Abstract
This paper studies the estimation of a general functional single index model, in which the conditional distribution of the response depends on the functional predictor via a functional single index structure. We find that the slope function can be estimated consistently by the estimation obtained by fitting a misspecified functional linear quantile regression model under some mild conditions. We first obtain a consistent estimator of the slope function using functional linear quantile regression based on functional principal component analysis, and then employ a local linear regression technique to estimate the conditional quantile function and establish the asymptotic normality of the resulting estimator for it. The finite sample performance of the proposed estimation method is studied in Monte Carlo simulations, and is illustrated by an application to a real dataset.
Similar content being viewed by others
References
Ait-Saïdi, A., Ferraty, F., Kassa, R., & Vieu, P. (2008). Cross-validated estimations in the single-functional index model. Statistics, 42(6), 475–494.
Cai, T. T., Hall, P., et al. (2006). Prediction in functional linear regression. The Annals of Statistics, 34(5), 2159–2179.
Cai, T. T., & Yuan, M. (2012). Minimax and adaptive prediction for functional linear regression. Journal of the American Statistical Association, 107(499), 1201–1216.
Cardot, H., Ferraty, F., & Sarda, P. (2003). Spline estimators for the functional linear model. Statistica Sinica, 2003, 571–591.
Chen, D., Hall, P., Müller, H.-G., et al. (2011). Single and multiple index functional regression models with nonparametric link. The Annals of Statistics, 39(3), 1720–1747.
Chen, K., & Müller, H.-G. (2012). Conditional quantile analysis when covariates are functions, with application to growth data. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 74(1), 67–89.
Crambes, C., Kneip, A., Sarda, P., et al. (2009). Smoothing splines estimators for functional linear regression. The Annals of Statistics, 37(1), 35–72.
Delaigle, A., Hall, P., Apanasovich, T. V., et al. (2009). Weighted least squares methods for prediction in the functional data linear model. Electronic Journal of Statistics, 3, 865–885.
Fan, J., & Gijbels, I. (1996). Local polynomial modelling and its applications. London: Chapman and Hall.
Fan, Y., James, G. M., Radchenko, P., et al. (2015). Functional additive regression. The Annals of Statistics, 43(5), 2296–2325.
Ferraty, F., Goia, A., Salinelli, E., & Vieu, P. (2013). Functional projection pursuit regression. Test, 22(2), 293–320.
Ferraty, F., Rabhi, A., & Vieu, P. (2005). Conditional quantiles for dependent functional data with application to the climatic" el niño" phenomenon. Sankhyā: The Indian Journal of Statistics, 2005, 378–398.
Ferraty, F., & Vieu, P. (2006). Nonparametric functional data analysis: Theory and practice. Berlin: Springer Science & Business Media.
Goia, A., & Vieu, P. (2014). Some advances in semiparametric functional data modelling. Contributions in Infinite-Dimensional Statistics and Related Topics, 2014, 135–140.
Goia, A., & Vieu, P. (2015). A partitioned single functional index model. Computational Statistics, 30(3), 673–692.
Hall, P., Horowitz, J. L., et al. (2007). Methodology and convergence rates for functional linear regression. The Annals of Statistics, 35(1), 70–91.
He, G., Müller, H.-G., Wang, J.-L., Yang, W., et al. (2010). Functional linear regression via canonical analysis. Bernoulli, 16(3), 705–729.
Hsing, T., & Eubank, R. (2015). Theoretical foundations of functional data analysis, with an introduction to linear operators. Hoboken: Wiley.
Kai, B., Li, R., & Zou, H. (2011). New efficient estimation and variable selection methods for semiparametric varying-coefficient partially linear models. Annals of statistics, 39(1), 305–332.
Kato, K. (2012). Estimation in functional linear quantile regression. The Annals of Statistics, 40(6), 3108–3136.
Kong, D., Xue, K., Yao, F., & Zhang, H. H. (2016). Partially functional linear regression in high dimensions. Biometrika, 103(1), 147–159.
Li, Y., Hsing, T., et al. (2010). Deciding the dimension of effective dimension reduction space for functional and high-dimensional data. The Annals of Statistics, 38(5), 3028–3062.
Li, Y., Wang, N., & Carroll, R. J. (2013). Selecting the number of principal components in functional data. Journal of the American Statistical Association, 108(504), 1284–1294.
Ma, S. (2016). Estimation and inference in functional single-index models. Annals of the Institute of Statistical Mathematics, 68(1), 181–208.
Ma, S., & He, X. (2016). Inference for single-index quantile regression models with profile optimization. The Annals of Statistics, 44(3), 1234–1268.
Müller, H.-G., Stadtmüller, U., et al. (2005). Generalized functional linear models. The Annals of Statistics, 33(2), 774–805.
Pollard, D. (1991). Asymptotics for least absolute deviation regression estimators. Econometric Theory, 7(2), 186–199.
Ramsay, G., & Silverman, H. (2005). Functional data analysis. New York: Springer.
Shin, H., & Lee, S. (2016). An rkhs approach to robust functional linear regression. Statistica Sinica, 2016, 255–272.
Wang, G., Feng, X.-N., & Chen, M. (2016). Functional partial linear single-index model. Scandinavian Journal of Statistics, 43(1), 261–274.
Wang, J.-L., Chiou, J.-M., & Müller, H.-G. (2016). Functional data analysis. Annual Review of Statistics and Its Application, 3, 257–295.
Yao, F., Lei, E., & Wu, Y. (2015). Effective dimension reduction for sparse functional data. Biometrika, 102(2), 421–437.
Yao, F., Müller, H.-G., & Wang, J.-L. (2005). Functional data analysis for sparse longitudinal data. Journal of the American Statistical Association, 100(470), 577–590.
Yao, F., Sue-Chee, S., & Wang, F. (2017). Regularized partially functional quantile regression. Journal of Multivariate Analysis, 156, 39–56.
Yao, F., Wu, Y., & Zou, J. (2016). Probability-enhanced effective dimension reduction for classifying sparse functional data. Test, 25(1), 1–22.
Yuan, M., Cai, T. T., et al. (2010). A reproducing kernel hilbert space approach to functional linear regression. The Annals of Statistics, 38(6), 3412–3444.
Zhu, H., Li, R., Zhang, R., & Lian, H. (2020). Nonlinear functional canonical correlation analysis via distance covariance. Journal of Multivariate Analysis, 180, 104662.
Zhu, H., Zhang, R., Yu, Z., Lian, H., & Liu, Y. (2019). Estimation and testing for partially functional linear errors-in-variables models. Journal of Multivariate Analysis, 170, 296–314.
Funding
Zhang’s research was partially supported by the National Natural Science Foundation of China (11971171), 111 Project (B14019), Project of National Social Science Fund of China (15BTJ027) and Key Laboratory of Advanced Theory and Application in Statistics and Data Science-MOE, China. Liu’s research was partially supported by Guangdong Basic and Applied Basic Research Foundation (2021A1515110443), and University Innovation Team Project of Guangdong Province (2020WCXTD018). Ding’s research was partially supported by the National Natural Science Foundation of China (11901286).
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
Appendix
1.1 A.1 Proof of Proposition 1
It suffices to prove that, for any constant \(u \in \mathbb {R}\), there exists a constant c such that \(\mathcal {L}_{\tau }(u, \beta )\ge \mathcal {L}_{\tau }(u, c\beta _{0})\). Specifically,
where \(c=\langle K(\beta _{0}), \beta \rangle /\langle K(\beta _{0}), \beta _{0}\rangle\) and the inequality follows from the convexity of \(\rho _{\tau }\) and Jensen’s inequality, and the last equality holds true by invoking the linearity condition (7). This completes the proof of Proposition 1.
1.2 A.2 Proof of Theorem 1
We begin by defining some notations to be used in the proofs. Let \(\xi _{i}=\left( 1,\xi _{i1},\ldots ,\xi _{im_{n}}\right) ^{\top }\), \({\widehat{\xi }}_{i}=\left( 1,{\widehat{\xi }}_{i1},\ldots ,{\widehat{\xi }}_{im_{n}}\right) ^{\top }\), \({\widetilde{\theta }}=\left( u_{\tau },\alpha _{\tau 1}, \ldots , \alpha _{\tau m_{n}}\right) ^{\top }\), \(A=\left( u, \alpha _{1}, \ldots , \alpha _{m_{n}}\right) ^{\top }\) and \(\zeta _{i}=\sum _{j=m_{n}+1}^{\infty }\alpha _{\tau j}\xi _{ij}\). Then the objective function in (9) can be rewritten as
Let
\(\Lambda =diag \left( 1, \lambda _{1},\ldots ,\lambda _{m_{n}}\right)\), \(\theta =\left( n^{1/2}(u-u_\tau ), U^{\top }\right) ^{\top }\), \({\widetilde{\xi }}_{i}=n^{-1/2}\Lambda ^{-1/2}{\widehat{\xi }}_{i}\), where
\(U=(u_{1},\ldots , u_{m_{n}})^{\top }=n^{1/2}\Lambda _{1}^{1/2}(b-\alpha )\), \(b=\left( \alpha _{1}, \ldots , \alpha _{m_{n}}\right) ^{\top }\), and \(\Lambda _{1}=diag \left( \lambda _{1},\ldots ,\lambda _{m_{n}}\right) .\) Let
It’s easy to see that minimizing (14) with respect to A is equivalent to minimizing (16) over \(\theta .\) Let \(\psi _{\tau }(t)=\tau -\mathbf {1}(t<0),\) \(\Omega =\{X_{1}, \ldots , X_{n}\}\), \(S_{ni}=\rho _{\tau }\left( W_{i}+\varepsilon _{i\tau }-{\widetilde{\xi }}_{i}^{\top }m_{n}^{1/2}\theta \right) - \rho _{\tau }\left( W_{i}+\varepsilon _{i\tau }\right)\), \(\Gamma _{ni}=\mathbb {E}(S_{ni}|\Omega )\), \(R_{ni}=S_{ni}-\Gamma _{ni}+m_{n}^{1/2}{\widetilde{\xi }}_{i}^{\top }\theta \psi _{\tau }\left( \varepsilon _{i\tau }\right)\), \(S_{n}=\sum _{i=1}^{n}S_{ni}\), \(\Gamma _{n}=\sum _{i=1}^{n}\Gamma _{ni}\), \(R_{n}=\sum _{i=1}^{n}R_{ni}\). Then
Lemma 1
Let \(Z_{1}, \ldots , Z_{n}\) be arbitrary scalar random variables such that \(\mathop {max }\limits _{1\le i \le n}\mathbb {E}\left( |Z_{i}|^{L_{0}}\right) <\infty\) for some \(L_{0} \ge 1\). Then, we have
Proof of Lemma 1
This inequality follows from the observation that
\(\square\)
Lemma 2
Under conditions (C1)–(C3) and (C5), we have
where \(\left\| {\widetilde{\xi }}_{i} \right\| _{2}=\left( {\widetilde{\xi }}_{i}^{\top } {\widetilde{\xi }}_{i}\right) ^{1/2}\).
Proof of Lemma 2
It is easy to show
By (5.21) and (5.22) in Hall et al. (2007), we have \(\left\| {\widehat{\phi }}_{j}-\phi _{j}\right\| ^{2}=O_{p}(j^{2}/n)\), uniformly in \(j \in \{ 1, \ldots , m_{n}\}\). Combining \(\mathbb {E}\left( ||X||^{4}\right) <\infty\) and Lemma 1, we have \(\mathop {max }\limits _{1\le i \le n} ||X_{i}||^{2}=O_{p}\left( n^{1/2}\right)\). Hence, \(\mathop {max }\limits _{1\le i \le n}2||X_{i}||^{2}\sum _{j=1}^{m_{n}}\lambda _{j}^{-1}\left\| {\widehat{\phi }}_{j}-\phi _{j}\right\| ^{2}\le O_{p}\left( n^{-1/2}\right) \sum _{j=1}^{m_{n}}j^2/\lambda _{j}\le O_{p}\left( m_{n}^{\nu +3}n^{-1/2}\right)\). By assumption (C2) and Lemma 1, we have \(\mathop {max }\limits _{1\le i \le n}\sum _{j=1}^{m_{n}}2\lambda _{j}^{-1} \xi _{ij}^{2}=O_{p}\left( m_{n}n^{1/2}\right)\). Since \(\nu >1\) and \(\zeta >\nu /2+1\), there exists a small constant \(C_0> 0\) (depending on \(\nu\) and \(\zeta\)) such that \(\mathop {max }\limits _{1\le i \le n}2||X_{i}||^{2}\sum _{j=1}^{m_{n}}\lambda _{j}^{-1}\left\| {\widehat{\phi }}_{j}-\phi _{j}\right\| ^{2}=O_p\left\{ n^{-C_0}\left( n/m_{n}\right) \right\}\) and \(\mathop {max }\limits _{1\le i \le n}\sum _{j=1}^{m_{n}}2\lambda _{j}^{-1} \xi _{ij}^{2}=O_p\left\{ n^{-C_0}\left( n/m_{n}\right) \right\}\), and complete the proof of Lemma 2. \(\square\)
Lemma 3
Under conditions (C1)–(C5), we have
where \(W_{i}\) are defined in (15).
Proof of Lemma 3
Following conditions (C1), (C4) and \(\left\| {\widehat{\phi }}_{j}-\phi _{j}\right\| ^{2}=O_{p}\left( j^{2}/n\right)\), uniformly in \(j \in \{ 1, \ldots , m_{n}\}\), we have
Therefore, by condition (C5), we obtain \(\max _{1 \le i \le n}\left| \sum _{j=1}^{m_{n}}\left( \xi _{ij}-{\widehat{\xi }}_{ij}\right) \alpha _{\tau j}\right| =o_{p}(1)\). Following conditions (C2)–(C5) and Lemma 1, we have
where \(v>1\) and \(\zeta >v/2+1.\) Thus, we have
Lemma 4
For \(k \in \left\{ 1,\ldots , m_{n}\right\}\), define the following expressions:
Under conditions (C1)–(C5), we have
where \(o_{p}\) and \(O_{p}\) are uniform for \(k \in \{1,\ldots , m_{n}\}.\)
Proof of Lemma 4
Please refer to the proof of Lemma 2 of Kong et al. (2016). \(\square\)
proof of Theorem 1
By simple calculation, we have
Therefore,
Following
and (5.2) in Hall et al. (2007), we get
where \(\theta _{1}\) is the first component of \(\theta\), \(0_{m_{n}}\) represents an \(m_{n}\)-dimensional vector whose components are all zero, and \({\widehat{\Delta }}=\left\{ \displaystyle {\iint _{\mathcal {I}^{2}}\left( {{\widehat{K}}}-K\right) ^{2}}\right\} ^{1/2}.\) By (5.9) in Hall et al. (2007) and conditions (C3)–(C5), we get \(\mathbb {E}\left( {\widehat{\Delta }}^{2}/\lambda _{m_{n}}^{2}\right) =o(1).\) Therefore,
Suppose that \(L_{1}\) is a positive constant. For \(\Vert \theta \Vert _{2}\le L_{1},\) following Lemmas 2 and 3, and \(\varepsilon _{i\tau }=O_{p}(1)\), we have
Combining (18) and (19), we obtain
It is easy to show
Since \(\mathbb {E}\left( {\widehat{\Delta }}^{2}/\lambda _{m_{n}}^{2}\right) =o(1),\) we have
Hence, by (17), we get
It is easy to prove
Following Lemma 4 and condition (C5), we conclude that
Thus,
By the proof of Lemma 2 of Yao et al. (2017) and condition (C6),
Following condition (C6), (22) and (23), we get
Combining (20), (21) and (24), for sufficiently large \(L_{1}\), we have
where \(c_{0}\) is a positive constant. This implies that
as \(n\rightarrow \infty .\) Hence, \(\mathbb {P}\left( \left\| {\widehat{\theta }}\right\| _{2}\le L_{1}m_{n}^{1/2}\right) \rightarrow 1\) as \(n\rightarrow \infty ,\) where \({\widehat{\theta }}\) is the minimizer of (16). Hence, \(\left\| {\widehat{\theta }}\right\| _{2}=O_{p}\left( m_{n}^{1/2}\right) .\) Therefore, we have
By some straightforward calculations, we get
By conditions (C3) and (C5), and (25), we obtain
Since \(\left\| {\widehat{\phi }}_{j}-\phi _{j}\right\| ^{2}=O_{p}\left( j^{2}/n\right)\), uniformly in \(j \in \{ 1, \ldots , m_{n}\}\), we have
Since \(\zeta >\nu /2+1\) and \(\nu >1\), we conclude that
By conditions (C4) and (C5), we get
Combining (26)–(28), we complete the proof of Theorem 1.
1.3 A.3 Proof of Theorem 2
Quadratic Approximation Lemma Pollard (1991): suppose \(\mathcal {L}_{n}(\delta )\) is convex and can be represented as \(\delta ^{\top }B\delta /2+u_{n}^{\top }\delta +a_{n}+R_{n}(\delta ),\) where B is symmetric and positive definite, \(u_{n}\) is stochastically bounded, \(a_{n}\) is arbitrary, and \(R_{n}(\delta )\) goes to zero in probability for each \(\delta\). Then \({\widehat{\delta }},\) the minimizer of \(\mathcal {L}_{n}(\delta ),\) is only \(o_{p}(1)\) away \(\mathrm {from}-B^{-1} u_{n} .\) If \(u_{n}\) converges in distribution to u, then \({\widehat{\delta }}\) converges in distribution to \(-B^{-1}u.\)
We use the Quadratic Approximation Lemma to complete the proof. Recall the notations \(z=\langle X_{0}\), \(\beta _{\tau }\rangle\), \({{\widehat{z}}}=\langle X_{0}\), \({\widehat{\beta }}_{\tau }\rangle\), \(Z_{i}=\langle X_{i}\), \(\beta _{\tau }\rangle\), \({{\widehat{Z}}}_{i}=\langle X_{i}\), \({\widehat{\beta }}_{\tau }\rangle\). Let \(a_{0}=Q_{\tau }(z)=Q_{\tau }(\langle X_{0}\), \(\beta _{\tau }\rangle )\), \(b_{0}=Q'_{\tau }(z)\), \({{\widehat{K}}}_{i}=K\left\{ \left( {{\widehat{Z}}}_{i}-{{\widehat{z}}}\right) /h\right\}\), \(K_{i}=K\left\{ \left( Z_{i}-z\right) /h\right\}\) and \(K'_{i}=K'\{(Z_{i}-z)/h\}\) for \(i=1, \ldots , n\). We first show that
Define
Then
Following
we have \(|\rho _{\tau }(x-y)-\rho _{\tau }(x)|\le 2|y|\). By Theorem 1, we get
Hence, we have \(R_{11}=O_{p}\left( n^{-\frac{2\zeta -1}{2(\nu +2\zeta )}}\right)\). It is easy to show that
Let \(Z_{i0}=(X_i-X_{0})/h\). Then
This implies that \(R_{12}=O_{p}\left( n^{-\frac{2\zeta -1}{2(\nu +2\zeta )}}\right)\). Following similar arguments, we have \(R_{13}=O_{p}\left( n^{-\frac{2\zeta -1}{2(\nu +2\zeta )}}\right) .\) Thus (29) holds. Following (29) and \(n^{\frac{\nu +1}{\nu +2\zeta }}h\rightarrow 0\), we get
This means that the estimate of \(Q_{\tau }(\cdot )\) based on \(\left\{ \left( \langle X_{i}, {\widehat{\beta }}_{\tau }\rangle , Y_{i}\right) , i=1,2, \ldots , n\right\}\) is asymptotically as efficient as that based on \(\left\{ \left( \langle X_{i}, \beta _{\tau }\rangle , Y_{i}\right) , i=1,2, \ldots , n\right\} .\)
Let \(\eta _{i}=(a-a_{0})+(b-b_{0})(Z_{i}-z)\). Following (30), we obtain
Let \(\varrho =(nh)^{1/2}\{a-a_{0}, h(b-b_{0})\}^\top\). In the sequel, we show the first quantity in (33) has approximately the form \(W_{n}^{\top }\varrho\), and the second quantity has the form \(\frac{1}{2}\varrho ^{\top }S_{n}\varrho\). Denote by \(F_{Y}(\cdot |Z)\) the conditional distribution of Y given \(\langle X, \beta _{\tau } \rangle =Z\). Then
Using Taylor’s expansion, we obtain
for \(|Z_{i}-z|\le h\), and
where \(f_{Y_{i}}(\cdot |Z_{i})\) is the conditional density function of \(Y_{i}\) given \(\langle X_{i}, \beta _{\tau }\rangle =Z_{i}\). Hence, we get
which is again of the form
The second term on the right side of (36) can be approximated as
which is again of the form
The second term on the right side of (33) can be approximated as
which is now of the form
Combining (34)–(37), we observe that (33) is now approximated by a quadratic function of \(\varrho\). By the Quadratic Approximation Lemma, \(\varrho\) is only \(o_{p}(1)\) away from \(-{S}_{n}^{-1}W_{n}\), where
It can be easily shown that, as \(n\longrightarrow \infty\), \(S_{n}\) converges in probability to
where \(\mu _{2}=\int _{-1}^{1} t^{2}K(t)dt\). By some straightforward calculations, we can show \((nh)^{-1}\sum _{i=1}^{n}\{\mathbf {1}(Y_{i}\le Q_{\tau }(Z_{i}))-\tau \}K_{i}\) converges in distribution to a normal distribution with zero mean and variance \(\tau (1-\tau )f_{\beta _{\tau }}(z)\int _{-1}^{1} K^{2}(t)dt\), and \(\frac{1}{2}(nh)^{-1}\sum _{i=1}^{n}f_{Y_{i}}(Q_{\tau }(Z_i)|Z_i)Q''_{\tau }(z)(Z_{i}-z)^{2}K_{i}\) converges in probability to \(\frac{1}{2}h^2 Q''_{\tau }(z)\mu _{2}f_{Y}(Q_{\tau }(z)|z)f_{\beta _{\tau }}(z).\) The proof of Theorem 2 is completed by combining the above results.
Rights and permissions
About this article
Cite this article
Zhu, H., Zhang, R., Liu, Y. et al. Robust estimation for a general functional single index model via quantile regression. J. Korean Stat. Soc. 51, 1041–1070 (2022). https://doi.org/10.1007/s42952-022-00174-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s42952-022-00174-4