1: \documentclass[epj]{svjour}
2: %\documentclass[epj,referee]{svjour}
3: \usepackage{
4: cite,
5: graphicx,
6: dsfont,
7: url
8: }
9:
10: \title{Stochastic analysis of different rough surfaces}
11:
12: \author{M.~Waechter\inst{1}\thanks{\email{matthias.waechter@uni-oldenburg.de}}
13: \and F.~Riess\inst{1}
14: \and Th.~Schimmel\inst{2}
15: \and U.~Wendt\inst{3}
16: \and J.~Peinke\inst{1}\thanks{\email{peinke@uni-oldenburg.de}}
17: }
18:
19: \institute{
20: Institute of Physics, Carl-von-Ossietzky University,
21: D-26111 Oldenburg, Germany
22: \and
23: Institute of Applied Physics,
24: University of Karlsruhe,
25: D-76128 Karlsruhe, Germany
26: \and
27: School of Materials Science,
28: Otto-von-Guericke University,
29: D-39016 Magdeburg, Germany
30: }
31:
32: \date{\today}
33:
34: \abstract{
35: This paper shows in detail the application of a new stochastic approach for
36: the characterization of surface height profiles, which is based on the
37: theory of Markov processes.
38: %
39: With this analysis we achieve a characterization of the scale dependent
40: complexity of surface roughness by means of a Fokker-Planck or Langevin
41: equation, providing the complete stochastic information of multiscale joint
42: probabilities.
43: %
44: The method is applied to several surfaces with different properties, for the
45: purpose of showing the utility of this method in more detail. In particular
46: we show evidence of the Markov properties, and we estimate the parameters of
47: the Fokker-Planck equation by pure, parameter-free data analysis. The
48: resulting Fokker-Planck equations are verified by numerical reconstruction
49: of the conditional probability density functions.
50: %
51: The results are compared with those from the analysis of multi-affine and
52: extended multi-affine scaling properties which is often used for surface
53: topographies. The different surface structures analysed here show in detail
54: the advantages and disadvantages of these methods.
55: %
56: \PACS{
57: {02.50.-r}{Probability theory, stochastic processes, and statistics} \and
58: {02.50.Ga}{Markov processes} \and
59: {68.35.Bs}{Surface structure and topography of clean surfaces}
60: }
61: }
62:
63: \newcommand{\un}[1]{\ensuremath{\,\mathrm{#1}}}
64: \newcommand{\mum}{\ensuremath{\,\mathrm{\mu m}}}
65: \newcommand{\Tik}{\rule[-3pt]{1pt}{4pt}}
66: \newcommand{\balken}[2]{%
67: \raisebox{-2ex}{\makebox[0pt][l]{\hspace{0.5em}\footnotesize #2}}%
68: \Tik\rule{#1}{1pt}\Tik%
69: }
70: \unitlength 0.008237\textwidth % entspricht 1mm mit a4paper.sty
71:
72: \begin{document}
73: \maketitle
74:
75:
76: \section{Introduction}
77: \label{sec:intro}
78:
79: Among the great variety of complex and disordered systems the complexity of
80: surface roughness is attracting a great deal of scientific interest
81: \cite{Sayles1978,Vicsek1992-e,Barabasi1995,Davies1999,Wendt2002a}. The physical
82: and chemical properties of surfaces and interfaces are to a significant degree
83: determined by their topographic structure. Thus a comprehensive
84: characterization of their topography is of vital interest from a scientific
85: point of view as well as for many applications
86: \cite{Dharmadhikari1999,Saitou2001,Sydow2003}.
87:
88: Most popular methods used today for the characterization of surface roughness
89: are based on the concepts of self-affinity and multi-affinity, where the
90: multifractal $f(\alpha)$ spectrum has been regarded as the most complete
91: characterization of a surface
92: \cite{Feder1988,Family1991,Vicsek1992-e,Barabasi1995}. One example of a
93: measure of roughness which is commonly used in this context is the rms surface
94: width $w_r(x)=\langle (h(\tilde{x})-\bar{h})^2 \rangle^{1/2}_r$, where
95: $h(\tilde{x})$ is the measured height at point $\tilde{x}$,
96: $\langle\,\cdot\,\rangle_r$ denotes the average over an interval of length $r$
97: around the selected point $x$, and $\bar{h}$ the mean value of $h(\tilde{x})$
98: in that interval. Thus the roughness is measured at a specific location $x$
99: and over a specific scale $r$. Then a scaling regime of the ensemble average
100: $\langle w_r \rangle$ in $r$, if it exists, is analyzed according to $\langle
101: w_r^\alpha \rangle \sim r^{\xi_\alpha}$, usually $\alpha\in\mathds{Q}$. Here,
102: $\langle\,\cdot\,\rangle$ denotes the mean over the available range in $x$.
103: For a more thorough introduction into scaling concepts we refer the reader to
104: the literature, e.g.\ \cite{Feder1988,Family1991,Vicsek1992-e,Barabasi1995}.
105: Terms concerning scaling concepts which are used in this paper are rapidly
106: introduced in section~\ref{sec:scaling}. Here, we have to note the following
107: points which concern stochastic aspects of roughness analysis: First, the
108: ensemble average $\langle w_r \rangle$ must obey a scaling law as mentioned
109: above, and second, the statistics of $w_r(x)$ are investigated over distinct
110: length scales $r$, thus possible correlations between $w_r(x)$ and $w_{r'}(x)$
111: on different scales $r,r'$ are not examined.
112:
113: In this paper we want to give a deeper introduction into a new approach to
114: surface roughness analysis which has recently been introduced by us
115: \cite{Friedrich1998a,Waechter2003_plus_preprint} and by \cite{Jafari2003}.
116: This method is based on stochastic processes which should grasp the scale
117: dependency of surface roughness in a most general way. No scaling feature is
118: explicitly required, and especially the correlations between different scales
119: $r$ and $r'$ are investigated. To this end we present a systematic procedure
120: as to how the explicit form of a stochastic process for the $r$-evolution of a
121: roughness measure similar to $w_r(x)$ can be extracted directly from the
122: measured surface topography. This stochastic approach turns out to be a
123: promising tool also for other systems with scale dependent complexity such as
124: turbulence \cite{Friedrich1998a,Renner2001,Lueck1999}, financial data
125: \cite{Friedrich2000a,Ausloos2003}, and cosmic background radiation
126: \cite{Ghasemi2003}. Also this stochastic approach has recently
127: enabled the numerical reconstruction of surface topographies
128: \cite{Jafari2003}.
129:
130: Here we demonstrate our ansatz by analysing a number of data sets from
131: different surfaces. The purpose is to show extensively the utility of this
132: method for a wide class of rough surfaces. The examples show different kinds of
133: scaling properties which are, in addition, briefly analysed. Among these
134: examples is a collection of road surfaces measured with a specially designed
135: profile scanner. Preliminary results of the analysis of one of these surfaces
136: have already been presented \cite{Waechter2003_plus_preprint}. AFM measurement
137: data from an evaporated gold film have already been analysed in an earlier
138: stage of the method \cite{Friedrich1998a}. Since then, the method has been
139: significantly refined and extended. Measurements of a steel crack surface were
140: taken by confocal laser scanning microscopy (CLSM) \cite{Wendt2002a}.
141:
142: As a measure of surface roughness we use the \emph{height increment}
143: \cite{CenteredInc}
144: \begin{equation}
145: \label{eq:h_r}
146: h_r(x) := h(x+r/2) - h(x-r/2)
147: \end{equation}
148: depending on the length scale $r$. For other scale dependent roughness
149: measures, see \cite{Family1991,Barabasi1995}. The height increment $h_r$ is
150: used because its moments, which are well-known as structure functions (see
151: section~\ref{sec:scaling}), are closely connected with spatial correlation
152: functions. Nevertheless, it should be pointed out that our method presented in
153: the following could be easily generalized to any scale dependent measure,
154: \emph{e.g.} the above-mentioned $w_r(x)$ or wavelet functions \cite{Haase2003}.
155: As a new ansatz, $h_r$ is regarded as a \emph{stochastic variable in $r$}.
156: Without loss of generality we consider the process as being directed from
157: larger to smaller scales. The focus of our method is the investigation of how
158: the surface roughness is linked between different length scales.
159:
160: In the remainder of this paper we will first summarize in
161: section~\ref{sec:Markov_theory} some central aspects of the theory of Markov
162: processes which form the basis of our analysis procedure. Details concerning
163: the measurement data are presented in section~\ref{sec:measurement_data},
164: their scaling properties are analyzed in section~\ref{sec:scaling}. The Markov
165: properties of our examples are investigated in section \ref{sec:markov_props}.
166: In section~\ref{sec:Dk_estimation} we estimate for each data set the
167: parameters of a Fokker-Planck equation. The ability of this equation to
168: describe the statistics of $h_r$ in the scale variable $r$ is then examined
169: in section~\ref{sec:veri_coeff}, followed by concluding remarks in
170: section~\ref{sec:conclusions}.
171:
172:
173: \section{Surface roughness as a Markov process}
174: \label{sec:Markov_theory}
175:
176: Complete information about the stochastic process would be available from the
177: knowledge of all possible $n$-point, or more precisely $n$-scale,
178: joint probability density functions (pdf) %
179: $p(h_1, r_1; h_2, r_2; \ldots ; h_n, r_n)$ %
180: describing the probability of finding simultaneously the increments $h_1$ on
181: the scale $r_1$, $h_2$ on the scale $r_2$, and so forth up to $h_n$ on the
182: scale $r_n$. Here we use the notation $h_i(x)=h_{r_i}(x)$, see
183: eq.~(\ref{eq:h_r}). Without loss of generality we take $r_1<r_2<\ldots<r_n$.
184: As a first question one has to ask for a suitable simplification. In any case
185: the $n$-scale joint pdf can be expressed by multiconditional pdf
186: \begin{eqnarray} \label{eq:condpdf}
187: \lefteqn{p(h_1,r_1;\ldots;h_n,r_n) = }\nonumber\\
188: &&p(h_1,r_1 | h_2,r_2;\ldots;h_n,r_n)\cdot p(h_2,r_2|h_3,r_3;\ldots;h_n,r_n)
189: \nonumber\\
190: &&{}\cdot\ldots\cdot p(h_{n-1},r_{n-1}|h_n,r_n)\cdot p(h_n,r_n)\,.
191: \end{eqnarray}
192: Here, $p(h_i,r_i| h_j,r_j)$ denotes a \emph{conditional probability} of
193: finding the increment $h_i$ on the scale $r_i$ under the condition that
194: simultaneously, \emph{i.e.}\ at the same location $x$, on a larger scale $r_j$
195: the value $h_j$ was found.
196: %
197: It is defined with the help of the joint probability $p(h_i,r_i; h_j,r_j)$ by
198: \begin{equation}
199: \label{eq:condpdf_def}
200: p(h_i,r_i| h_j,r_j) = \frac{p(h_i,r_i; h_j,r_j)}{p(h_j,r_j)}\;.
201: \end{equation}
202:
203: An important simplification arises if
204: \begin{equation}
205: \label{eq:markov_straight}
206: p(h_i, r_i | h_{i+1}, r_{i+1}; \ldots; h_n, r_n) =
207: p(h_i, r_i | h_{i+1}, r_{i+1})\;.
208: \end{equation}
209: This property is the defining feature of a Markov process evolving from
210: $r_{i+1}$ to $r_i$. Thus for a Markov process the $n$-scale joint pdf
211: factorize into $n$ conditional pdf
212: \begin{eqnarray}
213: \label{eq:markov1}
214: \lefteqn{p(h_1,r_1;\ldots;h_n,r_n) = p(h_1,r_1\, | h_2,r_2)}\nonumber\\
215: &&{}\cdot \ldots \cdot p(h_{n-1},r_{n-1}\, | h_n,r_n)
216: \cdot p(h_n,r_n) \;.
217: \end{eqnarray}
218: The Markov property implies that the $r$-dependence of $h_r$ can be regarded
219: as a stochastic process evolving in $r$, driven by deterministic and random
220: forces.
221: %
222: Here it should be noted that if condition (\ref{eq:markov_straight}) holds
223: this is true for a process evolving in $r$ from large down to small scales as
224: well as the reverse from small to large scales \cite{Renner2001Diss}.
225: %
226: Equation (\ref{eq:markov1}) also emphasizes the fundamental meaning of
227: conditional probabilities for Markov processes since they determine any
228: $n$-scale joint pdf and thus the complete statistics of the process.
229:
230: For any Markov process a Kramers-Moyal expansion of the governing master
231: equation exists \cite{Risken1984}. For our height profiles it takes the form
232: \begin{eqnarray}
233: \label{eq:KME}
234: \lefteqn{-r\,\frac{\partial}{\partial r}\;p(h_r,r|h_0,r_0)\,=}\nonumber\\
235: &&\sum_{k=1}^{\infty}
236: \left( -\frac{\partial}{\partial h_r} \right)^k
237: D^{(k)}(h_r,r)\, p(h_r,r|h_0,r_0) \,.
238: \end{eqnarray}
239: The minus sign on the left side of eq.~(\ref{eq:KME}) expresses the direction
240: of the process from larger to smaller scales, furthermore the factor $r$
241: corresponds to a logarithmic variable $\rho=\ln r$ which leads to simpler
242: results in the case of the scaling behaviour \cite{FPE}.
243: To derive the Kramers-Moyal coefficients $D^{(k)}(h_r,r)$, the limit $\Delta r
244: \rightarrow 0$ of the conditional moments has to be performed:
245: \begin{equation}
246: \label{eq:Dk_def}
247: D^{(k)}(h_r,r) = \lim_{\Delta r \rightarrow 0} M^{(k)}(h_r,r,\Delta r)\,,
248: \qquad\mbox{where}
249: \end{equation}
250: \begin{eqnarray}
251: \label{eq:Mk_def}
252: \lefteqn{M^{(k)}(h_r,r,\Delta r) =}\nonumber\\
253: &&\frac{r}{k!\Delta r}
254: \int_{\scriptscriptstyle-\infty}^{\scriptscriptstyle+\infty}
255: (\tilde{h}-h_r)^k \, p(\tilde{h},r-\Delta r | h_r,r) \, d\tilde{h} \,.
256: \end{eqnarray}
257: The moments $M^{(k)}(h_r,r,\Delta r)$ characterize the alteration of the
258: conditional probability $p(h_r,r|h_0,r_0)$ over a finite step size
259: $\Delta r=r_0-r$ and are thus also called ``transitional moments''.
260:
261: A second major simplification is valid if the noise included in the process is
262: Gaussian distributed. In this case the coefficient $D^{(4)}$ vanishes (from
263: eqs.~(\ref{eq:Dk_def}) and (\ref{eq:Mk_def}) it can be seen that $D^{(4)}$ is
264: a measure of non-gaussianity of the included noise). According to Pawula's
265: theorem, together with $D^{(4)}$ all the $D^{(k)}$ with $k\geq 3$ disappear
266: and the Kramers-Moyal expansion (\ref{eq:KME}) collapses to a Fokker-Planck
267: equation \cite{Risken1984}, also known as Kolmogorov equation
268: \cite{Kolmogorov1931}:
269: %
270: \begin{eqnarray}
271: \label{eq:FPE1}
272: \lefteqn{-r\,\frac{\partial}{\partial r}\;p(h_r,r|h_0,r_0)\,=}\\
273: &&\left\{ -\frac{\partial}{\partial h_r} D^{(1)}(h_r,r)
274: + \frac{\partial^2}{\partial h_r^2} D^{(2)}(h_r,r)
275: \right\} p(h_r,r|h_0,r_0)
276: \nonumber
277: \end{eqnarray}
278: %
279: The Fokker-Planck equation then describes the evolution of the conditional
280: probability density function from larger to smaller length scales and thus also
281: the complete $n$-scale statistics. The term $D^{(1)}(h_r,r)$ is commonly
282: denoted as the drift term, describing the deterministic part of the process,
283: while $D^{(2)}(h_r,r)$ is designated as the diffusion term, determined by the
284: variance of a Gaussian, $\delta$-correlated noise (compare also
285: eqs.~(\ref{eq:Dk_def}) and (\ref{eq:Mk_def})).
286:
287: By integrating over $h_0$ it can be seen that the Fokker-Planck equation
288: (\ref{eq:FPE1}) is also valid for the unconditional probabilities $p(h_r,r)$
289: (see also section \ref{sec:veri_coeff}). Thus it covers also the behaviour of
290: the moments $\langle h_r^n\rangle$ (also called structure functions) including
291: any possible scaling behaviour.
292: An equation for the moments can be obtained by additionally multiplying with
293: $h_r^n$ and integrating over $h_r$
294: \begin{eqnarray}
295: \label{eq:moments}
296: \lefteqn{-r \frac{\partial}{\partial r} \langle h_r^n \rangle =}\\
297: && n \langle D^{(1)}(h_r,r) h_r^{n-1} \rangle
298: + n(n-1) \langle D^{(2)}(h_r,r) h_r^{n-2} \rangle \;. \nonumber
299: \end{eqnarray}
300: For $D^{(1)}$ being purely linear in $h_r$ ($D^{(1)} = \alpha h_r$) and
301: $D^{(2)}$ purely quadratic ($D^{(2)}= \beta h_r^2$), the multifractal scaling
302: $\langle h_r^n\rangle \sim r^{\xi_n}$ with $\xi_n = n \alpha + n(n-1)\beta$ is
303: obtained from (\ref{eq:moments}). If in contrast $D^{(2)}$ is constant in
304: $h_r$, a monofractal scaling where $\xi_n$ are linear in $n$ may occur, see
305: \cite{Friedrich1998a}.
306:
307: Lastly, we want to point out that the Fokker-Planck equation (\ref{eq:FPE1})
308: corresponds to the following Langevin equation (we use It\^{o}'s definition)
309: \cite{Risken1984}
310: \begin{equation}\label{eq:Langevin}
311: -\frac{\partial h_r}{\partial r} =
312: D^{(1)}(h_r,r)/r+\sqrt{D^{(2)}(h_r,r)/r}\,\Gamma(r)\;,
313: \end{equation}
314: where $\Gamma(r)$ is a Gaussian distributed, $\delta$-correlated noise. The
315: use of this Langevin model in the scale variable opens the possibility to
316: directly simulate surface profiles with given stochastic properties, similar
317: to \cite{Jafari2003}.
318:
319: With this brief summary of the features of stochastic processes we have fixed
320: the scheme from which we will present our analysis of diverse rough surfaces.
321: There are three steps: First, the verification of the Markov property. Second,
322: the estimation of the drift and diffusion coefficients $D^{(1)}$ and $D^{(2)}$.
323: Third, the verification of the estimated coefficients by a numerical solution
324: of the corresponding Fokker-Planck equation, thus reconstructing the pdf which
325: are compared to the empirical ones.
326:
327:
328:
329: \section{Measurement data}
330: \label{sec:measurement_data}
331:
332: With the method outlined in section \ref{sec:Markov_theory} we analysed a
333: collection of road surfaces measured with a specially designed profile scanner
334: as well as two microscopic surfaces, namely an evaporated gold film and a
335: crack surface of a low-alloyed steel sample, as already mentioned in
336: section~\ref{sec:intro}.
337:
338: The road surfaces have been measured with a specially designed surface profile
339: scanner. The Longitudinal resolution was 1.04\un{mm}, the profile length being
340: typically 20\un{m} or 19000 samples, respectively. Between ten and twenty
341: parallel profiles with a lateral distance of 10\un{mm} were taken for each
342: surface, see fig.~\ref{fig:data_road_scaling}. The vertical error was always
343: smaller than 0.5\un{mm} but in most cases approximately 0.1\un{mm}. Details can
344: be found in \cite{Waechter2002a}.
345:
346: For the Au film data, the surface of four optical glass plates had been coated
347: with an Au layer of 60\un{nm} thickness by thermal evaporation
348: \cite{Friedrich1998a}. The topography of these films was measured by atomic
349: force microscopy at different resolutions, resulting in a set of images of
350: $256\times256$ pixels each, where every pixel specifies the surface height
351: relative to a reference plane, see fig.~\ref{fig:data_gold}. Out of these
352: images 99 could be used for the analysis presented here, resulting in about
353: $6.5\cdot 10^6$ data points. Sidelengths vary between 36\un{nm} and 2.8\un{\mu
354: m}.
355:
356: The sample of the crack measurements was a fracture surface of a low-alloyed
357: steel (german brand 10MnMoNi5-5). A detailed description of the measurements
358: can be found in \cite{Wendt2002a}. Three CLSM (Confocal Laser Scanning
359: Microscopy) images of size 512x512 pixels in different spatial resolutions were
360: available, see fig.~\ref{fig:Wendt_data}. Pixel sizes are 0.49, 0.98, and
361: 1.95\mum, resulting in image widths of 251, 502, and 998\mum, repectively.
362: Unavoidable artefacts of the CLSM method were removed by simply omitting for
363: each image those data with the smallest and largest height value, similar to
364: \cite{Wendt2002a}. Nevertheless this cannot guarantee the detection of all the
365: artefacts. The possible consequences are discussed together with the results.
366:
367: For the analysis in the framework of the theory of Markov processes, we will
368: normalize the measurement data by the quantity $\sigma_\infty$ defined by
369: \begin{equation}
370: \label{eq:sigma_inf}
371: \sigma_\infty^2 = \lim_{r \rightarrow \infty} \langle h_r^2 \rangle\,.
372: \end{equation}
373: Thus it is possible to obtain dimensionless data with a normalization
374: independent of the scale $r$, in contrast to e.g.\ $\sigma_r^2=\langle h_r^2
375: \rangle$. As a consequence the results, especially $M^{(k)}(h_r,r,\Delta
376: r)$ and $D^{(k)}(h_r,r)$ (cf.\ section~\ref{sec:Markov_theory}), will also be
377: dimensionless. From the definition it is easy to see that $\sigma_\infty$ can
378: be derived via
379: $\sigma_\infty^2=2\sigma_x^2=2\langle(h(x)-\bar{h})^2\rangle$ if
380: $h(x)$ becomes uncorrelated for large distances $r$.
381:
382:
383:
384: \section{Scaling analysis}
385: \label{sec:scaling}
386:
387: In this paper a number of examples was selected from all data sets under
388: investigation. Because most popular methods of surface analysis are based on
389: scaling features of some topographical measure, the examples were chosen with
390: respect to their different scaling properties as well as their results from our
391: analysis based on the theory of Markov processes.
392:
393: In the analysis presented here we use the well-known \emph{height increment}
394: $h_r(x)$, which has been defined in eq.~\ref{eq:h_r}, as a scale-dependent
395: measure of the complexity of rough surfaces \cite{CenteredInc}. Scaling
396: properties are reflected by the $r$-dependence of the so-called structure
397: functions
398: \begin{equation}
399: \label{eq:S^n}
400: S^n(r) = \langle |h_r^n| \rangle \,.
401: \end{equation}
402: If one then finds
403: \begin{equation}
404: \label{eq:scaling_def}
405: S^n(r) \sim r^{\xi_n}
406: \end{equation}
407: for a range of $r$, this regime is called the scaling range. In that range the
408: investigated profiles have \emph{self-affine} properties, i.e., they are
409: statistically invariant under an anisotropic scale transformation. If
410: furthermore the dependence of the exponents $\xi_n$ on the order $n$ is
411: nonlinear, one speaks of \emph{multi-affine} scaling. Those properties are no
412: longer identified by a single scaling exponent, but an infinite set of
413: exponents. A detailed explanation of self- and multi-affine concepts is beyond
414: the scope of this article. Instead, we would like to refer the reader to the
415: literature \cite{Feder1988,Family1991,Vicsek1992-e,Barabasi1995}.
416: %
417: The power spectrum, which often is used to determine scaling properties, can
418: easily be derived from the second order structure function. It is defined as
419: the Fourier transform of the autocorrelation function $R(r)$, which itself is
420: closely related to $S^2(r)$ by $R(r) = \langle h(x)^2\rangle -S^2(r)/2$,
421: by comparing eqs.~(\ref{eq:h_r}) and (\ref{eq:S^n}).
422:
423: In addition to the $r$-dependence of the structure functions, a generalized
424: form of scaling behaviour can be determined analogously to the Extended Self
425: Similarity (ESS) method which is popular in turbulence research
426: \cite{Benzi1993}. When the $S^n(r)$ are plotted against a structure function of
427: specific order, say $S^3(r)$, in many cases an extended scaling regime is
428: found according to
429: \begin{equation}
430: \label{eq:ESS}
431: S^n(r) \sim \left(S^3(r)\right)^{\zeta_n}\,.
432: \end{equation}
433: Clearly, meaningful results are restricted to the regime where $S^3$ is
434: monotonous. It is easy to see that now the $\xi_n$ can be obtained by
435: \begin{equation}
436: \label{eq:zeta_n}
437: \xi_n=\zeta_n\cdot\xi_3\,,
438: \end{equation}
439: cf.~\cite{Benzi1993}. While for turbulence it is widely accepted that by this
440: means experimental defiencies can be compensated to some degree, for surface
441: roughness the meaning of ESS lies merely in a generalized form of scaling
442: properties.
443:
444: It should be noted that the results of any scaling analysis may be influenced
445: by the method of measurement, by the definition of the roughness measure, here
446: $h_r(x)$ (or $w_r(x)$ as mentioned in section \ref{sec:intro}), as well as by
447: the algorithms used for the analysis \cite{Wendt2002a, Alber1998a}.
448: Nevertheless, this problem is not addressed here as the main focus of our
449: investigations is the application of the theory of Markov processes to
450: experimental data.
451:
452:
453: \subsection{Surfaces with scaling properties}
454: \label{sec:scaling_with}
455:
456: In fig.~\ref{fig:data_road_scaling} we present road surface data with
457: different kinds of scaling properties. For each data set a short profile
458: section is shown. Structure functions of order one to six on double
459: logarithmic scale are presented in fig.~\ref{fig:Sn_road_scaling}. Following
460: the arguments in \cite{Renner2001}, higher order structure functions cannot be
461: evaluated with sufficient precision from the given amount of data points.
462: %
463: The worn asphalt pavement (Road~1) is an example of a comparably large scaling
464: regime over more than one order of magnitude in $r$. A surface with similar
465: features, namely a cobblestone road, has already been presented
466: in~\cite{Waechter2003_plus_preprint}.
467: %
468: Two separate scaling regions are found for a Y-shaped concrete stone pavement
469: (Road~2). Additionally a sharp notch can be seen in the structure functions at
470: $r=0.2\un{m}$, indicating a strong periodicity of the pavement caused by the
471: length of the individual stones.
472: %
473: The third example, a ``pebbled concrete'' pavement (Road~3), consists of
474: concrete stones with a top layer of washed pebbles. This material is also known
475: as ``exposed aggregate concrete''. Here, the scaling region of the structure
476: functions is only small.
477: %
478: For the basalt stone pavement (Road~4), being the fourth example, scaling
479: properties are poor. We have nevertheless marked a possible scaling range
480: and derived the respective scaling exponents for comparison with the other
481: examples. Similar to the Y-shaped concrete stones, a periodicity can be found
482: at about 0.1\un{m} length scale.
483:
484: \newlength{\breite}\setlength{\breite}{0.8\linewidth}
485: \begin{figure}[htbp]\centering
486: % Beispiel fuer gutes scaling, ESS ok
487: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~1}}\end{picture}%
488: \includegraphics[width=\breite]{figures/figure_01a}
489:
490: % Beispiel fuer scaling mit 2 Bereichen, ESS ok
491: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~2}}\end{picture}%
492: \includegraphics[width=\breite]{figures/figure_01b}
493:
494: % Beispiel fuer schwaches scaling , ESS ok
495: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~3}}\end{picture}%
496: \includegraphics[width=\breite]{figures/figure_01c}
497:
498: % Beispiel fuer schwaches scaling , kein ESS
499: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~4}}\end{picture}%
500: \includegraphics[width=\breite]{figures/figure_01d}
501:
502: \caption{%
503: Measurement data from selected road surfaces with different kinds of
504: scaling properties. %
505: Pavements are worn asphalt (Road 1), Y-shaped concrete stones (Road 2),
506: pebbled concrete stones (Road 3), and basalt stones (Road 4), from top to
507: bottom. For each surface a short
508: section of the respective height profile is shown. }
509: \label{fig:data_road_scaling}
510: \end{figure}
511:
512: \begin{figure}[htbp]\centering
513: % Beispiel fuer gutes scaling, ESS ok
514: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~1}}\end{picture}%
515: \includegraphics[width=\breite]{figures/figure_02a}
516:
517: % Beispiel fuer scaling mit 2 Bereichen, ESS ok
518: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~2}}\end{picture}%
519: \includegraphics[width=\breite]{figures/figure_02b}
520:
521: % Beispiel fuer schwaches scaling , ESS ok
522: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~3}}\end{picture}%
523: \includegraphics[width=\breite]{figures/figure_02c}
524:
525: % Beispiel fuer schwaches scaling , kein ESS
526: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~4}}\end{picture}%
527: \includegraphics[width=\breite]{figures/figure_02d}
528:
529: \caption{%
530: Structure functions $S^n(r)$ of selected road surfaces (see
531: fig.~\ref{fig:data_road_scaling}) with different kinds of scaling
532: properties on a log-log scale (see text).
533: %
534: Symbols correspond to orders $n$; diamonds ($n=1$), triangles
535: ($n=2$), circles ($n=3$), squares ($n=4$), x signs ($n=5$), and plus signs
536: ($n=6$). }
537: \label{fig:Sn_road_scaling}
538: \end{figure}
539:
540: The results for the generalized scaling behaviour according to
541: eq.~(\ref{eq:ESS}) are shown in fig.~\ref{fig:data_road_scaling_ess}. It can be
542: seen that indeed for three of the surfaces in fig.~\ref{fig:data_road_scaling}
543: an improved scaling behaviour is found by this method. Only for Road~4 do the
544: generalized scaling properties remain weak.
545: %
546: In fig.~\ref{fig:data_road_scaling_exp} the scaling exponents $\xi_n$ of the
547: structure functions within the marked scaling regimes in
548: fig.~\ref{fig:Sn_road_scaling} were determined and plotted against the order
549: $n$ as open symbols. Additionally, values of $\xi_n$ were derived according to
550: eqs.~(\ref{eq:ESS}) and (\ref{eq:zeta_n}) and added as crosses. For Road~2 two
551: sets of exponents correspond to the two distinct scaling regimes in
552: fig.~\ref{fig:Sn_road_scaling}. Even though there is only one set of $\zeta_n$
553: found in fig.~\ref{fig:data_road_scaling_ess}, two sets of $\xi_n$ are
554: obtained due to the two different $\xi_3$, see eqs.~(\ref{eq:scaling_def}) and
555: (\ref{eq:zeta_n}).
556:
557: All surfaces from fig.~\ref{fig:data_road_scaling} show a more or less
558: nonlinear dependence of the $\xi_n$ on $n$, indicating multi-affine scaling
559: properties. The scaling exponents obtained via the generalized scaling
560: according to eq.~(\ref{eq:ESS}) are in good correspondence with the $\xi_n$
561: achieved by the application of eq.~(\ref{eq:scaling_def}). Deviations are seen
562: for Road~2 and at higher orders for Road~3, possibly caused by inaccuracies in
563: the fitting procedure. For Road~4 no generalized scaling is observed (compare
564: fig.~\ref{fig:data_road_scaling_ess}) and thus values of $\xi_n$ cannot be
565: derived from $\zeta_n$. From this we conclude that scaling properties for some
566: cases are questionable as a comprehensive tool to characterize the complexity
567: of a rough surface.
568:
569: \begin{figure}[htbp]\centering
570: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~1}}\end{picture}%
571: \includegraphics[width=\breite]{figures/figure_03a}
572:
573: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~2}}\end{picture}%
574: \includegraphics[width=\breite]{figures/figure_03b}
575:
576: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~3}}\end{picture}%
577: \includegraphics[width=\breite]{figures/figure_03c}
578:
579: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~4}}\end{picture}%
580: \includegraphics[width=\breite]{figures/figure_03d}
581:
582: \caption[Road surface scaling]{%
583: Generalized scaling analysis of the surfaces shown in
584: fig.~\ref{fig:data_road_scaling}.
585: Structure functions $S^n$ are displayed versus $S^3$ on a log-log scale.
586: Symbols correspond to orders $n$ as in fig.~\ref{fig:Sn_road_scaling}.
587: }
588: \label{fig:data_road_scaling_ess}
589: \end{figure}
590:
591: \begin{figure}[htbp]\centering
592: \begin{picture}(0,0)\put(10,29){\makebox{\small Road~1}}\end{picture}%
593: \includegraphics[width=\breite]{figures/figure_04a}
594:
595: \begin{picture}(0,0)\put(10,29){\makebox{\small Road~2}}\end{picture}%
596: \includegraphics[width=\breite]{figures/figure_04b}
597:
598: \begin{picture}(0,0)\put(10,29){\makebox{\small Road~3}}\end{picture}%
599: \includegraphics[width=\breite]{figures/figure_04c}
600:
601: \begin{picture}(0,0)\put(10,29){\makebox{\small Road~4}}\end{picture}%
602: \includegraphics[width=\breite]{figures/figure_04d}
603:
604: \caption[Road surface scaling]{%
605: Scaling exponents $\xi_n$ of the surfaces shown in
606: fig.~\ref{fig:data_road_scaling} achieved via eq.~(\ref{eq:scaling_def})
607: (open symbols) and those obtained via $\zeta_n$ from eq.~(\ref{eq:zeta_n})
608: (crosses). For Road~2 two sets of exponents were obtained from the two
609: scaling regimes found for $S^n(r)$ in fig.~\ref{fig:Sn_road_scaling}.}
610: \label{fig:data_road_scaling_exp}
611: \end{figure}
612:
613: An example for good scaling properties is the gold film surface (Au). To
614: increase statistical accuracy, increments are evaluated here in the direction
615: of the rows of the images as well as the columns.
616: %
617: In fig.~\ref{fig:data_gold} two of the 99 images under investigation are
618: shown. Figure~\ref{fig:Sn_gold} presents the structure functions $S^n(r)$,
619: derived from all images. The surface is randomly covered with granules which
620: show no typical diameter. A scaling regime of more than one order of
621: magnitude in $r$ is found for the structure functions $S^n(r)$ in
622: fig.~\ref{fig:Sn_gold}. Generalized scaling behaviour is clearly present
623: as shown in fig.~\ref{fig:data_gold_ess}(a). The scaling exponents $\xi_n$
624: presented in part (b) of the same figure are nearly linear in $n$, thus this
625: surface can not be regarded as multi-affine, but appears to be self-affine.
626: Here, the $\xi_n$ achieved via eq.~(\ref{eq:zeta_n}) match perfectly those
627: obtained from eq.~(\ref{eq:ESS}).
628:
629: \begin{figure}[htbp]
630: \setlength{\breite}{0.5\linewidth}
631: \begin{picture}(0,0)%
632: %\put(0,33){\makebox{\small Au}}%
633: \put(0, 2){\balken{0.4545\breite}{50\un{nm}}}%
634: \end{picture}%
635: \begin{picture}(0,0)%
636: %\put(30,33){\makebox{\small (b)}}%
637: \put(30, 2){\balken{0.4545\breite}{0.5\un{\mu m}}}%
638: \end{picture}%
639: \raisebox{2.5ex}{%
640: \includegraphics[width=\breite]{figures/figure_05a}%
641: \includegraphics[width=\breite]{figures/figure_05b}%
642: }%
643: \caption{%
644: AFM images of the Au film surface. Sidelengths are 110\un{nm} and
645: 1.1\un{\mu m}. The relative surface height is represented as gray level.
646: Maximum heights are 7.2\un{nm} and 13.3\un{nm}, respectively.
647: }
648: \label{fig:data_gold}
649: \end{figure}
650:
651: \begin{figure}[htbp]\centering
652: \setlength{\breite}{\linewidth}
653: \includegraphics[width=0.8\breite]{figures/figure_06}
654: \caption{%
655: Structure functions $S^n(r)$ of the Au film surface on a log-log scale.
656: Symbols correspond to orders $n$ as in fig.~\ref{fig:Sn_road_scaling}.
657: }
658: \label{fig:Sn_gold}
659: \end{figure}
660:
661: \begin{figure}[htbp]\centering
662: \begin{picture}(0,0)\put(8,29){\makebox{\small (a)}}\end{picture}%
663: \includegraphics[width=\breite]{figures/figure_07a}
664: \begin{picture}(0,0)\put(8,29){\makebox{\small (b)}}\end{picture}%
665: \includegraphics[width=\breite]{figures/figure_07b}
666: \caption{%
667: Generalized scaling properties (a) and scaling exponents (b) of the
668: Au film surface shown in fig.~\ref{fig:data_gold}. Scaling exponents
669: $\xi_n$ achieved via eq.~(\ref{eq:ESS}) are marked by open symbols, those
670: obtained via $\zeta_n$ from eq.~(\ref{eq:zeta_n}) by crosses. Compare also
671: with fig.~\ref{fig:data_road_scaling_ess}. }
672: \label{fig:data_gold_ess}
673: \end{figure}
674:
675:
676: \subsection{Surfaces without scaling properties}
677: \label{sec:scaling_without}
678:
679: To complete the set of examples, we present two surfaces without scaling
680: properties. The first one is a smooth asphalt road (Road~5), shown in
681: fig.~\ref{fig:data_road_no_scaling}. No power law can be detected for the
682: $S^n(r)$ but a generalized scaling is observed in
683: fig.~\ref{fig:data_road_no_scaling_ess}(a). The range of values of $S^3$,
684: however, is relatively small.
685:
686: \begin{figure}[htbp]\centering
687: \begin{picture}(0,0)\put(8,29){\makebox{\small (a)}}\end{picture}%
688: \includegraphics[width=\breite]{figures/figure_08a}
689: \begin{picture}(0,0)\put(8,29){\makebox{\small (b)}}\end{picture}%
690: \includegraphics[width=\breite]{figures/figure_08b}
691: \caption{%
692: Measurement data (a) and structure functions (b) from a road surface
693: without scaling properties (Road~5). The pavement is smooth asphalt. See
694: also figs.~\ref{fig:data_road_scaling} and \ref{fig:Sn_road_scaling}.}
695: \label{fig:data_road_no_scaling}
696: \end{figure}
697:
698: The second example lacking a scaling regime is the steel fracture surface
699: (Crack). One of the three CLSM images under investigation is shown in
700: fig.~\ref{fig:Wendt_data}(a). Figure \ref{fig:Wendt_data}(b) presents an
701: additional REM image at a higher resolution, which gives an impression of the
702: surface morphology.
703: %
704: For the structure functions in fig.~\ref{fig:Sn_Wendt} no
705: scaling properties are found, and the dependences of $S^n(r)$ on $S^3(r)$ in
706: fig.~\ref{fig:data_road_no_scaling_ess}(b) also deviate from proper power laws.
707: %
708: It should be noted that in general scaling properties not only depend on the
709: respective data set but also on the analysis procedure. Using other measures
710: than $h_r(x)$, in \cite{Wendt2002a} scaling regimes of those measures have
711: been found, and scaling exponents could be obtained.
712:
713: \begin{figure}[htbp]
714: \setlength{\breite}{0.5\linewidth}%
715: \begin{picture}(0,0)%
716: \put(0,33){\makebox{\small (a)}}%
717: \put(0, 2){\balken{0.4\breite}{200\un{\mu m}}}%
718: \end{picture}%
719: \begin{picture}(0,0)%
720: \put(30,33){\makebox{\small (b)}}%
721: \put(30, 2){\balken{0.7143\breite}{100\un{\mu m}}}%
722: \end{picture}%
723: \raisebox{2.5ex}{%
724: \includegraphics[width=\breite]{figures/figure_09a}%
725: \includegraphics[width=1.028\breite]{figures/figure_09b}%
726: }%
727: \caption{%
728: Measurement data from a steel crack surface (Crack). (a) CLSM image, side
729: length 502\mum, (b) REM image, side length 140\mum. }
730: \label{fig:Wendt_data}
731: \end{figure}
732:
733: \begin{figure}[htbp]\centering
734: \includegraphics[width=0.8\linewidth]{figures/figure_10}
735: \caption{%
736: Structure functions $S^n(r)$ of the CLSM images from a steel crack surface
737: (Crack) on a log-log scale. The symbols correspond to orders $n$ as in
738: fig.~\ref{fig:Sn_road_scaling}. }
739: \label{fig:Sn_Wendt}
740: \end{figure}
741:
742: \begin{figure}[htbp]\centering
743: \setlength{\breite}{0.8\linewidth}%
744: \begin{picture}(0,0)\put(8,29){\makebox{\small Road~5 (a)}}\end{picture}%
745: \includegraphics[width=\breite]{figures/figure_11a}
746: \begin{picture}(0,0)\put(8,29){\makebox{\small Crack (b)}}\end{picture}%
747: \includegraphics[width=\breite]{figures/figure_11b}
748: \caption{%
749: Generalized scaling properties of the surfaces shown in
750: figs.~\ref{fig:data_road_no_scaling} (Road~5 (a)) and \ref{fig:Wendt_data}
751: (Crack (b)). Compare also with
752: figs.~\ref{fig:data_road_scaling_ess} and \ref{fig:data_gold_ess}. }
753: \label{fig:data_road_no_scaling_ess}
754: \end{figure}
755:
756:
757: \subsection{Conclusions on scaling analysis}
758: \label{sec:scaling_conclusions}
759:
760: To conclude the scaling analysis of our examples, we have chosen surfaces with
761: a range of different scaling properties from good scaling over comparably wide
762: ranges, such as for Road~1 and Au, to the absence of scaling, such as for
763: Road~5 and Crack. The generalized scaling analysis, analogous to ESS
764: \cite{Benzi1993}, leads to the same scaling exponents as the dependence of the
765: structure functions $S^n(r)$ on the scale $r$, with some minor deviations.
766:
767:
768: \section{Markov properties}
769: \label{sec:markov_props}
770:
771: As outlined in sections \ref{sec:intro} and \ref{sec:Markov_theory}, we want to
772: describe the evolution of the height increments $h_r(x)$ in the scale variable
773: $r$ as realizations of a Markov process with the help of a Fokker-Planck
774: equation. Consequently, the first step in the analysis procedure has to be the
775: verification of the Markov properties of $h_r(x)$ as a stochastic variable in
776: $r$.
777:
778: For a Markov process the defining feature is that the $n$-scale conditional
779: probability distributions are equal to the single conditional probabilities,
780: according to eq.~(\ref{eq:markov_straight}). With the given amount of data
781: points the verification of this condition is only possible for three different
782: scales. Additionally the scales $r$ are limited by the available profile
783: length.
784: %
785: For the sake of simplicity we will always take $r_3-r_2=r_2-r_1=\Delta r$.
786: Thus we can test the validity of eq.~(\ref{eq:markov_straight}) in the form
787: \begin{equation}
788: \label{eq:markov_simple}
789: p(h_1,r_1|\,h_2, r_1\!+\!\Delta r)=
790: p(h_1,r_1|\,h_2, r_1\!+\!\Delta r; h_3=0, r_1\!+\!2\Delta r)\,.
791: \end{equation}
792: Note that in eq.~(\ref{eq:markov_simple}) we take $h_3=0$ to restrict the
793: number of free parameters in the pdf with double conditions.
794:
795: Three procedures were applied to find out if Markov properties exist
796: for our data. From the results of all three tests we will find a minimal
797: length scale $l_M$ for which this is the case. The meaning of this so-called
798: Markov length will be discussed below. In the following we will demonstrate
799: the methods using the example of the Au surface.
800:
801:
802: \subsection{Testing procedures}
803:
804: The most straightforward way to verify eq.~(\ref{eq:markov_simple}) is the
805: visual comparison of both sides, i.e., the pdf with single and double
806: conditions. This is illustrated in fig.~\ref{fig:Schimmel_markov} for two
807: different scale separations $\Delta r=17\un{nm}$ and 35\un{nm}. In each case a
808: contour plot of single and double conditional probabilities $p(h_1,r_1 |
809: h_2,r_2)$ and $p(h_1,r_1 | h_2,r_2; h_3\!\!=\!\!0,r_3)$ is presented in the top
810: panel of (a) and (b), respectively.
811: Below two one-dimensional cuts at fixed values of $h_2 \approx \pm
812: \sigma_\infty$ are shown, representing directly
813: $p(h_1,r_1|h_2\!\!=\!\!\pm\sigma_\infty,r_2;h_3\!\!=\!\!0,r_3)$.
814: %
815: It can be seen that in panel (a), for the smaller value of $\Delta r$, the
816: single and double conditional probability are different. This becomes clear
817: from the crossing solid and broken contour lines of the contour plot as well
818: as from the differing lines and symbols of the one-dimensional plots below.
819: Panel (b), for $\Delta r=35$\un{nm}, shows good correspondence of both
820: conditional pdf. We take this finding as a strong hint that for this scale
821: separation $\Delta r$ eq.~(\ref{eq:markov_straight}) is valid and Markov
822: properties exist. Following this procedure for all accessible values of
823: $\Delta r$, the presence of Markov properties was examined. For this surface
824: Markov properties were found for scale distances from $(25\pm 5)\un{nm}$
825: upwards.
826:
827:
828: \setlength{\breite}{0.66\linewidth}
829: \begin{figure}[htbp]\centering
830: \begin{picture}(0,0)\put(0,49){(a)}\end{picture}%
831: \includegraphics[width=\breite]{figures/figure_12a}\\
832: \begin{picture}(0,0)\put(0,49){(b)}\end{picture}%
833: \includegraphics[width=\breite]{figures/figure_12b}
834: \caption{%
835: Test for Markov properties of Au film data for two different scale
836: separations $\Delta r=14$\un{nm} (a) and $35$\un{nm} (b), where
837: \mbox{$\Delta r=r_3-r_2=r_2-r_1$} (see text). In both cases
838: $r_2=169\un{nm}$.
839: %
840: In each case a contour plot of conditional probabilities $p(h_1,r_1 |
841: h_2,r_2)$ (dashed lines) and $p(h_1,r_1 | h_2,r_2;h_3\!\!=\!\!0,r_3 )$
842: (solid lines) is shown in the top panel. Contour levels differ by a factor
843: of 10, with an additional level at $p=0.3$. Below the top panels in each
844: case, two one-dimensional cuts
845: at $h_2 \approx \pm \sigma_\infty$ are shown with $p(h_1,r_1 | h_2,r_2)$ as
846: dashed lines and $p(h_1,r_1 | h_2,r_2;h_3\!\!=\!\!0,r_3 )$ as circles. }
847: \label{fig:Schimmel_markov}
848: \end{figure}
849:
850: The validity of eq.~(\ref{eq:markov_simple}) can also be be quantified
851: mathematically using statistical tests. An approach via the well-known
852: $\chi^2$ measure has been presented in \cite{Friedrich1998b}, whereas in
853: \cite{Renner2001} the Wilcoxon test has been used. Next, we give a brief
854: introduction to this procedure, which will be used here, too. More detailed
855: discussions of this test can be found in
856: \cite{Bronstein1991,Renner2001,Renner2001Diss}. For this procedure, we
857: introduce the notation of two stochastic variables $x_i,i=1,\ldots,n$ and
858: $y_j,j=1,\ldots,m$ which represent the two samples from which both conditional
859: pdf of eq.~(\ref{eq:markov_straight}) are estimated, i.e.\
860: \begin{eqnarray}
861: \label{eq:wilcox_xy}
862: x(h_{r_2}, r_1, r_2) &=& h_{r_1}|_{h_{r_2}} \nonumber\\
863: y(h_{r_2}, h_{r_3}, r_1, r_2, r_3) &=& h_{r_1}|_{h_{r_2}; h_{r_3}} \;.
864: \end{eqnarray}
865: Here $\cdot|_{h_x}$ denotes the conditioning.
866: All events of both samples are sorted together in ascending order into one
867: sequence, according to their value. Now the total number of so called
868: inversions is counted, where the number of inversions for a single event $y_j$
869: is just the number of events of the other sample which have a smaller value
870: $x_i < y_j$. If eq.~(\ref{eq:markov_simple}) holds and $n,m\geq 25$, the
871: total number of inversions $Q$ is Gaussian distributed with
872: \begin{eqnarray}
873: \label{eq:Q_mean_sigma}
874: \langle Q\rangle &=& nm/2 \quad\mathrm{and} \nonumber \\
875: \sigma_Q &=& \sqrt{nm(n+m+1)/12}\; .
876: \end{eqnarray}
877: We normalize $Q$ with respect to its standard deviation and consider the
878: absolute value
879: \begin{equation}
880: \label{eq:wilcox_t}
881: t = | Q - \langle Q\rangle | / \sigma_Q \; .
882: \end{equation}
883: For its expectation value it is easy to show that $\langle
884: t\rangle=\sqrt{2/\pi}$ (still provided that (\ref{eq:markov_simple}) is valid),
885: where here the average $\langle\cdot\rangle$ is performed over $h_2$. If a
886: larger value of $\langle t\rangle$ is measured for a specific combination of
887: $r$ and $\Delta r$, we conclude that eq.~(\ref{eq:markov_simple}) is not
888: fulfilled and thus Markov properties do not exist. A practical problem with
889: the Wilcoxon test is that all events $x_i,y_j$ have to be statistically
890: independent. This means that the intervals of subsequent height increments
891: $h_r$ have to be separated by the largest scale involved. Thus the number of
892: available data is dramatically reduced.
893:
894: \setlength{\breite}{0.8\linewidth}
895: \begin{figure}[htbp]
896: \centering
897: \includegraphics[width=\breite]{figures/figure_13}
898: \caption{%
899: Wilcoxon test for the Au surface. The scale $r$ is 28\un{nm}. The
900: theroretically expected value $\langle t\rangle=\sqrt{2/\pi}$ is marked
901: with a horizontal line, the Markov length $l_M=(25\pm5)\un{nm}$ with a
902: vertical line. }
903: \label{fig:Schimmel_wit}
904: \end{figure}
905:
906: In fig.~\ref{fig:Schimmel_wit} we present for the Au surface measured values
907: of $\langle t(r,\Delta r)\rangle$ at a scale $r=28\un{nm}$. The Markov length
908: $l_M$ is marked where $\langle t(r,\Delta r)\rangle$ has approached its
909: theroretical value $\sqrt{2/\pi}$.
910:
911: Another method to show the validity of condition (\ref{eq:markov_simple}) is
912: the investigation of the well-known necessary condition for a Markovian
913: process, the validity of the Chapman-Kolmogorov equation \cite{Risken1984}
914: \begin{equation}
915: \label{eq:CKE}
916: p(h_1, r_1|h_3,r_3) =
917: \int_{-\infty}^{+\infty} p(h_1, r_1|h_2,r_2)p(h_2, r_2|h_3,r_3)\mbox{d}h_2
918: \;.
919: \end{equation}
920: We use this equation as a method to investigate the Markov properties of our
921: data. This procedure was used for example in
922: \cite{Friedrich1997a,Friedrich1997b,Ragwitz2001,Jafari2003} for the
923: verification of Markov properties. It also served to show for the first time
924: the existence of a Markov length in \cite{Friedrich1998b}.
925: %
926: The conditional probabilities in eq.~(\ref{eq:CKE}) are directly
927: estimated from the measured profiles. In
928: fig.~\ref{fig:Schimmel_markov_chapman} both sides of the Chapman-Kolmogorov
929: equation are compared for two different values of $\Delta r$. In an analogous
930: way to fig.~\ref{fig:Schimmel_markov}, for each $\Delta r$ the two conditional
931: probabilities are presented together in a contour plot as well as in two cuts
932: at fixed values of $h_2$. While for the smaller value $\Delta r=14$\un{nm}
933: both the contour lines and the cuts at fixed $h_2$ clearly differ, we find
934: a good correspondence for the larger value $\Delta r=35$\un{nm}.
935:
936: \setlength{\breite}{0.66\linewidth}
937: \begin{figure}[htbp]
938: \centering
939: \begin{picture}(0,0)\put(0,49){(a)}\end{picture}%
940: \includegraphics[width=\breite]{figures/figure_14a}
941: \begin{picture}(0,0)\put(0,49){(b)}\end{picture}%
942: \includegraphics[width=\breite]{figures/figure_14b}
943: \caption{%
944: A check of the Chapman-Kolmogorov equation (\ref{eq:CKE}) for the Au
945: surface for two different scale separations $\Delta r=14$\un{nm} (a) and
946: 35\un{nm} (b). In both cases $r_2=169$\un{nm}. The plots are organized in
947: the same way as fig.~\ref{fig:Schimmel_markov}. The pdf representing the
948: left side of (\ref{eq:CKE}) are shown with solid lines, the integrated pdf
949: of the right side of (\ref{eq:CKE}) as dashed lines and circles. }
950: \label{fig:Schimmel_markov_chapman}
951: \end{figure}
952:
953: A third method which we did not use here but which is reported in the
954: literature is based on the description of the stochastic process by a Langevin
955: equation. With this knowledge of the Langevin equation (\ref{eq:Langevin}) the
956: noise can be reconstructed and analyzed with respect to its correlation
957: \cite{Siefert2003,Marcq2001}.
958:
959:
960: \subsection{Conclusions on Markov properties}
961: \label{sec:markov_conclusions}
962:
963: The results of the methods described above were combined to determine
964: whether Markov properties of the height increment $h_r(x)$ in the scale
965: variable are present for our surface measurements.
966: We found Markov properties for all the selected examples of surface measurement
967: data.
968: %
969: \begin{table}[htbp]
970: \newcommand{\mm}{\un{mm}}
971: \begin{center}
972: \begin{tabular}{cc|cc|cc}
973: Surface & $l_M$ & Surface & $l_M$ & Surface & $l_M$ \\
974: \hline
975: Road~1 & 33\mm & Road~2 & 10\mm & Road~3 & 4.2\mm \\
976: Road~4 & 17\mm & Road~5 & 4.2\mm& Au & 25\un{nm}\\
977: Crack & 20\mum& & & &
978: \end{tabular}
979: \caption{Markov lengths $l_M$ for all the surfaces presented.}
980: \label{tab:l_M}
981: \end{center}
982: \end{table}
983: %
984: It is also common to all examples that these Markov properties are not
985: universal for all scale separations $\Delta r$ but there exists a lower
986: threshold which we call the Markov length $l_M$. It was determined in each
987: case by systematic application of the three testing procedures for all
988: accessible length scales $r$ and scale separations $\Delta r$.
989: The resulting values are listed in table~\ref{tab:l_M}. The presence of
990: Markov properties only for values of $\Delta r$ above a certain threshold has
991: also been found for stochastic data generated by a large variety of processes
992: and especially occurs in turbulent velocities
993: \cite{Friedrich1998a,Friedrich1998b,Friedrich2000b,Renner2001,Ghasemi2003}.
994:
995: The meaning of this Markov length $l_M$ may be seen in comparison with a mean
996: free path length of a Brownian motion. Only above this mean free path is a
997: stochastic process description valid. For smaller scales there must be some
998: coherence which prohibits a description of the structure by a Markov process.
999: If for example the description of a surface structure requires a second order
1000: derivative in space, a Langevin equation description (\ref{eq:Langevin})
1001: becomes impossible. In this case a higher dimensional Langevin equation (at
1002: least two variables) is needed. It may be interesting to note that the Markov
1003: length we found for the Au surface of about 30\un{nm} coincides quite well with
1004: the size of the largest grain structures we see in fig.~\ref{fig:data_gold}.
1005: Thus the Au surface may be thought of as a composition of grains (coherent
1006: structures) by a stochastic Markov process.
1007:
1008: In the case of Road~2 with its strong periodicity at 0.2\un{m} the Markov
1009: properties end slightly above this length scale. It seems evident that here
1010: the Markov property is destroyed by the periodicity. While some of the other
1011: surfaces also have periodicities, these are never as sharp as for Road~2. An
1012: upper limit for Markov properties could not be found for any of the other
1013: surfaces.
1014:
1015: Another interesting finding can be seen from figs.~\ref{fig:Sn_road_scaling},
1016: \ref{fig:Sn_gold}, \ref{fig:data_road_no_scaling}, and \ref{fig:Sn_Wendt}.
1017: There is no connection between the scaling range and the range where Markov
1018: properties hold. Regimes of scaling and Markov properties are found to be
1019: distinct, overlapping or covering, depending on the surface. Data sets which
1020: fulfill the Markov property do not in all cases show a scaling regime at all.
1021: Also, on the other hand, scaling features seem not to imply Markov properties,
1022: which has been indicated previously for some numerically generated data in
1023: \cite{Friedrich1998a}. While there is always an upper limit of the scaling
1024: regime, we found only one surface for which the Markov properties possess an
1025: upper limit.
1026:
1027:
1028: \section{Estimation of drift and diffusion coefficients}
1029: \label{sec:Dk_estimation}
1030:
1031:
1032: As a next step we want to concentrate on extracting the concrete form of the
1033: stochastic process, if the Markov properties are fulfilled.
1034: %
1035: As mentioned in section~\ref{sec:Markov_theory} our analysis is based on the
1036: estimation of Kramers-Moyal coefficients.
1037: %
1038: The procedure we use to obtain the drift ($D^{(1)}$) and diffusion coefficient
1039: ($D^{(2)}$) for the Fokker-Planck equation (\ref{eq:FPE1}) was already outlined
1040: by Kolmogorov \cite{Kolmogorov1931}, see also \cite{Risken1984,Renner2001}.
1041: %
1042: First, the conditional moments $M^{(k)}(h_r,r,\Delta r)$ for finite step sizes
1043: $\Delta r$ are estimated from the data via the moments of the conditional
1044: probabilities.
1045: %
1046: This is done by application of the definition in eq.~(\ref{eq:Mk_def}),
1047: which is recalled here:
1048: % (for enhanced clarity, we use $\tilde{h}$ for $\tilde{h}_{r-\Delta r}$)
1049: %
1050: \begin{equation}
1051: \label{eq:Mk_2}
1052: M^{(k)}(h_r,r,\Delta r) =
1053: \frac{r}{k!\Delta r}
1054: \int_{\scriptscriptstyle-\infty}^{\scriptscriptstyle+\infty}
1055: (\tilde{h}-h_r)^k \; p(\tilde{h},r-\Delta r | h_r,r) \; d\tilde{h}
1056: \end{equation}
1057: %
1058: The conditional probabilities in the integral are obtained by counting events
1059: in the measurement data as shown already in section~\ref{sec:markov_props}.
1060: Here, one fundamental difficulty of the method arises: For reliable estimates
1061: of conditional probabilities we need a sufficient number of events even for
1062: rare combinations of $\tilde{h}, h_r$. Consequently, a large amount of data
1063: points is needed.
1064: %
1065: This problem becomes even more important if one takes into account that a large
1066: range in $r$ should be considered. The number of statistically independent
1067: intervals $h_r$ is limited by the length of the given data set and decreases
1068: with increasing $r$.
1069:
1070: In a second step, the coefficients $D^{(k)}(h_r,r)$ are obtained from the
1071: limit of $M^{(k)}(h_r,r,\Delta_r)$ when $\Delta r$ approaches zero (see
1072: definition in eq.~(\ref{eq:Dk_def})).
1073: For fixed values of $r$ and $h_r$ a straight line is fitted to the sequence of
1074: $M^{(k)}(h_r,r,\Delta r)$ depending on $\Delta r$ and extrapolated against
1075: $\Delta r=0$.
1076: %
1077: The linear dependence corresponds to the lowest order term when the $\Delta
1078: r$-dependence of $M^{(k)}(h_r,r,\Delta r)$ is expanded into a Taylor series
1079: for a given Fokker-Planck equation \cite{Friedrich2002, Siefert2003}.
1080: %
1081: Our interpretation is that this way of estimating the $D^{(k)}$ is the most
1082: advanced one, and also performs better than first parameterizing the $M^{k}$
1083: and then estimating the limit $\Delta r\rightarrow 0$ for this
1084: parameterization, as previously suggested in \cite{Renner2001,Renner2001Diss}.
1085:
1086: There have been suggestions to fit functions to $M^{(k)}$ other than a straight
1087: line, especially for the estimation of $D^{(2)}$, see \cite{Renner2001Diss}.
1088: Furthermore it has been proposed to use particular terms of the above-mentioned
1089: expansion to directly estimate $D^{(k)}(h_r, r)$ without
1090: extrapolation \cite{Ragwitz2001}.
1091: %
1092: On the other hand, in \cite{Sura2002} it becomes clear that there can be
1093: manifold dependences of $M^{(k)}$ on $\Delta r$ which in general are not known
1094: for a measured data set. Consequently, one may state that there is still a
1095: demand to improve the estimation of $D^{(k)}$. At the present time we
1096: suggest to show the quality of the estimated $D^{(k)}$ by verification of the
1097: resulting Fokker-Planck equation, once its drift and diffusion coefficients
1098: have been estimated.
1099: %
1100: However, for our data neither nonlinear fitting functions nor correction terms
1101: applied to the $M^{(k)}$ resulted in improvements of the estimated $D^{(k)}$.
1102:
1103: A crucial point in our estimation procedure is the range of $\Delta r$ where
1104: the fit can be performed. Only those $\Delta r$ can be used where Markov
1105: properties were found in the scale domain. In section \ref{sec:markov_props} we
1106: showed that for our data Markov properties are given for $\Delta r$ larger than
1107: the Markov length $l_M$ (see table \ref{tab:l_M}). In order to reduce
1108: uncertainty, a large range of $\Delta r$ as the basis of the extrapolation is
1109: desirable. From eq.~(\ref{eq:Mk_2}) it can be seen, however, that $\Delta r$
1110: must be smaller than $r$. As a compromise between accuracy and extending the
1111: scale $r$ to smaller values, in many cases an extrapolation range of $l_M \leq
1112: \Delta r \leq 2\,l_M$ was used (cf.\ table \ref{tab:l_extrapol}). This
1113: procedure is shown in fig.~\ref{fig:drh_ex} for Road~1.
1114:
1115: \begin{table}[htbp]
1116: \centering
1117: \newcommand{\mm}{\un{mm}}
1118: \newcommand{\lmax}{\ensuremath{l_{\mathit{max}}}}
1119: \begin{tabular}{ccc|ccc}
1120: Surface& $l_M$ & \lmax& Surface& $l_M$ & \lmax \\
1121: \hline
1122: Road~1 & 33\mm & 67\mm& Road~2 & 10\mm & 21\mm \\
1123: Road~3 & 4.2\mm& 19\mm& Road~4 & 17\mm & 25\mm \\
1124: Road~5 & 4.2\mm&8.3\mm& Au &25\un{nm}&84\un{nm}\\
1125: Crack & 20\mum&44\mum& & &
1126: \end{tabular}
1127: \caption{Extrapolation ranges for all the presented surfaces. Listed are
1128: the smallest ($l_M$) and largest ($l_\mathit{max}$) values of $\Delta r$
1129: used for extrapolation of the $D^{(k)}(h_r, r)$. }
1130: \label{tab:l_extrapol}
1131: \end{table}
1132:
1133: \setlength{\breite}{0.8\linewidth}
1134: \begin{figure}[htbp]
1135: \centering
1136: \begin{picture}(0,0)\put(0,30){\makebox{\small (a)}}\end{picture}%
1137: \includegraphics[width=\breite]{figures/figure_15a}
1138: \begin{picture}(0,0)\put(0,30){\makebox{\small (b)}}\end{picture}%
1139: \includegraphics[width=\breite]{figures/figure_15b}
1140: \caption[Extrapolation procedure for $D^{(1)}$.]{%
1141: Extrapolation procedure for $D^{(1)}$ (a) and $D^{(2)}$ (b), illustrated
1142: for surface Road~1. Length scale $r$ is 108\un{mm}, $h_r$ is
1143: $-\sigma_\infty$. Values of $M^{(1)}$ and $M^{(2)}$ inside the range
1144: marked by broken lines were used for the extrapolation. The results
1145: $D^{(1)}(h_r,r)$ and $D^{(2)}(h_r,r)$ are marked with filled circles on
1146: the vertical axis. }
1147: \label{fig:drh_ex}
1148: \end{figure}
1149:
1150:
1151: \subsection{Estimation results}
1152: \label{sec:Dk_results}
1153:
1154: Following the procedure outlined above, $D^{(1)}(h_r,r)$ and $D^{(2)}(h_r,r)$
1155: were derived for the measurement data presented in section
1156: \ref{sec:measurement_data}, with the exception of Road~5 (see below).
1157: %
1158: For the road surfaces, estimations were performed for length scales $r$
1159: separated by ten measurement steps or 10.4\un{mm}, respectively, to reach a
1160: sufficient density over the range where the coefficients were accessible.
1161: %
1162: Figures~\ref{fig:D1_road_scaling} and \ref{fig:D2_road_scaling} show
1163: estimations of the drift coefficients $D^{(1)}$ and the diffusion coefficients
1164: $D^{(2)}$ for the road surfaces, each performed for one fixed length scale
1165: $r$. The error bars are estimated from the errors of $M^{(k)}(h_r,r,\Delta
1166: r)$ via the number of statistically independent events contributing to each
1167: value, assuming that each bin of $p(h_1,r_1|h_0,r_0)$ containing $N$ events
1168: has an intrinsic uncertainty of $\pm\sqrt{N}$.
1169: %
1170: Additionally, values of $D^{(4)}$ are added to the plots of $D^{(2)}$ which
1171: have been estimated in the same way. Thus it can be seen that in all cases
1172: $D^{(4)}$ is small compared to $D^{(2)}$, except for Road~4, and in most cases
1173: its statistical errors are larger than the values themselves. Negative values
1174: are not shown because the vertical axes start at zero. As $M^{(4)}$ is positive
1175: by definition, the occurence of negative values of $D^{(4)}$ results from the
1176: limit $\Delta r\rightarrow 0$ and should be only due to the statistical errors
1177: involved.
1178: %
1179: Even if there is no evidence that $D^{(4)}$ is identically zero, the presented
1180: values give a hint that its influence in the Kramers-Moyal expansion
1181: (\ref{eq:KME}) is rather small and the assumption of a Fokker-Planck equation
1182: (\ref{eq:FPE1}) is justifiable, with the possible exception of Road~4.
1183:
1184: \setlength{\breite}{0.8\linewidth}
1185: \begin{figure}[htbp]
1186: \centering
1187: \begin{picture}(0,0)\put(35,28){\makebox{\small Road~1}}\end{picture}%
1188: \includegraphics[width=\breite]{figures/figure_16a}
1189: \begin{picture}(0,0)\put(35,28){\makebox{\small Road~2}}\end{picture}%
1190: \includegraphics[width=\breite]{figures/figure_16b}
1191: \begin{picture}(0,0)\put(35,28){\makebox{\small Road~3}}\end{picture}%
1192: \includegraphics[width=\breite]{figures/figure_16c}
1193: \begin{picture}(0,0)\put(35,28){\makebox{\small Road~4}}\end{picture}%
1194: \includegraphics[width=\breite]{figures/figure_16d}
1195: \caption[Estimated drift coefficient $D^{(1)}(h_r, r)$]%
1196: {Estimated drift coefficients $D^{(1)}(h_r, r)$ of the Fokker-Planck
1197: equation for the road surfaces shown in fig.~\ref{fig:data_road_scaling}.
1198: Scales $r$ are 108\un{mm} (Road~1), 114\un{mm} (Road~2), 94\un{mm}
1199: (Road~3), and 104\un{mm} (Road~4). Parameterizations are shown as lines.}
1200: \label{fig:D1_road_scaling}
1201: \end{figure}
1202:
1203: \begin{figure}[htbp]
1204: \centering
1205: \begin{picture}(0,0)\put(23,28){\makebox{\small Road~1}}\end{picture}%
1206: \includegraphics[width=\breite]{figures/figure_17a}
1207: \begin{picture}(0,0)\put(23,28){\makebox{\small Road~2}}\end{picture}%
1208: \includegraphics[width=\breite]{figures/figure_17b}
1209: \begin{picture}(0,0)\put(23,28){\makebox{\small Road~3}}\end{picture}%
1210: \includegraphics[width=\breite]{figures/figure_17c}
1211: \begin{picture}(0,0)\put(23,28){\makebox{\small Road~4}}\end{picture}%
1212: \includegraphics[width=\breite]{figures/figure_17d}
1213: \caption[Estimated drift coefficient $D^{(1)}(h_r, r)$]%
1214: {Estimated diffusion coefficients $D^{(2)}(h_r, r)$ (circles) of the
1215: Fokker-Planck equation for road surfaces shown in
1216: fig.~\ref{fig:data_road_scaling}. Additionally the fourth Kramers-Moyal
1217: coefficients $D^{(4)}(h_r, r)$ are shown as squares. Scales $r$ are as in
1218: fig.~\ref{fig:D1_road_scaling}. Parameterizations are shown as lines.}
1219: \label{fig:D2_road_scaling}
1220: \end{figure}
1221:
1222: Estimated drift and diffusion coefficients $D^{(1)}$ and $D^{(2)}$ for the Au
1223: surface are shown in fig.~\ref{fig:Schimmel_D12} for $r=169$\un{nm}. Again,
1224: $D^{(4)}$ was added to the plot of $D^{(2)}$, in this case without error bars
1225: to enhance clarity. Errors of $D^{(4)}$ are in this case always much larger
1226: than the values themselves and would cover the values of $D^{(2)}$ as well as
1227: their errors. Also the error bars of $D^{(2)}$ appear to be quite large for
1228: the Au surface. The data here are measured as two-dimensional images, thus
1229: the number of statistically independent $h_r(x)$ decreases quadratically with
1230: increasing $r$, resulting in rather large error estimates. For the calculation
1231: of $D^{(k)}$ nevertheless all accessible $h_r(x)$ were used.
1232: %
1233: As the regime of Markov properties starts at $\Delta r=25\un{nm}$, the range
1234: $25\un{nm} \leq \Delta r \leq 84\un{nm}$ was used as basis for the
1235: extrapolation (see table~\ref{tab:l_extrapol}). For $r<84\un{nm}$ the upper
1236: limit was reduced in order to derive the coefficients also for smaller scales
1237: $r$ (compare also section \ref{sec:Dk_estimation}). In this way the drift and
1238: diffusion coefficients of the Au film could be worked out from 281 down to
1239: 56\un{nm}.
1240:
1241: \begin{figure}[htbp]
1242: \centering
1243: \begin{picture}(0,0)\put(40,28){\makebox{\small (a)}}\end{picture}%
1244: \includegraphics[width=\breite]{figures/figure_18a}
1245: \begin{picture}(0,0)\put(22,28){\makebox{\small (b)}}\end{picture}%
1246: \includegraphics[width=\breite]{figures/figure_18b}
1247: \caption{%
1248: Estimated drift (a) and diffusion (b) coefficient of the Au surface for
1249: $r=169$\un{nm}. Estimates of $D^{(4)}$ are added as squares in (b).}
1250: \label{fig:Schimmel_D12}
1251: \end{figure}
1252:
1253: In the same way as for the other surfaces, estimations of the Kramers-Moyal
1254: coefficients were performed for the steel crack. The results are shown in
1255: fig.~\ref{fig:Wendt_D1234}. Again, the estimates for $D^{(4)}$ are also
1256: presented, which are of the same order of magnitude as $D^{(2)}$ for
1257: higher values of $h_r$ ($|h_r|>0.5\sigma_\infty$).
1258:
1259: \begin{figure}[htbp]
1260: \centering
1261: \begin{picture}(0,0)\put(40,28){\makebox{\small (a)}}\end{picture}%
1262: \includegraphics[width=\breite]{figures/figure_19a}
1263: \begin{picture}(0,0)\put(22,28){\makebox{\small (b)}}\end{picture}%
1264: \includegraphics[width=\breite]{figures/figure_19b}
1265: \caption{%
1266: Estimated Kramers-Moyal coefficients %$D^{(1)}$, $D^{(2)}$, and $D^{(4)}$
1267: for surface Crack for a length scale $r=49$\mum. (a)
1268: $D^{(1)}(h_r,r)$, (b) $D^{(2)}(h_r,r)$ and $D^{(4)}(h_r,r)$.}
1269: \label{fig:Wendt_D1234}
1270: \end{figure}
1271:
1272: For the surface Road~5 (cf.\ fig.~\ref{fig:data_road_no_scaling}) drift and
1273: diffusion coefficients could not be estimated. The reason can be seen in
1274: fig.~\ref{fig:Dk_road_no_scaling_ex}. The diagram shows the dependence of
1275: $M^{(1)}(h_r,r,\Delta_r)$ and $M^{(2)}(h_r,r,\Delta r)$ on $\Delta r$ for fixed
1276: $r$ and $h_r$, in this case 104\un{mm} and $-0.6\sigma_\infty$. For $\Delta
1277: r>l_M$ it can be seen that $M^{(1)}$ and $M^{(2)}$ behave like $1/\Delta r$.
1278: This behaviour can be explained by the presence of some additional uncorrelated
1279: noise, where additional means independent of the stochastic process. A similar
1280: behaviour was found for financial market data \cite{Renner2001b}. In this case
1281: the integral in eq.~(\ref{eq:Mk_2}) will tend to a constant for small $\Delta
1282: r$, independent of the value of $\Delta r$. Because we divide the integral by
1283: $\Delta r$, the $M^{(k)}$ will then diverge as $\Delta r$ approaches zero.
1284: Note that within the same mathematical framework the presence of uncorrelated
1285: noise can be quantitatively determined \cite{Siefert2003}.
1286:
1287: \begin{figure}[htbp]
1288: \centering
1289: \begin{picture}(0,0)\put(40,28){\makebox{\small (a)}}\end{picture}%
1290: \includegraphics[width=\breite]{figures/figure_20a}
1291: \begin{picture}(0,0)\put(40,28){\makebox{\small (b)}}\end{picture}%
1292: \includegraphics[width=\breite]{figures/figure_20b}
1293: \caption[Extrapolation procedure for $D^{(1)}$.]{%
1294: Uncorrelated noise in the case of surface Road~5. Dependence of (a)
1295: $M^{(1)}(h_r,r,\Delta r)$ and (b) $M^{(2)}(h_r,r,\Delta r)$ on $\Delta r$
1296: for $r=104\un{mm}$ and $h_r=-0.6\sigma_\infty$. The Markov length $l_M$ is
1297: marked with a dashed line. For illustration a function proportional to
1298: $1/\Delta r$ is fitted to the $M^{(k)}$.
1299: }
1300: \label{fig:Dk_road_no_scaling_ex}
1301: \end{figure}
1302:
1303:
1304: \subsection{Conclusions on the estimation of drift and diffusion coefficients}
1305:
1306: Estimations of the drift and diffusion coefficients $D^{(1)}(h_r,r)$ and
1307: $D^{(2)}(h_r,r)$ have been performed for all the surfaces introduced in section
1308: \ref{sec:measurement_data}. An exception is Road~5, where the stochastic
1309: process in the scale variable, while still Markovian, appears to be dominated
1310: by additional uncorrelated noise. From eqs.~(\ref{eq:Dk_def}) and
1311: (\ref{eq:Mk_def}) it can be seen that this leads to diverging Kramers-Moyal
1312: coefficients $D^{(k)}$, as is the case for Road~5.
1313:
1314: As mentioned above, the magnitude of the fourth Kra\-mers-Moyal coefficient
1315: $D^{(4)}$ is of particular importance. If $D^{(4)}$ can be taken as zero, the
1316: whole scale dependent complexity can be described by a Fokker-Planck equation.
1317: Otherwise, if $D^{(4)}$ is not zero, an infinite set of $D^{(k)}$ is necessary.
1318: In terms of a Langevin equation (\ref{eq:Langevin}), for $D^{(4)}\neq 0$ no
1319: Gaussian noise is present. This case is related to unsteady stochastic
1320: processes \cite{Honerkamp1990}. As we see from the topographies in
1321: figs.~\ref{fig:data_road_scaling} and \ref{fig:Wendt_data}, jumps are more
1322: likely to be present for the Road~4 and Crack surfaces than for the remaining
1323: ones. This impression is consistent with the result that here we find
1324: $D^{(4)}\neq 0$.
1325: %
1326: As a consequence, in these cases the Fokker-Planck equation with a drift and
1327: diffusion coefficient is not sufficient to describe the stochastic process in
1328: the scale variable, because the higher coefficients cannot be neglected. The
1329: reconstruction of conditional probabilities (cf.\ section
1330: \ref{sec:veri_coeff}) failed for these surfaces.
1331:
1332: The range of scales where the drift and diffusion coefficients could be
1333: estimated varies for the different surfaces, depending on the Markov length on
1334: one side and on the length of the measured profiles on the other side. In
1335: the case of Road~2 an additional upper limit for the Markov properties was
1336: caused by the influence of a strong periodicity of the pavement.
1337:
1338:
1339: \section{Verification of the estimated Fokker-Planck equations}
1340: \label{sec:veri_coeff}
1341:
1342: In the previous section methods to estimate the Kramers-Moyal coefficients were
1343: discussed. We found that this estimation is not trivial.
1344: To prove the quality of the estimated $D^{(k)}$
1345: we now want to verify the corresponding Fokker-Planck equations.
1346:
1347:
1348: \subsection{Parameterization of Drift and Diffusion Coefficients}
1349: \label{sec:Dk_parameterization}
1350:
1351: With the estimations of the drift and diffusion coefficient from section
1352: \ref{sec:Dk_estimation} for each surface a Fokker-Planck equation
1353: (\ref{eq:FPE1}) is defined which should describe the corresponding process. For
1354: the verification of these coefficients it is additionally desirable to generate
1355: parameterizations which define $D^{(1)}(h_r, r)$ and $D^{(2)}(h_r, r)$ not only
1356: at discrete values but at arbitrary points in the $(h_r, r)$-plane.
1357:
1358: Such parameterizations have already been shown in
1359: figs.~\ref{fig:D1_road_scaling}, \ref{fig:D2_road_scaling},
1360: \ref{fig:Schimmel_D12}, and \ref{fig:Wendt_D1234}, as lines together with the
1361: estimated discrete values. For $D^{(1)}$ it can be seen that for all surfaces
1362: a straight line with negative slope was used, with additional cubic terms for
1363: Road~2, Road~4, and Crack.
1364: %
1365: The diffusion coefficients were in all cases parameterized as parabolic
1366: functions. The special shape of the diffusion coefficient for Road~2 was
1367: parameterized as one inner and one outer parabola for small and larger values
1368: of $h_r$, respectively (compare with fig.~\ref{fig:D2_road_scaling}).
1369: %
1370: We would like to note that both the drift and diffusion coefficients of the
1371: cobblestone road presented in \cite{Waechter2003_plus_preprint} are best
1372: fitted by piecewise linear functions with steeper slopes for larger $h_r$.
1373:
1374: It is easy to verify that with a linear $D^{(1)}$ and a constant $D^{(2)}$ the
1375: Fokker-Planck equation~(\ref{eq:FPE1}) describes a Gaussian process, while
1376: with a parabolic $D^{(2)}$ the distributions become non-Gaussian, also called
1377: intermittent or heavy tailed.
1378: %
1379: For the Au surface it can be seen in fig.~\ref{fig:Schimmel_D12} that
1380: $D^{(2)}$ has only a weak quadratic dependence on $h_r$ and possibly could
1381: also be interpreted as constant (we nevertheless kept the small quadratic term
1382: because it is confirmed by the verification procedure below). If $D^{(2)}$ is
1383: constant in $h_r$ the type of noise in the corresponding Langevin equation
1384: (\ref{eq:Langevin}) is no longer multiplicative but additive, which results in
1385: Gaussian noise in the process. Thus the statistics of $h_r$ in $r$ will always
1386: stay Gaussian, and all moments $\langle h_r^n\rangle$ with $n>2$ can be
1387: expressed by the first and second one. As a further consequence, the scaling
1388: exponents $\xi_n$ (see section~\ref{sec:scaling}) are obtained by $\langle
1389: h_r^n\rangle\sim \langle h_r^2\rangle^{n/2}$ as $\xi^n = \frac{n}{2}\xi_2$.
1390: %
1391: This linear dependence on $n$ denotes self-affinity
1392: rather than multi-affinity and is confirmed by the scaling analysis in
1393: section~\ref{sec:scaling_with}.
1394:
1395:
1396: \subsection{Reconstruction of empirical pdf}
1397:
1398: Next, we want to actually evaluate the precision of our results. Therefore we
1399: return to eq.~(\ref{eq:FPE1}). Knowing $D^{(1)}$ and $D^{(2)}$ it should be
1400: possible to calculate the pdf of $h_r$ with the corresponding Fokker-Planck
1401: equation.
1402: %
1403: Equation~(\ref{eq:FPE1}) can be integrated over $h_0$ and is then valid also
1404: for the unconditional pdf:
1405: %
1406: \begin{eqnarray}
1407: \label{eq:FPE2}
1408: \lefteqn{-r\,\frac{\partial}{\partial r}\;p(h_r,r)\,=} \\
1409: && \left\{ -\frac{\partial}{\partial h_r} D^{(1)}(h_r,r)
1410: + \frac{\partial^2}{\partial h_r^2} D^{(2)}(h_r,r)
1411: \right\} \, p(h_r,r) \nonumber
1412: \end{eqnarray}
1413: %
1414: Now at the largest scale $r_0$ where the drift and diffusion coefficients could
1415: be worked out the empirical pdf is parameterized
1416: % (compare figs.~\ref{fig:D1_road_scaling}, \ref{fig:D2_road_scaling},
1417: % \ref{fig:Schimmel_D12}, \ref{fig:Wendt_D1234})
1418: and used as the initial condition for a numerical solution of
1419: eq.~(\ref{eq:FPE2}). For several values of $r$ the reconstructed pdf is
1420: compared to the respective empirical pdf, as shown below in this section. If
1421: our Fokker-Planck equation successfully reproduces these single scale pdf,
1422: the structure functions $\langle h_r^n\rangle$ can also easily be obtained.
1423:
1424: A second verification is the reconstruction of the conditional pdf by a
1425: numerical solution of Fokker-Planck equation (\ref{eq:FPE1}) for the
1426: conditional pdf. Reconstructing the conditional pdf this way is much more
1427: sensitive to deviations in $D^{(1)}$ and $D^{(2)}$. This becomes evident by
1428: the fact that the conditional pdf (and not the unconditional pdf of
1429: figs.~\ref{fig:road_scaling_veri1} and \ref{fig:Schimmel_veri1}) determine
1430: $D^{(1)}$ and $D^{(2)}$ and thus the stochastic process, see
1431: eqs.~(\ref{eq:Dk_def}) and (\ref{eq:Mk_def}).
1432: %
1433: The knowledge of the conditional pdf also gives access to the
1434: complete $n$-scale joint pdf (eq.~\ref{eq:markov_straight}).
1435: %
1436: Here again the difference from the
1437: multiscaling analysis becomes clear, which analyses higher moments $\langle
1438: h_r^n\rangle = \int h_r^n\cdot p(h_r)\,\mathrm{d} h_r$ of $h_r$, and does not
1439: depend on the conditional pdf. It is easy to show that there are many different
1440: stochastic processes which lead to the same single scale pdf $p(h_r)$.
1441:
1442: For both verification procedures we use a technique which is mentioned in
1443: \cite{Risken1984} and has already been used in
1444: \cite{Renner2001,Waechter2003_plus_preprint}. An approximative solution of the
1445: Fokker-Planck eq.~(\ref{eq:FPE1}) for infinitesimally small steps $\Delta r$
1446: over which $D^{(k)}$ can be taken as constant in $r$, is known
1447: \cite{Risken1984}
1448: \begin{eqnarray}
1449: \label{eq:FPE_approx_sol}
1450: \lefteqn{ p(h_1, r-\Delta r| h_0,r) =
1451: \frac{1}{2\sqrt{\pi D^{(2)}(h_r,r)\Delta r}} } \nonumber\\
1452: &&\times\exp{\left(-\frac{(h_1-h_0-D^{(1)}(h_0,r)\Delta r)^2}
1453: {4D^{(2)}(h_r,r)\Delta r}\right)}\,.
1454: \end{eqnarray}
1455: %
1456: A necessary condition for a Markov process is the validity of the
1457: Chapman-Kolmogorov equation (\ref{eq:CKE}) \cite{Risken1984},
1458: which allows to combine two conditional pdf with adjacent intervals in $r$
1459: into one conditional pdf spanning the sum of both intervals. By an iterative
1460: application of these two relations we are able to obtain conditional
1461: probabilities \mbox{$p(h_i, r_0-i\Delta r| h_0,r_0)$} spanning large intervals
1462: in the scale $r$, given that for all involved scales $r_i$ the drift and
1463: diffusion coefficients are known.
1464:
1465: In the following the results of this verification procedure are shown for
1466: those surfaces where the drift and diffusion coefficients of the Fokker-Planck
1467: equation could be obtained.
1468:
1469:
1470: \subsection{Verification results}
1471: \label{sec:veri_results}
1472:
1473: The results of the reconstruction of the unconditional pdf for the road
1474: surfaces with scaling properties are presented in
1475: fig.~\ref{fig:road_scaling_veri1}. The pdf of Road~1 show at smaller scales a
1476: peak around 5\un{\sigma_\infty} which is not reproduced by our Fokker-Planck
1477: equation because in this regime of $h_r$ $D^{(1)}$ and $D^{(2)}$ could not be
1478: estimated with sufficient precision. Here it has to be noted that according to
1479: eq.~(\ref{eq:sigma_inf}) $\sigma_\infty>\sigma_r$ for any $r$, and thus
1480: 5\un{\sigma_\infty} is a large value for a pdf, denoting quite rare events (the
1481: $r$-dependence of $\sigma_r$ has been presented by $S^2(r)=\sigma_r^2$, see
1482: section \ref{sec:scaling}). The magnitudes of the estimated drift and
1483: diffusion coefficients had to be adjusted by a factor of 0.65 to give optimal
1484: results in the reconstruction.
1485: %
1486: For Road~2 it is likely that the correspondence between the emprical and
1487: reconstructed pdf could be improved by a more advanced parameterization of the
1488: nontrivial shape especially of the estimated drift coefficient (see
1489: fig.~\ref{fig:D1_road_scaling}). Here, the estimated drift and diffusion
1490: coefficients could be used without adjustment.
1491: %
1492: The reconstructed pdf for Road~3 are in perfect agreement with the empirical
1493: ones. A substantial adjustment factor of 0.20 for $D^{(1)}$ and 0.26 for
1494: $D^{(2)}$ was necessary to achieve the best result.
1495:
1496: \setlength{\breite}{0.66\linewidth}
1497: \begin{figure}[htbp]
1498: \centering
1499: \begin{picture}(0,0)\put(8,31){\makebox{\small Road~1}}\end{picture}%
1500: \includegraphics[width=\breite]{figures/figure_21a}
1501: \begin{picture}(0,0)\put(8,31){\makebox{\small Road~2}}\end{picture}%
1502: \includegraphics[width=\breite]{figures/figure_21b}
1503: \begin{picture}(0,0)\put(8,31){\makebox{\small Road~3}}\end{picture}%
1504: \includegraphics[width=\breite]{figures/figure_21c}
1505: \caption{%
1506: Numerical solution of the Fokker-Planck equation (\ref{eq:FPE2}) compared
1507: to the empirical pdf (symbols) for road surfaces with scaling properties.
1508: For each surface, the topmost solid line corresponds to an empirical pdf
1509: parameterized at the largest scale, and the dashed lines to the
1510: reconstructed pdf. Scales are (from top to bottom)
1511: %
1512: for Road~1: 316, 158, 79, 66\un{mm}, %
1513: for Road~2: 158, 79, 47, 20\un{mm}, %
1514: for Road~3: 188, 95, 47, 24\un{mm}. %
1515: Pdf are shifted in the vertical direction for clarity of presentation. }
1516: \label{fig:road_scaling_veri1}
1517: \end{figure}
1518:
1519: Reconstructed conditional pdf are shown in fig.~\ref{fig:road_scaling_veri2}
1520: for the road surfaces. While there are deviations for larger values of
1521: $h_0,h_1$, the overall agreement between the empirical and reconstructed pdf is
1522: good. Especially the rather complicated shape of the conditional pdf of Road~2
1523: appears to be well modelled by our coefficients $D^{(1)},D^{(2)}$. As mentioned
1524: above, an improved parameterization of $D^{(1)}$ may lead to even better
1525: results. The magnitudes of $D^{(1)},D^{(2)}$ were adjusted by the same
1526: factors as for the unconditional pdf above.
1527:
1528: \setlength{\breite}{0.64\linewidth}
1529: \begin{figure}[htbp]
1530: \centering
1531: \begin{picture}(0,0)\put(0,42){\makebox{\small Road~1}}\end{picture}%
1532: \includegraphics[width=\breite]%
1533: {figures/figure_22a}
1534: \begin{picture}(0,0)\put(0,42){\makebox{\small Road~2}}\end{picture}%
1535: \includegraphics[width=\breite]%
1536: {figures/figure_22b}
1537: \begin{picture}(0,0)\put(0,42){\makebox{\small Road~3}}\end{picture}%
1538: \includegraphics[width=\breite]%
1539: {figures/figure_22c}
1540: \caption[Numerical solution of Fokker-Planck equation]%
1541: {Numerical solution of the Fokker-Planck equation (\ref{eq:FPE1}) compared to
1542: the empirical pdf for road surfaces with scaling properties. Similar to
1543: fig.~\ref{fig:Schimmel_markov} in each case a contour plot of empirical
1544: (solid lines) and reconstructed pdf (broken lines) is shown on top, with
1545: contour levels as in fig.~\ref{fig:Schimmel_markov}. Below two cuts at
1546: $h_0\approx\pm\sigma_\infty$ are located. Here, empirical pdf are plotted
1547: as symbols.
1548: Scales are %
1549: $r_0=304\un{mm}$, $r_1=158\un{mm}$ (Road~1), %
1550: $r_0=158\un{mm}$, $r_1=112\un{mm}$ (Road~2), %
1551: and $r_0=188\un{mm}$, $r_1=92\un{mm}$ (Road~3). %
1552: }
1553: \label{fig:road_scaling_veri2}
1554: \end{figure}
1555:
1556:
1557: \setlength{\breite}{0.64\linewidth}
1558: \begin{figure}[htbp]
1559: \centering
1560: \begin{picture}(0,0)\put(8,30){\makebox{\small Au (a)}}\end{picture}%
1561: \includegraphics[width=\breite]{figures/figure_23a}
1562: \begin{picture}(0,0)\put(0,42){\makebox{\small Au (b)}}\end{picture}%
1563: \includegraphics[width=\breite]%
1564: {figures/figure_23b}
1565: \begin{picture}(0,0)\put(0,42){\makebox{\small Au (c)}}\end{picture}%
1566: \includegraphics[width=\breite]%
1567: {figures/figure_23c}
1568: \caption[Au: Numerical solutions of Fokker-Planck equations]{%
1569: Numerical solutions of the Fokker-Planck equations (\ref{eq:FPE1}) and
1570: (\ref{eq:FPE2}) compared to the empirical pdf for the Au surface. \\
1571: %
1572: (a) Results of the integrated equation (\ref{eq:FPE2}) presented as in
1573: fig.~\ref{fig:road_scaling_veri1}. Scales $r$ are 281, 246, 148, and
1574: 56\un{nm} (from top to bottom). \\
1575: %
1576: (b), (c) Numerical solution of equation (\ref{eq:FPE1}) for the
1577: conditional pdf compared to the empirical pdf at scales $r_0=183\un{nm}$,
1578: $r_1=155\un{nm}$ (b) and $r_0=281\un{nm}$, $r_1=56\un{nm}$ (c).
1579: The organisation of the diagram is as in fig.~\ref{fig:road_scaling_veri2}.
1580: }
1581: \label{fig:Schimmel_veri1}
1582: \end{figure}
1583:
1584: In the case of the Au film the drift and diffusion coefficients could be worked
1585: out from 281 down to 56\un{nm}, see section \ref{sec:Dk_estimation}. In
1586: contrast to this regime, the range of correlation between scales is only about
1587: 40\un{nm}, i.e., height increments on scales which are separated by at least
1588: 40\un{nm} are uncorrelated.
1589: %
1590: Nevertheless, both verification procedures outlined in section
1591: \ref{sec:veri_coeff} gave good results over the whole range from 281 to
1592: 56\un{nm} as shown in fig.~\ref{fig:Schimmel_veri1}. Here the estimated
1593: $D^{(1)}$ and $D^{(2)}(h_r,r)$ were multiplied by factors 1.3 and 2.2,
1594: respectively.
1595:
1596:
1597: \subsection{Discussion of the verification procedure}
1598:
1599: For the verification of the drift and diffusion coefficients estimated in
1600: section~\ref{sec:Dk_estimation} numerical solutions of the Fokker-Planck
1601: equations (\ref{eq:FPE1}) and (\ref{eq:FPE2}) have been performed using these
1602: estimations. The reconstructed pdf have been compared to the empirical ones to
1603: validate the descripition of the data sets as realizations of stochastic
1604: processes obeying the corresponding Fokker-Planck equation.
1605:
1606: Good results were obtained for most surfaces where the drift and diffusion
1607: coefficients could be derived. In the case of Road~4 and Crack we found that
1608: the higher Kramers-Moyal coefficients $D^{(3)}$ and $D^{(4)}$ were
1609: significantly different from zero, and the empirical pdf could not be
1610: reproduced with a Fokker-Planck equation (which only uses $D^{(1)}$ and
1611: $D^{(2)}$).
1612:
1613: It may be surprising that the correspondence between the empirical and
1614: reconstructed pdf seems better for the conditional rather than for the
1615: unconditional pdf in some cases (compare figs.~\ref{fig:road_scaling_veri1} and
1616: \ref{fig:road_scaling_veri2}). One reason may be that in
1617: fig.~\ref{fig:road_scaling_veri2} it is clear that the empirical pdf are not
1618: precisely defined for combinations of large $h_0$ and $h_1$. The eye
1619: concentrates on the central regions of the contour plots where the uncertainty
1620: of the empirical pdf is reduced, as well as deviations due to possible
1621: inaccuracies and uncertainties of our drift and diffusion coefficients. This
1622: effect is also confirmed by our mathematical framework where all steps in the
1623: procedure are based on the estimation and evaluation of the conditional (not
1624: the unconditional) pdf.
1625:
1626: The reconstruction procedure allows also to adjust the estimated coefficients
1627: in order to improve the above-men\-tioned description, thus compensating for a
1628: number of uncertainties in the estimation process.
1629: %
1630: While the functional form of $D^{(1)}$ and $D^{(2)}$ found in
1631: section~\ref{sec:Dk_estimation} for all surfaces could be confirmed, in most
1632: cases the magnitudes of the estimated values had to be adjusted to give
1633: satisfactory results in this reconstruction procedure.
1634: %
1635: We found this effect also when analysing turbulent velocities and financial
1636: data.
1637: %
1638: One reason may be the uncertainties of the estimation procedure.
1639: %could be guessed.
1640: %
1641: A second source of deviations may be that the dependence of
1642: $M^{(k)}(h_r,r,\Delta r)$ on $\Delta r$ is not always purely linear in the
1643: extrapolation range (see section~\ref{sec:Dk_estimation}). Thus fitting a
1644: straight line and extrapolating against $\Delta r=0$ may lead to coefficients
1645: $D^{(1)}$ and $D^{(2)}$ which still have the correct functional form in $h_r$
1646: but incorrect magnitudes. As mentioned in section~\ref{sec:Dk_estimation}, in
1647: our case no general improvements could be achieved by the use of nonlinear
1648: (i.e.\ polynomial) fitting functions or higher order terms of the corresponding
1649: Taylor expansion. It is possible that the range of $\Delta r<l_M$ where no
1650: Markov properties are given is in most cases large enough that approximations
1651: for small $\Delta r$ are inaccurate. We would like to note that there are also
1652: data sets which did not require any adjustment of the estimated coefficients,
1653: see Road~2 and \cite{Waechter2003_plus_preprint}.
1654: %
1655: A last remark concerns the latest results in the case of Road~1, see
1656: fig.~\ref{fig:drh_ex}. If the fraction of $M^{(k)}(\Delta r)$ which is
1657: proportional to $1/\Delta r$ is substracted before performing the
1658: extrapolation, the resulting $D^{(k)}$ are substantially improved in
1659: their magnitudes. This may be a way to correct the extrapolation of the
1660: $D^{(k)}$ in cases where uncorrelated noise is involved.
1661:
1662: In any case, whether an adjustment of $D^{(1)}$ and $D^{(2)}$ was needed or
1663: not, for the presented surfaces a Fokker-Planck equation was found which
1664: reproduces the conditional pdf. Together with the verification of the Markov
1665: property (\ref{eq:markov_straight}) thus a complete description of the
1666: $n$-scale joint pdf is given, which was the aim of our work.
1667:
1668:
1669: \section{Conclusions}
1670: \label{sec:conclusions}
1671:
1672: For the analysis and characterization of surface roughness we have presented a
1673: new approach and applied it to different examples of rough surfaces. The
1674: objective of the method is the estimation of a Fokker-Planck equation
1675: (\ref{eq:FPE1}) which describes the statistics of the height increment $h_r(x)$
1676: in the scale variable $r$. A complete characterization of the corresponding
1677: stochastic process in the sense of multiscale conditional probabilities is the
1678: result.
1679:
1680: The application to different examples of surface measurement data showed that
1681: this approach cannot serve as a universal tool for any surface, as it is also
1682: the case for other methods like those based on self- and multi-affinity. With
1683: given conditions, namely the Markov property and a vanishing fourth order
1684: Kramers-Moyal coefficient (cf.\ section~\ref{sec:Markov_theory}),
1685: a comprehensive characterization of a single surface is obtained. The
1686: features of the scaling analysis are included, and beyond that a deeper
1687: insight in the complexity of roughness is achieved.
1688: %
1689: As shown in \cite{Jafari2003} such knowledge about a surface allows the
1690: numerical generation of surface structures which should have the same
1691: complexity. This may be of high interest for many research fields based on
1692: numerical modelling.
1693:
1694: The precise estimation of the magnitudes of the drift and diffusion
1695: coefficients for surface measurement data still remains an open problem. While
1696: for other applications a number of approaches have been developed
1697: \cite{Renner2001,Friedrich2002,Ragwitz2001,Sura2002} in any case a verification
1698: of the estimated Fokker-Planck equation is necessary and may lead to
1699: significant adjustments, as it is the case for some of our data sets.
1700:
1701:
1702: \begin{acknowledgement}
1703:
1704: We enjoyed helpful and stimulating discussions with R.\ Friedrich, A.\
1705: Kouzmitchev and M.\ Haase. Financial support by the Volks\-wa\-gen
1706: Foundation is kindly acknowledged.
1707:
1708: \end{acknowledgement}
1709:
1710:
1711: \bibliography{obf_overview,mw}
1712: \bibliographystyle{unsrt}
1713:
1714: \end{document}
1715:
1716: %%% Local Variables:
1717: %%% mode: latex
1718: %%% TeX-master: t
1719: %%% End:
1720: