power law probability distribution with exponent αα. The power law exponent is determined for statistics of simulated TA signals, namely the box-counting fractal dimension D1D_1, energy spectrum exponent βTAβ_{TA}, and an intermittency exponent μTAμ_{TA}. For the binary TA signal with no variability in amplitude, the parameters are related linearly as D1=2−βTA=1−μTAD_1 = 2 - β_{TA} = 1 - μ_{TA}. The relations are unchanged if the crossing interval distribution has a finite power law region (i.e. inertial subrange) representing a flow with finite Reynolds number. However, the finite distribution yields statistics that are not truly scale-invariant, and distorts the linear relation between the statistic exponents and αα. The behavior is due to finite-size effects apparent from the survival function, or the complementary cumulative distribution, which for finite Reynolds number is only approximately self-similar and has an effective exponent differing from αα. An expression presented for the effective exponent recovers the expected relations between αα and the TA statistics. The findings demonstrate how a finite Reynolds number can affect indicators of self-similarity, fractality, and intermittency observed from single-point measurements.

Introduction

Turbulent fluid dynamics is one of relatively few fields where the existence of self-similarity (scale-invariance) is supported by both theory [1] and extensive observation [2]. The most well-known power law in turbulence describes self-similarity within the energy spectrum: for intermediate scales known collectively as the inertial subrange, the fluctuating energy decays as E∼f−βE \sim f^{-\beta}. Here ff is the frequency (or wavenumber) and β\beta is the spectral exponent. From a statistics perspective, mechanistic concepts underlying a power law include random walks [3, 4], fractal geometries [5, 6], and self-organized criticality [7, 8, 9].

Attempts to relate self-similarity in turbulence to concepts such as fractal geometries have provided both promising [e.g., 10, 11] and conflicting [12, 13, 14, 15] evidence, where the latter studies observed a scale-dependent fractal dimension for isosurfaces and iso-crossings. The conflicting evidence may be explained by a combination of several challenges in isolating specific self-similar features in turbulence. For isosurfaces and iso-crossings near the mean value [12, 13], diffusive events are over-represented compared to level sets farther from the mean, leading to a relatively narrower inertial subrange of scales [see, e.g., 16]. The same study [16] also demonstrated how scalar ramp-cliff patterns can influence the box-counting fractal dimension estimated from scalar concentration fields [12, 14, 15].

An additional factor – and the focus of the present study – is the restriction of the self-similar behavior to a finite range of scales based on the Reynolds number ReRe. The finite inertial subrange is a form of truncated power law, where finite-size effects cause cumulative statistics to deviate from a true power law [17, 18]. Previous works have briefly mentioned how finite-size effects can lead to an apparent scale-dependent fractal dimension [19, 20], but these effects have not otherwise been closely evaluated for power laws in turbulence. Specifically, it is not clear how the extent of the inertial subrange influences the relation between power law exponents of various statistics including the energy spectrum and fractal dimension.

One strategy to assess power law relations in turbulence is to reduce the measured fluctuating quantity to a one-dimensional crossing signal. The properties of isosurface geometries are ideally evaluated in three dimensions, but the reduction of the isosurface to its crossings of a one-dimensional transect is often an experimental necessity, particularly for point measurements in the high-Reynolds-number atmospheric surface layer. While the signal can be constructed based on crossings of any arbitrary level set value, the most common signal is defined using zero crossings of the velocity fluctuations [21, 22, 23]. The telegraph approximation (TA) signal [24, 25] based on these zero crossings is 1 when the fluctuating velocity is positive, and is 0 when the velocity is negative. The interval between crossings is known as the interpulse period [22, 25] or persistence [26, 27, 28]. The inertial subrange of the full velocity signal is similarly present in the zero-crossing TA signal. Among other statistics, the spectrum of the TA signal and the probability distribution of the interpulse period both exhibit self-similarity in the inertial subrange for sufficiently large ReRe [see, e.g., 24, 25, 29, 30].

Available measurements of simplified turbulent crossing signals present their own challenges for studying finite-size effects. First, a limited range of β\beta values is observed from turbulent measurements, which precludes empirical fits across the parameter space of β\beta and other exponent values. Second, even high-Reynolds-number flows have a narrow inertial subrange in the context of finite-size effects, as will be seen in later results. For instance, measurements in atmospheric flows typically have Taylor microscale Reynolds number Reλ∼O(103)Re_{\lambda} \sim O(10^{3}), corresponding to no more than three decades (i.e. orders of magnitude) of self-similar inertial subrange [2].

To properly explore the full parameter space of Reynolds number and power law exponents, synthetic TA signals are constructed here using stochastic simulations based on idealized interpulse distributions. Power law statistics are computed across a range of interpulse exponent values, and the effective Reynolds number is also varied by truncating the power law region of the interpulse distributions. The statistics evaluated here are the fractal box-counting dimension, the energy spectrum, and an intermittency parameter. The simulations are analogous to a Monte Carlo analysis, except the goal is to identify the ensemble average of statistics rather than their uncertainty. The idealized simulations are purely stochastic and do not directly model any governing physics. This approach assumes the original signal is self-similar and identifies the consequence of a finite Reynolds number (truncated power law) on crossing statistics of the signal. While the analysis is discussed in the context of turbulent flows, the findings are generally applicable to any finite self-similar process.

The study is organized into the following sections: Sec. II describes the stochastic simulations, Sec. III presents results of the simulations, Sec. IV introduces a correction for finite Reynolds number, and Sec. V summarizes the findings.

Stochastic simulations

The premise of the stochastic simulations is to create a synthetic TA signal s(t)s(t) defined by a sequence of “events”, where the interval τ\tau between events is governed by a power law probability density function (PDF). In the context of a turbulent zero-crossing signal, each event represents the position tt (in space or time) where the fluctuating quantity ss crosses zero. The design of the simulations is detailed in the sections below for both “unbounded” and truncated interpulse power laws. The unbounded power law is used as a control case, and the truncated power laws approximate the interpulse distribution for three Reynolds numbers spanning the range Reλ∼O(102−104)Re_{\lambda} \sim O(10^{2} - 10^{4}).

Unbounded power law

For the unbounded control case, the TA interpulse distribution is modeled as a power law that exceeds the extent of the simulation domain. The power law PDF is defined as

PDF(τ)=11−(τ2/τ1)1−αα−1τ1(ττ1)−α,(1)PDF(\tau) = \frac{1}{1-(\tau_{2}/\tau_{1})^{1-\alpha}}\frac{\alpha-1}{\tau_{1}}\left(\frac{\tau}{\tau_{1}}\right)^{-\alpha}, \tag*{(1)}

where τ1\tau_{1} and τ2\tau_{2} are the minimum and maximum values of the distribution, respectively, and α\alpha is the distribution exponent. The power law in Eq. (1) is equivalent to a Pareto distribution with exponent α−1\alpha-1. The integral of the PDF is equal to unity – as required by the PDF definition – only for α>1\alpha> 1 and if a minimum value is imposed. The integral is infinite and the PDF is not well-defined for α≤1\alpha\leq1. The minimum τ1=1\tau_{1}=1 is used here for simplicity. While a maximum value τ2\tau_{2} is typically ∞\infty for an unbounded power law, τ2=10300\tau_{2}=10^{300} is employed to avoid infinite values in the simulations. This value is close to the largest definable number in double-precision floating-point format, and yields 300 decades of power law. An example PDF is shown in Figure 1(a).

Example synthetic telegraph approximation signal

Figure 1. Example synthetic telegraph approximation (TA) signal constructed from a power law probability distribution PDF∼τ−αPDF \sim\tau^{-\alpha}. (a) Probability distribution for the interval τ\tau between events in a signal following Eq. (1). (b) Determination of a single τ\tau value using inverse transform sampling of the cumulative distribution CDFCDF and Eq. (3). (c) Position tt of events based on ten τ\tau values. (d) TA signal s(t)s(t) whose value 0 or 1 changes at each event.

Inverse transform sampling is used to select values of τ\tau from the distribution. In this approach, the cumulative distribution function

CDF(τ)=11−(τ2/τ1)1−α[1−(ττ1)1−α](2)CDF(\tau) = \frac{1}{1-(\tau_{2}/\tau_{1})^{1-\alpha}}\left[1-\left(\frac{\tau}{\tau_{1}}\right)^{1-\alpha}\right] \tag*{(2)}

is inverted to define τ\tau as a function of CDFCDF:

τ(CDF)=τ1[1−(1−(τ2τ1)1−αCDF)]11−α.(3)\tau(CDF) = \tau_1 \left[1-\left(1-\left(\frac{\tau_2}{\tau_1}\right)^{1-\alpha}CDF\right)\right]^{\frac{1}{1-\alpha}} . \tag*{(3)}

The CDFCDF value is simulated by selecting a random value between 0 and 1, and the corresponding interval τ\tau is determined from Eq. (3) as shown in Figure 1(b). A small sample of events is shown in Figure 1(c), where each event is separated by simulated intervals τ\tau.

The position tt of each event is used to build a synthetic TA signal s(t)s(t). Following the TA definition given in the introduction, the s(t)s(t) value alternates between 0 and 1 at each event as seen in Figure 1(d). The signal is constructed on a discrete domain between 0 and tmax⁡=106t_{\max}=10^6. The resolution between points on the domain is τ1=1\tau_1=1, thus allowing for six decades of statistics in the signal. While PDF(τ)PDF(\tau) represents 300 decades of power law, only six decades can be observed in s(t)s(t) due to the limited domain. However, based on later results, the effect of truncation by the domain size is negligible for this case such that it is considered unbounded for the purpose of the study.

Later results present simulations for one hundred α\alpha values between 1.02 and 3. The value α=1\alpha=1 is excluded because the PDF is not well-defined as discussed above. The approximate exponent for a turbulent TA signal is α≈1.5\alpha\approx1.5 [24]. For a given α\alpha, intervals τ\tau are simulated until tmax⁡t_{\max} is exceeded to ensure the signal is fully populated. Statistics are thereafter calculated using s(t)s(t), and the process is repeated until the ensemble average statistics are converged. The number of signals contributing to each statistic varies between 10210^2 for large α\alpha and 10410^4 for small α\alpha. The latter returns sparse signals, which require a larger number of realizations to converge statistics.

Truncated power law

For a turbulent TA signal, the interpulse PDF is only a power law for τ\tau values within the inertial subrange of scales [24, 25]. Smaller interpulse periods are well approximated by a lognormal distribution [32, 33], and larger τ\tau values follow an exponential cutoff in boundary layer flows [22, 25, 27]. The same PDF shape – a blend of lognormal, power law, and exponential distributions – is used here to simulate the effect of a truncated power law region. The case represents a weak power law because a power law expression does not describe the full range of values in the distribution [34].

Same as for the unbounded power law, inverse transform sampling is used to simulate s(t)s(t) based on a random selection of values for τ\tau. For simplicity, the lognormal portion of the distribution is defined using the parameters μ∗=1\mu_*=1 and σ∗2=1\sigma_*^2=1, which respectively correspond to the mean and variance of log⁡(τ)\log(\tau). From these parameters, the mode of the lognormal curve is fixed at τ1=1\tau_1=1. The resulting PDF for τ\tau is given by the piecewise function

PDF(τ)={C1τe−(log⁡(τ)−1)22τ≤eαC2τ−αeα<τ≤bC3e−λτb<τ≤τ2(4)PDF(\tau) = \begin{cases} \frac{C_{1}}{\tau}e^{-\frac{(\log(\tau)-1)^{2}}{2}} & \tau\le e^{\alpha} \\ C_{2}\tau^{-\alpha} & e^{\alpha}<\tau\le b \\ C_{3}e^{-\lambda\tau} & b<\tau\le\tau_{2} \end{cases} \tag*{(4)}

The transition from the lognormal to the power law curve occurs at eαe^{\alpha}. This point corresponds to dlog⁡(PDF)/dlog⁡(τ)=αd\log(PDF)/d\log(\tau)=\alpha along the lognormal curve, ensuring a smooth transition to the power law. The transition to the exponential cutoff is imposed at b=10xeαb=10^{x}e^{\alpha}, where xx is the desired number of power law decades. The exponential parameter λ=α/b\lambda=\alpha/b enforces a smooth transition to the exponential cutoff, i.e. dlog⁡(PDF)/dlog⁡(τ)=αd\log(PDF)/d\log(\tau)=\alpha at τ=b\tau=b. The factors are defined as

C1=[C4+eα2−121−αC5−e12(α2−1)+ααbα−1(e−λτ2−e−α)]−1C2=C1eα2−12C3=C2(eb)αC4=π2(erf⁡(α−12)+1)C5=b1−α−eα(1−α).(5)\begin{aligned} C_{1} &= \left[C_{4}+\frac{e^{\frac{\alpha^{2}-1}{2}}}{1-\alpha}C_{5}-\frac{e^{\frac{1}{2}(\alpha^{2}-1)+\alpha}}{\alpha b^{\alpha-1}}\left(e^{-\lambda\tau_{2}}-e^{-\alpha}\right)\right]^{-1} \\ C_{2} &= C_{1}e^{\frac{\alpha^{2}-1}{2}} \\ C_{3} &= C_{2}\left(\frac{e}{b}\right)^{\alpha} \\ C_{4} &= \sqrt{\frac{\pi}{2}}\left(\operatorname{erf}\left(\frac{\alpha-1}{\sqrt{2}}\right)+1\right) \\ C_{5} &= b^{1-\alpha}-e^{\alpha(1-\alpha)}. \tag*{(5)} \end{aligned}

The constants C2C_{2} and C3C_{3} are defined relative to C1C_{1} to ensure the amplitude of PDF(τ)PDF(\tau) is matched at the transition points, and C1C_{1} is defined to achieve ∫0τ2PDF(τ)=1\int_{0}^{\tau_{2}}PDF(\tau)=1. The cumulative distribution corresponding to Eq. (4) is

CDF(τ)={C1π2(erf⁡(log⁡(τ)−12)+1)τ≤eαC1C4+C21−α(τ1−α−eα(1−α))eα<τ≤bC1C4+C2C51−α+C3λ(e−λτ−e−α)b<τ≤τ2(6)CDF(\tau)= \begin{cases} C_{1}\sqrt{\frac{\pi}{2}}\left(\operatorname{erf}\left(\frac{\log(\tau)-1}{\sqrt{2}}\right)+1\right) & \tau\le e^{\alpha} \\ C_{1}C_{4}+\frac{C_{2}}{1-\alpha}\left(\tau^{1-\alpha}-e^{\alpha(1-\alpha)}\right) & e^{\alpha}<\tau\le b \\ C_{1}C_{4}+\frac{C_{2}C_{5}}{1-\alpha}+\frac{C_{3}}{\lambda}\left(e^{-\lambda\tau}-e^{-\alpha}\right) & b<\tau\le\tau_{2} \end{cases} \tag*{(6)}

Finally, the inversion of Eq. (6) yields the transform equation used to simulate the finite power law:

τ={exp⁡[2erfinv⁡(π2CDF−1)+1]CDF≤C1C4[1−αC2(CDF−C1C4)+eα(1−α)]11−αC1C4<CDF≤C1C4+C2C51−α−1λlog⁡[λC3(C1C4+C2C51−α−CDF)+e−α]CDF>C1C4+C2C51−α(7)\tau= \begin{cases} \exp\left[\sqrt{2}\operatorname{erfinv}\left(\sqrt{\frac{\pi}{2}}CDF-1\right)+1\right] & CDF\le C_{1}C_{4} \\ \left[\frac{1-\alpha}{C_{2}}(CDF-C_{1}C_{4})+e^{\alpha(1-\alpha)}\right]^{\frac{1}{1-\alpha}} & C_{1}C_{4}<CDF\le C_{1}C_{4}+\frac{C_{2}C_{5}}{1-\alpha} \\ -\frac{1}{\lambda}\log\left[\frac{\lambda}{C_{3}}\left(C_{1}C_{4}+\frac{C_{2}C_{5}}{1-\alpha}-CDF\right)+e^{-\alpha}\right] & CDF>C_{1}C_{4}+\frac{C_{2}C_{5}}{1-\alpha} \end{cases} \tag*{(7)}

In Eqs. (6) and (7), erf⁡(x)\operatorname{erf}(x) and erfinv⁡(x)\operatorname{erfinv}(x) refer to the error function and its inverse, respectively, and the two notations for the exponential function exe^{x} and exp⁡(x)\exp(x) are used interchangeably for readability.

Figure 2 shows a synthetic truncated power law. An example turbulent zero-crossing (interpulse) PDF is included for reference. The interpulse is estimated from hotwire anemometry measurements of boundary layer turbulence [31], and exhibits a similar shape to the simulated PDF with one decade of power law. Using the inverse transform of CDF in Figure 2(b) and Eq. (7), the truncated signals are simulated on the same domain and for the same range of α\alpha values as the unbounded case.

Power law distribution truncated by lognormal and exponential curves

Figure 2. Power law distribution whose extent is truncated by lognormal and exponential curves. The example exhibits one decade of power law, corresponding to Reynolds number Reλ∼102Re_{\lambda} \sim10^{2}. (a) Probability distribution for τ\tau following Eq. (4). The gray line is an example interpulse distribution from boundary layer turbulence measurements [31], shifted for visual comparison. (b) Inverse transform sampling of CDFCDF following Eq. (7).

While the synthetic and experimental distributions in Figure 2 appear to be qualitatively similar, the idealized distribution is not designed to reproduce aspects of a zero-crossing signal that are outside the scope of the work. For instance, the parameters for the lognormal region do not produce the correct scaling for the mean value of τ\tau [22], and the direct transitions between the different scaling regions may not accurately reflect experimental observations.

The breadth of the simulated power law region can be related directly to the Reynolds number. Assuming the inertial subrange spans O(10η)O(10\eta) to O(L)O(L) [35], where η\eta and LL are the Kolmogorov microscale and the integral scale, respectively, the extent of the inertial subrange is O(0.01Reλ3/2)O(0.01Re_\lambda^{3/2}) [2]. The number of decades in the inertial subrange is therefore 32log⁡10(Reλ)−2\frac{3}{2}\log_{10}(Re_\lambda)-2. Results for distributions with one (Reλ∼102Re_\lambda\sim10^2), three (Reλ∼103Re_\lambda\sim10^3), and five (Reλ∼104Re_\lambda\sim10^4) decades of power law are presented herein. The six decades of resolution in the simulated domain yield Reλ≳105Re_\lambda\gtrsim10^5 for the unbounded case.

Results

Fractal Dimension

The fractal dimension of the simulated signal is estimated here applying the box counting method [see, e.g., 6, 36, 37] to the event positions featured in Figure 1(c). In the box-counting approach, the domain is discretized into segments of size Δt\Delta t and the number of segments NN containing at least one event is counted. If the resulting dependency N(Δt)N(\Delta t) follows a power law

N(Δt)∼Δt−D1,(8)N(\Delta t) \sim\Delta t^{-D_1}, \tag*{(8)}

the signal is considered statistically self-similar. Whether the signal is also considered to be a fractal object depends on the definition, as fractality is sometimes reserved for geometric shapes. The subscript 1 is adopted for the fractal dimension D1D_1 in Eq. (8) because the estimate is made on a one-dimensional signal. The value for D1D_1 is bounded between 0 and 1. These limits correspond to a signal with τ>Δt\tau> \Delta t (for D1=0D_1 = 0) or τ<Δt\tau< \Delta t (for D1=1D_1 = 1) for all intervals τ\tau across the tested range of Δt\Delta t.

Example box-counting results are shown in Figure 3(a,Figure 3b) for a range of α\alpha and power law truncation. The expected power law in Eq. (8) is approximately observed for the unbounded distribution. However, in Figure 3(b) N(Δt)N(\Delta t) becomes increasingly dissimilar from a power law as the self-similar region in PDF(τ)PDF(\tau) is increasingly truncated. This trend is consistent with similar Monte Carlo simulations that showed D1(Δt)D_1(\Delta t) to vary with Δt\Delta t for any truncated power law due to finite-size effects [20]. As a result, a constant fractal dimension can only be achieved in an approximate sense for finite Reynolds number flows. The absence of a true power law in Figure 3(b) is further discussed in § 0.4.

Statistics for the fractal dimension $D_1$ estimated via box counting on simulated TA signals

Figure 3. Statistics for the fractal dimension D1D_1 estimated via box counting on simulated TA signals. (a) Box count N∼Δt−D1N \sim\Delta t^{-D_1} for the unbounded case and α=1.1\alpha= 1.1 (dotted line), α=1.5\alpha= 1.5 (solid), and α=2\alpha= 2 (dashed). (b) Box count for α=1.5\alpha= 1.5 and varying Reynolds number, where the vertical lines delineate the power law region for the Reλ∼102Re_\lambda\sim10^2 case. (c) Fractal dimension D1D_1 as a function of α\alpha with D1=α−1D_1 = \alpha- 1 (line) for reference. The shaded region in (c) corresponds to the typical value α≈1.5\alpha\approx1.5 for a turbulent TA signal. In this and later figures, the legend indicates the number of power law decades in PDF(τ)PDF(\tau) and the order of the equivalent Reynolds number ReλRe_\lambda based on the Taylor microscale.

Despite the departure from a power law, the dimension D1D_1 is estimated by fitting Eq. (8) to the Figure 3(a,Figure 3b) curves assuming D1D_1 is constant. The fit is performed within the range of Δt\Delta t corresponding to the self-similar region in PDF(τ)PDF(\tau). The fitted values for D1D_1 across the tested range of α\alpha are shown in Figure 3(c). The error bars correspond to the change in D1D_1 when the region where the power law is fitted is shifted by a factor of two in either direction. The error bars therefore increase as the dependence D1(Δt)D_1(\Delta t) increases. The large error bars corresponding to small α\alpha in Figure 3(c) reflect the dissimilarity from a true power law observed in Figure 3(b). The same method is used to calculate the error bars in later figures.

For the unbounded PDF case in Figure 3(c), D1(α)D_1(\alpha) follows a linear trend D1=α−1D_1 = \alpha- 1 up to approximately α≈1.5\alpha\approx1.5, where the linear relation matches previous predictions [19, 20]. Above α≈1.5\alpha\approx1.5, D1D_1 asymptotically approaches 1. The same asymptotic behavior is observed for the truncated distributions representing finite Reynolds number. However, the results for small α\alpha depart from the linear relation as the equivalent Reynolds number decreases. The reason for the departure is related to the trends in Figure 3(b) and is discussed in § 0.4.

Energy spectrum

The energy spectrum is defined as ETA(f)∼∣s^(f)∣2E_{TA}(f) \sim|\hat{s}(f)|^2, where s^(f)\hat{s}(f) is the Fourier transform of s(t)s(t) in frequency or wavenumber space. Prior to computing the transform, s(t)s(t) is multiplied by a Hamming window filter whose length matches the domain size tmax⁡t_{\max}. The signal is also zero-padded. The window filter and zero-padding mitigate aliasing in s^(f)\hat{s}(f).

Using the same format as Figure 3, the resulting energy spectra are shown in Figure 4. Spectra for the truncated distributions in Figure 4(b) exhibit stronger self-similarity than the box counts in Figure 3(b). The Fourier transform efficiently isolates local (in scale) contributions to the variance. In contrast, the box-counting measures a cumulative effect capturing all intervals smaller than the given Δt\Delta t. For the cumulative statistics, the non-power-law behavior is spread across scales to the expected self-similar region.

Statistics for the energy spectrum power law exponent $\beta_{TA}$ of simulated TA signals

Figure 4. Statistics for the energy spectrum power law exponent βTA\beta_{TA} of simulated TA signals. (a) Spectrum ETAE_{TA} for the unbounded case and α=1.1\alpha= 1.1 (dotted line), α=1.5\alpha= 1.5 (solid), and α=2\alpha= 2 (dashed). (b) Spectrum for α=1.5\alpha= 1.5 and varying Reynolds number, where the vertical lines delineate the power law region for the Reλ∼102Re_\lambda\sim10^2 case. (c) Exponent βTA\beta_{TA} as a function of α\alpha with βTA=3−α\beta_{TA} = 3 - \alpha (line) for reference. The shaded region in (c) corresponds to the typical value α≈1.5\alpha\approx1.5 for a turbulent TA signal.

The spectral exponent βTA\beta_{TA} is estimated by fitting the power law ETA∼f−βE_{TA} \sim f^{-\beta} to each individual spectrum. The

“TATA” subscript is adopted because βTA≈4/3\beta_{TA} \approx4/3 observed for turbulent flows differs from the value β≈5/3\beta\approx5/3 for the full signal [29]. The dependency of βTA\beta_{TA} on α\alpha is shown in Figure 4(c). The apparent “roughness” of the curves is attributed to the shape of s(t)s(t). Artificial oscillations appear in the energy spectrum when the sinusoidal basis functions of the Fourier transform are used to decompose the discontinuous signal. These oscillations may propagate to βTA\beta_{TA} as the fitted power law region varies with α\alpha.

As before, the error bars are largest for the lowest equivalent Reynolds number due to the curvature of ETA(f)E_{TA}(f) immediately adjacent to the expected self-similar region. The trend for small α\alpha observed in Figure 3(c) is similarly present for the spectrum exponent. The unbounded case follows a linear relation βTA=3−α\beta_{TA} = 3 - \alpha in Figure 4(c) up to α≈2\alpha\approx2. This relation is applicable to a superposition of Poisson processes [9], and appears similarly applicable to the self-similar process simulated here. For α>2\alpha> 2, the results slowly deviate from the linear relation and there is agreement across cases. It is assumed that βTA\beta_{TA} asymptotically approaches 0 as α\alpha increases, but this trend cannot be confirmed due to the limited tested range of α\alpha.

Intermittency

The intermittency is another quantitative measure of variability in the distribution of events. Intermittency can be parameterized using Obukhov’s local moving average [38]

χ(t,Δt)=1Δt∫tt+Δt∣ds2dt∣dt.(9)\chi(t,\Delta t) = \frac{1}{\Delta t}\int_t^{t+\Delta t}\left|\frac{\mathrm{d}s^2}{\mathrm{d}t}\right|\mathrm{d}t. \tag*{(9)}

Given the amplitude of s(t)s(t) is invariable, the integral corresponds to the number of events occurring within “windows” of size Δt\Delta t. The intermittency is quantified using the scaling [24, 39, 40]

⟨χ2⟩⟨χ⟩2∼Δt−μTA,(10)\frac{\langle\chi^2\rangle}{\langle\chi\rangle^2} \sim\Delta t^{-\mu_{TA}}, \tag*{(10)}

where angled brackets ⟨⋅⟩\langle\cdot\rangle indicate an ensemble average across tt. Equation (10) is defined here using the second-order moment, but the same principle can be applied to higher-order moments. The exponent μTA\mu_{TA} represents how the variability in number of events across windows changes as the window size is increased.

The intermittency parameter statistics are shown in Figure 5. Results for the truncated distributions in Figure 5(b) exhibit a lack of self-similarity. Same as for the box counting methodology, Eq. (9) accounts for all intervals smaller than Δt\Delta t, resulting in a cumulative metric.

Statistics for the intermittency exponent $\mu_{TA}$ of simulated TA signals

Figure 5. Statistics for the intermittency exponent μTA\mu_{TA} of simulated TA signals. (a) Intermittency parameter ⟨χ2⟩/⟨χ⟩2\langle\chi^2 \rangle/\langle\chi\rangle^2 for the unbounded case and α=1.1\alpha= 1.1 (dotted line), α=1.5\alpha= 1.5 (solid), and α=2\alpha= 2 (dashed). (b) Intermittency for α=1.5\alpha= 1.5 and varying Reynolds number, where the vertical lines delineate the power law region for the Reλ∼102Re_\lambda\sim10^2 case. (c) Exponent μTA\mu_{TA} as a function of α\alpha with μ=2−α\mu= 2 - \alpha (line) for reference. The shaded region in (c) corresponds to the typical value α≈1.5\alpha\approx1.5 for a turbulent TA signal.

Values for μTA\mu_{TA}, fitted using Eq. (10), are plotted as a function of α\alpha in Figure 5(c). The curves follow the same trends as D1D_1 and βTA\beta_{TA}. A linear relation μTA=2−α\mu_{TA}=2-\alpha is observed for the unbounded distribution and small α\alpha values. Truncating the power law distribution to represent a finite Reynolds number leads to a departure from the linear relation. For larger α\alpha, all cases deviate from μTA=2−α\mu_{TA}=2-\alpha as μ\mu asymptotically approaches zero.

Exponent relations

The relations between power law exponents D1D_1, βTA\beta_{TA} and μTA\mu_{TA} are plotted in Figure 6 for the tested range of α\alpha. The lines correspond to the linear trends in panel (c) of Figure 3, Figure 4, and Figure 5. For visual clarity, the error bars from previous figures are only reproduced in Figure 6 for the truncated case with one decade of self-similarity.

Relations between power law exponents of the presented statistics for simulated TA signals

Figure 6. Relations between power law exponents of the presented statistics for simulated TA signals. (a) Fractal dimension D1D_1 and intermittency exponent μTA\mu_{TA}, compared to μTA=D1−1\mu_{TA}=D_1-1 (line). (b) Dimension D1D_1 and spectrum exponent βTA\beta_{TA}, compared to βTA=2−D1\beta_{TA}=2-D_1 (line). (c) Exponents μTA\mu_{TA} and βTA\beta_{TA}, compared to βTA=μ+1\beta_{TA}=\mu+1 (line).

A robust inverse relation between μTA\mu_{TA} and D1D_1 is observed in Figure 6(a). The results are in close agreement with the prediction μTA=1−D1\mu_{TA}=1-D_1 [40, 41], which can be derived via the correlation dimension [42]. The linear trend is invariant to the Reynolds number (i.e. the truncation of the power law PDF), and the primary difference across cases is the observed range in μTA\mu_{TA} and D1D_1 values.

Linear trends βTA=2−D1\beta_{TA}=2-D_1 and βTA=μTA+1\beta_{TA}=\mu_{TA}+1 also exist for the spectral exponent in Figure 6(b,Figure 6c). However, the linearity is limited to βTA≈1.2\beta_{TA}\approx1.2. The behavior for smaller βTA\beta_{TA} is attributed to the slower rate at which βTA\beta_{TA} asymptotically approaches zero, compared to the corresponding rates for D1D_1 and μTA\mu_{TA}.

The results for PDF(τ)PDF(\tau) with Reλ∼102Re_\lambda\sim10^2 are visibly offset from the other cases in Figure 6(b,Figure 6c). The difference in βTA\beta_{TA} is approximately 0.2, which is within the extent of the error bars. The difference may therefore be due to the lack of self-similarity in the statistics and the precise range chosen to fit the power law exponents. The result highlights the challenge in recovering the expected relations when the Reynolds number yields a narrow inertial subrange and the cumulative statistics (D1D_1, μTA\mu_{TA}) lose the signature of self-similarity.

Aside from the offset, the relations in Figure 6 do not depend on the Reynolds number and the bounds of the power law PDF. Direct linear relations can be expected between power law statistics, even if the governing distribution PDF(τ)PDF(\tau) is self-similar across a finite range of values. The effect of Reynolds number on the power law exponent relations is therefore limited to the altered connection between the statistics and the underlying probability exponent α\alpha.

Effect of finite Reynolds number

The probability distribution definitions in Eqs. (1) and (2) impose a finite maximum value τ2\tau_2. As a result, intervals above τ2\tau_2 are under-sampled relative to an infinite power law distribution [17, 18]. Values within the power law region may be under- or over-sampled, depending on the shape of the cutoff regions bounding the power law.

The consequence of the sampling discrepancy is apparent in the survival function 1−CDF(τ)1-CDF(\tau), also known as the complementary cumulative distribution. For the power law in Eq. (2), the survival function can be expressed as

1−CDF(τ)=(τ/τ1)1−α−(τ2/τ1)1−α1−(τ2/τ1)1−α.(11)1-CDF(\tau)=\frac{\left(\tau/\tau_1\right)^{1-\alpha}-\left(\tau_2/\tau_1\right)^{1-\alpha}}{1-\left(\tau_2/\tau_1\right)^{1-\alpha}}. \tag*{(11)}

Equation (11) is only a power law for infinite Reynolds number with τ2/τ1→∞\tau_2/\tau_1\to\infty, and otherwise has a finite additive constant distorting the probability that the next interval exceeds a given value of τ\tau. The distortion has led to the use of more generalized distributions such as the Zipf–Mandelbrot law [43] to describe cumulative statistics of truncated power laws.

The survival functions for three values of α\alpha are shown in Figure 7(a,Figure 7b,Figure 7c). The trends are consistent with the previously observed dependencies on α\alpha: for finite Reynolds number with a truncated distribution the survival function only approximates a power law, and the departure from a power law is largest for small α\alpha values.

Effective exponent plots showing survival functions and associated exponents

Figure 7. Effective exponent αe\alpha_e resulting from the distortion of the survival function by the truncated power law, i.e. finite Reynolds number. Rows correspond to the survival function 1−CDF(τ)1-CDF(\tau) (a,b,c) and associated exponent αe\alpha_e (d,e,f) following Eq. (??). Columns correspond to values of α\alpha: 1.1 (a,d); 1.5 (b,e); and 2 (c,f).

To quantify the deviation from a true power law, the effective exponent αe\alpha_e can be calculated from the slope of the curves in Figure 7(a,Figure 7b,Figure 7c). Mathematically, the exponent is

1−αe=ddlog⁡(τ)[log⁡(1−CDF)].(12)1-\alpha_e=\frac{d}{d\log(\tau)}\left[\log(1-CDF)\right]. \tag*{(12)}

The chain rule can be used to simplify the derivative operation as d/dlog⁡(τ)=τd/dτd/d\log(\tau)=\tau d/d\tau. Using Eq. (11) to compute the derivative of log⁡(1−CDF)\log(1-CDF), the effective exponent can be expressed as

αe(τ)=1+τPDF(τ)1−CDF(τ).(13)\alpha_e(\tau)=1+\frac{\tau PDF(\tau)}{1-CDF(\tau)}. \tag*{(13)}

Importantly, the exponent changes as a function of τ\tau, reflecting the fact that the survival function is not a true power law for a truncated distribution. The effective exponent is shown in Figure 7(d,Figure 7e,Figure 7f) for each simulated case. The effective exponent is αe≈α\alpha_e\approx\alpha for the unbounded case, indicating the domain size is sufficiently large to approximately represent an infinite power law for this study.

Within the range of τ\tau where a power law is expected, αe\alpha_e in Figure 7 becomes increasingly larger than α\alpha as α\alpha decreases and as the Reynolds number decreases. The decrease in α\alpha and Reynolds number both represent an increase in the portion of the CDFCDF that is “missing” due to the truncated upper limit of the distribution. For a power law defined only in the range between τ1\tau_1 and τ2\tau_2, αe\alpha_e is

αe(τ)=1+(α−1)(τ/τ1)1−α(τ/τ1)1−α+(τ2/τ1)1−α.(14)\alpha_e(\tau)=1+(\alpha-1)\frac{(\tau/\tau_1)^{1-\alpha}}{(\tau/\tau_1)^{1-\alpha}+(\tau_2/\tau_1)^{1-\alpha}}. \tag*{(14)}

The result αe=α\alpha_e=\alpha is recovered when τ2/τ1→∞\tau_2/\tau_1\to\infty, i.e. Re→∞Re\to\infty. The equation for the truncated distribution in Figure 2 follows a similar form to Eq. (14), except the right-side term in the denominator depends on the prescribed cutoff behavior. In this sense, the lognormal and exponential cutoffs employed here introduce minor quantitative differences in the results, but the scale-dependence of the survival function is a direct consequence of finite-size effects resulting from the finite power law region.

In practice, αe\alpha_e in Eq. (13) can be estimated from discrete histograms approximating the TA interpulse distribution PDF(τ)PDF(\tau). For simplicity, a single representative value of αe\alpha_e is employed here using the average of αe(τ)\alpha_e(\tau) within the inertial subrange where power law statistics are fitted.

Figure 8 compares the power law exponents as a function of α\alpha and αe\alpha_e. Substituting for the effective exponent αe\alpha_e accounts for the trends observed in panel (c) of Figure 3, Figure 4, and Figure 5. Further, the corrected results follow the linear relations exhibited by the unbounded case. The remaining deviation in the Re⁡λ∼102\operatorname{Re}_{\lambda} \sim10^2 case is within the extent of the error bars and may be due to the strong departure from self-similar statistics previously discussed.

Power law exponent as a function of $\alpha$ and $\alpha_e$ for simulated TA signals and boundary layer turbulence measurements

Figure 8. Power law exponent as a function of α\alpha (transparent markers) and the effective exponent αe\alpha_e (opaque) for simulated TA signals. (a) Fractal dimension D1D_1. (b) Spectrum exponent βTA\beta_{TA}. (c) Intermittency exponent μTA\mu_{TA}. (d) Spectrum exponent βTA\beta_{TA} for boundary layer turbulence measurements of the streamwise (●) and wall-normal (×) velocity components.

As a practical example, Figure 8(d) shows the correction of α\alpha for measured turbulent TA signals. The signals were acquired from a range of positions within a wind tunnel boundary layer above both smooth and rough surfaces [31]. The results for the spectrum exponent βTA(αe)\beta_{TA}(\alpha_e) align with the expected linear relation and exhibit reduced scatter compared to βTA(α)\beta_{TA}(\alpha). The observed difference in βTA\beta_{TA} between the streamwise and wall-normal velocity components is due to the smaller integral length for the wall-normal component, leading to a narrower inertial subrange and a greater effect of truncation. The effective exponent αe\alpha_e in Figure 8(d) successfully accounts for this difference.

The discrepancy between βTA(α)\beta_{TA}(\alpha) in Figure 8(d) and βTA=3−α\beta_{TA}=3-\alpha has been previously explained as an effect of intermittency. Specifically, the correction βTA=3−μTA/2−α\beta_{TA}=3-\mu_{TA}/2-\alpha was proposed [24]. The present simulations demonstrate that βTA=3−α\beta_{TA}=3-\alpha is applicable to intermittent, self-similar processes governed by an unbounded power law. Adjusting the unbounded power law in Figure 4 for intermittency would lead to incorrect results. Rather, the relations between the power law statistics and PDF(τ)PDF(\tau) must consider the effective exponent αe\alpha_e in Eq. (13) resulting from a finite Reynolds number. The success of the correction in Figure 8 demonstrates that the distortion of the survival function propagates to statistics based on the simulated signals, and that the resulting parameters D1D_1, βTA\beta_{TA}, and μTA\mu_{TA} depend on the survival function and its effective exponent 1−αe1-\alpha_e.

Summary

For a binary stochastic process described by an unbounded power law probability distribution, linear equations exist to relate the power law exponent of various statistics. The fractal dimension D1D_1, energy spectrum exponent βTA\beta_{TA}, and intermittency exponent μTA\mu_{TA} are related as D1=2−βTA=1−μTAD_1=2-\beta_{TA}=1-\mu_{TA}. The relation between D1D_1 and μ\mu matches the predicted analytical solution. The statistics are also linearly related to the probability exponent α\alpha, e.g. as βTA=3−α\beta_{TA}=3-\alpha, for the range of α\alpha applicable to turbulent signals (α≈1.5\alpha\approx1.5). While a small selection of statistics are evaluated here, similar linear relations and trends are expected for other parameters like the correlation integral [44].

These relations are directly applicable to the telegraph approximation (TA) of turbulent crossing signals and estimates of the fractal dimension from single-point measurements. However, certain aspects of the results depend on the Reynolds number and the extent of the inertial subrange where power law behavior is expected, in which a finite Reynolds number yields truncated power law statistics. Specifically, a finite Reynolds number changes the effective exponent αe\alpha_e of the survival function for the TA interpulse τ\tau, which propagates to ensuing statistics. The original linear relations between α\alpha and the other exponents can be recovered by considering the effective exponent αe\alpha_e in Eq. (13). Yet, cumulative statistics such as the fractal dimension are not self-similar, as αe(τ)\alpha_e(\tau) is strongly scale-dependent when the Reynolds number is small. The departure from self-similarity may be even greater in practice, as the effect of an exponential cutoff on the scale invariance of the power law is not considered here [45]. These finite Reynolds number effects may help to explain experimental observations in turbulence. Previous findings on the scale dependence of the fractal dimension [12, 13] are likely due to a combination of the technical challenges discussed in the introduction [16] and the finite-size effects studied here. Additionally, deviations from βTA=3−α\beta_{\mathrm{TA}} = 3 - \alpha in turbulent TA signals (Figure 8) are well-described by the proposed correction derived from the survival function. The Reynolds number required for these statistical effects to become negligible, Reλ≳105\mathrm{Re}_{\lambda} \gtrsim10^{5}, is well beyond current numerical and laboratory capacities. These findings are specific to one-dimensional signals, as the extension of the statistical effects to higher dimensions is not explored here.

Based on the design of the simulations, the linear trends and effective exponent discussed above are purely statistical properties of truncated power laws. The findings are independent of the governing Navier-Stokes equations or underlying mechanisms such as self-organized criticality. The relations provide no information on causality, and the only prerequisite is for self-similarity to exist within the signal. In this regard, the conclusions apply to any binary process defined by a truncated power law probability distribution. Importantly, the linear expressions cannot be directly applied to the full turbulent signals exhibiting amplitude variability beyond 0 or 1. In this case, certain relations also depend on the phase of the signal [46].

Acknowledgments

The author is financially supported by the US National Science Foundation (NSF-AGS-2031312). The author is also grateful to Professors M. Chamecki and G. G. Katul for fruitful discussions that helped to shape this work.

References

  1. [1]A. N. Kolmogorov, The local structure of turbulence in an incompressible viscous fluid for very large Reynolds numbers, Dokl. Akad. Nauk SSSR 30, 299 (1941).
  2. [2]S. B. Pope, Turbulent Flows, 1st ed. (Cambridge University Press, 2000).
  3. [3]E. W. Montroll and G. H. Weiss, Random walks on lattices. II, J. Math. Phys. 6, 167 (1965).
  4. [4]M. E. J. Newman, Power laws, Pareto distributions and Zipf’s law, Contemp. Phys. 46, 323 (2007).
  5. [5]B. B. Mandelbrot, Fractals and the Geometry of Nature, 1st ed. (W. H. Freeman and Company, 1982).
  6. [6]R. F. Voss, Random fractal forgeries, in Fundamental Algorithms for Computer Graphics, edited by R. A. Earnshaw (Springer, 1985) pp. 805–835.DOI
  7. [7]P. Bak, C. Tang, and K. Wiesenfeld, Self-organized criticality: an explanation of the 1/f noise, Phys. Rev. Lett. 59, 381 (1987).
  8. [8]P. Bak, C. Tang, and K. Wiesenfeld, Self-organized criticality, Phys. Rev. A 38, 364 (1987).
  9. [9]H. J. Jensen, Self-Organized Criticality: Emergent Complex Behavior in Physical and Biological Systems (Cambridge University Press, 1998).
  10. [12]P. L. Miller and P. E. Dimotakis, Stochastic geometric properties of scalar interfaces in turbulent jets, Phys. Fluids 3, 168 (1991).DOI
  11. [13]A. A. Praskovsky, J. F. Foss, S. J. Kleis, and M. Y. Karyakin, Fractal properties of isovelocity surfaces in high Reynolds number laboratory shear flows, Phys. Fluids A 5, 2038 (1993).DOI
  12. [14]H. J. Catrakis and P. Dimotakis, Mixing in turbulent jets: Scalar measures and isosurface geometry, J. Fluid Mech. 317, 369 (1996).
  13. [15]E. Villermaux and C. Innocenti, On the geometry of turbulent mixing, J. Fluid Mech. 393, 123 (1999).
  14. [16]K. P. Iyer, J. Schumacher, K. R. Sreenivasan, and P. K. Yeung, Fractal iso-level sets in high-Reynolds-number scalar turbulence, Phys. Rev. Fluids 5, 044501 (2020).DOI
  15. [17]G. Pickering, J. M. Bull, and D. J. Sanderson, Sampling power-law distributions, Tectonophysics 248, 1 (1995).
  16. [18]S. M. Burroughs and S. F. Tebbens, Upper-truncated power laws in natural systems, Pure Appl. Geophys. 158, 741 (2001).DOI
  17. [19]H. J. Catrakis and P. Dimotakis, Scale distributions and fractal dimensions in turbulence, Phys. Rev. Lett. 77, 3795 (1996).
  18. [20]H. J. Catrakis, Distribution of scales in turbulence, Phys. Rev. E 62, 564 (2000).
  19. [21]H. W. Liepmann, Application of a theorem on the zeros of stochastic functions to turbulence measurements, Helv. Phys. Acta 22, 119 (1949).
  20. [22]K. R. Sreenivasan, A. Prabhu, and R. Narasimha, Zero-crossings in turbulent signals, J. Fluid Mech. 137, 251 (1983).
  21. [23]P. Kailasnath and K. R. Sreenivasan, Zero crossings of velocity fluctuations in turbulent boundary layers, Phys. Fluids 5, 2879 (1993).DOI
  22. [24]A. Bershadskii, J. J. Niemela, A. Praskovsky, and K. R. Sreenivasan, “Clusterization” and intermittency of temperature fluctuations in turbulent convection, Phys. Rev. E 69, 056314 (2004).
  23. [25]D. Cava, G. G. Katul, A. Molini, and C. Elefante, The role of surface characteristics on intermittency and zero-crossing properties of atmospheric turbulence, J. Geophys. Res. 117, D01104 (2012).DOI
  24. [26]P. Perlekar, S. S. Ray, D. Mitra, and R. Pandit, Persistence problem in two-dimensional fluid turbulence, Phys. Rev. Lett. 106, 054501 (2011).
  25. [27]M. Chamecki, Persistence of velocity fluctuations in non-Gaussian turbulence within and above plant canopies, Phys. Fluids 25, 115110 (2013).DOI
  26. [28]S. Chowdhuri, T. Kalmár-Nagy, and T. Banerjee, Persistence analysis of velocity and temperature fluctuations in convective surface layer turbulence, Phys. Fluids 32, 076601 (2020).DOI
  27. [29]K. R. Sreenivasan and A. Bershadsk ii, Clustering properties in turbulent signals, J. Stat. Phys. 125, 1145 (2006).
  28. [31]M. Heisel, G. G. Katul, M. Chamecki, and M. Guala, Velocity asymmetry and turbulent transport closure in smooth- and rough-wall boundary layers, Phys. Rev. Fluids 5, 104605 (2020).DOI
  29. [32]M. A. Badri Narayanan, S. Rajagopalan, and R. Narasimha, Experiments on the fine structure of turbulence, J. Fluid Mech. 80, 237 (1977).
  30. [33]A. Bershadsk ii and K. R. Sreenivasan, Clustering properties in turbulent signals, J. Stat. Phys. 125, 1141 (2006).
  31. [34]R. Percline, Strong, weak and false inverse power laws, Stat. Sci. 20, 68 (2005).DOI
  32. [35]S. G. Saddoughi and S. V. Veeravalli, Local isotropy in turbulent boundary layers at high Reynolds number, J. Fluid Mech. 268, 333 (1994).
  33. [38]A. M. Obukhov, Some specific features of atmospheric turbulence, J. Fluid Mech. 13, 77 (1962).
  34. [39]A. S. Monin and A. M. Yaglom, Statistical Fluid Mechanics: Mechanics of Turbulence, 1st ed., Vol. II (MIT Press, 1975).
  35. [40]K. R. Sreenivasan and R. A. Anton ia, The phenomenology of small-scale turbulence, Annu. Rev. Fluid Mech. 29, 435 (1997).DOI
  36. [41]U. Frisch, P.-L. Sulem, and M. Nelkin, A simple dynamical model of intermittent fully developed turbulence, J. Fluid Mech. 87, 719 (1978).
  37. [42]H. G. E. Hentschel and I. Procaccia, The infinite number of generalized dimensions of fractals and strange attractors, Physica D 8, 435 (1983).DOI
  38. [43]B. B. Mandelbrot, Information theory and psycholinguistics, in Scientific Psychology: Principles and Approaches, edited by B. B. Wolman and E. Nagel (Basic Books, Inc., 1965) pp. 550–562.
  39. [44]P. Grassberger and I. Procaccia, Characterization of strange attractors, Phys. Rev. Lett. 50, 346 (1983).
  40. [45]J. Laherrère and D. Sornette, Stretched exponential distributions in nature and economy: “fat tails” with characteristic scales, Eur. Phys. J. B 2, 525 (1998).DOI
  41. [46]T. Higuchi, Relationship between the fractal dimension and the power law index for a time series: a numerical investigation, Physica D 46, 254 (1990).

Paper details

vlnc:W_hdp5ep2n
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents