Abstract
This paper deals with extreme-value index estimation of a heavy-tailed distribution of a spatial dependent process. We are particularly interested in spatial rare events of a \(\beta\)-mixing process. Given a stationary real-valued multidimensional spatial process \(\left\{X_{\mathbf{i}},\mathbf{i}\in{\mathbb{Z}}^{N}\right\}\), we investigate its heavy-tail index estimation. Asymptotic properties of the corresponding estimator are established under mild mixing conditions. The particularity of the tail proposed estimator is based on the spatial nature of the sample and its unbiased and reduced variance properties compared to well known tail index estimators. Extreme quantile estimation is also deduced. A numerical study on synthetic and real datasets is conducted to assess the finite-sample behaviour of the proposed estimators.
REFERENCES
B. Basrak and A. Tafro, ‘‘Extremes of moving averages and moving maxima on a regular lattice,’’ Probability and Mathematical Statistics 34, 61–67 (2014).
A. Bassene, Contribution á la modélisation spatiale des événements extrêmes (PhD Thesis, Université Charles de Gaulle-Lille III; Université Gaston Berger de Saint-Louis (Sénégal), 2016).
J. Beirlant, Y. Goegebeur, J. Segers, and J. L. Teugels, Statistics of extremes: theory and applications (John Wiley and Sons, 2006).
J. Blanchet and A. C. Davison, ‘‘Spatial modeling of extreme snow depth,’’ The Annals of Applied Statistics 1699–1725 (2011).
S. Bobbia, R. Macwan, Y. Benezeth, A. Mansouri, and J. Dubois, ‘‘Unsupervised skin tissue segmentation for remote photoplethysmography,’’ Pattern Recognition Letters 124, 82–90 (2019).
S. Bobbia, R. Macwan, Y. Benezeth, K. Nakamura, R. Gomez, and J. Dubois, ‘‘Iterative boundaries implicit identification for superpixels segmentation: a real-time approach,’’ IEEE Access (2021).
C. Bolancé and M. Guillen, ‘‘Nonparametric estimation of extreme quantiles with an application to longevity risk,’’ Risks 9 (4), 77 (2021).
G. P. Bopp, B. A. Shaby, and R. Huser, ‘‘A hierarchical max-infinitely divisible spatial model for extreme precipitation,’’ Journal of the American Statistical Association 116 (533), 93–106 (2021).
R. C. Bradley, ‘‘Some examples of mixing random fields,’’ The Rocky Mountain Journal of Mathematics, 495–519 (1993).
V. Chavez-Demoulin and A. Guillou, ‘‘Extreme quantile estimation for \(\beta\)-mixing time series and applications,’’ Insurance: Mathematics and Economics 83, 59–74 (2018).
A. Daouia, S. Girard, and G. Stupfler, ‘‘Tail expectile process and risk assessment,’’ Bernoulli 26 (1), 531–556 (2020).
R. A. Davis, C. Klüppelberg, and C. Steinkohl, ‘‘Statistical inference for max-stable processes in space and time,’’ Journal of the Royal Statistical Society: SERIES B: Statistical Methodology, 791–819 (2013).
A. C. Davison, S. A. Padoan, and M. Ribatet, ‘‘Statistical modeling of spatial extremes,’’ Statistical Science 27 (2), 161–186 (2012).
L. De Haan and S. Resnick, ‘‘On asymptotic normality of the hill estimator,’’ Stochastic Models 14 (4), 849–866 (1998).
L. De Haan and A. Ferreira, Extreme Value Theory: An Introduction, vol. 21 (Springer, 2006).
L. de Haan, C. Mercadier, and C. Zhou, ‘‘Adapting extreme value statistics to financial time series: dealing with bias and serial dependence,’’ Finance and Stochastics 20 (2), 321–354 (2016).
J. Dedecker, P. Doukhan, G. Lang, L. R. J. Rafael, S. Louhichi, and C. Prieur, Weak Dependence. In Weak Dependence: With Examples and Applications (Springer, 2007), p. 9–20.
H. Drees, ‘‘Weighted approximations of tail processes for b-mixing random variables,’’ Annals of Applied Probability, 1274–1301 (2000).
H. Drees, ‘‘Extreme quantile estimation for dependent data, with applications to finance,’’ Bernoulli 9 (4), 617–657 (2003).
Y. Goegebeur and A. Guillou, ‘‘Asymptotically unbiased estimation of the coefficient of tail dependence,’’ Scandinavian Journal of Statistics 40 (1), 174–189 (2013).
Y. Goegebeur, A. Guillou, and A. Schorgen, ‘‘Nonparametric regression estimation of conditional tails: The random covariate case,’’ Statistics 48 (4), 732–755 (2014).
M. I. Gomes, L. De Haan, and L. Peng, ‘‘Semi-parametric estimation of the second order parameter in statistics of extremes,’’ Extremes 5 (4), 387–414 (2002).
B. M. Hill, ‘‘A simple general approach to inference about the tail of a distribution,’’ The Annals of Statistics 1163–1174 (1975).
T. Hsing, ‘‘On tail index estimation using dependent data,’’ The Annals of Statistics, 1547–1569 (1991).
D. Kurisu, K. Kato, and X. Shao, Gaussian approximation and spatially dependent wild bootstrap for high-dimensional spatial data. arXiv preprint arXiv:2103.10720 (2021).
D. M. Mason, ‘‘Laws of large numbers for sums of extreme values,’’ The Annals of Probability, 754–764 (1982).
P. Ndao, A. Diop, and J.-F. Dupuy, ‘‘Nonparametric estimation of the conditional tail index and extreme quantiles under random censoring,’’ Computational Statistics and Data Analysis 79, 63–79 (2014).
T. Opitz, Extrêmes multivariés et spatiaux, PhD Thesis, Université Montpellier 2 (Sciences et Techniques, 2013).
T. Opitz, ‘‘Modeling asymptotically independent spatial extremes based on laplace random fields,’’ Spatial Statistics 16, 1–18 (2016).
S. Resnick and C. Stǎricǎ, ‘‘Consistency of hill’s estimator for dependent data,’’ Journal of Applied probability, 139–167 (1995).
S. Resnick and C. Stǎricǎ, ‘‘Tail index estimation for dependent data,’’ Annals of applied Probability 8 (4), 1156–1183 (1998).
P. M. Robinson, ‘‘Asymptotic theory for nonparametric regression with spatial data,’’ Journal of Econometrics 165 (1), 5–19 (2011).
P. Sharkey and H. C. Winter, ‘‘A bayesian spatial hierarchical model for extreme precipitation in great britain,’’ Environmetrics 30 (1), e2529 (2019).
E. Thibaud, R. Mutzner, and A. C. Davison, ‘‘Threshold modeling of extreme spatial rainfall,’’ Water resources research 49 (8), 4633–4644 (2013).
K. F. Turkman, M. A. Turkman, and J. Pereira, ‘‘Asymptotic models and inference for extremes of spatio-temporal data,’’ Extremes 13 (4), 375–397 (2010).
J. Velthoen, C. Dombry, J.-J. Cai, and S. Engelke, Gradient boosting for extreme quantile regression. arXiv preprint arXiv:2103.00808 (2021).
I. Weissman, ‘‘Estimation of parameters and large quantiles based on the \(k\) largest observations,’’ Journal of the American Statistical Association 73 (364), 812–815 (1978).
ACKNOWLEDGMENTS
The authors acknowledge the Associate Editor and an anonymous reviewer for their helpful comments and suggestions that led to an improved version of this paper. Tchamiè Tchazino was supported by DAAD bursary grants implemented by Institut de Mathématiques et de Sciences Physiques (IMSP)-Benin. This publication was made possible through support provided by the IRD and AFD.
Author information
Authors and Affiliations
Corresponding authors
Appendices
APPENDIX
Proofs of the Main Results
To establish the proofs of the main results, we adopt [32]’s notation of the spatial locations (for seek of simplicity). That is the process \(\left\{X_{\mathbf{i}},\mathbf{i}\in\mathbb{Z}^{N}\right\}\) is written as \(\left\{X_{i},1\leq i\leq n=n_{1}\times n_{2}\times\cdots\times n_{N}\right\}\) using for instance a triangular array notation and a lexicographic ordering. For this notation the mixing conditions \(C_{M}\) and \(C_{R}\) (regularity) are written as.
Condition \(C^{\prime}_{M}\) (mixing condition). Let’s \((l_{n})_{n\in\mathbb{N}^{*}}\) be a sequence of integers such that \(1\leq l_{n}\leq{n}\); set \(\mathcal{B}_{m}^{j}=\sigma(X_{i},m\leq i\leq j)\) be \(\sigma\)-fields generated by the random variables \((X_{i})_{i}\) with \(m\leq i\leq j\). The \(\beta\)-mixing condition is given by:
See [18] for a discussion on the \(\beta\)-mixing and examples.
Condition \(C^{\prime}_{R}\) (regularity). There is \(\epsilon>0\), a function \(r:\mathbb{R}\times\mathbb{R}\rightarrow\mathbb{R}\), and \(({l}_{n})\) defined above is such that \(l_{n}=\text{o}(n/k_{n})\); and when \(n\rightarrow\infty\)
-
(a\({}^{\prime}\)) \(\frac{\beta(l_{n})}{l_{n}}n+l_{n}\frac{\text{log}^{2}k_{n}}{\sqrt{k_{n}}}\rightarrow 0\);
-
(b\({}^{\prime}\)) \(\frac{n}{l_{n}k_{n}}\text{Cov}\left({{\underset{i=1}{\overset{l_{n}}{\sum}}}}{1}_{\{X_{i}>F^{\leftarrow}(1-k_{n}x/n)\}},{{\underset{i=1}{\overset{l_{n}}{\sum}}}}{1}_{\{X_{i}>F^{\leftarrow}(1-k_{n}y/n)\}}\right)\rightarrow r(x,y),\) \(\forall\ 0\leq x,y\leq 1+\epsilon\);
-
(c\({}^{\prime}\)) there exists a constant \(C\) such that :
$$\frac{n}{l_{n}k_{n}}E\left[\left({{\underset{i=1}{\overset{l_{n}}{\sum}}}}{1}_{\{F^{\leftarrow}(1-k_{n}y/n)<X_{i}\leq F^{\leftarrow}(1-k_{n}x/n)\}}\right)^{4}\right]\leq C(y-x)\quad\forall\ 0\leq x<y\leq 1+\epsilon.$$
A. Proof of Theorem II.1
To establish the proof of the theorem, we need the following proposition.
Proposition VI.1. Let \(\left\{X_{\mathbf{i}},\mathbf{i}\in{\mathbb{Z}}^{N}\right\}\) be a \(\beta\) -mixing stationary spatial process with a distribution function \(F\) ; verifying \(C_{A}\) and \(C_{R}\) and \(K\) a function verifying \(C_{K}\) . Let \((k_{n})\) be an intermediate sequence such that \(\sqrt{k_{n}}\mathcal{A}(b(n/k_{n}))\to\lambda\) , as \(n\to\infty\) . For all \(\epsilon>0\) , by Skorohod construction, there exist a function \(\tilde{\mathcal{A}}\sim\mathcal{A}\) and a Gaussian centred process \((W(t))_{t\in[0,1]}\) with covariance function \(r\) such that, as \(n\rightarrow\infty\) ,
Proof of Proposition VI.1. Suppose the relationship (12) (from the condition \(C_{A}\)) hold. By applying Theorem B\(.2.18\) in [15], we get:
\(\forall\ \epsilon\), \(\delta>0\) \(\exists\ u_{0}=u_{0}(\epsilon,\delta)\) such that \(\forall\ ux\geq u_{0}\);
Set \(X_{i}=U(Y_{i})\) where \(Y_{i}\) follows a standard Pareto distribution. Then \((Y_{\mathbf{i}})_{\mathbf{i}\in\mathbb{Z}^{N}}\) is stationary and \(\beta\)-mixing satisfying the regular variation (\(C_{R}\)). Then, since \(Q_{n}(t)=U(Y_{n-\lfloor k_{n}t\rfloor,n})\) and according to Theorem \(2.1\) in Drees \((2003)\) and under Skorohod construction, there exists a centred Gaussian process \((W(t))_{t\in[0,1]}\) with covariance function \(r\) such that for all \(\epsilon>0\)
as \(n\rightarrow\infty\). The inequality \((38)\) gives, for all \(n>n_{0}(\epsilon,\delta)\):
So,
Since \(\frac{1}{b(n/k_{n})}Y_{n-\lfloor k_{n}t\rfloor,n}\geq 1\), by choosing \(\delta\in(0,-\rho)\) the right term tends to \(0\) when \(\epsilon\rightarrow 0\). Thus under the convergence \((39)\) the proof of the Proposition VI.1 is obtained as in [16].
Proof of Theorem II.1.: From Proposition VI.1, we deduce that:
Using an integration by part, we can write:
Hence
By taking \(0<\epsilon<1/2-\tau\), we get the convergence of \(\int_{0}^{1}t^{-1/2-\epsilon}d(tK(t))\). Thus, this ends the proof.
B. Proof of Corollary II.1
Indeed, the term \(\mathcal{A}(b(n/k_{n}))\int_{0}^{1}t^{-\rho}K(t)dt\) outcome from (13) is the bias of the estimator; and as \(\sqrt{k_{n}}\mathcal{A}(b(n/k_{n}))\longrightarrow\lambda\) then we get the asymptotic bias \(\lambda\mathcal{AB}(K)=\lambda\int_{0}^{1}t^{-\rho}K(t)dt\).
The variance \(\mathcal{AV}(K)\) is obtained from the Gaussian centered process \((W(t))_{t\in[0.1]}\) covariance function \(r\); that is
The proofs of Corollaries II.2 and II.3 are straightforward and follow the same lines as the Corollary II.1.
C. Proof of Theorem II.2
Let
where \(\hat{\alpha}\) and \(\hat{\beta}\) are consistent estimators of \(\alpha\) and \(\beta\), respectively.
Let us first, give the following decomposition
According to Corollary II.2, the first term on the right converges to Gaussian distribution. So it remains to prove that the second term tends to \(0\) in probability. The proof follows the same lines as those of Theorem \(2\) in [10]. The difference in our approach is in managing the assumptions of the \(C_{R}\) and \(C_{K}\) conditions since we assumed that \(K\) is not necessarily a kernel function. We have:
Let us evaluate the three terms.
Using Corollary II.1 and the fact that \(\hat{\alpha}\) and \(\hat{\beta}\) are consistent estimators of \(\alpha\) and \(\beta\), respectively, we have \(T_{1}=\text{o}_{\mathbb{P}}(1)\) and \(T_{2}=\text{o}_{\mathbb{P}}(1)\).
It remains to deal with Term \(T_{3}\) .
Noting that \(\log\frac{Q_{n}(t)}{Q_{n}(1)}=\log\frac{Q_{n}(t)}{\text{U}(b(n/k_{n}))}-\log\frac{Q_{n}(1)}{\text{U}(b(n/k_{n}))}\), the Proposition VI.1 gives for all \(\epsilon\in(0,1/2)\)
Thus we get:
The term \(A\) converges by using an integration by part.
The term \(B\) is
Let us consider \(\epsilon\in(0,1)\) and \(\tilde{\rho}\) a random value between \(\rho\) and \(\hat{\rho}_{k_{n\rho}}\). We have
The inequality \(\overset{*}{\leq}\) is justified by:
set \(h(t)=t^{\frac{1}{4}-\hat{\rho}_{k_{n,\rho}}}-t^{\frac{1}{4}-\rho}=t^{\frac{1}{4}-\tilde{\rho}}(t^{\tilde{\rho}-\hat{\rho}_{k_{n,\rho}}}-t^{\tilde{\rho}-\rho})\). A Taylor expansion of the term \((t^{\tilde{\rho}-\hat{\rho}_{k_{n,\rho}}}-t^{\tilde{\rho}-\rho})\) gives
and we get
In the same way we have:
Then we get \(B=\text{o}_{\mathbb{P}}(1)\).
By making an integration by part \(C\) is
and as \(\sqrt{k_{n}}\tilde{\mathcal{A}}(b(n/k_{n}))\rightarrow\lambda\) and \(\int_{0}^{1}t^{-\rho}\left(K_{2,\hat{\rho}_{k_{n,\rho}}}(t)-K_{2,\rho}(t)\right)dt\) converges so we have \(C=\text{o}_{\mathbb{P}}(1)\).
Similarly, an integration by part allows us to conclude that \(D=\text{o}_{\mathbb{P}}(1)\).
In short, we have \(T_{3}=\text{o}_{\mathbb{P}}(\hat{\beta})\) and since \(\hat{\beta}<1\) we have
This ends the proof of Theorem II.2.
D. Proof of Theorem III.1
To do this, we only need to show the asymptotic normality of \(\frac{\sqrt{k_{n}}}{\text{log}\frac{1}{pb(n/k_{n})}}\text{log}\frac{\hat{x}_{p,\xi}}{{x}_{p}}\).
We have the decomposition below:
Let us now look at the \(5\) terms.
Theorem II.2 ensures the asymptotic normality of the term \(T_{4}\)
Using Proposition VI.1 (for \(t=1\) and the fact that \(\log(x)\sim x-1\) when \(x\rightarrow 1\)), we can show
Indeed,
To prove that \(T_{6}=\text{o}(1)\), inequation (38) leads to:
for all \(0<\delta<-\rho\).
Note that the term \(T_{8}\) is a function of \(\xi\) which can be a canonical value or an consistent estimator of \(\rho\).
-
If \(\xi=\tilde{\rho}\) then, we have:
\(\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{1})-\hat{\gamma}_{k_{n}}(K_{2,\xi})]=\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{1})-\gamma]-\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{2,\xi})-\gamma]=\text{O}_{\mathbb{P}}(1)\) according to the Corollary II.1. This leads to \(T_{8}=\text{o}_{\mathbb{P}}(1)\).
-
If \(\xi=\hat{\rho}\)
$$T_{8}=\frac{(1-\hat{\rho})(1-2\hat{\rho})}{\hat{\rho}^{2}}\frac{\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{1})-\hat{\gamma}_{k_{n}}(K_{2,\hat{\rho}})]}{\text{log}\frac{1}{pb(n/k_{n})}}$$$${}\times\frac{\left(\frac{1}{pb(n/k_{n})}\right)^{{\rho}}-1}{{\rho}}+\frac{(1-\hat{\rho})(1-2\hat{\rho})}{\hat{\rho}^{2}}$$$${}\times\frac{\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{1})-\hat{\gamma}_{k_{n}}(K_{2,\hat{\rho}})]}{\log\frac{1}{pb(n/k_{n})}}\left(\frac{\left(\frac{1}{pb(n/k_{n})}\right)^{\hat{\rho}}-1}{\hat{\rho}}-\frac{\left(\frac{1}{pb(n/k_{n})}\right)^{{\rho}}-1}{{\rho}}\right).$$However, according to Corollary II.1 and Theorem II.2,
$$\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{1})-\hat{\gamma}_{k_{n}}(K_{2,\hat{\rho}})]=\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{1})-\gamma]-\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{2,\rho})-\gamma]$$$${}-\sqrt{k_{n}}[\hat{\gamma}_{k_{n}}(K_{2,\hat{\rho}})-\hat{\gamma}_{k_{n}}(K_{2,\rho})]=\text{O}_{\mathbb{P}}(1).$$The term \(T_{8}\) becomes,
$$T_{8}=\text{o}_{\mathbb{P}}(1)+\text{o}_{\mathbb{P}}(1)\left\{\frac{\left(\frac{1}{pb(n/k_{n})}\right)^{\hat{\rho}}-1}{\hat{\rho}}-\frac{\left(\frac{1}{pb(n/k_{n})}\right)^{{\rho}}-1}{{\rho}}\right\}$$$${}=\text{o}_{\mathbb{P}}(1)+\text{o}_{\mathbb{P}}(1)\int\limits_{0}^{\frac{1}{pb(n/k_{n})}}s^{\rho-1}(s^{\hat{\rho}-\rho}-1)ds.$$Inspired by [10], we get
$$\displaystyle\int\limits_{0}^{\frac{1}{pb(n/k_{n})}}s^{\rho-1}(s^{\hat{\rho}-\rho}-1)ds=\text{o}_{\mathbb{P}}(1),$$
which leads to the conclusion that \(T_{8}=\text{o}_{\mathbb{P}}(1)\) and therefore we get the proof of Theorem III.1.
About this article
Cite this article
Tchazino, T., Dabo-Niang, S. & Diop, A. Tail and Quantile Estimation for Real-Valued \(\boldsymbol{\beta}\)-Mixing Spatial Data. Math. Meth. Stat. 31, 135–164 (2022). https://doi.org/10.3103/S1066530722040044
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.3103/S1066530722040044