Dans ce manuscrit, nous avons étudié l’extension de l’analyse multi-résolution aux images couleurs par des transformées sur graphe. Dans ce cadre, nous avons déployé trois stratégies d’analyse différentes. En premier lieu, nous avons défini une transformée basée sur l’utilisation d’un graphe perceptuel dans l’analyse à travers la transformé en ondelettes spectrale sur graphe. L’application en débruitage d’image met en évidence l’utilisation du SVH dans l’analyse des images couleurs. La deuxième stratégie consiste à proposer une nouvelle méthode d’inpainting pour des images couleurs. Pour cela, nous avons proposé un schéma de régularisation à travers les coefficients d’ondelettes de la TOSG, l’estimation de la structure manquante se fait par la construction d’un graphe des patchs couleurs à partir des moyenne non locales. Les résultats obtenus sont très encourageants et mettent en évidence l’importance de la prise en compte du SVH. Dans la troisième stratégie, nous proposons une nouvelleapproche de décomposition d’un signal défini sur un graphe complet. Cette méthode est basée sur l’utilisation des propriétés de la matrice laplacienne associée au graphe complet. Dans le contexte des images couleurs, la prise en compte de la dimension couleur est indispensable pour pouvoir identifier les singularités liées à l’image. Cette dernière offre de nouvelles perspectives pour une étude approfondie de son comportement. / In our work, we studied the extension of the multi-resolution analysis for color images by using transforms on graphs. In this context, we deployed three different strategies of analysis. Our first approach consists of computing the graph of an image using the psychovisual information and analyzing it by using the spectral graph wavelet transform. We thus have defined a wavelet transform based on a graph with perceptual information by using the CIELab color distance. Results in image restoration highlight the interest of the appropriate use of color information. In the second strategy, we propose a novel recovery algorithm for image inpainting represented in the graph domain. Motivated by the efficiency of the wavelet regularization schemes and the success of the nonlocal means methods we construct an algorithm based on the recovery of information in the graph wavelet domain. At each step the damaged structure are estimated by computing the non local graph then we apply the graph wavelet regularization model using the SGWT coefficient. The results are very encouraging and highlight the use of the perceptual informations. In the last strategy, we propose a new approach of decomposition for signals defined on a complete graphs. This method is based on the exploitation of of the laplacian matrix proprieties of the complete graph. In the context of image processing, the use of the color distance is essential to identify the specificities of the color image. This approach opens new perspectives for an in-depth study of its behavior.

This thesis deals with optimizing extremal eigenvalues of weighted graph Laplacian matrices. In general, the Laplacian matrix of a (weighted) graph is of particular importance in spectral graph theory and combinatorial optimization (e.g., graph partition like max-cut and graph bipartition). Especially the pioneering work of M. Fiedler investigates extremal eigenvalues of weighted graph Laplacians and provides close connections to the node- and edge-connectivity of a graph. Motivated by Fiedler, Göring et al. were interested in further connections between structural properties of the graph and the eigenspace of the second smallest eigenvalue of weighted graph Laplacians using a semidefinite optimization approach. By redistributing the edge weights of a graph, the following three optimization problems are studied in this thesis: maximizing the second smallest eigenvalue (based on the mentioned work of Göring et al.), minimizing the maximum eigenvalue and minimizing the difference of maximum and second smallest eigenvalue of the weighted Laplacian. In all three problems a semidefinite optimization formulation allows to interpret the corresponding semidefinite dual as a graph realization problem. That is, to each node of the graph a vector in the Euclidean space is assigned, fulfilling some constraints depending on the considered problem. Optimal realizations are investigated and connections to the eigenspaces of corresponding optimized eigenvalues are established. Furthermore, optimal realizations are closely linked to the separator structure of the graph. Depending on this structure, on the one hand folding properties of optimal realizations are characterized and on the other hand the existence of optimal realizations of bounded dimension is proven. The general bounds depend on the tree-width of the graph. In the case of minimizing the maximum eigenvalue, an important family of graphs are bipartite graphs, as an optimal one-dimensional realization may be constructed. Taking the symmetry of the graph into account, a particular optimal edge weighting exists. Considering the coupled problem, i.e., minimizing the difference of maximum and second smallest eigenvalue and the single problems, i.e., minimizing the maximum and maximizing the second smallest eigenvalue, connections between the feasible (optimal) sets are established.

Cette thèse s'inscrit dans la théorie des matrices aléatoires, à l'intersection avec la théorie des probabilités libres et des algèbres d'opérateurs. Elle s'insère dans une démarche générale qui a fait ses preuves ces dernières décennies : importer les techniques et les concepts de la théorie des probabilités non commutatives pour l'étude du spectre de grandes matrices aléatoires. On s'intéresse ici à des généralisations du théorème de liberté asymptotique de Voiculescu. Dans les Chapitres 1 et 2, nous montrons des résultats de liberté asymptotique forte pour des matrices gaussiennes, unitaires aléatoires et déterministes. Dans les Chapitres 3 et 4, nous introduisons la notion de fausse liberté asymptotique pour des matrices déterministes et certaines matrices hermitiennes à entrées sous diagonales indépendantes, interpolant les modèles de matrices de Wigner et de Lévy. / The thesis fits into the random matrix theory, in intersection with free probability and operator algebra. It is part of a general approach which is common since the last decades: using tools and concepts of non commutative probability in order to get general results about the spectrum of large random matrices. Where are interested here in generalization of Voiculescu's asymptotic freeness theorem. In Chapter 1 and 2, we show some results of strong asymptotic freeness for gaussian, random unitary and deterministic matrices. In Chapter 3 and 4, we introduce the notion of asymptotic false freeness for deterministic matrices and certain random matrices, Hermitian with independent sub-diagonal entries, interpolating Wigner and Lévy models.

This thesis deals with optimizing extremal eigenvalues of weighted graph Laplacian matrices. In general, the Laplacian matrix of a (weighted) graph is of particular importance in spectral graph theory and combinatorial optimization (e.g., graph partition like max-cut and graph bipartition). Especially the pioneering work of M. Fiedler investigates extremal eigenvalues of weighted graph Laplacians and provides close connections to the node- and edge-connectivity of a graph. Motivated by Fiedler, Göring et al. were interested in further connections between structural properties of the graph and the eigenspace of the second smallest eigenvalue of weighted graph Laplacians using a semidefinite optimization approach. By redistributing the edge weights of a graph, the following three optimization problems are studied in this thesis: maximizing the second smallest eigenvalue (based on the mentioned work of Göring et al.), minimizing the maximum eigenvalue and minimizing the difference of maximum and second smallest eigenvalue of the weighted Laplacian. In all three problems a semidefinite optimization formulation allows to interpret the corresponding semidefinite dual as a graph realization problem. That is, to each node of the graph a vector in the Euclidean space is assigned, fulfilling some constraints depending on the considered problem. Optimal realizations are investigated and connections to the eigenspaces of corresponding optimized eigenvalues are established. Furthermore, optimal realizations are closely linked to the separator structure of the graph. Depending on this structure, on the one hand folding properties of optimal realizations are characterized and on the other hand the existence of optimal realizations of bounded dimension is proven. The general bounds depend on the tree-width of the graph. In the case of minimizing the maximum eigenvalue, an important family of graphs are bipartite graphs, as an optimal one-dimensional realization may be constructed. Taking the symmetry of the graph into account, a particular optimal edge weighting exists. Considering the coupled problem, i.e., minimizing the difference of maximum and second smallest eigenvalue and the single problems, i.e., minimizing the maximum and maximizing the second smallest eigenvalue, connections between the feasible (optimal) sets are established.

We consider the problem of maximizing the second smallest eigenvalue of the weighted Laplacian of a (simple) graph over all nonnegative edge weightings with bounded total weight. We generalize this problem by introducing node significances and edge lengths. We give a formulation of this generalized problem as a semidefinite program. The dual program can be equivalently written as embedding problem. This is fifinding an embedding of the n nodes of the graph in n-space so that their barycenter is at the origin, the distance between adjacent nodes is bounded by the respective edge length, and the embedded nodes are spread as much as possible. (The sum of the squared norms is maximized.) We proof the following necessary condition for optimal embeddings. For any separator of the graph at least one of the components fulfills the following property: Each straight-line segment between the origin and an embedded node of the component intersects the convex hull of the embedded nodes of the separator. There exists always an optimal embedding of the graph whose dimension is bounded by the tree-width of the graph plus one. We defifine the rotational dimension of a graph. This is the minimal dimension k such that for all choices of the node significances and edge lengths an optimal embedding of the graph can be found in k-space. The rotational dimension of a graph is a minor monotone graph parameter. We characterize the graphs with rotational dimension up to two.:1 Introduction 1.1 Notations and Preliminaries 1.2 The Algebraic Connectivity 1.3 Two applications 1.4 Outline 2 The Embedding Problem 2.1 Semidefinite formulation 2.2 The dual as geometric embedding problem 2.3 Physical interpretation and examples 2.4 Formulation without fifixed barycenter 3 Geometrical Operations 3.1 Congruent transformations 3.2 Folding a flat halfspace 3.3 Folding and Collapsing 4 Structural properties of optimal embeddings 4.1 Separator-Shadow 4.2 Separators containing the origin 4.3 The tree-width bound 4.4 Application to trees 5 The Rotational Dimension of a graph 5.1 Defifinition and basic properties 5.2 Characterization of graphs with small rotational dimension 5.3 The Colin de Verdi ere graph parameter List of Figures Bibliography Theses

O presente texto descreve a tese de doutorado intitulada Análise de Formas usando Wavelets em Grafos. O tema está relacionado à área de Visão Computacional, particularmente aos tópicos de Caracterização, Descrição e Classificação de Formas. Dentre os métodos da extensa literatura em Análise de Formas 2D, percebe-se uma presença menor daqueles baseados em grafos com topologia arbitrária e irregular. As contribuições desta tese procuram preencher esta lacuna. É proposta uma metodologia baseada no seguinte pipeline : (i) Amostragem da forma, (ii) Estruturação das amostras em grafos, (iii) Função-base definida nos vértices, (iv) Análise multiescala de grafos por meio da Transformada Wavelet Espectral em grafos, (v) Extração de Características da Transformada Wavelet e (vi) Discriminação. Para cada uma das etapas (i), (ii), (iii), (v) e (vi), são inúmeras as abordagens possíveis. Um dos desafios é encontrar uma combinação de abordagens, dentre as muitas alternativas, que resulte em um pipeline eficaz para nossos propósitos. Em particular, para a etapa (iii), dado um grafo que representa uma forma, o desafio é identificar uma característica associada às amostras que possa ser definida sobre os vértices do grafo. Esta característica deve capturar a influência subjacente da estrutura combinatória de toda a rede sobre cada vértice, em diversas escalas. A Transformada Wavelet Espectral sobre os Grafos revelará esta influência subjacente em cada vértice. São apresentados resultados obtidos de experimentos usando formas 2D de benchmarks conhecidos na literatura, bem como de experimentos de aplicações em astronomia para análise de formas de galáxias do Sloan Digital Sky Survey não-rotuladas e rotuladas pelo projeto Galaxy Zoo 2 , demonstrando o sucesso da técnica proposta, comparada a abordagens clássicas como Transformada de Fourier e Transformada Wavelet Contínua 2D. / This document describes the PhD thesis entitled Shape Analysis by using Wavelets on Graphs. The addressed theme is related to Computer Vision, particularly to the Characterization, Description and Classication topics. Amongst the methods presented in an extensive literature on Shape Analysis 2D, it is perceived a smaller presence of graph-based methods with arbitrary and irregular topologies. The contributions of this thesis aim at fullling this gap. A methodology based on the following pipeline is proposed: (i) Shape sampling, (ii) Samples structuring in graphs, (iii) Function dened on vertices, (iv) Multiscale analysis of graphs through the Spectral Wavelet Transform, (v) Features extraction from the Wavelet Transforms and (vi) Classication. For the stages (i), (ii), (iii), (v) and (vi), there are numerous possible approaches. One great challenge is to nd a proper combination of approaches from the several available alternatives, which may be able to yield an eective pipeline for our purposes. In particular, for the stage (iii), given a graph representing a shape, the challenge is to identify a feature, which may be dened over the graph vertices. This feature should capture the underlying inuence from the combinatorial structure of the entire network over each vertex, in multiple scales. The Spectral Graph Wavelet Transform will reveal such an underpining inuence over each vertex. Yielded results from experiments on 2D benchmarks shapes widely known in literature, as well as results from astronomy applications to the analysis of unlabeled galaxies shapes from the Sloan Digital Sky Survey and labeled galaxies shapes by the Galaxy Zoo 2 Project are presented, demonstrating the achievements of the proposed technique, in comparison to classic approaches such as the 2D Fourier Transform and the 2D Continuous Wavelet Transform.

The interaction between the study of geometric and analytic aspects of Riemannian manifolds and that of graphs is a very amazing subject. The study of synthetic curvature notions on graphs adds new contributions to this topic. In this thesis, we mainly study two kinds of synthetic curvature notions: the Ollivier-Ricci cuvature on locally finite graphs and the combinatorial curvature on infinite semiplanar graphs. In the first part, we study the Ollivier-Ricci curvature. As known in Riemannian geometry, a lower Ricci curvature bound prevents geodesics from diverging too fast on average. We translate this Riemannian idea into a combinatorial setting using the Olliver-Ricci curvature notion. Note that on a graph, the analogue of geodesics starting in different directions, but eventually approaching each other again, would be a triangle. We derive lower and upper Ollivier-Ricci curvature bounds on graphs in terms of number of triangles, which is sharp for instance for complete graphs. We then describe the relation between Ollivier-Ricci curvature and the local clustering coefficient, which is an important concept in network analysis introduced by Watts-Strogatz. Furthermore, positive lower boundedness of Ollivier-Ricci curvature for neighboring vertices imply the existence of at least one triangle. It turns out that the existence of triangles can also improve Lin-Yau\'s curvature dimension inequality on graphs and then produce an implication from Ollivier-Ricci curvature lower boundedness to the curvature dimension inequality. The existence of triangles prevents a graph from being bipartite. A finite graph is bipartite if and only if its largest eigenvalue equals 2. Therefore it is natural that Ollivier-Ricci curvature is closely related to the largest eigenvalue estimates. We combine Ollivier-Ricci curvature notion with the neighborhood graph method developed by Bauer-Jost to study the spectrum estimates of a finite graph. We can always obtain nontrivial estimates on a non-bipartite graph even if its curvature is nonpositive. This answers one of Ollivier\'s open problem in the finite graph setting. In the second part of this thesis, we study systematically infinite semiplanar graphs with nonnegative combinatorial curvature. Unlike the previous Gauss-Bonnet formula approach, we explore an Alexandrov approach based on the observation that the nonnegative combinatorial curvature on a semiplanar graph is equivalent to nonnegative Alexandrov curvature on the surface obtained by replacing each face by a regular polygon of side length one with the same facial degree and gluing the polygons along common edges. Applying Cheeger-Gromoll splitting theorem on the surface, we give a metric classification of infinite semiplanar graphs with nonnegative curvature. We also construct the graphs embedded into the projective plane minus one point. Those constructions answer a question proposed by Chen. We further prove the volume doubling property and Poincare inequality which make the running of Nash-Moser iteration possible. We in particular explore the volume growth behavior on Archimedean tilings on a plane and prove that they satisfy a weak version of relative volume comparison with constant 1. With the above two basic inequalities in hand, we study the geometric function theory of infinite semiplanar graphs with nonnegative curvature. We obtain the Liouville type theorem for positive harmonic functions, the parabolicity. We also prove a dimension estimate for polynomial growth harmonic functions, which is an extension of the solution of Colding-Minicozzi of a conjecture of Yau in Riemannian geometry.

O presente texto descreve a tese de doutorado intitulada Análise de Formas usando Wavelets em Grafos. O tema está relacionado à área de Visão Computacional, particularmente aos tópicos de Caracterização, Descrição e Classificação de Formas. Dentre os métodos da extensa literatura em Análise de Formas 2D, percebe-se uma presença menor daqueles baseados em grafos com topologia arbitrária e irregular. As contribuições desta tese procuram preencher esta lacuna. É proposta uma metodologia baseada no seguinte pipeline : (i) Amostragem da forma, (ii) Estruturação das amostras em grafos, (iii) Função-base definida nos vértices, (iv) Análise multiescala de grafos por meio da Transformada Wavelet Espectral em grafos, (v) Extração de Características da Transformada Wavelet e (vi) Discriminação. Para cada uma das etapas (i), (ii), (iii), (v) e (vi), são inúmeras as abordagens possíveis. Um dos desafios é encontrar uma combinação de abordagens, dentre as muitas alternativas, que resulte em um pipeline eficaz para nossos propósitos. Em particular, para a etapa (iii), dado um grafo que representa uma forma, o desafio é identificar uma característica associada às amostras que possa ser definida sobre os vértices do grafo. Esta característica deve capturar a influência subjacente da estrutura combinatória de toda a rede sobre cada vértice, em diversas escalas. A Transformada Wavelet Espectral sobre os Grafos revelará esta influência subjacente em cada vértice. São apresentados resultados obtidos de experimentos usando formas 2D de benchmarks conhecidos na literatura, bem como de experimentos de aplicações em astronomia para análise de formas de galáxias do Sloan Digital Sky Survey não-rotuladas e rotuladas pelo projeto Galaxy Zoo 2 , demonstrando o sucesso da técnica proposta, comparada a abordagens clássicas como Transformada de Fourier e Transformada Wavelet Contínua 2D. / This document describes the PhD thesis entitled Shape Analysis by using Wavelets on Graphs. The addressed theme is related to Computer Vision, particularly to the Characterization, Description and Classication topics. Amongst the methods presented in an extensive literature on Shape Analysis 2D, it is perceived a smaller presence of graph-based methods with arbitrary and irregular topologies. The contributions of this thesis aim at fullling this gap. A methodology based on the following pipeline is proposed: (i) Shape sampling, (ii) Samples structuring in graphs, (iii) Function dened on vertices, (iv) Multiscale analysis of graphs through the Spectral Wavelet Transform, (v) Features extraction from the Wavelet Transforms and (vi) Classication. For the stages (i), (ii), (iii), (v) and (vi), there are numerous possible approaches. One great challenge is to nd a proper combination of approaches from the several available alternatives, which may be able to yield an eective pipeline for our purposes. In particular, for the stage (iii), given a graph representing a shape, the challenge is to identify a feature, which may be dened over the graph vertices. This feature should capture the underlying inuence from the combinatorial structure of the entire network over each vertex, in multiple scales. The Spectral Graph Wavelet Transform will reveal such an underpining inuence over each vertex. Yielded results from experiments on 2D benchmarks shapes widely known in literature, as well as results from astronomy applications to the analysis of unlabeled galaxies shapes from the Sloan Digital Sky Survey and labeled galaxies shapes by the Galaxy Zoo 2 Project are presented, demonstrating the achievements of the proposed technique, in comparison to classic approaches such as the 2D Fourier Transform and the 2D Continuous Wavelet Transform.

In this thesis, we propose many developments in the context of Structural Similarity. We address both node (local) similarity and graph (global) similarity. Concerning node similarity, we focus on improving the diffusive process leading to compute this similarity (e.g. Commute Times) by means of modifying or rewiring the structure of the graph (Graph Densification), although some advances in Laplacian-based ranking are also included in this document. Graph Densification is a particular case of what we call graph rewiring, i.e. a novel field (similar to image processing) where input graphs are rewired to be better conditioned for the subsequent pattern recognition tasks (e.g. clustering). In the thesis, we contribute with an scalable an effective method driven by Dirichlet processes. We propose both a completely unsupervised and a semi-supervised approach for Dirichlet densification. We also contribute with new random walkers (Return Random Walks) that are useful structural filters as well as asymmetry detectors in directed brain networks used to make early predictions of Alzheimer's disease (AD). Graph similarity is addressed by means of designing structural information channels as a means of measuring the Mutual Information between graphs. To this end, we first embed the graphs by means of Commute Times. Commute times embeddings have good properties for Delaunay triangulations (the typical representation for Graph Matching in computer vision). This means that these embeddings can act as encoders in the channel as well as decoders (since they are invertible). Consequently, structural noise can be modelled by the deformation introduced in one of the manifolds to fit the other one. This methodology leads to a very high discriminative similarity measure, since the Mutual Information is measured on the manifolds (vectorial domain) through copulas and bypass entropy estimators. This is consistent with the methodology of decoupling the measurement of graph similarity in two steps: a) linearizing the Quadratic Assignment Problem (QAP) by means of the embedding trick, and b) measuring similarity in vector spaces. The QAP problem is also investigated in this thesis. More precisely, we analyze the behaviour of $m$-best Graph Matching methods. These methods usually start by a couple of best solutions and then expand locally the search space by excluding previous clamped variables. The next variable to clamp is usually selected randomly, but we show that this reduces the performance when structural noise arises (outliers). Alternatively, we propose several heuristics for spanning the search space and evaluate all of them, showing that they are usually better than random selection. These heuristics are particularly interesting because they exploit the structure of the affinity matrix. Efficiency is improved as well. Concerning the application domains explored in this thesis we focus on object recognition (graph similarity), clustering (rewiring), compression/decompression of graphs (links with Extremal Graph Theory), 3D shape simplification (sparsification) and early prediction of AD. / Ministerio de Economía, Industria y Competitividad (Referencia TIN2012-32839 BES-2013-064482)

The interaction between the study of geometric and analytic aspects of Riemannian manifolds and that of graphs is a very amazing subject. The study of synthetic curvature notions on graphs adds new contributions to this topic. In this thesis, we mainly study two kinds of synthetic curvature notions: the Ollivier-Ricci cuvature on locally finite graphs and the combinatorial curvature on infinite semiplanar graphs. In the first part, we study the Ollivier-Ricci curvature. As known in Riemannian geometry, a lower Ricci curvature bound prevents geodesics from diverging too fast on average. We translate this Riemannian idea into a combinatorial setting using the Olliver-Ricci curvature notion. Note that on a graph, the analogue of geodesics starting in different directions, but eventually approaching each other again, would be a triangle. We derive lower and upper Ollivier-Ricci curvature bounds on graphs in terms of number of triangles, which is sharp for instance for complete graphs. We then describe the relation between Ollivier-Ricci curvature and the local clustering coefficient, which is an important concept in network analysis introduced by Watts-Strogatz. Furthermore, positive lower boundedness of Ollivier-Ricci curvature for neighboring vertices imply the existence of at least one triangle. It turns out that the existence of triangles can also improve Lin-Yau\''s curvature dimension inequality on graphs and then produce an implication from Ollivier-Ricci curvature lower boundedness to the curvature dimension inequality. The existence of triangles prevents a graph from being bipartite. A finite graph is bipartite if and only if its largest eigenvalue equals 2. Therefore it is natural that Ollivier-Ricci curvature is closely related to the largest eigenvalue estimates. We combine Ollivier-Ricci curvature notion with the neighborhood graph method developed by Bauer-Jost to study the spectrum estimates of a finite graph. We can always obtain nontrivial estimates on a non-bipartite graph even if its curvature is nonpositive. This answers one of Ollivier\''s open problem in the finite graph setting. In the second part of this thesis, we study systematically infinite semiplanar graphs with nonnegative combinatorial curvature. Unlike the previous Gauss-Bonnet formula approach, we explore an Alexandrov approach based on the observation that the nonnegative combinatorial curvature on a semiplanar graph is equivalent to nonnegative Alexandrov curvature on the surface obtained by replacing each face by a regular polygon of side length one with the same facial degree and gluing the polygons along common edges. Applying Cheeger-Gromoll splitting theorem on the surface, we give a metric classification of infinite semiplanar graphs with nonnegative curvature. We also construct the graphs embedded into the projective plane minus one point. Those constructions answer a question proposed by Chen. We further prove the volume doubling property and Poincare inequality which make the running of Nash-Moser iteration possible. We in particular explore the volume growth behavior on Archimedean tilings on a plane and prove that they satisfy a weak version of relative volume comparison with constant 1. With the above two basic inequalities in hand, we study the geometric function theory of infinite semiplanar graphs with nonnegative curvature. We obtain the Liouville type theorem for positive harmonic functions, the parabolicity. We also prove a dimension estimate for polynomial growth harmonic functions, which is an extension of the solution of Colding-Minicozzi of a conjecture of Yau in Riemannian geometry.