Skip to Main content Skip to Navigation

Vers la résolution "optimale" de problèmes inverses non linéaires parcimonieux grâce à l'exploitation de variables binaires sur dictionnaires continus : applications en astrophysique

Abstract : This thesis deals with solutions of nonlinear inverse problems using a sparsity prior; more specifically when the data can be modelled as a linear combination of a few functions, which depend non-linearly on a "location" parameter, i.e. frequencies for spectral analysis or time-delay for spike train deconvolution. These problems are generally reformulated as linear sparse approximation problems, thanks to an evaluation of the nonlinear functions at location parameters discretised on a thin grid, building a "discrete dictionary". However, such an approach has two major drawbacks. On the one hand, the discrete dictionary is highly correlated; classical sub-optimal methods such as L1- penalisation or greedy algorithms can then fail. On the other hand, the estimated location parameter, which belongs to the discretisation grid, is necessarily discrete and that leads to model errors. To deal with these issues, we propose in this work an exact sparsity model, thanks to the introduction of binary variables, and an optimal solution of the problem with a "continuous dictionary" allowing a continuous estimation of the location parameter. We focus on two research axes, which we illustrate with problems such as spike train deconvolution and spectral analysis of unevenly sampled data. The first axis focusses on the "dictionary interpolation" principle, which consists in a linearisation of the continuous dictionary in order to get a constrained linear sparse approximation problem. The introduction of binary variables allows us to reformulate this problem as a "Mixed Integer Program" (MIP) and to exactly model the sparsity thanks to the "pseudo-norm L0". We study different kinds of dictionary interpolation and constraints relaxation, in order to solve the problem optimally thanks to MIP classical algorithms. For the second axis, in a Bayesian framework, the binary variables are supposed random with a Bernoulli distribution and we model the sparsity through a Bernoulli-Gaussian prior. This model is extended to take into account continuous location parameters (BGE model). We then estimate the parameters from samples drawn using Markov chain Monte Carlo algorithms. In particular, we show that marginalising the amplitudes allows us to improve the sampling of a Gibbs algorithm in a supervised case (when the model's hyperparameters are known). In an unsupervised case, we propose to take advantage of such a marginalisation through a "Partially Collapsed Gibbs Sampler." Finally, we adapt the BGE model and associated samplers to a topical science case in Astrophysics: the detection of exoplanets from radial velocity measurements. The efficiency of our method will be illustrated with simulated data, as well as actual astrophysical data.
Document type :
Complete list of metadatas

Cited literature [212 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Friday, October 11, 2019 - 5:13:06 PM
Last modification on : Thursday, October 15, 2020 - 4:07:15 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02314077, version 1



Mégane Boudineau. Vers la résolution "optimale" de problèmes inverses non linéaires parcimonieux grâce à l'exploitation de variables binaires sur dictionnaires continus : applications en astrophysique. Astrophysique [astro-ph]. Université Paul Sabatier - Toulouse III, 2019. Français. ⟨NNT : 2019TOU30020⟩. ⟨tel-02314077⟩



Record views


Files downloads