Section: New Results

Semi and non-parametric methods

Deep learning models to study the early stages of Parkinson's Disease

Participants : Florence Forbes, Veronica Munoz Ramirez, Virgilio Kmetzsch Rosa E Silva.

Joint work with: Michel Dojat from Grenoble Institute of Neuroscience.

Current physio-pathological data suggest that Parkinson's Disease (PD) symptoms are related to important alterations in subcortical brain structures. However, structural changes in these small structures remain difficult to detect for neuro-radiologists, in particular, at the early stages of the disease (de novo PD patients) [58], [43], [59]. The absence of a reliable ground truth at the voxel level prevents the application of traditional supervised deep learning techniques. In this work, we consider instead an anomaly detection approach and show that auto-encoders (AE) could provide an efficient anomaly scoring to discriminate de novo PD patients using quantitative Magnetic Resonance Imaging (MRI) data.

Estimation of extreme risk measures

Participants : Stephane Girard, Antoine Usseglio Carleve.

Joint work with: A. Daouia (Univ. Toulouse), L. Gardes (Univ. Strasbourg) and G. Stupfler (Univ. Nottingham, UK).

One of the most popular risk measures is the Value-at-Risk (VaR) introduced in the 1990's. In statistical terms, the VaR at level $\alpha \in (0,1)$ corresponds to the upper $\alpha$-quantile of the loss distribution. The Value-at-Risk however suffers from several weaknesses. First, it provides us only with a pointwise information: VaR($\alpha$) does not take into consideration what the loss will be beyond this quantile. Second, random loss variables with light-tailed distributions or heavy-tailed distributions may have the same Value-at-Risk. Finally, Value-at-Risk is not a coherent risk measure since it is not subadditive in general. A first coherent alternative risk measure is the Conditional Tail Expectation (CTE), also known as Tail-Value-at-Risk, Tail Conditional Expectation or Expected Shortfall in case of a continuous loss distribution. The CTE is defined as the expected loss given that the loss lies above the upper $\alpha$-quantile of the loss distribution. This risk measure thus takes into account the whole information contained in the upper tail of the distribution.

However, the asymptotic normality of the empirical CTE estimator requires that the underlying distribution possess a finite variance; this can be a strong restriction in heavy-tailed models which constitute the favoured class of models in actuarial and financial applications. One possible solution in very heavy-tailed models where this assumption fails could be to use the more robust Median Shortfall, but this quantity is actually just a quantile, which therefore only gives information about the frequency of a tail event and not about its typical magnitude. In [23], we construct a synthetic class of tail $L_p$-medians, which encompasses the Median Shortfall (for $p=1$) and Conditional Tail Expectation (for $p=2$). We show that, for $1<p<2$, a tail $L_p$-median always takes into account both the frequency and magnitude of tail events, and its empirical estimator is, within the range of the data, asymptotically normal under a condition weaker than a finite variance. We extrapolate this estimator, along with another technique, to proper extreme levels using the heavy-tailed framework. The estimators are showcased on a simulation study and on a set of real fire insurance data showing evidence of a very heavy right tail.

A possible coherent alternative risk measure is based on expectiles [6]. Compared to quantiles, the family of expectiles is based on squared rather than absolute error loss minimization. The flexibility and virtues of these least squares analogues of quantiles are now well established in actuarial science, econometrics and statistical finance. have recently received a lot of attention, especially in actuarial and financial risk man-agement. Their estimation, however, typically requires to consider non-explicit asymmetric leasts-quares estimates rather than the traditional order statistics used for quantile estimation. This makes the study of the tail expectile process a lot harder than that of the standard tail quantile process. Under the challenging model of heavy-tailed distributions, we derive joint weighted Gaussian approximations of the tail empirical expectile and quantile processes. We then use this powerful result to introduce and study new estimators of extreme expectiles and the standard quantile-based expected shortfall, as well as a novel expectile-based form of expected shortfall [22].

Both quantiles and expectiles were embedded in the more general class of $L_p$-quantiles [21] as the minimizers of a generic asymmetric convex loss function. It has been proved very recently that the only $L_p$-quantiles that are coherent risk measures are the expectiles. In [75], we work in a context of heavy tails, which is especially relevant to actuarial science, finance, econometrics and natural sciences, and we construct an estimator of the tail index of the underlying distribution based on extreme $L_p$-quantiles. We establish the asymptotic normality of such an estimator and in doing so, we extend very recent results on extreme expectile and $L_p$-quantile estimation. We provide a discussion of the choice of $p$ in practice, as well as a methodology for reducing the bias of our estimator. Its finite-sample performance is evaluated on simulated data and on a set of real hydrological data. This work is submitted for publication.

Conditional extremal events

Participants : Stephane Girard, Antoine Usseglio Carleve.

Joint work with: G. Stupfler (Univ. Nottingham, UK), A. Ahmad, E. Deme and A. Diop (Université Gaston Berger, Sénégal).

The goal of the PhD thesis of Aboubacrene Ag Ahmad is to contribute to the development of theoretical and algorithmic models to tackle conditional extreme value analysis, ie the situation where some covariate information $X$ is recorded simultaneously with a quantity of interest $Y$. In such a case, extreme quantiles and expectiles are functions of the covariate. In [13], we consider a location-scale model for conditional heavy-tailed distributions when the covariate is deterministic. First, nonparametric estimators of the location and scale functions are introduced. Second, an estimator of the conditional extreme-value index is derived. The asymptotic properties of the estimators are established under mild assumptions and their finite sample properties are illustrated both on simulated and real data.

As explained in Paragraph 7.2.2, expectiles have recently started to be considered as serious candidates to become standard tools in actuarial and financial risk management. However, expectiles and their sample versions do not benefit from a simple explicit form, making their analysis significantly harder than that of quantiles and order statistics. This difficulty is compounded when one wishes to integrate auxiliary information about the phenomenon of interest through a finite-dimensional covariate, in which case the problem becomes the estimation of conditional expectiles. In [74], we exploit the fact that the expectiles of a distribution $F$ are in fact the quantiles of another distribution $E$ explicitly linked to $F$, in order to construct nonparametric kernel estimators of extreme conditional expectiles. We analyze the asymptotic properties of our estimators in the context of conditional heavy-tailed distributions. Applications to simulated data and real insurance data are provided. The results are submitted for publication.

Estimation of the variability in the distribution tail

Participant : Stephane Girard.

Joint work with: L. Gardes (Univ. Strasbourg).

We propose a new measure of variability in the tail of a distribution by applying a Box-Cox transformation of parameter $p\ge 0$ to the tail-Gini functional. It is shown that the so-called Box-Cox Tail Gini Variability measure is a valid variability measure whose condition of existence may be as weak as necessary thanks to the tuning parameter $p$. The tail behaviour of the measure is investigated under a general extreme-value condition on the distribution tail. We then show how to estimate the Box-Cox Tail Gini Variability measure within the range of the data. These methods provide us with basic estimators that are then extrapolated using the extreme-value assumption to estimate the variability in the very far tails. The finite sample behavior of the estimators is illustrated both on simulated and real data. This work is submitted for publication [72].

Extrapolation limits associated with extreme-value methods

Participant : Stephane Girard.

Joint work with: L. Gardes (Univ. Strasbourg) and A. Dutfoy (EDF R&D).

The PhD thesis of Clément Albert (co-funded by EDF) is dedicated to the study of the sensitivity of extreme-value methods to small changes in the data and to their extrapolation ability. Two directions are explored:

(i) In [15], we investigate the asymptotic behavior of the (relative) extrapolation error associated with some estimators of extreme quantiles based on extreme-value theory. It is shown that the extrapolation error can be interpreted as the remainder of a first order Taylor expansion. Necessary and sufficient conditions are then provided such that this error tends to zero as the sample size increases. Interestingly, in case of the so-called Exponential Tail estimator, these conditions lead to a subdivision of Gumbel maximum domain of attraction into three subsets. In constrast, the extrapolation error associated with Weissman estimator has a common behavior over the whole Fréchet maximum domain of attraction. First order equivalents of the extrapolation error are then derived and their accuracy is illustrated numerically.

(ii) In [14], We propose a new estimator for extreme quantiles under the log-generalized Weibull-tail model, introduced by Cees de Valk. This model relies on a new regular variation condition which, in some situations, permits to extrapolate further into the tails than the classical assumption in extreme-value theory. The asymptotic normality of the estimator is established and its finite sample properties are illustrated both on simulated and real datasets.

Bayesian inference for copulas

Participants : Julyan Arbel, Marta Crispino, Stephane Girard.

We study in [16] a broad class of asymmetric copulas known as Liebscher copulas and defined as a combination of multiple—usually symmetric—copulas. The main thrust of this work is to provide new theoretical properties including exact tail dependence expressions and stability properties. A subclass of Liebscher copulas obtained by combining Fréchet copulas is studied in more details. We establish further dependence properties for copulas of this class and show that they are characterized by an arbitrary number of singular components. Furthermore, we introduce a novel iterative construction for general Liebscher copulas which de facto insures uniform margins, thus relaxing a constraint of Liebscher's original construction. Besides, we show that this iterative construction proves useful for inference by developing an Approximate Bayesian computation sampling scheme. This inferential procedure is demonstrated on simulated data.

Approximations of Bayesian nonparametric models

Participant : Julyan Arbel.

Joint work with: Stefano Favaro and Pierpaolo De Blasi from Collegio Carlo Alberto, Turin, Italy, Igor Prunster from Bocconi University, Milan, Italy, Caroline Lawless from Université Paris-Dauphine, France, Olivier Marchal from Université Jean Monnet.

For a long time, the Dirichlet process has been the gold standard discrete random measure in Bayesian nonparametrics. The Pitman–Yor process provides a simple and mathematically tractable generalization, allowing for a very flexible control of the clustering behaviour. Two commonly used representations of the Pitman–Yor process are the stick-breaking process and the Chinese restaurant process. The former is a constructive representation of the process which turns out very handy for practical implementation, while the latter describes the partition distribution induced. Obtaining one from the other is usually done indirectly with use of measure theory. In contrast, we propose in [25] an elementary proof of Pitman–Yor's Chinese Restaurant process from its stick-breaking representation.

In [17], we consider approximations to the popular Pitman-Yor process obtained by truncating the stick-breaking representation. The truncation is determined by a random stopping rule that achieves an almost sure control on the approximation error in total variation distance. We derive the asymptotic distribution of the random truncation point as the approximation error goes to zero in terms of a polynomially tilted positive stable random variable. The practical usefulness and effectiveness of this theoretical result is demonstrated by devising a sampling algorithm to approximate functionals of the-version of the Pitman–Yor process.

In [18], we approximate predictive probabilities of Gibbs-type random probability measures, or Gibbs-type priors, which are arguably the most “natural” generalization of the celebrated Dirichlet prior. Among them the Pitman–Yor process certainly stands out for the mathematical tractability and interpretability of its predictive probabilities, which made it the natural candidate in several applications. Given a sample of size $n$, in this paper we show that the predictive probabilities of any Gibbs-type prior admit a large $n$ approximation, with an error term vanishing as $o(1/n)$, which maintains the same desirable features as the predictive probabilities of the Pitman–Yor process.

In [18], we prove a monotonicity property of the Hurwitz zeta function which, in turn, translates into a chain of inequalities for polygamma functions of different orders. We provide a probabilistic interpretation of our result by exploiting a connection between Hurwitz zeta function and the cumulants of the exponential-beta distribution.

Concentration inequalities

Participant : Julyan Arbel.

Joint work with: Olivier Marchal from Université Jean Monnet and Hien Nguyen from La Trobe University Melbourne Australia.

In [19], we investigate the sub-Gaussian property for almost surely bounded random variables. If sub-Gaussianity per se is de facto ensured by the bounded support of said random variables, then exciting research avenues remain open. Among these questions is how to characterize the optimal sub-Gaussian proxy variance? Another question is how to characterize strict sub-Gaussianity, defined by a proxy variance equal to the (standard) variance? We address the questions in proposing conditions based on the study of functions variations. A particular focus is given to the relationship between strict sub-Gaussianity and symmetry of the distribution. In particular, we demonstrate that symmetry is neither sufficient nor necessary for strict sub-Gaussianity. In contrast, simple necessary conditions on the one hand, and simple sufficient conditions on the other hand, for strict sub-Gaussianity are provided. These results are illustrated via various applications to a number of bounded random variables, including Bernoulli, beta, binomial, uniform, Kumaraswamy, and triangular distributions.

Extraction and data analysis toward "industry of the future"

Participants : Florence Forbes, Hongliang Lu, Fatima Fofana.

Joint work with: J. F. Cuccaro and J. C Trochet from Vi-Technology company.

The overall idea of this project with Vi-Technology is to work towards manufacturing processes where machines communicate automatically so as to optimize the process performance as a whole. Starting from the assumption that transmitted information is essentially of statistical nature, the role of mistis in this context was to identify what statistical methods might be useful for the printed circuits boards assembly industry. A first step was to extract and analyze data from two inspection machines in an industrial process making electronic cards. After a first extraction in the SQL database, the goal was to enlighten the statistical links between these machines. Preliminary experiments and results on the Solder Paste Inspection (SPI) step, at the beginning of the line, helped identifying potentially relevant variables and measurements (eg related to stencil offsets) to identify future defects and discriminate between them. More generally, we had access to two databases at both ends (SPI and Component Inspection) of the assembly process. The goal was to improve our understanding of interactions in the assembly process, find out correlations between defects and physical measures, generate proactive alarms so as to detect departures from normality.

Tracking and analysis of large population of dynamic single molecules

Participant : Florence Forbes.

Joint work with: Virginie Stoppin-Mellet from Grenoble Institute of Neuroscience, Vincent Brault from Laboratoire Jean Kuntzmann, Emilie Lebarbier from Nanterre University and Guy Bendao from AgroParisTech.

In the last decade, the number of studies using single molecule approaches has increased significantly. Thanks to technological progress and in particular with the development of TIRFM (Total Internal Reflection Fluorescence Microscopy), biologists can now observe single molecules at work. However, real time single molecule approaches remain mastered by a limited number of labs, and challenging obstacles have to be overcome before it becomes more broadly accessible. One important issue is the efficient detection and tracking of individual molecules in noisy images (low signal-to-noise ratio, SNR). Considering for example a TIRFM movie where single molecules stochastically appear and disappear at random positions, the low SNR implies that each individual molecule has to be detected at sub-pixel resolution over its local background and that this operation has to be repeated on each frame of the movie, thus requiring considerable amount of calculations. Procedures to detect single molecules are available, but they are mostly applicable to immobile molecules, are not statistically robust, and they often require an image processing that alters the quantitative signal information. In particular the intensity of a signal might be modified so that it becomes difficult to know the number of molecules associated with a specific signal. Crucial information such as the stoichiometry of the molecular complexes are then lost. Another challenging issue concerns data processing. Molecule tracking generate traces of time-dependent intensity fluctuations for each molecule. But single traces contain limited amount of information, and thus a very large number of traces must be analysed to extract general rules. In this context, the first aim of the present project was to provide a general procedure to track in real time transient interactions of a large number of biological molecules observed with TIRF microscopy and to generate traces of time-dependent intensity fluctuations. The second aim was to define a robust statistical approach to detect discrete events in a noisy time-dependent signal and extract parameters that describe the kinetics of these events. For this task we gathered expertise from biology (Grenoble Institute of Neuroscience) and statistics (Inria Mistis, LJK and AgroParisTech) in the context of a multidisciplinary project funded by the Grenoble data institute for 2 years.