Castaing Scaling and Kernels for Theories of the Turbulent Cascade Miriam A. Forman Department of Physics and Astronomy State University of New York at Stony Brook NY11794-3800 USA Abstract. The Castaing cascade kernel function G(u;L,L’) relates the pdf of velocity fluctuations on scale L’ to those on scale L. The shape of G(u;L,L’) and how it depends on scale is a measure of intermittency alternative to measures such as the exponent of the structure function or multifractal spectra. This paper provides analytical forms and graphic illustrations of the Castaing kernels corresponding to six models of the turbulent cascade. The popular P-model and Poisson models of turbulence have discrete Castaing kernels, which seems non-physical. G(u). Such distributions can be conceptually useful particularly for looking at the dependence of the parameters on scale [5, 11, 13]. However, neither G(u) nor PL(u) can be precisely Gaussian; if they were, the scaling would be log-normal, and the third moment of the parallel velocity increment would be zero, both of which are refuted fundamentally by Frisch [6]. The Castaing pdf is only approximate and should be used with care. This paper deals only with Castaing scaling which is much more general. INTRODUCTION Castaing scaling [1,2,3,4] describes the cascade of turbulent energy in a fluid from scale L to smaller scale L’ by relating the probability distribution functions (pdfs) of the velocity increments δV at the different scales with a cascade kernel G(u:L’,L) such that +∞ P L' (δV ) = ∫ ( ) G (u ; L' , L )e − u PL e − u δV du . (1) −∞ RELATION OF CASTAING KERNEL G, TO THE STRUCTURE FUNCTION ζ The kernel G(u;L’,L) represents the probability that turbulence of amplitude e-uδV on scale L, cascades to turbulence of amplitude δV on scale L’. Castaing scaling is linear in δV, in the sense that the same G(u) relates e-uδV at the larger scale to δV at the smaller scale, for every -∞ < δV < ∞. Since Castaing kernels directly describe the distribution of turbulent amplitudes at larger scales that contribute to a given amplitude at small scales, they provide a graphic description of the cascade. Multifractality and intermittency appear in the cascade when the kernel G(u) is anything other than a delta-function. The details depend on the shape of G(u; L’,L) versus u, and its dependence on L’. Following equation 1, the scaling properties of all the moments of the pdf arise from the scaling of G(u;L,L). In particular, since the moment of order q at scale L’ is +∞ S(q , L' ) = ∫ ∞ d(δV )(δV )q −∞ ∫ ( ) G (u ; L' , L )e − u PL e − u δV du −∞ (2a) ∞ qu = G (u ; L' , L )e du S(q , L ) − ∞ ∫ If both G(u) and PL(δV) are Gaussian functions of their arguments, PL’(u) will be a winged “Castaing distribution” characterized by the mean and width of (2b) all of the dependence on L’ (in other words, the CP679, Solar Wind Ten: Proceedings of the Tenth International Solar Wind Conference, edited by M. Velli, R. Bruno, and F. Malara © 2003 American Institute of Physics 0-7354-0148-9/03/$20.00 550 unique parabola which has ζ(3) = 1, and ζ(q)⇒q/3 when µ⇒0. Some experts [e.g., 6] dismiss this model in incompressible fluid turbulence because the ζ(q) should not have a peak and its slope should be nonnegative at all q. A peculiarity of the KO62 kernel is that at positive u it is non-zero (although small) and increases with scale. This allows a non-zero probability for arbitrarily large fluctuations at smaller scales to arise from smaller-amplitude fluctuations at larger scales. These dominate the moments of large positive order, and confuse their scaling. Whether this is also strictly impossible in compressible and/or magnetized fluids such as the solar wind is unclear. Solar wind velocity increments do have ζ(q) with negative slopes in some scale ranges at which the intermittency is very large [7]. scaling) is contained in G(u:L’,L). Thus when the scaling ζ(q) = d ln S(q , L ′) occurs on scales from L’ to d ln L ′ L, (2) implies that the Fourier transform of G(u;L’,L) is equal to [4] ˆ (k ; L' , L ) = exp[− ln(L / L' )ζ (− ik )] G . (3) Scaling in actual data sets is usually described with ζ(q), and turbulence theories with a mathematical form for ζ(q). In practice ζ(q) can be calculated from finite and imperfect real data sets only over a limited range of q, because very large and very small δV will be missing. If (1) applies, the scaling is completely subsumed into the kernel G, and does not involve the pdf at large scale. Also, the scaling of the moments of the absolute value of δV is the same as the scaling of the moments of δV for integer q ≥ 0. The odd moments of the absolute values are not the same as the odd moments of δV itself, but they scale the same way if equation (1) is valid. The linear form of (1) assures that moments of the absolute value of δV at odd, noninteger and negative q provide additional useful information about G(u;L’,L). The binomial and Poisson models [8,9,10] avoid the negative slopes in ζ(q) and, since their kernels are zero at u>0, never let small δV at large scale lead to larger δV at smaller scales. Of these, DeBrulle’s [10] 2-parameter generalized Poisson model is the most flexible, and can have D(hmin) = 2 (∆ = 0.4, β = 0.6 works well). However, the binomial and Poisson kernels are uncomfortably discontinuous sums of delta functions. In the binomial and Poisson models, δV at small scale can arise only from certain discrete values of δV at larger scale, which is implausible. Although the idea of Castaing scaling underlies turbulence theory already, it is instructive to examine the actual kernels G(u;L’,L) for popular and reasonable models of the turbulent cascade, to better appreciate the differences between the models, and to encourage further turbulence theories to describe themselves with their kernels. The Universal Multifractal model [13] is a generalization of the KO62 model to a kernel that is a continuous Levy function of u, near the most probable values of u. Crucial truncations at extreme values of u arise from plausible self-organized criticality and limited data sets, preventing the slope of ζ(q) from becoming negative. Such modifications could be applied to the KO 1962 model as well. COMPARISON OF CASTAING KERNELS FOR TURBULENCE MODELS The intermittent models in Table 1 produce similarlooking ζ(q) forq-3<5 where ζ(q) might be reliably estimated from data. Table 1 lists in historical order seven models of inertial turbulence, with their ζ(q) and G(u;L’,L), minimum scaling exponent, and the dimension on the minimum scaling exponent. All models have ζ(3) =1 in accord with the Kologorov 4/5 law [6] for inertial turbulence. G(u;L’,L) was calculated by inverse Fourier transform of eq. (3) [4]. The minimum scaling exponent and dimension on the minimum scaling exponent were calculated from ζ(q) by the method given in [6]. 7 Debrulle, delta=0.4, beta=0.6 6 5 binomial, P=0.7 logLevy, 1.5 She-Leveque zeta(q) 4 3 lognormal, mu =0.2 2 1 0 -10 -5 0 5 10 15 20 25 30 -1 -2 q -3 -4 Intermittent turbulence implies a Castaing kernel of finite width; equivalently, a ζ(q) with negative curvature. The Kolmogorov-Obkuhov 1962 (KO62) form is the simplest extension of the K41 delta function, to a normal distribution in u. Its ζ(q) is the -5 FIGURE 1. Exponents of the structure function for six models of the intermittent turbulent cascade. Parameters for all models were chosen such that ζ(4)/4 -ζ(2)/2 = 0.022. 551 TABLE 1. ζ(q) and G(u) for models of the turbulent cascade Turbulence model ζ(q) Exponent of structure function <VLq> ∝ Lζ(q) G(u,η η) Castaing kernel for L0 > L u≡ln(δVL/ δVL 0) η≡ ln(L0/L) Kolmogorov 1941 q 3 Delta function: δ(u + η/3) Kolmogorov-Obukhov 1962 (log-normal) 2 q µ q q + − 3 2 3 3 Gaussian q q 3 3 1 − log2 P + (1 − P) binomial n 1 2n ∑ 0 m n−m n! δ u − ln(P ) − ln(1 − P ) 3 3 (n − m)! m! .5<P<1 She and Leveque 1994 (Poisson) De Brulle 1994 (Poisson) Poisson ∞ ∑ m =0 ∑ δ(u + aη + mW ) m =0 ln (bη)m e − bη m! 1 1 log2 3 P 2 1 b= -∞ −α Scale factor γ = 3 µ η 2(α − 1) As UM, except straight line for q > qD ∆ 1− β NA 1 µ 3 + 6(α −1) η Truncated Levy 1 q µ h min = + ( ) 1 − α D 3 6 α −1 3 552 3− ∆ 1− β Levy function of order α, α q µ q q − + 3 2(α − 1) 3 3 1<α<2 (most likely 1.5) 1− ∆ 3 1 β W= 3 Mean = − S +L’s UM + self-organized criticality and finite data set NA 1 9 ln1.5 (2η)m − 2η η e δ u + + m 9 3 m! 1− ∆ a= 3 (log-Levy) -∞ Poisson ∞ 0 < ∆, β < 1 Shertzer and Lovejoy’s (1997) Universal Multifractal 3 for (L0/L) = 2n q q 2 3 + 21 − 9 3 q ∆ 1− ∆ 3 − β q+ 1 3 1− β 1 3 2 9 3 η µ exp − u 1 + + 3 2 2πµη 2µη µ ≈ 0.2 Meneveau and Sreenivasan’s P-model 1991 (binomial) hmin and D(hmin) α −1 D(hmin) = 3− µ qD 2 3 α She-Leveque The non-linear processes causing self-organized criticality of extreme fluctuations may make Eq. (1) invalid at very extreme u anyway. 10 G(u), for L/L' = 4 Debrulle, delta=0.4, beta=0.6 binomial P=0.7 lognormal, mu =0.2 1 logLevy, alpha = 1.5, mu = 0.2 Kolmogorov 41 0.1 ACKNOWLEDGMENTS 0.01 It is a pleasure to acknowledge discussions with Leeonard F. Burlaga and colleagues at NASA/Goddard Spaceflight Center, USA, and the kindness of the Conference organizers. This work was supported by NASA grant NAG 510995. 0.001 -2 -1.5 -1 -0.5 0 0.5 u FIGURE 2. Castaing kernels (Eq. 1) for the models and parameters in Figure 1, for a scale L/L’ = 4. The formulas in Table 1 were used. The discrete models are shown as isolated values scaled by the ∆u between points, to compare with the continuous models. The integral under all curves = 1. The vertical line shows the location of the delta function implied by the original Kolmogorov 1941 theory. REFERENCES Agreement is good in the central parts of the G(u) for all models, corresponding to q-3<5, except for the strange discreteness in the binomial and Poisson models. Both (non-truncated) continuous models have wings at u>0, but the logLevy has much bigger wings than the lognormal. Shertzer et al.’s [13] assertion that self-organized criticality will modify ζ(q) at its extremes, will also modify the wings of the G(u) for the lognormal and logLevy models. 1. B. Castaing, Y. Gagne and E. Hopfinger, Physica D, 46, 177 (1990) 2. Chaubaud, A. Naert, J. Peinke, F. Chilla, B. Castaing, and B. Hebral, Phys. Rev. Lett., 73, 3277 (1994) 3. Arneodo, S. Roux, and J.F. Muzy, J. Phys. II France, 7, 363 (1997) 4. J.F. Muzy, J. Delour, and E. Bacry, arXiv:condmat/0005400 24 May 2000 5. L. Sorriso-Valvo, V. Carbone, P. Veltri, G. Consolini, and R. Bruno, Geophys. Rev. Lett., 26, 1801 (1999) 6. U. Frisch, Turbulence, The Legacy of A. N. Kolmogorov, Cambridge: Cambridge University press, 1995 7. P. Veltri and A. Mangeney, in Solar Wind Nine, edited by Habbal et al., AIP Conference Proceedings 471, Woodbury, NY: American Institute of Physics, 1999, pp. 543-546 8. C. Meneveau and K.R. Sreenivasan, Phys. Rev. Lett., 59, 1424 (1987) 9. Z–S. She and E. Leveque, Phys. Rev. Lett, 72, 336 (1994) CONCLUSIONS The Castaing kernels describe the distribution of amplitudes of turbulence on larger scales that give rise to turbulence of a given amplitude on a smaller scale, as a function of the ratio of amplitudes (ln(u)) and scale. The Kolmogorov 1941 theory has a single delta function as a kernel, but theories of intermittent turbulence have more spread-out kernels. The popular binomial (P) model and the Poisson models which have well-behaved structure functions, have discrete kernels that are sums of delta functions, implying that small-scale turbulence of a given amplitude can arise only from certain discrete amplitudes on larger scale. This discreteness seems unphysical. On the other hand, the full Kolmogorov 1962 lognormal theory and its extension to logLevy, have continuous Castaing kernels, but problematic wings and structure functions. The lognormal or logLevy theories, modified by selforganized criticality and effect of finiteness of data sets as proposed by Shertzer, et al. [13], avoids these problems and seems the most acceptable among the classes considered here. 10. B. Dubrulle, Phys. Rev. Lett, 73, 959 (1994) 11. C. Pagels and A. Balogh, Nonl. Proc. in Geophys., 8, 313 (2001) 12. Shertzer, D., S. Lovejoy, F. Schmitt, Y. Chigirinskya, and D. Marsan, Fractals, 5, 427 (1997) 13. Forman, M.A., and L.F. Burlaga, in Solar Wind Ten, edited by M. Velli, et al., AIP Conference Proceedings, Woodbury, NY: American Institute of Physics (this volume) 553
© Copyright 2024 Paperzz