Abstract
We propose a functional single-index model (FSiM) to study the link between a scalar response variable and multiple functional predictors, in which the mean of the response is related to the linear predictors via an unknown link function. The FSiM serves as a good tool for dimension reduction in regression with multiple predictors and it is more flexible than functional linear models. Assuming that the functional predictors are observed at discrete points, we use B-spline basis functions to estimate the slope functions and the link function based on the least-squares criterion, and propose an iterative estimating procedure. Moreover, we provide uniform convergence rates of the proposed spline estimators in the FSiM, and construct asymptotic simultaneous confidence bands for the slope functions for inference. Our proposed method is illustrated by simulation studies and by an analysis of a diffusion tensor imaging data application.
Similar content being viewed by others
References
Bunea, F., Ivanescu, A.E., Wegkamp, M.H. (2011). Adaptive inference for the mean of a Gaussian process in functional data. Journal of the Royal Statistical Society Series B, 531–558.
Cai, T., Hall, P. (2006). Prediction in functional linear regression. Annals of Statistics, 34, 2159–2179.
Cardot, H., Ferraty, F., Sarda, P. (2003). Spline estimators for the functional linear model. Statistica Sinica, 13, 571–591.
Carroll, R. J., Fan, J., Gijbels, I., Wand, M. P. (1997). Generalized partially linear single-index models. Journal of the American Statistical Association, 92, 477–489.
Chen, D., Hall, P., Müller, H. G. (2011). Single and multiple index functional regression models with nonparametric link. Annals of Statistics, 39, 1720–1747.
Claeskens, G., Krivobokova, T., Opsomer, J. D. (2009). Asymptotic properties of penalized spline estimators. Biometrika, 96, 529–544.
Crainiceanu, C. M., Staicu, A.-M., Ray, S., Punjabi, N. (2012). Bootstrap-based inference on the difference in the means of two correlated functional processes. Statistics in Medicine, 31, 3223–3240.
Csőrgő, M., Révész, P. (1981). Strong approximations in probability and statistics. New York: Academic Press.
de Boor, C. (2001). A practical guide to splines. New York: Springer.
Demko, S. (1986). Spectral bounds for \( \left|a^{-1}\right|_{\infty }\). Journal of Approximation Theory, 48, 207–212.
DeVore, R. A., Lorentz, G. G. (1993). Constructive approximation. Berlin: Springer-Verlag.
Gertheiss, J., Maity, A., Staicu, A.-M. (2013). Variable selection in generalized functional linear models. Stat, 2, 86–101.
Goldsmith, J., Bobb, J., Crainiceanu, C. M., Caffo, B., Reich, D. (2011a). Penalized functional regression. Journal of Computational and Graphical Statistics, 20, 830–851.
Goldsmith, J., Crainiceanu, C. M., Caffo, B., Reich, D. (2011b). Penalized functional regression analysis of white-matter tract profiles in multiple sclerosis. NeuroImage, 57, 431–439.
Goldsmith, J., Crainiceanu, C., Caffo, B., Reich, D. (2012). Longitudinal penalized functional regression for cognitive outcomes on neuronal tract measurements. Journal of the Royal Statistical Society Series C, 61, 453–469.
Hall, P., Horowitz, J. L. (2007). Methodology and convergence rates for functional linear regression. Annals of Statistics, 35, 70–91.
Huang, J. (2003). Local asymptotics for polynomial spline regression. Annals of Statistics, 31, 1600–1635.
Huang, J. Z., Wu, C. O., Zhou, L. (2004). Polynomial spline estimation and inference for varying coefficient models with longitudinal data. Statistica Sinica, 14, 763–788.
Ichimura, H. (1993). Semiparametric least squares (SLS) and weighted SLS estimation of single index models. Journal of Econometrics, 58, 71–120.
Jiang, C. R., Wang, J. L. (2011). Functional single index models for longitudinal data. Annals of Statistics, 39, 362–388.
Li, Y., Hsing, T. (2007). On rates of convergence in functional linear regression. Journal of Multivariate Analysis, 98, 1782–1804.
Ma, S., Yang, L., Carroll, R. (2012). Simultaneous confidence band for sparse longitudinal regression. Statistica Sinica, 22, 95–122.
McLean, M. W., Hooker, G., Staicu, A.-M., Scheipl, F., Ruppert, D. (2014). Functional generalized additive models. Journal of Computational and Graphical Statistics, 23, 249–269.
Müller, H. G., Stadtmüller, U. (2005). Generalized functional linear models. Annals of Statistics, 33, 774–805.
Ramsay, J., Silverman, B. W. (2005). Functional data analysis. Springer series in statistics (2nd ed.). New York: Springer.
Wu, Y. (2010). The partially monotone tensor spline estimation of joint distribution function with bivariate current status data. In Technical report, University of Iowa.
Xia, Y., Tong, H., Li, W. K., Zhu, L. (2002). An adaptive estimation of dimension reduction space (with discussion). Journal of the Royal Statistical Society Series B, 64, 363–410.
Xue, L., Qu, A., Zhou, J. (2010). Consistent model selection for marginal generalized additive model for correlated data. Journal of American Statistical Association, 105, 1518–1530.
Xue, L., Yang, L. (2006). Additive coefficient modeling via polynomial spline. Statistica Sinica, 16, 1423–1446.
Yao, F., Müller, H. G., Wang, J. L. (2005). Functional linear regression analysis for longitudinal data. The Annals of Statistics, 33, 2873–2903.
Zhou, S., Shen, X., Wolfe, D. A. (1998). Local asymptotics for regression splines and confidence regions. The Annals of Statistics, 26, 1760–1782.
Acknowledgments
The author thanks the Editor, the Associate Editor and the two referees for their insightful comments and suggestions that lead to substantially improve the paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
Ma’s research was partially supported by NSF grant DMS 1306972.
Appendix
Appendix
For any positive numbers \(a_{n}\) and \(b_{n}\), let \(a_{n}\sim b_{n}\) denote that lim\(_{n \rightarrow \infty }a_{n}/b_{n}=1\). For any vector \(\zeta = ( \zeta _{1},\ldots ,\zeta _{s} ) ^{\mathrm{T}}\in R^{s}\), denote its \(L_{r}\) norm as \( \Vert \zeta \Vert _{r}= ( \vert \zeta _{1} \vert ^{r}+\cdots \vert \zeta _{s} \vert ^{r} ) ^{1/r}\). For any symmetric matrix \(\mathbf{A} _{s\times s}\), denote its \(L_{r}\) norm as \( \Vert \mathbf{A} \Vert _{r}=\max _{\zeta \in s,\zeta \ne 0} \Vert \mathbf{A\zeta } \Vert _{r} \Vert \zeta \Vert _{r}^{-1}\). For any matrix \(\mathbf{A}= ( A_{ij} ) _{i=1,j=1}^{s,t}\), denote \( \Vert \mathbf{A} \Vert _{\infty }=\max _{1\le i\le s}\sum \nolimits _{j=1}^{t} \vert A_{ij} \vert \). The estimator \( \widehat{\dot{g}} ( u;\delta _{n} ) \) can be rewritten as \( \widehat{\dot{g}} ( u;\delta _{n} ) =\mathbf{B} _{1}^{q-1} ( u ) ^{\mathrm{T}}\mathbf{D}_{1}\widehat{\lambda } ( \delta _{n} ) \), where
are B-spline functions with order \(q-1\), and
Proof of Proposition 1
By (6), \(\widehat{\lambda } ( \delta _{n} ) \) can be decomposed as \(\widehat{\lambda } ( \delta _{n} ) =\widehat{\lambda }_{\varepsilon } ( \delta _{n} ) +\widehat{\lambda }_{g} ( \delta _{n} ) \), where
in which \(\varepsilon _{n}= ( \varepsilon _{1},\ldots ,\varepsilon _{n} ) ^{\mathrm{T}}\) and \(\mathbf{g}_{n}= \{ g ( \int \nolimits _{\mathcal {T}}\beta ( t ) ^{\mathrm{T}}\mathbf{X}_{i} ( t ) \mathrm{d}t ) ,1\le i\le n \} ^{\mathrm{T}}\). Correspondingly, \(\widehat{g} ( u;\delta _{n} ) \) is decomposed into \(\widehat{g} ( u;\delta _{n} ) =\widehat{g} _{\varepsilon } ( u;\delta _{n} ) +\widehat{g}_{g} ( u; \delta _{n} ) \), where \(\widehat{g}_{\varepsilon } ( u; \delta _{n} ) =\mathbf{B}_{1} ( u ) ^{\mathrm{T}} \widehat{\lambda }_{\varepsilon } ( \delta _{n} ) \) and \(\widehat{g}_{g} ( u;\delta _{n} ) =\mathbf{B} _{1} ( u ) ^{\mathrm{T}}\widehat{\lambda }_{g} ( \delta _{n} ) \). Thus,
where
By Conditions (C3) and (C4), we have that for all \(1\le i\le n\), there exists a constant \(0<C<\infty \) such that
and by (8) there exist constants \(0<C^{\prime },C^{\prime \prime }<\infty \) such that
Moreover, by (8) for all \(1\le i\le n\), there exists a constant \(0<C^{\prime \prime \prime }<\infty \) such that
By Theorem 5.4.2 of DeVore and Lorentz (1993) and Berstein’s inequality in Boor (2001), one has for large enough \(n\), there are constants \(0<c_{1}<C_{1}<\infty \), such that
with probability approaching \(1\), for \(J_{n,1}\log ( n ) /n=o ( 1 ) \),
and thus
By the above result and Demko (1986), it can be proved that with probability approaching \(1\) and for large enough \(n\),
for some constant \(0<C_{2}<\infty \). Therefore, by (14), (15), (16), (17) and (19), we have
Thus, we have
Let \(\mathbb {X=} ( \mathbf{X}_{1},\ldots ,\mathbf{X}_{n} ) \). By Condition (C5) and (18) for every \(u\in \mathcal {I}\), \(E \{ \widehat{g}_{\varepsilon } ( u;\delta _{n} ) \vert \mathbb {X} \} =0\), and
Thus, it can be proved by Berstein’s inequality in Boor (2001) that \( \sup \nolimits _{u\in \mathcal {I}} \vert \widehat{g}_{\varepsilon } ( u;\delta _{n} ) \vert =O_{p} ( \sqrt{\log ( n ) J_{n,1}n^{-1}} ) \). Therefore, we have
Result (i) is proved by the above result and (8). It is easy to prove that \( \Vert \mathbf{D}_{1} \Vert _{\infty }=O ( J_{n,1} ) \), where \(\mathbf{D}_{1}\) is defined in (13). Following the similar reasoning as the proof for \(\widehat{g} ( u; \delta _{n} ) \), the result in (ii) can be proved. \(\square \)
Lemma 1
Under Condition (C3), we have that there exists \( \widetilde{\delta }_{1,n}^{0}=\mathbf{(}\widetilde{\delta } _{r1,n}^{0}:1\le r\le J_{n,2})^{\mathrm{T}}\in R^{J_{n,2}}\) with \(\widetilde{ \delta }_{11,n}^{0}\le \cdots \le \widetilde{\delta }_{J_{n,2}1,n}^{0}\) such that \(\sup _{t\in \mathcal {T}} \vert \beta _{1} ( t ) - \widetilde{\beta }_{1,n} ( t ) \vert =O(J_{n,2}^{-\alpha })\), where \(\widetilde{\beta }_{1,n} ( t ) =\mathbf{B}_{2} ( t ) ^{\mathrm{T}}\widetilde{\delta }_{1,n}^{0}\).
Proof
By choosing \(\epsilon _{1}<\cdots <\epsilon _{J_{n,2}}\) , we define
which is monotone nondecreasing function in \(t\). By the fact that for \(t\) in \( [ \upsilon _{r_{1}},\upsilon _{r_{1}+1} ) \), \(\sum \nolimits _{r= \upsilon _{r_{1}}+1-q}^{\upsilon _{r_{1}}}B_{r,2}(t)=1\), we have
for \(\widetilde{t}\in [ \upsilon _{r_{1}},\upsilon _{r_{1}+1} ) \), and thus
Let \(h=\max _{q\le l\le J_{n,2}}(\upsilon _{r+1}-\upsilon _{r})\). Define
Then, \(\omega (\beta _{1};h)\) is a monotone and subadditivity function of \(h\) , that is, \(\omega (\beta _{1};h)\le \omega (\beta _{1};h_{1}+h_{2})\le \omega (\beta _{1};h_{1})+\omega (\beta _{1};h_{2})\) for \(h_{1}>0\) and \( h_{2}>0\). See Lemma 2.19 of Wu (2010) for the detailed proof. We choose \( \epsilon _{r}=\upsilon _{1}+(r-1)(\upsilon _{q+1}-\upsilon _{q})/q\) for \( r=1,\ldots ,q\) and \(\epsilon _{r}=\upsilon _{r}\) for \(r=q+1,\ldots ,J_{n,2}\) to guarantee that \(\epsilon _{r+1}-\epsilon _{r}>0\). Then, we have \( \vert \epsilon _{r}-\upsilon _{r} \vert \le h\) for \(r=1,\ldots ,J_{n,2}\). Moreover, for \(\widetilde{t}\in [ \upsilon _{r_{1}},\upsilon _{r_{1}+1} ) \) and \(r_{1}-q\le r\le r_{1}\), \( \vert \widetilde{t} -\epsilon _{r} \vert \le (q+1)h\). Therefore, we have
The last step follows from the subadditivity of \(\omega (\beta _{1};h)\). Let
Denote \(d(\beta _{1},G_{q})\) as the distance of \(\beta _{1}\) from \(G_{q}\). Following the reasoning as given in Lemma 2.19 of Wu (2010), it can be shown that for any \(g\in G_{q}\), we have
for some constant \(0<c<\infty \), and thus
where \(G_{q-1,q}= \{ \partial g/\partial t,g\in G_{q} \} \). Proceeding in this way, we can derive
Thus, the result in Lemma 1 follows from the above result and Condition (C3).\(\square \)
Lemma 2
Let \(\widehat{\delta }\) be the minimizer of \( \widetilde{L}_{n} ( \delta ) \) given in (7 ) subject to \(\delta _{1,1}\le \cdots \le \delta _{J_{n,2},1}\) satisfying \( \Vert \widehat{\delta }-\delta _{n}^{0}\Vert _{\infty }\le a_{n}\) with probability approaching \(1\), where \(\widetilde{ \delta }_{n}^{0}= ( \widetilde{\delta }_{1,n}^{0\mathrm{T }},\ldots ,\delta _{p,n}^{0\mathrm{T}} ) ^{\mathrm{T}}\), under the assumptions in Theorem 1, we have
Proof
Let \(\widehat{\delta }_{n}\) be the minimizer of \(\widetilde{L} _{n}( \delta ) \) and \(\Vert \widehat{\delta }_{n}-\delta _{n}^{0}\Vert _{\infty }\le a_{n}\). By Taylor’s expansion, we have
Moreover,
where
and
By Berstein’s inequality Boor (2001), it can be proved that \(\Vert \varTheta _{1}\Vert _{\infty }=O_{p}( ( \log n)\) \( ^{1/2}n^{1/2}J_{n,2}^{-1/2}) \). Next, we will show that \(\Vert \varTheta _{2}\Vert _{\infty }=o_{p}( n^{1/2}J_{n,2}^{-1/2}) \). By Proposition 1 and the assumption in Theorem 1, we have
By the law of large numbers, we have \(\sum \nolimits _{i=1}^{n}\Vert \varPsi _{i}\Vert _{\infty }=O_{p}( n^{1/2}J_{n,2}^{-1/2}) \). Therefore, \(\Vert \varTheta _{2}\Vert _{\infty }=o_{p}( n^{1/2}J_{n,2}^{-1/2}) \). Thus, we have \(\Vert -L_{n}( \widetilde{\delta }_{n}^{0}) /\partial \delta \Vert _{\infty }=O_{p}( ( \log n)\) \( ^{1/2}n^{1/2}J_{n,2}^{-1/2}) \). Moreover,
Therefore, we have \( \Vert \widehat{\delta }_{n}-\widetilde{ \delta }_{n}^{0} \Vert _{\infty }=O_{p} ( ( \log n ) ^{1/2}n^{-1/2}J_{n,2}^{1/2} ) \). Since \(\widetilde{\delta } _{11,n}^{0}\le \cdots \le \widetilde{\delta }_{J_{n,2}1,n}^{0}\), then with probability approaching \(1\), \(\widehat{\delta }_{n}=\widehat{ \delta }\). \(\Box \)
Lemma 3
Under the assumptions in Theorem 1,
Proof
By (6), we have \(\widehat{\lambda } ( \delta ) = \{ \mathcal {B} ( \delta ) ^{\mathrm{T} }\mathcal {B} ( \delta ) \} ^{-1}\mathcal {B} ( \delta ) ^{\mathrm{T}}\mathbf{Y}_{n}\), where \(\mathbf{Y} _{n}= ( Y_{1},\ldots ,Y_{n} ) ^{\mathrm{T}}\). Thus,
where
Let
Following similar reasoning as the proofs in Proposition 1, it can be shown that
Denote
By (19) and Berstein’s inequality in Boor (2001), we have
and thus with probability approaching \(1\), \(\sup _{1\le s\le J_{n,1}} \vert \dot{A}_{s} ( \widetilde{\delta } _{n}^{0} ) \vert \le C\) for some constant \(0<C<\infty \) by the fact that \(B_{s,1} ( u ) \) and \(\dot{B}_{s,1} ( u ) \) are functions with values bounded between \(0\) and \(1\). Hence, with probability approaching \(1\),
for some constant \(0<C^{\prime }<\infty \). By B-spline properties, we have \( \sum \nolimits _{1\le s\le J_{n,1}} \) \(\vert B_{s,1} ( \varPhi _{i}^{\mathrm{T}}\widetilde{\delta }_{n}^{0} ) \vert =O(1)\). Therefore,
Moreover, by Condition (C3), for every \(t_{ij}\in \mathcal {T}\), there exists \(\zeta _{n,k} ( t_{ij} ) \in R^{J_{n,1}}\) such that \( \vert E \{ X_{i,k} ( t_{ij} ) \vert U_{i} \} -\mathbf{B}_{1} ( U_{i} ) ^{\mathrm{T}}\zeta _{n,k} ( t_{ij} ) \vert =O ( J_{n,1}^{-1} ) \), and thus for every \(s\) and \(k\),
Let
where \(\mathcal {B}= [ \{ \mathbf{B}_{1} ( U_{1} ) ,\ldots , \mathbf{B}_{1} ( U_{n} ) \} ^{\mathrm{T}} ] _{n\times J_{n,1}}\). By (24) and Berstein’s inequality, we have
and thus
Furthermore, it can be proved that \( \Vert \widehat{\varOmega }_{1} ( \widetilde{\delta }_{n}^{0} ) ^{\mathrm{T}}-\widetilde{\varOmega } _{1}^{\mathrm{T}} \Vert _{\infty }=O ( J_{n,2}^{-\alpha }+m_{\min }^{-1} ) \). Therefore, we have
By (21), (22), (23) and (25), we have
Let
By the above result and Proposition (1), we have
where
In the following, we will prove that \( \Vert \varTheta _{i} \Vert _{\infty }=o_{p} ( n^{1/2}J_{n,2}^{-1/2} ) \) for \(i=2,3,4\). By (8), we have \( \vert \widehat{g} ( \varPhi _{i}^{\mathrm{T}} \delta _{n}^{0},\delta _{n}^{0} ) -\widehat{g} ( U_{i} ) \vert =O ( J_{n,2}^{-\alpha }+m_{\min }^{-1} ) \). Moreover, we have \( \Vert \varPhi _{i} \Vert _{\infty }=O ( J_{n,2}^{-1}+m_{\min }^{-1} ) \). Thus,
where \(\widetilde{\varTheta }_{2}=\widetilde{\varTheta }_{12}+\widetilde{\varTheta } _{22}\),
in which \(\widehat{g}_{g} ( U_{i} ) =\mathbf{B}_{1} ( U_{i} ) ^{\mathrm{T}} ( \mathcal {B}^{\mathrm{T}}\mathcal {B} ) ^{-1} \mathcal {B}^{\mathrm{T}}\mathbf{g}_{n}\) and \(\widehat{g}_{g} ( U_{i} ) =\mathbf{B}_{1} ( U_{i} ) ^{\mathrm{T}} ( \mathcal {B} ^{\mathrm{T}}\mathcal {B} ) ^{-1}\mathcal {B}^{\mathrm{T}}\varepsilon _{n}\). By law of large numbers and \( \vert g ( U_{i} ) - \widehat{g}_{g} ( U_{i} ) \vert =O_{p} ( J_{n,1}^{-\alpha }+J_{n,1}^{1/2}/n^{1/2} ) \), we have \( \Vert \widetilde{\varTheta } _{12} \Vert _{\infty }=o_{p} ( n^{1/2}J_{n,2}^{-1/2} ) \). Moreover,
Therefore, for \(n^{1/2}J_{n,2}^{-\alpha -1/2}=o ( 1 ) \), \( n^{1/2}m_{\min }^{-1}J_{n,2}^{-1/2}=o ( 1 ) \) and \(n^{1/2}m_{\min }^{-2}=o ( 1 ) \), we have \( \Vert \varTheta _{2} \Vert _{\infty }=o_{p} ( n^{1/2}J_{n,2}^{-1/2} ) \). Similarly, it can be proved that \( \Vert \varTheta _{3} \Vert _{\infty }=o_{p} ( n^{1/2}J_{n,2}^{-1/2} ) \). By Proposition 1 and (26), for \(n^{1/ ( 2\alpha +1 ) }\ll J_{n,2}\ll n^{1/3} ( \log n ) ^{-1}\), \(n^{1/ ( 2\alpha +3 ) }\ll J_{n,1}\ll J_{n,2}\ll J_{n,1}^{2}\), and \(n^{1/2}m_{\min }^{-1}J_{n,2}^{-1/2}=o ( 1 ) \), we have
Proof of Theorem 1
By (20), we have
and by (8) and \(n^{1/2}J_{n,2}^{-\alpha -1/2}=o ( 1 )\),
Therefore, result (i) in Theorem 1 is proved. By (27), \(\partial \widetilde{L}_{n} ( \widetilde{\delta }_{n}^{0} ) /\partial \delta _{n}=-2 ( \varPi _{1}+\varTheta _{3}+\varTheta _{4} ) ( 1+o_{p} ( 1 ) ) \), where
By (26),
Therefore,
By Taylor expansion, Lemma 1, Berstein’s inequality in Boor (2001) and the above result, we have
Result (ii) follows from Lindeberg–Feller Central Limit Theorem and Slutsky’s Theorem.\(\square \)
Proof of Theorem 2
By (20), Proposition 1 and the conditions in Theorem 2, we have \(\sup _{u\in \mathcal {I}}\) \( \vert \widehat{g} ( u;\widehat{\delta } ) -g ( u ) \vert =O_{p} \{ ( \log n ) ^{1/2}J_{n,2}^{1/2}n^{-1/2}+J_{n,1}^{-\alpha } \}\). \(\square \)
Proof of Theorem 3
Let \(\varXi _{i}=E ( \varPsi _{i}^{\otimes 2} ) \) and \(\varPi _{i}=E ( \sigma ^{2} ( U_{i} ) \varPsi _{i}^{\otimes 2} ) \). Let \(\mathbf{Z }_{1},\ldots ,\mathbf{Z}_{n}\) be independent random variables from \(\hbox {MVN} ( \mathbf{0},\mathbf{I}_{pJ_{n,2}\times pJ_{n,2}} ) \), where \( \mathbf{Z}_{i}= \{ Z_{i,sk} \} \). Define
By the fact that \(\widehat{\beta }_{k}-\beta _{k,n}=\sigma _{n,k}^{-1} ( t ) \mathbf{B}_{2} ( t ) ^{\mathrm{T}}\varvec{\Lambda } _{k} ( \widehat{\delta }-\widetilde{\delta } _{n}^{0} ) \), (8) and (28), we have
It is apparent that \(\eta _{k}^{0} ( t ) \) is a Gaussian process with \(E \{ \eta _{k}^{0} ( t ) \} \equiv 0\), Var\( \{ \eta _{k}^{0} ( t ) \} \equiv 1\), and covariance matrix given in (9). Therefore, we have
Next, we will prove that \(\sup \nolimits _{t\in \mathcal {T}} \vert \eta _{k} ( t ) -\eta _{k}^{0} ( t ) \vert =o_{p} ( 1 ) \). Let \(\mathbf{e}_{i}= \{ e_{i,sk} \} =\varPi _{i}^{-1/2}\varepsilon _{i}\varPsi _{i}\). Denote \(\varPi _{i}^{1/2}= \{ \xi _{i,s^{\prime }k^{\prime },sk} \} \). There exists a constant \( 0<C<\infty \), such that \(\sup \vert \xi _{i,s^{\prime }k^{\prime },sk} \vert \le CJ_{n,2}^{-1/2}\). Then, \(E ( \mathbf{e}_{i} ) = \mathbf{0}\) and Var\( ( \mathbf{e}_{i} ) =\mathbf{I}_{pJ_{n,2}\times pJ_{n,2}}\).There exist \(s,s^{\prime },k,k^{\prime }\) such that
For notation simplicity, let \(\xi _{i}=\xi _{i,s^{\prime }k^{\prime },s,k}\). Order all \(\xi _{i}\), \(1\le i\le n\), from the largest to the smallest such that \(\xi _{ ( 1 ) }\ge \) \(\xi _{ ( 2 ) }\ge \cdots \ge \xi _{ ( n ) }\). Moreover, \(Z_{i,sk}\) can be written as \( Z_{i,sk}=W ( i ) -W ( i-1 ) \), where \( \{ W ( s ) ,0\le s<\infty \} \) is a Wiener process that is a Borel function of \(Z_{i,sk}\). Let \(S_{i}=\sum \nolimits _{i^{\prime }=1}^{i}e_{i^{\prime },sk}\) and \(S_{0}=0\). Define \(M_{n}=\max _{1\le s\le n} \vert S_{s}-W(s) \vert \). By Theorem 2.6.2 in Csőrgő and Révész (1981), we have \(M_{n}=O_{p} ( \log n ) \). Then,
Therefore,
About this article
Cite this article
Ma, S. Estimation and inference in functional single-index models. Ann Inst Stat Math 68, 181–208 (2016). https://doi.org/10.1007/s10463-014-0488-3
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-014-0488-3