A review of goodness-of-fit tests for models involving functional data

A sizable amount of goodness-of-fit tests involving functional data have appeared in the last decade. We provide a relatively compact revision of most of these contributions, within the independent and identically distributed framework, by reviewing goodness-of-fit tests for distribution and regression models with functional predictor and either scalar or functional response.



There are no comments yet.


page 1

page 2

page 3

page 4


Goodness-of-fit tests for functional linear models based on integrated projections

Functional linear models are one of the most fundamental tools to assess...

Selecting the Derivative of a Functional Covariate in Scalar-on-Function Regression

This paper presents tests to formally choose between regression models u...

Modern Non-Linear Function-on-Function Regression

We introduce a new class of non-linear function-on-function regression m...

FuncNN: An R Package to Fit Deep Neural Networks Using Generalized Input Spaces

Neural networks have excelled at regression and classification problems ...

Bayesian Goodness of Fit Tests: A Conversation for David Mumford

The problem of making practical, useful goodness of fit tests in the Bay...

Exhaustive goodness-of-fit via smoothed inference and graphics

Classical tests of goodness-of-fit aim to validate the conformity of a p...

Smoothed inference and graphics via LP modeling

Classical tests of goodness-of-fit aim to validate the conformity of a p...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Since the earliest Goodness-of-Fit (GoF) tests were introduced by Pearson more than a century ago, there has been a prolific statistical literature on this topic. If we were to highlight a milestone in this period, that may be 1973, with the publication of Durbin (1973) and Bickel and Rosenblatt (1973)

, introducing a novel design of GoF tests based on distances between distribution and density estimates, respectively.

To set the context for the reader, assume that

is an identically and identically distributed (iid) sample of a random variable

with (unknown) distribution (or density , if that is the case). If the target function is the distribution , then the GoF testing problem can be formulated as testing vs. , where stands for a parametric family of distributions indexed in some finite-dimensional set

. A general test statistic for this problem can be written as

, with the functional

denoting, here and henceforth, some kind of distance between a nonparametric estimate, given in this case by the empirical cumulative distribution function

, and an estimate obtained under the null hypothesis

, in this case. Similarly, for testing the GoF of a certain parametric density model, the testing problem is formulated as vs. and can be approached with the general test statistic . In this setting, is the density estimate under and

denotes the kernel density estimator

by Parzen (1962) and Rosenblatt (1956), where , is the kernel function, and is the bandwidth.

The previous ideas were naturally generalized to the context of regression models in the 1990s. Consider a nonparametric, random design, regression model such that with , and and . Denote by an iid sample of satisfying such a model. In this context, the GoF goal is to test vs. , where represents a parametric family of regression functions indexed in . Continuing along the testing philosophies advocated by Durbin (1973) and Bickel and Rosenblatt (1973), the seminal works of Stute (1997) and Härdle and Mammen (1993) respectively introduced two types of GoF tests for regression models:

  1. Tests based on empirical regression processes, considering distances between estimates of the integrated regression function ( being the marginal distribution of under and ). Specifically, the test statistics are constructed as , with and .

  2. Smoothing-based tests, using distances between estimated regression functions, , with a smooth regression estimator. As a particular case, , with some weights depending on a smoothing parameter . Such an estimator can be obtained with Nadaraya–Watson or local linear weights (see, e.g., Wand and Jones (1995)).

A complete review of GoF for regression models was presented by González-Manteiga and Crujeiras (2013), who described the aforementioned two types of paradigms and focused on the smoothing-based alternative for discussing their properties (asymptotic behavior and calibration of the distribution in practice). The authors thoroughly checked references in the statistical literature for more than two decades, and they also identified some areas where GoF tests where still to be developed. One of these areas is functional data analysis.

The goal of this work is to round off this previous review, updating the more recent contributions in GoF for distribution and regression models with functional data. Consequently, the rest of the chapter is organized as follows: Section 2 is devoted to GoF for distribution models of functional random variables, and Section 3 focuses on regression models with scalar (Section 3.1) and functional response (Section 3.2).

2 GoF for distribution models for functional data

Owing to the requirement of appropriate tools for analyzing high-frequency data, and the boost provided by the books by Ramsay and Silverman (2005) and Ferraty and Vieu (2006) (or, more recently, by Horváth and Kokoszka (2012)), functional data analysis is nowadays one of the most active research areas within statistics.

Actually, the pressing needs of developing new statistical tools with data in general spaces have reclaimed separable Hilbert spaces as a very natural and common framework. However, given the generality of this kind of spaces, there is a scarcity of parametric distribution models for a Hilbert-valued random variable aside the popular framework of Gaussian processes.

Let denote a Hilbert space over , the norm of which is given by its scalar product as . Consider iid copies of the random variable , with

the probability space where the random sample is defined and

the Borel -field on . The general GoF problem for the distribution of consists on testing vs. , where is a class of probability measures on indexed in a parameter set , now possibly infinite-dimensional, and

is the (unknown) probability distribution of

induced over .

When the goal is to test the simple null hypothesis , a feasible approach that enables the construction of test statistics is based on projections , in such a way that the test statistics are defined from the projected sample . Such an approach can be taken on the distribution function: with and . Some specific examples are given by the adaptation to this context of the Kolmogorov–Smirnov, Cramer–von Mises, or Anderson–Darling type tests. As an alternative, and mimicking the smoothing-based tests presented in Section 1, a test statistic can also be built as with . It should be also noted that, when embracing the projection approach, the test statistic may take into account ‘all’ the projections within a certain space, e.g. by considering for a probability measure on the space of projections, or take just with being a randomly-sampled projection from a certain non-degenerate probability measure .

Now, when the goal is to test the composite null hypothesis , the previous generic approaches are still valid if replacing with . Within this setting, Cuesta-Albertos et al. (2006) and Cuesta-Albertos et al. (2007) provide a characterization of the composite null hypothesis by means of random projections, and provide a bootstrap procedure for calibration, as well as Bugni et al. (2009). As an alternative, Ditzhaus and Gaigall (2018) follows a finite-dimensional approximation. Note that, in the space of real square-integrable functions , as a particular case one may take , with . The previous references provide some approaches for the calibration under the null hypothesis of the rejection region , where .

A relevant alternative to the procedures based on projections is the use of the so-called ‘energy statistics’ Székely and Rizzo (2017). Working with a general Hilbert separable space (as it can be seen in Lyons (2013)), if and ( being the distribution under the null), then


with and iid copies of the variables with distributions and , respectively. Importantly, (1) equals if and only if , a characterization that serves as basis for a GoF test. The energy statistic in (1) can be empirically estimated from a sample as

with simulated from . This estimated energy can be compared with appropriate Monte Carlo simulation , designed to build an -level critical point using . Note that, under the null hypothesis, is simulated from . In the case of testing a composite hypothesis, then generation is done under with estimated using .

Due to the scarcity of distribution models for random functions, the Gaussian case is one of the most widely studied, as it can bee seen, e.g., in Kellner and Celisse (2019); Kolkiewicz et al. (2021) and in the recent review by Górecki et al. (2020) on tests for Gaussianity of functional data.

Finally, it is worth it to mention the two-sample problem, a common offspring of the simple-hypothesis one-sample GoF problem. Two-sample tests have also received a significant deal of attention in the last decades; see, e.g., the recent contributions by Jiang et al. (2019) and Qiu et al. (2021), and references therein.

3 GoF for regression models with functional data

We assume henceforth, without loss of generality and for the sake of easier presentation, that both the predictor and response are centered, so that the intercepts of the linear functional regression models are null.

3.1 Scalar response

A particular case of a regression model with functional predictor and scalar response is the so-called functional linear model. For

, this parametric model is given by


for some unknown indexing the functional form of the model. This popular model can be seen as the natural extension of the classical linear (Euclidean) regression model.

In general, there have been two mainstream approaches for performing inference on (2): (i) testing the significance of the trend within the linear model, i.e., testing vs. , usually with ; (ii) testing the linearity of , i.e., testing vs. .

For the GoF testing problem presented in (ii), given an iid sample , one may consider the adaptation to this setting of the smoothing-based tests, with a basic test statistic structure given by , where is a suitable estimator for and


is the Nadaraya–Watson estimator with a functional predictor. A particular smoothing-based test statistic is given by that of Delsol et al. (2011),

which employs a weighted distance between (3) and , the latter being a smoothed version of the parametric estimator that follows by replacing with in (3). Note that a crucial problem for implementing this test is the computation of the critical region , which depends on the selection of when a class of estimators for is used under the null. This class of smoothed-based tests were deeply studied in the Euclidean setting (see González-Manteiga and Crujeiras (2013)). Nevertheless, this is not the case in the functional context, except for the recent contributions by Maistre and Patilea (2020) and Patilea and Sánchez-Sellero (2020).

As also presented by González-Manteiga and Crujeiras (2013) in their review, it is possible to avoid the bandwidth selection problem using tests based on empirical regression processes. For this purpose, a key element is the empirical counterpart of the integrated regression function , where means that , for all . In this scenario, the test statistic can be formulated as , where , where . Deriving the theoretical behavior of an empirical regression process indexed by , namely is a challenging task. Yet, as previously presented, the projection approach over can be considered. The null hypothesis can be formulated as

which in turn is equivalent to replacing ‘for all ’ with ‘for all ’ or ‘for all for all ’, where

are infinite- and finite-dimensional spheres on , is an orthonormal basis for , and . As follows from García-Portugués et al. (2014), a general test statistic can be built aggregating all the projections within a certain subspace: with based on


for . In this case, is a probability measure defined in or , for a certain . Alternatively, the test statistic can be based on only one random projection: . More generally, may consider the aggregation of a finite number of random projections, as advocated in the test statistic of Cuesta-Albertos et al. (2019). Both types of tests, all-projections and finite-random-projections, may feature several distances for , such as Kolmogorov–Smirnov or Cramér–von Mises types.

Model (2) can be generalized to include a more flexible trend component, for instance, with an additive formulation. The functional generalized additive model (see McLean et al. (2015)) is formulated as


and it can be seen that (2) is a particular case of (5) with and . The functional can be approximated as


are unknown tensor product B-spline coefficients. Both for the

and components, cubic B-spline bases, namely and , are considered.

Model (5) can be written in an approximated way as a linear model with random effects (see Yasemin-Tekbudak et al. (2019)) using the evaluations of over a grid . Under the assumption of being a Gaussian process, the so-called restricted likelihood ratio test (RLRT) can be used, where testing the GoF of the functional linear model (2) against model specifications within (5

) is equivalent to test that the variance of the random effect is null.

Another generalization of the functional linear model is given by the functional quadratic regression model introduced by Horváth and Reeder (2013):


Clearly, when , (2) follows as a particular case of (6

). Using a principal component analysis methodology to approximate the covariance function

with and with

the eigenfunctions of

, model (6) can be written as a kind of linear model, were the null hypothesis is tested.

A recent contribution by Lai et al. (2020) is devoted to the testing a modified null hypothesis: is independent of and ”, using the recent results related with the distance covariance (see Székely et al. (2007), Lyons (2013), and Sejdinovic et al. (2013)). Consider and two semimetric spaces of negative type, where and are the corresponding semimetrics. Denote by

a random element with joint distribution

and marginals and , respectively, and take an iid copy of . The generalized distance covariance is given by

As noted by Lai et al. (2020), the generalized distance covariance can be alternatively written as

Note that if and only if and are independent. Given an iid sample of , an empirical estimator of is given by

with and . Taking and , is the absolute value and is the distance associated to . The test statistic is and is based on .

All the tests described in this section have challenging limit distributions and need to be calibrated with resampling techniques.

3.2 Functional response

When both the predictor and the response, and , are functional random variables evaluated in and , the regression model is related with the operator . Perhaps the most popular operator specification is a (linear) Hilbert–Schmidt integral operator, expressible as


for , which is simply referred to as the functional linear model with functional response. The kernel can be represented as , with and being orthonormal bases of and , respectively.

Similarly to the case with scalar response, performing inference on (7) have attracted the analogous two mainstream approaches: (i) testing vs. , usually with ; (ii) testing vs. . The GoF problem given in (ii) can be approached by considering a double-projection mechanism based on and . Given an iid sample , a general test statistic follows (see García-Portugués et al. (2020a)) as with , where and follows from (4) by replacing with , and and with and , respectively. In this case, is a probability measure is defined in or , for certain . The projection approach is immediately adaptable to the GoF of (7) with , and allows graphical tools for that can help detecting the deviations from the null, see García-Portugués et al. (2020b). An alternative route considering projections just for is presented by Chen et al. (2020).

The above generalization to the case of functional response is certainly more difficult for the class of tests based on the likelihood ratios. Regarding the smoothing-based tests, Patilea et al. (2016) introduced a kernel-based significance test consistent for nonlinear alternative. More recently, Lee et al. (2020) proposed a significance test based on correlation distance ideas.


The authors acknowledge the support of project MTM2016-76969-P, PGC2018-097284-B-100, and IJCI-2017-32005 from the Spain’s Ministry of Economy and Competitiveness. All three grants were partially co-funded by the European Regional Development Fund (ERDF). The support by Competitive Reference Groups 2017–2020 (ED431C 2017/38) from the Xunta de Galicia through the ERDF is also acknowledged.


  • Bickel and Rosenblatt (1973) Bickel, P. J. and Rosenblatt, M. (1973). On some global measures of the deviations of density function estimates. The Annals of Statistics, 1(6):1071–1095.
  • Bugni et al. (2009) Bugni, F. A., Hall, P., Horowitz, J. L., and Neumann, G. R. (2009). Goodness-of-fit tests for functional data. The Econometrics Journal, 12(S1):S1–S18.
  • Chen et al. (2020) Chen, F., Jiang, Q., Feng, Z., and Zhu, L. (2020).

    Model checks for functional linear regression models based on projected empirical processes.

    Computational Statistics & Data Analysis, 144:106897.
  • Cuesta-Albertos et al. (2007) Cuesta-Albertos, J. A., del Barrio, E., Fraiman, R., and Matrán, C. (2007). The random projection method in goodness of fit for functional data. Computational Statistics & Data Analysis, 51(10):4814–4831.
  • Cuesta-Albertos et al. (2006) Cuesta-Albertos, J. A., Fraiman, R., and Ransford, T. (2006). Random projections and goodness-of-fit tests in infinite-dimensional spaces. Bulletin of the Brazilian Mathematical Society, 37(4):477–501.
  • Cuesta-Albertos et al. (2019) Cuesta-Albertos, J. A., García-Portugués, E., Febrero-Bande, M., and González-Manteiga, W. (2019). Goodness-of-fit tests for the functional linear model based on randomly projected empirical processes. The Annals of Statistics, 47(1):439–467.
  • Delsol et al. (2011) Delsol, L., Ferraty, F., and Vieu, P. (2011). Structural test in regression on functional variables.

    Journal of Multivariate Analysis

    , 102(3):422–447.
  • Ditzhaus and Gaigall (2018) Ditzhaus, M. and Gaigall, D. (2018). A consistent goodness-of-fit test for huge dimensional and functional data. Journal of Nonparametric Statistics, 30(4):834–859.
  • Durbin (1973) Durbin, J. (1973). Weak convergence of the sample distribution function when parameters are estimated. The Annals of Statistics, 1(2):279–290.
  • Ferraty and Vieu (2006) Ferraty, F. and Vieu, P. (2006). Nonparametric Functional Data Analysis: Theory and Practice. Springer Series in Statistics. Springer, New York.
  • García-Portugués et al. (2020a) García-Portugués, E., Álvarez-Liébana, J., Álvarez-Pérez, G., and González-Manteiga, W. (2020a). A goodness-of-fit test for the functional linear model with functional response. Scandinavian Journal of Statistics, to appear.
  • García-Portugués et al. (2020b) García-Portugués, E., Álvarez-Liébana, J., Álvarez-Pérez, G., and González-Manteiga, W. (2020b). Goodness-of-fit tests for functional linear models based on integrated projections. In Aneiros, G., Horová, I., Hušková, M., and Vieu, P., editors,

    Functional and High-Dimensional Statistics and Related Fields

    , Contributions to Statistics, pages 107–114. Springer, Cham.
  • García-Portugués et al. (2014) García-Portugués, E., González-Manteiga, W., and Febrero-Bande, M. (2014). A goodness-of-fit test for the functional linear model with scalar response. Journal of Computational and Graphical Statistics, 23(3):761–778.
  • González-Manteiga and Crujeiras (2013) González-Manteiga, W. and Crujeiras, R. M. (2013). An updated review of goodness-of-fit tests for regression models. Test, 22(3):361–411.
  • Górecki et al. (2020) Górecki, T., Horváth, L., and Kokoszka, P. (2020). Tests of normality of functional data. International Statistical Review, 88(3):677–697.
  • Härdle and Mammen (1993) Härdle, W. and Mammen, E. (1993). Comparing nonparametric versus parametric regression fits. The Annals of Statistics, 21(4):1926–1947.
  • Horváth and Kokoszka (2012) Horváth, L. and Kokoszka, P. (2012). Inference for Functional Data with Applications. Springer Series in Statistics. Springer, New York.
  • Horváth and Reeder (2013) Horváth, L. and Reeder, R. (2013). A test of significance in functional quadratic regression. Bernoulli, 19(5A):2130–2151.
  • Jiang et al. (2019) Jiang, Q., Hušková, M., Meintanis, S. G., and Zhu, L. (2019). Asymptotics, finite-sample comparisons and applications for two-sample tests with functional data. Journal of Multivariate Analysis, 170:202–220.
  • Kellner and Celisse (2019) Kellner, J. and Celisse, A. (2019). A one-sample test for normality with kernel methods. Bernoulli, 25(3):1816–1837.
  • Kolkiewicz et al. (2021) Kolkiewicz, A., Rice, G., and Xie, Y. (2021). Projection pursuit based tests of normality with functional data. Journal of Statistical Planning and Inference, 211:326–339.
  • Lai et al. (2020) Lai, T., Zhang, Z., and Wang, Y. (2020). Testing independence and goodness-of-fit jointly for functional linear models. Journal of the Korean Statistical Society, to appear.
  • Lee et al. (2020) Lee, C. E., Zhang, X., and Shao, X. (2020). Testing conditional mean independence for functional data. Biometrika, 107(2):331–346.
  • Lyons (2013) Lyons, R. (2013). Distance covariance in metric spaces. The Annals of Probability, 41(5):3284–3305.
  • Maistre and Patilea (2020) Maistre, S. and Patilea, V. (2020). Testing for the significance of functional covariates. Journal of Multivariate Analysis, 179:104648.
  • McLean et al. (2015) McLean, M. W., Hooker, G., and Ruppert, D. (2015). Restricted likelihood ratio tests for linearity in scalar-on-function regression. Statistics and Computing, 25(5):997–1008.
  • Parzen (1962) Parzen, E. (1962).

    On estimation of a probability density function and mode.

    Annals of Mathematical Statistics, 33(3):1065–1076.
  • Patilea and Sánchez-Sellero (2020) Patilea, V. and Sánchez-Sellero, C. (2020). Testing for lack-of-fit in functional regression models against general alternatives. Journal of Statistical Planning and Inference, 209:229–251.
  • Patilea et al. (2016) Patilea, V., Sánchez-Sellero, C., and Saumard, M. (2016). Testing the predictor effect on a functional response. Journal of the American Statistical Association, 111(516):1684–1695.
  • Qiu et al. (2021) Qiu, Z., Chen, J., and Zhang, J.-T. (2021). Two-sample tests for multivariate functional data with applications. Computational Statistics & Data Analysis, 157:107160.
  • Ramsay and Silverman (2005) Ramsay, J. O. and Silverman, B. W. (2005). Functional Data Analysis. Springer Series in Statistics. Springer, New York.
  • Rosenblatt (1956) Rosenblatt, M. (1956). Remarks on some nonparametric estimates of a density function. Annals of Mathematical Statistics, 27(3):832–837.
  • Sejdinovic et al. (2013) Sejdinovic, D., Sriperumbudur, B., Gretton, A., and Fukumizu, K. (2013). Equivalence of distance-based and RKHS-based statistics in hypothesis testing. The Annals of Statistics, 41(5):2263–2291.
  • Stute (1997) Stute, W. (1997). Nonparametric model checks for regression. The Annals of Statistics, 25(2):613–641.
  • Székely and Rizzo (2017) Székely, G. J. and Rizzo, M. L. (2017). The energy of data. Annual Review of Statistics and its Application, 4(1):447–479.
  • Székely et al. (2007) Székely, G. J., Rizzo, M. L., and Bakirov, N. K. (2007). Measuring and testing dependence by correlation of distances. The Annals of Statistics, 35(6):2769–2794.
  • Wand and Jones (1995) Wand, M. P. and Jones, M. C. (1995). Kernel Smoothing, volume 60 of Monographs on Statistics and Applied Probability. Chapman & Hall, London.
  • Yasemin-Tekbudak et al. (2019) Yasemin-Tekbudak, M., Alfaro-Córdoba, M., Maity, A., and Staicu, A. M. (2019). A comparison of testing methods in scalar-on-function regression. AStA. Advances in Statistical Analysis, 103(3):411–436.