Cette thèse traite de l'étude de méthodes de diagonalisation conjointe de matrices complexes, en vue de la séparation de sources, que ce soit dans le domaine des télécommunications numériques ou de la radioastronomie. Après avoir présenté les motivations qui ont poussé cette étude, nous faisons un bref état de l'art dans le domaine. Le problème de la diagonalisation conjointe, ainsi que celui de la séparation de source sont rappelés, et un lien entre ces deux sujets est établi. Par la suite, plusieurs algorithmes itératifs sont développés. Dans un premier temps, des méthodes utilisant une mise à jour de la matrice de séparation, de type gradient, sont présentées. Elles sont basées sur des approximations judicieuses du critère considéré. Afin d'améliorer la vitesse de convergence, une méthode utilisant un calcul du pas optimal est présentée, et plusieurs variantes de ce calcul, basées sur les approximations faites précédemment, sont développées. Deux autres approches sont ensuite introduites. La première détermine la matrice de séparation de manière analytique, en calculant algébriquement les termes composant la matrice de mise à jour par paire à partir d'un système d'équations linéaire. La deuxième estime récursivement la matrice de mélange, en se basant sur une méthode de moindres carrés alternés. Afin d'améliorer la vitesse de convergence, une recherche de pas d'adaptation linéaire est proposée. Ces méthodes sont alors validées sur un problème de diagonalisation conjointe classique. Puis les algorithmes sont appliqués à la séparation de sources de signaux de télécommunication numérique, en utilisant des statistiques d'ordre deux ou supérieur. Des comparaisons sont également effectuées avec des méthodes standards. La deuxième application concerne l'élimination des interférences terrestres à partir de l'estimation de l'espace associé, afin d'observer au mieux des sources cosmiques, issues de données de station LOFAR. / This thesis deals with the study of joint diagonalization of complex matrices methods for source separation, wether in the field of numerical telecommunications and radioastronomy. After having introduced the motivations that drove this study, we present a brief state-of-the-art in the field. The joint diagonalization and source separation problems are reminded, and a link between these two themes is established. Thereafter, several iterative algorithms are developed. First, methods using a gradient-like update of the separation matrix are introduced. They are based on wise approximations of the considered criterion. In order to improve the convergence speed, a method using a computation of an optimal step size is presented, and variations around this computation, based on the previously introduced approximations are done. Two other approaches are then introduced. The first one analytically determines the separation matrix, by algebraically computing the terms composing the update matrix pairwise from a linear equation system. The second one recursively estimates the mixing matrix, based on an alternating least squares method. In order to enhance the convergence speed, a seek of an enhanced line search algorithm is proposed. These methods are then validated on a classical joint diagonalization problem. Aterwards, these algorithms are applied to the source separation of numerical communication signals, while using second or higher order statistics. Comparisons are also made with well-known methods. The second application relates to elimination of rterrestrial interferences from the estimation of the associated space in order to observe at best cosmic sources from LOFAR station data.

This thesis develops ways of localizing the Ritz values of non-Hermitian matrices. The restarted Arnoldi method with exact shifts, useful for determining a few desired eigenvalues of a matrix, employs Ritz values to refine eigenvalue estimates. In the Hermitian case, using selected Ritz values produces convergence due to interlacing. No generalization of interlacing exists for non-Hermitian matrices, and as a consequence no satisfactory general convergence theory exists. To study Ritz values, I propose the inverse field of values problem for k Ritz values, which asks if a set of k complex numbers can be Ritz values of a matrix. This problem is always solvable for k = 1 for any complex number in the field of values; I provide an improved algorithm for finding a Ritz vector in this case. I show that majorization can be used to characterize, as well as localize, Ritz values. To illustrate the difficulties of characterizing Ritz values, this work provides a complete analysis of the Ritz values of two 3 × 3 matrices: a Jordan block and a normal matrix. By constructing conditions for localizing the Ritz values of a matrix with one simple, normal, sought-after eigenvalue, this work develops sufficient conditions that guarantee convergence of the restarted Arnoldi method with exact shifts. For general matrices, the conditions provide insight into the subspace dimensions that ensure that shifts do not cluster near the wanted eigenvalue. As Ritz values form the basis for many iterative methods for determining eigenvalues and solving linear systems, an understanding of Ritz value behavior for non-Hermitian matrices has the potential to inform a broad range of analysis.

The analysis of similarity of matrices over fields, as well as integral domains which are not fields, is a classical problem in Linear Algebra and has received considerable attention. A related problem is that of simultaneous similarity of matrices. Many interesting algebraic questions that arise in such problems are discussed by Shmuel Friedland[1]. A special case of this problem is that of Simultaneous Unitary Similarity of hermitian matrices, which we describe as follows: Given a collection of m ordered pairs of similar n×n hermitian matrices denoted by {(Hl,Dl)}ml=1, 1. determine if there exists a unitary matrix U such that UHl U∗ = Dl for all l, 2. and in the case where a U exists, find such a U, (where U∗is the transpose conjugate of U ).The problem is easy for m =1. The problem is challenging for m > 1.The problem stated above is the algorithmic version of the problem of classifying hermitian matrices upto unitary similarity. Any problem involving classification of matrices up to similarity is considered to be “wild”[2]. The difficulty in solving the problem of classifying matrices up to unitary similarity is a indicator of, the toughness of problems involving matrices in unitary spaces [3](pg, 44-46 ).Suppose in the statement of the problem we replace the collection {(Hl,Dl)}ml=1, by a collection of m ordered pairs of complex square matrices denoted by {(Al,Bl) ml=1, then we get the Simultaneous Unitary Similarity problem for square matrices. Suppose we consider k ordered pairs of complex rectangular m ×n matrices denoted by {(Yl,Zl)}kl=1, then the Simultaneous Unitary Equivalence problem for rectangular matrices is the problem of finding whether there exists a m×m unitary matrix U and a n×n unitary matrix V such that UYlV ∗= Zl for all l and in the case they exist find them. In this thesis we describe algorithms to solve these problems. The Simultaneous Unitary Similarity problem for square matrices is challenging for even a single pair (m = 1) if the matrices involved i,e A1,B1 are not normal. In an expository article, Shapiro[4]describes the methods available to solve this problem by arriving at a canonical form. That is A1 or B1 is used to arrive at a canonical form and the matrices are unitarily similar if and only if the other matrix also leads to the same canonical form. In this thesis, in the second chapter we propose an iterative algorithm to solve the Simultaneous Unitary Similarity problem for hermitian matrices. In each iteration we either get a step closer to “the simple case” or end up solving the problem. The simple case which we describe in detail in the first chapter corresponds to finding whether there exists a diagonal unitary matrix U such that UHlU∗= Dl for all l. Solving this case involves defining “paths” made up of non-zero entries of Hl (or Dl). We use these paths to define an equivalence relation that partitions L = {1,…n}. Using these paths we associate scalars with each Hl(i,j) and Dl(i,j)denoted by pr(Hl(i,j)) and pr(Dl(i,j)) (pr is used to indicate that these scalars are obtained by considering products of non-zero elements along the paths from i,j to their class representative). Suppose i (I Є L)belongs to the class[d(i)](d(i) Є L) we denote by uisol a modulus one scalar expressed in terms of ud(i) using the path from i to d( i). The free variable ud(i) can be chosen to be any modulus one scalar. Let U sol be a diagonal unitary matrix given by U sol = diag(u1 sol , u2 sol , unsol ). We show that a diagonal U such that U HlU∗ = Dl exists if and only if pr(Hl(i, j)) = pr(Dl(i, j))for all l, i, j and UsolHlUsol∗= Dl. Solving the simple case sets the trend for solving the general case. In the general case in an iteration we are looking for a unitary U such that U = blk −diag(U1,…, Ur) where each Ui is a pi ×p (i, j Є L = {1,… , r}) unitary matrix such that U HlU ∗= Dl. Our aim in each iteration is to get at least a step closer to the simple case. Based on pi we partition the rows and columns of Hl and Dl to obtain pi ×pj sub-matrices denoted by Flij in Hl and Glij in D1. The aim is to diagonalize either Flij∗Flij Flij∗ and a get a step closer to the simple case. If square sub-matrices are multiples of unitary and rectangular sub-matrices are zeros we say that the collection is in Non-reductive-form and in this case we cannot get a step closer to the simple case. In Non- reductive-form just as in the simple case we define a relation on L using paths made up of these non-zero (multiples of unitary) sub-matrices. We have a partition of L. Using these paths we associate with Flij and (G1ij ) matrices denoted by pr(F1ij) and pr(G1ij) respectively where pr(F1ij) and pr(G1ij) are multiples of unitary. If there exist pr(Flij) which are not multiples of identity then we diagonalize these matrices and move a step closer to the simple case and the given collection is said to be in Reduction-form. If not, the collection is in Solution-form. In Solution-form we identify a unitary matrix U sol = blk −diag(U1sol , U2 sol , …, Ur sol )where U isol is a pi ×pi unitary matrix that is expressed in terms of Ud(i) by using the path from i to[d(i)]( i Є [d(i)], d(i) Є L, Ud(i) is free). We show that there exists U such that U HlU∗ = Dl if and only if pr((Flij) = pr(G1ij) and U solHlU sol∗ = Dl. Thus in a maximum of n steps the algorithm solves the Simultaneous Unitary Similarity problem for hermitian matrices. In the second chapter we also relate the Simultaneous Unitary Similarity problem for hermitian matrices to the simultaneous closed system evolution problem for quantum states. In the third chapter we describe algorithms to solve the Unitary Similarity problem for square matrices (single ordered pair) and the Simultaneous Unitary Equivalence problem for rectangular matrices. These problems are related to the Simultaneous Unitary Similarity problem for hermitian matrices. The algorithms described in this chapter are similar in flow to the algorithm described in the second chapter. This shows that it is the fact that we are looking for unitary similarity that makes these forms possible. The hermitian (or normal)nature of the matrices is of secondary importance. Non-reductive-form is the same as in the hermitian case. The definition of the paths changes a little. But once the paths are defined and the set L is partitioned the definitions of Reduction-form and Solution-form are similar to their counterparts in the hermitian case. In the fourth chapter we analyze the worst case complexity of the proposed algorithms. The main computation in all these algorithms is that of diagonalizing normal matrices, partitioning L and calculating the products pr((Flij) = pr(G1ij). Finding the partition of L is like partitioning an undirected graph in the square case and partitioning a bi-graph in the rectangular case. Also, in this chapter we demonstrate the working of the proposed algorithms by running through the steps of the algorithms for three examples. In the fifth and the final chapter we show that finding if a given collection of ordered pairs of normal matrices is Simultaneously Similar is same as finding if the collection is Simultaneously Unitarily Similar. We also discuss why an algorithm to solve the Simultaneous Similarity problem, along the lines of the algorithms we have discussed in this thesis, may not exist. (For equations pl refer the pdf file)

Robust states exist at the interfaces between topologically trivial and nontrivial phases of matter. These boundary states are expression of the nontrivial bulk properties through a connection dubbed the bulk-boundary correspondence. Whether the bulk is topological or not is determined by the value of a topological invariant. This quantity is defined with respect to symmetries and dimensionality of the system, such that it takes only quantized values. For static topological phases that are realized in ground-states of isolated, time-independent systems, the topological invariant is related to the properties of the Hamiltonian operator. In contrast, Floquet topological phases that are realized in open systems with periodical pumping of energy are topologically characterized with a unitary Floquet operator i.e., the time-evolution operator over the entire period. Topological phases of matter can be distinguished by the dimensionality of robust boundary states with respect to the protecting bulk. This dissertation concerns recently discovered higher-order topological phases where the difference between dimensionalities of bulk and boundary states is larger than one. Using analytical and numerical single-particle techniques, we focus on instances where static higher-order topology can be understood with insights from the mature field of Floquet topology. Namely, even though static systems do not admit a Floquet description, we find examples of higher-order systems to which certain unitary operators can be attributed. The understanding of topological characteristics of these systems is therefore conditioned by the knowledge on topological properties of unitary operators, among which the Floquet operator is well-known. The first half of this thesis concerns toy models of static higher-order topological phases that are topologically characterized in terms of unitary operators. We find that a class of these systems called quadrupole topological insulators exhibit a wider range of topological phases than known previously. In the second half of this dissertation, we study reflection matrices of higher-order topological phases and show that they can exhibit the same topological features as Floquet systems. Our findings suggest a new route to experimental realizations of Floquet systems, the one that avoids noise-induced decoherence inevitable in many other experimental setups.