Abstract
In this paper we propose statistical inference tools for the covariance operators of functional time series in the two sample and change point problem. In contrast to most of the literature, the focus of our approach is not testing the null hypothesis of exact equality of the covariance operators. Instead, we propose to formulate the null hypotheses in the form that “the distance between the operators is small”, where we measure deviations by the sup-norm. We provide powerful bootstrap tests for these type of hypotheses, investigate their asymptotic properties and study their finite sample properties by means of a simulation study.
Similar content being viewed by others
References
Aue, A., Dubart Norinho, D., Hörmann, S. (2015). On the prediction of stationary functional time series. Journal of the American Statistical Association, 110, 378–392.
Aue, A., Rice, G., Sönmez, O. (2018). Detecting and dating structural breaks in functional data without dimension reduction. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 80(3), 509–529.
Aue, A., Rice, G., Sönmez, O. (2020). Structural break analysis for spectrum and trace of covariance operators. Environmetrics, 31(1), e2617.
Billingsley, P. (1968). Convergence of probability measures. New York: Wiley.
Boente, G., Rodriguez, D., Sued, M. (2018). Testing equality between several populations covariance operators. Annals of the Institute of Statistical Mathematics, 70(4), 919–950.
Bosq, D. (2000). Linear Processes in Function Spaces: Theory and Applications, Lecture Notes in Statistics. New York: Springer.
Bücher, A., Kojadinovic, I. (2016). A dependent multiplier bootstrap for the sequential empirical copula process under strong mixing. Bernoulli, 22(2), 927–968.
Bücher, A., Kojadinovic, I. (2019). A note on conditional versus joint unconditional weak convergence in bootstrap consistency results. Journal of Theoretical Probability, 32, 1145–1165.
Cabassi, A., Pigoli, D., Secchi, P., Carter, P. A. (2017). Permutation tests for the equality of covariance operators of functional data with applications to evolutionary biology. Electronic Journal of Statistics, 11(2), 3815–3840.
Cárcamo, J., Rodríguez, L.-A., Cuevas, A. (2020). Directional differentiability for supremum-type functionals: Statistical applications. Bernoulli, 26(3), 2143–2175.
Carey, J. R., Liedo, P., Müller, H.-G., Wang, J.-L., Chiou, J.-M. (1998). Relationship of age patterns of fecundity to mortality, longevity, and lifetime reproduction in a large cohort of mediterranean fruit fly females. The Journals of Gerontology Series A, Biological Sciences and Medical Sciences, 53, B245-51.
Carlstein, E. (1986). The use of subseries methods for estimating the variance of a general statistic from a stationary time series. Annals of Statistics, 14(3), 1171–1179.
Dehling, H. (1983). Limit theorems for sums of weakly dependent banach space valued random variables. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 63(2), 393–432.
Dehling, H., Philipp, W. (2002). Empirical Process Techniques for Dependent Data (pp. 3–113). Boston, MA: Birkhäuser Boston.
Dette, H., Kokot, K., Aue, A. (2020a). Functional data analysis in the Banach space of continuous functions. Annals of Statistics, 48(2), 1168–1192.
Dette, H., Kokot, K., Volgushev, S. (2020b). Testing relevant hypotheses in functional time series via self-normalization. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 82(3), 629–660.
Ferraty, F., Vieu, P. (2010). Nonparametric functional data analysis. New York: Springer.
Fremdt, S., Steinebach, J. G., Horváth, L., Kokoszka, P. (2013). Testing the equality of covariance operators in functional samples. Scandinavian Journal of Statistics, 40(1), 138–152.
Gaenssler, P., Molnár, P., Rost, D. (2007). On continuity and strict increase of the cdf for the sup-functional of a gaussian process with applications to statistics. Results in Mathematics, 51(1), 51–60.
Guo, J., Zhou, B., Zhang, J.-T. (2018). Testing the equality of several covariance functions for functional data: a supremum-norm based test. Computational Statistics & Data Analysis, 124, 15–26.
Horváth, L., Kokoszka, P. (2012). Inference for functional data with applications. New York: Springer.
Hsing, T., Eubank, R. (2015). Theoretical foundations of functional data analysis, with an introduction to linear operators. New York: Wiley.
Janson, S., Kaijser, S. (2015). Higher moments of Banach space valued random variables (p. 238). Providence, RI: Memoirs of the American Mathematical Society.
Jarušková, D. (2013). Testing for a change in covariance operator. Journal of Statistical Planning and Inference, 143(9), 1500–1511.
Kraus, D., Panaretos, V. M. (2012). Dispersion operators and resistant second-order functional data analysis. Biometrika, 99(4), 813–832.
Künsch, H. (1989). The jacknife and the bootstrap for general stationary observations. Annals of Statistics, 17(3), 1217–1241.
Liebl, D., Reimherr, M. (2019). Fast and fair simultaneous confidence bands for functional parameters. arXiv:1910.00131.
Panaretos, V. M., Kraus, D., Maddocks, J. H. (2010). Second-order comparison of Gaussian random functions and the geometry of DNA minicircles. Journal of the American Statistical Association, 105(490), 670–682.
Paparoditis, E., Sapatinas, T. (2016). Bootstrap-based testing of equality of mean functions or equality of covariance operators for functional data. Biometrika, 103(3), 727–733.
Pigoli, D., Aston, J. A. D., Dryden, I. L., Secchi, P. (2014). Distances and inference for covariance operators. Biometrika, 101(2), 409–422.
Pilavakis, D., Paparoditis, E., Sapatinas, T. (2020). Testing equality of autocovariance operators for functional time series. Journal of Time Series Analysis, 41, 571–589.
Politis, D., Romano, J. (1994). The stationary bootstrap. Journal of the American Statistical Association, 89(428), 1303–1313.
Ramsay, J. O., Silverman, B. W. (2005). Functional data analysis (second). New York: Springer.
Sharipov, O. S., Wendler, M. (2020). Bootstrapping covariance operators of functional time series. Journal of Nonparametric Statistics, 32(3), 648–666.
Stoehr, C., Aston, J. A. D., Kirch, C. (2019). Detecting changes in the covariance structure of functional time series with application to fMRI data. arXiv:1903.00288.
Van der Vaart, A. W., Wellner, J. A. (1996). Weak convergence and empirical processes: With applications in statistics. New York: Springer.
Zhang, X., Shao, X. (2015). Two sample inference for the second-order property of temporally dependent functional data. Bernoulli, 21(2), 909–929.
Acknowledgements
This research was partially supported by the Collaborative Research Center “Statistical modeling of nonlinear dynamic processes” (Sonderforschungsbereich 823, Teilprojekt A1, C1) and the Research Training Group “High-dimensional phenomena in probability - fluctuations and discontinuity” (RTG 2131). The authors are grateful to Christina Stoehr for sending us the results of Stoehr et al. (2019) and to Martina Stein, who typed parts of this manuscript with considerable technical expertise. The authors are also grateful to the referees for their constructive comments on an earlier version of this paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: Proofs of main results
Appendix: Proofs of main results
1.1 A.1: Proof of Theorem 1
We apply the central limit theorem as formulated in Theorem 2.1 in Dette et al. (2020a) to the sequence of \(C(T^2)\)-valued random variables \(((Z_{j} - \mu )^{\check{\otimes }2})_{j\in {\mathbb {N}}} = (\eta _{j}^{\check{\otimes }2})_{j\in {\mathbb {N}}}\).
It can be easily seen that conditions (A1), (A2) and (A4) in this reference are satisfied. In order to see that the remaining condition (A3) also holds, we use the triangle inequality and Assumption 1 of the present work to obtain, for any \(j\in {\mathbb {N}}\) and \(s,t,s^\prime ,t^\prime \in T\),
where \(\mathbb {E}\big [ (\Vert \eta _{j}\Vert _\infty \, M )^J \big ] \le \tilde{K} < \infty \) by (A3). Now observe that
Here the error \(o_{\mathbb {P}}(1)\) refers to the supremum norm, because by Theorem 2.1 in Dette et al. (2020a) the sequence \(\big (\frac{1}{\sqrt{n}} \sum _{j=1}^n \eta _{j} \big )_{n \in {\mathbb {N}}} \) converges weakly in C([0, 1]) and by continuous mapping \( \Vert \frac{1}{\sqrt{n}} \sum _{j=1}^n \eta _{j}^{\check{\otimes }2} \Vert _\infty \) is of order \( O_{\mathbb {P}}(1)\), which yields \(\frac{1}{\sqrt{n}}\Vert (\frac{1}{\sqrt{n}} \sum _{j=1}^n \eta _{j})^{\check{\otimes }2} \Vert _\infty = o_{{\mathbb {P}}} (1)\). Moreover, as shown above, Theorem 2.1 in Dette et al. (2020a) can also be applied to the sequence \(( \eta _{j}^{\check{\otimes }2} )_{j\in {\mathbb {N}}}\), which yields the claim of Theorem 1. \(\square \)
1.2 A.2: Proof of Proposition 1
As the samples are independent, it directly follows from Theorem 1 that
in \(C([0,1]^2)^2\) as \(m,n\rightarrow \infty \), where \(Z_1\) and \(Z_2\) are independent, centred Gaussian processes defined by their long-run covariance operators (14) and (15). By the continuous mapping theorem it follows that
in \(C([0,1]^2)\) as \(m,n\rightarrow \infty \) (the error \(o_{\mathbb {P}}(1)\) refers again to the supremum norm), where Z is again a centred Gaussian process with covariance operator (13).
If \(d_\infty = 0\), the convergence in (50) together with the continuous mapping yield (12). If \(d_\infty > 0\), the asymptotic distribution of \({\hat{d}}_\infty \) can be deduced from Theorem B.1 in the online supplement of Dette et al. (2020a) or alternatively from the results in Cárcamo et al. (2020). \(\square \)
1.3 A.3: Proof of Theorem 2 and 3
Proof of Theorem 2. Using similar arguments as in the proof of Theorem 1, it follows that the process \( {\hat{B}}^{(r)}_{m,n}\) in (18) admits the stochastic expansion
where the remainder is defined by \(R_{m,n}=R^{(1)}_m - R^{(2)}_n\) with
Because both terms have a similar structure, we consider only the first one. Note that it is easy to see that \(\Vert \bar{\eta }_1^{\check{\otimes }2}\Vert = O_{\mathbb {P}}(\frac{1}{m})\) and therefore the third term in (51) is of order \(O_{\mathbb {P}} \big (\sqrt{\frac{l_1}{m}}\big ) = o_{\mathbb {P}}(1)\). The first and second term can be treated in the same way and we only consider the first one. It follows from the proof of Theorem 4.3 in Dette et al. (2020a) that the term
is of order \(O_{\mathbb {P}}(1) \). By Theorem 2.1 in the same reference the second factor of the tensor satisfies \(\Vert \bar{\eta }_1\Vert _\infty = O_\mathbb {P}(\frac{1}{\sqrt{m}})\), and therefore the first term in (51) is of order \(o_\mathbb {P}(1)\). Using similar arguments for the second term in (51) and the summand \(R^{(2)}_n\) yields
Next, note that the sequences \((\eta _{1,j}^{\check{\otimes }2})_{j\in {\mathbb {N}}}\) and \((\eta _{2,j}^{\check{\otimes }2})_{j\in {\mathbb {N}}}\) satisfy Assumption 2.1 in Dette et al. (2020a).
Thus, similar arguments as in the proof of Theorem 3.3 and 4.3 in the same reference yield
in \(C([0,1]^2)^{R+1}\) as \(m,n \rightarrow \infty \) where the process \(Z_{m,n}\) is defined in (50) and the random functions \(Z^{(1)},\ldots ,Z^{(R)}\) are independent copies of Z which is also defined in (50). Note that in this paper the authors prove weak convergence of a vector in \(C([0,1])^{R+1}\). The proof of weak convergence of the finite dimensional distributions can be directly transferred to vectors in \(C([0,1]^2)^{R+1}\) , while the proof of equicontinuity requires condition (A1) in Assumption 1, which reduces for the space \(C([0,1]^2)\) to (6). If \(d_\infty = 0\), the continuous mapping theorem implies
in \({\mathbb {R}}^{R+1}\) as \(m,n \rightarrow \infty \) where the statistic \(\hat{d}_\infty \) is defined by (11), the bootstrap statistics \(T_{m,n}^{(1)},\ldots ,T_{m,n}^{(R)}\) are defined by (19) and the random variables \(T^{(1)},\ldots ,T^{(R)}\) are independent copies of T which is defined by (12). Now, Lemma 4.2 in Bücher and Kojadinovic (2019) directly implies (21), that is,
For the application of this result, it is required that the distribution of the random variable T has a continuous distribution function, which follows from Gaenssler et al. (2007). In order to show the consistency of test (20) in the case \(d_\infty >0\), write
and note that, given (54) and (16), the assertion in (22) follows by simple arguments. \(\square \)
Proof of Theorem 3. First note that the same arguments as in the proof of Theorem 3.6 in Dette et al. (2020a) show that the estimators of the extremal sets defined by (23) are consistent that is
where \(d_H\) denotes the Hausdorff distance. Thus, given the convergence in (53), the arguments in the proof of Theorem 3.7 in the same reference yield
in \({\mathbb {R}}^{R+1}\) as \(m,n \rightarrow \infty \) where the statistic \(\hat{d}_\infty \) is defined by (11), the bootstrap statistics \(K_{m,n}^{(1)},\ldots ,K_{m,n}^{(R)}\) are defined by (24) and the random variables \(T^{(1)}({\mathcal {E}}),\ldots ,T^{(R)}({\mathcal {E}})\) are independent copies of \(T(\mathcal {E})\) which is defined by (16). Note that this convergence holds true under the null and the alternative hypothesis.
If \(\varDelta = d_\infty \), Lemma 4.2 in Bücher and Kojadinovic (2019) directly implies (26) and again the results in Gaenssler et al. (2007) ensure that the limit \(T(\mathcal {E})\) has a continuous distribution function.
If \(\varDelta \ne d_\infty \), write
Then, it follows from (55) and simple arguments that, for any \(R\in {\mathbb {N}}\),
if \(\varDelta > d_\infty \) and \(\varDelta < d_\infty \), respectively. This proves the remaining assertions of Theorem 3. \(\square \)
1.4 A.4: Proof of Proposition 2
Let \(C_{n,j}\) denote the covariance operator of \(X_{n,j}\) defined by \(C_{n,j}(s,t) = \text {Cov}(X_{n,j}(s),X_{n,j}(t))\) and consider the sequential process
which is an element of \(C([0,1], C([0,1]^2))\). Here the order \( o_{\mathbb {P}}(1)\) for the remainder is obtained by similar arguments as given at the beginning of the proof of Theorem 2 and the details are omitted for the sake of brevity. Note that \(\{\hat{\mathbb {V}}_n(s)\}_{s\in [0,1]}\) can equivalently be regarded as an element of \(C([0,1]^3)\) and we have the representation
where the processes \( \tilde{\mathbb {V}}_{1,n}, \tilde{\mathbb {V}}_{2,n} \in C([0,1]^3)\) are defined by
(\(s,t,u\in [0,1]\)) and
Recall the definition of the array (\(\tilde{\eta }_{n,j} :n\in {\mathbb {N}}, j = 1,\ldots , n\)) in (27). By Theorem 2.2 in Dette et al. (2020a) it follows that
in \(C([0,1]^3)\), where \(\mathbb {V}_l\) is a centred Gaussian measure on \(C([0,1]^3)\) characterized by the covariance operator
and the long-run covariance operator \(\mathbb C_l\) is defined in (32). From the continuous mapping theorem, we obtain
in \(C([0,1]^3)\), where \(\tilde{\mathbb {V}}_1, \tilde{\mathbb {V}}_2\) are centred Gaussian measures on \(C([0,1]^3)\) characterized by
with covariance operators
In the following we will show the weak convergence
in \(C([0,1]^3)\) as \(n\rightarrow \infty \), where \(\mathbb {V}\in C([0,1]^3)\) is a centred Gaussian random variable characterized by its covariance operator
and the long-run covariance operators \(\mathbb {C}_1, \mathbb {C}_2\) are defined by (32). The convergence in (57) implies that the processes \(\tilde{\mathbb {V}}_{1,n}, \tilde{\mathbb {V}}_{2,n}\) are asymptotically tight and the representation in (56) yields that \(\hat{\mathbb {V}}_{n}\) is asymptotically tight as well (see Section 1.5 in Van der Vaart and Wellner 1996). In order to prove the convergence in (58), it consequently remains to show the convergence of the finite-dimensional distributions. For this, we utilize the Crámer–Wold device and show that
for any \((s_1,t_1,u_1),\dots ,(s_q,t_q,u_q) \in [0,1]^3\), \(c_1,\ldots ,c_q \in {\mathbb {R}}\) and \(q\in {\mathbb {N}}\). Asymptotic normality of \(\tilde{Z}_n\) can be proved by the same arguments as in the proof of Theorem 2.1 in Dette et al. (2020a), and it remains to show that the variance of the random variable \(\tilde{Z}_n\) converges to the variance of \(\tilde{Z}\). Using (3.17) in Dehling and Philipp (2002) and assumptions (A2) and (A4) we obtain for any \((s,t,u),(s^\prime , t^\prime , u^\prime ) \in [0,1]^3\)
where the symbol “\(\lesssim \)” means less or equal up to a constant independent of n, and \(\Vert X\Vert _2 = \mathbb {E}[X^2]^{1/2}\) denotes the \(L^2\)-norm of a real-valued random variable X (also note that we implicitly assume \(\sum _{i=j}^k a_i = 0\) if \(k<j\)). Furthermore, assuming without loss of generality that \(s \le s^\prime \), we have
where the last equality follows by the same arguments as used in (59). For the remaining expression we use the dominated convergence theorem to obtain
which means that for any \((s,t,u),(s^\prime , t^\prime , u^\prime ) \in [0,1]^3\)
By similar arguments we obtain
and therefore we have
which finally proves (58).
Next we define the \(C([0,1]^3)\)-valued process
then the convergence in (58) and the continuous mapping theorem yield
in \(C([0,1]^3)\), where \(\mathbb {W}\) is centred Gaussian defined by \( \mathbb {W}(s,t,u)= \mathbb {V}(s,t,u) - s \mathbb {V}(1,t,u) \) with covariance operator given by (31). Finally, recall the definition of the process \((\hat{\mathbb {U}}_{n}:n\in \mathbb {N})\) in (28) and note that, in contrast to \(\hat{\mathbb {W}}_n\), this process is not centred. Consequently, if \(d_\infty = 0\), we have \(\sqrt{n}\, \mathbb {U}_n = \hat{\mathbb {W}}_n\) and the convergence in (61) and the continuous mapping theorem directly yield (30).
If \(d_\infty > 0\), assertion (33) is a consequence of the weak convergence in (61) and Theorem B.1 in the online supplement of Dette et al. (2020a) and also of the results in Cárcamo et al. (2020). \(\square \)
1.5 A.5: Proof of Theorem 4 and 5
Proof of Theorem 4. Recalling the definition of the bootstrap processes in (34) it can be shown by similar arguments as given at beginning of the proof of Theorem 2 that
(for \(r = 1,\dots , R\)), where
\( \tilde{Y}_{n,j} = \tilde{\eta }_{n,j}^{\check{\otimes }2}(t,u) - (\hat{C}_2 - \hat{C}_1) \mathbb {1}\{j > \lfloor \hat{s}n \rfloor \} \) (\(j=1,\dots ,n\)). The array \((\tilde{\eta }_{n,j}^{\check{\otimes }2} \, :n\in {\mathbb {N}}, ~ j = 1,\ldots , n)\) satisfies (A1), (A3) and (A4) of Assumption 2.1 in Dette et al. (2020a). The convergence in (61) and similar arguments as in the proof of Theorem 4.3 in the same reference show
in \(C([0,1]^3)^{R+1}\) as \(n \rightarrow \infty \), where the process \(\mathbb {V}\) is defined in (57) and \(\mathbb {V}^{(1)},\dots ,\mathbb {V}^{(R)}\) are independent copies of \(\mathbb {V}\). For the sake of completeness we repeat the necessary main steps here, which are proved using analogous arguments as given in Dette et al. (2020a). First we define \( {Y}_{n,j}(t,u) = \tilde{\eta }_{n,j}^{\check{\otimes }2}(t,u) - ({C}_2 - {C}_1) \mathbb {1}\{j > \lfloor {s}^{*} n \rfloor \} \) and show the approximation
where the process \( \bar{C}_n^{(r)}\) is defined by
In a second step we show
where the process \( \tilde{C}_n^{(r)}\) is defined by
In a third step one notes that
and shows the weak convergence
in \(C([0,1]^3)^{R+1}\) as \(n \rightarrow \infty \), where the process \(\mathbb {V}\) is defined in (57) and \(\mathbb {V}^{(1)},\ldots ,\mathbb {V}^{(R)}\) are independent copies of \(\mathbb {V}\). Observing (62), (64) and (65) then proves the weak convergence in (63) Finally, this result and the continuous mapping theorem yield
in \(C([0,1]^3)^{R+1}\) as \(n \rightarrow \infty \) where the process \(\hat{\mathbb {W}}_n\) is defined by (60), the bootstrap counterparts \(\hat{\mathbb {W}}_{n}^{(1)},\ldots ,\hat{\mathbb {W}}_{n}^{(R)}\) are defined by (36) and the random variables \(\mathbb {W}^{(1)},\ldots ,\mathbb {W}^{(R)}\) are independent copies of \(\mathbb {W}\) which is defined by its covariance operator (31).
If \(d_\infty = 0\), the continuous mapping theorem directly implies
in \({\mathbb {R}}^{R+1}\) as \(n \rightarrow \infty \) where the statistic \(\hat{\mathbb {M}}_n\) is defined by (29), the bootstrap statistics \(\check{T}_{n}^{(1)},\ldots ,\check{T}_{n}^{(R)}\) are defined by (37) and the random variables \(\check{T}^{(1)},\ldots ,\check{T}^{(R)}\) are independent copies of the random variable \(\check{T}\) defined by (30). Now the same arguments as in the discussion starting from Eq. (54) imply the assertions of Theorem 4. \(\square \)
Proof of Theorem 5. We first mention that it follows by similar arguments as given in the proof of Theorem 4.2 in Dette et al. (2020a) that the estimator of the unknown change location defined by (35) satisfies
whenever \(d_\infty >0\). Whenever \(d_\infty =0\), suppose that the estimate \(\hat{s}\) converges weakly to a \([\vartheta ,1-\vartheta ]\)-valued random variable which is denoted by \(s_{\max }\). Then, if \(d_\infty > 0\), the convergence in (33) and Slutsky’s theorem yield
where \(\tilde{D} (\mathcal {E})\) is the same as in (33) and the statistic \(\hat{d}_\infty \) is defined by (39).
The same arguments as in the proof of Theorem 3.6 in Dette et al. (2020a) again yield that the estimators of the extremal sets defined by (40) are consistent. The convergence in (66) and similar arguments as in the proof of Theorem 4.4 in the same reference then yield
in \({\mathbb {R}}^{R+1}\) as \(n \rightarrow \infty \) where the bootstrap statistics \(\check{K}_{n}^{(1)},\ldots ,\check{K}_{n}^{(R)}\) are defined by (41) and the random variables \(D^{(1)}(\mathcal {E}),\ldots ,D^{(R)}(\mathcal {E})\) are independent copies of \(D(\mathcal {E})\) which is defined by (67). The convergence in the preceding equation holds true under the null and the alternative hypothesis and now the same arguments as in the discussion starting from Eq. (55) imply the assertions made in Theorem 5. \(\square \)
About this article
Cite this article
Dette, H., Kokot, K. Detecting relevant differences in the covariance operators of functional time series: a sup-norm approach. Ann Inst Stat Math 74, 195–231 (2022). https://doi.org/10.1007/s10463-021-00795-2
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-021-00795-2