Abstract
We investigate the tail asymptotic behavior of the sojourn time for a large class of centered Gaussian processes X, in both continuous- and discrete-time framework. All results obtained here are new for the discrete-time case. In the continuous-time case, we complement the investigations of Berman (Commun Pure Appl Math 38(5):519–528, 1985a and Probab Theory Relat Fields 20(1):113–124, 1987) for non-stationary X. A by-product of our investigation is a new representation of Pickands constant which is important for Monte-Carlo simulations and yields a sharp lower bound for Pickands constant.
Similar content being viewed by others
References
Adler RJ (1990) An introduction to continuity, extrema, and related topics for general Gaussian processes, vol. 12 of Lecture Notes-Monograph Series. Institute of Mathematical Statistics, Hayward, California
Adler RJ, Taylor JE (2007) Random fields and geometry. Springer Monographs in Mathematics. Springer, New York
Aldous D (1989) Probability approximations via the Poisson clumping heuristic, vol. 77 of Applied Mathematical Sciences. Springer-Verlag, New York
Arendarczyk M (2017) On the asymptotics of supremum distribution for some iterated processes. Extremes 20(2):451–474
Bai L, Dȩbicki K., Hashorva E, Luo L (2018) On generalised Piterbarg constants. Methodol Comput Appl Probab 20(1):137–164
Berman SM (1973) Excursions of stationary Gaussian processes above high moving barriers. Annals Probab 1(3):365–387
Berman SM (1982) Sojourns and extremes of stationary processes. Annals Probab 10(1):1–46
Berman SM (1985a) Sojourns above a high level for a Gaussian process with a point of maximum variance. Commun Pure Appl Math 38(5):519–528
Berman SM (1985b) The maximum of a Gaussian process with nonconstant variance. Ann Inst H Poincaré Probab Statist 21(4):383–391
Berman SM (1987) Extreme sojourns of a Gaussian process with a point of maximum variance. Probab Theory Relat Fields 74(1):113–124
Berman SM (1992) Sojourns and extremes of stochastic processes. The Wadsworth & Brooks/Cole Statistics/Probability Series, Pacific Grove CA: Wadsworth & Brooks/Cole Advanced Books & Software
Berman SM, Kôno N (1989) The maximum of a Gaussian process with nonconstant variance: a sharp bound for the distribution tail. Ann Probab 17(2):632–650
Bingham NH, Goldie CM, Teugels JL (1989) Regular variation, vol. 27 of Encycolpedia of Mathematics and its Applications. Cambridge University Press, Cambridge
Chan HP, Lai TL (2006) Maxima of asymptotically Gaussian random fields and moderate deviation approximations to boundary crossing probabilities of sums of random variables with multidimensional indices. Ann Probab 34(1):80–121
Cheng D (2017) Excursion probabilities of isotropic and locally isotropic Gaussian random fields on manifolds. Extremes 20(2):475–487
Dȩbicki K, Kisowski P (2008) A note on upper estimates for Pickands constants. Statist Probab Lett 78(14):2046–2051
Dȩbicki K, Hashorva E (2017) On extremal index of max-stable stationary processes. Probab Math Stat 37(2):299–317
Dȩbicki K, Michna Z, Rolski T (2003) Simulation of the asymptotic constant in some fluid models. Stoch Models 19(3):407–423
Dȩbicki K, Hashorva E, Ji L (2015) Gaussian risk models with financial constraints. Scand Actuar J 2015(6):469–481
Dȩbicki K, Hashorva E, Ji L (2016) Parisian ruin over a finite-time horizon. Sci China Math 59(3):557–572
Dȩbicki K, Engelke S, Hashorva E (2017a) Generalized Pickands constants and stationary max-stable processes. Extremes 20(3):493–517
Dȩbicki K, Hashorva E, Liu P (2017b) Extremes of Gaussian random fields with regularly varying dependence structure. Extremes 20(2):333–392
Dȩbicki K, Hashorva E, Liu P (2017c) Uniform tail approximation of homogenous functionals of Gaussian fields. Adv Appl Probab 49(4):1037–1066
Dieker AB, Yakir B (2014) On asymptotic constants in the theory of extremes for Gaussian processes. Bernoulli 20(3):1600–1619
Dieker AB, Mikosch T (2015) Exact simulation of Brown-Resnick random fields at a finite number of locations. Extremes 18:301–314
Geluk JL, de Haan L (1987) Regular variation, extensions and Tauberian theorems, vol. 40 of CWI Tract. Stichting Mathematisch Centrum. Centrum voor Wiskunde en Informatica, Amsterdam
Gnedenko BV, Korolyuk VS (1950) Some remarks on the theory of domains of attraction of stable distributions. Dopovidi Akad Nauk Ukrain RSR 4:275–278
Harper AJ (2017) Pickands’ constant H α does not equal 1/Γ(1/α), for small α. Bernoulli 23(1):582–602
Hashorva E (2018) Representations of max-stable processes via exponential tilting, Stoch Process Appl, in press. https://doi.org/10.1016/j.spa.2017.10.003
Hashorva E, Hüsler J (2000) Extremes of Gaussian processes with maximal variance near the boundary points. Methodol Comput Appl Probab 2(3):255–269
Hüsler J (1990) Extreme values and high boundary crossings of locally stationary Gaussian processes. Ann Probab 18(3):1141–1158
Hüsler J (1995) A note on extreme values of locally stationary Gaussian processes. J Stat Planning Inference 45(1-2):203–213
Hüsler J, Piterbarg VI (1999) Extremes of a certain class of Gaussian processes. Stoch Process Appl 83(2):257–271
Hüsler J, Piterbarg VI (1999) Extremes of a certain class of Gaussian processes. Stoch Process Appl 83(2):257–271
Hüsler J, Piterbarg VI (2017) On shape of high massive excursions of trajectories of Gaussian homogeneous fields. Extremes 20(3):691–711
Michna Z (2017) Remarks on Pickands’ theorem. Probab Math Stat 37(2):373–393
Pickands III J (1969) Asymptotic properties of the maximum in a stationary Gaussian process. Trans Am Math Soc 145:75–86
Piterbarg VI (1972) On the paper by J. Pickands Upcrosssing probabilities for stationary Gaussian processes. Vestnik Moscow Univ Ser I Mat Mekh 27, 25-30. English Transl Moscow Univ Math Bull 27:25–30
Piterbarg VI (1996) Asymptotic methods in the theory of Gaussian processes and fields, vol. 148 of Translations of Mathematical Monographs. American Mathematical Society, Providence
Piterbarg VI (2004) Discrete and continuous time extremes of gaussian processes. Extremes 7(2):161–177
Piterbarg VI (2015) Twenty lectures about Gaussian processes. Atlantic Financial Press, London
Piterbarg VI (2016) High extrema of Gaussian chaos processes. Extremes 19 (2):253–272
Shao Q (1996) Bounds and estimators of a basic constant in extreme value theory of Gaussian processes. Stat Sin 6:245–257
Acknowledgments
The authors would like to thank Enkelejd Hashorva for his numerous remarks on all the steps of preparation of the manuscript and for ideas that led to the proof of Theorem 1.1. K.D. was partially supported by NCN Grant No 2015/17/B/ST1/01102 (2016-2019). X.P. thanks the Fundamental Research Funds for the Central Universities (ZYGX2015J102) and National Natural Science Foundation of China (71501025,11701070) for partial financial support. Financial support from the Swiss National Science Foundation Grant 200021-175752/1 is also kindly acknowledged.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
Let Ku be an index function of u, D be a compact set in \(\mathbb {R}^{n}\) and suppose without loss of generality that 0 ∈D. Further, let {ξu,k(t),t ∈D,k ∈ Ku} be a family of centered Gaussian random fields with a.s. continuous sample paths and variance function \(\sigma ^{2}_{\xi _{u,k}}\). For t such that \(\sigma ^{2}_{\xi _{u,k}}(\textit {\textbf {t}})>0\) define the standardised process
-
C0: {gk(u),k ∈ Ku} is a sequence of deterministic functions of u satisfying
$$\begin{array}{@{}rcl@{}} \lim\limits_{u\to\infty}\inf\limits_{k\in K_u}g_{k}(u)=\infty. \end{array} $$ -
C1: \(\sigma _{\xi _{u,k}}(\textbf {0})= 1\) for all large u and any k ∈ Ku, and there exists some bounded continuous function h on D such that
$$\begin{array}{@{}rcl@{}} \lim\limits_{u\to\infty}\sup\limits_{\textit{\textbf{t}}\in \textit{\textbf{D}},k\in K_u}|g_{k}^2(u)\left( 1- \mathbb{E}\{ \xi_{u,k}(\textit{\textbf{t}})\xi_{u,k}(\textbf{0})\} \right) - h(\textit{\textbf{t}})| = 0. \end{array} $$ -
C2: There exists a centered Gaussian random field \(\zeta (\textit {\textbf {t}}),\textit {\textbf {t}}\in \mathbb {R}^{n}\) with a.s. continuous trajectories such that for any s,t ∈D
$$\begin{array}{@{}rcl@{}} \lim\limits_{u\to\infty}\sup\limits_{k\in K_u}|g_{k}^2(u)\left( Var(\tilde\xi_{u,k}(\textit{\textbf{t}})-\tilde\xi_{u,k}(\textit{\textbf{s}}))\right) - 2Var(\zeta(\textit{\textbf{t}})-\zeta(\textit{\textbf{s}}))| = 0. \end{array} $$ -
C3: There exist positive constants C,ν,u0 such that
$$\begin{array}{@{}rcl@{}} \sup\limits_{k\in K_u} g_{k}^2(u)\mathbb{E}\{(\tilde\xi_{u,k}(\textit{\textbf{t}})-\tilde\xi_{u,k}(\textit{\textbf{s}}))^2\} \leq C ||\textit{\textbf{s}}-\textit{\textbf{t}}||^{\nu} \end{array} $$holds for all s,t ∈D,u ≥ u0, where \(||\textit {\textbf {t}} ||^{v}={\sum }_{i = 1}^{n}|t_{i}|^{v}\).
We present below an extension of Theorem 2.1 in Dȩbicki et al. (2017c). Recalling that μη(dt)/ηn denotes the counting measure on \(\eta \mathbb {Z}^n, \eta >0\) and μ0 is the Lebesgue measure on \(\mathbb {R}^{n}\).
Theorem 5.1
Leth,gk(u),ξu,k(t),t ∈D,k ∈ Kuandζbe such thatC0-C3hold. Then, forη ≥ 0
holds at x = 0 and all x ∈ (0,μη(D)) continuity points of \(\mathcal {B}^{h,\eta }_{\zeta }(\textit {\textbf {D}},\cdot )\), where
Proof of Theorem
5.1 Before beginning the proof, for notational simplicity we denote by Ru,k and ρu,k the covariance and the correlation function of ξu,k. Further set
and
Conditioning on ξu,k(0) and using that ξu,k(0) and ξu,k(t) − Ru,k(t,0)ξu,k(0) are mutually independent for all large u, we obtain
Let
Consequently, in order to show the claim it suffices to prove that
at x = 0 and all x ∈ (0,μη(D)) continuity points of \(\mathcal {B}^{h,\eta }_{\zeta }(\textit {\textbf {D}},\cdot )\).
For any s,t ∈D,
which together with C2-C3 yields
Hence, the finite-dimensional distributions of χu,k converge to that of \(\sqrt {2}\zeta (\textit {\textbf {t}}),\textit {\textbf {t}}\in \textit {\textbf {D}}\) uniformly with respect to k ∈ Ku. Let C(D) denote the Banach space of all continuous functions on D equipped with sup-norm. By Eq. 5.3 and C3, we know that
holds for large enough u and all s,t ∈D, thus the measures on C(D) induced by {χu,k(t),t ∈D,k ∈ Ku} are uniformly tight for large u. Therefore, {χu,k(t),t ∈D} converge weakly to \(\{\sqrt {2}\zeta (\textit {\textbf {t}}),\textit {\textbf {t}}\in \textit {\textbf {D}}\}\) as u →∞ uniformly with respect to k ∈ Ku.
Further, for large u
which together with C1-C2 implies
By C0-C1, for each \(z\in \mathbb {R}\)
Thus, for each \(z\in \mathbb {R}\), the probability measures on C(D) induced by \(\{{\sigma _{\xi _{u,k}}(\textit {\textbf {t}})}\chi _{u,k}(\textit {\textbf {t}}) + f_{u,k}(\textit {\textbf {t}},z),\textit {\textbf {t}}\in \textit {\textbf {D}}\}\) converge weakly, as u →∞, to that induced by \(\{\sqrt {2}\zeta (\textit {\textbf {t}})-h(\textit {\textbf {t}}) + z,\textit {\textbf {t}}\in \textit {\textbf {D}}\}\) uniformly with respect to k ∈ Ku. Therefore, for any \(\eta >0,z\in \mathbb {Z}\)
holds at all x ∈ (0,μη(D)) continuity points (depending on z) of \(\mathcal {I}(\cdot ,z)\) defined by
For η = 0, by Berman (1973)[Lemma 4.2] the set of discontinuity points under the uniform convergence on D of
is of measure 0 under the probability measure induced by \(\{ {\sqrt {2}\zeta (\textit {\textbf {t}})-h(\textit {\textbf {t}})} + z,\textit {\textbf {t}}\in \textit {\textbf {D}}\}\). Consequently, by the continuous mapping theorem we also have Eq. 5.7 for η = 0.
Next, we borrow an argument from Berman (1992) [Theorem 1.3.1] to verify Eq. 5.2 for all positive continuity points of \(\mathcal {B}^{h,\eta }_{\zeta }(\textit {\textbf {D}},\cdot )\). Let x0 ∈ (0,μη(D)) satisfy
Since for large M and all x ≥ 0 by Borell-TIS inequality
it follows from the dominated convergence theorem that
and thus by the monotonicity of \(\mathcal {I}(x,z)\) in x for each fixed z, x0 is a continuity point of \(\mathcal {I}(\cdot ,z)\) for almost all \(z\in \mathbb {R}\). Hence by Eq. 5.7 for almost all \(z\in \mathbb {R}\)
Further, by Eqs. 5.5 and 5.6 and Borell-TIS inequality, for sufficiently large u and M
where Eq. 5.10 follows from Sudakov-Fernique inequality (see, e.g., Adler (1990)[Theorem 2.9]), with conditions required therein being ensured by Eq. 5.4 and \(B_{\nu }^{(i)}\)’s being independent fBm’s with Hurst index ν/2.
Consequently, by Eqs. 5.8, 5.9 and 5.11, a direct application of the dominated convergence theorem yields that
verifying Eq. 5.1 for any x0 ∈ (0,μη(D)) a continuity point of \(\mathcal {B}^{h,\eta }_{\zeta }(\textit {\textbf {D}},\cdot )\).
The case x = 0,η = 0 is shown in Dȩbicki et al. (2017c). Since the case x = 0,η > 0 can be established by arguments similar to the presented above, we omit the details. This completes the proof. □
Let for any η ≥ 0,b > 0,S > 0, x ∈ [0,μη([−S,S]))
and
Lemma 5.1
-
i)
Let X be as in Theorem 2.1 andv(u) be as in Eq. ??. For anyη ≥ 0, S > ηandtusuch that\(\lim _{u \to \infty } t_{u}=t_{0}\in [0,T]\),we have
$$\begin{array}{@{}rcl@{}} \lim_{u\to\infty} {\Psi}^{-1}(u) \mathbb{P} \left\{ v(u){\int}_0^{S/v(u)} \left( X(t_u+s)\right) \mu_{\eta_u}(\mathrm{d} s)>x \right\} = \mathcal{B}_{\alpha,H(t_0)}^{\eta} (S,x) \end{array} $$atx = 0 and anyx ∈ (0,μη([0,S])) a continuitypoint of\(\mathcal {B}_{\alpha ,H(t_{0})}^{\eta } (S,\cdot )\).
-
ii)
Let X be as in Theorem 3.1,v(u),\(\mathcal {P}_{\alpha }^{b,\eta }(\cdot ,\cdot )\)and\(\mathcal {T}_{\beta }^{b,\eta }(\cdot ,\cdot )\)bedefined in Eqs. ??, 5.12and 5.13, respectively. Then for anyη ≥ 0 andS > 0,asu →∞
$$\begin{array}{@{}rcl@{}} \mathbb{P} \left\{ L^{\ast}_{\eta,u}[-S/v(u),S/v(u)] >x \right\} \sim {\Psi}(u)\times \left\{ \begin{array}{ll} \mathcal{P}_{\alpha}^{b,\eta}(S,x) & \text{if}\ \alpha=\beta\\ \mathcal{T}_{\beta}^{b,\eta}(S,x) & \text{if}\ \alpha>\beta \end{array} \right. \end{array} $$holds forx = 0 and anyx ∈ (0,μη([−S,S])) a continuity pointof\(\mathcal {P}_{\alpha }^{b,\eta }(S,\cdot )\)or\(\mathcal {T}_{\beta }^{b,\eta }(S,\cdot )\)respectively.
Proof of Lemma
5.1 i) For any x ≥ 0
Set D = [0,S],gk(u) = u,Ku = 1 and ξu,k(t) = X(tu + t/v(u)). We now check that C1-C3 are fulfilled with h(t) = H(t0)tα,t ∈ [0,S] and \( \zeta (t)=\sqrt {H(t_{0})}B_{\alpha }(t)\). First note that Eq. ?? implies
Further, by Eq. 2.2 and the Uniform Convergence Theorem (see, e.g., Bingham et al. (1989)[Theorem 1.5.2])
and thus
Similarly, for any s,t ∈ [0,S],
Next, according to Eq. 2.2 and Potter’s Theorem (see, e.g., Bingham et al. (1989)[Theorem 1.5.6]), for large enough u there exists some constant C > 1 such that
holds for all s,t ∈ [0,S]. Hence
where in the last inequality we have used Eq. 2.1 with \(\overline {h}=\max _{t\in [0,T]}H(t)\). Therefore, C0-C3 hold and thus our claim i) follows by Theorem 5.1 with \( \mathcal {B}^{h,\eta }_{\zeta }(\textit {\textbf {D}},x)=\mathcal {B}_{\alpha ,H(t_{0})}^{\eta } (S,x)\). The claim in ii) follows with similar arguments. □
Lemma 5.2
If\(\zeta (\textit {\textbf {t}})={\sum }_{i = 1}^{n} \sqrt {\lambda _{i}}B_{\alpha }^{(i)}(t_{i})\)for somepositive constantsλ1…λn,where\(B_{\alpha }^{(i)}\)’s are independentfBm’s with Hurst indexα/2,then for anyx,η ≥ 0,h(t) = Varζ(t) and\(\textit {\textbf {D}}={\prod }_{i = 1}^{n}[0,T_{i}]\)wehave
Proof of Lemma
5.2 For any x,η ≥ 0 and \(\textit {\textbf {D}}={\prod }_{i = 1}^{n}[0,T_{i}]\),
where in the last inequality we have used Hα([0,T]) ≤⌈T⌉Hα([0, 1]). □
Lemma 5.3
If X is a centered Gaussian process with unit variance fulfilling (??),v(u) is defined in Eq. ?? andKuis an index function of u, then for a sequence ofnumbers\(\{0<{S_{1}^{k}}<{S_{2}^{k}}<{T_{1}^{k}}<{T_{2}^{k}}<\infty , k\in K_{u}\}\)suchthat for someu0 > 0
we have
holds for sufficiently large u, where \(\underline {h}=\inf _{t\in [0,T]}H(t)>0\), \(\overline {h}=\sup _{t\in [0,T]}H(t)<\infty \) and
\(C(\alpha ,\overline {S},\overline {T})= 2\lceil (16\overline {h})^{1/\alpha }\rceil ^{2} \lceil \overline {S}\rceil \lceil \overline {T}\rceil H_{\alpha }^{2}([0,1])\).
Proof of Lemma
5.3 We borrow some arguments from the proof of Michna (2017)[Lemma 5]. Define
In view of Eq. 2.1, for sufficiently small ε we have
where the last inequality follows from the regularity of K. Therefore, by Eq. 5.17 for large enough u and then each k ∈ Ku
and
Then for every k ∈ Ku we have
where \(\tilde {Y}(s,t)=Y(s,t)/\sigma (s,t)\) and
Moreover, as in (Michna 2017), for any \((s_{i},t_{i})\in {A_{u}^{k}}\times {B_{u}^{k}}, i = 1,2, k\in K_{u}\) we get
Let \(Z(s,t):=\frac {1}{\sqrt {2}}(\vartheta _{1}(s)+\vartheta _{2}(t)), \)where 𝜗i,i = 1, 2 are mutually independent copies of a mean zero stationary Gaussian process 𝜗 with unit variance and covariance function satisfying
As mentioned in the proof of Theorem 2.1, the existence of such a Gaussian process is guaranteed by the Assertion in (Hüsler and Piterbarg 1999b)[p.265]. Hence by Slepian inequality and Eq. 5.17, for sufficiently large u(> u0) and then each k ∈ Ku we have
where the last inequality follows from stationarity of Z with \(\overline {S}\) and \(\overline {T}\) defined in Eq. 5.18.
Next, we apply Theorem 5.1 to give an uniform upper bound for the right hand side of Eq. 5.21. Setting \(\textit {\textbf {D}}=[0,\overline {S}]\times [0,\overline {T}], g_{k}(u)=u_{k}^{*}, k\in K_{u}\) and ξu,k(s,t) = Z(s/v(u),t/v(u)), we need to check that C0-C3 are fulfilled. First note \(\inf _{k\in K_{u}}u^{\ast }_{k}\geq u\) then C0 is naturally satisfied. Moreover, due to Eq. 5.17 and the regularity of K at 0
and thus
where in the last equality we have used Eqs. 5.14 and 5.20. Therefore, C1 holds with \(h(s,t)= 16\overline {h}(s^{\alpha }+t^{\alpha }), {(s,t)\in [0,\overline {S}]\times [0,\overline {T}]}\). Similarly, for any \((s_{i},t_{i})\in [0,\overline {S}]\times [0,\overline {T}],i = 1,2\)
verifying C2 with \(\zeta (s,t)=\sqrt {16\overline {h}}(B_{\alpha }^{(1)}(s)+B_{\alpha }^{(2)}(t))\), where \(B_{\alpha }^{(i)},i = 1,2\) are independent fBm’s with Hurst index α/2. Next, in view of Eqs. 5.15 and 5.22, for large enough u there exists some constant C > 1 such that
holds for all \(s,t\in [0,{\overline {S}\vee \overline {T}}]\). This together with Eq. 5.20 implies that for any two points \((s_{i},t_{i})\in [0,\overline {S}]\times [0,\overline {T}],i = 1,2\) and sufficiently large u
i.e., C3 is fulfilled with ν = α/2. Therefore, by Theorem 5.1 and Lemma 4.2,
holds for sufficiently large u. Further, by Potter’s theorem, there exists a small enough ε such that
which together with Eqs. 5.16 and 5.17 yields for sufficiently large u(> u0)
Hence
where the last inequality follows from Eq. 2.2. Therefore for large enough u and each k ∈ Ku
which combined with Eq. 5.23 yileds
Consequently, substitution of Eqs. 5.21 and 5.24 to Eq. 5.19 completes the proof. □
Lemma 5.4
Let W be anN(0, 1) random variable independent of\(\mathcal {Z}\)whichis exponentially distributed with parameter 1. For anyc > 0 we have
Proof of Lemma
5.4 Since Z > 0 almost surely, then
Let the random variable V be such that
It is well-known, see e.g., Hashorva (2018)[Lemma 7.1] that V has an N(c2/2,c2) distribution. Hence by the independence of Z and W
establishing the proof. □
Rights and permissions
About this article
Cite this article
Dȩbicki, K., Michna, Z. & Peng, X. Approximation of Sojourn Times of Gaussian Processes. Methodol Comput Appl Probab 21, 1183–1213 (2019). https://doi.org/10.1007/s11009-018-9667-7
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11009-018-9667-7