betaFeedback

WorksPhysics and Astronomy

Role of correlations in the maximum distribution of multiscale stationary Markovian processes

Abstract

We are interested in investigating the statistical properties of extreme values for strongly correlated variables. The starting motivation is to understand how the strong-correlation properties of power-law distributed processes affect the possibility of exploring the whole domain of a stochastic process (the real axis in most cases) when performing time-average numerical simulations and how this relates to the numerical evaluation of the autocorrelation function. We show that correlations decrease the heterogeneity of the maximum values. Specifically, through numerical simulations we observe that for strongly correlated variables whose probability distribution function decays like a power-law 1/xα1/x^α, the maximum distribution has a tail compatible with a 1/xα+21/x^{α+2} decay, while for i.i.d. variables we expect a 1/xα1/x^α decay. As a consequence, we also show that the numerically estimated autocorrelation function converges to the theoretical prediction according to a factor that depends on the length of the simulated time-series nn according to a power-law: 1/nαδ1/n^{α^δ} with δ<1δ<1, This accounts for a very slow convergence rate.

Introduction

Extreme value theory [1] is a research field that investigates a crucial issue in many disciplines ranging from finance [2, 3] to natural and social phenomena [4] to disordered systems in physics [5]. In a nutshell, the problem is to investigate the statistics of the maximum (or the minimum) of a set of stochastic variables. In the case when such variables are independent and identically distributed (i.i.d.) many analytical results are known [6]. Indeed, as long as the variables are weakly correlated, the results obtained for the i.i.d. case can be easily generalized. When the variables are strongly correlated little is known.

We are interested in investigating the statistical properties of extreme values for strongly correlated variables. Moreover, we will mainly focus on stationary processes that are also power-law distributed. The starting motivation is to understand how the strong-correlation properties of power-law distributed processes affect the possibility of exploring the whole domain of a stochastic process (the real axis in most cases) when performing time-average numerical simulations and how this relates to the numerical evaluation of the autocorrelation function of a stochastic process. We have already shown [7] how the simulated autocorrelation function RL(τ)R_L(\tau) converges to the predicted autocorrelation function R(τ)R(\tau) at a rate that essentially depends by the pdf tail computed in LL, where LL is the maximum values that is attained when numerically simulating the stochastic process. As such, here we are interested in understanding how LL is related to the length nn of the simulated time-series.

Without pretending to be too general, we will here consider stationary Markovian stochastic processes that can be described by a nonlinear Langevin equation, with a white-noise term, and that admit a Fokker-Planck equation [8, 9], as in Refs. [10, 11, 12]. We will leave for a future work the generalization to other classes of stochastic processes.

We will show that correlations modify the average maximum and the maximum distribution expected for i.i.d. variables in the direction of decreasing the heterogeneity of the maximum distribution. Specifically, through numerical simulations we observe that for strongly correlated power-law distributed processes the maximum distribution has a tail compatible with a 1/xα+21/x^{\alpha+2} decay, while for i.i.d. variables we expect a 1/xα1/x^\alpha decay, where α\alpha is the exponent of the power-law decaying probability distribution function (pdf). As a consequence, we also show that the error that one makes when numerically evaluating the autocorrelation function RL(τ)R_L(\tau) depends on nn according to a power-law: 1/nδ1/n^\delta with δ<1\delta< 1, This accounts for a very slow convergence rate.

The paper is organized as follows: in section II we investigate the statistical properties of the average maximum for strongly correlated variables. In section III we investigate the maximum distribution for strongly correlated variables. We discuss our results and draw our conclusions in section IV.

Average maximum of correlated power-law distributed processes

The importance of power-law distributed processes has been thoroughly elucidated in Ref. [13], that, despite its publication year, can be still considered as the reference paper for anyone interested in accessing the power-law world.

In particular, in section III.C the author shows a nice argument for obtaining the average maximum value as a function of the length nn of the considered time-series:

⟨xmax⁡⟩≈n1/(α−1)(1)\langle x_{\max}\rangle\approx n^{1/(\alpha-1)} \tag*{(1)}

where α\alpha is the exponent of the probability distribution function (pdf) power-law tail: p(x)≈1/xαp(x) \approx1/x^\alpha.

Indeed, the procedure sketched in Ref. [13] section III.C is quite general and can be also applied to processes that have a pdf with tails different from a power-law. In fact in Figure 1 we show ⟨xmax⁡⟩\langle x_{\max}\rangle as a function of nn, for an Ornstein-Uhlembeck process [17] with unitary diffusion coefficient and drift coefficient h(x)=−γxh(x)=-\gamma x whose pdf is given by:

Average maximum for an Ornstein-Uhlembeck process

Figure 1. Average maximum ⟨xmax⁡⟩\langle x_{\max}\rangle for an Ornstein-Uhlembeck process with γ=0.1\gamma=0.1. The red circles show the results of numerical simulations while the black solid lines show the theoretical results expected according to the procedure of Ref. [13] section III.C. Each circle represents the median and standard deviation computed over M=106M=10^6 simulations.

pOU(x)=γ2πexp⁡(−γ2x2).(2)p_{\mathrm{OU}}(x)=\frac{\sqrt{\gamma}}{\sqrt{2\pi}}\exp\left(-\frac{\gamma}{2}x^2\right). \tag*{(2)}

The red circles show the results of numerical simulations while the black solid lines show the theoretical results expected according to the procedure of Ref. [13] section III.C. Each circle represents the median and standard deviation computed over M=106M=10^6 simulations. The required integrations have been performed only numerically. The agreement between numerical simulations and theoretical prediction is quite good.

Let us now consider a stationary Markovian stochastic process described by a nonlinear Langevin equation with unitary diffusion coefficient and drift coefficient [7, 10, 11, 12]:

h(x)={−2V0tan⁡(V0x)if ∣x∣≤L,(1−1+4V1)/xif ∣x∣>L.(3)h(x)= \begin{cases} -2\sqrt{V_0}\tan\left(\sqrt{V_0}x\right) & \text{if } |x|\le L,\\ \left(1-\sqrt{1+4V_1}\right)/x & \text{if } |x|>L. \end{cases} \tag*{(3)}
V1=Ltan⁡(V0L)(1+Ltan⁡(V0L))V_1=L\tan\left(\sqrt{V_0}L\right)\left(1+L\tan\left(\sqrt{V_0}L\right)\right)

This process admits a stationary pdf with power-law tails p(x)≈x−αp(x)\approx x^{-\alpha} and a power-law decaying autocorrelation function ρ(τ)≈τ−β\rho(\tau)\approx\tau^{-\beta} with β=(α−3)/2\beta=(\alpha-3)/2. When α∈]3,5[\alpha\in]3,5[ the process is long-range correlated. When α>5\alpha>5 the process is short-range correlated. When α≤3\alpha\le3 the variance of the process is not defined. Hereafter we will always consider α>3\alpha>3. In Figure 2 we show ⟨xmax⁡⟩\langle x_{\max}\rangle as a function of nn, for α=4\alpha=4 (top-left panel), α=5\alpha=5 (top-right panel) and α=6\alpha=6 (bottom-left). The red circles show the results of numerical simulations while the black solid lines show the theoretical results expected according to the procedure of Ref. [13] section III.C. Each circle represents the median and standard deviation computed over M=106M=10^6 simulations. Numerical simulations are quite different from predictions. In fact, the average maximum still grows like a power-law ⟨xmax⁡⟩≈nζ\langle x_{\max}\rangle\approx n^\zeta, but with an exponent ζ\zeta much smaller than the one predicted in Ref. [13], see Eq. (1). In the bottom-right panel we show the fitted exponents as a function of α\alpha. The disagreement with the theoretical prediction is quite evident.

Average maximum for correlated power-law distributed processes

Figure 2. Average maximum ⟨xmax⁡⟩\langle x_{\max}\rangle for the process of Eq. (3) with α=4\alpha=4 (top-left panel), α=5\alpha=5 (top-right panel) and α=6\alpha=6 (bottom-left panel). The red circles show the results of numerical simulations while the black solid lines show the theoretical results expected according to the procedure of Ref. [13] section III.C. Each circle represents the median and standard deviation computed over M=106M=10^{6} simulations. In the bottom-right panel we show the fitted exponents ζ\zeta as a function of α\alpha.

Indeed we believe that this result clearly indicates that processes correlations play a central role in determining the properties of the maximum distribution and, in particular, of ⟨xmax⁡⟩\langle x_{\max}\rangle.

Maximum distribution of correlated processes

Hereafter, we would like to show how correlations modify the properties of the maximum distribution for which exact results are known in the case of i.i.d. stochastic variables Ref. [6]. In fact, one main result is that for i.i.d. gaussian distributed stochastic variables the maximum distribution is the Gumbel-type distribution [14] while for i.i.d. power-law distributed variables we have a Frechet-type distribution [15, 16]. In particular, by following the lines of section 3 in Ref. [6] for i.i.d. processes one gets:

pgauss(x)=12πσ2exp⁡(−x22σ2)↦G(X)=n2πσ2exp⁡(−X22σ2)[1−12Erf⁡(X2σ2)]n(4)p_{\mathrm{gauss}}(x)=\frac{1}{\sqrt{2\pi\sigma^{2}}}\exp\left(-\frac{x^{2}}{2\sigma^{2}}\right)\mapsto G(X)=\frac{n}{\sqrt{2\pi\sigma^{2}}}\exp\left(-\frac{X^{2}}{2\sigma^{2}}\right)\left[1-\frac{1}{2}\operatorname{Erf}\left(\frac{X}{\sqrt{2\sigma^{2}}}\right)\right]^{n} \tag*{(4)}
ppow(x)≈A01xα↦F(X)≈A0(α−1)n1Xα(5)p_{\mathrm{pow}}(x)\approx A_{0}\frac{1}{x^{\alpha}}\mapsto F(X)\approx A_{0}(\alpha-1)n\frac{1}{X^{\alpha}} \tag*{(5)}

where, in the case of power-law distributed processes, A0A_{0} is an appropriate normalization constant and we consider the asymptotic behaviour only.

In the top-left panel of Figure 3 we show G(x)G(x) (solid lines) for different values of nn together with the empirical distributions obtained trough numerical simulations of an Ornstein-Uhlenbeck process with γ=0.1\gamma=0.1. We have already seen in the previous section that there is a good agreement between theoretical predictions and numerical simulations for ⟨xmax⁡⟩\langle x_{\max}\rangle. However, at the level of the maximum distribution one can notice that in the OU process xmax⁡x_{\max} values seem closer to ⟨xmax⁡⟩\langle x_{\max}\rangle than for the i.i.d. case. Somehow, the effect of correlations is that of introducing a sort of viscosity that bound the xmax⁡x_{\max} values to their averages. Such an effect is even more pronounced in the case of the power-law correlated power-law distributed process of Eq. (3). The two right panels of Figure 3 show, for different values of nn, the empirical maximum distributions obtained trough numerical simulations of the process of Eq. (3) with α=6\alpha=6 (top-right panel) and α=4\alpha=4 (bottom-right panel). We see that in this case the tails of the maximum distribution are compatible with a power-law decay 1/Xα+21/X^{\alpha+2}, while for i.i.d. processes we expect a decay F(X)≈1/XαF(X)\approx1/X^{\alpha}. All distributions have been computed over M=106M=10^{6} different realizations of the process, each realization having length nn.

Maximum distributions for four correlated processes

Figure 3. Maximum distribution for an Ornstein-Uhlembeck process with γ=0.1\gamma= 0.1 (top-left panel), the process of Eq. (3) with α=6\alpha= 6 (top-right panel), the process of Eq. (3) with α=4\alpha= 4 (bottom-right panel) and a power-law correlated process normally distributed (bottom-left panel). The solid lines in the two left panels show the G(x)G(x) function of Eq. (4) at different values of nn. The black solid lines in the right panels show a reference power-law function with exponent indicated in the legends. The empirical distributions are computed over M=106M = 10^{6} realizations. Different colours refer to simulations with different values of nn: Red for n=103n = 10^{3}, Blue for n=104n = 10^{4}, Green for n=105n = 10^{5}, Magenta for n=106n = 10^{6},

In order to better elucidate the role of correlation in shaping the maximum distribution, we have considered a coordinate transformation x↦yx \mapsto y that starting from the process of Eq. (3) leads to a gaussian process with null average and unitary variance in the yy domain, see section III of [10] for more details. We have shown in [10] and [11] that this is a power-law correlated gaussian process. In the bottom-left panel of Figure 3 we show G(x)G(x) (solid lines) for different values of nn together with the empirical distributions obtained trough numerical simulations of such process. Since the process is gaussian, we do not observe power-law tails in the maximum distribution. However, the simulated distributions are much narrower than G(x)G(x), thus confirming the viscosity effect mentioned above. Indeed, the effect is much more pronounced here than for the OU process, which again can be considered as a signature of the role of strong correlations in shaping the maximum distribution.

The simulations with n=103n = 10^{3}, n=104n = 10^{4} and n=105n = 10^{5} shown in the above figures have all been obtained with the same seed and the same starting point x0=0.1x_{0} = 0.1. At the end of each of the MM realizations, the final point of the process’ realization was used as the starting point of the realization of the next iteration. In the case of n=106n = 10^{6}, in order to reduce the computational time needed to perform such simulations, we parallelized the computation by splitting the MM realizations into 10 bunches each one composed of M′=105M^{\prime} = 10^{5} realizations of length n=106n = 10^{6}. In order to better illustrate the effect of this choice, in the four left panels of Figure 4, for the same processes as in Figure 3, we show the maximum distribution for n=105n = 10^{5} when we perform a unique simulation (green circles) and when we split the simulation in 10 bunches each one composed of M′=105M^{\prime} = 10^{5} realizations of length n=105n = 10^{5}. In the case when we perform a unique simulation we can attain much larger maximum values. This is again a signature of the fact that the more a process is correlated the smaller the possibility of attaining large values. In fact, in the right panel of Figure 4 we show the same quantities for an i.i.d. gaussian distributed process: when we perform a unique simulation of M=106M = 10^{6} realizations each one of length n=106n = 10^{6} maxima are distributed in the interval [13.31, 20.73] while when we split the simulation in M′=10M^{\prime} = 10 bunches of M=105M = 10^{5} realizations each one of length n=106n = 10^{6} maxima are distributed in the interval [13.50, 20.72]

Maximum distributions for correlated and i.i.d. processes

Figure 4. Left panels - Maximum distribution for an Ornstein-Uhlembeck process with γ=0.1\gamma= 0.1 (top-left panel), the process of Eq. (3) with α=6\alpha= 6 (top-right panel), the process of Eq. (3) with α=4\alpha= 4 (bottom-right panel) and a power-law correlated process normally distributed (bottom-left panel). The solid lines in the two left panels show the G(x)G(x) function of Eq. (4) for n=105n = 10^{5}. The black solid lines in the right panels show a reference power-law function with exponent indicated in the legends. Right panel - Maximum distribution for a gaussian i.i.d. process. The solid line shows G(x)G(x) of Eq. (4) for n=106n = 10^{6}. In all panels green circles refer to simulations over M=106M = 10^{6} realizations and violet circles refer to simulations done in 10 bunches each one composed of M′=105M' = 10^{5} realizations. The length of each simulation is n=105n = 10^{5} in the left panels and n=106n = 10^{6} in the right panel.

In summary, we might speculate that the effect of correlations is that of bounding the xmax⁡x_{\max} values to their average values, although probably the type of distribution is of the same as for i.i.d. variables.

Conclusions

The viscosity effect mentioned above has already been observed in Ref. [11] and Ref. [12] where the mean first passage time and the first passage time distribution where respectively investigated. In fact, we had shown that the more the process is correlated the larger the average time Tx(Λ)T_x(\Lambda) needed to reach a certain boundary Λ\Lambda, thus indicating that correlations enhance the time spent in a given position and therefore diminishes the likelihood that it reaches nearby positions by chance.

The results shown above, are also relevant for determining the error that one makes when numerically evaluating the autocorrelation function due to the fact that any simulated time series X={x1,x2,⋯xn}\mathcal{X} = \{x_1, x_2, \cdots x_n\} is bounded, i.e. ∣xi∣≤L|x_i| \le L, which in turn is a function of nn. In fact, in Ref. [7] we have shown that the simulated autocorrelation function RL(τ)R_L(\tau) converges to the predicted autocorrelation function R(τ)R(\tau) at a rate that essentially depends by the pdf tail computed in LL. Clearly ⟨xmax⁡⟩\langle x_{\max}\rangle is a very reasonable proxy for LL. In the case of the process of Eq. (3), RL(τ)≃L1−α/2R(τ)R_L(\tau) \simeq L^{1-\alpha/2} R(\tau) [7]. We have shown above that ⟨xmax⁡⟩≃nβ\langle x_{\max}\rangle\simeq n^\beta where β\beta depends on α\alpha according to the empirical law illustrated in the bottom-right panel of Figure 2. As such the simulated autocorrelation function converges to the predicted autocorrelation function R(τ)R(\tau) at a rate that is given by 1/nβ(α/2−1)1/n^{\beta(\alpha/2-1)}. By inspecting the bottom-right panel of Figure 2 one may argue that β\beta is such that the convergence rate is of the order of 1/nδ1/n^\delta with δ<1\delta< 1, i.e. a very slow convergence rate. This indicates that by increasing the length of the time series the advantage in simulating the autocorrelation function is small.

The investigation of the way the maximum distribution depends on nn gives us indication about the way the process domain is explored during the time evolution. This is particularly relevant when performing numerical simulations as the capability of attaining large values in often crucial in the estimation of many statistical quantities such as autocorrelation function of power-law tails that are relevant for out understanding of the process we are considering. However, it is worth mentioning that the topics we are addressing here are related but different from the issues related to ergodicity [5, 19, 20]. In fact, we are not giving here any prescription about the fact that the process be or not be ergodic. We are rather giving a few technical indications about the effectiveness of time-average numerical simulations, which is preliminary to the investigation of the ergodic properties of a certain process.

Despite the specific motivation that inspired this work, we believe that our results are general enough to be of relevance in many disparate research fields. On our side we leave for future work the generalization of these results to other classes of stochastic processes as well as the investigation of the role of correlation properties in aging phenomena in physical and non physical complex systems.

References

References

  1. [1]P. Embrechts, C. Kluppelberg and T. Mikosch, Modelling Extremal Events, (Springer–Verlag, Berlin, 1997).
  2. [2]H. Markowitz, Portfolio selection, Journal of Finance 7, 77 (1952).
  3. [3]B. O. Bradley, M. S. Taqqu, An extreme value theory approach to the allocation of multiple assets, International Journal of Theoretical and Applied Finance 7, 1031 (2004).DOI
  4. [4]S. Alberio, V. Jentsch, H. Kantz, Extreme events in nature and society, (Springer Berlin 2006)
  5. [5]J. P. Bouchaud, Weak ergodicity breaking and aging in disordered systems, J. Phys. I France 2, 1705 (1992).DOI
  6. [6]S. N. Majumdar, A. Pal, G. Schehr, Extreme value statistics of correlated random variables: A pedagogical review, Physics Reports 840, 1 (2020).
  7. [7]S. Miccichè, A numerical recipe for the computation of stationary stochastic processes’ autocorrelation function, Chaos Solitons & Fractals, 171, 113458 (2023).DOI
  8. [8]H. Risken, The Fokker-Planck Equation, (Springer, Berlin, 1989).
  9. [9]C.W. Gardiner, Handbook of Stochastic Methods, (Springer Verlag, Berlin, 1985).
  10. [10]S. Miccichè, Modeling long-range memory with stationary Markovian processes, Phys. Rev. E 79, 031116 (2009).pubmed.ncbi.nlm.nih.gov/19391911
  11. [11]S. Miccichè, Role of conditional probability in multiscale stationary Markovian processes, Phys. Rev. E 82, 011104 (2010).pubmed.ncbi.nlm.nih.gov/20866562
  12. [12]S. Miccichè, First passage time distribution of stationary Markovian processes, EPL 92, 50011 (2010).DOI
  13. [13]M. E. J. Newman, Power laws, Pareto distributions and Zipf’s law, Contemporary Physics 46, 323 (2005).DOI
  14. [14]E. J. Gumbel, Statistics of Extremes, Columbia University Press, New York, 1958).
  15. [15]M. Fréchet, Sur la loi de probabilité de l’écart maximum, Ann. Soc. Polon. Math. 6, 93 (1927).
  16. [16]R. A. Fisher, L.H.C. Tippett, Limiting forms of the frequency distribution of the largest or smallest member of a sample, Proceedings of the Cambridge Philosophical Society 24, 180 (1928).DOI
  17. [17]G. E. Uhlenbeck, L. S. Ornstein, On the Theory of the Brownian Motion, Phys. Rev. 36, 823 (1930).DOI
  18. [19]G. Gallavotti, Ergodicity, ensembles, irreversibility in Boltzmann and beyond, Journal of Statistical Physics 78, 1571 (1995); P. Walters, An Introduction to Ergodic Theory. (Springer, Berlin, 1982).DOI
  19. [20]F. Flandoli, http://users.dma.unipi.it/~flandoli/AUTCap2.pdf, (in italian).

Paper details

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

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents