Hybrid regularisation and the (in)admissibility of ridge regression in infinite dimensional Hilbert spaces

We consider the problem of estimating the slope function in a functional regression with a scalar response and a functional covariate. This central problem of functional data analysis is well known to be ill-posed, thus requiring a regularised estimation procedure. The two most commonly used approaches are based on spectral truncation or Tikhonov regularisation of the empirical covariance operator. In principle, Tikhonov regularisation is the more canonical choice. Compared to spectral truncation, it is robust to eigenvalue ties, while it attains the optimal minimax rate of convergence in the mean squared sense, and not just in a concentration probability sense. In this paper, we show that, surprisingly, one can strictly improve upon the performance of the Tikhonov estimator in finite samples by means of a linear estimator, while retaining its stability and asymptotic properties by combining it with a form of spectral truncation. Specifically, we construct an estimator that additively decomposes the functional covariate by projecting it onto two orthogonal subspaces defined via functional PCA; it then applies Tikhonov regularisation to the one component, while leaving the other component unregularised. We prove that when the covariate is Gaussian, this hybrid estimator uniformly improves upon the MSE of the Tikhonov estimator in a non-asymptotic sense, effectively rendering it inadmissible. This domination is shown to also persist under discrete observation of the covariate function. The hybrid estimator is linear, straightforward to construct in practice, and with no computational overhead relative to the standard regularisation methods. By means of simulation, it is shown to furnish sizeable gains even for modest sample sizes.

Sparsely Observed Functional Time Series: Theory and Applications

Tomas Rubin

Functional time series is a temporally ordered sequence of not necessarily independent random curves. While the statistical analysis of such data has been traditionally carried out under the assumption of completely observed functional data, it may well happen that the statistician only has access to a relatively low number of sparse measurements for each random curve. These discrete measurements may be moreover irregularly scattered in each curve's domain, missing altogether for some curves, and be contaminated by measurement noise. This sparse sampling protocol escapes from the reach of established estimators in functional time series analysis and therefore requires development of a novel methodology. The core objective of this thesis is development of a non-parametric statistical toolbox for analysis of sparsely observed functional time series data. Assuming smoothness of the latent curves, we construct a local-polynomial-smoother based estimator of the spectral density operator producing a consistent estimator of the complete second order structure of the data. Moreover, the spectral domain recovery approach allows for prediction of latent curve data at a given time by borrowing strength from the estimated dynamic correlations in the entire time series across time. Further to predicting the latent curves from their noisy point samples, the method fills in gaps in the sequence (curves nowhere sampled), denoises the data, and serves as a basis for forecasting. A classical non-parametric apparatus for encoding the dependence between a pair of or among a multiple functional time series, whether sparsely or fully observed, is the functional lagged regression model. This consists of a linear filter between the regressors time series and the response. We show how to tailor the smoother based estimators for the estimation of the cross-spectral density operators and the cross-covariance operators and, by means of spectral truncation and Tikhonov regularisation techniques, how to estimate the lagged regression filter and predict the response process. The simulation studies revealed the following findings: (i) if one has freedom to design a sampling scheme with a fixed number of measurements, it is advantageous to sparsely distribute these measurements in a longer time horizon rather than concentrating over a shorter time horizon to achieve dense measurements in order to diminish the spectral density estimation error, (ii) the developed functional recovery predictor surpasses the static predictor not exploiting the temporal dependence, (iii) neither of the two considered regularisation techniques can, in general, dominate the other for the estimation in functional lagged regression models. The new methodologies are illustrated by applications to real data: the meteorological data revolving around the fair-weather atmospheric electricity measured in Tashkent, Uzbekistan, and at Wank mountain, Germany; and a case study analysing the dependence of the US Treasury yield curve on macroeconomic variables. As a secondary contribution, we present a novel simulation method for general stationary functional time series defined through their spectral properties. A simulation study shows universality of such approach and superiority of the spectral domain simulation over the temporal domain in some situations.

EPFL2021

La détection de périodicités cachées

Jean-Marc Nicoletti

This thesis is a small part of the preparation of the launch of the Gaia mission, a satellite of the European Space Agency (ESA). One of the goals of the mission is to perform a classification among variable stars considering different attributes. Periodic behavior in the observed light curve is such an attribute. It is of importance to determine these hidden cycles with as high an accuracy as possible. A key difficulty is connected to the fact that observations are taken at irregularly distributed time points. Classical methods of frequency analysis do not work in this situation. In a first step, we made a catalogue of potential solutions and applied them to real and simulated data. The performances have been compared in order to select the best method. In a second step, we considered the asymptotic performance of estimators based on regression methods. We derived the asymptotic distribution under the hypothesis that the observation model contains a periodic signal with period P0 > 0, continuous and square integrable on [0; P0), with Gaussian correlated errors. The irregular sampling has to satisfy the asymptotic property that the whole interval [0, P0) can be observed. We have also considered the special case of a periodic signal with period P0 > 0, piecewise constant on [0; P0) and the asymptotic distribution of the estimator. If an irregular sampling scheme does not satisfy the asymptotic property that the signal can be fully observed, the period can still be estimated reliably. We determined the asymptotic distribution of an estimator in a particular situation and under the assumption that the observation model contains a periodic signal with period P0 > 0, continuous and square integrable on [0; P0), and is observed with additive errors that are independent and identically distributed.

EPFL2012

Functional data analysis by matrix completion

Marie-Hélène Descary

Traditional approaches to analysing functional data typically follow a two-step procedure, consisting in first smoothing and then carrying out a functional principal component analysis. The idea underlying this procedure is that functional data are well approximated by smooth functions, and that rough variations are due to noise. However, it may very well happen that localised features are rough at a global scale but still smooth at some finer scale. In this thesis we put forward a new statistical approach for functional data arising as the sum of two uncorrelated components: one smooth plus one rough. We give non-parametric conditions under which the covariance operators of the smooth and of the rough components are jointly identifiable on the basis of discretely observed data: the covariance operator corresponding to the smooth component must be of finite rank and have real analytic eigenfunctions, while the one corresponding to the rough component must have a banded covariance function. We construct consistent estimators of both covariance operators without assuming knowledge of the true rank or bandwidth. We then use them to estimate the best linear predictors of the the smooth and the rough components of each functional datum. In both the identifiability and the inference part, we do not follow the usual strategy used in functional data analysis which is to first employ smoothing and work with continuous estimate of the covariance operator. Instead, we work directly with the covariance matrix of the discretely observed data, which allows us to use results and tools from linear algebra. In fact, we show that the whole problem of uniquely recovering the covariance operator of the smooth component given the one of the raw data can be seen as a low-rank matrix completion problem, and we make great use of a classical relation between the rank and the minors of a matrix to solve this matrix completion problem. The finite-sample performance of our approach is studied by means of simulation study.