Abstract
Langevin models are widely used to model various stochastic processes in different fields of natural and social sciences. They are adapted to measured data by estimation techniques such as maximum likelihood estimation, Markov chain Monte Carlo methods, or the non-parametric direct estimation method introduced by Friedrich et al. (Phys Lett A 271(3):217, 2000). The latter has the distinction of being very effective in the context of large data sets. Due to their \(\delta \)-correlated noise, standard Langevin models are limited to Markovian dynamics. A non-Markovian Langevin model can be formulated by introducing a hidden component that realizes correlated noise. For the estimation of such a partially observed diffusion a different version of the direct estimation method was introduced by Lehle et al. (Phys Rev E 97(1):012113, 2018). However, this procedure requests that the correlation length of the noise component is small compared to that of the measured component. In this work, we propose a direct estimation method without this restriction. This allows one to effectively deal with large data sets from a wide range of examples. We discuss the abilities of the proposed procedure using several synthetic examples.
Graphic Abstract
Similar content being viewed by others
Data Availability Statement
This manuscript has no associated data or the data will not be deposited. [Authors’ comment: The data that support the findings of this study are available from the corresponding author upon reasonable request.]
Notes
In the case that \(B^k\) contains a fixed point, the mean \(\langle Y_t | X_t\in B^k \rangle \) is approximately zero (thanks to an anonymous referee for this suggestion) and consequently \(\langle X_{t+\tau } - X_t | X_t\in B^k \rangle \approx D^{(1)}_x(x^k) = 0\). Thus, all bins \(B^k\) where \(\langle X_{t+\tau } - X_t | X_t\in B^k \rangle \approx 0\) holds are candidates for fixed points. We suppose that this is even an equivalence relation, yet we cannot justify this rigorously. This additional information can be useful for determining the initial guess during optimization (cf. Sect. 3). A single fixed point provides an indication for the parameter \(X^{*}\) of the linear model. If multiple fixed points are to be expected, their location could be a basis for an initial guess of higher order than linear.
References
R. Friedrich, J. Peinke, M. Sahimi, M.R.R. Tabar, Phys. Rep. 506(5), 87 (2011)
H. Risken, T. Frank, The Fokker–Planck Equation, 2nd edn. (Springer, Berlin, 1996)
P.E. Kloeden, E. Platen, Numerical Solution of Stochastic Differential Equations (Springer, Berlin, 1992)
H. Bauer, R.B. Burckel, Probability Theory (De Gruyter, Berlin, 1996)
D. Kleinhans, R. Friedrich, Phys. Lett. A 368(3), 194 (2007)
W.V.D. Linden, V. Dose, U.V. Toussaint, Bayesian Probability Theory: Applications in the Physical Sciences (Cambridge University Press, Cambridge, 2014)
D.S. Sivia, J. Skilling, Data Analysis: A Bayesian Tutorial (Oxford University Press, New York, 2006)
R. Friedrich, S. Siegert, J. Peinke, S. Lück, M. Siefert, M. Lindemann, J. Raethjen, G. Deuschl, G. Pfister, Phys. Lett. A 271(3), 217 (2000)
S. Siegert, R. Friedrich, J. Peinke, Phys. Lett. A 243(5), 275 (1998)
M. Ragwitz, H. Kantz, Phys. Rev. Lett. 87(25), 254501 (2001)
R. Friedrich, C. Renner, M. Siefert, J. Peinke, Phys. Rev. Lett. 89(14), 149401 (2002)
M. Ragwitz, H. Kantz, Phys. Rev. Lett. 89(14), 149402 (2002)
I. Horenko, C. Hartmann, C. Schütte, F. Noe, Phys. Rev. E 76(1), 016706 (2007)
K. Schmietendorf, J. Peinke, O. Kamps, Eur. Phys. J. B 90(11), 222 (2017)
B. Lehle, J. Peinke, Phys. Rev. E 97(1), 012113 (2018)
T. Pesch, S. Schröders, H.J. Allelein, J.F. Hake, New J. Phys. 17(5), 055001 (2015)
B. Schäfer, C. Beck, K. Aihara, D. Witthaut, M. Timme, Nat. Energy 3(2), 119 (2018)
A.P. Dempster, N.M. Laird, D.B. Rubin, J.R. Stat, Soc. Ser. B 39(1), 1 (1977)
A. Dembo, O. Zeitouni, Stoch. Process. Appl. 23(1), 91 (1986)
F. Campillo, F. Le Gland, Stoch. Process. Appl. 33(2), 245 (1989)
M.A. Tanner, W.H. Wong, J. Am. Stat. Assoc. 82(398), 528 (1987)
L. Tierney, Ann. Stat. 22(4), 1701 (1994)
A. Golightly, D.J. Wilkinson, Biometrics 61(3), 781 (2005)
A. Golightly, D. Wilkinson, Comput. Stat. Data Anal. 52(3), 1674 (2008)
N. Reinke, A. Fuchs, D. Nickelsen, J. Peinke, J. Fluid Mech. 848, 117 (2018)
R.C. Aster, B. Borchers, C.H. Thurber, Parameter Estimation and Inverse Problems, 3rd edn. (Elsevier, Amsterdam, 2019)
P. Milan, M. Wächter, J. Peinke, Phys. Rev. Lett. 110(13), 138701 (2013)
C. Willers, O. Kamps, in preparation
S. Lück, C. Renner, J. Peinke, R. Friedrich, Phys. Lett. A 359(5), 335 (2006)
C. Renner, J. Peinke, R. Friedrich, J. Fluid Mech. 433, 383 (2001)
A.P. Nawroth, J. Peinke, D. Kleinhans, R. Friedrich, Phys. Rev. E 76(5), 056102 (2007)
L.R. Gorjão, M. Anvari, H. Kantz, C. Beck, D. Witthaut, M. Timme, B. Schäfer, IEEE Access 8, 43082 (2020)
V. Klippenstein, M. Tripathy, G. Jung, F. Schmid, N.F.A. van der Vegt, J. Phys. Chem. B 125(19), 4931 (2021)
J. Wouters, V. Lucarini, J. Stat. Phys. 151(5), 850 (2013)
F. Ferretti, V. Chardès, T. Mora, A.M. Walczak, I. Giardina, Phys. Rev. X 10(3), 031018 (2020)
M. Takahashi, Financ. Eng. Jpn. Mark. 3(2), 87 (1996)
C. Willers, O. Kamps, Exemplary python implementation of the estimation procedures proposed in “Non-parametric estimation of a Langevin model driven by correlated noise” (2021). https://doi.org/10.5281/zenodo.4986257
Author information
Authors and Affiliations
Contributions
CW developed the estimation procedure for the non-Markovian Langevin model and worked out the validation examples. OK supervised the work. CW wrote the manuscript in consultation with OK.
Corresponding author
Appendices
Appendix A: Analytic solution of the linear model
Here, we calculate the stationary solution of the SDE model defined by the equations (\(\alpha ,\beta ,\theta >0\))
in an analytic way. In general, the solution of the initial value problem defined by the Langevin equation
and the initial value \(X_0 = x_0\) is given by [2, 3]
In our model, we have \(X_t = (X^{(1)}_t, X^{(2)}_t)^\top \),
As we are interested in the stationary solution that emerges for large values of t, we end up with \(X_t^{\mathrm{st}} = \int _0^t e^{-B(t-t')}\sigma \, dW_{t'}\) and need to evaluate the matrix exponential. Here, we have to distinguish two cases: \(\alpha \ne \theta \) and \(\alpha =\theta \).
In the case \(\alpha \ne \theta \), we evaluate the matrix exponential via the diagonalization of B,
where \(T=\left( \begin{array}{cc} 1 &{} \sqrt{\beta }\frac{\alpha \theta }{\alpha -\theta }\\ 0 &{} 1 \end{array} \right) \) and \(D=\left( \begin{array}{cc} \frac{1}{\alpha } &{} 0\\ 0 &{} \frac{1}{\theta } \end{array} \right) \):
We obtain
for the first component of the solution and
for the second component.
In the case \(\alpha =\theta \), we calculate the matrix exponential via the Jordan–Chevalley decomposition of B,
where \(A=\left( \begin{array}{cc} \frac{1}{\alpha } &{} 0\\ 0 &{} \frac{1}{\alpha } \end{array}\right) \) is diagonal, \(N=\left( \begin{array}{cc} 0 &{} -\sqrt{\beta }\\ 0 &{} 0 \end{array} \right) \) is nilpotent, and \(AN=NA\):
We obtain
for the first component of the solution. The second component is the same as in the latter case.
Appendix B: Variance and ACF of the linear model
At first, we calculate the autocovariance function \(\text {Cov}(X_t,X_s)\) for the first component of the stationary solution of the linear model in the case \(\alpha \ne \theta \) (cf. Appendix A; for the sake of readability, we write \(X_t\) instead of \(X_t^{(1)}\) and \(W_t\) instead of \(W_t^{(1)}\)):
Due to \(\langle X_t \rangle =0\), we have
where we define
Representative for the four summands, we calculate \(\langle I_{0,t}^{t,\alpha }I_{0,s}^{s,\theta }\rangle \). Hereby we use the relation \(I_{a,b}^{c,d} = I_{a,b'}^{c,d} + I_{b',b}^{c,d}\) for \(a<b'<b\). Further, \(I_{a,b}^{c,d}\) and \(I_{a',b'}^{c',d'}\) are stochastically independent, if \(a<b\le a'<b'\), and \(\langle I_{a,b}^{c,d} \rangle = 0\). Without any loss of generality, we assume \(t>s\). All in all, we obtain
Now, we replace the Itô integrals by approximating sums. With a partition \(0=t_0<t_1<...<t_N=s\), we have
\((W_{t_{i+1}}-W_{t_i})\) and \((W_{t_{j+1}}-W_{t_j})\) are stochastically independent, if \(i\ne j\). Further, \(\langle (W_{t_{i+1}}-W_{t_i}) \rangle = 0\). Consequently, when expanding the sums, the mixed terms vanish and we end up with
The latter expression is an approximation of the Riemann integral
which, in the stationary case \(t\rightarrow \infty \) and \(s\rightarrow \infty \), yields \(\frac{\alpha \theta }{\alpha +\theta } e^{-\frac{1}{\alpha }(t-s)}\). All in all,
Consequently, for the covariance function we obtain
From this, we can infer that
Further, the autocorrelation function \(\text {ACF}(\tau )\), which is the normed autocovariance function (\(\text {ACF}(\tau ) =\text {Cov}(X_{\tau },X_0)/\text {Var}(X_0)\)), reads
In the case \(\alpha =\theta \), we have (again we write \(X_t\) instead of \(X_t^{(1)}\) and \(W_t\) instead of \(W_t^{(1)}\))
where
With similar calculations as in the latter case, we obtain
for the stationary case. From this, we can infer the relations
Appendix C: Taylor length \(\varLambda \) and correlation length L of the linear model
As introduced in Sect. 3, we obtain the Taylor length by a second-order series expansion of the autocorrelation function \(\text {ACF}(\tau )\) (cf. Appendix B) at \(\tau =0\) and the correlation length L by its integral. First, we regard the case \(\alpha \ne \theta \). From
we obtain
The correlation length L can be calculated as follows:
Similarly, we obtain \(\varLambda = \alpha \) and \(L=2\alpha \) in the case \(\alpha =\theta \).
Rights and permissions
About this article
Cite this article
Willers, C., Kamps, O. Non-parametric estimation of a Langevin model driven by correlated noise. Eur. Phys. J. B 94, 149 (2021). https://doi.org/10.1140/epjb/s10051-021-00149-0
Received:
Accepted:
Published:
DOI: https://doi.org/10.1140/epjb/s10051-021-00149-0