Abstract
We study limit theorems for a wide class of multi-parameter stochastic processes which are driven by a noise process which may be weakly or even long-range dependent. The processes under study arise in diverse areas and fields such as functional data analysis, life science, engineering and finance. It turns out that under fairly weak conditions on the underlying noise process the limiting law of the corresponding partial sum process is a consequence of the weak convergence of the sequential empirical Kiefer process. The asymptotic limit theory covers the classical large sample situation as well as a general change-point model which extends the location-scale model often considered in change-point analysis. The scope of the results is illustrated by various applications.
Similar content being viewed by others
References
Andersen TG, Bollerslev T, Diebold FX, Labys P (2003) Modeling and forecasting realized volatility. Econometrica 71(2):579–625
Apostol TM (1974) Mathematical analysis, 2nd edn. Addison-Wesley Publishing Co, Reading
Berkes I, Hörmann S, Schauer J (2009) Asymptotic results for the empirical process of stationary sequences. Stoch Process Appl 119(4):1298–1324
Bosq D (2000) Linear processes in function spaces: theory and applications, vol 149. Lecture notes in statistics. Springer, New York
Csörgő M, Révész P (1975) A strong approximation of the multivariate empirical process. Studia Sci Math Hungar 10(3–4):427–434
Dedecker J, Merlevède F, Rio E (2013) Strong approximation results for the empirical process of stationary sequences. Ann Probab 41(5):3658–3696
Dehling H (1983) Limit theorems for sums of weakly dependent Banach space valued random variables. Z Wahrsch Verw Gebiete 63(3):393–432
Dehling H, Philipp W (2002) Empirical process techniques for dependent data. In: Empirical process techniques for dependent data. Birkhäuser Boston, Boston, pp 3–113
Dehling H, Taqqu MS (1989) The empirical process of some long-range dependent sequences with an application to \(U\)-statistics. Ann Stat 17(4):1767–1783
Doukhan P, Oppenheim G, Taqqu MS (eds) (2003) Theory and applications of long-range dependence. Birkhäuser Boston Inc., Boston
Dudley RM, Norvaiša R (2011) Concrete functional calculus. Springer monographs in mathematics. Springer, New York
Embrechts P, Klüppelberg C, Mikosch T (1997) Modelling extremal events for insurance and finance, vol 33. Applications of mathematics (New York). Springer, Berlin
Giraitis L, Kokoszka P, Leipus R (2001) Testing for long memory in the presence of a general trend. J Appl Probab 38(4):1033–1054
Henze N, Klar B, Meintanis SG (2003) Invariant tests for symmetry about an unspecified point based on the empirical characteristic function. J Multivar Anal 87(2):275–297
Horváth L, Kokoszka P (2012) Inference for functional data with applications. Springer series in statistics. Springer, New York
Hušková M, Meintanis SG (2008) Tests for the multivariate \(k\)-sample problem based on the empirical characteristic function. J. Nonparametric Stat. 20(3):263–277
Koul HL (1992) Weighted empiricals and linear models. Institute of Mathematical Statistics lecture notes—monograph series, 21. Institute of Mathematical Statistics, Hayward, CA
Koul HL, Staudte RG Jr (1972) Weak convergence of weighted empirical cumulatives based on ranks. Ann Math Stat 43:832–841
Kurtz TG, Protter P (1991) Weak limit theorems for stochastic integrals and stochastic differential equations. Ann Probab 19(3):1035–1070
Maejima M, Tudor CA (2007) Wiener integrals with respect to the Hermite process and a non-central limit theorem. Stoch Anal Appl 25(5):1043–1056
Motulsky H, Christopoulos A (2004) Fitting models to biological data using linear and nonlinear regression. Oxford University Press, Oxford
Müller DW (1970) On Glivenko-Cantelli convergence. Z Wahrscheinlichkeitstheorie und Verw Gebiete 16:195–210
Philipp W, Pinzur L (1980) Almost sure approximation theorems for the multivariate empirical process. Z Wahrsch Verw Gebiete 54(1):1–13
Politis D, Romano JP, Wolf M (1999) Weak convergence of dependent empirical measures with application to subsampling in function spaces. J Stat Plan Inference 79(2):179–190
Politis DN, Romano JP, Wolf M (1999) Subsampling. Springer series in statistics. Springer, New York
Robinson PM (1991) Testing for strong serial correlation and dynamic conditional heteroskedasticity in multiple regression. J Econom 47(1):67–84
Shorack GR, Wellner JA (1986) Empirical processes with applications to statistics. Wiley series in probability and mathematical statistics: probability and mathematical statistics. Wiley, New York
Sklar M (1959) Fonctions de répartition à \(n\) dimensions et leurs marges. Publ Inst Statist Univ Paris 8:229–231
Steland A, Rafajłowicz E (2014) Decoupling change-point detection based on characteristic functions: methodology, asymptotics, subsampling and application. J Stat Plan Inference 145:49–73
Straf ML (1972) Weak convergence of stochastic processes with several parameters. In: Proceedings of the 6th Berkeley symposium mathematical statistics and probability, vol 2. University of California (1970), pp. 187-221
Whitt W (1980) Some useful functions for functional limit theorems. Math Oper Res 5(1):67–85
Whitt W (2002) Internet supplement of Stochastic-process limits. An introduction to stochastic-process limits and their application to queues. http://www.columbia.edu/~ww2040/allpapers.html
Whitt W (2002) Stochastic-process limits. An introduction to stochastic-process limits and their application to queues. Springer series in operations research. Springer, New York
Yoshihara K (1979) Note on an almost sure invariance principle for some empirical processes. Yokohama Math J 27(2):105–110
Acknowledgments
The author thanks Dr. A. Prause for careful proof-reading of a previous and the final version, and, in particular, for a proposal for a simplified argument to show (6.8) under the weak total variation assumption.
Author information
Authors and Affiliations
Corresponding author
Appendix: Proofs
Appendix: Proofs
Let \( l \in {\mathbb {N}} \). It is worth recalling the following facts on the relevant Skorohod spaces. For brevity of exposition, we consider the case \( D( [0, 1]^d; {\mathbb {R}} ^l ) \), the required modifications for the spaces \( D( [0,1] \times [a, b]; {\mathbb {R}} ^l ) \), \( -\infty \le a < b \le \infty \) being straightforward. In particular, the case \( [a,b] = [-\infty , \infty ] \) is treated by noting that \( [-\infty , \infty ] \) is a complete, separable and metric space when equipped with the metric \(\tilde{d}(x, y) = | G(x) - G(y) | \), \( x, y, \in [-\infty , \infty ] \), where G is strictly increasing, continuous and bounded, e.g., \( G(x) = \text {atan}(x) \). Let \( \Lambda \) denote the class of homeomorphisms of [0, 1] onto itself with \( \lambda (0) = 0 \). In general, for any homeomorphism \( \lambda \) of, say, \( [0,1]^d \), we may define the slope norm by
where \( \Vert \bullet \Vert \) denotes an arbitrary vector norm on \( {\mathbb {R}} ^d \). We consider the group of homeomorphisms of \( [0, 1]^d \) given by \( \Lambda ^d = \Lambda \times \cdots \times \Lambda \), where \( \lambda (t) = (\lambda _1(t), \dots , \lambda _d(t) ) \), \( t = (t_1, \dots , t_d) \in [0, 1]^d \), for \( \lambda \in \Lambda ^d \). A partition \( {\mathcal {D}}\) being invariant under \( \Lambda ^d \) is given by \( {\mathcal {D}}= \{ \lambda ( \Pi _k ) : \lambda \in \Lambda ^d, k \in {\mathbb {N}} _0 \} \), where \( \Pi _k \) is the partition of cells \( A_{i_1, \dots , i_d} = [ (i_1-1)/2^k, i_1/2^k ) \times \cdots \times [(i_d-1)/2^k, i_d/2^k) \), \( 1 \le i_1, \dots , i_d \le 2^k \), \( k \in {\mathbb {N}} _0 \). The simple functions are those functions assuming a constant value on each cell \( D \in {\mathcal {D}}\) and \( D([0,1]^d; {\mathbb {R}} ^l) \) then consists, by definition, of the uniform limits of sequences of such simple functions. The (finite slope) Skorohod metric on \( D([0,1]^d; {\mathbb {R}} ^l) \) is now defined by
By Straf (1972, Theorems 3.7, 3.10), \( ( D([0,1]^d; {\mathbb {R}} ^l), d ) \) is a complete and separable metric space.
Further, recall that the space \( D( [0,1] \times {\mathbb {T}}; {\mathbb {R}} ^l ) \) of càdlàg functions is defined analogously, since \( E = [0,1] \times {\mathbb {T}}\) is a complete metric space.
1.1 Proofs of Section 3
Proof of Theorem 3.1
Observe that \( W_N( u, t ) \) can be represented via the sequential Kiefer process in the form
for \( (u,t) \in [0,1] \times {\mathbb {T}}\). By assumption,
as \( N \rightarrow \infty \). Since the Skorohod space \(D( [0,1] \times [-\infty ,\infty ]; {\mathbb {R}} )\) is Polish, we may apply the Skorohod/Dudley/Wichura representation theorem in general metric spaces to obtain equivalent versions of \( {\mathcal {F}}, {\mathcal {F}}_N \), \( N \ge 1 \), defined on a new probability space, which converge a.s. in the Skorohod metric. For simplicity of notation, we denote these equivalent versions again by \( {\mathcal {F}}_N \) and \( {\mathcal {F}}\), such that
By definition of the metric d , there exist sequences \( \{ \lambda _{Nk} : N \in {\mathbb {N}} \} \), \( k = 1,2 \), of strictly increasing continuous mappings on [0, 1] and \( [-\infty , \infty ] \), respectively, such that
and
as \( N \rightarrow \infty \). Notice that for \( {\mathcal {F}}\in C( [0,1] \times [-\infty , \infty ]; {\mathbb {R}} ) \) (6.2) and (6.3) hold true with \( \lambda _{N1} = {\text {id}}_{[0,1]} \) and \( \lambda _{N2} = {\text {id}}_{[-\infty , \infty ]} =: {\text {id}}\), \( N \ge 1 \).
To treat the càdlàg case, we need the following preparation. Since the strictly increasing transformations \( ( {\text {id}}_{[0,1]}, \lambda _{N2}^{-1} ) \) converge to the strictly increasing and continuous mapping \( ({\text {id}}_{[0,1]}, {\text {id}}_{[-\infty ,\infty ]} ) \), as \( N \rightarrow \infty \), Whitt (1980, Th. 3.1) yields
in \( D( [0,1] \times [-\infty ,\infty ]; {\mathbb {R}} ) \), as \( N \rightarrow \infty \). (The proof of Whitt carries over with appropriate norms: The mapping \( \mu _n \) constructed there is now constructed for both coordinates). Thus, by invoking again the Skorohod/Dudley/Wichura representation theorem, we may and shall assume in what follows that, additionally,
as \( N \rightarrow \infty \), a.s.
Let us first consider the case \( {\mathcal {F}}\in C( [0,1] \times [-\infty , \infty ]; {\mathbb {R}} ) \) a.s., to prepare the more involved arguments for the càdlàg case. Observe that, by definition of the Stieltjes integral,
where \( W_N^{(A,B)}(u,t) = \int _A^B \phi (x; t) \, {\mathcal {F}}_N(u,dx) \) and \( W^{(A,B)}(u,t) = \int _A^B \phi (x; t) \, {\mathcal {F}}(u,dx) \), \( -\infty < A < B < \infty \). Observe that
Thus, provided we can show that \( \left| W_N^{(A,B)}(u,t) - W^{(A,B)}(u,t) \right| = o(1) \), a.s., uniformly in \( u \in [0,1] \) and \( t \in {\mathbb {T}}\), where in the càdlàg case we shall have to implement appropriately chosen homeomorphisms, the result follows.
Thus, fix \( - \infty < A < B < \infty \). Integration by parts yields
The right-hand side can be bounded by
which is o(1) , as \( N \rightarrow \infty \), a.s. Thus the result is shown. Having in mind the corollary, notice that if \( \Vert {\mathcal {F}}_N - {\mathcal {F}}\Vert _\infty \rightarrow 0 \), as \( N \rightarrow \infty \), a.s., on the original probability space, we obtain the stronger result of a.s. uniform convergence, \( \Vert W_N - W \Vert _\infty \rightarrow 0 \), as \( N \rightarrow \infty \), a.s. In the general case, the weak convergence \( W_N \Rightarrow W \), as \( N \rightarrow \infty \), follows by the second half of the Skorohod/Dudley/Wichura representation theorem.
To establish the result for the càdlàg case, let \( \lambda _{N1} : [0,1] \rightarrow [0,1] \) and \( \lambda _{N2} : [-\infty , \infty ] \rightarrow [-\infty , \infty ] \), \( N \ge 1 \), be homeomorphisms such that
as \( N \rightarrow \infty \), a.s. For \( -\infty < A < B < \infty \) define
\( u \in [0,1] \), \( t \in {\mathbb {T}}\). We shall show that
as \( N \rightarrow \infty \), a.s., which yields
a.s., as \( N \rightarrow \infty \), i.e., \( W_N \Rightarrow W \) for the original versions.
To verify (6.5) put \( \phi _A^B(x;t) = \phi (x;t) { 1}_{[A,B]}(x) \). Recalling that \( {\mathcal {F}}_N(u,x) = r_N {\lfloor Nu \rfloor } [ F_{ {\lfloor Nu \rfloor } }(x) - F(x) ] \), we have
Recall that \( \lambda _{N1} \) and \( \lambda _{N2} \) are homeomorphisms of [0, 1] and \( [-\infty , \infty ] \), respectively, which ensure (6.2). Since
we have
for \( (u,t) \in [0,1] \times [a,b] \), where
Notice that \( \phi _N^{(A,B)} = \phi _N \) on \( [A,B] \times {\mathbb {T}}\), where
since \( \phi _A^B = \phi \) on \( [A,B] \times {\mathbb {T}}\) and \( \lambda _{N2} \), \( N \ge 1 \), are homeomorphisms of \( [-\infty ,\infty ] \). It follows that
We will now show that
as \( N \rightarrow \infty \), a.s. To do so, we employ the decomposition
as \( N \rightarrow \infty \), a.s., where
as \( N \rightarrow \infty \). Let us first consider \( I_{N2} \). Recalling that \( \phi _N(x;t) = \phi ( \lambda _{N2}(x); t ) \), \( (x,t) \in [A,B] \times {\mathbb {T}}\) and the change of variable formula for Stieltjes integrals,
see e.g., Apostol (1974, Th. 7.4), we obtain
Thus,
such that, by virtue of (6.4), a.s.
as \( N \rightarrow \infty \), where \( C = 2 ( C_{TV} + C_{\sup } ) \).
To handle \( I_{N1} \), let us verify that \( x \mapsto \phi _N(x;t) \) is of uniformly bounded variation on [A, B] . Let \( A = x_0 < \cdots < x_n = B \) be an arbitrary partition of [A, B] . Notice that \( \{ \lambda _{N2}( x_i ) : i = 1, \dots , n \} \) forms again a partition of [A, B] (that depends on N), since the \( \lambda _{N2} \) are increasing bijections. But then we have
a.s., as \( N \rightarrow \infty \), which verifies that, a.s.,
But now the required estimate of \( I_{N1} \) follows easily again by integration by parts.
which is \(o_P(1)\) as \( N \rightarrow \infty \), where \( C^{\prime } = 2( C_{TV} + C_{\sup } ) \).
This proves (6.6) which in turn implies that
as \( N \rightarrow \infty \), a.s., i.e., (6.5) is established. \(\square \)
Proof of Corollary 3.2
By periodicity, we have
Let \( \varepsilon > 0 \). Since \( f(x) = O (|x|^{-\gamma }) \), if \( |x| \rightarrow \infty \), we may find \( k_0 \) such that
Obviously, if \( t \ge t_0 > 0 \) for some \( t_0 \), the O term is uniform in \( t \in {\mathbb {T}} \cap [t_0,\infty ) \). \( \Box \)
Proof of Lemma 3.1
By boundedness of \( \phi \) and Davydov’s inequality, we have
Hölder’s inequality and dominated convergence thus imply
as \( a \rightarrow \infty \). \(\square \)
Proof Theorem 3.2
We may assume that \( \{ X_n : n \in {\mathbb {Z}} \} \) is a strictly stationary sequence of random curves satisfying the assumptions, from which we observe \( X_1, \dots , X_N \). We equip \(L_2[a,b] \) with the Borel \( \sigma \)-field associated to the metric induced by the inner product. Recall that the \( \alpha \)-mixing coefficients are given by \( \alpha (k) = \sup _{A \in {\mathcal {F}}_{-\infty }^0, B \in {\mathcal {F}}_k^\infty } | P(A \cap B) - P(A) P(B) | \), \( k \in {\mathbb {N}} _0 \), where \( {\mathcal {F}}_{-\infty }^0 = \sigma ( X_i : i \le 0 ) \) and \( {\mathcal {F}}_k^{\infty } = \sigma ( X_i : i \ge k ) \). Clearly, the strictly stationary time series \( Y_{i1} = \int _a^b X_i(u) \varphi _1(u) \, du, i \ge 1 \), inherits those mixing coefficients. We have
such that Assumption (A) holds true by virtue of Yoshihara (1979). We may put \( {\mathbb {Y}} = [-L,L] \) with \(L = K \Vert \varphi _1 \Vert _{L_2[a,b]} \Vert \varphi _1 \Vert _\infty \). By continuity of the covariance operator, \( \Vert \varphi _1 \Vert _\infty < \infty \) holds true. Further, \( \sup _{t \in {\mathbb {T}}} \widetilde{ \phi } ( \bullet ; t ) \) is bounded (on \({\mathbb {Y}}\)) and \( \Vert \widetilde{ \phi } ( \bullet ; t ) \Vert _{TV,{\mathbb {Y}}} \le 2 K \Vert \varphi _1 \Vert _{L_2[a,b]} \Vert \varphi _1 \Vert _\infty \), such that \( C_{TV} < \infty \), a.s. This completes the proof. \(\square \)
1.2 Proofs of Section 4
We show Theorem 4.1, which provides the asymptotic null distribution of \( T_N(u;t) \), and Theorem 4.3, which shows that under the general transformation change-point model the underlying process \( {\mathcal {F}}_N(s,x) \), which induces \( T_N(u;t) \) via the representation
is affected by a drift which yields the detection power.
Proof of Theorem 4.1
Again we apply the Skorohod/Dudley/Wichura representation theorem, assume w.l.o.g. that \( {\mathcal {F}}_N \) converges to \( {\mathcal {F}}\) in the Skorohod metric and fix \( - \infty < a < b < \infty \). Notice that for \( u \in [1/N,1] \) and \( t \in {\mathbb {T}}\)
where
for \( N \ge 1 \). Since under the assumptions of the theorem \( {\mathcal {F}}_N = r_N {\lfloor Nr \rfloor } [ F_{ {\lfloor Nr \rfloor } }(x) - F_0(x) ] \), by using the fact that
we obtain the representation (5.9). Define
Note that for any functions f and g
Combining (5.10) with the reverse triangle inequality yields
a.s., as \( N \rightarrow \infty \), uniformly in \( u \in (0,1] \) and \( t \in [a,b]\), again using the integration by parts techniques elaborated in the previous proofs. Using similar arguments, it follows that
a.s, as \( N \rightarrow \infty \), uniformly in \(u \in (0,1] \) and \( t \in [a,b] \). Lastly, we show that \( | T_{N2} - T_{N1} | \) is negligible. Notice that
where
for \( 0 \le s \le u \). By virtue of (5.10) it follows that
as \( N \rightarrow \infty \), again uniformly in \( u \in (0,1] \) and \( t \in [a,b] \). This verifies
a.s., as \( N \rightarrow \infty \). Hence the result follows. \(\square \)
Proof of Theorem 4.3
We argue similarly as in the proof of Theorem 3.1 and make use of the Skorohod/Dudley/Wichura representation theorem. Thus, by defining equivalent versions on a new probability space, we can assume that under \( P_0 \)
as \( N \rightarrow \infty \). Let us also introduce an additional time series \( \{ Z_t \} \) satisfying
where \( {\mathcal {L}}( \{ \xi _t \}; Q ) \) denotes the law of a time series \( \{ \xi _t \} \) under the probability Q. This means, under \( P_\vartheta \), when there is a change-point, \( \{ Z_t \} \) is distributed as \( \{ Y_t \} \) under the no-change probability measure \( P_0 \). Also define the associated empirical sequential Kiefer process
Clearly, under \( P_\vartheta \) we have \( \widetilde{ {\mathcal {F}} } _N \Rightarrow \widetilde{ {\mathcal {F}} } \), as \( N \rightarrow \infty \), for some stochastic process \( \{ {\mathcal {F}}(u,x) : u \in [0,1], x \in [-\infty , \infty ] \} \) with \( {\mathcal {L}}( \widetilde{ {\mathcal {F}} } ; P_\vartheta ) = {\mathcal {L}}( {\mathcal {F}}; P_0 ) \), and w.l.o.g. we again may assume that the convergence is a.s. w.r.t. the supnorm. By assumption, under \( P_\vartheta \) the law of
can be obtained by transforming the after-change observations, i.e., the law of
under \( P_\vartheta \) equals the law of
under \( P_0 \). Using
we therefore obtain, for those parts,
and
as well as jointly. Notice that we centered the terms corresponding to time instants after the change at their expectations which introduces a non-random correction term. For large enough N that result can be written as
where \( \widetilde{ {\mathcal {F}} } _N^{(\vartheta )}(u,x) \) is given by
for \( u \in [\vartheta , 1] \) and \( x \in [a,b] \). Obviously, under \( P_\vartheta \), the process
converges weakly to
as \( N \rightarrow \infty \), which shows the first assertion. The second assertion now follows easily. \(\square \)
Rights and permissions
About this article
Cite this article
Steland, A. Asymptotics for random functions moderated by dependent noise. Stat Inference Stoch Process 19, 363–387 (2016). https://doi.org/10.1007/s11203-015-9130-0
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11203-015-9130-0