1: \subsection[Joint statistics of longitudinal and transverse increments]{Joint multipoint-statistics of longitudinal and transverse increments}

2: \label{sec:2d}

3: 

4: In the preceding section we have analyzed the statistics of longitudinal and transverse increments separately. But this separation is restrictive because the dynamics of both components come from one velocity field and are therefore connected. They can not be separated due to the nonlinear advection term in the Navier-Stokes equation. Therefore we extend the above analysis and examine the joint stochastic properties.  We chose the both increments as one state with the scale parameter $r$, see Eq. (\ref{ustate}). The aim is to estimate a Fokker-Planck equation (\ref{FoplaCond}) in these two variables. We proceed similar to the one dimension analysis. The main difference is that one need much more data and one has to estimate much more coefficients, because the drift-coefficient is now a vector and the diffusion coefficient a matrix.

5: 

6: The verification of the Markov properties for two variables is doubtful because one has to estimate the double conditioned probability function for a two dimensional process, i.e. a six-dimensional function with a finite number of data points. One needs approximately $10^4$ times more data points in comparison to the one dimensional case if the results should be similar significant. The limiting factor is the duration of the measurement and the amount of data. But we know separately for both components that the the Markov properties are valid. In general, if two variables have Markov properties, also the joint statistics have Markov properties (the reverse is not true in general \cite{Risken}). Therefore we assume that the combined process is Markovian as well.

7: 

8: The next step is to estimate the Kramers-Moyal coefficients. First, one has to calculate the approximation of the drift vector $M^{(1)}_{i}(u,v,r,\Delta r)$ and of the diffusion matrix $M^{(2)}_{ij}(u,v,r,\Delta r)$ in dependence of $\Delta r$, see Eq. (\ref{MkDef}).

9:  As in the one dimensional case we calculate the approximation of the drift and diffusion coefficients by fitting a linear polynomial to these functions in dependence of $\Delta r$ above the Markov length. In doing so, we have to consider two different Markov length for both components, one for each component.

10: 

11: The shape of the drift  and diffusion coeffients are shown in Figs. \ref{fig:D1_3D} and \ref{fig:D2_3D} for the length scale $L/4$. The drift coefficients have a simple form: they depend only linearly on one variable, see Fig. \ref{fig:D1_3D}. That means the process decouples in the deterministic part. Remark if this was not the case the increments could not be Markovian seperately in contradiction to our findings. The diagonal coefficients of the diffusion matrix are shown in Figs. \ref{fig:D2_3D}a) and \ref{fig:D2_3D}b). Roughly spoken they consist of a curved surface shifted upwards with a minimum at $u>0$ and $v=0$. The off-diagonal coefficient $D^{(2)}_{12}$ (it is $D^{(2)}_{12}\equiv D^{(2)}_{21}$) has a saddle shape, see Fig. \ref{fig:D2_3D}c).

12: \begin{figure}[tbp]

13: \begin{center}

14: \includegraphics[width=3.5in]{WK_D1}% Here is how to import EPS art

15: \end{center}

16: \caption{The $u$ and $v$ dependence of the drift vector for the scale $r=L/4$. a) The drift coefficient $D^{(1)}_1$, b) the drift coefficient $D^{(1)}_2$. Note that the vertical axis is rotated for an better comparison between a) and b). Both coefficients are pure linear and depend only on one variable.\label{fig:D1_3D} }

17: \end{figure}

18: \begin{figure}[tbp]

19: \begin{center}

20: \includegraphics[width=5in]{WK_D2}% Here is how to import EPS art

21: \end{center}

22: \caption{The $u$ and $v$-dependence of the diffusion matrix for the scale $r=L/4$. a) The coefficient $D^{(2)}_{11}$, b) the coefficient $D^{(2)}_{22}$. It can be seen that the diagonal coefficient are not constant but have a parabolic form, which is more pronounced for the $D^{(2)}_{22}$ coefficient (multiplicative noise). Both coefficients are symmetric under reflection with respect to $v\to -v$, but not for $u\to -u$. c) The saddle-formed off-diagonal coefficient. \label{fig:D2_3D} }

23: \end{figure}

24: 

25: To summarize, we can approximate the form of the drift and diffusion coefficients by a low-order polynomial in $u$ and $v$:

26: \begin{eqnarray}\label{ds}

27: D^{(1)}_1(u,v,r)&=&d^u_1(r)u\\\nonumber

28: D^{(1)}_2(u,v,r)&=&d^v_2(r)v\\\nonumber

29: D^{(2)}_{11}(u,v,r)&=&d_{11}(r)+d^u_{11}(r)u+d^{uu}_{11}(r)u^2+d^{vv}_{11}(r)v^2\\\nonumber

30: D^{(2)}_{22}(u,v,r)&=&d_{22}(r)+d^u_{22}(r)u+d^{uu}_{22}(r)u^2+d^{vv}_{22}(r)v^2\\\nonumber

31: D^{(2)}_{12}(u,v,r)&=&d_{12}(r)+d^v_{12}(r)v+d^{uv}_{12}(r)uv\nonumber.

32: \end{eqnarray}

33: The $d$-coefficients contain the $r$-dependence. Their lower index labels the associated Kramers-Moyal coefficient, the upper index the order of the coefficient with respect to $u$ and $v$. In the Fokker-Planck equation, the coefficients occur symmetric with respect to reflection $v\to -v$. Thus a reflection $v\to -v$ does not change anything, whereas this symmetry is violated for the longitudinal increment. Of course, the Kramers-Moyal coefficients can be better approximated using higher order in $u$ and $v$, but their contributions are small and their value are not well defined at the edges of the available data range, because high velocities are too rare to ensure a good statistics. Note, the significance of higher order terms in $D^{(1)}$ and $D^{(2)}$ is important for the closure of Eqn. (\ref{StrukturGl}) but their investigation should not be addressed in this article.

34: 

35: \begin{figure}[tbp]

36: \begin{center}

37: \includegraphics[width=3.5in]{d_r}% Here is how to import EPS art

38: \end{center}

39: \caption{The coefficients $d$ of Eq. (\ref{ds}), which reflect the dependence of the Kramers-Moyal equation on $u$ and $v$ in dependence of the linear scale variable  $r$. In Figs. a), b) and c), the argument of the transversal coefficient is rescaled by the factor 2/3. The corresponding longitudinal and transversal coefficients coincides which each other apart from the intermittency term $d_{22}^{vv}$.  e) and f): the coefficients for the off-diagonal diffusion coefficients.

40: \label{fig:d_r} }

41: \end{figure}

42: 

43: In order to describe the statistics with the Fokker-Planck equation completely, the $r$-dependence of the $d$-coefficients has to be estimated, see Fig. \ref{fig:d_r}. It can be approximated by

44: \begin{eqnarray}\label{d_r}

45: &&D^{(1)}_1:\;\;\;\;d^u_1(r)=\alpha^u_1+\beta^u_1r\;\;\;\\\nonumber

46: &&D^{(1)}_2:\;\;\;\;d^v_1(r)=\alpha^v_1+\beta^v_1r+\gamma^v_1r^2\\\nonumber

47: &&D^{(2)}_{11}:\;\;\;\;d_{11}(r)=\beta_{11}r,\;\;\;d^u_{11}(r)=\beta^u_{11}r,\;\;\;d^{uu}_{11}(r)=\alpha^{uu}_{11},\;\;\;d^{vv}_{11}(r)=\alpha^{vv}_{11}\\\nonumber

48: &&D^{(2)}_{22}:\;\;\;\;d_{22}(r)=\beta_{22}r,\;\;\;

49: d^u_{22}(r)=\beta^u_{22}r,\;\;\;d^{uu}_{22}(r)=\alpha^{uu}_{22},\;\;\;\\&&\;\;\;\;\;\;\;\;\;\;\;\;\;\,

50: d^{vv}_{22}(r)=\alpha^{vv}_{22}+\beta^{vv}_{22}r+\gamma^{vv}_{22}r^2\\\nonumber

51: &&D^{(2)}_{12}:\;\;\;\;d_{12}^v(r)=\beta_{12}^vr,\;\;\;\alpha^{uv}_{12}+\beta^{uv}_{12}r+\gamma^{uv}_{12}r^2\nonumber.

52: \end{eqnarray}

53: 

54: Before we interpret the results, it has to be shown that the Fokker-Planck equation together with the coefficients (\ref{ds}) and (\ref{d_r}) can reproduce the statistics of the measured data. There are in principal two ways to verify the estimated drift and diffusion coefficients. One way is to solve the Fokker-Planck equation, the other way is to calculate the structure functions, which can be done for example with the hierarchical equation for the structure functions (\ref{StrukturGl}). In both cases the results can be compared with the corresponding quantities estimated directly from the data. In Fig. \ref{fig:verific} the solution of the hierarchical structure function equation (\ref{StrukturGl}) is shown for $n=0$ and $m=1,\,\dots,6$. It is in good agreement with the structure functions. To show that also the joint probability distributions can be reproduced, we solve the Fokker-Planck equation by calculating the path-integral (\ref{pathint}) numerically. We use the estimated Kramers-Moyal coefficients and start the simulation on the integral scale $r=L$ with a Gaussian distribution for $p(\mathbf{u}_0,r=L)$ and integrate down to $r=2\lambda$. Figs. \ref{fig:verific2} and \ref{fig:verific3} show the results. Because of the good accordance we assume that the $d$-coefficients can be used to characterize the statistics of longitudinal and transverse increments.

55: 

56: \begin{figure}[tbp]

57: \begin{center}

58: \includegraphics[width=3.5in]{Verifikation}% Here is how to import EPS art

59: \end{center}

60: \caption{\label{fig:verific}The even longitudinal structure functions up to order 6 calculated directly from data (symbols) compared to that of the Fokker-Planck equation using Eq. (\ref{StrukturGl}) (solid line).}

61: \end{figure}

62: \begin{figure}[tbp]

63: \begin{center}

64: \includegraphics[width=5in]{SimContour}% Here is how to import EPS art

65: \end{center}

66: \caption{Solution of the Fokker-Planck equation. a) Contour plot of the initial condition in logarithmic scale. The simulation starts at the integral length $r=L$ with a Gaussian distribution fitted to the data. b) The contour plots in logarithm scale of the simulated probability distribution on the scale $r=2\lambda$. The distance between the contour lines is chosen in logarithmic scale and correspond to a factor 10. Dashed lines are the probability calculated directly from data, the full lines are the simulation ones. The simulation reproduces well the properties of the data.\label{fig:verific2}}

67: \end{figure}

68: \begin{figure}[tbp]

69: \begin{center}

70: \includegraphics[width=5in]{SimCut}% Here is how to import EPS art

71: \end{center}

72: \caption{Cuts through the probability distributions shown in Fig. \ref{fig:verific2}. Black circles are the pdfs of the data, the straight line is the simulation and the dashed line is the initial condition at $r=L$. a) The distribution in dependence of $u$ for $v=0$. b) The distribution in dependence of $v$ for $u=0$. Notice that the skewness in Fig. a) and the intermittency for both components are well reproduced.\label{fig:verific3}}

73: \end{figure}

74: 

75: The number of coefficients in (\ref{d_r}) can be reduced by a remarkably symmetry. If one multiplies the $r$-dependence for the transverse increments with a factor $2/3$, the related coefficients of longitudinal and transverse increments coincides, i.e. $d_\textrm{long}(r)\approx d_\textrm{trans}(2r/3)$. The only exception is the diffusion term $d^{vv}_{22}(r)$. Thus the number of independent coefficients is halved.

76: This symmetry will be examined in more detail below. 

77: 

78: