1: \begin{abstract}

2: We consider the sampling problem for functional PCA (fPCA), where

3: the simplest example is the case of taking time samples of the

4: underlying functional components. More generally, we model the

5: sampling operation as a continuous linear map from $\Hil$ to

6: $\reals^m$, where the functional components to lie in some

7: Hilbert subspace $\Hil$ of $L^2$, such as a reproducing kernel

8: Hilbert space of smooth functions. This model includes time and

9: frequency sampling as special cases. In contrast to classical

10: approach in fPCA in which access to entire functions is assumed,

11: having a limited number $m$ of functional samples places

12: limitations on the performance of statistical procedures. We study

13: these effects by analyzing the rate of convergence of an

14: $M$-estimator for the subspace spanned by the leading components in

15: a multi-spiked covariance model. The estimator takes the form of

16: regularized PCA, and hence is computationally attractive. We

17: analyze the behavior of this estimator within a nonasymptotic

18: framework, and provide bounds that hold with high probability as a

19: function of the number of statistical samples

20: $n$ and the number of functional samples $m$. We also

21: derive lower bounds showing that the rates obtained are minimax

22: optimal.

23: \end{abstract}